FUNCTIONAL ANALYTIC

METHODS FOR
PARTIAL DIFFERENTIAL EQUATIONS
 PURE AND APPLIED MATHEMATICS

A Program of Monographs, Textbooks, and Lecture Notes

EXECUTIVE EDITORS

Earl J. Taft Zuhair Nashed

Rutgers University University of Delaware
New Brunswick, New Jersey Newark, Delaware

EDITORIAL BOARD

M S. Baouendi Ani/ Nerode

University of California, Cornell University
San Diego
Donald Passman
Jane Cronin University of Wisconsin,
Rutgers University Madison

JackK. Hale Fred S. Roberts

Georgia Institute of Technology Rutgers University

S. Kobayashi Gian-Carlo Rota

University ofCalifornia, Massachusetts Institute of
Berkeley Technology

Marvin Marcus David L. Russell

University of California, Virginia Polytechnic Institute
Santa Barbara and State University

W S. Massey Walter Schempp

Yale University Universitiit Siegen

Mark Teply
University of Wisconsin,
Milwaukee
 MONOGRAPHS AND TEXTBOOKS IN
PURE AND APPLIED MATHEMATICS

1. K. Yano, Integral Formulas in Riemannian Geometry (1970)

2. S. Kobayashi, Hyperbolic Manifolds and Holomorphic Mappings (1970)
3. V. S. Vladimirov, Equations of Mathematical Physics (A. Jeffrey, ed.; A. Littlewood,
trans.) ( 1970)
4. B. N. Pshenichnyi, Necessary Conditions for an Extremum (L. Neustadt, translation
ed.; K. Makowski, trans.) (1971)
5. L. Narici eta/., Functional Analysis and Valuation Theory (1971)
6. S. S. Passman, Infinite Group Rings ( 1971)
7. L. Dornhoff, Group Representation Theory. Part A: Ordinary Representation Theory.
Part B: Modular Representation Theory (1971, 1972)
8. W. Boothby and G. L. Weiss, eds., Symmetric Spaces (1972)
9. Y. Matsushima, Differentiable Manifolds (E. T. Kobayashi, trans.) (1972)
10. L. E. Ward, Jr., Topology (1972)
11. A. Babakhanian, Cohomological Methods in Group Theory (1972)
12. R. Gilmer, Multiplicative Ideal Theory (1972)
13. J. Yeh, Stochastic Processes and the Wiener Integral (1973)
14. J. Barros-Neto, Introduction to the Theory of Distributions ( 1973)
15. R. Larsen, Functional Analysis (1973)
16. K. Yano and S. Ishihara, Tangent and Cotangent Bundles (1973)
17. C. Procesi, Rings with Polynomial Identities ( 1973)
18. R. Hermann, Geometry, Physics, and Systems (1973)
19. N. R. Wallach, Harmonic Analysis on Homogeneous Spaces (1973)
20. J. Dieudonne, Introduction to the Theory of Formal Groups (1973)
21 . I. Vaisman, Cohomology and Differential Forms (1973)
22. B.-Y. Chen, Geometry of Submanifolds (1973)
23. M. Marcus, Finite Dimensional Multilinear Algebra (in two parts) (1973, 1975)
24. R. Larsen, Banach Algebras (1973)
25. R. 0. Kujala and A. L. Vitter, eds., Value Distribution Theory: Part A; Part B: Deficit
and Bezout Estimates by Wilhelm Stoll (1973)
26. K. B. Stolarsky, Algebraic Numbers and Diophantine Approximation (1974)
27. A. R. Magid, The Separable Galois Theory of Commutative Rings (1974)
28. B. R. McDonald, Finite Rings with Identity (1974)
29. J. Satake, Linear Algebra (S. Koh et al., trans.) (1975)
30. J. S. Golan, Localization of Noncommutative Rings (1975)
31. G. Klambauer, Mathematical Analysis (1975)
32. M. K. Agoston, Algebraic Topology (1976)
33. K. R. Goodearl, Ring Theory (1976)
34. L. E. Mansfield, Linear Algebra with Geometric Applications (1976)
35. N.J. Pullman, Matrix Theory and Its Applications (1976)
36. B. R. McDonald, Geometric Algebra Over Local Rings (1976)
37. C. W. Groetsch, Generalized Inverses of Linear Operators (1977)
38. J. E. Kuczkowski and J. L. Gersting, Abstract Algebra (1977)
39. C. 0. Christenson and W. L. Voxman, Aspects of Topology (1977)
40. M. Nagata, Field Theory (1977)
41 . R. L. Long, Algebraic Number Theory (1977)
42. W. F. Pfeffer, Integrals and Measures (1977)
43. R. L. Wheeden and A. Zygmund, Measure and Integral (1977)
44. J. H. Curtiss, Introduction to Functions of a Complex Variable (1978)
45. K. Hrbacek and T. Jech, Introduction to Set Theory (1978)
46. W. S. Massey, Homology and Cohomology Theory (1978)
4 7. M. Marcus, Introduction to Modern Algebra ( 1978)
48. E. C. Young, Vector and Tensor Analysis (1978)
49. S. B. Nadler, Jr., Hyperspaces of Sets (1978)
50. S. K. Segal, Topics in Group Kings (1978)
51. A. C. M. van Rooij, Non-Archimedean Functional Analysis (1978)
52. L. Corwin and R. Szczarba, Calculus in Vector Spaces (1979)
53. C. Sadosky, Interpolation of Operators and Singular Integrals (1979)
 54. J. Cronin, Differential Equations ( 1980)
55. C. W. Groetsch, Elements of Applicable Functional Analysis (1980)
56. /. Vaisman, Foundations of Three-Dimensional Euclidean Geometry (1980)
57. H. I. Freedan, Deterministic Mathematical Models in Population Ecology (1980)
58. S. B. Chae, Lebesgue Integration (1980)
59. C. S. Rees eta!., Theory and Applications of Fourier Analysis (1981)
60. L. Nachbin, Introduction to Functional Analysis (R. M. Aron, trans.) (1981 I
61 . G. Orzech and M. Orzech, Plane Algebraic Curves ( 1981)
62. R. Johnsonbaugh and W. E. Pfaffenberger, Foundations of Mathematical Analysis
(1981)
63. W. L. Voxman and R. H. Goetschel, Advanced Calculus (1981)
64. L. J. Corwin and R. H. Szczarba, Multivariable Calculus (1982)
65. V. I. lstra[escu, Introduction to Linear Operator Theory ( 1981 I
66. R. D. Jarvinen, Finite and Infinite Dimensional Linear Spaces (1981 I
67. J. K. Seem and P. E. Ehrlich, Global Lorentzian Geometry ( 1981 I
68. D. L. Armacost, The Structure of Locally Compact Abelian Groups (1981)
69. J. W. Brewer and M. K. Smith, eds., Emmy Noether: A Tribute (1981 I
70. K. H. Kim, Boolean Matrix Theory and Applications (1982)
71. T. W. Wieting, The Mathematical Theory of Chromatic Plane Ornaments (1982)
72. D. B.Gauld, Differential Topology (1982)
73. R. L. Faber, Foundations of Euclidean and Non-Euclidean Geometry (1983)
74. M. Carmeli, Statistical Theory and Random Matrices (1983)
75. J. H. Carruth et al., The Theory of Topological Semigroups (1983)
76. R. L. Faber, Differential Geometry and Relativity Theory (1983)
77. S. Barnett, Polynomials and Linear Control Systems ( 1983)
78. G. Karpilovsky, Commutative Group Algebras (1983)
79. F. Van Oystaeyen and A. Verschoren, Relative Invariants of Rings (1983)
80. I. Vaisman, A First Course in Differential Geometry (1984)
81 . G. W. Swan, Applications of Optimal Control Theory in Biomedicine ( 1984)
82. T. Petrie and J.D. Randall, Transformation Groups on Manifolds (1984)
83. K. Goebel and S. Reich, Uniform Convexity, Hyperbolic Geometry, and Nonexpansive
Mappings (1984)
84. T. Albu and C. Nastasescu, Relative Finiteness in Module Theory ( 1984)
85. K. Hrbacek and T. Jech, Introduction to Set Theory: Second Edition (1984)
86. F. Van Oystaeyen and A. Verschoren, Relative Invariants of Rings (1984)
87. B. R. McDonald, Linear Algebra Over Commutative Rings ( 1984)
88. M. Namba, Geometry of Projective Algebraic Curves (1984)
89. G. F. Webb, Theory of Nonlinear Age-Dependent Population Dynamics (1985)
90. M. R. Bremner eta/., Tables of Dominant Weight Multiplicities for Representations of
Simple Lie Algebras (1985)
91. A. £ Fekete, Real Linear Algebra (1985)
92. S. B. Chae, Holomorphy and Calculus in Normed Spaces (1985)
93. A. J. Jerri, Introduction to Integral Equations with Applications (1985)
94. G. Karpilovsky, Projective Representations of Finite Groups (1985)
95. L. Narici and E. Beckenstein, Topological Vector Spaces (1985)
96. J. Weeks, The Shape of Space (1985)
97. P.R. Gribik and K. 0. Kortanek, Extremal Methods of Operations Research (1985)
98. J.-A. Chao and W. A. Woyczynski, eds., Probability Theory and Harmonic Analysis
(1986)
99. G. D. Crown eta/., Abstract Algebra (1986)
100. J. H. Carruth eta/., The Theory of Topological Semigroups, Volume 2 ( 1986)
101 . R. S. Doran and V. A. Belfi, Characterizations of C *-Algebras ( 1986)
102. M. W. Jeter, Mathematical Programming (1986)
103. M. Altman, A Unified Theory of Nonlinear Operator and Evolution Equations with
Applications ( 1986)
104. A. Verschoren, Relative Invariants of Sheaves ( 1987)
105. R. A. Usmani, Applied Linear Algebra (1987)
106. P. Blass. and J. Lang, Zariski Surfaces and Differential Equations in Characteristic p
> 0 (1987)
107. J. A. Reneke eta/., Structured Hereditary Systems ( 1987)
108. H. Busemann and B. B. Phadke, Spaces with Distinguished Geodesics (1987)
109. R. Harte, lnvertibility and Singularity for Bounded Linear Operators (1988)
11 0. G. S. Ladde et at., Oscillation Theory of Differential Equations with Deviating
Arguments (19871
111. L. Dudkin et at., Iterative Aggregation Theory ( 1987)
112. T. Okubo, Differential Geometry (1987)
113. D. L. Stancl and M. L. Stancl, Real Analysis with Point-Set Topology (1987)
114. T. C. Gard, Introduction to Stochastic Differential Equations ( 1988)
115. S. S. Abhyankar, Enumerative Combinatorics of Young Tableaux (1988)
116. H. Strade and R. Farnsteiner, Modular Lie Algebras and Their Representations (1988)
117. J. A. Huckaba, Commutative Rings with Zero Divisors (1988)
118. W. D. Wallis, Combinatorial Designs (1988)
119. W. Wi~law, Topological Fields (1988)
120. G. Karpilovsky, Field Theory ( 1988)
121 . S. Caenepeel and F. Van Oystaeyen, Brauer Groups and the Cohomology of Graded
Rings (19891
122. W. Kozlowski, Modular Function Spaces (1988)
123. E. Lowen-Colebunders, Function Classes of Cauchy Continuous Maps (1989)
124. M. Pavel, Fundamentals of Pattern Recognition (1989)
125. V. Lakshmikantham et at., Stability Analysis of Nonlinear Systems (1989)
126. R. Sivaramakrishnan, The Classical Theory of Arithmetic Functions (1989)
127. N. A. Watson, Parabolic Equations on an Infinite Strip (1989)
128. K. J. Hastings, Introduction to the Mathematics of Operations Research (1989)
129. B. Fine, Algebraic Theory of the Bianchi Groups (1989)
130. D. N. Dikranjan et at., Topological Groups (1989)
131 . J. C. Morgan II, Point Set Theory ( 1990)
132. P. Biter and A. Witkowski, Problems in Mathematical Analysis ( 1990)
133. H. J. Sussmann, Nonlinear Controllability and Optimal Control (1990)
134. J.-P. Florens eta/., Elements of Bayesian Statistics (19901
135. N. Shell, Topological Fields and Near Valuations (1990)
136. B. F. Doolin and C. F. Martin, Introduction to Differential Geometry for Engineers
(1990)
137. S. S. Holland, Jr., Applied Analysis by the Hilbert Space Method ( 1990)
138. J. Okniriski, Semigroup Algebras (1990)
139. K. Zhu, Operator Theory in Function Spaces (1990)
140. G. B. Price, An Introduction to Multicomplex Spaces and Functions ( 1991)
141 . R. B. Darst, Introduction to Linear Programming ( 1991)
142. P. L. Sachdev, Nonlinear Ordinary Differential Equations and Their Applications
(1991)
143. T. Husain, Orthogonal Schauder Bases (1 991)
144. J. Foran, Fundamentals of Real Analysis (1 991)
145. W. C. Brown, Matrices and Vector Spaces ( 19911
146. M. M. Rao and Z. D. Ren, Theory of Orlicz Spaces (1991)
147. J. S. Golan and T. Head, Modules and the Structures of Rings ( 19911
148. C. Small, Arithmetic of Finite Fields (1991)
149. K. Yang, Complex Algebraic Geometry (1991)
1 50. D. G. Hoffman et at., Coding Theory ( 19911
151. M. 0. Gonzalez, Classical Complex Analysis (1992)
152. M. 0. Gonzalez, Complex Analysis (19921
153. L. W. Baggett, Functional Analysis (1992)
154. M. Sniedovich, Dynamic Programming (1992)
155. R. P. Agarwal, Difference Equations and Inequalities (19921
156. C. Brezinski, Biorthogonality and Its Applications to Numerical Analysis ( 1992)
157. C. Swartz, An Introduction to Functional Analysis (19921
158. S. B. Nadler, Jr., Continuum Theory (19921
159. M.A. AI-Gwaiz, Theory of Distributions (19921
160. E. Perry, Geometry: Axiomatic Developments with Problem Solving (19921
161. E. Castillo and M. R. Ruiz-Cobo, Functional Equations and Modelling in Science and
Engineering (19921
162. A. J. Jerri, Integral and Discrete Transforms with Applications and Error Analysis
(1992)
163. A. Charlier et at., Tensors and the Clifford Algebra (19921
164. P. Biter and T. Nadzieja, Problems and Examples in Differential Equations (1992)
165. E. Hansen, Global Optimization Using Interval Analysis (1992)
166. S. Guerre-Delabriere, Classical Sequences in Banach Spaces (1992)
167. Y. C. Wong, Introductory Theory of Topological Vector Spaces (1992)
168. S. H. Kulkarni and B. V. Limaye, Real Function Algebras (1992)
169. W. C. Brown, Matrices Over Commutative Rings (1993)
170. J. Loustau and M. D1'llon, Linear Geometry with Computer Graphics (1993)
171. W. V. Petryshyn, Approximation-Solvability of Nonlinear Functional and Differential
Equations ( 1993)
172. E. C. Young, Vector and Tensor Analysis: Second Edition (1993)
173. T. A. Bick, Elementary Boundary Value Problems ( 1993)
174. M. Pavel, Fundamentals of Pattern Recognition: Second Edition (1993)
175. S. A. Albeverio et al., Noncommutative Distributions (1993)
176. W. Fulks, Complex Variables (1993)
177. M. M. Rao, Conditional Measures and Applications (1993)
178. A. Janicki and A. Weron, Simulation and Chaotic Behavior of a.-Stable Stochastic
Processes (1994)
179. P. Neittaanmiiki and D. Tiba, Optimal Control of Nonlinear Parabolic Systems (1994)
180. J. Cronin, Differential Equations: Introduction and Qualitative Theory, Second Edition
(1994)
181. S. Heikkila and V. Lakshmikantham, Monotone Iterative Techniques for Discontinu-
ous Nonlinear Differential Equations (1994)
182. X. Mao, Exponential Stability of Stochastic Differential Equations (1994)
183. B. S. Thomson, Symmetric Properties of Real Functions (1994)
184. J. E. Rubio, Optimization and Nonstandard Analysis (1994)
185. J. L. Bueso eta/., Compatibility, Stability, and Sheaves (1995)
186. A. N. Michel and K. Wang, Qualitative Theory of Dynamical Systems (1995)
187. M. R. Darnel, Theory of Lattice-Ordered Groups (1995)
188. Z. Naniewicz and P. D. Panagiotopoulos, Mathematical Theory of Hemivariational
Inequalities and Applications (1995)
189. L. J. Corwin and R. H. Szczarba, Calculus in Vector Spaces: Second Edition (1995)
190. L. H. Erbe eta/., Oscillation Theory for Functional Differential Equations (1995)
191. S. Agaian eta/., Binary Polynomial Transforms and Nonlinear Digital Filters (1995)
192. M. I. G1'l', Norm Estimations for Operation-Valued Functions and Applications (1995)
193. P. A. Grillet, Semigroups: An Introduction to the Structure Theory (1995)
194. S. Kichenassamy, Nonlinear Wave Equations (1996)
195. V. F. Krotov, Global Methods in Optimal Control Theory (1996)
196. K. I. Beidar eta/., Rings with Generalized Identities (1996)
197. V.I. Arnautov eta/., Introduction to the Theory of Topological Rings and Modules
(1996)
198. G. Sierksma, Linear and Integer Programming (1996)
199. R. Lasser, Introduction to Fourier Series ( 1996)
200. V. Sima, Algorithms for Linear-Quadratic Optimization (1996)
201. D. Redmond, Number Theory (1996)
202. J. K. Beem eta!., Global Lorentzian Geometry: Second Edition (1996)
203. M. Fontana eta/., PrOfer Domains ( 1997)
204. H. Tanabe, Functional Analytic Methods for Partial Differential Equations (1997)

Additional Volumes in Preparation

FUNCTIONAL ANALYTIC
METHODS FOR
PARTIAL DIFFERENTIAL EQUATIONS

Hiroki Tanabe
Otemon Gakuin University
Osaka, Japan

n
MARCEL

DEKKER
MARCEL DEKKER, INc. NEW YoRK • BAsEL
Library of Congress Cataloging-in-Publication Data

Tanabe, Hiroki.
Functional analytic methods for partial differential equations I
Hiroki Tanabe.
p. em. -(Monographs and textbooks in pure and applied
mathematics ; 204)
Includes bibliographical references (p. - ) and index.
ISBN 0-8247-9774-4 (alk. paper)
I. Functional analysis. 2. Differential equations, Partial.
I. Title. II. Series.
QA32l.T36 1996
515'.353-dc20
96-31585
CIP

The publisher offers discounts on this book when ordered in bulk quantities. For
more information, write to Special Sales/Professional Marketing at the address
below.

This book is printed on acid-free paper.

Copyright © 1997 by MARCEL DEKKER, INC. All Rights Reserved.

Neither this book nor any part may be reproduced or transmitted in any form or
by any means, electronic or mechanical, including photocopying, microftlming,
and recording, or by any information storage and retrieval system, without
permission in writing from the publisher.

MARCEL DEKKER, INC.

270 Madison Avenue, New York, New York 10016

Current printing (last digit):

10987654
Preface

This book is devoted to the functional analytic method for partial differ-
ential equations, and includes both classical theory and up-to-date results.
The first main object is to present a self-contained proof of Agmon-Douglis-
Nirenberg's LP estimates for elliptic boundary value problems. For that
purpose it is required to discuss the theory of singular integrals of A. P.
Calderon and A. Zygmund. Although this theory is classical, it still main-
tains its importance, and its extension is now being studied vigorously. The
Hilbert transform, which is the case of the singular integral of a single inde-
pendent variable, is also described completely. The Hilbert transform plays
an important role in the study of maximal regularity results. It is one of
the objects of this book to describe such concepts, which generally are not
found in other books on our subject. The second half of the book is de-
voted to evolution equations in Banach spaces. Recent results on equations
of parabolic type and of hyperbolic type, and retarded functional differen-
tial equations together with control theory related to them, are described.
This book is concerned only with linear equations; however, some notes on
nonlinear equations as well as degenerate equations, which are also out of
our scope, are given in the bibliographical remarks.
Chapter 1 includes preliminary materials which will be used in subsequent
chapters.
Chapter 2 is devoted to the most elementary part of the theory of singular
integrals of Calderon and Zygmund and the basic theory of the Hilbert
transform.
Chapter 3 is devoted to the theory of function spaces, such as interpolation
inequalities, the Gagliardo-Nirenberg inequality and Sobolev's imbedding
theorems.
Chapter 4 include Agmon-Douglis-Nirenberg's LP estimates for solutions
of elliptic boundary value problems with a slightly simplified proof of the
original one.
In Chapter 5 adjoint boundary value problems are described following
M. Schechter, and their solvability is discussed following S. Agmon. The

iii
iv

estimates of the kernels of the resolvents of elliptic operators and the semi-
groups generated by them are established by a functional analytic method,
and the result is applied to the problem in £ 1 space.
In Chapter 6 parabolic evolution equations are described following P.
Acquistapace and B. Terreni. Some remarks on A. Yagi's results are also
given.
Chapter 7 is devoted to the study of hyperbolic evolution equations orig-
inated by T. Kato and developed by himself, K. Kobayashi, A. Yagi, N.
Okazawa, and A. Unai.
Chapter 8 is devoted to retarded functional differential equations in Hilbert
spaces. First the solvability is described following G. Di Blasia, K. Kunisch
and E. Sinestrari. Next, the results on control problems of C. Bernier, M. C.
Delfour, and A. Manitius for equations in finite dimensional spaces and of S.
Nakagiri for equations in reflexive Banach spaces are extended to equations
in Hilbert spaces with delay terms that are as unbounded as the main term.
Prerequisites are Lebesgue integration and the elementary concepts of
Banach and Hilbert spaces.
I express my profound gratitude to Professor Angelo Favini of the Univer-
sity of Bologna and Professor Atsushi Yagi of Osaka University who urged
me to write this book and constantly encouraged me during its prepara-
tion. I also thank Professor Yasushi Shizuta of Nara Women's University,
who motivated me to greatly improve the contents of Chapter 3, Profes-
sor Noboru Okazawa of Science University of Tokyo, who kindly offered me
his unpublished paper, and Professor Shin-ichi Nakagiri of Kobe University,
who read the manuscript of Chapter 8 carefully and gave me kind advice.
Thanks are also due to Professor Yoshihiko Yamamoto of Osaka Univer-
sity for his cordial instruction on how to operate the computer to prepare
the manuscript by LATEX.
The contents of this book considerably overlap with my book Functional
Analysis, II, published in Japanese by Jikkyo Shuppan Publishing Com-
pany, Tokyo, in 1981. I express my deep gratitude to Jikkyo Shuppan for
permitting me to reproduce the relevant parts of the book.
I also would like to express gratitude to the staff of Marcel Dekker, Inc.,
in particular to Ms Maria Allegra for their assistance.

Hiraki Tanabe
Contents

1 Preliminaries 1
1.1 Preliminaries on Measure and Integration 1
1.2 Preliminaries from Functional Analysis · 3
1.3 Semigroups · · · · · 3
1.4 Interpolation Spaces · · · · · · · · · · · 7

2 Singular Integrals 13
2.1 Singular Integrals of A. P. Calderon and A. Zygmund 13
2.2 J. Marcinkiewicz's Interpolation Theorem 15
2.3 Case of Bounded Kernels · 18
2.4 Case of Continuous Kernels 21
2.5 Hilbert 'fransform · · · · · 29
2.6 Equimeasurable Functions · 31
2. 7 Hilbert 'fransform (Continued) 37
2.8 Case of Odd Kernels · 40
2. 9 Riesz Kernels · · · · · 43
2.10 Case of Even Kernels · 47
2.11 General Case · · · · · 64
2.12 Derivatives of Homogeneous Functions 65

3 Sobolev Spaces 67
3.1 Sobolev Spaces · · · · · · · · 67
3.2 Interpolation Inequalities (1) 68
3.3 Interpolation Inequalities (2) 72
3.4 Interpolation Inequalities (3) 75
3.5 Sobolev Spaces in General Domains 99
3.6 Uniformly Regular Open Sets · · · · 102
3. 7 Sobolev Spaces in Uniformly Regular Open Sets · 105
3.8 Embedding Theorems · · · · · · · · · · · · · 110

v
vi CONTENTS

3.9 Additional Topics · · · · · · · · · · · · · · · · · · · · · · · · · 118

4 Elliptic Boundary Value Problems 121

4.1 Fundamental Solutions of Elliptic Operators · 121
4.2 Assumptions and Main Result· · · · · · · · 130
4.3 Preliminaries from the Theory of Ordinary
Differential Equations · · · · 131
4.4 Case of Constant Coefficients · · 135
4.5 Poisson Kernels · · · · · · · · · · 137
4.6 Preliminaries on Integral Kernels 145
4. 7 Nonhomogeneous Boundary Conditions 155
4.8 Problems in Uniformly Regular Open Sets · 164

5 Elliptic Boundary Value Problems(Continued) 169

5.1 Adjoint Boundary Conditions · · · · · · · · · · · 169
5.2 Existence of Solutions · · · · · · · · · · · · · · · 180
5.3 Estimates of the Kernels of exp( -tAp) and (Ap- .A)- 1 · 190
5.4 Boundary Value Problems in £ 1 • • • • • • • • • • 210
5.5 Boundary Value Problems in Spaces of Continuous
Functions · · · · · · · · · · · · · · · · · · 215
5.6 An Example of a Higher Order Operator· · · · 218

6 Parabolic Evolution Equations 221

6.1 Equations with Coefficients Differentiable in t · 221
6.2 Preliminaries(!) · · 224
6.3 Classical Solutions 234
6.4 Preliminaries(2) · · 238
6.5 Strict Solutions · · 241
6.6 Maximal Regularity 245
6. 7 An Example · · · · · 250
6.8 Equations with Coefficients Holder Continuous in t · 256
6.9 Preliminary Lemmas and Remarks · · · · · · · · · · 258
6.10 Representation of Solutions by Fundamental Solution · 262
6.11 Approximate Equations 270
6.12 Existence of Solutions 275
6.13 Application · · · · · · · 278

7 Hyperbolic Evolution Equations 285

7.1 Admissible Subspaces · · · · · · · · · · · · 285
7.2 Stable Families of Infinitesimal Generators · 289
7.3 Construction of Evolution Operator (1) 293
7.4 Uniqueness of Regular Solutions · · · · 296
7.5 Construction of Evolution Operator (2) 297
CONTENTS vii

7.6 Existence of Regular Solutions 307

7. 7 Equations in Hilbert Spaces 309
7.8 Remark on Applications · · · · 320

8 Retarded Functional Differential Equations 323

8.1 Maximal Regularity Result 323
8.2 Assumptions and Notations · 328
8.3 Solvability and Regularity 329
8.4 Solution Semigroup · · · · · 334
8.5 Mild Solutions · · · · · · · 340
8.6 Regularly Accretive Operators 342
8. 7 Semigroup Approach Revised and Control Problem · 346
8.8 Structural Operator · · · · · · · 351
8.9 Controllability and Observability 355
8.10 Characterization of Ker(). - A)l · 358
8.11 A Special Case · · · · · · · · 361
8.12 Eigenmanifold Decomposition · 364
8.13 Second Structural Operator · · 371
8.14 A Special Case (Continued) · · 379
8.15 F-Completeness of Generalized Eigenfunctions 383
8.16 Example and Remarks · · · · · · · · · · · · · · 387

Bibliographical Remarks 391

Bibliography 397

List of Symbols 411

Index 413
FUNCTIONAL ANALYTIC
METHODS FOR
PARTIAL DIFFERENTIAL EQUATIONS
Chapter 1

Preliminaries

1.1 Preliminaries on Measure and

Integration
In this chapter we collect some preliminary results from the theory of mea-
sure and integration.
Let E be a Lebesgue measurable subset of the v-dimensional Euclidean
space Rv. The Lebesgue measure of E is denoted by lEI. The supremum of
IEI/IQI over all cubes Q containing E is called the parameter of regularity
of E and is denoted by r(E).
A sequence {En} of measurable sets is called a regular sequence if there
exists a positive number a: such that r(En) ~ a: for all n = 1, 2, ....
For a nonempty open subset n of Rv the notation Lloc(n) stands for the
set of all functions which are Lebesgue integrable on each compact subset
ofn.
Theorem 1. 1 For a function f E Lfoc(Rv) there exists a null set N such
that for any x ~ N and for any regular sequence {en} tending to x

lim - 1
n---+oo 1en 1
1
en
f(y)dy = J(x).

Especially for any x ~ N

lim v; 1 v
p--+O
r
vP Jlx-yi<P
f(y)dy = f(x),

where Vv is the volume of the unit ball of Rv.

More strongly the following theorem holds.

1
2 CHAPTER 1. PRELIMINARIES

Theorem 1. 2 For a function f E Lloc(Rv)

lim
p->O
v. 1 v r
vP J1x-yj <p
lf(y)- f(x)ldy = 0 (1.1)

at almost every point x in Rv.

Definition 1. 1 A point at which (1.1) holds is called a Lebesgue point of

f.
Lemma 1.1 (Young's inequality) If a 2 0, b 2 0,1 < p < oo, q = pj(p-1)

The equality holds if and only if aP = bq.

Proof. The assertion is obtained by comparing the sum of the area of the
figure surrounded by the three curves x = a, y = 0, y = xP- 1 and that of
the figure surrounded by x = 0, y = b, x = yq- 1 with ab.

Lemma 1. 2 (Hausdorff-Young's inequality) For f E L 1 (Rv),g E £P(Rv),

1 :::; p :::; oo, we have

where f * g is the convolution off and g:

(! * g)(x) = r f(x- y)g(y)dy,
lw
and II llv is the norm of LP(Rv).

Proof. Integrating both sides of

I(!* g)(x)IP = lln f(x- y)g(y)dylp

: :; (ln lf(x- Y)l 1/p'lf(x- Y)I 11 Pig(y)ldy) P

: :; (Ln lf(x- Y)ldy) Ln lf(x- Y)llg(y)IPdy

pjp'

f lf(x- Y)llg(y)IPdy
= llfllf- 1 }Rn
over Rn we readily obtain the desired result.
1.3. SEMIGROUPS 3

1.2 Preliminaries from Functional Analysis

Let T be a linear operator with domain in a Banach space X and range in
another Banach space Y. The domain and the range of T are denoted by
D(T) and R(T) respectively. For a linear closed operator T from a Banach
space X into itself the spectrum and the resolvent set ofT are denoted by
a(T) and p(T) respectively, and the point spectrum ofT by ap(T). The
adjoint space of X is denoted by X*. The adjoint operator of a densely
defined linear operator T is denoted by T*.
Theorem 1. 3 Let T be a densely defined closed linear operator from a
Banach space X to another Banach space Y.
(i) A necessary and sufficient condition in order that R(T) = Y is that T*
has a continuous inverse.
(ii) A necessary and sufficient condition in order that R(T*) =X* is that
T has a continuous inverse.
For the proof see K. Yosida [166: Corollary 1 of VII. 5].
The following lemma is known as Dini's theorem.
Lemma 1. 3 If f(t) is a strongly continuous function defined in the inter-
val [a, b] taking values in a Banach space X and has a strongly continuous
right derivative D+ f(t) in (a, b), then f is strongly continously differentiable
in (a, b) and df(t)jdt = D+ f(t).
We denote the set of all bounded linear operators from a Banach space X
into another Banach spaceY by £(X, Y). We make this space a Banach
spaces with the operator norm unless otherwise stated. We write £(X)
instead of £(X, X). We denote by

x = w- lim Xn
n--+<XJ

that a sequence {xn} of a Banach space converges weakly to x.

1.3 Semigroups
Let X be a Banach space with norm II · II·
Definition 1. 2 A family of bounded linear operators {T(t); t > 0} is
called a semigroup in X if the following conditions are satisfied:
(i) T(t + s) = T(t)T(s) for t > 0, s > 0.
(ii) T(t) is strongly continuous in t > 0.
If moreover the following condition is satisfied, it is called a Co -semigroup.
(iii) limt--+o T(t) =I in the strong operator topology.
4 CHAPTER 1. PRELIMINARIES

Definition 1. 3 Let {T(t); t 2:: 0} be a C0 -semigroup in X. The operator

A defined by
. T(t)x-x
A X = l Im --'--"----- (1.2)
t-+0 t
for x E D(A) = {x E X; the limit (1.2) exists} is called the infinitesimal
generator of the Co-semigroup {T(t); t 2:: 0}.

Theorem 1. 4 Let {T(t); t 2:: 0} ba a C0 -semigroup. Then there exists a

constant M 2:: 1 and a real number {3 such that

IIT(t)lls;Mef3t, t2::0. (1.3)

The infinitesimal generator A of {T(t); t 2:: 0} is a densely defined closed

linear operator satisfying the following conditions:
(i) p(A) :) { -\; Re,\ > {3};
(ii) 11(,\- A)-nil :::; M(Re,\- j3)-n, Re,\ > {3, n = 1, 2, ....
Conversely if A is a densely defined closed linear operator satisfying (i)
and (ii), then there exists a C 0 -semigroup {T(t); t 2:: 0} with infinitesimal
generator A. The semigroup {T(t); t 2:: 0} is uniquely determined by A and
satisfies (1.3).
Moreover we have

(1.4)

for Re,\ > {3.

The set of all infinitesimal generators which generate C0 -semigroups in X

satisfying (1.3) is denoted by G(X, M, {3). The semigroup with infinitesimal
generator A is denoted by etA or exp(tA).
Theorem 1. 5 Let {T(t); t 2:: 0} be a Ca-semigroup in a Banach space X
with infinitesimal generator A. Then

T(t) = lim (1- !A)-n Hille's representation (1.5)

n-+oo n
T(t) = n-+oo
lim exp (tAn) Yosida's representation, (1.6)

uniformly in any bounded subinterval of [0, 00) in the strong operator topol-
ogy of X, where An = A(1 - n- 1 A)- 1 .

Theorem 1. 6 Let A be a closed linear operator with not necessarily dense

domain. Suppose that there exist an angle (}0 E ( 1r /2, 1r] and positive con-
stants cl, c2 such that
1.3. SEMIGROUPS 5

(i) p(A) ::J ~={A; -Oo < arg,\ < Oo, 1,\1 > C1},
(ii) 11(,\- A)- 11 ::; C2/l,\l, ,\ E ~.
Let r be a smooth path running from ooe-i0 to ooei0 in ~~ where rr /2 < () <
Oo. Set
(1.7)

Then T(t) is a semigroup. T(t) is extended as an analytic function oft in

the sector {t; Iarg tl < Oo - rr /2} and the inequality

II! I
T(t) = IIAT(t) II ::; c~wt' t >0 (1.8)

holds for some constants C and w. If A is densely defined, then {T(t)} is a

C 0 -semigroup, and A is the infinitesimal generator of {T(t)}.
Definition 1. 4 The sernigroup {T(t)} of the above theorem is called an
analytic semigroup.
Remark 1. 1 Let A be a closed linear operator in a Banach space X.
Suppose that (j3,oo) C p(A) and 11(,\- j3)(,\- A)- 11 is bounded for,\ E
(j3, oo ). Furthermore if X is reflexive, then D(A) is dense as is shown below.
Let x be an arbitrary element of X. Then ,\(,\-A)- 1 x is bounded as,\~ oo.
Therefore, there exists a subsequence An ~ oo such that An(An - A)- 1 x
converges weakly to some element y EX. Since
A(,\n - A)- 1 x = An(An - A)- 1 x- X~ y- X

weakly, (,\n- A)- 1 x ~ 0 and A is closed, it follows that y- x = 0. Since

,\n(An- A)- 1 x E D(A), we conclude x = y E D(A). Therefore an analytic
sernigroup in a reflexive Banach space is a Co-semigroup.
The sernigroup (1.7) is also denoted by etA or exp(tA). More generally we
have for n = 0, 1, 2, ...

!!:_T(t) = -1 { ,\ne-'t(,\- A)- 1 d,\ (1.9)

dtn 2rri Jr '
ll:t:T(t)ll = IIAnetAII::; c~:wt. (1.10)

The inequality (1.10) is proved as follows. If we choose a sufficiently large

positive number k if necessary, the conditions (i),(ii) of Theorem 1.6 hold for
A 1 =A- k with C1 = 0, and T(t) = etkT1 (t), where T1(t) is the semigroup
generated by A 1 • It suffices to show that (1.10) holds for T 1 (t) with w = 0:

(1.11)
6 CHAPTER 1. PRELIMINARIES

The case n = 0 is established by choosing r as

{.:\; >. = re-i0 , r ~ 1/t} U {.:\; 1>.1 = 1/t, Iarg >.I :::; B} U {.:\; >. = rei0 , r ~ 1/t},

where 1r /2 < e < 00 • The case n ~ 1 is proved by choosing

r = {.:\; >. = re-i0 , -oo < r:::; 0} U {.:\; >. = rei 11 , 0:::; r < oo}.
Theorem 1. 7 Let etA be an analytic semigroup. Then etAx ---+ x as t ---+ 0
if and only if x E D(A).

Proof. Since etA is uniformly bounded for t bounded, it suffices to show

that for x E D(A) etAx---+ x as t---+ 0. If x E D(A)

etAX = ~
27r2 Jr
r eAt(>.- A)- 1 A- 1Axd>.

=~ {e-At {(>.-A)- 1 +A- 1 }Axd>.

l l
21l"2 Jr >.

=-.
1 e-At
-(>.-A)- 1Axd>.+-.
1 e-At
-xd>.
21r2 r >. 21r2 r >.
1 r e-At
= 27ri Jr T(>.- A)-1 Axd>. + x.

Hence choosing r as in the proof of (1.11) for n = 0 we get

if t is sufficiently small. Hence the assertion follows.

Lemma 1. 4 If A genemtes a C 0 -semigroup or an analytic semigroup in

X, and 0 E p(A), then for any x EX and t > 0 we have

1t esAxds =(etA- I)A-1x.

Proof. For 0 < E < t we have

If etA is a C0 -semigroup, then the right hand side goes to (etA - I) A - 1x as

t ---+ 0. If etA is an analytic semigroup, the same holds in view of Theorem
1.7.
1.4. INTERPOLATION SPACES 7

Lemma 1. 5 Suppose that A generates an analytic semigroup.

etAx-x --
(i) For x E D(A) lim =Ax if and only if AxE D(A).
t->0 t
(ii) For x E D(A) we have lim tAetAx = 0.
t->0

Proof. (i) By the proof of Lemma 1.4 we have

etAx- x 11t 8
D(A) 3 = - e AAxds.
t t 0
Hence the assertion follows from Theorem 1. 7.
(ii) If x E D(A), then tAetAx = tetA Ax ---+ 0 as t ---+ 0. The same fact
remains valid for x E D(A), since in view of (1.8) lltAetAII is bounded fort
bounded.
For the details for semigroups we refer to E. Hille and R. S. Phillips [74].

1.4 Interpolation Spaces

Let X andY be a couple of Banach spaces with norm II · llx and II · IIY
respectively. We assume that both X and Y are continuously imbedded
in some locally convex linear topological space E. Such a pair is called an
interpolation pair.
For 1 ::; p::; oo a(x) is the set of all functions with values in X which
are defined and strongly measurable in (0, oo) such that

lluiiL~(X) = { (}r= llu(t)ll~tdt) 11

0
p
1::;p<oo (1.12)
ess SUPtE(O,oo) llu(t) llx p=oo
are finite. As is easily seen a(x) is a Banach space with norm (1.12).
For 0 < () < 1, 1 ::; p::; oo let V((),p; X, Y) be the set of all functions u
defined in (0, oo) and taking values in E such that
t 8 u E L~(X) and t 8 u' E L~(Y),
where u' is the derivative in the sense of distributions. V((), p; X, Y) is a
Banach space with norm
p p ) 1/p
+ lit
(J (J I
llullv(8,p;X,Y) = (
lit uiiL~(X) u IIL~(Y) ·
It is easily seen that if u E V((), p; X, Y) then the trace u(O) of u at t = 0 is
well defined as an element of X + Y. We denote the set of all these traces
by (X, Y)8,p:
(X, Y)8,p = { u(O); u E V((), p; X, Y)}.
8 CHAPTER 1. PRELIMINARIES

(X, Y)o,p is a Banach space with norm

llallcx,Y)s,p = inf{llullvco,p;X,Y); u(O) =a, u E V(O,p; X, Y)}.
The space (X, Y)o,p, 0 < 0 < 1,::::; p::::; oo is called a real interpolation space
between X and Y. For the details of interpolation spaces we refer to P. L.
Butzer and H. Berens [25], J. L. Lions and E. Magenes [99], J. L. Lions and
J. Peetre [100] and H. Ttiebel [154].
Let A be a densely defined closed linear operator in a Banach space X.
A is called a positive operator if ( -oo, OJ c p(A) and there exists a positive
constant C such that for each t > 0
ll(t+A)- 111 :::;Cj(t+1).
We make D(A) a Banach space with the graph norm of A:
lluiiD(A) = IIAullx + llullx·
The proof of the following theorems are found in [25: Chapter 3] and [154:
1.14].
Theorem 1. 8 Let A be a positive operator in a Banach space X and x
an element of X. Then x E (D(A), X)o,p if and only if the functions t f--->
t 1 - 9 A(A + t)- 1 x belongs to a(x). The norm of (D(A), X)o,v is equivalent
to

1:::;p<oo
p= 00

If A generates a Co-semigroup in X, then for some real number k, k- A is a

positive operator. For the sake of simplicity of notations we assume k = 0.
Theorem 1. 9 Let A generate a C0 -semigroup in X. Then an element x
of X belongs to (D(A), X)o,p if and only if the function t f---> t 9 - 1(etAx- x)
belongs to a(X). The norm of (D(A), X)o,p is equivalent to

{ (Jor= 11t0-1(etAx- x)II~Tdt) 11 p

+ llxllx
ess suptE(O,oo)llt9 - 1(etAx- x)lix + llxllx p= 00

Suppose that A generates an analytic C0 -semigroup in X. Then there exist

positive constants C1, C2 and an angle Oo E ( 1r/2, n] such that
p(A) :J E ={A; -Oo < arg ,\ < Oo, 1,\1 > C1},
11(,\- A)- 111 : : ; C2/l,\l, ,\ E E.
For the sake of simplicity we assume that C1 = 0 and -A is a positive
operator.
1.4. INTERPOLATION SPACES 9

Theorem 1. 10 If A generates an analytic C 0 -semigroup in X, then an

element x of X belongs to (D(A), X)o,p if and only if the function t ~---+
t 0 AetAx belongs to Lf(X). The norm of (D(A), X)o,p is equivalent to

{ (1 llt0 AetAxll)c ~t) l/p + llxllx

00
1:s;p<oo

ess suptE(O,oo)llt0 AetAxllx + llxllx p= 00

Furthermore if x E (D(A), X)o,p, then for any E E (0, Oo)

p= 00

We denote the norm of C(X, Y) by II · llccx,Y)·

Theorem 1. 11 Let (X1, Y1 ) and (X2 , }'2) be two interpolation pairs. If
T E £(X1, X2) n C(Yt, Y2), then

T E C((Xt, Yt)o,p, (X2, Y2)o,p), 0 < 0 < 1, 1::; p::; oo,

and

Remark 1. 2 In the interpolation theory we often have an inequality of

the type (1.13) with the right hand side multiplied by some positive constant
co depending on 0. In the present case we have (1.13) with co = 1 which
can be shown if we use

This equality follows from the fact that if u(O) =a, u E V(O, p; X, Y), then
u>.(O) =a, U>. E V(O,p; X, Y), A> 0, where U>.(t) = u(.At), and Lemma 1.1.

If 1 < p < oo, then 0 < 1/p < 1 and

V(1jp, p; X, Y) = {u; u E £P(O, oo; X), u' E £P(O, oo; Y)}, (1.14)

where LP(O, oo; X) is the set of all strongly measurable functions in (0, oo)
with values in X such that

lluiiLP(O,oo;X) = (Jor= llu(t)ll)cdt) 1/p

< oo. (1.15)
10 CHAPTER 1. PRELIMINARIES

As is easily seen LP(O, oo; X) is a Banach space with norm (1.15). The norm
of V(1jp,p; X, Y) is

p I p )1/p
llullvc1/p,p;X,Y) = ( lluiiLP(O,oo;X) + llu IILP(O,oo;Y) ·

If u E V(1/p,p; X, Y), then for any t ~ 0 u(t) E (X, Yhfv,P' and

llu(t)llcx,Yh;p.p::; llu(t + ·)llv(1/p,p;X,Y)::; llullvc1/p,p;X,Y)· (1.16)

Furthermore

llu(t)- u(s)llfx,Yh;p,p ::; llu(t + ·)- u(s + ·)llt(lfp,p;X,Y)

= 1 00
(llu(t + T)- u(s + T)ll)c + llu1(t + T)- u 1(s + T)ll~) dT--> 0
as t --> s. Hence we have

V(1jp,p; X, Y) c C([O, oo)); (X, Y)Ifp,p), (1.17)

where C([O, oo); (X, Yh!v,p) is the set of all continuous functions in [0, oo)
with values in (X, Yh!v,v·
We denote by W 1,P(O, oo; Y) the set of all functions u such that u,u 1 E
LP(O, oo; Y). W 1 ,P(O, oo; Y) is a Banach space with norm

p I p ) 1/p
llullw 1 .P(O,oo;Y) = ( lluiiLP(O,oo;Y) + llu lbco,oo;Y) ·

If X and Y are Hilbert spaces, then £ 2 (0, oo; X) and W 1 ,2 (0, oo; Y) are
Hilbert spaces.
If X is a dense subspace of Y and the imbedding X C Y is continuous,
then in view of (1.14),(1.17) we have

LP(O, oo; X) n W 1'P(O, oo; Y) C C([O, oo); (X, Y)I/p,p) (1.18)

and
p p )1/p
sup llu(t) llcx,Yh;p,p ::; ( lluiiLP(O,oo;X) + llullwl.P(O,oo;Y)
tEIO,oo)
=lluiiLP(O,oo;X)nWl.P(O,oo;Y) (1.19)

for each p E (1, oo).

Let A be a positive operator. Then p(A) contains a closed sector E =
{A; 00 ::; arg).::; 27r-Bo}U{O}, where 0 < Bo < 71", and II(.\- A)- 1 11::; 1 + I.AI
c
1.4. INTERPOLATION SPACES 11

in E, where Cis a positive constant. Let p > 0. The fractional power AP

of A is defined as the inverse of the operator

where r is the path running from ooe-iiJo to ooeiiJo along the boundary of
E.
Lemma 1. 6 lfO < p < 1- e, then

(D(A), X)o,v c D(A~')

for any 1:::; p:::; oo.

Proof. For ,\ E E

IIA(>.- A)- 1llc(X)

1 c
:::; C, IIA(,\- A)- lic(D(A),X) :::; 1 + 1>-1"

Hence in view of Theorem 1.11

The conclusion follows from (1.20) and

APx = ~
27f2
r
Jr
,\P- 1 A(,\- A)- 1 xd>..

The following lemma is known as the moment inequality.

Lemma 1. 7 Let A be a positive operator. Then for 0 :::; a < f3 < "' there
exists a constant Ca.,f3,-y such that for x E D(A"~)

In particular for 0 :::; a :::; 1, ,\ E E we have

II Aa.(>. A)-111 < Ca. (1.21)

- (1 + 1>-1)1-a.

For the details on fractional powers see H. Triebel [154].

Chapter 2

Singular Integrals

2.1 Singular Integrals of A. P. Calderon and

A. Zygmund
In this chapter the most fundamental part of the theory of the singular
integrals of A. P. Calderon and A. Zygmund

v.p. { K(x- y)f(y)dy = lim { K(x- y)f(y)dy

J R" <->O J1x-yl ><

is described. The results are due to A. P. Calderon and A. Zygmund [27].

Throughout this chapter it is assumed that the integral kernel K(x) satisfies
the following assumption.
Assumption 1 K(x) is a complex valued measurable function defined in
Rn, homogeneous of degree -n:
K(>-.x) = ;._-nK(x) for A> 0, X ERn, (2.1)
absolutely integrable on the unit sphere I;= {x ERn; lxl = 1} and satisfies

h K(x)d(}" = 0, (2.2)

where d(}" is the surface element of I;.

The Fourier transform is denoted by :F:

(:F f)(~) = (27r)-n/2 { e-ixf. f(x)dx. (2.3)

}R"
The Lebesgue measure of a subset A of Rn is denoted by IAI. The norm of
£P(Rn) is denoted by II liP in this chapter.

13
14 CHAPTER 2. SINGULAR INTEGRALS

Lemma 2. 1 For f E £P(Rn), 1 < p < oo, f > 0 the integral

j,(x) = { K(x- y)f(y)dy (2.4)

Jlx-yl><
exists almost everywhere in Rn.
Proof. By a suitable change of independent variables

{ K(x -y)f(y)dy = { K(y)J(x-y)dy

1 1
Jlx-yl>< Jlvl><
= { 00
K(ta)f(x- ta)tn- 1dtda = { K(a) 00
f(x- ta) dt da.
}E < }E < t

Hence, it suffices to show that for any bounded measurable subset S of Rn

(2.5)

is finite. With the aid of Holder's inequality

1s 1 if(x- ta)i ~t dx ~ C,,p 1s (/_: if(x- ta)!Pdt)

00 1
/p dx

~ C,,pjBj 1- 11P (l /_: if(x- ta)!Pdtdx) 1/P, (2.6)

where C,,p = (p- 1)CP- 1l/Pc 1/P. For a fixed a E 2: write x = z- sa, s E
Rl, (z, a) = z1a1 + · · · + Znan = 0. ChooseS' C Rn- 1 and a, b so that

S C {x = z- sa; z E S', a< s < b}.

Then the last side of (2.6) does not exceed

The numbers a, b can be chosen independently of a, and hence (2.5) does

not exceed
C,,pjSj1-1/P(b- a)1/PIIJIIP i
IK(a)jda.

Remark 2. 1 In the above proof (2.2) was not used.

2.2. J. MARCINKIEWICZ'S INTERPOLATION THEOREM 15

2.2 J. Marcinkiewicz's Interpolation

Theorem
Let (0, B, ¢),(0,/3,~) be two measure spaces. The norms of LP(O), LP(O)
are both denoted by II llv·
Definition 2. 1 Let 1 :::; p :::; oo, 1 :::; q :::; oo.
(i) A not necessarily linear mapping T from LP(O) to Lq(O) is said to be
of type (p, q) if there exists a constant K such that IITfllq :::; Kllfllv for any
f E £P(O).
(ii) Let T be a not necessarily linear mapping from £P(O) to a set of mea-
surable functions defined in 0. In case q < oo Tis said to be of weak type
(p, q) if there exists a constant K such that
~({y E 0; I(Tf)(y)l > s}):::; (KIIfllv/s)q (2.7)
for any f E £P(O) and s > 0. In case q = oo a mapping of type (p, oo) is
said to be of weak type (p, oo).
As is easily seen a mapping of type (p, q) is of weak type (p, q).
The totality of functions which are expressed as a sum of a function in
LP(O.) and a function in Lq(O.) is denoted by LP(O) + Lq(O.). Suppose
1 :::; p < q < r < oo and f E U(O). Let /1,/2 be functions such that
/l(x) = 0, fz(x) = f(x) if lf(x)l:::; 1 and /l(x) = f(x), fz(x) = 0 if lf(x)l >
1. Then f = /1 + Jz, lft(x)IP :::; lf(xW, l!z(xW :::; lf(xW. Hence /1 E
LP(O.), fz E U(O). Therefore U(O) c LP(O.) + U(O).
Lemma 2. 2 Let f be a nonnegative function defined in 0., and 1 :::; p < oo.
Set E(T) = {x E 0.; f(x) > T} forT 2: 0. A necessary and sufficient
condition in order that f E LP(O) is

1 00
TP- 1 qy(E(T))dT < 00. (2.8)

In this case
(2.9)

Proof. Suppose (2.8) is true. Then it can be shown that TPqy(E(T))---+ 0 as

T---+ 0, oo as follows. Suppose TPfjJ(E(T)) does not tend to 0 as T---+ 0. Then
there exists a sequence {Tj} such that Tj¢(E(Tj)) 2: {j > O,Tj+l < Tj/2.

1::
Then we have

p 1
TP- 1 qy(E(T))dT 2: qy(E(Tj))(Tj- Tj+l)

2: Tj
{j ( Tjp - p
Tj+l ) = {j ( 1- (Tj+l
--:;:; )p) 2: ({j 1- 21p ) > 0,
16 CHAPTER 2. SINGULAR INTEGRALS

which contradicts {2.8). If TP¢>(E(T)) does not tend to 0 as T -+ oo, then

we can find a sequence {Tj} such that Tj ¢>(E(Tj)) 2 8 > 0, 2Tj < Tj+1· This
leads us to a contradiction analogously. Therefore we obtain with the aid
of an integration by parts

Conversely suppose f E LP(O). Since

we see that TP</>(E(T))-+ 0 as T-+ oo. Consequently

r
Jn f(x)P</>(dx) =-lim
,_.o
1'
00
TPd¢>(E(T))

=!!To [EP</>(E(E)) + p 1 00
TP- 1¢>(E(T))dT] 2 p 1 00
TP- 1</>(E(T))dT,

from which (2.8) follows.

Theorem 2. 1 Let 1 ::::; p < q < r and T be a linear mapping from LP(O) +
U{O) to a set of measurable functions defined in 0. If T is of weak type
(p,p) and also of weak type (r,r), then Tis of type (q,q).

Proof. For a measurble functions defined in n and T > 0 put

{f)T = </>({x E fl; if(x)i > T}).

We use the same notation for a measurable function defined in with ¢ in n

place of ¢>. By the assumption there exists a positive constant K such that
for any T > 0, f E LP(O) or f E U(O)

Let f be a function belonging to 0(0). Put

if(x)i ::::; T
ft(x) = { f(x) , h(x) = f(x)- ft(x),
Tsignf(x) if(x)i > T

Then Tf = Tft +Th, and hence

{y; i(Tf)(y)i > T} C {y; i(Tft)(y)i > T/2} U {y; i(Tf2)(y)i > T/2}.
2.2. J. MARCINKIEWICZ'S INTERPOLATION THEOREM 17

Therefore by (2.10)

Noting (ft)a = (f)a for a < T, (fda = 0 for a ~ T, (h)a = (f)a+r for any
a~ 0, we get in view of Lemma 2.2

1
lift II~= r 00
ar-l(ft)adn =r lr ar-l(f)ada, (2.12)

llhll~ = 1 P
00
aP- 1 (h)ada

= 1
p
00
aP- 1 (f)a+rda =p 100
(a- T)P- 1 (f)adn. (2.13)

With the aid of (2.11),(2.12),(2.13)

Again using Lemma 2.2

18 CHAPTER 2. SINGULAR INTEGRALS

2.3 Case of Bounded Kernels

In this section we assume that the kernel K(x) satisfies the following as-
sumption.
Assumption 2 K(x) satisfies Assumtion 1 and is bounded on the unit
sphere E.
For a function f E L 2 (Rn) we consider the following integral

lim j,(x) =lim { K(x- y)f(y)dy. (2.14)

<->O <->O Jlx-yl><
For 0 < E < f1- put
K(x) E < lxl < fl- K(x) lxl > f
K,,Jl(x) = { 0 K,(x) = { . (2.15)
otherwise 0 lxl:::; E
Then
j,(x) = {
}Rn
K,(x- y)f(y)dy = (K, * f)(x).
We calculate the Fourier transform of K,,w Note that K,,J.L E L 1 (Rn).
Write lxl = r, 1~1 = p,x~ = rpcos¢,x = rrJ for x,~ E Rn. In view of
(2.1),(2.2)

(2.16)

Write
J.LP (e-iscos,P - e-s) ds = Il + I2i, (2.17)
1 p s
J.LP ds 1J.LP ds
I1 = 1 (cos(scos¢)- e- 8 ) - , h =- sin(scos¢)-.
p s <p s
For 0 <a< b
b coss 1b+1r/ 2 cos(s- 7r/2) 1b+1r/ 2 sins
1 --ds = ds= ds
a+1r/2 S-7f/2 a+1r/2 S-7f/2
1 - -1)
S

b+'lr 12 sin s 1b+1r / 2

--ds +
(
= 1 sin sds. (2.18)
+1r/2 S a+1r/2 S- 7f/2 S
2.3. CASE OF BOUNDED KERNELS 19

. .
Smce hm
b->oo
Hence if cos¢
1b·
smx . .
--dx eXIsts, we see from (2.18) that hm
f= 0,
X b->oo a
1b cos8 .
--d8 eXIsts.
8

exists. Next we obtain the estimates of (2.17) independent of E, p,. Let A

be a constant such that
(2.19)

for any a> 0. As is easily seen II 2 1::; 2A. In view of (2.18),(2.19)

lia
b cos 8
--d8
8
I< 2A + lb+n/
- a+n/2
2
( 1
8-7r/2
- -1) d8
8

b(a+K/2) ( 7r) (2.20)

= 2A + log a(b + 7r / 2) ::; 2A +log 1 + 2a .

First assume cos¢ > 0. Then

1 1
icos(8cos¢)- e- 8 1-
d 8

1
8

:S 1
1 ds +
(1- cos(scos¢))-
1
(1- e- 8 )d8
-

1
8 0 8

::; 1
1 d8 +
(1- coss)-
8 0
1
(1- e- 8 )d8
- =B.
8
(2.21)

Hence
(i) if p,p < 1, II11 ::; B,
(ii) if Ep > 1, in view of (2.20)

lhl= 11
J.Lpcos </> cos 8
-d8-
1J.LP e -s -d8 I
<pcos<f> 8 <p 8

::; 2A + log ( 1 + 1r ) + 1oo C d8

8 -
2Epcos¢

1
1 8

00
::; 2A +log ( 1 + -1r- ) + d8,
e- 8 -
2cos¢ 1 8

(iii) if Ep < 1 < p,p, in view of (2.20),(2.21)

II1I = 1(1; + 1J.LP) (cos(8cos¢) -e-s) d:l

20 CHAPTER 2. SINGULAR INTEGRALS

< B+ I1J-tpcos¢ coss dsl + roo e_8ds

- cos¢ S J1 S

::; B+2A+log (1+ 2 c:s¢) + 1 e- 8 ~ •

00 8

The case cos¢ < 0 can be estimated similarly. Summing up we see that
there exists a constant C such that

11<p 1-'P
(e-i8cos¢- e-8)- ::;
s
dsl c +log--.
1
Icos¢1
(2.22)

Consequently

(27r)2nl(.rX<J-t)(~)l::; k l:PIK(o-) (e-i8cos¢- e-8) ~s Ido-

::; sup IK(o-)1 r

aEE }E (c 1 "'I) do-.
+log - 1
COS 'I'

Thus there exists a constant M such that for any E, f.J,, ~

(2.23)
As f..t--+ oo, K<J-t converges strongly to K. in L 2 (Rn), and hence FK<J-t tends
to FK. strongly in L 2 (Rn). On the other hand in view of(2.16),(2.22) for
anyC~ 0

lim (27r)nf2(FK.~-')(~) = { K(o-) 1oo (e-i8cos¢- e-8) ds do-,

J-t--->oo jE <p S
I}<proo (e-i8cos¢ - e-8) dss I: ; c +log _1_.
I cos¢1
(2.24)

Therefore

In view of (2.24), (2.25) for any ~ i- 0

lim(FK.)(~) = (27r)-nl 2 { K(o-) {oo (e-i 8cos¢- e- 8) ds do-. (2.26)
<--->0 JE Jo s
The right hand side of (2.26) is defined to be the Fourier transform of K(x),
and is denoted by (FK)(~). Obviously FK is a homogeneous function of
order 0, and by virtue of (2.23) we have
I(FK.)(OI::; M, I(FK)(OI::; M. (2.27)
2.4. CASE OF CONTINUOUS KERNELS 21

a. e. and also in the strong topology of £ 2 (Rn ). Therefore as It --+ 0 {Kt 11 *!}
is a Cauchy sequence in L 2 (Rn). The limit is Kt * f since Kt 11 * f tends to
Kt * f a.e., and

Since :FKt <Ff is strongly convergent to :FK ·:Ff in L 2 (Rn) as E--+ 0, Kt*f
is also strongly convergeent in L 2 (Rn). We denote the limit by K *f. Then

:F(K *f)= l.i.m.{-to:F(Kt *f)= (27r)ni 2:FK · :FJ.

Consequently

IlK* fll2 = IIF(K * !)112 = (27r)n 12 II:FK · Ffll2

:::; (27r)n/2MII:F fll2 = (27r)n/2Mllfll2·

Summing up we have established the following theorem.

Theorem 2. 2 Let K(x) be an integral kernel satisfying Assumption 2.

For f E L 2 (Rn), E > 0 put

.f.(x) = { K(x- y)f(y)dy.

Jlx-yl>t

Then there exists a constant C independent of E and f such that llftll2 :::;
Cllfll2·{it} is stron11ly convergent in L 2 (Rn) as E --+ 0. The mapping
f H K *f =lim{-to ft is a bounded linear operator from L 2 (Rn) to itself
The Fourier transform of K * f is expressed as

(2.28)

where :FK is an essentially bounded measurable function defined by the right

hand side of (2.26). :FK is a homogeneous function of order 0.

2.4 Case of Continuous Kernels

In this section we assume
22 CHAPTER 2. SINGULAR INTEGRALS

Assumption 3 In addition to Assumption 2 K(x) satisfies the following

conditions. There exists a function O(x) homogeneous of order 0 such that

K(x) = 0 c:l) lx~n, x E Rn \ {0}, (2.29)

IO(x)- O(y)l ~ w(lx- yl) for x, y E E, (2.30)

where w(t) is a nonnegative continuous increasing function defined in [0, oo)

satisfying
{1 w(t) dt < oo w(t) >Co> 0. (2.31)
Jo
t ' t -
The following lemma is known as the Calder6n-Zygmund decomposition
([27]).
Lemma 2. 3 Suppose f 2:: 0, f E LP(Rn), 1 ~ p < oo. Then for any
positive number s there exists a sequence of non-overlapping cubes {h}
such that
s ~ l~llk
f(x)dx < 2ns. (2.32)

Put Ds = Ukh· Then f(x) ~ s a. e. outside D 8 , and

(2.33)

Proof. Let I be a cube. Since

as III ---+ oo, there exists a positive number v such that if III 2:: v

l~ll f(x)dx < s.

Let Rn = Uk'= 1 Io,k be a decomposition of Rn into a sum of cubes of volume
v. Then for each k
- _/ I f f(x)dx < s.
1
O,k Jro,k
Denote by {h,k} the sequence of cubes obtained by dividing each Io,k into
2n equal parts. We classify {h,k} as follows:

~ f f(x)dx < s.
lr"
- I
1
I1 , k l,k
2.4. CASE OF CONTINUOUS KERNELS 23

Denote by {I2,k} the sequence of cubes obtained by dividing each I~:k into
2n equal parts, and we classify it as follows:

I I'~2,k I r,
J12,k
f(x)dx < s.

Repeating this process we get a sequence of cubes {Im,k} such that

{Im,k} = {I:n,d u {I~.d. -II/I r,

m,k JI=,k
f(x)dx;::: s, II} I
m,k JI=.k
r,
f(x)dx < s

for each m. Then

If m;::: 1, each I:n,k is obtained by dividing some I~_ 1 , 1 , where I~ 1 = Io,l·

Hence II:n,kl = 2-nii~- 1 , 1 1. Therefore

II~,k~1~.k f(x)dx::; 1Il:1,tl1::,_ 1 .l f(x)dx < 2ns.

Ler {Ik} be the totality of I:n,k, m = 1, 2, ... , k = 1, 2,. ... Then (2.32)
holds for any k. Put D 8 = Ukh· Since

lhl::;- 11
s h
f(x)dx::; 1 (1
-lhi 1- 1/P
s h
f(x)Pdx ) 1
/P ,

we get

Hence

It is easy to show that (2.33) holds. Let x 0 be a Lebesgue point off which
does not belong to D 8 • Then for any m ;::: 1 there exists an integer k(m)
such that Xo E I~,k(m)" Since {I~,k(m)} is a sequence of regular closed sets
tending to x 0 , we get by Theorem 1.1

s> II'' 1
m,k(m)
1
1,
=,k(=l
f(x)dx---+ f(xo).

Thus we conclude f(xo) ::; s.

24 CHAPTER 2. SINGULAR INTEGRALS

Lemma 2. 4 Let K(x) be an integral kernel satisfying Assumption 3. Sup-

pose 0:::; f E LP(Rn), 1 :::; p:::; 2, E > 0. Let K, be the function defined by
*
(2.15), and let j, = K, f be the function defined by (2.4). Let s be an
arbitrary positive number. If we put Es = {x; l],(x)l > s }, then there exist
positive constants C1, C2 independent of j, E, S SUCh that

(2.34)

where Ds is the set in Lemma 2.3, and [f(x)]s is the function defined by

f(x) :::; s
[f(x)]s = { ~(x)
f(x) > s.

Proof. We denote various constants independent of J, E, s by C. Let {h}

be the sequence of cubes of Lemma 2.3, and g, h be the functions defined
by

h(x) = { II~Ihk f(y)dy xEh

, g(x) = f(x) - h(x).
f(x) X~ Ds

Then s :::; h(x) :::; 2ns in D 8 , and h(x) = f(x) = [f(x)]s in Rn \ D 8 • For
each k
r
g(y)dy = o,
Jh
(2.35)

and g(x) = 0 in Rn \ D 8 • If we put h, = K, * h, g, = K, * g,

then
E 8 CE1UE2. (2.36)
With the aid of the proof of Theorem 2.2

f ih,(x)l 2 dx =
}Rn
(2n)n f I(.1"K,)(~)(Fh)(~)l 2 ~
}Rn
:::; (2n)nM 2 r I(Fh)(~Wd~ =
}Rn
(2n)nM 2 r
}Rn
h(x) 2 dx.

Hence
(2.37)
2.4. CASE OF CONTINUOUS KERNELS 25

On the other hand

{ h(x) 2 dx = { h(x?dx + { h(x) 2 dx ~ 22 ns2 1Dsl + { [f(x)];dx.

}Rn lv. }R.. \D. }Rn
Combining this with (2.37) we get

IEll ~ 22 n+2 (27r)n M 2 1Dsl + ;

S
(27r)n M 2 r
}Rn
[J(x)];dx. (2.38)

Letting Sk be the ball centered at the center of h and of radius the diameter
of h we put Ds = uksk, s~ = Rn \ sk, b~ = Rn \ Ds. Then

(2.39)

Noting

we have

Let x be fixed in 8~. Denote the center of h by Yk· In case where {y; ly-xl ~
E} n h is empty we get in view of (2.35)

{ K<(x- y)g(y)dy ={ K(x- y)g(y)dy

~k ~k
= { (K(x- y) - K(x- Yk))g(y)dy. (2.41)
Jh
We estimate each term of the rightmost side of

(2.42)
26 CHAPTER 2. SINGULAR INTEGRALS

Ify E h, we have lx-ykl/2:::; lx-yl:::; 3lx-ykl/2since IY-Ykl:::; lx-ykl/2.

Hence

Noting

Ilx - Yly
X- X - Yk I
lx - Yk I
< - I
X-- y - X- y
- lx- Yl lx- Ykl
I+ I X- y
lx- Ykl
- I
X- Yk < 2ly- Ykl < --'vnlhl
lx- Ykl - lx- Ykl - lx- Ykl '
1/n
-:----'-----'----:-

we have in view of (2.30)

In ( lxx-y)
1
(X-Yk)l (Vnlhl /n) (2.44)
- Yl - n lx - Yk I :::; w lx - Yk I .
Combining (2.42),(2.43),(2.44) and using the latter inequality in (2.31) we
get
IK(x- y)- K(x- Yk)l
11 1
< Clhil/n + 1 w (vfnllkl n) < C w (Vnlhl /n) .
- lx- Ykin+l lx- Ykin lx- Ykl - lx- Ykln lx- Ykl
It follows from this inequality and (2.41) that

Ijhr K.(x- y)g(y)dyl:::; IX - cYk In w (fihll~n)

X - Yk
r lg(y)idy.
jh
(2.45)

Next consider the case where {y; IY - xi :::; f} n h is not empty. Note that
in this case
h C {y; IY- xi :::; 3f}. (2.46)
Let "f be the characteristic function of the interval [0,3]. If y E h, then
"'(ix- yjjf) = 1 in view of (2.46). Hence, noting that IK.(x- y)i:::; ccn

Ilk K.(x- y)g(y)dyl:::; cf_-n lk"' ('X~ Yl) lg(y)ldy. (2.47)

2.4. CASE OF CONTINUOUS KERNELS 27

It follows from (2.45),(2.47) that

Ijhr K<(x- y)g(y)dyl:::; IX - cYk In w (flhll/ln)

X- Yk J1kr lg(y)ldy
+CE-n lk 'Y ( lx ~ Yl) lg(y)ldy (2.48)

holds for X E s~. By virtue of (2.40),(2.48)

( lg<(x)ldx:::;'""' f I c I w (flhll~n) f lg(y)ldydx

)b,• L...J}s'
k k
x-ykn x-yk J1 k
+2: f,ccn f "~(lx-yl)lg(y)ldydx
k Jsk Jh E

= 2:k Js~f IX-Yk

c lnw (flhll~n) dx f lg(y)ldy
X-Yk }h
+CE-n2: { lg(y)l { 'Y(Ix-yl)dxdy. (2.49)
k J1k ls~ E

Since

we obtain from (2.49) that

(
lb~
jg<(x)ldx:::; c2: J1kr lg(y)ldy = c Jv.
k
r lg(y)ldy. (2.50)

On the other hand

{ lg(y)ldy:::; { f(y)dy + { h(y)dy = 2 { f(y)dy:::; 2n+ 1 siDsl·

Jv. Jv. Jv. Jv.
Combining this with (2.50) we get

( lg<(x)ldx:::; CsiDsl·
Jv~
28 CHAPTER 2. SINGULAR INTEGRALS

This implies ID~ n E2l ::; CIDsl· Hence noting (2.39)

IE21::; JD~ n E21 +IDs n E21::; CJDsl· (2.51)

We conclude (2.34) combining (2.36),(2.38),(2.51).
Theorem 2. 3 Suppose K(x) is an integral kernel satisfying Assumption
3. Let f E LP(Rn), 1 < p < oo. ForE> 0 let j,(x) be the function defined
by (2.4). Then there exists a constant Cp independent of E, f such that

(2.52)

As E --. 0 {].} is strongly convergent in LP(Rn). If we denote the limit

by K * f, then the mapping f f---7 K * f is a bounded linear operator from
LP(Rn) to itself
Proof. Write T.j = j,. In view of Theorem 2.2 there exists a constant
C such that IIT./11 2 ::; Cll/11 2 for f E L 2 (Rn), E > 0. Therefore T, is of
type (2,2) uniformly with respect to E. If 0 ::; f E L 1 (Rn), we have by
(2.33),(2.34) and [f(x)];::; sf(x)

l{x; I(T.J)(x)l > s}l::; Cllflltfs.

If f is an arbitrary element of L 1 (Rn), then noting f = ft - /2, ft
max{!, 0}, h = -min{!, 0} we get
l{x; J(T.J)(x)l > s}l
s/2}1 + l{x; I(T,h)(x)l > s/2}1
::; l{x; I(T.jt)(x)l >
::; 2CIIftllt/s + 2CIIhllt/s = 2CIIflltfs.
Therefore T, is of weak type (1, 1) uniformly in E. Hence by virtue of Theo-
rem 2.1 and its proof we have for any f E LP(Rn), 1 < p::; 2, E > 0
(2.53)
for some constant Cp. Considering the adjoint operator we see that (2.53)
holds also for 2 < p < oo.
NeXt, let f E LP(Rn), 1 < p < oo. If g E CJ(Rn), i.e. g is a continuously
differentiable function having a compact support in Rn, and E, {j > 0, then
we have
liT./- To!IIP ::; 2Cpll/- 9llp + IIT.g- TogiiP (2.54)
by virtue of (2.53). If 0 < E < 1, then by (2.2)

(T,g)(x) ={
Jlx-yl>l
K(x- y)g(y)dy + 1 <<lx-vl<l
K(x- y)(g(y)- g(x))dy.
(2.55)
2.5. HILBERT TRANSFORM 29

Since K E £P( { x; Ixi > 1} ), g E £ 1 (Rn), the first term on the right of (2.55)
belongs to LP(Rn) in view of Lemma 1.2. The second term on the right
of (2.55) is a function with support contained in a fixed compact set, and
converges uniformly as E ---+ 0 since g satisfies a uniform Lipschitz condition.
Hence it converges strongly in LP(Rn). Combining this with (2.54) we
conclude that Td is strongly convergent in LP(Rn) as E---+ 0. Hence using
(2.53) we obtain
IlK* fllv ::; Cvllfllv·

2.5 Hilbert Transform

The mapping f 1---t K *f in the previous sections essentially reduces to

!
7r
1 00
f(t) dt =
-oo X - t
!
7r
lim
E--+0
r
llx-tl>•
f(t) dt
X- t
(2.56)

when n = 1. This is also expressed as

where Pf.1/ x is the distribution defined by

(pr.!) (¢) =
X
lim
<--++O
r
llxl>•
¢(x) dx.
X

Definition 2. 2 The function defined by (2.56) is called the Hilbert trans-

form of f.

In what follows we calculate the Fourier transform of 7r- 1Pf.1/x following

the proof of Theorem 2.2. For 0 < E < ft we put

1/(7rx) E < lxl < tt 1/(7rx) lxl > E

h.,/1-(x) = { h.(x) ={ .
0 otherwise 0 lxl ::; E
Then

(Fh.~)(~) = _1_1
...j2if •<lxl <,u
e-ixe dx =-
7rX ;: 7r 1"' sin(x~) dx
v~!._ X

~<0
30 CHAPTER 2. SINGULAR INTEGRALS

Letting fJ,--. oo we get

(2.57)

Put h = 7r- 1Pf.1/x. Letting E--. 0 in (2.57) yields

(:Fh)(E;) = -(27r)- 112 isign£;. (2.58)

In view of (2.57) we have

where A is the constant in (2.19). Hence for f E L 2 (R)

llh, * Jll2:::; 4A7r- 1 IIJII2·
According to Theorem 2.2

g(x) = -1 hm
.
1f f--->0
1 lx-tl>•
J(t)
--dt
X- t
(2.59)

exists in the strong topology of L 2 (R), and in view of (2.28),(2.58)

(:Fg)(E;) = -i(:FJ)(E;)sign£;. (2.60)

Consequently
llgll2 = 11!112, (2.61)
(:FJ)(E;) = i(:Fg)(E;)sign£;. (2.62)
By virtue of (2.62)

J(x) = --1 hm
.
1f f--->0
1 lx-tl>•
g(t)
-
X-
dt
t
(2.63)

holds in the strong topology of L 2 (R).

Suppose 1 < p < oo. The Hilbert transform is a bounded linear operator
from £P(R) to itself in view of Theorem 2.3. Let f E LP(R) and g be its
Hilbert transform. If fn E £P(R)nL 2 (R), fn--. fin £P(R),gn is the Hilbert
transform of fn, then gn E £P(R) n L 2 (R),gn--. gin LP(R). Since

- fn(x) =! lim
1f f--->0
r
Jlx-tl>•
gn(t) dt
X- t

holds in the strong topology of L 2 (R) and the right hand side is the Hilbert
transform of gn, the above limit exists also in the strong topology of LP(R).
2.6. EQUIMEASURABLE FUNCTIONS 31

Hence letting n ---+ oo we see that (2.63) holds also in the strong topology
of LP(R).
Suppose f E LP(R), IE LP' (R),p' = pf(p- 1), and g,g are the Hilbert
transforms off, I respectively. Let fn E LP(R) n L 2 (R), In E u' (R) n
L 2 (R), fn ---+ f in LP(R), In ---+ I in u' (R), and gn, §n be the Hilbert
transforms of fn, In respectively. Then, gn E LP(R) n L 2 (R), fin E u' (R) n
L 2 (R),gn---+ gin LP(R),fln---+ gin LP' (R). Since the Hilbert transform is
a unitary operator in L 2 (R) in view of (2.61) and (2.63), we have

/_: fn(x)f~(x)d:c = /_: gn(x)§n(x)dx.

Letting n---+ oo we get

/_: f(x)l(x)dx = /_: g(x)g(x)d:c. (2.64)

Summing up we have proved the following theorem.

Theorem 2. 4 For 1 < p < oo the Hilbert transform is a bounded linear
operator from LP(R) to itself If g is the Hilbert transform off E LP(R),
then the inverse transform is given by (2.63), and the limits in the right
hand sides of (2.59) and (2.63) exist in the strong topology of LP(R). If
f E LP(R), IE LP' (R),p' = pf(p- 1) and g,g are the Hilbert transforms
off, I respectively, then the equality (2.64) holds. In particular the Hilbert
transform is a unitary operator in L 2 (R), and (2.60) holds if g is the Hilbert
transform of f.

2.6 Equimeasurable Functions

Let f(x) be a nonnegative measurable function defined in Rn. For each
T ~ 0 put
(2.65)
In this section we consider only functions such that m(T) < oo for any
T > 0 and lim,-__. 00 m( T) = 0. It is easy to see that functions f such that
0::; f E LP(Rn), 1::; p < oo, satisfy this condition.
Definition 2. 3 Let f be a nonnegative measurable function defined in
Rn. A nonnegative measurable function J* defined in [0, oo) is called an
equimeasurable function of f if

m*(T) =: i{t; j*(t) > T}l = m(T) (2.66)

holds for any T ~ 0.

32 CHAPTER 2. SINGULAR INTEGRALS

Lemma 2. 5 The function m(T) defined by (2.65) is monotone decreasing

and right continuous.
Proof. It is obvious that m(T) is monotone decreasing. The right continuity
follows from

Put
to= m(O), To= sup{T; m(T) > 0}. (2.67)
Clearly 0 < to :::; oo, 0 < To :::; oo.
Lemma 2. 6 The function J* defined by
j*(t) = inf{T;t 2:: m(T)}, j*(O) =To
is the unique monotone decreasing, right continuous equimeasurable function
of f.
Proof. First we show that T 2:: f*(t) and t 2:: m(T) are equivalent. It is
obvious that t 2:: m( T) implies T 2:: f* (t). Conversely, if T > f* (t) then
t 2:: m(T) by the definition of J*. Suppose T = f*(t). If Tk l T, then
t 2:: m(Tk)· Since m(T) is right continuous, we get t 2:: m(T). Thus, we see
that T < f*(t) and t < m(T) are equivalent. Hence for any T > 0
{t;j*(t) > T} = [O,m(T)).
Consequently if m*(T) is the function defined by (2.66), then we have
m*(T) = m(T). Clearly J*(t) is monotone decreasing. For each T > 0
we have t 0 2:: m(T), or T 2:: f*(t 0), and hence J*(t 0) = 0. Therefore, if
to < oo, then f*(t) = 0 for any t > t 0 • If t < to, then t < m(T) for
some T > 0. Since T < J*(t) then, we have J*(t) > 0. Hence, J*(t) > 0
for 0 < t < to. Suppose 0 < t < to and tk l t. If 0 < T < J*(t),
then t < m(T). Hence tk < m(T) if k is sufficiently large, which implies
T < J*(tk):::; J*(t). This shows f*(t) = limk-texl f*(tk)· Consequently!* is
right continuous. Finally suppose g* is another monotone decreasing right
continuous equimeasurable function of f. Let t 1 be an arbitrary positive
number. Put T = J* (t1), t2 = min{ t; J* (t) = T} = inf{ t; J* (t) = T}. Since
{t; f* (t) > T} = [0, t 2), we have
i{t;g*(t) > T}l = m(T) = l{t; j*(t) > T}l = t2.
Since g* is monotone decreasing and right continuous, we have {t; g* (t) >
T} = [0, t 2). Hence
g*(t1):::; g*(t2):::; T = j*(t1).
Similarly we can show J* (t1) :::; g* (t1). Thus J* (t) = g* (t).
2.6. EQUIMEASURABLE FUNCTIONS 33

Lemma 2. 7 If J* is an equimeasumble function off, then for 1::; p < oo

r f(x)Pdx = roo j*(t)Pdt. (2.68)

}Rn Jo
Proof. In view of (2.66)

r f(x)Pd:J; =- roo rPdm(T) =- roo rPdm*(r) = roo j*(t)Pdt.

}Rn Jo Jo Jo
Lemma 2. 8 If A is a measumble subset of Rn of finite measure, then

r
}A f(x)dx::;
riAl
Jo f*(t)dt. (2.69)

Proof. Let h be the function such that h(x) = f(x) for x E A and
h(x) = 0 for x f/:. A. Since h(x)::; f(x) we have
m1(r) =l{x; h(x) > r}l::; l{x; f(x) > r}l = m(r).
Therefore if ft is the monotone decreasing right continuous equimeasurable
function of h, then

f;(t) = inf{r;t ~ m 1(r)}::; f*(t).

Since {x; h(x) > r} C {x; h(x) > 0} C A for any T > 0, we have m1(r)::;
IAI. Hence fi(t) = 0 fort~ IAI. With the aid of Lemma 2.7

1 A
fWd:c= 1 Rn
hWd:c= la
o
oo flW&= !alAI flW&::; !alAI f*W~
0 0

In what follows in this section we suppose that 0::; f E LP(Rn), 1 ::; p <
oo, and J* is the monotone decreasing, right continuous equimeasurable
function of f. For t > 0 we put

{31 (t) =! t f*(s)ds. (2.70)

t Jo
The right hand side of (2.70) is finite since J* E £P(O, oo). Since

!f3t(t) = t; lat(f*(t)- f*(s))ds::; 0,

if f3j(t 1 ) = 0 for some t 1 > 0, then f* (t) = J* (t1) in (0, h), and hence
f3t(t) = f* (t1) there. Therefore we see that either f3t(t) is strictly decreasing
34 CHAPTER 2. SINGULAR INTEGRALS

in (O,oo) or there exists a positive number t 1 such that f3J(t) is constant in

(0, t1] and strictly decreasing in [t1, oo). In the former case we put t1 = 0.
Put f3J(O) = 81 = limt--->o f3J(t). As is easily seen 0 < 81 ::; oo and 81 =
f3J(tl)· It is easy to see that f3J(t) ---? 0 as t---? oo.
We denote the inverse function of 8 = f3J(t) by t = (31(8). If t1 = 0, 81 =
oo, then f31 (8) is uniquely determined in 0 < 8 < oo. If t 1 = 0, 81 < oo, we
put f31 (8) = 0 for 8 > 81. If t 1 > 0, we put f31 (8 1) = t1 and f31 (8) = 0 for
8 > 81.
Lemma 2. 9 Iff E LP(Rn), 1 < p < oo, we have

(2.71)

Proof. By virtue of Lemma 2.7 J* E LP(O, oo). For 0 < a< b we get by
integrating by parts

1b f3J(t)Pdt =
1b (Jot
a CP j*(8)d8
)p dt::; pal-p (r
_ 1 Jo j*(8)d8
)p
+p ~ 1 1 b
t 1-p f*(t)
(
1 t
j*(8)d8
)p-1 dt. (2.72)

The first term of the right hand side of (2. 72) tends to 0 as a ---? 0 since

1-p (1a j*(8)d8)p ::; -1- 1a j*(8)Pd8

_a_
p-1 0 p-1 0

by Holder's inequality. Hence letting a ---? 0, b ---? oo in (2. 72) we obtain

roo f3J(t)Pdt::; _P_ roo j*(t)f3J(t)P-ldt

Jo P -1 Jo
::; p ~ 1 (1oo j*(t)Pdt) 1/p (1oo f3J(t)Pdt) 1-1/p'
from which (2. 71) readily follows.
Lemma 2. 10 Suppose 0 ::; f E LP(R), 1 < p < oo. For E > 0 put

F,(x) = - 11'
f_ 0
J(x + y)dy, G(x) = supF,(x).
<>0
(2.73)

Then

1 00

-00
G(x)Pdx::; (
p
~ 1 )P 1 00

-00
f(x)Pdx. (2.74)
2.6. EQUIMEASURABLE FUNCTIONS 35

Proof. First note that G(x) is measurable since the supremum of the right
hand side of the second equality of (2. 73) may be taken only over rational
c:'s. Put H(T) = {x; G(x) > T} forT> 0. Since

y>x
1
G(x) =sup--
Y- X
1y
X
f(t)dt,

X E H(T) if and only if

ly f(t)dt- YT > lx f(t)dt- XT

for some y > x. Hence, if we set F(x) = J;J(t)dt- XT, we have

H(T)={x;F(y)>F(x) forsome y>x}. (2.75)

Since H(T) is an open set, it is the sum of disjoint open intervals: H(T) =
U(ak, bk)· With the aid of HOlder's inequality we see

lim F(x) = :r=oo. (2.76)

x->±oo

We show -oo < ak < bk < oo for any k. Suppose ak = -oo for some k.
Then ( -oo, bk) C H(T). By (2.76) we have F(c) > F(bk) for some c < bk.
Let c1 be such that F(c1) = maxc:S:x:S:h F(x), c1 E [c, bk]· Then c S c1 < bk
since F(cl) 2 F(c) > F(bk)· Hence c1 E ( -oo, bk) C H( T). If c1 < x S bk,
then F(x) S F(c1). If x > bk, then F(x) S F(bk) < F(cl) since bk rt. H(T).
Hence F(x) S F(c 1) for any x > c1 . This contradicts c1 E H(T). Next
suppose bk = oo. Then (ak,oo) C H(T). Let c > ak. In view of (2.76)
F(c1) = maxx;:::c F(x) for some c1 2 c. Then F(x) S F(c1) for x > c1.
This is a contradiction since c1 E ( ak, oo) C H (T). Thus we have proved
-oo < ak < bk < oo.
Next we show F(ak) = F(bk) for any k. Since ak rt. H(T) and bk > ak,
we have F(bk) S F(ak)· Suppose F(bk) < F(ak) for some k. Then F(bk) <
F(c) for some c E (ak, bk)· Let c1 be such that F(cl) = maxc:s:xSbk F(x), c S
c1 S bk. Since F(c 1) 2 F(c) > F(bk), we have c S c1 < bk. Hence
c1 E (ak, bk) C H(T). If c1 < x S bk, F(x) S F(c1). If x > bk, F(x) S
F(bk) < F(c1) since bk rt. H(T). Hence F(x) S F(c1) for any x > c1. This
contradicts c1 E H(T). Thus we conclude F(ak) = F(bk)·
Consequently
36 CHAPTER 2. SINGULAR INTEGRALS

and hence

Therefore

In view of Lemma 2.8

IH(T)IT = L(bk- ak)T = L 1 bk

f(t)dt =
i f(t)dt::;
1IH(r)l
j*(t)dt.
k H(r) 0

Consequently
1 {IH(r)l
T::; IH(T)I Jo j*(t)dt = ,BJ(IH(T)I),

which implies IH(T)I::; ,Bf(T). From this and Lemma 2.2

1 00

-oo
G(x)Pdx = p roo Tp-liH(T)IdT
Jo
::; p 1oo Tp-l,Bf(T)dT 181 ,Bf(T)dTP.= (2.77)

As was shown in the proof of Lemma 2.9

lim t,BJ(t)P =lim tl-p (

t-.o t->O
rt j* (s)ds)p =
}0
Q.

Hence making the change of the independent variable t = ,Bf (T) in the last
integral of (2. 77)

1-00
00
G(x)Pdx::; -1 h
00
td,B1(t)P =- [t,B1 (t)J: + 1h
00
,BJ(t)Pdt

::; tlfiJ(t!)P + roo ,B1(t)Pdt

Jh
= 1h fiJ(t)Pdt + 1 00

1
fiJ(t)Pdt = 1
00
fiJ(t)Pdt.

With the aid of this inequality and Lemma 2.9 we obtain (2.74).
2.7. HILBERT TRANSFORM (CONTINUED) 37

Lemma 2.11 Under the assumptions of Lemma 2.10 set

f(x) =sup -2
e>O E
11€ -E
f(x + y)dy.

1
Then

1 00

-00
f(x)Pdx:::; 2 (
p
~
1
)P
00

-00
J(x)Pdx. (2.78)

Proof. Set E(T) = {x; f(x) > T} forT> 0, and

Let H(T) be the set in the proof of Lemma 2.10 and H-(T)
= {x;G-(x) > T}. Then by the proof of Lemma 2.10 we get IH-(T)I:::;
fj/(T). Since f(x) :::; (G(x) + c-(x))/2 we have E(T) c H(T) n H-(T).
Hence

Following the proof of the previous lemma we conclude (2.78).

2.7 Hilbert Transform (Continued}

Let¢ be an even function in C 1 (R) such that ¢(x) = 1 for x :2:: 3/4, ¢(x) = 0
for 0 :::; x :::; 1/4 and 0 :::; ¢(x) :::; 1 in R. We denote by 1/J(x) the Hilbert
transform of ¢(x)jx. Since

d r ¢(t) dt d ¢(x- t) dt r
dx Jlx-tl>e -t-X-t = dx }ltl>e X- t t
r f) ¢(x- t) dt d ¢(t) dt r
= lltl>e OX X- t t = llx-tl>e dt -t-X-t'
we see that
_!!:_ 1/J (X)
dx
= lim .!.
E--+0 7r
r .!!_ f/J (t) __!:!___
llx-tl>e dt t X- t

exists and is continuous. It is also easily seen that 1/J is an even function.
Next we show that

(2.79)
38 CHAPTER 2. SINGULAR INTEGRALS

as x ---+ ±oo. Suppose x > 1. Then,

1r'lj;(x) =
. lim
N--->oo, (1-1 10
t--->0 -N + _ +
11o
+ lx-t + 1N ) -
¢(t) dt
x+t t - - ,t
1- 1-
1 1 X-

1 1
¢(t)-dt
- -= dt = --1 [log--
N +log(x+ 1)] ,

1
-N t x - t -N t(x - t) x x+N

1° -1
¢(t)- -
- dt -
t X- t
--
0
1 ¢(t)- -
-
t
dt
X+ t
,

l x-t -
</J(t) dt
-- =
t X- t
lX-t dt =-1 [ log--
1 t(x - t) X
X-
+log(x -1) ,
E
E ]

= .! (log__!!__ -log X+ E) .
1x+tN<jJ(t) __!!:!:__ = 1N
t X- t
dt
x+t t(x - t) X N -X E

Hence we have

1
1r'lj;(x) =-log--+
x x+1
X- 1 10
1
(
2¢(t)
)(
x-t x+t
)dt

for x > 1, from which (2.79) follows.

Theorem 2. 5 Suppose f E LP(R), 1 < p < oo. Set

for E > 0. Then, there exists a constant Cp depending only on p such that

As E---+ 0, {]t} converges both almost everywhere and in the strong topology
of LP(R).

Proof. Let ¢, 'ljJ be functions as above. Then

2.7. HILBERT TRANSFORM (CONTINUED) 39

The Hilbert transform of¢ (x- t) - E X-

1- considered as a function oft for
t

a fixed x is -~'1/J ( x ~ t). Hence, if we denote the Hilbert transform off

by g, then in view of Theorem 2.4

h- ' .(x) = 1
--
7rE
1 '1/J (X-- t)
_
00

00 E
g(t)dt

= _ __!__
7rE
roo '1/J (!) (g(x- t) + g(x + t))dt.
lo E
(2.82)

Set
I(x; t) = 1 t
-t
g(x + y)dy, g(x) =sup -2 11<
<>0 E -<
ig(x + y)idy.

According to Lemma 2.11 g E £P(R). Let x be a point such that g(x) < oo.
Then

II(t; x)l $ J:t ig(x + y)idy $ 2tg(x), (2.83)

dl(t; x) = (g(x- t) + g(x + t))dt. (2.84)

Hence, in view of (2.82),(2.84)

]I,.(x) = - :E 1 '1/J ( ~) dl(x; t).

00

Integrating by parts and noting (2. 79)

lt,.(x) = 7r~2 1 00 'l/! ~) I(x; t)dt.

1
( (2.85)

By virtue of (2.83),(2.85)

IA,.(x)l $ 7r~2 1 00 1'1/J' 0) 12tg(x)dt = ~g(x) 1oo ti'l/J'(t)idt. (2.86)

On the other hand

li2,<(x)l $.! 1.
7r </4<lx-tl<<
¢(X- I I E
t)
X-
j(t) dt $
t
_!
7rE
r
l1x-tl<<
IJ(t)idt.

Hence if we put
f(x) =sup -21
<>0 E -<
if(x + y)idy, 1<
40 CHAPTER 2. SINGULAR INTEGRALS

we have
- 8 -
l/2,.(x)l:::; - f(x). (2.87)
7r

From (2.81),(2.86),(2.87) it follows that

- 2
sup lf.(x)l:::; -g(x)
100 tl'l/J'(t)ldt +-8-J(x).
<>0 7r 0 7r

Applying Theorem 2.4 and Lemma 2.11 to the right hand side of the above
inequality we conclude (2.80).
The strong convergence of {f.} in LP(R) was already shown. Let f be~
arbitrary element of LP(R) and g E CJ (R). Set h = f- g. If we define g., h.
as we defined]., then by the argument at the end of the proof of Theorem
2.3 we see that g. converges uniformly. Consequently
lim sup ].(x) -liminf ].(x)
<-->0 <-->0

= limsuph.(x)
<-->0
-liminf ii.(x):::; 2sup jh.(x)l
<-->0 <>0
(2.88)

1: 1:
By virtue of (2.80),(2.88) we get

lli~_!l~P ].(x) -li~~f ].(x)IPdx:::; 2Cp lh(x)IPdx.

The right hand side of this inequality can be made arbitrarily small by
choosing g appropriately. Therefore the integrand of the left hand side
vanishes almost everywhere.

2.8 Case of Odd Kernels

In this section we make the following hypothesis.
Assumption 4 K(x) is an odd function satisfying Assumption 1.
If K(x) is an odd function, (2.2) clearly holds.
Theorem 2. 6 Suppose K(x) is a kernel satisfying Assumption 4. For
f E LP(Rn), 1 < p < oo, E > 0 set

]._(x) = { K(x- y)f(y)dy.

Jlx-yJ><
Then

Ln ~~~ l].(x)IPdx:::; (~cp r (k IK(a)idu) p Ln lf(x)IPdx, (2.89)

2.8. CASE OF ODD KERNELS 41

where Cv is the constant in Theorem 2.5. The singular integral

(2.90)

exists both almost everywhere and in the strong topology of £P(Rn). The
mapping f 1-t K * f is a bounded linear transformation from LP(Rn) to
itself
Proof. As was shown in the proof of Lemma 2.1 we have

j,(x) = 1 1
E
K(a)
,
00 dt
f(x- ta)-da.
t

Since K(x) is an odd function we get

- = 211 1
f,(x)
E
K(a)
ltl>•
dt
f(x- ta)-da.
t
With the aid of Holder's inequality

lf,(x)l:::; ~ JErIK(a)i 11p'IK(a)l 1 /p Iliti>•

r f(x -ta)dtlda
t

: :; ~ (l IK(a)ida)
1/p' (
h Ihi>•
IK(a)l f(x- ta) ~t
p
da
) 1/p

Hence

sup lf,(x)IP:::; -21

<>0 P
(1 E
IK(a)ida )p-11 E
IK(a)l sup
<>0
11 ltl>•
dt
f(x- ta)-
t
p da.

Integrating both sides over Rn

r sup lf,(x)!Pdx:::; 21P ( }ErIK(a)ida)p-

}Rn <>0
1

d
r IK(a)l }Rn
r sup r f(x- ta)_!:.t
p
X dxda. (2.91)
jE <>0 jiti>•

Making the change of the variables x = y +sa, s E R, Y1a1 + · · · + Ynan = 0

for a fixed a E E and applying Theorem 2.5

r sup 1r
}Rn <>0 lltl><
f(x _ ta) dt lp dx
t
42 CHAPTER 2. SINGULAR INTEGRALS

= r
}Rn-1 1 00
sup 1
<>0
r
lttl><
J(y + (s _ t)a) dt lp dsdy
t

L"
-00

:::; (?rCp)P Ln- I: 1 IJ(y + sa)IPdsdy = (1rCp)P IJ(x)IPdx.

Combining this with (2.91)

Ln ~~~ j],(x)IPdx:::; (~cp r (h IK(a)jda) p Ln if(x)IPdx.

Thus (2.89) has been shown. Let f be a function in CJ(Rn). If f(x) = 0

for lxl > M, then j,(x) is independent off for lxl 2::: M + 1, 0 < f < 1. If
lxl <M +1

j,(x) = 1 K(x- y)f(y)dy

1
<<lx-yi<2M+l

= K(x- y)(f(y)- J(x))dy.

<<lx-yi<2M+l

If we set L = max j\7J(x) j, then

r
ltx-yl<<
IK(x- y)(f(y)- J(x))jdy

:::; L r
ltx-yl<<
lx- YIIK(x- y)jdy = fL
JE
r IK(a)jda.
Hence, j, converges uniformly to some function j as f ---+ 0. By virtue of
(2.89)

}Rn
r j](x)jPcb;:::; }Rnr sup jf,(x)jPcb; < 00.
<>0

Consequently j E LP(Rn) and

jj],- ill~= r
ltxi<M+l
j],(x)- ](x)jPcb;---+ 0.

In case f is an arbitrary element of LP(Rn), let g E CJ(Rn) and h = f- g.

As in the last part of the proof ofTheorem 2.5 we can show that f, converges
almost everywhere to some function which we denote by K *f. The first
term of the right hand side of
2.9. RIESZ KERNELS 43

tends to 0 as E, 8---+ 0. Noting in addition that

lihe- h61ip::::; 2 ( r sup lhe(x)!Pdx) p::::; 7rCp rIK(cr)ldallhllp,

}Rne>O
11
J~
we see that {]e} converges in LP(Rn) as E----. 0. By virtue of (2.89) we have

IlK* fliP::::; ~Cp i IK(cr)ldcrii!IIP"

2. 9 Riesz Kernels
The Riesz kernels are functions defined by
r((n + 1)/2) Xj
Rj(x) =- 7r(n+1)/2 lxln+1 (2.92)

for j = 1, 2 .... , n. Clealy the lliesz kernels satisfy Assumptions 3 and 4.

We calculate the Fourier transforms of the lliesz kernels. Let 0 < E < 1 <
p,. With the aid of an integration by parts

(1- n) r e-ixt;___!!j__dx = r e-ixt;~lxl1-ndx

Je<lxl <1 lxln+l Je<lxl <1 OXj

= r e-ixt; Xjn dS- r e-ixt; Xjn dS + r i~je-ixt;lxl1-ndx

Jlxl=1 lxl Jlxl=< lxl Je<lxl<1
= r Xje-ixt;dS- en r Xje-ixt;dS + i~j r e-ixt;lxl 1-ndx.
Jlxl=1 Jlxl=< }<<lxl <1
(2.93)
As is easily seen

lim E-n { Xje-ixt;dS = lim { CTje-iart;dCT = { CTjda = 0. (2.94)

<--->0 Jlxl=< e--+0 J~ J~
With the aid of an orthogonal transformation which maps~ to (1~1, 0, ... , 0)

r Xje-ixt;ds = i_!_ r e-ixt;ds = i_!_ r cos(x~)dS

Jlxl=1 O~j Jlxl=1 O~j Jlxl=1
= i 0~. { cos(l~lx 1 )dS = -i ~~~ { x1 sin(l~lx 1 )dS, (2.95)
<.,3 Jlxl=1 <., Jlxl=1
lim r
<--+O J<<lxl <1
e-ixt;lxl1-ndx =
J1x1 <1
r
e-ixt;lxl1-ndx

= r
Jlxl<1
lxl 1-n cos(x~)dx = r
Jlxl<1
lxl 1-n cos(l~lxt)dx. (2.96)
44 CHAPTER 2. SINGULAR INTEGRALS

From (2.93),(2.94),(2.95),(2.96) it follows that

rIm 1
E->O E<lxl<l
ln+l dx = -1 -1 ( - z
e -ixe I Xj
X - n
. ,- ~j 1 . (I 'I )dB
1<, 1 lxl=l x 1 sm ., x 1

+i~j r
Jlxl<l
lxll-n cos(l~lxl)dx). (2.97)

Hence, if we set

C1 = -1 i (- { X1 sinx1d8 + { lxll-n cosx1dx) ,

- n Jlxl=l Jlxl<l
the equality
.
l 1m1
E->0 E<lxl<l
e
-ixe xi d _
IX ln+l X - cl<,j.:: (2.98)

holds for 1~1 = 1. Next

(2.99)

Evidently
(2.100)

Hence
2.9. RIESZ KERNELS 45

- !:_ (- f x 2 e-ixedS- f e-ixe~___!!_L_dx) (2.101)

- ~j Jlxl=l j Jlxl>l OXj lxln+l

uniformly on any compact subset of {~ E Rn; ~i "# 0}. By virtue of

(2.99),(2.100),(2.101)

{ le-~~:e1 dx E C 1 ({~ E Rn;6 ·· ·~n "# 0}), (2.102)

Jlxl>l X

r { -ix~ ___!j_dx- . _!!____ { e-ixe dx (2.103)

JL~"Jo Jl<lxl<JL e lxln+l 2
- O~j Jlxl>l lxln+l
for j = 1, ... , n. Set

/(1~1) = { e-~:~1 dx = { cos~:~) dx = { cos(l!l~l) dx.

Jlxl>l lxl Jlxl>l lxl Jlxl>l lxl
Then f(IW is of class C 1 in {~; 6 ... ~n -t 0}. Hence f E C 1 ((0, oo)), and
f(l~l) is continuously differentiable in Rn \ {0}. In view of (2.103)

.
lIm 1
JL->00 l<ixJ <JL
e -ix~ I Xj
X
ln+l d x- a
2£lC. f(ICI)-
- . . ~j !'(ICI)
'> - 2-lcl
U<,J
'> •
':.

Hence if we set c2 = if' (1) we see that

iim f e-ixe I xl~+l dx = c2 ~i (2.104)
JL->OO }l<lxl<JL X

holds when 1~1 = 1. From (2.98) and (2.104) it follows that

lim { e-ix~___!!_L_dx = c3~· (2.105)

e->O,JL->oo Je<lxl<JL lxln+l 3

holds when 1~1 = 1 where c3 = c1 + c2. Since the left hand side of (2.105) is
a homogeneous function of ~ of degree 0 we have

lim
e->O,JL->oo
1 <I xi <JL
. X·
3 -dx
e- 2 x~--
lxln+l
= c3 ...1...
~·

1~1
(2.106)

when~"# 0. In order to obtain the value of c3 we let j = 1, ~ = (1, 0, ... , 0).

Then

(2.107)

(2.108)
46 CHAPTER 2. SINGULAR INTEGRALS

Since
.
hm
t-+O,IJ.-+oo
1 J1. •
sm ra1 1r .
- - d r = -2 s1gna1,
r

we get from (2.108)

If we denote the area of the unit sphere of K' by on,,

1 a1do- =On {
}o
1
t(1- t 2 )(n-J)j 2 dt

11
a1>0

1
=-On (1- s)(n-3)/2ds = o
_n_.
2 0 n-1
Hence
. On 1ri 27r(n-1)/2 7r(n+1)/2i
ca = -7r'l-- = --- = - .
n -1 n- 1 f((n -1)/2) f((n + 1)/2)
Thus the Fourier transform of the Riesz kernels are

(2.109)

Let f E L 2 (Rn) and

gj(X) =-lim
t-+O Jix-yi>t
r Rj(X- y)f(y)dy = -(Rj * f)(x).
Then in view of (2.109)
n n
(:FLRi * gj)(~) = (27r)n/2 L(:FRj)(~)(:Fgj)(~)
j=1 j=1
n
= -(27r)n L(:FRj)(~)2(:Ff)(~) = (:Ff)(~).
j=1

Hence
n
- L Rj * (Rj *f) = f. (2.110)
j=1

By virtue of Theorem 2.3 we see that (2.110) also holds for f E LP(Rn), 1 <
p< oo. Thus we have established the following theorem.
2.10. CASE OF EVEN KERNELS 47

Theorem 2. 7 The Fourier tmnsforms of the Riesz kernels are expressed

as (2.109). The equality (2.110) holds for f E LP(Rn), 1 < p < oo.
Riesz kernels are used to transform even kernels to odd kernels so that we
are able to establish the boundedness of operators with general kernels with
the aid of Theorem 2.6.

2.10 Case of Even Kernels

In this section we make the following assumption.
Assumption 5 K(x) is an even function satisfying Assumption 1 and

(2.111)

Here
logx x>l
log+x ={ 0 (2.112)

Lemma 2.12 For a> O,b > 0

ab::; a log+ a+ eb- 1 • (2.113)

Proof. Let ¢(x), 'lj;(y) be functions defined by

logx+ 1 x>l ey-1 y>l

¢(x) = { 'lj;(y) = {
1
0 o::;y::;1
Since 'ljJ is the inverse function of ¢ in the region x > 1, y > 1, we get
following the proof of Lemma 1.1 that

ab::; la ¢(x)dx + lb 'lj;(y)dy.

The inequality (2.113) follows from this inequality and

loa ¢(x)dx =a log+ a, 1b 'lj;(y)dy::; eb- 1 •

Lemma 2. 13 Let K(x) be a kernel satisfying Assumption 3. Suppose f

is a measumble function with compact support in Rn satisfying

L. lf(x)llog+ lf(x)ldx < oo. (2.114)

48 CHAPTER 2. SINGULAR INTEGRALS

If S is a measurable set of Rn with JsuppjJ < JSJ < oo, where suppf stands
for the support of j, then there exist constants C3, C4 SUCh that forE> 0

(2.115)

where fe is the function defined by (2.4).

Proof. Since

r IJ(x)Jdx =
}Rn
1
lfl>e
lf(x)Jdx + 1lfl~e
lf(x)Jdx

s r lf(x)Jlog+ IJ(x)Jdx + eJsuppfJ, (2.116)

}Rn
we see that f E L 1 (Rn). We begin with the case f 2 0. Set

Es = {x; J.f.(x)J > s}, E~ = Es ns.

Let f31 ,j3f be the functions defined in section 6 and s0 = f3J(JSJ). Then by
Lemma 2.2

r lie(x)Jdx =
Js Jo
to IE~Jds s JSJso + 1 00

so
IEslds. (2.117)

In view of Lemma 2. 7

JSJso = JSlf3J(JSJ) = l iSI

j*(s)ds S
1R"
f(x)dx. (2.118)

From Lemmas 2.2 and 2.8 we get IDs IS j3f(s). In view of this and Lemma
2.4

1so
00
JEsJdssC1 roo~
Jo S
r [j(x)J~dxds+C2 rooj3f(s)ds,
JR, Jso
(2.119)

roo
}o S
~ r [j(x)j~dxds = r
}Rn }Rn }o
roo S
~ [j(x)]~dsdx

= r [ rf(x) ds + roo ~ f(x) 2 ds] dx = 2 r f(x)dx. (2.120)

}Rn }o }f(x) S }Rn

Let s 1 = f3J(t 1 ) be the number defined in section 6. If we make the change

of the variable t = j3f (s), then noting s = f3J(t), j3f (so) = JSJ, j3f (s1) = t1 S
JsuppfJ < JSJ we get

1 00
j3f(s)ds = 1St j3f(s)ds = ltl tdf3 (t) 1
so so lSI
2.10. CASE OF EVEN KERNELS 49

l r + 1 ,6j(t)dt
iS I [ t lSI
] it
= [t,6J(t)Ji11 + ,6j(t)dt = j*(s)ds
1 }o lSI t1

t1 liS! 11SI
=l j*(s)ds- f*(s)ds+ ,61 (t)dt.
0 h

Since J*(s) = ,61 (s) = ,61 (t 1 ) for 0 < s :::; h, the last side of the above
equalities

{lSI {lSI {lSI

= Jo ,6J(t)dt- Jo j*(s)ds:::; Jo ,61 (t)dt

=l iSI 11t
-
t 0
f*(s)dsdt
0
dt
= liSI j*(s) 1ISI -ds
8 t
lSI II lSI (I 1)1/2
= 1 f*(t)log ~ dt=21 j*(t)log ~ dt.

With the aid of Lemma 2.12 we can show that this does not exceed

2tl [f'(t)log+ f'(t) +exp (log C~'r' -I)] dt

: :; 21 oo
f* (t) log+ f* (t)dt + ~ 1
lSI (1 1)1/2
~ dt

= 2 { f(x) log+ J(x)dx +~lSI.

}R" e
From this inequality and (2.116),(2.117),(2.118),(2.119),(2.120) we obtain

ls lfoo(x)ldx

:::; (1 + 2C1 + 2C2) L,. f(x) log+ f(x)dx + ( (1 + 2C1)e + 4 ~2 ) lSI,

In the general case dividing f into its positive and negative parts we can
show that (2.115) holds with

C3 = 1 + 2C1 + 2C2, C4 = 2(1 + 2C1)e + 8C2fe.

Lemma 2.14 For a> 0,,6 > 0

+ log+ ,6,
log+ (a,6) :::; log+ a (2.121)

(a+ ,6) log - 2 - :::; a log + a+ ,6log + ,6.

+a+,6 (2.122)
50 CHAPTER 2. SINGULAR INTEGRALS

Proof. If af3 :::; 1, (2.121) clearly holds. If af3 > 1,

log+(af3) = log(af3) =log a+ logf3:::; log+ a+ log+ (3.
The inequality (2.122) follows from the fact that x log+ x is a convex func-
tion.
Lemma 2. 15 Suppose that K(x) is a kernel satisfying Assumption 3. If
f is a function satisfying the hypotheses of Lemma 2.13, then ], converges
in Lloc(Rn) as E---. 0.
Proof. Let S be a bounded measurable set with lsuppfl < lSI. Fork > 0
set
f(x) lf(x)l 5: k
{
[f(x)]k = ksignf(x) lf(x)l > k
For 0 < D < 2 choose k so that

Ln If- !J]kjlog+ If- [f]kldx < ~' (2.123)

r o
}Rn If- !J]kldx < 4log+(2/b).
1
(2.124)

With the aid of Lemma 2.14 and (2.123),(2.124)

r
}Rn
lf-!f]kllog+lf-!f]kldx-5,
D/2 Dj2 }Rn D/2
r
lf-[f]kllog+lf-[f]kldx

r
+ }Rn
1f -b/2[f]k 1 log+ -gdx
2
< 1. (2.125)

Let g be a function in CJ(Rn) such that

lsuppgl <lSI, Ln lg- [f]kl 2dx < (~) 2

•

Noting a log+ a:::; a 2

r
}Rn
lg-[f]kllog+lg-[f]kldx<
D/2 D/2 - }Rn
lg-[f]kl2dx<l.
D/2
r (2.126)

Applying (2.122) to a= If- !f]kl/(b/2), f3 = l[f]k- gl/(b/2)

j
lf0 2gllog+ If~ gl 5: (a+ (3) log+ a ; f3 5: a log+ a+ (3log+ f3
< If- [f]kllog+ If- [f]kl + l!f]k- gllog+ l!f]k- 91. (2. 127)
- b/2 b/2 b/2 b/2
2.10. CASE OF EVEN KERNELS 51

Seth= f -g. From (2.125),(2.126),(2.127)

r
}Rn 0/2
~log+ 0ldx < 2.
0
(2.128)

As was shown in the proof of Theorem 2.3 g,(x) converges uniformly as

E---+ 0. Since lsupphl <lSI we get from Lemma 2.13 and (2.128)

ls l~f Idx ~ c3 Ln '~'log+ 1~1 dx + C41SI ~ c3 + C4ISI.

Therefore

fs1J,-f,,ldx~ /slg,-g,,ldx+ /slh,ldx+ /slh,,ldx

~ ls lg, - g,, ldx + 26( C3 + C4ISI).

Thus we conclude

In what follows we suppose that K(x) is a kernel satisfying Assumption

5. Let ¢(t) be a continuously differentiable function satisfying

~
0 ~ t ~ 1/4
¢(t) ={
t > 3/4
Let R(x) = (R 1 (x), ... , Rn(x)) be the Riesz kernels. Set

K 1 (x) =lim lim { R(x- y)K(y)dy, (2.129)

<---+O I5---+0 J1x-yl ><,I vi >15
K 2(x) = lim { R(x- y)K(y)¢(1yl)dy. (2.130)
<---+O Jlx-yl><

Lemma 2.16 The right hand side of(2.129) converges in L[0 c(Rn \ {0})
and that of (2.130) in Lioc(Rn). K1 and K2 are both odd functions, and
K1 is homogeneous of degree -n. If K E Lq(E) for some 1 < q < oo, then
K1 E U(E) and for some constant Cq

(2.131)
52 CHAPTER 2. SINGULAR INTEGRALS

Proof. Suppose 1/2:::; lxl :::; 1, 0 < E < 1/4. Then

{ R(x- y)K(y)dy = lim { R(x- y)K(y)dy

Jlx-yi>< 6-+ 0 Jlx-yi>•,IYI>6

= lim { R(x- y)K(y)dy + { R(x- y)K(y)dy

6 -> 0 }6<IYI<1/4 Jlx-yl>•,1/4<lyl<2

+ { R(x- y)K(y)dy = h + h +h. (2.132)

}IYI>2

Since for lxl ~ 1/2, IYI < 1/4

IR(x- Y)- R(x)l:::; CIYIIxl-n-1, (2.133)

we have

1111 =I r }lyl<1/4
(R(x- y)- R(x))K(y)dyl

:::; I 1~+1
x
r
}IYI<1/4
IYIIK(y)ldy = I 1~+1
X J~
IK(o-)ldo-. r (2.134)

Noting lx- Yl ~ IYI/2 if 1/2:::; lxl :::; 1, IYI > 2, we get

lhl :::=; 1 IYI>2

I
X-
c
y
In IK(y)ldy :::=; C
1
IYI>2
IK(y)l
- 1 - 1n-dy.
y
(2.135)

If we set
1/4 < lxl < 2
f(x) = { :(x)
otherwise
then

r
Jn,.
IJ(x)pog+ lf(x)ldx = r
J1/4<lxl<2
IK(x)llog+ IK(x)ldx

= h1~
2

1/4
dt
IK(o-)llog+ (eniK(o-)1) -do-
t

:::; h1~ 1/4

2 dt
IK(o-)1 (log+ IK(o-)1 +log+ en) -do-
t

= h ~
IK(o-)llog+ IK(o-)loo 1 h
2
-dt +
1/4 t ~
IK(o-)loo
2

1/4
log+ en_
dt
t
1
:::; C ~ IK(o-)llog+ IK(o-)ldo- + C ~ IK(o-)ldo- < oo.
2.10. CASE OF EVEN KERNELS 53

Hence by Lemma 2.15 h converges in Lfoc(Rn) as f ---+ 0. Consequently

we see that (2.132) converges in L 1 ({x; 1/2 ::; lxl ::; 1}). Next suppose
1::; lxl ::; 2. Then 1/2::; lx/21 ::; 1 and

{ R(x- y)K(y)dy = 2~ { R (~2 - y) K(y)dy.

Jlx-yi>E Jlx/2-yi>E/2
Hence we see that (2.132) converges in L 1 ({x; 1::; lxl::; 2}). Repeating this
process we conclude that (2.132) converges in Lloc(Rn \ {0}). It is easy to
show that K 1 is homogeneous of degree -n.
Suppose K E U(E), 1 < q < oo. Then in view of (2.134),(2.135)

r
Jl/2<lxl<l
IIllqdx, r
Jl/2<lxl<l
lhlqdx::; c r IK(o-Wdo-.
JE
(2.136)

By virtue of Theorem 2.3 or 2.6

r
JR"
II2Iqdx::; Cq r
jl/4<lxl<2
IK(xWdx = Cq r IK(o-Wdo-.
JE
(2.137)

From (2.136),(2.137) it follows that

r
Jl/2<lxl<l
IKl(xWdx::; Cq rIK(o-Wdo-.
JE
From this inequality and the homogeneity of K 1 the inequality (2.131) fol-
lows.
Finally we investigate the convergence of (2.130). Let N be an arbitrary
positive number. If lxl < N, 0 < f < 1, then

{ R(x- y)K(y)¢(1yl)dy
Jlx-yi>E

= { R(x- y)K(y)¢(1yl)dy + { R(x- y)K(y)dy

Jlx-yi>E,IYI<N+l }IYI>N+l
= I4 +Is.
Since the function defined by
K(x)¢(1xl) lxl <N+ 1
f(x) ={
0 lxl >N+1
satisfies the assumption of Lemma 2.15, I4 converges in Lloc (Rn) as f ---+ 0.
We have

lis I::; r
}IYI>N+l
I cYIn IK(y)ldy
X-

::; c r
}lyi>N+l
(NI ~n1 )n IK(y)ldy::;
Y
eN r IK(o-)ldo-.
}E
54 CHAPTER 2. SINGULAR INTEGRALS

Thus we conclude that (2.130) converges in Lloc(Rn).

Lemma 2. 17 There exists a constant C such that for !x! ;::: 1

IK1(x)- Kz(x)!::; Clxl-n- 1 h !K(o-)!da. (2.138)

There exists a function G(x) which is homogeneous of degree 0 and integrable

on :E such that !K2 (x)!::; G(x) for !x!::; 1. If K E U(:E) for some 1 < q <
oo, then
(2.139)

for some constant Cq, and (2.130) holds in the strong topology of Lq(Rn).
Proof. In view of Lemma 2.16 K1 E Lloc(Rn \ {0} ), Kz E Lloc(Rn), and in
Lloc(Rn \ {0})

K1(x)- K 2 (x) =lim

€->O
1 lx-yl>£
R(x- y)K(y)(1- 4>(!yl))dy. (2.140)

Suppose lxl > 1. Then if IYI < 3/4

!R(x- y)- R(x)!::; Clyllxl-n- 1.
Hence

IK1 (x)- K 2 (x)! = IL. (R(x- y)- R(x))K(y)(1- 4>(!yl))dyl

::; Clxl-n- 1 { !YIIK(y)!dy::; C!x!-n- 1 { !K(o-)!da.

}IYI<3/4 }r:;
Thus (2.138) is established.
If lxl ::; 1/8,

!Kz(x)! = r R(x- y)K(y)¢(!yl)dyl

I}IYI>1/4
::; C { !x- Y!-n!K(y)!dy
}IYI>1/4

::; C { !Y!-n!K(y)!dy::; C { !K(o-)!do-. (2.141)

}IYI>1/4 }r:;
If 1/8 ::; lxl ::; 1, letting x be the characteristic function of the interval [0, 1]

1lx-yl>£
R(x- y)K(y)</>(!yl)dy =¢(!xi) 1 lx-yl>£
R(x- y)K(y)dy
2.10. CASE OF EVEN KERNELS 55

+ { R(x- y)K(y)(</>(lyl)- </>(lxl))dy

Jlx-yl>e
= ¢(1xl) { R(x- y)K(y)dy
Jlx-yl>e

+ { (R(x- y)- x(lyi)R(x))K(y)(</>(lyl)- </>(lxl))dy

Jlx-yl>e
+R(x) { x(lyi)K(y)(</>(lyl)- </>(lxl))dy. (2.142)
Jlx-yl>e
If 0 < E < 1/16

lx-yl>e x(lyi)K(y)(¢(1yl)- </>(lxl))dy

= f K(y)(</>(lyl)- ¢(1xl))dy
}IYI<1/16
+ { K(y)(</>(lyl)- </>(lxl))dy
J1x-yl >e,1/16< IYI <1
= f K(y)(</>(lyl)- ¢(1xl))dy
Jlx-yl>e,1/16<IYI<1
~ { K(y)(¢(1yl)- </>(lxl))dy =0
}1/16<IYI<1
as E~ 0. Hence in view of (2.142) if 1/8 :::; lxl :::; 1

K2(x) = ¢(1xi)K1 (x) + Ln (R(x- y)- x(lyi)R(x))K(y)(</>(lyl)- </>(lxl))dy.

(2.143)
If 1/8 :::; lxl :::; 1, IYI :::; 1, then lx - Yl :::; 9lxl and hence
IR(x- y)- x(lyi)R(x)l = IR(x- y)- R(x)l:::; CIYIIx- Yl-n-I,
IR(x- y)- x(lyi)R(x)l:::; IR(x- y)l + IR(x)l:::; Glx- YI-n·
Therefore

IR(x- y)- x(lyi)R(x)l

:::; C(IYIIx- Yl-n- 1) 112 lx- Yl-n/ 2 = CIYI 112 IX- Yl-n- 1/ 2 (2.144)

if 1/8 :::; lxl :::; 1, IYI :::; 1. If 1/8 :::; lxl :::; 1, IYI > 1, then lx- Yl :::; 2lyl, and
hence

IR(x- y)- x(lyi)R(x)l = IR(x- y)l

= Clx- YI-n:::; Clyl1/21x- Yl-n-1/2. (2.145)
56 CHAPTER 2. SINGULAR INTEGRALS

Combining (2.144) and (2.145) we get that if 1/8::; lxl::; 1

IR(x- y)- x(lyi)R(x)l::; CIYI 112 Ix- Yl-n- 112 .
Therefore in view of (2.143)

IK2(x)l::; IK1(x)l + r IYI 112 IK(y)llx- Yl-n+l/ 2dy

c }Rn
::; 8nlxlniK1(x)l + 8n-1Cixln-1 r IYI1/21K(y)llx-
}Rn
Yl-n+l/2dy

(2.146)
holds for 1/8 ::; lxl ::; 1. Now we show that

r
}1/2<lxl<3/2
r
}Rn
IYI1/21K(y)llx- Yl-n+1/2dydx < ()(). (2.147)

From this result it follows that the last integral of (2.146) is finite for almost
every x. We have

r r
}1/2<lxl<3/2 }1/4<IYI<3
IYI 112 IK(y)llx- Yl-n+l/ 2dydx

r <3 IYI1/21K(y)l Jlx-yl<9/2

::; }1/4<IYI r lx- Yl-n+1/2dxdy

::; c r IYI 112 IK(y)ldy < oo,

}1/4<IYI<3
r r
}1/2<lxl<3/2 }IYI<1/4
IYI1/21K(y)llx- Yl-n+1/2dydx

::; 4n-1/2 r}1/2<lxl <3/2

dx r
}IYI <1/4
IYI1/21K(y)ldy < oo,

r r
}1/2<lxl<3/2 }iyl>3
IYI 112 IK(y)llx- Yl-n+l/ 2dydx

::; 2n-1/2 r.
J1/2<lxl<3/2
dx r
JIYI>3
IYI1-niK(y)ldy < ()().

Combining these three inequalities we obtain (2.147). If we set

G(x) ~.Co ( /, IK(o)lda + lxi"IK,(x)l

+lxln- 1 ln IYI 112 IK(y)llx- Yl-n+ 112 dy) (2.148)
2.10. CASE OF EVEN KERNELS 57

with some suitable constant C0 , then G(x) is homogeneous of degree 0 and

by virtue of (2.141),(2.146) IK2(x)l ~ G(x) holds for Jxl ~ 1. Furthermore
by (2.147) we have
( G(x)dx < oo, (2.149)
J1/2<JxJ <3/2
from which G E £ 1 (~) follows.
Finally suppose K E U (~) for 1 < q < oo. In order to show (2.139) it
suffices to investigate the last term in the bracket of (2.148). Integrating
both sides of

r IYI 112 IK(y)Jix-

(}JyJ<l/4 Yl-n+ 112 dy)q

: :; (}JyJr <1/4 1Yil/21K(y)Jdy) q-1 }JyJr <1/4 IYI1/21K(y)Jix- Yl(-n+1/2)qdy

over 1/2 < JxJ < 3i2 we get

r (r
}1/2<JxJ<3/2 }JyJ<1/4
IYI1/21K(y)JJx- Yl-n+1/2dy)q dx

~ (}JyJ<1/4
r IYI1/21K(y)Jdy)q J1/4<Jxl<7/4
r Jxl(-n+1/2)qdx

:::; C (liK(cr)Jdcr) q ~ C lJK(crWdcr. (2.150)

Since for 1/2 < Jxl < 3/2

r IYI1/21K(y)JJx-
(}1/4<JyJ<3 Yl-n+l/2dy)q

~ ( (1YI 112 JK(y)J)q Jx- Yl-n+ 112 dy

}1/4<JyJ<3

X ( ( Jx- yJ-n+1/2dy) q-1

}1/4<JyJ<3

~ r
}1/4<JyJ<3
(1Yil/21K(y)l) q lx- yrn+1f2dy ( r
}JyJ<9/2
IYI-n+1/2dy)q-1

we have

( ( ( JyJ1f2JK(y)JJx _ yJ-n+1/2dy)q dx
J1/2<lxJ<3/2 Jl/4<JyJ<3
58 CHAPTER 2. SINGULAR INTEGRALS

r
: :; Jl/4<lvl<3 (1Yi 112 IK(y)i)q dy ( r
Jlvl<9/2
IYI-n+l/ 2dy)q

:::; C h IK(aWda. (2.151)

Analogously

r
Jl/2<lxl<3/2
(r Jlvl>3
IYI1/21K(y)ilx- Yi-n+l/2dy) q dx

:::; c r
Jl/2<lxl<3/2
(rJlvl>3
IYil-niK(y)jdy)q dx

:::; C (l IK(a)jda) q:::; hIK(aWda. C (2.152)

Combining (2.150),(2.151),(2.152) we conclude (2.139).

Lemma 2.18 For 0:::; f E £P(R), 1 < p < oo,

1 r j(t + s)sn-lds)p dt:::; ( ~ 1 )p 1 j(t)Pdt,

00 00
(sup f-n (2.153)
-oo e>O Jo P -oo

1 -oo
00
(supf1
e>O e
00
J(t+s)s- 2ds)P dt:::; (
P
~
1
)Pjoo f(t)Pdt.
-oo
(2.154)

Proof. For f > 0 define the functions Fe and G by (2.73). Then we have

E-n ie f(t + s)sn- 1ds:::; c 1 ie f(t + s)ds = Fe(t):::; G(t). (2.155)

With the aid of Lemma 2.10 and (2.155) we obtain (2.153). Since EFe(t):::;
fl-l/pilfllv, we have
00
f 1 f(t + s)s- 2ds = f 100 d (sF8 (t)) s- 2ds
e ds

= f [s- 1F8 (t)]:' + 2f 1 00

s- 2F8 (t)ds

= -Fe(t) + 2f 1 00
s- 2F8 (t)ds:::; 2EG(t) 1 00
s- 2ds = 2G(t). (2.156)

The inequality (2.154) follows from Lemma 2.10 and (2.156).

Theorem 2. 8 Suppose K(x) is a kernel satisfying Assumption 5. For
f E £P(Rn), 1 < p < oo, set
}e(x) = { K(x- y)J(y)dy.
Jlx-yl>e
2.10. CASE OF EVEN KERNELS 59

Then there exists a constant Cp such that

(2.157)

The singular integml K * f = lim,_. 0 j, exists both almost everywhere and

in the strong topology of LP(Rn). The mapping f t--t K * f is a bounded
linear opemtor from V(Rn) to itself

Proof. Let K 1 (x), K 2 (x) be the functions defined by (2.129),(2.130) respec-

tively. Let g = (g 1 , ••• , 9n) be the set of functions defined by

g(x) = -(R * f)(x) = -lim f R(x- y)f(y)dy,

i--->O Jlx-yl >•

where R( x) = ( R 1 ( x), ... , Rn (x)) are the Riesz kernels. Then in view of
Theorems 2.6 and 2. 7

almost everywhere and in the strong topology of LP(Rn). Now we show

~n K(x- y)¢ ( jx ~ yj) J(y)dy =E-n ~n K2 (X~ Y) g(y)dy, (2.158)

where ¢ is the function in the definition of K 2 (x). First suppose g E

C6(Rn)n and suppg C {x; \x\ ::; N}. In {y; \y\ ::; N + 1}

f R(y- z)g(z)dz = f R(y- z)(g(z)- g(y))dz

}ly-zl>6 }ly-zl>6

converges to f(y) uniformly as {j--. 0. If jyj > N + 1, 0 < {j < 1, we have

60 CHAPTER 2. SINGULAR INTEGRALS

Hence

lim { K(x- y)¢ ( lx- Yl) { R(y- z)g(z)dzdy

ti-+0 }Rn E }ly-z/>ti

= Ln K(x- y)¢ ex~ Yl) f(y)dy. (2.160)

If IY - zl > 6, we have for some constant C0 depending on 6

IK(x- y)¢ ( lx ~ Yl) R(y- z)g(z)l

::::; CoiK(x _ y)l¢ ( lx- Yl) lg(z)l . (2.161)
E (IY- zl + 1)n

Since for lzl ::::; N

(N + 1)(1Y- zl + 1) ~ IY- zl + N + 1 ~ IYI-Izl + N + 1 ~ IYI + 1,

the right hand side of (2.161) does not exceed

Co(N+1)niK(x-y)1¢(1x-yl) lg(z)l .
E (IYI + 1)n

By virtue of (2.159) this is integrable in (y, z) ERn x Rn. Therefore in view

of Lemma 2.16 the left hand side of (2.160) is equal to

lim
ti-+0
j"}ly-zl>ti
{ K(x- y)¢ ( lx- Yl) R(y- z)dyg(z)dz
E

= lim j" r K(Ey)¢(1yi)R(x- EY- z)Endyg(z)dz

ti-+O Jlx-<y-zl>ti

= E-n lim Jr
ti-+O
r
Jl(x-z)/<-yl>ti/<
K(y)¢(1yi)R (X- E
z - y) dyg(z)dz

=E-n Ln Kz (X~ z) g(z)dz.

Thus we see that (2.158) holds if g E CJ(Rn)n. Furthermore from the proof
it is clear that (2.158) holds if we define f = R * g for a given g E CJ (Rn )n.
In the general case we choose a sequence {gk} C CJ(Rn)n so that
00

2::.:: llgk+l - gkiiP < 00

k=l
2.10. CASE OF EVEN KERNELS 61

and put fk = R*gk. Then by virtue of Theorem 2.3 or 2.6

00

llfk -fliP - t 0, 2::::: llik+l- !kllp < oo.

k=l
If we set

lgl(x)l + L
00

g(x) = lgk+l(x)- gk(x)l,

k=l

= lh(x)l + L
00

](x) lfk+l(x)- !k(x)l,

k=l
then g, J E LP(Rn), lgk(x)l :::; g(x), lfk(x)l :::; f(x),gk(x)---> g(x), fk(x)--->
f(x) almost everywhere. In view of Lemma 2.1

r IK(x- y)l¢> (lx- Yl) j(y)dy:::; Ax-y/></4

}Rn E
r IK(x- y)IJ(y)dy < 00

almost everywhere. With the aid of (2.138)

The first term of the right hand side of this inequality is finite for almost
every x by virtue of Lemma 1.2 and

The second and third terms are also finite almost everywhere by Lemma 2.1
and 1.2 respectively. Hence applying (2.158) to fk, gk and letting k ---> oo
we see that (2.158) holds in the general case. In view of (2.158)

j<(x) = { K(x- y)f(y)dy = { K(x- y)¢> ( lx- Yl) f(y)dy

Jlx-yl>< Jlx-yl>< E

= ln K(x- y)¢> ex~ Yl) f(y)dy

- { K(x- y)¢> ( lx- Yl) f(y)dy
Jlx-yl<< E
62 CHAPTER 2. SINGULAR INTEGRALS

}nn
(x-
=E-n { K2
E
y) g(y)dy- {
Ax-yl<€
K(x- y)¢ ('x- Yl)
E
f(y)dy

=en r (X-
Jlx-yl>€
Kl
E
y) g(y)dy

+en { (x-
lix-yi>€
(K2
E
y) - K1 (x-
E
y)) g(y)dy

+E-n r (X-
Jlx-yl<€
K2
E
y) g(y)dy
4
- { K(x- y)¢ ( lx ~ Yl) f(y)dy = :2.: Ii.
Jlx-yl<€ i=l

Since K 1 is homogeneous of degree -n

I1 ={ K1(x -y)g(y)dy.
Jlx-yl>€

By virtue of Theorem 2.6

(2.162)

With the aid of (2.138)

Denoting the area of the unit sphere of Rn by On

Hence
2.10. CASE OF EVEN KERNELS 63

Making the change of the variables x = y-ta, 2:;= 1 y1a1 = 0, and applying
Lemma 2.18

00
r (supE1 s- 2 ig(x- sa)ids)p dx

1-oo
}Rn t>O t

= r
00
(sup E1
00
s- 2 ig(y- (t + s)a)ids)p dtdy
} Rn-1 t>O t

: :; (~)P f
p-1 }Rn-1
1oo ig(y- ta)IPdtdy =
-00
(~)P f
p-1 }Rn
ig(x)IPdx.

Hence we obtain

Therefore

Next

lhl:::; en { G(x- y)ig(y)idy

Ax-yi<t
= ~ G(a)en 1t ig(x- sa)lsn- 1dsda

(r
:::; }1:; G(a)da
)1-1/p[r ( r
}1:; G(a) en }o ig(x- sa)lsn- 1ds
)p do-]1/p .
By virtue of Lemma 2.18

Hence

r
}Rn t>O
sup lhiPdx:::; ( - p
p- 1
)p (}1:; r G(a)da)p r
}Rn
ig(x)IPdx. (2.164)
64 CHAPTER 2. SINGULAR INTEGRALS

Finally

1141 =I {
}</4<lx-yl<<
lx- Yl-nK ( 1x = yl) ¢ (lx- Yl)
X Y f
f(y)dyl

: ; (~)nl
f lx-yl<<
IK(Ix=yl)llf(y)ldy
X Y

From (2.162),(2.163),(2.164),(2.165) we conclude (2.157). The proof of the

remaining part is anologous to that of Theorem 2.5.

2.11 General Case

The following is the main theorem of this chapter.
Theorem 2. 9 Suppose that K(x) is a measumble, homogeneous function
of degree -n defined in Rn satisfying

hIK(x)llog+ IK(x)lda < oo, h K(x)da = 0, (2.166)

where :E is the unit sphere of Rn and da is the areal element of :E. For
f E LP(Rn), 1 < p < oo, E > 0 set

}.(x) = 1
lx-yl><
K(x- y)f(y)dy.

Then there exists a constant Cp which does not depend on f such that

{ sup l}.(x)IPdx::; Cp { lf(x)IPdx.

}Rn<>O }Rn
*
The singular integml K f = lim._.o }. exists both almost everywhere and
in the strong topology of LP(Rn), and the mapping f f-1 K * f is a bounded
linear opemtor from LP(Rn) to itself.
Proof. We have only to apply Theorem 2.6 and 2.8 to the odd part and the
even part of K respectively.
2.11. DERIVATIVES OF HOMOGENEOUS FUNCTIONS 65

2.12 Derivatives of Homogeneous Functions

Theorem 2.10 Suppose that Lis a homogeneous function of degree 1-n
belonging to C 1 (Rn \ {0} ). Then fori= 1, ... , n, DiL(x) = (ajaxi)L(x) is
a homogeneous function of degree -n and satisfyies (2.2). For 1 < p < oo
there exists a constant Cp such that for f E CJ (Rn)

(2.167)

Proof. Clearly DiL is homogeneous of degree -n. Let 0 < a < b < oo.
Since

we have

1 a<ixl<b
DiL(x)dx = {
Jlxl=b
L(x) xbi dS- {
Jlxl=a
L(x) Xi dS
a
= 0.
On the other hand using the polar coordinates x = rCJ

1 <lxl<b
DiL(x)dx ={
JE a
1b DiL(rCJ)rn- 1 drdn

= { DiL(CJ)dn
}E
1b
a
dr
r
= { DiL(CJ)dCJ log~.
}E a
Hence we see that DiL satisfies (2.2). We have for f E CJ(Rn)

Di { L(x- y)f(y)dy = Di { L(y)f(x- y)dy

Jlx-yi>< }IYI><
= { L(y)Dd(x- y)dy = - { L(y) !:la f(x- y)dy
}IYI>< }IYI>< vYi
= { L(y)f(x- y) IYil dS + f DiL(y)f(x- y)dy
}iyi=< Y }IYI><
= { L(CJ)f(x- w)CJidCJ + { DiL(x- y)f(y)dy.
JE Jlx-yi><
The first term on the last side converges to JE L(CJ)CJidCJ f(x) as E ____. 0.
The second term converges to DiL *f. In view of Theorem 2.9 we have
IIDiL *flip ::; Cpllfllp· Hence

Di { L(x- y)f(y)dy
}Rn
= j'E{ L(CJ)CJidCJ f(x) + (DiL * f)(x),
and (2.167) holds.
Chapter 3

Sobolev Spaces

3.1 Sobolev Spaces

In this chapter we use the notations D = (D 1, ... , Dn) = (8jax 1 , .•. , ajoxn)
and Da = Df 1 • • • D~", lnl = n 1 + · · · + nn for a vector n = (n1, ... , nn)
with integral components ni ~ 0. We often write Dm to denote mth order
derivatives, i.e. Dm is one of Da with lnl = m. Let 0 be a nonempty open
subset of Rn and m be a nonnegative integer. Then cm(O) denotes the
set of all functions whose derivatives of order up to m are all continuous
in 0, and Clf'(O) the totality of functions belonging to cm(O) and with
compact support in 0. We denote by Bm(O) the set of all functions which
are bounded and continuous in 0 together with their derivatives of order
up to m. For 0 < h < 1 we denote by Bm+h(O) the set of all functions
belonging to Bm(O) whose mth order derivatives are all uniformly Holder
continuous in 0 with exponent h. Similarly the sets Bm(n) and Bm+h(O)
are defined replacing 0 by 0. For u E Bm(O) we put

lulm,oo,n = max sup IDau(x)l, (3.1)

lal=mxEfl
llullm,oo,n = J=O,
. max lulj,oo,fl· (3.2)
... ,m

In particular
llullo,oo,n = lulo,oo,n =sup lu(x)l. (3.3)
xEfl
If a = m + h with an integer m and 0 < h < 1 we set

Iu Ia,oo,n = IDau(x) - Dau(y)l

max sup
lal=m x,yEfl,x'f;y
IX - Y lh (3.4)

iiulla,oo,n = max{llullm,oo,n' lula,oo,n}. (3.5)

67
68 CHAPTER 3. SOBOLEV SPACES

Bm(O.), Bm(O.), Ba(O.), Ba(O) are Banach spaces with norm (3.2) or (3.5).
If 0. = Rn we write I lm,CXJ, II llm,CXJ, I la,CXJ, II lla,CXJ instead of I lm,CXJ,R",
I llm,CXJ,R", I la,CXJ,R", I lla,CXJ,R" respectively.

3.2 Interpolation Inequalities (1)

In this section we prove interpolation inequalities in case 0. = Rn.
Theorem 3. 1 For 0 :::; a < b < c < oo there exists a constant "fa,b,c such
that for u E Bc(Rn) the following inequality holds:

lui b,oo < "' juj(b-a)j(c-a)

- ra,b,c c,oo
juj<c-b)/(c-a).
a,CX) (3.6)

Proof. For the sake of simplicity we write I Ia and "' omitting oo and a, b, c
in I la,CXJ and "fa,b,c respectively.
(i) Case 0 :::; a < b < c:::; 1. Since

ju(x) -u(y)j = (iu(x) -u(y)i)(b-a)/(c-a) (iu(x) -u(y)i)(c-b)/(c-a)

lx - yjb lx - yjc lx - yja
we get if 0 < a, c < 1
iulb :::; juj~b-a)j(c-a) juj~c-b)/(c-a).

For u E B 1 (Rn)

ju(x)- u(y)i = 11 !1
u(y + t(x- y))dtl

= 11 t 1
Diu(y + t(x- y))(xi- Yi)dtl
n
: :; 2:.:: IDiulolxi- Yil:::; iuhvnlx- yj,
i=l

which implies
iuh :::; sup ju(x)- u(y)j :::; vnluh. (3.7)
xi-Y lx- Yi
From this and ju(x) -u(y)i:::; 2julo we see that (3.6) also holds in case a= 0
or c = 1.
(ii) Case 0:::; a< b = 1 < c:::; 2. If n = 1, we have for p > 0

u'(x) =-
11x+p (u'(x)- u'(y))dy + -(u(x
1
+ p)- u(x)),
p X p
3.2. INTERPOLATION INEQUALITIES(1) 69

lulaPa a>O
lu(x + p) - u(x)l :::; {
2lulo a=O
lu'(x)- u'(y)l:::; lulc(Y- x)c-l.

Hence if a> 0

Minimizing the last side with respect to p we obtain

lull :::; 'Yiul~l-a)/(c-a) lul~c-1)/(c-a). (3.8)

The case a = 0 is similarly proved. If n > 1, we obtain the desired result

applying (3.8) to each independent variable.
(iii) Case a= k, b = j, c =mare integers. What should be established is

luli :::; 'Yiul~-k)/(m-k) lulkm-j)/(m-k). (3.9)

We have only to consider the case k = 0, since (3.9) is implied by

ID ku I·J-k -< 'Y IDk u l(j-k)/(m-k)

m-k
IDk l(m-j)f(m-k)
u0 •

Hence we shall prove

(3.10)

for 0 < j < m. First we consider the case j = m - 1:

(3.11)

If m = 2 (3.11) is
11/2 luo
lull:::; Clu2
11/2
, (3.12)
which follows from (ii) with a = 0 and c = 2. Suppose (3.11) is true for
m - 1 in place of m. Then

IUm-2- l(m-2)/(m-1) IUo

I < Cl Um-1 ll/(m-1)
· (3.13)

Applying (3.12) to vm- 2 u

1Dm-2uh :::; qvm-2ui~/21Dm-2ul~/2'

which implies
(3.14)
70 CHAPTER 3. SOBOLEV SPACES

(3.11) follows from (3.13) and (3.14). Finally we consider the general case.
Suppose (3.10) is true with j replaced by j + 1. Then

(3.15)

In view of (3.11)
luJ·J<
-
CJuJi/Ci+l)JuJ
i+1
1/U+ 1)
0 • (3.16)
The desired inequality follows from (3.15) and (3.16).
(iv) Case 0:::; a< b = 1 < c = m, man integer. We proceed by induction.
The case m = 2 is already proved in (ii). Suppose the desired inequality is
true form:
JuJt :::; CJuJ~-a)j(m-a) JuJ~m-1)/(m-a). (3.17)
By the previous step we have

IUIm<- Cl U lm+1
(m-1)/mJ J1/m
U 1 •
The desired result for m + 1 follows from this and (3.17).
(v) Case 0 =a< b = j < c:::; j + 1,j an integer. The case j = 1 is proved
in (ii). Suppose j > 1. Since 1 < c- j + 1 :::; 2 we get by (ii)

JDi-1uJ1:::; CJDi-1uJ~~~+f+1)JDi-1uJ6c-j)j(c-i+1)'

which implies
(3.18)
In view of (iii)
Juli-1 :::; CJuJ)i-1)/iJuJ~Ii.

The desired result follows from this and (3.18).

(vi) Case 0 :::; a < b = 1 < c. If c :::; 2 or c is an integer, the result is already
established in (ii) or (iv) respectively. So we assume that cis not an integer
and c > 2. Put [c] = m. Since

o:::;a<b=1<m<c<m+1

we have (3.17). Since 0 < m- 1 < c- 1 < m we have by (v)

JDuJm-1:::; CJDuJ~~;:-1)/(c-1)JDuJ6c-m)/(c-1)

or
JuJm :::; CJuJ~m-1)/(c-1) JuJ~c-m)/(c-1).
The desired inequality follows from this and (3.17).
(vii) Case 0 :::; a < b = j < c :::; j + 1, j an integer. If j - 1 :::; a, we
3.2. INTERPOLATION INEQUALITIES(l) 71

have 0 :::; a - j + 1 < 1 < c - j + 1 :::; 2, and hence applying the result of
(ii) to ni- 1 u we get the desired inequality. Suppose a < j - 1. If we put
k = [a]+ 1, then 0 < j - k < c- k:::; j - k + 1. Hence, applying the result
of (v) to Dku we get

luli :::; Clul~j-k)/(c-k) lulkc-j)/(c-k). (3.19)

Noting 0:::; a- k + 1 < 1 < j - k + 1 we have by (iv)

lulk :::; Clul)k-a)/(j-a) lul~-k)/(j-a). (3.20)

The desired inequality follows from (3.19) and (3.20).

(viii) Case b = j is an integer. If j -1:::; a, then 0:::; a- j+ 1 < 1 < c- j+ 1.
Hence, we get the desired inequality applying the result of (vi) to Di- 1 u.
The case c:::; j + 1 is shown in (vii). In case a < j - 1, c > j + 1let k, m be
integers satisfying k - 1 :::; a < k, m < c :::; m + 1. Then k < j < m. In view
of (iii) we have

lub :::; Clulg,-k)/(m-k) lulkm-j)/(m-k). (3.21)

Noting 0:::; a- k + 1 < 1 < j - k + 1 and applying (iv) to nk- 1u we get

lulk :::; Clul)k-a)/(j-a) lul~-k)/(j-a) · (3.22)

Since 0 < m - j < c - j :::; m - j + 1 we get

lulm :::; Clul~m-j)/(c-j) lul)c-m)/(c-j) (3.23)

applying (v) to Diu. The desired result follows from (3.21),(3.22),(3.23).

(ix) The general case. The case where b is an integer was established in
the previous step. So we assume that b is not an integer. Let j = [b], then
j < b < j + 1. If a< j and c > j + 1, we have in view of (viii)

luli :::; Clul~j-a)/(c-a) lul~c-j)/(c-a)' (3.24)

lui· < Clul(j+1-a)/(c-a)lul(c-j-1)/(c-a)
J+1- c a ' (3.25)

and in view of (i)

b-j I lj+l-b
Iu Ib-< Iu li+1 (3.26)
uj •

The desired result follows from (3.24),(3.25),(3.26). If a;::: j and c > j + 1,

we have j :::; a< b < j + 1 < c. Hence, it follows from (i) and (viii) that

lui b-< Clul\b-a)f(j+1-a)lul(i+1-b)/(j+1-a)

J+1 a ' (3.27)
luiH1 :::; Clul~j+l-b)/(c-b) lul~c-j-1)/(c-b) (3.28)
72 CHAPTER 3. SOBOLEV SPACES

respectively. The desired result follows from (3.27),(3.28). Analogously, if

a < j and c $ j + 1, we obtain the desired result from

jujb $ Clul~b-j)f(c-j) juj]c-b)j(c-j)' jujj $ Cjuj~j-a)j(b-a) juj~b-j)j(b-a)'

which follow from (i) and (viii). The case a;:::: j, c $ j + 1 is reduced to (i).

3.3 Interpolation Inequalities (2)

l
Let 0 be a nonempty open subset of Rn. Following A. Friedman [65] we
define
(fniujP£tt) lfp O<p<oo
(3.29)
jujp,n = lul-nfp,<Xl,n -()() < p < 0
lulo,<Xl,n p=±oo
When n= Rn' we write I IP instead of I jp,R".

Lemma 3. 1 There exists a constant C such that for -oo < p $ -n, q > 0

lulo,<Xl S Cjul~/(p-q) lul:f(q-p) · (3.30)

Proof. Suppose first p < -n. Then 0 < -njp < 1, and hence

lu(x)- u(y)j $ lul-n/p,<Xllx- Yl-nfp = julplx- yj-nfp.

Therefore

ju(x)j $ ju(x) - u(y)j + ju(y)j $ jujpjx- yj-nfp + ju(y)j. (3.31)

When q ;:::: 1, integrating (3.31) over {y; lx- Yl < p} and applying Holder's
inequality to the last term we get

where Vn and On are the volume of the unit ball and the area of the unit
shere in Rn. Dividing both side of (3.32) by VnPn and taking the supremum
of the left hand side we get

(3.33)
3.3. INTERPOLATION INEQUALITIES(2) 73

Minimizing the right hand side of (3.33) with respect top we obtain (3.30).
When q < 1, the result follows analogously starting from

lu(x)l::; lulvlx- Yl-nfp + lulb~lu(yW,

,
which is a simple consequence of (3.31). If p = -n, we can show (3.30) in
the same manner using (3. 7).
Lemma 3. 2 There exists a constant C such that for -oo < r < p ::;
-n,q > 0
lulr ::; Clul~(r-q)jr(p-q) lul~(r-p)jr(q-p). (3.34)

Proof. Noting that 0 < pjr < 1, 0 < -njr < -njp::; 1 and (3.7)

lu(x)- u(y)l = (lu(x)- u(y)l )p/r lu(x) _ u(y)l(r-p)/r

lx _ yl-nfr lx _ Yl-nfp

lul~/r (2lulo,<Xl)(r-p)/r p< -n

{ (3.35)
::; ( vnlulp)pfr (2lulo,<Xl)(r-p)/r p= -n

We obtain (3.34) combining (3.30) and (3.35).

Lemma 3. 3 For -oo < p::; -n, 0 < q < r the inequality (3.34) holds.
Proof. The result follows from (3.30) and

lulr = ( / lulr dx) /r

1
= ( / lulr-q lulq dx) /r
1

::; lul6:~q)/r ( / lulqdx) 1/r = lul6:~q)jrlul~/r.

Theorem 3. 2 For A. < p, < v, then there exists a constant 'Y>.,J-t,v such
that for u E C0 (Rn)

luh; 1-' ::; 'Y>.,J-t,v luli/~1-')/(v->.) lult~>.)j(v->.) · (3.36)

Proof. (i) Case 0::; A< p, < v. If we put p = p,(v- A.)/A.(v- p,), then
1 < p::; oo,p' = pj(p -1) = p,(v- A.)/v(p,- A.). We obtain (3.36) applying
Holder's inequality in

lui1/J-t =(/ lul1/1-'dx)I-'= (/ lul(v-J-t)/J-t(v->.)lui(J-t->.)/J-t(v->.)dx)I-'.

(ii) Case A.< p, < v::; 0. What is to be proved is
IU I-nJ-t,<Xl -< "/ IU I-n>.,<XJ
(v-J-t)J(v->.) I I(J-t->.)J(v->.)
U -nv,<XJ '
74 CHAPTER3. SOBOLEVSPACES

which is already shown in Theorem 3.1.

(iii) Case -1/n:::; A< f.t:::; 0 < v. We put p = 1/A,r = 1/t-t,q = 1/v.
Then -oo :::; r < p :::; -n, q > 0. The result follows from Lemma 3.1 or 3.2
according as f.t = 0 or f.t < 0.
(iv) Case A< -1/n < f.t:::; 0 < v. From the previous case with A= -1/n
it follows that

IU I1/JL <- Cl U 1n(v-JL)/(nv+1)

-n
IU l(nJL+1)/(nv+1)
1/v • (3.37)

In view of Theorem 3.1

IU I-n =I U I1,oo <- Cl U l(nJL+1)/n(JL-A)I

-nA,oo l(nA+1)/n(A-JL)
U -nJL,oo
_ Cl U l(nJL+l)/n(JL-A)
- 1/A IU l(nA+1)/n(A-JL)
1/JL • (3.38)

Combining (3.37) and (3.38) we obtain (3.36).

(v) Case A < f.t = -1/n < 0 < v. Since -nA > 1, we get by virtue of
Theorem 3.1

IU I-n -_ 1U 11,oo -< ClU ~-1/nA

-nA,oo IU 11+1/nA
O,oo -_ ClU ~-1/nAI
1/A U 1O,oo
1+1/nA
.
By Lemma 3.1
lui O,oo -< Cl U lnvj(nv+1)
-n
IU 11/(nv+1)
1/v •

The desired inequality follows from these two inequalities.

(vi) Case A< f.t < -1/n < 0 < v. By virtue of (v) with A replaced by f.t

I < Cl u1/JL
IU_n_ l(nv+1)/n(v-JL) I ~-(nJL+1)/n(v-JL)
u1jv •

Noting 1 < -nt-t < -nA we get in view of Theorem 3.1

I < Cl u1/A
Iu1/JL- l(nJL+1)j(nA+1) I 1-n(JL-A)/(nA+1)
u_n •

We obtain the desired result from these two inequalities.

(vii) Case -1/n :::; A < 0 < f.t < v. Since 0 < 1/v < 1/t-t, 1/A :::; -n, we
obtain (3.36) with the aid of Lemma 3.3.
(viii) Case A< -1/n < 0 < f.t < v. We have in view of (v) and (vii)

IU I-n <- IU l(nJL+1)/n(JL-A)

c'
1/A IU l-(nA+1)/n(JL-A)
1/Jl '

Iu1/JL-
I < Cl 1n(v-JL)/(nv+1)l
u_n
l(nJL+1)j(nv+1)
u1/v '

respectively. Combining these two inequalities we obtain the desired in-

equality.
3.4. INTERPOLATION INEQUALITIES(3) 75

3.4 Interpolation Inequalities (3)

The purpose of this section is to establish the Gagliardo-Nirenberg inequal-
ity for Rn. We mainly follow L. Nirenberg [118].
Let n be a nonempty open subset of Rn. For 1 ::; p ::; oo and a nonnega-
tive integer m we put
1/p

u _
I lm,p,fl -
{
(

L2:supxEniD,u(x)l
max ess
IDauiPdx
lal=m
)
1::; p < oo,
p= oo,
(3.39)

lal=m
m ) 1/p
l!ul!m,p,n = {
(
~ lu11,v,n 1::; p < oo,
(3.40)
max
k=O, ... ,m
lulk,oo,n p=oo,

if the values of the right hand sides are defined and finite. If 1 ::; p <
oo, lulo,v,n coincides with lulv.n defined by (3.29). For u E Bm(n), lulm,oo,n
coincides with the one defined by (3.1). When n = Rn, we write lulm,p,
l!ul!m,p short for lulm,p,Rn, l!ul!m,p,Rn·
The totality of functions whose distribution derivatives of order up to m
all belong to LP(n) is denoted by wm,v(n) which is a Banach space with
norm (3.40).

Lemma 3. 4 Let 1 ::; p < oo, 1 ::; q < oo, 1/r = (1/q + 1/p)/2, and
I = [a, b] be a finite interval with III = b - a. Then for any function
u E C 2 (I)

(llu'lrdx) 1/r::; IIIH1/r-1/p (llu"IPdx) 1/p

+8III-1-1/r+1/v (fr1ulqdx) 1/q. (3.41)

Proof. Let u = v + iw with real valued functions v and w. Let a =

(b- a)/4, a ::; x1 ::; a+ a, a+ 3a ::; x 2 ::; b. Then by virtue of the mean
value theorem
76 CHAPTER 3. SOBOLEV SPACES

Hence, for a ::; x ::; b

v'(x) = v'(x12) + 1x v"(y)dy = v(x2)- v(xl) + 1x v"(y)dy,

X12 X2 - Xl X12

w'(x) = w'(Yl2) + 1x
Yt2
w"(y)dy = w(x 2)- w(xl)
X2 - X1
+ r
lY12
w"(y)dy.

From these two equalities we get

iu'(x)i::; ju(x2) -u(xl)i + 11x v"(y)dy+i1x w"(y)dy,. (3.42)

X2 - X1 Xt2 Yl2

The first term of the right hand side of (3.42) does not exceed (ju(x2)l +
iu(xl)i)j2a since x2- x1 ~ 2a.
Consider the case x ::; x 12 < y 12 . The second term of the right hand side
of (3.42) is not greater than

11xX12 u"(y)dy+i 1Y12

x w (y)dy
12
I 11

::;1X
iu"(y)jdy +
Xt2 1Y12 lw"(y)jdy::; 1Y12 iu"(y)jdy::; 1b iu"(y)jdy.
X12 X a

Other cases are handled similarly, and we see that the inequality

is true for a ::; x ::; b. Integrating both sides over a ::; x 1 ::; a +a and
a + 3a ::; x 2 ::; b we get

a2iu'(x)i::; ~ 1b iu(y)jdy + a21b iu"(y)jdy. (3.43)

With the aid of Holder's inequality we get from (3.43)

S ~(b-a)1f.+1-1/• (llul'dx) 1/• +<>"(b-a)1/d1-1/p (llu"l'dx) 1/p

Dividing by a 2 and noting 1/q = 2/r- 1/p we obtain (3.41).
3.4. INTERPOLATION INEQUALITIES(3) 77

Lemma 3. 5 Let 1 :::; p:::; oo, 1 :::; q :::; oo, 1/r = (1/q + 1/p)/2. Then for
any function u in C 2 ( [0, oo)) vanishing outside some bounded set

(1= lu'lrdx) l/r:::; 4J2 (1= lu"IPdx)

112
P (1= lulqdx) 112
q (3.44)

Hence, for functions u E C8"(-oo,oo)

(!-==lu'lrdx) 1/r
:::; 4V2
(!=-= lu"IPdx
) (J=-= lulqdx)
l/2p l/2q (3.45)

Proof. First consider the case 1 < p < oo, 1 :::; q < oo. Let l be a positive
number such that u(x) = 0 for x > l, and k be a natural number;::: 2. If
the first term of the right hand side of (3.41) with I = [0, ljk] is greater
than the second term, we let It = [0, a 1 ] = [0, ljk]. Otherwise, we choose
It = [0, a 1 ] so that It :::> [0, ljk] and both terms of the right hand side of
(3.41) with I= It are equal. That is possible since 1 + 1/r -1/p > 0. Then
we have

( l) 1+1/r-1/p ( }Ir lu"IPdx)1/p

u {
1/r 2k
lu'l"dx) :0 4v'2 Ulu"l') 'i':·u, lul'dx) 'i'•
The proof of (3.44) is complete if a 1 2 l. If a 1 < l, we define I2 =
[a1, a2], ... , Ii = [ai-l, ai], ... as follows. Suppose Ii-1 = [ai-2, ai-l] is de-
fined. If the first term of the right hand side of (3.41) with I= [ai-l, ai-l+
ljk] is greater than the second term, we set Ii = [ai-l, ai] = [ai-l, ai-l +
ljk], and otherwise we define Ii = [ai-l, ai] so that Ii :::> [ai-l, ai-l+ ljk]
and both terms of the right hand side of (3.41) with I= Ii are equal. After
proceeding finite times in this manner, we get ai-l < l :::; ai· Denote this
number i by j. Then for i = 1, ... , j

(
1. lu'lrdx
)
1/r

( l) Hl/r-1/p (1 lu"lvdx) 1/p

<{
2 -
k I;
-
4V2 (1. lu"IPdx)
1~

(1. lulqdx) 1~

Since j:::; k, rjp- r < 0, we have ask---+ oo

i (z)r+l-r/p(f )r/p
A:= ~2r k }I, iu"IPdx
78 CHAPTER 3. SOBOLEV SPACES

On the other hand, noting r j2q + r j2p = 1 and using Holder's inequality

Combining these two statements and the inequality

we obtain (3.44). Letting p ---+ 1, we get the result of the case p = 1. If

p = oo or q = oo, we obain (3.44) letting p---+ oo or q---+ oo.

Lemma 3. 6 If n < p < oo, there exists a constant C depending only on

n and p such that for any u E CJ(Rn)

(3.46)

Proof. Let x, y be two distinct points of Rn and d = jx- yj. Set

G = { z E Rn; jz - xj ::; d, jz - yj ::; d}.

Then the volume of G is cdn with some positive constant c depending only
on n. Integrating both sides of

ju(x)- u(y)j ::; ju(x)- u(z)j + ju(z)- u(y)j

with respect to zinG, we get

c~ju(x)- u(y)j ::; !a ju(x)- u(z)jdz +!a ju(z)- u(y)jdz. (3.47)

Expressing z by polar coodinates about x:

z= x + pJ, 0 < p < oo, J = (J 1 , ... , Jn) E E =the unit sphere,

3.4. INTERPOLATION INEQUALITIES(3) 79

we get

faiu(x)- u(z)ldz::::; lad llu(x + pO")- u(x)ldO"pn- 1dp,

lu(x + w) - u(x)l = liP! u(x + tO")dtl

= IJor t•= O"iDiu(x + tO")dtl ::::; Jor IY'u(x + tO")Idt.

1

Hence, if p < oo, with the aid of Holder's inequality

faiu(x)- u(z)ldz::::; lad liP IY'u(x + tO")IdtdO"pn- 1dp

=lad l i p IY'u(x + tO")It(n-1)fpt(1-n)/PdtdO"pn-1dp
: : ; lad (liP IY'u(x + tO")IPtn-1dtdO") 1/p

X (lip t(1-n)/(p-1)dtdO") (p-1)/p pn-1dp

= (p = nn)
P
1
n
(p-1)/p
{ pn-nfp
Jo
d (
{
Jlz-xi<P
IY'u(z)IPdz
) 1/p
dp

p- 1
::::; ( --On
) (p-1)/p ~+1-n/p
/
(1 IY'u(z)IPdz
)
1/p
.
p-n n+l-np Rn

Estimating the second term of (3.47) analogously we obtain (3.46)0

Lemma 3. 7 If 1::::; p < n, then for any u E CJ(Rn)

p(n - 1) rrn 1/n

lulnv/(n-p) ::::; 2( _ ) IDiulv · (3.48)
n P i=1
Proof. We begin with the case p = 1 and prove
lulnf(n-1) ::::; ~IT IDiuli 1n. (3.49)
i=1

Since

u(x) = u(x1, o o o, xn) = /_: Dju(x1, ... , Xj-1, t, Xj+1, ... , Xn)dt
= -1 Xj
00
Dju(x1, oo., Xj-1, t, Xj+1, ... , Xn)dt,
80 CHAPTERS. SOBOLEVSPACES

I:
we have

2lu(x)l:::; 1Dju(x1,···,Xj-1,t,xi+1, ... ,xn)ldt.

Therefore
(2lu(x)l)n/(n-1) = [(2lu(x)l)n]1/(n-1)

: :; IT (loo IDju(x1, ... , Xj-1, t,

n )1/(n-1) .
j=1 -oo Xj+1, ... , Xn)ldt

Integrating both sides with respect to x 1 and applying Holder's inequality

I:I: (21ul)nf(n- 1)dx1dx2:::;}] (I: I: IDiuldx1dx2)

1 1
/(n- )

X Jl (I: I:I: IDiuldx1dx2dXj)

1 1
/(n- ) •

Repeating this process we obtain (3.49). When p > 1, we put v = lul(n- 1)P/(n-p).
Since (n -1)pf(n- p) > 1, v E CJ(Rn) and
IDivl = 1Di(lul 2 )(n- 1)p/ 2 (n-p) I
3.4. INTERPOLATION INEQUALITIES(3) 81

= (n- 1)plul(np+p-2n)/(n-p)IRe(uDi'u)l
n-p
::; ( n - 1)p luln(p-1)/(n-p) IDiul.
n-p
Hence
( )
IDivl1 ::; n- 1 p (
n-p
r
hn lulnpf(n-p)dx )
(p-1)/p (
r
hn IDiuiPdx
) 1/p

- (n -1)pl ln(p-1)/(n-p)ID· I
- n- p u np/(n-p) ~up·

Applying (3.49) to v and noting

IV In/(n-1) = IU lnpf(n-p)
(n-1)p/(n-p)

we obtain (3.48).
Now we prove Gagliardc:rNirenberg's inequality ([67], [68], [118]).
Theorem 3. 3 Let 1 ::; p ::; oo, 1 ::; q ::; oo, and let j, m be integers sat-
isfying 0 ::; j < m. If m- j- njp is not a nonnegative integer, then for
any a satisfying j j m ::; a ::; 1 there exists a constant 1 depending only on
n, m, j, p, q, a such that

IDiulr ::; 1lul~,plul6~a (3.50)

for any u E C[J'(Rn), where

(3.51)

lfm- j- njp is a nonnegative integer, (3.50) holds for a= jjm.

Remark 3. 1 Since a 2: jjm

1 m
-:;: = ;:
(jm ) + p+
-a
a 1-a
-q- ::;
a 1-a
p + -q- ::; 1.
Hence, either r < 0 or r 2: 1. If r 2: 1 (3.50) is
lulj,r ::; llul~,p lul6~a.

If in particular p = q, a= jjm, this reduces to

Iu Ij,p-< I Iu lm,p
j/ml 11-j/m
u O,p • (3.52)
82 CHAPTER 3. SOBOLEV SPACES

Remark 3. 2 Further results in case m- j - njp is a nonnegative integer

will be given later.

Proof of Theorem 3.3. (i) Case a= jfm. This case (3.51) reduces to

!r = i!
mp
+ (1- m !.
q
i)
Since r ~ 1 what should be proved is

_ "f IU lj/ml
IU Ij,r < 11-j/m .
m,p U o,q (3.53)

This is obvious if j = 0, and so we assume 0 < j < m. We begin with the

case j = 1, and show

(3.54)

by induction with respect to m. If m = 2, this reduces to

(3.55)

for 1/r = (Ijp + Ijq)/2. This is due to Lemma 3.5 if n = 1. If n > 1,

writing x = (x1, x'), x' = (x 2 , ••. , Xn), we apply Lemma 3.5 to u(·, x'). If

1:
p < oo,q < oo,

ID1u(x1,x'Wdx1

:::; (4v'2t (1: IDfu(xl,x')IPdxl)r;zv (1: lu(x1,x'Wdx1r;zq

With the aid of Holder's inequality

ID1ul~,r = ln-l1: ID1u(x1,x'Wdx1dx 1

:::; (4J2t (ln IDfuiPdx) rjZp (ln lulqdx) rjZq

We can estimate IDzulo,r, ... , IDnulo,r analogously. The inequalities in case

of p = oo or q = oo are established similarly, and the proof of (3.54) in case
of m = 2 is complete.
Suppose (3.54) is true with m - 1 in place of m. Let

! = _!_! +
r mp
(1- _!_) !,
m q
_!_- _1_! + (1- _I_)!
P1 - m - 1 p m- 1 r·
3.4. INTERPOLATION INEQUALITIES(3) 83

In view of the induction hypothesis

ID U I1,p1 <- CID U 11/(m-1)1D

m-1,p
l(m-2)/(m-1)
U O,r '

or

lui 2,pt <

_
Clul1/(m-1)1ul(m-2)/(m-1)
m,p 1,r · (3.56)

Noting 1/r = {1/p1 + 1/q)/2 we get in view of (3.55)

lul1,r ~ Clu1~::1 1u1~:: · (3.57)

Combining (3.56),(3.57) we obtain (3.54). Thus (3.53) has been proved

when j = 1.
Suppose (3.53) is known to hold for 0, ... , j - 1 instead of j. If we set

!r = imp
!+ (1- i) !,
m q

then

1 j-11 ( j-1)1
r= m - 1 p+ 1 - m - 1 q1 .

Therefore, making use of the induction hypothesis with m - 1 in place of m

ID U I·J-1,r -< CID U l(j-1)/(m-1)1D

m-1,p
l(m-j)/(m-1)
U O,qt '

or

lui·J,r <
-
Clul(j-1)/(m-1)1ul(m-j)/(m-1).
m,p 1,q1 (3.58)

In view of (3.54)

(3.59)

The inequality (3.53) follows from (3.58),(3.59).

(ii) Case a= 1. Since m- j - njp is not a nonnegative integer, we have
1 ~ p < oo. What is to be shown is

(3.60)

for1/r = jfn+ 1/p- mjn. We begin with the case j = 0, i.e. with showing

(3.61)
84 CHAPTER 3. SOBOLEV SPACES

for 1/r = 1/p- mjn assuming m- njp is not a nonnegative integer. If

m = 1, n f= p < oo since 1- njp is not a nonnegative integer. If p > n, then
r = npj(n- p) < 0. In view of Lemma 3.6 we get

lulr = iul-n;r,oo = iuh-n/p,oo : : :; Cluh,v-

If p < n, then r > p. In view of Lemma 3. 7

- - p(n- 1) rrn 1/n p(n- 1)

lulr - lulo,r - iulnv/(n-p) : : :; 2(n _ p) i=liDiulo,p : : :; 2(n _ p) iuit,P"

Hence, we see that (3.61) holds if m = 1.

Suppose (3.61) has been established up tom- 1. Set
1 1 m 1 1 m-1 1 1
- = - - - - =- + -. (3.62)
r p n r1 p n r n

Noting that m-1-n/p is not a nonnegative integer we apply the induction

hypothesis to Du. Then we get

(3.63)

If r > 0, then r > r 1 > p ~ 1. So, with the aid of (3.61) for the case of
m=1weget

(3.64)

We obtain (3.61) from (3.63) and (3.64). Next consider the case r < 0. If
r1 determined by (3.62) is positive, then r1 > n. In view of Lemma 3.6 we
get

lulr = iul-n;r,oo = lull-n/r 1 ,oo :::::; Cluh,rt·

Combining this with (3.63) we obtain (3.61). If r 1 < 0, we set l = [-njr], h =
~njr- [-n/r]. Since -njr = m- njp is not a nonnegative integer, h > 0.
Noting -n/r1 = l- 1 + h we get

The inequality (3.61) follows from this and (3.63). The proof of the case
j = 0 is complete.
3.4. INTERPOLATION INEQUALITIES(3) 85

Suppose j > 0 and (3.60) is known to hold with j replaced by j - 1.

If 1/r = jjn + Ijp- mjn, then 1/r = (j- I)jn + Ijp- (m- I)jn and
(m- I) - (j- 1) - njp is not a nonnegative integer. Therefore, by the
induction hypothesis

Thus, the proof of the case a = 1 is complete.

(iii) Case jjm <a< 1. If we define r 1 , r 2 , 0 by

~
r1
= 1_! +
mp
(1- ~) !, m q
1
-r2 =-+-
n p
j
--,
n
1 m I-a
0=1 -J· m
;'

then by virtue of (i) and (ii)

'D u j I < Cj u lj/mj

rt -
11 -j/m
m,p U O,q ' (3.65)

Since 0 < 0 < 1, 1/r = Ojr 1 + (1- O)jr2 , we have in view of Theorem 3.2

(3.66)

The desired result follows from (3.65),(3.66).

Lemma 3. 8 Let m and j be integers satisfying 0 ::; j < m. Let ¢ be a

function in C(f(Rn) such that ¢ 2:: 0, ¢(x) = 1 if lxl ::; 1/2, ¢(x) = 0 if
jxj > 1. Then, for any function u in cm(Rn) and j3 with 1!31 = j

D 13 u = '2:: Ka * Dau + L * u, (3.67)

lal=m

where

xa)
Ka(x) = (m- 1l)!On ¢(x)Df3 ( lxln (3.68)

and L(x) is a continuous function with support contained in the set

{x; 1/2::; lxl::; 1}.
Proof. Let a E :E. By induction we can show

u(x) = ¢(0)u(x) = (-l)m lo

{'XJ tm-1
(m -1)!¢(ta)
(aat )m u(x -ta)dt
m-
-'"'
1

L..J a
k=O
1 = ( a ) k(tk a ) · u(x- ta)dt
-
at --¢(ta)
k! at
86 CHAPTER3. SOBOLEVSPACES
rXJ tm-1
= Jo (m _ 1)! ¢>(ta) 2: a 0 D 0 u(x- ta)dt
lal=m

-2:
m-11oo
k=O
.J (-{))
2: (k) 0 j=O
k

{)t
k-j tk
1 . -
k.
( {)

{)t
)j+1
¢>(ta)·u(x-ta)dt

roo tm-1
= Jo (m _ 1)! ¢>(ta) 2: a 0 D 0 u(x- ta)dt
lal=m

2:1 2: 1oo (k). "":j"ti 2:

_m- k
a 0 D 0 ¢>(ta) · u(x- ta)dt.
k=O j=O O J J· ial=i+1

Integrating over E

Onu(x) = h1 L: 0
oo
(
tm-1
m
_ )I ¢>(ta)
1.
2:
lal=m
a 0 D 0 u(x- ta)dtda

- 2: 2: (k). "":j"
m-
1

k=O j=O
k

J
11 IYI-n
J· Rn
2:
ial=i+1
Y0 D 0 c/>(y) · u(x- y)dy.

Dividing by On and letting Df3 operate

Df3u(x) = ( _\)'fl 2: f nP (¢>(y)yaiYI-n). D 0

u(x- y)dy
m . n iai=m}Rn
3.4. INTERPOLATION INEQUALITIES(3) 87

Hence (3.67) holds with K" defined by (3.68) and with

By the definition (3.68) of Ka there exists a constant C such that

(3.69)

Next we consider the case m- j - njp is a nonnegative integer. We begin

with the following lemma.
Lemma 3. 9 Let A and B be measumble subsets of Rn, and let G be an
integml operator

(Gf)(x) = LG(x, y)f(y)dy

with kernel G(x, y) which is a measurable function defined in Ax B. Let

1::; p,q, r::; oo and 1/r = 1/p+ 1/q -1. If

for ally E B

for all x E A,

then

(i 1
I(Gf)(xWdx) /r::; K (L lf(y)IPdy)
1
/P.

If q = oo, r = oo or p = oo, the corresponding integmls are understood as

the essential supremum.
Proof. We follow the proof of Lemma 2.6.1 of N. Kerzman [94]. We may
assume G, f 2:: 0. We consider only the case in which p, q, r are all finite,
since the proof of other cases is immediate. We denote the LP norm by
I llv- With the aid of Holder's inequality
(Gf)(x) = L Gqfr fPirc(r-q)/r f(r-p)frdy
88 CHAPTER3. SOBOLEVSPACES

::=; (h Gq fPdy) 1/r (h Gqdy) (r-q)jqr llfll};-p)jr

::=; K(r-q)jr (h G(x, y)q f(y)Pdy) 1/r llfll};-p)/r.

Taking the rth power and integrating over A we get

jjGfjj~ ::=; Kr-q hi G(x,y)qdxf(y)Pdylifll~-p ::=; Krllfll~·

Theorem 3. 4 Let m, j be integers satisfying 0 :::; j < m, and let 1 :::; p <
oo, 1 :::; q < oo. Suppose m- j- njp = 0, or p = nj(m- j). Then for
any a satisfying jfm:::; a < 1 there exists a constant 'Y depending only on
m, j, q, a, n such that for any u E C()"(Rn)
(3.70)
where

~r = in +a(~-
p
m)
n
+q 1- a = (1- a) (i +~) .
nq
(3.71)

Proof. Put s = mnqj(m- j)(n + jq). Then the set of values of r when a
ranges over [j jm, 1) is [s, oo ). So we are going to show (3. 70) for s:::; r < oo
with a defined by (3.71). In view of Theorem 3.3 we know that

lui·J,s-< Cluli/mlul(m-i)/m
m,p O,q (3.72)
is true (the case a= jfm). Here we note that s ~ 1 by Remark 3.1. Since
1/s is a convex combination of 1/p and 1/q, we have
s:::; max{p, q}. (3.73)
Suppose max{p,q} ::=; r < oo. We use Lemma 3.8:

Df3u= L Ka *D 0 u+L*u,
iai=m
where 1,61 = j. Let q1 be a real number defined by 1/r = 1/p + 1/q1 - 1.
Then, we have

1:::; q1 < nj(n- m + j). (3.74)

By virtue of (3.69) and (3.74)

r
}Rn
IKa(xWldx:::; c r
Jlxl$.1
ixi(m-n-j)qldx < oo.
3.4. INTERPOLATION INEQUALITIES(3) 89

Therefore, in view of Lemma 3. 9

JDPuJo,r :::; 2: JK,. * D"'ulo,r + IL * uJo,r :::; C ( 2: ID"'ulo,p + Julo,q) ·

l<>l=m lal=m
Hence, we get
lulj,r :::; C(lulm,p + lulo,q ). (3.75)
Applying this inequality to the function u-\(x) = u(.Ax), where .A > 0, and
noting
Ju-\Jj,r = >.i-nfrJulj,r, Ju-\lm,p = >.m-nfPJulm,p, lu-\lo,q = >.-nfqlulo,q (3.76)
we obtain
lui·3,r <_ C (>.m-nfp-j+nfrJuJ m,p + _A-nfq-j+nfrlul O,q ) .

Minimizing the right hand side with respect to .A we conclude (3. 70).
Next, supposes< r < max{p,q}. Lett be a number such that t ;:::
max {p, q}. Then, by what was just shown above
luli,t :::; Clul~,p lul6~b, (3. 77)
1
-=(1-b)
t
(j-+-.
n
1) q
(3.78)

Since s < r < t we have with the aid of Holder's inequality

. <
IU IJ,T -
IU lt.(r-s)fr(t-s)
J,t
IU ~~(t-r)/r(t-s)
J,S • (3.79)
Combining (3.72),(3.77),(3.79) we get
1 _ bt(r- s) j s(t- r)
lulj,r:::; Clul~,plulo~a, (
a- rt-s ) +- ( ).
mrt-s
In view of (3.78)

j
-+a
n
(1
---
p
m) + - - a) (j- +-
n
1- a = (1-
q n q
1)

= (t(r-s) +s(t-r) -bt(r-s) _j_s(t-r)) (1.+!)

r(t- s) r(t- r(t- s) r(t- s)
s) n m q
= t(r- s) ( 1 _b) (i +!) +s(t- r) ( 1 _ j_) (i +!)
r(t- s) n r(t- s)
q n m q
r-s
-- r(t- s)
+ t-r
- t-s --1
r(t- s) - r(t- s) - r ·

Next consider the case m- j - njp > 0. In the following theorem m-

j - njp need not be an integer.
90 CHAPTER3. SOBOLEVSPACES

Theorem 3. 5 Let m, j be integers satisfying 0 ::::; j < m, and let 1 ::::; p ::::;
oo, 1 ::::; q::::; oo. Suppose m- j - njp > 0. Then, for any a satisfying

! :=;a:=;(~+~) I(:-~+~) (3.80)

there exists a constant 'Y depending only on n, m, j, p, q, a such that

lulj,r ::::; 'Yiul~,vlul6~a (3.81)

for any u E Clf(Rn), where

!
r
= L
n
+a (! - m) + 1 - a.
p n q
(3.82)

Remark 3. 3 The rightmost side of (3.80) is less than 1. (3.80) is the

range of a for which r defined by (3.82) ranges over [s, oo], where s is the
value of r corresponding to a= j jm:

!s = i_!
mp
+ (1- j_) !.
m q

Proof of Theorem 3.5. If we prove (3.81) for r = s and r = oo, then

the case r E (s, oo) can be established following the proof of Theorem 3.4,
where s is the number defined in the above remark. The case r = s is
proved in Theorem 3.3. The case r = oo can be established with the aid of
(3.67),(3.69) and Lemma 3.9.

1) I (m-;;:- - p1+ q1) <

Next we consider the case

j
( ;;: + q a < 1, (3.83)

or
j 1 m 1
-n +---
p n
<- < 0.
r
(3.84)

Lemma 3. 10 Let m be a positive integer and 1 ::::; p ::::; oo, 1 ::::; q ::::; oo.
Suppose m- njp > 0. Then for each integer k satisfying 0::::; k < m- njp
there exists a constant C such that for u E Clf(Rn)
(n+kq)pf(mpq+np-nq) I l(mp-n-kp)q/(mpq+np-nq)
IU l m,p U O,q
1 ::::; p < oo, 1 ::::; q < 00
lulk,oo ::::; C lul~,~q)f(n+mq) lulb:';-k)qf(n+mq) P = oo, 1 ::::; q < 00

kp/(mp-n) I l(mp-n-kp)/(mp-n)
IU l m,p U O,oo 1 ::::; p < oo, q = 00

k/m I l(m-k)/m
IU l m,oo p=q=oo
U O,oo
3.4. INTERPOLATION INEQUALITIES(3) 91

Proof. In view of Lemma 3.8 we have

nf3u = L Ka. * Da.u+ L*u

[a.[=m

for 1,61 = k, and

IKa.(x)l:::; C¢(x)!xlm-n-k.
With the aid of Holder's inequality

!Df3u(x)!:::; L !Ka.lo,v'!Da.ulo,p + ILio,q'lulo,q·

[a.[=m
If 1 < p :::; oo, then

(m- n- k)p' > (-n + njp)p' = -n.

HenceKa. E £P' (Rn). Ifp = 1, thenk:::; m-n-1, andhenceKa. E L 00 (Rn).
Therefore
!Df3u(x)! :::; C(lulm,p + lulo,q)·
Using the argument by which we deduced (3.70) from (3.75) we obtain the
desired result.
Lemma 3. 11 Let m be a positive integer and 1 :::; p :::; oo, 1 :::; q :::; oo.
Suppose that m- njp is a positive integer. Then there exists a constant C
depending only on m, n, p, q such that for u E Clf(Rn) and I,BI = m-n/p-1

!Df3u(x)- Df3u(y)!

:::; C!x- Yl {[ 1 + ( log+ lx _ Yl

2 )(p-l)/pl lulm,p + lulo,q } (3.85)
holds for x, y ERn, where log+ t = logt ift > 1 and 0 if t:::; 1.

Proof. According to Lemma 3.8

Df3u = L Ka. * Da.u + L * u

[a.[=m

for I,BI = m- njp- 1, where

Ka.(x) = ¢(x)Pa.(x)lxl-k, (3.86)

Pa.(x) is a homogenous polynomial of degree 2m- njp- 1, k = n +2m-

2njp- 2, and hence
(3.87)
92 CHAPTER 3. SOBOLEV SPACES

Consequently for x, y ERn

Df3u(x)- Df3u(y) = L
lal=m
1
R"'
(Ka(x- z)- Ka(Y- z)) Dau(z)dz

+{ (L(x- z)- L(y- z)) u(z)dz.

}Rn
Noting that IY - zi :$ 3lx- Yl/2 in the region lx- zi :$ lx- Yl/2, we see if
1 < p :$ oo that

{
IJl:z:-zl~l:z:-yl/2 (Ka(x- z)- Ka(Y- z)) Dau(z)dzl

:$ r
Jl:z:-zl~l:z:-yl/2
IKa(X- z)DO!u(z)idz

+ r
}ly-zl ~3lx-ylf2
IKa(Y- z)DO!u(z)idz
1/p'

:$ ( r
Jl:z:-zl~l:z:-yl/2
IKa(X- z)IP' dz) IDO!ulo,p

1/p'

+( r
}ly-zl91x-yl/2
IKa(Y- z)IP' dz) IDO!ulo,p· (3.88)

Noting (njp- n + 1)p' = -n + p'

1/p'
(
lx-zl~lx-yi/21Ka(X- z)IP' dz )

1/p'

:$ C ( { izi-n+v' dz) :$ Clx- Yi·

Jlzl~l:z:-yl/2
Estimating the last integral of (3.88) analogously we get

{
IJl:z:-zl~l:z:-yl/2 (Ka(x- z)- Ka(Y- z)) Dau(z)dzl :$ Clx- Yiiuim,p·
(3.89)
It is easy to see that this inequality is true also when p = 1. By the same
method as above

{ (Kcx(x- z)- Ka(Y- z))Dau(z)dzl :$ Clx- Yiiulm,p·

}ly-zl~l:z:-yl/2
(3.90)
3.4. INTERPOLATION INEQUALITIES(3) 93

Next, we estimate

I{
Ax-zl >lx-yl/2,ly-zl >lx-yl/2
(Ka(x- z)- Ka(Y- z))Dau(z)dzl.

In the region

lx- zl > lx- Yl/2 and IY- zl > lx- Yl/2 (3.91)
we have IY- zl/3 < lx- zl < 3ly- zl. In what follows for the time being
we consider in the region (3.91).

Pa(X- z)
Ka(x-z)-Ka(y-z)=(¢(x-z)-¢(y-z)) Ix-z lk
+¢(y _ z) (Pa(X- z) _ Pa(Y- z)), (3 .92)
lx - zlk IY - zlk
IPa(x- z)l/lx- zlk :::; Glx- zlnfp+l-n. (3.93)

Since "V Pa is homogeneous of degree 2m- njp- 2 ;::: 0, it is easy to see that

IPa(x- z)- Pa(Y- z)l ::=; Clx- YIIY- zl 2 m-n/P- 2 •

It is also easily seen that

Hence, we see that

IPa(x- z)lx- zl-k- Pa(Y- z)IY- zl-kl:::; Clx- YIIY- zlnfp-n. (3.94)

From (3.92),(3.93),(3.94) it follows that

IKa(x- z)- Ka(Y- z)l:::; Cl¢(x- z)- ¢(y- z)llx- zlnfp+l-n

+C¢(y- z)lx- YIIY- zlnfp-n. (3.95)

Suppose lx- Yl :::; 2. With the aid of (3.95) we see that if 1 < p:::; oo

{ (Ka(x- z)- Ka(Y- z)) Dau(z)dz

Jlx-zl>lx-yl/2,ly-zl>lx-yl/2

:::; C { l¢(x- z)- ¢(y- z)llx- zln/p+l-niDau(z)idz

Jlx-zl9
+Cix- Yl { ¢(y- z)IY- zlnfp-niDau(z)ldz
}ly-zl >lx-yl/2
94 CHAPTER3. SOBOLEVSPACES

::::; Glx- yj { jx- zjnfpH-njDo:u(z)jdz

Ac-zl9
+Cjx- yj { jy- zjnlv-njDo:u(z)jdz
A~:-vl/2<lv-zl <1
1/p'

::::; Glx - yj ( r
Jlx-zl<3
jx - zjP'-ndz) julm,p

1/p'

+Cjx- yj ( { jy- zj-ndz) juim,v

Jlx-vl/2< lv-zl <1
1/p'
::::; Glx - yjjulm,p + Glx - yj (log jx : yj) julm,p·

It is easy to see that this inequality is true also when p = 1. Hence, com-
bining (3.89),(3.90) and this inequality we conclude that if jx- yj ::::; 2

linn (Ko:(x- z)- Ko:(Y- z)) Do:u(z)dzl

2 ) (p-1)/p
::::; Cjx - yjjulm,p + Cjx - yj ( log jx _ yj julm,p· (3.96)

As is easily seen

linn (L(x- z)- L(y- z)) u(z)dzl

::::; Cjx- yj { ju(z)jdz::::; Cjx- yiiuio,q· (3.97)

Jlx-zlllly-zi:S:1

It follows from (3.96) and (3.97) that if jx- yj ::::; 2

jD!3u(x)- Df3u(y)j

::::; Cjx- yj { [ 1 + (log jx: yj) (p- )/p]lulm,p + jujo,q}.

1

Finally consider the case jx- yj > 2. In view of Lemma 3.10 or its proof

jD!3u(x)- Df3u(y)j::::; jD!3u(x)j + jD!3u(y)j

::::; C (jujm,p + jujo,q)::::; Cjx- yj(jujm,p + jujo,q).
Thus the proof of the lemma is complete.
3.4. INTERPOLATION INEQUALITIES(3) 95

Theorem 3. 6 Let m, j be integers satisfying 0 :::; j < m, and let 1 :::; p :::;
oo, 1 :::; q :::; oo. Suppose that m- j- njp is a positive integer. Then for
any a satisfying

( ;;:+-q
j 1) I (m 1 1) <a<1
~-p+q (3.98)

there exists a constant "' depending only on m, j, n, p, q, a such that for any
u E C[J"(Rn)
(3.99)
where
! = j_ +a(!_ m) + 1- a. (3.100)
r n p n q

Proof. In the present case the range of r is (3.84), and

JDiuJr = JDiul-n/r,oo :::; Julj-njr,oo·

In view of (3.100)

j- n/r = a(m- njp + njq)- njq < m- njp. (3.101)

If -n/r is an integer, then by Lemma 3.10

Juli -n/r,oo
:::; CJul~,p Jul6~a,
where a= (n + jq- nq/r)pj(mpq + np- nq), with a suitable modification
when p = oo or q = oo. It is easily seen that a satisfies (3.98) and (3.100).
Next consider the case where -n/r is not an integer. First suppose that

j + [-n/r] + 1:::; m- n/p -1.

Set k = j + [-n/r]. Then

k < j- n/r < k + 1 :::; m- njp- 1. (3.102)

By virtue of Lemma 3.10

Julk,oo :::; CJul~,plul6~b, (3.103)

Julk+l,oo :::; CJul~,pluJ6~c, (3.104)

where

b = (n + kq)p/(mpq + np- nq),

c = (n + kq + q)p/(mpq + np- nq) = b + pqj(mpq + np- nq).
96 CHAPTER 3. SOBOLEV SPACES

By Theorem 3.1 and (3.102)

.
IU IJ-n/r,oo <
-
Cl lj-nfr-kl lk+l-j+n/r
U k+l,oo U k,oo · (3.105)

Combining (3.103),(3.104),(3.105) we obtain

where

a= c(j- njr- k) + b(k + 1- j + njr)

= [b+ pqj(mpq + np- nq)] (j- njr- k) +b(k + 1- j + njr)
= b + pq(j- njr- k)j(mpq + np- nq)
= [(n + kq)p + pq(j- njr- k)] j(mpq + np- nq)
= (n + jq- nqjr)pj(mpq + np- nq).
Finally suppose that

j + [-njr] + 1 > m- njp -1. (3.106)

It follows from (3.101) and (3.106) that j + [-njr] = m - njp- 1. Set

h = -njr- [-njr]. Then -njr = [-njr] +hand 0 < h < 1.

IDiul-n/r,oo =sup IDi+l-n/rlu(x)- Di+l-n/rlu(y)l/lx- Ylh

xf.y

=sup 1Dm-nfp-1u(x) - Dm-n/p-lu(y)l/lx- Ylh·

xf.y

If lx - Yl :'S 2, then in view of Lemma 3.11

1Dm-n/p-1u(x)- Dm-nfp-lu(y)l/lx- Ylh

2 ) (p-l)/pl }
:'S Gjx- Yll-h { [ 1 + ( log lx _ Yl lulm,p + lulo,q

:'S C(lulm,p + lulo,q)·

If lx- Yl 2 2, then by Lemma 3.10 or its proof

1Dm-nfp-1u(x)- Dm-n/p-lu(y)l/lx- Ylh

:'S C(IDm-nfp-lu(x)l + 1Dm-nfp- 1u(y)i) :'S C(lulm,p + lulo,q)·
Hence

1Dm-n/p- 1u(x)- Dm-nfp-lu(y)lflx- Ylh :'S C(lulm,p + lulo,q)

3.4. INTERPOLATION INEQUALITIES(3) 97

whenever x f:. y. The desired result follows if we apply the argument by

which we deduced (3. 70) from (3. 75).
It is easy to verify that the inequalities established so far hold also for
functions for which the seminorms in the inequalities are all finite.
Theorem 3. 7 Let m, j be integers satisfying 0 :::; j < m, and let 1 :::; p <
oo. Suppose m- j- njp is not a nonnegative integer. Then, if j jm :::; a :::; 1
and 1/r = jjn + 1/p- amjn ~ 0, we have wm•P(Rn) C Wi,r(Rn) and

Jlullj,r:::; 'Y (lul~,plul6:;;a + lulo,p) · (3.107)

If a= jjm in particular

llullj,p :::; 'Y (lul!,i':;lul~:;;j/m + lulo,p) . (3.108)

This holds also when m- j- njp is a nonnegative integer.

Proof. It follows from the hypothesis that r ~ p. Applying Theorem 3.3
with q = p,
lulj,r :::; Glul~,p lul6:;;a · (3.109)
If a= jjm in particular, r = p, and hence

IU Ij,p < j/ml 11-j/m,

_ Cl U lm,p U O,p (3.110)
which holds also when m- j- njp is a nonnegative integer. For an integer
k satisfying 0:::; k < j put b = (am- j + k)j(m- j + k). Then

k < b < a< 1 ! = ~ +! - b(m- j + k).

m-j+k- - - ' r n p n
Hence, again by Theorem 3.3 with q = p

lulk,r:::; Glul~-j+k,plul6:;;b. (3.111)

Replacing j by m - j + k in (3.110)
lui m-3+k,p-
. < Clul(m-J+k)/mlul(j-k)/m
m,p O,p •

Substituting this into (3.111) and applying Young's inequality

(am-j+k)/ml U l(m-am+j-k)/m
Iu Ik,r -< Cl U lm,p O,p

= C (lul~,plul6:;;a)(am-J+k)/am lul6~;k)/am (3.112)

:::; C (lul~,plul6:;;a + lulo,p) ·
The inequality (3.107) follows from (3.109),(3.112). Analogously (3.108) is
established.
98 CHAPTER 3. SOBOLEV SPACES

Corollary 3. 1 If m is a positive integer, 1 ::; p < oo and m- njp < 0,

then wm,P(Rn) c Lr(Rn) for 1/r = 1/p- mjn.

Theorem 3. 8 Let m, j be integers satisfying 0 ::; j < m and 1 ::; p < oo.
Suppose m- j- njp = 0. Then for any r satisfying p::; r < oo, we have
wm,P(Rn) c Wi,r(Rn) and

(3.113)

where a is a number such that 1/r = jjn + 1/p- amjn, i.e. jjm ::; a =
1-njmr < 1.

Proof. Applying Theorem 3.4 with q = p we get

lulj,r ::; Clul~,p lul6;;;a · (3.114)

Let k be an integer such that 0::; k < j. Put b = m/k+n(1/p-1jr)jm.

Then,

b < a, kjm::; b < kjm + njmp, 1/r = k/n + 1/p- bmjn.

Hence, in view of Theorem 3.5 and Lemma 1.1 (Young's inequality)

lulk,r::; Clul~,plul6;;;b::; C (lul~,plul6;;;a)b/a lul~:;b/a

::; C (lul~,plul6;;;a + lulo,p) · (3.115)

The inequality (3.113) follows from (3.114) and (3.115).

Theorem 3. 9 Let m, j be integers satisfying 0 ::; j < m, and let 1 ::; p ::;
oo. Suppose m- j- njp > 0. Then for any r satisfying p::; r::; oo we have
wm,P(Rn) C Wi,r(Rn) and for any u E wm,P(Rn)

llulli,r ::; 'Y (lul~,plul6;;;a + lulo,p),

where a is a number satisfying 1/r = jjn + 1/p- amjn.
Proof. The theorem can be shown with the aid of Theorem 3.5 in the same
manner as Theorem 3. 7.

Remark 3. 4 Under the assumption of Theorem 3.9 a satisfies jjm::; a::;

jjm + njmp < 1.

Theorem 3. 10 Let m be a positive integer and 1 ::; p < oo. Suppose

m-njp> 0.
3.5. SOBOLEV SPACES IN GENERAL DOMAINS 99

(i) If njp is not an integer, then wm,p(Rn) C Bm-n!P(Rn). Put l =

m - [njp] - 1. Then, for each integer j satisfying 0 :::; j :::; l

llullj,oo :::; 'Y (iul~,tjp)jmpiulb;-n-jp)jmp + iulo,p) · (3.116)

For 0:::; h:::; [njp] + 1- njp

'Dl U
1 < 1 l(n+lp+hp)jmpl U l(mp-n-lp-hp)/mp
h,oo _ 'Y U m,p o,p · (3.117)
Hence for 0:::; s :::; m - njp
llulls,oo :::; 'Y (iul~,tsp)jmplulb;p-n-sp)jmp + iulo,p) · (3.118)

In particular
llullm-njp,oo :::; 'YIIullm,p· (3.119)

(ii) If njp is an integer, then for each s satisfying 0:::; 8 < m- njp we have
wm,p(Rn) c B 8 (Rn). (3.116) holds for each integer j such that 0:::; j :::;
l = m- njp- 1. For 0:::; h < 1 (3.117) holds with a constant 'Y depending
also on h. For 0:::; 8 < m -njp (3.118) holds with a constant 'Y depending
also on 8.

Proof. (i) We apply Theorem 3.3 with j = 0 and q = p. For an integer k

satisfying 0:::; k < m- njp we take a= (n + kp)jmp. Then njmp:::; a< 1
and r defined by (3.51) is -njk. Hence

luik ,oo = juj -n/k -< Cjuj<n+kp)jmplul(mp-n-kp)fmp

m,p O,p ·
Thus the inequality (3.116) is shown with the aid of the argument by which
we derived (3.107) from (3.109), (3.112). We next apply Theorem 3.3 with
j = l and p = q. If we take a = (n + lp + hp)jmp, then ljm < a :::; 1 and
r = -njk. Hence we obtain (3.117). (3.118) is a consequence of (3.116)
and (3.117).
(ii) We simply apply Theorem 3.5 and 3.6 instead of Theorem 3.3 in the
proof of (i).

3.5 Sobolev Spaces in General Domains

In this section we prove some preliminary results concerning Sobolev spaces
in general nonempty open sets of Rn.
Theorem 3. 11 If 0 is a non empty open set of Rn, m is a nonnegative
integer and 1:::; p < oo, then wm,P(O) n C 00 (0) is dense in wm,P(O).
100 CHAPTER3. SOBOLEVSPACES

Proof. We follow the proof of Theorem 6.3 of A. Friedman [65]. For a> 0
set
fla = {x E 0; dist(x, ofl) > a-1, lxl <a}. (3.120)
For each natural number k let ak, a~ be the sets defined by
a1 = 03, ai = fls/3• ak = nk+2 \ Ok, a~= nk+S/3 \ Ok+1/3 for k 2:: 2.
It is easy to show that {au is an open covering of n, G~ c ak and any
three of {ak} have an empty intersection. In case ai. = 0 8 ; 3 is empty
we renumber {Ok} appropriately. Let 'f/k be a function in Cff'(Rn) with
support contained in ak such that 'f/k = 1 in a~ and 0 :::; 'f/k :::; 1 in Rn.
Set 'fJ = 2::~1 'f/k· For each X E n at least one of {au contains X and at
most two of {ak} contain x. Hence, we have 1 :::; ry(x) :::; 2. Therefore if
we set (k(x) = 'f/k(x)jry(x), then (k E Cff'(O), supp(k C ak, 2::~ 1 ((x)
i.e. {(k} is a partition of unity subordinate to {ak}· Let u be an arbitrary
1, =
element of wm,P(fl) and f_ > 0. Set Wk = P<k * ((ku), where P< is a mollifier.
If Ek is sufficiently small, then the support of Wk is contained in fl4 if k = 1
and in nk+3 \ nk-1 if k ;::: 2, and
llwk - (kullm,p,O < 2(2 -k)/p- 2 E. (3.121)
Set w = 2::~1 Wk. Since five of the sets nk+3 \ nk-1 have an empty inter-
section, wE C 00 (0) and for lal ::'S m

ID"u(x) - D"w(x) I' = ~~ D" ((,(x)u(x) - w,(x)) I'

:::; 4P- 1 2: IDa((k(x)u(x)- Wk(x))IP.
k=1
With the aid of this and (3.121) we obtain

2: li(ku- wkii~,p,n
00

llu- wll~,p,n ::'S 4P- 1

k=1
00 00

< 4p-1 2:22-k-2pf_p = 2:2-kEp = EP,

k=1 k=1
and the proof is complete.
By Hm,oo(n) we denote the set of all functions whose derivatives of order
up to m - 1 are all bounded and uniformly Lipschitz continuous in n.
Hm,oo(n) is a Banach space with norm
I IDau(x)- nau(y)l
llullm,oo,n = llullm-1,oo,O + ial:s;m-1
max sup
x,yEO,xf-y
IX- YI . (3.122)
3.5. SOBOLEV SPACES IN GENERAL DOMAINS 101

If n = Rn, we simply write llull~,oo short for llull~,oo,Rn·

Lemma 3. 12 For any nonempty open set n c Rn,
Hm,oo(n) c wm,oo(n).
Proof. It suffices to show in case m = 1. Let u E H 1 •00 (n) and suppose for
all x, yEn
lu(x)- u(y)l :::; Llx- Yl· (3.123)
Let e3 be the n-dimensional vector whose jth component is 1 and other
components are 0. Let ¢J E Cg>(n). If lhl is sufficiently small, then by
(3.123)

lin u(x)(ifJ(x + he3 )- ifJ(x))dxl

= lin (u(x- he3)- u(x))ifJ(x)dxl :::; Llhl in lifl(x)ldx.

Therefore

I(Dju)(ifl)l = lin u(x)D3¢J(x)dxl

lim I ru(x)h- (ifJ(x + hej)- ifJ(x)) dxl:::; L rlifl(x)ldx.
= h~o 1
Jn Jn
This means that D3u E L 00 (n) since Cg>(n) is dense in L 1 (n).
Lemma 3.13 Let n be a nonempty open set of Rn and u E W 1•00 (n).
Then, if we modify the values of u on some null set, u is Lipschitz continuous
in each compact subset of n. If x E n and lx - Yl < dist(x, an)

lu(x)- u(y)l S (t, ID,ul5,oo,n) 'i' lx- Yl· (3.124)

Proof. For E > 0 set

= {x En; dist(x, an) > E}.
n<
Let P< be a mollifier and set u. = P< * u. Then u. E C 00 (n.). Let X En
and lx - Yl < dist(x, an). If E is SO small that 0 < E < dist(x, an) and
lx- Yl < dist(x, an) - E, then

u.(y)- u.(x) = 1!1

u.(x + t(y- x))dt

= 1~Diu.(x+t(y-x))(yi-xi)dt.
1 n
102 CHAPTER 3. SOBOLEV SPACES

n
lu,(x) - u,(y)l ::; :2::: IDiu<lo,oo,n. lxi- Yil
i=l

~ t ID;ulo,oo,nlx;- Yll ~ (t ID;ul:,oo,n) l/> lx- Yl·

The desired result is obtained by letting E _, 0.

We get from Lemmas 3.12 and 3.13 the following theorem.
Theorem 3. 12 For any positive integer m

wm,oo(Rn) = Hm,oo(Rn).
Remark 3. 5 From the proof of Lemma 3.13 it follows that wm,oo(n) =
Hm,oo(n) is true for any convex open set n.

3.6 Uniformly Regular Open Sets

Definition 3. 1 Let n be a nonempty open subset of Rn. If for any point
a of the boundary an there exists a neighborhood 0 of a and a homeomor-
phism <I> of class em from 0 to the open unit ball {y; IYI < 1} of Rn such
that

<I>( a) = 0, <1>(0 n n) = {y ERn; IYI < 1, Yn > 0},

<I>(Onan) = {y ERn; IYI < 1,Yn = 0}, (3.125)

then n is called an open set of class em.

If n is an open set of class em with bounded boundary, then there exist
finite points a1, ... , aN on the boundary an and a neighborhood Oi of ai,
and a homeomorphism <I>i of class em from oi to the open unit ball of
Rn, i = 1, ... , N satisfying the conditions of Definition 3.1 for a= ai, i =
1,· .. ,N,and
u~l<I>- 1 ({y; IYI < 1/2}):) an.
When an is not bounded, we consider following F. E. Browder [21] open
sets with the following properties:
Definition 3. 2 Suppose n is a nonempty open subset of Rn whose bound-
ary is not bounded. Then n is said to be uniformly regular of class em if
there exist a family of open sets {Oi; i = 1, 2, ... } and of homeomorphisms
3.6. UNIFORMLY REGULAR OPEN SETS 103

{4>i} of Oi onto the unit ball {y; IYI < 1} in Rn, and an integer N such that
the following conditions are satisfied:
(1) Let OI = 4>i 1 ( {y E Rn; IYI < 1/2} ). Then U~ 1 OI contains the N- 1
neighborhood of 80..
(2) For each i
4>i(Oi n 0.) = {y; IYI < 1, Yn > 0}, (3.126)
4>i(Oi n 80.) = {y; IYI < 1, Yn = 0}. (3.127)
(3) Any N + 1 distinct sets of { Oi} have an empty intersection.
(4) The family { Oi} is locally finite, i.e. only a finite number of Oi have a
nonempty intersection with some neighborhood of each point of Rn.
(5) Let wi = 4>i 1 be the inverse mapping of 4>i. Then for each i = 1, 2, ...
and IYI < 1
(3.128)
Let 4>ik(x), wik(Y) be the kth components of4>i(x), wi(Y) respectively. Then
1Da4>ik(x)l:::; M, IDawik(Y)I:::; M, l4>in(x)l:::; Mdist(x,80.) (3.129)
for Ia I :::; m, X E oi, IYI < 1, k = 1, .. . 'n, and i = 1, 2, ....
When m 2:: 2 or also when m = 1 if 8wik/8Yi are equicontinuous, then (4)
is implied by (3) and (5) as is shown in the following lemma.
Lemma 3.14 Let f(y) = (ft(y), ... , fn(Y)) be of class C 1 in the unit ball
of Rn and suppose the Jacobian D f / Dy is different from 0 at y = 0. Then
there exists a constant 8 depending only on the norm of the inverse of the
Jacobian matrix fv = 8ff8y at y = 0 and the moduli of continuity of
8fif8yj, i, j = 1, ... 'n, at y = 0 such that
{f(y); IYI < 1}:) {x; lx- f(O)I < 8}.
Proof. Denote the norm of the matrix A be IAI. Let E be a positive number
such that lfv(Y)- fv(O)I:::; 1/(2lfv(0)- 1 1) for IYI:::; E. Set
g(y) = y- fv(o)- 1 (f(y)- x),
f(y) - f(z) = fv(O)(y- z) + h(y, z)(y- z),

h(y, z) = 1 1
(fy(z + t(y- z))- fv(O)) dt.

Then lh(y, z)l :::; 1/(2lfv(o)- 1 1) for IYI:::; E, lzl:::; E. Put 8 = E/(2lfv(o)- 1 1).
If lx- f(O)I :::; 8, then for each y such that IYI :::; E
lg(y)l= IY- fv(o)- 1 (f(y)- f(O))- fv(o)- 1 (!(0)- x)l
= 1- fv(o)- 1 h(y, O)y- fv(o)- 1 (!(0)- x)l
:::; lfv(o)- 1 llh(y, O)IIYI + lfv(o)- 1 18:::; f_,
104 CHAPTER 3. SOBOLEV SPACES

Furthermore if IYI ::; E, lzl ::; E, then lg(y)- g(z)i ::; IY- zl/2. Hence, if
lx- f(O)I::; 8,g has a unique fixed pointy which clearly satisfies x = f(y).

Applying Lemma 3.14 to \IIi we see that each Oi contains a ball Bi of

some fixed radius. In view of (3.128) the sets Oi which have a nonempty
intersection with some bounded neighborhood of a point are located in some
bounded region. If there exist infinite number of such Oi, then a limit point
of the centers of the balls Bi contained in these Oi is contained in infinite
number of Bi and hence of Oi. This contradicts (3).
Let 0 be an open set uniformly regular of class em, and {Oi}, {<l>i},N
be as in Definition 3.2. Put

Oo = {X E 0; dist(x, 80) > (4N)- 1},

0-1 = {x rj. O;dist(x,80) > (4N)- 1}.

Then {Oi; i = -1, 0, 1, 2, ... } is an open covering of Rn. Let >. be a function
in C{f"(Rn) satisfying supp>. C {y; IYI < 1}, >.(y) = 1 for IYI < 1/2,0 ::;
>.(y) ::; 1 in Rn. If we set

for i = 1, 2,.... Then T/i E C[f"(Rn), T/i(x) = 1 in 0~, SUPPT/i c oi and

0::; 'fli(x) ::; 1 in Rn. Let 'flo and 11- 1 be functions in cm(Rn) such that

SUPPT/o c 0, SUPP'Tl-1 c Rn \ 0,
'flo(x) = 1 for x E 0 and dist(x, 80) > 3j(4N),
'flo(x)=O for xEO and dist(x,80)<1/(2N),
T/-1(x) = 1 for X rj. 0 and dist(x,80) > 3/(4N),
'fl- 1 (x)=0 for xrj.O and dist(x,80)<1/(2N).

Put 11(x) = 2::::_ 1 T/i(x). Since {Oi; i = -1, 0, 1, 2, ... } is an open covering
of Rn such that at most N + 2 of Oi have a nonempty intersection, we have
1::; 'fl(X) ::; N + 2 in Rn and 'fl E cm(Rn). If we put
(3.130)

then (i E cm(Rn),supp(i C Oi,I:;:_ 1 (i(x) = 1 and {Da(i} is uniformly

bounded for each lal ::; m.
In what follows in this chapter {(i} ~- 1 will always stand for this partition
of unity.
3.7. UNIFORMLY REGULAR OPEN SETS 105

3. 7 Sobolev Spaces in Uniformly Regular

Open Sets
Let cm(!J) be the set of all m times continuously differentiable functions in
n, and con(ri) be the totality of functions in cm(!J) which vanish outside
a bounded set. Put

R~ = {x = (x1, ... ,xn) E Rn;xn > 0},

R~ = {x = (x1, . .. ,xn) E Rn;xn ~ 0}.
For x = (x1, ... , Xn) ERn we write x' = (x1, ... , Xn-1) and x = (x', Xn)·
Theorem 3. 13 Let 0 be an open subset of Rn uniformly regular of class
em. Then CQ"(O) is dense in wm·P(f!) if 1:::; p < oo.
Proof. According to Theorem 3.11 it suffices to show that each function u
in Wm•P(f!) n C 00 (f!) can be approximated by a sequence of functions in
CQ"(O) in the norm of wm,P(f!). Let¢> be a function in CQ"(Rn) such that
¢(x) = 1 for !xi < 1 and ¢(x) = 0 for !xi > 2, and put ¢>R(x) = ¢>(x/ R)
for R > 0. Since ¢>Ru ---+ u in wm·P(f!) as R ---+ oo, we may assume that u
vanishes outside some bounded set. If we put ui = (iu for i = 0, 1, 2, ... ,
then we have u = L~=o Ui for some integer k. If we put

IYI < l,Yn > 0

IYI ~ l,yn > 0
for i = 0, 1, ... 'k, then Vi E wm·P(R~) n cm(R~). Let 'lj;(t) be a function
in C 00 (-oo,oo) satisfying '1/J(t) = 1 fort~ 0 and '!j;(t) = 0 fort:::; -1/2.
Put
Yn > -8
Yn:::; -8
Evidently Vi,6 E wm·P(Rn) n cm(Rn). Since Vi,o(Y) = Vi(y',Yn + 8) for
y E R~, we have vi,6 ---+vi in wm•P(R~) as 8---+ 0. If 8 is sufficiently small,
then the support of vi,6 is contained in the unit ball. Put

Then Ui,6 E wm·P(Rn) n cm(Rn), suppui,6 c oi and Ui,6---+ Ui in wm·P(f!)

as 8 ---+ 0. Hence if we put u0 = uo + L~=l ui,6 after extending uo as 0
outside 0, u0 E Wm•P(Rn) n Cm(Rn) and u0 ---+ u in Wm•P(f!) as 8 ---+ 0.
We obtain a desired sequence by appropriately numbering the restriction of
106 CHAPTER3. SOBOLEVSPACES

P< * Uo to n where P< is a mollifier.

Next we state a method of extending a function defined in R~ to Rn
following L. Nirenberg [118].
Lemma 3. 15 For a nonnegative integerm there exist real numbers ,\1, ... ,
Am+1 such that

(3.131)
k=1

for j = 0, 1, ... , m. For a function u defined in R~ put

Xn <0
(3.132)
Then
(i) u E cm(Rn) ifu E cm(R~),
(ii)u E C[f(Rn) if u E C(f(R+)·
There exists a constant Cm depending only on m such that for any 1 :::; p :::; oo
and 0:::; j:::; m
(3.133)
Proof. The existence of ,\ 1 , ... , Am+l follows from the nonvanishing of the
determinant of Vandermonde. The remaining part of the lemma is easily
shown.
Theorem 3. 14 Let m be a natuml number. If n is an open set uniformly
regular of class C 1 , then wm·=(n) = Hm·=(n).
Proof. We may confine ourselves to the case m = 1. In view of Lemma
3.12 it suffices to show that W 1 ·=(n) c H 1 ·=(n). Let u be an arbitrary
element of W 1 ·=(n). Then by Lemma 3.13 u is Lipschitz continuous on
each compact subset of n if we modify the values of u on a null set, and
(3.124) holds for ix - Yi < dist(x, &n). Let X and y be two points of n
belonging to some oi, i ~ 1. If we set

then 'Y = {x(t); 0:::; t:::; 1} is a path of class C 1 connecting X andy in Oi.
In view of (3.129) there exists a constant C independent of i such that

I'YI =the length of 'Y:::; Cjx- Yi· (3.134)

3.7. UNIFORMLY REGULAR OPEN SETS 107

Taking points xo = x,x1,·· ·,xk-1,xk = yon "f so that lxi+1- xil <
dist('Y, 0) we get in view of Lemma 3.13 that

Hence,
k-1
iu(y)- u(x)l ::; E lu(xH1)- u(xi)l
i=O
k-1
::; A E lxi+1- xi I ::; AI'YI ::; ACix- Yl· (3.135)
i=O

Suppose that u is not uniformly Lipschitz continuous. Then there exist

sequences { X 11 }, {y 11 } of points of 0 so that X 11 #- y 11 and

lim iu(xv) - u(yv)l = oo. (3.136)

v-+oo lxv - Yvl
If there exists a positive number 8 such that lx 11 - Yvl 2:: 8 for any v, then
iu(xv)- u(yv)l < 2lulo,oo,!l
lxv -Yvl - {j '

which contradicts (3.136). Hence replacing by a subsequence if necessary

we may assume that lxv - Yvl ---+ 0. If dist(x 11 , 80) 2:: 8 > 0 for any v, then
lxv- Yvl < 8 if vis large. In view of Lemma 3.13 we have

which contradicts (3.136). Hence we may assume dist(x 11 , 80) < N- 1. By

the condition (1) of the uniform regularity of class C 1 there exists for each
v a number i such that Xv E 0~. Since lxv- Yvl---+ O,infdist(0~,80i) > 0,
we have Yv E Oi for large v. Hence, by virtue of (3.135)

iu(xv)- u(yv)l::; ACixv- Yvl

which contradicts (3.136). Thus (3.136) does not hold and we conclude
u E H 1 •00 (0).
By virtue of Theorem 3.14 and the closed graph theorem there exist pos-
itive constants C, c depending only on m and 0 such that

(3.137)
108 CHAPTER 3. SOBOLEV SPACES

Lemma 3. 16 If u is the function defined by (3.132) for u E Hm,oo(R+),

then u E Hm,oo(Rn), and there exists a constant e independent of u such
that
llull~ ' oo :::; e11u11~ ' oo ' R"+ · (3.138)

Proof. Since u E em- 1 (R'+), it follows from Lemma 3.15 that u E cm- 1 (Rn).
If X= (x1, ... , Xn), y = (yl, ... , Yn), Xn < 0, Yn;::: 0, Ia I :::; m- 1

=L
m+l
Dau(x)- Dau(y) .>..k( -k)a" (Dau(x1, ... , Xn-1, -kxn)- Dau(y)).
k=1

The inequality (3.138) follows from this and

1- kxn- Ynl :::; -kxn + Yn:::; k( -Xn + Yn) = kixn- Ynl·

Theorem 3. 15 Let m be a nonnegative integer, and 1 :::; p :::; oo. If n

is an open set of Rn uniformly regular of class em, then there exists a
bounded linear operator E from wm,P(O) to wm,p(Rn) such that for each
u E wm,P(O) the restriction of Eu to 0 coincides with u.

Proof. First consider the case 1 :::; p < oo. In view of Theorem 3.13 it
suffices to define Eu for u E e 0 (0). Let { Oi}, { oa, {
<I>i}, {wi}, N, M be
as in the definition of uniform regular domains of class em, and {(i} be the
family of functions defined by (3.130). Put Ui = (iu. Let vi, i > 0, be the
functions defined by

IYI < 1, Yn;::: 0

IYI;::: 1, Yn;::: 0
Then, vi E ef!"(R+), and by virtue of (3.129)

llvillm,p,R+:::; el!uillm,p,O (3.139)

with a constant e independent of i. If we define the function Vi by

Yn > 0

Yn <0

then in view of Lemma 3.15 viE ef!"(Rn), suppvi C {y; IYI < 1},
(3.140)
3.7. UNIFORMLY REGULAR OPEN SETS 109

Next, if we put

then it follows from (3.129) that

Ilui llm,p ::; Clivi llm,v• (3.141)

where C is a constant independent also of i. Let u0 be the extension of uo

toRn by 0 outside 0. Then u0 E C[f(Rn) and

lluollm,p = lluollm,p,O· (3.142)

Set u =E:o ui. Let In I ::; m. Since for each x only at most N + 1 of
D<>ui(x) are different from 0,

(3.143)

It is easy to show that there exists a constant C independent of i such that

(3.144)

From (3.139), ... ,(3.144) it follows that

llull~,p = r L
}Rn lal::;m
IDau(x)IPdx

::; (N + 1)P- 1 ~ L,. l,t=m IDaui(x)IPdx = (N + 1)P- 1 ~ lluill~,p

(3.145)

Since N +2 of {Oi} have an empty intersection, the last side of (3.145) does
not exceed
c rL
Jnlal:c;m
jDau(x)IPdx = Cllull~,p,fl'
Evidently uln = u. Thus, by putting Eu = u, we complete the proof of the
case 1 ::; p < oo.
Next suppose p = oo. For u E wm,=(n) let the functions Vi, Vi, Ui be
110 CHAPTER 3. SOBOLEV SPACES

defined as in the case 1::; p < oo. Since u E Hm,oo(O) in view of Theorem
3.14, it is clear that ViE Hm,oo(R'+), and

(3.146)

with a constant C independent ofi. By virtue of Lemma 3.16 ViE Hm,oo(Rn)

and
llvill~,oo::; Cllvill~,oo,R+· (3.147)

It is evident that Ui E Hm,oo(Rn) and

(3.148)

Put u= :L::~o ui. Since for x, y E Rn, IaI ::; m - 1

IDau(x)l ::; (N + 1) sup IDaui(x)l,

i
IDau(x)- Dau(y)l ::; 2(N + 1) sup IDaui(x)- Daui(Y)I,
i

we have
(3.149)

By virtue of Theorem 3.14, (3.146),(3.147),(3.148),(3.149) we get

llullm,oo $ Clliill~ ' oo $ C sup

t.
lluillm,oo,n $ Cllullm,oo,n·
Thus Eu = u satisfies the assertion of the theorem.

3.8 Embedding Theorems

It will be shown that the results of section 3.4 hold for open sets other than
Rn.

Theorem 3. 16 Let 0 be an open set of Rn uniformly regular of class C 2 •

If m, j are integers satisfying 0 ::; j < m and 1 ::; p ::; oo, then there exists
a constant 'Y such that for u E wm,P(O)

llulli,p,n::; 'Y (1ul;,(;,nlul~~:£m + lulo,p,n) · (3.150)

Proof. We begin with the case j =m - 1:

(3.151)
3.8. EMBEDDING THEOREMS 111

Let E be the operator of Theorem 3.15 in the case m = 2. Then by Theorem

3.7

Hence

llulh,p,n:::; C (llull~;:,nlul~;:,n + lulo,p,n)

1/2 112 112 I 1112
:::; C ( lu 12,p,nlulo,p,n + llu 1l1,p,n u o,p,n + Iu Io,p,n ) '

from which it follows that

llulh,p,n:::; C (lul~;:,nlul~;:,n + lulo,p,n) · (3.152)

Suppose (3.151) is true with m replaced by m- 1:

(m-2)/(m-1) I 11/(m-1)
II U II m-2,p,!l:::; C (I U lm-1,p,!l U O,p,!l + IU IO,p,!l ) · (3.153)

Applying (3.152) to Dm- 2 u we get

lulm-1,p,n :::; C (1ul;,(~,nlul;,(:_2,p,n + lulm-2,p,n) ·

Combining this with (3.153)

lulm-1,p,n:::; c(lul;,(~,nlu1~--;~;;~(m- 1 )lul~::.~- 1 )

1/2 I 1112 I 1cm-2)/(m-1)l 11;cm-1)
+ IU 1m,p,!l U O,p,!l + U m-1,p,!l U O,p,!l + IU IO,p,!l ) · (3.154)

With the aid of Young's inequality

lul1/2 lul(m-2)/2(m-1) lul1/2(m-1)

m,p,!l m-1,p,!l O,p,!l
_ ( -(m-2)/ml l(m-1)/ml l1/m )m/2(m- 1) ( I I )(m-2)/2(m-1)
- E U m,p,!l U O,p,!l E U m-1,p,!l
m -(m-2)/ml l(m-1)/ml 11/m m- 2 I I
:::; 2(m - 1) E U m,p,!l U O,p,!l + 2(m - 1) EU m-1,p,!l,
lul1/2 lul1/2 = (1ul(m-1)/mlul1/m)m/2(m-1) lul(m-2)/2(m-1)
m,p,!l O,p,!l m,p,!l 0 0

m I cm-1)/ml 111m m- 2 I I
:::; 2(m- 1) u 1m,p,n u o + 2(m- 1) u o,p,n,

lcm-2J/cm-1JI
Ium- 1 p" 11/(m-1)
Uop"
<
_
m- 2 I I
--EUm-1p!l + - -1E2-ml Uop!l·
I
, ,o~~ ' ,u; m- 1 ' ' m- 1 ' '
112 CHAPTER3. SOBOLEVSPACES

Combining these three inequalities and (3.154) and letting E be sufficiently

small we get

(3.155)

Applying Young's inequality to the first term in the bracket of the right
hand side of (3.153) we get

llullm-2,p,n:::; C (Eiulm-1,p,n + ~-mlulo,p,n + lulo,p,n) · (3.156)

The inequality (3.151) follows from (3.155) and (3.156).

Finally we prove (3.150) by induction on j. Suppose (3.150) is true with
j + 1 in place of j:

(j+1)/ml U 11-(j+l)/m
II U II j+1,p,!l:::; C (I U lm,p,!l O,p,!l + IU IO,p,!l ) . (3.157)

Replacing m by j + 1 in (3.151)
j/(j+1) 1/(i+1) + lulo,p,n ) ·
llullj,p,n :::; C ( luli+l,p,nlulo,p,n (3.158)

The combination of (3.157) and (3.158) yields (3.150).

Theorem 3. 17 Let 0 be an open set of Rn uniformly regular of class C 2 •
Let m, j be integers satisfying 0 :::; j < m, and let 1 :::; p < oo. Suppose
that m- j - njp is not a nonnegative integer. If jjm :::; a :::; 1, 1/r =
jfn+1/p-amjn 2:0, then wm,P(O) C Wi,r(n) and there exists a constant
'Y such that for u E Wm,P(O)

(3.159)

Proof. We begin with the case m = 1. Let E be the operator of Theorem

3.15 form= 1. Suppose 1- njp is not a nonnegative integer. In view of
Theorem 3.7 if 0:::; a:::; 1 and 1/r = 1/p- ajn 2: 0, then

IEulo,r:::; C (IEultPIEul6:;a + IEulo,p) ·

Hence

lulo,r,n :::; C (11ulltp,nlul6:;~n + lulo,p,n) :::; C (1ultp,nlul6:;~n + lulo,p,n) ·

Thus the proof of the present case is complete.
Suppose the assertion of the theorem is true for m - 1. We are going to
show that it is also true for m. Let j, p, r, a be as in the assumption of the
theorem. First we consider the case j > 0. If we set b = (am - 1) / (m - 1),
3.8. EMBEDDING THEOREMS 113

then (j - 1)/(m - 1) ~ b ~ 1. It is easily seen that the hypothesis of

the theorem is satisfied with m, j, a replaced by m - 1, j - 1, b. Hence, by
induction assumption

11Dulli-1,r,n ~ C (1Dul~-1,p,n1Dul6:;~n + IDulo,p,n), (3.160)

llulli-1,r,n ~ C (1ul~-1,p,nlul6:;~n + lulo,p,n) · (3.161)

By Theorem 3.16

IDulo,p,n:::; luh,v,n:::; C (1u1;,:;,n1u16::n1)/m + lulo,p,n) ·

Substituting this in the right hand side of (3.160)

11Dullj-1,r,n:::; C [1u1~,p,n (1ul;,:,;,nlul6::n1)/m + lulo,p,n) 1-b

+lul;,:;,nlul6::n1)/m + lulo,p,n]

-< c(lulb+(1-b)/mlul(m-1)(1-b)/m
m,p,O O,p,O + lulbm,p,O lul1-b
O,p,O
+lul;,:;,nlul6::n1)/m + lulo,p,n)
Since b + (1- b)/m =a, 1- b = m(1- a)j(m- 1), we get

11Dulli-1,r,n :::; C (1ul~.v,nlul6:;~n

b IU 11-b
+IU lm,p,O O,p,O + IU 11/m
m,p,O IU lcm-1)/m
O,p,O + IU IO,p,O ) · (3.162)

Noting b:::; a, 1/m :::; a and applying Young's inequality to the right hand
sides of

b IU 11-b
IU lm,p,O O,p,O = (1 U Iam,p,O IU 11-a
O,p,O )b/a IU l(a-b)/a
O,p,O '
1/m (m-1)/m = ( lulm,v,nlulo,p,n
lulm,v,nlulo,p,n a 1-a ) 1/am lulo,p,n
(am-1)/am
'
we get from (3.162)

11Dulli-1,r,n ~ C (1ul~.v.nlul6:;~n + lulo,p,n) · (3.163)

114 CHAPTER 3. SOBOLEV SPACES

Similarly from (3.161) we get

lulo,r,n :::; llullj-l,r,n

:::; C [ (1ui:;,J2 1 mlul~;p7n + lulo,p,n) b lul6:;;~n + lulo,p,n]

:::; C (lul(m-l)b/mlulb/m+l-b
m,p,O O,p,O
+ lulo ,p,n)

:::; C [(lui am,p,O lull-a

O,p,O
) (m-l)b/ma lul(m(a-b)+b)/ma
O,p,O
+ lulo ,p,n]

:::; C ( lul~,p,n lul6:;;~n + lulo,p,n) ·

Combining this with (3.163) we obtain (3.159).
Next we consider the case j = 0. What is to be proved is that if m- njp
is not a nonnegative integer, then for 0:::; a:::; 1, 1/r = 1/p- amjn ~ 0

lulo,r,n :::; I ( lul~,p,n lul6:;;~n + lulo,p,n) · (3.164)

Case 1/m:::; a:::; 1, r < oo. Lets be a number such that 1/s = 1/n + 1/r.
Then 1/s = 1/n + 1/p- amjn,p :::; s < r. Since m- 1- njp is not a
nonnegative integer, by what is already proved in case j > 0

llulh,s,n:::; C (1ul~,p,nlul6:;;~n + lulo,p,n) · (3.165)

On the other hand from the result of the case m = 1 (note that 1 - nj s =
-njr < 0)
lulo,r,n :::; Cllulh,s,n· (3.166)
The inequality (3.164) follows from (3.165) and (3.166).
Case 1/m :::; a :::; 1, r = oo. In this case a = njmp < 1 since otherwise
we would have m- njp = 0 contrary to the hypothesis. Let 1/s = 1/n +
1/p- mjn. Then n < s < oo. Applying Theorem 3.3 with 1, 0, s,p, njmp
as m,j,p, q, a we get

IEulo,oo:::; CIEui~:SmpiEul~;;nfmp,
where E is the operator of Theorem 3.15 for m = 1. Hence

Iu Io,oo,n :::; ell U l n/mpl 1-n/mp·

l,s,n u 1o,p,n (3.167)

By virtue of the result of the case j > 0 which is already proved

llullt,s,n :::; C (lulm,p,n + lulo,p,n) · (3.168)

3.8. EMBEDDING THEOREMS 115

Combining (3.167) and (3.168) we obtain the desired inequality

iulo,oo,n ::; C (lui~;,~ iui~~~mp + iulo,p,.n) ·

Case 0 ::; a < 1/m. Put b = amj(m- 1). Then 0 ::; b < 1 and 1/r =
1/p- b(m- 1)/n. Hence by induction assumption

lulo,r,n::::; C ( lulm-1,p,nlulo:;;,n
b 1 b
+ lulo,p,n ) · (3.169)

In view of Theorem 3.16

iulm-1,p,.n : : ; C (iui;:,;,i{ 1miui;::.n + iulo,p,.n) · (3.170)

The desired inequality (3.164) follows from (3.169) and (3.170).

Theorem 3. 18 Let n be an open set of Rn uniformly regular of class C 2 .
Let m, j be integers satisfying 0 ::; j < m and let 1 ::::; p < oo. Suppose
m- j- njp = 0. Then for any r satisfying nj(m- j) = p::; r < oo, we
have wm,v(n) c Wi,r(n) and

iiullj,r,.n ::; 'Y (iul~,p,nlui6:;;~n + lulo,p,n), (3.171)

where a is the number satisfying 1/r = jjn + 1/p- amjn = mjn- amjn,
i.e. jjm::::; a= 1- njmr < 1.
Proof. The proof of the case m = 1 and of the case j > 0 under the
assumption that the conclusion of the theorem is true for m - 1 is the same
as that of the previous theorem. Supposing that the theorem has been
proved for m - 1 we are going to show that the assertion of the theorem is
true for m and j = 0:

(3.172)

for njm = p ::; r < oo. First we consider the case r 2 nj(m- 1). Let
1/s = 1/n + 1/r. Noting 1/m ::::; 1- njmr we apply Theorem 3.17 with
= =
j 1,p njm, a= 1- njmr to obtain

iiull1,s,.n ::; C ( iui~~~:::::-.n iui~:n7:,.n + iulo,njm,.n) ·

The inequality (3.172) follows from this inequality and (3.166). The result
of the case njm::; r < nj(m-1) follows from that of the case r = nj(m-1)
and the interpolation inequality
m-nfr IU 1O,njm,.n
IU IO,r,.n : : ; IU lO,n/(m-1),.0 n/r-m+1 · (3.173)
116 CHAPTER 3. SOBOLEV SPACES

The following two theorems are verified with the aid of Theorem 3.9,
Theorem 3.10 and Theorem 3.16.
Theorem 3. 19 Let 0 be an open set of Rn uniformly regular of class
C 2 • Let m, j be integers satisfying 0 :::; j < m, and let 1 :::; p :::; oo.
Suppose m - j - njp > 0. Then for any r satisfying p :::; r :::; oo we
have wm,P(O) C Wi,r(O) and for any u E wm,P(O)

JJullj,r,n :S 'Y (luJ~,p,nJuJ6:;~n + Julo,p,n), (3.174)

where a is a number satisfying 1/r = jjn + 1/p- amjn.

Theorem 3. 20 Let 0 be an open set of Rn uniformly regular of class C 2 •

Let m be a positive integer and 1:::; p < oo. Suppose m- njp > 0.
(i) If njp is not an integer, then wm,P(O) c Bm-n!P(fl). Put l = m-
[njp] - 1. Then, for each integer j satisfying 0:::; j :::; l

l uJJ·J,oo,n <_ 'Y (JuJ(n+jp)fmpJuJ(mp-n-jp)fmp

m,p,!l o,p,n + JuJ o,p,!l ) . (3.175)

For 0:::; h:::; [njp] + 1- njp

JDluJh,oo,n :S 'Y (iuJ~,~~~+hp)/mpJuJ6:;~~n-lp-hp)fmp + JuJo,p,n). (3.176)

Hence for 0 :::; s :::; m - njp

l uJJs ,oo ,n :S 'Y (JuJ(n+sp)fmpJuJ(mp-n-sp)fmp

m,p,O O,p,O
+ JuJo ,p,n) . (3.177)

In particular
JJuJJm-n/p,oo,n :S 'YJJuJJm,p,!l·
(ii) Ifnjp is an integer, then for each s satisfying 0:::; s < m-njp we have
wm,P(O) c B 8 (0). (3.175) holds for each integer j such that 0:::; j:::; l =
m- njp- 1. For 0:::; h < 1 (3.176) holds with a constant depending also
on h. For 0:::; s < m- njp (3.177) holds with a constant 'Y depending also
on s.

In order to show only the inclusion relation wm,P(O) c Wi,r(o) in the

above theorems it suffices to assume that n is uniformly regular of class C 1
as is shown in the following theorem.
Theorem 3. 21 Suppose that 0 is an open set of Rn uniformly regular of
class C 1 . Let m, j be integers satisfying 0 :::; j < m, and let 1 :::; p < oo.
3.8. EMBEDDING THEOREMS 117

(i) If m- j -njp is not a nonnegative integer, p::; r ::; oo, j jn+ 1/p-mjn::;
1/r, then wm·P(O) c Wi,r(O), and there exists a positive constant 'Y such
that for u E Wm•P(O)
(3.178)

(ii) If m- j - njp is a nonnegative integer, then wm·P(O) c Wi,r(O) for

p::; r < oo and (3.171) holds.

Proof. (i) It suffices to prove the theorem for j = 0. We begin with the case
m = 1. What is to be shown is that if 1- njp is not a nonnegative integer,
p::; r::; oo and 1/p- 1/n::; 1/r::; 1/p, then W 1•P(0) c U(O). In view
of Theorem 3.7 we have W 1 ·P(Rn) c U(Rn). Let E be the operator of
Theorem 3.15 form= 1. If u E W 1 ·P(0), then Eu E W 1·P(Rn) C U(Rn).
Hence u E Lr(O). Next, suppose that the conclusion of the theorem is
true for m - 1, i.e. if m - 1 - njp is not a nonnegative integer, p ::; s ::;
oo, 1/p-(m-1)/n::; 1/s, then wm-l,P(O) c £8(0). Assume that m-njp
is not a nonnegative integer, p ::; r ::; oo, 1/p- mjn ::; 1/r. Then, clearly
m - 1 - njp is not a nonnegative integer. Put

!
s
= max {! - 1,!} .
p
m-
n r

(a) If 1/p- (m- 1)/n > 1/r, then 1/s = 1/p- (m- 1)/n < 1/p. By
the induction hypothesis we have wm-l,p(O) C £8(0). Hence Wm•P(O) C
W 1•8 (0). Since 1- n/s = m- njp is not a nonnegative integer and 1/s-
1/n = 1/p- mjn::; 1/r < 1/s, we have W 1•8 (0) c Lr(O) by the result
shown in case m = 1. Hence wm,P(O) c Lr(O).
(b) If 1/p- (m- 1)/n ::; 1/r, then wm-l,P(O) C Lr(O) by the induction
hypothesis. Hence wm,P(O) C Lr(O).
(ii) The conclusion in case m = 1 is established with the aid of Theorem
3.8 in place of Theorem 3.7 in the proof of (i). Suppose that the theorem
has been proved form- 1. Assume that m- njp is a nonnegative integer
and p::; r < oo. If m- 1- njp is a nonnegative integer, then Wm•P(O) C
wm-l,P(O) c Lr(O). If m- njp = 0, then m- 1- njp = -1 is not a
nonnegative integer. Hence in view of the first part (i) we have wm,P(O) C
W 1 •n(O). If r ~ n, W 1 •n(O) C U(O) by Theorem 3.8. Hence wm,P(O) C
U(O). It is evident that this inclusion relation also holds for njm = p::;
r < n since wm,p(O) c LP(O).

Remark 3. 6 Analogously to Theorem 3.21 it can be shown that a re-

sult corresponding to Theorem 3.20 also holds for an open set 0 uniformly
reugular of class C 1 .
118 CHAPTER 3. SOBOLEV SPACES

3.9 Additional Topics

Other important materials on Sobolev spaces are Rellich's theorem and
traces on the boundary. We state them without proof.
Let n be an open subset of Rn uniformly regular of class C 1. If m, j are
integers satisfying 0 :::; j < m, 1 :::; p < oo and
j 1 m 1
1:::; r:::; oo, -+---<- (3.179)
n p n - r'
then by Theorem 3.21 we have
wm,p(n) c wi,r(n). (3.180)

If moreover n is bounded and the inequality holds in the second inequality

of (3.179), then the imbedding (3.180) is compact.
Theorem 3. 22 (Rellich's theorem) Suppose that n is a bounded open set
of Rn of class C 1 . If m, j are integers satisfying 0 :::; j < m, 1 :::; p < oo, 1 :::;
r:::; oo,jjn + 1/p- mjn < 1/r, then the imbedding wm,p(n) c Wi,r(n) is
compact.
Corollary 3. 2 Suppose n is a bounded open set of Rn of class C 1 . If
m is a positive integer and 1 :::; p < 00 1 then the imbedding Wm,p(n) C
wm- 1,P(n) is compact.
If n is an open set uniformly regular of class em' then the" boundary values"
(ajav)iujan,j = 0, ... , m- 1, are defined for functions u E Wm,p(n).
Theorem 3. 23 Let n be an open set of Rn uniformly regular of class
em, m > 0, and 1 :::; p :::; oo. Then for j = 0, 1, ... , m- 1, there exists
a bounded linear mapping 'Yi from wm,v(n) to LP(an) such that if u E
Wm,p(n) n Cm(fi), then "jjU E cm-i(an) and

('Yiu)(x) = (:v Y u(x) (3.181)

for X E an, where V is the outward normal vector of an.

Definition 3. 3 'YoU is called the trace of u on the boundary an.

Usually we simply write u, (ajav)iu in place of 'You, "/jU respectively.

Let n be an open set of Rn uniformly regular of class em, and 1 :::; p < oo.
We denote by wm- 1 /P,P(an) the totality of the traces offunctions belonging
to wm,v(n). The norm of an element g of wm- 1 /P,P(an) is defined by

[gJm-1/p,p,oO = inf{jjullm,p,n; U E Wm,p(n), 'YoU= g}. (3.182)

3.9. OTHER MATERIALS 119

Theorem 3. 24 wm-l/v,v(an) is a Banach space with the norm (3.182).

C(f(an) is dense in wm-lfv,v(an).

n
For a nonempty open set of Rn, a nonnegative integer m and 1 ::; p < oo,
we denote by ~·P(n) the closure of C(f(n) in wm,v(n).
Theorem 3. 25 Let n be an open set of Rn uniformly regular of class
em, m > 0, and 1::; p < oo. An element u ofWm,v(n) belongs to ~·P(n)
if and only if
"foU ="flU=···= "fm-lU = 0.

If the following conditions are satisfied, then n is said to have the re-
stricted cone property.
There exist an open covering { Oi; i = 1, 2, ... } of an and a sequence of
open cones {Ci; i = 1, 2, ... } with vertices at the origine such that x+Ci C n
for each i and X E n n oi. There exists a positive number r such that for
each X E an the ball {y; IY -xi ::; r} is contained in some oi. There exists
a natural number N such that N + 1 of Oi has an empty intersection. If
we denote the opening and the height of Oi by fh and hi respectively, then
inf ei > 0, inf hi > 0, sup hi < oo and diamOi ::; hi.
If n has the restricted cone property, then imbedding theorems and Rel-
lich's theorem hold if 1 < p < oo. The proof is based on Calderon's extension
theorem which is established with the aid of Theorem 2.10.
Chapter 4

Elliptic Boundary Value

Problems

4.1 Fundamental Solutions of Elliptic

Operators
The contents of this chapter are the LP estimates by S. Agmon, A. Douglis
and L. Nirenberg [11] for solutions of general elliptic boundary value prob-
lems. We begin with the fundamental solutions of elliptic operators with
constant coefficients by F. John [84].
The notations
D = (Dt, ... , Dn) = (ojoxt, ... , ojoxn),
Da = Df ·D~",f,a = f,f ···f.~", lal = a1 +'''+an
1 • • 1

for a= (at, ... , an)

are also used in this section as in the previous section.
A linear differential operator

L(x,D) = 2::::: aa(x)Da (4.1)

lal::=;m
of order m with coefficients defined in an open set n c Rn is called elliptic
inn if
2::::: aa(x)f,a-:/- 0 (4.2)
lal=m
for any x E 0, 0 -:f- f, ERn. A function K(x, y) defined in 0 x 0 possibly ex-
cept the diagonal { (x, x); x E 0} is called a fundamental solution of L(x, D)

121
122 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

if
L(x, D) kK(x, y)f(y)dy = f(x) (4.3)

holds for any f E CQ"(O). If L(x, D) = L(D) is an operator in Rn with

constant coefficients, a fundamental solution is of the form K(x - y). In
this case K(x) is called a fundamental solution of L(D).
As is well known
lxl 2-n /(2- n)On n>2
K(x) ={ (4.4)
(2n)- 1 log lxl n=2

is a fundamental solution of .t., where On is the surface area of the unit

sphere in Rn.
F. John [83],[84] constructed a fundamental solution for elliptic operators
with smooth coefficients. In this chapter we describe a case of operators
with only the highest order part of constant coefficients following [84].
Lemma 4. 1 Let q be a nonnegative integer such that n +q is even. Then
ifn is odd,

and if n is even,

is a fundamental solution of .t. (n+q)/ 2.

Proof. Suppose first that n is odd. Since

_t.(n+q)/2-llxlq = Cn,qlxl2-n,

where

and (4.4) is a fundamental solution of 6., we get for f E CQ"(Rn)

_t.(n+q)/2 r
}Rn
lx-ylqf(y)dy=.t.
}Rn
r _t.(n+q)/2-llx-ylqf(y)dy

= Cn,q.t. f lx- Yl 2-nf(y)dy = Cn,q(2- n)Onf(x).

}Rn
4.1. FUNDAMENTAL SOLUTIONS OF ELLIPTIC OPERATORS 123

Consequently (4.5) is a fundamental solution of _6.(n+q)f 2.

Analogously we can show that (4.6) is a fundamental solution of .6. (n+q)/ 2
in case n is even using
.6. (n+q)/2-1 (ixlq log jxi)

= (-l)n/22n+q-3r ( q; 2) r (n; 2) r (n; q) lxl2-n

if n > 2 and

.6.qf2(ixlq log jxi) = 2qr +2)

(~ 2
log lxl + const
if n = 2.
In order to construct a fundamental solution for elliptic operators F. John
[84] expressed the fundamental solution of .6. (n+q)/ 2 in the previous lemma
in terms of plane waves.
For the logarithm we always take the principal branch in the complex
plane slit along the negative real axis.
Lemma 4. 2 If q is a nonnegative integer such that n + q is even, then

- (2 \n I { (x~)q log X~2 M~ (4.7)

7r 2 q. }1~1=1

is a fundamental solution for .6. (n+q)/ 2.

Proof. First we show that if q is a nonnegative integer

q is odd
(4.8)
q is even

This follows from

r
}1~1=1
(x~)q log X~ da~ =
2
r
}l~l=1,x~>O
(x~)q (log(x~) - ii) M~
+ r
Jl~l=1,x~ <0
(x~)q (log lx~l + ii) da~
= f (x~)q (log(xO- ii) da~
Jl~l=1,x~>0
+( -l)q r
}l~l=1,x~>O
(x~)q (log(x~) + ~i) da~
2
124 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

= (1 + ( -1)q) r
Jl~l=1,x~>O
(x~)q log(x~)do-~

+((-1)q-1)ii r
Jl~l=1,x~>O
(x~)qdo-~

= 1+ ~1)q r
}1~1=1
lx~lqlogjx~ldo-~ + ((-1)q -1) 7r4i r
}1~1=1
lx~lqdo-~.
Next we show

and

r
}1~1=1
lx~lq log lx~ldo-~
= 27r(n- 1)/2r ( q; 1 ) lxlq (log lxl + Cn,q) j r ( n; q), (4.10)

where Cn,q is a constant depending only on n and q. With the aid of an

orthogonal transformation which maps x~ to lxl6 for a fixed x

r
}1~1=1
lx~lqdo-~ = lxlq r
}1~1=1
l6lqdo-~. (4.11)

We obtain (4. 9) with the aid of (4.11) and

r
}1~1=1
l6lqdo-~ = 2 r
}1~1=1,6>0
~~do-~

= 20n-1 1 1 tq(1- t 2 )(n- 3 )12 dt = On-1 1 1 s(q- 1)1 2 (1- s)(n- 3 )12 ds

= 0 n_ 1 B ( q; 1, n; 1) = 27rcn-1)/2r ( q; 1) I r ( n; q) .
Analogously (4.10) follows from

r
}1~1=1
lx~lq log lx~ldo-~ = r
}1~1=1
(lxll6i)q log(lxll61)do-~

= lxlq (log lxl r l6lqdo-~ + }1~1=1

}1~1=1
r l6lq log l6ldo-~) .
Now we turn to the proof that (4.7) is a fundamental solution of ~(n+q)/ 2 •
We show it only in the case n is even, since the case n is odd is easier. We
4.1. FUNDAMENTAL SOLUTIONS OF ELLIPTIC OPERATORS 125

denote (4.6),(4.7) by K 1(x), K(x) respectively. In view of (4.8),(4.10) and

r (- 2
1)
q +- r (q 2) - 1
-+2- -n-12q2-
q!

we have
K(x) = K1(x) + c~,qlxlq,
where c~,q is a constant depending only on n and q. By virtue of Lemma 4.1
K1(x) is a fundamental solution of b.(n+q)/2. Since lxlq is a polynomial of
degree q, we have b. (n+q)f 2 lxlq = 0. Hence K(x) is a fundamental solution
of b. (n+q)/2.
By virtue of Lemma 4.2 we have for¢ E C 0 (Rn)
1
</>(x) =- (27!"2")n q.1

xb.Cn+q)/ 2 { ¢(y)({ ((x-y)~)qlog(x-_y)~d~~)dy.(4.12)

}Rn 11~1=1 2

Let L(D) = Elo:l=m aaD" be an elliptic operator of order m with constant

coefficients consisting only of the principal part. Put L(~) = Elo:l=m aa~"
for~= (6, ... , ~n) ERn. Let q be a nonnegative integer such that n + q is
even. We are going to show that the function K(x) defined by
K(x) =b. (n+q)/ 2W(x), (4.13)

W(x) =
1 r (x.;)m+q x.;
(4.14)
(2ni)n(m + q)! }1~1=1 L(~) log Td~~

is a fundamental solution of L(D). If we put

s
F(s) = sm+q log "7
2

for real s, then

W(x) = - 1 { F(x~) da . (4 15)

(2ni)n(m + q)l }l~l=l L(~) ~ ·

(-dd)m
Since
(m+q)l s
F(s) = sq log + constsq,
1 "7
s q. 2

we have

L(D)F(x~) = L(~) [(m ~ q)l (x~)q log x~ + const(x~)q] .

q. 2
126 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Hence,

L(D)W(x) =

where P(x) is a homogeneous polynomial of degree q. Therefore for ¢ E

C0 (Rn)

L(D) { K(x- y)¢(y)dy = L(D)i:l. (n+q)/ 2 { W(x- y)¢(y)dy

}Rn }Rn

=fl. (n+q)/ 2 r [- (2nz~)nq! r

} Rn Jlel=l
[(x- y)~]q log (x ~ y)~ dae
z

+P(x- y)] ¢(y)dy

=- ~n i).(n+q)/ 2
(2nz) q!
r ¢(y) r
} Rn }lf.l=l
[(x-y)~]qlog(x~y)~duedy.
z

By (4.12) the last side is equal to ¢(x). Thus K(x) is a fundamental solution
of L(D).
Next we show that K(x) is analytic in x f:- 0. Let Tf be a fixed vector in
Rn such that I'Tfl = 1. For x such that lxl + X'Tf > 0 put

2(Tf ((x + lxi'Tf)

T((, x) = ( + ~x- lxl(lxl + X'Tf) (x + lxi'Tf).
Since~ = T((, x) is an orthogonal transformation for a fixed x and x~ =
lxi(Tf,
r (x~)m+q X~ r (lxi(Tf)m+q ( ('T/)
Jlel=l L(~) log --:;:due= Jlc';l=l L(T((,x)) log lxl--:;: due;

-
- lxl
m+q r ((Tf)m+q
log lxl }lc';l=l L(T((, x)) dac; + lxl
m+q r ((Tf)m+q ('T/
}lc';l=l L(T((, x)) log i dac;.

This show that W (x) is analytic in lx I + X'Tf > 0. Since Tf is arbitrary, W (x)
is, and hence K(x) is analytic in x f:- 0.
We denote by Kq(x) the fundamental solution constructed by (4.13),(4.14),
and investigate its dependence on q. Letting one of fl. operate under the
integral sign

K (x) = _ 1 f:l.(n+q-2)/2 { _1_

q (2ni)n(m + q)! }lf.l=l L(~)
4.1. FUNDAMENTAL SOLUTIONS OF ELLIPTIC OPERATORS 127

X [(m + q)(m + q- 1)(xt;)m+q-2 log ~~+(2m+ 2q- 1)(x~)m+q- 2 ] dn~

= Kq-2(x) + 'lj;q(x),
where

'lj;q(x) =
2m+ 2q - 1 6. (n+q-2)/2 r
(x~)m+q-2 d
(4.16)
(2ni)n(m + q)! }l~l=l L(~) a~.
If n is odd, then so is q- 2. Hence, the integral of (4.16) vanishes since the
integrand is an odd function then. If n is even and m < n, the right hand
side of (4.16) vanishes, since the integral is a polynomial of degree m + q- 2.
If n is even and m ~ n, 'lj;q is a homogeneous polynomial of degree m - n.
Therefore, in any case we have

(4.17)

for any derivative Dm of order m.

Theorem 4. 1 The fundamental solution K(x) of L(D) constructed by
(4.13), (4.14) has the following form:

K(x) = 'lj;(x) + p(x) log lxl,

=
where 'ljJ is a homogeneous function of degree m- n, p(x) 0 if either n is
odd or m < n, and p is a homogeneous polynomial of degree m- n if n is
even and m ~ n.
Proof. As in the proof of Lemma 4.2
128 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

q is odd

q is even

Hence, if n is odd

Wx =
1
( ) 4(2ni)n- 1 (m + q)!
1 lel=l
jx~jm+q
L(~) dCJ,
e
(4.18)

and if n is even

1 r (x~)m+q
(4.19)
W(x) = - (2ni)n(m + q)! Jlel=l L(~) log lx~jdrJe.

Therefore, if n is odd, W(x) is homogeneous of degree m + q, and hence

K(x) is homogeneous of degree m- n. Suppose n is even. Then

Hence if we put

we have
W(x) = '~h(x) + 'I/J2(x) log lxl,
'I/J1(x) is homogeneous of degree m + q, and 'lj;2(x) is a homogeneous poly-
nomial of degree m + q. Thus, we conclude the proof of the theorem.

Theorem 4. 2 The mth order derivatives of the fundamental solution K(x)

constructed by (4.13), (4.14) are homogeneous functions of degree -n, and
satisfy (2.2).
4.1. FUNDAMENTAL SOLUTIONS OF ELLIPTIC OPERATORS 129

Proof. The assertion of the theorem is obvious if n is odd in view of Theo-

rems 4.1 and 2.10. Ifn is even, we may assume q = 0 by (4.17), and hence
in view of (4.19)

- 1
K(x)-- (2 ')n ,~
n/21 (x~)m I I
L(C) log x~ da~. (4.20)
7r'l m. 1~1=1 .,
Since for lo:l = m
Da [(x~) m log lx~IJ = (m! log lx~ I + const )~a,

1
we get

Da 1 (x~)m
-(C) log lx~ld~ =
1~1=1 L ., 1~1=1
(m! log lx~l
~a
+ const) L(C) da~
.,

= m! 1 1~1=1
~a
L(C) log lx~lda~
.,
+ const

= m!
{ ~a { ~a I I
Jl~l= 1 L(~) da~ log lxl + m! Jl~l= 1 L(~) log j;f~ da~ + const.
X

With the aid of this relation and (4.20), and noting ~nf 2 log lxl = 0 we
obtain
a
D K x -
1_An/2
() _ _ _
(27ri)nu
{ L1 ~~C~d
Jl~l=1 L(~) og lxl"' a~.
The right hand side of this equality is obviously homogeneous of degree -n,
and satisfies (2.2) by Theorem 2.10.
Next, we show that for u E Clf(Rn)

f K(x- y)L(D)u(y)dy = u(x). (4.21)

JR"
As is easily seen K( -x) is a fundamental solution of L( -D). Hence for
¢ E CQ"(Rn)

f f K(x- y)L(D)u(y)dy¢(x)dx
JR" JR"
= f L(D)u(y) f K(x- y)¢(x)dxdy
JR" JR"
= f u(y)L(-D) f K(x- y)¢(x)dxdy = f u(y)¢(y)dy.
JR" JR" JR"
This implies that (4.21) is true.
We use the same notation I · lm,p,n, etc. as those introduced in section
3.4.
130 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Theorem 4. 3 For a natural number l such that l ~ m and 1 < p < oo

there exists a constant Cl,p such that for u E Cb(Rn)

(4.22)

Proof. With the aid of Theorem 4.1 and (4.21)

Dm- 1 u(x) = { Dr;- 1 K(x- y)L(D)u(y)dy.

}Rn
Applying Theorem 2.10 to Dm- 1K we get the conclusion for l = m:

(4.23)

If l > m, the conclusion is obtained by applying (4.23) to D 1-mu.

4.2 Assumptions and Main Result

Let L = L(x, D) = Lio:I:Sm aa(x)Da: be an elliptic operator of order m in
an open set n of Rn. The principal part of L(x, D) is denoted by L 0 (x, D):

L 0 (x,D) = L aa(x)Da:.
lo:l=m

Let ~' 'f/ be linearly independent real vectors. Owing to the ellipticity of L
the polynomial L 0 (x, ~ + T'fl) of the variable T has no real roots. In what
follows we assume the following conditions.
SMOOTHNESS CONDITION on 0. 0 is uniformly regular of class em.
SMOOTHNESS CoNDITION on L. For lal = m aa: is bounded and uniformly
continuous inn, and for lal < m aa: is bounded and measurable inn.
ELLIPTICITY CoNDITION. L is uniformly elliptic in n, i.e. there exists a
positive constant c such that for any X E n and ~ E Rn

(4.24)

RooT CoNDITION. For every pair of linearly independent real vectors~. 'f/
the polynomial L 0 (x, ~ + T'fl) of the variable T has equal number of roots
with positive imaginary part and with negative imaginary part.
Due to the Root Condition the order m of L is even, and L 0 (x, ~ + T'fl)
has exactly m/2 roots with positive imaginary part. It is easy to verify that
if n ~ 3 all elliptic operators satisfy the Root Condition, since if T is a root
for~' 'f/ then -Tis a root for -~, -ry and a sphere in Rn- 1 is connected.
4.3. ORDINARY DIFFERENTIAL EQUATIONS 131

Let Bj(x,D) = Ll.al:s;m3 bj,a(x)D.6,j = 1, ... ,m/2, be a set of linear dif-

ferential operators defined on 80. We assume mi < m for j = 1, ... , m/2.
SMOOTHNESS CONDITION on {Bj}';l;. For 1.8J :::; mj, j = 1, ... , m/2, bj,a is
m- mi times differentiable on 80 and the derivatives of order up to m- mi
are all bounded and uniformly continuous on 80.
CoMPLEMENTING CONDITION. Let x be an arbitrary point on 80 and v
be the outward normal unit vector to 80 at x. For each tangential vector
~ -1- 0 to 80 at x let T1 (x, ~), ... , Tm; 2 (x, 0 be the roots of the polyno-
mial L 0 (x, ~+Tv) with positive imaginary part. Then the polynomials,
in T, {BJ(x,~ + Tv)f;l; are linearly independent modulo the polynomial
n;l; (T- Tj(x,~)), i.e. a linear combination of {BJ(x,~ + Tv)}';l; is di-
visible by n;l; (T- Tj(X, ~))if and only if all the coefficients vanish, where
BJ is the principal part of Bi.
Following S. Agmon, A. Douglis and L. Nirenberg [11] it will be shown
that under the above conditions the following a priori estimate holds:

llullm,p,n :::; Cp [IlL(·, D)ullo,p,n

m/2
+ ~[Bj(·,D)u]m-m3 -l/p,p,an + llullo,p,n] (4.25)

for 1 < p < oo.

4.3 Preliminaries from the Theory of

Ordinary Differential Equations
Let l(T) = E'7=o am-jTi be a polynomial of order m. Suppose that l(T)
has exactly m/2 roots with negative real part. Let bj(T) = E~,; 1 bjkTk,j =
1, ... , m/2, be a polynomial of order mi. We are interested in solutions
belonging to £ 2 (0, oo) of the initial value problem

l (d) =~
dt u
m

J=O
diu
am-j dti = 0, t > 0, (4.26)

j = 1, ... ,m/2. (4.27)

Let .A 1 , ... , Ap and Ap+l, ... , Ap+q be the distinct roots of l(T) with negative
and positive real part respectively, and Vj be the multiplicity of Aj· The
132 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

general solution of (4.26) is

p+q Vj

u(t) = l:I:>j,ktk-le>.jt. (4.28)

j=lk=l

As is easily seen u(t) given by (4.28) belongs to £ 2 (0, oo) if and only if
Cj,k = 0 for j > p:
p Vj

u(t) = LLci,ktk-le>.jt. (4.29)

j=lk=l
If we put
p

l(T) = II (T- Ajtj'

j=l

then (4.29) is a general solution of

l(djdt)u(t) = 0. (4.30)

If the m/2 coefficients cj,k in (4.29) are uniquely determined by the initial
conditions (4.27), then the problem (4.30),(4.27) has a unique solution. We
wish to express this condition in an easily visible form.

Lemma 4. 3 Let p(T) be a polynomial of order f.t whose coefficient of the

highest order term is equal to 1:
p.

p(T) = l:a:JL-kTk, ll:o = 1.

k=O
For j = 0, 1, ... , f.t - 1 we put

Pj(T) = 2::: ll:j-kTk.

k=O
lf'Y is a rectifiable Jordan curve enclosing all the roots ofp(T) in its interior,
then for 0 ::::; j, k ::::; f.t - 1

_1_1
27r2. "!
PJL-1-j(T) kd _
( ) T T -
p T
$:.
UJk• (4.31)

Proof. If k < j, then pJl_ 1 _j(T)Tk is a polynomial of order not exceeding

f.t- 2. Hence, choosing 'Y to be a circle ITI = R and letting R -+ oo we
4.3. ORDINARY DIFFERENTIAL EQUATIONS 133

obtain (4.31). If k = j,

Hence, the right hand side of (4.31) is equal to

_1_1 dT - _.!.__ 1-1- ~a T" dT

27ri -y T 27ri -y Tp(T) ~ J-L-K •

The desired result in this case is obtained as in the case k < j. If k > j,
( ) k J-L-1-j
PJ-L-1-j T k T ~ K
p(T) T = p(T) ~ (XJ-L-1-j-KT

k-j-1 J-L- 1 -j k-j-1 J-L

_ T ~ K+j+1 _ T ~ K
- -(T) L...,; (XJ-L-1-j-KT -(T) ~ (XJ-L-KT
p K=O p K=J+1

k-j-1 ( j ) k-j-1 j
= ~(T) p(T)- l:aJ-L-KT" = Tk-j- 1 - ~(T) l:aJ-L-KT".
p K=O p K=O

Therefore, the right hand side of (4.31) is equal to

1
- 27ri
1 k-j-1
-y T p(T)
j
~ (XJ-L-KT"dT.
Hence, we obtain the desired resuslt also in this case as in the case k < j.
We set
m/2
~A k
= L...,; am/2-kT
A

l(T) ,
k=O
and for j = 0, ... , m/2- 1
j
A

lj(T) = L...,; aj-kT k •

~A

k=O

In view of Lemma 4.3 we have for j, k = 1, . .. , m/2

-21 . 1 1-lmj2-k(T)T3"-1 dT
A
= Djk, (4.32)
1r'l -y [ ( T)
-A
134 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

where "f is a rectifiable Jordan curve enclosing all the roots of [ Conse-
quently the function u defined by

(4.33)

is the solution of the initial value problem

i (!) u(t) = 0, -oo < t < oo, (4.34)

(
d
dt
)j-l u(O) = Cj, j = 1, ... , m/2. (4.35)

bj(T) = Pj(T)l(T) + bj(T). (4.36)

With the aid of (4.32) and (4.36) we get

Therefore if c1 , ... , cm; 2 satisfy

m/2
:2::: ckbJk = gj, j = 1, ... , m/2, (4.37)
k=l

then (4.33) is the solution of (4.30),(4.27). Hence

det (bjk) -1- 0 (4.38)

is a necessary and sufficient condition in order that the initial value problem
(4.26),(4.27) has a unique solution belonging to £ 2 (0, oo) for any g1, ... , gm/2·
The condition (4.38) is equivalent to the condition that bi (T), ... , b~ 12 (T)
are linearly independent or b1(T), ... , bm; 2(T) are linearly independent mod-
ulo the polynomial l(T), i.e. a linear combination ..\1b1 (T)+ · ·+.Am;2bm;2(T)
4.4. CASE OF CONSTANT COEFFICIENTS 135

is divisible by Z(T) if and only if .A 1 = · · · = .Am;2 = 0. Let (bik) be the

inverse matrix of (b.ik)· Then (4.37) implies ck = Ejl; bkJgj. Substituting
this into (4.33) we get

(4.39)

It is easy to show
(4.40)

4.4 Case of Constant Coefficients

In this section we consider the problem in a half space for operators which
have only the principal part with constant coefficients:

L(D)u(x) =L aaD"'u(x) = 0, x E R~, (4.41)

lal=m
Bj(D)u(x)l:~n=O =L bj,aD.6u(x)l:~n=O
l.61=mj
= gj ( x 1) , J· = 1, ... , m /2 , x' E Rn- 1 , (4.42)

where

R~ = {x = (x', Xn); x' = (x1, ... , Xn-1) E Rn-1, Xn > 0}.

We denote the Fourier transform of functions g E L 2 (Rn- 1 ) by

where f,' = (~1, ... • ~n- 1 ) E Rn- 1. The partial Fourier transform of func-
tions u E L 2 (R~_) with respect to x' is defined by

u(f,', Xn) = (27r)( 1 -n)/2 r

}Rn-1
e-ix'( u(x 1, Xn)dx 1•
136 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

If gi E L 2 (Rn- 1 ),u E Wm• 2 (R+.), then the problem (4.41),(4.42) is trans-

formed to
= 0, xn > 0,
L(i~ 1 , Dn)u(e, xn) (4.43)
BJ(ie, Dn)u(e, 0) = gJ(~ 1 ), j = 1, ... , m/2. (4.44)
Owing to the Root Condition the polynomial L(~ 1 , T) has exactly m/2 roots
with positive imaginary part, which we denote by T 1 (e), ... , Tm; 2 (e). We
define the following functions.
m/2 m/2
M+((, T) = II (T- Tj(e)) = :E a!,/2-k(e)Tk' (4.45)
j=1 k=O
j
Mf(e, T) = :E aj_k(e)Tk, j = 0, ... , m/2 -1, (4.46)
k=O
m/2
Bj((, T) = 2::: bjk(~ 1 )Tk- 1 is the remainder
k=1
of Bj(e, T) divided by M+(e, T),
(bik(~ 1 )) is the inverse matrix of (bjk(O),
m/2
Nj(e, T) = :E bkj(~ 1 )M;:: 12 _k(~ 1 , T). (4.47)
k=1
The coefficients at(e), k = 0, ... , m/2, are analytic functions of~~ i- 0 as
is shown below. For a fixed e
let 'Y be a rectifiable Jordan curve having
T1(e), ... ,Tm;2(e) inside and all the other roots of L(e,T) outside. Then

1
2ni
1
"Y
T
k g
U
L(~ 1 , T) aTL(~ 'T)dT
1
m~
~
= ~ Tj(~)
I k
(4.48)

for k = 0, 1, ... , m/2. We may assume that when e

is in a small neigh-
borhood of each nonzero point, 'Y can be chosen fixed. Then the left
e
hand side of (4.48) is an analytic function of in such a small neighbor-
hood. Hence, the elementary symmetric functions at (~ 1 ), k = 0, ... , m/2,
of T1 (~I), ... , Tm/2 (e) are also analytic.
We apply the result of the previous section to (4.43),(4.44). In place
of l(T),bJ(T) we have L(ie,T),BJ(ie,T). The polynomials l(T),lJ(T) are
replaced by im/ 2 M+(e, -iT), ii Mf((, -iT), and bj(T) is by
m/2
im3 Bj(~l' -iT) = :E imj-k+1bjk (~I)Tk-1.
k=1
4.5. POISSON KERNELS 137

Hence in place of the matrix (lJ.ik) we have

(imj-k+lbjk(~')rl = w-mk-llJ.ik(e))'
and hence im/ 2 -m3- 1Nj(e, -iT) in place of nj(T). Let 'Y be a rectifiable
Jordan curve enclosing all the roots of M+(e, T) for a fixed i- o. Then i'Y e
encloses all the roots of M+(~', -iT). Applying (4.39) to the present case
we get

By the change of the variable T ---+ iT

u(e
'
X)=~
n
___!__1 Nj(e,T) ei:r:,.TdT·i-m3g·(~').
L...J 21ri M+(CI T) J
(4.49)
j=l "( "' '

Corresponding to (4.40) we have

___!___
21ri
11
Ni(e, T)Bk(~', T) dT
M+(~',T)
= O·k·
3 (4.50)

4.5 Poisson Kernels

In this and the following two sections we consider the boundary value prob-
lem (4.41),(4.42). Changing the notations we denote by x = (x1, ... ,Xn-l)
and t generic points of Rn-l and R 1 respectively, and in consequence
R+. = {(x, t); x E Rn-1, t > 0},
Di=-,
a i=1, ... ,n-1, Dt=-,
a
axi at
Dx = {D1, ... , Dn-l), D = (Dx, Dt) = {D1, ... , Dn-l, Dt)·
In terms of these notations the problem {4.41),(4.42) is expressed as
L(D)u(x, t) = 0, x E Rn-l, t > 0, (4.51)
Bj(D)u(x, 0) = gj(x), x E Rn-l, j = 1, ... , m/2. (4.52)
The adjoint variable of x = (x1, ... , Xn-d is denoted by~= (6, ... , ~n-1)·
The relation (4.50) is written as

(4.53)
138 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

where"' is a Jordan contour enclosing all the roots of M+(.;, r). If u(x, t) is
the solution of (4.51),(4.52) belonging to wm• 2 (R~), then its partial Fourier
transform with respect to x is by (4.49)

The functions Kj, j = 1, ... , m/2, called Poisson kernels, are defined in R~
as follows: for mj 2 n - 1

Kj(X, t) = - 1
(2ni)n(mj - n
r
+ 1)! Jlel= 1
dne

x 1 N~(.;, r) (x.; + trr3-n+ 1 log

"' M (.;, r)
x.;-: tr dr,
z
(4.55)

and for mj < n - 1

Kj(X, t) = ( -1)n-1-mj (n- m~ - 2)! { dae

(2nz )n Jlel=1
X 1N~(.;,
"'M (.;, r)
r) (x.; + tr)m3-n+ 1dr, (4.56)

where dae is the area element of the unit sphere in Rn- 1 and"' is a Jordan
contour in the half plane Imr > 0 enclosing all the roots of M+(.;, r) for all
1.;1 = 1.
Let C>.,J.L be the constant such that
:>.

( _!}__)
dz
[z>-+J.L (log~z + C>. ,J.L )] = (A+
~-t!
~-t)! zJ.L log~z (4.57)

for nonnegative integers A, f.t. The existence of such a constant is easily seen.
It is also easy to show that for f.t < 0, A+ f.t 2 0

(d~)>. (z>-+J.LlogD = (-1)1+J.L(A+~-t)!(-1-~-t)!zJ.L (4.58)

For j = 1, ... , m/2 and a nonnegative integer q, let Kj,q(x, t) be a function

defined by: if mj 2 n - 1

Kj,q(x, t) = (2ni)n~j + q)! lel=1 dne (4.59)

X 1"'
Nj (.;, T) m +q (
M+(.;, r) (x.; + tr) 3 log
x.; + tr
i
) d
+ Cn+q-1,m;-n+1 T,
4.5. POISSON KERNELS 139

where Cn+q-l,mJ-n+l is a constant defined by (4.57), and if mj < n- 1

It is easy to verify that L(D)Kj,q(x, t) = 0, and if q is a nonnegative integer

with the same parity as n- 1
6_(n+q-l)/ 2 K·
z J,q
(x ' t)-
-
K·(x t)
J ' • (4.61)

We continue to use the notation Dk to denote kth derivatives. If k = 1

we simply write D.
Lemma 4. 4 Let F(x, t) be a function in c'l(R~_) such that

!Dk F(x, t)l :=:; Cjt, k = 0, 1, 2, (4.62)

on the hemishpere :E = {(x, t) E Rn; lxl 2 + t 2 = 1, t > 0}. Then F(x, t) is

bounded on :E.
Proof. Let (x, t) E :E, x f:. 0. Since

DF(O, 1)- DF(x, t) = 1 1

:rDF ( )1- r2 l:l, r) dr

= 1~ [}; ~~~DiDF()1-r i:l,r) 2

+DtDF( )1-r2 1: ,r) ]dr,

1

we get with the aid of (4.62)

IDF(O, 1) - DF(x, t)i :=:; const Jtt (~

1- r
+ _!) dr :=:; const (1 +log_!) .
r 2 t
Hence
IDF(x, t)i :=:; const(1 +!log ti). (4.63)
Repeating the same argument using (4.63) instead of (4.62) we obtain

IF(O, 1)- F(x, t)i :=:; const,

from which the conclusion follows.

140 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Lemma 4. 5 The kernels Kj,q are infinitely differentiable in except at R+

the origin. For each nonnegative integer l there exists a constant Ct such
that
ID1Kj,q (x, t) I ~ Ct (lxl 2 + t 2 ) (mj+q-l)/ 2 [ 1 + Ilog (lxl 2 + t 2 ) 112 1] . (4.64)
If l 2:: mj + q + 1, then D 1Kj,q is a homogeneous function of order mj + q -l,
and the logarithmic term in (4.64) may be omitted.
Proof. Let 'Y be a closed curve consisting of the arc {r; lrl = C, Imr 2:: c- 1 }
and the segment joining both of its end points, where C is a large positive
constant. Since Nj(f,, r)jM+(t;,, r) is bounded on { (f,, r); It;, I = 1, T E 'Y }, it is
not difficult to show that (4.64) holds on the unit sphere { (x, t); lxl 2 +t2 = 1}
if l ~ mj + q. From the definition of Kj,q it easily follows that (4.64) holds
also for general points (x, t). Suppose l 2:: mj + q + 1. Since

(! Y [zm3+q (log T+ Cn+q-1,mrn+1)] = (! Y(zm3+q log T)

= (-1)l-mrq+ 1(mj + q)! (l- mj- q -1)!zm3+q-l,

1
we have
D 1Kj,q(X, t) = r F(f,, r)(xf, + tr)mj+q-lda-edT,
l1e1=1 "~
_ ( -1)l-m3-q _ . _ _ I Nj(f,, T) l
F(f,, T)- (27ri)n (l mJ q 1). M+(f,, r) (f,, r) '
where (f,, r) 1 is the value obtained by replacing Di, Dt by f,i, Tin D 1 respec-
tively. Hence D 1Kj,q(x, t) is homogeneous of degree mj + q- l. The proof
will be complete if we show that this is bounded on the hemishere :E. For
that purpose it suffices to show in view of Lemma 4.4 that the inequalities
IDk D 1Kj,q(x, t) I~ Cjt, k = 0, 1, 2
holds on :E. Replacing k + l by l we are going to show that for l 2:: mj + q + 1
ID 1Kj,q(x, t)l ~ Cjt, (x, t) E :E. (4.65)
If t 2:: 1/2, then (4.65) clearly holds. Suppose 0 < t < 1/2. Let ( be a
function in c=([-1, 1]) such that ((r) = 1 for lrl ~ 1/2 and ((r) = 0 for
3/4 ~ lrl ~ 1. We write
D 1Kj,q(x, t) = I1 + I2,
11 = r
llel=1 "'
1(~(f,, T)) ((xf,)dnfne,
+ tT
1(~(f,,
X J.1.

h = r
llel=1 "' X
T\ (1- ((xf,))dTdrYe,
+ tT J.1.
4.5. POISSON KERNELS 141

where f-L = l - mi - q. Since lx~ +tTl ~ c for some constant c > 0 if

lx~l > 1/2, T E /,

lhl:::.; const r }-yrIF(~. T)lldTida~;.

l1e1=1
(4.66)

For a fixed x = (x 1, ... , Xn- 1) let Tx be an orthogonal transformation which

maps (x1, ... , Xn-d to (lxl, 0, ... , 0). Making the change of the variable
'f1 = Tx~ in the integral of I 1

I1 = 11 1711=1 'Y
F(Tx-1'fl,T) (I I ) d
(I I
X 'T/1 + tT
Y' ( x 'T/1 dT a71 •
Since 0 < t < 1/2 we have lxl ~ ..j3!4. Hence the integrand of the above
integral vanishes for I'T/1I > ..j3!4. Therefore writing 'f1 1 = ('T/2, ... , 'T/n-1) we
have

I1 - -1 171'1=1
da
71 ,
1..j374
-v'3;4
1 d'T/1
-y
F(T;- 1 iJ, T)((lxi'T/1)
(lxi'T/1 + tT)
J-L
( 1 _ 'T/2)n/2-2 dT,
1

r,
where = ('T/1. (1- ,.,n 112'T/'). Integrating by parts f-L- 1 times with respect
to 'T/1 yields

II11 = 1
(t-L- 1)!jxl~-'- 1 }171'1=1
1 r c1a 71' r-.fi74 d,1
J--.fi74
X i (lxi'T/1+tT)- 1 (a~JJ-L- 1 [F(T;- 1 fJ,T)((Ixi'T/1)(1-'fl~r/2 - 2 ] dT'
Noting lxl ~ ..j3!4, llxi'T/1 +tTl > timT and that (8 ;a,.,l)~-'- 1 [ ] is bounded
we obtain III :::.; C ft. Combining this with (4.66) we conclude (4.65).

Theorem 4. 4 ForgiE c~+L-mJ+ 1 (Rn- 1 ), j = 1, ... , m/2, l ~ maxmi

Ln-t
m/2 m/2
u(x, t) =~ Kj(X- y, t)gj(y)dy = ~ Kj * gj (4.67)

is a solution of (4.51), (4.52) belonging to c1(R+J

Proof. It is evident that the function defined by (4.67) satisfies L(D)u = 0
in t > 0. Set Uj = Ki * gi for j = 1, ... , m/2. We are going to show that

(4.68)
142 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Let q be the largest integer with the same parity as n - 1 and satisfying
q ::; l - mi + 2. Since n + q - 1 ::; n + l - mi + 1 we get with the aid of
(4.61) and integration by parts

Uj(X, t) = r
}Rn-1
~~n+q- 1 )1 2 Kj,q(x- y, t)gj(y)dy
= { Kj,q(x- y, t)~ (n+q- 1)1 2gi(y)dy. (4.69)
}Rn-1

Hence for 0 ::; k ::; l

Dkuj(x, t) = { Dk Kj,q(x- y, t)~ (n+q- 1)1 2gj(y)dy.

}Rn-1

Since q = l - mi + 1 or q = l - mi + 2, we see that mi + q - k > 0. Therefore

Dk Kj,q(x, t) is continuous in Rf- by Lemma 4.5, and hence Uj E C 1(Rf-).
By the definition of Kj,q

Bk(D)Kj,q(x, t) = (2 ') ( - 1 )I { dCJf.

7rZ n mj + q - mk . }lf.l=1

X 1N~(~,
"' M
r)
(~, r)
Bk(~, r)(x~ + tr)m3+q-mk (log x~: tr +canst)
z
dr.

Hence with the aid of (4.53)

(4.70)

It follows from (4.69)

Bk(D)uj(X, 0) = Ln- 1
Bk(D)Kj,q(X- y, 0)~ (n+q- 1)f 2gj(y)dy

= { Bk(D)Kj,q(y, 0)~ (n+q- 1)1 2 gj(X- y)dy

}Rn-1

= ~(n+q-1)/2 f gi(x- y)Bk(D)Kj,q(Y, O)dy

}Rn-1

= ~(n+q-1)/2 f gi(y)Bk(D)Kj,q(x- y, O)dy.

}Rn-1
4.5. POISSON KERNELS 143

Therefore in view of (4.70) we see that Bk(D)uj(x, 0) = 0 if k -::fi j. For

k = j by virtue of (4.12)

Bj(D)uj(x, 0) = - (2 .)~- 1 1~ (n+q- 1 )12 { gj(y)dy

7rZ q. }Rn-1

X { ((x- y)_;)q (log (x ~ y)_; + const) 00~

}1~1=1 z
= 1
(27ri)n-1q!
~(n+q-1)/2
}Rn-1 J
r
g·(y)

X { ((x- y)_;)q log (x ~ y)_; da~dy = gj(x).

}1~1=1 z
The proof of the theorem is complete.
It follows from (4.61) and Lemma 4.5 that

IDl Kj(X, t) I : :; Ct (lxl2 + t2/mJ-n+l-l)/2 [1 + llog (lxl2 + t2) 1/21] .

(4.71)
If l > mj - n + 1, the logarithmic term on the right hand side may be
omitted. Next we show that for x -::fi 0

Bk(D)Kj(x, 0) = 0. (4.72)

Let <P E C0 (Rn- 1 ) and 0 cf. supp¢. By (4.68) we see that

lim { Bk(D)Kj(x- y, t)¢(y)dy

t-+0 }Rn-1

=lim Bk(D) { Kj(x- y, t)¢(y)dy = Dkj¢(x).

t-+0 }Rn-1

Hence noting Kj E c= (.R:~ \ {0}) we get

Ln- 1
Bk(D)Kj(X- y, 0)</J(y)dy = 0.

Letting x = 0 we see that (4. 72) holds for x -::fi 0.

It follows from (4.71) and the mean value theorem that

!Bk(D)Kj(x, t)i:::; const · t (lxl 2+ t2)(mJ-mk-n)/ 2 [1 + llog (lxl 2+ t2)112 1]

(4.73)
In particular
144 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Lemma 4. 6 Set m 0 = rnaxmj. Suppose that for j = 1, ... , m/2 gj E

cm+n+Hmo-mj (Rn- 1 ) and fork = 0, 1, ... , m + n + 1 + mo - mj

Dkgj(x) =0 (lxlm-n-mj-k log lxl)

as lxl --+ oo. If we set

then iij E C 0 (il~), iij(x, 0) = Dm-m3gj(x).

Proof. Integrating by parts we get

Let ( be a function in CQ" (Rn- 1 ) such that ((x) = 1 for lxl < p where p is
a positive constant. If we set

W1 (x, t) = Ln-
m/2
~ 1
Kk(X- y, t)((y)Dm-mj gk(y)dy,

= ~ Ln-
m/2
Wz(X, t) 1
Bj(D)Kk(X- y, t)(1- ((y))Dm-mjgk(y)dy,

then
(4.74)
Since
(Dm-m3gk E c;+Hmo-mk+mj(Rn-1) C c;:+Hmo-mk(Rn-1),

we see Bj(D)w 1 E C0 (.R~) in view of Theorem 4.4, and

Bj(D)w1(x, 0) = ((x)Dm-m3gj(x).
In particular for lxl <p
Bj(D)w1(x,O) = nm-mjgj(x). (4.75)

Suppose lxl < p/2. If 1- ((y) # 0, then IYI 2:: p, and hence IYI/2 :::; lx- Yl :::;
3IYI/2. Therefore, in view of (4.73)

IBj(D)Kk(x- y, t)l :::; const tiylmk-mrn(I +I log IYII).

4.6. PRELIMINARIES ON INTEGRAL KERNELS 145

Consequently

lw2(x, t)l::; const I:> r

m/2

k=1 }IYI>P
1Yimk-mrn(1 + pog IYID

X 1Yimrmk-n(1 +I log lyll)dy = const t r

}IYI>P
IYI-2n(1 +I log IYII) 2 dy.

This shows that w2(x, t) -+ 0 as t -+ 0 for lxl < pj2. Combining this
with(4.74),(4.75) we obtain Uj(x, t)-+ Dm-m3gj(x) in lxl < p/2. Since pis
arbitrary we complete the proof.

4. 6 Preliminaries on Integral Kernels

Let K(x, t) be the following integral kernel:

where no is a bounded measurable function defined on the unit sphere

E = {x ERn-\ lxl = 1} and satisfying
h il(x)da- = 0. (4.77)

We investigate the partial Fourier transform of K(x, t) with respect to x.

For f.L > 0 set

K(x, t)
Kp,(x, t) = { 0 (4.78)

Using the polar coordinates as in Chapter 2 we write x = ro-, 1.;1 = p, x.; =

rpcos¢. Fort> 0

{ Kp,(x, t)e-ix€dx

l
}Rn-1

= il(a-)da- la0p, e-irpcoscj> (r2 + t2)(n-1)/2

r
n-2d
r

l 1
L:

11-P . 8 n-2ds
= il(a-)da- e-Mcoscj> _ _ _--:---::.,..-;=-
L: 0 (s2 + t2p2)(n-1)/2
= l L:
il(a-)da- 1o
11-P .
(€-MCOScj> _ €-8)
(s2
sn-2ds

+ t2p2/n-1)/2
• (4.79)
146 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Dividing the inner integral into the real and imaginary parts

1
o
J.LP .
(e-Mcos¢- e-s)
(s2
sn-2ds
+ t2p2/n-1)/2
= 11 + I2i, (4.80)

1
~ ~~~
It= (cos(scos¢)- e-s) ( 1);2 ,
(s2 + t2p2) n-

12 =- 1 J.LP
sin(scos¢)
sn-2ds
( 1);2 ·
(s2 + t2p2) n-

1
If p,p:::; 1
II1I:::;
1
icos(scos¢)- e-sl d: :=:; B, (4.81)

where B is the constant defined by (2.21). If p,p > 1,

II1I:::; B +I}1 {J.LP (cos(scos¢)- e-s) sn- 2 ~8

(s2+t2p2)n
_ 1)/2 1

+roo
:::; B
} 1
e_ 8 ds
s
+I t"p
}1
cos(scos¢) sn- 2 ~:-1)/21·
(s2 + t2p2)
(4.82)

Suppose cos¢ > 0. Making the change of the variable s ---+ s /cos¢ and
setting b = tp cos ¢

1 J.LP
cos(scos¢)
8 n-2ds
( _ 1)/ 2 =
(s2 + t2p2) n cos¢
1J.Lpcos¢
coss
8 n-2ds

(s2 + b2)
(n- 1)/2 .
(4.83)
If p,pcos¢:::; 1, the absolute value of the right hand side of (4.83) does not
exceed
{1 ds = log _1_.
lcos¢ s cos¢
If p,pcos¢ > 1, we divide the integral of the right hand side of (4.83) into
the part over (cos¢, 1) and that over (1, p,pcos¢). As for the first part

l1:s/oss(s2:::;~8- 1 )/2 1:::; 1:s¢ d: =loge~¢· (4.84)

Setting a = p,p cos¢ we write the second part as

1 J.LP cos</>
cos s
8 n-2 ds

(s2 + b2/n-1)/2
= 1
1
a ds
cos s - -
s +b

+ r cos s ( (s2 + sn-

}1
2
b2)(n-1)/2
- - 1-) ds.
s +b
(4.85)
4.6. PRELIMINARIES ON INTEGRAL KERNELS 147

By virtue of (2.19),(2.20) the first term of the right hand side of (4.85) is
estimated as

l +ba+b cos(s-b)-dss = cosb

la+b coss-+sinb
ds la+b sins-ds I
1+b s l+b s

::; 2A +log ( 1 + 2 ( 1: b)) + 2A::; 4A +log ( 1 + ~).

The second term does not exceed

lo oo I---...,---;-;-;::-
n-2
8
(s2 + b2)(n-1)/2
- -1- ds -
s+ b -
I loooo ----;----;-;-;::-
Tn-2
(T2 + 1)(n-1)/2
- - - IdT
+ 1
T
1
'

which is easily seen to be convergent. Hence we get

ll 1
J.LP
cos(scos¢)
8 n-2ds I 1
( )/ ::; const +log --:i:·
(s2 + t2rJl-) n- 1 2 cos 'I'

Estimating analogously in case cos¢> < 0 we obtain

1
lh I ::; const + log Icos ¢I' (4.86)

In case cos ¢> > 0

I2 = - lo asin s ----,----,-,--=-
sn-2ds
(s2 + b2)(n-1)/2

- - r sin s-d_s_
s+b }
- r sins ( sn- 2
(s2 + b2)(n-1)/2
- - 1-) ds
s+ b '

1r·
- }0 0

II sins 8 ~ bl ~ sin(s- b) ~I
= Icos b lb a+b sins-;
ds la+b cos s-;ds I
-sin b b

::; 4A +I sinbllog ( 1 + ; ) ::; const,

loa sins ( (s2 + b2)(n-1)/2

n-2 8
- -1- ) ds
+b 8

< lo oo I n-2 8
- -1- Ids = const.
- (s2 + b2)(n-1)/2 +b 8
148 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Hence II2I :::; const. Thus we have shown that

rp (e-is
Jo
cos c/>
-e
-s) sn- 2ds 1
(s2+t2p2/n-1)/2 :::;const+loglcos¢1' (4.87)

Combining this with (4. 79)

ILn- 1 KJL(x,t)e-ixedxl:::; ~ IO(CT)I (const+log lc~¢l) dCT:::; const.

(4.88)
Since
•
I1m
JL--+oo
1 /LP
( e -iscosc/> - e -s) S
n-2d
(s2 + t2p2)(n-1)/2
S

=lim
JL--+oo
1 JLP .
(e-Mcosc/>_e-s)-
ds

!)
S

+ lim rJLP (e-iscosc/>- e-s) ( sn-2 - ds

JL--+oo } 0 (s2 + t2p2)(n-1)/2 8

= roo ( e-iscosc/> - e-s) ds

Jo s
+ roo (e-iscosc/>
}0 -
e-s) ( sn-2
(s2+t2p2)(n-1)/2--;
1) ds
'

exists if cos¢ f:. 0, we conclude with the aid of (4.79),(4.87),(4.88) that

r
} Rn- 1
K(x, t)e-ixedx = lim
JL--+OO lxl <JL
1
K(x, t)e-ixedx

exists and is bounded.

It is easily seen that the integral fRn- 1 K(x - y, t)f(y)dy is absolutely
convergent for f E L 2 (Rn- 1 ), t > 0. Arguing as in section 3 of Chapter 2
we can establish the following theorem.
Theorem 4. 5 For f E L 2(Rn-l)

(K(·, t) * f)(x) = r
} Rn-1
K(x- y, t)f(y)dy = lim r
JL--+00 } Rn-1
KJL(x- y, t)f(y)dy

exists in the strong topology of L 2(Rn-l ). The mapping f 1-+ K(·, t) * f

is a bounded linear transformation from L 2(Rn-l) to itself and its norm is
bounded in t > 0:

IK(·, t) * flo,2,Rn-1 :::; Clflo,2,Rn-1.

4.6. PRELIMINARIES ON INTEGRAL KERNELS 149

Next we consider the following more general integral kernel defined in the
half space R'J:_:

K(x, t) = 0 ( jPj'
X t ) /( lxl 2 + t 2)(n-1)/2 ,
jPj (4.89)

where P = (x, t), IPI = (lxl 2 + t 2) 112 . O(x, t) is a function defined and
continuous on the hemisphere lxl 2 + t 2 = 1, t ~ 0, and satisfies a uniform
Holder condition at points on the planet= 0: for lxl = 1, IYI 2 +t2 = 1, t ~ 0

IO(x, 0)- O(y, t)l ::; c1 (lx- Yl 2 + et12 ' c1 > 0, 0 < p::; 1. (4.90)
Furthermore it is assumed that

h O(x, O)do- = 0, (4.91)

where :E is the unit sphere of Rn- 1 • Set

K1 ( x, t ) = 0 ~· 0
X ) /(
lxl 2 + t 2)(n-1)/2 ,

(X~,0)] I (lxl 2 +t2)(n-1)/2 .

(

K2(x,t)= [ 0 (
X t ) -0
jPj'jPj
Then
K(x, t) = K1(x, t) + K2(x, t). (4.92)
Since

we have

Hence

r
}Rn-1
IK2(x, t)l dx::; 2P/ 2C1tP r
}Rn-1
d~n-1+
(lxl2 + t2)
)/2 ::; c2.
P
(4.93)

For f E L 2 (Rn- 1 ) we set

u(x, t) = Ln- 1
K(x- y, t)f(y)dy.
150 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Then

u(x, t) = kn- K1(x- y, t)f(y)dy +

1
kn-
Kz(X- y, t)f(y)dy
1

= (K1(·, t) * f)(x) + (Kz(·, t) * f)(x).

By virtue of Theorem 4.5 we have

IK1(·, t) * flo,z,Rn-1 :::; Glflo,z,Rn-1.

With the aid of (4.93)

IKz(·, t) * flo,z,Rn-1 :::; IKz(·, t)lo,1,Rn-1 lflo,z,Rn- 1 :::; Czlflo,z,Rn-1.

Hence we have established the following theorem.
Theorem 4. 6 Let K(x, t) be the integral kernel of (4.89). Then for f E
Lz(Rn-1)
IK(·, t) * flo,z,Rn-1 :::; Clflo,z,Rn-1.
Here C is a constant independent off, t.
For a positive integer k and 1 < p < oo let wk- 1/P,P(Rn- 1) be the space of
the traces on { (x, 0) E Rn; x E Rn- 1} of functions belonging to Wk,P(R~)
(Chapter 3, section 10). For f E wk- 1/P,P(Rn- 1) set

lflk-1/p,p,Rn-1 = inf{lvlk,p,Rf_; v E Wk,p(R~), v(·, 0) = f}. (4.94)

Here by definition

Lemma 4. 7 Let G(x, t) be a measurable function defined in the half space

R~. Sets+ = {(x, t) E Rn; lxl 2 + t 2 = 1, t > 0}. Suppose that for some
nonnegative measurable function f!(x, t) defined on s+ and satisfying

r f!(x, t)dax,t <

Js+
00

we have
IG(x, t)l < f!(x/IPI, t/1~1)' p = (x, t).
- (lxl2 + tzt 2
For v E LP(R~), 1 < p < oo, set

u(x, t) = J inn G(x- y, t

+
+ s)v(y, s)dyds.
4.6. PRELIMINARIES ON INTEGRAL KERNELS 151

Then u E LP(R't-) and the following inequality holds:

Jkn +
iu(x, t)IPdxdt:::; C Jkn +
lv(x, t)IPdxdt.

Proof. SetS= {(x, t) ERn; lxl 2 + t2 = 1} and

O(x/IPI, t/IPI) (lxl 2 + t2 ) -n/2 t>O

M(x, t) = {
-0( -x/IPI, -t/IPI) (lxl 2 + t2 ) -n/2 t<O
then M(x, t) is homogeneous of order -nand

1s IM(x, t)idax,t < oo.

If we extend v(x, t) as 0 outside R'f- to the whole of Rn, then for t >0

Jk.n
iu(x, t)l:::; M(x- y, t + s)iv(y, s)idyds

= Jkn
+
M(x- y, t + s)iv(y, s)idyds = u*(x, t).

By virtue of Lemma 2.1 the above integral exists almost everywhere in R'f-.
Applying Theorem 2.6 we conclude

Jk.n+
iu(x, t)IPdxdt:::; Jk.n iu*(x, t)IPdxdt:::; const Jk.n +
lv(x, t)IPdxdt.

Theorem 4. 7 Let K(x, t) be the integml kernel of (4.89). Suppose more-

over that K(x, t) has continuous derivatives bounded on the hemisphere
lxl 2 + t2 = 1, t 2:: 0. For f E W 1 - 1/P.P(Rn- 1 ), 1 < p < oo, set

u(x, t) = kn-l K(x- y, t)f(y)dy.

Then for some constant C independent off the following inequality holds:

(4.95)

Proof. By the definition of the seminorm (4.94) there exists a function

v E W 1 ·P(R'f-) such that

v(x, 0) = f(x), lvh,v,R+ :::; 2lfh-1/p,p,Rn-1. (4.96)

152 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Since

v(x, t) = f(x) +it D 8 v(x, s)ds,

r Irt Dsv(x, s)dslp dx:::::; }Rn-1

}Rn-1 Jo
r tP- 1 Jot IDsv(x, sW dsdx
1 r ID v(x sW dxds < tP- IviP
-< tP- }Rn
1
s ' - 1,p,R+'
+

we have

(Ln- 1
1
lv(x, t)IPdx) /p:::::; (Ln- 1
1
lf(x)IPdx) /p + t 1- 11Pivh,v,R+. (4.97)

Noting (4.96)

u(x, t) = kn- 1
K(x- y, t)v(y, O)dy

=- { {T D 8 {K(x- y, t + s)v(y, s)}dsdy

}Rn-1 Jo

+ kn- 1
K(x- y, t + T)v(y, T)dy
=- { {T D 8 K(x- y, t + s) · v(y, s)dsdy
}Rn-1 Jo
- { {T K(x- y, t + s)D8 v(y, s)dsdy
}Rn-1 Jo
+ kn- 1
K(x - y, t + T)v(y, T)dy.

By a suitable change of variable we get

Diu(x, t) = - { {T D 8 K(x- y, t + s) · Div(y, s)dsdy

}Rn-1 }o
T
- { { DiK(x- y, t + s) · D 8 v(y, s)dsdy
}Rn-1 }o
+ kn- 1
DiK(x- y, t + T) · v(y, T)dy. (4.98)

Since
(4.99)
4.6. PRELIMINARIES ON INTEGRAL KERNELS 153

it follows that

(r
1/p'
IDiK(y, t + T)lp' dy) :::; const(t + T)- 1-(n- 1)/P.
}Rn-1

Combining this with (4. 97) we get

I.Ln-1 DiK(x- y, t + T) . v(y, T)dyl = 0 ( r-nfp)

as T ~ oo. Hence letting T ~ oo in (4.98) we find

Diu(x, t) = - J.L. D 8 K(x- y, t + s) · Div(t, s)dsdy

-J.L.
+
DiK(x- y, t + s) · D 8 v(y, s)dsdy. (4.100)
+
Analogously to (4.99) we have

Hence applying Lemma 4. 7 we obtain

Jk+
1/p
( )
IDiu(x, tW dxdt :::; constlvh,v,R+ :::; constlfh-1/p,p,R"-t.

Remark 4. 1 In the proof of Theorem 4.7 the condition (4.91) is not nec-
essary. It is shown in [11] that if DlK(x, t) exists and is bounded on the
hemisphere :E, then the inequality (4.95) also holds for Dtu.
We verify that the theorems of this section can be applied to the kernel

K(x ' t) -- Dm3+q+n- 1 K J,q

· (x ' t) • (4.101)

By Lemma 4.5 this kernel K(x, t) is homogeneous of order 1 - n. It is

easily seen that K(x, t) is expressed in the form (4.89) with a function
Q(x, t) which is infinitely differentiable on the hemisphere S. Since L(D)
is elliptic, the coefficient of Df in L(D) is different from 0. Hence the
relation L(D)K(x, t) = 0 can be rewritten as

+L
n-1
Df' K(x, t) ckDkDa(k) K(x, t) = 0. (4.102)
k=1
154 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Here ck is a constant and la(k)l = m- 1. Since

lno:(k) K(x, t) I: ; const (lxl 2 + t 2 ) (2 -m-n)/2

in view of Lemma 4.5, we see

{ DkDo:(k) K(x, t)dx = { Da(k) K(x, t) xk dS---+ 0

Axi5.R Jlxi=R R
as R ---+ oo. Hence
{ Df K(x, t)dx = 0. (4.103)
}Rn-l
Set for j = 1, 2, ...

As is easily seen gi(t) = gj(1)ri. Therefore

gj(t) = Dt9j-1(t) = Dt (9j-1(1W-i)

= (1- j)gj-1(1)ri = (1- j)r 1gj-1(t). (4.104)

Since 9m(t) =0 by (4.103), we get g (t) =0 from (4.104), or

1

{ DtK(x, t)dx = 0. (4.105)

}Rn-l

If we decompose K(x, t) as (4.92), then K 2 (x, t) is absolutely integrable in

Rn- 1 with respect to x and the value of the integral does not depend on t.
Hence from (4.105) we get

{ DtK1(x, t)dx = 0.
}Rn-l

On the other hand

{
j Rn-t
DtK1 (x, t)dx = {
j Rn-l
0 (-lxl,
X
o) Dt (lxl 2 + t 2 ) ( 1 -n)/2 dx

= (1- n)t { n(~.

lxl
o) (lxl + 2 t 2 ) -(n+ 1)/ 2 dx

hO(a, la +
}Rn-l
oo rn-2dr
= (1- n)t O)da ( +1 );2 •
~ o (r2 t2) n

h
Hence we conclude
O(a, O)da = 0.
4. 7. NONHOMOGENEOUS BOUNDARY CONDITIONS 155

4. 7 Nonhomogeneous Boundary Conditions

Let u be a function in C0 (il'+) satisfying

L(D)u(x, t) = f(x, t) t > 0, (4.106)

Bj(D)u(x, 0) = gj(x) j = 1, ... ,m/2. (4.107)

Let K(x, t) be the fundamental solution of L(D) stated in section 1. As in

section 7 of Chapter 3 we extend the function f to Rn by

f(x, t) t;:::o
X t - {
f( ' ) - t;
N+1
>..kf(x, -kt) t<O
(4.108)

N+1
2)-k)i>..k = 1, j = 0, 1, ... ,N,
k=1

and N is taken sufficiently large. If we put

v(x, t) = JL. K(x- y, t- s)f(y, s)dyds, (4.109)

then v E cN+m- 1 (Rn) and satisfies L(D)v =f. Furthermore, vis infinitely
differentiable outside suppf and by Theorem 4.1

(4.110)

as lxl 2 + t2 ---+ oo. If we set

(4.111)

then hi E cN+m- 1-mj (Rn- 1 ). Moreover, hJ is infinitely differentiable

outside some bounded set and

(4.112)

as lxl ---+ oo. Since

L(D)(u-v) = 0, t > 0,
Bj(D)(u-v)=gJ-hj, j=1, ... ,m/2, t=O,
156 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

one might expect that the following relation holds:

Ln-
m/2
u(x, t) = v(x, t) +~ 1
Kj(X- y, t) (gj(Y)- hj(y)) dy.

However, the convergence of the integral of the right hand side is not evident.
Therefore, we show the following theorem instead.
Theorem 4. 8 Let u be a function in CQ"(R~) satisfying (4.106), (4.107).
Then the relation

Dmu(x, t) = Dmv(x, t) + ~
m/2
Ln- 1
DmKj(X- y, t) (gj(Y)- hj(y)) dy

(4.113)
holds in the half space t > 0. Here v,hj are functions defined by (4.109),
(4.111) respectively.
If we put
u- v = w, gj - hj = BjWit=O = Wj, (4.114)
then what is to be shown is

Ln-
m/2
Dmw(x, t) = ~ 1
DmKj(X- y, t)wj(y)dy. (4.115)

Since in view of Lemma 4.5

DmKj(X, t) = 0 ({lxl2 + t2)(mrm-n+l)/2log (lxl2 + t2)1/2)' (4.116)

as lxl 2 + t 2 ---+ oo, and by (4.112)

Dkwj(x) =0 (lxlm-n-mJ-k log lxl) (4.117)

as lxl ---+ oo for k = 1, 2, ... , N + m- 1- mj, the integral on the right of

(4.115) converges.
In order to prove Theorem 4.8 we prepare the following two lemmas.
Lemma 4. 8 Let u be a function belonging to cm(_R~) and satisfying

L(D)u(x, t) = 0, t > 0, (4.118)

Bj(D)u(x, 0) = 0, j = 1, ... ,mj2. (4.119)

Suppose that u and its derivatives of order up tom are absolutely integrable
on each hyperplane t = canst > 0, and those integrals are uniformly con-
vergent in t in every finite interval 0 < E :::; t :::; R. Moreover suppose that u
4. 7. NONHOMOGENEOUS BOUNDARY CONDITIONS 157

and its derivatives of order up tom belong to L 2 (Rn-l) on each hyperplane

t = const 2 0 and their norms

0 :::; k :::; m, t>0

are bounded. If in additon u E L 2 (Rf-), then u = 0.

Proof. Let u(~, t) be the partial Fourier transform of u with respect to x.
The assumptions imply that u and its derivatives in t of order up to m
are continuous functions of(~, t) in t > 0. It follows from (4.118) and the
hypotheses that
L( i~, Dt)u(~, t) = 0 (4.120)
fort> 0. For almost every~ we have J000 1'11(~, t)1 2 dt < oo. Hence by what
was stated in section 3 we get for these values of ~
M+ (i~, Dt) u(~, t) = 0. (4.121)
Since the left hand side of (4.121) is a continuous function of (~,t), this
holds for every ~- It also follows that u(~, t) is infinitely differentiable in
t 2 0 for each fixed f If we verify that for almost every ~
Bj(i~, Dt)u(~, 0) = 0, j = 1, ... , m/2, (4.122)
then we conclude by the argument of section 3 that u(~, t) 0, and hence
u =
=
0. By hypothesis {Bj(D)u(·, t)} is a bounded subset of L 2 (Rn- 1 ).
Therefore noting that u E cm(Rf-) and (4.119) we see that {Bj(D)u(·,t)}
converges weakly to 0 in £ 2 (Rn-l) as t --+ 0, and hence so does
{Bj(i·, Dt)u(·, t)}. Consequently there exists a subsequence {tv} tending
to 0 such that some sequence of convex combinations of {Bj(i·, Dt)u(·, tv)}
converges to 0 almost everywhere. On the other hand for every fixed ~ we
have Bj(i~, Dt)u(~, t)--+ Bj(i~, Dt)u(~, 0) as t--+ 0 since u(~, t) is a smooth
function of t 2 0 for each fixed ~ as was remarked above. Thus (4.122)
follows and the proof is complete.
Lemma 4. 9 Let {Wi 1 , ... ,i,.; i 1, ... , im = 1, ... , n} be a set of functions be-
longing to C 1(Rf-) which are invariant under permutations of i 1, ... , im. If
for every i1, ... , im+l = 1, ... , n

then there exists a function g E cm+l(Rf-) such that for every il, 0 0 0 'im =
1, ... ,n
wit, ... ,i,. = Dit ... Dim. g.
We can choose such a function g so that all of its derivatives of order up to
m-1 vanish at the origin, and under this condition g is uniquely determined.
158 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Proof. (i) Case m = 1. Suppose that w 1 , ... , Wn are functions in C 1 (R+.)

satisfying DjWi = DiWj for i, j = 1, ... , n. If we put

g(x)
nlax;
= g(x1, ... , Xn) = 2.:::::: wi(x1, ... , Xi-1, Yi, 0, ... , O)dyi,
i=1 °
then

Djg(x) = Wj(X1, ... , Xj-1, Xj, 0, ... , 0)

+ ~lax;
LJ Djwi(x1, ... ,xi-1,Yi,o, ... ,o)dyi
i=j+1 °
= Wj(X1, ... , Xj-1, Xj, 0, ... , 0)
+ ~lax;
LJ Diwi(x1, ... , xi-1, Yi, 0, ... , O)dyi
i=j+1 °
= Wj(X1, ... , Xj-1, Xj, 0, ... , 0)
n
+ 2.:::::: {wj(X1, ... ,Xi-1,Xi,0, ... ,0) -Wj(X1, ... ,Xi-1,0,0, ... ,0)}
i=j+1
= Wj (x1, .. . , Xn)·
This function g satisfies g(O) = 0. It is easy to show that g is uniquely
determined by this condition.
(ii) Suppose the assetion of the lemma is true for m - 1. Then for each
i = 1, ... 'n there exists a function gi E cm(R+.) such that

and all derivatives of gi of order up to m - 2 vanish at the origin. For

i,j = 1, ... , n

Hence if m = 2 Djgi = Digi. If m 2 3, Djgi - Digj is a polynomial of

degree m- 3 at the most, and all of its derivatives of order up to m- 3
vanish at the origin. Hence Djgi = Digj also holds in this case. By virtue
of the result of (i) there exists a function g such that gi = Dig and g(O) = 0.
It is easily seen that this function g is the one sought.
Proof of Theorem 4.8. Let q be an integer with the same parity as n- 1
such that mi + q 2 2m + 1 for j = 1, ... , m/2, and N an integer satisfying
N 2 mo + q - m + n, where mo = maxi= 1, ... ,m;2 mi. Then

N 2 m + n + 1 2 m + 3, N 2 mo + n + 2. (4.123)
4. 7. NONHOMOGENEOUS BOUNDARY CONDITIONS 159

Form::; l::; 2m+ 1 we get with the aid of (4.61),(4.64),(4.117)

Ln- 1
D 1Kj(X- y, t)wj(y)dy

= { D 1 .6.~n+q- 1 )1 2 Kj,q(x- y, t)wj(y)dy

}Rn-1
= { D 1Kj,q(x-y,t)·.6.(n+q- 1 )1 2 wi(y)dy. (4.124)
}Rn-1
HerewenoteWj E cN+m- 1 -mj(Rn- 1 ) c cn+q- 1 (Rn- 1 ). Sincemj+q-l2
0, we find with the aid of (4.64) that the last side of (4.124) is continuous
in R+. Hence, if we set

r Dit ... Di.,Kj(X- y, t)wj(y)dy,

m/2
Wit, ... ,i.,(X, t) = L }Rn-1
j=1

then wit, ... ,i., E cmH(.R+) and we can apply Lemma4.9 to the set {wit, ... ,i,J
to find that there exists a function g E C2m+ 1 (R'+) satisfying

Ln-
m/2
Dmg(x, t) =~ 1
Dm Kj(X- y, t)wj(y)dy. (4.125)

Hence it suffices to show that

Dm(w-g)=O. (4.126)

In view of (4.117),(4.123) we can apply Lemma 4.6 to obtain

(4.127)

Suppose that the following statements are true:

(i) form::; l::; 2m+ 1 D 1g(·, t) E L2 (Rn- 1 ) and its norm ID1g(·, t)l 0 2 Rn-1

is bounded in t 2 0, ' '

(ii)
JLn +
IDxDmg(x, t)l 2 dxdt < 00,

ID
(iii) for m + 1 ::; l ::; 2m+ 1 the integral fRn- 1 1g(x, t)l dx converges
uniformly in any finite interval 0 < E ::; t ::; R < oo.
Then we can prove (4.126) as follows. Since in view of (4.110) we have
160 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

form:::; l:::; 2m+ 2(:::; N + m- 1), the above statements (i),(ii),(iii) holds
also for w. If we set h = w- g

L(D)h(x, t) = 0, t > 0, (4.128)

D;"-mJ Bj(D)h(x, 0) = 0, j = 1, ... , m/2. (4.129)

Hence noting hE C2 m+ 1 (R~) we have

L(D)D;"+ 1 h(x, t) = 0, t > 0,

Bj(D)D;"+ 1 h(x, 0) = D;"J+ 1 D;"-mJ Bj(D)h(x, 0) = 0, j = 1, ... , m/2.

We can apply Lemma 4. 8 to Dr;"+ 1 h to obtain nr;+ 1 h = 0. This implies that

Dr;"h depends only on t. However, Dr;"h E L 2 (Rn- 1 ) on each hyperplane
t = const in view of (i). Hence
Dr;"h = 0. (4.130)

From (4.128) it follows that DtD';- 1 L(D)h = 0. Rewriting this like (4.102)
we have

0= D,D'J:-' ( D;" + ~ c,D,Da(k)) h

n-1
= Df+ 1Dr;"- 1 h + I: ckDtDa(k) Dr;"- Dkh 1

k=1

and using (4.130) we get

(4.131)

Next rewriting D~Dr;- 2 L(D)h = 0 as

0 = Di D;"- 2 ( Df + ~ ckDkDa(k)) h

n-1
= Df+2 Dr;"- 2 h +I: ckDi Da(k) D;"- Dkh 2

k=1

and using (4.131) we get Df+2 Dr;"- 2 h = 0. Combining this with (4.131) we
obtain Dm+2Dr;"- 2 h = 0. Proceeding in this manner we conclude D 2 mh =
0. This implies that his a polynomial. However, in view of (i) Dmh belongs
to L 2 (Rn- 1 ) on each hyperplane t = const. This implies Dmh = 0, and
(4.126) is established.
4.7. NONHOMOGENEOUS BOUNDARY CONDITIONS 161

Thustheproofwill becompleteifweshow (i),(ii),(iii). Form::; l::; 2m+1

we have in view of (4.125)

Dlg(x, t) =
m/2
~ kn- 1
Dl Kj(X- y, t)wj(y)dy.

Hence setting

we are going to prove

r
}Rn-1
IIj(X, t)l 2 dx::; const, t ~ 0, (4.132)

Jinn +
IDxlj(X, t)l 2 dxdt < 00, (4.133)

r
}Rn-1
IDij(X, t)i dx is uniformly convergent
in 0< E ::; t ::; R if m ::; l ::; 2m. (4.134)

In case l - mj is even, we choose q so that n + q - 1 - l + mj is even (this

q may be different from the one which appeared at the beginning of the
proof) we set
K(x t) -Dlb.,(n+q- 1-l+mj)f2 K· (x t)
' - X J,q ' •

The theorems of the previous section can be applied to this kernel according
to the remark at the end of the section. With the aid of (4.64),(4.117) we
can integrate by parts in

to derive
Ij(X, t) = r
}Rn-1
K(x- y, t)w(y)dy

where w = b.,Cl-mj)f2 wj(y). Note that wE C 1(Rn- 1) since

Wj E cN+m-1-mj(Rn-1) C cl-mj+1(Rn-1)

in view of (4.123). Furthermore by (4.117)

w(y) = 0 (IYlm-n-l), Dw = 0 (IYlm-n-l-1) (4.135)
162 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

as IYI oo. Hence w E W 1•2 (Rn- 1 ). Letting ( be a smooth function

---+
such that ((0) = 1 and ((t) = 0 for t ~ 1 set vo(y, t) = w(y)((t). Then
vo E W 1 • 2 (R~_). This implies wE W 112 •2 (Rn- 1 ). Applying Theorems 4.6,
4.7 we conclude that (4.132),(4.133) are true. In view of Lemma 4.5 we have

IDK(x, t)i $ const (lxl 2 + t 2 ) -n/2 •

It follows from (4.135) that wE L 1 (Rn- 1 ). Hence

r
Ax!?_>.
IDij(X, t)i dx = r Ir
Ax!?.>. }Rn-1
DK(x- y, t)w(y)dyl dx

$ C { { (lx - Yl 2 + t 2 ) -n/2 iw(y) idydx

Jlx!?.>. }Rn-1

$ C { {
}!y!::;_l-' Jlx!?.>.
(lx- Yl 2 + t 2 rn/ 2 dxiw(y)idy

+C { { (lx- Yl 2 + t 2 ) -n/ 2 dxiw(y)idy = I1 + }z.

j!YI ?.1-< J1x1 ?.>.
By an obvious change of variable

Hence I2 can be made arbitrarily small uniformly in t ~ E by letting fJ,

sufficiently large. If .:\ > f.J,,

The right hand side is independent of t and tends to 0 as .:\ ---+ oo. Thus
(4.134) is established. In case l - mi is odd, we set

K-(x
t !
t) = DlD·fl(n+q-l+mj-
t X
2 )1 2 K· (x t)
J,q ' '
Wi(Y) = Dill~-mj- 1 )1 2 wi(y).

=I: h
Then
n-1
Ij(x, t) ki(x- y, t)wi(y)dy.
R n-1
i=1

Hence we can show (4.132),(4.133),(4.134) also in this case.

4. 7. NONHOMOGENEOUS BOUNDARY CONDITIONS 163

Theorem 4. 9 Let u E wm,P(Rn), 1 < p < oo, l ;::: m. Furthermore sup-

pose that u vanishes outside a bounded set. If we set
L(D)u = J,
Bj(D)uit=O = gj, j = 1, ... , m/2,
then there exists a constant Cl,p which does not depend on u such that

Here the seminorm I ·il-mrl/p,p,R"-1 is defined by (4.94).

Proof. We may assume u E CQ'(R'f_). In accordance with Theorem 4.8 we
write
m/2
D 1u(x, t) = D 1v(x, t) +L Ij(x, t),
j=l

lj(x, t) = { D 1Kj(x- y, t)wj(y)dy,

} Rn-1
where v and wi = gi -hi are functions defined by (4.109) and (4.114)
respectively. In view of Theorem 2.10 and its proof we get

lvil,p,Rf_ ::::; Cl,vlfil-m,p,Rf_ ·

Hence
lhill-mrl/p,p,R"-1::::; 1Bj(D)vil-m3 ,p,Rf_::::; Cl,vlvil,p,Rf_::::; Cl,vlfil-m,p,Rf_·
In case l - mi is even we get with the aid of an integration by parts
n-1
Ii(x, t) = L Dklj,k(x, t),
k=l
where
Ij,k(x, t) = { D 1L\~n+q-l+mrl)/ 2 Kj,q(x- y, t)Dkl\ (l-mj- 2 )12 wi(y)dy.
} Rn-1
By virtue of Theorem 4.7
164 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

In case l - mi is odd we can proceed analogously putting

I·J,k (x l t) = 1
Rn-1
DlD k jj.(n+q-l+mr2)/2K·
X J,q
(x-y l t)jj.(l-mj-1)/2w·(y)dy
J '

4.8 Problems in Uniformly Regular Open

Sets
We assume that the assumptions listed in section 4.2 are satisfied by an
open set nand the operators L = L(x,D) and {Bif;l; = {Bj(x,D)}';l;.
For a positive number R we set

For a function u which is a trace on aR of a function belonging to W 1·P(ER)

and vanishing near lxl = R we set

(4.136)

Lemma 4. 10 Let L(x, D) be an operator of order m defined in ER, and

Bj(x, D) be operators of order mj, j = 1, ... , m/2, defined on aR. Suppose
that the assumptions of section 4.2 are satisfied by ER, L, { Bj} jl;. Then for
1 < p < oo there exists a positive number p < R such that for u E Wm•P(ER)

l
vanishing for lx I > p the following inequality holds:

llullm,p,L:R : : ; C [ IILullo,p,L:R + m~
~[Bju]m-mr1/p,p,O'R + llullo,p,L:R ·

(4.137)

Proof. Let F, Gi be functions defined by

L 0 (0, D)u = Lu + (L 0 (0, D)- L 0 (x, D)) u- L 1(x, D)u = F,

BJ(o, D)u = Biu + (BJ(O, D)- BJ(x, D)) u- B 1(x,D)u = Gj,

where L 0 , BJ are the principal parts of L, Bj and L 1, BJ are the lower order
parts of L, Bi respectively. Then in view of Theorem 4.9

(4.138)
4.8. PROBLEMS IN UMFORMLY REGULAR OPEN SETS 165

As is easily seen

IFio,p,W; :::; ILulo,p,ER

+max max laa:(x)- aa:(O)IIulm,p,ER + Cllullm-1,p,ER' (4.139)
la:l=mxEEp
IGilm-mj-1/p,p,R"- 1 :::; [Bju]m-mr1fp,p,o-R
+ max max lbjp(x)- bjp(O)IIulm,p,ER + Cllullm-1,p,ER· (4.140)
l.61=mj xEo-p
If we set
m/2
E = m~ max laa:(x)- aa:(O)I + ~ m~ max lbjp(x)- bjp(O)I, (4.141)
1<>1-mxEEp ~ l.6l-m3· xEo-p
J=1
then E ---+ 0 as p ---+ 0. With the aid of Theorem 3.16 and Young's inequality
there exists a positive number c. for each E > 0 such that

llullm-1,p,ER :::; Elulm,p,ER + C.lulo,p,ER' (4.142)

Substituting (4.139),(4.140),(4.141),(4.142) in (4.138)

lulm,p,ER = lulm,p,R+ :::; Ct,p

X [ILulo,p,ER + 'f[Bju]m-mr1/p,p,o-R + 2Eiulm,p,ER + C.lulo,p,ER]·

J=1
Letting p be so small that 2ECt,p :::; 1/2 we get

The desired inequality follows from this inequality and (4.142).

The following lemma is due to Browder [21].
Lemma 4. 11 If 0 is an open set in Rn uniformly regular of class em,
then there exist a natuml number N and positive numbers M, Do such that
for 0 < D < Do there exist an open covering { Oi} ~ 1 of 0 and for each i
a homeomorphism <I>i of class em from Oi to the open ball B6 of radius D
centered at the origin satisfying the following conditions:
(a) N+l distinct sets of {Oi} have an empty intersection;
(b) if oi and an have a nonempty intersection

<I>i(Oi n 0) = {y; IYI < D,Yn > 0},

<I>i( oi n 80) = {y; IYI < D, Yn = 0};
166 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

(c) ifil!i,k(x), Wi,k(Y) are the kth components ofil!i(x), wi(Y) = il!i 1 (y), then
for any i, k, lnl:::; m,x E Oi,Y E Bo

(4.143)

Proof. Since 0 is uniformly regular of class of em, there exists a locally

finite sequence {0i,o}~ 1 of open sets, a homeomorphism if!i,O of class em
from each Oi,O to the unit ball of Rn, a natural number N 0 and a positive
number M 0 such that the conditions (1),(2),(3),(4),(5) of Definition 3.2 are
all satisfied. Replacing M 0 by another constant if necessary we assume that

l8if!i,o(x)j8xl :::; Mo, 18W"i,o(y)j8yl :::; Mo

also hold, where wi,O = if!i:"J· Then

lif!i,o(x') -if!i,o(x)l :::; Molx' -xl, l'~~i,o(y')- Wi,o(Y)I :::; MolY' -YI· (4.144)

For 0 < d < min{Mo/No, 1/2} we set sd = {y ERn; IYI < 1/2, IYnl < d}.
Then U~ 1 wi,o(Sd) contains the d/Mo neighborhood of 80 as is shown as
follows. Let dist(x, 80) < d/M0 • Since d/Mo < N0- 1 we have x E 0~ 0 =
Wi,o( {y; IYI < 1/2}) for some i. If we set y = if!i,o(x), then '

IYI < 1/2, IYnl = lif!i,o,n(x)l :::; Modist(x, 80) <d.

Hence y E Sd and x = Wi,o(Y) E Wi,o(Sd)·

Let Bo be the unit ball and Po = Bon {y; Yn = 0}, and B1 = {y; IYI <
1/2}, P1 = B1 n {y; Yn = 0}. Let 0 < d1 < n- 112 min{1/2, Mo/No}. Set

Y' = {rl' = ('T/1, ... , 'TJn-l, 0); 'T/i are integers,

d1TJ 1 = (dl'T/1, ... , dl'T/n-1, 0) E P1},
and for ry' E Y'
STJ' = {y; IY- dl'T/ 1 1< y'ndl}·
Then as is easily seen STJ' C B 0 • We show that
(4.145)

If y E Sdu then y' = {Yl, ... , Yn-l, 0} E P 1. Let d1TJ' be the nearest point
toy' with ry' E Y'. Then IY'- d1TJ'I < -./n- 1dl. Hence

IY- d1TJ'I = {Y; + IY'- d1TJ'I 2}112 < {~ + (n -1)dn 112 = vnd1,
or y E STJ'·
In view of (4.144) we have diamwi,o(STJ'):::; 2y'nd1Mo. We have also

U~ 1 UTJ'EY' Wi,o(STJ') ::) U~ 1 Wi,o(Sd 1 ) ::) the dl/Mo neighborhood of 80.

4.8. PROBLEMS IN UNIFORMLY REGULAR OPEN SETS 167

Next, set d2 = dtf(2Mo..fii), and

Y = {7J = (7J1, ... , 7Jn); 7Ji are integers, fk7J En, dist(d27J, on) > d1/Mo}.
Then
= {x; ix- rk7JI < vnd2} c n.
8 11
Let {Oi} be the totality of wi,o(S11 ' ), 7] 1 E Y', i = 1, 2, ... , and 8 71 , 7J
E Y.
Then there exists a homeomorphism of class em from each of Oi to the ball
with center the origin and radius ..fiid1 • We conclude the proof without
difficulty putting 8o = min{1/2, M 0 /N0 } and 8 = ..fiid1 •
n
Theorem 4. 10 Let be an open set in nn
uniformly regular of class em.
Suppose that L(x, D), {Bj(x, D)} ';J;
satisfy the conditions listed in section
4.2. Then for each p E (1, oo) there exists a positive constant eP such that
the inequality (4.25) holds for u E Wm,p(n):

iiullm,p,n ~ eP [IlL(·, D)uiio,p,n

+ ~[B;(·, D)u[m-m;->/v,v,an + l!u[[o,,,n]·

Proof. For 8 > 0 let {Oi}, {<l>i} be the families of open sets and mappings
satisfying the conditions (a), (b), (c) of Lemma 4.11. We construct the par-
tition of unity {(i} subordinate to the families {Oi}, {<l>i} as was done in
section 3.6 with {y; IYI < 8} in place of {y; iYi < 1}. We set ui = (iu for
each i. In case Oi n an is not empty the images of the operators under
the homeomorphism x = wi(Y) satisfy the conditions of section 4.2 in the
half ball {y; iYi < 8, Yn > 0}. Hence, if 8 is sufficiently small we can apply
Lemma 4.10 to ui(wi(Y)), and pulling back to the original coodinates x we
get

If Ui is a function in wm,v(n) coinciding with Bj(·, D)ui on an, then

m-1

[BjUi]m-m;-1/p,p,an ~ li(igilim-m3 ,p,n +e L IIDkuiiLP(supp(;)·

k=1

Hence following the proof of Theorem 3.15 we conclude the proof of the
present theorem.
168 CHAPTER 4. ELLIPTIC BOUNDARY VALUE PROBLEMS

Remark 4. 2 In [11] it is assumed that n is bounded. However, it is not

assumed that mi < m, and is proved that for an integer

l
l 2:: max:j=l, ... ,m/2{ m, mi + 1} the estimate

JJuJJt,p,O::::; Ct,p [ JJLuJJt-m,p,O + m~

~[Biult-mJ-1/p,p,ao + JJuJJo,p,o

holds under suitable smoothness conditions on the boundary 80 and the

coefficients of L(·, D), {Bj(·, D)}7l;.
Chapter 5

Elliptic Boundary Value

Problems (Continued)

5.1 Adjoint Boundary Conditions

Let 0 be an open subset of Rn uniformly regular of class em. Let L(x, D)
be a uniformly elliptic operator of order m with coefficients defined in n
and satisfying the Root Condition of section 4.2. Let {B;(x, D)}j}; be a
set of differential operators with coefficients defined on ao. Assume that
the order m; of B; is less than m.
The set { B; }j}; is called normal if the following conditions are satisfied:
(i) {m;}j}; are distinct, i.e. m; =/:- mk if j =/:- k,
(ii) 80 is nowhere characteristic with respect to each of B;, i.e. for any
X E ao BJ(x, v) =/:- 0, j = 1, ... 'm/2, where v is the outward normal vector

to ao at X and BJ is the principal part of Bj.

Suppose that the coefficients of L and {B;} j}; are sufficiently smooth
and {B3} ;1; is normal. Let L' (x, D) be the formal adjoint of L(x, D): if

L(x, D)= 2: a0 (x)D 0 , (5.1)

lai:Sm
then
L'(x,D)v= 2: (-D) 0 (iia(x)v). (5.2)
lai:Sm

In this section following M. Schechter [133] we describe how to construct

the adjoint boundary operators.

169
170 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Theorem 5. 1 Under the hypotheses stated above there exists a normal set
of boundary operators {Bj(x, D)};f; such that a necessary and sufficient
condition in order that u E cm(n) satisfies

Bj(x, D)u = 0, j = 1, ... , m/2, on an (5.3)

is that the equality

(L(x,D)u,v) = (u,L'(x,D)v) (5.4)

holds for any v E Cfr'(O) satisfying

Bj(x, D)v = 0, j = 1, ... , m/2, on an. (5.5)

Conversely, v E Cfr'(O) satisfies (5.5) if and only if (5.4) holds for any
u E cm(fi) satisfying (5.3).
Remark 5. 1 In the above theorem u E cm(f!), v E Cfr'(O) may be re-
placed by u E Cfr'(O), v E cm(f!).
A normal set of m boundary operators of order less than m is called a
Dirichlet set of order m. If {Cj }J= 1 is a Dirichlet set of order m and f.ti is
the order of Cj, then {f.ti}j= 1 = {0, 1, ... , m- 1}. If a;av is the normal
derivative, then {(a jav)i- 1 }J= 1 is a Dirichlet set of order m.
Let a be a point of an. We construct the set { Bj} of Theorem 5.1 in
some neighborhood of a. By a coordinate transformation we suppose that
the part of n in the neighborhood is the half ball
I;R = {x ERn; lxl < R,xn > 0},
and the part of an is the disk

aR = {x ERn; lxl < R,xn = 0}.

In the following two lemmas we consider only smooth functions in !:R.

Lemma 5.1 Let {Cj}j= 1 and {Cj}j= 1 be a couple of Dirichlet sets of
order m. If Ci and Cj are both of order j - 1 for j = 1, ... , m, then there
exist differential operators Aik of order j - k containing only the tangential
derivatives D1, ... , Dn-1 such that
j

Cj = l:AjkCk on aR. (5.6)

k=1

Aii is a function which vanishes nowhere on aR.

5.1. ADJOINT BOUNDARY CONDITIONS 171

Proof. If the lemma is proved in case Cj = D~- 1 , j = 1, ... , m, then we

have
j

D~- 1 = l:AitCt. (5.7)

!=1

With some tangential differential operators r ik of order at most j - k, Cj

is expressed as
j

cj = l::rjkD~-1, (5.8)
k=1

rii being a nonvanishing function. It follows from (5.7),(5.8) that

j k j j j
Cj = l:rik l:AktCt = l:l:rikAktCt = l:SitCt.
k=1 1=1 !=1 k=l !=1

Here ej! = Ei=! r jkAkl is a tangential differential operator of order at

most j - l, and eii = r iiAii is a nonvanishing function. Hence it suffices
to show the lemma only in case Cj = D~- 1 , j = 1, ... , m.
Each Ck is expressed as
k
ck = l::rktD~-1,
!=1

where rk! are operators with the same property as those of (5.8). Therefore
k-1
D nk-1 -_ r-1c
kk k
_ r-1 ~ r
kk L....t kl
nt-1
n ·
!=1

Hence, if n~- 1 is expressed as (5. 7) for l = 1, ... , k- 1, then so is D~- 1 • If

j = 1, (5. 7) is clear. Hence for any j = 1, ... , m n~- 1 is expressed as (5. 7).

Lemma 5. 2 J!{Cj}j; 1 is a Dirichlet set and {gJ}'}; 1 is a set of functions

defined on JR, then there exists a function v E cm(f:R) such that Civ =
gj,j = 1, ... ,m on JR.

Proof. In view of Lemma 5.1 there exist tangential differential operators

r ik, Akt such that
j k
ci = l::rjkn~-1, D~- 1 = l:AktCt.
k=1 !=1
172 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

It is easy to show
j

LrikAkt = 6il· (5.9)

k=l
If v is a function defined in ER satisfying
k
D~- 1 v = LAkWl
1=1

on CJR, then with the aid of (5.9) we get that on CJR

j j k j j

civ = Lrikn~- 1 v = Lrik LAktgl = LLrikAktgl = gi.

k=1 k=1 1=1 1=1 k=l

Proof of Theorem 5.1. If u, v E cm(f;R) and v vanishes near lxl = R, then

by an integration by parts

where dx' = dx1 · · · dxn-1· We write the operator Las

L = L ap.kDJ.L D~,
IJ.LI+kS:m

where f.t = (f.t 1 , ... , f.tn- 1 , 0). With the aid of (5.10)

(Lu, v)- (u, L'v)

L
IJ.LI+kS:m
r [D~u·(-D)P.(ap.kii)-u·(-Dn)k(-D)J.L(ap.kv)]dx
}ER

L L
k

IJ.LI+kS:m i=1
1 D~- u·
O'R
1 (-Dn)k-i(-D)P.(ap.kii)dx'

L L 1 D~- 1 u · L (-D)P.( -Dn)k-i(ap.kii)dx'

m k
=-
k=1 i=1 O'R IJ.LIS:m-k

(5.11)
5.1. ADJOINT BOUNDARY CONDITIONS 173

where
m
Niv = - L L (-D)~'( -Dn)k-i(a~'kv).
k=i II' I:5;m-k
Ni is of order at most m - i. The ellipticity of L implies aom -::/:- 0. Hence
N i _-(- 1)m-i-1-aom Dm-i
n + ... (5.12)

is of order m- i and O"R is not characteristic with respect toNi. Adding

m/2 boundary operators {Bj}J!=m/2+1 to {Bj}-;!;, we obtain a Dirichlet
set {Bj}j= 1. We renumber it so that {Bj}j= 1 = {Bi}f= 1 and Bj is of order
j - 1 for j = 1, ... , m. We denote the order of Bj,j = m/2 + 1, ... , m, by
mi. Then {mi}f= 1 = {0, 1, ... , m}, and Bm3 +1 = Bj. By virtue of Lemma
5.1 we have
i

Dni-1 = "" -
L...JAilBl, (5.13)
!=1
where Au are tangential differential operators of order i -l at the most and
Aii is a nonvanishing function. Denoting the formal adjoint of Au by A~ 1 ,
we set
m
Cj = L A~,m-j+1Ni (5.14)
i=m-j+1
for j = 1, ... , m. Ci is of order at most j - 1, and by (5.12)
I ' I - '-1
Ci = Am-J+1,m-i+1 Nm-i+1 + · · · = ( -1)3 Am-i+1,m-J+1 aomDh + ····
Hence Cj is of order j - 1 and O"R is not characteristic with respect to Cj.
In view of (5.11),(5.13),(5.14)

1R ~AuB1u
m i
(Lu, v)- (u, L 1v) = ~ · Nivdx 1

= fl
!=1 fiR
B1u· tA~Nivdx1 =
i=l
fl
!=1 fiR
B1u· Cm+1-lvdx 1•

Hence we obtain

(Lu,v)- (u,L 1v) = fl j=1 fiR

Bju·Cm-m3 vdx1• (5.15)

Suppose that (5.3) holds. If

Cm-m3 v=0, j=m/2+1, ... ,m, on O"R, (5.16)

174 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

then in view of (5.15) we see that (5.4) holds. Conversely, suppose u is

such that (5.4) holds for any v satisfying (5.16). Let ( be a real valued
function belonging to CQ"'({x; lxl < R}). In view of Lemma 5.2 there exists
a function v E Cm(:f::R) satisfying

j = 1, ... ,m/2
on rJR.
j = m/2 + 1, ... , m

We can choose v so that v = 0 near lxl = R. Hence with the aid of (5.15)

1R ( 1R
m/2 m/2

~ 2 2
1Biul dx' =~ Bju · Cm-m3 vdx'

= t1
j=l IJR
Bju · Cm-m3 vdx' = (Lu,v)- (u, L'v) = 0.

Consequently we get Biu = 0, j = 1, ... , m/2, on rJR· Thus {Bjf;l;

{ Cm-m3 }~m/2+l meets our requirement in ER. Returning to the original
coordinate system and using the partition of unity we obtain a desired set
of boundary operators.

Definition 5. 1 Two sets {Bj}j:_ 1 and {Bi}f= 1 of boundary operators are

said to be equivalent if the following condition is satisfied:

Biv = 0, j = 1, ... , m, on an
if and only if
Biv = 0, j = 1, ... , m, on an.
Lemma 5. 3 Let {Bi}J= 1 , {Bj}J-; 1 be normal sets of boundary operators.
Denote the orders of Bi, Bi by mi, mi. If these two sets are equivalent,
then m = m and there exist tangential differential operators Aik such that
m
Bi = 2::.:: AikBk, j = 1, ... , m. (5.17)
k=l

The set {Aik} satisfies the following conditions:

(i) Aik is of order at most mi - mk,
(ii) if mi = mk, then Ajk is a nowhere vanishing function.
Proof. Let l be a positive integer such that

{mifi=l u {mi}f=l c {0, 1, ... , l- 1},

5.1. ADJOINT BOUNDARY CONDITIONS 175

and {Bj}~=l a Dirichlet set containing {Bi}J=l· Then in view of Lemma

5.1 each Bj is expressed as
l
Bj = 2:AjkBk. (5.18)
k=l
Denote the order of Bk for k > m by mk. Then Aik is of order at most
mi - mk. For each j = 1, ... , m there exists k such that mk = mi and
for such k Aik is a nonvanishing function. If Aji -::p 0 for some i > m, then
there exists a function g defined on an such that Aii9 -::p 0. By Lemma 5.2
there exists a function v such that Biv = g and Bkv = 0 for k -1 i on aR.
In view of (5.18) Bjv = Aji9 -::p 0 on aR. Since Bkv = 0 fork= 1, ... , m on
aR, this contradicts the equivalence of {Bj} and {Bj}· Hence Aji = 0 for
i > m. Hence we conclude (5.17) and {m'i}f= 1 c {mj}j;: 1 . The opposite
inclusion is established analogously.
Definition 5. 2 Let {Bj(x,D)}71; be the set of boundary operators of
Theorem 5.1. Then (L'(x, D), {Bj(x, D)}71i, n) is called the adjoint bound-
ary value problem of (L(x, D), {Bj(x, D)} 71;, n).
It is easy to see that the adjoint boundary value problem of the Dirichlet
problem for L(x, D) is the Dirichlet problem for L' (x, D).
Lemma 5. 4 Suppose that L, { Bi}71; satisfies the Complementing Con-
dition and {Bi}71i is normal. If { Bi}71i is a normal set equivalent to
{Bi}71i, then L, { Bi}71; satisfies the Complementing Condition.
Proof. By virtue of Lemma 5.3 we have
m/2
Bj = 2: AjkBk·
k=l
- o -o o
We denote the principal part of Bj, Bj, Aik by Bi, Bi, Aik" Then
m/2
-o ~ o o
Bi (x, ~+TV)= L.., Aik(x, ~)Bk(x, ~+Tv),
k=l
where ~' V are a tangential and the outward normal vector to an at X E
an respectively. Since {BJ(x,~ + Tvn;1: is a linearly independent set of
polynomials ofT, the matrix ( A~k(x, ~)) is nonsingular. If
m/2 m/2m/2
~ -o ~~ o o _
L.., CjBj (x, ~+TV)= L.., L.., Cji\jk(x, ~)Bk(x, ~+TV)= 0
j=l k=lj=l
176 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

modulo M+ (x, ~ + rv), then E;l~ ciA~k(x, ~) = 0, k = 1, . .. , m/2. Hence

Cl = · · · = Cmj2 = 0.

Next we show that if L, {Bj};l~ satisfy the Complementing Condition,

then so does its adjoint L', { Bj} ;l~. For complex variables z,( we set

k
ak(z, () = 2::>k-i(i-1, ao(z, () = 0.
i=l
=
For a polynomial P(z) 'E~=o akzk of a complex variable z and a complex
vector w = (wo,wl, ... ,wm) set
m
P(w) = 2::: akwk.
k=O

If we set m
R(z, () = P(a(z, ()) = 2::: akak(z, (), (5.19)
k=O
as is easily seen

R(z, () = R((, z), P(z)- P(() = (z- ()R(z, (). (5.20)

If we set
k
f)
(k) -
Rz (z, () - ( f)z ) R(z, (), pk(z) = ( :z) k P(z),
then we have

(z- ()R~k)(z, () = p(k)(z)- kR~k-l)(z, (), (5.21)

Lemma 5. 5 Let P 1 ,P2 be polynomials of order m1,m2 of a complex vari-

able. Set P = P1P2 and m = m 1 +mz. Let {QiYj~ 1 be linearly independent
polynomials of order less than m, and w = (w 0 , w1, ... , Wm-1) a complex
vector. The conditions

P(a(·,w)) =0 modulo Pz, (5.22)

Qi(w)=O, j=1, ... ,ml (5.23)

imply w = 0 if and only if {Qi}}~ 1 is linearly independent modulo P1.

5.1. ADJOINT BOUNDARY CONDITIONS 177

Proof. By definition
m
P(CT(z,w)) = :~:::>kCTk(z,w) = R(z,w).
k=O
Let { zi} be the distinct roots of P2, the multiplicity of Zi being denoted by
vi. The condition (5.22) holds if and only if

R\"(z;, w) ~ (:z )' R(z, w( , . ~ ( :z)' P(a(z, w)t" ~ 0 (5.24)

for 0 :S k < vi. As is easily seen

(k) . _
Rz (z~, ()- (( _ Zi)k+l
k!P(()
for 0 :S k <Vi.

Hence if we put Pik(() = P(()/(( -zi)k+\ then Pik is a polynomial of order

less than m and R~k)(zi, () = k!Pik((). Hence (5.24) is equivalent to
(5.25)
Thus the only vector w satisfying (5.22) and (5.23) is 0 if and only if them
polynomials Pik, Qi are linearly independent. Set
Mik(() = P2(()j((- zi)k+l for 0 :S k <vi.
Suppose that {Qi} j~ 1 is linearly independent modulo P1 , and that suppose
for some constants aik, bi

2:::: aikpik + 2:::: bjQj =o.

Since Pik
Hence
= P1Mik, 2::::: biQi =0 modulo P1. Therefore b1 = · · · = bm = 0.
1

P1 L aikMik = L aikpik =0,

which implies 2::::: aikMik =
0. Since {Mik} are linearly independent we
obtain aik = 0. Thus Pik, Qi are linearly independent.
Conversely suppose that Pik, Qi are linearly independent, and that for
some constants b1 , ... , bm 1 and a polynomial H we have

LbiQi +HPl = 0.
Since H is of order less than m 2 , there exist constants aik such that H =
2::::: aikMik· Hence
2::::
bjQj + aikpik 2::::
o, =
which implies b1 = · · · = bm = 0. 1
178 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Theorem 5. 2 Suppose that the assumptions of Theorem 5.1 are satisfied.

If L, { Bj} ';l;
satisfies the Complementing Condition, then so does the ad-
joint L', {Bj} ";!i. The converse is also true.

Proof. We consider in the situation of the proof of Theorem 5.1 and follow
the notations there. We denote the principal parts of L, L' by L 0 , (L') 0 :

L0 = .2:: aJ.LkDJ.LD~, (L') 0 = .2:: aJ.LkDJ.LD~.

IJ.LI+k=m IJ.LI+k=m

Let ~ = (6, ... , ~n-1, 0) and v = (0, ... , 0, -1) be a tangential and the
outward normal vector at the origin, and set

L 0(T) = L 0 (0,~ +TV)= I: aJ.Lk(o)eTk,

IJ.LI+k=m

(L') 0(T) = (L') 0 (0,~ +Tv)= I: aJ.Lk(o)eTk.

IJ.LI+k=m

Then (L') 0(T) = L0(T), where £0 is the polynomial whose coefficients

are the complex conjugates of the corresponding coefficients of L 0 • Let
T1, ... , Tm; 2 and Tm/ 2+1, ... , Tm be the roots of L 0 with positive and nega-
tive imaginary parts respectively. Set
m/2 m
M+(T) = II (T- Tj), M-(T) = II (T- Tj)·
j=1 j=m/2+1

We write
l i
-
Bt = """"
LJ rliDni-1 , l = 1, ... , m, n~- 1 =I: AikBk, i =1, ... 'm,
i=1 k=1

where { Bt H~ 1 be the operators in the proof of Theorem 5.1. Then

l
I:rliAik = Dtk· (5.26)
i=k

If we set "fli = rpi (0, ~), Aik = Afk (0, ~), then by (5.26) we have
l
.2:: "fliAik = Dtk· (5.27)
i=k
5.1. ADJOINT BOUNDARY CONDITIONS 179

The principal parts of Ni, Cj are

m
Nf = ( -1)i-1l: l:
iiJJ-kDJJ- D~-i,
k=i IJJ-I=m-k
m m
cJ=(-1)i l: l: l: iiJJ-kA?,m-i+1DilD~-i.
i=m-j+1 k=i IJJ-I=m-k
If we set

then
m
N?(T) = (-1)i- 1 l : l : aJlk(o)eTk-i, (5.28)
k=i IJJ-I=m-k
m
CJ(T) = (-1)j l: (-1)i- 1,\i,m-J+1Nf(T). (5.29)
i=m-j+1
Suppose that

=
m
l:.Bie!-J+ 1(T) 0 modulo !Vr(T), (5.30)
j=1
.Bm3 +1 = 0 for j = 1, ... , m/2. (5.31)
If we show that ,81 = · · · = .Bm = 0, then the proof will be complete. For
i = 1, ... ,m set

From (5.28),(5.29) it follows that

m m m
l:.BjC~-J+1(T) = l:.Bj(-1)m-J+1l:(-1)i-1,\ijNf(T)
j=1 j=1 i=j
m i m
= l: l:< -1)i- 1,Bj,\ij( -1)i- 1N?(T) = l:< -1)i- 1wi-1N?(T)
i=1 j=1 i=1
m m
= l:wi-1l: l: aJlk(o)eTk-i
i=1 k=i IJJ-I=m-k
m k
= l: l: iiJJ-k(O)~Jll: Tk-iwi-1
k=1IJJ-I=m-k i=1
180 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

L iillk(O)~~"a-k(T,w) = £ 0(a-(T,w)).
l~£1+k=m

Combining this with (5.30) we get

L 0(a-(T,w)) =
0 modulo M-(T). (5.32)

On the other hand by (5.27)

l l i
Bf(w) = L"YliWi-1 = L"YtiL(-1)i- 1,8jr\ii
i=1 i=1 i=1
l l l
= L( -1)i- 1 Pi L "YliAii = L( -1)i- 1 Pi bit = (-1)1- 1Pt·
j=1 i=j j=1

Hence by (5.31) we get for j = 1, ... , m/2

BJ(w) = B~3 + 1 (w) = (-1)m3 Pm3 +1 = 0. (5.33)

In view of (5.32),(5.33), Lemma 5.5 and the hypothesis of the theorem we
get wi-1 = 0 for i = 1, ... , m. Since the matrix (,\ij) is nonsingular, we
conclude {31 = · · · = !3m = 0.

5.2 Existence of Solutions

Let n be an open subset of Rn uniformly regular of class em, and L(x, D),
jl;
{ Bj (x, D)} be operators satisfying the conditions stated in section 4.2.
For 1 < p < oo we define the operator Ap as follows:

D(Ap) = {u E wm,P(fl); Bj(·,D)u = 0 on arl,j = 1, ... ,m/2},

(5.34)
for u E D(Ap) Apu = L(·, D)u in the sense of distributions.
In view of Theorem 4.10 the following inequality holds:

llullm,p,n ::; Cp (IIAvullo,p,O + llullo,p,n) · (5.35)

In this section we use the notations of Chapter 3 to denote norms of Sobolev

spaces.
Lemma 5. 6 The operator Ap defined by (5.34) is closed in LP(fl).
Proof. Suppose that ui E D(Ap), Uj ~ u, Avui ~ f in LP(fl). By virtue
of (5.35)

llui- ukllm,p,n ::; Cp (IIAvui- Apukllo,p,n + llui- Ukllo,p,n) ·

5.2. EXISTENCE OF SOLUTIONS 181

Hence u E wm,p(n), lluj -ullm,p,n ---+ 0. Therefore u E D(Ap), and Apuj ---+
Apu, Apu= f.

In this section following S. Agmon [9] we show that under some natural
assumptions the resolvent set p(Ap) of Ap is not empty.
Let fJ be an angle such that 0 :::; fJ < 271". Introducing an auxiliary real
variable t we set

Q = n X R = {(x, t); x En, -00 < t < oo}, (5.36)

£(x, Dx,Dt) = L(x,Dx)- (-1)mf 2 ei11 Df. (5.37)

We make the following assumptions:

£(x, Dx, Dt) is elliptic in Q, (5.38)

£(x, Dx, Dt), {Bj(x, Dx)}~,!i satisfyies the
Complementing Condition in Q. (5.39)

The condition (5.38) holds if and only if

(i) For any X En and a real vector~ f:- 0 arg{(-1)mf 2 £ 0 (x,O} f:- fJ.
By the definition of the Complementing Condition the second condition
(5.39) is
(ii) Let X be an arbirary point of on, ~ be tangential to on at x, v the
outward normal vector to on at X and A a complex number with argument
fJ. Suppose moreover (~, >.) f:- 0. Let ri(~, >.), ... , r;:;_12 (~, .X) be the roots
of the polynomial L 0 (x, ~ + rv)- ( -1)m/2 ,\with positive imaginary parts.
Then the polynomials {BJ(x, ~ + rv)}~,f; are linearly independent modulo
TI~if(r- rf(~, >.)).

Remark 5. 2 If,\ = 0 the condition (ii) reduces to the Complementing

Condition for L, {B3 }~,ff. If~ = 0 in the condition (ii), we get that
Tm3 BJ(x, v) are linearly independent modulo fl~,f;(r-rf(x, v)), and hence
linearly independent. This implies that {B3}~,!; is normal.

The following lemma is a slight extension of Theorem 2.1 of S. Agmon [9].

Lemma 5. 7 Let n be an open set in Rn uniformly regular of class em.

Suppose that L, {Bj}~,ff satisfies the conditions of section 4.2 and the as-
sumptions (i),(ii) of this section. Then for any 1 < p < oo there exists
a constant Cp such that for any u E wm,p(n), g3 E wm-m3,P(n),j =
182 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

1, ... , m/2, satisfying Bi(·, D)u = 9i on an, and a complex number.\ sat-
isfying arg.\ = 0, l.\1 > Cp the following inequality holds:

~ 1.\l(m-i)/mllulli.v,n::; Cv [II(L- .\)ullo,p,n

m/2 m/2
+ ~ l.\l(m-mj)/mll9illo,p,O + ~ ll9illm-mj,p,O
l
· (5.40)

Proof. Let (be a function inc=( -oo, oo) such that ((t) = 0 for ltl > 1 and
((t)= 1 for ltl < 1/2. For r > 0, u E wm,p(n) set v(x, t) = ((t)eirtu(x).
Then applying Theorem 4.10 to £, {Bj} j!i in Q we get

llvllm,p,Q o> C [IICvllo,,,Q + ~[B;v[m-m;-'I•·•·"Q + llvllo.v,Q]· (5.41)

With the aid of Leibnitz' formula

Cv(x, t) = ((t)eirt (L- rmei9 ) u(x)
-( -1)mf2ei9 E (~)
k=O
((m-k)(t)( -1)kf2rkeirtu(x).

Hence
m-1

IICvllo,p,Q::; II(£- rmei 9)ullo,p,n +C L rkllullo,p,O· (5.42)

k=O
Since X E an if (x, t) E aQ, we have on aQ
Bjv(x, t) = ((t)eirt Bju(x) = ((t)eirtgj(x).
Thererfore with the aid of the moment inequality (3.150)

[BjvJm-mr1/p,p,8Q ::; ll(eirt9illm-m3 ,p,Q

m-m3
<
_ ~ L...J rm-mj -k llg j II k,p,O
k=O
::; C (rm-m3 ll9jllo,p,O + ll9illm-m3 ,p,n) · (5.43)

On the other hand

llvll~.p,Q = L 1 Jnr ID~Dtv(x,

00

lal+kS::m -oo
tW dxdt
5.2. EXISTENCE OF SOLUTIONS 183

~ L 11/2 r IDfeirtn~u(x)lp lb:dt

l<>l+k::;:m -1/2 Jn
= t
k=O
rPk
lal:-s:m-k Jn
t
L r ID<>u(xW dx = rp(m-j)llull~,p,fl'
i=D
(5.44)

Combining (5.41),(5.42),(5.43),(5.44) we get

t, rm-illulli.v,n :s; c[II(L- rmei 0 )ullo,v,n +}; rkllullo,v,n

+ ~ {rm-m;II9JIIo,p,n + II9JIIm-m;,p,n) + llullo,p,nl

If r is sufficiently large, the second and fourth terms in the bracket of the
right hand side of this inequality are small compared with the left hand
side. Thus putting A= rmei 0 , we complete the proof.
Letting gi = 0 in Lemma 5. 7 we obtain the following theorem.

Theorem 5. 3 Let 0 be an open set of Rn uniformly regular of class em.

Suppose that L, {Bif:;l;satisfy the conditions of section 4.2 and (i),(ii) of
this section. Then, for any 1 < p < oo there exists a constant Cp such that
if argA = (), IAI > Cp, the following inequality holds for any u E D(Ap):
m
L IAI(m-j)/mllulli,p,n :s; Cvii(Ap- A)ullo,p,O.· (5.45)
j=O

Remark 5. 3 In view of the moment inequality (3.150) we see that the

inequality (5.45) is equivalent to the following inequality:

IAIIIullo,p,n + llullm,p,n :s; Cvii(Ap- A)ullo,p,n (5.46)

replacing Cp by another constant if necessary.

The following definition is due to S. Agmon [9].

Definition 5. 3 Let A be a linear closed operator in some Banach space

and () an angle. If for some positive constant C the half line {A; argA =
(), IAI > C} is contained in the resolvent set p(A) of A and the inequality
II(A-A)- 1 11 :s; C/IAI holds for A on this half line, then the half line argA = ()
is called a my of the minimal growth of the resolvent of A.
184 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

In view of Theorem 5. 3 if

R(Ap- .\) = LP(n), arg.\ = (), l.\1 ~ Cv (5.47)

is shown to hold, we see that arg.\ = () is a ray of the minimal growth of the
resolvent of AP' In [9] it is announced that the proof of (5.47) will be given
somewhere. However, the proof does not seem to have appeared. In Chapter
3 of [149] the proof of (5.47) is given in case n is bounded. The outline is as
follows. If the coefficients are sufficiently smooth, (5.47) can be shown using
M. Schechter's result on the Fredholm alternative between L- .\, { Bj}, and
its adjoint L'- ..\,{Bj} ([134]). In the general case we approximate the
operators L, {Bj} by a sequence of operators £Ck), {BY)} with smooth
coefficients. Then (5.47) holds with constant Cp = C~k) possibly depending
also on k. However, (5.40) holds for £(k), {BY)} with constants independent
of k. Hence, the inequality

holds for arg.\ = (), l.\1 > C~k) with some constant C independent of k,
where A~k) is the operator defined by (5.34) with L(k), {B?)} in place of L,
{ Bj }. Consequently with the aid of Neumann series expansion the resolvent
(A~k) - .\)- 1 can be continued analytically to a fixed sector

with uniform estimate there. Thus letting k ~ oo it can be shown that

p(Av) contains the set Ep.
Definition 5. 4 Let L(x, D) be a differential operator of order min n. If
for any X E 0, ~ =/:- 0

then L(x, D) is called strongly elliptic.

For l = 1, 2, ... , ( -~) 1 is strongly elliptic.
Theorem 5. 4 A strongly elliptic operator satisfies the Root Condition.
Proof. Let L(x, D) be strongly elliptic of order m. Since L 0 (x, -~)
(-1)mL 0 (x,~), m is even. For 0 ~ s ~ 1

Ls(x, D) = sL(x, D) + (1- s)( -~)m/ 2

5.2. EXISTENCE OF SOLUTIONS 185

is strongly elliptic. If~. 77 are linearly independent real vectors, then the
roots of the polynomial L~(x, ~ + T'f7) are continuous functions of s, and
if s = 0 the exactly m/2 of them have positive imaginary parts. Hence
L(x, ~ + rry) = L 1 (x, ~ + rry) has exactly m/2 roots with positive imaginary
parts.
That L(x, D) is strongly elliptic is equivalent to saying that for () E
[11"/2, 371" /2]
(5.48)
is elliptic. In what follows in this section we assume that
(iii) L(x, D) is strongly elliptic in 0,
(iv) for any() E [11"/2, 371"/2] Co(x, Dx, Dt), {Bj(x, Dx)}jJ; satisfy the Com-
plementing Condition in Q.
From Theorem 5.3 we obtain the following theorem.
Theorem 5. 5 Let n be an open set of Rn uniformly regular of class em.
Assume that L(x, D), {Bj(x, D)}jJ; satisfy the conditions listed in section
4.2 and (iii),(iv) of this section. Then for each 1 < p < oo there exists a
positive constant Cp such that for any u E wm,P(O), gj E wm-m3,P(0)
satisfying Bj(x, D)u = gj, j = 1, ... , mj2, on 80, and a complex number,\
satisfying Im,\ :=::; 0, 1,\1 > Cp the following inequality holds:

~ 1,\l(m-j)/mllullj,p,n :::_; Cp [II(L(·, D)- ,\)ullo,p,!l

m/2 m/2
+ ~ l,\l(m-mj)/mll9iiiO,p,!l + ~ llgillm-mj,p,!l
l
· (5.49)

Especially if u E D(Ap) we have the following inequality

1,\lllullo,p,n + llullm,p,n :::_; Cvii(Ap- ,\)ullo,p,!l· (5.50)

If the assumptions of Theorem 5.5 are satisfied and furthermore if for some
,\satisfying Im,\:::.; 0, 1,\1 > Cp (5.47) holds, then -Av generates an analytic
semigroup in LP(O). Note that D(Ap) is dense in LP(O) since it contains
C[f(O). By what was stated after Definition 5.3 -Av generates an analytic
semigroup if n is bounded.
We make the following assumption in order to consider the adjoint bound-
ary value problem.
(v) The coefficients of the lower order terms of the formal adjoint L'(x, D)
of L(x, D) are bounded and measurable.
186 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Namely the formal adjoint of L constructed regardless of the di:fferentia-

bilty of the coefficients of L satisfies the Smoothness Condition of section
4.2. Note that the coefficients of the highest order terms of L' are the com-
plex conjugates of those of the highest order terms of L, and hence uniformly
continuous in 0.
Next assume the adjoint boundary operators { Bj} ';li
constructed in sec-
tion 5.1 also satisfy the conditions in section 4.2.
(vi) There exists a normal set { Bj} ;!i
of boundary operators satisfying the
following conditions. Let mj be the order of Bj. Then mj < m and the
coefficients of Bj have bounded and uniformly continuous derivatives up to
order m- mj on an. Let v E wm• 2 (n). Then (Lu, v) = (u, L'v) holds for
any u E wm• 2 (n) satisfying Bj(·,D)u = O,j = 1, ... ,mj2, on an if and
only ifv satisfies Bj(·,D)v = O,j = 1, ... ,m/2, on an, and the assertion
also holds interchanging u, {Bj} and v, {Bj}.
The adjoint boundary value problem of

(L(x, Dx)- ( -1)ml 2 ei0 D;n, {Bj(x, Dx)}';ii, Q)

is
(L'(x, Dx)- ( -1)ml 2 e-i0 D;n, {Bj(x, Dx)}';ii, Q). (5.51)
Hence (5.51) satisfies the Complementing Condition for any() E [n /2, 3n /2].
The operator A~ is defined just as Ap was defined:

D(A~) = {u E wm,v(n); Bj(·, D)u = O,j = 1, ... , m/2, on an},

(5.52)
for u E D(A~) A~u = L'(·, D)u in the sense of distributions.
By virtue of Theorem 5.5 the following statement holds. For 1 < p < oo set
p' = pj(p- 1). If we replace Cv by another constant if necessary, then for
any u E D(A~,) and .\ satisfying Im.\ :::; 0, l.\1 > Cp

l.\lllullo,p',n + iiullm,p',n:::; Cvli(A~- .\)uiio,p,fl· (5.53)

In what follows we assume that the coefficients of Bi and Bj are extended
to the whole of 0 so that they belong to Bm-mJ (0) and Bm-mj (0) respec-
tively for each j = 1, ... , m/2.
The following definition is also due to F. E. Browder [21].
Definition 5. 5 Let n be an open set in Rn. If for any a E an the part
of n, an in some neighborhood of a is expressed as

respectively for some i = 1, ... 'n and a C 2 m function 1/J, then n is called
an open set locally regular of class c 2 m.
5.2. EXISTENCE OF SOLUTIONS 187

It is possible to show the following result following the idea of Browder [21].
In addition to the assumptions made so far we assume that 0 is locally
regular of classcz.m. Then for ImA:::; 0, IAI > Cp

R(Ap- A) = £P(O), R(A~, - :\) = £P' (0). (5.54)

The proof of this statement is very lengthy, and so we state only the out-
line. Since by Theorem 5.5 Ap -A has a continuous inverse, R(Ap -A)
is closed. Hence it suffices to show that R(Ap - A) is dense. This is the
same for R(A~, - :\). If we approximate the coefficients of L, {Bj} by
smooth functions, then the inequality (5.40) holds with common constants
for this approximating sequence. Hence it suffices to consider only the case
of smooth coefficients. Let p be in a compact subinterval of (1, oo) contain-
ing 2. For a complex number A satisfying ImA :::; 0 and with sufficiently
large absolute value we replace L by L - A and show
(5.55)
In this case for u E Wm•P(O), 9i E wm-mj,P(O) such that Bju = 9i on 80
for j = 1, ... , m/2 the inequality

(5.56)

holds. As for the adjoint problem we have for u E D(A~,)

llullm,p',O :S CIIA~,ullo,p',O· (5.57)

Lemma 5. 8 Let u be a function in £P(O). Assume that for some neigh-
borhood U of each point of 0 u E wm,P(U) and Lu E £P(O). If more-
over u satisfies the boundary conditions Biulao = 0, j = 1, ... , m/2, then
u E wm,p(O).
Browder [21: Theorem 3] proved this lemma in the case of Dirichlet bound-
ary conditions. The proof remains valid in the general case owing to Theo-
rem 4.10.
Let f E C0 (0). In view of (5.57) A~ has a continuous inverse. Hence
by Theorem 1.3 R((A~)*) = £ 2 (0). Therefore there exists u E £ 2 (0) such
that f = (A~)*u. This function u is a weak solution of the boundary value
problem

Lu = f in 0, Bju = 0, j = 1, ... , m/2, on 80. (5.58)

By virtue of a regularity result of weak solutions of elliptic boundary value
problems (Schechter [134] in case 0 is bounded or Browder [21] in case
188 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

of Dirichlet boundary conditions in an unbounded domain) we have u E

cm(O). Hence u is a classical solution of (5.58). Applying Lemma 5.8
we get u E wm, 2 (0), and hence u E D(A2), f = A 2u. Thus (5.55) is
established for p = 2.
Next suppose that 1/2 < 1/p :::; 1/2 + 1/n. Let ¢ be a function in
Cf!"(Rn) such that ¢(x) = 1 for lxl :::; 1 and ¢(x) = 0 for lxl 2:: 2, and set
¢R(x) = ¢(x/R) for R 2:: 1. Let f E LP(O) n L 2 (0). By what was shown
just above there exists u E D(A2) such that f = A2u. Applying (5.56) to
¢Ru with gi = Bi(¢Ru)- ¢RBiu we get

m/2 )
li¢Rullm,p,O:::; C ( IIL(¢Ru)llo,p,O + ~ IIBi(¢Ru)- ¢RBiullm-mj,p,O ·

(5.59)
By Leibnitz' formula

L(¢Ru) = ¢Rf + L Caf3Da¢R · Df3u,

!a!+!f3!S::m,af:O

With the aid of Holder's inequality

iicaf3Da¢R · D 13 uiio,p,o :::; C ( in 1Da¢R · D 13 uip dx )

1/p

:::; C ( k. 1Da¢Ri2p/(2-p) dx )
(2-p)/2p (
in 1Df3ul2 dx
) 1/2

= CRn(l/p-1/2)-!al ( k. 1Da¢12p/(2-p) dx )
(2-p)/2p
11Df3ullo,2,0·

Noting n(1/p- 1/2) - lal :::; 1- lal :::; 0 we get

(5.60)
Analogously

IIBi(¢Ru)- ¢RBiullm-mj,p,O :::; Cllullm-1,2,0· (5.61)

Substituting (5.60),(5.61) in (5.59) we get

li¢Rullm,p,O :::; C (lifllo,p,O + llullm-1,2,0) · (5.62)

Letting R-+ oo we get u E wm,p(O), and hence u E D(Ap), f = Apu. With
the aid of the same argument we show that (5.55) is true for p satisfying
5.2. EXISTENCE OF SOLUTIONS 189

1/p1 < 1/p:::; 1/pl + 1/n with some P1 satisfying 1/2 < 1/P1 ::; 1/2 + 1/n.
Continuing this process we can verify that (5.55) holds for 1 < p:::; 2.
Next, if 1/2 - mjn :::; 1/p, 2 < p < oo, then in view of Theorem 3.19
Wm• 2 (0) C LP(fl). Hence if for f E C8"(r2) f = Azu, then u E LP(O) n
cm(O). By virtue of Lemma 5.7 u E wm,P(fl), and hence u E D(Ap), f =
Apu. If m < n/2, set 1/p1 = 1/2 - mjn, and we can prove by the same
argument as above that for p satisfying 1/PI - mjn :::; 1/p :::; 1/Pl (5.55)
holds. Continuing this process we can show that (5.55) is true for 2 < p <
oo.
In view of (5.50),(5.53),(5.54) we see that

p(Ap) :J {A; Im.A:::; 0, I.AI > Cp}, (5.63)

p(A~,) :J {A; Im.A:::; 0, I.AI > Cp}· (5.64)

Since as is seen without difficulty (Apu, v) = (u, A~,v) for u E D(Ap), v E

D(A~, ), we have
(Ap)* :J A~,. (5.65)

If u is an arbitrary element of D( (Ap)*) and Im,\ :::; 0, I.AI > Cp, then in
view of (5.64) there exists a function v E D(A~,) such that

(5.66)

Combining (5.65),(5.66) we get

((Ap)* - 5-)u = ((Ap)* - 5-)v. (5.67)

Since,\ E p(Ap) implies 5. E p((Ap)*), we obtain from (5.67) that u = v E

D(A~, ). Thus we conclude

(5.68)

Especially if Lis formally self-adjoint, i.e. L = L', and if {Bj} is equivalent

to {BJ}, then A 2 is self-adjoint.
Summing up we have established the following theorem.

Theorem 5. 6 If in addition to the hypothesis of theorem 5.5 r2 is lo-

cally regular of class C 2m, then -Ap, -A~, generate analytic semigroups
exp( -tAp), exp( -tA~,) in LP(O), LP' (0) respectively, and

exp( -tAp)* = exp( -tA~, ). (5.69)

190 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

5.3 Estimates of the Kernels of exp( -tAp) and

(Ap- A)- 1
In this section under the assumptions of Theorem 5.6 we obtain the esti-
mates of the kernels of exp( -tAp) and (Ap - >.)- 1. We follow the idea of
R. Beals [19] of establishing the asymptotic behavior of the resolvent ker-
nel of A 2 in case A 2 is self-adjoint under mild smoothness assumptions on
the coefficients. We also follow L. Hi::irmander's idea of using exponential
functions to derive sharp interior estimates of resolvent kernels ([75]). For
direct construction of the Green function for the parabolic initial boundary
value problem with coefficients depending also on t see S. D. Eidel'man and
S. D. Ivasisen [57], S. D. Ivasisen [78], [79] and V. A. Solonnikov [143].
If the assumptions (i),(ii) of section 5.2 are satisfied, then the same
assumptions are also satisfied for angles sufficiently close to B. Hence,
we assume that (5.49) holds for >. satisfying arg>. ~ ( -00 , 00 ) for some
Bo E (0, 1r /2).
Lemma 5. 9 Iff E £P(O) n U (0), arg >. ~ ( -Bo, Bo), 1>-1 > max{ Cp, Cq},
then
(5.70)
Proof. Assume p < q. Replacing L by L- >.we show

(5.71)

As was stated in the proof of (5.54) the inequality (5.56) holds. Suppose
1/p- 1/n::; 1/q. Let¢ be a function in C(f(Rn) such that ¢(x) = 1 for
lxl ::; 1 and ¢(x) = 0 for lxl 2: 2. Set ¢R(x) = ¢(x/ R) for R 2: 1. Set
u = Aq- 1f. If we apply (5.56) to ¢Ru, then by the same calculation used in
deriving (5.62) we get

II¢Rullm,p,n ::; C(llfllo,p,n + llullm-1,q,n).

Letting R ~ oo we obtain u E wm,P(O), and hence u E D(Av)· Thus (5.71)

is verified. In the general case we choose P1, ... , Pk-1 so that p = Po < P1 <
· · · < Pk = q, l/Pi-1- 1/n::; 1/Pi· Then f E LP•(O), i = 0, 1, ... , k. By the
same argument as above we can show

A-1f
P
= A-1f
Pt
= ... = A-1 f = A-1f.
Pk-1 q

Lemma 5. 10 Let 0 be a nonempty open subset of Rn, and S, T bounded

linear opemtors from L 2 (0) to itself Suppose that R(S) C B(O.), R(T*) C
B(f1.), where B(f1.) is the set of bounded and continuous functions in 0., and
5.3. ESTIMATES OF KERNELS 191

that {Sg; ll9llo,2,n :::; 1} and {T*h; llhllo,2,n :::; 1} are equicontinuous in 0.
Then, ST is an integral operator with kernel K(x, y):

(STJ)(x) = L K(x, y)f(y)dy. (5.72)

K(x, y) is bounded, uniformly continuous in 0 X 0 and satisfies

Proof. By assumption there exists a strictly increasing continuous function

w defined in [0, oo) such that w(O) = 0 and
I(Sg)(x)- (Sg)(x')l:::; ll9llo,2,nw(lx- x'l), (5.74)
I(T*h)(y)- (T*h)(y')l:::; llhllo,2,nw(IY- y'l). (5.75)
Since for f E £ 2 (0), X E 0
I(STJ)(x)l:::; IISTfllo,oo,n:::; IIBIIB(£2,L"")IITfllo,2,n
:::; IIBIIB(£2,L"") IITIIB(£2,£2) llfllo,2,n, (5. 76)
there exists a function K(x, y) such that K(x, ·) E £ 2 (0) and

(STJ)(x) = L K(x, y)f(y)dy, (5.77)

(L IK(x, YWdy) 112 :::; IIBIIB(£2,L"")IITIIB(£2,£2)· (5.78)

Since forf E £ 2 (0) n £ 1 (0),g E £ 2 (0)

I(Tf,g)l = l(f,T*g)l:::; llfllo,l,niiT*gllo,oo,n
:::; llfllo,l,niiT* IIB(£2,£"") ll9llo,2,n,
we get
liTfllo,2,n :::; liT* II B(£2,L"") llfllo,l,O· (5.79)
Hence
I(STJ)(x)l:::; IIBIIB(£2,L"")IIT*IIB(£2,L"")II!IIo,l,n· (5.80)
From (5.77),(5.80) it follows that

ess supyEniK(x, Y)l :::; IIBIIB(£2,£"") liT* IIB(£2,£"") =

'Y· (5.81)
In view of (5.74),(5.79) for x,x' E 0

IL (K(x, y)- K(x', y))f(y)dyl = I(STJ)(x)- (STJ)(x')l

:::; 11Tfllo,2,nw(lx- x'l):::; liT* IIB(£2,L"")II!IIo,l,nw(lx- x'l).
192 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Hence

ess supyEniK(x, y)- K(x', y)l ::; liT* IIB(£2,L=)w(lx- x'i). (5.82)

Let {xi} be a countable dense subset of 0. Then, in view of (5.81),(5.82)

there exists a null set N C 0 such that for y rf. N

Let x be an arbitrary point of 0. By virtue of (5.83) for a subsequence {Xj1 }

tending to x the limit

K(x, y) = lim K(xin y)

!-+oo

exists for any y rf. N. It is obvious that the limit does not depend on the
choice of {xiJ· For x, x' E 0, y rf. N we have

IK(x, y)l ::; "f, IK(x, y)- K(x', y)l ::; liT* IIB(£2,L=)w(lx- x'i). (5.84)

It is easily seen that K(x,y) is measurable in 0 x 0, and by (5.78)

(5.85)

Let f E £ 2 (0) n £ 1 (0) and x E 0. Let again {xiJ be a sequence tending

to x. Then, in view of (5.83) we get from (5. 77) with Xj 1 in place of x

(STJ)(x) = k K(x, y)f(y)dy. (5.86)

By virtue of (5.85) this holds for any f E £ 2 (0).

Let f E L 2 (0),g E £ 2 (0) n £ 1 (0). Then

l(f,S*g)l = I(SJ,g)l::; IISfllo,oo,nllgllo,1,n

::; IISIIB(£2,£=) ll!llo,2,nllgllo,1,n·
Hence the restriction of S* to £ 2 (0) n £ 1 (0) can be extended continuously
to a bounded linear operatorS' from £ 1 (0) to £ 2 (0), and for any g E £ 1 (0)

liS'gllo,2,n ::; II Sll B(£2 ,L=) llg llo,1,n· (5.87)

As is easily seen for any f E L 2 (0),g E £ 1 (0)

(f,S'g) = (Sj,g). (5.88)
5.3. ESTIMATES OF KERNELS 193

Iff E L 2 (0) n £ 1 (0), g E £ 1 (0), then with the aid of (5.84),(5.86),(5.88)

(f,T*S'g) = (Tf,S'g) = (STJ,g)

=In In K(x, y)f(y)dyg(x)dx = kf(y) kK(x, y)g(x)dxdy.
Hence if we set K(x, y) = K'(x, y), then we have

(T*S'g)(y) =in K'(x,y)g(x)dx (5.89)

almost everywhere in 0. Therefore, if {gi} is a countable dense subset of

£ 1 (0), then there exists a null set N' :) N such that for any y (j. N' and j

(T* S'gi)(y) = k K'(x, y)gj(x)dx.

Since T* is a bounded linear operator from £ 2 (0) to B(O) by the hypothesis

and the closed graph theorem, we see noting (5.84) that (5.89) holds for any
g E £ 1 (0) andy (j. N'.
Let y, y' (j. N' and g E £ 1 (0). Then it follows from (5.89),(5.75),(5.87)
that

lin (K'(x,y) -K'(x,y'))g(x)dxl = \(T*S'g)(y)- (T*S'g)(y')\

:::; \\S'g\\o,2,nw(\y- y'\):::; I\SI\B(£2,L"")\\g\\o,1,nw(\y- y'\).

Hence noting that K(x, y) is continuous in x we obtain

sup \K(x, y)-K(x, y')\ =sup \K'(x, y)-K'(x, y')\ :::; \\S\\B(L2,L"")w(\y-y'\).
xEO xEO

Therefore K(x, y) is extended to the whole of Ox 0 as a continuous function,

and

\K(x, y)- K(x', y)\:::; \\T* \\B(£2,L"")w(\x- x'\),

\K(x, y) - K(x, y')\ :::; \\S\\B(£2,L"")w(\Y- y'\).

Therefore rewriting K(x, y) as K(x, y) we see that (5.72) holds. Using (5.80)

lin K(x,y)f(y)dyl = \(STJ)(x)\:::; \\S\\B(L2,L"")\\T*\\B(L2,L"")\\f\\o,1,n,

which implies (5.73).
194 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

For ann-dimensional complex vector 'T/ = (ry 1 , ••. , 'T/n) the operators A~, A'~
are defined as follows.
D(A~) = { u E wm,P(O); Bj(·, D + ry)u = 0, j = 1, ... , m/2, on 80},
(5.90)
for u E D(A~) A~u + ry)u in the sense of distributions,
= L(·, D
D(A'~) = {u E wm,P(O); Bj(·, D + ry)u = O,j = 1, ... , m/2, on 80},
(5.91)
for u E D(A'~) A'~u = L'(·, D + ry)u in the sense of distributions.

Since

we have

e-x11 L(x, D)(ex 11 u(x)) = L(x, D + ry)u(x), (5.92)

e-x11 Bj(x, D)(ex 11 u(x)) = Bj(x, D + ry)u(x). (5.93)

From the proof of Theorem 5.1 we see that there exist normal sets of bound-
ary operators {Cj(x,D)}j!f~, {Cj(x,D)}Ti~ such that

In (L(x,D)u · v- u · L'(x,D)v) dx
m/2
= L { Bi(x, D)u · Ci(x, D)vdS
i=l Jan
m/2
+ L f Cj(x,D)u · Bj(x,D)vdS. (5.94)
i=l Jan
By virtue of (5.92),(5.93),(5.94)

In (L(x, D + ry)u · v - u · L' (x, D - ry)v) dx

=In (e-x11 L(x,D)(ex'~~u)·ii-u·ex'f1L1 (x,D)(e x11v))dx
=in ( L(x, D)(ex 11 u) · e-xfiv- ex'~~u · L'(x, D)(e-xfiv)) dx

:E r
m/2
= Bj(x,D)(eX11u). Cj(x,D)(e-X11v)dS
i=l Jan
m/2
+ :E { Cj(x,D)(ex'~~u) · Bj(x,D)(e-xfiv)dS
i=l Jan
5.3. ESTIMATES OF KERNELS 195

m/2
=:E { Bi(x,D+77)u·Ci(x,D-77)vdS
i= Jan1

r Cj(x, D + 77)U. Bj(x, D- 77)vdS.

m/2
+ L Jan
i= 1

This shows that the adjoint of L(x, D + 77), { Bi (x, D + 77)}7li is

L'(x, D- fj), {Bj(x, D- iJ)}7ii.
Lemma 5. 11 For 1 < p < oo there exists a positive number {j such that for
arg A f/: (-Oo, Oo), I-XI > Cp, 1771 :::; 6I-XI 1 /m' u E wm,v(n), hj E wm-mj,P(n),
Bi(x, D + 77)u =hi, j = 1, ... , m/2 on an the following inequality holds:

~ 1-XI<m-k)/mllullk,v,n:::; C~ [11(L(x, D + 77)- -X)ullo,v,n

m/2 m/2
+ ~ 1-XI(m-mj)/mllhillo,p,n + ~ llhillm-mj,p,n] · (5.95)

Especially for u E D(A~)

1-XIIIullo,p,n + llullm,p,n:::; C~II(A~- -X)ullo,v,n· (5.96)

Analogously, for u E D(A';,)

1-XIIIullo,p,n + llullm,p,n:::; C~II(A';, - ,\)ullo,p,n· (5.97)

Proof. Set
gi = (Bi(x, D)- Bj(x, D + 77))u +hi.
Then 9i E wm-mj,P(n) and 9i = Bi(x, D)u on an. We have

II (L(x, D) - -X)ullo,p,n
:::; II (L(x, D + 77) - -X)ullo,p,n + II (L(x, D + 77) - L(x, D) )ullo,p,n
m-1
:::; II(L(x, D + 77)- -X)ullo,p,n + C 2: l77lm-kllullk,p,n,
k=O
ffij-1

II9JIIo,v,n:::; C L l77lmj-kllullk,p,n + llhJIIo,p,n,

k=O
m-1

II9JIIm-mj,p,n:::; C L l77lm-kllullk,p,n + llhJIIm-mj,p,n·

k=m-m3
196 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Hence with the aid of (5.49)

~ I.XI(m-k)/mllullk,p,O :::; c~ [11(L(x, D + 11) - .\)ullo,p,O

m-1 m/2 mj-1

+L 1171m-kllullk,p,n + L I.XI(m-mj)/m L l11lmrkllullk,p,O
k=O j=1 k=O
m/2 m/2
+ ~ I.XI(m-mj)/mllhillo,p,O + ~ llhillm-mj,p,n] ·

If 0 < {j :::; 1, 1111 :::; til.\1 1/m, then

L I.XI(m-k)/mllullk,p,n
m

k=O

: :; c; [11 (L(x, D + 11) - .\)ullo,p,O + ~ I.XI(m-k)/mllullk,p,O

{j

m/2
+ L I.XI(m-mj)/m{j L
mi-1
I.XI(mrk)/mllullk,p,O
j=1 k=O
m/2 m/2
+ ~ I.XI(m-mj)/mllhillo,p,O + ~ llhillm-mi,p,n]

: :; c; [II(L(x, D + 11)- .\)ullo,p,O + ~ I.XI(m-k)/mllullk,p,O

l
{j

m/2 m/2
+ ~ I.XI(m-mj)/mllhillo,p,O + ~ llhillm-mi,p,O ·

If {j is sufficiently small, the inequality (5.95) follows from this inequality.

Similarly to (5.68) we have
(5.98)
Hence with the aid of Theorem 1.3 we obtain the following lemma.
Lemma 5.12 Let 1 < p < oo and o be as in Lemma 5.11. Then ifarg.\ f/:.
( -Bo, Bo), I.XI > Cp, 1171 :::; ol.\1 1/m, then,\ E p(A~), ~ E p(A'~, ), and for any
f E LP(O) or f E LP' (0)
IICA;- .\)- 1!llo,p,n:::; C'llfllo,v,n/I.XI, (5.99)
5.3. ESTIMATES OF KERNELS 197

II(A~- >.)- 1 fllm,p,n:::; c;llfllo,p,n, (5.100)

II(A';,- 5.)- 1 fllo,p',n:::; c;llfllo,p',n/1>-1, (5.101)
II(A';,- 5.)- 1 fllm,p',n:::; c;llfllo,p',n· (5.102)

If 'fJ is pure imaginary, then

(A~_ >,)-1f = e-x1J(Ap _ ,\)-1(eX77f). (5.103)

Proof. That,\ E p(A~), 5. E p(A'~,) and (5.99)-(5.102) follow from (5.96)-

(5.98) and Theorem 1.3. Noting that if 'fJ is pure imaginary, u E wm,p(n)
if and only if ex7Ju E wm,p(n), we can easily show (5.103) with the aid of
(5.92),(5.93).
Remark 5. 4 If n is bounded, (5.103) holds for an arbitrary complex vec-
tor 'fl·

Lemma 5. 13 Let L(x, D), { Bj (x, D)} ;'l: be opemtors satisfying the con-
ditions of section 4.2. Suppose in addition that {Bj};'l: is normal and for
any u E wm,p(n) the following inequality holds:

II u[[ m,p,n :S: C [IIL(x, D )ullo,p,n + %[ B; (x, D )uim-m;-1/p,p,an] , (5.104)

Moreover suppose that 0 E p(Ap) for the opemtor Ap defined by (5.34). Then
for any f E LP(n), 9i E wm-mr 11P,P(an), j = 1, ... , m/2, the solution of
the boundary value problem

L(x, D)u = f in n, (5.105)

Bi(x, D)u = gi, j = 1, ... , m/2, on an (5.106)

exists and is unique.

Proof. In view of Theorem 3.27 there exits a sequence {9j,v} C C[f(an)

such that 9i,v --+ 9i in wm-mJ- 1/P,P(an). By virtue of Lemma 5.2 there
exists a function wj,v E Clf(O) satisfying Bj(x, D)wj,v = 9j,v on an. If we
set
Uv = A; 1(!- Lwj,v) + Wj,v,
then Uv E wm,p(n), Luv = f in n, Bjuv = 9j,v, j = 1, ... 'm/2 on an.
Applying (5.104) to U 11 - u~" we have
m/2
lluv- u~"llm,p,n:::; C~)9i,v- 9i,JJ-lm-mr1/p,p,&n·
j=1
198 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Hence {Uv - uJl} is a Cauchy sequence in wm,p(n). The limit is obviously

the unique solution of (5.105),(5.106).
Lemma 5.14 Let 1 < p < oo, A, 'TJ be as in Lemma 5.12. Then, for any
f E LP(n),gi E wm-mr 11P,v(an),j = 1, ... , m/2, the boundary value
problem

(L(x, D + TJ) - .A)u = f in n, (5.107)

Bj(x, D + ry)u = gi> j = 1, ... , m/2, on an. (5.108)

has a unique solution.

Proof. In view of Lemma 5.11 there exists a constant Cp,>. depending on p, .A
such that for u E wm,v(n), hi E wm-mJ,P(n) satisfying Bj(x, D + TJ) = hi
on an

m/2 l
llullm,p,O ::S Cp,>. [ II (L(x, D + TJ) - .A)ullo,p,n + ~ I hi llm-mJ ,p,n ·

Replacing the right hand side by the greatest lower bound with respect to
hi yields

llullm,v,n ::S Cv,>. [II(L(x, D + TJ)- .A)ullo,p,n

+ ~[B;(x, D + •J•Jm-m,->/,,,,an]·

Hence using Lemma 5.13 we conclude that the solution of (5.107),(5.108)

exists and is unique.
Lemma 5.15 Under the assumptions of Lemma 5.11 (A~-.A)- 1 is a holo-
morphic function of 'TJ in ITJI :::; 6I.AI 1/m.
Proof. For f E LP(n) set u'f/ = (A~ - .A)- 1 f. Then
(L(x, D + TJ) - .A)u'fl = f in n, (5.109)
Bi(x, D + ry)u'f/ = 0, j = 1, ... , m/2, on an. (5.110)

If we formally differentiate both sides of (5.109),(5.101) with respect to 'T/1

and set

g(x) = -(aLjaE:,t)(x, D + ry)u'f/(x),

gj(x) = -(aBijaE:,t)(x, D + TJ)~(x),
5.3. ESTIMATES OF KERNELS 199

then we obtain

(L(x, D + ry)- >..)8uTJ(x)j8ry1 = g(x) in 0,

Bj(x, D + ry)8uTJ(x)j8ry 1 = gj(x), j = 1, ... , m/2, on 80.
Evidently g E W 1 ·P(0) C £P(O), gj E wm-mj+ 1,P(0) C wm-mj,P(O).
Hence in view of Lemma 5.14 the solution of the boundary value problem

(L(x, D + ry) - >..)v = g in 0,

Bj(x, D + ry)v = gj, j = 1, ... , m/2, on 80,

exists. If we set uTJ,h = (u<:- uTJ)jh, where ( = (ry 1 + h, 'f/2, ... , 'f/n), then

(L(x, D + ry)- >..)(uTJ,h- v) = -h- 1 (L(x, D + ()- L(x, D + ry))u<:- g

in 0, and

Bj(x, D + ry)(uTJ,h- v) = -h- 1 (Bj(x, D + ()- Bj(x, D + ry))u<: - gj

on 80. Thus applying (5.95) to uTJ,h - v and letting h -+ 0 we obtain

lluTJ,h - vllm,p,n -+ 0.
We choose natural numbers l, sand exponents 2 = q1 < · · · < q < qs+1 =
8

oo, 2 = r1 < · · · < rl-s < rl-s+1 = oo as follows:

(i) in case m > n/2, l = 2, s = 1, i.e. 2 = q1 < q2 = oo, 2 = r1 < rz = oo;
(ii) in case m < n/2, s > nj2m, l - s > nj2m,
1 m 1
_!_ - - 1- < m, j = 1, ... , s - 1, ->->-, (5.111)
qj qi+1 n qs-1 n qs
1 m 1
_!_ - - 1- < m, j = 1, ... , l - s - 1, -->->-, (5.112)
Tj Tj+1 n T!-s-1 n T!-s
and m- njq8 , m- n/rl-s are not integers;
(iii) in case m = n/2, l = 4, s = 2, 2 = q1 < q2 < q3 = oo, 2 = r1 < rz <
r3 = oo.
Set
aj = (njm)(1/qi- 1/qi+ 1), j = 1, ... , s.
Then 0 < aj < 1 and in view of Theorems 3.19, 3.20 we have for j =
1, . .. ,s -1
(5.113)
and for 0::; h ::; m- nfqs
l ull h,oo,n-
< Cllulla.+h/mllull
m,q.,U
1 -a.-h/m.
O,q.,U (5.114)
200 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Replacing L by L+ Ao for some sufficiently large positive number Ao if neces-

sary we assume that (5.49) holds for p = q1, ... , q8 , r1, ... , rl-s and for any A
satisfying argA fj. (-Oo,00 ). Set Co= max{Cp;p= ql, ... ,q8 ,rl,····rl-s},
and denote by 8 the minimum value of 8 of Lemma 5.11 for p = q1, ... , qs,
r 1, ... , rl-s· Let Aj, j = 1, ... , l, be complex numbers such that arg AJ fj.
( -Oo, Oo), and 1] a complex vector such that 1171 ~ 8min{IA1I 1/m, ... , 1Ad 1 m}.
Set
S = (~- A 8 )- 1 (A~- As-1)- 1 ···(A~- A2)- 1 (A~- A1)-l, (5.115)
1 1 1
T =(A~- As+l)- (A~- As+Z)- · ··(A~- Al-1)- (A~- Al}- . (5.116)
1

By (5.111) and Theorem 3.17

R((A~- A1 )- 1 ) c wm• 2 (0) = wm,q 1 (0) c U 2 (0).
Hence in view of Lemma 5.9 (A~- A2 )- 1 on the right hand side of (5.115)
may be replaced by (A~2 - A2 )- 1. Repeating this yields

S =(A~.- A 8 )- 1 (A~•-t- As-1)- 1 · · · (A~2 - A2)- 1 (A~- A1)- 1. (5.117)

In view of Theorem 3.20 and (5.111)
R((A~.- As)- 1 ) c wm,q•(O) c Bm-nfq·(O).
Therefore
(5.118)
By virtue of (5.98) we get
T * -_ (A'ii2 - ')-l(A'ii2
Al '
-Al-l )-1 · · · (A'ii2 - '
As+2 )-l(A'ii2 '
- As+l )-1 ·

Analogously to (5.117),(5.118) we obtain

T* = (A'; 1_.- 5.t)- 1 · · · (A';2 - >-s+2)- 1 (A'~- >-s+d-1, (5.119)

R(T*) C Bm-n/rl-•(0). (5.120)

In view of Lemma 5.12

II(A~j - Aj)- 1 fllo,qj,n ~ Cllfllo,qj,n/IAil, (5.121)

II(A~j - Aj)- 1 fllm,qj,n ~ Cllfllo,qj,n (5.122)

for j = 1, ... ,s. By virtue of (5.113),(5.121),(5.122) we get

II(A~3 - Aj)- 1 fllo,qHt,n

~CII(A q
77 . -A·)- 1fllaj. 1-aj
1JII O,q
3 3 m,q3 ,n II(Aq773. -A·)-
3 3 ,n

~ CIIJII~:qj ,n (llfllo,qj ,n/IAj I) l-aj = CIAilaj - l llfllo,qj ,n

5.3. ESTIMATES OF KERNELS 201

for j = 1, ... , s - 1. Hence

s-1
IISJIIo,q.,n:::; CIAsl- 1 II 1Aila 3- 1ll/llo,2,n, (5.123)
j=1
s-1
IISJIIm,q.,n:::; C II 1Ajla 3- 1ll/llo,2,!l· (5.124)
j=1
It follows from (5.114),(5.123),(5.124) that for 0:::; h:::; m- njq8

IISJIIh , oo ,n:::; CIISJIIa.+h/miiSJII

m,qs,n
1 -a.-h/m
O,q,,n
s
:::; CIAslh/m II 1Ajlar111/llo,2,!l· (5.125)
j=1
Analogously for 0:::; h' :::; m- nfrz-s
l
liT* fllh',oo,n:::; CIAzlh'/m II 1Ajla3 - 1ll/llo,2,n, (5.126)
j=s+1

where ai = (njm)(1jrj-s - 1jri-s+1) for j = s + 1, ... , l. Therefore for

0 < h < min{m- njq8 , 1}, 0 < h' < min{m- nfrz-s, 1}
s
I(Sg)(x)- (Sg)(x')l:::; CIAslh/m II 1Ajlar1llgllo,2,nlx- x'lh, (5.127)
j=1

II
l
I(T*h)(y)- (T*h)(y')l:::; CIAtlh'/m IAjlar 1llhllo,2,niY- y'lh'. (5.128)
j=s+1

Owing to (5.118),(5.120),(5.127),(5.128) we can apply Lemma 5.10 to the

operator
ST =(A~- A1)- 1 (A~- A2)- 1 ···(A~- Az)- 1
to conclude that ST is an integral operator whose kernel we denote by
K1,, ... ,.>./x, y). In view of (5.73) and (5.125),(5.126) with h = 0, h' = 0
respectively
l
IK11, ... ,.>.1 (x, Y)l :::; C II 1Ajlar1. (5.129)
j=1
If TJ is pure imaginary, we have by Lemma 5.12
202 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Hence if we denote the kernel of (A2 -.-\ 1)- 1 · · · (A 2 -.-\t)- 1 by K>. 1 , ... ,>.1 (x, y),
then we have
K At, .. -,>.t (x ' y) -- eCx-y)n K At,
71 (x y)
... ,>.t ' (5.130)
for a pure imaginary 'f/· By Lemma 5.15 ST is a holomorphic function of
'f/ in l'fll::; 8min{l.-\1l 1/m, ... , l.-\tl 1/m}, and hence so is K1
1 , ••• ,>. 1 (x,y) as is
easily seen using Morera's theorem. Therefore (5.130) holds also if 'f/ is
not pure imaginary. In what follows we consider only real 'f/· Owing to
Remark 5.4 if n is bounded, we need not consider nonreal 'f/· It follows from
(5.129),(5.130) that
l
IK>.t, ... ,>.t (x, y)l :=; Ce(x-y)n II l.-\ilar1.
j=1
Minimizing the right hand side of this inequality with respect to 'f/ we obtain

l
:=; C exp [ -8 min (l.-\1! 1/m, · · ·' l.-\tl 11m) lx- Yl] II l.-\ilar 1
j=1
l l
::; c:Eexp ( -8l.-\kl 11 mlx- Yl) II
l.-\ila3 - 1· (5.131)
k=1 j=1
The following lemma is easily shown.
Lemma 5. 16 For any positive number a there exists a constant C such
that ta :=;Get for any t ~ 0.
Let r be a smooth contour running from ooe-i9o to ooei9o in the closed
angular domain .E = {A; arg ,\ ~ ( -00 , 00 )}. Then

exp( -tA2) = -21 . f e->.t(A2 - ,\)- 1d,\. (5.132)

1rz Jr
Hence,

exp( -ltA2) = [exp( -tA2)] 1

= _1_.- { ... { e->.tt-···->-tt(A2- ,\1)-1 ... (A2- At)-1d,\1 ... d.-\t.
(27rz )l Jr Jr
Consequently, exp( -tA2) has a kernel G(x, y, t), and

G(x,y,lt) = ( 2 ~i)l t . .t e->. 1 t-···->-ttK>. 1 , ... ,>. 1 (x,y)d,\1"'d,\l (5.133)

5.3. ESTIMATES OF KERNELS 203

holds for X' y E n, I arg t I :::; 7r I 2 - Bo. For fixed X' y' t we deform the integral
contour r to the following path r x,y,t:
r x,y,t = {-\;I arg -\I = Bo, l-\1 2:: a} u {A; A= aei(J' IBI 2:: Bo},
where
lx- Ylm/(m-1) Ep lx- Ylm/(m-1)
a= E ltlm/(m-1) = ltf' p= ltjl/(m-1) ' E > 0. (5.134)

If we put _\ = re±iOo for I arg -\I = Bo, we get

Re-\t = Re-\ ·Ret- Im-\ · Imt

= rRe t (cos () 0 =f sin ()0 ~::) 2:: rRe t (cos () 0 - sin Bo ~~: :1) .
Hence if t is in the closed sector
lim tl < ( 1 _Eo) cos Bo (5.135)
Ret - sin 00 '
where 0 <Eo < 1, then Re-\t ~ rRet · Eocos00 • Therefore there exists a
positive constant c such that fort in the sector (5.135) and_\= re±iOo

Re-\t 2:: crltl. (5.136)

From (5.131),(5.133) it follows that

IG(x, y, Zt)l :::; C [ ... [ e-Re>.1t-···-Re>.zt

::z:,y,t x,y,t

l l
X I: exp ( -<5l-\kl 1/mlx- Yl) IJ l-\ilaj- 1ld-\1 · · · d-\tl. (5.137)
k=1 j=1
The term with k = 1 in the right hand side of (5.137) is

C t. y,t exp ( -Re-\1t- <51-\111/mlx- Yl) l-\1la~-1ld-\11

l
X ]][,.,y,t e-Re>.jtl-\ilar1ld-\il· (5.138)

Using (5.136) we get for j = 2, ... , l

t. y,t e-Re>.jtl-\ilar1ld-\il

= r
J1>.j I=a
+ r : :;
J1 Aj I?_a
ealtlaaj-127ra + 21a
00
e-crltlraj-1dr.
204 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Using the definition (5.134) of a, p and with the change of the variable
s = rltl in the last integral we get with the aid of Lemma 5.16

hx,y,t e-Re>.jtl.\jlartld.\jl
< 27r (Ep)aj e<P +
- ltl
_2_
1tla3
roo e-cssartds -< Ce2<p.
Jo IW 3
(5.139)

Next if we put

h,,y,t exp ( -Re.\tt- D'l.\tll/mlx- Yl) 1.\tlal-tld.\tl

= r
JJ>.1J=a
+ r
JJ>.1J>a
=It+ I2, (5.140)

then
It :::; 27raa 1 exp(altl - oatfmlx- Yl)

= 27r G~) a1 exp(Ep- &tfmp):::; ltfu 1 exp(2Ep- &tfmp), (5.141)

00
12 :::; 21 exp( -crltl - ortfmlx- yl)ra 1 -tdr (5.142)

Making the change of the variable r = as in the right hand side of (5.142)
and substituting (5.134)

I2 :::; 2aa 1 1 00
sa 1-t exp( -cEps- & 11m pstfm)ds

:::; 2aal exp(-DpEtfm) 1oo sal-te-c<psds

:::; 2 ( E
__!!_ 1 loo ~a1-te-c~~
)a1 exp(-DpEtfm)--
ltl (Ep)al o
= Cltl-a 1 exp( -DpEtfm). (5.143)
From (5.140),(5.141),(5.143) it follows that

h,,y,t exp ( -Re.\tt- D'l.\111/mlx- Yl) l.\tlal-1ld.\tl

:::; Cltl-a 1 exp(2Ep- &tfmp). (5.144)

In view of (5.139),(5.144) and E~= 1 ai = njm (5.138) does not exceed

Cltl-njm exp [(2lE - &tim) p] .

5.3. ESTIMATES OF KERNELS 205

The terms with k = 2, ... , l in the right hand side of (5.138) are estimated
similarly. Hence, choosing E so small that 2lE- & 11m < 0 we find that there
exist positive constant C, c such that in the sector (5.135) the following
inequality holds:
C ( lx- Ylm/(m-1))
IG(x, y, t)l::; ltlnfm exp -c ltjlf(m-1) . (5.145)

By an analogous argument we can show that for lal < m

C ( lx- Ylm/(m- 1))
IDaG x t < e
x ( 'y, ) I - ltl(n+lal)/m xp
-c ltjl/(m-1) . (5.146)

The outline of the proof of (5.146) is as follows. If we choose q1 , ... , q8 so

that q8 > n in addition to (i),(ii),(iii), then in view of Theorem 3.20 and
(5.114) we have

and
llullk,oo,n::; Cllull~,~~.~mllull~~~;nk/m, k = 0, ... , m- 1. (5.147)
Hence by (5.114),(5.123),(5.124) we have for lal <m
s
IID~Sfllo,oo,n::; IISJIIIal,oo,n::; Cl>-sllal/m II 1>-ilar1llfllo,2,fl· (5.148)
j=1
Since D~K1 1 ... ,>Jx,y) is the kernel of D~ST we get from (5.126) with
h' = 0, (5.148) and Lemma 5.10
l
ID~Kt ... ,.x 1 (x,y)l::; Cl>-sllal/m II l>-ilar 1.
j=1
Therefore

ID~K.xt, ... ,.xz(x,y)l = ~D~ [e(x-y)1JK1t, ... ,>)x,y)] I

= """'
L.J
(3S,a
(a) f3 Da-(3
X
e(x-y)1J · Df3X K ..\1,
11
... ,..\z
(x, Y)

L II 1>-jlaj-1
l
::; C l77lla-f31e(x-y)1JI).sllf31/m
(35,a j=1

L l>-sllal/me(x-y)1J II 1>-ilar1.
l
::; C (5.149)
(35,a j=l
206 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Following the argument by which we derived (5.145) we can show that

(5.146) holds. Noting that G(y, x, t) is the kernel of exp( -tA~) we can
establish an analogous estimate for D~G(x, y, t), I,BI < m:

/3 C ( lx- Yim/(m-1))
IDu G(x, y, t) I :::; ltiCn+l/31)/m exp -c ltjlf(m-1) . (5.150)

Since exp(-2tA2 ) = {exp(-tA2 )} 2 , we have

G(x, y, 2t) = k
G(x, z, t)G(z, y, t)dz. (5.151)

Using
lx- zlm/(m-1) + IY- zlm/(m-1) ~ 2-1/(m-1) lx- Ylm/(m-1)'
lx- zlm/(m-1) + IY- zlm/(m-1) ~ lx- zlm/(m-1)

we obtain
r (-c lx-1t11/(m-1)
Jo. exp
zlm/(m-1)) ( IY- zlm/(m-1))
exp -c IW/Cm-1) dz

Jo.r{exp [- 2ltj1/(m-1) (lx- zlm/(m-1) + IY- zlm/(m-1))] }2 dz

= c

<
-
exp ( -
c
2m/(m-1)
lx- Ylm/(m-1))
ltjlf(m-1)
1 (
R"
exp --
2
c lx- zlm/(m-1))
ltjlf(m-1)
dz

= itinfm exp (- c lx- Yim/(m-1)) { exp (-~ lt;lm/(m-1)) dl;.

2m/(m-1) ltjl/(m-1) }Rn 2
(5.152)

From (5.146),(5.150),(5.151),(5.152) it follows that for lal < m, I,BI < m

C ( lx- Ylm/(m-1))
I : :;
a {3
IDxDy G(x, y, t) ltl(n+lal+l/31)/m exp -c lip/(m-1) (5.153)

with another positive constant c. Since

d
-exp(-tA
dt
2) = --.
1
21r2
J
r
.\e ->.t (A 2 - ,\ )-1 d.\,

we can verify similarly

1
8 I C
8t G(x, y, t) :::; 1tln/m+1 exp -c
( lx- Ylm/(m-1))
ltjl/(m-1) . (5.154)
5.3. ESTIMATES OF KERNELS 207

Next we estimate the kernel of (A 2 - .A)- 1. Letting 0 < 01 < 1r /2 - 00 , 0 <

E1 < 1 we consider in the region

Im.A > 0, IRe.AI < (I _ E1) sin 81 . (5.155)

Im.A - cos 81
For t = reifh

Re.At = Re.A · Ret - Im.A · Imt

= Re.A · rcos01- Im.A · sin01::::; -~:1rsin01 · Im.A.
Hence there exists a positive constant c such that
Re.At ::::; -cri.AI. (5.156)
Therefore if .A is in the region (5.155) we can integrate along the half line
t = rei1h, 0 < r < oo, in the following integral

(A 2 - .A)- 1 = j e->.t exp( -tA )dt. 2 (5.157)

If we denote the kernel of (A 2 - .A)- 1 by K;>..(x, y), then by (5.157)

K;>..(X, y) = 1=re; 61
e>.tG(x, y, t)dt. (5.158)

In view of (5.145),(5.156)

IK>.(X, y)j ::::;

{oo e-cri>.i rn/m
Jo
C ( lx - Ylm/(m-1))
exp -c r1/(m-1) dr. (5.159)

In case n > m changing the variable as r = lx- ylms and putting h =

(I.AI1/mlx- yl)1-m

1oo r-nfme-cri>.i exp( -clx- Ylm/(m-1)r-1/(m-1))dr

= lx- Ylm-n 1oo s-n/m exp( -cs-1/(m-1)- ci.AIIx- yjms)ds
::::; jx- yjm-n 1h s-n/m exp( -cs-1/(m-1))ds

+lx- Ylm-n 1oo s-n/m exp( -ch-m/(m-1)s)ds = 11 +h.

If I.AI 1/mlx- Yl < 1, using

1h s-n/m exp( -cs-1/(m-1))ds::::; 1oo s-n/m exp( -cs-1/(m-1))ds < oo,

208 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

1h s-n/m exp( -cs-1/(m-1))ds:::; C exp( -I.AI1/mlx- yl).

If I.AI 1/mlx- Yl 2:: 1, noting h:::; 1 and using Lemma 5.16 we have for some
positive constant c'

lh s-n/m exp( -cs-1/(m-1))ds

:::; C 1h exp(-c's- 1/(m- 1))ds:::; Chexp(-c'h- 1/(m- 1))

:::; C exp( -c' h- 1/(m- 1)) = C exp( -c'I.AI 1/mlx- yl).

Consequently for some constants C, c

l1 :::; Glx - Ylm-n exp( -ci.AI 1/m lx - yl). (5.160)

Again with the aid of Lemma 5.16

1 00
s-n/m exp( -ch -m/(m- 1) s )ds

:::; exp( -ch-1/(m-1)) 1oo s-n/mds

= __!!!_h-(n-m)/m exp( -ch-1/{m-1))
n-m
:::; C exp( -c"h- 1/(m- 1)) = C exp( -c"I.AI 1/mlx- yl). (5.161)

Therefore we have shown that if n > m the following inequality holds:

(5.162)

In case n = m with the aid of the change of the variable r = lx- Ylms

1oo r-1e-cri.AI exp( -clx- Ylm/(m-1)r-1/{m-1))dr

= 1 00
s- 1 exp(-cs- 1/(m- 1))exp(-ch-m/(m- 1)s)ds. (5.163)

If I.AI 1/mlx- Yl < 1, put b = (I.AIIx- Ylm)- 1 = hm/(m- 1). Then b > 1. The
right hand side of (5.163) does not exceed

lab s- 1 exp( -cs- 1/(m- 1))ds + 1 00

1
s- 1 exp( -cb- s)ds
5.3. ESTIMATES OF KERNELS 209

::=; 11 8-1 exp( -C8-1/(m-1))d8 + 1b 8-1d8 + 1oo e-cs ~8

:::; C [1 + log(l>-11/mlx- yl)-1) .

If 1>-1 1/mlx- Yl ~ 1, h:::; 1. The left hand side of (5.163) does not exceed

1h 8- 1 exp( -c8- 1/(m- 1l)d8

+ 100
8- 1 exp( -c8- 1/(m- 1)) exp( -ch-m/(m- 1) 8)d8

:::; exp( -2- 1ch- 1/(m- 1)) 1h 8- 1 exp( -2- 1c8- 1/(m- 1l)ds

+ exp( -2- 1ch- 1/(m- 1l)

X [1 1
8- 1 exp( -c8- 1f(m- 1))d8 + 100
8- 1 exp( -2- 1 c8)d8]

:::; Cexp(-2- 1cl>-l 1/mlx- yl).

Therefore if n = m we have

Finally in case n < m making the change of the variable r = 8/1>-1 and
setting h = 1>-1 1/mlx- Yl

1oo r-nfme-cri.XI exp( -clx- Ylm/(m-1)r-1/(m-1))dr

= 1>-ln/m-1 1 00
8 -nfme-csexp(-chm/(m-1) 8 -1j(m-1))d8

h
::=; l>-ln/m-1exp(-ch) 1 8-nfme-csd8

+l>-ln/m-1e-chf2100 8 -nfme-csf2d8 :::; Cl>-ln/m-1e-ch/2.

Hence if n < m we have

(5.165)

Analogously in the region lm,\ < 0, IRe>-1/IIm>.l :::; (1 - ~: 1 ) sin Otf cos 01
we deform the integral path in the integral (5.158) to the half line t =
re-i 0 t, 0 < r < oo, to obtain the same estimates as (5.162),(5.164),(5.165).
210 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

In the set - Re.A/ lim .A I > (1 - f 1) sin lh /cos 81 we can establish the esti-
mates of the same form integrating along the positive real axis in the integral
(5.158). If we use (5.153) we can obtain the estimates of D~D~K_x(x, y), lal <
m, I.BI < m. It is clear that G(x, y, t), K,x (x, y) are also the kernels of
exp( -tAp), (Ap - >.)- 1 respectively. Thus recalling that we replaced L by
L + >.0 we have proved the following theorem.
Theorem 5. 7 Suppose that the hypothesis of Theorem 5.6 are satisfied.
Then the kernels G(x, y, t), K_x(x, y) of exp( -tAp), (Ap- >.)- 1, 1 < p < oo,
exist, and there exist positive constants C, c, a real number w and an angle
8o E (0, 1r /2) such that
(i) for lal < m, I.BI < m, (x, y) En X n, I argtl < n/2- 8o,

C ( lx- Ylm/(m-1))
ID~ DeG(x, y, t) I ::::; ltl(n+lai+I.BI)/m exp -c ltjl/(m-1) ewltl'
(5.166)

I0 I C ( lx- Ylm/(m- 1))

at G(x, y, t) ::::; 1tln/m+1 exp -c IW/(m-1) ewltl. (5.167)

(ii) for arg A ~ ( -8o, 8o), I.AI > c, Ia I < m, I.BI < m, (x, y) E nX n

m < n + Ia I + I.BI
m > n + lal + I.BI . (5.168)
m = n + Ia I + I.BI
Since for I arg tl < 1r /2 - 8o, I arg sl < 1r /2 - 8o
exp( -(t + s)Ap) = exp( -tAp) exp( -sAp),
we have
G(x, y, t + s) = k G(x, z, t)G(z, y, s)dz. (5.169)

5.4 Boundary Value Problems in L 1

Suppose that the assumptions of Theorem 5.6 are satisfied. It will be shown
that the realization -A 1 of the operator -Lin L 1(0) under the boundary
conditions Biulon = 0, j = 1, ... , m/2, generates an analytic semigroup
in L 1(0). In this connection we mention H. Amann [13] and D. Guidetti
[72]. Amann obtained the desired result for second order operators using
the theory of dual semigroups. Guidetti established directly the estimate of
the resolvent kernel for systems without considering the adjoint problem.
5.4. BOUNDARY VALUE PROBLEMS IN L 1 211

Set
lxlm/(m-1))
H(x, t) = exp ( -c ltjl/(m- 1)
so that by Theorem 5. 7
cewltl
IG(x, y, t)l :::; ltln/m H(x- y, t). (5.170)

With the aid of the change of the variable x = 1t1 1fmy we get

{ H(x, t)ds = ltln/m f exp( -clylm/(m- 1l)dy. (5.171)

Jnn }nn

(G(t)f)(x) =in G(x, y, t)f(y)dy. (5.172)

In view of (5.170),(5.171) we have for I argtl < 7f/2- Oo

in I(G(t)f)(x)ldx:::; in in~~:~: H(x- y, t)dxlf(y)ldy

cewltl r r
:::; ltln/m Jnn H(x, t)dx Jn IJ(y)ldy

= cewltl r
}nn
exp( -clylm/(m- 1))dylflo,1,fl· (5.173)

Hence G(t) is a bounded linear operator from £ 1(0) to itself. By virtue of

(5.169) we have

G(t+s)=G(t)G(s), largtl<7f/2-0o, largsl<7f/2-0o. (5.174)

Let f E Co(O) and f(x) = 0 for lxl > N. Then

r I(G(t)f)(x)ldx:::; r rcii:~:H(x- y, t)IJ(y)ldydx

1
Axi>N+l J!x!>N+1 Jn t
cewlt!l
:::; lflo,oo,n I ln/m H(x- y, t)dydx

1
t !x!>N+l !Y!<N
cewlt!1
:::; lflo,oo,fl ltln/m H(x- y, t)dxdy
IYI<N !x-y!>1

:::; l!lo,oo,nCewltl { dy { exp( -clxlm/(m-1l)dx.

}!Y!<N J!x!>!t!- 11=
212 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Therefore

lim r
t->O Jlxi>N+l
I(G(t)f)(x)- f(x)ldx =lim
t->O Jlxi>N+l
r
I(G(t)f)(x)ldx = 0.
(5.175)
On the other hand with the aid of Holder's inequality

r
Jlxi<;,N+l
I(G(t)f)(x)- f(x)ldx

::::; l{x E it; lxl::::; N + 1}ll-l/PIG(t)f- flo,p,n ---t 0 (5.176)

as t ---t 0, where 1 < p < oo. From (5.175),(5.176) it follows that
lim IG(t)f- flo 1 n ---t 0. (5.177)
t->0 ' '
Since C 0 (!1) is dense in £ 1 (!1) and IIG(t)ll.ccu,£1) is uniformly bounded if
tis bounded in view of (5.173), we conclude that (5.177) holds for every
f E £ 1 (!1). By virtue of (5.167) G(t)f is differentiable in t and
l(djdt)G(t)flo,l,f!::::; Cewtlflo,l,n/t, t > 0,
for any f E £ 1 (!1). Thus we have shown that G(t) is an analytic C0 -
semigroup in £ 1 (!1). The infinitesimal generator of G(t) is denoted by A 1 :
G(t) = exp( -tA 1 ). (5.178)
If K is a complex number, the kernel of exp( -t(Ap + K)) = e-~<t exp( -tAp)
is e-~<tG(x, y, t). Hence replacing L(x, D) by L(x, D)+ K for a sufficiently
large positive number K, we assume that (5.166),(5.167) hold for w < 0.
Then, in view of Lemma 1.2 we have for every 1 < p < oo, f E £P(i1), t > 0

Iexp( -tAp)flo,p,f!::::;
cewt r
tn/m JR .. H(x, t)dxlflo,p,f!

= Cewt
{ exp( -clxlm/(m-ll)dxlflo,p,n, (5.179)
JR,.
and hence 0 E p(Ap)· Similarly by (5.173) 0 E p(AI). By Theorem 5.7 the
kernel K(x, y) = K 0(x, y) of Ap- 1 exists and there exist positive constants
C, c such that for let I < m, I,BI <m
ID~D~K(x,y)l::::; ce-clx-yl
lx _ Ylm-n-lal-1.61 m- n - Ia I - I,BI <0
X{ 1 m- n -lai-I,BI > 0 . (5.180)
1 +log+ lx - Yl- 1 m- n- lal -I,BI =0
Since A1 1 = J0= G(t)dt, K(x, y) is also the kernel of A;:- 1 •
5.4. BOUNDARY VALUE PROBLEMS IN L 1 213

Theorem 5. 8 The domain D(A 1) of A 1 coincides with the totality of

functions u satisfying the following three conditions:
(i) u E wm- 1,q(O) for any q satisfying 1:::; q < nj(n- 1);
(ii) L(x, D)u E £ 1 (0) in the sense of distributions;
(iii) for any p such that 0 < (njm)(1- 1/p) < 1 and v E D(A~,)

(Lu, v) = (u, L'v). (5.181)

FUrthermore, foru E D(Al), A 1u = L(x, D)u and ifmi < m-1, Bj(x,D)u =
0 on an.

Proof. First we verify that both sides of (5.181) are meaningful. Since
m-njp' > 0, we have v E wm,p' (0) c £=(n) by Theorem 3.19. Therefore
the left hand side makes sense. As for the right side it suffices to show that
u satisfying (i) belongs to £P(O). If n;:::: m, set 1/q = 1/r + (m- 1)/n for
r satisfying {1- (m -1)/n}- 1 :::; r < nj(n- m). Then 1:::; q < nj(n- 1).
Since m- 1- njq < 0, wm- 1 ,q(O) c U(O) in view of Theorem 3.17.
Hence u E U(O) for any 1:::; r < nj(n- m). Since 1 < p < nj(n- m), we
conclude u E £P(O). If n < m, we have wm- 1,q(O) c wn,q(O) c U(O)
for any q:::; r < oo by Theorem 3.18 or 3.19. Hence, u E U(O) for any
1:::; r < oo. Especially u E £P(O).
Suppose u E D(Al). Set A1 u = f. Then

u(x) = k K(x, y)f(y)dy. (5.182)

By virtue of (5.180) for Jal = m- 1

JD~K(x, y)J:::; CJx- yJ1-ne-clx-yl.

Hence, if 1:::; q < nj(n- 1) we get by Lemma 1.2 that

(JnrIDau(xWdx)1/q :::; C [L, (JxJ1-ne-clx1) q dx

] 1/q
lflo,1,n < oo.
Therefore
(5.183)
Let
0:::; (n/m)(1- 1/p) < 1. (5.184)
If m < n, we have in view of (5.180)
214 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Since 1::; p < nj(n- m),

k. (ixlm-ne-cfxlr dx < oo.

Hence, we get u E LP(O) applying Lemma 1.2. If m > n, IK(x, y)i <
Ce-cfx-yf. Therefore we readily obtain u E LP(O). If m = n, we have

IK(x, y) I ::; ce-cfx-yf (1 +log+ lx - Yl- 1),

k. [e-cfxf(1+log+lxl- 1 )r dx<oo,

and we see u E LP(O) also in this case. Consequently u E LP(O) in any

case. Since (5.184) is equivalent to 1 ::; p < nj(n- m) if m < n and to
1 ::; p < oo if m ~ n, we see that u E U(O) for 1 ::; q < nj(n- 1), and
hence combining this with (5.183) conclude that u satisfies (i).
Next suppose that 0 < (n/m)(1 - 1/p) < 1. Then u E LP(O) as was
shown above. Hence

(5.185)

As t --> 0 G(t)u --> u,LG(t)u = G(t)A 1 u --> A 1 u in L 1 (0). Therefore,

Lu = A 1u E L 1 (0) in the sense of distributions, and u satisfies (ii). If
v E D(A~, ), 0 < (n/m)(1- 1/p) < 1,

(5.186)

Since v E wm,p' (0) c L 00 (0) as was remarked at the beginning of the

proof, we obatin (5.181) letting t --> 0 in (5.186). Thus we have proved
that u E D(A 1 ) implies that u satisfies (i),(ii),(iii). Conversely suppose
that u satisfies (i),(ii),(iii). If v is the solution of A 1v = Lu, v also satisfies
(i),(ii),(iii), and L(u- v) = 0. Let g be an arbitrary element of LP' (0), 0 <
(n/m)(1- 1/p) < 1 and w the solution of A~,w = g, then by (iii)

0= (L(u-v),w) = (u-v,L'w) = (u-v,g).

Hence u = v E D(A 1 ).
Suppose u E D(A 1 ), 1 ::; q < nj(n- 1). If Ia: I ::; m- 1, 0 < s::; 1,

{ (
[ }R" C H(x s))q dx] 1/q < C snfmq-(n+faf)/m
s(n+faf)/m ' - q '

..!!:._ _ n + Ia: I > -1.

mq m
5.5 PROBLEMS IN SPACES OF CONTINUOUS FUNCTIONS 215

Hence as t ----. 0

IIG(t)u- ullm-1,q,.l1 =II Jort G(s)A1udsll

m-1,q,.l1
<
-
Cq tnfmq-(n-1)/miA 1 ui 0,1,.11 ___. 0.

Therefore, if mj < m - 1 we have Bju = 0 on 80.

5.5 Boundary Value Problems in Spaces of

Continuous Functions
In this section we assume that 0 is a bounded open subset of Rn of class
C2 . We consider elliptic boundary value problems in the space C(O). The
related results are H. B. Stewart [144],[145].
Let
L =- n 2: -
8 (
. . 1 8x3·
8
8x3· . 1
8
aii- ) + n b i - + c
8xi
2:
~.J= 2=

be an elliptic differential operator of second order satisfying the assumptions

of the previous section 5.4. Let A be the operator defined by

D(A) = {u E W 2 •P(O) for any p E (1, oo), u = 0 on 80, Lu E C(O)},

or

D(A) = {u E W 2 •P(O) for any p E (1, oo),8uj8nL = 0 on 80,Lu E C(O)},

and
Au= Lu for u E D(A)
according as we are concerned with the Dirichlet or Neumann boundary
condition respectively. Let Av be the operator defined by (5.34) for the
operator L and the boundary conditions mentioned above. Then by virtue
of Sobolev's imbedding theorem we have

Let G(x, y, t) and K>-.(x, y) be the kernels of the semigroup exp( -tAp) and
the resolvent (Av - >.)- 1 respectively. Then by Theorem 5. 7 there exist
positive constants C, c such that

jG(x, y, t) I :::; t~z exp ( -c ix ~ Yiz) ewt (5.187)

216 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

(5.188)

for x, yEn, t > 0, ,\ E E, where E = {,\; Bo < arg,\ < 27r-Bo, I-XI~ Cl}, 0 <
Bo < 7r /2.
Remark 5. 5 In the above the factor ewt and the restriction I-XI ~ cl
would be removed, but we are not interested in this problem here.
Let f be an arbitrary element of C(O). Since n is bounded, f E LP(n) for
any p E (1, oo). For,\ E Ewe put

u(x) = k K>.(x, y)f(y)dy. (5.189)

Then u E D(Ap) and (Ap - ,\)u = f. Hence Lu = f + ,\u E C(O) owing to

Sobolev's imbedding theorem. Consequently u E D(A) and

(A- ,\)u =f. (5.190)

Conversely if u is an element of D(A) such that (A - ,\)u = f. Then

(Ap - ,\)u = f for 1 < p < oo. Therefore (5.189) is a unique element
satifying (5.190). Hence we conclude p(A) ~ E. By virtue of (5.188)

lu(x)l ::; C k lx- Yl2-ne-ci>.il/21x-yidylflo,oo,n

::; c r IYI
}Rn
2 -ne-ci>.i 112 IYI dylflo,oo,n ::; Clflo,oo,n/1-XI

if n > 2. If n = 1, 2, we can prove the same result. Therefore

I(A- -X)- 1 flo,oo,n ::; Clflo,oo,n/1-XI.

Consequently -A generates an analytic semigroup exp( -tA) in C(O). The
kernel of exp( -tA) is

-21 . f e->.tK>.(x,y)d-X=G(x,y,t).
7rZ Jn
In the case of the Dirichlet boundary condition D(A) is not dense in C(O)
since
D(A) c C0 (0) = {u E C(O); u = 0 on an}.
5.5 PROBLEMS IN SPACES OF CONTINUOUS FUNCTIONS 217

Iff E C5(!1), then f E D(Av) for any 1 < p < oo. Therefore choosing p so
that p > n/2

Iexp( -tA)f- flo,oo,n = Iexp( -tAv)f- flo,oo,n

:=:; CIAv(exp( -tAv)f- f)lo,p,n = Cl exp( -tAv)Avf- Avflo,p,O----> 0
as t ----> 0. Since C5(!1) is dense in C0 (0), we see that for any f E Co(O)
exp( -tA)f----> f in C(O) as t----> 0.
In the case of the Neumann boundary condition we use the fact

{u E C 2 (0); 8uj8nL = 0 on 8!1} is dense in C(O) (5.191)

without proof. Iff belongs to the space (5.191), then f E D(Ap) for any
1 < p < oo. Hence we can show as in the case of the Dirichlet bound-
ary condition that for any f E C(O), exp( -tA)f ----> f in C(O) as t ----> 0.
Therefore, D(A) is dense in C(O).
Theorem 5. 9 The operator A defined at the beginning of this section gen-
erates an analytic semigroup in C(O). In the case of the Dirichlet boundary
condition D(A) = C0 (0), and in the case of the Neumann boundary condi-
tion D(A) is dense in C(O).
Remark 5. 6 We can prove an analogous result for operators of arbitrary
order.

Remark 5. 7 In the case of the Neumann boundary condition we can

prove that D(A) is dense in C(O) also in the following manner. Owing
to H. Triebel [154: Theorems 1.15.3, 4.9.1] and To Seeley [135],[136] we
have D(A~) = H; 9(!1) if 0 < 2() < 1 + 1/p, where H; 9(!1) is defined
through complex interpolation (A. P. Calderon [26]) between the Sobolev
spaces Wk,P(!1), k = 0, 1, 2, 0 In view of Sobolev's imbedding theorem
0 0.

H; (!1) C C(O) if 1/p < 2()/no Therefore, first choosing p so large that
9

nfp < 1 + 1/p and then() so that nfp < 2() < 1 + 1/p we have

D(A~) = H; 9 (!1) c C(O)o

Iff E D(A:), then

Iexp( -tA)f- flo,oo,n :=:; CIA~(exp( -tAv)f- f)lo,p,n

= Cl exp( -tAv)A~f - A~flo,p,n ----> 0

as t----> Oo Since C 2 (0) c W 2 •P(!1) c H; 9 (!1) = D(A:) and C 2 (0) is dense

in C(O), D(A~) is also dense in C(O)o Hence we conclude exp( -tA)f----> f
in C(O) for any f E C(O)o
218 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

5.6 Example of a Higher Order Operator

In this section we give an example of 6th order elliptic operator which is
not variational and generates an analytic semigroup.
Let

(5.192)

(5.193)

be elliptic operators of second and fourth oder respectively with smooth

coefficients in a bounded domain 0 in Rn with smooth boundary. Here
we assume A(x) = (aij(x)) is a real, positive definite symmetric matrix for
each X E n, and Llal=4 a,(x)t;<> > 0 for 0 -::j:. .; E Rn. For the sake of
simplicity we assume that a,(x), lal = 4, are real valued and the principal
part of 8 1 is formally selfadjoint. Let B 0 , B 1 be the realizations of Bo, 81 in
LP(O), 1 < p < oo, under the Dirichlet boundary conditions. We will show
that any half line starting from the origin except the positive real axis is a ray
of minimal growth of the resolvents of B 0 B 1 and B 1 B 0 • This is an extension
of what is given in A. Favini and H. Tanabe [62] where it is assumed that
81 = 136, with proof almost unaltered. It can be seen that these operators
are not variational by considering the case Bo = -t::.., 8 1 = tl 2 . By taking
the adjoint it suffices to consider only B 1 B 0 . Then the boudary conditions
are u = Bou = (8j8v)B0 u = 0 on 80.
Let x E 80 be fixed. Let v be the outward unit vector at x, t; E Rn be
tangential to 80 at x, and r ;::: 0, 0 < () < 27r, r 2 + lt;l 2 > 0. Let 8 1 , 8 2 ,83 be
the roots of the polynomial

.2:: a,(x)(8v + t;)a · (A(x)(8v + t;), 8V + t;)- rei 9 (5.194)

1<>1=4

with positive imaginary parts. In what follows we write A for A(x). What
is to be shown is that the polynomials

1, (A(8v + t;), 8V +.;), 8 (A(8v + t;), 8V + t;)

are linearly independent modulo (8- 81}(8- 82)(8- 83), namely

c1 + c2 (A(8v + t;), 8V + t;) + c38 (A(8v + t;), 8V + t;)

= c3(Av, v)(8- 81)(8- 82)(8- 83) (5.195)
5.6. AN EXAMPLE OF HIGHER ORDER OPERATOR 219

if and only if c1 = c2 = c3 = 0. Comparing the coefficients of both sides of

(5.195) we obtain

c1+ c2(A~, ~) = -c3(Av, v)818283,

2c2(Av, ~) + c3(A~, ~) = c3(Av, v)(8283 + 8381 + 8182), (5.196)
c2(Av, v) + 2c3(Av, ~) = -c3(Av, v)(81 + 82 + 83).

Calculating the determinant of the system of linear equations (5.196) for

the unknowns c1 , c2 , c3 we reduce the problem to showing

4(Av, ~) 2 + 2(Av, ~)(Av, v)(81 + 82 + 83)

-(Av, v)(AC ~) + (Av, v) 2 (8283 + 8381 + 8182) f:- 0. (5.197)

We use the following classical theorem on polynomials by Hermit and Biehler

(T. Takagi [147: Chapter 2]).

Lemma 5. 17 Let f(z) be a polynomial whose roots have imaginary parts

of the same sign. Decomposing the coefficients of f(z) into the sum of the
real and imaginary parts we write f(z) = g(z) + ih(z). Then the roots of
g(z) and h(z) are real distinct and separate one another.

Proof. We denote the degree of f(z) by n, and suppose that the imaginary
parts of the roots a 1, ... , an of f(z) are all positive. Let /(z) be the poly-
nomial whose coefficients are the complex conjugates of the corresponding
coefficients of f(z). Then we have

g(z) = (f(z) + /(z))/2, h(z) = (f(z) -f(z))/2i.

Let f(z) = ao(z- a1) ···(z-an)· Then f(z) = iio(z- a1) ···(z-an)· Put
z= (z - a1) · · · (z - an)
K=--.
iio
(z - a1) ... (z - an)' ao

Then g(z) = 0 and h(z) = 0 are equivalent to Z = K and Z = -K respec-

tively, and l"-1 = 1. Let Ima > 0. Then when z goes along the real axis from
-oo to oo, z - ~ goes along the unit circle from 1 to 1 counterclockwisely.
z-a
Hence when z goes along the real axis from -oo to oo, Z goes along the
unit circle n times counterclockwisely. When Z reaches K, g(z) = 0 holds,
and when Z reaches -K, h(z) = 0 holds. This is repeated n times.

Remark 5. 8 Either g(z) or h(z) may be of degree n - 1. In this case

letting one of the roots of g(z) or h(z) be oo, the conclusion of the theorem
holds.
220 CHAPTER 5 ELLIPTIC BOUNDARY PROBLEMS (CONTINUED)

Suppose that (5.197) is false. Then we have

4(Av, ~) 2 + 2(Av, ~)(Av, v)(81 + 82 + 83)
-(Av, v)(AC ~) + (Av, v) 2(8283 + 8381 + 8182) = 0. (5.198)
We apply the theorem of Hermit-Biehler to the polynomial
f(z) = (z- 81)(z- 82)(z- 83).
Then
g(z) = z 3 - Re(81 + 82 + 83) · z2
+Re(8283 + 8381 + 8182) · z- Re818283, (5.199)
h(z) = -Im(81 + 82 + 8 3 ) • z2
+Im(8283 + 8381 + 8182) · z- lm818283. (5.200)
In view of Lemma 5.17 and Remark 5.8 we have
g(z) = (z- a:1)(z- o:2)(z- o:3), (5.201)
h(z) = -Im(8 1 + 82 + 8 3 ) • (z- ,Bt)(z- ,82)
(note Im(8 1 + 82 + 83)-:/: 0) for some o:1, o: 2, o:3, fit. ,82 satsifying
0:1 < ,81 < 0:2 < ,82 < 0:3. (5.202)
It follows from (5.201),(5.202) that
2 2
g' ( 0:1 ; 0:2) = - ( 0:1 ; 0:2 ) < 0, g' ( 0:2 ; 0:3) = - ( 0:2 ; 0:3) < 0.

Hence g'(z) < 0 for o: 1 ; o: 2 < z < o: 2 ; a: 3 . Since (5.202) implies

0:1 + 0:2 < fi1 + ,82 < 0:2 + 0:3 , we h ave g I ( fi1 + 132 ) < 0. f'rom (5 .200) ,
2 2 2 2
(5.198) it follows that
fi1+fi2 lm(8283+8381+8182) (Av,~)
2 2Im(81 + 82 + 83) - - (Av, v)'
and hence
g' (- (Av, ~)) < 0. (5.203)
(Av, v)
On the other hand differentiating both sides of (5.199), substituting
z= - i1~: ~~, using the assumption (5.198) and Schwarz' inequality we get
g' (- (Av, ~)) = (A~,~) (Av, v) - (Av, ~) 2 > O.
(Av, v) (Av, v)2 -
This contradicts (5.203).
Chapter 6

Parabolic Evolution
Equations

This chapter is devoted to the study of evolution equations of parabolic type

du(t)jdt = A(t)u(t) + J(t), t E (0, T], (6.1)

u(O) = x, (6.2)

in a Banach space X. The main results are due to P. Acquistapace and

B. Terreni [3],[6],[8] and P. Acquistapace [1]. "Parabolic" means that A(t)
generates an analytic semigroup for each t. It is not assumed that A(t) is
densely defined. A. Yagi [161] gave another proof using fractional powers of
A(t). We also note that the result of A. Yagi [159] which is a very general
one in the case where A(t) is densely defined is independent of the above
results of [1],[3],[6],[8].

6.1 Equations with Coefficients

Differentiable in t
We denote the resolvent (.\- A)- 1 of an operator A by R(A, A). We state
the assumptions of this section:
(P1) There exists an angle ()0 E (n /2, n] such that
(i) p(A(t)) :J I;= {A; Iarg .\I < Bo} U {0} for each t E [0, T],
(ii) there exists a positive constant M such that

IIR(.\,A(t))ll:::; M/l.\1, ,\ E I;, t E [0, T]. (6.3)

221
222 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

(P2) For each A E ~ the operator valued function R(A, A(t)) oft is contin-
uously differentiable in [0, T] in the uniform operator topology, and there
exist positive constants L 1 and p E (0, I] such that

11:tR(A,A(t))ll:::; ~~~P' A E ~. t E [O,Tj; (6.4)

(P3) There exist positive constants L 2 and a E (0, I] such that

II !A(t)- 1 - !A(T)- 1 11:::; L2lt- Tla, t, T E [0, T]. (6.5)

In view of Theorem I.6 A(t) generates an analytic semigroup {erA(t); T ;::: 0}

under the assumption (PI). The following equalities hold:

r
erA(t) = _2I. e'TA R(A, A(t))dA,
nz Jr
T > 0, t E [0, T], (6.6)

nz Jr
r
A(t)erA(t) = _2I" Ae'TAR(A,A(t))dA, T > 0, t E [O,T], (6.7)

where r is a smooth path connecting ooe-i0 and ooei0 , 1r /2 < () < Bo, in ~.
The norm of X is denoted by 11·11· We denote by C([O, T]; X) the Banach
space consisting of all continuous functions defined in the interval [0, T] and
taking values in X with norm

llullcuo,TJ;X) = sup llu(t)ll·

tE[O,T[

For a E (0, I) ca ([0, T]; X) is the set of functions belonging to C([O, T]; X)
and Holder continuous with exponent a with norm
llu(t)- u(s)ll
llullca(lo,T];X) = llullc{[O,TJ;X) + sup
It- sla
t,sE(O,TJ,t#s
and C 1 ([0, T]; X) is the set of all continuously differentiable functions de-
fined in [0, T] and with values in X with norm

llullc 1 (JO,TJ;X) = llullcuo,TJ;X) + llu'llcuo,TJ;X)·

For a E (0, I)

cHa ([0, T]; X) = C\[0, T]; X) n { u; u' E ca([O, T]; X)}

with norm

llullc 1 +a([O,TJ;X) = llullc 1 ([0,TJ;X) + llu'llca(lo,T];X)·

6.1. EQUATIONS WITH COEFFICIENTS DIFFERENTIABLE 223

In addition we denote by

C((O, T]; X), C"((O, T]; X) C 1 ((0, T]; X) Cl+"((O, T]; X)

the set of functions belonging to

C([E, T]; X), C"([E, T]; X), C 1 ([E, T]; X), Cl+"([E, T]; X)

for each E > 0.

For 1 :::; p :::; oo, LP(O, T; X) is the set of all functions u with values in
X such that u is strongly measurable in (0, T) and llu(t)IIP is integrable in
(0, T) if p < oo and essentially bounded in (0, T) if p = oo. The norm of
LP(O, T; X) is defined by

(
T )1/p p < oo,
lluiiLP(O,T;X) = { 111u(t)11Pdt
ess suptE[O,TIIIu(t)ll p=oo.
The set of all bounded linear operators from a Banach space X into itself
is denoted by £(X). The norm of £(X) is also denoted by I · II· For two
real numbers a, f3 we write a 1\ f3 = min{ a, /3}.
Let f E C([O, T]; X) and x E D(A(O)).
Definition 6. 1 u is a classical solution of (6.1),(6.2) if u E C([O, T]; X) n
C 1 ((0, T]; X), u(t) E D(A(t)) for any t E (0, T] and

du(t)jdt = A(t)u(t) + f(t) t E (0, T], u(O) = x.

Definition 6. 2 u is a strict solution of (6.1),(6.2) if u E C 1 ([0, T]; X), u(t) E
D(A(t)) for any t E [0, T] and
du(t)jdt = A(t)u(t) + f(t) t E [0, T], u(O) = x.
Remark 6. 1 In addition to strict and classical solutions in the above def-
initions solutions in the following sense called strong solutions are consid-
ered in Acquistapace and Terreni [3]: u E C([O, T]; X) and there exists a
sequence {un} C C([O, T]; X) such that un(t) E D(A(t)) for any t E [0, T],
A(·)un(·) E C([O,T];X) and
(i) Un --+ u in C([O, T]; X),
(ii) u~- A(·)un = fn E C([O, T]; X) and fn--+ fin C([O, T]; X),
(iii) Un(O)--+ x in X.
Hypothesis (P3) is not necessary so long as strong solutions are concerned.
It is used only to prove the existence and regularity of strict and classical
solutions.
224 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

6.2 Preliminaries(!)
The following lemma is easily proved as in section 3 of Chapter 1.
Lemma 6. 1 Under the assumptions (P1),(P2) we have
(i) lie'~" A(t) II 5, C, t, T E [0, T],
(ii) IIA(t)e'~"A(t)ll5, CjT, t E [O,T], T E (O,T].
We define an operator valued function P(t, s), 0 5, s < t 5, T, by

P(t s) = __.:!:__ { e>-(t-s)~R(>. A(t))d>. = (~ + ~) e<t-s)A(t). (6.8)

' 21Ti lr at ' at as
Arguing as in the proof of (1.11) we can show that the following inequality
holds with some positive constant K:
K
IIP(t, s)ll 5, (t _ s)l-P, 0 5, s < t 5, T. (6.9)

Lemma 6. 2 Under the hypothses (P1),(P2) P(t, s) is continuous in 0 5,

s < t 5, Tin the norm topology of .C(X).
Proof. The assertion follows since the integral (6.8) converges in the norm
topology of .C(X) uniformly in { (t, s); t, s E [0, T], t - s ~ E} for any E E
(0, T).
Lemma 6. 3 Under the hypothesis (P1),(P2),(P3) we have for>. E I;, t, T E
[O,T]

II :t R(>., A(t)) - :T R(>., A( T)) 115, C(lt - Tla + 1>.1 1-Pit- Ti). (6.10)

Proof. Since

:tR(>., A(t)) = -A(t)R(>., A(t)) !A(t)- 1 • A(t)R(>., A(t)),

we have
a
otR(A, A(t))- 07 R(>., A(T))
a

= -[A(t)R(>., A(t)) - A(T)R(>., A(T))]! A(t)- 1 • A(t)R(>., A(t))

-A(T)R(>., A(T)) [! A(t)- 1 - d~ A(T)- 1] A(t)R(A, A(t))

-A(T)R(>., A(T))! A(T)- 1 • [A(t)R(A, A(t)) - A(T)R(A, A(T))].

6.2 PRELIMINARIES(l) 225

Combining this with

IIA(t)R(.\ A(t))- A(T)R(.\ A(T))II = p.[R(.\ A(t))- R(.A,A(T))JII

= IIA 1t :(JR(.\ A(CI))do-11 ::; Cit- TII.AI 1 -P,

we complete the proof.
Lemma 6. 4 Under the hypothses (P1),(P2),(P3) we have for 0::; s < T <
t ::; T and 0 < 6 ::; a 1\ p
IIP(t, s)- P(T, s)ll
::; C[(t- T)a(t- s)- 1 + (t- T)(t- s)- 1(T- s)P- 1] (6.11)
::; C(t- T) 0 (T- s)a/\p-o- 1. (6.12)

Proof. Since

P(t,s)- P(T,s) = 2 !i t eA<t-s) [:tR(.A,A(t))- :TR(.A,A(T))] d.A

+-21·1t-s r.AeACT ~ R(.A, A(T))d.Ado-,

lit
7r'l T-S }[' uT

.AeAa :TR(.A,A(T))d.\11::; CCJP-2, (6.13)

we get with the aid of Lemma 6.3

IIP(t, s)- P(T, s)ll
<C
-
[(t- T)a
t-s
+ t- T
(t-s) 2-P
+ 1t-s (Jp-2dCJ]
r-s
. (6.14)

Using (6.14) and

t- T 1t-s 1
-:-------:-:::-- <
(t-s) 2 -p
CJP- 2dCJ =- - [(T- s)P- 1 -
1-p
(t- s)P- 1 ]

_1 (T- [1- (T-

T-S

=
1-p
s)P-1
t-s
s) 1-pl
< - -1 ( T - s)P -l (
1--- T- S) = --(T-s)P
1 _ - T
t-- 1
1-p t-s 1-p t-s
we obtain (6.11). The inequality (6.12) is an easy consequence of (6.11).
Proposition 6. 1 If the assumptions (P1),(P2) are satisfied, then P(·, O)x
E C((O,T];X)nL 1 (0,T;X) forx EX. Ifmoreover the assumption (P3) is
satisfied, then P(·, O)x E ca((O, T]; X).
226 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Proof. The assertion follows from Lemma 6.2, (6.9) and Lemma 6.4.
Proposition 6. 2 Suppose that the assumptions (P1),(P2) are satisfied. Then,
we have
(i) The operator valued function etA(t) is continuously differentiable in (0, T]
in the norm topology of £(X) and
(djdt)etA(t) = A(t)etA(t) + P(t, 0). (6.15)
(ii) Let x EX. Then etA(t)x E C([O, T]; X) if and only if x E D(A(O)). In
this case etA(t)xlt=O = x.
(iii) If moreover (P3) is satisfied, then etA(t)x E Cl+o:((O, T]; X) .
Proof. The assertion (i) follows from

.!!_etA(t)
dt
= .!!._ [~
dt 2nz Jr
r e).tR(>.. ' A(t))d>..]
= -21 . { >..e:>.tR(>..,A(t))d>..+ -21 . { e:>.t ~ R(>..,A(t))d>..
nz Jr nz Jr ut
= A(t)etA(t) + P(t, 0).
The assertion (ii) follows from
etA(t)x _ x = (etA(t) _ etA(O))x + etA(O)x _ x,

lletA(t) -etA(D) II = 112:i ke:>.t(R(>.., A(t)) - R(>.., A(O)) )d).. II

<c
- Jr
reRe:>.t_t_id>..i
IA.IP
< CtP
-
and Theorem 1. 7.
(iii) By Proposition 6.1 P(·, O)x E co:((O, T]; X). Hence in view of (6.15) it
suffices to show that A(t)etA(t)x E Co:((O, T]; X). For 0::; T < t::; T
IIA(t)etA(t)- A(T)erA(r)ll
::; IIA(t)etA(t)- A(T)etA(r)ll + IIA(T)etA(r)- A(T)erA(r)ll
= ~~2~i t >..e:>.t(R(>..,A(t)) -R(>..,A(T)))d>..ll +lilt :sA(T)esA(r)dsll

::; C t IA.IeRe:>.t t,~; IdA. I+ C 1t ~= :; C [(t- T)tp-Z + (t- T)(tT)- 1].

Thus the proof is complete.

Next we investigate the operator P defined by

(P J)(t) =lot P(t, s)f(s)ds. (6.16)

6.2 PRELIMINARIES(l) 227

Proposition 6. 3 Suppose that the assumptions (P1),(P2) are satisfied.

(i) P E .C(L 1 (0, T; X)) n .C(C([O, T]; X)).
(ii) Iff E C((O, T]; X) n L 1 (0, T; X), then Pf E C((O, T]; X).
Suppose moreover that (P3) is satisfied.
(iii) Iff E C((O,T];X) n£ 1 (0,T;X), then Pf E C 6 ((0,T];X) for any
8 E (0, a) n (0, p].
Proof. (i) Using (6.9) and Fubini's theorem we easily obtain
KTP
IIPflluco,T;X) ~ -P-II!IIuco,T;X)·

That P E .C(C([O, T]; X)) is an easy consequence of (6.9) and the continuity
of P(t, s) in 0 ~ s < t ~ T.
(ii) For t > T > E > 0

II(PJ)(t)- (PJ)(T)II =lilt P(t,s)f(s)ds

+ lr (P(t, s)- P(T, s))f(s)ds + 1< (P(t, s)- P(T, s))f(s)dsll

~ C(t- T)P <~~ET llf(s)ll + lr IIP(t, s)- P(T, s)llds <~~ET IIJ(s)ll
+ 0 ~~~< IIP(t, s)- P(T, s)ll1< llf(s)llds. (6.17)

Each term of the last side tends to 0 as t ---> T or T ---> t.

(iii) With the aid of (6.17), Lemma 6.4 and its proof
II(PJ)(t)- (PJ)(r)ll ~ C(t -r)P sup llf(s)ll
<-5,8-5,T

+C r [(t- r)a(t- s)- 1 + lt- 8

aP- 2 dn] ds <-5,8-5,T
sup llf(s)ll

]1<
}< T-8

(t-r)a t-T
+C sup [ + - ( r - s)P- 1 llf(s)llds.
o-5, 8 -5,< t- S t- S 0

Since

lr [<t- r)a (t- s)- + 1~~8 aP- da] ds

1 2

t- E 1
= (t- r)a log--+ [(r- E)P + (t- r)P- (t- E)P]
p(1- p) t- T
1 t- E
<(t-r)alog--+ (t-r)P,
t- T p(1- p)
228 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

we obtain

I!(PJ)(t)- (Pj)(T)II::::; C(t -T)P sup IIJ(s)ll

<"5os"5oT
t-E
+C(t- T)"' l o g - sup IIJ(s)ll
t - T <"5os"5oT

+C [(t- T)"'(t- E)- 1 + (t- T)(t- E)- 1 (T- E)P- 1 ] 1< IIJ(s)ilds,

which implies the desired result.

Let Q be the operator defined by

(6.18)

Since Pis an integral operator of Volterra type, the inverse of 1 + P exists

and is given by
(QJ)(t) = f(t) + l t R(t, s)f(s)ds, (6.19)

where
00

R(t,s) = 2)-1)nPn(t,s), (6.20)

n=1
P 1(t,s)=P(t,s), (6.21)

Pn(t,s)= 1tP(t,T)Pn-1(T,s)dT, n=2,3,.... (6.22)

By induction we can easily show that

IIPn(t, s)ll ::::; (~~;:;r (t- s)np-1, (6.23)

Pn(t, s) = 1t Pn-1(t, T)P(T, s)dT, n = 2, 3, ... , (6.24)

where K is the constant in (6.9). By virtue of (6.23) the series (6.20) is

uniformly convergent in any compact set of { (t, s); 0 ::::; s < t ::::; T} in the
norm topology of .C(X). Hence R(t, s) is continuous in 0 ::::; s < t ::::; T in
the norm topology, and satisfies

c (6.25)
IIR(t,s)ll::::; (t-s)l-P
6.2 PRELIMINARIES(l) 229

With the aid of (6.22),(6.24) it is easy to show that

P(t, s) + R(t, s) + 1t P(t, T)R(T, s)dT = 0, (6.26)

P(t, s) + R(t, s) + 1t R(t, T)P(T, s)dT = 0. (6.27)

In view of (6.25) and the continuity of R(t, s) in (t, s) the assertions of

Proposition 6.3 (i),(ii) hold with P(t, s) replaced by R(t,s).
Proposition 6. 4 Under the assumptions (P1),(P2) we have
(i) Q E .C(L 1 (0, T; X)) n .C(C([O, T]; X)).
(ii) Iff E C((O, T]; X) n £ 1 (0, T; X), then Qf E C((O, T]; X).
Under the assumptions (P1),(P2),(P3) we have
(iii) Iff E L 1 (0,T;X)nC6 ((0,T];X), o E (O,a)n(O,p], thenQf E
C 6 ((0, T]; X).
Proof. The assertions (i),(ii) are consequences of Proposition 6.3 (i),(ii)
with R(t, s) in place of P(t, s). We now prove (iii). If we set g = Qf =
f + Rf, where (Rf)(t) = J~ R(t, s)f(s)ds, then Rf = -Pg. In view of
Proposition 6.3 (ii) with P replaced by R we get Rf E C((O, T]; X). Hence
g E C((O,T];X). Clearly g E L 1 (0,T;X). Hence Pg E C 6 ((0,T];X) in
view of Proposition 6.3 (iii). Thus we conclude
Qf = f + Rf = f- Pg E C 6 ((0, T]; X).

Let T be the operator defined by

(Tf)(t) = l t e(t-s)A(t) f(s)ds. (6.28)

Proposition 6. 5 Suppose that the assumptions (P1),(P2) are satisfied.

(i) Iff E L 1 (0,T;X), then Tf E C([O,T];X) and (Tf)(O) = 0.
(ii) Iff E £ 1 (0, T; X) n C 6 ((0, T]; X), o E (0, 1], then (TJ)(t) E D(A(t))
fortE (0, T] and

A(t)(Tf)(t) = l t A(t)eCt-s)A(t)(j(s)- f(t))ds + (etA(t)- 1)f(t). (6.29)

(iii) Iff E £ 1 (0, T; X) n C 6 ((0, T]; X), o E (0, 1], then Tf E C 1 ((0, T]; X)
and

d
-(Tf)(t) = l t A(t)e(t-s)A(t)(j(s)- f(t) )ds+etA(t) f(t)+ l t P(t, s)f(s)ds.
ili 0 0
(6.30)
230 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

If moreover (P3) is satisfied, we have also

(iv) Iff E £l(O,T;X)nC6 ((0,T];X),o E (O,a)n(O,p], thenTf E
C 1H((O, T]; X).
Proof. The assertion (i) is a simple consequence of Lebesgue's dominated
convergence theorem.
(ii) For 0 < E < t

lot-< e(t-s)A(t) f(s)ds =lot-< e(t-s)A(t)(J(s)- f(t))ds

+lot-< A(t)e(t-s)A(t) A(t)- 1f(t)ds

=lot-< e(t-s)A(t)(J(s)- f(t))ds + (etA(t)- e<A(t))A(t)-1 f(t).

Letting E---> 0 we get

(Tf)(t) =lot e(t-s)A(t)(J(s)- f(t))ds + (etA(t) -1)A(t)- 1f(t). (6.31)

Since
C,(t- s) 6 - 1 E<s<t
IIA(t)e(t-s)A(t)(J(s)- f(t))ll :::; { C
-llf(s)-
t-E
f(t)ll O<s<E

each term of the right hand side of (6.31) belongs to D(A(t)) and (6.29)
holds.
(iii) Suppose 0 < E < T < t:::; T. Then

(Tf)(t)- (TJ)(r) = 1t e(t-s)A(t) f(s)ds

+lor (e(t-s)A(t)- e(r-s)A(r))f(s)ds

= 1t e(t-s)A(t)(J(s)- f(t))ds + 1t e(t-s)A(t) f(t)ds

+lor (e(t-s)A(t) _ e(r-s)A(t))f(s)ds

+lor (e(r-s)A(t) _ e(r-s)A(r))f(s)ds

= 1t e(t-s)A(t)(J(s)- f(t))ds + (e(t-r)A(t)- 1)A(t)-1 f(t)

6.2 PRELIMINARIES(1) 231

+ 1-r (e<t-s)A(t)- e<-r-s)A(t))(J(s)- f(T))ds

+ 1-r (e<t-s)A(t) _ e(-r-s)A(t))j(T)ds

+ 1-r (eC-r-s)A(t)- e(-r-s)A(-r))f(s)ds.

The fourth term in the last side is equal to
(etA(t) _ e(t--r)A(t) _ e-rA(t) + 1)A(t)-1j(T)
= (e(t--r)A(t)- 1)(e-rA(t) -1)A(t)-1 j(T)
= (e(t--r)A(t)- 1)(e-rA(t)- 1)A(t)- 1(j(T)- j(t))
+(e(t--r)A(t) _ 1)(e-rA(t) _ etA(t))A(t)-1 f(t)
+(e(t--r)A(t)- 1)(etA(t)- 1)A(t)-1 f(t).

Hence
(Tf)(t)- (Tf)(T) = _1_ t
e(t-s)A(t)(J(s)- f(t))ds
t- T t- T 1-r
eCt--r)A(t) _ 1
+ A(t)- 1 f(t)
t-T
{T e(t-s)A(t) _ e(-r-s)A(t)
+ Jo t_T (J(s)- f(T))ds
eCt--r)A(t) - 1
+ (e-rA(t)- 1)A(t)-1(j(T)- f(t))
t-T
eCt--r)A(t) _ 1
+ (e-rA(t)- etA(t))A(t)-1 f(t)
t-T
e(t--r)A(t) - 1
+ (etA(t) -1)A(t)-1 f(t)
t-T
{-r e(-r-s)A(t) _ eC-r-s)A(-r) _ 7
+ Jo t_T f(s)ds- ~Ii.

It is easily seen that I 1 , I4 , I5 go to 0 as T---+ t, and by Lemma 1.5 (i)

eCt--r)A(t) _ 1
lz + I6 = A(t)-1etA(t) f(t) ---+ etA(t) f(t).
t-T
By Lebesgue's dominated convergence theorem Ir ---+ J~ P(t, s)f(s)ds. Fi-
nally we investigate h. We have for 0 < E < t - 'T/ < T < t::::; T

h - i t A(t)e(t-s)A(t)(J(s)- f(t))ds
232 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

t-'TI [e(t-s)A(t) _ e<r-s)A(t) ]

= 1 - A(t)e<t-s)A(t) (J(s)- f(T))ds
t-T
+ 1t-'TI A(t)e<t-s)A(t)ds(J(t)- f(T))
r e(t-s)A(t) _ e(r-s)A(t)
+l (J(s)- f(T))ds
t-Tl t- T

- l t A(t)e<t-s)A(t)(f(s)- f(t))ds. (6.32)

t-Tl

Since
lle(t-s)A(t) _ e<r-s)A(t) II

= lilt A(t)e(a-s)A(t)dCT" ::; C 1t CT d~ s =Clog!=:' (6.33)

t-S =log ( 1 + -
log-- t-T)
- t-T
::; --, (6.34)
T-S T-S T-S

the norm of the third term in the right hand side of (6.32) does not exceed

The norm of the last term is also dominated by C,'f/6 • The second term is
equal to
(etA(t)- e'TIA(t))(j(t)- f(T)),

hence its norm is dominated by C,(t- T) 6 . The integrand of the first term
tends to 0 as T---+ t for each fixed s. Its norm does not exceed

c (-1- + - 1- ) ilf(s)- f(T)II::; __!!____ (IIJ(s)ll + sup 11!(7 )11) ·

T- S t - S T- S <5or5,T

Hence by Lebesque's dominated convergence theorem the first term goes to

0 as T ---+ t for each fixed 'f/· Thus

h---+ 1t A(t)e<t-s)A(t)(J(s)- f(t))ds

as T ---+ t. Summing up we have established that T f has a left derivative

which is equal to the right hand side of (6.30).
We prove now the continuity of the left derivative of T f. For 0 < T <
6.2 PRELIMINARIES(l) 233

t:::; Twe have

1t A(t)e<t-s)A(t)(f(s)- f(t))ds -1-r A(T)e<-r-s)A(-r)(f(s)- f(T))ds

= 1t A(t)e<t-s)A(t)(f(s)- f(t))ds

+ 1-r [A(t)e<t-s)A(t)(f(s)- f(t))- A(T)e<-r-s)A(-r)(f(s)- f(T))] ds

= 1t A(t)e<t-s)A(t)(f(s)- f(t))ds
+ 1-r [(A(t)e(t-s)A(t)- A(T)e<t-s)A(-r))(f(s)- f(t))

+(A(T)e<t-s)A(-r)- A(T)e<-r-s)A(-r))(f(s)- f(T))

+A(T)e(t-s)A(-r)(f(T)- f(t))] ds

= 1t A(t)e<t-s)A(t)(f(s)- f(t))ds
+ 1-r (A(t)e<t-s)A(t)- A(T)e<t-s)A(-r))(f(s)- f(t))ds

+ 1-r (A(T)e<t-s)A(-r)- A(T)e<-r-s)A(-r))(f(s)- f(T))ds

+(etA(-r)- e(t--r)A(-r))(j(T)- j(t)). (6.35)
Using
IIA(t)e<t-s)A(t)- A(T)e(t-s)A(-r)ll:::; C(t- T)(t- s)P-2,
IIA(T)e(t-s)A(-r)- A(T)e(-r-s)A(-r)ll:::; C t- T
(t-s)(T-s)
we see that for 0 < E < T < t:::; T the norm of the left hand side of (6.35)
does not exceed

C, 1t (t- s) 6 - 1ds + C, 1-r (t- T)(t- s)P+li- 2ds

+C r(t- T)(t- s)P- 2 IIf(s)- f(t)llds + C, 17" t- T (T- s) 6 - 1ds

Jo , t- s
+C 1 <
(
t-S
t-T
)(
T-S
) IIJ(s)- f(T)IIds
{j
+ C,(t- T) . (6.36)

Using t- T:::; (t- T) 6 (t- s) 1 - 6 in the second term of (6.36) and

-r (T- s)li-1 1-r sli-1
1 -'---____;_-ds < ds
t-s 0 t-T+s
234 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

= 1 rj(t-r) 0"6-1
--cfn(t- T) 0 - 1 <
1= 0"6-1
--cfn(t- T) 0 - 1 (6.37)
0 1+0" 0 1+0"
in the fourth term of (6.36) we get

ll1t A(t)e(t-s)A(t)(J(s)- f(t))ds -1r A(T)e<r-s)A(r)(f(s)- j(T))dsll

:::; C,(t- T) 6 + C (t-t -)

f
T2 _
p
1'
0
IIJ(s)- f(t)llds

+C (
t-
t
f
~TT - f ) } r0 iif(s)- j(T)jjds. (6.38)

On the other hand we have

lletA(t) f(t)- erA(r) j(T)ii :::; lletA(tl(j(t)- j(T))Ii
+ii(etA(t)- etA(rl)j(T)Ii + ii(etA(r)- erA(r))j(T)ii
:::; C,(t- T) 6 + C(t- T)tP- 1 IIf(T)II +Clog .!llf(T)II· (6.39)
T

Thus the left derivative of Tf is continuous, and applying Lemma 1.3 we

complete the proof of (iii).
(iv) The assertion is a direct consequence of (6.38),(6.39) and Proposition
6.3 (iii).

6.3 Classical Solutions

Using the results in the previous section we can prove the existence of clas-
sical solutions of (6.1),(6.2).
Theorem 6. 1 Suppose that the assumptions (P1),(P2),(P3) are satisfied.
If x E D(A(O)) and f E C([O, T]; X) n C 17 ((0, T]; X), CT E (0, 1], then the
function u defined by

u(t) = etA(t)x + 1t e<t-s)A(t) [(I+ P)- 1 (!- P(·, O))x] (s)ds (6.40)

is a classical solution of (6.1),(6.2). Moreover u E Cl+aAo((O, T], X) for

oE (0, a) n (0, p].
Proof. In view of Proposition 6.2 (ii) etA(t)x E C([O, T]; X) and etA(t)xlt=O =
x. Since P(·, O)x E £ 1 (0, T; X) by Proposition 6.1, we have f- P(·, O)x E
£ 1 (0, T; X). Hence by Proposition 6.4 (i)

g := (1 + P)- 1 (!- P(·, O)x) E £ 1 (0, T; X). (6.41)

6.3 CLASSICAL SOLUTIONS 235

Therefore by Proposition 6.5 (i) we have Tg E C([O, T]; X) and (Tg)(O) = 0.

Thus u E C([O, T]; X) and u(O) = x. Let 6 E (0, a) n (0, p]. In view of
Proposition 6.1
f- P(·, O)x E Ca 11 a((O, T]; X) C Ca 116 ((0, T]; X).
Hence by Proposition 6.4 (iii)
g E Ca 116 ((0, T]; X). (6.42)
Therefore applying Proposition 6.5 (ii) we get (Tg)(t) E D(A(t)) for t E
(0, T]. Thus u(t) E D(A(t)) fortE (0, T] and
A(t)u(t) = A(t)etA(t)x

+ 1t A(t)e(t-s)A(t)(g(s)- g(t))ds + (etA(t)- 1)g(t).

By virtue of Proposition 6.2 (iii) etA(t)x E Cl+a((O, T]; X). From (6.41),(6.42)
and Proposition 6.5 (iv) it follows that Tg E Cl+allo((O, T]; X), and by
Proposition 6.5 (iii)

.:!:_(Tg)(t) = rt A(t)e(t-s)A(t)(g(s)- g(t))ds + etA(t)g(t) + rt P(t, s)g(s)ds

dt Jo Jo
fortE (0, T]. Hence u E Cl+a/\6((0, T]; X) and

du(t)jdt- A(t)u(t) = P(t, O)x + 1t P(t, s)g(s)ds + g(t) = f(t)

since (1 + P)g = f- P(·, O)x. Thus the proof is complete.
Next we prove the uniqueness of a classical solution. We set

P(t,s) =~ r e'>.(t-s)_Q_R(-\,A(s))d,\ = (!!_ + _Q_) e(t-s)A(s)

21f2 lr as at as
for 0::::; s < t::::; T. Then P(t, s) has the same property as P(t, s). Therefore
for the operator P defined by

(P f)(t) = 1t P(t, s)f(s)ds

the assertions of Propositions 6.3, 6.4 hold. Similarly for the operator T
defined by
(Tf)(t) = 1t e(t-s)A(s) f(s)ds
we have the assertions of Proposition 6.5.
236 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Theorem 6. 2 Suppose that the assumptions (Pl),(P2) are satisfied. If u

is a classical solution of (6.1),(6.2) for x E D(A(O)) and f E C([O, T]; X),
then we have
u = (1- P}- 1 (e·A(O)x + Tf). (6.43)
Hence a classical solution is uniquely determined by x and f.
Proof. Let u be a classical solution of (6.1),(6.2). For a fixed t E (0, T] set
v(s) = e<t-s)A(s)u(s) for 0:::; s:::; t. Then
v'(s) = e<t-s)A(s)u'(s)- e<t-s)A(s) A(s)u(s) + P(t, s)u(s)
= e<t-s)A(s) J(s) + P(t, s)u(s).

Integrating both sides from 0 to t - E, 0< E < t, we obtain

v(t- E)- v(O) = 1t-< e<t-s)A(s) f(s)ds + 1t-< P(t, s)u(s)ds. (6.44)

Writing
v(t- E)= e<A(t-<)u(t- E)= e<A(t-<)(u(t- E)- u(t))
+(e<A(t-<)- e<A(t))u(t) + e<A(t)u(t),

we investigate the behavior of each term of the right hand side as E---+ 0. In
view of Lemma 6.1 the first term goes to 0. As in the proof of Proposition
6.2 (ii) we have
lle<A(t-<) - e<A(t) II :::; CEP.

Hence the second term also goes to 0. Since u(t) E D(A(t)) the third term
tends to u(t) by Theorem 1.7. Hence v(t- E) ---+ u(t). Letting E---+ 0 in
(6.44) and noting v(O) = etA(D)x we obtain

u(t) = etA(O)x + (Tf)(t) + (Pu)(t),

from which (6.43) follows.
Remark 6. 2 If we put the initial condition at t =s E (0, T) instead of
t = 0, then the classical solution is expressed as

u(t) = e<t-s)A(t)x + 1t e<t-r)A(t) [(I+ P 8 )-

1 (!- P(·, s)x)] (T)dT, (6.45)

where
6.3 CLASSICAL SOLUTIONS 237

By virtue of (6.26) or (6.27)

((1 + P 8 )-l f) (t) = f(t) + 1t R(t, T)f(T)dT.

Hence
[(1 + P 8 ) - 1 (f- P(·, s)x)] (t)

= f(t)- P(t, s)x + l t R(t, T)(j(T)- P(T, s)x)dT

= f(t) + l t R(t, T)f(T)dT + R(t, s)x.

Substituting this in (6.45) we obtain
u(t) = e(t-s)A(t)x
+ l t e(t-r)A(t) [f(T) + 1 7
R(T, J)f(a")dJ + R(T, s)x] dT

= e(t-s)A(t)x + 1t e(t-r)A(t) R(T, s)xdT

+ l t e(t-r)A(t) f(T)dT + l t 1t e(t-r)A(t) R(T, J)dT f(J)dJ

= U(t, s)x + 1t U(t, T)f(T)dT,

where
U(t, s) = e(t-s)A(t) + 1t e(t-r)A(t) R(T, s)dT
is the evolution operator or fundamental solution constructed by T. Kato
and H. Tanabe [93) in case where A(t) is densely defined. If f(t) 0, then =
following the proof of Theorem 6.1
u'(t) = A(t)e(t-s)A(t)x
+ l t A(t)e(t-r)A(t)(R(T, s)x- R(t, s)x)dT + (e(t-s)A(t)- 1)R(t, s)x.
Hence we can show following the argument of [93) that for x E D(A(s))

II :t U(t, s)xll = IIA(t)U(t, s)xll ::; t ~ )xll·

By Remark 1.1 if X is reflexive, then A(t) is densely defined. Therefore
Theorems 6.1 and 6.2 reduce to Theorems 3.1 and 4.2 of [93) except the
precise Holder continuity of the derivative of the solution.
238 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

6.4 Preliminaries(2)
In this section we establish some preliminary results which will be needed in
the next section to show that the classical solution constructed in Theorem
6.1 is a strict solution under some additional conditions on the data.
Proposition 6. 6 Suppopse that the assumptions (P1),(P2) are satisfied.
Let x E D(A(O)). Then
(i) P(·, O)x E £ 00 (0, T; X) and as t----; 0

P(t,O)x= (1-etA(o))· dd A(t)- 1 1 A(O)x+o(1);

t t=O
(ii) As t----; 0

etA(t) - 1
---x =
etA(O) - 1 d
x- (etA(o)- 1) · -A(t)- 1
I
A(O)x + o(1);
t t dt t=O
(iii) As t ----; 0

A(t)etA(t)x = etA(o) A(O)x- tA(O)etA(O) · ~A(t)- 1 1 A(O)x + o(1);

dt t=O
(iv) If (P3) is also satisfied, then as t - T ----; +O

P(t, O)x- P(T, O)x

=- [etA(r)- erA(r)] · ~A(t)- 1 ~ A(O)x + 0 ((t- T)P!\a).
dt t=O
Proof. (i) Differentiating both sides of

with respect tot we get

:tR(>.., A(t)) · A(t)- 1

1 f) 1 d -1 d . -1
= >:fJtR(A,A(t))+>:dtA(t) -R()..,A(t))dtA{t) . (6.46)

Multiplying both sides of this equality by e>.t and integrating along r we

obtain
6.4 PRELIMINARIES(2) 239

Consequently

:i t e~t
P(t, O)x = P(t, O)(A(o)- 1 - A(t)- 1)A(O)x + P(t, O)A(t)- 1A(O)x

= P(t, O)(A(o)- 1 - A(t)- 1)A(O)x + 2 :tR()..,A(t))A(O)xd)..

+(1- etA(t))~A(t)- 1 A(O)x

dt .
Combining this with (6.4),(6.5),(6.9) and
llesA(t)- esA(o)ll:::; Ctsp-1 (6.47)
we obtain the desired result.
(ii) By a standard calculation
etA(t) - 1
---x =
etA(o) - 1 d
x- (etA(o) -1) · -A(t)- 1 A(O)x
I
t t & ~0

+ etA(t:- 1 [A(0)- 1 - A(t)- 1 + t · !A(t)-t=J A(O)x

etA(t) 1 etA(O) 1 ]
+[ t - A(t)- 1 - t - A(o)- 1 A(O)x

-(etA(t)- etA(O)) · ~A(t)- 1 1 A(O)x. (6.48)

dt t=O
With the aid of (6.47) we get
li(etA(t) -1)A(t)-1- (etA(O) -1)A(0)-111

= lilt (esA(t) - esA(O))dsll :::; CtP+1. (6.49)

Hence the norms of the fourth and fifth terms of the right hand side of (6.48)
do not exceed CtP. Evidently the third term goes to 0 as t ---> 0. Thus the
proof of (ii) is complete.
(iii) A standard calculation shows

A(t)etA(t)x = etA(O) A(O)x- tA(O)etA(o) · ~A(t)- 1 1 A(O)x

dt t=O
+tA(t)etA(t) [A(o)-1- A(t)-1 + ~A(t)-11 ] A(O)x
t dt t=O
+(etA(t) _ etA(O))A(O)x

- (tA(t)etA(t)- tA(O)etA(O)) · ~A(t)- 1 1 A(O)x. (6.50)

dt t=O
240 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

With the aid of Lemma 6.1 we see that the third term of the right hand side
of (6.50) goes to 0 as t---+ 0. By (6.47) the fourth term is O(tP). Analogously
it is easily seen that the last term is also O(tP). Thus the proof of (iii) is
complete.
(iv) For 0 ::; T < t ::; T

P(t,O)x -P(T,O)x = 2~i t e>-.t [:tR(,\,A(t))- : 7 R(,\,A(T))] xd,\

+-
21 .
1r2
llt
r r
8
AeMda-~R(A,A(T))xd,\.
uT
(6.51)

With the aid of (6.46)

[:tR(,\,A(t))- : 7 R(,\,A(T))] A(0)- 1

= :tR(A, A(t)) · (A(o)- 1 - A(t)- 1 ) + :tR(A, A(t)). A(t)- 1

8 8
- 87 R(,\,A(T)). (A(o)- 1 - A(T)- 1 ) - 87 R(A, A(T)). A(T)- 1

= [:tR(,\,A(t))- : 7 R(,\,A(T))] (A(o)- 1 -A(t)- 1 )

+:T R(,\, A(T))(A(T)- 1 - A(t)- 1 ) + ~ [ :tR(,\, A(t))- :T R(A, A(T))]

+~ [ !A(t)- 1 - !A(T)- 1 ] - [R(A, A(t))- R(A, A(T))] d~A(T)- 1
-R(,\,A(t)) [!A(t)- 1 - !A(T)- 1 ],

: 7 R(A, A(T))A(0)- 1 = : 7 R(,\, A(T))(A(0)- 1 - A(T)- 1 )

1 8 1 d -1
+~ 87 R(A,A(T)) + ~ dTA(T)
-R(,\,A(T)) [dd A(T)- 1 - dd A(t)- 1 1 ] -R(,\,A(T))· dd A(t)- 1 1 .
T t t=O t t=O

Substituting these in (6.51) we get

t
P( t, O)x - P(T, O)x

= 2 ~i e>-.t { [ :t R(,\, A(t)) - :T R(,\, A(T))] (A(o)- 1 - A(t)- 1 )

+ : 7 R(,\, A(T))(A(T)- 1 - A(t)- 1 ) + ~ [:tR(,\, A(t))- : 7 R(,\, A(T))]

6.5 STRICT SOLUTIONS 241

+~ [! A(t)- 1 - d~ A(r)- 1] - [R(-X, A(t))- R(-X, A(r))]! A(r)- 1

-R(-X,A(t)) [!A(t)- 1 - !A(r)- 1] }A(O)xd,\

+ 2~i t 1t ,\e>.cr da{ :T R(,\, A( r) )(A(o)-1 - A(r)-1)

1 8 1 8 -1
+~ 87 R(-X, A(r)) + ~ 87 A(r)

-R(-X,A(r)) [!A(r)- 1 - !A(t)-t=J }A(O)xd,\

- [etA(r)- erA(r)] . iA(t)-11 A(O)x.

dt t=O

Using this it is easy to show that

P(t, O)x- P(r, O)x

= 0 ((t- r)P) + 0 ((t- r)a)- [etA(r)- erA(r)] · iA(t)- 11 A(O)x.

dt t=O

Proposition 6. 7 Suppose that the assumptions (P1),(P2),(P3) are satis-

fied. Iff E L 00 (0, T; X), then Pf E C 6 ([0, T]; X) for any 6 E (0, a) n (0, p].
Proof. Letting E= 0 in the proof of Proposition 6.3 (iii) we obtain

II(P J)(t)- (P J)(r)ll ~ C [(t- r)P + (t- r)a log _t_]

t- T
sup IIJ(s)ll,
O~s~T

from which the conclusion readily follows.

6.5 Strict Solutions

In this section it is shown that under some additional conditions on the ini-
tial value x and the inhomogeneous term f the classical solution constructed
in Theorem 6.1 is a strict solution. The results of this and next sections are
due toP. Acquistapace and B. Terreni [3: section 5].
Theorem 6. 3 Suppose that the assumptions (P1),(P2),(P3) are satisfied.
Let x E D(A(O) ), f E ccr ([0, T]; X), CJ E (0, 1], and suppose that

A(O)x + f(O)- ! A(t)- 1 1t=o A(O)x E D(A(O)). (6.52)

Then the function u defined by (6.40) is a unique strict solution of(6.1),(6.2).

Moreover, u E Cl+crA 6 ((0, T]; X) for each 6 E (0, a) n (0, p].
242 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Proof. In order to show that u is a strict solution it suffices to verify that

u'(O) exists and u' E C([O, T]; X). Put g = (1 + P)- 1(!- P(·, O)x). Then
by (6.40)
u(t) - x etA(t) - 1 1
----'--'-t- = t x + t(Tg)(t). (6.53)
Since
g = f- P(·, O)x- Pg, (6.54)
we have
Tg = T(f- f(O))- TP(·, O)x- TPg + Tf(O). (6.55)
Since f E ca([o, T]; X)
(T(f- f(O)))(t) = O(ta+ 1). (6.56)
In view of Proposition 6.6 (i)
P(·, O)x E L 00 (0, T; X), (6.57)
which implies f - P(·, O)x E L 00 (0, T; X). Hence by Proposition 6.4 (i)
g E L 00 (0, T; X). Therefore
(Pg)(t) = O(tP), (6.58)
and hence
(TPg)(t) = O(tP+ 1). (6.59)
By virtue of Proposition 6.6 (i)

(TP(·, O)x)(t) =lot e<t-s)A(t) P(s, O)xds

=lot (e<t-s)A(t)- e<t-s)A(O))P(s, O)xds

Jo
+ t
e<t-s)A(O) [(1-
esA(O)). ~A(t)- 1 1 A(O)x + O(sp/\a)] ds.
dt t=O
Since P(·, O)x E L (0, T; X) and by (6.47)
00

lle<t-s)A(t) - e<t-s)A(O) II :::; Ct(t- s)P-1,

we have
(TP(·, O)x)(t)
=
Jo
t (e<t-s)A(O)- etA(O)) · ~A(t)-
dt
1 1 A(O)xds + O(tp/\a+l)
t=O
= (etA(o) - 1)A(o)- 1 • ~A(t)- 1 1 A(O)x
dt t=O
-tetA(O) · ~A(t)- 1 1 A(O)x + O(tpAa+ 1). (6.60)
dt t=O
6.5 STRICT SOLUTIONS 243

Using (6.49)

(Tf(O))(t) =lot e<t-s)A(t) f(O)ds = (etA(t) -1)A(t)-1 f(O)

= (etA( D) - 1)A(o)- 1f(O) + O(tP+l ). (6.61)

From (6.53),(6.55),(6.56),(6.59),(6.60),(6.61) and Proposition 6.6 (ii) it fol-

lows that

u(t)- x = etA(o)- 1 x- (etA(O)- 1). ~A(t)-11 A(O)x

t t dt t=O
- e
tA(O) 1 d
- A(o)- 1 · -d A(t)- 1
I A(O)x
t t t=O
d
+etA(O) · -A(t)- 1
I A(O)x + e
tA(O) 1
- A(0)- 1f(O) + o(1)
dt t=O t
=
etA(o)- 1 d
x + -d A(t)- 1
I A(O)x
I
t t t=O
tA(O) 1 d tA(O) 1
- e - A(o)- 1 . -d A(t)- 1 A(O)x + e - A(o)- 1/(0) + o(1)
t t t=O t
= etA(O) - 1 A(o)- 1 ( A(O)x + J(O)- -A(t)- d 1I A(O)x)
t dt t=O
+! A(t)- 11t=o A(O)x + o(1) ---> A(O)x + f(O)

as t ---> 0 by (6.52) and Lemma 1.5. Hence we have established that u'(O)
exists and
u'(O) = A(O)x + f(O). (6.62)
Next we show that u' is continuous at t = 0. As was shown in the proof of
Theorem 6.1

u'(t) = A(t)etA(t)x + P(t, O)x + lt A(t)e<t-s)A(t)(g(s)- g(t))ds

+etA(t)g(t) +lot P(t, s)g(s)ds. (6.63)

Since g E L 00 (0, T; X), Pg E C 6 ([0, T]; X), 8 E (0, a) n (0, p] by Proposition

6. 7. It is easily seen that

IIA(t)e<t-s)A(t)- A(s)e(t-s)A(s)ll::::; C(t- s)P-1.

244 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Using these and (6.54),(6.57) and applying Proposition 6.6 (iv) we get

lt A(t)e(t-s)A(t)(g(s)- g(t))ds

= - l t A(t)e(t-s)A(t) [(Pg)(s)- (Pg)(t)] ds

+lot A(t)e(t-s)A(t)(J(s)- f(t))ds

- l t [A(t)e(t-s)A(t)- A(s)e(t-s)A(s)] [P(s, O)x- P(t, O)x]ds

- {t A(s)e(t-s)A(s) [(etA(s)- esA(s)). ~A(t)-11 A(O)x

Jo dt t=o
+0 ((t- s)PAa)] ds = O(t6 ) + O(t>.) + O(tP)
+ {t A(O)(etA(O)- e(2t-s)A(O)). ~A(t)-11 A(O)xds
Jo dt t=O
+ lt [A(s)(etA(s) _ e(2t-s)A(s)) _ A(O)(etA(O) _ e(2t-s)A(O))]

· dd A(t)- 1 1 A(O)xds + O(tpl\a). (6.64)

t t=O
Since
IIA(s)(etA(s) _ e(2t-s)A(s)) _ A(O)(etA(O) _ e(2t-s)A{O))II
$ IIA(s)etA(s) - A(O)etA(O) II + IIA(s)e(2t-s)A(s) - A(O)e(2t-s)A(O) II
$ CstP- 2 + Cs(2t- s)P- 2 ::; CstP- 2,

the norm of the fifth term of the last side of (6.64) is O(tP). Hence

lt A(t)e(t-s)A(t)(g(s)- g(t))ds = o(l)

+ [tA(O)etA(o)- etA(O)(etA(o) -1)] · ~A(t)- 1 1 A(O)x. (6.65)

dt t=O
Next, by (6.47),(6.54),(6.57),(6.58) and Proposition 6.6 (i) we have

etA(t)g(t) = etA(t)( -(Pg)(t) + f(t)- P(t, O)x)

= -etA(t)(Pg)(t) + etA(t)(J(t)- f(O)) + (etA(t)- etA(O))f(O)
+etA(o) f(O)- (etA(t) - etA(O))P(t, O)x- etA(o) P(t, O)x
6.6 MAXIMAL REGULARITY 245

= etA(O) f(O) + etA(O)(etA(O)- 1). ~A(t)-11 A(O)x

dt t=O
+O(tP"a) + O(t 17 ). (6.66)
From (6.63),(6.65),(6.66),(6.58) and Proposition 6.6 (i),(iii) it follows that

u'(t) = etA(o) [A(O)x + f(O)- ~A(t)- 1 1 A(O)x]

dt t=O

+dd A(t)- 1 1 A(O)x + o(1)----> A(O)x + f(O) = u'(O).

t t=O

6.6 Maximal Regularity

If A is a linear closed not necessarily densely defined operator which gener-
ates an analytic semigroup, then the interpolation space (D(A), Xh-6,=,
0 < (} < 1, between D(A) and X is defined as in the case of densely defined
operators:
(D(A), Xh-6,= = { u(O); t 1 - 6u(t), t 1 - 6Au(t), t 1 - 6u' (t) E L =(0, T; X)}.
It is clear that
D(A) c (D(A), Xh-6,= c D(A). (6.67)
Let A' be the restriction of A to the subspace
Z = {x E D(A); AxE D(A)}.
If x E D(A), then as 0 < >. ----> oo
II.A(.A- A)- 1 x- xll =II(.\- A)- 1 Axll:::; CIIAxii/.A----> 0. (6.68)
Since >.(>.- A)- 1 is uniformly bounded as >. ----> +oo , we see from (6.68)
that .\(.\- A)- 1 x ----> x as >.----> +oo if x E D(A). Furthermore if x E D(A),
then .\(.\- A)- 1 x E Z since
A(>.(>.- A)- 1 x) = .\2 ( . \ - A)- 1 x- .AxE D(A).
Hence D(A') = Z is dense in D(A). Therefore A' generates an analytic
C0 -semigroup in D(A) which is a Banach space with norm of X.
Lemma 6. 5 (D(A'), D(A)h-=,6 = (D(A), Xh-6,=·
Proof. It is evident that (D(A'), D(A)h-6,= C (D(A), Xh-6,=· Con-
versely suppose that x E (D(A), Xh-6,=· Then there exists a function u
with values in X such that
t 1 - 6u(t), t 1 - 6Au(t), t 1 - 6u' (t) E L= (0, oo; X), u(O) = x.
246 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

If we put w(t) = (1- tA)- 1 u(t), then w(t) E D(A') for t > 0,
llt 1 - 9w(t) II ::; Ct 1 - 9 llu(t) II, 11t 1 - 9Aw(t) II ::; Ct 1 - 9 IIAu(t) II,
11t 1 - 9w'(t)11 = t 1 - 911(1- tA)- 1 u'(t) + (1- tA)- 2 Au(t)ll
::; Ct 1 - 9 llu'(t)ll + Ct 1 - 9 IIAu(t)ll,
llw(t)- xll::; 11(1- tA)- 1 (u(t)- x)ll + 11(1- tA)- 1 x- xll
::; Cllu(t)- xll + 11(1- tA)- 1 x- xll-+ 0

as t-+ 0 since x E D(A) in view of (6.67). Hence x E (D(A'), D(A)h-9,CXJ·

In what follows we write D A(B, oo) instead of (D(A), X)t_ 9,(X). Since A'
is densely defined, the characterizations of D A ( e' 00) stated in Chapter 1
holds for A owing to Lemma 6.5:

{x EX; supr 9 lletAx- xll < oo}, (6.69)

t>O
{x E X;sup-\ 9 IIAR(A,A)xll < oo} (6.70)
>.>0
{x EX; sup I-\I 9 IIAR(-\, A)xll < oo}. (6.71)
.AEE

Now we return to the situation where the assumptions (P1),(P2) or (P1),

(P2),(P3) are satisfied.

Lemma 6. 6 Suppose that the assumptions (P1),(P2) are satisfied. If x E

D A(O) ({3, 00 ), {3 E (0, 1), then

(6.72)

for 0::; T ::; t::; T, 0 < t, where Cx is a constant depending on x.

Proof. We have

A(T)etA(r)x = A(T)etA(r)[x- etA(O)x

+(A(o)- 1 - A(T)- 1 )A(O)etA(o)x] + etA(r) A(O)etA(o)x.

Hence (6.72) follows from Lemma 6.1, (6.69) and the inequality

IIA(O)etA(O)xll ::; Cxtf3-l

which is a consequence of (6.71).

Propostion 6. 8 Suppose that the assumptions (P1),(P2) are satisfied. If

f3 E (O,p], then etA(t)x E Cf3([0,T];X) if and only ifx E DA(o)(f3,oo).
6.6 MAXIMAL REGULARITY 247

Proof. Let x E D A(o) (/3, oo) and 0 ::::; T <t ::::; T. We have

etA(t)x _ e-rA(-r)x = (etA(t) _ etA(-r))x

+ [<etA(-r) _ e-rA(-r)) _ (etA(O) _ e-rA(O))] X

+(etA(O)- e-rA(O))x. (6.73)

The first term of the right hand side of (6.73) is equal to

(6.74)

With the aid of (6.46)

~R().. A(a))x = ~R().. A(a)) [x- eaA(O)x

aa ' aa '
+(A(o)- 1 - A(a)- 1 )A(O)eaA(O)x + A(a)- 1 A(O)eaA(O)x]

= :aR().., A(a)) [x- eaA(O)x + (A(o)- 1 - A(a)- 1 )A(O)eaA(O)x]

+ [± !R()..,A(a)) + ±d~A(a)- 1 - R()..,A(a)) d~A(a)- 1 ] A(O)eaA(O)x.

Substituting this in (6. 74)
(etA(t) _ etA(-r))x

= ~ r e)o.t rt ~R().., A(a)) [x- eaA(O)x

2nz Jr 1-r aa
+(A(0)- 1 - A(a)- 1 )A(O)eaA(o)x] dad)..

+-.
1
2nz
1e>.t 1t [--R()..,A(a))
r 7" aa
1 f)
)..
+ --A(a)-
).. da
1 d 1

-R().., A( a))! A(a)- 1 ] A(O)eaA(o)xdad)... (6.75)

Similarly applying the same argument to the second term of (6. 73) which
is equal to
248 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

= l f) -1.
r -
fJr 21r2
frr (e-Xt- e.xr)R().., A(r))xd.Adr

= -21 .
11"2
fr
I'
1t .Ae.xo-
r
loro f) R()..,
-f)
r
A(r))xdrdCJd.A,

we obtain
[(etA(r) _ erA(r)) _ (etA(O) _ erA(O))]x

= -1.
211"2 r
fr
1t .Ae.Xa
I'
-R(.A,
f)
0 fJr
lr
A(r)) [x- erA(O)x

+(A(o)- 1 - A(r)- 1)A(O)erA(o)x]drdo-d.A

1
+ 2ni Jr Jr
r rt
.Ae
.Xa r [~1 fJrR().., A(r)) + ~1 drA(r)
Jo
f) d -1

-R().., A(r)) :r A(r)- 1 ] A(O)erA(o)xdrdCJd.A. (6. 76)

By (6.69) and Lemma 6.6 the norm of the first term of the right hand side
of (6. 75) is dominated by

CtP- 1 1t CJ{3d(J:::; CtP 1t (Jf3- 1dCJ:::; CtP(tf3- Tf3):::; CtP(t- T)f3,

and that of the second term by C(t- T)f3. The norm of the first term of
the right hand side of (6. 76) does not exceed

CT/3+ 11t (JP- 2 d(J:::; CTP 1t (Jf3- 1d(J:::; CTP(t- T)f3,

and the second term does not exceed

CTf31og ~ = CTf31og (1 + t ~ 7
) :::; CTf3 C~ T)f3 / f3 = C(t- T)f3 j/3.
Finally by Lemma 6.6

lletA(O)x- erA(O)xll =lilt A(O)esA(o)xdsll :::; C 1t sf3-1ds:::; C(t- T)f3.

Summing up we obtain etA(t)x E Cf3 ([0, T]; X).

Conversely suppose that etA(t)x E Cf3 ([0, T]; X). Then by Proposition
6.2 (ii) we have x E D(A(O)) and etA(t)xJt=0 = x. From this and (6.47) we
obtain
etA(O)x _ x = (etA(o) _ etA(t))x + (etA(t)x _ x) = O(tP) + O(tf3).
Hence we conclude x E D A(O) (/3, oo) with the aid of (6.69).
6.6 MAXIMAL REGULARITY 249

Remark 6. 3 The "if' part of Proposition 6.8 holds for f3 E (0, 1].

Proposition 6. 9 Suppose that the assumptions (P1),(P2),(P3) are satis-

fied. Iff E C 6 ([0, T]; X), 6 E (0, a) n (0, p], then

Qf = (1 + P)- 1 f E C 6 ([0, T]; X).

Proof. Q f = f + Rf = f- P(f + Rf) and by Proposition 6. 7 P(f + Rf) E

C 6 ([0, T]; X). Hence we conclude Qf E C 6 ([0, T]; X).
Proposition 6. 10 Suppose that the assumptions (P1),(P2),(P3) are sat-
isfied. Iff E C 6 ([0,T];X), 6 E (O,a) n (O,p] and f(O) = 0, then Tf E
C 1H([O, T]; X) and (Tf)'(O) = 0.
Proof. The assertion is obtained by letting E = 0 in (6.38),(6.39) in the
proof of Proposition 6.5 (iii), noting

t
llf(r)ll ::; Cr6' log:;:= log ( + t-T)
1 -T- ::; 61 (t-7) 0
-T- '

and using Proposition 6. 7.

Theorem 6. 4 Suppose that the assumptions (P1),(P2),(P3) are satisfied.
Let x E D(A(O)), f E C 6 ([0, T]; X), 6 E (0, a) n (0, p], and suppose that u is
a strict solution of (6.1),(6.2). Then u E C1+6 ([0, T]; X) if and only if the
following condition is satisfied:

A(O)x+f(O)- dd A(t)- 1 1 A(O)xEDA(o)(6,oo).

t t=O

Proof. Let z be the solution of

z'(t) = A(t)z(t) + A(O)x + f(O)- ! A(t)-t=o A(O)x,

z(O) = 0.

By Theorem 6.3 z is a strict solution, and is expressed as

z(t) = 1t e(t-s)A(t)h(s)ds, (6.77)

where h is the solution of

h(t) + t P(t, s)h(s)ds = A(O)x + f(O)- ddt A(t)-

Jo
11
t=o
A(O)x.
250 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

We set
w(t) = u(t) - A(t)- 1 A(O)x- z(t). (6.78)
Then w is a strict solution of

w'(t) = A(t)w(t) + j(t)- j(O)- [ !A(t)- 1 - :tA(t)-t=J A(O)x,

w(O) = 0,

and represented by
w(t) = lt e(t-s)A(t)k(s)ds,

where

k(t) + lt P(t, s)k(s)ds

= j(t) - f(O) - [ :t A(t)- 1 - ! A(t)- 1 LJ A(O)x.

In view of Proposition of 6.9 k E C 6 ([0, T]; X), k(O) = 0. Hence by Proposi-

tion 6.10 wE C1+ 6 ([0, T]; X). Therefore in view of(6. 78) u E C1+6 ([0, T]; X)
if and only if z E C1+6 ([0, T]; X). By Proposition 6.9 hE C 6 ([0, T]; X). By
(6.77) and Proposition 6.5 (iii)

! z(t) = ! (T(h- h(O)))(t) + ! (Th(O))(t)

=! (T(h- h(O)))(t) + etA(t)h(O) + (Ph(O))(t). (6.79)

By Proposition 6.10 the first term of the right hand side of (6.79) belongs
to C 6 ([0, T]; X). By virtue of Proposition 6. 7 the last term also belongs
to C 6 ([0, T]; X). Hence z E C 1+6([0, T]; X) if and only if etA(t)h(O) E
C'5 ([0, T]; X). Therefore using Proposition 6.8 we conclude that
u E C1+6 ([0, T]; X) if and only if

A(O)x+ f(O)- dd A(t)- 1 1 A(O)x = h(O) E DA(o)(o,oo).

t t=O

6. 7 An Example
Following Acquistapace and Terreni [3] we state an example satisfying the
assumptions (P1), (P2), (P3) of section 6.1.
6. 7. AN EXAMPLE 251

Let X= C((O, 1]) with norm !lull = supxE[O,l[ iu(x)i. We define the ope-
rator A(t) by
D(A(t)) = {u E d([O, 1]); u(O) = 0, a(t)u(1) + ,6(t)u'(1) = 0},
(6.80)
A(t)u = u" for u E D(A(t)).
Here a(·) and ,6(·) are real valued functions in C 1 ((0, T]) satisfying
a(t) ~ 0, ,6(t) ~ 0, a(t) + ,6(t) > 0 for t E (0, T]. (6.81)
Proposition 6. 11 We have
. { { u E C((O, 1]); u(O) = u(1) = 0} ,6(t) = 0,
(l) D(A(t)) = { u E C((O, T]); u(O) = 0} ,6(t) of 0.
In particular D(A(t)) is not dense in X.
(ii) a(A(t)) C ( -oo, OJ. If 0 < () < 1r and>. E Eo = {>. E C \ {0}; I arg >.i :::;
B}, then>. E p(A(t)) and

IIR(>., A(t))il :::; ( 1 + tan2 ~) /1>-1. (6.82)

Proof. (i) The assertion is easily shown, but for the sake of completeness we
give the proof of the latter case ,6(t) of 0. We write a, ,6 short for a(t), ,6(t)
for a fixed t. Suppose u E C((O, 1]) and u(O) = 0. Then forE> 0 there exists
a function wE C 2 ((0, T]) such that w(O) = 0 and supxE[O,l[ iu(x)- w(x)i :::;
E/2. Let {j be a positive number satisfying

{j < 1, {j ~~w(1) + w'(1)1:::; ~·

and¢> a function in d(-oo,oo) such that ¢(1) = 1,0 :::; ¢>(x) :::; 1 for
x E ( -oo, oo) and ¢>(x) = 0 for lx- 11 ~ 6. Set

v(x) = w(x) + ¢>(x)(l- x) (~w(l) + w (1)). 1

Then v E D(A(t)) and

iv(x)- w(x)i:::; ¢(x)l1- xll~w(1) + w'(1)1:::; 6 ~~w(l) + w'(1)1:::; ~·

Hence llu- vii :::; E.
(ii) Let >. E Eo, 0 < () < 1r. Then for f EX the boundary value problem

>.u- u" = J, x E (0, 1],

u(O) = 0, a(t)u(l) + ,6(t)u1(l) = 0
252 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

has the unique solution

u(x; t) = 1 1
K(x, y; t)f(y)dy, (6.83)

where taking the branch Rev>: 2:: 0

K(x, y; t)
sinh V):y a(t) sinh V3:(1- x) + Y>:f3(t) cosh V3:(1- x)
y ::; x,
_ { V): a(t) sinh V): + Y>:f3(t) cosh V):
- sinh Y>:x a(t) sinh V3:(1- y) + Y>:f3(t) cosh V3:(1- y)
V): a(t) sinh V): + Y>:f3(t) cosh V):

Clearly u(·;t) = R(>-.,A(t))f and in view of (6.83) we have

IIR(>-., A(t))/11 ::; II! II sup ( IK(x, y; t)idy. (6.84)

xE[O,ll }o
If we set p = Rev>:, a = Im v>:, then p > 0 and

IK(x,y; t)l
coshpy
-=-COS
h (1
p -X
) + IY>:if3(t)
a(t)
y ::; x,
{ IY>:i ln(t) sinh -J>: + Y>:f3(t) cosh Y>:i
::; coshpx h (1 ) a(t) + IY>:If3(t)
-~-V):=>-.-1 cos p - y ln(t) sinh V): + Y>:f3(t) cosh Y>:i

Hence

11 IK(x, y; t)idy
::; Isinh px cosh p( 1 - x) + cosh px sinh p( 1 - x) I
+ i-J>:if3(t)
a(t)
x--~--~~=-~~----~
PIY>:IIa(t) sinh Y>: + Y>:f3(t) cosh Y>:i
_ sinhp a(t) + IY>:If3(t)
- Pi-J>:iia(t) sinh V>.. + Y>:f3(t) cosh V>..i'

With the aid of a direct calculation we get

ln(t) sinh Y>: + V3:{3(t) cosh Y>:i

= Ie; {[(a+ pf3) cos a- af3 sin a]+ i[(a + pf3) sin a+ af3 cos a]}
6. 7. AN EXAMPLE 253

+ e;P { [(p,8 - a) cos CJ + CJ ,8sin CJ] + i[(a - p,8) sin CJ + CJ,8 cos CT]} I
21 e; {[(a+ p,8) cosCJ- CJ,8sinCJ] + i[(a + p,8) sinCJ + CJ,8cosCJ]}I
-I e;P {[(p,8- a) cosCJ + CJ,8 sinCJ] + i[(a- p,8) sinCJ + CJ,8 cos CT]} I

~ e~ 1~
2 [(a+ p,8)2 + CJ2 ,82] _ T [(a _
1/2
= p,8)2 + CJ2 ,82]

2 ~ (eP- e-P) [(a+ p,8) 2 + CJ 2,82] 112 2 sinhp · (a+ p,8).

Hence we have
la(t) sinh V>.. + V">..,LJ(t) cosh V">..l 2 sinhp · (a(t) + p,8(t)). (6.85)
Therefore we obtain
( K x . t d < a(t) + IV">..I,LJ(t) (6.86)
}0 I ( 'y, )I y- PIVXI(a(t) + p,8(t)).

Since Iarg VXI ~ () /2 < 1r /2, we have

()) 1/2
o< P ~ IV">.. I ~ ( 1 + tan2 2 P· (6.87)

With the aid of (6.86),(6.87) we obtain

1 1
IK(x,y;t)ldy ~ (1+tan2 ;) l~l·
Combining this and (6.84) we conclude (6.82).
Proposition 6.12 If l.\1 2 E > 0 and A- E E ~o, then R(.\, A(·)) E
C 1 ([0, T]; .C(X)) and satisfies

ll%tR(.\,A(t))ll ~ l.\~~2 . (6.88)

Proof. With the aid of a direct calculation we find

[R(.\,A(t))/](x) = sin~x
a(t) J01 sinh V>-.(1- y)f(y)dy + V}..,LJ(t) J01 cosh VX(1- y)f(y)dy
x--~-------------=~-=--~--~----------
a(t) sinh V).. + VX,LJ(t) cosh VX
1
- VX
r
Jo sinh V">..(x- y)f(y)dy.
254 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Hence we have

[ :t R(.\, A(t))f] (x) =sinh ..f>.x

X
a(t)f3'(t)- a'(t)f3(t)
(a(t) sinh..;>.+ ..fif3(t) cosh -/).) 2 o
1
1 f( ) . h 1\ d
y Sill YAY y. (6.89)

Since
Isinh ..f>.xl = (sinh 2 px + sin2 O"X) 112 ,

we have Isinh ..fix I ::; sinh p + 1 for 0 ::; x ::; 1, and

11I sinh ..f>.yldy :=:;

11
0
1
(sinhpy + 1)dy = -(coshp-
p
1) + 1.

Hence recalling (6.85) we obtain

l[%tR(.\,A(t))f] (x)l
::; (sinhp + 1) . 2 C
smh p · (a(t) + p/3(t)) 2
[~(coshp-
p
1) + 1] 11!11·

If I.\ I ~ E, .\ - E E ~o, then p is bounded away from 0, and hence so

are sinhp, a(t) + pf3(t), sinhp/ coshp. Thus, using also (6.87), we conclude
(6.88).

Proposition 6.13 If a,/3 E Cl+h([O,T)),h E (0, 1), then R(1,A(·)) E

Cl+h([O, T]; £(X)).

Proof. The assertion is an easy consequence of (6.89).

Owing to Propositions 6.11, 6.12, 6.13 the results of the previous sections
are applied to the problem

Ut(X, t) = Uxx(x, t) + f(x, t), (x, t) E [0, 1] X [0, T],

u(O, t) = 0, t E [O,T],
a(t)u(1, t) + /3(t)ux(1, t) = 0, t E [O,T],
u(x, 0) = ¢(x), X E [0, 1],

where f E C([O, 1] x [0, T]), ¢ E C([O, 1)). Concerning the condition (6.52)
of Theorem 6.3 there exists the following result.
6. 7. AN EXAMPLE 255

Proposition 6. 14 Let ¢J E D(A(O)) and f E C([O, T]; X). Then, a nec-

essary and sufficient condition in order that the condition (6.52) holds with
A(O) replaced by A(O) - 1 is that

f(O, 0) + ifl" (0) = 0 if ,8(0) f- 0,

f(O, 0) + qy" (O) = f(O, 1) + ifl" (1) + ~~] ifl' (1) = 0 if ,B(O) = 0.

Proof. We examine the condition

(A(O) -1)¢J + f(O, ·) + !R(1,A(t))L0 (A(O) -1)¢J E D(A(O)). (6.90)

In view of (6.89) and the boundary condition ifJ(O) = 0 the value of the
function of (6.90) at x = 0 is equal to ifJ"(O) + f(O, 0). If ,8(0) = 0, then the
value of the same function at x = 1 is

¢Y"(1) + f(O, 1) + a(~,~~~h 1 1 1

(qy"(y)- qy(y)) sinhydy, (6.91)

since ¢J(1) = 0. Integrating by parts twice we obtain

1 1
(qy" (y) - qy(y)) sinh ydy = ifl' (1) sinh 1.

Hence (6.91) is equal to

¢Y"(1) + f(O, 1) + ~(~j ¢Y'(1).

Consequently the result follows from Proposition 6.11 (i).
Remark 6. 4 The spaces D A(t) ( {}, oo) are characterized as follows: if 0 <
{} < 1/2
DA(t)(B, oo) = C20 ([0, 1]) n D(A(t)),
if 1/2 < {} < 1

DA(t)(B,oo) = {u E C 20 ([0, 1]);u(O) = a(t)u(1) + ,B(t)u'(1) = 0},

if{}= 1/2

DA(t)(1/2,oo)
{u E C*• 1 ([0, 1]); u(O) = 1, sup lu(x) - u(1)1/lx- 11 < oo}
xE[0,1)
{ ,B(t) > 0,
{u E C*• 1 ([0, 1]);u(O) = u(1) = 0}, ,B(t) = 0,
256 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

where C*• 1 ([0, 1]) is the Zygmund class of functions defined by

C*• 1 ([0, 1])

_ {
- uE
C([O 1])·
, , sup
lu(x) + u(y)- I + y)/2)1 < oo } .
I 2u((x
x,yE[0,1j,x#y X - Y

The proofs are found in P. Acquistapace and B. Terreni [4],[5], G. Da Prato

and P. Grisvard [39] and A. Lunardi [101]. See alsoP. Acquistapace and B.
Terreni [7].

6.8 Equations with Coefficients Holder

Continuous in t
In the papers [1],[6],[8] P. Acquistapace and B. Terreni proved under very
general hypotheses the existence and uniqueness of the solutions of (6.1),
(6.2). It is also shown that the hypothses are independent of those of section
6.1. One of the hypotheses reads
(P4) There exist a positive constant L, a positive integer k and real numbers
a1, ... , ak, {31, ... , f3k with 0:::; !3i < ai :::; 2 such that

k
IIA(t)R(,\, A(t))(A(t)- 1 - A(s)- 1 )11:::; L l:<t- s)a'l-\l~'•- 1 (6.92)
i=1

for ,\ E ~ = {A; Iarg -\I < Oo} \ {0} and 0:::; s < t:::; T. Set

{j =min{. min (ai- !3i), 1}. (6.93)

~=1, ... ,k

In [1],[6] it is shown that the evolution operator U(t, s) of (6.1),(6.2) is given

by

U(t, s) = e<t-s)A(s) + 1t Z(T, s)dT, (6.94)

Z(t, s) = A(t)e<t-s)A(t) - A(s)e(t-s)A(s)

+ 1t R(t, T)(A(T)e(r-s)A(r)- A(s)e(r-s)A(s))dT

+ 1t (R(t, T)- R(t, s))A(s)e<r-s)A(s)dT

+R(t, s)(e<t-s)A(s)- 1), (6.95)
6.8. COEFFICIENTS HOLDER CONTINUOUS 257

where R(t, s) is the solution of the integral equation

R(t, s) = Q(t, s) + 1t R(t, T)Q(T, s)dT

= Q(t, s) + 1t Q(t, T)R(T, s)dT (6.96)

and Q(t, s) is the operator valued function

Q(t, s) = A(t) 2 e<t-s)A(t) (A(t)- 1 - A(s)- 1 ). (6.97)

By a formal calulation we explain a general idea which suggests that U (t, s)

given above is an evolution operator of (6.1),(6.2).
Let u be a strict solution of

du(t)jdt = A(t)u(t), t E [0, T].

For a fixed 0 :s; s < t :s; T set
v(a) = e<t-a)A(t)u(a), a E [s, t].

Then
v'(a) = e<t-a)A(t) A(a)u(a)- A(t)e<t-a)A(tlu(a)
= A(t)e(t-a)A(t)(A(t)- 1 - A(a)- 1 )A(a)u(a). (6.98)

Integration from s to t yields

u(t) = e<t-s)A(tlu(s) + 1t A(t)e<t-a)A(tl(A(t)- 1 - A(a)- 1 )A(a)u(a)da.

Letting A(t) operate on both sides

A(t)u(t) = A(t)e<t-s)A(t)u(s) + 1t Q(t, a)A(a)u(a)da. (6.99)

From (6.96),(6.99) we conclude

1t Q(t, T)A(T)u(T)dT = 1t R(t, T)A(T)eC-r-s)A(-rlu(s)dT.

Substituting this in the right hand side of (6.99) and rewriting the result
appropriately

A(t)u(t) = A(t)e(t-s)A(t)u(s) + 1t R(t, T)A(T)e(-r-s)A(-r)u(s)dT

258 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

= A(s)e(t-s)A(s)u(s) + (A(t)e(t-s)A(t) - A(s)e(t-s)A(s))u(s)

+ 1t R(t, r)(A(r)e(r-s)A(r) - A(s)e(r-s)A(s))u(s)dr

+ 1t (R(t, r)- R(t, s))A(s)e(r-s)A(s)u(s)dr

+R(t, s)(e(t-s)A(s)- 1)u(s)
= A(s)e(t-s)A(s)u(s) + Z(t, s)u(s).

Integrating the above from s to t we conclude

u(t) = e<t-s)A(s)u(s) + 1t Z(r, s)u(s)dr = U(t, s)u(s). (6.100)

Thus one might expect that U(t, s) defined by (6.94)-(6.97) would be a

desired evolution operator.

6.9 Preliminary Lemmas and Remarks

Lemma 6. 7 For 0 ::::; s < t ::::; T, A E E
k
IIR(A, A(t))- R(A, A(s))ll ::::; C L(t- s)ai IAIJ]i- 1• (6.101)
i=1

Proof. This is a simple consequence of (6.92) and

R(A, A(t))- R(_\, A(s))
= A(t)R(_\, A(t))(A(s)- 1 - A(t)- 1 )A(s)R(A, A(s)). (6.102)
Lemma 6. 8 For a nonnegative integer m, a positive number rJ and 0 ::::;
s < t::::; T,

IIA(t)meaA(t)- A(s)meaA(s)"::::; ct (~::~~i' (6.103)

ct ~tm~;:::.
i=1

IIA(t)meaA(t)- A(s)meaA(s)llc(D(A(s)),X)::::; (6.104)

i=1
Proof. These are simple consequences of
A(t)meaA(t) _ A(s)meaA(s)

= ~ f _\me>.a(R(A, A(t))- R(A, A(s)))d_\ (6.105)

2?r~ Jr
and Lemma 6.7.
6.9. PRELIMINARY LEMMAS AND REMARK 259

Corollary 6. 1 For 0 $ s <t$ T

k
!IA(t)e(t-s)A(t)- A(s)e(t-s)A(s)ll $ C2)t- s)"'i-.Bi-1 $ C(t- s)6-1.
i=1
(6.106)

Recall that 8 is the number defined by (6.93).

Lemma 6. 9 (i) For 0$ s < t $ T

k
IIQ(t, s)!l $ C2:(t- s)<>i-.Bi- 1 :::; C(t- s)6-1. (6.107)
i=1

(ii) For 0 $ s < T <t $ T and 'fJ E (0, 8)

!IQ(t, s)- Q(t, T)ll $ C11 (T- s)11(t- 7) 6 -11- 1 . (6.108)

Proof. (i) (6.107) readily follows from (6.92) and

Q(t,s) = ~ { .\e>.(t-s)A(t)R(A,A(t)) (A(t)- 1 - A(s)- 1) d.\.

2n~ Jr

(ii) Using

Q(t, s) - Q(t, T)
= (A(t)2e(t-s)A(t) _ A(T)2e(t-s)A(T))(A(T)-1 _ A(s)-1)
+A(T) 2e(t-s)A(.,-)(A(T)- 1 - A(s)- 1 )
+(A(t)2e(t-s)A(t) _ A(t)2e(t-T)A(t)) (A(t)-1 _ A(T)-1)
=h +lz +h,
where

h = ~ { .\2e>-Ct-s) A(t)R(.\, A(t))

2n~ Jr
x(A(T)- 1 - A(t)- 1)A(T)R(.\,A(T))(A(T)- 1 - A(s)- 1)d,\,
lz = -2
1
n~ Jr
0 rAe>.(t-s) A(T)R(.\, A(T))(A(T)- 1 - A(s)- 1 )d.\,

h = ~ { .\(e>.(t-s)- e>-(t-T))A(t)R(.\, A(t))(A(t)- 1 - A(T)- 1)d.\

2n~ Jr

= -21 ·1t-s { .\2e>.a A(t)R(.\, A(t))(A(t)- 1 - A(T)- 1)d.\do-,

n~ t-T Jr
260 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

we get

IIQ(t, s)- Q(t, T)ll

k (T-s)a; [ k (t-T)a'] k lt-s(t-T)a;
-<C~
&;t (t- s)f3•+ 1 1+~
&;t (t- s)f3• +C~
&;t t-r O"f3•+2
dO".

Noting

(t- T)a; < (t- s)a;-{3; (T- s)a; (T- s)a;-{3; C(T- s)li
--'----'c-,,- ' , :::; :::; '
(t-s)~"'- (t-s)~"•+ 1 t-s t-s
t-s (t _ T)a; lt-s dO"
l -'--;:-:-'::----- dO" < (t - T) ai - {3; -
-r (]"{3;+ 2 - t-r 0" 2
= (t- T)a;-{3;- 1 (T- s) < C(t- T) 6 - 1 (T- s)
t-s - t-s '
we can easily show (6.108).
Lemma 6. 10 (i) For 0 :::; s < t :::; T

IIR(t,s)ll:::; C(t-s) 6 - 1• (6.109)

(ii) For 0:::; s < T < t:::; T and 'TJ E (0, 8)

IIR(t, s)- R(t, T)ll:::; c,(T- s) 17 (t- T)li-17- 1 . (6.110)
Proof. (i) (6.109) is easily obtained by estimating each term of the successive
approximation series for the solution of the integral equation (6.96).
(ii) This is a simple consequence of (6.96) and (6.108).
It was pointed out in P. Acquistapace and B. Terreni [8] that the sets
of the assumptions (P1),(P2),(P3) and (P1),(P2),(P4) are independent of
each other. In order to show that it suffices to verify that (P1),(P2),(P3)
do not imply (P4).
Lemma 6. 11 Suppose in addition to the assumptions (P1),(P2),(P4) that
A(-)- 1 E C 1 ([0, T]; £(X)). Then we haveR ( !A(t)- 1 ) C D(A(t)) for
t E [O,T].
Proof. For u E X and 0 :::; t < t + h :::; T we have
6.9. PRELIMINARY LEMMAS AND REMARK 261

In view of Lemma 6.7 and the assumption (P4)

k k
III1II ::; C 2::: ha;-f\ III411 ::; 2::: ha;-jj; ·
i=1 i=1

Clearly I2 ----+ 0 as h ----+ 0. Hence

lim h E D(A(t)).
ddt A(t)- u = h-+0
1

In the example of section 6.7 let a(t) = 1, ,6(t) = t. Then

[A(t)- 1 f](x) = 1x (x- y)f(y)dy- [1 1

f(y)dy- 1 : t 1 1
yf(y)dy] x.

Therefore
(6.111)

Suppose that the assumption (P4) holds. Then by Lemma 6.11 we have

R (! A(t)- 1 1t=J c D(A(O)) = {u E C([O, T]); u(O) = u(1) = 0}.

However, if we take f(x) =1 in (6.111), we have
[ -A(t)
d
dt
-11 t=O
]
f (x) = --,
X
2

and the function does not belong to D(A(O)), which is a contradiction.

Therefore (P4) does not hold.
262 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

6.10 Representation of Solutions by

Fundamental Solution
Lemmas 6.7, 6.8, 6.9, 6.10 show that all the integrals in (6.94),(6.95),(6.96)
are convergent. Hence the operator valued function U(t, s) is well defined
by these equalities and (6.97), and we have

IIU(t, s)ll ~ C, II :t U(t, s)JI ~ t ~ 8 , IIZ(t, s)ll ~ C(t- s) 6 - 1 • (6.112)

For r E [0, T) and g E C([r, T]; X) set

(Qrg)(t) = 1t Q(t, s)g(s)ds. (6.113)

By virtue of Lemma 6.9 (i) Qr E .C(C([r, T]; X)) and the inverse of I- Qr
is given by

((I- Qr)- 1 f)(t) = j(t) + 1t R(t, T)j(T)dT. (6.114)

If r = 0, we write Qr = Q. Using (6.114) Z(t, s) is also given by

Z(t, s) = [(1- Q 8 )-
1 (A(·)e<·-s)A(·)- A(s)e<·-s)A(s))] (t)

+ lt (R(t, T) - R(t, s))A(s)e<r-s)A(s)dT

+R(t, s)(e(t-s)A(s)- 1). (6.115)

Lemma 6. 12 Let f E C([O, T]; X) and x E D(A(O)). If a strict solution

u of (6.1),(6.2) exists, then A(O)x + f(O) E D(A(O)).
Proof. By a direct calculation we have
6.10. REPRESENTATION OF SOLUTIONS 263

With the aid of Lemma 6. 7 we get

k
llhii::SCI:t<>•-il• sup llu(t)-xll·
i=l tE[O,Tj t

and by (6. 92)

k
IIIsll:::; Lt<>•-il•IIA(O)xll·
i=1
Clearly lz ---+ 0, !4 ---+ 0 as t ---+ 0. Hence

A(O)x + f(O) = u'(O) =lim u(t)- x

t->0 t

=lim
t->0
~R
t
(~,t A(o)) (A(O)x + f(O)) E D(A(O)).

The following theorem is due toP. Acquistapace and B. Terreni [6].

Theorem 6. 5 (i) Let f E C([O, T]; X), x E D(A(O)) and suppose that
A(O)x + f(O) E D(A(O)). If u is a strict solution of (6.1),(6.2), then u is
given by
u(t) = U(t, O)x +it U(t, s)f(s)ds, t E [0, T]. (6.116)

(ii) Let f E C([O, T]; X) and x E D(A(O)). If u is a classical solution of the

problem (6.1),(6.2), then (6.116) holds fortE (0, T].
Consequently a classical solution, and hence a strict solution also, is
uniquely determined by the inhomogeneous term f and the initial value x.
Proof. (i) Let u be a strict solution of (6.1),(6.2). For 0 < t:::; T set

v(s) = e(t-s)A(t)u(s), s E [0, t].

Differentiating both sides we get this time instead of (6.98)

v'(s) = A(t)e(t-s)A(t)(A(t)- 1 - A(s)- 1 )A(s)u(s) + e(t-s)A(t) f(s). (6.117)

Let 0 < f < 1. Integrating both sides of (6.117) from 0 tot- Et we get

v(t- Et) - v(O)

= Jo
t-d A(t)e(t-s)A(t)(A(t)- 1 - A(s)- 1 )A(s)u(s)ds

+ Jot-d e(t-s)A(t) f(s)ds. (6.118)

264 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Letting A(t) operate on both sides of (6.118) and using the definition of v(·)
we get
A(t)e<tA(t)u(t- Et) - A(t)etA(t)x

= lt-<t Q(t, s)A(s)u(s)ds + lt-<t A(t)e<t-s)A(t) f(s)ds.

We rewrite this as follows:

A(t)u(t)- lt Q(t, s)A(s)u(s)ds = G,(t), (6.119)

where

G,(t) = -l~t< Q(t, s)A(s)u(s)ds + A(t)u(t)

rt-t<
-A(t)e<tA(t)u(t- tE) + A(t)etA(t)x + A(t) Jo e<t-s)A(t) f(s)ds

=- t Q(t,s)A(s)u(s)ds
lt-<t
-(e<tA(t) -1- EtA(t)e<tA(t))(A(t)u(t) + f(t)) + (e<tA(t)- 1)f(t)
-A(t)e<tA(t)(u(t- Et)- u(t) + Etu'(t)) + (A(t)etA(t)- A(O)etA(O))x

+A(O)etA(O)x + Jo
t-<t (A(t)e<t-s)A(t) - A(s)e(t-s)A(s))f(s)ds

rt-<t
+ Jo A(s)e<t-s)A(s) f(s)ds = t-;Ii.
8

In view of (6.113) with r = 0 we obtain from (6.119)

A(t)u(t) = [(I- Q)- 1 G,) (t),
and hence
u'(t) = [(I- Q)- 1 G,) (t) + f(t).
Integrating this from Et to t

u(t) = u(Et) +it f(s)ds +it [(I- Q)- 1G,) (T)dT. (6.120)
<t <t

Using (6.114),(6.109),(6.107) we get

l IIt [(1- Q)- 1(I1 + I2 + I4)] (T)ildT:::;

t
c t
Jo
III1(T) + I2(T) + I4(T)IIdT,
6.10. REPRESENTATION OF SOLUTIONS 265

III1(T)II ~ r
h-ff
IIQ(T, s)A(s)u(s)llds ~ C(a) 6 sup IIA(s)u(s)ll,
~~~
III2(T)II ~ CIIA(T)u(T) + f(T)II·
By the proof of Lemma 6.12 we see that A(t)u(t) + f(t) E D(A(t)). Using
this and Lemma 1.5 (ii) we see that llh(T)II---+ 0 as E---+ 0. Noting

u(t- Et)- u(t) + Etu'(t) = Et 1 1

(u'(t)- u'(t- wt))dn,

we obtain

III4(T)II = IJEtA(t)e<tA(t) 1 1
(u'(t)- u'(t- wt))dcrll

~C 1 1
llu'(t)- u'(t- wt)lldcr---+ 0

as E---+ 0, and
III4(T)II ~ C sup llu'(s)ll.
sE[O,T[

Therefore by Lebesque's dominated convergence theorem we find that

(6.121)

Since lllsll ~ Ct6 - 1llxll by Corollary 6.1, we have

lim
<--->0
lt [(I-
<t
Q)- 1 Is) (T)dT

=lot [u- Q)- 1(A(·)e·A(·)- A(O)e·A(o))x] (T)dT. (6.122)

Similarly

lim
<---> 0
lt [(I-
<t
Q)- 1 h] (T)dT =
Jo
t [(I- Q)- 1

(1. (A(·)e(·-s)A(·)- A(s)e(·-s)A(s))f(s)ds)] (T)dT. (6.123)

Using (6.114) in the right hand side of (6.123) we find that

lim
<--->0
lt [(I-
<t
Q)- 117 ) (T)dT
266 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

=lot [loT(A(r)e<T-s)A(T)- A(s)e(T-s)A(s))f(s)ds

+loT R(r, a) loa (A(a)e(a-s)A(a)- A(s)e<a-s)A(s))f(s)dsda] dr

=lot loT (A(r)e(T-s)A(T)- A(s)e(T-s)A(s))f(s)dsdr

+lot loT 1T R(r, a)(A(a)e(a-s)A(a) - A(s)e(a-s)A(s))da f(s)dsdr

=lot 1t (A(r)e(T-s)A(T)- A(s)e<T-s)A(s))dr f(s)ds

+it 1t 1T R(r,a)(A(a)e<a-s)A(a) -A(s)e(a-s)A(s))dadrf(s)ds

=lot 1t [(I- Q 8 )-
1 (A(·)e<·-s)A(·)- A(s)e<·-s)A(s))] (r)drf(s)ds.

(6.124)

Since

1[(I-t
tt
Q)- 1 16] (r)dr

= i: [A(O)eTA(O)x +loT R(r, s)A(O)esA(O)xds] dr

= (etA(O)- e<tA(O))x +it

<t
r
Jo
(R(r, s)- R(r, O))A(O)e 8 A(O)xdsdr

+it R(r, O)(eTA(O)- 1)xdr,

<t

we see

limit [(I- Q)- 116 ] (r)dr

<--->0 <t

= (etA(O)- 1)x +it loT (R(r, s)- R(r, O))A(O)esA(O)xdsdr

+it R(r,O)(eTA(O) -1)xdr. (6.125)

Finally as for h, Is we have

6.10. REPRESENTATION OF SOLUTIONS 267

= i: {(I- Q)- 1 [<e<·A(·) -l)f +~a·-<· A(s)e<·-s)A(s) f(s)ds]} (r)dr

= 1: [u- Q)- 1 (e'·A(·) -l)f) (r)dr

+ t r-<7' A(s)e<r-s)A(s) f(s)dsdT

J<t Jo
+it r R(r, a) r-<0' A(s)e(a-s)A(s) f(s)dsdadT
<t Jo Jo
=it
<t
[u- Q)- 1 (e<·A(·) -l)f) (r)dr

+ rt r-<7' A(s)e<r-s)A(s) f(s)dsdT

J,t Jo
+it(" r-<a(R(r,a) -R(r,s))A(s)e(a-s)A(s)f(s)dsdadr
<t Jo Jo
+it r r-<0' R(r, s)A(s)e<a-s)A(s) f(s)dsdadT
<t Jo Jo

= i: [u- Q)- 1 (e'·A(·) -l)f) (r)dr

+ 1 (1-<)<t it
<t
A(s)e<r-s)A(s)dT f(s)ds

+ it-<t 1t A(s)e<r-s)A(s)dT f(s)ds

(1-<)<t s/(1-<)

+it r r-<0' (R(r, a) - R( T, s))A(s)e(a-s)A(s) f(s)dsdadr

a Jo Jo
+ t r-<7' R(r, s) 17'
J<t Jo s/(1-<)
A(s)e(a-s)A(s)da f(s)dsdT

= i: [u- Q)- 1 (e'·A(·) -l)f) (r)dr

{(1-<)<t
+ Jo (e<t-s)A(s)- e<<t-s)A(s))f(s)ds
t-<t
+i (e(t-s)A(s) _ e(s/(1-<)-s)A(s))J(s)ds
(1-<)<t

+it (" r-<a(R(r,a) -R(r,s))A(s)e<a-s)A(s)f(s)dsdadr

<t Jo Jo
268 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

+ 1t r-er R(T, s)(e<r-s)A(s)- e<s/(1-<)-s)A(s))f(s)dsdT

<t Jo
= i: [(I- Q)-1(e<·A(·) f- e(<·/(1-<))A(·) n] (T)dT
+ 1t e<<r/( 1-<))A(r) f(T)dT + 1t r R(T, s)e(<s/( 1-<))A(s) J(s)dsdT
<th
1
d
t 1(1-<)<t
- [(I- Q)- 1f] (T)dT + (e<t-s)A(s)- e<<t-s)A(s))J(s)ds
t 0
t-<t lt-<t
+l e<t-s)A(s) f(s)ds- e(<s/(1-<))A(s) J(s)ds
(1-<)<t (1-<)<t

+1t r r-<a (R(T,J)- R(T, s))A(s)e(a-s)A(s) J(s)dsdJdT

<t Jo Jo
+it r-<r R(T, s)e<r-s)A(s) f(s)dsdT
<t Jo
-1: lr-<r R(T, s)e(<s/(1-<))A(s) J(s)dsdT

=it [(I- Q)-1(e<·A(·) f- e(<·/(1-<))A(·) n] (T)dT

<t
+ [1t e(<s/(1-<))A(s) f(s)ds -Jt-<t e<<s/(1-<))A(s) J(s)ds]
<t (1-<)<t

+ 1t lr R(T, s)e(<s/( 1-<))A(s) f(s)dsdT -1t [(I- Q)- 1 f] (T)dT

r-er

1
d <t
(1-<)<t lt-<t
+ (e(t-s)A(s) - e(<t-s)A(s))J(s)ds + e<t-s)A(s) J(s)ds
(1-<)d

+it r r-<a (R(T, J)- R(T, s))A(s)e(a-s)A(s) f(s)dsdJdT

<t Jo Jo
+ 1 tlr-<r R(T, s)e(r-s)A(s) f(s)dsdT. (6.126)
t 0

Since
6.10. REPRESENTATION OF SOLUTIONS 269

we see that the first term of the last side of (6.126) goes to 0 as E --> 0. The
second term also tends to 0 since it is equal to

( {t - rt ) e(<s/(1-<))A(s) f(s)ds.
lt-<t 1(1-<)<t

It is easily seen that the third and the fifth terms also go to 0. Hence we
obtain

lim i t [(I- Q)- 1 (h +Is)] (T)dT

<--->0 <t

=-!at [(I- Q)- 1 f] (T)dT +!at e<t-s)A(s) f(s)ds

+!at lor 1" (R(T, a)- R(T, s))A(s)e<a-s)A(s) f(s)dsdndT

+!at lor R(T, s)e(r-s)A(s) f(s)dsdT

= -!at f(s)ds- !at lor R(T, s)f(s)dsdT +!at e<t-s)A(s) f(s)ds

+!at lor 1" (R(T, a)- R(T, s))A(s)e(a-s)A(s) f(s)dsdndT

+!at lor R(T, s)e(r-s)A(s) f(s)dsdT

=!at e<t-s)A(s) f(s)ds -!at f(s)ds

+!at 1t R(T, s)(e<r-s)A(s)- 1)dT f(s)ds

+!at 1t 1r (R(T, a) - R(T, s))A(s)e<a-s)A(s)dadT f(s)ds. (6.127)

From (6.120),(6.121),(6.122),(6.124),(6.125),(6.127) and noting (6.115) we

obtain (6.116).
(ii) If u is a classical solution, then it is a strict solution in [E, T] for any
E E (0, T). Since A(E)u(E) + j(E) E D(A(E)) in view of Lemma 6.12, we have
applying (i) in [E, T]

u(t) = U(t, E)u(E) +it U(t, s)f(s)ds, t E [E, T].

Letting E--> 0 for a fixed t E (0, T] we obtain (6.116).

270 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

6.11 Approximate Equations

We consider a sequence of equations

d
dt Un(t) = An(t)Un(t) + f(t), t E [0, T] (6.128)
Un(O) =X, (6.129)

and n = 1, 2, ... , where

An(t) = nA(t)R(n, A(t)) = A(t)(1- n- 1 A(t))- 1

is the Yosida approximation of A(t).

Lemma 6. 13 For A. E E and positive integers n we have

1 3 1 } (6.130)
lA. + nl :::; mm (1 + IA.I) sinllo' nsinllo ·
0 {

Proof. The assertion is easily shown if we note that ReA. > lA. Icos Bo for
A.E E.

Lemma 6. 14 For A E E, positive integers n and 0:::; s < t:::; T we have

(i) p(An(t))::) E and

R(A., An(t)) = A.~ n (n- A(t))R (A.~n, A(t))

=(A.: nY R (A~n'A(t)) + A.~n' (6.131)

(ii)
c
IIR(A., An(t))ll :::; 1 + lA. I, (6.132)

(iii)

R(A, An(t))- R(A., A(t))

= A.~nA(t)R (A.~n,A(t)) A(t)R(A,A(t)), (6.133)

(iv)
n
An(t)R(\An(t))= A.+nA(t)R ( A.+n'A(t)
A.n ) , (6.134)
6.11. APPROXIMATE EQUATIONS 271

(v)

R(\An(t)) -R(A.,An(s)) =- (A.:nr

xA(t)R (A.-:n' A(t)) (A(t)- 1 - A(s)- 1 )A(s)R (A.-:n, A(s)).

(6.135)
Proof. (i) As is easily seen we have

A.- An(t) =(A.+ n) (A.-:n - A(t)) (n- A(t))- 1 • (6.136)

Hence p( An (t)) ::) ~ follows if we note that 0 # ,\ E ~ if and only if ,\- 1 E ~

and , A.n E ~if and only if,\ - 1 +n- 1 E ~. (6.131) is a simple consequence
"'+n
of (6.136).
(ii) This is a simple consequence of (i) and Lemma 6.13.
(iii),(iv) can be shown straightforwardly using (i), and (v) is established
with the aid of (i) and (6.102).
Lemma 6. 15 ForT > 0, t E [0, T], m = 0, 1, and positive integers n
(i)
(6.137)
(ii) for 0 ~ ry ~ 1

IIAn(t)merAn(t)- A(t)merA(t)ll ~ n77T~+71' (6.138)

ct
(iii)

IIAn(t)erAn(t)- An(s)erAn(s)ll ~ (tT::r·i. (6.139)

i=1

Proof. (i) (6.137) is a simple consequence of Lemma 6.14 (ii).

(ii) With the aid of Lemma 6.14 (iii) we have
c
IIR(A.,An(t)) -R(A.,A(t))jj:::; jA.+nj' (6.140)

The inequality (6.138) is a simple consequence of (6.140) and Lemma 6.13.

(iii) With the aid of Lemma 6.13, Lemma 6.14 (v) and the hypothesis (P4)
we get
k
IIR(A., An(t))- R(A.,An(s))ii:::; ci::(t- s)"iiA.I{j;-1, (6.141)
i=1
272 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

from which (6.139) follows easily.

Lemma 6. 16 ForT> 0, m = 1, 2, 0::::; s < t::::; T and positive integers n

(i)

IIAn(t)meTAn(t)(A(t)-1- A(s)-1)11::::; ct ~~~::~:'

i=1
(6.142)

(ii) for 0 ::::; rJ ::::; 1

Proof. (i) With the aid of Lemma 6.14 (iv) and the hypothesis (P4) we have

(6.144)

The result follows from (6.144), Lemma 6.13 and

An(t)merAn(t)(A(t)-1- A(s)-1)
1.
= -2
7rz
r
Jr
;.m- 1eAT An(t)R(>., An(t))(A(t)- 1 - A(s)- 1)d>..

(ii) We have by Lemma 6.14 (iii) and the assumption (P4)

II (R(>., An(t)) - R(>., A(t))) (A(t)- 1 - A(s) - 1 ) I

= II>.~ n A(t)R (). >: n, A(t)) A(t)R(>., A(t))(A(t)- 1 - A(s)- 1 )11

::::; C I>.~ n I t,(t- s)"'' 1>.1 11'- 1 . (6.145)

The inequality (6.143) follows from (6.145) and

(An(t)merAn(t)- A(t)merA(t))(A(t)-1- A(s)-1)

= -21 . { >.me-\r(R(\An(t)) -R(>.,A(t)))(A(t)- 1 -A(s)- 1)d>..

7rz Jr
6.11. APPROXIMATE EQUATIONS 273

Set for 0 ::; s <t ::; T and n = 1, 2, ...

Qn(t, s) = An(t) 2 e(t-s)A,.(t)(An(t)- 1 - An(s)- 1 ). (6.146)
Since An(t)- 1 = A(t)- 1 - 1jn, we have
Qn(t, s) = An(t) 2 e(t-s)A,.(t)(A(t)- 1 - A(s)- 1 ). (6.147)
Lemma 6. 17 (i) For 0::; s < t::; T and positive integers n
IIQn(t, s)JI ::; C(t- s) 6 - 1• (6.148)
(ii) For 0::; s < T < t::; T, 71 E (0, 6) and positive integers n

IIQn(t, s)- Qn(t, T)li ::; c,.,(T- sf1(t- T)0-11- 1 • (6.149)

Proof. (i) (6.148) is a direct consequence of (6.147) and Lemma 6.16 (i).
(ii) Following the proof of Lemma 6.9 (ii) and using Lemma 6.13 and (6.144)
we obtain (6.149).
Lemma 6. 18 For 0::; s < t::; T, positive integers n and 0::; 71::; 1

IIQn(t, s)- Q(t, s)ll ::; ~ (t- s)o-n- 1 • (6.150)

Proof. This is an easy consequence of Lemma 6.16 (ii).

Let Rn(t, s), 0::; s < t::; T, be the solution of the integral equation

Rn(t, s) = Qn(t, s) + 1t Rn(t, T)Qn(T, s)dT

= Qn(t, s) + 1t Qn(t, T)Rn(T, s)dT. (6.151)

Lemma 6. 19 (i) For 0::; s < t::; T and positive integers n

IIRn(t,s)ll::; C(t- s) 6 - 1 • (6.152)
(ii) For 0 ::; s < T < t ::; T, 71 E (0, 6) and positive integers n
IIRn(t, s)- Rn(t, T)li::; C 11 (T- s) 11 (t- T)o-n- 1. (6.153)
Proof. The results are established exactly as Lemma 6.10.
We have
Rn(t, s)- R(t, s)

= Qn(t, s)- Q(t, s) + 1t (Qn(t, T)- Q(t, T)) Rn(T, s)dT

+ 1t Q(t,T)(Rn(T,s) -R(T,s))dT. (6.154)

274 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Hence

lt R(t, a-) (Rn(o-, s)- R(o-, s)) do-= lt R(t, a-) (Qn(o-, s)- Q(o-, s)) do-

+ l t R(t, a-) la (Qn(o-, T)- Q(o-, T)) Rn(T, s)dTdo-

+ lt 1t R(t,o-)Q(o-,T)do-(Rn(T,s) -R(T,s))dT. (6.155)

Using (6.96) we obtain from (6.155)

lt Q(t, T) (Rn(T, s)- R(T, s)) dT = l t R(t, a-) (Qn(o-, s)- Q(o-, s)) do-

+ lt R(t,o-) la (Qn(o-,T) -Q(o-,T))Rn(T,s)dTdo-. (6.156)

Lemma 6. 20 For 0::; s < t ::; T, 'f/ E (0, 6) and positive integers n

IIRn(t, s)- R(t, s)ll::; ~~ (t- s) 6 -77- 1 . (6.157)

Proof. The inequality (6.157) is established with the aid of (6.154), (6.150),
(6.152),(6.156),(6.109).
We define

Un(t,s) = e(t-s)A,..(s) + lt Zn(T,s)dT, (6.158)

Zn(t, s) = An(t)e(t-s)A,..(t) - An(s)e(t-s)A,..(s)

+ lt Rn(t, T)(An(T)e(r-s)A,..(r)- An(s)e(r-s)A,..(s))dT

+ lt (Rn(t, T)- Rn(t, s)) An(s)e(r-s)A,..(s)dT

+Rn(t, s)(e(t-s)A,..(s) -1). (6.159)

Then Un(t, s) is the evolution operator of (6.128),(6.129). The solution of

(6.128),(6.129) is given by

Un(t) = Un(t, O)x +lot Un(t, s)f(s)ds. (6.160)

In view of Lemma 6.15 (iii) we get

IIAn(t)e(t-s)A,..(t) - An(s)e(t-s)A,..(s) II ::; C(t - s)6-1. (6.161)
6.12. EXISTENCE OF SOLUTIONS 275

With the aid of (6.161), Lemmas 6.15, 6.19 we obtain

IIUn(t, s)ll :::; C, IIZn(t, s)ll :::; C(t- s) 6 - 1. (6.162)

Lemma 6. 21 As n --+ oo

Un(t, s) --+ U(t, s), 0:::; s:::; t:::; T, (6.163)

Zn(t, s)--+ Z(t, s), 0:::; s < t:::; T, (6.164)

in the norm of .C(X).

Proof. In view of (6.161), Corollary 6.1 and Lemma 6.15 (ii) we have
IIAn(T)e(r-s)A,(r) _ An(s)e(r-s)A,(s) _ A(T)e(r-s)A(r)

+A(s)e(r-s)A(s)ll <min {c(T- s).5-1 C }

- ' n(T- s) 2 '

which implies
IIAn(T)e(r-s)A,(r)- An(s)e(r-s)A,(s)- A(T)e(r-s)A(r)

+A(s)e(r-s)A(s) II :::; ~ (T- s)b-'f/-1-.571 (6.165)

for 0 :::; "1 :::; 1. By Lemma 6.10 (ii), Lemma 6.19 (ii) and Lemma 6.20 we
have

IIRn(t, T)- Rn(t, s)- R(t, T) + R(t, s)ll

:::; min { ~~ (t- 7) 6-11-1, C11 (T- s)11(t- 7) 6-71- 1}

for "1 E (0, 8). Hence

IIRn(t, T)-Rn(t, s)-R(t, T)+R(t, s)ll:::; n~72 (t-T) 6 - 11 - 1 (T-s)111 2 • (6.166)

Using (6.161),(6.165),(6.166), Lemmas 6.10, 6.15, 6.20 we obtain the desired

results without difficulty.

6.12 Existence of Solutions

In this section we show the existence of classical and strict solutions of
(6.1),(6.2) under the assumptions (P1),(P2),(P4). The definitions of these
solutions is the same as Definitions 6.1 and 6.2 of section 6.1.
276 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Lemma 6. 22 For 0 ::; s < t ::; T

k
ii[A(t)e(t-s)A(t)- A(s)e(t-s)A(s)]A(s)- 1 11::; C_2_)t- sy:.;-/3;. (6.167)
i=1

Proof. From (6.102) and the assumption (P4) it follows that

k
II (R(-X, A(t)) - R(,\, A(s)))A(s) - 1 11 ::; C ,2)t- s)a' I-XI 13'- 2 • (6.168)
i=1

The inequality (6.167) follows from (6.168).

Lemma 6. 23 For 0 ::; s < t ::; T
IIZ(t, s)A(s)- 111 ::; C(t- s) 6 • (6.169)
Proof. The inequality (6.169) follows from Lemmas 6.10, 6.22 and

ll(e(t-s)A(s)- 1)A(s)-111 = lilt-s erA(s)dTII::; C(t- s).

Proposition 6.15 Assume f E ca([O, T]; X) for some a E (0, 1) and

x E D(A(O)). Set

u(t) = U(t, O)x + lt U(t, s)f(s)ds, t E (0, T]. (6.170)

Then u E C 1 ((0, T]; X) and limt---+o u(t) = x. If moreover x E D(A(O)) and

A(O)x + f(O) E D(A(O)), then

lim dd u(t) = A(O)x + f(O). (6.171)

t-+0 t
Proof. It is obvious that
d
dt U(t, O)x = A(O)etA(o)x + Z(t, O)x, (6.172)

! lt 1t Z(T, s)dT f(s)ds = lt Z(t, s)f(s)ds. (6.173)

For 0 < E < t we have

.!!:._ t-• e(t-s)A(s) f(s)ds
dt } 0
= e•A(t-<) f(t- E)+ lt-< A(s)e(t-s)A(s)(f(s)- f(t))ds

+ lt-< (A(s)e(t-s)A(s) - A(t)e(t-s)A(t))f(t)ds

-e•A(t) f(t) + etA(t) f(t).
6.12. EXISTENCE OF SOLUTIONS 277

With the aid of Lemma 6.8 we have

k
Jle<A(t-<)- e<A(t)ll::; C2>o:;-fl;.
i=l

Hence
lle<A(t-<) j(t- E)- e<A(t) j(t)ll
::; lle<A(t-<)(J(t- E)- j(t))ll + ll(e<A(t-<)- e<A(t))j(t)ll
k
::; Cllf(t- E)- f(t)ll + ci:>o:,-fi,JIJ(t)ll-> 0
i=l

as E ---> 0. Therefore we get

d it-E e(t-s)A(s) f(s)ds

lim- = i t A(s)e(t-s)A(s)(J(s)- f(t))ds
•~o dt 0 o
+it (A(s)e(t-s)A(s)- A(t)e(t-s)A(t))f(t)ds + etA(t) J(t). (6.174)

Combining (6.172),(6.173),(6.174) yields

! u(t) = A(O)etA(O)x + Z(t, O)x +it A(s)e(t-s)A(s)(J(s)- f(t))ds

+it (A(s)e(t-s)A(s)- A(t)e(t-s)A(t))J(t)ds

+etA(t) j(t) +it Z(t, s)J(s)ds. (6.175)

It is easy to see that u(t) ---> x as t---> 0. It is also easily seen that the third,
fourth and last terms of the right hand side of (6.175) tend to 0 as t ---> 0.
Suppose that x E D(A(O)) and A(O)x + f(O) E D(A(O)). Then by virtue of
Lemma 6.23 Z(t, O)x - t 0 as t - t 0. In view of Lemma 6.8 we have

L to:;-;3;.
k
JletA(t) - etA(O) II ::; C
i=l

Hence
A(O)etA(O)x + etA(t) f(t)
= etA(O)(A(O)x + f(O)) + (etA(t)- etA(O))f(t) + etA(O)(J(t)- f(O))
-t A(O)x + f(O)
as t ---> 0. Therefore (6.171) is established.
278 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Theorem 6. 6 Suppose that the assumptions (P1),(P2),(P4) are satisfied.

Let f E Co:([O, T]; X) for some a E (0, 1] and x E D(A(O)). Then a classical
solution of (6.1), (6.2) exists and is unique, and is given by (6.170). If
moreover x E D(A(O)) and A(O)x + f(O) E D(A(O)), then the classical
solution is a strict solution.

Proof. By virtue of Theorem 6.5 and Proposition 6.15 it remains to show

that the function u given by (6.170) satisfies the equation (6.1) in (O,T].
Let Un be the solution of (6.128),(6.129). Then Un is given by (6.160) and
analogously to (6.175) we have
d
dt Un(t) = An(O)etA,(O)X + Zn(t, O)x

+it An(s)e<t-s)An(s)(f(s)- f(t))ds

+ it(An(s)e<t-s)A,(s)- An(t)e<t-s)A,(t))f(t)ds

+etAn(t) f(t) +it Zn(t, s)f(s)ds. (6.176)

Using Lemmas 6.15, 6.21 and (6.161),(6.162),(6.175),(6.176) it can be easily

shown that un(t)--> u(t), dun(t)jdt--> du(t)jdt fortE (0, T]. Since

Un(t) = An(t)- 1 (! Un(t) - f(t))

= ( A(t)- 1 - ~) (! Un(t)- f(t)) --> A(t)- 1 (! u(t)- f(t)),

we have u(t) = A(t)- 1 (du(t)jdt- f(t)). Therefore u(t) E D(A(t)) and (6.1)
holds.
Remark 6. 5 A maximal regularity result analogous to Theorem 6.4 holds.
See Acquistapace and Terreni [8: Theorem 6.1].

6.13 Application
In [1] Acquistapace showed that the hypotheses (P1),(P2),(P4) are satisfied
by the realizations of strongly elliptic operators in the space C(O), where
0 is a bounded open set of Rn, under suitable boundary conditions making
use of the result of H. B. Stewart [145]. The results of [1],[6],[8] are powerful
for such equations having operators which are not densely defined. However
in this section we consider the same kind of problem in £ 1 (0). This is a
6.13 APPLICATION 279

generalization ofD. G. Park [122] where the differentiability in t of the

coefficients is assumed.
Let n be a bounded open set of Rn of class C 2m. Let

L(x, t, Dx) = 2.: a0 (x, t)D~

l<>l~m

be a linear differential operator of order m with coefficients defined in 0 X

[0, T]. We assume that L(x, t, Dx) is uniformly strongly elliptic in 0 x [0, T],
i.e.
(-1)m/2 Re 2.:
aa(x,t)t;";::: eolt;lm, eo> 0,
l<>l~m

for any (x, t) E 0 X [0, Tj and t; ERn. Let

Bj(x, t, Dx) = 2.: bif3(X, t)D~, j = 1, ... , m/2,

lf31~mj

be a set of boundary operators with coefficients defined on an X [0, T]. We

assume that all the hypotheses of section 3 of Chapter 5 are satisfied by
L(x, t, Dx) and {Bi(x, t, DxH'7l; for each fixed t E [0, T].
Let

L'(x, t, Dx) = 2.: a~(x, t)D~,

l<>l~m

Bj(x, t, Dx) = 2.: b~f3(x, t)D~, j = 1, ... , m/2

lf31~mj

be the adjoint system of L(x, t, Dx), {Bj(x, t, Dx)}-;l;. Let all the coeffi-
cients of Bj(x, t, Dx), Bj(x, t, Dx),j = 1, ... , m/2, be extended to the whole
of 0 x [0, Tj. We assume that

aa, a~, la:l :::; m, DJbjf3, 1!11:::; mj, h'l :::; m- mj,
DJb~f3' 1!11 :::; m~, bl :::; m- m~, j = 1, ... , m/2
are uniformly Holder continuous in t of order h. We denote by A(t) the
realization of L(x, t, Dx) in L 1 (!1) under the boundary conditions
Bj(X, t, Dx)ulan = 0, j = 1, ... , m/2,
and those of L(x, t, Dx) and L'(x, t, Dx) in LP(O), 1 < p < oo, under the
boundary conditions

Bj(x, t, Dx)ulan = 0, j = 1, ... , m/2, and

Bj(x, t, Dx)ulan = 0, j = 1, ... , m/2,
280 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

by Ap(t) and A~(t) respectively. We assume for the sake of simplicity that
0 E p(A(t)) n p(Ap(t)) for all p E (1, oo). Following the argument of section
3 of Chapter 5 we establish the estimate of the kernel of A(t)- 1 - A(T)- 1
and its derivatives.
As in Chapter 5 for a real vector 'filet A~ (t) and A~ 11 ( t) be the realizations
of L(x, t, Dx+'fl) and L' (x, t, Dx+'fl) in LP(n) under the boundary conditions
Bj(x, t, Dx + 'fl)ulan = 0, j = 1, ... , m/2, and
Bj(x, t, Dx + 'fl)ulan = 0, j = 1, ... , m/2,
respectively. The assertions of Lemmas 5.11-5.15 hold for each t E (0, T]
with constants independent oft. Hence there exists an angle ()0 E (0, 1r /2)
such that if we set
E = {A; Oo ::; arg A ::; 27r - Oo} U {0}
the following assertion holds: for 1 < p < oo there exist positive constants
o and C such that if A E E,JAI ~ C,J'fll::; oJAJ 1 fm,t E (O,T], then A E
p(A~(t)) n p(A~, 11 (t)) and
m
2: IAI(m-k)/mll (A~(t) - A)- 1 fllk,p,n ::; Cllfllo,p,n, (6.177)
k=O
m
2: IAI(m-k)/mii(A~,'TI(t)- A)- 1 fllk,p',n::; Cllfllo,p',n (6.178)
k=O

for f E LP(n) or f E LP' (n).

Lemma 6. 24 Let u(t) = (A~(t) - A)- 1 f for 1 < p < oo, A E E with IAI
sufficiently large and l'fll ::; oJAJ 1 fm. Then fort, T E (0, T]
m
2: IAI(m-k)/mllu(t)- u(T)llk,p,n::; CJt- Tlhllfllo,p,n· (6.179)
k=O

Proof. By the definition of u(t) we have

(L(x, t, Dx + 'fl) - A)u(t) = f in n,
Bj(x, t, Dx + 'fl)u(t) = 0 on an, j = 1, ... , m/2.
Therefore
(L(x, t, Dx + 'fl)- A)(u(t)- u(T))
= (L(x, T, Dx + 'fl)- L(x, t, Dx + 'fl))u(T) in n,
Bj(x, t,Dx +'T/)(u(t)- u(T))
= (Bj(X, T, Dx + 'T/)- Bj(X, t, Dx + 'T/))u(T) on an.
6.13 APPLICATION 281

By Theorem 5.5 and (6.177)

m
L j>.j(m-k)/mjju(t)- u(T)ilk,p,.n
k=O
::; c[II(L(x, T, Dx + ry)- L(x, t, Dx + ry))u(T)IIo,p,.n
m/2
+L j>.j(m-mj)/mjj(Bj(X, T, Dx + ry)- Bj(X, t, Dx + ry))u(T)jJo,p,.n
j=1
m/2
+ ~ jj(Bj(X, T,Dx + ry)- Bj(X, t,Dx + ry))u(T)ilm-mj,p,.n]

::; Cjt- Tjh [~ Jryjm-kllu(T)jJk,p,.n

m/2 mj
+L j>.j(m-mj)/m L jryjmrkllu(T)ilk,p,.n
j=1 k=O
m/2 m
+~ L i"'im-kllu(T)ilk,p,.n]
J=1 k=m-mj
m
:'S Cjt- Tjh L j>.j<m-k)/mjju(T)jJk,p,.n :'S Cjt- Tjhllfllo,p,.n·
k=O
For the sake of simplicity of notations we write in what follows A(t), A 71 (t),
A' (t), A 171 (t) for Av(t), A~ (t), A~(t), A~ 77 (t).
We choose exponents q1, q2, ... , q8 +1, r 1, r2, ... , rl-s+ 1 of section 3 of Chap-
ter 5 so that we have q8 > n, rl-s > n. Then, in view of (5.114) we have for
k= o, ... ,m-1

ljujjk ' oo ,.n ::; Cjjujj(n+kq.)jmq.Jjujj(mq.-n-kq.)jmq

m,qs,n O,qs,n
•. (6.180)

We denote the kernels of the operators (A(t) - >. 1)- 1 · · · (A(t) - .At)- 1 and
(A77(t) - >.1)- 1 · · · (A 71 (t) - >.t)- 1 by K>.. 1 , ... ,>.. 1 (x, y; t) and K~ 1 , •.• ,>.z (x, y; t)
respectively. If we set

S(t) = (A 71 (t)- >.s)- 1 · · · (A 71 (t)- >.1)- 1,

T(t) = (A 71 (t)- >.s+1)- 1 · · · (A 71 (t)- >.t)- 1 ,

then
K At, ... ,A 1 ·ex ' y·' t) K At, ... ,A 1 (x ' y·' T)
77 77
-
282 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

is the kernel of

S(t)T(t)- S(T)T(T) = (S(t)- S(T))T(t) + S(T)(T(t)- T(T)). (6.181)

From Lemma 6.24, (5.113), (6.180) it follows that

II [(A71(t)- Aj)-1- (A71(T)- Aj)-1] !llo,qHz,n

::; Cit- Tlhl>-jlaj- 1 1Jfllo,qj,n, j = 1, ... , s -1, (6.182)
II [(A71(t)- As)-1- (A71(T)- As)-1] Jllk,ex>,!l
::; Cit- Tlhl>-sla,-l+kfmiJfllo,q.,n, k = 0, ... , m- 1. (6.183)

Using
s
S(t)- S(T) = 2)A71(T)- ). 8 ) - 1 · .. (A71(T)- >.i+l)- 1
j=1
x ((A (t)- >.j)- 1 - (A 71 (T)- Aj)- 1] (A 71 (t)- Aj_l)- 1 · · · (A 71 (t)- >.1)- 1
71

and (6.182),(6.183) we obtain

IIS(t)- S(T)II.c(£2(!l),Wk,oo(n))
s
::; Cit- Tlhl>-slk/m II 1>-jlarl, k = 0, ... , m- 1. (6.184)
j=l

With the aid of Lemma 5.10, (6.184) and (5.126) with T replaced by T(t)
we can estimate the derivatives in x of order up to m - 1 of the kernel of
the first term of the right hand side of (6.181). Those of the kernel of the
second term can be estimated similarly, and we get

l
::; Cit- Tlhl>-sl 1"' 11 m II 1>-jlarl, lnl < m. (6.185)
j=1
With the aid of the argument by which we deduced (5.131) from (5.129) we
obtain

ID~ (K>. 1 , ••• ,>. 1 (x, y; t)- K>. 1 , ••• ,>.1 (x, y; T))l ::; Cit- Tlhl>-sllaJ/m
l l
X:E exp ( -81>-kl 11mlx- Yl) II 1>-jlaj-1, lnl < m. (6.186)
k=1 j=1
6.13 APPLICATION 283

We denote the kernel of A(t)- 1 by K(x, y; t). We follow the argument by

which we derived (5.168) from (5.131) in Chapter 5 using the same notations
there and omitting the details. The kernel of D~(e-laA(t) -e-laA(r)) is equal
to

and in view of (6.186) it is dominated by

Cit- Tlha--1<>1/m-n/m exp(2lEp- &1/m p)

in absolute value. We choose E so small that 2lE - & 11m < 0. Since

we obtain

m< n+ lo:l
m > n + lo:l (6.187)
m = n+ lo:l
In [122] P. Grisvard's result on interpolation in L 1 (0) with boundary condi-
tions was used. However, the corresponding result in LP(O), 1 < p < oo, an-
swers the purpose there and also in what follows. Let m 0 = minj= 1 , ... ,m;2 mj.
Suppose
h >max { m-mo-1 , -1}
m m
Then there exist numbers p, (}satisfying
m-mo-1 1
h>p>O> , (} -m
> -. (6.188)
m
Let (1- O)m- m 0 < 1/q < 1, 1- 1/n < 1/q < 1. In view of P. Grisvard
[71] (see also H. 'friebel [154; pp.329-321]) we have
(D(Aq(t)), U(O))o,q = w(l-O)m,q(O), (6.189)

and by Lemma 1.6

(6.190)
284 CHAPTER 6. PARABOLIC EVOLUTION EQUATIONS

Since (J;::: 1/m, we have

wm-1,q(O) C W(1-9)m,q(O). (6.191)

Combining (6.189),(6.190),(6.191) yields

wm- 1 ,q(0) C D(Aq(t) 1 -P). (6.192)

Using (1.21),(6.192) and that n is bounded we get

IIA(t)R(,\, A(t))(A(t)- 1 - A(T)- 1)/IIo,1,n
= IIA(t)P R(,\, A(t))A(t) 1 -P(A(t)- 1 - A(T)- 1)/IIo,1,n
::::; CI-\IP- 11!A(t) 1-P(A(t)- 1 - A(T)- 1)/IIo,1,n
::::; CI-\IP- 1IIA(t) 1-P(A(t)- 1 - A(T)- 1)/IIo,q,n
::::; CI-\IP- 1II(A(t)- 1 - A(T)- 1)/IIm-1,q,n·

With the aid of (6.187), (n- 1)q <nand Lemma 1.1 we get

Hence we obtain

IIA(t)R(-\, A(t))(A(t)- 1 - A(T)- 1)Ikcucn)) ::::; Cit- Tlhi-\IP- 1

to conclude that the assumption (P4) or rather (II),(III) of A. Vagi [161]

are satisfied.
Remark 6. 6 It can be shown with the aid of Theorem 2.3 of Acquistapace
and Terreni [7] that a similar result holds in the space C(O) of continuous
functions.
Chapter 7

Hyperbolic Evolution
Equations

This chapter is devoted to evolution equations of hyperbolic type. The

results are due to T. Kato [88], Y. Kobayashi [95], N. Okazawa [119], N.
Okazawa and A. Unai [120]. See also A. Pazy [127: Chapter 5] and J. R.
Dorroh [56].

7.1 Admissible Subspaces

Throughout this section X denotes a Banach space with norm II · II, and
{T(t); t;::: 0} a C0 -semigroup in X with infinitesimal generator A. In view
of Theorem 1.4 there exist a constant M ;::: 1 and a real number f3 such that
IIT(t)ll 5: Mef3t, t;::: 0. (7.1)
A subspace Y of X is said to be an invariant subspace of the semigroup
{T(t); t;::: 0} if Y is invariant under the mapping T(t) for any t ;::: 0.
The realization Av of A in a subspace Y of X is defined by
D(Av) = {y E D(A) n Y; AyE Y}, (7.2)
Ayy = Ay for y E D(Ay ). (7.3)
In what follows in this chapter we assume that Y is a subspace of X and
furthermore that it is a Banach space with norm II ·llv which is stronger
than that of X. The norm of .C(Y) is also denoted by II·IIY·
Definition 7. 1 A subspace Y of X is called A-admissible if it is an in-
variant subspace of the semigroup {T(t); t ;::: 0} and the restriction of T(t)
toY is a C0 -semigroup in Y.

285
286 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Clearly Y is A-admissible if and only ifT(·)y E C([O, oo); Y) for each y E Y.

Proposition 7. 1 A necessary and sufficient condition in order that Y is

A-admissible is that the following two conditions are satisfied:
(i) there exists a real number {3 such that Y is invariant under the mapping
R(A, A) for any ,\ > max{{3, ~},
(ii) the realization Av of A in Y generates a Co-semigroup in Y.
Moreover, if Y is A-admissible, then Av is the infinitesimal generator of
the restriction of {T(t); t ~ 0} toY.
Proof. Suppose that Y is A-admissible. Then the restriction of T(t) toY
is a C0 -semigroup in Y whose infinitesimal generator we denote by A. By
virtue of Theorem 1.4 there exists a constant M ~ 1 and a real number ~
such that
IIT(t)llv:::; Meffit, t ~ 0. (7.4)
For any y E Y and,\> max{{3, ~}

R(A, A)y = looo e->-tr(t)ydt = R(,\, A)y

both in the strong topology of X and that of Y. Hence R(,\, A)y E Y, i.e. Y
is invariant under R(,\, A). It follows from the definition of the infinitesimal
generator that if y E D(A) then y E D(A) n y and Ay = Ay E Y. This
shows that y E D(Av) and Avy = Ay. Hence A CAy. On the other hand
if y E D(Av ), then Ay = Ayy E Y, and

T(t)y- y = lot T(s)Ayds = lot T(s)Avyds (7.5)

holds in the strong topology of Y. Dividing (7.5) by t_ and letting t---+ 0 we

find that y E D(A) and Ay = Ayy. Hence Av C A. Therefore we have
proved Av =A.
Conversely suppose that (i) and (ii) are satisfied. Let ,\ > max{{3, ~}, y E
Y, and set x = R(,\,A)y. Then by (i) x E Y, and so Ax= ,\x- y E Y.
Hence x E D(Av) and y = (,\ - Ay )x. Clearly x is uniquely determined
by y, and so ,\ E p(Av ), R(,\, Av )y = R(,\, A)y. Therefore for t > 0 and
njt > max{{3, ~} we have

( 1- ~Ay) -n y = ( 1- ~A) -n y.
Letting n---+ oo we get etAyy = T(t)y. This shows that Y is invariant under
T(t) and the restriction of T(t) toY is a C0 -semigroup.
7. 1. ADMISSIBLE SUBSPACES 287

Corollary 7. 1 Y is A-admissible if and only if there exist a constant M ~

1 and a real number /3 such that
(i) Y is invariant unde! R( .\, A) for Re). > max{,8, /3},
(ii) for Re.\ > max{,B, ,8}, n = 1, 2, ...

IIR(.\,A)niiY:::; M(Re.\- /J)-n, (7.6)

(iii) for). > max{,B, /3}, R(.\, A)Y is dense in Y.

Proof. IfY is A-admissible, then it follows from Proposition 7.1 and its proof
that there exist a constant M ~ 1 and a real number /3 such that (7.4) holds,
T(t)iY = etAy for t ~ 0 and R(.\, A)IY = R(.\, Ay) for ). > max{,B, /J}.
Furthermore for ). > max{,B, /3}, R(.\, A)Y = R()., Ay )Y = D(Ay ). Hence
(i),(ii),(iii) hold. Conversely suppose that (i),(ii),(iii) are satisfied. Then
from the proof of Proposition 7.1 ). E p(Ay) and R()., A)y = R(.\, Ay )y for
). > max{,B, /3} andy E Y. Hence (ii) is rewritten as

Furthermore by (iii) D(Ay) is dense in Y. Hence in view of Theorem 1.4

Ay generates a C0 -semigroup in T. Therefore Y is A-admissible.

Remark 7. 1 If Y is reflexive, then in Corollary 7.1 the condition (iii)

follows from (i) and (ii). Indeed it follows from the resolvent equation

R(.\, A) ~ R(tt, A) = (tt- .\)R(.\, A)R(tt, A)

that D = R(.\, A)Y is independent of). E p(A). In view of (7.6) with n = 1
for each y E Y, .\R(.\, A)y is bounded in Y as).~ oo. If Y is reflexive, there
exists a subsequence An ~ oo such that AnR(.\n, A)y is weakly convergent
in Y to so some element z E Y. Since .\R()., A)y ~ y strongly in X as
). ~ oo, we have y = z. Since AnR(.\n, A)y E D, y is in the weak closure
and hence in the strong closure of D.

Proposition 7. 2 Let Y be the closure ofY in X. LetS be an isomorphism

from Y onto Y. Then Y is A-admissible if and only if A1 = SAS- 1 is the
infinitesimal generator of a C0 -semigroup {T1 (t); t ~ 0} in Y. In this case
we have
(7.7)

Proof. By the definition of A 1 we have

D(Al) = {x E Y; s- 1 x E D(A), AS- 1 x E Y}

= {x E Y; s- 1 x E D(Ay)} = SD(Ay ).
288 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Hence D(A 1) is dense in Y if and only if D(Ay) is dense in Y. For x E D(A 1)

we have

Therefore

{ (,\- A1)x; x E D(A1)} = 8 { (,\ - Ay )y; y E D(Ay )},

{x; (,\- A 1)x = 0} = {0} if and only if {y; (,\- Ay)y = 0} = {0}.

These imply that p(A 1) = p(Ay) and

R(A, A1) = 8R(-X, A)8- 1 = 8R(-X, Ay )8- 1 for ,\ E p(Al) = p(Ay ).

(7.9)
Next we claim that for ,\ E p(A), ,\ E p(A 1) if and only if Y is invariant
under R(,\, A). Suppose first that A E p(A 1). Then for any y E Y there
exists an element x E D(A1) such that

which implies y = (A- A)8- 1x, and hence R(-X, A)y = 8- 1x E Y. Conse-
quently R(,\, A)Y C Y. Conversely suppose that R(-X, A)Y C Y. Let x E Y.
Set y = R(-X, A)8- 1x. Then y E Y n D(A) and Ay = AY- 8- 1x E Y. This
implies y E D(Ay ), and hence 8y E D(A 1). Therefore with the aid of (7.8)

x = 8(,\- A)y = 8(,\- A)8- 18y = (,\- A1)8y E R(,\- A1).

With the aid of (7.8) we see that (,\ - A 1 )x = 0 implies x = 0. Thus we

conclude A E p(A1).
Suppose that A1 generates a C0 -semigroup in Y. Then D(A 1 ) is dense
in Y, and there exist a constant M 1 ~ 1 and a real number /31 such that
p(A1) ::J {A; Re,\ > /31} and for Re,\ > /31

IIR(\ A1)nllv:::; M1(Re,\- /31}-n, n = 1, 2,.... (7.10)

Since D(A1) is dense in Y, D(Ay) is dense in Y. If Re,\ > /31, then in view
of (7.9),(7.10) A E p(A 1) = p(Ay) and

IIR(-\,Ay)nliY = II8- 1R(-\,A1)n8lly

:::; ll8- 1ll.q¥,Y)II8IIc(Y,Y)M1(Re,\- /31)-n, n = 1, 2, ....

This implies that Ay generates a C0 -semigroup. By the previous step for

,\ > max{/3, /31}, Y is invariant under R(-X, A). Therefore by Proposition
7.1 Y is A-admissible.
7.2. STABLE FAMILIES OF INFINITESIMAL GENERATORS 289

Conversely suppose that Y is A-admissible. Then by Corollary 7.1 there

exist a constant M ;::: 1 and a real number ~ such that (i),(ii),(iii) in the
corollary hold. Since D (Ay) is dense in Y, D(A 1 ) is dense in Y. If Re>. > ~,
then >. E p(Ay) = p(A1) and by virtue of (7.9),(7.6)

IIR(>., A1)nllv = IISR(>., A)ns- 1 llv

S IIBII.c(Y,V)IIS- 1 II,qv,Y)M(Re.A- ~)-n, n = 1, 2, ....

Thus A 1 generates a C0 -semigroup in Y. Finally (7.7) follows from (7.9)

and Theorem 1.5.

7.2 Stable Families of Infinitesimal

Generators
In this section we consider a family { A(t); t E [0, T)} of infinitesimal gener-
ators of C0 -semigroups in a Banach space X.
Definition 7. 2 A family {A(t); t E [0, T]} of infinitesimal generators of
C0 -semigroups in X is called stable with stability constants M ;::: 1, f3 if

p(A(t)) ::) ([3, oo), t E [0, T] (7.11)

and for every finite sequence 0:::; t 1 :::; t2 :::; • • • :::; tk :::; T, k = 1, 2, ... , and
.-\>[3

iJ R<A. A(<;ll J s M<>-- m-•. (7.12)

In (7.12) and in what follows products containing {tj} are always "time
ordered", i.e. a factor with a larger tj stands to the left of ones with smaller
tj.
If A(t) E G(X, 1, [3) for any t E [0, T], then the family {A(t); t E [0, T]} is
clearly stable with stability constants 1, [3.
Proposition 7.3 A family {A(t); t E [0, T]} of infinitesimal generators of
Co-semigroups in X is stable if and only if there exist constants M ;::: 1
and f3 such that (7.11) holds and either one of the following conditions is
satisfied: for any finite sequence 0 :::; t1 :::; t2 :::; · · · :::; tk :::; T, k = 1, 2, ... ,
and Sj ;::: 0, j = 1, ... , k

(7.13)
290 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

or for any finite sequence 0 ::::; t 1 ::::; t2 • • • ::::; tk ::::; T, k = 1, 2, ... , and
AJ > ~,j = 1, ... ,k,
k k
II R(Aj,A(tj)) ::::; MII(Aj- ~)- 1 • (7.14)
j=l j=l

Proof. Suppose that {A(t); t 2 0} is stable with stability constants M, ~·

Let 0::::; h ::::; t 2 ::::; • • • ::::; tk ::::; T and Sj,j = 1, ... , k, be positive rational
numbers. We write Bj = mi/N with a common denominator N. Set m =
r:;=l mj = NL:;=l Bj· Then in view of (7.12)

k k
= II[NR(N,A(tj))]mJ =Nm IIR(N,A(tj))mj
j=l j=l

Letting N --+ oo and using Theorem 1.5 we obtain (7.13). Since C0 -

semigroups are strongly continuous, we conclude that (7.13) holds also for
positive real numbers Sj.
Suppose that (7.13) holds. Let Aj > ~' j = 1, ... , k. With the aid of (1.4)
we have

Making use of (7.13)

Thus (7.14) holds. It is obvious that (7.14) implies (7.12).

A useful criterion of the stability is given in the following proposition.
7.2. STABLE FAMILIES OF INFINITESIMAL GENERATORS 291

Proposition 7. 4 Let {A(t); t E [0, T]} be a stable family of infinitesimal

generators of Co-semigroups in X with stability constants M and (3. Let
B(t), t E [0, T], be bounded linear operators in X. If IIB(t)ll :::; K for any
t E [0, T], then {A(t) + B(t); t E [0, T]} is a stable family of infinitesimal
generators with stability constants M and (3 + K M.
Proof. If A> (3 + KM, then A E p(A(t) + B(t)) and
00

R(A,A(t) + B(t)) = l:R(A,A(t))[B(t)R(A,A(t))t.

n=O

Consequently for 0 :::; t1 :::; tz :::; · · · :::; tk :::; T

k
}] R(A, A(tj) + B(ti)) =}] k {
~ R(A, A(tj)) [B(ti)R(A, A(ti))t .
00 }

(7.15)
If we expand the right hand side of (7.15), a general term is of the form

the norm of which does not exceed Mn+ 1 Kn(A- (3)-n-k in view of the
stability condition (7.12), where n = 2:~= 1 nj. The number of the terms
in which 2:~= 1 ni =n is equal to the coefficient of an in the expansion of
(:E:=o am)k. Since

this coefficient is equal to (k + ~- 1} Therefore

.}] R(A,A(tj) + B(tj)) :::; ~ (k + ~ - 1)Mn+1 Kn(A _ (3)-n-k

= M(A- (3)-k ~ c + ~- 1) [MK(A- (3)-1]n

= M(A- (3)-k (1- MK(A- (3)- 1

rk = M(A- (3- MK)-k,
and the proof is complete.
292 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Proposition 7. 5 Let Y be a Banach space which is densely and continu-

ously embbeded in X, and {S(t); t E [0, T]} a family of isomorphisms from
Y onto X satisfying the following conditions:
(i) IIS(t)ll.c(Y,X) and IIS(t)- 1ll.c<x,Y) are uniformly bounded,
(ii) S(t) is of bounded variation in [0, T] in the norm of .C(Y, X).
Let {A(t); t E [0, T]} be a family of infinitesimal generators in X and set
A1(t) = S(t)A(t)S(t)- 1. If {A1(t); t E [0, T]} is a stable family of infinites-
imal generators in X, then Y is A(t)-admissible for each t E [0, T] and the
family {Ay(t); t E [0, T]} of the realizations of A(t) in Y is a stable family
of infinitesimal generators in Y.
Proof. From Proposition 7.2 it follows that Y is A(t)-admissible for each
t E [0, T], and in view of Proposition 7.1 Ay(t) is the infinitesimal generator
of a C0 -semigroup in Y. As in the proof of Proposition 7.2 D(A1 (t)) =
S(t)D(Ay(t)) and R(A,Ay(t)) = S(t)- 1R(A,A 1(t))S(t) for large enough
real numbers>.. Therefore
k k
IT R(>.,Ay(tj)) =IT S(tj)- 1R(>.,A(tj))S(tj)· (7.16)
j=1 j=1
If we set Pi= (S(tj)- S(ti_ 1))S(tj_l)-1, then right hand side of (7.16) is
equal to
S(tk)- 1R(>., A1 (tk))(1 + Pk) · · · (1 + P2)R(>., A1 (t1))S(t1). (7.17)
Let M 1 and {31 be the stability constants of {A 1(t); t E [0, T]}. When we
estimate the norm of a term containing m of the Pj in the expansion of
(7.17), we need only m+ 1 of M 1. Hence the norm of (7.17) does not exceed
k
IIS(tk)- 1llc(X,Y)M1(>.- {31)-k IT (1 + M1IIPill) IIS(t1)llc(Y,X)·
j=2

If we put

C =max { sup IIS(t)ll.c<Y,X), sup IIS(t)- 1llc(x,Y)},

tE]O,T] tE]O,T]

we obtain
k
IT R(A, Ay(tj))
j=1
::; C 2 M1(>.- {31)-k exp (M1 t IIPill)

ct,
y J=2

"'C' M, (>, -,11,)-k exp [ M, IIS(t;)- S(l;-dllc(Y,X)l

<
-
c2 M 1 eMtCV(>.- {31 )-k '
7.3. CONSTRUCTION OF EVOLUTION OPERATOR(l) 293

where Vis the total variation of S(t). Hence {Ay(t)} is stable.

7.3 Construction of Evolution Operator (1)

In this section we construct an evolution operator for the equation of hy-
perbolic type

d
dt u(t) = A(t)u(t) + f(t), t E [0, T], (7.18)
u(O) = x. (7.19)

Let X andY be Banach spaces with norm 11·11 and ll·lly respectively. We
assume that Y is a dense subspace of X and the norm of Y is stronger that
that of X.
We state the basic assumptions of this section.
(H1) {A(t); t E [0, T]} is a stable family of infinitesimal generators of C0 -
semigroups in X.
(H2) Y is A(t)-admissible for any t E [0, T], and the family {Ay(t); t E
[0, T]} of the realizations Ay(t) of A(t) in Y is stable in Y.
(H3) For t E [0, T], D(A(t)) :::::> Y, and A(t) is a bounded linear operator
from Y into X, and A(t) is continuous in [0, T] in the norm topology of
.C(Y, X).
We denote the stability constants of {A(t); t E [0, T]} by M, (3, and those of
{Ay(t); t E [0, T]} by M, ~.

Theorem 7. 1 Under the assumptions (H1),(H2),(H3) there exists an evo-

lution operator U(t, s), 0 :::; s :::; t :::; T, of (7.18), (7.19) which is a strongly
continuous family of bounded linear operators in X satisfying

U(t, r)U(r, s) = U(t, s), 0:::; s:::; r:::; t:::; T, (7.20)

U(s, s) =I, s E [0, T], (7.21)
IIU(t, s)ll :::; M ef3(t-s), 0:::; s:::; t:::; T, (7.22)

( ;t) + U(t, s)ylt=s = A(s)y, 0:::; s < T, y E Y, (7.23)

8
08 U(t, s)y = -U(t, s)A(s)y, 0:::; s:::; t:::; T, y E Y, (7.24)

where the right derivative in (7.23) and the derivative in (7.24) are in the
strong topology of X Moreover, the evolution operator having the above
properties is unique.
294 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Proof. For a positive integer n lett~= (kjn)T, k = 0, 1, ... , n. Set

A (t) = { A(t~) t~::; t < t~+l• k = 0, 1, ... , n- 1,

(7.25)
n A(T) t=T.
Since A(·) is continuous in the norm of .C(Y, X), we have
IIA(t)- An(t)il.ccY,X)---+ 0 as n---+ oo (7.26)
uniformly in t E [0, T]. Clearly {An(t); t E [0, T]} is stable in X with
stability constants M, {3, and {An,Y(t); t E [0, T]} which is a family of the
realizations of An(t) in Y is stable with stability constants M, ~.
We define an operator valued function Un(t, s), 0::; s::; t::; T, by

Un(t, s) = (7.27)

It is easy to show that

Un(t, s) = Un(t, r)Un(r, s), 0::; s::; r::; t::; T, (7.28)
Un(s, s) =I, (7.29)
and Un(t, s) is strongly continuous ins, tin 0::; s::; t::; T. From (H1) and
Proposition 7.3 it follows that
IIUn(t, s)ll::; Mef3(t-s), 0::; s::; t::; T, (7.30)
and from (H2) and Proposition 7.3

Un(t,s)YcY and 11Un(t,s)IIY::;.Me.6<t-s), o::;s::;t::;T. (7.31)

Since D(A(t)) :> Y we have for y E Y
f)
at Un(t, s)y = An(t)Un(t, s)y, t -:f- t~, k = 0, 1, ... , n, (7.32)
f)
08 Un(t, s)y = -Un(t, s)An(s)y, s -:f- t~, k = 0, 1, ... , n. (7.33)

With the aid of (7.32),(7.33) we have for y E Y, 0::; s < t::; T

Un(t, s)y- Um(t, s)y = -1t :r Un(t, r)Um(r, s)ydr

= 1t Un(t, r)(An(r)- Am(r))Um(r, s)ydr. (7.34)

7.3. CONSTRUCTION OF EVOLUTION OPERATOR(l) 295

Setting 'Y = max{,B, jj} we get from (7.34)

IIUn(t, s)y- Um(t, s)yli ::::; M Me'Y(t-s)IIYIIY 1t IIAn(r) - Am(r)iic(Y,x)dr.

(7.35)
From (7.26),(7.35) it follows that Un(t, s)y converges uniformly in 0::::; s ::::;
t::::; Tin the strong topology of X. Since Y is dense in X and {Un(t, s)} is
uniformly bounded in £(X) in view of (7.30), Un(t, s)x converges uniformly
in 0::::; s::::; t::::; Tin the strong topology of X also for x EX. Let
U(t, s)x = lim Un(t, s)x,
n->oo
0::::; s::::; t::::; T, x E X. (7.36)

Clearly U(t, s)x is continuous in 0 :=:; s :=:; t :=:; T in the strong topology of
X, and it follows from (7.28) and (7.29) that (7.20) and (7.21) hold. Next
we show (7.23),(7.24). For y E Y, 0::::; s < t::::; T, 0::::; T ::::; T we have

Un(t, s)y- e(t-s)A(r)y = -1t :r Un(t, r)e(r-s)A(r)ydr

= 1t Un(t, r)(An(r)- A(T))e(r-s)A(r)ydr. (7.37)

Hence

IIUn(t, s)y- e(t-s)A(r)Yii::::; MMe'Y(t-s)IIYIIY 1t IIAn(r)- A(T)iic(Y,x)dr.

(7.38)
Letting n---+ oo and dividing by t - s we get
U(t, s)y- y - e(t-s)A(r)y- y II
II t-s t-s
- (t ) 1 1t (7.39)
::::; MMe" -s IIYIIY t _ s 8 IIA(r)- A(T)ii.c(Y,x)dr.

Choosing T = s in (7.39) and letting t ---+ s we obtain (7.23). If we choose

T= t in (7.38) and let s ---+ t we obtain

( -a)-u( t, s ) y I . ____:.-'----'-'------'--'-~
= 1lm U(t,t)y-U(t,s)y
as s=t sTt t- s
-l'
- lm y- U(t, s)y -- -A(t) y. (7.40)
sTt t- S
For 0::::; s < t::::; T we get with the aid of (7.23)

(a)+
as U(t, s)y = ~ffl h1 [U(t, s + h)y- U(t, s)y]
296 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

= limU(t, s +h) Y- U(sh+ h, s)y

hlO

= -U(t, s) ( :t) + U(t, s)yL 8 = -U(t, s)A(s)y, (7.41)

and for 0 ::::; s ::::; t ::::; T by (7.40)

( :s)- U(t, s)y = ~ffl ~[U(t, s)y- U(t, s- h)y]

. y- U(s, s- h)y
= hmU(t, s) h = -U(t, s)A(s)y. (7.42)
hlO

Combining (7.41) and (7.42) we obtain (7.24).

Finally we show the uniqueness of an evolution operator. Suppose that
V(t, s) is an evolution operator satisfying (7.20),(7.21),(7.22),(7.23),(7.24).
Then using (7.24) we get for 0::::; s < t::::; T, y E Y

1 t [)
V(t, s)y- Un(t, s)y =- s or V(t, r)Un(r, s)ydr

= 1t V(t, r)(A(r)- An(r))Un(r, s)ydr.

Therefore

IIV(t, s)y- Un(t, s)yll::::; MMe-y(t-s)IIYIIY 1t IIA(r)- An(r)llc(Y,x)dr.

(7.43)
Letting n -? oo in (7.43) and using (7.26) we obtain V(t, s)y = U(t, s)y.
Since Y is dense in X we conclude V(t, s) = U(t, s).

7.4 Uniqueness of Regular Solutions

In this section we consider the initial value problem (7.18),(7.19) assuming
that f E C([O, T]; X) and x E Y. Following A. Pazy [127] we call a function
u a Y-valued solution of (7.18),(7.19) if u E C 1 ([0, T]; X) n C([O, T]; Y) and
(7.18),(7.19) hold. A Y-valued solution is a more restrictive solution than
a strict one since it is required to satisfy u(t) E Y C D(A(t)).
Theorem 7. 2 Suppose that (H1),(H2),(H3) are satisfied. Let x E Y and
f E C([O, T]; X). If u is a Y-valued solution of (7.18),(7.19), then it is
represented by
u(t) = U(t, O)x +lot U(t, s)f(s)ds, (7.44)
CONSTRUCTION OF EVOLUTION OPERATOR(2) 297

where U(t, s) is the evolution operator constructed in Theorem 1. Hence a

Y-valued solution is unique.

Proof. Let Un(t, s) be the operator valued function defined by (7.27). Then
Un(t, r)u(r) is a continuously differentiable function of r except at a finite
number of points and
[)
or Un(t, r)u(r) = -Un(t, r)(An(r)- A(r))u(r) + Un(t, r)f(r). (7.45)

Integrating (7.45) from 0 to t we get

u(t) = Un(t, O)x- lt Un(t, r)(An(r)- A(r))u(r)dr + lt Un(t, r)f(r)dr.

(7.46)
Since

lilt Un(t, r)(An(r)- A(r))u(r)drii

:::; M ef3t sup llu(r) llv

rE[O,Tj
t
Jo
IIAn(r) - A(r)ll.c(Y,x)dr---> 0

as n---> oo, we obtain (7.44) from (7.46).

7.5 Construction of Evolution Operator (2)

In [88] T. Kato proved replacing (H2) by the following more restrictive
assumption
(H2') There is a family {S(t); t E [0, T]} of isomorphisms of Y onto X such
that S(t) is strongly continuously differentiable in [0, T] and

S(t)A(t)S(t)- 1 = A(t) + B(t), t E (0, T], (7.47)

where B(t) is strongly continuous in (0, T] to .C(X)

that the evolution operator constructed in section 7.3 has stronger regularity
properties so that it provides with a Y-valued solution. In this section
following K. Kobayashi [95] we prove a similar result replacing the norm
continuity of A(t) in t by the strong continuity:
(H3') D(A(t)) :::::> Y for each t E [0, T], and A(t) is strongly continuous in
(0, T] to .C(Y, X).
298 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Theorem 7. 3 Under the assumptions (H1), (H2'), (H3') there exists a

unique evolution operator U(t,s),O::; s::; t::; T, of (7.18), (7.19) which
is a strongly continuous family of bounded linear operators in X satisfying
(7.20), (7.21), (7.22), (7.24) and

U(t, s)Y c Y, and U(t, s) is strongly continuous

in 0::; s::; t::; T to .C(Y), (7.48)
a
at U(t, s)y = A(t)U(t, s)y, y E Y, 0::; s::; t::; T, (7.49)

where the derivatives in (7.24), (7.49) are in the strong topology of X.

Proof. We divide the proof into several steps.
Step 1. Wefirstnotethat IIA(t)llccY,X), IIS(t)llccv,x), IIS(t)- 1 llccx,Y)• IIB(t)ll
are bounded in t E [0, T]. By vitue of Proposition 7.4 and the assumption
(H2') S(t)A(t)S(t)- 1 is stable in X, and so applying Proposition 7.5 we see
that Y is A(t)-admissible for each t E [0, T] and Av(t) is stable in Y. We
denote the stability constants of {Ay (t)} by M, [3.
Let P = {tk; k = 0, 1, 2, ... } be a sequence such that 0::; t 0 < t 1 < · · · <
tk < · · · ::; T and t 00 = limk---+oo tk. For this sequence P we define the
operator U(t, s; P), t 0 ::; s::; t < t 00 , by

(7.50)

For an operator valued function F(t), t E [0, T], let F(t; P) be the step
function defined by

F(t; P) = F(tk), t E [tk, tk+l), k = 0, 1, 2, ....

By the above remark U(t, s; P) leaves Y invariant and

IIU(t,s;P)II :::;MefJ(t-s), IIU(t,s;P)IIY :::;Mejj(t-s). (7.51)

Step 2. Suppose that P = {tk; k = 0, 1, 2, ... } an infinite sequence. Let

tk ::; t~ < tk+l, k = 0, 1, 2, .... Then we have

lim U(t~, to; P)x existsin X for any x EX, (7.52)

k---+oo
lim U(t~, t 0 ; P)y exists in Y for any y E Y. (7.53)
k---+oo
CONSTRUCTION OF EVOLUTION OPERATOR(2) 299

Proof. If x E Y, then by (7.51)

Jl(djdt)U(t, to; P)xJI = IIA(t; P)U(t, to; P)xll

::; JIA(t; P)llc<v,x)IIU(t, to; P)JivJixJiv::; Cllxllv·
Hence limk->oo U(t%, t 0 ; P)x exists in X. The same result remains valid for
x EX since JIU(t, s; P)ll is uniformly bounded by (7.51).
In order to prove (7.53) we first show that
JIS(tk)U(t%, ti; P)S(ti)- 1x- U(t%, ti; P)xll ::; C(t%- ti)llxJJ (7.54)
for x E X and 0 ::; i ::; k. It suffices to show (7.54) for x E Y since Y is
dense in X. Using (7.47) we get forti ::; a < tj+ 1, j = i, ... , k- 1,
f)
aa [S(t%)U(t%, a; P)S(a; P)- 1U(a, ti; P)x]
= -S(t%)U(t%, a; P)A(ti)S(ti)- 1U(a, ti; P)x
+S(t%)U(t%, a; P)S(ti)- 1A(ti)U(a, ti; P)x
= -S(tk)U(t%, a; P)S(ti)- 1 B(ti)U(a, ti; P)x.

Hence we have

:a [S(t%)U(t%, a; P)S(a; P)- 1 U(a, ti; P)x]

= -S(t%)U(t%, a; P)S(a; P)- 1 B(a; P)U(a, ti; P)x (7.55)
for ti ::; a < tk. The same equality also holds for tk ::; a ::; t%. Integrating
(7.55) from tj to tH 1 and rewriting the obtained equality appropriately we
get

S(t%)U(t%, tH1; P)S(tHt)- 1U(tH1, ti; P)x

+S(t%)U(t%, ti+ 1; P)(S(ti)- 1 - S(tH 1)- 1)U(tH1, ti; P)x
-S(t%)U(t%, ti; P)S(ti)- 1 U(tj, ti; P)x

=- l tj+l
t·
3
S(tk)U(t%,a;P)S(a;P)- 1B(a;P)U(a,ti;P)xda. (7.56)

Adding (7.56) for j = i, ... , k - 1 yields

S(tk)U(tk, tk; P)S(tk)- 1 U(tk, ti; P)x- S(t%)U(t%, ti; P)S(ti)- 1x
k-1
+ LS(t%)U(t%, ti+1; P)(S(tj)- 1 - S(tH1)- 1)U(tj+1, ti; P)x
j=i

= - l t;
tk
S(t%)U(t%, a; P)S(a; P)- 1 B(a; P)U(a, ti; P)xda. (7.57)
300 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Integrating (7.55) from tk tot~

S(t~)S(tk)- 1 U(t~, ti; P)x- S(t~)U(t~, tk; P)S(tk)- 1 U(tk, ti; P)x
t"
= - { S(t~)U(t~, o-; P)S(o-; P)- 1 B(o-; P)U(o-, ti; P)xdo-. (7.58)
k
ltk
From (7.57) and (7.58) we obtain
S(t~)U(t~, ti; P)S(ti)- 1x- U(t~, ti; P)x
= S(t%)(S(tk)- 1 - S(t~)- 1 )U(t~, ti; P)x
k-1
+ LS(t~)U(t~, ti+1; P)(S(tj)- 1 - S(ti+l)- 1)U(ti+1• ti; P)x
j=i

+ L t"
k S(t~)U(t~, o-; P)S(o-; P)- 1 B(o-; P)U(o-, ti; P)xdo-. (7.59)

Since S(t)- 1 is Lipschitz continuous in the norm of £(X, Y), we readily

obtain (7.54) from (7.59) using (7.51).
For x EX set Wi = S(ti)U(ti, to; P)S(to)- 1 x, and
W(t, s; P) = S(t)U(t, s; P)S(s)- 1 - U(t, s; P).
Then by (7.51) and (7.54) we have
IIW(t~, ti; P)will = ii(S(t~)U(t~, ti; P)S(ti)-1 - U(t~, ti; P))will
$ C(t~- ti)llwill $ C(t~- ti)llxll· (7.60)
Since
S(t~)U(t~, t 0 ; P)S(to)- 1x
= S(t~)U(t~, ti; P)S(ti)- 1wi = W(t~, ti; P)wi + u_(t~, ti; P)wi,
it follows from (7.60) that
ak,i =IIS(tj')U(tj', to; P)S(to)- 1x- S(t~)U(t~, to; P)S(to)- xll 1

$ IIW(tj', ti; P)will + IIW(t~, ti; P)will

+IIU(tj', ti; P)wi- U(t~, ti; P)will
$ C(tj- ti + t~- ti)llxll + IIU(tj, ti; P)wi- U(t%, ti; P)will·
Since limk--+oo U(t~, ti; P)wi exists in X in view of (7.52) (from the proof it
is obvious that the same result remains valid with to replaced by any other
value in (to, t 00 ) in (7.52) ), we obtain
limsupak,j $ C(too -ti)llxll
k,j--+oo
CONSTRUCTION OF EVOLUTION OPERATOR(2) 301

for each i. Hence we get limsupk,j->oo ak,j = 0, and

lim S(t~)U(t~, to; P)S(to)- 1x

k->oo
exists in X. Therefore for any y E Y
U(t~, to; P)y = S(t~)- 1 · S(t%)U(t%, to; P)S(to)- 1 · S(to)Y
converges in Y for any y E Y.
Step 3. For each E > 0, y E Y and s E [0, T) there exists a finite partition
P( E, s, y) : s =to < t 1 < · · · < tN = T of the interval [s, T] such that
tk+1 - tk :::; E, k = 0, 1, ... 'N - 1, (7.61)
II(A(t') -A(t))U(t,s;P)yii:::; E for t,t' E [tk,tk+1],
k = 0, 1, ... , N- 1. (7.62)
Proof. Set t 0 = s and define tk+ 1 inductively in the following manner. If
tk = T, then set tk+ 1 = tk; and if tk < T, then set tk+l = tk + hk where hk
is the largest number such that the following conditions hold:
0 < hk :::; E, tk + hk :::; T, (7.63)
II(A(t')- A(t))uk(t- tk)ll :::; E fort, t' E [tk, tk + hk], (7.64)
where
k-1
Uk(t) = etA(tk) II e(tHt-t3)A(t3)y.
j=O

Since (A(t')- A(t))uk(t- tk) is strongly continuous in X in t, t' by virtue

of the assumption (H3'), we see that hk > 0. If we verify that tN = T for
some N, then the proof will be complete. On the contrary suppose that
tk < T for all k. Clearly hk -+ 0 ask-+ oo. Set P' = {tk; k = 0, 1, 2, ... }.
Since (A(t') - A(t))uk(t - tk) is strongly continuous in t, t' in X, we can
show without difficulty that if k is so large that hk < E, then there exist
t~, t~ E [tk, tk+1) such that

(7.65)
According to (7.53) limk->oo U(t%, t 0 ; P')y exists in Y. Hence by (H2') we
have

k-1
= (A(tU _ A(t~)) e(t~ -tk)A(tk) II e(tJ+t-tj)A(t3 )y
j=O
= (A(tk) - A(t%)) U(t%, to; P')y -+ 0
302 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

in X as k ----+ oo. Therefore letting k ----+ oo in (7.65) we obtain 0 2: E/2,

which is obviously a contradiction.
Step 4. Let Ei > 0, si E [0, T) and Yi E Y, i = 1, 2, and let~ = P(Ei, si, Yi) be
a partition of [si, T] satisftying (7.61),(7.62) with Ei, si, Yi in place of E, s, y.
Let Pi be any partition of [si, T] which is a refinement of Pi· Then we have
IIU(tl, s1; P1)Y1 - U(t2, s2; P2)Y21i
:S C [IIYl - Yll + IIY2- Yll + E1 + E2 + (itl - t2l + ls1 - s2I)IIYIIYJ
(7.66)
for all ti E [si, T], i = 1, 2, andy E Y.
Proof. Differentiating U(ti, CT; Pi)U(CT, si; Pi)Yi in CT and integrating over
[si, ti] we obtain
U(ti, si; Pi)Yi- U(ti, si; A)Yi

= 1:' U(ti, CT; Pi)(A(CT; Pi)- A(CT; Pi))U(CT, si; Pi)YidCT. (7.67)

Since Pi is a refinement of~ we have in view of (7.62)

II(A(CT; ~)- A(CT, Pi))U(CT, si; Pi)Yill

:::; II(A(CT;~) -A(CT))U(CT,si;Pi)Yill + li(A(CT,Pi) -A(CT))U(CT,si;Pi)Yill
:::; 2Ei
for CT E (si, T]. Hence from (7.67) we get

IIU(ti, si; Pi)Yi- U(ti, si; A)Yill :::; CEi, i = 1, 2. (7.68)

Therefore

IIU(t1, s1; P1)Y1- U(t2, s2; P2)Y2II

:::; C(E1 + E2) + IIU(t1, s1; Pl)yl- U(t2, s2; P2)Y2II
:::; C( El + E2) + 11 + 12 + h, (7.69)
where

h = IIU(tt, St; Pt)Yl -U(tl, St; P3)Ytll,

12 = IIU(tl, s1; P3)Yl - U(t2, s2; P3)Y2II,
h = IIU(t2, s2; P3)Y2- U(t2, s2; P2)Y21i,
and P3 is the superposition of P 1 and P2. By virtue of (7.68) we have
(7.70)
CONSTRUCTION OF EVOLUTION OPERATOR(2) 303

Therefore it remains to estimate 12 . We may assume s2 ::; s 1 • By (7.51) we

have

12 ::; M e.BT CIIY1 - Yli + IIY2- Yli) + IIU(t1, s1; P3)y- U(t2, s2; P3)Yii· (7.71)

Noting li(djdt)U(t, s2; P3)Yii ::; CIIYIIv, we obtain

IIU(t1, s1; P3)y- U(t2, s2; P3)Yii

::; IIU(t1, s1; P3)y- U(t1, s2; P3)Yii + IIU(t1, s2; P3)y- U(t2, s2; P3)Yii
::; IIU(t1, s1; P3)(1- U(s1, s2; P3))yi1 + Clt1 - t21iiYIIY
::; Cli(1- U(s1, s2; P3))yi1 + Clt1- t21iiY1iv
::; C(ls1- s2i + it1- t2i)i1Yiiv· (7.72)

Combining (7.69), (7.70), (7.71), (7.72) we conclude (7.66).

Step 5. Let x E X and 0::; s ::; t ::; T be fixed. Let {sn}, {tn}, {Yn} be
sequences such that 0 ::; Bn ::; tn ::; T, Yn E Y, Bn ---+ s, tn ---+ t, Yn ---+ x in
X. For each n let Pn be a partition of [sn, T] satisfying (7.61), (7.62) with
E, s, y replaced by 1/n, sn, Yn respectively. Then in view of (7.66) we have
for each y E Y

n,m---+oo

::; lim C[IIYn- Yli

n,m---+oo
+ IIYm- Yli + 1/n + 1/m
+(itn- tml + lsn- Bmi)IIYIIv] = 2CIIx- Yli·
Since Y is dense in X, this implies

U(t, s)x = n---+oo

lim U(tn, sn; Pn)Yn (7.73)

exists in X, and furthermore using (7.66) we see that the limit U(t, s)x does
not depend on the choice of the sequences {sn}, {tn}, {Yn}, {Pn}· We are
going to show that
For each s, t, U(t, s) is a bounded linear opemtor in X satisfying

IIU(t, s)ll ::; Me,B(t-s), (7.74)

and is strongly continuous in 0 ::; s ::; t ::; T to ..C(X). Furthermore it
satisfies (7.20), (7.21).
From Step 4 or (7.68) in its proof it follows that

U(t, s)x = lim U(tn, sn; Pn)Yn

n---+oo
304 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

for any refinement Pn of Pn. Hence it is easily seen that U(t, s) is a linear
operator in X. The inequality (7.74) is a direct consequence of (7.51). Let
s ::; r ::; t and r n be a point of Pn such that r n -+ r as n -+ oo. If we set
Zn = U(rn, sn; Pn)Yn, then Zn E Y and Zn-+ U(r, s)x. Let P~ = Pnn[rn, T].
Then P~ satisfies (7.61), (7.62) with E, s, y replaced by 1/n, rn, Zn, since

U(t, rn; P~)Zn = U(t, rn; P~)U(rn, sn; Pn)Yn = U(t, sn; Pn)Yn· (7.75)

Therefore letting n-+ oo in (7.75) with t = tn we obtain

U(t, r)U(r, s)x = U(t, s)x.

With the aid of Step 4 we can show without difficulty that for 0 ::; s ::; t ::;
T, 0 ::; s' ::; t' ::; T and x E X, y E Y

IIU(t, s)x- U(t', s')xll ::; C [2llx- Yll +(It- t'l +Is- s'I)IIYIIv].

Hence U (t, s )x is continuous in 0 ::; s ::; t ::; T in the strong topology of X.

Step 6. We prove (7.24) in this step. The statement is established with the
aid of an argument similar to that used in the proof of the same statement
in Theorem 7.1 via (7.39), (7.40), (7.41) if we have the following inequality:
for 0 ::; s ::; t ::; T, r E [0, T] and y E Y

IIU(t, s)y- e(t-s)A(r)YII ::; C 1t II(A(o-)- A(r))e(a-s)A(r)ylldo-. (7.76)

Proof of (7.76). Let Pn = P(1jn, s, y) be a partition of [s, T] as in Step 3.

Differentiating U(t, a-; Pn)e<a-s)A(r)y in a- and integrating over [s, t] yields
U(t, s; Pn)Y- e(t-s)A(r)y

= 1t U(t, a-; Pn)(A(o-; Pn)- A(r))e(a-s)A(r)ydo-.

Hence with the aid of (7.51) we obtain

IIU(t, s; Pn)Y- e(t-s)A(r)YII

::; M ef3T 1t II(A(o-; Pn) - A(r))e(a-s)A(r)YIIdo-. (7. 77)

The inequality (7. 76) is obtained by letting n -+ oo in (7. 77).

We also note that if we have (7. 76) we can show (7.23) since

s)y- y - e(t-s)A(s)y- y II ::; _1_11U(t, s)y- e<t-s)A(s)YII

II U(t,t-s t-s t-s
CONSTRUCTION OF EVOLUTION OPERATOR(2) 305

:::; t ~s 1t \\(A(a)- A(s))e(u-s)A(s)y\\da,

1\(A(a)- A(s))e(u-s)A(s)y\\
:::; \\(A(a)- A(s))(e(u-s)A(s)y- y)ll + 1\(A(a)- A(s))y\1
:::; Clle(u-s)A(s)y- Yllv + II(A(a)- A(s))YII---> 0

as a----> s.
Step 7. We complete the proof of the theorem in this step. We begin with
the proof of (7.48). Let s E [0, T) andy E Y. Let Pn = P(ljn, s, y) be a
partition of [s, T] as in Step 3. Set V(t, r) = U(t, r)S(r)- 1 for s:::; r:::; t:::; T.
Then using(7.24)

8
or V(t, r)U(r, s; Pn)Y = -U(t, r)A(r)S(r)- 1 U(r, s; Pn)Y
d
+U(t, r) drS(r)- 1 • U(r, s; Pn)Y + U(t, r)S(r)- 1 A(r; Pn)U(r, s; Pn)Y
d
= -U(t, r)A(r)S(r)- 1U(r, s; Pn)Y + U(t, r) drS(r)- 1 • U(r, s; Pn)Y

+U(t, r)S(r)- 1 (A(r; Pn)- A(r))U(r, s; Pn)Y

+U(t, r)S(r)- 1 A(r)U(r, s; Pn)Y
d
= -U(t, r)S(r)- 1 B(r)U(r, s; Pn)Y + U(t, r) drS(r)- 1 • U(r, s; Pn)Y
+U(t, r)S(r)- 1 (A(r; Pn)- A(r))U(r, s; Pn)Y
= V(t, r)C(r)U(r, s; Pn)Y
+V(t, r)(A(r; Pn)- A(r))U(r, s; Pn)Y, (7.78)

where
C(r) = S(r) drd S(r)- 1 - B(r) = S(r)S(r)-
.
1 - B(r),

which is strongly continuous in [0, T] to ,C(X). Integration of (7.78) from s

tot yields

S(t)- 1 U(t, s; Pn)Y- V(t, s)y = 1t V(t, r)C(r)U(r, s; Pn)ydr

+ 1t V(t,r)(A(r;Pn) -A(r))U(r,s;Pn)ydr. (7.79)

Since
II(A(r; Pn)- A(r))U(r, s; Pn)YI\ :::; 1/n
306 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

in view of (7.62) we see that the second term of the right hand side of (7. 79)
tends to 0 as n --+ oo. Hence we obtain

S(t)- 1 U(t, s)y- V(t, s) = 1t V(t, r)C(r)U(r, s)ydr. (7.80)

Since Y is dense in X, (7.80) implies

V(t, s) = S(t)- 1 U(t, s) -1t V(t, r)C(r)U(r, s)dr. (7.81)

Let W(t, s) be the solution of the integral equation

W(t, s) = U(t, s) -1t W(t, r)C(r)U(r, s)dr. (7.82)

This equation can be solved by successive approximation and W(t, s) is

continuous in the strong topology of£( X). Since S (t) - 1 W (t, s) also satisfies
the integral equation (7.81) whose solution is unique, we get

S(t)- 1 W(t, s) = V(t, s) = U(t, s)S(s)- 1 •

Hence we obtain
U(t, s)IY = S(t)- 1W(t, s)S(s),
from which (7.48) follows.
From (7.23) which was proved in the last step and (7.48) it easily follows
that
( :t) + U(t, s)y = A(t)U(t, s)y (7.83)

for y E Y. Since the right hand side of (7.83) is continuous in 0 ::::; s ::::; t ::::; T
in the strong topology of X, we conclude (7.49) with the aid of Lemma 1.3.
Thus the proof of Theorem 7.3 is complete.

Remark 7. 2 The partition P(E, s, y) of Step 3 depends onE, s, y. In [160]

A. Yagi obtained a partition P of the whole inteval [0, T] independent of s
such that we have

sup II(A(t)- A(t; P))U(t, s; P)yll ::::; E.

O~s~t~T

This enables us to construct the evolution operator also by using the Yosida
approximation of A(t). But the proof is rather delicate and laborious, and
we prefered the simpler proof of Kobayashi [95].
7.6. EXISTENCE OF REGULAR SOLUTIONS 307

Theorem 7. 4 Suppose that {A(t); t E [0, T]} is a stable family of in-

finitesimal generators of Co-semigroups in X. If D(A(t)) = D is inde-
pendent oft and for every y E D, A(t)y is strongly continuously differ-
entiable in X, then there exists a unique evolution operator U(t, s) satisfy-
ing (7.20),(7.21),(7.22),(7.24),(7.48),(7.49), where Y is the spaceD equipped
with the graph norm of A(O):

IIYIIY = IIYII + IIA(O)yll·

Proof. Let M, {3 be the stability constants of {A(t); t E [0, T]}. If we set
S(t) = .\0 - A(t) for some .\0 > {3, then S(t) is an isomorphism of Y onto X
and the assumption (H2') is satisfied with B(t) = 0. Clearly the assumption
(H3), and hence (H3') is satisfied. Therefore the assertion of the theorem
follows from Theorem 7.3.

7.6 Existence of Regular Solutions

This section is concerned with the initial value problem (7.18), (7.19).
Theorem 7. 5 Suppose that the assumptions of Theorem 7.3 are satisfied.
If x E Y, f E C([O, T]; Y), then the function u given by

u(t) = U(t, O)x +lot U(t, s)f(s)ds (7.84)

is the unique Y-valued solution of (7.18), (7.19).

Proof. In view of (7.21), (7.49) U(t, O)x is a Y-valued solution of (7.18),
(7.19) with f(t) = 0. By virtue of (7.48), U(t, s)f(s) is continuous in
0::; s::; t::; T to Y and with the aid of (7.49) we see that

w(t) =lot U(t, s)f(s)ds

is continuously differentiable in X and

dw(t)jdt = A(t)w(t) + f(t).
Hence u(t) given by (7.84) is a desired Y-valued solution. The uniqueness
is proved just as Theorem 7. 2.
Theorem 7. 6 Suppose {A(t); t E [0, T]} is a stable family of infinitesimal
generators of Co-semigroups in X such that D(A(t)) = D is independent of
t and for each y E D, A(t)y is continuously differentiable in X. If x E X,
f E C 1 ([0, T]; X), then the function given by (7.84) is the unique strict
solution of (7.18), (7.19).
308 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Proof. We use the notation of the proof of Theorem 7.4. By assumption

is a continuous function taking values in Y. By virtue of (7.24) we have

U(t, s) = aU(t, s)jas · S(s)- 1 + >.oU(t, s)S(s)- 1 •

Using this and integrating by parts we get

lot U(t, s)f(s)ds

=lot :s U(t, s) · S(s)- f(s)ds + Ao lot U(t, s)S(s)- f(s)ds
1 1

= S(t)- f(t) - U(t, O)S(0)- f(O) -lot U(t, s)! (S(s)- f(s))ds
1 1 1

+>. lot U(t, s)S(s)- f(s)ds.

0 1 (7.85)

From (7.85) it follows that

d t
dt Jo U(t, s)f(s)ds = -A(t)U(t, O)S(o)- 1 f(O)
rt d
- Jo A(t)U(t, s) ds (S(s)- 1 f(s))ds + >.0 S(t)- 1 f(t)

+>.o lot A(t)U(t, s)S(s)- f(s)ds, 1

A(t) lt U(t, s)f(s)ds = A(t)S(t)- 1 j(t)- A(t)U(t, O)S(0)- 1 f(O)

-lt d
A(t)U(t, s)-(S(s)-
ds
1 f(s))ds + >.o
o
lt
A(t)U(t, s)S(s)- 1 f(s)ds.

Therefore we obtain

d
dt t t
Jo U(t, s)f(s)ds- A(t) Jo U(t, s)f(s)ds
= >.oS(t)- 1 j(t) - A(t)S(t)- 1 f(t) = f(t).
7.7. EQUATIONS IN HILBERT SPACES 309

=
Since the first term of the right hand side of (7.84) is a strict solution with
f(t) 0, we complete the proof.

7. 7 Equations in Hilbert Spaces

In this section following N. Okazawa and A. Unai [120] we state how to
solve hyperbolic equations in Hilbert spaces.
Let X be a Hilbert space with norm II · II and innerproduct (·, ·), and
{A(t); t E [0, T]} be a family of closed linear operators in X. Let S be a
nonnegative selfadjoint operator in X, and set

Then Y is a Hilbert space embedded densely and continuously in X. We

make the following assumptions:
(H4) There exists a constant a ;::: 0 such that

IRe(A(t)y, y)l :S niiYII 2 , y E D(A(t)), t E [0, T].

(H5) Y c D(A(t)), t E [0, T].

(H6) There exists a constant f3 2:: a such that

IRe(A(t)x, Sx) I ::; f3IIS 112 xll 2 , x E D(S), t E [0, T].

(H7) A(t) is continuous in [0, T] in the norm of .C(Y, X).

Proposition 7. 6 Let A and S be densely defined closed linear operators

in a Banach space X such that D(S) c D(A) and D(S*) is dense in X*.
Suppose that
(i) there exists a complex number~ such that for sufficiently large positive
integer n

R(~S+A+~) =X,
and hence for any y EX there exists a sequence {xn} C D(S) such that

1
-Sxn
n
+ Axn + ~Xn = y, (7.87)

(ii) for every y E X there exists a sequence {xn} C D(S) satisfying (7.87)
such that {xn}, {Axn} are bounded.
Then (A+ ~)D(S) and hence R(A + ~) is dense in X.
310 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Proof. Let f EX* be such that for any x E D(S), ((A+ .;)x, f) = 0. Let y
be an arbitrary element of X and a sequence { Xn} be as above. Then

(y, f)= ( ~SXn + Axn + .;xn, f) = ~(Sxn, f).

1 1
Iff E D(S*), then -(Sxn.J) = -(xn, S* f)----+ 0 as n----+ oo. Since D(S*)
n n
is dense in X* and { ~Sxn} is bounded, ~(Sxn, f)----. 0 for any f E X*.
This implies (y, f)= 0, and hence f = 0.

Definition 7. 3 A linear operator A in a Hilbert space X is called accretive

if
Re(Ax, x) 2:: 0 (7.88)
for any x E D(A). An accretive operator A is called m-accretive if R(fJ, +
A)= X for some fJ, > 0. If -A is accretive, then A is called dissipative, and
if -A ism-accretive, then A is called m-dissipative.

From (7.88) it follows that

/I(A + >.)xll 2:: Re>-llxll·

Hence if A is accretive and Re>. > 0, then A + >. has a continuous inverse.
Therefore an m-accretive operator is closed. If A is accretive and R(A +f.L) =
X, fJ, > 0, then -fJ, E p(A), and y = (>. + A)x is equivalent to

Consequently with the aid of a fixed point theorem we can easily show that
an accretive operator A is m-accretive if and only if R(>. + A) = X for
any >. such that Re>. > 0. If A is m-accretive, then for any x E X and a
positive integer n we have lln(n + A)- 1xll ::; llxll· Hence some subsequence
of {n( n + A) - l x} converges weakly to some element z E X. Since

A(n + A)- 1 x = x- n(n + A)- 1 x,

(n + A)- 1x ----. 0 and A is closed, it follows that x = z E D(A). Therefore
D(A) is dense in X. Thus we have shown that A ism-accretive if and only
if -A generates a contraction C0 -semigroup: llctAII ::; 1. It is easily seen
that a bounded accrretive operator is m-accretive.
A subset Z of X is called a core for an operator A if Z C D(A) and A
coincides with the smallest closed extension of Alz, i.e. for any x E D(A)
there exists a sequence {xn} C Z such that Xn----+ x and Axn----+ Ax.
7. 7. EQUATIONS IN HILBERT SPACES 311

Propostion 7. 7 In Proposition 7.6 if A is accretive and~> 0, then A is

m-accretive and D(S) is a core for A.

Proof. By Proposition 7.6 R(A + ~) is dense. Since A+~ has a continuous

inverse and A is closed, R(A + ~) is closed. Therefore R(A + ~) = X,
and hence A is m-accretive. Let x E D(A). Then there exists a sequence
{Xn} C D(S) such that (A+ ~)xn --+ (A+ ~)x. Since A+~ has a bounded
inverse, Xn--+ x and hence Axn--+ Ax. This means that D(8) is a core for
A.

Lemma 7. 1 Let A be a li,;,ear closed operator in X such that A + a is

accretive for some a ~ 0. Let 8 be a nonnegative self-adjoint operator in
X such that D(8) c D(A). Suppose that there exist nonnegative constants
j3 and 'Y such that

Re(Ax, 8x) ~ -'YIIxll 2- f311xllll8xll

for any x E D(8). Then A+ a ism-accretive and D(8) is a core for A.

Proof. Let ~ > j3. Set E = ~- 1 • Then (A+ a) 1 + -;;,8 ( E)-1

is a bounded
linear operator for any positive integer n, and for any x E X we have

Re(((A+a)(1+~8)- 1 +~)x,x)
= Re ((A+a) (1+ ~8) - 1 x, (1+ ~8)- 1 x)
+~Re (A ( 1 + ~8) - 1 x, 8 ( 1 + ~8) - 1 x)
+~a ( ( 1 + ~8) - 1 x, 8 ( 1 + ~8) - 1 x) + ~llxll 2
2 -~ ['Y II ( 1+ ~8) -1 xll2 + /311 ( 1+ ~8) -1 xllll8 ( 1 + ~8) -1 xll]
+~llxll 2 2 ( ~- f3 - ~'Y) llxll 2·

Therefore if n is so large that~- j3- -;;,'

E
Y > 0, then (A+ a) ( 1 + -;;,8
E ) -1
+~
ism-accretive. Consequently
312 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

is surjective, and

1 ) -1 1
( -S+A+a+~ ::; ~ - f3 -eyn
/ .
n

Hence for any y EX there exists a sequence {xn} such that

(~S+A+a+~) Xn =y
and {xn} is bounded. Since

{ ~ Sxn} is bounded. Therefore in view of Proposition 7. 7 A + a is m-

accretive and D(S) is a core for A.
Since A+ a ism-accretive under the assumptions of Lemma 7.1, we have
p(A) :::::J { \ Re.A <-a}, and

jj(A+ .-\)- 1 11::; Re.-\1- a for Re.A >a. (7.89)

Furthermore -A generates a C0 -semigroup satisfying

(7.90)

Let s< = S(l + t:S)- 1 = c 1 (1- (1 + ES)- 1 ), E > 0, be the Yosida approxi-
mation of S.

Lemma 7. 2 Let A and S be as in Lemma 7.1. Suppose that there exists

a nonnegative constant f3 such that f3 2: a and

Re(Ax, Sx) 2: -f3(x, Sx), x E D(S). (7.91)

Then
Re(Ax, S<x) 2: -f3(x, S<x), x E D(A). (7.92)
7. 7. EQUATIONS IN HILBERT SPACES 313

Proof. For x E D(A)

(Ax, Sex)= (Ax- A(1 + E8)- 1x, Sex)+ (A(1 + E8)- 1x, Sex)
= E(ASex, Sex)+ (A(1 + ES)- 1x, 8(1 + E8)- 1x).

Hence

Re(Ax, Sex) ~ -miiBexll 2 - ,8((1 + E8)- 1x, 8(1 + E8)- 1x)

= -aEIIBexll 2 - ,8((1 + E8)- 1x, Sex)
= (,8- a)EIIBexll 2 - ,B(x, Sex)~ -,B(x, Sex).

In what follows S will always denote a nonnegative selfadjoint operator

andY the space defined by (7.86).
Lemma 7. 3 Let A and S be as in Lemma 7.1. Suppose that (7.91) is
satisfied. Then Y is -A-admissible. Furthermore for any y E Y we have
1181/2(1 + >.A)-1YII s (1- >.,e)-111S1/2YII, 0 <A< ,e-1, (7.93)
IIS1f2e-tAYII S e~'tii81/2YII, t ~ 0. (7.94)

Proof. Let y E D(S112) and y(>.) be the solution of

y(>.) + >.Ay(>.) = y, 0 < >. < a- 1 . (7.95)

Let Be, E > 0, be the Yosida approximation of S. Then with the aid of (7.92)
we get

11Bi 12 y(>.)ll 2 = (y(>.), Bey(>.))= (y- >.Ay(>.), Bey(>.))

S Re(s;l 2y, s;l 2y(..\)) + ,8..\(y(>.), Bey(>.))
s IIBi 12 YIIIIBi 12 Y(>-)II + .B>-11Bi 12 Y(>-)II 2·

This implies
(7.96)

for 0 < >. < ,e- 1 S a- 1. Hence {si1 2y(>.)} is bounded as E-+ 0, and so it
contains a weakly convergent subsequence. Since (1 + ES)- 112y(>.)-+ y(..\)
as E-+ 0 and si 12 = 8 112(1 + ES)- 112, it follows that y(>.) E D(S112) and

s1f2y(>.) = w-lim
e->0
s;f2y(>.).

From (7.95) and (7.96) it follows that Y is invariant under (1 + ..\A)- 1 and
(7.93) holds. A repeating application of (7.93) yields

1181/2(1 + >.A)-nYII S (1- ,B>.)-nlls1/2YII (7.97)

314 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

for any positive integer n. Let t > 0. If n is so large that n > {3t, then we
get from (7.97)

Letting n---+ oo we conclude that Y is invariant under e-tA and (7.94) holds.
In order to prove that Y is -A-admissible it remains to verify that for
YE Y
IISlf2etAY- sl/2YII ---+ 0 (7.98)
as t---+ 0. With the aid of (7.94) we can easily show that
Slf2y = W-lim Slf2e-tAy. (7.99)
t--+0

Hence
IIB112 YII $ li~_}~f IIB 112 e-tAYII·
On the other hand we see from (7.94)

lim sup IIB112 e-tAYII $ IIB 112 YII·

t--+0

Therefore
(7.100)

We conclude (7.98) from (7.99) and (7.100).

From (7.93),(7.94) and {3 ~ a we obtain

11(1 + ,\A)- 1IIY $ (1- {3,\)-1, 0 < ,\ < {3-1, (7.101)

lie-tAllY$ ef3t, t ~ 0. (7.102)

We are going to prove the following theorem.

Theorem 7. 7 Under the assumptions (H4),(H5),(H6),(H7) there exists a
unique evolution operator U(t, s), (s, t) E [0, T] x [0, T], for the initial value
problem (7.18),(7.19) satisfying (7.20),(7.21),(7.22),(7.24),(7.48),(7,49) for
(s, t) E [0, T] x [0, T]. For each x E Y, f E C([O, T]; Y), s E [0, T]

u(t) = U(t, s)x + 1t U(t, T)f(T)dT (7.103)

is a unique Y-valued solution of (7.18) in [0, T] satisfying the initial condi-

tion u(s) = x.
7. 7. EQUATIONS IN HILBERT SPACES 315

In view of Lemmas 7.1, 7.3 and (7.101),(7.102), {A(t);t E [0, T]} is stable
in X with stability constants 1, a, Y is A(t)-admissible for each t E [0, T]
and {Ay(t); t E [0, T]} is stable with stability constants 1, (3. Hence, we
can apply Theorem 7.1 to find that there exists a unique evolution operator
U(t, s), 0::; s ::; t ::; T, satisfying (7.20),(7.21),(7.22),(7.23),(7.24) with 1,
a in place of M, (3. Let {Un(t, s)} be the approximating sequence defind
by (7.27). Then for each x E X, {Un(t, s)x} converges to U(t, s)x in X
uniformly in 0::; s ::; t::; T, and in view of (7.30),(7.31),(7.94) we have
IIUn(t, s)ll ::=; e'x(t-s), IIUn(t, s)IIY ::=; ef3(t-s), (7.104)
IIS1/2Un(t, s)yll::; ef3(t-s)IIS1/2YII for y E Y. (7.105)
Lemma 7. 4 Let y E Y. Then
(i) U(t, s)Y C Y and
IIU(t, s)IIY::; ef3(t-s), (7.106)
IIS1/2u(t, s)yll::; ef3Ct-s)IIS1/2YII, (7.107)
(ii) S 112 U(t, s)y is weakly continuous in 0::; s::; t::; T,
(iii) for to E [0, T]

S 112 U(t, s)y---+ S 112 y as (t, s)---+ (to, to),

(iv) fortE (0, T], U(t, ·)y E C([O, t]; Y),
(v) for s E [0, T), S 112 U(·, s)y is right continuous in [s, T).
Proof. (i) and (ii) can be established by means of the standard argument
with the aid of (7.104),(7.105).
(iii) In view of (ii) it suffices to show that

(7.108)

as (t, s)---+ (to, to). Since S 112 y =w-limcs,t)->(to,to) S 112 U(t, s)y we have

IIS 112yll ::; liminf IIS 112U(t, s)yll.

(s,t)->(to,to)

On the other hand in view of (i)

limsup IIS 112U(t,s)yll::; IIS 112yll·

(s,t)->(to,to)

Hence (7.108) follows.

(iv) Let y E Y, 0 ::; s ::; s' ::; t ::; T. Then as s' ---+ s or s ---+ s'

IIS 112U(t, s')y- S 112 U(t, s)yll

= IIS 112U(t, s')(y- U(s', s)y)ll::; ef3(t-s')IIS 112 (y- U(s', s)y)ll---+ 0
316 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

by virtue of (7.107) and (iii).

(v) Let to E [s, T). Then it follows from (iii) that as t Lto
IIS 112 U(t, s)y- S 112 U(to, s)y\\ = IIS 112 (U(t, to) - 1)U(to, s)yli -+ 0.
Lemma 7. 5 Let y E Y. Then
(i) we have
A(t)U(t, s)y = w- lim An(t)Un(t, s)y,
n---+oo
0:::; s:::; t:::; T, (7.109)

where {An(t)} is the approximating sequence defined by (7.25), and

\\A(t)U(t, s)yli:::; Mefi(t-s)\\Y\\Y, (7.110)
where M = sup{\\A(t)\k(Y,x); t E [0, T]},
(ii) A(t)U(t, s)y is weakly continuous in 0:::; s:::; t:::; T,
(iii) for each s E [0, T), A(·)U(·, s)y is right continuous in [s, T].
Proof. (i) In view of (7.26), (7.104)
\\An(t)Un(t, s)y- A(t)Un(t, s)yli
:::; \\An(t)- A(t)\\c(Y,x)\\Un(t, s)yliy-+ 0.
Hence in order to prove (7.109) it suffices to show that A(t)Un(t, s)y -+
A(t)U(t, s)y weakly in X. This can be shown with the aid of
\\A(t)Un(t, s)y\\:::; MIIUn(t, s)y\\Y:::; MefiTIIY\\Y
and the usual argument. (7.110) follows from (7.109) and
\\An(t)Un(t,s)y\\:::; Mefi(t-s)\\Y\\Y· (7.111)
(ii) Let y E Y and 0:::; s 0 :::; t0 :::; T. Then
\\A(t)U(t, s)y- A(to)U(t, s)y\\:::; \\A(t)- A(to)\ic(Y,x)\\U(t, s)yliy-+ 0
(7.112)
as t -+ t 0 • Hence it suffices to show
A(to)U(t, s)y-+ A(to)U(to, so)Y
weakly in X as (s, t)-+ (to, s0 ). But this follows from Lemma 7.4 (ii) since
A(to)(1 + S)- 112 E C(X).
(iii) Let t 0 E [s, T). Then in view of (7.112) it suffices to show that
A(t0 )U(t0 , s)y = limA(to)U(t, s)y. (7.113)
tlto
Since
\\A(to)(U(t, s)y- U(t 0 , s)y)\\:::; M\\U(t, s)y- U(to, s)yjjy,
(7.113) follows from Lemma 7.4 (v).
We followed T. Kato [88; section 5] in the proof of the above two lemmas.
7.7. EQUATIONS IN HILBERT SPACES 317

Lemma 7. 6 Let y E Y. Then

(i) for each s E [0, T) and z EX, (U(·, s)y, z) E C 1 ([s, T]) and
a
at (U(t, s)y, z) = (A(t)U(t, s)y, z),

(ii) for each s E [0, T), A(·)U(·, s)y is Bochner integrable in [s, T], and

U(t, s)y = y- 1t A(r)U(r, s)ydr, t E [s, T], (7.114)

(iii) for each s E [0, T), U(·, s)y is right differentiable in [s, T), and

( :t) + U(t, s)y = A(t)U(t, s)y, t E [s, T),

(iv)
a
at U(t, s)y = A(t)U(t, s)y a.e. t E (s, T).

Proof. (i) Letting n ---+ oo in

(Un(t, s)y, z) = (y, z) + 1t (An(r)Un(r, s)y, z)dr,

we obtain with the aid of Lemma 7.5 (i) and (7.111)

(U(t, s)y, z) = (y, z) + 1t (A(r)U(r, s)y, z)dr. (7.115)

Hence the assertion follows from Lemma 7.5 (ii).

(ii) Since A(·)U(·, s)y is weakly continuous in [s, T], it is Bochner integrable
there. Therefore (7.114) follows from (7.115) (see K. Yosida [166], p.133).
(iii) is a consequence of (7.114) and Lemma 7.5 (iii).
(iv) is a consequence of (7.114), since the assertion of Theorem 1.2 remains
valid for functions with values in a Banach space.
With the aid of Lemma 7.2 we obtain
IRe(A(t)x, S<x)l ~ .BIIS 1 12 xll 2 , x E D(A(t)), t E [0, T], E > 0. (7.116)
Lemma 7. 7 We have
(i) for s E [0, T) andy E Y, U(·, s)y E C([s, T]; Y),
(ii) U(t, s) is strongly continuous in 0 ~ s ~ t ~ T to ,C(Y),
(iii) for s E [0, T) andy E Y, U(·, s)y E C 1 ([s, T]; X) and
a
at U(t, s)y = A(t)U(t, s)y.
318 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

Proof. (i) Since U(·, s)y E C([s, T]; X) it suffices to show that S 112U(·, s)y E
C([s, T]; X). Let to E [s, T]. Since S 112 U(·, s)y is weakly continuous in [s, T]
by virtue of Lemma 7.4 (ii), it remains to show that

(7.117)

In view of Lemma 7.6 (iv) we have forE> 0

&
or IIS11 2U(r, s)yll 2 = 2Re(A(r)U(r, s)y, S,U(r, s)y) a.e. (s, t).

Integrating both sides from t 0 tot and using (7.116),(7.107)

I11B1" 12 U(t, s)yll 2 -IIS1" 12 U(to, s)yll 2 1

= 211: Re(A(r)U(r, s)y, S,U(r, s)y)drl

:$ 2{311: IISlf2U(r, s)yll2drl :$ 2{3e2.6TIIS1/2YII21t- tal·

Letting E - t 0 we obtain

lllsl/2u(t, s)yll2 -llsl/2u(to, s)yll21::; 2f3e2.6TIIS1/2YII21t- tal,

from which (7.117) follows.

(ii) Let 0 :$so < to :$ T. Choosing a E (so, to) we write for (s, t) sufficiently
close to (so, to)
U(t, s)y- U(to, so)Y
= U(t, a)(U(a, s)y- U(a, so)Y) + (U(t, a)- U(to, a))U(a, so)y.
Hence the assertion follows from (i) and Lemma 7.4 (i),(iv).
(iii) is a direct consequence of (ii) and Lemma 7.6.
So far we have proved that there exists a unique evolution operator sat-
isfying (7.20),(7.21),(7.22),(7.24),(7.48),(7.49) in 0 :$ s :$ t :$ T. We
can directly construct the desired evolution operator U(t, s) for (s, t) E
[0, T] x [0, T] by defining the approximating sequence {Un(t, s)} for (s, t) E
[0, T] x [0, T]. However, we can also define U(t, s) for 0 :$ t :$ s :$Tin the
following manner. Since { -A(T- t); t E [0, T]} satisfies (H4),(H5),(H6),
(H7), we can construct an evolution operator V(t, s), 0 :$ s :$ t :$ T, for the
equation
du(t)jdt = -A(T- t)u(t)
as above. Set
U(t, s) = V(T- t, T- s)
7. 7. EQUATIONS IN HILBERT SPACES 319

for 0 ::; t ::; s ::; T. Then for y E Y

a
at U(t, s)y = A(t)U(t, s)y, 0::; t::; s::; T.

Hence for 0 < s < t and y E Y we have

a
as U(t, s)U(s, t)y = 0.

Therefore we obtain
U(t, s)U(s, t) =I. (7.118)
With the aid of (7.118) we can easily show that U(t, s), (s, t) E [0, T] x [0, T]
satisfies the assertions of Theorem 7. 7.

Remark 7. 3 In an unpublished paper [119] N. Okazawa showed that the

conclusion of Theorem 7.4 remains valid replacing the norm continuity of
t f---t A(t) by its strong continuity. The method is to construct the evolu-
tion operator U(t, s) as the limit of the sequence of the evolution operators
Un(t, s) for the equations with A(t) replaced by the Yosida approximations
An(t) = A(t)(1- n- 1 A(t))- 1 . In the proof of the convergence of {Un(t, s)}
Okazawa uses the method of Y. Komura [97] and T. Kato [87]. Namely, if
we set
Unm(r, s) = Un(r, s)y- Um(r, s)y,
then
1 a 2
2ar llunm(r, s)ll
= Re(An(r)Un(r, s)y- Am(r)Um(r, s)y, Unm(r, s))
= Re(An(r)Un(r, s)y- Am(r)Um(r, s)y, Unm(r, s)- Wnm(r, s))
+Re(An(r)Un(r, s)y- Am(r)Um(r, s)y, Wnm(r, s)) =h + h
where

Wnm(r, s) = Jn(r)Un(r, s)y- Jm(r)Um(r, s)y,

Jn(r) = (1- n- 1 A(r))- 1 •

We have

I1 = Re( An(r)Un(r, s)y- Am(r)Um(r, s)y,

1
--An(r)Un(r, s)y + -Am(r)Um(r,
1 s)y)
n m
320 CHAPTER 7. HYPERBOLIC EVOLUTION EQUATIONS

and in view of the accretivity of a- A(r)

I2 ::::; nllwnm(r, s)ll 2 •

Combining this with

IIS112 Un(t, s)yll : : ; exp{,8(1- n- 1,a)- 2 (t- s)}IIS112 yll,

the convergence of {Un(t, s)y} is obtained.

7.8 Remark on Applications

F. Jo Massey III [110] applied the results of T. Kato [88] to the mixed
problem for the following symmetric hyperbolic system of partial differential
equations:

au ~ au
at + f;t aj(X, t) axj + b(x, t)u = f(x, t), x E n, t E [0, T], (70119)

u(x, 0) = </>(x), x E n, (70120)

u(x, t) E P(x, t), x E an, t E [0, T]. (7.121)
The unknown u = (u 1 , UN) is a real vector-valued function, the coef-
o •• ,

ficients, ai and b, are real N x N matrix-valued functions, and the ai are

symmetric. It is assumed that ai E C 2 (0 x [0, T]) and b E C 1 (0 X [0, T]).
n is a bounded open subset of Rm of class C 3 0
It is also assumed that the boundary matrix
m
an(x, t) = 2:nj(x)aj(x, t), x E an, t E [0, T],
j=l

is nonsingular on an X [0, T], where n = (nl, ... 'nm) is the exterior unit
normal to an.
The boundary subspace P(x, t) is a linear subspace of RN which varies in
a C 3 manner with (x, t) E an x [0, T]. P(x, t) is maximal nonnegative for
each x, t, i.e.
(an(x, t)u, u) ;::: 0, u E P(x, t),
7.8. REMARK ON APPLICATIONS 321

and P(x, t) is not a proper subset of any other subspace of RN having this
property.
Let A(t) be the smallest closed extension of the operator A 0 (t) defined
by
f)
:E aj(x, t) f):. + b(x, t)u
N
Ao(t) =
j=l J

with

D(A0 (t)) = H~, (0) = {u E H 1 (0)N; u(x) E P(x, t) a.e. 80}.

Massey proved the following theorems.

Theorem Ml There exists an isomorphism S(t) from H~, (0) onto L2 (0)N
such that
S(t)A(t)S(t)- 1 = A(t) + B(t),
where B(t) is a bounded linear operator from L 2 (0)N to itself.
Theorem M2 If P(x, t) = P(x) is independent oft, then S(t) in Theorem
M1 may be chosen so that S(-) E C 1 ([0, T); C(Hj,(O), L 2 (0)N)) and B(·) E
C([O, T]; £(L 2 (0)N).
Theorem M3 Suppose that¢ E H~0 (0) and f E C([O,T];H 1 (0)) so that
f(t) E H~, (0) for t E [0, T]. Then (7.119),(7.120),(7.121) has a unique
solution u such that u E C 1 ([0, T]; L 2 (0)N) and u(t) E H~, (0) fortE [0, T].

If P(x, t) is independent oft so that the conclusion of Theorem M2 is true,

then Theorem 7.3 (or Theorem 6.1 of T. Kato [88]) may be applied to
the present family {A(t)} taking Y = H~(O). Moreover Massey shows
that the general case where P(x, t) varies with t may be reduced to the
case P(x, t) = P(x, 0) by an orthogonal transformation of the dependent
variables.
There are also a number of applications to nonlinear hyperbolic equations,
T. Kato [90),[91),[92), K. Kobayashi and N. Sanekata [96), N. Okazawa and
A. Unai [121), N. Sanekata [131] and also see the Reference of [92).
Chapter 8

Retarded Functional
Differential Equations

This chapter is concerned with the following retarded functional differential

equations

! u(t) = Aou(t) + A1u(t- h)+ j_oh a(s)A2u(t + s)ds + f(t),

tE[O,T], (8.1)

u(O) =g0 , u(s) =g 1(s), s E [-h,O) (8.2)

in a Hilbert space. First following G. DiBlasio, K. Kunisch and E. Sinestrari
[51] the solvability of (8.1), (8.2) is described. After that we state the control
theory by C. Bernier and A. Manitius [20], M. C. Delfour and A. Manitius
[44],[45], A. Manitius [107],[108],[109] etc. for equations in finite dimensional
spaces and by S. Nakagiri [112], J.-M. Jeong, S. Nakagiri and H. Tanabe [82],
S. Nakagiri and H. Tanabe [114], etc. for equations in infinite dimensional
spaces.

8.1 Maximal Regularity Result

We begin with a maximal regularity result concerning equations in a Hilbert
space. Let A be a densely defined closed linear operator which generates
an analytic semigroup in a Hilbert space H. Hence there exist positive con-
stants M, Co and an angle() E (7r/2, 1r] such that for A. E :E ={A; I arg.A\ <
o, \.AI> Co}
1\.AR(.A,A)\\ ~ M, (8.3)

323
324 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

where R(.A,A) =(.A- A)- 1 . In this section the semigroup generated by A

is denoted by U(t): U(t) =etA, and we use the notation

(U * f)(t) =lot U(t- s)f(s)ds.

We always endow D(A) with the graph norm of A denoted by I · IID(A) so
that D(A) is a Hilbert space. The following proposition is due to J. L. Lions
and E. Magenes (99].
Proposition 8. 1 If u = U * f, f E L 2 (0, T; H), then

u E L 2 (0, T; D(A)) n W 1•2 (0, T; H), (8.4)

and hence
u E C((O, T]; (D(A), Hh;z,z). (8.5)
There exists a constant c1 such that

lluii£2(0,T;D(A))nW1,2(0,T;H) 5 ClllfiiP(O,T;H) · (8.6)

We have

du(t)jdt=Au(t)+f(t) a.e. (O,T), (8.7)

u(O) = 0. (8.8)

Furthermore u is a unique function satisfying (8.4), (8.7), (8.9).

Proof. First we note that (8.5) follows from (8.4) in view of (1.18). We
begin with the proof of the uniqueness. Let u satisfy (8.4),(8. 7),(8.8) and
0 < t < T. If we integrate the equality
(fJjfJs)(U(t- s)u(s)) = U(t- s)(u'(s)- Au(s)) = U(t- s)f(s)

from 0 tot, we obtain u = U *f. However, if we want to avoid verifying

the absolute continuity of U(t- s)u(s) in (0, t), we use the mollifier Pn(t) =
np(nt), where 0 5 p E C 00 (-oo,oo),f~00 p(t)dt = 1,p(t) = 0 for ltl ~ 1.
Let

Un(t) =loT Pn(t- s)u(s)ds, fn(t) =loT Pn(t- s)f(s)ds.

Then by a standard argument we see that Un ----+ u in L 2 (0, T; D(A)) and

fn ----+ f in L 2 (0, T; H). Since

u~(t) = -Pn(t- T)u(T) +loT Pn(t- s)u'(s)ds,

8.1. MAXIMAL REGULARITY RESULT 325

we have u~---+ u' in L 2 (0, T-E; H) for any E > 0. Let 0 < t < T. Integrating
both sides of
(EJjos)(U(t- s)un(s)) = U(t- s)(u~(s)- Aun(s))
from 0 to t we get

un(t)- U(t)un(O) = lt U(t- s)(u~(s)- Aun(s))ds.

Letting n---+ oo we obtain

u(t) = lt U(t- s)(u'(s)- Au(s))ds = lt U(t- s)f(s)ds.

Clearly this equality also holds for t = T. Thus the proof of the uniqueness
is complete.
We denote the Laplace transform of a function f defined in ( -oo, oo) and
taking values in H by]:

](>.) = rrc
1
v2n
100
-oo
e-t>. f(t)dt.

Since for real variables .;, 71 ](.; + iry) is the Fourier transform of e-t~ f(t),

1:
we have in view of Plancherel's theorem

Ill<.;+ iry)ll 2d'f/ = 1: e- 2t~llf(t)ll 2 dt.

Let f E L 2 (0, T; H). We extend f to ( -oo, oo) putting f(t) = 0 outside
[0, T]. Let w(t) be the function such that
w(>.) = R(>.,A)j(>.), Re>. >Co.
Then for .; > Co

e-t~w(t) = - 1 -1 00
..ffff -00
eit11 w(.; + iry)dry

1
= . tn=
100 e' 't
71 R(,; + iry, A)J(.; + iry)dry.
A

v2n -oo

1: 1:
Hence using (8.3)

lle-t~w(t)II~CA)dt = A)](~+ iry)II~(A)d'f/

: ; (~ y1: 1:
IIR(.; + iry,

II](.;+ iry)ll 2dry = ( ~y e-u~llf(t)ll 2 dt

= ( ~) 21T e-2t~llf(t)ll2dt.
326 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Since the rightmost side goes to 0 as ~ ~ +oo, we see that w(t) = 0 for
t < 0. Also we have
wE L?oc( -oo, oo; D(A)). (8.9)
Differentiating both sides of

w(t) =- 1 -1
..j2ii
00

-00
et<~+i77) R(~ + i'TJ, A)](~+ i'TJ)d'TJ
we get

~
ddt w(t) = y27r 1 et<~+i7J)(~ + i'TJ)R(~ +
00

-00
i'TJ, A)](~+ i'TJ)d'TJ. (8.10)

1: 1:
Hence

2
II(~+ i'TJ)R(~ + i'TJ, A)](~+ i'TJ)II 2d'f/
J:
lle-t.;! w(t)ll dt =
::; M2 II](~+ i'TJ)II2d'f/ = M21T e-2t~llf(t)112dt.
(
Consequently w E W[~; -oo, oo; H). Combining this with (8.9) we get

wE C( -oo, oo; (D(A), H)l/2,2)·

From (8.10) and

Aw(t) = - 1
..j2ii
-1 00

-00
et<~+i7J) AR(~ + i'TJ, A)j(~ + i'TJ)d'TJ,
it follows that

!!:_w(t) - Aw(t)
dt
= - 1-
..j2ii
Joo et<~+i7J) ](~ + i'TJ)d'TJ = f(t).
-00

Therefore if we define u as the restriction of w to [0, T], then u satisfies

(8.4),(8. 7), (8.8).
Proposition 8. 2 If u = U(· )x, x E (D(A), Hh;2,2, then

u E L 2 (0, T; D(A)) n W 1 •2 (0, T; H),

and there exists a constant c2 such that

lluiiP(O,T;D(A))nW 1 •2(0,T;H) ::; c2llxii(D(A),Hh;2,2" (8.11)

8.1. MAXIMAL REGULARITY RESULT 327

Proof. This is a direct consequence of the definition of (D(A), Hh; 2 , 2 and

Theorem 1.10.
Combining Propositions 8.1 and 8.2 we obtain

Proposition 8. 3 Iff E £ 2(0, T; H) and x E (D(A), Hh; 2,2, then the

function u defined by
u=U(·)x+U*f
belongs to £ 2 (0, T; D(A)) n W 1 •2 (0, T; H) and satisfies
du(t)jdt = Au(t) + f(t) a. e. (0, T), (8.12)
u(O) = x, (8.13)
lluii£2(0,T;D(A))nWt,2(0,T;H)
::::; cdfiiPco,T;H) + c2llxllcncA),Hh 12,2" (8.14)

We shall need the following regularity result when the data are more regular.

Proposition 8. 4 (i) Iff E W 1 •2 (0, T; H), x E D(A) and Ax+ f(O) E

(D(A), Hh; 2 ,2 , then the solution u of (8.12),(8.13) satisfies

u E W 1 •2(0, T; D(A)) n W 2•2(0, T; H) C C 1 ([0, T]; (D(A), Hh;2,2),

llu'll £2(0,T;D(A) )nWt,2(0,T;H)
::::; clllf'IIPco,T;H) + c2IIAx + f(O)IIcncA),Hh 12,2 (8.15)

and we have
du(t)/dt = U(·)(Ax + f(O)) + U * f'. (8.16)

(ii) Iff E £ 2(0, T; D(A)), x E D(A) and Ax E (D(A), H)1; 2,2, then the
solution u of (8.12),(8.13) satisfies

u E £ 2 (0, T; D(A2 )) n W 1 •2 (0, T; D(A)),

u, Au E C([O, T]; (D(A), Hh;2,2),
IIAull L 2(0,T;D(A))nW 1 • 2(0,T;H)
::::; cliiAfiiPco,T;H) + eziiAxllcn(A),Hht2,2' (8.17)

and we have
Au(·)= U(·)Ax + U * Af(·). (8.18)

Proof. (i) Let Un} be a sequence in C 1 ([0, T]; H) such that fn ----> f in
W 1 •2 (0, T; H). Set
Un = U(·)x + U * fn·
328 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Differentiating both sides of

Un(t) = U(t)x + lt U(s)fn(t- s)ds

we get
u~(t) = U(t)(Ax + fn(O)) + lt U(t- s)f~(s)ds.
As n---+ oo
Un---+ u, u~---+ U(·)(Ax + f(O)) + U * !'
in T; H). Consequently u E
L 2 (0, T; H) and (8.16) holds. Hence
W 1 •2 (0,
the remaining part of the assertion is an easy consequence of Proposition
8.3.
(ii) It is easy to verify (8.18). Hence we complete the proof applying
Proposition 8.3.

8.2 Assumptions and Notations

Let H be a Hilbert space with norm II · II. We state the assumptions.
(R1) A 0 is a densely defined closed linear operator and generates an analytic

=
semigroup in H. Hence there exist positive constants M, Co and an angle
() E (11"/2,11"] such that p(Ao) :::> :E {>.; JargA.J < B, JA.J >Co} and for A. E :E
the following inequality holds

IIA.R(A., Ao)ll ::; M. (8.19)

(R2) A 1 , A2 are bounded linear operators from D(Ao) to H.
(R3) a(·) is a real valued function belonging to £ 2 (-h,O), where his a
positive number.
Let T be a fixed positive number. For a function u defined in [-h, T] and
taking values in H we set
ut(s) = u(t + s) s E [-h, OJ, t E [0, T].
Hence for each t E [0, T] Ut is a function with values in H defined in [-h, OJ.
For u E L 2 ( -h, 0; D(Ao)) set

Lu = loh a(s)A 2 u(s)ds. (8.20)

Then L E £(L 2 ( -h, 0; D(Ao) ), H) and

IlLII ::; llall£2(-h,o)IIA2il.ccncAo),H)· (8.21)

8.3. SOLVABILITY AND REGULARITY 329

Since
Lut = /_: a(s)A2ut(s)ds = j_oh a(s)A2u(t + s)ds,

the problem (8.1),(8.2) is rewritten as

d
dt u(t) = Aou(t) + A1u(t- h)+ Lut + f(t), t E [0, T], (8.22)
u(O) = g0 , u(s) = g 1(s) s E [-h, 0). (8.23)

We are going to find a solution of (8.22),(8.23) in the space

£ 2 ( -h, T; D(Ao)) n W 1' 2 (0, T; H) c C([O, T]; (D(Ao), Hh;2,2).

Hence we assume that

g0 E (D(Ao),Hh;2,2, g 1 E £ 2(-h,O;D(Ao)), f E L 2(0,T;H).

8.3 Solvability and Regularity

In this section we prove the existence and uniqueness of solutions of (8.22),
(8.23) together with some regularity property.

Theorem 8. 1 Suppose that the assumptions (R1),(R2),(R3) are satisfied.

Then for any g0 E (D(Ao), Hh;2, 2 , g 1 E £ 2 ( -h, 0; D(Ao)), f E £ 2 (0, T; H)
there exists a unique solution u of (8.1),(8.2) belonging to L 2 ( -h, T; D(Ao))n
W 1 ,2 (0, T; H). Moreover there exists a constant c3 such that the following
inequality holds:

Ilull £2(0,T;D(A 0 ))nW 1·2(0,T;H)

~ c3 (llg0 llcvcAo),Hh 12,2 + llg1 IIL2(-h,O;D(Ao)) + II!IIL2(0,T;H)) · (8.24)

Proof. By Proposition 8.1 for u(t) = J~ eCt-s)Ao f(s)ds we have

lluiiP(O,T;D(Ao))nW1,2(0,T;H) ~ clllfiiP(o,T;H), (8.25)

and by Proposition 8.2 for x = etAox we have

llull £2(0,T;D(Ao))nW 1·2(0,T;H) ~ c2llxiiCD(Ao),Hht2,2' (8.26)

We note that the constants c1 , c2 are taken independently ofT in a bounded

interval. For the sake of simplicity we denote the operator norm all by II ·II·
330 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

First we consider the case 0 < T :::; h. Then

T
IIA191(·- h)lli2(0,T;H) = 111Alg1(t- h)ll 2dt
= 1
-h
T-h
IIA1gl(t)ll2dt:::; 1-h
0
IIA1gl(t)112dt

:::; IIA1II 2 j_oh 119 1(t) llt(Ao)dt = IIAlll 2ll91lli2(-h,O;D(Ao))'

Hence

Similarly

Hence
IILu.IIP(O,T;H) :::; IILIIVTIIull£2( -h,T;D(Ao))· (8.28)
For a given u E L 2 (0, T; D(A0 )) define

g\t) t E [-h, 0)
u(t) ={ (8.29)
u(t) tE [O,T]

Then u E L 2 ( -h, T; D(A0 )). Noting u(t- h)= g 1 (t- h) fortE (0, T] since
we are assuming T:::; h, we set for 0:::; t:::; T

(Su)(t) = etAog0 +lot e(t-s)Ao (A1g 1 (s- h)+ Lus + f(s)) ds. (8.30)

By Proposition 8.3 and (8.27),(8.28) Sis a mapping from L 2 (0, T; D(Ao))

into itself. We are going to apply a fixed point theorem to S. Let ii1, ii2 be
two elements of L 2 (0, T; D(Ao)) and set

t E (-h, 0)
' i = 1,2. (8.31)
tE [O,T]
8.3. SOLVABILITY AND REGULARITY 331

Then
(Sih)(t)- (Su2)(t) =lot e<t-s)Ao L(uls- U2 8 )ds. (8.32)

In view of (8.25),(8.28)

IISu1- Su211£2(o,T;D(Ao)):::; c1IIL(u1.- u2.)11Pco,T;H)

:::; c1IILII Vrllu1. - u2.11 L~( -h,T;D(Ao))
= cliiLIIVTIIul- u2ii£2(o,T;D(Ao))· (8.33)

Hence if c1IILIIVT < 1, Sis a strict contraction and there exists a unique
fixed point u. The function defined as (8.29) is a unique solution of (8.1),(8.2).
The magnitude ofT depends only on c1 and IlLII, and is independent of the
initial values. Since in (0, T)

u(t) = etAogO +lot e<t-s)Ao (Algl(s- h)+ Lus + f(s)) ds,

we have by (8.25),(8.26),(8.27),(8.28)

lluii£2(0,T;D(Ao))nW 1 ·~(0,T;H)
:::; c2llg 0 llcvcAo),Hh 1 ~.~ + c1IIA1g 1(·- h)+ Lu. + !IIP(o,T;H)
:::; c2llg 0 llcvcAo),Hh 1 ~.~ + cliiAlllllg 1II£2C-h,O;D(Ao))
+cliiLIIVTIIuii£2(-h,T;D(Ao)) + cliiJIIL~(o,T;H)
:::; c2llg 0ll (D(Ao),Hht~.~ + cliiA1IIIIg 1ll£2( -h,O;D(Ao))
+cliiLIIVrllg 1IIL2C-h,o;D(Ao)) + cliiLIIVTIIull£2(o,T;D(Ao))
+clllfiiPco,T;H)·
Hence

iiuiiL~(O,T;D(Ao))nW 1 ·~(0,T;H)

:::; 1 _ c 1 I~LIIVT [c2llg0 llcvcAo),Hh 1 ~.~

+c1 (IIA1II + IlLII VT) llglll£2( -h,O;D(Ao)) + clllfii£2(0,T;H)] · (8.34)

Next we continue the solution to an arbitrary time interval. Suppose that

t1 > 0, t2 > 0 and u is a solution in [-h, t1 + t 2]. If t E [t1, t1 + t2]

u(t) = etAog0 + (!ott+ 1:) e<t-s)Ao (A 1 u(s- h)+ Lu 8 + f(s)) ds

332 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

= e<t-tt)Ao [ettAogO + 1tt e<tl-s)Ao (Alu(s- h)+ Lus + J(s)) ds]

+ lt e<t-s)Ao (A1u(s- h)+ Lu + f(s)) ds

8
tt
= e<t-tt)Aou(tl)
rt-t1
+ Jo e<t-tt-s)Ao (A1u(t1 + s- h)+ Lutt+s + j(t1 + s)) ds.
Hence fortE [0, tz]

u(t1 +t) = etAou(t1)+ 1t e<t-s)Ao (A1 u(t1 + s- h)+ Lutt +s + j(t1 + s)) ds.
This shows that u(t 1+t) is a solution in [-h, t 2 ] with initial values u(t1), ut 1
and the inhomogeneous term j(t1 + t).
Conversely suppose that v is a solution in [-h, t 1 ] and w is a solution in
[-h,t2 ] with initial values v(tl),vtt and the inhomogeneous term j(t1 + ·).
Set
v(t) -h::::; t::::; tl
u(t) ={ (8.35)
w(t- t1) t1 ::::; t::::; t1 + tz
We are going to verify that u is a solution in [-h, t1 + tz]. If -h::::; t < 0,
w(t) = Vtt (t) = v(t1 + t) = u(t1 + t), and if 0 ::; t ::::; tz, w(t) = u(t1 + t).
Hence
w(t) = u(t 1 +t) for - h::::; t::::; t 2 • (8.36)
If t1 < t::::; h + tz, using (8.36) we have
u(t) = w(t- t1) = e(t-tt)Aov(t1)
rt-tt
+ Jo e(t-tt -s)Ao (A1w(s- h)+ Lw8 + j(t1 + s)) ds

= e<t-tt)Ao [ettAogO + 1tl e<tt-s)Ao (Alv(s- h)+ Lvs + f(s)) ds]

t-tt e<t-t 1-s)Ao (A1u(t1 + s- h)+ LUt1+s + j(t1 + s)) ds
+ Jo

= etAogO + Jotl e<t-s)Ao (Alu(s- h)+ Lus + f(s)) ds

+ lt e<t-s)Ao (A1u(s- h)+ Lu8 + f(s)) ds
tt
= etAogo + 1t e<t-s)Ao (Alu(s- h) +Lus + f(s)) ds.
8.3. SOLVABILITY AND REGULARITY 333

This implies that u is a solution in [-h, t 1 + t 2 ].

Therefore if we have a solution u in [-h, T] where 0 < T < (clll£11)- 2,
then choosing t 1 = t 2 =Tin the above we obtain a solution u* in [-h, T]
with the initial values u(T), uT. If we set u(t) = u* (t- T) in [T, 2T], then u
is a solution in [-h, 2T]. Continuing this process we obtain a global solution.
In the next section we shall need the following regularity result.
Theorem 8. 2 If g0 E D(Ao), g1 E W 1•2( -h, 0; D(A0 )), g1(0) = g0 ,
f E W 1 •2 (0, T; H) and
Aog 0 + A1g 1( -h)+ Lg 1 + f(O) E (D(Ao), Hh;2,2,
then the solution of (8,1),(8.2) obtained in Theorem 8.1 belongs to

W 1•2( -h, T; D(Ao)) n W 2•2(0, T; H) c C 1([0, T]; (D(Ao),Hh;2,2)·

Proof. First consider the case T::::; h, c1IILIIVT < 1. Set

K = {u E W 1 •2 (0, T; D(Ao)); u(O) = g 0 }.
We are going to show that the mapping S defined by (8.30) is a strict
contraction in K. Let u E K and define u by (8.29). Then from the
assumptions it follows that u E W 1 •2 ( -h, T; D(Ao)). Set
](s) = A1g 1(s- h)+ Lus + f(s).
Then j E W 1 •2 (0, T; H) and using (8.27), (8.28) we get

llfllwl.2(o,T;H) ::::; IIA1IIII9 1IIw1.2( -h,O;D(A 0 ))

+IlL II Vrllull W 1·2(-h,T;D(A0 )) + II /II W 1·2{0,T;H) ·
Since

we have in view of Proposition 8.4

SuE W 1•2(0, T; D(Ao)) n W 2•2(0, T; H).
It is clear that (Su)(O) = g0 • Hence Su E K and S maps K into itself.
Next, let u1, u 2 E K and define u 1, u 2 by (8.31). Since L(uw - u2o) =
L(g 1 - g 1 ) = 0, we get differentiating both sides of (8.32)
d
dt ((Su1)(t)- (Su2)(t))

=lot e(t-s)Ao :s L(uls- U2s)ds =lot e(t-s)Ao L(uis- u~s)ds.

334 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Therefore applying (8.33) we get

IISih - Su211~n.~co,T;D(Ao))
= IIBu1- Su211i~co,T;D(Ao)) + II(Su1- Su2)'11i~co,T;D(Ao))
::; ciiiLII 2T (11u1 - U2 lli~(-h,T;D(Ao)) +!lui - u~lli~c -h,T;D(Ao)))
= c~IIL1i 2 TIIu1- u211~n.~co,T;D(Ao))·
Consequently S is a strict contraction in K, and the solution u belongs
to W 1 ,2 (0, T; D(A 0 )). Using the argument by which we deduced (8.34) we
obtain

llu'IIL2(0,T;D(Ao))nWt·~(O,T;H)
1 [
::; 1- c111LIIVT c211Aog
0
+ A1g 1 (-h)+ Lg 1 + f(O)IIcncAo),Hht2.~
+c1(IIA1II + IILIIVT)IIg 11 IIL2(-h,O;D(Ao)) + c1llf'IIL2(0,T;H)] ·
Therefore we can proceed as in Theorem 8.1 to obtain the conclusion for an
arbitrary T > 0.

8.4 Solution Semigroup

In this section we consider the problem (8.1),(8.2) with j(t) =0:
d u(t) = Aou(t)
dt + A1u(t- h)+ 10
-h a(s)A2u(t + s)ds, t E [0, oo), (8.37)

u(O) = g 0 , u(s) = g 1(s) s E [-h, 0). (8.38)

Set
Z = (D(Ao), Hh; 2,2 x £ 2(-h, 0; D(Ao)) (8.39)
and for g = (g 0 , g 1 ) E Z
1/2
ligll~ = ( llg0 1i~D(Ao),Hht~.~ + llg 1 11i~(-h,O;D(Ao)) ) · (8.40)
Then Z is Hilbert space with norm (8.40). By virtue of Theorem 8.1 for
g = (g 0 , g 1) E Z there exists a unique solution u of (8.37),(8.38) in [-h, oo ),
and using (1.17) we have for each T > 0

Cr 1 tE[O,T]
max llu(t)II(D(Ao) Hht~ ~
' '
::; llull£2(-h,T;D(Ao))nwt.~(O,T;H) ::; CTIIgllz, (8.41)
8.4. SOLUTION SEMIGROUP 335

where CT is a positive constant depending on T.

We denote the solution of (8.37),(8.38) by u(t; g) = u(t; g0 , g 1 ). Set
S(t)g = (u(t; g), Ut(·; g)). (8.42)
Theorem 8. 3 {S(t); t 2 0} is a C0 -semigroup in Z.
Proof. In view of (8.41) for g = (g 0 ,g 1 ) E Z
IIS(t)gll~ = llu(t;g)IIZvcAo),H)t 12 ,2 + llut(·,g)lli2(-h,O;D(Ao))
= llu(t; g)IIZvcAo),Hh 12,2 + llu(·; g) lli 2Ct-h,t;D(Ao))
:::; llu(t; g)IIZvcAo),Hh 12,2 + llu(·; g)lli2c-h,t;D(Ao)) :S C'f(C'f + 1)llgll~·
Hence S(t) E £(Z, Z). For t, T E [0, oo)
IIS(t)g- S(r)gll~
= llu(t; g) - u(r; g) IIZD(Ao),H)t 12,2 + llut(·; g) - u,-(·; g)lli2(-h,O;D(Ao))
= llu( t; g) - u( r; g) IIZD(Ao ),Hh12,2

+ j_oh llu(t + s; g)- u(r + s; g)ll'tcAo)ds - t 0 as T --tt.

Consequently S(t)g is strongly continuous in [0, oo ). Choosing t = 0 in the

above we obtain that S(r)g - t gas T - t 0.
Let t 1 > 0, t 2 > 0. Let v be the solution of (8.37),(8.38) in [-h, t1] with
the initial values g = (g 0 ,g 1), and w be the solution in [-h, t 2 ] with the
initial values (v(tt), Vt 1 ). Then as was shown in the proof of Theorem 8.1
the function defined by (8.35) is the solution in [-h, t 1 +t 2 ] with the initial
values g 0 , g 1 , i.e.
v(t) = u(t; g) - h:::; t:::; it, w(t) = u(t; v(t1), Vt 1 ) - h:::; t:::; t2,
u(t) = u(t; g) -- h:::; t:::; t1 + t2.
Hence
u(tt + t2; g) = u(t1 + t2) = w(t2) = u(t2; v(t1), VtJ
and by (8.36) for s E [-h, 0]
u(t1 + t2 + s; g) = w(t2 + s) = u(t2 + s; v(t1), Vtt) = ut 2(s; v(t1), VtJ·
Therefore
S(t1 + t2)g = (u(t1 + t2; g), Utt +t 2(·;g))
= (u(t2; v( t1), Vtt ), Ut 2 (·; v(t1), VtJ) = S(t2)(v(tt), Vtt)
= S(t2)(u(t1; g), Utt (·;g))= S(t2)S(t1)g.
Thus we have shown S(t1 + t2) = S(t2)S(t1) and the proof is complete.
336 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Definition 8. 1 The semigrooup S(t) in Theorem 8.3 is called the solution

semigroup of (8.1),(8.2).

Theorem 8. 4 The infinitesimal generator A of the semigroup S(t) is given

by

D(A) = {g = (g 0 ,g 1) E Z;g 1 E W 1'2 (-h,O;D(Ao)),

g0 = g 1(0), Aog0 + A 1g 1( -h)+ Lg 1 E (D(Ao), Hh;2,2}, (8.43)
Ag = (Aog 0 + A1g 1( -h)+ Lg 1, g11 ). (8.44)

Proof. Let A be the operator defined by (8.43),(8.44), and A the infinitesi-

mal generator of S(t). We first show that
(i) S(t)D(A) c D(A) for any t 2 0,
(ii)D(A) is dense in Z,
A c A,
(iii)
(iv)A is closed in Z.
Proof of (i). Let g E D(A). In view of Theorem 8.2 we have for any T > 0

u(·; g) E W 1 ' 2( -h, T; D(Ao)) n W 2'2 (0, T; H)

c C 1 ([0, T]; (D(Ao), Hh;2,2). (8.45)

Hence for any t 2 0 u(t;g) E D(A0 ) and Ut(·;g) E W 1,2 (-h,O;D(Ao)).

Clearly Ut(O;g) = u(t;g). By the equation (8.37) satisfied by u(·;g) and
(8.45) we have

Aou(t; g)+ A1ut( -h; g)+ Lut(·; g)

= Aou(t; g)+ A1u(t- h; g)+ Lut(·;g) = u'(t;g) E (D(Ao), Hh;2,2·
Therefore S(t)g = (u(t; g), Ut(·; g)) E D(A).
Proof of (ii). Let g E Z. Since c 1 J0' S(t)gdt - t g in Z as E - t 0, it suffices

to show that J; S(t)gdt E D(A) forE> 0. Note that

l' S(t)gdt = (l' u(t;g)dt, l' ut(·;g)dt).

For -h::::; s < 0

! (l' ut(·,g)dt) (s) = :s 1' u(t + s; g)dt

= l' u'(t + s; g)dt = u(E + s; g) - u(s; g)= u(E + s; g)- g 1(s).

8.4. SOLUTION SEMIGROUP 337

Hence

0
= j_h IIAo(u(E + s; g)- g (s))ll 2ds 1

::::; 2 (j_oh IIAou(E + s; g)ll 2 ds + j_oh IIAog 1 (s)ll 2 ds) < oo.

Therefore J; ut(·; g)dt E W 1 •2 ( -h, 0; D(Ao)). Clearly

(if Ut(·;g)dt) (0) =if u(t;g)dt.

By the equation (8.37)

Ao if + (if
u(t; g)dt A1 Ut(·; g)dt) (-h)+ L if Ut(·; g)dt

= if + if
A0 u(t; g)dt A1 u(t- h; g)dt + ifL Ut(·; g)dt

=if (Aou(t; g)+ A1u(t- h; g)+ Lut(·; g)) dt

=if = u'(t; g)dt u(E; g)- u(O; g)= u(E; g)- g0 E (D(Ao), Hh;2,2·

Consequently J;
S(t)gdt E D(A).
Proof of (iii). Suppose that g E D(A). By virtue of Theorem 8.2 (8.45)
holds for any T > 0. Therefore as t ----> 0

l ~(u(t;g) -g 0) -u'(O;g)ll
(D(Ao),Hh/2,2

= II ~(u(t; g) - u(O; g)) - u' (0; g) II ---t 0. (8.46)

(D(Ao),H)t/2,2

Since
1 1
t (ut(·;g) -g 1 ) (s) = t(u(t+s;g) -u(s;g))
1 {1 d {1
=t}o dOu(Bt+s;g)dB= Jo u'(Bt+s;g)dO,
338 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

we have

o
1 -h t
II!
(ut(·;g) -g1) (s) -u'(s;g)ll2
D(Ao)
ds

=1°
-h Jo
I{ 1
(u'(Bt+s;g)-u'(s;g))d011
D(Ao)
ds
2

1
:::; 1° { iiu'(Ot+s;g) -u'(s;g)ii~(Ao)dOds
-hlo
1 0
= { 1iiu'(Ot+s;g)-u'(s;g)ii~(Ao)dsd0.
Jo -h

For each fixed 0 E (0, 1)

I= 1° llu'(Ot + s; g)- u'(s; g)II~(Ao)ds---+ 0

-h
as t ---+ 0. Furthermore

I:::; 2[ h llu'(Ot+s;g)II~(Ao)ds+2l: llu'(s;g)ii~(Ao)ds

0

:::; 2 [h liu'(s;g)ii~(Ao)ds + 2loh llg1(s)II~(Ao)ds

is bounded as t ---+ 0. Therefore

°
1 -h ~~~(ut(·;g)-g 1 )(s)-u'(s;g)ll 2D(Ao) ds---+0. (8.47)

Combining (8.46),(8.47) we obtain

~(S(t)g- g)=(~ (u(t;g)- g0 ), ~ (ut(·;g)- g1 ) )

---+ (u'(O; g), u'(·; g)) = ( Aog0 + A 1g 1( -h)+ Lgl, g 1') = Ag

in Z as t---+ 0. Therefore g E D(A), Ag = Ag.

Proof of (iv). Let gn = (g~, g~) be a sequence of elements of D(A) such
that gn---+ g = (g 0 ,g 1),Agn---+ w = (w 0 ,w 1) in Z. By the definition and
assumptions we have

g~ E W 1•2 ( -h, 0; D(Ao)), g~(O) = g~,

Aog~ + A1g~( -h)+ Lg~ E (D(Ao), Hh;2,2
8.4. SOLUTION SEMIGROUP 339

g~---+ g0 in (D(Ao), Hh; 2 ,2 , g~---+ g 1 in L 2 ( -h, 0; D(Ao)), (8.48)

Aog~ + A1g~ (-h)+ Lg~ ---+ w 0 in (D(Ao), Hh;z,z, (8.49)
gn1 I ---+ w1 m
. L2 ( - h , 0; D (A o)) . (8.50)
From (8.48),(8.50) it follows that g 1 E W 1 • 2 (-h,O;D(A 0 )),g~ ---+ g 1 in
W 1 •2 ( -h, 0; D(Ao)) and
g1' = w1. (8.51)
Therefore g~ = g~(O)---+ 1
g~( -h)---+ g (-h) in D(A 0 ). Combining this
g 1 (0),
with (8.48) we get g0 = g 1 (0), and g~---+ g0 in D(A0 ). Therefore
Aog~ + A1g~( -h)+ Lg~---+ Aog0 + A 1g 1 ( -h)+ Lg 1
in H. This and (8.49) implies
Aog 0 + A1g1( -h)+ Lg 1 = w 0 E (D(Ao), Hh;z,z. (8.52)
Therefore g E D(A), and by virtue of (8.51),(8.52) Ag = w. Thus the proof
of (i),(ii),(iii),(iv) is complete.
Let g E D(A). Then in view of (i) S(t)g E D(A). By (iii)
AS(t)g = AS(t)g = S(t)Ag = S(t)Ag,
which implies that AS(t)g is continuous in t E [0, oo), and AS(t)g
dS(t)gjdt. Combining this with (iv) we obtain

A lt S(T)gdT = lt AS(T)gdT = lt !s(T)gdT = S(T)g- g.

Let g be an arbitrary element of Z. Then by (ii) there exists a sequence
{gn} C D(A) such that gn---+ gin Z. Then

A it S(T)gndT = S(t)gn- gn---+ S(t)g- g, it S(T)gndT---+ it S(T)gdT.

Hence, in view of (iv)

lt S(T)gdT E D(A) and A lt S(T)gdT = S(t)g- g.

Therefore if g E D(A),

A!
t
t
Jo
S(T)gdT = S(t)g- g
t
---+ Ag, lit
-
t 0
S(T)gdT---+ g

as t ---+ 0. This implies that g E D(A) and Ag = Ag. Consequently we

conclude A = A.
340 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

8.5 Mild Solutions

Suppose that g = (g 0 ,g 1) E D(A) and f E W 1 •2 (0,T; (D(Ao),Hh;2,2)·
Then

g 1 E W 1•2( -h, 0; D(Ao)), g 0 = g 1(0),

Aog0 + A1g 1( -h)+ Lg 1 E (D(Ao), Hh;2,2, f(O) E (D(Ao), Hh;2,2·

Hence the assumptions of Theorem 8.2 are satisfied. Therefore the solution
u of (8.1),(8.2) satisfies

u E W 1•2( -h, T; D(Ao)) n W 2 •2 (0, T; H)

c C\[0, T]; (D(Ao), Hh;2,2). (8.53)

Set x(t) = (u(t), Ut) for t 2 0. In view of (8.53) u(t) E D(Ao), Ut E

W 1·2( -h, 0; D(Ao)). Clearly Ut(O) = u(t). By (8.1),(8.53)

Aou(t) + A1ut( -h)+ Lut = u'(t)- f(t) E (D(Ao), Hh;2,2·

Hence x(t) E D(A) and

Ax(t) = (Aou(t) + A1u(t- h)+ Lut, (ut)') = (u'(t)- f(t), (ut)'), (8.54)

where (ut)' is the function

d d . '
ds (ut)(s) = ds u(t + s) = u'(t + s) = (u')t(s). (8.55)

Lemma 8. 1 For u E W 1 •2 ( -h, T; H) we have

dutf dt = (u')t, t E [0, T].

Proof. For 0 ::; t ::; T, -h ::; s ::; 0

u(t + s)- u(s) = l t u 1 (T + s)dT = l t (u')r(s)dT = (lt (u')rdT) (s).

Hence
Ut- Uo = l t (u')rdT

from which the conclusion follows.

By Lemma 8.1 and (8.54),(8.55) we get

Ax(t) = (u'(t)- f(t),dutfdt) = x'(t)- (f(t),O).

8.5. MILD SOLUTIONS 341

Therefore if we put F(t) = (f(t), 0), x(t) satisfies

dx(t)jdt = Ax(t) + F(t), 0:::; t:::; T, (8.56)
x(O) =g. (8.57)
It is easy to see that x', Ax are continuous functions with values in Z in
[0, T]. Namely x is a strict solution of (8.56),(8.57).
Conversely suppose that x(t) = (x 0 (t), x 1 (t)) is a strict solution of (8.56),
(8.57). Since x(t) E D(A) we have
x 1(t) E W 1•2( -h, 0; D(Ao)), x 0 (t) = x\t)(O), (8.58)
Aox0 (t) + A1x1(t)(-h) + Lx 1(t) E (D(Ao),Hh;2,2· (8.59)
Furthermore
d
dtx 0 (t) = Aox0 (t) + A1x 1,it)( -h)+ Lx 1(t) + f(t), (8.60)

( !x 1 (t)) (s) = ! x 1 (t)(s). (8.61)

By (8.61) x 1(t)(s) is a function oft+ s, and we write x 1(t + s) instead of

x 1(t)(s). By virtue of the second part of (8.58) we have x 0 (t) = x 1(t) for
t ~ 0 in the new notation for x 1 • Therefore if we define the function u by
x 0 (t) t ~ 0,
u(t) = {
";.'.' g 1 (t) -h :::; t < 0,
-I:~'}
then wehaveu(t) = x 1(t).fort ~-h. Clearly (8.60) implies (8.1), and (8.57)
implies (8.2). Since x(t) = (u(t), ut) is a strict solution of (8.56),(8.57),

ll(u')tlli2(-h,O;D(Ao)) = /_: ll(u')t(s)llbcAo)ds

= jo
-h
llu'(t + s)llb(Ao)ds = lt
t-h
llu'(s)llbcAo)ds

is bounded in [0, T]. Therefore u E W 1·2( -h, T; D(Ao)). From the equa-
tion (8.1) it also follows that u E W 2 •2 (0, T; H). Hence u is a solution of
(8.1),(8.2) satisfying (8.53).
The strict solution of (8.56),(8.57) is represented by

u(t) = S(t)g + 1t S(t- r)F(r)dr. (8.62)

For arbitrary g E Z and F = (f, 0), f E L 2( -h, 0; (D(Ao), Hh; 2 ,2 ) the

right hand side of (8.62) is meaningful, and is called a mild solution of
(8.56),(8.57).
342 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

8.6 Regularly Accretive Operators

If we take as Ao in (8.1) a regularly accretive operator whose definition will
be given below, we can develop a semigroup treatment for (8.1),(8.2) in a
more easily handled space. Such a treatment is convenient in applications
to control problems. As a preparation we consider operators associated with
sesquilinear forms on Hilbert spaces.
Let H, V be two Hilbert spaces such that V is a dense subspace of H and
the imbedding V C H is continuous. We denote the innerproduct and norm
of H by(·,·) and 1·1 respectively. The norm of Vis denoted by 11·11· Then
by assumption there exists a positive constant Co such that

lui:::; Collull, u E V. (8.63)

In this and the following sections H*, V* stand for the sets of all conjugate
linear continuous functionals defined in H, V respectively. By assumption
the mapping which maps elements of H* to their restrictions to V is an
injection from H* to V*. Therefore identifying an element of H* and its
restriction to V we may consider H* C V*. Using Riesz's theorem we
identify H and H*. Hence we may consider

V c H c V*. (8.64)

The pairing between V and V* is also denoted by (·, ·): for l E V*, u E
V, (l, v) is the value of l at v. In particular if l E H, then (l, v) is the
innerproduct of l and v considered as elements of H. The norm of V* is
denoted by ll·lk
IIlii. = sup l(l, v)l/llvll· (8.65)
O#vEV
As is easily seen 11!11• :::; Colfl for any f E H. Hence the imbedding H C V*
is also continuous.
In what follows in this chapter we denote the complex number field by C.
Let a(u, v) be a sesquilinear form defined on V x V: for u, v E V a(u, v) E
C and

a(u1 + u2, v) = a(u1, v) + a(u2, v),

a(u, v1 + v2) = a(u, v1) + a(u, v2)
a(>.u, v) = >.a(u, v), a(u, >.v) = 5.a(u, v).
We assume that there exist positive constants C 1 , eo and a real number k
such that

la(u, v)l:::; Clllullllvll, u, v E V, (8.66)

Rea(u,u) ~ eollull 2 - klul 2 , u E V. (8.67)
8.6. REGULARLY ACCRETIVE OPERATORS 343

The inequality (8.67) is called Garding's inequality. In view of (8.66) for

each fixed u E V the mapping V 3 v 1--t a( u, v) E C is continuous and
conjugate linear. Therefore it defines an element Au E V* such that

(Au, v) = a(u, v), v E V. (8.68)

Since
I(Au, v)l = la(u, v)l::; Clllullllvll,
we have IIAull* ::; Clllull· Therefore A E C(V, V*) and IIAII ::; cl. Let A be
the realization of A in H:

D(A) = {u E V; Au E H}, Au= Au for u E D(A). (8.69)

Then it is well known that -A generates an analytic C0 -semigroup in H.

Such an operator is called regularly accretive (T. Kato [85]). It is shown in
[149] that -A generates an analytic C0 -semigroup in V*. For the sake of
convenience we reproduce the proof here.
Lemma 8. 2 (Lax-Milgram's theorem) Let X be a Hilbert space with in-
nerproduct (·, ·) and norm 11·11· Let B[u, v] be a sesquilinear form on X x X,
and suppose that there exist positive constants C, c such that

IB[u,v]l::; Cllullllvll, u,v EX, (8.70)

IB[u, u]l;::: cllull 2 , u EX. (8.71)

Then for any F E X* there exists an element u of X such that for any
vEX we have F(v) = B[u, v].
Proof. Let u E X be fixed. Then the functional X 3 v 1--t B[u, v] E C is
conjugate linear and continuous. Therefore there exists an element Su of
X such that B[u, v] = (Su, v) for any vEX. Sis a linear mapping from X
to X and by (8. 70)

I(Su,v)l = IB[u,v]l::; Cllullllvll

which implies IISII ::; C. In view of (8.71)

cllull 2 ::; IB[u, u]l = I(Su, u)l ::; IISullllull,

from which it follows that cllull ::; IISull, and hence S has a continuous
inverse. Therefore R(S) is closed. If v is orthogonal to R(S), then for any
u EX we have (Su, v) = 0. Choosing u = v we have

cllvll 2 ::; IB[v,v]l = I(Sv,v)l = 0,

344 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

which implies v = 0. Consequently we have shown that S is an isomorphism

from X onto X. Suppose F E X*. Then, in view of Riesz's theorem there
exists an element w of X such that F(v) = (w, v) for any vEX. If u is an
element of X such that w = Su, then

F(v) = (Su, v) = B[u, v], vEX.

If there exists another element u' E X such that F(v) = B[u', v] for any
vEX, then B[u- u', v] = 0 for any vEX. Hence

cllu- u'll 2 :::; B[u- u',u- u'] = 0,

which implies u = u'.

Lemma 8. 3 V is dense in V*.

Proof. It suffices to show that if l E V** and l(v) = 0 for any v E V,

then l = 0. Since V is reflexive, this reduces to proving that if u E V and
(u, v) = 0 for any v E V, then u = 0. This follows from the denseness of V
in H.

Theorem 8. 5 If Re,\ :::; -k, then A - ,\ is an isomorphism from V to V*,

and for any f E V* or E H we have

I(A- ,\)- 1/l:::; C21>.1- 11/l, (8.72)

I(A- >.)- 1/l:::; Cal>.l- 112 11/11*' (8.73)
II(A- >.)- 1!II:::; Cal>.l- 112 1/1, (8.74)
II(A- >.)- 1!II:::; c0 1ll/ll*' (8.75)
II(A- ,\)- 1 !II* :::; C21>.1- 111/ll*' (8.76)

where C2 = 1 + Cl/eo, Ca = vC1+eo/eo. Especially -A, -A generate

analytic Co-semigroups in H, V* respectively.

Proof. Let Re,\:::; -k. For u, v E V set

B[u,v] =((A- >.)u,v) = a(u,v)- >.(u,v).

Then

IB[u, v]l
:::; Clllullllvll + l>.llullvl :::; ( cl + I>.IC5) llullllvll,
ReB[u, u] = Rea(u, u)- Re>.lul 2
~ eollull 2 - (k + Re>.)lul 2 ~ eollull 2. (8.77)
8.6. REGULARLY ACCRETIVE OPERATORS 345

Hence we can apply Lemma 8.2 to conclude that A - .A is an isomorphism

from V onto V*.
Let (A- .A)u = f for f E V* orE H. Then
a(u, v)- .A(u, v) = (f, v), v E V. (8.78)
Letting v = u in (8.78) we get
a(u, u)- .Aiul 2= (f, u). (8.79)
Therefore
Rea(u,u)- Re.Aiul 2= Re(f,u). (8.80)
Combining (8.77) and (8.80)
eollull 2: : ; Re(f, u) ::::; IIIII* llull, (8.81)
which implies
(8.82)
or (8.75). From (8.79) and (8.82) it follows that
I.AIIul 2= la(u, u)- (f, u)l : : ; C1llull 2+ 11!11* llull
: : ; C1llfii;/Co + 11!11;/eo = Cillfll;,
which implies (8.73). Again from (8.77) and (8.80)
eollull 2: : ; Re(f, u) ::::; l!llul. (8.83)
Combining (8.79) and (8.83) we get
I.AIIul 2= la(u,u)-(f,u)l::::; C1llull 2+lfllul::::; C1lfllul/eo+lfllul = C2l!llul,
which implies
(8.84)
or (8.72). From (8.83) and (8.84) it follows that
llull 2: : ; lfllul/eo::::; C2I.AI- 1Ifl2/eo= Cii.AI- 11!12,
which implies

or (8.74). From (8.78) and (8.82)

I.AII(u,v)l::::; la(u,v)l + l(f,v)l::::; C1llullllvll + llfll*llvll
: : ; C1llfll*llvll/eo + llfll*llvll = C2llfll*llvll
for any v E V. Therefore we conclude

or (8.76).
346 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

8. 7 Semigroup Approach Revised and

Control Problem
We study a semigroup theoretical treatment of (8.1),(8.2) on different spaces
from those in the preceding sections. These spaces are more useful in con-

1:
sidering the following control problem for the equation (8.1):

! u(t) = Aou(t) + A1u(t- h)+ a(s)A2u(t + s)ds + <l>ow(t),

t E [0, T], (8.85)
u(O) = g0 , u(s) = g 1(s), s E [-h, 0), (8.86)
where <1> 0 is a controller. Following C. Bernier and A. Manitius [20], M. C.
Delfour and A. Manitius [44],[45], M. C. Manitius [107], S. Nakagiri [112],
etc. and using the result of G. Di Blasio, K. Kunisch and E. Sinestrari
[51],[52] we rewrite the problem (8.55),(8.56) as an equation in the space
Z = (D(Ao), Hh;2, 2 x L 2( -h, 0; D(Ao)):

d
dtx(t) = Ax(t) + <l>w(t), 0 ~ t ~ T, (8.87)
x(O) = g, (8.88)

where A is the infinitesimal generator of the solution semigroup associated

with the problem (8.1),(8.2) and <l>w = (<1> 0 w, 0). When we discuss the
controllability, observability, etc. we consider also the adjoint equation

dz(t)jdt=A*z(t), z(O)='¢.

This is an equation in Z* = (D(A 0 ), H)~ 12 , 2 x L 2 ( -h, 0; D(Ao)*). However,

D(A0 )* is not a space easily dealt with, namely D(Ao) is too small a space
to consider its adjoint space. Therefore in what follows we consider the
equation (8.85) in the space V* in case -Ao is a regularly accretive operator
associated with a sesquilinear form satisfying (8.66),(8.67). Then Z = H x
L 2 ( -h, 0; V), since the domain of A 0 in this situation coincides with V
and (V, V*h; 2 ,2 = H by virtue of Proposition 8.5 below, and hence Z* =
H x L 2 (-h,O;V*).
We state our fundamental hypothesis.
Let Hilbert spaces H, V and a sesquilinear form a(·,·) be as in section
8.6. We denote the operators A, A of section 8.6 both by the same notation
-A 0 , and D(Ao) denotes the domain of its realization in H:

(-Aou,v) = a(u,v) (8.89)

 8.7 SEMIGROUP APPROACH REVISED AND CONTROL 347

for u E D(Ao) orE V and v E V, and D(Ao) = {u E V; A 0 u E H}. In

this notation Ao generates an analytic semigroup both in H and V*. We
assume that
(R2') A 1 , A 2 are bounded linear operators from V to V*, and and their
restrictions to D(Ao) are bounded linear operators from D(Ao) to H.
As for a(·) we assume (R3) of section 8.6. <1> 0 is a bounded linear operator
from some Banach space U to H.
Let Ai, i = 0, 1, 2, be the adjoint operators of Ai:
(A;u,v) = (u,Aiv), i = 0, 1,2, (8.90)
where the right hand side is the pairing between u E V and Aiv E V*. It
is well known that A 0 is the operator associated with the sesquilinear form
a*(u, v) = a(v, u) adjoint to a(u, v), which also satisfies (8.66), (8.67). The
· ,domain D(A0) of the realization of A 0 in H coincides with the domain of
the adjoint operator of the realization of A 0 in H. We assume
(R4) The restrictions of Ai, A2 to D(A0) are bounded linear operators from
D(A0) to H.
Lemma 8. 4 If u E L 2 (0, T; V), u' E L 2 (0, T; V*), T > 0, then
u E C([O, T]; H). If v is another function having the same property, then
(u(t),v(t)) is absolutely continuous in [O,T] and
d(u(t), v(t))jdt = (u'(t), v(t)) + (u(t), v'(t)). (8.91)
···.Proof. If we set u(t) = u( -t} ·for -a < t < 0, u(t) = u(2T - t) for
T < t < T + a, where 0 < a < T, then u E L 2 ( -a, T + a; V) and u' E
L 2 (-a, T +a; V*). Let 'ljJ be a continuously differentiable real valued function
such that 'lj;(t) = 1 fortE [0, T] and 'lj;(t) = 0 fort in some neighborhoods
of -a and T +a, and set w(t) = 'lj;(t)u(t). Let Pn be a mollifier, and set

Wn(t) = 1 T+a
-a Pn(t- s)w(s)ds.

Then as n ---+ oo Wn ---+ w in L 2 (-a, T + a; V) and w~ ---+ w' in L 2 (-a, T +

a; V*). Noting that Wn E C 1 ([-a, T +a]; V), we obtain

lwn(t)- Wm(tW = [ta !lwn(s)- Wm(s)l ds 2

= [ta 2Re (w~(s)- w:n(s), Wn(s)- Wm(s)) ds

:::; 1T+a
-a llw~(s)- w:n(s)ll~ds + 1-a
T+a
llwn(s)- Wm(s)ll 2 ds.
348 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Hence {wn} is a Cauchy sequence in C([-a, T +a]; H). Therefore if we

modify the values of u in some null set, we have lwn(t) -u(t)l-+ 0 uniformly
in [0, T]. Thus we conclude u E C([O, T]; H). Next we prove (8.91). If
¢ E CQ"(O, T), then

loT (wn(t), v(t))¢' (t)dt

=loT (! (¢(t)wn(t)), v(t)) dt -loT ¢(t) (w~(t), v(t)) dt

=-loT ¢(t)(wn(t),v'(t))dt-loT ¢(t)(w~(t),v(t))dt.
Letting n-+ oo
T T T
lo (u(t),v(t))¢'(t)dt= -lo ¢(t)(u(t),v'(t))dt-lo ¢(t)(u'(t),v(t))dt.

This means that the distribution derivative of (u(t), v(t)) coincides with the
right hand side of (8.91). Since this is integrable in [0, T], the conclusion of
the lemma follows.
Proposition 8. 5 (V, V*h;2,2 =H.
Proof. Suppose x E H. If we put u(t) = etAox, then
1d
"2dtlu(tW =Re(u'(t),u(t))
= Re(Aou(t), u(t)) = -Rea(u(t), u(t)):::; -eollull 2 + kju(tW.
Integrating this differential inequality we obtain

~lu(tW +Co lot e2k(t-s)llu(s)ll2ds:::; ~e2ktlxl2·

Therefore for any T > 0 u E £ 2 (0, T; V), u' = A 0 u E £ 2 (0, T; V*).
Hence x = u(O) E (V, V*h; 2,2. Conversely suppose x E (V, V*h;2, 2. If
we set u(t) = etAox, then in view of Theorem 1.10 u E £ 2 (0, T; V) and
u' E £ 2 (0, T; V* ). By Lemma 8.4 u E C([O, T]; H). Hence x = u(O) E H.
As was stated in the beginning of this section we consider the problem
(8.85), (8.86) in V*. In view of Proposition 8.5 the associated solution semi-
group S(t) is a C0 -semigroup in H x L 2 ( -h, 0; V). In what follows in this
chapter we denote this space by Z:

Z=H x L 2 (-h,O;V). (8.92)

8. 7 SEMIGROUP APPROACH REVISED AND CONTROL 349

The problem (8.85),(8.86) is rewritten as

d
dtx(t) = Ax(t) + <I>w(t), 0::; t::; T, (8.93)
x(O) = g, (8.94)

where x(t) = (u(t),ut),

<I>w = (<I> ow, 0) (8.95)
and g = (g 0 , g 1 ) E Z. The adjoint problem of (8.93),(8. 94) is
d
dt z(t) =A* z(t), (8.96)
z(O) = 1/J E Z*. (8.97)

Since Z is reflexive, S*(t) is a Co-semigroup in

Z* = H x L 2 ( -h, 0; V*), (8.98)

and the solution of (8.96),(8.97) is represented by

z(t) = S* (t)'ljJ. (8.99)

We consider also the following problem in addition to (8.85),(8.86) and

(8.96),(8.97):

! v(t) = Abv(t) + Aiv(t- h)+ loh a(s)A~v(t + s)ds, t E [0, T], (8.100)

v(0)=¢0 , v(s)=¢ 1 (s) sE[-h,O). (8.101)

The associated solution semigroup is denoted by ST(t), and its infinitesimal
generator by AT. In view of Theorem 8.4 the infinitesimal generator A of
S (t) is characterized by

D(A) = {g = (g 0 , g1 ) E Z; g1 E W 1 •2 ( -h, 0; V),

g0 = g 1 (0), A 0 g0 + A 1 g1 ( -h)+ Lg 1 E H}, (8.102)
Ag = ( Aog0 + Atg1 ( -h)+ Lg 1 , g1' ) , (8.103)

D(AT) = {g = (g 0 ,g1 ) E Z; g1 E W 1 •2 ( -h, 0; V),

g0 = g 1 (0), A 0g0 + Aig 1 ( -h)+ L* g1 E H}, (8.104)
ATg = ( A0g0 + Aig 1 (-h) + L*g\g 1' ) , (8.105)
350 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

where

Lg 1 = j_oh a(s)A2g 1(s)ds, L*g 1 = j_oh a(s)A;g 1(s)ds.

For.>.. E C set

.6.(.>..) = .>..- Ao- e->.h A 1 -1 -h

0
e>. 8 a(s)A2ds, (8.106)

.6.T(.>..) = .>..-A~ - e->.h A;' -1° e>.ha(s)A~ds. (8.107)

-h

Then .6.(.>..), .6.T(.>..) are bounded linear operators from V to V*, and .6.(.\)* =
.6.(5.). As is easily seen if Re.>.. is sufficiently large, then .6.(.\), .6.T(.>..) have
bounded inverses .6.(.>..)- 1 , b..T(.>..)- 1 , which are called retarded resolvents of
A and AT respectively.

Lemma 8. 5 Suppose g = (g0 ,g1), f = (f 0 ,jl) E Z. Then

g E D(A) and (.>.. -A)g = f (8.108)

if and only if

go E V, (8.109)

.6.(.\)go = fo + j_oh e->.(h+r)Ad1(T)dT

+1° a(s)1° e>.(s-r)A2f 1(T)dTds, (8.110)

-h 8

g1(s) = e>.sgo + 10 e>.(s-r) f1(T)dT. (8.111)

Consequently if Re.\ is sufficiently large, then.>.. E p(A) n p(AT ).

Proof. First assume that (8.108) holds. Then in view of (8.102),(8.103)

.\go - Aogo - A1g1 (-h) - Lg1 = fo' (8.112)

.\g1 _ g1' = l, g1(0) =go E V. (8.113)

Solving the initial value problem (8.113) we obtain (8.111). Substituting

this in (8.112) we conclude (8.110).
Conversely suppose that (8.109),(8.110),(8.111) hold. From (8.111) we
get
(8.114)
8.8. STRUCTURAL OPERATOR 351

It follows from (8.110),(8.111) and the definition of~(.\)

. Aog0 + A1g1( -h)+ j_oh a(s)A2g 1(s)ds

=Aogo+Al (e--\hgo+ j_oh e--\(h+r)fl(r)dr)

+ j_oh a(s)A2 (e,\ 8 g0 + 1° e-\(s-r) f 1(r)dr) ds

= Aogo + e--\h Alga+ 1o e--\(h+r) Adl(r)dr

1°
-h

+ j_oh e,\ 8 a(s)A2g0 ds + j_oh a(s) e-\(s-r) A2f 1(r)drds

= Aogo + e--\h Alga+ 1o e,\sa(s)A2gods +~(.\)go- fo

-h
= .\go - fo E H.

This and (8.114) imply that g E D(A) and

(.\ _ A)g = .\g _(.\go_ Jo, .\gl _ fl) =f.

8.8 Structural Operator

Following M. C. Delfour and A. Manitius (44],(45) we define the struc-
tural operator associated with the problem (8.85),(8.86) as follows: for
f = (! 0 , f 1 ) E Z

Ff = (J 0 ,Ad 1(-h- ·) + [h a(r)A2l(r- ·)dr). (8.115)

We denote the first and second components of F f by [F f] 0 and (F jjl re-

spectively:
[Ff] 0 = J0 , (8.116)

[Ff] 1(s) = A1l( -h- s) + j_sh a(r)A2l(r- s)dr. (8.117)

It is clear that FE ,C(Z, Z*). It is easy to see that the adjoint operator F*
ofF is given by

(8.118)
352 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

namely F* is the operator with Ai, A2 in place of A 1 , A 2 in the definition of

F. Clearly F* E C(Z, Z* ). We find it to convenient to define the following
operators introduced by M. C. Delfour and A. Manitius [44],[45]: for,\ E C
E:>..EC(V,Z), [E:>..u] 0 =u, [E:>..ujl(s)=e:>.. 8u for uEV, (8.119)

T:>.. E C(Z, Z), [T:>..g] 0 = 0, [T:>..g] 1 (s) = 1° e:>..( 8-r)g 1(r)dr,

for g = (g 0 ,g 1) E Z, (8.120)

H:>.. E C(Z*, V*), H:>..J = f 0 + 1° e:>.. 8f 1(s)ds

-h
for J=(J 0 ,l)EZ*, (8.121)
8
K:>..EC(Z*,Z*), [K:>..f] 0 =0, [K:>..f] 1(s)=1 e:>..<r- 8)l(r)dr
-h
for f = (J0 ,J 1 ) E Z*. (8.122)
Lemma 8. 6 For ,\ E C
(i) FT:>.. = K:>..F, (ii) F*T:>.. = K:>..F*, (iii) E~ = Hx, (iv) H{ =Ex,
(v) T; = Kx, (vi) K{ = Tx.
Proof. (i) and (ii) We have forgE Z
[FT:>..g] 0 = [T:>..g] 0 = 0,
[K:>..Fg] 0 = 0,
8
[FT;>..gjl(s) =A1[T:>..gjl(-h-s)+ j_ h a(r)Az[T:>..g] 1(r-s)dr

= Al j_oh-8 e:>..(-h-8-r)gl(r)dr + j_8h a(r)Az1~8 e:>..(r-8-a)gl(CJ)dcrdr,

8
[K:>..Fg] 1(s) = j_ h e:>..(r- 8)[Fg] 1(r)dr

8
= j_ h e:>..(r- 8) [A1g 1(-h- r) + j_rh a(CJ)Azg 1(CJ- r)dcr] dr

= A 11
8 e:>..(r- 8)g 1( -h- r)dr + 1 8 e:>..(r- 8) 1r a(CJ)Azg 1(CJ- r)dcrdr
-h -h -h
= A11o e:>..(-h-r-8)gl(r)dr + 18 a(CJ) 18 e:>..(r-8) Azgl(CJ- r)drdCJ
-h-8 -h a
= Al j_oh-8 e:>..(-h-r-8)gl(r)dr + j_8h a(CJ)Az1~8 e:>..(a-r-8)gl(r)drdCJ
= [FT:>..g] 1 (s).
Thus (i) is established. Since F* has the same property as F, and T:>.., K:>..
are independent of A, i = 0, 1, 2, (ii) is a consequence of (i).
8.8. STRUCTURAL OPERATOR 353

(iii) and (iv) We have for f E Z* and u E V

(E~J,u) = (f,E>.u) = (f 0 ,u) + [oh (l(s),e>- 8 u)ds

= (! +
0 [oh e>. 8 l(s)ds,u) = (Hxf,u),
which implies (iii). Taking adjoints we obtain (iv).
(v) and (vi) We have for fEZ* and g E Z

(T;J,g) = (!, T>.g) = [oh (t1(s), 10 e>-(s-r)g1(T)dT) ds

= [oh ([""h e>.(s-r) f1(s)ds,g1(T)) dT

= [oh ([Kxf] 1(T),g 1(T)) dT = (Kxf,g),

which shows that (v) holds. Taking adjoints we get (vi).

The following theorem was proved by M. C. Delfour and A. Manitius [44]
for finite dimensional equations, and by S. Nakagiri [112] and J.-M. Jeong,
S. Nakagiri and H. Tanabe [82] for equations in infinite dimensional spaces.
Theorem 8. 6 (i) Fort 2: 0

FS(t) = ST(t)* F, S(t)* F* = F* ST(t). (8.123)

(ii) If Re>. is sufficiently large,

F(>.- A)- 1 =(.A- Ar)- 1 F, (.A- A*)- 1 F* = F*(.A- AT)- 1 . (8.124)

Remark 8. 1 The second statement of (i) is the relation which connects

two kinds of adjoints of S(t).

Proof of Theorem 8.6. Let f = (! 0 , J 1 ) be an arbitrary elements of Z. Let

Re.A be so large that both 6.(>.), 6.T(.X) have bounded inverses. Then in
view of Lemma 8.5 there exists an element g = (g 0 , g 1 ) of D(A) such that
(.A- A)g =f. By virtue of Lemma 8.5

H>.FJ=J 0 + [oh e>.s [Ad 1(-h-s)+ [sh a(T)Ad 1(T-s)dT] ds

= f0 + 1°
-h
e>.s Al/ 1 ( -h- s)ds + 1° 1
-h
e>- 8
8

-h
a(T)Ad 1(T- s)dTds
354 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Consequently we have

Again by Lemma 8.5

E>.i:l.(>.)- 1H>.Ff + T>.f = (g 0 , e>-·g0 ) + ( 0, 1° e>.(·-r) f 1(T)dT)

= (go, e>-·gO + 10 e>.(·-r) f1(T)dT) = (go,g1) = g,

and so
(8.125)
Analogously
(.\.- AT)- 1 = Exl:l.T(.X)- 1HxF* + Tx. (8.126)
Taking adjoint of (8.126) and using Lemma 8.6
(>.- Ar)- 1 = FE>.i:l.(>.)- 1H>. + K>.. (8.127)
From (8.125),(8.127) and Lemma 8.6 it follows that
F(>.- A)- 1 = FE>.i:l.(>.)- 1H>.F + FT>.
= FE>.i:l.(>.)- 1H>.F + K>.F = (>.- Ar)- 1F. (8.128)
Hence the first equality of (8.124) is established. Taking the adjoint and
replacing .X by >. we obtain the second one. From (8.128) it follows that if
f E D(A), then
Ff = F(>.- A)- 1(>.- A)f = (>.- Ar)- 1F(>.- A)f E D(Ar)
and
(>.- Ar)Ff = F(>.- A)J,
and so
ArFf = FAJ. (8.129)
If we put u(t) = FS(t)f for f E D(A), then applying (8.129) to S(t)f E
D(A) we get

u'(t) = FAS(t)f = ArFS(t)f =.Aru(t), u(O) = Ff,

8.9. CONTROLLABILITY AND OBSERVABILITY 355

which implies
u(t) = ST(t)Ff.
Hence we see that
FS(t)f = ST(t)Ff (8.130)
holds iff E D(A). Since FS(t) and ST(t)F are both bounded linear oper-
ators from Z to Z* we see that (8.130) holds for any fEz. Thus the first
part of (8.123) is established. The second part follows by taking adjoint.
From the theorem or its proof (cf. (8.129)) we get
Corollary 8.1 Iff E D(A), then Ff E D(AT) and ATFf = FAf, and
iff E D(AT), then F* f E D(A*) and A* F* f = F* ATf·
Definition 8. 2 p(tl) = {,\;b..( A) is an isomorphism from V onto V*},
a-(tl) = C \p(tl). Replacing b..( A) by LlT(A) the sets p(b..T) and a-(b..T) are
defined.
The following corollary was first proved by M. C. Delfour and A. Manitius
[45: Proposition 4.2 (ii)]
Corollary 8. 2 (i) If A E p(Ll), then A E p(A) and

(A- A)- 1 = E-Atl(A)- 1 H.AF + T.A.

(ii) IfAEp(tlT), thenAEp(AT) and

(A- AT)- 1 = E.Ab..T(A)- 1 H.AF* + T.A.
Proof. See (8.125) and (8.126) in the proof of Theorem 8.6.

8.9 Controllability and Observability

In this section we study the controllability problem for (8.93),(8.94). For
the sake of simplicity we consider the case g = 0:

(S) { :tx(t) = Ax(t) + <l>w(t) t;:::: 0,

x(O) = 0.

Definition 8. 3 The system (S) is said to be approximately controllable if

Cl {1t S(t- T)<l>w(T)dT; wE £ 2 (0, t; U), t > 0} = Z,

where Cl means the closure in Z.

356 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

The observability of the system

(ST) { :ty(t) = ATy(t) t ~ 0,

y(O) = ¢ E Z

is defined as follows.

Definition 8. 4 The system (ST) is said to be observable if <fl0[ST(t)¢] 0

0 implies ¢ = 0.
=
=
The observability means that <fl0[ST(t)¢ 1 ] 0 <PO[ST(t)¢2]0 implies ¢ 1 = ¢2,
i.e. the initial value ¢ is observable through the measured values <I>0v(t),
where v(t) is the solution of (8.100),(8.101).

Theorem 8. 7 Suppose F is an isomorphism from Z to Z*. Then (S) is

approximately controllable if and only if (sT) is observable.

Proof. (S) is approximately controllable if and only if

fEZ*, (f, lt S(t- T)<I>w(T)dT) = 0 for any wE L 2 (0, t; U), t >0

implies f = 0, or
<fl* S(t)* f =0 implies f = 0. (8.131)
Since F* is an isomorphism from Z to Z* under our assumption, (8.131) is
equivalent to saying that

¢ E Z, <I>* S(t)* F*¢ =0 implies ¢ = 0. (8.132)

By Theroem 8.6 this holds if and only if

¢ E Z, <I>* F* ST(t)¢> =: 0 implies ¢ = 0. (8.133)

As is easily seen

Hence (8.133) is equivalent to the observability of (ST).

A sufficient condition in order that the assumption of Theorem 8. 7 is
satisfied is given by the following proposition.

Proposition 8. 6 If A 1 is an isomorphism from V to V*, then F is an

isomorphism from Z to Z*.
8.9. CONTROLLABILITY AND OBSERVABILITY 357

Proof. The relation F g = f is rewritten as

9 o = fo,

A1g 1(-h- s) +ish a(T)A2g 1(T- s)dT = l(s). (8.134)

We want to solve (8.134) for a given p E L 2 ( -h, 0; V*). Letting A;:- 1

operate on both sides and making a suitable change of variable we see that
(8.134) is equivalent to

g 1(-s)+ 1 8
b(s-T)A1 1A 2g 1(-T)dT=A1 1f 1(s-h), sE [O,h],

where b(s) = a(s- h), s E [0, h]. Hence the problem is reduced to solving
the following integral equation

¢(s) + 1 8
b(s- T)A;:- 1A2¢(T)dT = 1/J(s), s E [0, h], (8.135)

where 1/J E L 2 (0, h; V). Let a* b denote the convolution

(a* b)(s) = 1 8
a(s- a-)b(a-)du.

The resolvent kernel of (8.135) is the function R with values in ..C(V, V)

satisfying

R+bA1 1A2 + R* bA;:- 1A2 = R+ bA;:- 1A2 + bA;:- 1A2 * R = 0,

and is constructed by successive approximation:
<Xl
R = ~.)-1)nbn (A;:- 1 A2r,
n=1

where bn = b * ··· * b is the convolution of n of b. Since bE L 2(0, h), b2 is

bounded: ess suplb2(s)l = II~II<Xl < oo. By induction we can show that
8 n-1
lb2n(s)l::; llb21i~ (n _ 1)!,

lb2n+1(s)l::; llb21i~ Jo
r (s- a-)n-1
(n _ 1)! lb(a-)loo,

for n = 1, 2, .... Therefore R E L 2 (0, h; ..C(V, V)). The unique solution of

(8.135) is given by¢= 1/J + R * 1/J.
The results of this section are due to A. Manitius [109] and S. Nakagiri
and M. Yamamoto [115].
358 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

8.10 Characterization of Ker(.A- A)z

Supposing that >. E ap(A) =the set of point spectra of A, we characterize

Ker(>.- A) 1 = {¢ E Z; (>.- A) 1¢ = 0}.

Under this hypothesis t,.(>.) is not invertible by Lemma 8.5.

Theorem 8. 8 For l = 1, 2, · · ·

Ker(>.- A) 1 = { (¢g,e>-· ~ (~? ¢?);

l-1 (-1)i-j (i-j+1) 0 - . -
}
I : ( i - · 1)!t,. (>.)¢i-O,J-1, ... , l , (8.136)
i=j-1 J+

where t,.(i-i+l)(>.) = (djd>.)Ci-i+ 1)t,.(>.). The same result holds with A and
t,. replaced by AT and t,.T respectively.
Proof. In case l = 1 the assertion is

(8.137)

This is a direct consequence of Lemma 8.5.

We denote the set of the right hand side of (8.136) by K 1• We first show
that K 1 C Ker(>.- A) 1• Let

¢= (¢g,e>-·~(~?¢?) EKt,
l-1 ..
~ ( - 1 )~-J A (i-j+1) (')A.~ = 0 • 1 { (8.138)
L....t (i - )• + 1)! L.l. A '/"~ ' ) = '• • •' •
i=j-1
Then

where
8.10. CHARACTERIZATION OF KER(>-. - A)1 359

= Ao¢g + (e--\h A1 + [oh e-\ 8 a(s)dsA2) ¢g

l-1
+ 2: ~ (-oi,1- (-1)i~Ci)(>-.)) ¢?
i=1

= (>-. - ~(>-.) )¢g - ¢~ - ~ ( -.~)i

z.
~ (i) (>-.)¢?' (8.139)
i=1
where we used that for i = 1, 2, ...

( -1)i~ (i) (>-.) = -Di,1 - hie--\h A1 -1° ( -h

-s)ie-\8 a(s)dsA2. (8.140)

Using (8.138) for j = 1 we get from (8.139)

'1/Jo = >-.¢g- ¢~.

Hence
l-2(-1)i )
(>-.-A)¢= ( ¢~,e-X·~ ~¢?+1 ·

The condition (8.138) for j = 2, ... , l is rewritten as

l-2 . .
~ ( -1Y-J A (i-j+1) ( ')"'~ = 0
i~1(i-j+1)!u "''~-'2+1 'j=1, ... ,Z-1.

Consequently we see that (>-. -A)¢ E Kt_ 1 • Repeating this we obtain

(>-.- A)1- 1 ¢ E K 1 = Ker(>-.- A), or¢ E Ker(>-.- A)1•
Next supposing that Ker(>-.- A) 1 C K 1 holds, we are going to show that
the same inclusion relation holds for l+ 1 in place of l. Let¢ E Ker(>-.-A) 1+1 .
Then(>-.- A)¢ E Ker(>-.- A) 1• Hence

(>-.-A)¢= ( '1/Jg, e-X·~ ( ~? '1/Jf) , (8.141)

l-1 . .
~ ( - 1Y-J ~(i-j+1)(>-.)·'·~ = 0 J. = 1, ... ' l. (8.142)
L.J (i - J. + 1)! '1-' 2 '
i=j-1
In view of Lemma 8.5
360 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

From (8.143) it follows that

(8.145)

where we again used (8.140), and from (8.144)

(8.146)

If we set <1>8 = ¢0 , </>~ = 'l/J8, ... , ¢f = '1/Jf_v then (8.145),(8.146) imply

(8.147)

(8.148)

By (8.142) we have

l (- ) t-J
. .
L . 1. A(i-J+l)(')"'o
)1 L.l.
(Z-J+1. A 'l'i =
0' J. = 2'• • •' l+1 • (8.149)
i=j-1

Combining (8.147),(8.148),(8.149) we conclude </> E Kl+ 1 to complete the

proof of the theorem.
8.11. A SPECIAL CASE 361

8.11 A Special Case

We consider the special case where

(8.150)

1 being a real number. In this case

.6.(A) = A- m(A)Ao, (8.151)

m(A) = 1 + 1e->.h + 1° -h
e>. 8 a(s)ds. (8.152)

As is easily seen m(A) is an entire function of A and

m(A) ____. 1 as ReA ____. oo. (8.153)

In G. DiBlasio, K. Kunisch and E. Sinestrari [52] and S. Nakagiri and H.

Tanabe [114] the relations between the spectrum of A and that of .6.(·) are
discussed in detail.
We prepare some results from the spectral theory of closed linear operators
in a Banach space.
Definition 8. 5 Let T be a closed linear operator in some Banach space,
and A be an isolated point of the spectra i7(T) of T. Then the spectral
projection of T at A is the operator P defined by

P = -21 .
7r 2
i.
IJ.L->.1=<
(p, - T) -1 dp,,

where E is a positive number such that A is the only spectrum of T in the

closed disk {p,; IM- A\ :S E}.
In what follows in this chapter we denote by ImT the image (range) of
the operator T.
The following lemma is rather well known (cf. T. Kato [86: p.180]).
Lemma 8. 7 Let T be a closed linear operator in a Banach space X, and
A an isolated point ofi7(T). Then (p,-T)- 1 has the Laurent expansion near
A: 00 00

(8.154)
n=O n=1

An and Bn are bounded linear operators and

Bn = (A- T)n- 1 P, n = 1, 2, ... , (8.155)

362 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

where P is the spectral projection ofT at .A. If moreover .A is a pole of

(f.k- T)- 1 of order v, then .A is an eigenvalue ofT and

Ker((.A- T)n- 1 ) # Ker((.A- T)n),

n = 1, 2, ... , v,
Ker((.A- T)n) = Ker((.A- Tt) =ImP, n = v + 1, v + 2, ... ,
Im((.A- T)n- 1 ) # Im((.A- T)n), n = 1, 2, ... , v,
Im((.A- T)n) = Im((.A- Tt) = Im(I- P), n = v + 1, v + 2, ....

Hence
X= Ker((.A- Tt) ~ Im((.A- Tt). (8.156)
The first part of the lemma is proved by a direct calculation, and the second
part is proved by using that .A is the only spectrum of T in the space ImP
and the spectra ofT in the space Im(I- P) coincides with a(T) \{A}.
If the imbedding V c H is compact, then by the Riesz-Schauder theory
the spectra of A 0 consists only of discrete eigenvalues:

a(Ao) = {f.kj; j = 1, 2, ... }, fki --too as j---> oo. (8.157)

It is easily seen that the spectrum of A 0 considered as an operator in V*

coincides with that of its realization in H.
Theorem 8. 9 (J.-M. Jeong [80]) Suppose that (8.150) holds for some real
constant"'(, and that the imbedding V C H is compact. Then we have

a(A) = ae(A) U ap(A), (8.158)

where ae(A) ={A; m(.A) = 0} and ap(A) ={A; m(.A) # 0, Ajm(.A) E a(A0 )}.
Each nonzero point of ae(A) is not an eigenvalue of A and is a limit point
of a(A). ap(A) consists only of discrete eigenvalues.
Suppose m(O) = 0. Then 0 is an eigenvalue of A of infinite multiplicity.
The point 0 is an isolate point of a(A) if it is a simple zero of m(·), and is
a limit point of a(A) if it is a multiple zero of m(· ).
Proof. If .A E p(A), then in view of Lemma 8.5 the equation (.A-m(.A)A 0 )g0 =
f 0 has a unique solution g 0 E V for any f 0 E H. Hence m(.A) # 0 and
Ajm(.A) E p(A 0 ). Conversely if .A satisfies this condition, then the equation
(.A- A)g = f has a unique solution g for any fEZ in view of Lemma 8.5.
Therefore we have

p(A) = {A; m(.A) # 0, .A/m(.A) E p(Ao)}.

Suppose .A0 # 0 is a zero of m(.A) of order k. Then there exists a function

h(.A) which is holomorphic and does not vanish in some neighborhood of .A 0
8.11. A SPECIAL CASE 363

such that m(>.)/ ,\ = (>.- >. 0 )kh(>.)k. Applying the inverse function theorem
to (>.- >. 0 )h(>.) and noting 1-Li --+ oo we see that for sufficiently large j there
exists a complex number ,\i such that (Aj- >. 0 )h(>.j) = p,j 1 /k and ,\i--+ >.0 •
Then >.ifm(>.j) = f-Li E CT(A 0 ). Hence ,\ 0 is a limit point of {>.i} C CT(A).
Next suppose that ,\0 E CTp(A), i.e. m(>. 0 ) f:- 0, A0 /m(>. 0 ) E CT(Ao). If
there exists a sequence { ,\i} such that ,\0 f:- Aj E CT(A), m(>.j) f:- 0 and >.i --+
>.0 . Then >.ifm(Aj) --+ >. 0 /m(,\ 0 ), Aj/m(Aj) E CT(Ao). Since CT(Ao) consists
only of isolated points we have Aj/m(Aj) = >.0 jm(>. 0 ) for sufficiently large
j. By virtue of the theorem of identity we have >.jm(>.) = >. 0 /m(,\ 0 ), which
contradicts (8.153). Next we show that ,\0 is a pole of(>.- A)- 1 • From the
theorem of identity and (8.153) it follows that if,\ f:- >.0 lies in a sufficiently
small neighborhood of >.0 , then >.jm(>.) E p(A0 ), and hence >. E p(Ll). By
the Riesz-Schauder theory >. 0 /m(,\ 0 ) is a pole of the resolvent of Ao whose
order we denote by v. Therefore if we denote by k the order of Ao as a zero
of >.jm(>.) - >. 0 /m(,\ 0 ), then ,\0 is a pole of

1 ( ,\ ) -1
l:l(>.)-1 = m(>.) m(>.) - Ao

of order kv. Hence the result follows from Corollary 8.2 or (8.125) since all
other functions in the right hand sides of (8.125) are entire functions of>..
It is easily seen that the elements ¢8, ¢~, · · ·, ¢f_ 1 of the right hand side of
(8.136) satisfy

Moreover Kerll(>. 0 )v is finite dimensional. Therefore Ker(,\ 0 - A)kv is finite

dimensional.
If m(O) = 0, then for any g0 E V g = (g 0 , g0 ) satisfies Ag = 0 in view of
Lemma 8.5. Hence 0 is an eigenvalue of A of infinite multiplicity. Suppose 0
is a simple zero of m(>.). Then there exists a function h(>.) which is holomor-
phic and does not vanish in some neighborhood of 0 such that m(>.) = >.h(>.).
If there exists a sequence 0 f:- ,\i E CT(A) such that ,\j --+ 0, then for large
j m(>.j) f:- 0, >.ifm(>.i) E CT(A0 ) and >.ifm(>.i) = h(>.i)- 1 --+ h(o)- 1 . Since
CJ(Ao) consists only of isolated points we have Aj/m(>.j) = 1/h(O) for large
j. Hence m(>.) = >.h(O), which contradicts (8.153). Next suppose that 0 is
a zero of m(>.) of order k > 1. Then there exists a function h(>.) with the
same property as above such that m(>.) j ,\ = >,k- 1 h(>.)k- 1 . Applying the in-
verse function theorem to >.h(>.) we find that for large j there exists a point
>.i such that Ajh(>.j) = p,j 1 /(k- 1\ ,\i --+ 0. Then >.ifm(>.j) = f-Li E CT(Ao).
Hencce 0 f:- Aj E CJ(A). This means that 0 is a limit point of CJ(A).
364 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

8.12 Eigenmanifold Decomposition

The adoint operator A* of A is characterized by the following theorem, the
proof of which is due to S. Nakagiri [112].
Theorem 8. 10 We have for the operator A*

D(A*) = {j = (fo, f1 ); fo E V, f1 E W1,2( -h, 0; V*),

A~f 0 + f 1 (0) E H, f 1 ( -h)= Aif0 }, (8.159)
A* f = (A~f 0 + f 1(0),a(·)A;f 0 - f 11 ) for f E D(A*). (8.160)
Proof. Suppose f E D(A*) and A* f = 'lj;. Then for any g E D(A) we have
(Ag, f)= (g, '¢), or

( Aog0 + A1g 1( -h)+ ioh a(s)A 2 g 1(s)ds, f 0 )

+ ioh (g1' (s), f1(s))ds =(go, 'lj;o) + ioh (g1(s), 'lj;1(s))ds. (8.161)

Let¢ be a real valued function such that¢ E W 1 •2 ( -h, 0), ¢(-h)= 0, ¢(0) =
1, J~h a(s)¢(s)ds = 0. Then for any g0 E D(Ao) (g 0 , ¢(·)g0 ) E D(A). Sub-
stituting this in (8.161) we get

or

Since the right hand side is continuous in g 0 in the topology of V, we find

that f 0 E V. Set

M(s) =ish '¢ 1(T)dT, N(s) = Aif0 +ish a(T)A;f 0 dT.

Then M, N E W 1 •2 ( -h, 0; V*). Let y E W 1 •2 ( -h, 0; D(A0 )). Then

1 0
-h (y'(s), M(s))ds
0
= (y(O), M(O)) - i h (y(s), 'lj; 1(s))ds, (8.162)

loh (y'(s), N(s))ds

8.12. EIGENMANIFOLD DECOMPOSITION 365

= (y(O),N(O))- (y(-h),N(-h))- j_oh (y(s),a(s)AH 0 )ds

= (y(O),N(O))- (y(-h),AU 0 )
- (/_: a(s)A 2 y(s)ds, f 0 ) • (8.163)

Substituting (y(O), y) E D(A) in (8.161) we get

( Aoy(O) + A1y( -h) + j_oh a(s)Azy(s)ds, f 0 )

+ j_oh (y'(s), f 1(s))ds = (y(O), 'lj; 0 ) + j_oh (y(s), 'lj; 1(s))ds. (8.164)

Combining (8.162),(8.163),(8.164) we obtain

(Aoy(O) + A1y( -h), f 0 ) + (y(O), N(O))- (y( -h), A~f 0 )

- j_oh (y'(s), N(s))ds + j_oh (y'(s), l(s))ds

= (y(O), 'lj; 0 ) + (y(O), M(O))- j_oh (y'(s), M(s))ds,

and hence using (A1y( -h), f 0 ) = (y( -h), Aif0 )

j_oh (y'(s), f 1(s)- N(s) + M(s))ds

= (y(O), 'lj; 0 - N(O) + M(O)) - (Aoy(O), f 0 ). (8.165)

From this it follows that

=N(s)- M(s).
f 1(s) (8.166)

Therefore j1 E W 1, (-h,O;V*),J 1(-h)

2 N(-h) -M(-h) Aif and
= = 0

'lj; 1 = a(·)A2f 0 - p'. Letting y(s) =g D(A we get from (8.165)

0 E 0)

(g 0 , 'lj; 0 - N(O) + M(O)) = (Aog 0 , ! 0 )

which implies A;J 0 = 'lj; 0 -N(O)+M(O). From this and (8.166) we conclude

A;J 0 + J 1(0) = A;J0 + N(O)- M(O) = 'lj; 0 E H.

Conversely iff belongs to the right hand side of (8.159) and 'lj; is equal to
the right hand side of (8.160), then by a direct calculation we can easily
verify that (f,Ag) = ('lj;,g) for any g E D(A).
366 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Lemma 8. 8 Let f, g E Z*. Then

f E D(A*) and (A*- >.)f = g (8.167)

if and only if

f 0 E V and .D..T(>.)f0 = -H>..g, f = F* E>.f 0 - K>.g. (8.168)

The same result remains valid replacing A*, .D..T(>.), F* by A.j,, .D.(>.), F re-
spectively.
Proof. Suppose (8.167) holds. Then by Theorem 8.10 f0 E V,
j1 E W 1 ,2 (-h,O;V*) and
A;fo + f1(0) = Vo +go, (8.169)
a(s)A;f 0 - f 11 (s) = Aj 1(s) + g 1(s), (8.170)
f1(-h) = Aifo. (8.171)

-i:
We find from (8.170) that

e>.s f 1(s) = e->..h f 1( -h)+ ish e>.r a(T)A;f 0 dT e>.r g 1(T)dT. (8.172)

This equality with s = 0 and (8.169),(8.171) imply

0 0
{ e>- 7 a(T)A;f 0 dT=>.f 0 +g 0 -A;f0 -e->..hAif0 + { e>..rg 1(T)dT.
1-h 1-h
Substituting this in the right hand side of

.D..T(>.)fo = Afo- A;fo- e->..h Aifo- {o e>-sa(s)A;fods

1-h
we obtain
.D..T(A)fo =-go- {o e>.r g1(T)dT = -H>..g.
}_h
From (8.171),(8.172) we get

f 1(s) = e->..(h+s) Aif0 + r e>..(r-s)a(T)A;f0 dT

1-h
-ish e>..(r-s)g 1(T)dT = [F* E>.f°F(s)- [K>..gp(s). (8.173)

Clearly [F* E>.f0 - K>..g] 0 = f 0 • This and (8.173) imply (8.168). The
converse statement can be proved analogously.
8.12. EIGENMANIFOLD DECOMPOSITION 367

We use the following notations:

a-v(.6.) ={A; .6.(A.) is not invertible},
a-p(.6.r) ={A; .6.r(A.) is not invertible}.
Corollary 8. 3 (i) a-p(.6.) = a-v(A:Z,) = a-p(A), (ii) a-v(.6.r) = a-v(A*) =
a-v(Ar ).
Proof. The result follows from Lemmas 8.5 and 8.8.
In addition to Lemma 8.6 we need the following lemma (S. Nakagiri [112:
Proposition 6.1]).
Lemma 8. 9 For A. E C and k = 1, 2, ...
(8.174)

(8.175)

(8.176)

(8.177)

Proof. The equalities (8.174) can be shown with the aid of Lemma 8.6 (i),(ii)
inductively on k. (8.175) is established by a straightforward calculation. We
show (8.176). Let u E V. In view of (8.140)
(-1)kA(k+l)(')-1': hk+l ->.hA
(k+ 1)!u /\U-uk,ou+(k+ 1)1e 1u

+ (k ~ 1)! j_oh ( -s)k+ 1 e>.sa(s)A2uds. (8.178)

On the other hand

H>.FTfE>.u = [FTfE>.u] 0 + j_oh e>.s[FTfE>.ujl(s)ds. (8.179)

In view of (8.175)

[FTfE>.u] 0 = [TfE>.u] 0 = (-~)k [Eik)u] 0 = ok,ou, (8.180)

[FTfE>.ujl(s) = At[TfE>.ujl(-h- s) +ish a(T)A2[TfE>.u] 1 (T- s)dT

= (-k~)k A 1 [E?)ujl(-h-s)+ (-k~)k ish a(T)A2[E?)u] 1 (T-s)dT

1
= k!(h+s)ke->.(h+s)A 1
1 u+ k!
Js-h a(T)(s-T)ke>.(r-s)A2udT.
368 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

1:
Hence

j_oh e>. 8 [FTfE>.ujl(s)ds = :,e->.h (h + s)kdsA1u

+ :! j_oh j_sh e>.r a(T)(s- T)k A 2 udTds

- hk+1 ->.h
- (k + 1)! e
( -1)k+1 10
k+1 >.r
A1u + (k + 1)! -h T e a(T)A2 udT.

Combining this with (8.178),(8.179),(8.180) we complete the proof.

Proposition 8. 7 (i) For A E ap(A) and l = 1, 2, ... , F is injective on

Ker(A- A) 1•
(ii) For A E ap(Ar) and l = 1, 2, ... , F* is injective on Ker(A- Ar)i.
Proof. Suppose ¢ E Ker(A- A) and F¢ = 0. By Theorem 8.8 ¢ =
(¢0 , e>-· ¢ 0 ). Then ¢ 0 = [F¢] 0 = 0, and hence ¢ = 0. If¢ E Ker(A- A) 1, l >
1, and F¢ = 0, then (A- A) 1- 1 ¢ E Ker(A- A). By virtue of Corollary 8.1
F¢ E n~-:;1 D((Ar)k) and

F(A- A) 1- 1 ¢ =(A- Ar) 1- 1 F¢ = 0.

From the result of the case l = 1 it follows that (A-A) 1- 1¢ = 0. Continuing

this process we conclude ¢ = 0. The following theorem was first established
by M. C. Delfour and A. Manitius [45] for equations in finite dimensional
equations, and by S. Nakagiri [112] and S. Nakagiri and H. Tanabe [114] for
equations in infinite dimensional spaces.

Theorem 8.11 (i) If A E ap(~) = av(Ar) = ap(A), then for l = 1, 2, ...

Ker(A- A;-,) 1 = FKer(A- A) 1, (8.181)

dimKer(A- A;-,) 1 = dimKer(A- Ar (8.182)

(ii) If A E ap(~r) = ap(A*) = ap(Ar), then for l = 1, 2, ...

Ker(A- A*) 1 = F*Ker(A- Ar) 1, (8.183)

dimKer(A- A*)1 = dimKer(A- Arr (8.184)

Proof. It suffices to show only (ii). With the aid of Corollary 8.1 we can
inductively show that

(A- A*) 1F* = F*(A- Ar) 1 on D(A!r).

8.12. EIGENMANIFOLD DECOMPOSITION 369

Therefore we have

To prove the converse inclusion relation let 1/J E Ker(.\- A*) 1. Set <Pi =
(.\- A*)i1j;, j = 0, ... , l-1. Then A*<Pi- 1 = .\</Ji- 1 - ¢i, j = 1, ... , l, where
<Pt = 0. By Lemma 8.8

b.T(.\)¢~-1 = H>.</Jj, (8.185)

<Pi-1 = F* E>.¢~_ 1 + K>.¢i, (8.186)

for j = 1, ... , l. Hence <Pt-1 = F* E>.¢f_ 1 . Using (8.186) we can show

inductively that
l-1
<Pj-1 = L Ki-j+l F* E>.¢?
i=j-1
l-1
= L F*'T1-i+ 1E>.¢?, j = 1, ... 'l, (8.187)
i=j-1

where we used (8.174). Substituting this in the right hand side of (8.185)
and using (8.177) we get

b. (.\)"'C!
T 'I'J-1
= H >. ~
~
F*Ti-iE "'~ = ~ (-1)i-i t::,.Ci-i+1)(.\)"'~
>. A 'I'~ ~ (i _ · + 1)! T 'I'~'
~=J ~=J J

which is rewritten as
l-1 . .
~ (-1)~-J t::,.Ci-i+1)(.\)"'~ = 0. (8.188)
. L...J. (i- . + 1)! T '!'~
~=j-1 J

On the other hand

l-1
1/J =<Po = L F*T{E>.¢?
= F* L F* ¢8, e>-· L -:; ~ ¢? ·
l-1( )' ( l-1( ) ' )
-.~~Eli)¢?=
'/,, '/,,
(8.189)
i=O i=O

Theorem 8.8 and (8.188),(8.189) imply

1/J E F*Ker(.\- ATr

370 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

The assertion {8.184) follows from {8.183) and Proposition 8.7.

Let >.be an isolate point of a(A). Then 5. is an isolate point of a(A*).
We denote the corresponding spectral projections by P>. and Px:

Analogously if >. is an isolate point of a(AT ), then 5. is an isolate point

of a(Aj.). The corresponding spectral projections are denoted by P'f and
P'{*. As is easily seen

(8.190)
We set
M>. =ImP>., Mx = IrnPx, Mf = ImPJ, Mf* =ImP'{*. (8.191)
Theorem 8.12 (i) If>. is an isolate point of a(A) and a(Aj.), then
FP>. = P'f* F. (8.192)
(ii) If>. is an isolate point ofa(AT) and a(A*), then
F* P'f = P;F*. (8.193)
Proof. The assertions are simple consequences of Theorem 8.6.
The following proposition follows from Lemma 8.7 and (8.191).
Proposition 8. 8 (i) Let>. be a pole of (p,- A)- 1 of order k>.. Then 5. is
a pole of (p,- A*)- 1 of order k>.. We have
M>. = Ker(>.- Al;". Mx = Ker(5.- A*l>-,
dimM>. = dimM1::; oo.
(ii) Let>. be a pole of(p,-AT)- 1 of order kf. Then 5. is a pole of(p,-Aj.)- 1
of order kf. We have

Mf = Ker(>. -AT )k'f, Mf* = Ker(5. - Aj. li,

dimMf = dimMf*::; oo.
Remark 8. 2 In the special case of section 8.11 we have a(A) = a(AT) =
a(A•) = a(Aj.) in view of Theorem 8.9 (note that 1 is real and a(·) is real
valued), and (8.124) holds in p(A) or p(A). If moreover 0 E p(Ao) and
1 ::f:- 0, then by virtue of Proposition 8.6 F is an isomorphism from Z onto
z•. Therefore by (8.124) we have kf = k>.. Hence by Theorem 8.11 and
Proposition 8.8 we have
F M>. = M'{'*, F* Mf = Mx· (8.194)
8.13. SECOND STRUCTURAL OPERATOR 371

If A is an isolate point of o-(A) and dimimPA < oo, then A is a discrete

spectrum of A. We denote by o-d(A) the set of discrete spectra of A. If
A E o-d(A), then 5. E o-d(A*). If moreover 5. E o-d(AT), then A E o-d(AT),
and by Lemma 8. 7 and Theorem 8.11 we have

dA =dimMA = dimMI* = dimM1 = dimMf < oo.

Let {(h, ... , ¢d>.} and {1/J1, ... , '1/Jd>.} be the bases of MA and Mf respec-
tively. Since M1 = F* Mf, the dA x dA matrix {(¢i, F*'I/Ji)} is nonsingular.
Hence we may suppose (¢i, F*'I/Ji) = oi,j· If we set forgE Z
d>.
F\g = :E(g, F*'I/Ji)¢i,
i=l

then as is easily seen Iml\ = MA and Keri\ = Im(A- A)k>.. Hence we

have F\ = PA:
d>.
PAg = :E(g, F*'I/Ji)¢i· (8.195)
i=l

8.13 Second Structural Operator

Let g0 E H. We represent the solution of (8.1),(8.2) with g 1 (s) 0, f(t) 0
by u(t) = W(t)g 0 . Then by Theorem 8.1 the mapping g0 f---1 W(·)g 0 is a
= =
bounded linear operator from H to

L 2 (-h, T; V) n W 1 •2 (0, T; V*) c C([O, T]; H)

for any T > 0. W(t) is continuous in t E [0, oo) in the strong topology of H
and
W(O) =I, W(s) = 0 s E [-h, 0). (8.196)
W(·) is called the fundamental solution of (8.1),(8.2).
Set V = (D(A0 ),Hh;2 ,2 . In many cases V = V. In view of Theorem 8.1
=
if g = (g 0 , g1 ) E Z V x L 2 ( -h, 0; D(A 0 )), then the solution of (8.1),(8.2)
satisfies

iiuii£2(-h,T;D(Ao))nwt.2(0,T;H) ::; CT (ligll.z + llfll£2(o,T:H)) · (8.197)

Hence if g0 E V, then

W(·)g 0 E L 2 (0, T; D(Ao)) n W 1 •2 (0, T; H) c C((O, T]; V).

Analogously the fundamental solution WTO of the transposed equation
(8.100),
372 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

(8.101) is defined.
Let u be the solution of (8.1),(8.2) with g = (g 0 , g 1 ) E Z, f E L 2 (0, T; V*)
and let¢ E H. Then u, WT(·)¢ E L 2 ( -h, T; V) n W 1,2(0, T; V*). Hence in
view of Lemma 8.4 (u(T), WT(t- T)¢) is absolutely continuous in T E [0, t]
for any t > 0, and

1:
d
dT (u(T), WT(t- T)¢) = (u'(T), WT(t- T)¢)- (u(T), W.f,(t- T)¢)
= (Aou(T)+A1u(T-h)+ a(s)A2u(T+s)ds+f(T),

WT(t- T)¢) - ( u(T), A~WT(t- T)¢ + AiWT(t- T- h)¢

+ j_oh a(s)A~WT(t- T + s)¢ds)

~oh a(s)A2u(T+s)ds+f(T),WT(t-T)¢)
J:
= (A1u(T-h)+

- ( u(T), AiWT(t- T- h)¢+ a(s)A~WT(t- T + s)¢ds).

Integrating this equality from 0 to t we get

(u(t), ¢)- (g 0, WT(t)¢) =it (A 1u(T- h), WT(t- T)¢)dT

+lt (j_oh a(s)A2u(T + s)ds, WT(t- T)¢) dT

+lt (J(T), WT(t- T)¢)dT- it (u(T), AiWT(t- T- h)¢) dT

-it ( u(T), [oh a(s)A~WT(t- T + s)¢ds) dT = i;?i· (8.198)

In view of the initial conditions (8.1),(8.196)

l t-h
I1 + I4 = -h (A 1u(T), WT(t- T- h)¢)dT

-it (A1u(T), WT(t- T- h)¢)dT

= j_oh (A 1g 1(T), WT(t- T- h)¢) dT. (8.199)

8.13. SECOND STRUCTURAL OPERATOR 373

Analogously

I2 +Is= 1°-hlo{t (a(s)A2u(r + s), WT(t- r)if>)drds

-1°-hJot (u(r), a(s)A;WT(t- r + s)if>) drds

= Ioh 1t+s (a(s)A2u(r), WT(t- r + s)if>)drds

-1°-hlot (a(s)A2u(r), WT(t- r + s)if>)drds

=I: 1° (a(s)A2g 1(r), WT(t- r + s)if>) drds. (8.200)

From (8.198),(8.199),(8.200) it follows that

(u(t), if>)= (g 0 , WT(t)if>) + J:h (A1g 1(r), WT(t- r- h)if>) dr

+ lohlrh (a(s)A2g 1(r),WT(t-r+s)if>)dsdr

+ 1t (J(r), WT(t- r)if>)dr. (8.201)

In particular if g 1 = 0, f = 0, we have

for any g0 , if> E H. Therefore we obtain

W(t)* = WT(t). (8.202)

If g E Z and f E £ 2 (0, T; H), then we have

u E £ 2 ( -h, T; D(Ao)) n W 1 •2 (0, T; H) c C([O, T]; V).

Hence by (8.201),(8.202)

(u(t),if>) = (W(t)g0 ,if>)+ [ 0

h(W(t-r-h)A1g 1(r),¢>)dr

+I: Irh (a(s)W(t- r + s)A2g 1(r), if>) dsdr + 1\w(t- r)f(r), if>)dr.
374 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Consequently we have

u(t) = W(t)g 0 + 1°
-h
Ut(r)g 1(r)dr + {t W(t- r)f(r)dr,
Jo
(8.203)

where

Ut(T) = W(t- T- h)A1 + j_rh a(s)W(t- T + s)dsA2. (8.204)

By Theorem 8.1 we have

lluJI£2{-h,T;V)nW1,2(o,T;V•)::; CT (llgllz + JI!JIP(o,T;V•)) · (8.205)

Letting g = 0 we find that the mapping

f ~ W *f = 1· W(·- r)f(r)dr,

which is a bounded linear operator from L 2 (0, T; H) to L 2 ( -h, T; D(Ao))

nW 1 ,2 (0, T; H) can be extended to a bounded linear operator from
L 2 (0, T; V*) to L 2 ( -h, T; V) n W 1 ,2 (0, T; V*), where we set (W * f)(t) = 0
for t < 0 and Tis an arbitrary positive number. This extended operator is
also denoted by f ~ W *f.
Let g 0 E V and u be the solution of (8.1),(8.2) with g 1 = 0, f = 0. Let
t1 > 0, t2 > 0. Then u(t + t 2) is the solution of (8.1),(8.2) with the initial
value (u(t 2 ), Ut 2 ) E Z and the inhomogeneous term 0. Hence in view of
(8.203)
u(t + t2) = W(t)u(t2) + j_oh Ut(r)u(t2 + r)dr.
Since u(t) = W(t)g 0 , we have for g 0 E V

W(t1 + t2)g0= W(tl)W(h)g 0+ j_oh Utt (r)W(t2 + r)g 0dr. (8.206)

Let f 0 E V, p E L 2 ( -h, 0; H), t > 0. Suggested by A. Manitius [107:

p.7] we set

Gtf = ([Gtf] 0 , [Gtf] 1 (·)), (8.207)

[Gtf] 1 (s) = W(t + s)f0+ j_oh W(t + s + r)f 1(r)dr, s E [-h, 0],

(8.208)

(8.209)
8.13. SECOND STRUCTURAL OPERATOR 375

Since W(·)f 0 E £ 2 ( -h, T; D(A 0 )) n W 1 •2 (0, T; H) for any T > 0, we have

W(t + ·)/0 E £ 2 ( -h, 0; D(Ao)) n W 1•2 (( -h) V ( -t), 0; H),

and furthermore

IIW(t + ·)f0ll£2(-h,O;V)nwt.2((-h)v(-t),o;V•)
= IIW(·)f0 IIL2(t-h,t;V)nwt.2((t-h)vo,t;V•) $ CTI/01 (8.210)

for 0 $ t $ T. If we extend jl(s) to ( -oo, 0] putting jl(s) = 0 for s < -h,

then

(8.211)

where jl(T) = jl( -T). Since W * j1 E £ 2 ( -h, T; D(A0)) n W 1•2(0, T; H)

for any T > 0 we have

j_: W(t + · + T)f 1(T)dT E £ 2 ( -h, O; D(Ao)) n W 1•2(( -h) v (-t), O; H).

Furthermore

II(W * f )(t + ·)ll£2(-h,O;V)nwt.2((-h)V(-t),O;V•)

.1

.1
= IIW * f IIL2(t-h,t;V)nW1 •2((t-h)VO,t;V*)
, 1
$ CTII/ IIL 2(0,T;V•) $ CTII/ 11£2(-h,O;V•)· (8.212)

In view of (8.208),(8.210),(8.211),(8.212) we obtain

II [Gtfjl II PC -h,O;V)nWl,2(( -h)V( -t),o;v•) $ CTII!IIz• (8.213)

I[Gt/] 01 = I[Gtfjl(O)I $ CTII/IIz•. (8.214)

Consequently
IIGtfllz $ CTII/IIz•, 0$ t $ T. (8.215)
Therefore Gt which is a bounded linear operator from V x £ 2 ( -h, 0; H) to
V x £ 2 ( -h, 0; D(Ao)) can be extended to a bounded linear operator from
Z* to Z. This extended operator is again denoted by Gt.

Definition 8. 6 G = Gh is called the second structural operator.

376 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Analogously starting from the operator ct defined by

ctg = ([Gtg] 0 , [GtgJl(·)),
[Gig] 1 (s) = WT(t + s)g0+ j_oh WT(t + s + T)g 1(T)dT, s E [-h, 0],

[Gig] 0 = [Gig] 1 (0)

for g = (g 0 ,g 1) E V x L 2(-h,O;H), we obtain an operator Gt E .C(Z*,Z).

Since as is easily seen for j, g E V x L 2 ( -h, 0; H) we have

(Gtf,g) =(!,Gig),

we find that
G+-c*
t - t· (8.216)
Therefore if we define c+ = ct, then

c+ = c*. (8.217)

If we set u(h + s) = [Gj]l(s), s E [-h, 0], for f E V x L 2( -h, 0; H), then

u(t) = W(t)f 0 +lot W(t- T)j 1( -T)dT, t E [0, h].

Therefore

! u(t) = A 0 u(t) + A1u(t- h)+/_: a(T)A2u(t + T)dT + f 1 ( -t),

t E [0, h], (8.218)

u(O) = j0, u(s) =0 s E [-h, 0). (8.219)

By the initial condition (8.219) we get from (8.218)

ddt u(t) = A 0 u(t) + 1°

-t
a(T)A2u(t + T)dT + f 1 ( -t), t E [0, h]. (8.220)

The solution u of (8.218),(8.219) exists also for fEZ* and we have Gf =

(u(h), u(h + ·)). Since u E L 2(0, h; V) n W 1 •2 (0, h; V*), we have

ImG c W( -h, 0)
= {¢ E Z; ¢ 1 E L 2(-h, 0; V) n W 1 •2 ( -h, 0; V*), ¢ 0 = ¢ 1 (0) E H}.
Conversely for ¢ E W( -h, 0) we can find f E Z* satisfying G f = ¢, since
if we set u(t) = ¢ 1 (t- h), t E [0, h], then f = (! 0 , j 1) with f 0 = u(O) and
8.13. SECOND STRUCTURAL OPERATOR 377

Jl defined by the equation (8.220) is a desired one. Moreover it is easily

seen that this f is a unique element satisfying this condition. Therefore the
inverse c- 1 exists and is given by
[G-1¢]o = ¢1( -h), (8.221)
[G- 1¢] 1(s) = ¢ 11 (-s- h)- Ao¢ 1(-s- h)

-1° a(T)A2¢ 1 (-s+T-h)dT. (8.222)

Thus we have established the following proposition.

Proposition 8. 9 G is an isomorphism from Z* to W( -h, 0) and the in-
verse c- 1 is given by (8.221),(8.222). A similar result remains valid for G*
with Ao, A2 replaced by A 0, A2.
From the definition of S(t) and (8.203) it follows that for f EZ

[S(t)Jjl(s) = W(t + s)f0 + j_oh Ut+s(T)f 1 (T)dT. (8.223)

Theorem 8. 13 S(t)G = GST(t)*, G* S(t)* = ST(t)G*.

Proof. Let f E V x L 2(-h,O;V) c V x £ 2(-h,O;H). Then Gf E Z =
V x £ 2 ( -h, 0; D(Ao)),
and with the aid of (8.223) and (8.208),(8.209) with
t=h

[S(t)Gf] 1 (s) = W(t + s)[G/] 0 + /_: Ut+s(CJ)[Gfjl(CJ)da

= W(t + s) ( W(h)f 0 + j_oh W(h + T)l(T)dT)

+ /_: Ut+s(CJ) ( W(h + CJ)f 0 + j_oh W(h + + T)l(T)dT)

(J dCJ

= W(t + s)W(h)f0 + j_oh Ut+s(CJ)W(h + CJ)f 0 da

+ j_oh W(t + s)W(h + T)f 1 (T)dT

0 0 4

+ j_h Ut+s(CJ) j_h W(h + + T)f 1 (T)dTda = ~li·

(J

Using (8.206) we get

378 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Since
/Ut+s(cr)W(h + cr + T)j 1 (T)/ ::; CI/W(h + cr + T)j 1 (T)I/D(Ao)•
and by (8.197)

loh 1/W(h + CT + T)jl(T)I/b(Ao)do-

= I/W(·)Jl(T)I/i2(r,h+r;D(Ao)) :=; C//Jl(T)/1~,

we can apply Fubini's theorem to 14 to obtain

h + 14 = loh [W(t + s)W(h + T)j 1 (T)

+ loh Ut+s(cr)W(h+cr+T)l(T)do-]dT

= loh W(t + s + h + T)j 1 (T)dT,

where we again used (8.206). Therefore
[S(t)GJF(s)

= W(t + s + h)f 0 + loh W(t + s + h + T)j 1 (T)dT = [Gt+sf] 1 (s),

which implies
S(t)GJ = Gt+hf, (8.224)
holds for any f E V x 0; V). We can easily show that this equality
L 2 ( -h,
holds also for f E Z* by approximating f by a sequence in V x L 2 ( -h, 0; V).
Analogously we can show
Sr(t)G* = G;+h· (8.225)
The assertion of the theorem follows from (8.224) and (8.225).

We denote by cr0 (A) the set of the poles of the resolvent of A.

Theorem 8.14 (i) Suppose that cr0 (A*) = cr 0 (Ar). Then
crd(A*) = crd(Ar), (8.226)
and for each,\ E crd(A*) = crd(Ar) and l = 1, 2, ...
G*Ker(-\-A*l =Ker(-\-Ar)1, (8.227)
dimKer(,\- A*) 1 = dimKer(,\- Ar) 1 < oo. (8.228)
8.14. A SPECIAL CASE (CONTINUED) 379

In particular
G*M~=MI. (8.229)
(ii) Suppose 0" 0 (AT) = 0"0 (A). Then

O"d(AT) = O"d(A), (8.230)

and for each.\ E O"d(AT) = O"d(A) and l = 1, 2, ...

GKer(.\ -AT )1 = Ker(.\ - A) 1, (8.231)

dimKer(.\- A:Z.,) 1 = dimKer(.\- A) 1 < oo. (8.232)

In particular
GMI* =M>.. (8.233)
Proof. It suffices to prove only (i). Suppose .\ E O"d(A*). Then .\ E
(A*) = 0"0 (AT). Hence.\ E O"p(AT) and by Theorem 8.11 (8.228) holds.
0"0
Therefore the order of .\ as a pole of (.\-A*) - l coincides with that as a pole
of(.\- AT)-1, and.\ E O"d(AT)· Thus we proved O"d(A*) C O"d(AT)· The
converse inclusion relation is proved similarly. From Theorem 8.13 it follows
that if 'ljJ E D(A*) then G*'lj; E D(AT) and ATG*'lj; = G* A*'lj;. Hence

G*Ker(.\- A* )1 C Ker(.\- AT )1, l = 1, 2, .... (8.234)

By Proposition 8.9 G* is injective. Therefore we conclude (8.227) from

(8.228),(8.234).
The results of this section were first established by A. Manitius [107] for
equations in finite dimensional spaces, and extended to equations in infinite
dimensional spaces by S. Nakagiri [112], J.-M. Jeong, S. Nakagiri and H.
Tanabe [82] and S. Nakagiri and H. Tanabe [114].

8.14 A Special Case (Continued}

In the situation of section 8.11 we first consider in this section the problem
of the completeness of generalized eigenvectors of the infinitesimal generator
of the associated solution semigroup.
Definition 8. 7 Let A be a closed linear operator in a Banach space X.
The system of generalized eigenvectors of A is said to be complete if they
span a dense subspace of X.
We use the result of A. Manitius [107] for equations in a finite dimensional
space.
Let A0 , At, A 2 be n x n matrices and consider the problem (8.1),(8.2)
380 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

(or equations with more complicated delay terms). The associated solution
semigroup S(t) =etA is a Co semigroup in M 2 = Rn X L 2 ( -h, 0; Rn) . .6..(A)
defined by (8.106) is an n x n matrix valued entire function and in view of
Corollary 8.2 a( A) is the set of A at which .6..(A) is singular. Therefore a(A)
consists only of poles of the resolvent of A. The structural operator F is a
bounded linear operator from M 2 to itself.
Theorem of A. Manitius The system of generalized eigenvectors of A is
complete if and only ifF* is injective.
Suppose that A 0 is nonsingular, and A 1 = /'Ao, I' is a real number f:-
0, A2 = Ao, and that m(O) f:- 0, where m(A) is the function defined by
(8.152). If m(A) = 0, then .6..(A) = A is nonsingular, since A f:- 0 then
by assumption. Therefore .6..(A) is singular if and only if m(A) f:- 0 and
Ajm(A) E a(Ao), and hence

a(A) = {A; m(A) f:- 0, Ajm(A) E a(A0 )}.

Now we return to our situation of section 8.11.

Theorem 8. 15 (J.-M. Jeong [80]) Suppose 0 E p(Ao), I' f:- 0, m(O) f:- 0,
and that the immedding V C H is compact. If the system of generalized
eigenvectors of A 0 is complete in H, then the system of generalized eigen-
vectors of A is complete in Z.

Proof. By virtue of Proposition 8.6 F* is an isomorphism from V to V*.

Let a(Ao) = {ttn; n = 1, 2, ... } and Pn the spectral projection of Ao at
Mn· Set Hn = PnH and Aon = AoJH,.· Then Pn V = Hn· Set Zn =
Hn X L 2( -h, 0; Hn)· We denote the solution semigroup of

d u(t) = Aonu(t) + /'Aonu(t- h)+

dt 10 a(s)Aonu(t + s)ds,
-h

u(O) = g0 , u(s) = g 1 (s) s E [-h, 0),

where g = (g 0 ,g1 ) E Zn, and its infinitesimal generator by Sn(t) and An

respectively. Then by the commutativity of A 0 and Pn we can easily verify
that Sn(t) = S(t)Jz,., D(An) = D(A) n Zn, An = AJv(A,.)· By virtue of
Manitius' theorem the system of generalized eigenvectors of An is complete
in Zw Since a(Aon) = {tLn}, we have

a(An) = {A; m(A) f:- 0, Afm(A) = Mn}· (8.235)

We write for¢ E Z, '1/J E Z* Pn¢ = (Pn¢0 , Pn¢ 1 (·)), P~'ljJ = (P~'IjJ 0 ,P~'IjJ 1 (-)).
Suppose that 'ljJ E Z* is orthogonal to all generalized eigenvectors of A.
8.14. A SPECIAL CASE (CONTINUED) 381

Let ¢ E Zn be a generalized eigenvector of An. Then ¢ is a generalized

eigenvector of A and¢= Pn¢· Therefore
(P~'lj;, ¢) = ('1/J, Pn¢) = ('1/J, ¢) = 0,
from which P~ 'lj; = 0 follows for any n. This means that 'lj; 0 , 'lj; 1 ( s) are
orthogonal to all generalized eigenvectors of A 0 , and hence 'lj; = 0.
Lemma 8. 10 Suppose that the assumptions of Theorem 8.15 are satisfied
and the system of generalized eigenvectors of A 0 is complete in H. Then if
PJ f = 0 for any,\ E o-p(AT), then f = 0.
Proof. By Theorem 8.9 PJ is expressed as (8.195):
d>.
P'[ f = I:u, F'l/Jxi)<P>d,
i=1

where <P>..i, i = 1, ... , d>.., is a basis of Mf, '1/J)-.i, i = 1, ... , d>.., is a basis
of Mx, and d>.. = dimMf = dimMx. Therefore, if P'[ f = 0 for any
,\ E o-p(AT ), then (F* f, '1/J>.i) = 0 for any i = 1, ... , d>.. and 5. E o-(A). By
Theorem 8.15 we conclude F* f = 0, and hence f = 0 since F* is injective
under the present hypothesis.
Theorem 8.16 Suppose that the assumptions of Theorem 8.15 are satis-
fied and that the system of generalized eigenvectors of Ao is complete in H.
Then the problem (ST) is observable if and only if for any ,\ E o-p(AT)
Ker<l>~ n Ker~T(A.) = {0}. (8.236)
Proof. Suppose that (8.236) holds and <I>i)[ST(t)¢] 0 =0. Since
(f.L- AT)- 1 = 100
e-J.ttST(t)dt

if Ref.L is sufficiently large, we have

<I>~[(f.L- AT)- 1 ¢] 0 = 1 e-~'t<I>;[ST(t)¢] 0 dt

00
= 0 (8.237)

for these values of ft. By virtue of Theorem 8.9 we see that (8.237) holds
for any f.L E p(AT)· Let,\ Eo-p(AT) and set Qf =(AT- A.)P'[. Then
(QI}j =(AT- A.)JP'[, (QI)k>- = 0,
where k>.. is the order of,\ as a pole of (f.L- AT)- 1 . In view of (8.237) we
obtain

<I>~[(QI}J¢] 0 = -21 . { (f.L- ,\)J<f>;[(f.L- AT)- 1 ¢] 0 df.L = 0 (8.238)

7rZ jlll->..1=<
382 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

for j = 0, ... , k>. - 1. Set ¢ 1 = (QI)k,. - 1 ¢. Then

Therefore by Theorem 8.8

(8.239)

We have by (8.238)

(8.240)

Hence from (8.236),(8.239),(8.240) it follows that¢~ = 0, and hence ¢ 1 = 0.

Next set ¢ 2 = (QI)k>-- 2¢ = 0. Then ¢ 2 E Ker(A- AT)· By the same
argument as above, we find that ¢ 2 = 0. Continuing this procedure we
obtain P'[ ¢ = 0 for any A E CTp(AT ). Therefore by Lemma 8.10 we conclude
¢= 0.
Conversely suppose (ST) is observable and ¢ 0 E Kercl> 0 n Ker6.T(A) for
some A E CTp(AT)· Then¢= (¢0 , e>.·¢0 ) E Ker(A- AT)· Therefore

ST(t)¢ = e>.t¢,
and hence
cl>~[ST(t)¢]o = <P~(e>.t¢o) = e>.tc:p~(¢o) = 0.

From the hypothesis we conclude ¢ = 0, and hence ¢ 0 = 0.

We owe the above proof of the sufficiency part to S. Nakagiri and M.
Yamamoto [115] and T. Suzuki and M. Yamamoto [146].
Finally in this section we consider the case where the control space U is
a finite dimensional space eN. Then the controller <Po is represented by
N
<Pow =L wib~,
i=1

where w = (w1, ... ,WN) E eN and b? E H,i = 1, ... ,N. The adjoint
operator <I>(} : H ~ eN is given by

We suppose that the basis ¢>.i, i = 1, ... , d>., of Mf are arranged so that
¢>.i, i = 1, ... , n>., n>. =dim Ker(A- AT), is the basis of Ker(A- AT). Then
fori= 1, ... ,n>. ¢>.i = (¢~i,eA·¢~i), and ¢~i,i = 1, ... ,n>., is a basis of
Ker6.T(A).
8.15. F-COMPLETENESS 383

We refer to the Rank Condition introduced by H. 0. Fattorini [58]:

RANK CoNDITION: For any>. E ap(AT)

rank ( .(:.~1.'.~~!... ::: ... ~~~~: ~~). ) = n.x.

(4>~n>.' b~) . . . (4>~n>.' bfj.)
Theorem 8. 17 Suppose that the hypotheses of Theorem 8.16 are satisfied.
Then the problem (ST) is observable if and only if the Rank Condition is
satisfied.
Proof. Let ¢ 0 E Ker~T(>.) for some>. E ap(AT ). Then ¢ 0 = E~~ 1 ci¢~i for
some Ci E C, i = 1, · · ·, n.x, and we have

<t>;¢o = ((¢o,b~), ... , (¢o,bfj.))

= (~ Ci(¢~i• b~), ... ' ~ Ci(¢~i• bfj.))
= (c1, · · ·, Cn>J ( .(:.~1.'.~~.)... ::: ... ~~~~: ~~~ ) .
(</>~n>.'b~) ··· (</>~n>.'bfj.)
Therefore if the Rank Condition is satisfied, if>Q¢0 = 0 implies c1 = · · · =
en"' = 0, or ¢ 0 = 0. Hence (ST) is observable in view of Theorem 8.16. The
proof of the converse is similar.
Remark 8. 3 An example of an operator for which the Rank Condition is
satisfied is the realization of ~ in a disk in R 2 under the Dirichlet boundary
condition (see p.517 of S. Nakagiri [111]). In this case n.x = 2 for any
eigenvalue >. by G. N. Watson [158: 15.28].

8.15 F-completeness of Generalized

Eigenfunctions
Throughout this section we assume

a0 (~) = a 0 (AT) = a 0 (A), (8.241)

a0 (~T) = a 0 (A*) = aa(AT), (8.242)

where a 0 (~), a 0 (A), etc. are the sets of poles of ~0-1, the resolvent of A,
etc. Since>. E a 0 (A) if and only if,\ E a0 (A*), we have

(8.243)
384 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Lemma 8. 11 (i) If A is an isolate point of o-(A), then

KerP.x = {g E Z; ll(t-t)- 1 HJJ-Fg is holomorphic at f-t =A}.

(ii) If A is an isolate point of o-(AT), then

KerPJ = {g E Z; l:l.T(t-t)- 1 HJJ-F* g is holomorphic at f-t =A}.
(iii) If A is an isolate point of o-(A*), then

KerP~ = {! E Z*; l:l.T(t-t)- 1 H11-f is holomorphic at f-t =A}.

(iv) If A is an isolate point of o-(A:r), then

KerPJ* = {! E Z*; ll(t-t)- 1 H11-f is holomorphic at f-t =A}.

Proof. (i) In view of Lemma 8.7 g E KerP.x if and only if (t-t- A)- 1 g is
holomorphic at f-t = A. By (8.125) this is equivalent to the statement that
EJJ-ll(t-t)- 1 HJJ-Fg is holomorphic at f-t =A. By the definition (8.119)
Ell-ll(t-t)- 1 Hll-Fg = {ll(t-t)- 1 Hll-Fg, ell-·ll(t-t)- 1 Hll-Fg).
Thus the assertion follows.
(iii) f E KerP~ if and only if (t-t- A*)- 1 f is holomorphic at f-t =A. Analo-
gously to (8.127) we have

(t-t- A*)- 1 = F* EJJ-l:l.T(t-t)- 1 Hil- + Kw

Hence (t-t - A*) - 1 f is holomorphic at f-t = A if and only if the same holds
for F* EJJ-llT(t-t)- 1 HJJ-f· In view of (8.118),(8.119)

F* EJJ-l:l.T(t-t)- 1 H11-f

= ([EJJ-llT(t-t)- 1 Hil-f] 0 ,A~[EJJ-llT(t-t)- 1 Hil-fjl(-h- ·)

+ f_h a(T)A;[EJJ-l:l.T(t-t)- 1 Hll-fF(T- ·)dT)
= ( llT(t-t)- 1 Hll-f, ell-(-h-·) A;'l:l.T(t-t)- 1 H11-f

+ [h a(T)ell-(T-·) A;l:l.T(t-t)- 1 HJJ-fdT).

Thus the conclusion follows.
Assume that o-p(A) = o-0 (6.). Then by (8.241)

o-v(A) = o-0 (A). (8.244)

8.15. F-COMPLETENESS 385

It follows from Corollary 8.3 (i), (8.241),(8.244) that

(8.245)

Therefore we can define

(8.246)

If>. E crp(A), then by (8.244) >. is a pole of the resolvent of A, and hence >.
is a pole of the resolvent of A*. Therefore Mi can be defined. Similarly in
view of (8.245) we can define Mf.
Analogously, if crp(AT) = cr 0 (~T), then

and we can define

(8.247)

Moreover for>. E crp(AT) = crp(A*) we can define Mf*,Mx.

Proposition 8. 10 (i) If crp(A) = CT0 (~), then

Mj_ = {! E Z*; ~T(f.l)- 1 HI-'f is holomorphicin p(~T) U CT0 (~T)},

(MT*)j_ = {g E Z; ~T(f.l)- 1 HI-'F* gisholomorphicinp(~T) U CT0 (~T)}.

(ii) If CTp(AT) = CT0 (~T), then

(MT)j_ = {! E Z*; ~(f.l)- 1 Hl-'f is holomorphic in p(~) U cr 0 (~)},

(M*)j_ = {g E Z; ~(f.l)- 1 H~-'Fg is holomorphicin p(~) U cr 0 (~)}.

Proof. (i) As is easily seen f E M~ if and only if f E KerP5;. Hence

f E Mj_ if and only if f E KerP5; for any >. E crp(A). By Lemma 8.11
(iii) f E KerP5; if and only if ~T(f.l)- 1 Hl-'f is holomorphic at f.l = >.. By
assumption >. E CTp(A) if and only if>. E cr0 (~), and in view of (8.243) this
is equivalent to >. E cr0 (~T ). Thus the first assertion follows. Similarly
g E (MI*)j_ if and only if g E KerP{. Therefore g E (MT*)j_ if and only
if g E KerP[ for any >. E crv(AT). By Lemma 8.11 (ii) g E KerP[ if and
only if ~T(f.l)- 1 H~-'F* g is holomorphic at JL = >.. By Corollary 8.3 (i), the
assumption crp(A) = cr 0 (~), and (8.243), we have crp(AT) = cr 0 (~T)· Hence
,\ E crp(AT) if and only if>. E cr 0 (~T)· Thus the second assertion follows.
The proof of (ii) is similar.
386 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Propostion 8. 11 (i) If ap(A) = a 0 (A), then

(MT*).L = (FM).L. (8.248)

(ii) If ap(AT) =a 0 (AT), then

(M*).L = (F*MT).L. (8.249)

Proof. As is easily seen g E (FM).L if and only if F* g E M.L, which by

Proposition 8.10 (i) is equivalent tog E (MT* ).L. The proof of (ii) is similar.
Delfour and Manitius introduced the following concept ofF-completeness
and explained its importance in their papers [45] and [107].
Definition 8. 8 The system of generalized eigenfunctions of A (resp. AT)
is said to be F-complete (resp. F*-complete) ifFM (resp. F* MT) spans
a dense subspace of lmF (resp. ImF*).

Remark 8. 4 The system of generalized eigenfunctions of A (resp. AT)

is F-complete (resp. F*-complete) if and only if (F M).L = (ImF).L (resp.
(F* MT).L = (ImF*).L.

Definition 8. 9 The system of generalized eigenfunctions of A* (resp. AT)

is said to be complete in Cl(lmF*) (resp. Cl(ImF)) if M* (resp. MT*)
spans a dense subspace of lmF* (resp. ImF).

Remark 8. 5 By Theorem 8.11 or Remark 8.2 we have M* C lmF* and

MT* C lmF. The system of generalized eigenfunctions of A* (resp. AT)
is complete in ImF* (resp. ImF) if and only if (M*).L = (ImF*).L (resp.
(MT*).L = (ImF).L.

Theorem 8. 18 (i) Ifap(A) = a0 (A), then the system of generalized eigen-

functions of A is F-complete if and only if the system of generalized eigen-
functions of AT is complete in Cl(lmF).
(ii) If ap(AT) = a 0 (AT), then the system of generalized eigenfunctions of
AT is F* -complete if and only if the system of generalized eigenfunctions of
A* is complete in Cl(lmF*).

Proof. The assertions follow from Propositions 8.10 and 8.11.

Thorem 8.18 is due to M. C. Delfour and A. Manitius [45] in finite di-
mensional case and S. Nakagiri in infinite dimensional case (unpublished).
8.16. EXAMPLE AND REMARK 387

8.16 Example and Remarks

Let n be a bounded domain in Rn with smooth boundary. We set H =
L 2 (0) and V = HJ(O) = W~· 2 (0). Let a(·,·) be the sesquilinear form in
HJ(n) x HJ(n) defined by

a(u, v) =Jr.f[~ 8u8V ~ 8u

L....,; aii(x) Bx· Bx. + L....,; bi(x) Bx· v + c(x)uv] dx. (8.250)
n i,j=1 ~ J i=1 ~

We assume that the coefficients aij, bi, care real valued and satisfy
aii E C 1 (0), bi E C 1 (0), c E L 00 (0),
aii = aii• i, j = 1, ... , n, and a uniform ellipticity
n
2: ~i(x)t;it;i ~ eol£;1 2, t; = (6, ... , l;n) ERn
i,j=1

for some positive constant eo. As is well known this sequilinear form satis-
fies (8.66),(8.67). The associated operator A 0 E .C(HJ(O), H- 1 (0)), where
H- 1 (0) = HJ(O)*, has the following realization in L 2 (0). Let

Ao =-"-8( 8x·8) + "Mx)-

n
8x · aii(x)-
L....,;
8
8x· + c(x).
n
L....,;
i,j=1 J ~ i=1 ~

be the associated differential operator. Then the realization of A 0 in L 2 (0)

under the Dirichlet boundary condition is the restriction of Ao to D(Ao) =
H 2 (0) n HJ(O), where H 2 (0) = W 2 •2 (0).
Next, let A, L = 1, 2, be the restriction to HJ(O) of the second order

t
linear differential operator A., L = 1, 2, given by

A= - it1 8~i ( aii(x) a~J + bi(x) 8~i + c"(x),

where
aii = aji E C 1 (0), bi E C 1 (0), c" E L 00 (0).
Clearly A" E .C(HJ(O),H- 1 (0)) n.C(D(Ao),L 2 (0)) for L = 1,2.
We consider the following parabolic partial functional differential equation
8u(t,x)
Bt =Aou(t,x)+A1u(t-h,x)

+ [oh a(s)A2u(t + s, x)ds + f(t, x), t E (0, T), x E 0, (8.251)

u(t,x) = 0, t E (O,T), x E 80, (8.252)

u(O, x) = g 0 (x), u(s, x) = g 1 (s, x) s E [-h, 0), x E 0. (8.253)
388 CHAPTER 8. RETARDED FUNCTIONAL DIFFERENTIAL

Here we assume that

f E L~0 c([O, oo); H- 1 (0)), (8.254)

g 0 E L 2 (0), g 1 E L 2 ( -h, 0; H6(0)). (8.255)

Under these assumptions the problem (8.251)-(8.253) is written in the form

of the abstract equation (8.1),(8.2) in the space H- 1 (0), and the hypotheses
of the preceding sections are satisfied.
Next we consider the transposed system. The adjoint operator E A;
£(HJ(O), H- 1 (0)) of A;
is defined by

(u,A;v) = -a(u,v), u,v E HJ(O).

The realization of A; in L 2 (0) is characterized as the restiction of the formal
adjoint of A

n [)( [)) n[)

A*=-~-
LJ 8xJ·
aij(x)-
8x·t
+ LJ
~ -(bi(x) ·) + c(x)
8x·
i,j=1 i=1 t

to H 2 (0) nHJ(O). The adjoint operator A: of A., i = 1, 2, is the restriction

of the formal adjoint A:
of A to HJ(O). The transposed system of (8.251)-
(8.253) is given by

8v(t x)
8~ = A;v(t, x) + Aiv(t- h, x)
+ loh a(s)A2v(t + s, x)ds + 'l,b(t, x), t E (0, T), x E 0, (8.256)

v(t,x) = 0, t E (O,T), x E 80, (8.257)

v(O, x) = ¢P(x), v(s, x) = q'> 1 (s,x) s E [-h, 0), X E 0. (8.258)

This equation is also formulated as an abstract equation in the space H- 1 (0).

In order to treat equations with an L 1 -valued controller cl>0 E £(U, £1(0) ),
J.-M. Jeong [81] investigated the problem (8.251)-(8.253) in the space
w- 1,P(0) = W 1·P' (0)*, 1 < p < nj(n- 1), noting that we may consider
L 1 (0) c w- 1 ,P(0) by identifying f E L 1 (0) with the continuous conjugate
linear functional

(C0 (0) :J) wJ·v' (O) 3 v ~--> k fvdx E c. (8.259)

He first proved that the operator Ao E £(WJ'P(O), w- 1 ,P(0)) defined by

a(u, v) = -(Aou, v}, u E W~'P(O), v E W~·p' (0),

8.16. EXAMPLE AND REMARK 389

where a(·,·) is the same sesquilinear form as (8.250), generates an analytic

semigroup in w- 1 ,P(fl) using the results on complex interpolation of A. P.
Calderon [26] and T. Seeley [136]. Next with the aid of the result ofT.
Seeley [135] (See also J. Priiss and H. Sohr [130]) he proved that

\\(-Ao)i 6 \\.qw-l.v(n))::; Ce-rlsl, sER

for some constant 1 E (0, 7f/2). Hence applying the maximal regularity
result of G. Dore and A. Venni [55] and following the method of G. DiBlasio,
K. Kunisch and E. Sinestrari [51] he constructed the solution semigroup
associated with the problem (8.251)-(8.253), which is a C0 -semigroup in the
space
Zp,q = Hp,q x U(-h, 0; W~'P(O))
where Hp,q = (W~'P(O), w- 1 ,P(fl)h;q,q,q E (l,oo). The enlarged system
of the transposed equation (8.100) is an equation in the space
, 1 ,
Zp',q' = Hp',q' x Lq ( -h, 0; W ,p (0)).

The space Hp,q was characterized by J.- Y. Park and J.-M. Jeong [126]
except some critical cases as follows: if 2/q- 2 + 1/p i- 0

H _ { the closure of CJ(O) in B~~ 2 fq(O) 2/q < 1- 1/p,

(8.260)
p,q- B~~ 2 fq(O) 2/q > 1- ljp.
This is a special case of a more general result on the characterization of
the interpolation space (W~'P(O), w- 1 ,P(O))o,q as a Besov space, which is
considered to be a result of independent interest and is proved with the aid
of the reiteration theorem by J. L. Lions and J. Peetre [100] and the result
of P. Grisvard [71: Theorem 7.3] (see also H. 'friebel [154: Theorem 4.3.3])
after showing that LP(O) is of class K 1 ; 2 (W~'P(fl), w- 1 ,P(0)) in the sense
of Lions and Peetre [100]. If 1 < p < n/(n- 1), then p' > n, and hence
we can choose q so that 1- 2/q' > njp'. Then by (8.260) and Sobolev's
imbbeding theorem we have Hp',q' C C0 (0). Therefore we may consider
L 1 (0) C H;',q' = Hp,q by a similar identification to (8.259), and hence
c~>; E C(Hp',q', U*). The solution of the transposed equation (8.100) is a
continuous function with values in Hp',q', and so it is possible to establish
analogous results in the preceding sections for equations with an £ 1-valued
controller.
Bibliographical Remarks

Chapter 1. For the theory of semigroups refer to E. Hille and R. S. Phillips

[74), J. Goldstein [70), K. Yosida [166), S. D. Zaidman [167). Goldstein's
book [70) contains abundant applications including those to mathematical
physics. S. Agmon and L. Nirenberg [12) is a far-reaching study of solu-
tions of ordinary differential equations in Banach spaces with applications
to partial differential equations. Zaidman's book [167) inclides the same
kind of topics. The Hille-Yosida theory is extended to the case of infinites-
imal generators with nondense domain by G. Da Prato and E. Sinestrari
[43), Sinestrari [138),[139) and to the case in which infinitesimal generators
are multivalued linear operators by A. Yagi [164).
Chapter 2. Singular integrals of several variables are due to A. P. Calderon
and A. Zygmund [27). We owe the exposition of Hilbert transforms to E.
C. Titchmarsh [153) and A. Zygmund [168) in addition to [27). Hilbert
transforms for functions with values in a Banach space play an important
role. A Banach space with the property that the Hilbert transform for func-
tions with values in it is bounded is characterized as (-convex. See D. L.
Burkholder [22) and also H. Amann [16: Chapter III) for this topic. Using
this result G. Dore and A. Venni [54),[55) established a very useful maximal
regularity result for evolution equations. See also G. Dore [53), M. Giga, Y.
Giga and H. Sohr [69) and papers listed in its References. In the proof of
these results the boundedness of pure imaginary powers of operators is im-
portant. For this topic see R. Seeley [135), J. Pruss and H. Sohr [129), [130)
and A. Venni [155).
Chapter 3. The proof of the Gagliardo-Nirenberg inequality is due to L.
Nirenberg [118). See also A. Friedman [65: Part 1). The proofs of the inter-
polation inequalities are also found in S. Agmon [10: section 3), A. Friedman
[65: Part 1) and Nirenberg [116). A very useful inequality of Sobolev type
[140),[141) is established in Nirenberg [117).
Chapter 4. This chapter consists of a slightly simplified exposition of the
LP-estimates for general elliptic boundary value problems by S. Agmon, A.

391
392 BIBLIOGRAPHICAL REMARKS

Douglis and L. Nirenberg [11]. In [11] the Schauder estimates are also es-
tablished.
Chapter 5. Section 5.1 is due toM. Schechter [132],[133],[134]. Adjoint
boundary conditions were introduced and discussed by N. Aronszajn and
A. N. Milgram [17]. The proof of Theorem 5.2 is due to Schechter [133]. In
Schechter [134] the regularity of weak solutions and a theorem on the Fred-
holm alternative are established for general elliptic boundary value prob-
lems. Theorem 2.1 of S. Agmon [9] which is Theorem 5.3 in this book is
very famous. There is also a related result in S. Agmon and L. Nirenberg
[12]. N. Ikebe [76] proved that (5.47) holds under the assumption that the
coefficients are Holder continuous in the case of boundary value problems in
contiguous two domains. The paper [21] of F. E. Browder is very useful in
the study of elliptic boundary value problems in unbounded domains. In the
estimate of the kernel of the semigroup exp( -tA2 ) in section 5.3 we followed
R. Beals [19] who deduced an asymptotic behavior of the resolvent kernel
of a selfadjoint realization A of an elliptic operator from that of the kernel
of (A1 - >.)- 1 = (A- >. 1 )- 1 ···(A- >.t)- 1 where Aj,j = 1, ... , l, are the
l th roots of >. and l is an integer such that [l /2] > n/2m. The coefficients
of A are not assumed to be so smooth that A 1 is a differential operator in
the ordinary sense. The use of the operators L(·, D + rt), Bj(·, D + rt) was
suggested by L. Hormander [75] who established a sharp estimate for the
resolvent kernel of a selfadjoint realization of an elliptic operator. Since
the paper [75] is concerned with interior estimates, the boundary operators
Bj(·, D + rt) do not appear. The results of this section were announced
in [148]. This method was used in the study of parabolic equations in L 1
space in D. G. Park [122],[123] and D. G. Park and H. Tanabe [125]. G.
Di Blasio [46] proved the regularity of solutions in the maximal sense in
interpolation spaces in an LP setting for parabolic initial boundary value
problems with time independent coefficients. In [48] Di Blasio proved that
(L 1 (0), D(A1))o,v ~ (L 1 (0), ~· 1 (0) n W~' 1 (0))o,p, 0 < () < 1, 1::; p < oo,
for the realization A 1 of a second order elliptic operator in L 1 (0) under the
Dirichlet boundary condition, and characterized the interpolation spaces
D A 1 (0, 1) in terms of Besov spaces. The example in the last section of this
chapter was investigated in connection with parabolic equations of higher
order in the time variable by A. Favini and H. Tanabe [62]. The generation
of analytic semigroups in various function spaces is established by S. Cam-
panato [28],[29], P. Cannarsa, B. Terreni and V. Vespri [31], P. Cannarsa
and V. Vespri [32],[33],[34], F. Colombo and V. Vespri [36]. In D. G. Park
and S. Y. Kim [124] an example such that J01 IIAetAxlldt = oo for some
analytic semigroup etA is constructed.
Chapter 6. The results of this chapter are due toP. Acquistapace and B.
BIBLIOGRAPHICAL REMARKS 393

Terreni [1],[3],[6], [8] except for the last section. In [1],[8] various function
spaces are introduced and solutions are estimated in these spaces. Re-
lated results are in A. Yagi [159] and [161]. In [161] under the assumptions
(P1),(P2) and (P4) with k = 1 Yagi constructed the fundamental solution
using the fractional power of A(t) with a simpler proof. In this connection
we refer also to Yagi [162],[163].
Classical results on parabolic evolution equations are found in the books
R. W. Carroll [35], S. G. Krein [98] and H. Tanabe [149].
A. Buttu [23] constructed the evolution operator assuming that the map-
ping t 1-t A(t) is merely continuous in case where D(A(t)) is not necessarily
dense and an interpolation space DA(t)(() + 1,oo) between D(A(t)) and
D(A(t) 2 ) is independent oft. In [24] replacing D A( D)((), oo ), D A(o)(() + 1, oo)
by the continuous interpolation spaces D A(O) (e), D A(o) (() + 1) a further reg-
ularity result is obtained. The result is applied to the initial boundary value
problems for parabolic differential equations in noncylindrical domains.
H. Amann [14] constructed the fundamental solution under the assump-
tion that D(A(t).B) is constant for some {3 E (0, 1) and t 1-t A(t) is Holder
continuous. The method is to extend A(t) to an operator A(t) in an ex-
tended space called an extrapolation space so that D(A(t)) is constant.
Owing to the weakness of the smoothness hypothesis for A(t) in t this re-
sult can be applied to quasilinear equations. He continued this study in
[15]. See also Amann [16: Chapter V]. Also by considering the equation in
an extrapolation space G. Da Prato [37] and G. Da Prato and P. Grisvard
[40] showed the existence of a unique solution with the maximal regularity
property for a merely continuous inhomogeneous term under the assump-
tions that D(A(t)) is constant, t 1-t A(t) is only continuous and A(t)- 1 A(s)
has a bounded extension. G. DiBlasio [49] introduced the space D(O,p)
and considering the equation in it established the existence and uniqueness
of a solution with maximal regularity in the £P sense.
A. Lunardi and V. Vespri [106] established the optimal Holder regularity
of solutions of variational problems for second order parabolic equations
using the method of analytic semigroups. Lunardi [102] constructed an
evolution operator using new maximal regularity results in case D(A(t))
is constant but not necessarily dense and t 1-t A(t) is Holder continuous.
See also her book [104]. Based on the estimates of this evolution operator
various regularity results are established. In the subsequent paper [103]
she established further regularity results in case A(t) is of class cHa in t
with an application to nonhomogeneous initial boundary value problems for
parabolic differential equations.
G. Da Prato and E. Sinestrari [42] proved maximal Holder regularity re-
sults in case D(A(t)) is constant without using the fundamental solution but
using some sharp regularity results in the autonomous case. Sinestrari [138]
394 BIBLIOGRAPHICAL REMARKS

is a systematic study of autonomous parabolic equations with infinitesimal

generator not necessarily densely defined. He proved the unique solvability
with maximal regularity, and applied it to initial boundary value problems
for parabolic differential equations in the space of continuous functions.
The result of A. Yagi [159] is the most general one in the case where
domains are densely defined but totally variable. His main assumption is
that for some -1:::; ai < f3i:::; 1

JIA(t)(,\- A(t))- 1 dA(t)- 1 jdt- A(s)(>.- A(s))- 1 dA(s)- 1 jdsll

k
:::; C 2: 1>.1"; It- slf'; ·
i=1

Using the result of A. Yagi [164] some linear degenerate equations are
solved by A. Favini and A. Yagi [63],[64]. For degenerate equations see also
A. Favini and P. Plazzi [59],[60],[61] as well as papers quoted there.
A. Lunardi, E. Sinestrari and W. von Wahl [105] proved the unique solv-
ability with the optimal Holder regularity of initial-boundary value problems
with inhomogeneous boundary conditions for parabolic equations of first or-
der in the time variable and of arbitrary order in the space variables by a
semigroup method. W. von Wahl [157] established an estimate which is
applicable to parabolic equations with merely continuous coefficients under
the Dirichlet boundary conditions. He also proved an LP-Lq estimate.
Chapter 7. The first systematic study of evolution equations of hyperbolic
type was done by T. Kato [88]. J. R. Dorroh [56] simplified the original proof
of [88]. In Kato [89] the stability condition was weakened to quasistability:

k k
II R(>.j, A(tj)) :::; M II (,\j- f3(tj))- 1

j=1 j=1

for ,\1 > f3(t 1 ), • • ·, ,\k > f3(,\k), where f3 is an upper-integrable function,
and the conditions on B(·) and S(·) in the relation

S(t)A(t)S(t)- 1 = A(t) + B(t)

is weakened to the hypotheses that B(·) is strongly measurable with IIBOII
upper-integrable and S(-) is the indefinite strong integral of S(-) with
JISOIIc(Y,X) upper-integrable. Such a generalization is useful in the study
of nonlinear equations. The study of nonlinear hyperbolic evolution equa-
tions was originated by T. Kato [88] and developed by Kato [89],[90], K.
Kobayashi and N. Sanekata [96], Sanekata [131]. For other papers on non-
linear hyperbolic evolution equations see the references of [92]. Assuming
BIBLIOGRAPHICAL REMARKS 395

that {A(t)} is quasistable and only a certain kind of measurability on the

mapping t f-t A(t) S. Ishii [77] constructed the evolution operator using the
Yosida approximation. G. Da Prato and M. Iannelli [41] proved the unique
solvabilty under somewhat different assumptions. P. Cannarsa and G. Da
Prato [30] established a result which can be applied to Kolmogoroff equa-
tions. A. Arosio [18] proved that under the assumption that X is reflexive
and A E BV(O, +oo; B(Y, X)) the evolution operator exists and Duhamel's
principle holds iff E BVtoc([O, +oo); X).
Chapter 8. The well-posedness of retarded functional differential equa-
tions with highest order delay was established by G. Di Blasio, K. Ku-
nisch and E. Sinestrari [51]. In [50] they discuss properties of the solu-
tion semigroup, and show conditions for its noncompactness, differentia-
bility and nondifferentiability. In [137] Sinestrari solved the problem in
the space of Holder continuous functions with maximal regularity in case
a(·) E L 1 ( -h, 0) and the coefficient operators are not necessarily densely de-
fined. Nonautonomous equations were solved by DiBlasio [47]. P. Vernole
[156] solved second order parabolic differential equations with time depen-
dent coefficients in Holder spaces. Theorem 8.5 was proved in Y. Fujie and
H. Tanabe [66]. In [152] the fundamental solution with strong regularity
property is constructed in case where a(·) is Holder continuous applying the
result of J. Pruss [128] in each interval [nh, (n + l)h], n = 0, 1, 2, .... But no
maximal regularity of solutions is obtained there. Using this fundamental
solution the second structural operator is directly defined and its relation
with the first structural operator F is established in [82],[114]. S. Nakagiri
[113] improved the results of [80], [115] especially in an important case of
finite dimensional control spaces. For quasilinear equations we refer to J.
Yong and L. Pan [165]. J. K. Hale's book [73] provides excellent information
on retarded ordinary functional differential equations.
Bibliography

[1] P. Acquistapace: Evolution operators and strong solutions of abstract

linear parabolic equations, Diff. Int. Eqns. 1 (1988), 433-457.

[2] P. Acquistapace: Abstract linear nonautonomous parabolic equations:

A survey, Proc. Bologna Conference on Differential Equations in Ba-
nach Spaces, Lecture Notes in Pure and Appl. Math. 148, ed. G.
Dore, A. Favini, E. Obrecht, A. Venni, Marcel Dekker, New York-
Basel-Hong Kong, 1993, 1-19.

[3] P. Acquistapace and B. Terreni: Some existence and regulaity results

for abstract non-autonomous parabolic equations, J. Math. Anal. Appl.
99 (1984), 9-64.

[4] P. Acquistapace and B. Terreni: Characterization of Holder and Zyg-

mund classes as interpolation spaces, preprint, Dipartimento di Math.,
Univ. Padova, 61 (1984).

[5] P. Acquistapace and B. Terreni: Maximal space regularity for abstract

linear non-autonomous parabolic equations, J. Func. Anal. 60 (1985),
168-210.

[6] P. Acquistapace and B. Terreni: On fundamental solutions for ab-

stract parabolic equations, Differential Equations in Banach spaces,
Bologna, 1985, Lecture Notes in Math. 1223, ed. A. Favini, E.
Obrecht, Springer-Verlag, 1986, 1-11.

[7] P. Acquistapace and B. Terreni: Holder classes with boundary condi-

tions as interpolation spaces, Math. Z. 195 (1987), 451-471.

[8] P. Acquistapace and B. Terreni: A unified approach to abstract linear

non-autonomous parabolic equations, Rend. Sem. Univ. Padova 78
(1987), 47-107.

397
398 BIBLIOGRAPHY

[9] S. Agmon: On the eigenfunctions and on the eigenvalues of general

elliptic boundary value problems, Comm. Pure Appl. Math. 15 (1962),
119-147.

[10] S. Agmon: Lectures on Elliptic Boundary Value Problems, D. van

Nostrand, Princeton-Toronto-New York-London, 1965.

[11] S. Agmon, A. Douglis and L. Nirenberg: Estimates near the boundary

for solutions of elliptic partial differential equations satisfying general
boundary conditions, I, Comm. Pure Appl. Math. 12 (1959), 623-727.

[12] S. Agmon and L. Nirenberg: Properties of solutions of ordinary differ-

ential equations in Banach space, Comm. Pure Appl. Math. 16 (1963),
121-239.

[13] H. Amann: Dual semigroups and second order linear elliptic boundary
value problems, Israel J. Math. 45 (1983), 225-254.

[14] H. Amann: On abstract parabolic fundamental solutions, J. Math.

Soc. Japan 39 (1987), 93-116.

[15] H. Amann: Parabolic evolution equations in interpolation and extrap-

olation spaces, J. Func. Anal. 78 (1988), 233-270.

[16] H. Amann: Linear and Quasilinear Parabolic Problems, I, Abstract

Linear Theory, Monographs in Math. 89, Birkhauser Verlag, Basel-
Boston-Berlin, 1995.

[17] N. Aronszajn and A. N. Milgram: Differential operators on Rieman-

nian manifolds, Rend. Circ. Mat. Palermo, Ser.2 2 (1953), 266-325.

[18] A. Arosio: Duhamel's principle for temporally inhomogeneous evo-

lution equations in Banach space, Nonlinear Anal. T.M.A. 8 (1984),
997-1009.

[19] R. Beals: Asymptotic behavior of the Green's function and spectral

function of an elliptic operator, J. Func. Anal. 5 (1970), 484-503.

[20] C. Bernier and A. Manitius: On semigroups in Rn x LP corresponding

to differential equations with delays, Canad. J. Math. 30 (1978), 897-
914.

[21] F. E. Browder: On the spectral theory of elliptic differential operators,

I, Math. Anal. 142 (1961), 22-130.
BIBLIOGRAPHY 399

[22] D. L. Burkholder: A geometrical condition that implies the existence

of certain singular integrals of Banach-space-valued functions, Con-
ferece on Harmonic Analysis in Honour of Antoni Zygmund, Chicago,
1981, ed. W. Beckner, A. P. Calderon, R. Fefferman and P. W. Jones,
270-286, Belmont, Cal., 1983, Wadsworth.
[23] A. Buttu: On the evolution operator for a class of non-autonomous
abstract parabolic equations, J. Math. Anal. Appl. 170 (1992), 115-
137.
[24] A. Buttu: A construction of the evolution operator for a class of
abstract parabolic equations, Dyn. Syst. Appl. 3 (1994), 221-234.
[25] P. L. Butzer and H. Berens: Semigroups of Operators and Approxi-
mation, Springer-Verlag, New York-Berlin-Heidelberg-Tokyo, 1967.
[26] A. P. Calderon: Intermediate spaces and interpolation, the complex
method, Studia Math. 24 (1964), 113-190.
[27] A. P. Calderon and A. Zygmund: On the existence of certain singular
integrals, Acta. Math. 88 (1952), 85-139.
[28] S. Campanato: Generation of analytic semigroups, in the Holder
topology, by elliptic operators of second order with Neumann boundary
condition, Le Matematiche 35 (1980), 61-72.
[29] S. Campanato: Generation of analytic semigroups by elliptic operators
of second order in Holder spaces, Ann. Sci. Norm. Sup. Pisa 8 (1981),
495-512.
[30] P. Cannarsa and G. Da Prato: Some results on abstract evolution
equations of hyperbolic type, Proc. Bologna Conference on Differential
Equations in Banach Spaces, Lecture Notes in Pure and Appl. Math.
148, ed. G. Dore, A. Favini, E. Obrecht, A. Venni, Marcel Dekker,
New York-Basel-Hong Kong, 1993, 41-50.
[31] P. Cannarsa, B. Terreni and V. Vespri: Analytic semigroups generated
by non-variational elliptic systems of second order under Dirichlet
boundary conditions, J. Math, Anal. Appl. 112 (1985), 56-103.
[32] P. Cannarsa and V. Vespri: Analytic semigroups generated on Holder
spaces by second order elliptic systems under Dirichlet boundary con-
ditions, Ann. Mat. Pura Appl. (IV), 140 (1985), 393-415.
[33] P. Cannarsa and V. Vespri: Generation of analytic semigroups by
elliptic operators with unbounded coefficients, SIAM J. Math. Anal.
18 (1987), 857-872.
400 BIBLIOGRAPHY

[34] P. Cannarsa and V. Vespri: Generation of analytic semigroups in the

LP topology by elliptic operators in Rn, Israel J. Math. 61 (1988),
235-255.

[35] R. W. Carroll: Abstract Methods in Partial Differential Equations,

Harper & Row, New York-Evanston-London, 1969.

[36] F. Colombo and V. Vespri: Generation of analytic semigroups in

Wk,P(O) and Ck(O), Diff. Int. Eqns. 9 (1996), 421-436.

[37] G. Da Prato: Abstract differential equations and extrapolation spaces,

Infinite-Dimensional Systems, Lecture Notes in Math. 1076 ed. A.
Dold and B. Eckmann, Springer-Verlag, 1983, 53-61.

[38] G. Da Prato and P. Grisvard: Sommes d 'operateurs lineaires et

equations differentielles operationelles, J. Math. Pures Appl. 54
(1975), 305-387.
[39] G. Da Prato and P. Grisvard: Equations d'evolution abstraites non-
lineares de type parabolique, Ann. Mat. Pura Appl. (IV) 120 (1979),
329-396.

[40] G. Da Prato and P. Grisvard: Maximal regularity for evolution equa-

tions by interpolation and extrapolation, J. Func. Anal. 58 (1984),
107-124.

[41] G. Da Prato and M. Iannelli: On a method for studying abstract

evolution equation in the hyperbolic case, Comm. Partial Diff. Eqns.
1 (1976), 585-608.

[42] G. Da Prato and E. Sinestrari: Holder regularity for non-autonomous

abstract parabolic equations, Israel J. Math. 42 (1982), 1-19.

[43] G. Da Prato and E. Sinestrari: Differential operators with nondense

domains, Ann. Scuola Norm. Sup. Pisa, Ser.4 14 (1987), 285-344.
[44] M. C. Delfour and A. Manitius: The structural operator F and its role
in the theory of retarded systems, I, J. Math. Anal. Appl. 73 (1980),
466-490.

[45] M. C. Delfour and A. Manitius: The structural operator F and its role
in the theory of retarded systems, II, J. Math. Anal. Appl. 74 (1980),
359-381.

[46] G. DiBlasio: Linear parabolic evolution equations in LP -spaces, Ann.

Mat. Pura Appl. (IV) 138 (1984), 55-104.
BIBLIOGRAPHY 401

[47] G. Di Blasio: Nonautonomous functional differential equations in

Hilbert spaces, Nonlinear Anal. T.M.A. 9 (1985), 1367-1380.

[48] G. Di Blasio: Analytic semigroups generated by elliptic operators in

L 1 and parabolic equations, Osaka J. Math. 28 (1991), 367-384.

[49] G. Di Blasio: Interpolation and extrapolation spaces and parabolic

equations, Proc. Bologna Conference on Differential Equations in Ba-
nach Spaces, Lecture Notes in Pure and Appl. Math. 148, ed. G.
Dore, A. Favini, E. Obrecht, A. Venni, Marcel Dekker, New York-
Basel-Hong Kong, 1993, 51-58.

[50] G. Di Blasio, K. Kunisch and E. Sinestrari: Retarded abstract equa-

tions in Hilbert spaces, Infinite-Dimensional Systems, Lecture Notes
in Math. 1076 ed. A. Dold and B. Eckmann, Springer-Verlag, 1983,
71-77.

[51] G. DiBlasio, K. Kunisch and E. Sinestrari: L 2 -regularity for parabolic

partial integrodifferential equations with delay in the highest-order
derivatives, J. Math. Anal. Appl. 102 (1984), 38-57.

[52] G. DiBlasio, K. Kunisch and E. Sinestrari: Stability for abstract linear

functional differential equations, Israel J. Math. 50 (1985), 231-263.

[53] G. Dore: LP regularity for abstract differential equations, Lecture

Notes in Math. 1540, ed. H. Komatsu, Springer-Verlag, 1991, 25-38.

[54] G. Dore and A. Venni: On the closedness of the sum of two closed
operators, Math. Z. 196 (1987), 189-201.

[55] G. Dore and A. Venni: Maximal regularity for parabolic initial-

boundary value problems in Sobolev spaces, Math. Z. 208 (1991), 297-
308.

[56] J. R. Dorroh: A simplified proof of a theorem of Kato on linear

evolution equations, J. Math. Soc. Japan 27 (1975), 474-478.

[57] S. D. Efdel'man and S.D. lvasisen: Investigation of the Green matrix

for a homogeneous parabolic boundary value problem, Trudy Moscow
Mat. Obsc. 23 (1970), 179-234 (in Russian); English translation:
Trans. Moscow Math. Soc. 23 (1970), 179-242.

[58] H. 0. Fattorini: On complete controllablity of linear systems, J. Diff.

Eqns. 3 (1967), 391-402.
402 BIBLIOGRAPHY

[59] A. Favini and P. Plazzi: On some abstract degenerate problems of

parabolic type -1:the linear case, Nonlinear Anal. T.M.A. 12 (1988),
1017-1027.
[60] A. Favini and P. Plazzi: On some abstract degenerate problems
of parabolic type -2:the nonlinear case, Nonlinear Anal. T.M.A. 13
(1989), 23-31.
[61] A. Favini and P. Plazzi: On some abstract degenerate problems of
parabolic type -3: applications to linear and nonlinear problems, Osaka
J. Math. 27 (1990), 323-359.
[62] A. Favini and H. Tanabe: Linear parabolic differential equations of
higher order in time, Proc. Bologna Conference on Differential Equa-
tions in Banach Spaces, Lecture Notes in Pure and Appl. Math. 148,
ed. G. Dore, A. Favini, E. Obrecht, A. Venni, Marcel Dekker, New
York-Basel-Hong Kong, 1993, 85-92.
[63] A. Favini and A. Yagi: Space and time regularity for degenerate evo-
lution equations, J. Math. Soc. Japan 44 (1992), 331-350.
[64] A. Favini and A. Yagi: Multivalued linear operators and degenerate
evolution equations, Annali Mat. Pura Appl. (IV) 163 (1993), 353-
384.
[65] A. Friedman: Partial Differential Equations, Krieger, Huntington,
New York, 1976.
[66] Y. Fujie and H. Tanabe: On some parabolic equations of evolution in
Hilbert space, Osaka J. Math. 10 (1973), 115-130.
[67] E. Gagliardo: Proprieta di alcune classi di funzioni in piu variabili,
Ricerche di Mat. 7 (1958), 102-137.
[68] E. Gagliardo: Ulteriori proprietd di alcune classi di funzioni in piu
variabili, Ricerche di Mat. 8 (1959), 24-51.
[69] M. Giga, Y. Giga and H. Sohr: LP-estimates for the Stokes system,
Lecture Notes in Math. 1540, ed. H. Komatsu, Springer-Verlag, 1991,
55-67.
[70] J. A. Goldstein: Semigroups of Linear Operators and Applications,
Oxford Univ. Press, New York & Clarendon Press, Oxford, 1985.
[71] P. Grisvard: Equations differentielles abstraites, Ann. Sci. Ecol. Norm.
Sup., 4e serie 2 (1969), 311-395.
BIBLIOGRAPHY 403

[72] D. Guidetti: On elliptic systems in Ll, Osaka J. Math. 30 (1993),

397-429.
[73] J. K. Hale: Theory of Functional Differential Equations, Springer-
Verlag, New York-Heidelberg-Berlin, 1977.
[74] E. Hille and R. S. Phillips: Functional Analysis and Semi-groups,
Amer. Math. Soc., Providence, 1957.
[75] L. Hormander: On the Riesz means of spectral functions and eigen-
function expansions for elliptic differential opertors, Some Recent
Advances in the Basic Sciences 2 (Proc. Annual Sci. Conf., Belfer
Grad. School Sci., Yeshiva Univ., New York, 1965-1966), 155-202.
[76] N. Ikebe: On elliptic boundary value problems with discontinuous
coefficients, Mem. Fac. Sci., Kyushu Univ. 21 (1967), 167-184.
[77] S. Ishii: Linear evolution equations duj dt + A(t)u = 0: a case where
A(t) is strongly uniform-measurable, J. Math. Soc. Japan 34 (1982),
413-424.
[78] S. D. Ivasisen: Green's matrices of boundray value problems for
Petrovski{ parabolic systems of general form, I Mat. Sbornik (N.S.)
114(156) (1981), 110-116 (in Russian); English translation: Math.
USSR-Sbornik 42 (1982), 93-144.
[79] S. D. Ivasisen: Green's matrices of boundary value problems for Petro-
vski{ parabolic systems of general form, II (Russian), Mat. Sbornik
(N.S.) 114(156) (1981), 523-565, 654-655 (in Russian); English trans-
lation: Math. USSR-Sbornik 42 (1982), 461-498.
[80] J.-M. Jeong: Spectral properties of the operator associated with a
retarded functional differential equation in Hilbert space, Proc. Japan
Acad. 65 (1989), 98-101.
[81] J.-M. Jeong: Retarded functional differential equations with L 1 -valued
controller, Funkcial. Ekvac. 36 (1993), 71-93.
[82] J.-M. Jeong, S. Nakagiri and H. Tanabe: Structural operators and
semigroups assiciated with functional differential equations in Hilbert
spaces, Osaka J. Math._ 30 (1993), 365-395.
[83] F. John: General properties of solutions of linear elliptic partial dif-
ferential equations, Proc. Symp. Spectral Theory and Differential
Problems, Oklahoma Agricultural and Mechanical College, Stillwa-
ter, Oklahoma, 1951, 113-175.
404 BIBLIOGRAPHY

[84] F. John: Plane Waves and Sherical Means Applied to Partial Differen-
tial Equations, Interscience Publishers, Inc., New York & Interscience
Publishers, Ltd., London, 1955.

[85] T. Kato: Fractional powers of dissipative operators, J. Math. Soc.

Japan 13 (1961), 246-274.

[86] T. Kato: Perturbation Theory for Linear Operators, Springer-Verlag,

Belin-Heidelberg-New York, 1966.

[87] T. Kato: Nonlinear semigroups and evolution equations, J. Math.

Soc. Japan 19 (1967), 508-520.

[88] T. Kato: Linear evolution equations of "hyperbolic" type, J. Fac. Sci.

Univ. Tokyo, Sec. I, 17 (1970), 241-258.

[89] T. Kato: Linear evolution equations of "hyperbolic" type, II, J. Math.

Soc. Japan 25 (1973), 648-666.

[90] T. Kato: The Cauchy probelm for quasi-linear symmetric hyperbolic

systems, Arch. Rat. Mech. Anal. 58 (1975), 181-205.
[91] T. Kato: Quasi-linear equations of evolution, with applications to
partial differential equations, Lecture Notes in Math. 448, Springer-
Verlag, 1975, 25-70.

[92] T. Kato: Abstract evolution equations, linear and quasilinear, revis-

ited, Lecture Notes in Math. 1540, ed. H. Komatsu, Springer-Verlag,
1991, 103-125.
[93] T. Kato and H. Tanabe: On the abstract evolution equation, Osaka
Math. J. 14 (1962), 107-133.

[94] N. Kerzman: Holder and LP estimates for solutions of Bu = f in

strongly pseudo convex domains, Comm. Pure Appl. Math. 24 (1971),
301-379.
[95] K. Kobayashi: On a theorem for linear evolution equations of hyper-
bolic type, J. Math. Soc. Japan 31 (1979), 647-654.

[96] K. Kobayashi and N. Sanekata: A method of iterations for quasi-

linear evolution equations in nonreflexive Banach sapces, Hiroshima
Math. J. 19 (1989), 521-540.

[97] Y. Komura: Nonlinear semigroups in Hilbert space, J. Math. Soc.

Japan 19 (1967), 493-507.
BIBLIOGRAPHY 405

[98] S. G. Krein: Linear Differential Equations in Banach Space, N auka,

Moscow, 1967 (in Russian); English translation: Transl. Math. Mono-
graphs 29, Amer. Math. Soc., Providence, 1972.
[99] J. L. Lions and E. Magenes: Problemes aux limites non homogenes
et applications, Dunod, Paris, vol. 1, 2, 1968, vol. 3, 1970.
[100] J. L. Lions and J. Peetre: Sur une classe d'espaces d'interpolation,
Inst. Hautes Etudes Sci. Publ. Math. 19 (1964), 5-68.
[101] A. Lunardi: Interpolation spaces between domains of elliptic opera-
tors and spaces of continuous functions with applications to nonlinear
parabolic equations, Math. Nachr. 121 (1985), 295-318.
[102] A. Lunardi: On the evolution operator for abstract parabolic equations,
Israel J. Math. 60 (1987), 281-314.
[103] A. Lunardi: Differentiability with respect to (t, s) of the parabolic
evolution operator, Israel J. Math. 68 (1989), 161-184.
[104] A. Lunardi: Analytic Semigroups and Optimal Regularity in
Parabolic Problems, Progress in Nonlinear Diff. Eqns. Appl. 16,
Birkhiiuser, Basel-Boston-Berlin, 1995.
[105] A. Lunardi, E. Sinestrari and W. von Wahl: A semigroup approach
to the time dependent parabolic initial-boundary value problem, Diff.
Int. Eqns. 5 (1992), 1275-1306.
[106] A. Lunardi and V. Vespri: Holder regularity in variational parabolic
non-homogeneous equations, J. Diff. Eqns. 94 (1991), 1-40.
[107] A. Manitius: Completeness and F-completeness of eigenfunctions
associated with retarded functional differential equations, J. Diff. Eqns.
35 (1980), 1-29.

[108] A. Manitius: Necessary and sufficient conditions of approximate con-

trollability for general linear retarded systems, SIAM J. Control Op-
tim. 19 (1981), 516-532.
[109] A. Manitius: F-controllability and observability of linear retarded sys-
tems, Appl. Math. Optim. 9 (1982), 73-95.
[110] F. J. Massey III: Abstract evolution equations and the mixed prob-
lem for symmetric hyperbolic systems, Trans. Amer. Math. Soc. 168
(1972), 165-188.
406 BIBLIOGRAPHY

[111] S. N akagiri: Identifyability of linear systems in Hilbert spaces, SIAM

J. Control and Optim. 21 (1983), 501-530.

[112] S. Nakagiri: Structural properties of functional differential equations

in Banach spaces, Osaka J. Math. 25 (1988), 353-398.

[113] S. Nakagiri: Controllability and identifiability for linear time-delay

systems in Hilbert space, preprint.

[114] S. Nakagiri and H. Tanabe: Structural operators and eigenmanifold

decomposition for functional differential equations in Hilbert spaces,
to appear in J. Math. Anal. Appl.

[115] S. Nakagiri and M. Yamamoto: Controllability and observability of

linear retarded systems in Banach space, Int. J. Control. 49 (1989),
1489-1504.

[116] L. Nirenberg: Remarks on strongly elliptic partial differential equa-

tions, Comm. Pure Appl. Math. 8 (1955), 648-674.

[117] L. Nirenberg: Estimates and existence of solutions of elliptic equa-

tions, Comm. Pure Appl. Math. 9 (1956), 509-530.

[118] L. Nirenberg: On elliptic partial differential equations, Ann. Scuola

Norm. Sup. Pisa, Ser. 3 13 (1959), 113-161.

[119] N. Okazawa: Remarks on linear evolution equations of hyperbolic type

in a Hilbert space, preprint.

[120] N. Okazawa and A. Unai: Linear evolution equations of hyperbolic

type in Hilbert space, SUT J. Math. 29 (1993), 51-70.

[121] N. Okazawa and A. Unai: Abstract quasilinear evolution equations

in a Hilbert space, with applications to symmetric hyperbolic systems,
SUT J. Math. 29 (1993), 263-290.

[122] D. G. Park: Initial-boundary value problem for parabolic equation in

£1, Proc. Japan Acad. 62 (1986), 178-180.

[123] D. G. Park: Regularity in time of the solution of parabolic initial-

boundary value problem in L 1 space, Osaka J. Math. 24 (1987), 911-
930.

[124] D. G. Park and S. Y. Kim: Integrodifferential equations with time

delay in Hilbert space, Comm. Korean Math. Soc. 7 (1992), 189-207.
BIBLIOGRAPHY 407

[125] D. G. Park and H. Tanabe: On the asymptotic behavior of solutions

of linear parabolic equations in £ 1 space, Ann. Scuola Norm. Sup.
Pisa 14 (1987), 587-611.
[126] J.-Y. Park and J.-M. Jeong: Supplement to the paper "Retarded func-
tional differential equations with £ 1 -valued controller", Funkcial. Ek-
vac. 38 (1995), 267-275.
[127] A. Pazy: Semigroups of Linear Operators and Applications to Partial
Differential Equations, Springer-Verlag, New York-Berlin-Heidelberg-
Tokyo, 1983.

[128] J. Pruss: On resolvent operators for linear integrodifferential equations

of Volterra type, J. Int. Eqns. 5 (1983), 211-236.

[129] J. Pruss and H. Sohr: On operators with bounded imaginary powers

in Banach spaces, Math. Z. 203 (1990), 429-452.

[130] J. Pruss and H. Sohr: Imaginary powers of elliptic second order differ-
ential operators in £P-spaces, Hiroshima Math. J. 23 (1993), 161-192.

[131] N. Sanekata: Abstract quasi-linear equations of evolution in nonre-

ftexive Banach spaces, Hiroshima Math. J. 19 (1989), 109-139.

[132] M. Schechter: Integral inequalities for partial differential operators

and functions satisfying general boundary conditions, Comm. Pure
Appl. Math. 12 (1959), 37-66.
[133] M. Schechter: General boundary value problems for elliptic partial
differential equations, Comm. Pure Appl. Math. 12 (1959), 457-486.

[134] M. Schechter: Remarks on elliptic boundary value problems, Comm.

Pure Appl. Math. 12 (1959), 561-578.
[135] R. Seeley: Norms and domains of the complex power A~, Amer. J.
Math. 93 (1971), 299-309.
[136] R. Seeley: Interpolation in £P with boundary conditions, Studia Math.
44 (1972), 47-60.

[137] E. Sinestrari: On a class of retarded partial differential equations,

Math. Z. 186 (1984), 223-246.
[138] E. Sinestrari: On the abstract Cauchy problem of parabolic type in
spaces of continuous functions, J. Math. Anal. Appl. 107 (1985), 16-
66.
408 BIBLIOGRAPHY

[139] E. Sinestrari: On the Hille- Yosida operators, Evolution Equations,

Control Theory, and Biomathematics, Han sur Lesse, 1991, Lecture
Notes in Pure Appl. Math. 155, Marcel Dekker, 1994, 537-543.
[140] S. L. Sobolev: On a theorem of functional analysis, Mat. Sbornik
(N.S.) 4 (1938), 471-497.
[141] S. L. Sobolev: Application of Functional Analysis in Mathematical
Physics, Leningrad National Univ., Leningrad, 1950 (in Russian);
English translation: Transl. Math. Monographs 7, Amer. Math. Soc.,
Providence, 1963.

[142] P. E. Sobolevskii: Equations of parabolic type in Banach space, Trudy

Moscow Mat. Obshch. 10 (1961), 297-350 (in Russian); English trans-
lation: Amer. Math. Soc. Transl. 49 (1965), 1-62.
[143] V. A. Solonnikov: On boundary value problems for linear parabolic
systems of differential equations of general form, Trudy Mat. Inst.
Steklov 83 (1965), 3-163 (in Russian); English translation: Proc.
Steklov Inst. Math. 83 (1965), 1-184.
[144] H. B. Stewart: Generation of analytic semigroups by strongly elliptic
operators, Trans. Amer. Math. Soc. 199 (1974), 141-162.

[145] H. B. Stewart: Generation of analytic semigroups by strongly elliptic

operators under general boundary conditions, Trans. Amer. Math.
Soc. 259 (1980), 299-310.
[146] T. Suzuki and M. Yamamoto: Observability, controllability and feed-
back stabilizability for evolution equation I, Japan J. Appl. Math. 2
(1985), 211-228.

[147] T. Takagi: Lectures on Algebra, Kyoritsu Shuppan Publ. Company,

Tokyo, 1946 (in Japanese).
[148] H. Tanabe: On Green's functions of elliptic and parabolic boundary
value problems, Proc. Japan Acad. 48 (1972), 709-711.

[149] H. Tanabe: Equations of Evolution, Iwanami Shoten, Tokyo, 1975 (in

Japanese); English translation: Pitman, London, 1979.
[150] H. Tanabe: Functional Analysis, II, Jikkyo Shuppan Publ. Company,
Tokyo, 1981 (in Japanese).
[151] H. Tanabe: Structural operators for linear delay-differential equations
in Hilbert space, Proc. Japan Acad. 64 (1988), 265-266.
BIBLIOGRAPHY 409

[152] H. Tanabe: Fundamental solutions for linear retarded functional dif-

ferential equations in Banach space, Funkcial. Ekvac. 35 (1992), 149-
177.
[153] E. C. Titchmarsh: Introduction to the Theory of Fourier Integrals,
Clarendon Press, Oxford, 1937.
[154] H. Triebel: Interpolation Theory, Function Spaces, Differential Oper-
ators, North-Holland, Amsterdam-New York-Oxford, 1978.
[155] A. Venni: A counterexample concerning imaginary powers of linear
operators, Lecture Notes in Math. 1540, ed. H. Komatsu, Springer-
Verlag, 1991, 381-387.
[156] P. Vernole: A time dependent parabolic initial boundary value delay
problem, J. Int. Eqns. Appl. 6 (1994), 427-444.
[157] W. von Wahl: The equation u' + A(t)u = f in a Hilbert space and
LP -estimates for parabolic equations, J. London Math. Soc., Ser. II 25
(1982), 483-497.
[158] G. N. Watson: A Theatise on the Theory of Bessel Functions, 2nd
edition, Cambridge Univ. Press, 1958.
[159] A. Yagi: On the abstract evolution equation of parabolic type, Osaka
J. Math. 14 (1977), 557-568.
[160] A. Yagi: Remarks on proof of a theorem of Kato and Kobayashi on
linear evolution equations, Osaka J. Math. 17 (1980), 233-244.
[161] A. Yagi: Fractional powers of operators and evolution equations of
parabolic type, Proc. Japan Acad. 64 (1988), 227-230.
[162] A. Yagi: Parabolic evolution equations in which the coefficients are
the generators of infinitely differentiable semigroups, Funkcial. Ekvac.
32 (1989), 107-124.

[163] A. Yagi: Parabolic evolution equations in which the coefficients are the
generators of infinitely differentiable semigroups, II, Funkcial. Ekvac.
33 (1990), 139-150.
[164] A. Yagi: Generation theorem of semigroup for multivalued linear op-
erators, Osaka J. Math. 28 (1991), 385-410.
[165] J. Yong and L. Pan: Quasi-linear parabolic partial differential equa-
tions with delays in the highest order spatial derivatives, J. Austral.
Math. Soc. (Series A) 54 (1993), 174-203.
410 BIBLIOGRAPHY

[166] K. Yosida: Functional Analysis, Springer-Verlag, Berlin-Heidelberg-

New York, 1966.
[167] S. D. Zaidman: Abstract Differential Equations, Research Notes in
Math. 36, Pitman, San Francisco-London-Melbourne, 1979.
[168] A. Zygmund: Trigonometrical Series, Cambrigde Univ. Press, 1968.
List of Symbols
cm(o), 67 II lla,oo, 68
C[}"(O), 67 I lv,n, 72
Bm(O), 67 I lv, 72
Ba(O), 67 I lm,p,n, 75
cm(n), 105 II llm,p,n, 75
C[}"(O), 105 I lm,p, 75
Bm(n), 67 II llm,p' 75
Ba(n), 67 II II~ oo n, 10o
wm,P(O), 75 II 11~·00: 101
W~'P(O), 119 [ ]m_'1/p,p,an, 118
Hm,oo(O), 100 II II*' 342
H 2 (0), 387 ( )T, 16
HJ(O), 387 !*(·), 31,32
H- 1 (0), 387 f3J(·), 33
w- 1,P(0), 388 (3f(·), 34
wm- 1fP,P(80), 118 log+, 47
L[ocCO), 1 supp, 48
C([O, T]; X), 222 Rf-,105
ca([O, T]; X), 222 On, 46
C 1 ([0, T]; X), 222 Vv, 1
Cl+a([O, T]; X), 222 :F, 13
C((O, T]; X), 223 E,108
ca((O, T]; X), 223 "/j, 118
C 1 ((0, T]; X), 223 D(T), 3
cHa((O, T]; X), 223 R(T), 3
LP(O, T; X), 223 a(T), 3
I lm,oo,n, 67 p(T), 3
II llm,oo,n, 67 ap(T), 3
I la,oo,n, 67 R().., A), 221
II lla,oo,n, 67 £(X, Y), 3
I lm,oo, 68 £(X), 3
I llm,oo, 68 G(X,M,/3), 4
I la,oo, 68 I IIL~(X)l 7

411
412 LIST OF SYMBOLS

(X, Y)o,p, 7
D A((), oo), 246
~(·), 350
~T(·), 350
(}(~), (}(~T ), 355
p(~), p(~T ), 355
(}p(~), (}p(~T ), 367
(/d(A), 371
(}o(A), 378
(}o(~), (}o(~T ), 383
E>., 352
T>., 352
H>., 352
K>., 352
c~ 355
Im, 361
Ker, 358
P>., 370
P[, 370
M>., 370
M~, 370
MI, 370
M T*>. '
370
M, 385
M*, 385
Index
A-admissible subspace, 285 of evolution equation, 237
Accretive operator, 310 of retarded functional
m-, 310 differential equation, 371
regularly, 343
Adjoint boundary value problem, Hilbert transform, 29
175 Hille's representation, 4
Approximately controllable system,
355 Inequality:
Gagliardo-Nirenberg's, 75, 81
Calder6n-Zygmund decomposition, GB.rding's, 343
22 Hausdorff-Young's, 2
Complementing Condition, 131 Young's, 2
Completeness: Infinitesimal generator, 4
F-, 386 Interpolation:
in Cl(ImF), 386 complex, 217
of generalized eigenvectors, 379 pair, 7
Convolution, 2 real----space, 8
Invariant subspace, 285
Dirichlet set, 170
Dissipative operator, 310 Lebesgue point, 2
m-, 310 Linearly independent polynomials
modulo a polynomial, 131
Ellipticity Condition, 130
Elliptic operator, 121 Mapping:
strongly, 184 of type (p, q), 15
uniformly, 130 of weak type (p, oo), 15
Equimeasurable function, 31 of weak type (p, q), 15
Equivalent boundary operators, 174 Maximal nonnegative subspace, 320
Evolution operator, 237
Normal boundary conditions, 169
Fractional power, 11
Fundamental solution: Observable system, 356
of differential equation, 121 Open set:

413
414 INDEX

locally regular of class ()2m, 186 Yosida's representation, 4

of class em' 102 Yosida approximation, 270
uniformly regular of class em'
102 Zygmund class, 256

Parameter of regularity, 1
Positive operator, 8

Rank Condition, 383

Ray of the minimal growth of the
resolvent, 183
Realization of infinitesimal
generator, 285
Regular sequence, 1
Restricted cone property, 119
Retarded resolvent, 350
Riesz kernel, 43
Root Condition, 130

Semigroup, 3
Co-, 3
analytic, 5
Smoothness Condition, 130, 131
Solution:
classical, 223
mild, 341
strict, 223
strong, 223
Solution semigroup, 336
Spectral projection, 361
Stability constants, 289
Stable family of operators, 289
Structural operator, 351
second, 375

Theorem:
Dini's, 3
Lax-Milgram's, 343
Manitius', 380
Rellich's, 118
Trace on the boundary, 118

Y-valued solution, 296