Sie sind auf Seite 1von 365

Monographs in Mathematics

Vol. 89

Managing Editors:
H.Amann
UniversiHit Zurich, Switzerland
K. Grove
University of Maryland, College Park
H. Kraft
UniversiUit Basel, Switzerland
P.-L. Lions
Universite de Paris-Dauphine, France

Associate Editors:
H. Araki, Kyoto University
J. Ball, Heriot-Watt University, Edinburgh
E Brezzi, Universita di Pavia
K.C. Chang, Peking University
N. Hitchin, University of Warwick
H. Hofer, Universitat Bochum
H. Knorrer, ETH Zurich
K. Masuda, University of Tokyo
D. Zagier, Max-Planck-Institut Bonn
Herbert Amann

Linear
and Quasilinear
PamboHc Problems
Volume I
Abstract Linear Theory

1995 Birkhauser Verlag


Basel· Boston· Berlin
Author:
Institut fUr Mathematik
Universitat Ziirich
Winterthurerstrasse 190
8057 Ziirich
Switzerland

A CIP catalogue record for this book is available from the Library of Congress,
Washington D.C., USA

Die Deutsche Bibliothek - CIP-Einheitsaufnahme


Amann, Herbert:
Linear and quasilinear parabolic problems I Herbert Amann. -
Basel; Boston; Berlin: Birkhauser
Vol. 1. Abstract linear theory. - 1995
(Monographs in mathematics; Vol. 89)
ISBN-13: 978-3-0348-9950-5 e-ISBN-13: 978-3-0348-9221-6
DOl: 978-3-0348-9221-6

This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, re-use of illustrations, broadcasting,
reproduction on microfilms or in other ways, and storage in data banks. For any kind of use permis-
sion of the copyright owner must be obtained.

© 1995 Birkhauser Verlag Basel, P.O. Box 133, CH-4010 Basel, Switzerland
Printed on acid-free paper produced of chlorine-free pulp
Softcover reprint of the hardcover I st edition 1995
987654321
Aber - so fragen wir - wird es bei der Ausdehnung des mathematischen Wissens
fUr den einzelnen Forscher nicht schliel3lich unmoglich, alle Teile dieses Wissens zu
umfassen? Ich mochte als Antwort darauf hinweisen, wie sehr es im Wesen der
mathematischen Wissenschaft liegt, daB jeder wirkliche Fortschritt stets Hand in
Hand geht mit der Auffindung scharferer Hilfsmittel und einfacherer Methoden,
die zugleich das Verstandnis friiherer Theorien erleichtern und umstandliche altere
Entwicklungen bcseitigen, und daB es daher dem einzelnen Forscher, indem er sich
diese scharferen Hilfsmittel und einfacheren Methoden zu eigen macht, leichter
gelingt, sich in den verschiedenen Wissenszweigen der Mathematik zu orientieren,
als dies fUr irgend eine andere Wissenschaft der Fall ist. 1

David Hilbert {1862-1943}

1 But - so we ask - given tne expansion of mathematical knowledge, will it eventually not be
impossible for the individual researcher to encompass all parts of this knowledge? As an answer
I want to point out how much it lies in the character of mathematical science that all real progress
is intimately tied to the discovery of sharper tools and simpler methods that, at the same time,
facilitate the comprehension of earlier theories and remove complicated older developments, and
that therefore the researcher, adopting these sharper tools and simpler methods, succeeds in
getting more easily acquainted with the diverse branches of mathematics than this is the case
for any other field of science.
Preface

In this treatise we present the semigroup approach to quasilinear evolution equa-


tions of parabolic type that has been developed over the last ten years, approxi-
mately. It emphasizes the dynamic viewpoint and is sufficiently general and flexible
to encompass a great variety of concrete systems of partial differential equations
occurring in science, some of those being of rather 'nonstandard' type. In partic-
ular, to date it is the only general method that applies to noncoercive systems.
Although we are interested in nonlinear problems, our method is based on
the theory of linear holomorphic semigroups. This distinguishes it from the theory
of nonlinear contraction semigroups whose basis is a nonlinear version of the Hille-
Yosida theorem: the Crandall-Liggett theorem. The latter theory is well-known
and well-documented in the literature. Even though it is a powerful technique
having found many applications, it is limited in its scope by the fact that, in
concrete applications, it is closely tied to the maximum principle. Thus the theory
of nonlinear contraction semigroups does not apply to systems, in general, since
they do not allow for a maximum principle. For these reasons we do not include
that theory.
Our approach is strongly motivated by the concept of weak solutions of dif-
ferential equations. In fact, as one of the applications of our general results we
cventually develop a theory of weak solutions for noncoercive quasilinear parabolic
systems in divergence form in an Lp-setting. This is in contrast to the standard
L 2 -setting for coercive problems, that is, unfortunately, not suitable for noncoer-
cive systems. Moreover, even in regular situations, where, in principle, we could
work directly within the framework of strong solutions, the theory of weak so-
lutions is of great importance. For instance, in connection with global existence
it allows for a priori estimates in 'weak norms', which facilitates the establish-
ing of those bounds considerably. For this reason we develop a general 'reflexive'
theory taylored for applications in an Lp-setting. We treat evolution equations in
spaces of continuous functions only marginally. An exposition of the latter theory,
emphasizing fully nonlinear problems and strong solutions, can be found in the
monograph by A. Lunardi [LungS].
In order to obtain results that are sufficiently general and flexible to be ap-
plicable to a wide variety of concrete problems, we need a considerable amount of
viii Preface

preparation. For this reason our treatise is divided in three volumes, carrying the
respective titles:

Abstract Linear Theory,


Function Spaces and Linear Differential Operators,
Nonlinear Problems.

In the first volume we give a thorough discussion of linear parabolic evolution


equations in general Banach spaces. This is the abstract basis for the nonlinear
theory. The second volume is devoted to concrete realizations of linear parabolic
evolution equations by general parabolic systems. There we discuss the various
function spaces that are needed and useful, and the generation of analytic semi-
groups by general elliptic boundary value problems. The last volume contains the
abstract nonlinear theory as well as various applications to concrete systems, il-
lustrating the scope and the flexibility of the general results. Of course, each one
of the three volumes contains much material of independent interest related to our
main subject.
In writing this book I had help from many friends, collegues, and students. It
is a pleasure to thank all of them, named or unnamed. I am particularly indebted
to P. Quittner and G. Simonett, who critically and very carefully read, not only
the whole manuscript of this first volume but also many earlier versions that were
produced over the years and will never be published, and pointed out numerous
mistakes and improvements. Large parts of the first volume, and of earlier versions
as well, were also read and commented on by D. Daners, J. Escher, and P. Guidotti.
Their constructive criticism, observations, and suggestions for improvements were
enormously helpful. Of course, I am solely responsible for all remaining mistakes.
My son Andreas gave expert advice for taming the computer and kept open
an emergency line. Finally, this book would never have appeared without the
invaluable help of my 'comma sniffer', whose contributions are visible on every
page. My heartfelt thanks go to both of them.
Through many years I obtained financial support by Schweizerischer Na-
tionalfonds, that is gratefully acknowledged. It enabled me to maintain an active
group in this field of research and to bring in visitors from outside. These contacts
were enormously beneficial for my work. I also express my gratitude to Birkhauser
Verlag, in particular to Th. Hintermann, for the agreeable collaboration.

Zurich, December 1994 Herbert Amann


Contents

Preface . . . . vii

Introduction xiii

Notations and Conventions


1 Topological Spaces . . . 1
2 Locally Convex Spaces . . . . 2
3 Complexifications . . . . . . 4
4 Unbounded Linear Operators 6
5 General Conventions . . . . . 7

Chapter I Generators and Interpolation


1 Generators of Analytic Semigroups
1.1 Properties of Linear Operators 10
1.2 The Class H(Ell Eo) .. 11
1.3 Perturbation Theorems 14
1.4 Spectral Estimates . . . 15
1.5 Compact Perturbations 20
1.6 Matrix Generators . . . 21
2 Interpolation FUnctors
2.1 Definitions . . . . . . . . 24
2.2 Interpolation Inequalities 25
2.3 Retractions . . . . . . . . 26
2.4 Standard Interpolation Functors 28
2.5 Continuous Injections 30
2.6 Duality Properties . . 30
2.7 Compactness..... 31
2.8 Reiteration Theorems 31
2.9 Fractional Powers and Interpolation 32
2.10 Semigroups and Interpolation . . . 33
2.11 Admissible Interpolation Functors . 35
x Contents

Chapter II Cauchy Problems and Evolution Operators


1 Linear Cauchy Problems
1.1 Holder Spaces . . . . . . . . 40
1.2 Existence and Regularity Theorems 43

2 Parabolic Evolution Operators


2.1 Basic Properties . . . . . . . . . 45
2.2 Determining Integral Equations. . . 47

3 Linear Volterra Integral Equations


3.1 Weakly Singular Kernels. . . . 48
3.2 Resolvent Kernels . . . . . . . 50
3.3 Singular Gronwall Inequalities 52
4 Existence of Evolution Operators
4.1 A Class of Parameter Integrals . . . . 53
4.2 Semigroup Estimates . . . . . . . . . 55
4.3 Construction of Evolution Operators . 57
4.4 The Main Result . . . . . . . . . . 63
4.5 Solvability of the Cauchy Problem 66
5 Stability Estimates
5.1 Estimates for Evolution Operators 68
5.2 Continuity Properties of Mild Solutions 71
5.3 Holder Estimates . . . . . . . . 72
5.4 Boundedness of Mild Solutions 74

6 Invariance and Positivity


6.1 Yosida Approximations . . . . . . . . . 75
6.2 Approximations of Evolution Operators 77
6.3 Invariance. . . . . . . . . 80
6.4 Orderings and Positivity. . . . . . . . 84

Chapter III Maximal Regularity


1 General Principles
1.1 Sobolev Spaces . . . . . . . . . . . 88
1.2 Absolutely Continuous Functions . 89
1.3 Generalized Solutions . . . . 91
1.4 Trace Spaces . . . . . . . . . 92
1.5 Pairs of Maximal Regularity 94
1.6 Stability . . . . . . . . . . . . 96
Contents xi

2 Maximal HOlder Regularity


2.1 Singular HOlder Spaces 98
2.2 Semigroup Estimates 102
2.3 'frace Spaces . . . . 106
2.4 Estimates for K A . . . 109
2.5 Maximal Regularity 113
2.6 Nonautonomous Problems. 117
3 Maximal Continuous RegUlarity
3.1 Necessary Conditions . . . . . . . 121
3.2 Higher Order Interpolation Spaces 123
3.3 Estimates for K A . . 124
3.4 Maximal Regularity . . . . . . . 126

4 Maximal Sobolev Regularity


4.1 Temperate Distributions . . . . . 128
4.2 Fourier 'fransforms and Convolutions 130
4.3 The Hilbert 'fransform . . . . . . . . . 135
4.4 UMD Spaces and Fourier Multipliers. 141
4.5 Properties of UMD Spaces 144
4.6 Fractional Powers . . . . . 147
4.7 Bounded Imaginary Powers 162
4.8 Perturbation Theorems . . 168
4.9 Sums of Closed Operators . 173
4.10 Maximal Regularity . . . . 180

Chapter IV Variable Domains


1 Higher Regularity
1.1 Properties of Differentiable Functions . . . . . . 194
1.2 General Solvability Results for Cauchy Problems 195
1.3 Estimates for Evolution Operators . . . . . . 198
1.4 Evolution Operators on Interpolation Spaces 204
1.5 The Cauchy Problem . . . . . . . . . . . . . 207
2 Constant Interpolation Spaces
2.1 Semigroup and Convergence Estimates. 211
2.2 Assumptions and Consequences. . . . 214
2.3 Construction of Evolution Operators . 218
2.4 Estimates for Evolution Operators . 227
2.5 The Cauchy Problem . . . . . . . . 230
2.6 Abstract Boundary Value Problems 233
xii Contents

3 Maximal Regularity
3.1 Abstract Initial Boundary Value Problems 242
3.2 Isomorphism Theorems . . . . . . . . . . . 245

Chapter V Scales of Banach Spaces


1 Banach Scales
1.1 General Concepts 250
1.2 Power Scales . . . 255
1.3 Extrapolation Spaces 261
1.4 Dual Scales . . . . . . 267
1.5 Interpolation-Extrapolation Scales 275
2 Evolution Equations in Banach Scales
2.1 Semigroups in Interpolation-Extrapolation Scales. 286
2.2 Parabolic Evolution Equations in Banach Scales 294
2.3 Duality . . . . . . . . . . 297
2.4 Approximation Theorems . . 300
2.5 Final Value Problems . . . . 304
2.6 Weak Solutions and Duality. 307
2.7 Positivity . . . . . . . . . . . 312
2.8 General Evolution Equations 314
Bibliography . . 321
List of Symbols 329
Index . . . . . . . 333
Nichts setzt clem Fortgang cler Wis-
senschaft mehr Hinclernis entgegen
als wenn man zu wissen glaubt, was
man noch nicht weiB.l
Georg Christoph Lichtenberg
(1742-1799)

Introduction

Partial differential equations of parabolic type are encountered in a variety of


problems in mathematics, physics, chemistry, biology, and many other scientific
subjects in which irreversible processes can be adequately described by mathemat-
ical models. For this reason parabolic equations have been thoroughly studied and
there is a considerable mathematical literature in this field. However, most of the
research has been concentrated on the study of a single second order parabolic
equation for one scalar-valued unknown - at least as far as nonlinear equations
are concerned - and on certain particular systems for a vector-valued unknown
describing specific physical situations. The Navier-Stokes equations of hydrody-
namics are among the most eminent representatives of the latter class.
During the last two or three decades, so-called reaction-diffusion equations
have become a much favored object of study by application-oriented analysts. In
contrast to the classical investigations in the theory of partial differential equations,
that concentrate on questions of existence and uniqueness, there has been devel-
oped a more qualitative, dynamical-systems-type approach to reaction-diffusion
equations. The basic idea of this method is to interprete the partial differential
equation as an ordinary differential equation in an infinite-dimensional Banach
space. This assigns a predominant role to the time variable and relegates the spa-
cial dependence to the set-up, that is, to the correct choice of the underlying
function spaces and to the properties of the operators representing the partial
differential equations. Having found the right frame for this description one can
try to mimic the finite-dimensional theory of ordinary differential equations to ob-
tain information on the long-time behavior of solutions, their stability properties,
bifurcation phenomena, etc., questions of paramount interest in applications.
The ordinary-differential-equations-approach to time-dependent partial dif-
ferential equations has proven to be very powerful. It is by no means restricted
to simple semilinear reaction-diffusion equations as they are studied in the litera-
ture most often. In fact, it is one of the main purposes of this treatise to extend
this approach to general quasilinear parabolic systems encompassing a great vari-
ety of concrete equations from science. In addition, by our abstract approach we
1 Nothing impedes the progress of science more than believing to know what one does not
know yet.
xiv Introduction

are rewarded with a general flexible theory that is also applicable to many other
problems not belonging to the class of parabolic systems in the narrow sense.
In the following, we describe our approach, as nontechnically as possible, and
indicate the difficulties and problems that have to be resolved. By this way we are
weaving a silver thread leading the reader through our treatise.

Semilinear Reaction-Diffusion Equations


Let X be a bounded open subset of R n with smooth boundary ax, lying locally
on one side of X. Most naturally, reaction-diffusion equations are derived from
conservation laws of the form

atu + div j = r in X , t > 0, (1)


by specifying the 'flux vector' j by means of phenomenological constitutive rela-
tions like
j:= j(u) := -Dgradu - du. (2)
Here r, the 'reaction rate', is a given smooth function of (x, t) E X X R+ and
the scalar-valued unknown u. The 'diffusion matrix' D: X ---+ R nxn and the 'drift
vector' d: X ---+ R n are also smooth, and D(x) is symmetric and positive definite,
uniformly with respect to x E X.
In concrete situations u may represent a concentration, a density, a temper-
ature, or some other physical or mathematical quantity. Then (1) amounts to a
mathematical formulation of the law of conservation of mass, if u is a concentra-
tion or a density, or of energy, if u is a temperature (and certain simplifications
and constitutive assumptions for the entropy are imposed), etc. Moreover, in the
very special case that D is a positive multiple of the identity matrix and d = 0,
the constitutive relation (2) reduces to Fick's law or Darcy's law (depending on
the model) if u is a concentration or a density, and to Fourier's law if u is the
temperature, etc.
In addition to (1) and (2), the behavior of u on the boundary of X has to be
specified. This can be done by prescribing the value of u on ax. By normalizing
the boundary values, this condition can be formulated as a homogeneous Dirichlet
condition:
u =0 on ax, t> 0 . (3)
Another possibility, being of chief importance in applications, consists of prescrib-
ing the flux through the boundary. Denoting by v the outer unit-normal vector
field on ax, the simplest situation occurs at an insulated boundary modeled by
the 'no-flux' condition
v . j(u) = 0 on ax , t > o. (4)
Of course, there are situations where (3) occurs on a part aox of ax only
and on the remaining part, a1x := aX\aoX, the no-flux condition (4) is effective.
Introduction xv

We always assume that OjX is open and closed in ax for j = 0,1. This situation
can also be described by fixing a continuous map 8: ax --+ {O, I}, a 'boundary
characterization map', and by letting

j = 0, 1.

Note that either one of the boundary parts ooX and OlX may be empty. Then
we can formulate the more general boundary condition, thereby encompassing (3)
and (4), as
-8v·j(u)+(1-8)u=0 onoX, t > O. (5)
Lastly, in order to determine the time-evolution of u from (1), (2), and (5),
that is, the functions 1L(', t) : X --+ IR for t > 0, we have to specify its initial dis-
tribution:
onX. (6)
By substituting (2) in (1) and (5), we can rewrite (1), (2), (5), and (6) as an
initial-boundary value problem:

OtU + Au = r(·, t, u)
Bu=O
in X
on ax } t > 0,
(7)
u(',O) = u o onX.

Here we have put

Au := - div(D grad u + du) (8)

and
Bu := 6(v. D grad u + (v· d)u) + (1 - 8)u . (9)
Of course, the 'boundary operator' B has to be interpreted in the sense of traces.
Note that
v· Dgradu = Dv· gradu = ODvU
is the derivative with respect to the outer co-normal Dv on ax. Thus, in the
very special case tha 0 D is the identity matrix, d = 0, and r is independent of t,
system (7) reduces to an initial-boundary value problem for the autonomous semi-
linear heat equation:

= r(·, u)

}
OtU - t:.u in X
u=O on ooX t > 0,
(lO)
ovu =0 on OlX
u(',O) = uO on X,

a problem having attracted much attention in the literature.


xvi Introduction

The Banach Space Formulation


In order to reformulate (7) as an ordinary differential equation we have to choose
our basic Banach space Eo in which we want to analyze the problem. Of course,
Eo will be a Banach space of distributions on X, that is, Eo is a Banach space
such that 2
Eo '-> D'(X) (11)
Next we define a lincar operator A in Eo by
dom(A) := { v E Eo ; Av E Eo and Bv = O} , Av:= Av. (12)
We also identify
u: X x lR+ --> lR and [t ~ u(·, t)] : lR+ --> lRx
and denote by f the Nemytskii operator induced by r, that is, we put
f(t,u) :=r(-,t,u(·)) , (t,u) E lR+ x lR x .
Then the initial-boundary value problem (7) can formally 3 be rewritten as an
initial value problem for an ordinary differential equation in Eo:
u + Au = f(t, u), t>0 , u(O) = u o . (13)
This follows from the fact that the boundary condition Bu = 0 has been incorpo-
rated in the domain of the linear operator A.
Of course, in order to render this procedure meaningful and to get a treatable
abstract initial value problem (13) we have to impose certain minimal require-
ments. As for the linear operator A, we request that

A is closed and densely defined in Eo,


having a nonempty resolvent set. } (14)

Then, denoting by El the domain of A, endowed with its graph norm, we see that
d
El '-> Eo ,
that is, (Eo, E l ) is a 'densely injected Banach couple'. As for the nonlinearity f,
we require that
f E C(lR+
}
X E l , Eo), and
f (t, .) : El --> Eo is (locally) Lipschitz continuous, (15)
uniformly with respect to t in bounded subintervals of lR+.
2Cf. the sections 'Notations and Conventions' and 'List of Symbols' for the notations and
definitions used without explanation in this introduction.
30bserve that our notation is inconsistent as far as we exhibit the time variable in the non-
linearity. Thus (13) is a formal relation only. In order to give it a precise meaning we have to
define what is meant by a solution. Formal notations of this type are very suggestive and useful
in the theory of differential equations and we use them throughout without fearing confusion.
Introduction xvii

These minimal conditions impose restrictions on the choice of the Banach space Eo.
Observe from (12) that the distributions in dom(A) , that is, in E 1 , have to be
regular enough to admit the traces v f--+ vlooX and v f--+ OAvV on ooX and 01X,
respectively. Hence the Banach space E1 has to consist of sufficiently regular dis-
tributions. Since A is supposed to have a nonempty resolvent set, this requires,
in turn, the Banach space Eo to be not 'too large'. This stipulation is reinforced
by the minimal requirements for f.
Except for the above somewhat implicit restrictions we are free in the choice
of Eo. Of course, we have to keep in mind that, by fixing the space Eo, we may not
rediscover all solutions of problem (7) as solutions of the abstract equation (13).
This can be the case if we choose Eo, and thus E 1 , to be 'too small', that is, if we
require the elements of E1 to be too regular. Of this danger one has to be aware, in
particular, in the case where X is an unbounded domain, say X = ]R.n (a case not
considered in this introduction), since the very definition of the Banach space Eo
often incorporates restrictions on the behavior of its elements 'near infinity'.

Using the relative freedom in the choice of Eo, we opt for simplicity. This
means that we select spaces that are easy to describe and handle. At first sight
the space C := C(X) of continuous functions on X seems to be a good candidate.
However, letting Eo := C, there is no better description of dom(A) than the one
of (12). In other words, although it is true that A is closed and has a nonempty
resolvent set, the space E1 does not coincide with any of the known function spaces.
In particular, E1 does not coincide with

C~ := { v E C 2 (X) ; Bv = o} ,
but is a proper superspace thereof. Moreover, E1 is not dense in Eo if ooX -=J 0.
(The density condition is not indispensable for some parts of the general theory
(cf. [Lun95]), but it is essential for others.) In addition, the domain of A depends on
the diffusion matrix D in the sense that, in general, distinct (even constant) diffu-
sion matrices D1 and D2 give rise to distinct domains dom(At} -=J dom(A 2 ) of the
corresponding operators induced by (8), (9), and (12) (cf. [Sob89]). Lastly, though
the space C is rather simple from the analytical point of view, it is nonreflexive
and thus lacks a very desirable and useful functional-analytical property.

The next class of simple spaces that comes to our mind is the class of Lebesgue
spaces Lp(X), 1:::; p:::; 00. Since the spaces L 1 (X) and Loo(X) show essentially
the same 'deficiencies' as the space of continuous functions (d. [Gui93]), we are
naturally led to put

for some p E (1,00) . (16)


In this case it turns out that the minimal requirement (14) is satisfied. Moreover,
the space E1 can be described explicitly by
(17)
XVIII Introduction

Note that El is independent of A if 8 0 X = 8X. But it does depend on A if 8 1 X i- 0


(through the condition 8 Dv u + (v· d)u = 0 on 81 X).
As for the minimal requirement (15), we first recall Sobolev's embedding
theorem:

if 11p:::- 11q :::- lip - sin,


(18)
if 0::; p ::; s - nip and s > nip,

where s, q E lR+ and p < s - nip if s - nip E N. Second, given a continuous func-
tion g : X x lR --> lR, it is known that the Nemytskii operator induced by g maps Lq
into Lp iff it satisfies an estimate of the form

(x,O E X x lR . (19)

Moreover, to guarantee the Lipschitz continuity of this superposition operator


one needs a polynomial growth restriction for 82 g in addition. Thus we see from
El <-t Wp2 and (18), (19) that, given t E lR+, the minimal requirement (15) implies
a polynomial bound for the map ~ f-t r(·, t, ~), if P ::; n12. On the other hand, super-
position operators have good properties in C. More precisely, if g E C k (X x lR, lR)
for some kEN,
[u f-t g(-'u(·))] E Ck(C(X),C(X)) . (20)

Thus, if we impose the condition p > n12, it follows from El <-t C <-t Eo that the
minimal requirement (15) is met.

Semilinear Parabolic Evolution Equations

The proof of the validity of (14) can be arranged to give more. Namely, it can
be shown that the resolvent set of - A contains a half-plane [Re A :::- AO 1 for some
AO E lR and that there exists a constant K such that

IAIliulio ::; K II(A + A)ull o , u E E1 , Re A :::- Ao , (21)

where 11·110 is the norm in Eo. This, together with (14), is equivalent to the asser-
tion that -A generates a strongly continuous analytic semigroup {e- tA ; t:::- O}
on Eo. Thus
(22)

that is, A E C(El' Eo) and -A, considered as an unbounded linear operator in Eo
with domain E 1 , is the infinitesimal generator of a strongly continuous analytic
semigroup on Eo.
Introduction xix

Now suppose that J is a perfect compact subinterval of jR+ containing 0 and


that u is a solution of (13) on J. This means that, letting j := J\ {O},

(23)

and u satisfies (13) point-wise, that is, u(t) + Au(t) = f(t, u(t)) for t E j and
u(O) = u o. Then ~ similarly as in the theory of ordinary differential equations ~
the variation-of-constants formula implies that

t E J . (24)

Note, however, that this similarity is a formal one only, since formula (24) does
not make sense, given our assumptions. Indeed, it follows from (23) and (15) that

By combining this with the strong continuity of the semigroup {e~tA ; t ~ 0 }


on Eo, we see that

(25)

But this does not guarantee the existence of the integral in (24).
Clearly, there is an easy remedy. Namely, it suffices to strengthen the concept
of a solution by replacing (23) by

u E C(J, Ed n C1(J, Eo) . (26)

This requires, of course, that we choose the initial value u O in E 1 • From (26) we
infer that (25) holds if (0, t] is replaced by [0, t]. Then the right-hand side of (24)
is well-defined and equation (24) holds in Eo.

Observe that (24) is a fixed point equation for u. More precisely, define

by
<T?(u)(t) := e~tAuO + 10 t e~(t~r)Af(T,U(T)) dT, t E J .

Then, since El '-+ Eo implies C(J,E 1 ) '-+ C(J,Eo), we see from (24) that

u=<T?(u) (27)
in C(J, Eo), provided u satisfies (26) and is a solution of (13) on J. Now, simi-
larly as in the theory of ordinary differential equations, we would like to use this
fixed point equation as a basis for an existence theory for the initial value prob-
lem (13). The advantage of such an approach is obvious: whereas (13) involves the
xx Introduction

unbounded linear operator A and the derivative u of u, in (27) there occur bounded
linear operators only and no derivative at all. As a result, the nonlinear map <I>
is rather easy to deal with. In fact, it is a simple consequence of (15) and (22)
that <I> is a Lipschitz continuous map from C (J, E I ) into C (J, Eo). Furthermore,
in a neighborhood of the constant map (t 1-+ UO) E C(J, E I ), where <I> is uniformly
Lipschitz continuous, the Lipschitz constant can be controlled and made smaller
than 1 by making the interval J sufficiently small. Thus it is tempting to prove
the existence of a fixed point of <I> by means of the method of successive itera-
tions. However, this approach is bound to fail already at the first step since the
image, UI, of a function Uo E C(J, E I ) under <I> is only known to lie in C(J, Eo), so
that UI is not necessarily in the domain of <I> , and <I> ( UI) is not defined, in general.

At this stage we recall that the semigroup generated by -A is analytic,


a property that has not been used so far. Analyticity implies that the image of e- tA
is contained in the domain of A whenever t > 0. More precisely, the map

(28)

is well-defined and analytic. This is an important smoothing property of analytic


semigroups. It is the abstract counterpart to the well-known fact that parabolic
differential equations regularize the initial data. This effect is seen most clearly in
the case of the heat equation on JRn;

BtU - b.u = 0 in JR n , t > °, U("O)=Uo.

Here the solution, being represented by the heat semigroup as u(·, t) = etnuo for
uO belongs to Lp(JR n ) only.
t ::::: 0, is analytic in JRn x (0,00), even if
Of course, the map (28) cannot be continuous on JR+. But it is known that
there exist constants c and W E JR such that

II e -tAil C(Eo,Ed <


_ C t- e wt
I
, t > 0, (29)

and this estimate is sharp, in general. At first sight this is not of much help;
although it follows from (15) and (28) that

(30)

for U E C(J, E I ), thanks to (29) we cannot guarantee that (30) is integrable on J.


Thus our minimal assumptions do not guarantee that <I> maps C(J, E I ) into itself.
Suppose, however, there exist 0: E (0,1) and a Banach space EO! satisfying

(31)

such that the estimate

II e -tAil C(Eo,EaJ <_ C t-a e wt , t > 0, (32)


Introduction xxi

is valid and the semigroup {e- tA ; t 2: O} on Eo restricts to a strongly contin-


uous semigroup on Eo:. Also suppose that we strengthen the minimal require-
ment (15) - then calling it (15)a - by replacing E1 by Ea. Then it is easy to
see that <P is a well-defined Lipschitz continuous self-map of C(J, Eo:) and that the
fixed point approach indicated above ean be carried through without difficulties.
This amounts to a completely satisfactory theory of abstract semilinear parabolic
evolution equations of the form (13). In particular, it can be shown that (13) gen-
erates a (local) semiflow (t, uo) f-+ u(t, uo) on the phase space E a , where u(., uo)
is the unique maximal solution of (13).
It is well-known that Eo: := D((w + A)O:) , where (w + A)O: is the a th 'frac-
tional power' of w + A and w is a sufficiently large real number, satisfies the above
requirements. In general, the domains of (w + A)O: are not explicitly known. How-
ever, it can be shown that

D((w + A)") '-> W; , 0::; s < 2a . (33)

Thus, choosingp > n/2 and a E (n/(2p), 1), we infer from (18) that Ea '-> C '-> Eo.
Hence (20) implies the validity of the modified minimal requirement (15)0: in
the case of problem (7) without any growth restriction for r. Furhermore, if we
fix p > n, we deduce from (18) and (33) that Ea '-> C 1 '-> C'-> Eo, provided
a E (nip, 1). In this case the reaction rate r can also be allowed to depend smoothly
upon grad u, that is,
r = r(x, t, u, grad u) . (34)
Indeed, again denoting by f the corresponding ~emytskii operator, that is,

f(t,u):= r(-,t,u(·),gradu(·)) , (t,u) E jR+ x C 1 (X) , (35)

we infer from (20) that f satisfies the modified minimal requirement (15)" in this
case as well.

Quasilinear Reaction-Diffusion Systems


The above approach to semilinear parabolic initial-boundary value problems is
well-known and well-documented in the literature, most notably in the rather
influential monograph of Henry [Hen81]. It has been the basis for numerous inves-
tigations of reaction-diffusion equations, in particular of semi linear heat equations
of the form (10), where u is sometimes admitted to be N-vector-valued.
However, most reaction-diffusion systems of interest in science are consider-
ably more intricate than the simple laws (1), (2), and (5):
• First, as a rule, realistic models involve several species, say N, each of whom
satisfies a conservation law (1). Thus we have to consider a system of equations:

(36)
xxii Introduction

• Second, the constitutive relations for the flux vectors ji depend on all
components of the vector
U:= (11,1, ... ,UN) .
Under rather general circumstances (e.g., [deGM84]) this dependence is of the form

N
ji(U) := - 2)Dik grad Uk + dikuk) , l::;i::;N, (37)
k=1

where Dik and dik are (n x n)-matrices and n-vectors, respectively.


• Third, for each species i there is a separate boundary characterization
map 8i . For simplicity, the boundary value of Ui on the 'Dirichlet boundary, 8;1 (0),
of species i' is again normalized to zero. On the remaining part of ax, that is,
on 8;1(1), the no-flux condition is replaced by the more general 'prescribed flux
condition'
-v· ji(U) = Si
with a 'surface reaction rate' Si. Thus there are the N boundary conditions:

(38)

• Fourth, the matrices Dik, the vectors dik , and the functions ri depend
upon (x,t) E X x lR+, the functions Si on (y,t) E ax x lR+, and all of them on
the solution vector u. Moreover, all these dependences are smooth, as a rule.

Of course, the initial condition (6) has to be replaced by N relations of


that type:
Ui(-,O) = u? ,
l::;i::;N. (39)

To rewrite system (36)-(39) as an initial-boundary value problem of a form


similar to (7), we introduce the (N x N)-matrices

k ] l<ik<N'
aa{3(t,u):= [ D~{3(.,t,u) aa(t,u):= [dik
a (.,t,u) ] l<ik<N

for 1 ::; (x, {3 ::; n, where D~~ and d~k are the entries and components of the (n x n)-
matrices Dik and n-vectors dik , respectively, and the N-vectors 4

!(t,u):= (rl(·,t,U), ... ,rN(·,t,U))

and

4Unless absolutely necessary, we do not distinguish between row and column vectors. Also we
often use the same notation for a fuuction and the Nemytskii operator induced by it, provided
it is clear from the context which interpretation has to be chosen.
Introduction xxiii

We also employ the summation convention for a, (3 E {I, ... , n} and put

b(t,u) := vQ;aa(t,u) ,

where v = (v 1 , ... , vn). Then, using euclidean coordinates, we define differential


operators A(t, u) in X and boundary operators 8(t, u) on 8X by

(40)

and
8(t, u)v := 6(v a aa(3(t, 71)8(3v + b(t, u)u) + (1 - 6)u , (41)
respectively, where fj is now the diagonal matrix with entries fjl, ... ,6N. Then the
reaction-diffusion system (36)~(39) is rewritten as a quasilinear initial-boundary
value problem:

8t u + A(t, u)u = f(t, u)


8(t,u)u = g(t,u)
in X
on8X } t > 0,
(42)
u(.,O) = U o onX,

where UO := (u~, ... ,u~).


Note that (42) is a strongly coupled system in the sense that the 'diffusion co-
efficients' aa(3 are fully occupied (N x N)-matrices, in general. Furthermore, these
matrices depend nonlinearly on the solution vector u, that is, the system (42) is
quasilinear, and it involves nonlinear boundary conditions. Clearly, the quasilinear
initial-boundary value problem (42) for a vector-valued unknown is considerably
more intricate than the semilinear problem (7) for a scalar unknown. In particu-
lar, (42) does not satisfy a maximum principle, in general, which means that we
are lacking a tool that has proven to be very powerful in the study of a single
parabolic differential equation. On the other hand, it is important to observe that
(42) is of divergence form, a fact that turns out to be crucial for handling nonlinear
boundary conditions.

Homogeneous Boundary Conditions


First we consider the case g = 0 and proceed as in the semilinear case. Thus
we begin by interpreting (42) as an ordinary differential equation in the Banach
space Eo. Of course, here and below it is always understood that all fUIlction
spaces consist of N-vector-valued functions, e.g., Lp = Lp(X, ]Rn) etc., if we refer
to systems, unless explicitly stated otherwise.
For the time being, by a solution of (42) on J we mean a fUIlction
xxiv Introduction

satisfying (42) in the obvious 'point-wise' sense. Given such a solution u, we put
A(t,u(t)) := A(t,u(t))I~~B(t,u(t)) ,
thereby using the notation of (17). Then u is, at least formally, a solution on J of
the quasilinear Cauchy problem
u+ A(t,u)u = f(t,u) , t E j , u(O) = Uo . (43)
Letting
Au(t):= A(t,u(t)) , tEJ ,
u is also a solution of the nonautonomous semilinear Cauchy problem
V+Au(t)v=f(t,v) , tEj, v(O) = uo . (44)
Hence it can be expressed by the variation-of-constants formula

u(t) = Uu(t,O)uO + It Uu(t,T)f(T,U(T)) dT, t E J, (45)

where now the semigroup in (24) is replaced by the evolution operator Uu of the
family {Au(t) ; t E J}. Consequently, denoting by <I> the right-hand side of (45),
we are again led to the fixed point equation (27) that we want to make the basis
for our study of the quasilinear problem (43).
Now the situation is much more complex than in the semilinear case. First
of all, to give a meaning to these formal considerations we have to guarantee
the existence of the evolution operator Uu , given a function u possessing suitable
regularity properties. Clearly, the existence and the properties of an evolution
operator are intimately related to the solvability of the linear Cauchy problem
v + A(t)v = f(t) , tEj , v(O) = UO , (46)
that is obtained from (44) by ignoring the u- and v-dependence of A and f, re-
spectively, for the moment. This Cauchy problem corresponds to a Banach space
formulation of the nonautonomous linear initial-boundary value problem
8tv + A(t)v = f(t) in X } t > 0,
B(t)v = 0 on8X (47)
v(·,O) = uO onX,
where now the coefficients in (40) and (41) are independent of u.
In the simplest case of a scalar unknown, that is, for N = 1, we know already
from (17) and (22) that
A(t) E 1t(El(t), Eo) , (48)
where
El (t) ~ ~~B(t) . (49)
Note that (48) means that each A(t) is the negative generator of an analytic
semigroup on Eo, whose domain is t-dependent, in general.
Introduction xxv

In the case of systems, that is, for N> 1, as well as in the general abstract
situation, condition (48) is a hypothesis. If A(t) is induced by the linear initial-
boundary value problem (47), it turns out that a necessary condition for (48) to
be satisfied is that

a+ A(t) is Petrowskii parabolic for t E J . (50)

This means that, given t E J and ~ = (e, ... , ~n) E lR n \ {O},

u(aa(3(t)~a~(3) C [Rez > 0] ,


or, in other words, that each eigenvalue of the (N x N)-matrices aa(3(x, t)~a~(3 has
a positive real part, given any x E X and ~ E lR n \ {O}.
In general, condition (50) is not sufficient, but there are additional require-
ments for the boundary symbols v""aa(3(t)e to be satisfied, that are too com-
plicated to be given here. However, there are important special cases, that are
of particular importance for applications, in which condition (50) already implies
(48) and (49). For cxample, this is the case if B(t) equals the 'Dirichlet boundary
operator', that is, B(t)u = ulax, for t E J. To give a further example, suppose
that, given t E J,

(51)

and

aa(3(t) = a(t)8a(3, 1:S: a,/3:S: n , (52)

and that
(53)
Note that this implies the existence of a boundary decomposition ax = aox u a1x
such that

B(t)u = { ~(t)a,/U + b(t)u (54)

Then (48) and (49) are satisfied iff

u(a(t)) C [Rez > 0] , (55)

that is, iff a + A(t) is Petrowskii parabolic for t E J. Observe that (55) does not
imply that the symmetric part of the matrix a(t) is positive definite. Hence a
Petrowskii parabolic system is not strongly parabolic, in general, that is, not
coercive.

In the second volume of this treatise we discuss in detail the question un-
der what conditions systems of partial differential operators (of arbitrary order),
together with appropriate sy~tems of boundary operators, generate analytic semi-
groups on Lebesgue and other spaces.
xxvi Introduction

Constant Domains
Now let conditions (48) and (49) be satisfied. Then, to guarantee the existence of
an evolution operator U for the linear Cauchy problem (46), one has to impose
regularity assumptions on the function t f--+ A(t). It turns out that the required
amount of regularity is proportional to the 'strength of variability' of the domains
of A(t), that is, the spaces El (t), if t varies over J.
The simplest case occurs if the domains of A(t) are independent of t E J,
that is, if El(t) ~ El(O) =: El for t E J. Note that this is the case if B(t) equals
the Dirichlet boundary operator for t E J. It is also true if conditions (51)-(55)
are satisfied and b = 0, since in this case the homogeneous boundary condition
B(t)u = 0 is equivalent to the constant boundary condition 80v u + (1 - 8)u = 0
on aX.
If the operators A(t) have a constant domain E l , it suffices to require that

(56)

for some p E (0,1). Then there exists a unique evolution operator U for the family
{A(t) ; t E J} such that each solution of (46) can be represented by the variation-
of-constants formula

u(t) = U(t,O)uO + lot U(t,T)f(T)dT, t E J .

Furthermore, the evolution operator U possesses properties that are similar to the
ones of the semigroup {e- tA ; t 2:: O} of the autonomous problem. In particular,
the following analogue of (29) is valid:

IIU(t, 8)llc(Eo,EIl ::; c(t - 8)-1 , t,8 E J, 8 <t .

Recall that our interest belongs to the quasilinear equation (43), and thus to
the fixed point equation (45). Since the dependence of A on u in (43) implies a
dependence of U on u in (45), it is necessary to study carefully the dependence
of U on A. The construction of the evolution operator, as well as the investigation
of its dependence on A, is the content of the second chapter of this volume. There
the general setting of an arbitrary densely injected Banach couple is considered,
which greatly enhances the applicability of our results.
Of course, treating the quasilinear problem (43) by reducing it to the fixed
point equation (45), we encounter the same difficulties we found when we had to
give a meaning to the fixed point map <I> in the semilinear case (13). Thus we
have to consider again an 'intermediate' space Ea satisfying (31) such that each
one of the semigroups {e-tA(s) ; t 2:: O}, 8 E J, restricts to a strongly continuous
semigroup on Ea and (32) is replaced by

IIU(t, 8)llc(Eo,Eu) ::; c(t - s)-a , t,8 E J, 0::; 8 <t .


Introduction XXVll

As in the semilinear case, D(AO!(O)) seems to be a natural candidate for Ea.


However, this choice is not the best one, mainly since D(AO!(t)) is not known ex-
plicitly for a E (0,1). Thus, for example, it is not elear if D(AO!(t)) is independent
of t E J, in general.
Although it is possible to develop a general theory of abstract quasilinear
Cauchy problems by basing it on the theory of fractional powers - this has been
done by P.E. Sobolevskii in [Sob66] (also cf. [Fri69] for an exposition of parts of that
work) - , this approach does not give optimal results. Indeed, by Sobolevskii's
approach it is not possible to prove that the solutions of (43) generate a semiflow,
that is, a nonlinear local semigroup, since one has to require more smoothness for
the initial datum U O than should be expected. In other words, the initial value U O
has to belong to a slightly smaller space than the natural phase space of (43).
The semiflow property of the solutions of the abstract quasilinear Cauchy
problem is both natural and fundamental for a qualitative study of the latter.
Thus any good general theory has to produce this property. Therefore we are
forced to look for other candidates for EO!. Fortunately, there are plenty of them:
most of the interpolation spaces of exponent a between Eo and El turn out to
lead to the desired results. Thus at this stage interpolation theory comes into play
naturally.
Together with the theory of analytic semigroups, interpolation theory plays a
predominant role in the abstract theory of linear and quasilinear Cauchy problems
of 'parabolic type'. For this reason we collect in Chapter I the basic facts from
these two fields, that are being used throughout our treatise. Furthermore, concrete
characterizations of interpolation spaces are studied in detail in the second volume.
In particular, the following is shown: let

{ v E WP" ; B(t)v = O} , 1 + lip < 8 < 00 ,


{ v E WP" ; (1 - 8)v18X = O} , lip < 8 < 1 + lip, (57)
Wp", 0::;: 8 < lip,

and put
W;'S:= W;'S(t) , S E [0,1 + IIp)\{llp}·
Then, given a E (0,1) with 2a ~ {lip, 1 + lip}' there exists an interpolation func-
tor (-, .)" of exponent a such that

(58)

where Eo := Lp and El(t) := ~~S(t) = dom(A(t)). Note that E" := E,At) is in-
dependent of t E J if 0 < 2a < 1 + lip and 2a i= lip.
The assumption of constant domains for A (t, u( t)), which entails the inde-
pendence of the interpolation spaces Eo(t) of t E J, simplifies the approach to
quasilinear parabolic problems considerably and results in a rather flexible theory.
XXVlll Introduction

This is essentially due to the fact that the existence of an evolution operator can
be guaranteed by presupposing (56) with an arbitrarily small Holder exponent p.
As a rule, A(t,u(t)) is well-defined only if 11,(t) belongs to E(3 for (3 sufficiently
large. Indeed, to guarantee (49) we have to assume that
a,(3E{l, ... ,n}, tEJ.

Thus we certainly have to require that u E C 1 . Consequently, we infer from (56)


that we have to assume that
(59)
where E(3 ~ C 1 . By (18), (57), and (58), for this embedding to be valid we
have to impose the condition (1 + nip) < 2(3 < 2, which entails p> n. Now we
fix a E ((3,1) and choose Ea as a phase space, that is, we study the fixed point
equation (45) in C(J, Ea). Then <I> maps C(J, Ea) into C(J, EaJ n CP(J, E(3), pro-
vided p < a - (3. Hence U1 := <I>(11,) satisfies (59) whenever 11, E C(J, Ea). So the
iteration process of the method of successive approximations can be carried out in-
definitely, and it can indeed be shown that <I> has a unique fixed point in C(J, Ea),
provided J is sufficiently short. Note that this imposes the restriction 2p < 1 - nip.
It should also be remarked that, up to this point, the divergence structure of (40),
(41) is not needed.

Extrapolation Theory
The approach outlined above uses in an essential way the fact that the domains
of A(t) are constant. More precisely, it is based on the fact that the interpolation
spaces E(3 and Eo: are independent of t for 1 + nip < 2,6 < 2a < 2. By (57) and
(58), this is not the case if dom(A(t)) = ~~S(t) depends on t E J.
On the other hand, thanks to its divergence structure, problem (42) has a
natural weak formulation. For this put

(v, wI := Lv, W dx , (v,W) E Lp' x Lp .

Furthermore, given t E J and 'U E C, let

Then a(t,11,) is a continuous bilinear form on W~ p,


s x Wp1s,
,
the 'Dirichlet form'
belonging to (A(t, 11,), B(t, 11,)). Thus, letting

-2 + lip < s < 0, s I- -1 + lip, (60)


with respect to the duality pairing naturally induced by the Lp-duality pair-
ing (', 'I, it follows that a(t,11,) induces
A(t, 11,) E .c(Wp~s, WP~J) (61)
Introduction xxix

via
(v,A(t,u)w):= a(t,u)(v,w) , (v, w) E Wd',B x ~:,B . (62)
We abo define a nonlinear map F by

(v,F(t,u)):= (v,f(t,u))+ r v·g(t,u)dIJ,


Jax v E Y~~,B . (63)

Then, thanks to the divergence structure of (40) and (41), the natural weak for-
mulation of the initial-boundary value problem (42) takes the form:

find u E C(J, Wp~B) satisfying

1{ -(v, u) + a(t, u)( v, u)} dt = 1 (v, F(t, u)) dt + (v(O), uo) (64)

for each v E cl(J, ~~,B) vanishing near the right endpoint of J.

It is easily seen that every solution of the initial-boundary value problem (42) is a
weak solution in the sense of (64). Furthermore, (64) is equivalent to the abstract
quasilinear Cauchy problem in p, W-J:
u+A(t,u)u=F(t,u) , tE j, u(O) = Uo . (65)

Observe that, thanks to (61), the domains of the operators A(t, u) are all equal
to Wp~B' Thus (65) is amenable by the techniques described above for the constant
domain problem (43), provided

(66)

with suitably regular dependence on (t, u).

It is a consequence of the divergence theorem that A(t, u) is an extension of

A(t, u) := A(t, U)IW;'B(t,u) .

This fact can be used to prove (66). To outline this method, we first fix a negative
generator B of a strongly continuous analytic semigroup on an arbitrary Banach
space Fo := (Fo, 11·110) and assume, for simplicity, that B has a bounded inverse
defined on Fa. Then

defines a norm on dom(B) that is equivalent to the graph norm. This implies that
FI := (dom(B), 11'111) is a Banach space such that (Fo,F1 ) is a densely injected
Banach couple, and B E H(Fl' F o).
The Banach space Fo can be recovered from Fl by observing that Fo is a
completion of Fl in the norm u f--> IIB11'ulil = Ilullo, where Bl is the Fl-realization
xxx Introduction

of B. This suggests to introduce a superspace F-l := (F-l' 11·11-1) of Fo by choos-


ing for
F-l a completion of Fo in the norm u t-+ IIB- 1 ullo =: Ilull-I.
Then (F-1, Fo) is a densely injected Banach couple as well, and it is not difficult
to show that Bo := B extends continuously to an operator B-1 E H(Fo, F-1). If
B is not invertible, we simply replace B by w + B for a sufficiently large w E lR+.
Now, given () E (0,1), we choose an interpolation functor (., ·)0 of expo-
nent () with the property that G l is dense in (Go, Gt}o whenever (Go, G 1 ) is a
densely injected Banach couple, an 'admissible interpolation functor'. Then we put
Fe := (Fo, Ft)o and FO-l := (F-l' Fo)o. This defines a 'scale of Banach spaces'

-1<,6<0<1.
Furthermore, denoting the Fa_I-realization of B-1 by B a - 1 for 0 E [0,1], it follows
that
B a - 1 E H(Fa , Fa-t) ,
These extensions are 'natural' in the sense that e-tB"'-l is the restriction to F a - 1
of e- tBj3 - 1 for 0 :S f3 < 0 :S l.
This 'interpolation-extrapolation' technique is rather flexible, has many ap-
plications, and is crucial for our approach to quasilinear parabolic evolution equa-
tions. It allows to measure very precisely regularity properties of operators and
solutions. Scales of Banach spaces, interpolation-extrapolation techniques, and
evolution equations in these scales are studied in detail in Chapter V.

Now we return to our quasilinear parabolic problem. In this case, given (t, u),
we can construct an interpolation-extrapolation scale by starting with A(t, u) on
Eo = Lp- Then, using suitable admissible interpolation functors, this scale, denoted
by Ea(t, u) for 0 E [-2,1]' can be explicitly represented as
Ea(t, u) ~ ~~B(t,u) , 20 E (-2 + lip, 2]\(Z + lip) ,
in the range given. Thus we infer from (57) and (60) that
20 E (-2 + lip, 1 + 1/p)\(Z + lip) . (67)
Therefore the interpolation-extrapolation spaces are independent of (t, u), that is,
they are constant, if 0 belongs to the range given in (67). Moreover, denoting
by A a - 1 (t,u) the 'extrapolated operators', the general theory of Banach scales
tells us that
Aa-l (t, u) E H(E", , E a-l) , lip < 20 < 1 + lip· (68)
Lastly, it can be shown that
A(t, u) = A- 1 / 2 (t, u) , (69)
which, thanks to (67) and (68), proves (66).
Introduction xxxi

Regularity
The interpolation-extrapolation technique outlined above enables us to solve the
initial-boundary value problem (42) in its weak form (64), provided a and Fare
suitably regular. In fact, it suggests to replace the quasilinear evolution equa-
tion (65) by the family

u+Aa-l(t,u)u=F(t,u) , tEj, u(O) = uo (70)

for l!p < 2a < 1 + l!p. In particular, we can fix 2a E (1,1 + l!p) and can choose
E l / 2 = Wp~B as a phase space for the constant domain problem (70). Then, given
uo E El/2' problem (70) has a unique maximal solution

'Un :=
° .
u(·, u ) E G(Jo, E l / 2) n G(Jo, En) n G 1 (J.o, E a- l ) . (71)
It follows from (69) and A-l/2 ~ A a - l that Uo is a weak solution of (42) in the
sense of (64). But (71) shows that Uo possesses better regularity properties.
This observation suggests a bootstrapping procedure. For this we observe
that Uo is a solution of the linear Cauchy problem

v + A, -1 ( t, Uo (t )) = F (t, Uo (t )), t E j °, v(O) = uo

in En-I. Recall that A,-l(t,uo(t)) ~ A,6-l(t,U(t)) if {3 > a. Thus, if the linear


Cauchy problem

v + A,6-l (t, uo(t))v = F(t, uo(t)) , t E jo , v(O) = uo (72)

is well-posed in E,6-l for some {3 E (a,l], its solution coincides with un. Note
that a solution u of (72) satisfies u(t) E dom(AB-l(t,Uo(t))) for t E jo. Hence it
follows that
uo(t) E dom(A,6-l(t,UO(t))) C Wp2,6-2 (73)
for t E jo, provided (72) is well-posed.

By these considerations we are led to study the linear Cauchy problem

iJ + A,6-l(t)V = F(t) , t E j ,

in E,6-l, where we are particularly interested in the case 2{3 > 1 + l!p, so that
dom(A,6-l(t)) is no longer constant. Since Aa-l(t) :J A,6-l(t), it is to be expected
that the restriction of the evolution operator Un - l of the family { A a - l (t) ; t E J}
to the space E,6-l '--+ E a - 1 is an evolution operator for { A iJ - l (t) ; t E J} on E,6-l.
This is true in fact, provided

p>{3-a, (74)

as follows from the general results derived in Chapter IV.


xxxii Introduction

It turns out that Aa-1(t,uo(t)) can be characterized, similarly as A- 1/ 2(t)


in (69), through the bilinear form a(t,uo(t)), the latter now being continuout; on
Wp~~B2a X ~~B' A careful study of point-wise multipliers on generalized Sobolev
spaces, that is carried out in the second volume, reveals that

(75)

if (J > 20: - 1. From (71) we infer, using suitable embedding theorems, that

Uo E CP (Jo , C(1) , (76)

provided 2p < 20: - (J - nip. By combining (75) and (76) we see that

(77)

for 2p < 1 - lip. Hence A a - 1 (" uo(-)) satisfies (74) for

2(3 < 20: +1- nip < 2 - (n - l)lp .

This way, and using the theory developed in Chapter IV, it follows that
.
Wp2/3 ) n C l(J,
' 2/3 2
Uo E C(Jo, Wp,B - ) '--+ CP(J,. C(1) ,

provided 2(3 < 2 - (n - l)/p and (J + 2p < 2(3 - nip < 2 - (2n - l)/p. Thus, since
we can choose an arbitrarily large value for p, we obtain

(J + 2p < 2. (78)

At this stage we can invoke the classical Holder theory for general parabolic sys-
tems, for example, to deduce that Uo is in fact a classical solution, even a Coo_
solution, of the initial-boundary value problem (42).

Unfortunately, the situation is more complicated than just indicated. First,


in the preceding outline we neglected the difficulties coming form F(t,uo(t)). The
above procedure requires t f--+ F( t, uo(t)) to belong to C(1 (J, E/3-1) for (3 arbitrarily
close to 1. This does not hold, in general, unless the boundary nonlinearity 9 satis-
fies appropriate compatibility conditions and the problem is suitably modified. For
the latter modification one has to have a good theory of nonhomogeneous linear
parabolic boundary value problems in W; for 1 + lip < s ::; 2 under mild regular-
ity conditions on the coefficients. These questions are investigated in Volume Two.
Second, it is by no means clear that Aa_1(t,uo(t)) E 1-i(Ea, E a - 1) ifuo sat-
isfies (76). Indeed, to derive (68) we started with A(t, u) E 1-i (E1 (t, u), Eo). In our
concrete situation this can only be established if the coefficients of .A and B are
of class C 1 . To guarantee this we would already have to know that Uo (t) E C 1 .
Thus we have to prove directly that -A",_1(t,U) generates a strongly continuous
analytic semigroup on E a - 1 if .A and B have Holder continuous coefficients only.
Introduction xxxiii

This problem is also studied in detail in the second volume by making use of the
interpolation-extrapolation technique for smooth operators.
Having established the necessary ingredients, the above bootstrapping tech-
nique can be carried through, in principle. However, the technical details are some-
what delicate, and it is not possible, in general, to derive (78) in one step. Instead,
a chain of bootstrapping arguments is needed. These arguments are given in detail
in the third volume on the basis of the results of Volume Two.
Besides of existence and regularity theorems, in the last volume a qualitative
theory of quasilinear parabolic evolution equations is developed as well. It is shown
that they generate smooth local semiflows on appropriate natural phase spaces.
Furthermore, questions of global existence, stability of critical points and periodic
orbits, bifurcation phenomena, etc., are studied in the general abstract setting. In
addition, those results are applied to a variety of concrete systems of partial dif-
ferential equations stemming from diverse applications in physics, chemistry, etc.

Maximal Regularity
The theory of linear and quasilinear parabolic evolution equations, outlined so far,
uses crucially the smoothing property of parabolic equations. This regularizing
effect is the ultimate reason for the validity of global existence theorems based on
a priori estimates in 'weak' norms, that is, norms involving little regularity of the
solutions only.
On the other hand, it restricts the applicability of the general abstract theory
to quasilinear reaction-diffusion systems whose 'diffusion matrices' depend on u
only, at least in the divergence form case. If grad u also occurs in the diffusion
matrices, this theory is not applicable since it uses the fact that the dependence
of Ac,,-l (t, u) E £( E et , E a - l ) on (t, u) is smooth with respect to the topology of an
interpolation space (E a - l , Ea)o that is strictly weaker than the topology of Ea.
If the 'intermediate' smoothness requirement is not satisfied, e.g., if the diffu-
sion matrices depend on grad u, we have to invoke maximal regularity results. The
maximal regularity approach is different in spirit from the evolution operator ap-
proach discussed so far. Whereas the latter is essentially a dynamical method,
that is the infinite-dimensional analogue to the theory of ordinary differential
equations, the former is basically a stationary technique. This means that the
linear Cauchy problem (46) is interpreted as an 'algebraic' equation in a suit-
able Banach space lEo such that lEo '--+ Ll (J, Eo). To be more precise, we consider
the case of constant domains and suppose that lEl is a Banach space such that
lEI '--+ Ll (J, E l ) n wI (J, Eo) and 8 + A E £(lEl' lEo). Of course,

(8 + A)u(t) := u(t) + A(t)u(t) , u E lEo, a.a. t E J .

Moreover, putting rU := u( 0) for u E lEI, let r lEI be the 'trace space' of r such
that El '--+ rlEl '--+ Eo and r is a continuous surjection from lEI onto rlEl. Then,
xxxiv Introduction

given (j, u O) E lEo X /,lEI, we rewrite the linear Cauchy problem (46) in the form:

find u E lEI such that (8 + A, /,)u = (j, U O).

Finally, the pair of Banach spaces (lEo, lEI) is said to be a pair of maximal regularity
for {A(t) ; t E J}, provided (8 + A, 'I) is a topological isomorphism from lEI onto
lEo x /,lE l . If this is the case, it is to be expected that we can apply linearization
techniques based on the implicit function theorem to solve the quasilinear Cauchy
problem (43), that now takes the form of the nonlinear operator equation

In fact, this way 'fully nonlinear problems' can be handled. However, in the
quasilinear situation we can combine the maximal regularity approach with the
interpolation-extrapolation method to deal with the case of nonconstant domains
as well. Indeed, in the qualitative study of quasilinear evolution equations a com-
bination of the dynamical evolution operator approach and the maximal regularity
technique, used in the third volume, gives the best results.
It turns out that the requirement that (lEo, lEI) be a pair of maximal regu-
larity for {A(t) ; t E J} imposes restrictions on the spaces and on the operators,
in general. These questions are carefully discussed in Chapter III, where the most
important and useful abstract maximal regularity results are derived. Concrete re-
alizations necessitate rather deep information on various function spaces, Nikols'kii
spaces, for example, and on linear elliptic differential operators. This is provided
in the second volume.

The Scope of the Theory

By now the reader surely feels that our approach to the quasi linear initial-boundary
value problem (42) is rather lenghty, highly complicated, and requires a lot of tech-
nical details. This being certainly true, one might wonder if it is worthwhile at all.
Of course, there are simpler methods for proving the existence of weak so-
lution to problem (42). First, the theory of monotone operators can sometimes
be applied. However, it requires rather stringent conditions for the operators and
nonlinearties and applies to a small class of problems only, leaving aside most
systems of interest.
Second, there is the Galerkin approximation approach, that is perhaps most
often applied to produce weak solutions of problem (42). That method is basically
a Hilbert space technique, being based on coercivity estimates. However, coercivity
estimates do not hold for general Petrowskii parabolic systems. To guarantee coer-
civity estimates one has to restrict the consideration to certain 'strongly parabolic'
systems, a class that is too narrow to cover many interesting problems occurring in
science. Similar statements hold for other (space or time) discretization methods.
Introduction xxxv

Even if the system is strongly parabolic, so that the Galerkin method is


available, that approach gives weak Wi-solutions only, that is, solutions in the
sense of (64) with p = 2. However, it is well-known (e.g., [Gia83]) that weak wi-
solutions of systems are not regular, in general. In fact, it is not even possible, as
a rule, to prove that they are HOlder continuous. This lack of regularity renders it
rather difficult - if not impossible - to develop a dynamical systems theory and
to use techniques from nonlinear analysis to establish qualitative properties of the
solution set.
In contrast, our approach applies to noncoercive problems as well and gives
weak Wp1-solutions, where p > n. Thus, thanks to (18), the solutions we find are
already known to be Holder continuous from the very beginning. As indicated
above, this implies that all our solutions are classical if all data are sufficiently
smooth. Of course, there may be singular solutions - in particular, if the coeffi-
cients of A and B lack regularity - that we 'do not see' in our approach. However,
we obtain a completely satisfactory local theory in the class of regular functions. It
tells us, roughly speaking, that the quasilinear Cauchy problem (43) has the same
behavior - as far as local existence and continuous dependence of the solutions
on the data are concerned - as an ordinary differential equation:iJ = 'P(t, y) in the
finite-dimensional setting. In addition, since our solutions are weak Wi-solutions
as well, they coincide on the common interval of existence with weak vVi-solutions
obtained by any other method, provided there is a uniqueness theorem for weak
Wi-solutions.
In addition, our approach is general. It is by no means restricted to second
order reaction-diffusion systems, but applies to quasilinear systems of Petrowskii
parabolic type of arbitrary order. Furthermore, the coefficients of A and B, as
well as the nonlinearities, can be nonlocal functions. This allows to deal rather
easily with elliptic-parabolic systems, for example, by reducing them to abstract
parabolic evolution equations. The general abstract theory can also be applied
to produce nontrivial results for free boundary value problems, and to systems
with 'dynamic boundary conditions'. It can handle 'highly degenerate' problems
that are not even Petrowskii parabolic, as well as quasilinear parabolic problems
involving pseudodifferential operators. Thus, at the end, we are rewarded by a
rather flexible general theory that requires the verification of a few hypotheses
only. Those can be met in many concrete situations occurring in applications, as
is demonstrated by the examples given in the third volume.

\Venn Du wissen willst, was bewiesen


wurde, schau den Beweis an. 5
Ludwig Wittgenstein
(1889-1951)

5If you want to know what has been proven, study the proof.
Notations and Conventions
In this section we collect some of the definitions and notations that we use throughout
this treatise. The reader is expected to have a working knowledge of the basic theory
of locally convex spaces and (unbounded) linear operators in Banach spaces. Thus we
review here only briefly the most basic facts from linear functional analysis in order to
fix the notations. We refer to the literature (in particular, to [Ber74], [DuS57], [Edw65],
[Golb66], [HiP57], [Hor66], [JarSl], [Sch71], [Tre67], and [Yos65]) for proofs and many
more details.

1 Topological Spaces
Let X and Y be nonempty sets. Then Y x is the set of all maps u: X ...... Y. If A is a
subset of X, the characteristic function of A (in X) is denoted by XA, that is,

XA(X) = 1 if x E A and XA(X) = 0 if x E A C := X\A .

Occasionally, we put 1 := Ix := Xx.


We denote by
N:= {O, 1,2, ... }
the set of all nonnegative integers, and N := N U {oo}. Similarly, i := JR U {±oo}, etc.
We put
x fly:= min{x,y}, x V y:= max{x,y} , x,yEJR,
and
x+ := x V 0, x-:= (-x) V 0
are the positive and negative parts, respectively, of x E JR, more generally, of x E JR x .
The set of all x E X in which definitions and relations hold is often denoted by [ ... ],
where ... stand for the definitions and relations. For example, if u E JRx then

[u :::: 0] := { X EX; u(x) :::: 0 }

etc.
Let X and Y be topological spaces. Then C(X, Y) is the set of all continuous maps
in yX. We write
X '-> Y or i: X '-> Y ,
if X is continuously injected in Y, that is, X C Y and the natural injection

i:X ...... Y, x>-+x

d d
is continuous. If X is a dense subset of Y, we write X C Y. Thus X '-> Y means that
X is densely and continuously injected in Y.
We often write Ix or simply 1 for the identity mapping, idx : X ...... X, x >-+ x, if
no confusion seems likely.
2 Notations and Conventions

Let X := (X, d) be a metric space and let Ai be a nonempty subset of X. Then

lE(M,s) := lEx(M,s) := {x EX; dist(x,M) < s}


is the open s-neighborhood of M in X and JE(M, s) := { x EX; dist(x, M) ::; s} is
the corresponding closed s-neighborhood. If l'vf = {x}, we write lE(x,s) and JE(x,s) for
lE({x},s) and JE({x},s), respectively, and call these sets open, resp. closed, balls with
center at x and radius s.
Let A and B be subsets of some topological space X. We usc the symbols if, cl(A),
or clx(A) to denote the closure of A in X. Similarly, B,
intB, or intx(B) stands for the
interior of Bin X. Then A is compactly contained in B, in symbols:

AcCB,
_ 0
if A is compact and contained in R.

2 Locally Convex Spaces


By a vector space, we always mean a vector space over K where K = lR or K = C. If 111 is
a subset of a vector space, we put

!VI:= M\{O} .

Let X be a nonempty set and let Y be a vector space. Then Y x is given the
canonical vector space structure defined by

(mL + f3v)(x) := (m(x) + f3v(x) , x EX, u, v E yX, a, f3 E K .

If X is also a vector space (over the same field), Rom(X, Y) is the vector subspace of yX
of all linear maps and End(X) := Rom(X, X), the space of endomorphisms.
Let X and Y be topological vector spaces (TVSs). Then LeX, Y) is the vector
subspace of Rom(X, Y) consisting of all continuous linear maps, and

LeX) := LeX, X) .

In this case X '-> Y means also that X is a vector subspace of Y, that is, i E £( X, Y).
Moreover,

£is(X, Y) := { T E LeX, Y) ; T is bijective and T- 1 E LeY, X) }

is the set of all topological linear (toplinear) isomorphisms from X onto Y and

£aut(X) := £is (X, X)

is the group of all top linear automorphisms of X, the general linear group, also denoted
by Q£(X).
By a locally convex space (LCS) we mean a Rausdorff locally convex TVS. If
X is a LCS, there exists a separating family of continuous semi norms on X inducing the
topology. Conversely, each separating family of seminorms P on a vector space X induces
Locally Convex Spaces 3

a coarsest locally convex Hausdorff topology on X such that each pEP is continuous
(cf. [Hor66, II.4]). Occasionally, we shall write X := (X, P) if X is a LCS whose topology
is induced by the family of semi norms P.
Let P and Q be two families of seminorms on the same vector space X. Then P is
said to be stronger than Q (and Q is weaker than P) if

(X, P) '-7 (X, Q) .

If P is stronger and weaker than Q, that is, if P and Q induce the same topology, then
P and Q are equivalent. In this case we write

(X, P) ='= (X, Q) .

Let X and Y := (Y, Q) be LCSs. We write Ls(X, Y) if L(X, Y) is given the simple
convergence topology induced by the family of seminorms

{T f-+ q(Tx) ; x E X, q E Q} ,
whereas L(X, Y) means that this vector space is always equipped with the bounded
convergence topology defined by the family of semi norms

{T f-+ SUPxEB q(Tx) ; Be X bounded, q E Q}


Observe that Ls(X, Y) and L(X, Y) are LCSs and

L(X, Y) '-7 Ls(X, Y) .

If X and Yare normed vector spaces, the topology of Ls(X, Y) is called strong (op-
erator) topology and the one of L(X, Y) uniform operator topology. Observe that
L(X, Y) is then a normed vector space with the uniform operator norm defined by

IITII := sup{ IITxl1 ! Ilxll ; x EX}

(provided X i= 0, of course). Moreover, L(X, Y) is a Banach space iff Y is one [Golb66,


Theorem 1.3.5 and Corollary 1.5.8].
Let X and Y be LCSs. A linear map T: X --t Y is compact if it maps bounded
sets into compact sets. We put

K(X, Y) := { T E L(X, Y) ; T is compact} ,

and JC(X) := K(X,X). Hence T E K(X, Y) iff T E L(X, Y) and T(B) is relatively com-
pact for each bounded subset B of X. Observe that the last condition implies the bound-
edness ofT, hence its continuity, if X is metrizable. (It should be noted that our definition
of a compact operator is not the one usually employed in the general theory of TVSs
(e.g., [Jar81]).
We say that X is compactly injected in Y,

X '-<--+ Y ,

if i: X '-7 Y and i E K(X, Y).


4 Notations and Conventions

If X is a normed vector space, we denote the open [resp. closed] unit ball in X
by lE := lEx [resp. i]. Moreover, x + c:lE := lE(x, c:) and lEn := lElRn, where JR:.n - more
generally, K n - is always given the euclidean norm, 1·1, unless explicitly stated otherwise.
Let X be a TVS. Then the dual (space) X' of X is given by X' := £'(X, K). Ac-
cording to our conventions X' is always endowed with the bounded convergence topology
that is called strong topology of X' in this context if X is a LCS. We often denote
the value of x' E X' at x E X by (X',X)X, or simply by (x', x), and call the bilinear map

(., . ) x := (., .) : X' X X -+ K

duality pairing between X' and X.


Let X be a LCS. We put X:V. := £.s(X,K) since the simple convergence topology
is called w* -topology of X' in this context. The w· -topology of X' is also denoted
by U(X',X).
Due to the Hahn-Banach theorem [Hor66, IILI] the family

Px' := {x ...... 1(x',x)1 ; x' E X'}

is a separating family of seminorms on X. Hence (X, P X') is a LCS and the topology in-
duced by P X ' is called weak topology of X and denoted by u(X, X'). We also write XU!
for (X,U(X,X')). It is obvious that

X '-> Xw and X' '-> X:" •.

Sometimes we use ~ [resp. ~] to denote weak [resp. weak star] convergence (that is,
convergence with respect to the weak [resp. w*-] topology).

3 Complexifications
Let X be a real vector space. We denote by Xc the additive group X x X equipped with
the multiplication by complex numbers (00,13) = a + ij3 E C given by

(a, j3)(x, y) := (ax - j3y, j3x + ay) .

We identify xEX with (x,O)EXxX and find that (O,I)(x,O) = (O,x). Thus each
z E X x X is uniquely represented as z = x + iy with x, Y E X. Hence X + iX := Xc
is a complex vector space, the complexification of X, where X + iX is to be un-
derstood as the direct sum over JR:. of the real vector spaces X and i X. In particular,
x + iy ...... x =: Re(x + iy) and x + iy ...... Y =: Im(x + iy) are JR:.-linear projections of Xc
onto X. Observe that X is identified with an JR:.-linear subspace of Xc, the real subspace.
If X is a real LCS , Xc is a complex LCS and X '-> Xc (as an JR:.-linear injection).
Let X, Y and Z be real vector spaces. Given T E Hom(X, Y), define the complex-
ification, Tc, of T by

Tc(x + iy) := Tx + iTy , x + iy E Xc .

Then Tc E Hom(Xc, Yc) and Tc(X) C Y. If no confusion seems likely, we write T for Tc,
that is, we consider Hom(X, Y) as an JR:.-vector subspace of Hom(Xc, Yc), the real vector
Complexifications 5

subspace of real linear maps in Hom(Xc, Ye). The elements R of this space are char-
acterized by R(X) C Y. Given S E Hom(Xc, Yc), there exist unique SI, S2 E Hom(X, Y)
such that Sx = SIX + iS2X for X EX. Hence

Sz = Sx + iSy = SIX + iS2x + i(SIY + iS2Y) (3.1)


= SIX + iSl y + i(S2X + iS2Y)
for z = X + iy E Xc. Thus Sz = SIZ + iS2Z for z E Xc, where we have used the identi-
fications Sj = (Sj)c. Hence, letting ReS:= SI and ImS := S2,

S = Re S + i 1m S , S E Hom(Xc, Yc) , (3.2)

where ReS and 1m S are real linear maps in Hom(Xc, Yc). Observe that S is a real linear
map in Hom(Xc, Ye) iff 1m S = o. Also observe that

(ST)c = ScTc , T E IIom(X, Y), S E Hom(Y, Z) .

Given S, T E Hom(X, Y) C Hom(Xc , Ye), it follows that

S+iTE [Hom(X,Y)]c.

Hence we deduce from (3.2) that, by identifying Hom(X, Y) with the vector subspace of
real linear maps in Hom(Xc, Ye), we can - and will - identify Hom(Xc, Yc) with the
complexification of Hom(X, Y), that is,

Hom(Xc, Yc) = [Hom(X, Y)]c . (3.3)

Now suppose that X and Yare real LeSs. Then

TEL:(X, Y) ¢=> Tc E L:(Xc, Yc)

and
T E K(X, Y) ¢=> Tc E K(Xc, Yc) .
Moreover, (3.3) easily implies

L:(Xc, Yc) = [L:(X, Y)]c .

In particular, since JR.c = C we see that

(Xe)' = (X')c (3.4)

and it follows from (3.1) that the duality pairing (., ·)xc : X~ x Xc ~ C is given by

(f + ig,x + iy)xc = (f,x) - (g,y) + i ((J,y) + (g,x))


Also observe that
(Tc)' = (T')c , TEL:(X, Y) .
Now suppose that (X, 1I·11l is a (complete) normed vector space over R Then Xc is
a (complete) normable vector space over C It is easily verified that

Ilzllxc:= sup II Re(eictz)11 = sup Ilxcosa + ysinall , z = X +iy E Xc,


O:S;a:S;2rr 0::; 0:::; 27r
6 Notations and Conventions

defines a norm on Xc and that X '-+ Xc is a real linear isometry. If Y is a second real
normed vector space, it is easily verified that

£(X, Y) '-+ £(Xc, Yc) is a real linear isometry


provided, of course, £(Xc, Yc) is given the uniform operator norm induced by 1I·llxc
and II· llYn:·

3.1 Remark Let (F, 11·11 F) be a complex Banach space of complex-valued func-
tions and let (X, 11·11) be the real subspace of IR-valued functions. Then Xc ~ F but
IHxc =1= II·IIF, in general. For example, let F:= (1[2,1'11)' where 1·1t stands for the £1-
norm. Then X = (IR2, I·h). Let x,y E X be given by x:= (2,0) and y:= (0,1). Then
x + iy = (2, i) E 1[2 and Ix + iYll = 2. On the other hand,
Ix+iylx" = sup 1(2cosa,sina)11

=sup{2lcosal+lsinal; OSaS27r}=V5. •
Although (3.4) is true, it follows (from Remark 3.1, for example) that, in general,
the dual norm of (Xc)' is different from the norm 11·11 (X'k if X is a normed vector space.
Thus (3.4) has to be replaced by the more precise relation
(Xc)' ~ (X')c ,
if X is a real normed vector space.
~ ow suppose that (X, (,1 .») is a real inner product space. Then (Xc, (·1·) xc) is a
complex inner product space where

(x + iYlu + iv)x: := (xlu) + (ylv) + i [(ylu) - (xlv)] .

In this case the inner product norm Ilx + iY11 := (11x11 2 + IlyI12)1/2 is a norm on Xc such
that X '-+ Xc is a real linear isometry.

4 Unbounded Linear Operators


Let X and Y be LeSs. If A is a linear operator with domain, dom(A), in X and range
in Y, we write
A: dom(A) C X --+ Y .
Observe that dom(A) is a vector subspace of X and A E Hom( dom(A), Y). In general,
the domain and the image of a map f are denoted by dom(f) and im(f), respectively.
We denote by C(X, Y) the set of all closed linear operators A : dom(A) eX --t Y,
and C(X) := C(X,X). If X and Yare normed vector spaces,

D(A) := (dom(A),II'IIA) ,
where
IlxiiA := IIAxl1 + Ilxll , x E dom(A) ,
is the graph norm of A. If X and Yare Banach spaces, D(A) is a Banach space iff
A E C(X, Y).
General Conventions 7

Let X and Y be LCSs such that X '---+ Y, and let A: dom(A) C Y -+ Y be linear
in Y. Then the X-realization, Ax, (the part of A in X, or the maximal restriction of A
to X) is the linear operator in X, defined by

dom(Ax) := { x E X n dom(A) ; Ax EX} , Axx:= Ax.

If no confusion seems likely, we often omit the index X. Observe that Ax E C(X)
if A E C(Y).

5 General Conventions
Let X be a real LCS. If in a given formula there are implicit or explicit references to
complex numbers, it is always understood that this formula is interpreted in its complex-
ified version. Thus, for example, if X is a (lK- )Banach space and A is a linear operator
in X, we mean by o-(A) and peA) the spectrum and the resolvent set, respectively, of A
if lK = C and of Ac if lK = lit

We often use c to denote constants, that may differ from occurrence to occurrence, but
are always independent of the specific free variables occurring at a given place. Thus we
use c very much in the same way as the Landau symbol O. If the equations under con-
sideration depend on additional parameters, say a, (3, ... , we sometimes write c(oo, (3, ... )
to indicate this.
Chapter I

Generators and
Interpolation

The abstract theory of linear parabolic problems, developed in the first part of
this treatise, rests on two cornerstones: on the theory of analytic semigroups and
on interpolation theory. In the first section of this chapter we discuss in some
detail the classes of generators of analytic semigroups that are used throughout.
In the second section we review the fundamentals of interpolation theory. We also
introduce the class of 'admissible interpolation functors' that, on the one hand, is
flexible enough to build an easy and general abstract theory of evolution equations
on it, and, on the other hand, is general enough to cover all applications occurring
in practice.

Throughout this chapter E, F, Ejl and Fj , j = 0,1,2, ... , denote Banach spaces.

1 Generators of Analytic Semigroups


In this section we study the properties of generators of analytic semigroups which
are basic for the whole treatise. Although these results are well-known, in princi-
ple, our approach has some novel aspects. In particular, we develop quantitative
versions of perturbation theorems and related results which show that much of the
theory can be controlled by two numerical parameters, which -- as we shall see in
later chapters ~ are readily accessible in concrete applications.
We assume that the reader is familiar with the basic facts about strongly
continuous semigroups on Banach spaces. We refer to [ClH+S7], [DavSO], [FatS3],
[GolsS5], [HiP57]' [Kat66], [KraZ76]' and [PazS3] for these and many more details.
10 I Generators and Interpolation

1.1 Properties of Linear Operators


For the reader's convcnience we include an easy proof of the following useful per-
turbation theorem concerning bounded invertibility.

1.1.1 Proposition Let X be a connected metric space, let

BEC(X,L(E,F)) ,

and suppose that there exists (3 > 0 such that

IIB(x)ell 2: (3llell , XEX, eEE. (1.1.1)

Then B(X) n Lis(E, F) of 0 implies B(X) C Lis(E, F).

Proof It follows from (1.1.1) that B(x) is injective for x E X. Thus, thanks to
Banach's homomorphism theorem, it suffices to prove that B(x) is surjective for
each x E X.
Let Xo := {x EX; B(x) E Lis(E, F) }. Then Xo of 0 by assumption. Since
Lis(E, F) is open in L(E, F), the continuity of B implies that Xo is open in X.
Let (x J ) be a sequence in Xo with Xj --> x in X and let f E F be given. Letting
ej := B-1(xj)f, it follows that

B(xj)(ej - ed = f - B(Xj)ek = [B(Xk) - B(Xj)]ek .

Hence we deduce from (1.1.1) that

which shows that (ej) is a Cauchy sequence in E. Thus ej --> e for some e E E and

B(x)e - f = B(x)(e - ej) + [B(x) - B(xj)]ej --> 0 .

This shows that Xo is closed in X. Thus Xo = X by the connectedness of X .•

Suppose that El '--7 Eo. Given a linear map A: El -+ Eo, we can interpret it
as a linear operator, Au, in Eo with domain E 1 . If no confusion seems likely, we
write again A for AD.

1.1.2 Lemma Let A: El --> Eo be linear. Then A E L(El' Eo) n C(Eo) iff the
graph norm (of Ao) is an equivalent norm for E 1 .

Proof Let A E L(El' Eo) n C(Eo) and let i: El '--7 Eo. Then

IIXIIA = IIAxllo + Ilxllo ::; (II All + Ilill) Ilxlll ,


and, consequently, El '--7 D(A). Since A is closed, D(A) is complete. Hence Ba-
nach's homomorphism theorem implies D(A) '--7 E 1 .
1.1 Generators of Analytic Semigroups 11

Conversely, if the graph norm is equivalent to 11·111, there exists ex > 0 such that

Hence A E £(El, Eo). Since D(A) ~ E l , we see that D(A) is complete. Conse-
quently, A E C(Eo) .•

1.2 The Class H(El' Eo)


By a densely injected Banach couple we mean a pair of Banach spaces, (Eo, E l ),
d
such that El '----+ Eo.

Throughout the remainder of Section 1 we s'uppose that (Eo, Ed is a densely in-


jected Banach couple.

We denote by
H(El,Eo)
the set of all A E £(El' Eo) such that -A, considered as a linear operator in Eo
with domain E l , is the infinitesimal generator of a strongly continuous analytic
semigroup {e- tA ; t::::: O} on Eo, that is, in £(Eo).
In order to derive uniform estimates for these semigroups and related oper-
ators, it is important to possess quantitative descriptions of H(El' Eo). For this
purpose, given K, ::::: 1 and w > 0, we write

if w +A E £is(EJ, Eo) and

K,-l < liP, + A)xllo Re>':::::w, (1.2.1)


- 1>'lllxllo + Ilxlli
where 11·llj is the norm in E j .

1.2.1 Remarks (a) Inequality (1.2.1) has the advantage of being symmetric
with respect to Ilxlli and IAlllxllo. However, it is easily verified that it suffices to
require less in order that a linear map A: EI ----+ Eo belongs to H(El' Eo, K"w) for
some K, ::::: 1 and w > O. In fact, suppose that there are K, ::::: 1 and w > 0 such that
w + A E Lis(E1 , Eo) and

1>.lllxll o s: K, 11(>' + A)xll o , x E E1 , Re A ::::: w . (1.2.2)

Then, letting
12 I Generators and Interpolation

it follows that A E H (EI' Eo, 1'>:1, w). Similarly, if there are I'>: ~ 1 and W > 0 such
that w + A E Lis(EI , Eo) and

IlxliI : : : I'>: II(A + A)xll o , x E E I , Re A ~ w , (1.2.3)


then A E H(EI' Eo, 1'>:2,W), where 1'>:2 := 1+ I'>: + I'>: IIAII.c(E E)' On the other hand,
1, 0
if A E H(EI,Eo,l'>:,w) for some I'>: ~ 1 and w > 0, estimates (1.2.2) and (1.2.3) are
trivially true .
Proof Let (1.2.2) be satisfied. Then we infer from
X=(W+A)-I[(A+A)x+(W-A)X] ,
and from W ::::: IAI for Re A ~ W that

IlxliI : : : II(w + A)-III.c(Eo,E,) (1 + 21'>:) II(A + A)xll o ' x E EI , Re A ~ w .


Hence
IAlllxlio + IlxliI : : : 1'>:(1 + 311(w + A)-III.c(Eo,E,)) II(A + A)xll o
: : : 1'>:1 II(A + A)xll o
for x E EI and Re A ~ w. Moreover,

II(A + A)xllo::::: IAlllxlio + IIAII.c(E"Eo) Ilxil I (1.2.4)


: : : (1 V IIAllqE"Eo )) (IAlllxllo + IlxlII)
for x E EI and Re A ~ w. This implies the first assertion.
Similarly, if (1.2.3) is satisfied,

IAlllxll o : : : II(A + A)xllo + IIAII.c(EloEO) Ilxil I


: : : (1 + I'>: IIAllqE"Eo )) II(A + A)xll o
Hence
IAlllxlio + IlxilI : : : 1'>:211(A + A)xll o
for x E EI and ReA ~ w. By taking (1.2.4) into consideration, the second assertion
follows as well. _

(b) Suppose that EI = Eo =: E. Then H(E, E) = £(E). Moreover,


A E H(E, E, 2, 1+ 21IAII), A E £(E) .

Proof Since H(E, E) c C(E), the closed graph theorem implies H(E, E) c £(E).
If A E £(E) then, given A E [Re z ~ 1+ 211AII],
(1/2)(IAlllxll + Ilxll) : : : (IAI-IIAII) Ilxll : : : II(A + A)xll
: : : (IAI + IIAII) Ilxll : : : 2(IAlllxll + Ilxll)
for x E E. Hence (1.2.1) is satisfied with I'>: := 2 and w := 1+ 211AII. Lastly, we
infer from O"(A) C [izi : : : IIAII] that w + A E £aut(E). _
1.1 Generators of Analytic Semigroups 13

The basic result concerning the classes H(E 1, Eo, 1\;, w) is the following:

1.2.2 Theorem H(E 1, Eo) = U H(E 1, Eo, I\;,w).


,,::::1
w>o

Proof Suppose that A E H(E 1 , Eo, I\;,w). Lemma 1.1.2 implies A E C(Eo). From
(1.2.1) and Proposition 1.1.1 we easily deduce that
[Rez ~ w] C p(-A)
and
Re..\ ~w. (1.2.5)
This is well-known to imply A E H(E 1, Eo) (e.g., [Fat83, Section 4.2]).
Suppose that A E H(El, Eo). Then (cf. [Fat83]' for example) A E C(Eo) and
there exist W > 0 and I\; ~ 1 such that (1.2.5) is true. Now the assertion follows
from Remark 1.2.1(a) .•

Since H(E 1 , Eo) C C(Eo), the dual A' E C(EiJ) is defined for A E H(El' Eo).
It is convenient to put Eg := EiJ and Art := A'. Then

E~ := E~(A~) := D(A~)

is a Banach space such that E~ '-+ Eg. If Eo is reflexive, A~ is densely defined so


that (Eg, E~) is a densely injected Banach couple.

1.2.3 Proposition Suppose that Eo is reflexive and A E H(El,EO,I\;,w). Then


A~ E H(EL Eg, 11:~,w), where 11:" := 1011:2 /(w AI).

Proof Since p(A) = p(Art) and (..\ + Art)-l = [(..\ + A)-lJ'for..\ E p(-AU), it fol-
lows from (1.2.5) that [Rez~w]cp(-A") and 1..\llIxUllo::;I1:II("\+A")x~lh for
xrt E E~ and Re..\ ~ w, where we denote the norm in EJ again by 11·llj. Hence we
infer from Remark 1.2.1(a) that Art E H(Ef, E~, 11:1,W), where

11:1 :=1\;(1+311(w+Art)-11I,cc Eg,E1l)'

thanks to the fact that II Art 11.c(E~ E~) ::; 1. Note that
" 0

IIx~lh ::; lI(w + Art)xrtllo + (1 + w) IIx~llo ::; (1 + 11:) II(w + A")x"lIo + IIx"llo

for x" E E~. Hence


II(w + AU)-1x~111 ::; (1 + 11:) Ilx~lIo + lI(w + A~)-lXUlio ::; (1 + 11: + l1:/w) Ilx"lIo
::; (311:/(w AI)) IIxUlio
for x~ E Eg. Now the assertion follows .•
14 I Generators and Interpolation

The class H(E I , Eo) has been introduced by the author in [Ama88a] and,
independently, in [ClH+87, Chapter 5], where it has been called Hol(E1,Eo). The
quantitative versions H(E1,Eo,K,W) are implicitly introduced in [Ama88a] since
it is easily verified that a subset U of H(E I , Eo) is regularly bounded (in the sense
of [Ama88a]) iff there exist K:::: 1 and w > 0 with U c H(EI,Eo,K,W).

1.3 Perturbation Theorems


Thanks to the quantitative information contained in Theorem 1.2.2 it is easy to
give quantitative versions of important perturbation theorems. Some of them are
collected in the following:

1.3.1 Theorem (i) H(E 1 , Eo) is open in £(EI' Eo). In fact, given r E (O,I/K)
and A E H(EI,Eo,K,W),

(ii) Let A E H(EI,Eo,K,W), 0 < r < I/K, and!3:::: O. Then, given BE £(EI,Eo)
satisfying
IIBxllo :::; r Ilxll l + !3llxllo ,
it follows that
A+BEH(E I ,Eo,K/(I-Kr),wV!3/r) .

Proof (i) Given B E £(EI' Eo) with IIBII :::; r, we see that

(K- I - r)(I>'lllxll o + IlxllI) :::; K-I(I>'lllxll o + IlxllI) - r Ilxll i


:::; 11(>' + A)xllo -IIBxllo :::; 11(>' + A + B)xllo :::; 11(>' + A)xllo + IIBxllo
:::; K(I>.lllxll o + IlxllI) + r Ilxlll :::; (K + r)(I>.lllxll o + IlxllI)
for x E El and Re>. :::: w. Since K + r :::; K/(1 - Kr) =: KI, it follows that A +B
satisfies (1.2.1) with K replaced by KI. From this we deduce that

x E EI , 0:::; t :::; 1.

Now Proposition 1.1.1 implies w + A + BE Lis(E I , Eo). Hence

(ii) Observe that


I.l Generators of Analytic Semigroups 15

Hence

(1\;-1 - r)(I'\lllxll o+ Ilx111) ::; (1\;-1 - JJlI,\I) l'\lllxllo + (1\;-1 - r) IIxl1 1


::; I (,\ + A)xllo - IIBxllo ::; I (,\ + A + B)xllo ::; I (,\ + A)xllo + IIBxllo
::; (I\; + JJlI'\1) l'\lllxllo + (I\; + r) IIxl11 ::; (I\; + r)(I'\lllxll o+ Ilxlh)

for x E E1 and Re'\ ::::: w V (3/ r. Now the assertion follows from the last step of the
proof of (i). _

1.3.2 Corollary Let A be a compact subset ofH(E1, Eo). Then there exist I\; ::::: 1
and w > 0 such that H(E1' Eo, 1\;, w) is a closed neighborhood of A in .c(E1' Eo).

Proof From Theorem 1.3.1(i) we know that each A E A has a neighborhood of


the form H(E1,Eo,I\;A,WA). Thus the compactness of A implies the existence of
I\;j ::::: 1 and Wj > 0, 0::; j ::; m, with

Since 1\;' : : : I\; ::::: 1 and w' ::::: W > 0 imply H(E1' Eo, 1\;', w') :J H(E1' Eo, 1\;, w), the as-
sertion follows, since it is easily verified that each Ji(E1' Eo, 1\;, w) is closed. _

Theorem 1.3.1 is a quantitative formulation of well-known perturbation the-


orems for generators of analytic semigroups (e.g., [Paz83, Theorem 3.2.1]). The
fact that H(E 1,Eo) is open in .c(E1,Eo) has also been observed in [ClH+87,
Theorem 5.3].

1.4 Spectral Estimates


It is a consequence of Theorem 1.2.2 and known properties of generators of analytic
semigroups that, given A E H(E1' Eo), there exist wand a sector

~19 := [I arg zl ::; 19 + 1l,/2] U {O} ,

where 0 < 19 < 7r /2, such that w + ~19 belongs to the resolvent set of -A. The
following proposition gives the sharper result that such a 19 can be chosen uniformly
for all A E H(E 1,Eo,l\;,w).

1.4.1 Proposition Given 1\;::::: 1 andw > 0, there arewo E (O,w) and 19 E (0,7r/2)
+ ~19 C p( - A) and
such that Wo

(104.1)
16 I Generators and Interpolation

Proof Put

0:= arcsin [(2 - )2)/(2",)] /\ arctan [(2 - )2)/4] E (0,7f/2) .


Suppose that A E H(E1' Eo, ""w) and>' E 2w + ~(I with Re>. :::; w. Then
Iarg(>. - 2w)1 = 'P + 7f/2
with 0 < 'P :::; 0 and 1>'1 ~ 11m AI ~ w cot O. Let Ao := 2w + i7], where 7] is defined
by 1771 := I>' - 2wl / cos 'P and sign( 7]) = sign(Im A) .

Then
II(A + A)xllo ~ 11(>'0 + A)xllo -I>' - >'olllxli o
(1.4.2)
~ (",-ll>'ol-IA - >'01) Ilxllo + ",-1 IIxl1 1
Since
17]1 = I>' - Aoi/ sin'P = IA - 2wl/ cOS'P ~ I>' - 2wl
and IAol ~ 17]1, it follows that
",-11>'01-1>' - Aol ~ 17]1 (",-1 - sin'P) ~ I>' - 2wl (",-1 - sinO)
(1.4.3)
~ IAI (1- 2w/I>'I)/()2",) .

Since Re>. :::; w, we see that 2w/I>'1 :::; 2tanO :::; 1- 1/\1"2. By inserting the latter
estimate in (1.4.3) we deduce from (1.4.2) that

On the other hand,


11(>' + A)xllo :::; II(w + A)xllo + I>' - wlllxll o
:::; ",(w Ilxllo + IIxlld + (IAI + w) Ilxlio (1.4.4)
:::; 3",(I>'llIxll o + Ilxlld
1.1 Generators of Analytic Semigroups 17

for>. E 2w + ~e with Re>. ::; w, since K 2: 1 and

1>'12: 2w cos B = 2wJ1- [(2 - V2)/(2K)] 2 2: V2w .


Hence estimate (1.4.1) is true if >. E 2w + ~(i satisfies Re>. ::; w.
Put 6 := 2w /(5K) and suppose that w - 6 ::; Re>. ::; w. Let >'0 := w + i 1m >.
and observe that w - 6 = w(l- 2/(5K)) 2: w/2 implies
(1)'01- K I>' - >'01)/1>'1 2: 1 - K6/(w - 6) 2: 1 - 2K6/w = 1/5 .
Hence it follows from (1.4.2) that
11(>' + A)xllo 2: 1>.1/(5K) Ilxll o + K- 1 II x l1 1 2: (5K)-1(1>.lll x ll o + Ilxll1) .
On the other hand, w - 62: w/2 implies w ::; 21>'1 for w - 6::; Re>. ::; w. Hence we
deduce from (1.4.4) that

11(>' + A)xllo ::; (KW + 1>'1 + w) Ilxll o + K IIxl1 1 ::; 5K(I>.llI x ll o + Ilxlh) .

o w-6 Wo w 2w
Consequently, estimate (1.4.1) is satisfied for>. E [w - 6 ::; Re z ::; w]. It is trivially
satisfied if Re >. 2: w. Put
6 tan B )
Wo := w - 6/2 and f}:= arctan ( 2(w + 6)

and observe that estimate (1.4.1) is satisfied for>. E Wo + ~ij. Now the assertion
is a consequence of Proposition 1.1.1. •

If A is a closed linear operator in E, we define its spectral bound, s(A), by


s(A) := sup{ ReA; >. E (T(A) } ,
where sup 0 := -00. Note that
s(a+A) =Rea+s(A) , aEe.
Also note that
s(-A)<w, AE1i(E1,Eo,K,W) ,
where, of course, A is interpreted as a linear operator in Eo.
18 I Generators and Interpolation

In many cases ~ in connection with stability problems for evolution equa-


tions, for example ~ it is important information if s( -A) < O. Such information is
not contained in the notation H(EI' Eo, K, w), since w > 0 by definition (which is a
convenient assumption for an casy handling of the class H(EI' Eo, K, w)). However,
we have the following important inclusion:

1.4.2 Proposition Given AD E H(E1,Eo,K,W), any WI > 0, and an arbitrary


(5 > s( - AD)! there exist K1 2': 1 and a neighborhood A of AD in £( E 1 , Eo) such that
(5+AcH(E 1 ,Eo,Kl,W1) .
If (52': w then H(El! Eo, K,W) C A.

Proof By Proposition 1.4.1 we know that w + ~D C p( - AD) and

II(A + A O)-lIIC(Eo ,E, ) ::; 5K , AEW+~O , (1.4.5)

for some {}:= {}(K,W) E (0,7f/2). Since


o
A := [ReA 2': (5] n (w + ~D)C
is a compact subset of p( -AD) and (A f--t A + AD) E C(A, Lis(E1 , Eo)),

max{ II (A + AO)-lllc(Eo,E,) ; AE A} < 00 , (1.4.6)


where max 0 := -00. From (1.4.5) and (1.4.6) we deduce the existence of a constant
M 2': 1 such that

II(A+Au)-lIIC(EO,E,)::; M, ReA 2': (5 . (1.4.7)

Put r := 1/(2M) and observe that


A+Ao+B= [l+B(A+A o)-l](A+Ao)
and 1 + B(A + AO)-l E £aut(Eo) with

I [1 + B(A + Ao)-lr11Ic(Eo ) ::; 2


for ReA 2': (5 and B E rlBlC(E"E o), thanks to (1.4.7). Hence
A + Au + BE Lis(E1 , Eo)
and

Consequently, we deduce from (1.4.8) that [ReA 2': 0] c p( -((5 + A)) and

IlxliI ::; 2M I (A + ((5 + A))xll o '


Now the first part of the assertion follows easily from Remark 1.2.1(a).
1.1 Generators of Analytic Semigroups 19

If u ::=: w then (1.2.1) implies

Ilxlll::; KII((>--u)+u+A)xll o ' Re(>- - u) ::=: w - u, x E El ,


for each A E H(El' Eo, K, w). Thus the last part of the assertion again follows from
Remark 1.2.1(a). _

Let u > s(-Ao) be given. Fix Ul E (s(-Ao),u) and put Wl:= u - Ul. Then
Proposition 1.4.2 implies the existence of a neighborhood A of Ao in £(El' Eo)
such that

s(-A) = Ul + s( -(Ul + A)) < Ul +Wl = U, AEA, (1.4.9)

Moreover, the following corollary is valid:

1.4.3 Corollary Given Ao E H(El) Eo) K,W) and U > s( -Ao), there exist M > 0,
a neighborhood A of Ao in £(El) Eo), and {} E (0, 1f /2) such that

u+~19cp(-A)

and
IIU + AIIL(E"Eo ) + (1 + I>-I/- j II [A + (u + A) rlIIL(Eo,E j ) ::; M
for>- E ~19, j = 0, 1, and A E A. If U ::=: W then H(El' Eo, K, w) c A.

Proof Fix (j E (s( -Ao), u) and put E := u - (j. By Proposition 1.4.2 there exist
Kl ::=: 1 and a neighborhood A of Ao in £(El) Eo) such that
(j + A C H(El' Eo, Kl, E) .
If u ::=: W then H(El' Eo, K,W) C A. Moreover, Proposition 1.4.1 implies the exis-
tence of {} E (0, 1f /2) such that E + ~19 C p( -A) and

for A E H (El' Eo, Kl, E). By replacing in this estimate A by (j + A for A E A, we


see that

II(u + A)xllo = II(E + (j + A)xllo ::; 5Kl(E Ilill + 1) Ilxlll , x E El ,

where again i: El '----+ Eo, and that

for x EEl, >- E ~19, and A E A. Now the assertion follows with

since M E lR.+. _
20 I Generators and Interpolation

Proposition 1.4.2 and Corollary 1.4.3 will be important for getting uniform
estimates for parabolic evolution operators for time-dependent linear evolution
equations. However, a first application of Corollary 1.4.3 will lead to the pertur-
bation result of Theorem 1.5.1.

1.5 Compact Perturbations


A family Jt of linear operators from E to F is collectively compact if

U{ K(lffiE) ; K E Jt} is relatively compact in F .

Observe that Jt is a bounded subset of the Banach space K(E, F) of all compact
linear maps from E to F. Using this definition we can formulate the following
theorem about compact perturbations.

1.5.1 Theorem Suppose that Jt c £(El' Eo) is collectively compact. Given Ii ~ 1


and w > 0, there exist iiI ~ Ii and WI ~ w such that

Proof By Corollary 1.4.3 there exists M such that

(1.5.1)

for x EEl, and

Re).~w, (1.5.2)

for A E 7t(El,Eo,Ii,W). From (1.5.1) and (1.5.2) we deduce that

11'\('\ + A)-lXlll :S M 11).(w + A)('\ + A)-lxllo


(1.5.3)
:S M211(w + A)xll o :S M311xlll
for x E El and Re'\ ~ w.
Let K be the closure of U{ K(lffiEJ ; K E Jt} in Eo. Since El is dense
in Eo, given Xo E K, there exists Xl E lffiEo (xo, 1/(61i)) n E l . Hence we deduce from
(1.5.3) and Remark 1.2.1(a) that

11(,\ + A)-lxlll = 11('\ + A)-l [Xl + (xo - Xl) + (x - xo)] III


:S 1).1- M311xIill
1 + 1/6 + Ii Ilx - xollo
for x E Eo, Re'\ ~ w, and A E7t(El' Eo, Ii, w). Consequently, there exists w(xo) ~w
such that
1.1 Generators of Analytic Semigroups 21

for A E H(El' Eo,~, w). By the compactness of J( we find Xo, . .. ,X m E J( such that
{lIllEo (Xj, 1/(6~)) ; j = 0, ... , m} is a covering of J(. Thus, letting
WI := max{ W(Xj) ; 0 :::; j :::; m} ,
we see that
Re A 2: WI, X E J( ,

which implies
II(A + A)-l KIIc(E 1 ) :::; 1/2 , Rc A 2: WI, K Eft. (1.5.4)
Since
A+A+K= (A+A)(l+(A+A)-lK) ,
we obtain from (1.5.4) that A E p( -(A + K)) and

II(A+A+K)-111c(Eo,E1 ):::; 2~
for A E H(El,Eo,~,w), K E ft, and ReA 2: WI' Now the assertion follows from
Remark 1.2.1(a). _

Theorem 1.5.1 - in the case that ft consists of one operator only - is proven
in [ClH+S7, Theorem 5.6] and, independently, by Desch and Schappacher [DeS88],
although it was around before as a 'folk theorem'. Our extension to a collectively
compact set of linear operators is trivial of course but useful.

1.6 Matrix Generators


Let (Fa, F l ) be a densely injected Banach couple. Then (Eo x Fa, El x Fd is a
densely injected Banach couple as well. Suppose that

A = [~~~ ~~~] E £(El X FI, Eo x Fa) ,

where we use obvious matrix notation. For definiteness, in the following theorems
we use the l'I-norm on the product Banach spaces.

1.6.1 Theorem Suppose that


All E H(El' Eo, ~l, WI) and A22 E H(Fl' Fa, ~2, W2) .

Put
~ := ~1(1+ ~21IA2111.C(El,Fo») V ~2 and W:= WI V W2

and suppose that there are r E (0, 1/~) and {3 2: 0 such that

IIA12yllE o :::; r IIyllF1 + {3llyliFo ' (1.6.1)

Then A E H(EI X F l , Eo x Fa, ~/(1- ~T),W V (3/T).


22 I Generators and Interpolation

Proof Put
C:= [A
All AD ] . (1.6.2)
21 22
Given (~, TJ) E Eo x Fo and>' E [Re z ::;, w], the equation (>. + C)(x, y) = (~, TJ) has
the unique solution

Hence it follows from (1.2.1) that

and
1>'lllyllFo+ IIYllFl :::; K211TJ - A 2l (>' + All)-l~IIFo
:::; K2(IITJIIFo + IIA2111 Klll~IIEo)
:::; (K2 V KI K211A21 11) II(~, TJ)IIEoXFo
On the other hand,

11(>' + C)(x,Y)IIEoxFo = 11(>' + An)xllEo + IIA 21 X + (>. + A22)YIIFo


:::; Kl(I>'lll xI Eo + IlxllEl) + IIA2l llllxll E1 + K2(1)'lllyllFo+ IlylIF,)
:::; [(Kl + IIA2111) V K2J (1).III(x,y)IIEoXFo+ II(x,y)IIE1XF,) .
From these estimates we deduce that

(1.6.3)

Put
B:= [~ A~2] (1.6.4)

and note that A = C +B and, thanks to (1.6.1),

IIB(x, y)IIEoxFo = I A12YIIEo :::; T II(x, y)IIE,XF, + jJ II(x, Y)IIEoXFo


for (x, y) EEl X Fl. Now (1.6.3) and Theorem 1.3.1(ii) imply the assertion .•

1.6.2 Remarks (a) Suppose that All and A22 satisfy the hypotheses of Theo-
rem 1.6.1 and replace K by

Then A E H(EI X H, Eo x Fo, ;;)(1- KT),W V 13), provided

Proof This follows from Theorem 1.6.1 by replacing the densely injected Banach
couple (Eo x Fo, El x F l ) by (Fo x Eo, Fl X El) .•
1.1 Generators of Analytic Semigroups 23

(b) The hypotheses that All E 1i(El' Eo) and A22 E 1i(Fl,Fo) are necessary for
the validity of Theorem 1.6.1 if A12 = 0.
Proof Suppose that (1.6.3) is true. Given (~, T)) E Eo x Fo and)" with Re).. ~ w,
there exists a unique (x, y) E El X Fl satisfying

Hence).. + All E Cis(El' Eo) by Banach's theorem. Letting ~ = 0, the second equa-
tion reduces to
().. + A 22 )y = T) .
Thus).. + A22 E Cis(Fl' Fo). Moreover, it follows from (1.6.3) that, letting T) := 0,

IlxllEl ~ II(x,y)IIE,xF, ~ K:11()"+All)xII Eo '

and, letting ~ := 0,

yEH,

Now the assertion follows from Remark 1.2.1(a). _

1.6.3 Corollary Suppose that Fl = Fo =: F. Then

A E 1i(El X F, Eo x F) iff All E 1i(El' Eo) .

A E 1i(El x F,Eo x F,K:,w) )

Proof Suppose that All E 1i(El' Eo, K:l, Wi)' Thanks to Remark 1.2.1(b) we
know that A22 E 1i(Fl' Fo, 2,1 + 211A2211). Let

1£:= K:l(l + 211A2111) V 2 and w:= Wi V (1 + 211A2211)


and note that (1.6.1) is satisfied with r := 1/(21£) and (3 := IIAdl. Now the second
part of the assertion follows from Theorem 1.6.1.
Let Band C be defined by (1.6.4) and (1.6.2), resp. Since B E £(Eo x F)
and A = B + C, it follows from Theorem 1.3.1(ii) that A E £(El X F, Eo x F)
iff C E £(El X F, Eo x F). Now Remark 1.6.2(ii) shows that All E 1i(E1 , Eo) is
implied by A E 1i(El X F, Eo x F). _
24 I Generators and Interpolation

It should be remarked that throughout this whole section we have not tried to
find the best choices for K and w. In fact, we have tried to choose simple estimates
for these numbers. It is the only purpose of the explicit formulas for K and w in
Theorem 1.3.1, and in the theorems of this subsection, to exhibit the dependence
of these constants on the various quantities involved. This information is useful if
one has to consider subsets of Ji and uniform estimates are needed.
If A12 = 0, a qualitative version of Theorem 1.6.1 has been proven - by a
different method - in [Nag89, Corollary 3.3]. A quantitative version of Corol-
lary 1.6.3 appears in [Ang90, Lemma 2.6] and a qualitative form of it in [Ama91].
For the sake of simplicity and for the importance in applications we restrict
ourselves to the case of densely injected Banach couples (Eo, El). Everything
proven - with the notable exception of Theorem 1.5.1 - remains valid without
the density assumption, provided we define Ji(El,Eo) by the formula of Theo-
rem 1.2.2. On the other hand, since Ji(El' Eo) c C(Eo), a result of Kato [Kat59]
implies that El is dense in Eo if Ji(El' Eo) i= 0 and either El or Eo is reflexive.
(Note that El is reflexive iff Eo is reflexive in the case that Ji(El' Eo) i= 0 since
then Lis(El' Eo) i= 0.) We refer to the monograph of Lunardi [Lun95] for a de-
tailed exposition of the theory of analytic semigroups with non dense domains of
the generators.

2 Interpolation Functors
In this section we collect the basic facts from the theory of interpolation spaces
that we use freely in the remainder of this work. For proofs and many more details
we refer to [BerL76], [BenS88], [ButB67]' [KrPS82], and [Tri78].

2.1 Definitions
The pair (Eo, Ed is said to be an interpolation couple if there exists a LCS X
such that E j '--t X, j = 0, 1. In this case Eo n El and Eo + El are well-defined
Banach spaces. Observe that Eo n El ~ El and Eo + El ~ Eo if El '--t Eo so that
we can choose X to be Eo. If (Eo, E 1 ) is an interpolation couple and

then E is said to be an intermediate space with respect to (Eo, El).


Let 8 be the category of (IK- )Banach spaces. Thus the objects of 8 are the
IK-Banach spaces, the morphisms of 8 are the bounded linear operators, and the
composition is the usual composition of maps. We denote by 8 1 the category of
interpolation couples, that is, the objects of 8 1 are the interpolation couples, the
morphisms of 8 1 are the elements A of £(Eo+El, Fo+F1 ) satisfying A E £(Ej, F i ),
1.2 Interpolation Functors 25

j = 0,1, where (Eo, E 1 ) and (Fo, Fd are interpolation couples, and the compo-
sition is the natural composition of maps. We write A: (Eo, El) -+ (Fo, Fl) if
(Eo, E 1 ) and (Fo, F1 ) are interpolation couples and A is a morphism of B 1 .
Let (Eo, E 1 ) and (Fo, F 1 ) be interpolation couples. Then E and F are said
to be interpolation spaces with respect to (Eo, Ed and (Fo, Fd if E and F
are intermediate spaces with respect to (Eo, E 1 ) and (Fo, F 1 ), respectively, and
A E £(E, F) whenever A: (Eo, E 1 ) -+ (Fo, Fl)' Moreover, E and F are said to be
interpolation spaces of exponent 0, where 0 < () < 1, with respect to (Eo, E 1 )
and (Fo, F1 ) if there exists c((}) > 0 such that

IIAIIL(E,F) :::; c(fJ) IIAI11(~o,Fo) IIAII~(El,F,) (2.1.1)

for A: (Eo, E 1 ) -+ (Fo, Fl)' If c(fJ) = 1 then E and F are exact interpolation spaces
of exponent fJ with respect to (Eo, Ed and (Fo, Fl)'
Lastly, a covariant functor J from Bl into B is said to be an [exact] in-
terpolation functor [of exponent 0] if, given interpolation couples (Eo, Ed and
(Fo, F 1 ), it follows that J(Eo, E 1 ) and J(Fo, F I ) are [exact] interpolation spaces
[of exponent fJ] with respect to (Eo, E 1 ) and (Fo, F1 ) and if

2.2 Interpolation Inequalities


The following lemma shows that the norms of intermediate spaces 'generated by
interpolation methods' possess an important convexity property.

2.2.1 Proposition Let J e be an interpolation functor of exponent fJ. Given an


interpolation couple (E o,El), putEe :=J e (Eo,El)' Then

x E Eo n El . (2.2.1)

Proof Since (lK, lK) is an interpolation couple, lKe ~ lK. Given x E Eo n E l , define
A: (lK,lK) -+ (Eo,E1 ) by AA:= AX. Observe that

j = 0,1.

Hence we deduce from (2.1.1) the existence of a constant c(fJ) > 0 such that

which implies the assertion. _


26 I Generators and Interpolation

Estimate (2.2.1) is also said to be an interpolation inequality. It follows


from (2.2.1) and Young's inequality, that is, from
x, Y E lR+, 0 <0<1 ,
that
IlxllEe :::; C IIxl11 + C(O)c IJ /(IJ-I) Ilxllo , x E Eo n E 1 , c>0, (2.2.2)

where c(O) := (1 - O)OIJj(l-e)c~/(I-e). Kote that Ce = 1 in (2.2.1) if J e is an exact


interpolation functor such that J Ii (lK, lK) = K

2.3 Retractions
An object, Y, in a given category is a retract of the object X if there are mor-
phisms, r: X --+ Y, a retraction onto Y, and r C: Y --+ X, a coretraction for r,
such that the diagram
id
Y Y

(2.3.1 )

X
is commutative. In other words: the morphism r: X --+ Y is a retraction of X
onto Y if there exists a right inverse for r in the given category.

2.3.1 Lemma Let r E £(X, Y) be a retraction in the category of Less and let
r C E £(Y, X) be a coretraction for r. Then p := rCr E £(X) is a projection and
X is a topological direct sum,
X = Xp EB X 1 - p , Xp:= im(p) , X 1 - p := im(l - p) .
Furthermore, r C E £is(Y,Xp) and X 1 - p = ker(p) = ker(r).

Proof Observe that p2 = (rCr)(rCr) = rCr = p. Moreover,


im(l - p) = ker(p) :J ker(r) .
Since p(x) = 0 implies 0 = rp(x) = r(x), it follows that ker(p) = ker(r). Now the
assertion is obvious .•

2.3.2 Proposition Let r: (Eo, Ed --+ (Fo, Fd be a retraction in the category 8 1


of interpolation couples, and let r C: (Fo, F 1 ) --+ (Eo, E 1 ) be a coretraction for r.
Suppose J is an arbitrary interpolation functor from 8 1 into 8. Then

is a retraction and r C is a coretraction for r = J(r).


I.2 Interpolation Functors 27

Proof This is an immediate consequence of (2.3.1) and the functorial properties


of J .•

As an application of the considerations we now show - by a proof which we


owe essentially to D. Daners (also cf. [DaK92, Proposition 3.4]) - that, in case
we are dealing with product spaces, we can often interpolate 'component-wise'.

2.3.3 Proposition Let (Eo, E I ) and (Fa, F I ) be interpolation couples such


that EI '---+ Eo and FI '---+ Fa. Let J be an arbitrary interpolation functor from BI
into B. Then (Eo x Fa, EI x F1 ) is also an interpolation couple, and

Proof It is obvious that (Eo x Fa, EI x F I ) is an interpolation couple with

EI x FI '---+ Eo x Fa .
Let rE: Eo x Fa -+ Eo be the natural projection and iE : Eo -+ Eo x Fa the nat-
ural injection. Observe that

and
iE: (Eo, EI) ---> (Eo x Fa, EI x FI) .
Moreover, rE is a retraction and iE is a coretraction for rE in the category B I .
Thus it follows from Lemma 2.3.1 and Proposition 2.3.2 that

J(Eo x Fa, El x Fd = im(p) EB ker(p) ,


where

and that
iE E Lis(J(Eo, E I ), im(p)) .
Let rF: Eo x Fa -+ Fa be the natural projection and iF: Fa -+ Eo x Fa the natu-
ral injection. Of course, r F and iF enjoy properties that are completely analoguous
to the ones of rE and iE, respectively. Moreover, iFrF = 1 - p and, consequently,
since ker(p) = im(l - p),

iF E Lis (J(Fo,Fr) , ker(p))


From this we deduce that

Now the assertion is obvious .•


28 I Generators and Interpolation

2.4 Standard Interpolation Functors


In this subsection we introduce briefly the 'classical' interpolation functors which
are of utmost importance in concrete applications.

2.4.1 Example: Real Interpolation Functors Let (Eo, E I ) be an interpolation


couple. Given x E Eo + E I , define the K-functional by

K(t, x) := K(t, x, Eo, Ed := inf{ IlxollEo + t IlxIIIEl ; X = Xo + Xl}


for t > O. Put

and
(Eo, EI)e,q := ({ x E Eo + El ; Ilxlle,q < 00 }, 11·lle,q)
for 0 < 8 < 1, 1 ~ q ~ 00. Let
~e,q(Eo, E l ) := (Eo, Ede,q and ~e,q(A):= A
for A: (Eo, EI ) ----> (Fo, FI)' Then, given any q E [1,00] and 8 E (0, I), it follows
that ~e,q is an exact interpolation functor of exponent 8. Henceforth we denote
it by
(', . )e,q
and call it the real interpolation functor of exponent 8 and parameter q . •

2.4.2 Example: Complex Interpolation Functors Suppose that lK = C and


let (Eo, E I ) be an interpolation couple. Let F(Eo, E I ) be the set of all bounded
and continuous functions f from S into Eo + E l , where 8:= [0 < Rez < 1], such
that fl8 is holomorphic and

j = 0, 1,

where 8 j := [Re z = j] and Co is the space of continuous functions vanishing at


infinity. Then F(Eo, Ed is a Banach space with the norm

IlfIlF(Eo,E,) := sup Ilf(it)IIE V sup Ilf(1 + it)IIEl


tEIR ° tEIR

Given 8 E (0, I), put

[Eo, El]e := ({ x E Eo + EI ; f(8) = x for some f E F(Eo, El ) }, 11·lle) ,


where
Ilxllll := inf{ 1IJIIF(Eo,E,) ; f(8) = x} .
Letting ~e(Eo, E l ) := [Eo, EI]e and ~/J(A) := A for A: (Eo, EI ) ----> (Fo, FI), it fol-
lows that ~ e is an exact interpolation functor of exponent 8.
I.2 Interpolation Functors 29

Now suppose that][{ = R Given an interpolation couple (Eo,E l ), we put

Thus, in order to determine 'J(}(Eo, E l ) we complexify Eo and E l , apply the in-


terpolation functor 'J () for complex spaces, and 'decomplexify' afterwards, that is,
restrict ourselves to the real subspace of the interpolation space [(Eok, (Elklo.
Defining 'Jo(A):= A for A: (Eo,E l ) -> (Fo, F l ), it follows that 'J o is an exact
interpolation functor of exponent 8 in this case as well.
In the general case, that is, if][{ = lR or C, we denote the interpolation functor
introduced by
[', 'le
and call it the complex interpolation functor .•

2.4.3 Remark Let (E, II· liE) be a complex Banach space of complex-valued func-
tions and let (F, 11·11) be the real subspace of real-valued functions. Then we know
from Remark 3.1 that Fe ~ E but II·IIF" -I-II·IIE, in general. This fact has to be
kept in mind when employing the complex interpolation functor in the real case.•

2.4.4 Example: Continuous Interpolation Functors We denote by 8 2 the


subcategory of 8 1 of all densely injected Banach couples, that is, 8 2 is the sub-
category of 8 1 whose objects are the densely injected Banach couples. Given two
such couples (Eo, E l ) and (Fo, H), it follows from the closed graph theorem that
A: (Eo,El ) -> (Fo,Fl ) iff A E .c(Eo,Fo) and A(El) C Fl. On the category 8 2 we
now define another important interpolation functor.

Let (Eo, E l ) be a densely injected Banach couple. Given 8 E (0, I), put

(Eo, EJ)~,CX) := closure of El in (Eo, El)(},oo .


Let (Fo) F l ) be a second object of 8 2 and suppose that A: (Eo, E l ) -> (Fo, Fl)'
Since A(El ) C F l , it follows that

A E .c((Eo, El)~,CX), (Fo, Fr)~,CX») .


Hence, putting 'Je(Eo,El):= (Eo,El)~oo and 'Je(A) :=A, we obtain an exact
interpolation functor of exponent 8. He~ceforth we denote it by

(', ')~,CX)
and call it the continuous interpolation functor. It is known that

(2.4.1 )

It is also known (e.g., [DorFS7]) that (" .)oIJ ,00 coincides with the 'continuous inter-
polation method' of Da Prato and Grisvard [DaPG79l .•
30 I Generators and Interpolation

2.5 Continuous Injections


Suppose that (Eo, E 1) is an interpolation couple. Given () E (0,1), we define an
intermediate space with respect to (Eo, E 1) by

(2.5.1)

Then the following important injections are valid:

d d d
Eo n E1 '---+ (Eo, E 1)f"q '---+ (Eo, E 1)1),1 '---+ [Eo, E1]1)
(2.5.2)
d o d
'---+ (Eo, E 1)1),00 '---+ (Eo, E 1)1),00 '---+ (Eo, E1kq '---+ Eo + E1
for 1 ::; q < 00 and 0 < ( < TJ < ~ < 1. It is also true that

(2.5.3)

for 1 ::; q < r < 00, 0 < () < 1. Moreover,

(2.5.4)

for 0 < () < 1 and 1 ::; q ::; 00.

2.6 Duality Properties


Given q E [1,00]' we define the 'dual exponent' q' E [1,00] by

1 1
-+-=1.
q q'

Then we have the following duality properties:

(2.6.1)

and
(2.6.2)
with respect to the duality pairing naturally induced by (".) EonE, , provided
d .
Eo n El '---+ E j for J = 0, 1.
d
If either Eo or El is reflexive and Eo n El '---+ Ej , j = 0, 1, then

[Eo, Ed~ ~ [Eb, E~]e , (2.6.3)

with respect to the duality pairing naturally induced by (.,.) EonE,' Moreover,
[Eo, E 1 ]e is reflexive for 0 < () < 1.
I.2 Interpolation Functors 31

2.7 Compactness
In the following, we put

j = 0, 1, 1:S; q :s; 00 . (2.7.1)

Suppose that E1 '-<......+ Eo. Then


(2.7.2)

2.8 Reiteration Theorems


As for the stability of the interpolation methods, we have the following reiteration
theorem for the real methods: Suppose that

(Eo, EdeJ ,l '---+ Fj '---+ (Eo,E1)ej ,oo , j = 0,1, (2.8.1)

where 0 :s; OJ :s; 1 and 00 i= 01 . Then


(Fo, F1)"f/,q ~ (Eo, E1)(1~"f/)eo+"f/e,.q , (2.8.2)

Similarly, as in (2.7.1), we put

j = 0,1. (2.8.3)

Then assuming, for simplicity, that (Eo, E 1) is a densely injected Banach couple,
the reiteration theorem for the complex method says that

[[Eo, E1leo , [Eo, E1le']"f/ = [Eo, Ed(l~"f/)eo+"f/II, , (2.8.4)

Observe that (2.5.2) and (2.8.1), (2.8.2) imply

([Eo, E1]e o, [Eo, E1]e,) "f/,q ~ (Eo, El)(1~"f/)eo+"f/e"q (2.8.5)

for 00 ,0 1 ,7] E (0, 1) with 00 i= 01 and 1 :s; q :s; 00. On the other hand, it can be
shown that

[(Eo, E1)eo ,qO, (Eo, E1)e"q']"f/ ~ (Eo, E1)(1~"f/)eo+"f/e"q (2.8.6)

for 00 ,01 ,7] E (0,1) with 00 i= 01 and 1 :s; qj :s; 00, where

and where the case qo = q1 = 00 is excluded. (We refer to [Gri66] and [MS90] for
further commutation properties for interpolation functors.)
32 I Generators and Interpolation

Let (2.8.1) be satisfied. Then (2.8.2) implies that (Fo, FI)g,DO equals the clo-
sure of Fo n FI in (Eo, E I )(1_ry)lI o +ryll, ,00' From (2.5.2), (2.5.3), and (2.8.2) we
deduce that

Fo n FI !:.. (Fo, FI)ry,1 ~ (Eo, EdCI-ry)lIo+ryll"l !:.. (Eo, EI)Cl-ry)lIo+ryll"oo .

Hence
0<7)<1. (2.8.7)

2.9 Fractional Powers and Interpolation


Let A E C(E) be such that lE.+ C p( -A) and there exists a constant M with

). E lE.+ . (2.9.1)

Assumption (2.9.1) is easily seen to imply the existence of constants c := c(M)


and <p := <p(M) E (0,7r) such that

R<p:= [Iargzl:::; <p] U [Izl:::; <p] C p(-A) (2.9.2)

and
(2.9.3)
Let r be any piece-wise smooth simple curve in R<p running from ooe-i<p to ooei<p
and avoiding lE.+. Then, given any z E C with Re z < 0, put

A Z := ~ f(_).)zP.,+A)-ld)'. (2.9.4)
27r2 Jr
Observe that AZ is a well-defined bounded linear operator in E, thanks to esti-
mate (2.9.3).

If A is densely defined, it is well-known and will be proven in Section II1.4.6


that the definition of the 'fractional power' AZ of A can be extended to all z E C
in such a way that AZ E C(E) with dense domain, A Z IA z2 x = A Z l+ Z2 X for any
x E dom(A2'Tn), where mEN with m ~ 2 and Re Zl V Re Z2 < m, and such that
AZ is the 'ordinary power' if z E Z. It follows that

(2.9.5)

and that z f--> AZx is continuous at Zo E C if x E dom(Azo+I), for example. Fur-


thermore (e.g., [Tri78 , Theorem 1.15.2]),

(E, D(A'Tn)) Re z/'Tn,1 !:.. D(AZ) !:.. (E, D(A'Tn))~ez/'Tn'oo (2.9.6)

for 0 < Rez < m and mEN.


1.2 Interpolation Functors 33

If the purely imaginary powers of A are locally uniformly bounded, that is,
Ait E £(E) for t E [-1,1] and there exists a constant K such that

It I ~ 1, (2.9.7)

then
[D(A<», D(Ai3) lo ~ D(A(1-0)<>+Oi3) (2.9.8)
for 0 ~ Re a < Re;3, ° < e < 1. More precisely, let
E A := (dom(A<», IIA<>·II) , Rea2:0.

Then, given e E (0,1), there exists a constant c := c(K) 2: 1 such that


c- 1 11·II E(l-e)a+Bi3
A
~ 11·11 [Ea
A'
Ei3]
A f)
~ C 1I·IIE(l-e)a+ei3
A
(2.9.9)

°
for ~ Rea < Re;3 and all densely defined A E C(E) satisfying (2.9.7). This fol-
lows by an obvious modification of the proof of [See71, Theorem 3].

Suppose that H is a Hilbert space and A is a self-adjoint linear operator in H.


Denoting by {E).. ; ,\ E ~} the spectral resolution of A, we can define AZ by

zEC, (2.9.10)

provided a(A) c ~+, that is, A 2: 0. If A is positive definite, i.e., A 2: a for some
a > 0, it follows that {AZ ; z E C} are the fractional powers of A in the sense of
the definition. If A 2: 0, we have the following improvement over (2.9.8):

(2.9.11)

for ° ~ Rea < Re;3 and 0< e < 1 (e.g., [Tri78 , Theorem 1.18.10]).
Lastly, we recall that, given mEN,

(2.9.12)

°
is an equivalent norm on (E,D(Am))o,q for < e < 1 and 1 ~ q ~ 00, uniformly
with respect to every densely defined A E C(E) satisfying (2.9.1). This is a conse-
quence of (the proof of) [Tri78 , Theorem 1.14.3].

2.10 Semigroups and Interpolation


Next we give some estimates for the K-functional in terms of semigroups. For this
subsection we refer, in particular, to [Tri78] and [Lun95].
34 I Generators and Interpolation

Given M ~ 1 and (J E JR., we denote by

the set of all A E C(E) such that -A is the infinitesimal generator of a strongly
continuous semigroup { e- tA ; t ~ 0, } on E satisfying

t ~ o.
Moreover,
Q(E):= U Q(E, M, (J)
(TElR
M;::-l

is the set of all negative generators of strongly continuous scmigroups on E.


Given A E Q(E),

type(-A):= inf{ (J E lR; there exists M ~ 1 with A E Q(E,M,(J)}

is the type of -A (or the 'exponential growth bound' of -A).

Suppose that AI, ... , An E Q(E, M, (J) for some (J :::; 0 and that the semi-
groups {e- tAj ; t ~ O}, j = 1, ... ,n, are pair-wise commuting. Put

E:= nn

j=1
D(Aj) (2.10.1)

and

(2.10.2)

Then (E, E) is a densely injected Banach couple and {e- sA ; s E (JR+)n} is a


strongly continuous n-parameter semigroup on E that has E as an invariant sub-
space. Put

w(t, x):= sup 11(1 - e-SA)xIIE , t>O, xEE. (2.10.3)


sE(O,t)n

Then there exists a constant c := c(M) ~ 1 such that

c- 1 K(t, x, E, E) :::; (1 II t) II xii + w(t, x) :::; cK(t, x, E, E) (2.10.4)

for t > 0 and x E E. If (J < 0, there exists a constant c := c(M, (J) ~ 1 such that

c- l K(t, x, E, E) :::; w(t, x) :::; cK(t, x, E, E) (2.10.5)

for t > 0 and x E E. Using these estimates it is obvious how to get equivalent
characterizations and equivalent norms for the real interpolation spaces (E, E)e,q
as well as for the continuous interpolation spaces (E, E)~ 00'
1.2 Interpolation Functors 35

Lastly, let (Eo, E 1 ) be a densely injected Banach couple and A E H(El' Eo).
Also suppose that there exists M 2': 1 such that
. tA
II(tA)1e- II.C(E o):::; M , t > 0, j = 0,1 . (2.10.6)

Given () E (0,1) and 1 :::; q :::; 00, put

x E Eo. (2.10.7)

Then
(Eo, E1)e,q = {x E Eo ; IIXlle,q,A < 00 }

and there exists a constant c := c(M) 2': 1 such that

c- 1 Ilxlle,q :::; Ilxllo + IIxlle,q,A :::; c IIxlle,q , (2.10.8)

whenever A E H(El' Eo) satisfies (2.10.6).


Given (J" < 0, there exists a constant c := c(M, (J") 2': 1 such that
c- 1 IIxlle,q :::; Ilxlle,q,A :::; c Ilxlle,q , (2.10.9)

for all A E H(El' Eo) satisfying (2.10.6) with type( -A) :::; (J". Finally,

(Eo, Ed~,oo = {x E Eo ; lim ee IltAe-tAxllo = O} (2.10.10)


t->O

if A E H(El' Eo) satisfies (2.10.6).


Estimate (2.10.9) has the following generalization. Suppose mEN and (J" <0
and, given A E H(El' Eo) with type( -A) :::; (J", put

Then, given M 2': I, there exists a constant c := c(M) 2': 1 such that

c- 1 11'II(EO ,EA')8,q :::; Ile me IltmAme-tA'lloIILq(IR+,dt/t) (2.10.11 )


:::; c 11·llc E o,EA')8,Q

for all A E H(El' Eo) satisfying type( -A) :::; (J" and (2.10.6) for j = 0, 1, ... , m.
This follows from the proof of [Tri78 , Theorem 1.14.5].

2.11 Admissible Interpolation Functors


Suppose that 0 < () < 1. By an admissible interpolation functor, denoted by (" ')(1,
we mean an interpolation functor of exponent () for the category of densely injected
Banach couples such that

El is dense in (Eo, El)e ,


36 I Generators and Interpolation

whenever (Eo, Ed is such a couple. Observe that the real interpolation func-
tors (', ')e,p, 1:::; p < 00, the complex interpolation functor [', ']e, and the 'contin-
uous' interpolation functor (', .)Oe ,00 are admissible. If no confusion seems possible,
we put
Ee := (Eo, E 1)e
and denote the norm in Ee by 11·lle. In the following, we fix for each () E (0,1) an
admissible interpolation functor arbitrarily.

2.11.1 Theorem If (Eo, Ed is a densely injected Banach couple,

0<,8<0:<1. (2.11.1)

Given 0:,,8, 'Y E [0,1] with 0:::; 'Y < ,8 < 0: :::; 1, there is c:= c(o:,,8, 'Y) such that

Ilxll{3 :::; cIlxll~n-(3)/(n-,) Ilxll~-,)/(n-,) , x E En. (2.11.2)

If E1 <.L-.-.7 Eo then
En <.L-.-.7 Ej3 , 0 :::; ,8 < 0: :::; 1 . (2.11.3)
Conversely, E1 <.L-.-.7 Eo if En <.L-.-.7 E{3 for some 0:,,8 E [0,1] with,8 < 0:.
Proof It follows from the extremal property of the real method (e.g., [BerL76,
Theorem 3.9.1]) that

(Eo, Ede,l :!.. Ee :!.. (Eo, Ed~,oo , O<(}<I. (2.11.4)

Hence (2.11.1) is an immediate consequence of (2.5.2). Thus (E{3, En) is a densely


injected Banach couple whenever 0 :::; ,8 < 0: :::; 1. Letting

we infer from (2.11.4) and the reiteration property (2.8.2) of the real method that

(E" Eo.)1),l '--> (Eo, E 1){3,1 '--> E{3 .


Now (2.11.2) follows from Proposition 2.2.1 and Example 2.4.1.
Assertion (2.11.3) is a consequence of (2.7.2), (2.11.4), and (2.5.2). The last
part of the statement follows from E1 '--> En <.L-.-.7 E{3 '--> Eo and the fact that
compositions of linear operators are compact provided at least one of them is
compact .•

2.11.2 Remarks (a) Let (Eo, Ed be a densely injected Banach couple. Then
the scale of Banach spaces {En ; 0 :::; 0: :::; I} possesses the following almost re-
iteration property:

if 0:::; 0: < ,8 :::; 1 and 0 < 'f/- < 'f/ < 'f/+ < 1 then
(En, E(3)1)+ '--> E(1_1))et+1){3 '--> (Eet, E(3)'1_ .
1.2 Interpolation Functors 37

Proof It follows from (2.5.2), (2.11.4), and (2.8.2) that

and
(Eo,Ed(l-O:)1'}+1'}/3,l ~ ((Eo,E1)o:,1,(Eo,E1kl)1'},1 '-+ (Eo:,E/3)1'}
for °< 1] < 1. This implies the assertion. _
(b) Of course, it is a consequence of the reiteration theorems for the real and
complex methods that

(2.11.5)

if (', ')11 is one of the real interpolation functors or (', ')11 = (', .)~ for each () E (0, 1),
IX)

or if (', ')11 = [', ·lll for each () E (0,1). Moreover, (2.8.5) and (2.8.6) imply the va-
lidity of (2.11.5) for appropriate other choices of (', ')11 as well.
(c) Let (Eo, E 1 ) be a densely injected Banach couple such that 1i(El' Eo) i= 0.
Then it is easily verified that El '-'----7 Eo iff there exists A E 1i (El' Eo) possessing a
compact resolvent. This is the case iff each A E 1i(El' Eo) has a compact resolvent
(considered as 'unbounded' linear operators in Eo, of course). _
Chapter II

Cauchy Problems and


Evolution Operators

Any theory of abstract quasilinear parabolic problems requires, of course, a good


understanding of the theory of linear parabolic evolution equations. In this chapter
we develop that part of the linear theory which is based upon the concept of evo-
lution operators. The latter correspond to the fundamental matrices in the theory
of ordinary differential equations. This 'classica1' theory is particularly well-suited
for the study of quasilinear parabolic problems exhibiting smoothing phenomena
since it exploits the fact that the solution of a linear parabolic evolution equation
has in general better regularity properties than its initial value.
Besides its importance for the study of quasilinear parabolic problems, the
concept of evolution operators is of intrinsic interest for many questions concerning
the behavior of solutions to nonautonomous linear problems - in particular, under
periodicity assumptions.
In the first section of this chapter we essentially describe the basic results
about linear parabolic Cauchy problems. In Section 2 we introduce the concept
of a parabolic evolution operator. Section 3 contains some technical results about
operator-valued Volterra integral equations. These are needed for the construction
of parabolic evolution operators in Section 4. Section 5 is devoted to the proof
of some stability estimates that are basic for the study of abstract quasilinear
parabolic problems. In Section 6 we investigate invariance properties of evolution
operators. In particular, we derive sufficient conditions for the positivity of these
operators in the case of ordered Banach spaces. These results can be considered
to be abstract versions of the maximum principle for second order elliptic and
parabolic differential equations.

Throughout this chapterE, F, E j , and Fj , j = 0, 1, ... , are Banach spaces.


40 II Cauchy Problems and Evolution Operators

1 Linear Cauchy Problems


In this section we collect some of the basic existence and regularity results for
linear parabolic evolution equations whose proofs are given in later sections. For
a concise formulation we first introduce some classes of function spaces.

1.1 Holder Spaces


Let X be a nonempty set. We write
B(X,E)
for the Banach space of all bounded maps from X to E, equipped with the supre-
mum norm 11·1100' If X is a topological space,
BC(X,E)
is the closed linear subspace of B(X, E) consisting of all bounded and continuous
functions.
Let X = (X, d) be a metric space. Then
BUC(X,E)
is the closed linear subspace of BC(X, E) consisting of all bounded and uniformly
continuous functions. Given p E (0,1], we put
lIu(x) - u(y) II
[u]p := [u]p,x := [u]CP(X,E):= x~~x [d(x,y)]P (1.1.1)
xiy

for u: X ....... E, and


1I·lb := 11·1100 + [']p .
Then
BUCP(X, E) := ({ u E B(X, E) ; [u]p < oo}, 1I·lIcp ) (1.1.2)p
is the Banach space of all bounded and uniformly p-H6lder continuous E-valued
functions on X. If p = 1, we write BUC1-(X, E) for (1.1.2h in order to avoid confu-
sion with spaces of differentiable functions. Of course, the elements of BUC 1-(X, E)
are the bounded and uniformly Lipschitz continuous functions from X to E. In gen-
eral, we use 1- to denote Lipschitz continuity. However, in numerical calculations
we identify 1- with the number 1. In particular, s < 1- for each s E lR. with s < 1.
Given p E (0,1) U {l-}, we denote by
CP(X, E)
the set of all u E C(X, E) such that each point in X has a neighborhood Y such
that ulY E BUCP(Y, E), that is, if u is (locally) p-HOlder continuous (Lipschitz
continuous if p = 1-).
11.1 Linear Cauchy Problems 41

1.1.1 Lemma Given u E CP(X, E) and K cc X, there exists a neighborhood Y


of K in X such that ulY E BUCP(Y, E). Thus Holder (resp. Lipschitz) continuous
functions are uniformly Holder (resp. Lipschitz) continuous on compact sets.

Proof This follows from an obvious modification of the proof of [Ama90b, Propo-
sition 6.4j .•

Lemma 1.1.1 implies that {11·llcP(K,E) ; K cc X} is a separating family


of seminorms on CP(X, E) inducing a locally convex Hausdorff topology. Thus
CP(X, E) is a LCS. It is well-known and not difficult to prove that

(1.1.3)

for p, IJ, T E [0,1) U {1-} with T :s; IJ :s; p, where BUCo := BUC and CO := C.

Now let X be an open subset of F or a perfect subset of K Then, given


p E [0,1) U {1-}, we denote by

the Banach space of all continuously differentiable functions u in BUC(X, E) such


that 8u E BUC(X, £(F, E)), and [8ujp,x < 00 if P > 0, equipped with the norm

u f-+ IluIIBUCl+P(X,E) := Ilulb+p := ~~1118julloo,x + [8ujp,x ,

where 8 denotes the Frcchet derivative and 8° := id, and where the last term is
omitted if p = 0. Similarly, u E C1+P(X, E) if each point in X has a neighbor-
hood Y such that ulY E BUC1+ P(X, E). This space is a LCS with respect to the
obvious topology. Lastly, we put 2- := 1 + 1-.
Of course,

for IJ E [0,2) U {l-, 2-}, etc., if Y is a nonempty subset of E.


Holder spaces occur naturally by interpolation as is witnessed by the following
'embedding'theorem.

1.1.2 Proposition Let (Eo, E 1 ) be a densely injected Banach couple and let
(', ')f) be an admissible interpolation functor'. Let X be open in F OT a convex
perfect subset of lK.. Then
42 II Cauchy Problems and Evolution Operators

Proof Given Xo EX, there exists a convex neighborhood Y of Xo in X such that


ulY E BC(X,E1 ) and 8uIY E BC(Y,£(F,Eo )). Hence the mean-value theorem
implies

Ilu(x) - u(y)lIo ::; 11 8ullc(Y,.c(F,Eo)) Ilx - yll , x,yEY.

Now we deduce from the interpolation inequality (1.2.11.2) that

Ilu(x) - u(y)llo ::; c Ilu(x) - u(y)II~-o lIu(x) - u(y)ll~


::; 2°c 118ull~(t,.c(F,Eo» lIull~(y,Ed IIx _ yill-O
::; c(O) lIullc(y,E1)ncl(y,Eo) IIx - yill-O

for x, y E Y. Now the assertion is obvious .•

A bilinear and continuous map of norm at most one,

(1.1.4)

is said to be a multiplication. If S is a nonempty set, we define point-wise mul-


tiplication
Ef x E~ -+ E S , (Ul,U2) f---t UleU2
induced by (1.1.4) by

Ul e ~(8) := Ul (8) e U2(8) , 8 E S.

It is obvious how to extend this definition to more than two factors.

The following proposition shows that HOlder spaces are 'stable' under point-
wise multiplication.

1.1.3 Proposition Let X be a metric space and let p E [0,1) U {l-}. Then the
point-wise multiplication induced by (1.1.4) is a bilinear and continuous map:
(i) BUCP(X, E 1 ) x BUCP(X, E 2 ) -+ BUCP(X, E);
(ii) CP(X, E 1 ) x CP(X, E2) -+ CP(X, E).

Proof Observe that, given a nonempty subset Y of X,

(1.1.5)

and, if p > 0,

(1.1.6)

for Uj E BUCP(Y, E j ). Now the assertion is obvious .•


ILl Linear Cauchy Problems 43

Observe that (1.1.5) and (1.1.6) imply that the map (i) has norm at most
one. In particular, if El = E2 = E = lK, it follows that

BUCP(X) := BUCP(X, lK)

is a Banach algebra (with unit 1).

1.2 Existence and Regularity Theorems


Let J C lR be a perfect interval containing its left endpoint s := min J. For each
t E J let A(t) be a linear operator in E, and let f: J --> E. Bya solution of the
linear evolution equation

it + A(t)u = f(t) , t E J\ {s} , (1.2.1)

where it := OtU, we mean a function u E C 1 (J\ {s}, E) such that u(t) E dom(A(t))
and it(t) + A(t)u(t) = f(t) for t E J\{s}. If, in addition, u E C(J, E) and u(s) = x
then u is a solution of the linear Cauchy problem (in E)

it+A(t)u=f(t) , tEJ\{s}, u(s) =x. (1.2.2)(s,x,A,J)

Lastly, u is a strict solution of (1.2.1) if u E C1(J, E) and u(s) E dom(A(s)).


The linear evolution equation (1.2.1) and the corresponding linear Cauchy
problem (1.2.2)(s,x,A,f) are said to be parabolic if -A(t) is the infinitesimal gen-
erator of a strongly continuous analytic semigroup on E for each t E J.

As for the solvability of parabolic linear Cauchy problems, we have the fol-
lowing fundamental existence and regularity result:

1.2.1 Theorem Suppose that (Eo, E 1 ) is a densely injected Banach couple and

(x, A, f) = (x, (A, f)) E Eo x CP(J, H(El' Eo) x Eo)

for some p E (0,1). Then the Cauchy problem (1.2.2)(s,x,A,f) possesses a umque
solution u := u(·, s, x, A, f), and

(1.2.3)

If x E El then 1L is a strict solution.

The existence of a unique solution, which is strict if x E E 1 ) has been proven


by Sobolevskii [Sob66] and, independently, by Tanabe [Tan60]. A different proof
has more recently been given by Acquistapacc and Tcrreni [AcT85]. The reg-
ularity result (1.2.3) is contained in [Sob66] and [AcT85]. A complete proof of
Theorem 1.2.1 is given in Subsections 4 and III.2.6.
44 II Cauchy Problems and Evolution Operators

It is an important property of parabolic evolution equations that they 'regu-


larize'. In fact, (1.2.3) shows that the solution is more regular than its initial value
in general. The following 'automatic regularity theorem' implies that the solution
of (1.2.2)(s,x,A,f) has even better regularity properties if f is more regular.

1.2.2 Theorem Suppose that (Eo, El) is a densely injected Banach couple. Given
e E (0,1),
let (., ·)0 be an admissible interpolation functor and put

Eo := (Eo, Ede .

Suppose that 0 < "/ < p < 1 and 0 < E < 1 - ,,/, and that

and
f E C"(J, E"{) + C(J, E"{+E) .
Then the Cauchy problem (1.2.2)(s,x,A,f) has a unique solution u, and

It is also a solution of the parabolic equation

iJ + A"{(t)v = f(t) , tE J\ {s} , (1.2.4)

in E,,{, where A"{ is the E"{-realization of A. If x E E"{ then u E C(J, E,,{), and if
x E dom(k, (s)) , it is a strict solution of (1.2.4).

This theorem is due to the author (cf. [Ama88c, Theorem 8.2]). Its proof is
given in Subsection IV.1.5.

It should be noted that the domains, dom(A"{(t)) , of the linear operators A"{(t)
depend on t E J, in general, whereas the operators A(t), t E J, have 'constant
domains'. Also observe that the assertion that u is a solution of (1.2.4) implies
u(t) E dom(A"{(t)) for t E J\{s}. This means that u(t) has, in general, better
'spacial' regularity. This will be clear from concrete applications of Theorem 1.2.2
where it is often possible to characterize dom(A"{(t)) explicitly.
As we shall see, parabolic evolution equations with 'constant domains' are
particularly well-suited for studying quasilinear problems. This is not so for prob-
lems with variable domains. In later sections show how - by extrapolation tech-
niques - we often can associate with a given evolution equation with 'nonconstant
domains' a generalized problem possessing 'constant domains'. Then we can apply
the general theory for evolution equations with 'constant domains', developed be-
low, to the generalized problem. Afterwards Theorem 1.2.2 can be used to carry
out an abstract 'bootstrapping' argument to prove that the generalized solutions
are, in fact, already solutions of the original problem.
II.2 Parabolic Evolution Operators 45

2 Parabolic Evolution Operators


Let J be a perfect subinterval of lEt Then

h := { (t, s) E J x J ; sst}
and we put
J!:,. := { (t, s) E h ; s < t} .
If 'P is an operator-valued function on J!:,. and t/J is such a function on J, we write

t/J'P(t, s) := t/J(t)'P(t, s), t.pt/J(t, s) := 'P(t, s)t/J(s)


if these compositions are meaningful and no confusion seems likely.

2.1 Basic Properties


Suppose that F '----+ E and that

A: J -+ C(E) with dom(A(t») C F, t E J . (2.1.1)

Then a map U: J!:!. -+ £(E) is said to be a parabolic evolution operator (prop-


agator, fundamental solution) for A with regularity subspace F if

U E C(h,£s(E») nC(J!:,.,£(E,F») (2.1.2)

and

U(t, t) = 1 , U(t,s) = U(t,T)U(T,S), SSTSt, (t,s)Eh, (2.1.3)

if
im(U(t, s») C dom(A(t») C F , (t,s)EJ!:,., (2.1.4)
and
AU E C(J!:,., £(E»)
with
sup (t - s) IIAU(t,s)II.c(E) < 00 , Icc J, (2.1.5)
(t,s)EI;"

if
U(·,s) E C1(Jn(s,00),£(E») , s E J , (2.1.6)
and
81U = -AU, (2.1. 7)
and if
U(t,·) E c 1 (Jn(-00,t),£s(F,E») , t E J , (2.1.8)
and
82 U:) UA . (2.1.9)
46 II Cauchy Problems and Evolution Operators

2.1.1 Example: Analytic Semigroups Suppose that (Eo,E1 ) is a densely


injected Banach couple and A E 'J-l(E1' Eo). Put J := jR+ and

(t,s) E h.
Then UA is a parabolic evolution operator for the constant map J -+ C(E), t ~ A
with regularity subspace E 1 . •

2.1.2 Remarks (a) Suppose that f E c(Jn [s,oo),E) and u is a solution of


the Cauchy problem

it + A(t)u = f(t) , t E In (s,oo) , u(s) = x. (2.1.10)(s,x,A,f)

Also suppose that V E C( h, £s(E)) satisfies V(t, t) = 1 and

V(t,.) E c 1 (Jn (s,t),£s(F,E))


with 82 V -:J VA for t E In (s,oo). Then

u(t) = V(t, s)u(s) + it V(t, 7)f(7) d7 , t E J n [s, (0) . (2.1.11)

Proof It is easily verified that, given t E J n (s, (0),

V(t,·)u E C 1 ((s,t),E) nC([s,tJ,E)

and that 82 (Vu)(t, 7) = Vf(t,7) for s < 7 < t. Hence

V(t,t')u(t') - V(t,s')u(s') = L t'


V(t,r)f(r)dr, s < s' < t' < t ,

and the assertion follows by letting t' -+ t and s' -+ 8 . •

(b) Supppose that there exists a parabolic evolution operator U for A. Then, given
(s, x) E J x E and f E C( J n [8,(0), E), the Cauchy problem (2.1.10)(s,x,A,f) has
at most one solution. It is given by the variation of constants formula (2.1.11)
with V:= U.
Proof This is an immediate consequence of (a) .•
(c) There exists at most one parabolic evolution operator for A. More precisely: if
Uj is a parabolic evolution operator for A with regularity subspace Fj for j = 1,2
then U1 = U2 =: U and F1 n F2 is a regularity subspace for U.
Proof Given (s,x) E J x E, it follows from (2.1.2), (2.1.3), (2.1.6), and (2.1.7)
that Uj(·,s)x is a solution of (2.1.1O)(s,x,A,0)' Now the assertion is an easy conse-
quence of (b). •
II.2 Parabolic Evolution Operators 47

(d) Let UA be a parabolic evolution operator for A with regularity subspace F.


Given a E lK, put

(t,s) E 1.6..
Then Ua+A is a parabolic evolution operator for a + A with regularity subspace F.
(e) Suppose there exists a (necessarily unique) parabolic evolution operator UA
for A. Given
(x, f) E E x L1,loc(J, E) ,
where J c lR.+ with 0 E J, put

u(t,x,A,f):= UA(t,O)x+ !at UA(t,T)!(T)dT, t E J.

Then u(·, x,!, A) is a mild solution of the linear Cauchy problem (2.1.1O)(o,x,A,f)'
Observe that
u(·, x, A, f) E C(J, E) .
Theorems 1.2.1 and 1.2.2 contain sufficient conditions for! to guarantee that the
mild solution is in fact a solution. Note that, thanks to (d),

u(·, x, A, f) = eatu(·, x, a + A, e- at f) , aElR.,

where eatv := [t f-+ ecttv(t)] .•

2.2 Determining Integral Equations


Let (Eo, E 1 ) be a densely injected Banach couple and suppose that J is a perfect
subinterval of lR.+ containing O. Also suppose that

(2.2.1)

for some p E (0,1). Lastly, suppose that UA is a parabolic evolution operator for A
with regularity subspace E 1 . Then it follows from (2.1.2), (2.1.6)-(2.1.9), and
Example 2.1.1 that

and that

for (t, s) E J't:" and s < T < t. By integrating we find that UA solves the integral
equation

UA(t,S) = e-(t-s)A(s) - i t UA(t,T) [A(T) - A(s)]e-(r-s)A(s) dr (2.2.2)


48 II Cauchy Problems and Evolution Operators

for (t, s) E h. Similarly,

in Ls(Eo) for ° <:.:: s< T < t with t E J, and, consequently,

UA(t, s) = e-(t-s)A(t) + it e-(t-T)A(t) [A(t) - A(T)] UA(T, s) dT (2.2.3)

in Ls(Eo) for (t, s) E h.


The integral equations (2.2.2) and (2.2.3) are the starting point for the proof
of Theorems 1.2.1 and 1.2.2. Namely, we shall show that these integral equations
are uniquely solvable, that their solutions coincide, and that the so-defined function
is a parabolic evolution operator for A. Afterwards, we discuss the solvability of
the Cauchy problem (2.1.lO)(o,x,A,f)'

3 Linear Volterra Integral Equations


Motivated by the considerations at the end of the preceding section, we prove some
simple facts about linear Volterra integral equations in spaces of bounded linear
operators.

3.1 Weakly Singular Kernels


Suppose that J is a perfect subinterval of JR+ containing °and put
Jr:= In [O,T] , T E JR+ .

Given a E JR, denote by


5t(E, F, a)
the Frechet space of all k E C( J~, L(E, F)) satisfying

Ilkll(a),T := Ilkll(a),T,.C(E,F):= sup (t - s)'" Ilk(t,s)IIC(E,F) < 00 ,


O<C;s<t<C;T

equipped with the topology induced by the seminorms { 11·II(a),T ; T E j}. More-
over, 5t(E, a) := 5t(E, E, a). Observe that

a>{3, TEj, (3.1.1)

so that
5t(E, F, (3) '---> 5t(E, F, a) , a> {3 . (3.1.2)
11.3 Linear Volterra Integral Equations 49

We put
Ilkll(a):= sup (t - s)a Ilk(t, s)II.c(E F)
(t,s)EJ;;' ,

and denote by
Jt oo (E, F, 0:)
the Banach space consisting of all k E Jt(E, F, 0:) satisfying Ilkll(a) < 00, equipped
with the norm 1I·II(a)' Observe that
Jt oo (E, F, 0:) '---t Jt( E, F, 0:) (3.1.3)
and
Jtoo(E,F,O) = BC(J'6,£(E,F)) . (3.1.4)
If 0: < 0, each k E Jt(E, F, 0:) can be continuously extended over J!:;. by putting
k(t, t) = 0 for t E J so that
Jt(E, F, 0:) '---t C( Jt;., £(E, F)) , 0: < 0 . (3.1.5)

If E = lK, we identify £(lK, F) naturally with F via

£(lK, F) :3 B +----+ B·1 E F .

Then k E Jt(lK, F, 0:) iff k E C(JX, F) and

sup (t - s)'" Ilk(t, s)IIF < 00 ,


Os,s<ts,T

In particular,
(3.1.6)
by the identification

CU,F):3 u +----+ [(t,s) 1-+ u(t)] E C(JX,F). (3.1.7)

Let G be a Banach space. Suppose that k E Jt(E, F, 0:) and hE Jt(F, G, fJ)
with 0:, fJ E (-00,1), and put

h * k(t, s) := it h(t, T)k(T, s) dT , (t, s) E Jt;. .

It is easily verified that

(3.1.8)
and

(3.1.9)

where B is the Euler beta function. It is a consequence of Fubini's theorem that


the operation * is associative.
50 II Cauchy Problems and Evolution Operators

3.2 Resolvent Kernels


Suppose that k E j{(E, Q) for some Q E [0,1). By an easy induction argument we
see that

I k * k * ... * k(t, s)IIc(E) :s:


~
[f(1 - Q) Ilkll(a),T
f(n(I-Q))
r (t _ s)n(l-a)-l (3.2.1)
n

for n E Nand 0 :s: s < t :s: T. Put


00

w:=L~' (3.2.2)
j=l j

Then we have the following:

3.2.1 Lemma w E j{(E, Q) and, given E > 0,


(t - s)<> Ilw( t, oS) I C(E) :s: c( Q, E )me(HE)m 1
/(1-a) (t-s)

for 0 :s: oS < t :s: T and T E j, where m := f(l- Q) Ilkll(<»,T'


Proof Let fJ := 1 - Q E (0,1]. Thanks to (3.2.1) it suffices to prove
00 xj-1 1/(3
~ - - < c(8 E)e(1+c)x x >0. (3.2.3)
~ r(fJj) - ., ,

Stirling's formula implies the existence of 8(t) E (0,1) such that


f(t) = V21rt t - 1/ 2e- t +8(t)/(12t) , t > O.
From this we deduce for j EN that
f(j + 1)11 [jf(j)]11 (3-1 Ji.. 1/2j(1+/1)/2
f(fJj) = r(fJj) :s: (27r) 2 e 12 fJ fJl1j

Hence, by Holder's inequality,


co xj 00 x j r(j + 1)11
~ f(fJj) = ~ (j!)11 f(fJj)

< [~ (7)X1/11).i] (3 [~j(1+(3)/(2(1-(3))] 1-(3


- c(/'J) L j! L (7)fJ).i(3/(l-(3)
J=l J=l
where 7) > 0 is arbitrary. Since the last series converges for 7) > 1/ fJ,
~
00

L....,
x J-
.

f(fJ .)
1

:s: c(fJ,7))
(e'lX - 1) :s: c(fJ,7))e
1/(3

1/(3
(3
(3
'Ix
1/(3

j=l J 7)X
for x > 0 and 7) > 1/ fJ. Now the assertion follows by setting 7) : = (1 + E) / fJ .•
I1.3 Linear Volterra Integral Equations 51

Now it is easy to prove the following existence and uniqueness theorem for
abstract linear Volterra equations.

3.2.2 Theorem Suppose that a, (3 E [0,1) and k E .ft(E, a). Then the linear
Volterra equations
u=a+u*k, v=b+k*v (3.2.4)
possess for each a E .ft(E, F, (3) and b E .ft(F, E, (3) unique solutions

u E .ft(E, F, (3) and v E .ft(F, E, (3) ,

respectively. They are given by

(3.2.5)

respectively, where w, the resolvent kernel of (3.2.4), belongs to .ft(E, a) and is


given by (3.2.2).

Proof We consider the first equation in (3.2.4). The second one can be treated
analogously.
Define W by (3.2.2) and u by (3.2.5), resp., and observe that W E .ft(E, a)
and u E .ft(E, F, (3) by Lemma 3.2.1 and by (3.1.2) and (3.1.8), respectively. It is
obvious that u solves (3.2.4).
Let T E j be fixed. By replacing J by Jr, it follows from (3.1.1), (3.1.2),
(3.1.8), (3.1.9), and (3.2.1) that * k E L(.ftoo(E, F, (3)) and that the spectral radius
of this operator equals zero. Hence (3.2.4) has at most one solution 'on JT' for
each T E f This proves the assertion. _

3.2.3 Remark In the definition of .ft(E, F, a) we can replace the assumption


that k E C(JX,L(E,F)) by k E Lao,loc(JX,L(E,F)). Then everything remains
true provided:
(i) sUPO:S;s<t:S;T is replaced by ess-suPO:S;s<t:S;T everywhere.
(ii) (3.1.5) is replaced by

.ft(E,F,a) n C(JX,L(E,F)) '-+ C(Jc.,L(E,F)) , a < O.

(iii) C(JX,F) is replaced by Lao,loc(JX,F) in the interpretation of .ft(IK,F,a).


(iv) BC is replaced by Lao in (3.1.4) and (3.1.6).
Observe that with this new definition of .ft(E, F, a), and by using obvious notation,

.ft(F, G, (3) * .ft(E, F, a) '-+ .ft(E, G, a + (3 - 1) n C( Jc., L(E, G))

if a + (3 < 1.
These generalizations are occasionally useful - in Theorem 3.3.1, for example. _
52 II Cauchy Problems and Evolution Operators

3.3 Singular Gronwall Inequalities


As a simple application of Theorem 3.2.2 we prove the following generalized Gron-
wall inequality.

3.3.1 Theorem Given a, (3 E [0,1) and E > 0, there exists a positive constant
c := c(a, (3, E) such that the following is true:
If u: J --+ IR satisfies

[t f-+ t,6u(t)] E Loo,loc(J,IR) (3.3.1 )

1t
and
u(t) ::; AC,6 +B (t - T)-aU(T) dT , a.a. t E j , (3.3.2)

where A and B are positive constants, then

a.a. t E j , (3.3.3)

where f.t(a, B) := (r(1- a)B) 1/(1-0:).

Proof Let E := F := IR and put k(t, s) := B(t - s)-a for (t, s) E Ji:,.. Then we
see that k E ~(E, a) and Ilkll(o:),T = B for T E f Let a(t) := AC,6 and observe
that (3.1.7) implies a E ~(E, (3) and lIall(,6),T = A for T E f Since U E ~(E, (3) by
(3.3.1) and Remark 3.2.3, it follows from (3.1.8) and (3.1.2) that

b:= a + k *U - U E ~(E, (3) .

Hence U = a - b + k * u, and Theorem 3.2.2 implies that

U = (a - b) + w * (a - b) .

Observe that b 2: 0 by (3.3.2) and that k 2: 0 implies w 2: O. Thus u ::; a + w * a,


1t
that is,
U(t) ::; AC,6 +A w(t - T)T-,6 dT , a.a. t E j .

By Lemma 3.2.1,

t>O, E>O.

Since
1t eV(t-Tl(t - T)-O:T-,6 dT::; evt 1t (t - T)-"T-,6 dT

= B(l - u, 1 - (3)t 1 - a -,6e vt


for t > 0 and v 2: 0, the assertion follows .•
11.4 Existence of Evolution Operators 53

3.3.2 Corollary Let (3.3.1) and (3.3.2) be satisfied. Then, given c: > 0, there
exists a constant c := c(c:, Ct, (J, B) such that

a.a. t E j .

3.3.3 Remarks (a) It should be noted that, in general, the constant c( Ct, c:) in
the estimate of Lemma 3.2.1- and, consequently, the constants c in Theorem 3.3.1
and Corollary 3.3.2 as well - go to infinity if c: -> O. Of course, if Ct = 0 then c: = 0
is possible and c(O, 0) = 1. In this case the constant c of Theorem 3.3.1 equals
1/(1 - (J) and (3.3.3) is then a consequence of the classical Gronwall inequality
(e.g., [Ama90b, Corollary (6.2)]).
(b) Of course, the factor e(Hc)/L(t-s) in Lemma 3.2.1 and in Theorem 3.3.1
and its corollary - where J.t:= m1/(1-a) in Lemma 3.2.1 - can be replaced
by e(/L+c)(t-s) . •

It is well-known that the generalized Gronwall inequality is a very use-


ful tool in the theory of semilinear parabolic evolution equations. It has been
proven - in a form somewhat less precise than the one of Theorem 3.3.1 - by
the author in [Ama78, Lemma 2.3] and, independently, by means of Laplace trans-
form techniques, by Henry [Hen81, Lemma 7.1.1]. The trick, used in the proof of
Lemma 3.2.1 for estimating the majorant of the series (3.2.2) by means of Stirling's
formula, is taken from [vW85, p. 6].

4 Existence of Evolution Operators


In this section we carry out the construction of a parabolic evolution operator,
given the classical Sobolevskii-Tanabe assumptions. It should be observed that we
pay particular attention to clarifying the dependence of our estimates on a few
explicitly specified parameters.

4.1 A Class of Parameter Integrals


In order to prove some useful uniform estimates for the semigroups generated by
the elements of 'H(E l , Eo) we use the following technical lemma. Recall that

En := En \ {O} = [I arg zl ~ 19 + 1T /2]

for 19 E [0, 1T /2].


54 II Cauchy Problems and Evolution Operators

4.1.1 Lemma Given {J E (0, 7f 12), let r be an arbitrary piece-wise smooth simple
curve in til running from ooe- i (iI+1I"/2) to ooei (iI+1I"/2), and let X be a metric space.
Suppose that the map f: til X X X lR.+ -+ E has the following properties:
(i) f(', x, t) : til -+ E is holomorphic for (x, t) E X X lR.+.
(ii) f(z,·,·) E C(X X lR.+, E) for z E til.
(iii) There are constants a E lR. and M > ° such that
.
(z, x, t) E ~iI X
'+ •
X x lR.

Then
[(x,t) I-> hf(z,x,t)dZ] EC(XxlR.+,E)

and
II h f(z,x, t) dzll ~ cMc a , (x, t) E X X lR.+ .

Proof Let (x, t) E X x (0,00) be fixed and put r = r; + St + rt, where


rt := {re±i(ii+1I"/2) ; t- I ~ r < 00 }
and

Then

1f(z,x,t)dz=
r
(1- + Js, + Jr,
r,
f fJf(z,x,t)dz

= 1= {f(re- i (iI+1I"/2) , x, t)e- iil + f(re i (iI+1I"/2) , x, t)ei 1?}i dr

+ J iI+1I"/2

-il-1I"/2
f(t-Iei'P,x, t)iCIei'P dr.p .

By substituting s := tr sin {J in the first integral after the second equality sign, it
transforms into

1 00

sin iI
{(-iSe-
f .
iil
t sm {J
,x,t)e-'iI+f
. (iSe iil
- - , x , t)e'iI
t sin {J
. } -ds ,-.
t sm {J
Since Re( ±i se±iilIt sin {J) = -sit, it follows from (iii) that

Now the assertion is an easy consequence of the theorem on the continuity of


parameter integrals and of Cauchy's theorem .•
II.4 Existence of Evolution Operators 55

It should be remarked that estimates of the type given in this lemma are
standard in the theory of analytic semigroups (e.g., [Tan79]).

4.2 Semigroup Estimates


Now let (Eo, E 1) be a densely injected Banach couple, J a perfect subinterval
of ~+ containing 0, and p E (0,1). Suppose that

A c CP(J, 'H.(E1' Eo)) and there are constants M, 'rf E ~+ and


rJ E (0, 7r /2) such that

[Ajp,J ::; 'rf , AEA,

that ~1'1 C p(-A(s)), and that (4.2.1)

IIA(05) IIL(E,,Eo) + (1 + 1>.1)1-j II (>. + A(05)) -lIIL(Eo,E j ) ::; M

for (s, >., A) E J X ~1'1 X A and j = 0, 1.

Observe that (4.2.1) implies

(05, X, A) E J X El X A . (4.2.2)

Assumption (4.2.1) also implies uniform estimates for the semigroups gener-
ated by -A(s), as well as their continuous dependence on s E J, as is seen from
the following result:

4.2.1 Lemma The estimate

II [tA(s)]ke-tA(s)IIL(E +t II [tA(s)] ke-tA(s) IIL(Eo,E


j
) 1) ::; c(k) (4.2.3)

is valid for kEN, (t, 05, A) E ~+ x J x A, and j = 0, 1. Moreover,

[(t,s) f---+ e-tA(s)] E C(~+ x J,.cs(Ek,Ej )) nC(lR+ x J,.c(Ej,E k ))


for j, k E {O, 1} with j ::; k.

Proof It is well-known from the theory of analytic semigroups that

Ak(05)e-tA(s) = (-It
27rZ it
r >.ket'\(>. + A(s)r 1 d>' , t>o, kEN,

r being a piece-wise smooth curve running in t1'l from ooe- i (1'I+71-j2) to ooe i (1'I+7r/2).
From the resolvent estimates of (4.2.1) and from Lemma 4.1.1 we deduce that

(t,s,A) E lR+ x J x A, kEN. (4.2.4)


56 II Cauchy Problems and Evolution Operators

Now (4.2.3) is an easy consequence of (4.2.2) and the fact that a semigroup
commutes with its generator. Since A E CP(J,£(E l , Eo)) and the inversion map
B f--> B- 1 is smooth (in fact: analytic), it follows that

j,kE{O,l}, j~k,

where we use again (4.2.2) and the fact that a closed linear operator commutes
with its resolvent. Now Lemma 4.1.1 implies
[(t,s) f--> e-tA(s)] E C(lR+ x J,£(Ej,E k )) (4.2.5)
for j, k E {a, I} with j ~ k.
Given x E E l , we deduce from

e-tA(s)x - x = fat 8Te- TA (S)xdr = -fat e-TA(S) A(s)xdr

and from (4.2.2) and (4.2.3) that


Ile-tA(s)x - xllo ~ ct IIxill , t E lR+, s E J .

Moreover,

e-tA(s) _ e-tA(s') = ~
27TZ
r etA [(>. + A(s)) (>. + A(s')r l ] d>'
Jr
-1 -

= ~ r etA (>, + A(s)r l [A(s') - A(s)](>' + A(s,))-1 d>' ,


27TZ Jr
(4.2.1), (4.2.2), and Lemma 4.1.1 imply
IIe-tA(s) - e-tA(s') II£(E 1 ,Eo ) ~ ct Is - s'I P , t E lR+, s, s' E J .
Thus we see from
e-tA(s)x _ e-t'A(s')x
(4.2.6)
= (e-tA(s) _ e-tA(s'»)x + (e-tA(s')x - x) _ (e-t'A(s')x - x)

and (4.2.5) that


(4.2.7)
Now we obtain
[(t,s) f--> e-tA(s)] E C(lR+ x J,£s(Eo)) (4.2.8)
from (4.2.7), from (4.2.3) with j = k = 0, and from the density of El in Eo. Lastly,
note that (4.2.2) and (4.2.8) imply
IIe-tA(s)x - xlh ~ M IIe-tA(s) A(s)x - A(s)xllo -> °
°
for t -> and x E E l . Thus we infer from (4.2.5) and (4.2.6) the validity of (4.2.8)
with Eo replaced by E l .•
II.4 Existence of Evolution Operators 57

4.3 Construction of Evolution Operators


We put
aA(t, s) := e-(t-s)A(s) , kA(t, s) := -[A(t) - A(s)]aA(t, s)

for (t, s) E 1'1.. and A E A. Then we deduce from (4.2.2) and Lemma 4.2.1 that
(4.3.1)
and that
AEA, (4.3.2)
where 11·11 is the norm of the space appearing in (4.3.1). Note that (4.2.2) and
Lemma 4.2.1 imply
kA E ~(Eo, 1- p) n ~(El' Eo, -p) (4.3.3)
and
(t - S)l- j - p IlkA(t, s)IIL(Ej,E o) :::; c (4.3.4)
for (t, s) E J'A, A E A, and j = 0,1.
Lastly, we put
=
WA:= LkA*···*kA (4.3.5)
n=l'-...--'"
n
and
(4.3.6)
In the following computations we often omit the index A if no confusion seems
likely.

4.3.1 Lemma The function WA is well-defined and satisfies


WA E ~(Eo, 1 - p) n ~(El' Eo, -p)
with
(t - s)l-j-p IIWA(t, S)IIL(Ej,Eo ) :::; c(c)e(p+e)(t-s) (4.3.7)j
for c > 0, (t, s) E J'A, A E A, and j = 0, 1.

Proof It follows from (4.3.4), Lemma 3.2.1, and Remark 3.3.3(b) that
WA E ~(Eo, 1 - p)
and that (4.3.7)0 is true. It is an obvious consequence of (4.3.5) that
w=k+k*w=k+w*k. (4.3.8)
Thus we deduce from (4.3.4), (4.3.7)0, (3.1.1), and (3.1.8) that
w E ~(El,Eo, -p) (4.3.9)
and that (4.3. 7h is valid .•
58 II Cauchy Problems and Evolution Operators

Note that (4.3.1), (4.3.3), and Theorem 3.2.2 imply that the Volterra integral
equation
(4.3.10)
possesses a unique solution

(4.3.11)

and that it is given by


(4.3.12)
Observe that the integral equation (4.3.10) coincides with (2.2.2). Thus UA
of (4.3.12) is our candidate for the parabolic evolution operator for A. In the
remainder of this section we show that UA has the desired properties. In addition,
we derive important uniform estimates.

4.3.2 Lemma Put

eA(t, s) := A(t)e~(t~s)A(t) _ A(s)e~(t~s)A(s) , (t,s) E J~ .

Then eA E Jt(Eo, 1 - p) and

(t - s)l~p IleA(t, s)II.c(Eo) ::; c , (t, s) E J~, A EA.

Proof It is a consequence of (4.2.1) that

II(A+A(t))~l_ (A+A(s))~lll.c(Eo)

::; I (A + A(t)) ~lll.c(Eo) IIA(t) - A(s) IIC(E] ,Eo) II (A + A( s) r111.c(Eo,Ed


::; c(t - sy IAI~l

for A E 2::1'!, 0::; S < t ::; T, and A E A. From this and from Lemma 4.1.1 we
deduce that
IIA(t)e~(t~s)A(t) - A(s)e~(t~s)A(s) 11.c(Eo)

= II~
27rZ
1Ae'>"(t~s)
r
[(A + A(t)r 1 - (,\ + A(s)r 1 ] dA11
.c(Eo)
::; c(t - sy~l

for (t, s) E J~ and A E A. Now the assertion follows from Lemma 4.2.1. •

The next technical lemma gives estimates for WA(t, s) - WA(T, s) as a func-
tion of t, T, s with s < T < t. Here and in the following, it is always understood
that (t,s) E J~.
11.4 Existence of Evolution Operators 59

4.3.3 Lemma Suppose that °< f3 < p. Then

IIWA(t, s) - WA(T, s)llc(Ej,Eo)


(4.3.13)
:::; c(c){ 81,j(t - T)P + (t - T)f3(T - s)}+p-/3-1 }e(Jl+c)(t-s)

jorc>O, S<T<t, AEA, andjE{O,l}.


Proof First we derive a bound for

k(t,s)-k(T,S) = [A(T)-A(t)]a(t,s)- [A(T)-A(s)][a(t,s)-a(T,s)] , (4.3.14)

where °: :; s < T < t. From (4.3.2) it follows that


(t - S)l- j II [A(t) - A(T)] a(t, s)llc(Ej,Eo) :::; c(t - T)P (4.3.15)

for j E {a, I}. Since

a(t,s) -a(T,s) = -it A(s)e-(0"-s)A(8)dO" ,

we deduce from Lemma 4.2.1 that

Ila(t, s) - a(T, s)llc(Ej,E1 ) :::; M it (0" - S)j-2 dO" :::; M(t - T)(T - S)j-2

for j E {a, I}. Hence

II [A(T) - A(s)] [a(t, s) - a(T, s)] Ilc(Ej,Eo) :::; c(t - T)(T - S)p+j-2 (4.3.16)

for j E {a, I}. On the other hand,

II [A(T) - A(s)] [a(t, s) - a(T, s)] Ilc(Ej,Eo)

:::; C(T - s)p{lla(t, s)IIC(Ej,E1 ) + Ila(T, s)llc(Ej,El)} (4.3.17)

:::; c((t - s)j-1(T - s)P + (T - s)P-Hj) :::; C(T - s)p-Hj .

By combining (4.3.16) and (4.3.17) we see that

II [A(T) - A(s)] [a(t, s) - a(T, s)] Ilc(Ej,Eo)


:::; c(t - T)P(T - S)p(p-2+j)(T - s)(p-Hj)(l- p) (4.3.18)
= c(t - T)P(T - S)j-1 .

Thus we obtain from (4.3.14), (4.3.15), and (4.3.18) the estimate

(T - s)l- j Ilk(t, s) - k(T, s)llc(Ej,Eo) :::; c(t - T)P , j = 0,1. (4.3.19)


60 II Cauchy Problems and Evolution Operators

On the other hand,

Ilk(t, s) - k(T, S)II.c(EO ) ::; C[(t - S)p-I + (T - Sy-I] ::; C(T - sy-I (4.3.20)

by (4.3.4). Hence it follows from (4.3.19) and (4.3.20) that

Ilk(t, s) - k(T, S)IIL:(Eo) ::; c[(t - T)P(T - s)-l]/1/P(T - s)(p-I)(l-/1/p)


(4.3.21)
= c(t - T)/1(T - sy-/1-1 .

Observe that (4.3.8) implies

w(t, s) - W(T, s) = k(t, s) - k(T, s)

+ 17 [k(t,O") - k(T,O")]W(O",S)dO" (4.3.22)

+ it k(t,O")w(O",s)dO" .

Thus it is an easy consequence of (4.3.21), (4.3.4), (3.1.9), and Lemma 4.3.1 that

Ilw(t, s) - W(T, s)IIL:(Eo)


::; c(c) { (t - T)/1 (T - S)p-/1-1 + [1 7
(t - T)/1 (T - 0" y-f1- 1 (0" - Sy-l dO"

+ it (t - O"y-l(O" - S)p-I dO"]e(JL+£)(t-s)}

::; c(c) [( t - T)/1 (T - S)P-/1-1 (1 + (t - sy) + (t - T)P (T - Sy-I ] e(JL+£)(t-s)

for c > O. From this we conclude the validity of the assertion for j = O. (Straightfor-
ward, we obtain in the last inequality the exponent (/-l + 2c)(t - s), for example.
Since this is true for every c > 0, we can replace 2c by c by changing the con-
stant c(c), of course. This fact is frequently used in the following without further
mention.)
Similarly, thanks to (4.3.19) and (4.3.22),

Ilw(t,s) - W(T, s)IIL:(E1,Eo)


::; c(c){ (t - T)P + (t - T)/1(T - s)2 p-/1 + (t - T)P(t - s)P}e(JL+£)(t-s)
::; c(c){ (t - T)P + (t - T)/1(T - s)P-/1}e(JL+£)(t-s)

for c > 0, which proves the assertion for j = 1. •

I
Put
t -£
d£(t,s):= s a(t, T)W(T,S) dT , O<c<t-s,
II.4 Existence of Evolution Operators 61

and dc:(t, s) := 0 for s ::; t ::; s + E, and note that

as E ---+ 0. (4.3.23)

Moreover,

[hde:(t, s) = e-c:A(t-c)w(t - E, s) _I t -e:A (r)a(t, r)w(r, s) dr (4.3.24)

for t > s + E in .c(Eo). Since A(t)e-(t-r)A(t) = aTe-(t-r)A(t) , we see that


Ehde:(t, s) = e-e:A(t-e:)w(t - E, s) - [e-e:A(t) - e-(t-s)A(t)]w(t, s)

+ i t ce
- (t,r)w(r,s)dr (4.3.25)

_It-cA(t)e-(t-T)A(t) [w(r, s) - w(t, s)] dr

for t > s + E. Put


dA(t,s):= e-(t-s)A(t)w(t,s)

+ it e(t,r)w(r,s)dr (4.3.26)

+ it A(t)e-(t-r)A(t) [w(t, s) - w(r, s)] dr .

It is an easy consequence of Lemma 4.2.1 and of Lemmas 4.3.1 to 4.3.3, letting


j3 := p/2 in Lemma 4.3.3, that

(4.3.27)

and
(t - s)I-j-p IldA(t, s)llc(Ej,Eo ) ::; c(E)e(I-'+c:)(t-s) (4.3.28)
for E > 0, (t, s) E J'6., A E A, and j = 0, 1, and that

aIde: ---+ dA in C(J'6.,L.s(Eo)) as E ---+ 0 . (4.3.29)

Thus we deduce from (4.3.23) that

aA*wA(',S) ECI(Jn(s,oo),.cs(Eo)), s E J ,

and
(4.3.30)
Consequently,

aA * w(t, s) - aA * w(t', s) = t dA(r, s) dr ,


Jt'
s < t' < t , (4.3.31)
62 II Cauchy Problems and Evolution Operators

in £8(Eo). By employing (4.3.27) and (4.3.28) we infer that (4.3.31) even holds
in £(Eo), which, in turn, implies
s E J. (4.3.32)

Now it is not difficult to prove the following technical assertions:

4.3.4 Lemma Given A E A,


UA(·,S) = aA(·,s) +aA *WA(·,S) E c1(Jn (s, 00), Eo) , s E J,

and
(4.3.33)
Moreover,
(4.3.34)
and
(t - s)k- j - p IlaA *wA(t,s)II.c(Ej,Ek) ~ c(c)e(Jl+c)(t-s) (4.3.35)
jorc>O, (t,S)EJ'A, AEA, andj,kE{0,1}.
Proof Put

A , s) .=
UC(t . e-(t-s)A(s) +d (t s)
E:, , o<c<t-s.
Then it is easily verified that
8 1 Ul(t, s) + A(t)Ul(t, s)
= -kA(t, s) + e-cA(t-c)WA(t - c, s) _I t c
- kA(t, T)WA(T, s) dT

°
for < c < t - s. Letting c
c10sedness of A(t) that
--+ 0, we infer from (4.3.23), (4.3.29), (4.3.30), and the

on J'A. Since WA - kA = kA * WA by (3.2.2), the first assertion is true.


If k = 0, the second assertion is an easy consequence of (4.3.1), (4.3.2), of
Lemma 4.3.1, and of (3.1.8) and (3.1.9). Observe that, thanks to (4.3.33),
A(aA *w) = AUA - AaA = -81 UA - AaA
= -81 aA - AaA - 81 (aA *WA) = kA - 8 1 (aA *WA) .

Hence we deduce from (4.2.2) that

liaA *wA(t,s)II.c(Ej,E 1 ) ~ M(llkA(t,s)II.c(Ej,Eo ) + 1i81 (aA *wA)(t,s)II.c(Ej,Eo))


for (t, s) E JX, A E A, and j E {O, 1}. Now (4.3.34) and (4.3.35) with k = 1 are
consequences of (4.3.3), (4.3.4), (4.3.27), (4.3.28), and (4.3.30) .•
II.4 Existence of Evolution Operators 63

4.4 The Main Result


After these preparations we can prove the first main result of this section, namely
the following existence and regularity result:

4.4.1 Theorem Let assumption (4.2.1) be satisfied. For each A E A there exists
a unique parabolic evolution operator UA possessing E1 as a regularity subspace
and satisfying

(4.4.1)

There exists a constant c(p) > 0, which is independent of T), such that, putting

(4.4.2)

the estimate

(4.4.3)

is valid for E > 0, (t,S)EJt:" AEA, andj=O,1.

Proof Define UA by (4.3.12). Then it follows from Lemmas 4.2.1 and 4.3.4 and
from (3.1.5) that UA satisfies (2.1.2), (2.1.4)-(2.1.7) with E:= Eo and F:= E l ,
as well as (4.4.1) and (4.4.3). Of course,

t E J . (4.4.4)

Now let A E A be fixed, suppose that A E C 1 (J, £(El' Eo)), and that J is
compact. Then
(4.4.5)

and O[A + A(·)r 1 = -(A + A)~lOA(A + A)~1 for A E ~19. Thus (4.2.1) implies

(A,s)E~19XJ. (4.4.6)

Now it is an easy consequence of Lemma 4.1.1 that

and
02a(t, s) = A( s )e~(t-s)A(s) + m(t, s) , (4.4.7)
where
m(t,s):= ~
21fz
r e'\(t~s)os(A+A(s)rldA.
J['
Observe that (4.4.5), (4.4.6), and Lemma 4.1.1 imply
64 II Cauchy Problems and Evolution Operators

Hence we infer from Theorem 3.2.2 the existence of a unique solution

v E Jt(Eo, 0) (4.4.8)

of the integral equation

Put
V:= a+ v*a
and note that, thanks to Lemma 4.2.1, (4.4.8), and (3.1.5),

(4.4.9)

and
V(t, t) = 1 , t E J. (4.4.10)
Given t E j, it is easily seen that

and that
82 V = 8 2 a - v + (v * a)A + v * m
(4.4.11 )
= (a+v*a)A+m+v*m-v = VA
in Ls(El, Eo). Observe that U(" s)x is for each (s, x) E J x Eo a solution of the
Cauchy problem

u+A(t)u=O, tEJn(s,oo), u(s) = x . (4.4.12)(s,x)

Hence we obtain from (4.4.9)-(4.4.11) and Remark 2.1.2(a) that V = U and that
(4.4.12)(s,x) has a unique solution. This shows that (2.1.8) and (2.1.9) are satis-
fied. Now (2.1.3) follows easily from the unique solvability of (4.4.12)(s.x) for each
(s, x) E J x Eo. This proves the theorem, provided each A E A is continuously
differentiable.
To remove this latter assumption we can again assume that J is compact.
Then we extend an arbitrarily fixed A E A trivially, that is, by zero, over lR and put

A" := ('P" * A)IJ ,


where {'Pc ; c > O} is a mollifier on lR (cf. Section III.4.2, where we recall the
basic facts about convolutions). Then A" E c= (J, L(El' Eo)) and

IIA"IIC(J,C(E1,Eo)) ::; IIAllc(J,C(E"Eo)) , c > 0, (4.4.13)

by Young's inequality. Moreover,


11.4 Existence of Evolution Operators 65

implies
[Ag]p,J:::; 'TJ, c: > 0 . (4.4.14)
Since, as is well-known and easily seen,

as c: ---- 0 , (4.4.15)

it follows from (4.4.14) and Theorem 1.1.3.1(i) that there exists c:o > 0 such that

0< c: < c:o .


Note that the spectral bound s( -A) of -A is negative. It is an easy consequence of
(4.4.15) and of Corollary 1.1.4.3 that there exist M' > 0, {}' E (0, 7r/2) , and c:' > 0
such that the family
A' := {A" ; 0 < c: < c:' }
satisfies assumption (4.2.1) (with M and {} being replaced by M' and {}', respec-
tively.)
Thus, by what has already been shown, we know that there exists a unique
parabolic evolution operator Uc for A g , 0 < c: < c:'. Given (t, s) E JX,

aT [Uc(t, r)U(r, s)] = Uc(t, r)[Ae(r) - A(r)] U(r, s) .


Upon integrating this identity,

U(t, s) - Ue(t, S) = it Uc(t, r) [Ae(r) - A(r)] U(r, s) dr.

From this we deduce that

O<c:<c:'.

Now (4.4.15), the fact that Uc is uniformly bounded in C(Eo) for 0 < c: < c:', and
the density of El in Eo imply

as c: ---- 0 . (4.4.16)

Since

Ug(t, s)x - Ue(t, s')x = 1


s'
8
aTUc(t, r)x dr = 1 8

Sf
Ue(t, r)Ae(r)x dr

for s' < s < t, we obtain from (4.4.15) and (4.4.16) that

U(t,s) - U(t,s') = 1
8'
8
U(t,r)A(r)dr, 0:::; s' < s < t,

in C8 (E 1 , Eo). From this it follows that U satisfies (2.1.8) and (2.1.9). Conse-
quently, (2.1.3) is again a consequence of Remark 2.1.2(a). _
66 II Cauchy Problems and Evolution Operators

4.4.2 Corollary Suppose that

for some p E (0,1). Then there exists a unique parabolic evolution operator UA
for A possessing E1 as a regularity subspace.

Proof First let J be compact. Then Theorem 1.1.2.2 and Corollary 1.1.4.3 imply
the existence of (J E lR such that A := {(J + A} satisfies assumption (4.2.1). Hence
the assertion follows in this case from Theorem 4.4.1 and Remark 2.1.2(d). If J is
not compact, we obtain the assertion by applying this result to every compact
subinterval of J .•

4.5 Solvability of the Cauchy Problem


Now it is easy to give a proof of the first part of the basic existence and regularity
result of this section:

Proof of Theorem 1.2.1: Part 1 In this part of the proof we show that the
Cauchy problem (1.2.2)(s,x,A,f) has a unique solution that is strict if x E E 1 . The
proof of the asserted Holder regularity is postponed to Section 111.2.6.
First suppose that J C lR+ with 0 E J, that s = 0, and that A := {A} satisfies
condition (4.2.1). Thanks to Remarks 2.1.2(a) and (b) it suffices to show that the
mild solution
u:= (Ux+U*f)(-,O)
is a solution of (1.2.2)(o,x,A,f), where U:= UA. Since U(',O)x is a solution of
(1.2.2)(o,x,A,O) by (2.1.2), (2.1.6) and (2.1.7), in order to prove that u is a solution
it remains to show that U * f := U * f(-, 0) is a solution of (1.2.2)(o,O,A,f)'
Recall that U = a + a*w. It is an easy consequence of (4.3.28), (4.3.30), and
the continuity of a * w on J b. that

Ot 1t (a * w) (t, T) f (T) dT = 1t 01 (a * w)( t, T) f (T) dT , (4.5.1)

which shows, thanks to (4.3.28) and (4.3.30), that

(a * w) * f(', 0) E C 1 (J, Eo) . (4.5.2)

Put
Vs(t) := 1t-ce-(t-T)A(T) f(T) dT , E>O, tEJ, t?E,

with vE(t) := 0 for 0 ~ t < E and note that

vs(t) ---+ v(t) := a * f(t, 0) in Eo as E---+O. (4.5.3)


11.5 Stability Estimates 67

Also observe that

and that

It-EA(T)e-(t-T)A(T) f(T) dT

= - I t - Ee (t,T)f(T)dT+ I t - cA (t)e-(t-T)A(t)f(T)dT.

Since the last integral can be rewritten as

It-cA(t)e-(t-T)A(t) (1(T) - f(t)) dT - [e-cA(t) - e-tA(t)]f(t) ,

we deduce from the Holder continuity of f and from Lemmas 4.2.1 and 4.3.2 that
there exists W E C(J, Eo) such that

avc ---7 W in C(J, Eo) as c: ---7 0.

Thus, thanks to (4.5.3), we see that v E Cl(J, Eo) which, together with (4.5.2),
implies
U * f(', 0) E Cl(J, Eo) . (4.5.4)
Note that the closed ness of A(t) and (2.1.7) imply

at I t c
- U(t, T)f(T) dT + A(t) I t c
- U(t, T)f(T) dT = U(t, t - c:)f(t - c:) .

From this, the closedness of A(t), and the strong continuity of U on Jt::,. we infer,
by letting c: ---7 0, that U * f(·, 0) is a solution of (1.2.2)(o,O,A,f)'
Suppose that x EEl. Then from (4.4.1), that is, from U E C( Jt::,., £s(E l )), we
infer U(·,O)x E C(J,Ed. Thus AU(-,O)x E C(J, Eo) since A E C(J,£(El,Eo)).
Now we see from
a[u(·,o)x] = -AU(·,O)x
that U(·,O)x is a strict solution of (1.2.2)(o,x,A,O)' By combining this with (4.5.4)
we see that u is a strict solution of (1.2.2)(o,x,A,f)' This proves the part under
consideration of the theorem in this case. Now the general case follows by an
obvious translation argument and by Remark 2.1.2(d) .•

The proof of Theorem 4.4.1 and the proof of part 1 of Theorem 1.2.1 follow
essentially Tanabe [Tan60] and [Tan79]. It should be remarked, however, that the
regularity statement (4.4.1) and the uniform estimates (4.4.3), which will be crucial
in the following, are contained in these references implicitly at most.
68 II Cauchy Problems and Evolution Operators

5 Stability Estimates
Let (Eo, Ed be a densely injected Banach couple, let J be a perfect subinterval
of lR+ containing 0, and let p E (0,1). Throughout this section we assume that

A c CP( J, £(El' Eo)) and there exist constants 1] :::: 0,


w > 0, r;,:::: 1, and (7 E lR such that

[AJp,J ~ 1] , AEA, (5.0.1 )


and

(7 + A c C(J, 1t(ElJ Eo, r;"w)) .

In the following, we prove some continuity and decay estimates for parabolic evo-
lution equations and for mild solutions of linear Cauchy problems. These results
are of basic importance for our study of quasilinear problems.

5.1 Estimates for Evolution Operators


As an easy consequence of the results of Sections 1.1 and 4 we obtain the following
fundamental estimates:

5.1.1 Theorem There exists a constant co(p) > 0, which is independent of 1],
such that, letting
1/ := CO(p)1]l/p + +w
(7 , (5.1.1)
the following is true: for each A E A there exists a unique parabolic evolution
operator UA for A possessing El as a regularity subspace, and

for (t,8) E J~, A E A, and j = 0, 1.

°
Proof It follows from (5.0.1), Proposition 1.1.4.1, and Corollary 1.1.4.3 that there
w
exist < w, {} E (0, 7r /2), and M > such that

and

for (8, A, A) E J X ~19 X A and j = 0, 1. Thus Theorem 4.4.1 implies the existence
of a constant co(p) > 0, being independent of 1], such that for each A E A there
II.5 Stability Estimates 69

exists a unique parabolic evolution operator Ua +w+A for CT + W+ A possessing E1


as a regularity subspace and satisfying

(5.1.2)

such that

for (t, s) E J'l:,., A E A, and j = 0,1, where I-l:= Co (p)1]l/p . Since, by Remark 2.1.2(d),

UA(t , s) .=
. - A(t , s)
e(o+w)(t-s)Ua+w+ , (t, s) E h , (5.1.3)

is a parabolic evolution operator for A with regularity subspace E 1 , the assertion


follows. _

5.1.2 Remark Suppose that A E 1i(E1 , Eo) and II> s( -A). Then

Ile-tAllc(Ej ) + t Ile-tAIIc(Eo,E 1 ) :::; ce vt (5.1.4)

for t > 0 and j = 0, 1. Hence

s( -A) = type( -A) , (5.1.5)

that is, the spectral bound and the exponential type of a generator of a strongly
continuous analytic semigroup coincide.

Proof Fix CT E (s( -A), II) and put w := II - CT. Then Proposition 1.1.4.2 implies
the existence of K 2: 1 such that CT + A E 1t(E 1 , Eo, K, w). Hence A := {A} sat-
isfies (5.0.1) with 1] = o. Consequently, (5.1.4) follows from Theorem 5.1.1 and
Example 2.1.1. Now it is obvious that type(-A) :::; s(-A). On the other hand, the
Hille-Yosida theorem implies the converse estimate. _

Throughout the remainder of Section 5 we denote by UA the parabolic funda-


mental solution for A E A, and II is the number defined in (5.1.1). Now we study
the behavior of UA in the interpolation spaces Eo., 0 < 0: < 1.

5.1.3 Lemma Suppose that 0:::; (L :::; fJ :::; 0: :::; 1. Also assume that fJ- < fJ iff
o < fJ < 0: < 1 and there does not exist an interpolation functor J of exponent fJ / 0:
such that J (Eo, Eo.) ~ E(3. Then

UA E C(h,.cs(E",)) nC(J'l:,.,.c(E(3,E",))

and
IIUA(t, s)llc(E",) + (t - s)o;-(3- IIUA(t, s)ll c (E(3,E",) :::; cev(t-s)

for (t, s) E J'l:,., and A E A.


70 II Cauchy Problems and Evolution Operators

Proof If 00,/3 E {O, I}, this follows from Theorem 5.1.1 and from (5.1.2) and
(5.1.3). From this we easily obtain the assertion by interpolation if either /3 = 0
and 0:::; a :::; 1, or 0 :::; /3 :::; 1 and 00= 1, or if /3 = a. (For the proof of the strong
continuity one uses in an essential way the fact that E1 is dense in Eo}
Suppose that 0 < a < 1. By interpolating the morphisms

by means of an interpolation functor ~ of exponent ~ it follows that

and
(t - S)(l-I;)" IIUA(t, s)11 ( ) :::; cel/(t-s)
£ J(Eo,Ea),Ea
for (t, s) E J6.., A E A, and 0 < ~ < 1. If thcre does exist an interpolation func-
tor ~ of exponent /3/00 such that ~(Eo, EaJ ~ E f3 , the assertion follows by letting
~ := /3/00. Otherwise we put ~ := (-, ')E with ~ := /3-/00 and obtain the desired
result from the almost" reiteration property of Remark I.2.11.2(a). _

The following lemma gives continuity estimates for UA as a function of A EA.


5.1.4 Lemma Suppose that 0 :::; /3 < 1 and 0 < a :::; 1. Then

(t - s)f3-O: II(UA - UB)(t, s)II£(Ea,E/3)


:::; cev(t-s) max IIA(r) - B(r)II£(E"Eo)
sSrSt

for (t, s) E J6..! and A, BE A.

Proof By replacing in the argument leading to the integral equation (2.2.2) the
function e-(r-s)A(s) by UB(r, s), it follows that

(t, s) E h .

Now we deduce from Lemma 5.1.3 that

II(UA - UB)(t, s)II£(Ea,E/3) :::; cev(t-s) it (t - r)-f3(r - S)O:-l dr IliA - Bill

= cB(a, 1 - (3) IliA - Bill (t - s)",-f3ev(t-s) ,


where
IliA - Bill := max II(A - B)(r)llqE"Eo)
sSrSt
and B is the Euler beta function. _
II.5 Stability Estimates 71

5.2 Continuity Properties of Mild Solutions


Given
(x, A, f) E Eo x A x Ll,loc(J, Eo) ,
we consider the linear Cauchy problem

u + A(t)u = f(t) , t E j, u(o) = x . (5.2.1)(x,A,f)

Recall from Remark 2.1.2(e) that the mild solution of (5.2.1)(x,A,f) is given by

u(·, x, A, f) := (UAX + UA * f)(·, 0) . (5.2.2)

In the following, we prove some stability estimates that are of importance


for the study of quasilinear problems. The next theorem implies, in particular,
that the mild solutions depend continuously upon all data, provided f is locally
bounded.

5.2.1 Theorem Suppose that 0 ::; f3 ::; n ::; 1 with n > 0 and f3 < 1 and 0 < 'Y ::; 1.
Then

llu(t, xo, Ao,jo) - u(t, Xl, AI, h) 11!3

::; c{ t"-fJ IIAo - AtllC([O,tj,,C(E,,Eo )) [llxolla + t l - n +1' IlfoIIL=((o,t),K,)]


+ llxo - xIII!3 + t l -!3llfo - h IIL=((o,t),Eo ) }e vt

fort E J and (xj,Aj,fj) E En X A x Loo,loc(J,EI')'


Proof Put Uj := u(·,xj,Aj,iJ) and U:i := UA j and observe that

Now the assertion is an easy consequence of Lemmas 5.1.3 and 5.1.4 and the
continuous embedding of En in E!3' •

5.2.2 Remark Suppose that 0 ::; f3 ::; n ::; 1 with n > 0 and f3 < 1 and 0 < 'Y ::; 1.
Let X ~ , ~ E {n, f3, 'Y}, be Banach spaces satisfying

and
EI '----+ XI' '----+ EI' ,
respectively. Observe that the X-spaces are not supposed to be interpolation spaces
between Eo and E I . Also suppose that A E A.
72 II Cauchy Problems and Evolution Operators

It follows from Lemma 5.1.3 and the factorization

X~ '---+ Eo UA
--+ El '---+
X~

that
~ E {a,,8,,} .
Similarly,

and Lemma 5.1.2 imply

and
(t - s) f3 IIUA(t,s)IIL:(x.."x/3) :S cev(t-s)
for (t, s) E J6.. Hence we deduce from the representation (5.2.2) that
u(·,x,A,f) E C(j,Xf3)

for x E Xf3 and f E Loo,loc(J,Xf3)'


Suppose, in addition, that
IIUA(t, s)llqx/3) :S cev(t-s) , (t, s) E J6., A EA.
Then it is obvious from the proof of Theorem 5.2.1 and factorization arguments
of the type used that
Ilu(t, xo, Ao, fo) - u(t, xl, A l , fdllx/3
:S c{ t"- f3 I1 A o - A11Ic([0,t],L:(E 1 ,Eo)) [ Ilxo Ilx" + tl-",+'Y lifo IILoo((O,t),x..,)]
+ Ilxo - xIIiX/3 + t l - f3 llfo - it IILoo((o,t),Eo) }evt
for t E j and (Xj, A j , fj) E X", x A x Loo,loc(J,X'Y)' This simple generalization of
Theorem 5.2.1 is useful in some applications. _

5.3 HOlder Estimates


Our next theorem provides uniform Holder estimates for mild solutions.

5.3.1 Theorem Suppose that 0 :S ,8 :S a < 1.


If x E E", and f E Loo,loc(J, Eo) then
u:= u(·,x,A,f) E C"'-f3(J,Ef3 ) . (5.3.1 )
More precisely,

Ilu(t) - u(s)lIf3 :S c(t - s)"'-f3evt (llxll", + 1IfIILoo((O,t),Eo)) (5.3.2)


for (t,s) E Jt:,. and (x,A,f) E E", x A x Loo,loc(J,Eo).
U.5 Stability Estimates 73

Proof Observe that

(5.3.3)

Hence it follows from the proof of Theorem 5.1.1 and from Lemma 5.1.3 that there
exists E > 0 such that

(5.3.4)

for (t, s) E J!:,. and A E A. Using the first half of this estimate, Lemma 5.1.3 with
f3 = 0, and the strong continuity of UA, the assertion is obvious if a = o. Hence we
can assume from now on that a > o.
Put v(t, s) := UA(t, s)x for (t, s) E J2,. and x E Ec" and observe that

v(-, s) E C( J n [s, 00), E",) n C 1(J n (s, 00), Eo) (5.3.5)

with
01V(t, s) = -A(t)UA(t, s)x .
Hence it follows from (5.3.4) that

v(t, s) - v(r, s) = - i t (T - s)l-a A(T)UA(T, S)X(T - S)",-l dT

and
Ilv(t, s) - v(r, s)llo ::; c(t - r)"'e(v-e)(t-s) Ilxli a (5.3.6)
for s::; r::; t and (t,s) E h. Note that (5.3.4) also implies

Ilv(t, s) - v(r, s)ll", ::; Ilv(t, s)ll" + Ilv(r, 8)11", ::; ce(l/-e)(t-s) Ilxll a (5.3.7)

Thus we deduce from (5.3.6), (5.3.7), the interpolation inequality (1.2.11.2), and
the definition of v that

(5.3.8)

for s::; r::; t, (t,8) E h, and A E A.


Put w := UA * fe, 0) and observe that

for (t,s) E J2,.. Thanks to Lemma 5.1.3 the first integral is estimated in the norm
of E{3 by
74 II Cauchy Problems and Evolution Operators

Since (t - s)l-;3 ::; t 1 -"'(t - s)",-;3, we see that the first integral in (5.3.9) is esti-
mated in the norm of E;3 by

c(t - s)",-;3e vt IlfIlLoo«o,t),E o) (5.3.10)

From Lemma 5.1.3 and (5.3.8) we deduce that

IIUA(t, T) - UA(S, T)II£(Eo,EIl)


::; IIUA(t, s) - UA(S, s)llc(En,EIl ) IIU(s, T)llc(Eo,E a )
::; c(t - 05)",-;3(05 _ T)-"'e(V-E)(t-r)

for T ::; 05 ::; t. This estimate shows that the second integral in (5.3.9) is bounded
in the E;3-norm by
(5.3.11)
for (t,s) E Jb., A E A, and f E Loo,loc(J, Eo). Now estimate (5.3.2) is a conse-
quence of (5.3.8)-(5.3.11). Since it is easily verified that W E C(J,E;3) (if a = (3),
it follows from (5.3.5) that U E C(J, E(3). This proves the theorem .•

5.4 Boundedness of Mild Solutions


Lastly, we prove a useful boundedness result that again illustrates the smooth-
ing effects of parabolic evolution equations. For this we assume without loss of
generality that v =F o.

5.4.1 Theorem Suppose that 0 ::; f3- ::; f3 ::; a ::; 1. Also assume that f3- < f3
iff 0 < f3 < a < 1 and there does not exist an interpolation functor ~ of exponent
f3/ a such that ~(Eo, En) ~ E;3. Also suppose that 'Y = 0 if a < 1 and 0 < 'Y < 1 if
a = 1. Letting
(Irlt
b"'-7(t, r) := Irl"-7- 1 io C-"'e'~sign(r) d~

for t E lR+ and r E lR, it follows that

Ilu(t, x, A, f) II", ::; c{ t;3- -"'evt Ilxll;3 + b"'-7( t, v) IlfIILoo«o,t),E-,)}

fort E J and (x,A,f) E E;3 x A x L oo ,loc(J,E7 ).

Proof This is an easy consequence of Lemma 5.1.3 .•

5.4.2 Corollary Let the assumptions of Theorem 5.4.1 be satisfied and suppose
v < o. Then

Ilu(t,x,A,f)II",::; c{t;3--"'e vt Il xll;3 + IlfIILoo«o,t),E-,)} (5.4.1)

for t E J and (x,A,f) E E;3 x A x L oo ,loc(J,E7 ).


11.6 1nvariance and Positivity 75

6 Invariance and Positivity


It is the main purpose of this section to show that the evolution operator U := UA
of A E CP ( J, 'H (E 1 , Eo)) enjoys essentially the same invariance properties as each
one of the semigroups {e-tA(s) ; t ~ O} for s E J. In particular, if Eo is an ordered
Banach space, we derive a useful criterion for A guaranteeing the positivity of UA.

6.1 Yosida Approximations


Suppose that
AEC(E) and (w,oo)cp(-A) (6.1.1)
for some w ~ O. Then
AE := A(l + SA)-l E £(E) (6.1.2)
for 0 < s < l/w, and
{A€ ; 0 < s < 1/w }
is the Yosida approximation of A.

6.1.1 Lemma Suppose that

A>W. (6.1.3)

Then
lim(l
£--->0
+ SA)-lX = X, x E dom(A) , (6.1.4)

and
lim A€x = Ax, x E dom(A) . (6.1.5)
£--->0

Proof ;\l"ote that (6.1.3) implies

0< s < l/w . (6.1.6)

Hence it follows from

x E dom(A) )

that
11(1 + SA)-lx - xii.,; EK IIAxl1 ) x E dom(A) .
Thus (1 + EA)-lX --> x for x E dom(A) as s --> O. Now (6.1.4) is an easy conse-
quence of (6.1.6). Since Acx = (1 + SA)-l Ax for x E dom(A), the last assertion
follows from (6.1.4) .•
76 II Cauchy Problems and Evolution Operators

Let (Eo, E 1 ) be a densely injected Banach couple and suppose that

A E H(El,Eo,""w) (6.1.7)

for some '" ~ 1 and W > O. Then (6.1.1) is satisfied (with E := Eo). Thus Ac E £( Eo)
is well-defined for 0 < E < l/w. Since

11(1 + EA)-lll.c(Eo,E, ) :::; "'/E , 0< E < l/w , (6.1.8)

by Remark 1.1.2.1(a), it follows that

IIA,lIc(Eo) :::; E- 1 ", II A ll c (E" Eo) , 0< E < l/w. (6.1.9)

Hence
Ac E H(Eo, Eo, 2,1 + 2E- 1 ", IIAllc(E"Eo))
by Remark 1.1.2.1(b). The following lemma implies, however, that A(A + Ac)-l
exists in [Rez ~ w/(I- EW)] and can be estimated independently of E.

6.1.2 Lemma Suppose that A E C(E) and there exist "', wE jR+ such that
[ReA ~ w] C p( -A) and

ReA ~ w . (6.1.10)

Then
[Rez ~ w/(I-Ew)] C p(-Ac)
and
IIA(A+A c )-lll:::; "'+ 1, ReA ~ w/(1 - EW) , (6.1.11)
for 0 < E < l/w.
Proof Note that

for A E <C. It is easily verified that

ReA ~ w/(I- EW) implies Re(A/(1 + EA)) ~ W .

+Ar
Hence
1
(A+Ac)-l = 1:c>.(I+EA)C:EA
(6.1.12)
_ _1_ [E + _1_ (_A_ + A) -1]
- 1 + c>. 1 + EA 1 + EA
for ReA ~ w/(l- EW) and 0 < E < l/w. Consequently, thanks to (6.1.10),

II( A+ A c )-111 < 1 E IAI + '"


- p:f 11 + EAI '
ReA ~ w/(1 - EW) . (6.1.13)
11.6 Invariance and Positivity 77

If c IAI ~ 1 and Re A ~ 0 then

c IAI + K < c IAI + K < 1 (6.1.14)


11+cAI - clAI - K+ .
Now the assertion follows since 11 + cAl ~ 1 for Re A ~ O. _

6.1.3 Corollary Suppose that A E C(E), [Rez ~ 0) c p(-A), and there exists
K > 0 such that
ReA ~ 0 .
Then there are 1):= 1)(K) E (0,7r/2) and c:= C(K) > 0 such that ~~ c p(-As) and

forc > O.

Proof This is an easy consequence of the first part of the proof of Proposi-
tion 1.1.4.1 by observing that for the analogue of (1.1.4.2) (where the term con-
taining Ilxlh is omitted) estimate (6.1.11) suffices. _

6.2 Approximations of Evolution Operators


Let (Eo, E 1 ) be a densely injected Banach couple, let J be a perfect subinter-
val of 1R+ containing 0, and suppose that p E (0,1), K ~ 1, and w > O. Also
suppose that
(6.2.1)
Then the Yosida approximations {Ae(s) E £(Eo) ; 0 < c < l/w } are well-defined
for s E J. Since the maps

1t(El' Eo, K,W) ~ £(Eo, El), B f-+ (1 + cB)-l


are analytic for 0 < c < l/w, it follows from (6.2.1) that Ae E CP (J, £(Eo») for
0< c < l/w. Hence, thanks to Remark 1.1.2.1(b) and Corollary 4.4.2, there ex-
ists a unique parabolic evolution operator UA. for As possessing Eo as a regu-
larity subspace. It is the purpose of the following considerations to show that,
given (t, s) E JA.,

as c ~ O.
Let
B:=w+A
and note that (6.1.2) implies

[Rez ~ 0) c p(-B(s») , s E J, (6.2.2)


78 II Cauchy Problems and Evolution Operators

and

Re>'::::O, sEJ. (6.2.3)

Hence the Yosida approximations

{BE(S) := (B(s)t ; I'; > O}


are well-defined for s E J. The following lemma gives uniform estimates for the
semigroups generated by -BE(s).

6.2.1 Lemma There exists c > 0, depending on K and w only, such that

jort>O, sEJ, 1';>0, andkE{O,l}.

Proof Corollary 6.1.3 and (6.2.3) imply the existence of {) E (0, 7r /2) and of a
constant c with

>. E ~!9, S E J, I'; > °.


Since
[BE(s)te-tBe(s) = (-2 1t
7rZ
r>.ket'>'(>'+BE(s))-ld>',
lr
r being a piece-wise smooth curve running in t!9 from ooe- i (!9+rr/2) to ooei (!9+rr/2),
the assertion follows from Lemma 4.1.1. •

By (6.2.1) and Remark I.1.2.1(a),

s E J.

Hence ° E P(BE(S)) and

sEJ, 1';>0,

so that [BE(S) - BE(t)]B;l(S) is well-defined for s, t E J and I'; > 0.

6.2.2 Lemma Given s, t E J,

I'; > 0.
II.6 Invariance and Positivity 79

Proof Observe that EBe = 1 - (1 + EB)-l for E > O. Hence


Be(s) - Be(t) = E- 1 [(1 + EB(t))-l - (1 + EB(s)r 1]
= (1 + EB(t))-\B(s) - B(t))(l + EB(s)r 1 .

Consequently,

[Be(s) - Be(t)]B;l(s) = (1 + EB(t)) -1 [A(s) - A(t)]B-1(S) .


Since (6.2.3) and Remark I.1.2.1(a) imply

11(1 +EB(t))-lll£(Eo ) V IIB- 1 (t)II£(Eo ,E 1 ):::; K,

the assertion follows .•

After these preparations we can prove the following uniform estimate for the
evolution operator UBE .

6.2.3 Lemma Suppose that [Ajp,J :::; TJ. Then there exist positive constants
c:= C(K,W,TJ) and J.l:= J.l(K,W,7},p) such that
IIUBE(t,s)IIc(E o):::; ceJl(t-s) , (t, s) E h, E> 0 .

Proof Using the notations of Section 4.3, it follows from Lemma 6.2.1 that there
exists Co:= CO(K,W) such that aBE E Jt(Eo, 0) and
(t,s)Eh, E>O. (6.2.4)

Hence, putting J.lo := (f(p)COTJ)l/ P, we infer from

kBE(t, s) = - [Be(t) - B e(s)]B;l(s)B e(s)e-(t-s)B e(S)


and from Lemmas 6.2.1 and 6.2.3 that kBE E Jt(Eo, 1 - p) and
(t, s) E Ji::,., E >0 .

Thus (cf. the proof of Lemma 4.3.1) WB E E Jt(Eo, 1 - p) and


(t - s)l- p IlwBJt, s)II£(Eo ) :::; ce",(t-s) , (t,s) E Ji::,., E > 0, (6.2.5)

where J.l := J.lo + 1. Now we infer from (6.2.4), (6.2.5), (3.1.8), and (3.1.9) that
aBE * WB E E Jt(Eo, -p) and
(t -s)-P lIaB E*wBe(t,s)II£(E o ):::; ce",(t-s) , (t,s) E h, E > O. (6.2.6)

Finally, the assertion follows from (6.2.4) and (6.2.6) since UBE = aBE + aBe * WB E
by the proof of Theorem 4.4.1. •
80 II Cauchy Problems and Evolution Operators

Now we are ready for the proof of the main result of this section, namely the
approximation assertion:

6.2.4 Theorem Given (t, s) E h.,

as E ----> O.

Proof Thanks to Remark 2.1.2(d) it suffices to show that UB,(t,S) ----> UB(t,S)
in .Ls(Eo) as E ----> 0. Moreover, we can assume that (t, s) E J'A and, by replacing J
by [0, t], that J is compact.
By integrating the identity

S<T<t,
it follows that

UB,(t,S)X - UB(t,S)X = -it UB.(t,T)[B",(T) - B(T)]UB(T,S)xdT (6.2.7)

for x E Eo. Note that B",( T) - B(T) = [(1 + EB(T) r l - 1] B(T) and (4.4.3) imply

Hence we deduce from Lemma 6.2.3 that the integrand in (6.2.7) is bounded
in the norm of Eo, uniformly for E E (0, l/w), provided x EEl. Consequently,
given x EEl,
UB,(t,s)x ----> UB(t,S)X in Eo
as E ----> 0, thanks to (6.1.5) and the dominated convergence theorem. Now the
assertion follows from Lemma 6.2.3 and the density of El in Eo .•

It should be remarked that the 'Yosida approximation technique' was first


used by Kato [Kat61] to construct parabolic evolution operators (cf. Subsec-
tion IV.2.3). Moreover, Theorem 6.2.4 can be improved to show that the con-
vergence is locally uniform with respect to (t, s) E J", (cf. [DaK92]).

6.3 Invariance
In this section we are interested in conditions guaranteeing that a closed convex
subset X of Eo is mapped into itself by the evolution operator UA, that is,

UA(t, s)(X) eX,


For this we first prove the following simple technical result:
II.6 Invariance and Positivity 81

6.3.1 Lemma Let Sl be a nonempty set and let JL be a probability measure on


(a a-algebra of subsets oj) Sl. Let X be a nonempty closed convex subset of some
Banach space E. If f E £1 (Sl, E) and f(w) E X for JL-a.a. w E Sl then f dJL E X. Jo
Proof Suppose that e' E E' and a E lR are such that X C [Re(e', x) ::; a]. Then

that is, J0 f dJL E [Re (e' , x) ::; a]. Now the assertion follows since, by the Hahn-
Banach theorem, X is the intersection of all closed real half-spaces [Re(e', x) ::; a]
containing X .•

For the reader's convenience we include the following well-known invariance


theorem from semigroup theory (e.g., [HiP57, Section 11.7]):

6.3.2 Theorem Suppose that A E Q(E,M,a). Then

t ~ 0, (6.3.1)

iff
(1 + SA)-l(X) eX, (6.3.2)

Proof Recall from semigroup theory that

(1 + SA)-lx = 1 CXJ
e-te-cAtxdt , x E E, 0 < s < l/a+ . (6.3.3)

Thus, letting JL := e-tdt on Sl := lR+, Lemma 6.3.1 and (6.3.1) imply (6.3.2). On
the other hand,

t~O, xEE. (6.3.4)

Hence (6.3.2) implies (6.3.1) .•

Throughout the remainder of this section we assume that

(Eo, E 1 ) is a densely injected Banach couple, }


J is a perfect subinterval of lR+ containing zero,
and X is a nonempty closed convex subset of Eo.

The following invariance theorem is a partial extension of Theorem 6.3.2 to the


case of parabolic evolution operators. The main idea of its proof is essentially due
82 II Cauchy Problems and Evolution Operators

to Daners and Koch (cf. [DaK92, Proposition 10.9]). Recall that s( -A) denotes
the spectral bound of A E C( Eo).

6.3.3 Theorem Suppose that A E CP(J, H(El' Eo)) for some p E (0,1) and that

s( -A(t)) :S -(J <0 , t EJ . (6.3.5)

Then
(1 + cA(t)f\X) eX, tEJ , 10 > 0, (6.3.6)
implies
UA(t, s)(X) eX, (t, s) E h . (6.3.7)

Proof Let (tt, s) E Ji::. be fixed and replace J by [0, ttl, again denoted by J.
Then A(J) is compact in H(El' Eo). Thus, thanks to Corollary I.1.3.2, there exist
'" ;::: 1 and Wo > 0 such that A E CP(J, H(E l , Eo, ""wo)).
Put w := (J /2. Then, given t E J, Proposition I. 1.4.2 guarantees the existence
of "'t ;::: 1 and of an open neighborbood At of A(t) in L(El' Eo) such that

-w + At C H(EI' Eo, "'t, w) .


By the compactness of A( J) we can find to, ... ,tn E J such that A( J) c UA tj .
Thus, letting'" := max "'t J , it follows that

A(t) := -w + A(t) E H(EI' Eo, ""w) , t E J .

Theorem 6.2.4 and Remark 2.1.2(d) imply that

UA)t, s) = U(w+Al,,(t, s) --> Uw+A(t, s) = UA(t, s)


in Ls(Eo) as 10 --> o. Thus it suffices to prove that (6.3.6) implies

(t,s)Eh,c>O.
Recall that u(·,x) := UAJ,s)x is, for each s E [O,tl) and x E Eo, the unique so-
lution of the initial value problem

u + A,,;(t)u = 0, s < t :S tl , u(s) = x .

Since AE = C l (1 - (1 + cA)-I), we see that u(·, x) solves

u+C I U=c- l (l+cA(t))-lu, S:St<tl, u(s)=x.


Hence u(·, x) is given by the variation-of-constants formula with respect to the
semigroup {e-t/E1Eo ; t;::: O}, that is,
II.6 Invariance and Positivity 83

Let v(t) := e-(t-S)/Ex for tEI:= [s, tIl and denote by K the Volterra integral
operator in C(I, Eo) with kernel

(t,T)Eh.
Since Ilk(t,T)II£(Eo) ::; Kif for (t,T) E h, it follows from (3.2.1) that K is quasi-
nilpotent. Hence (1 - K)-l exists and is given by the Neumann series. Thus
00

u(·,x) = (1- K)-lv = LKnv .


k=O

From this we deduce, letting

f(w):=v+Kw, wE C(I,Eo) ,
that u(·, x) = lim Uk in C(I, Eo), where

uo:=v, Uk+1=f(uk) , kEN.

Note that
f(w) = ax + (1 - a)Mw ,
where a(t) := a(t, s) := e-(t-S)/E and

Mw(t) := it <p(t, T) (1 + fA( T)) -IW(T) dT


with

Hence, letting /-Lt := <p(t,') dT on n := [s, tl for tEl, it follows from (6.3.6) and
Lemma 6.3.1 that Mw(t) E X for tEl, provided w E C(I,X). Thus

f(C(I,X)) e C(I,X) if xEX,


and, consequently, u(t, x) = lim Uk(t) E X for tEl if x E X .•

A nonempty subset C of a vector space is a cone if jR+C e C. Thus every


cone contains O. A cone is proper if C n (-C) = {O}.

The following theorem shows that we can omit condition (6.3.5) if X is a cone.

6.3.4 Theorem Suppose that A E CP (J, H(El' Eo)) for some p E (0,1). If X is
a cone,
t~O, sEJ, (6.3.8)
implies
UA(t,s)(X) eX, (t, s) E h .
84 II Cauchy Problems and Evolution Operators

Proof Having fixed (tl' s) E Jf)., we can assume that J is compact. Hence, by
invoking Corollary 1.1.3.2, we can assume that A E CP(J, H(El' Eo, fl"w)) for some
fl, ~ 1 and w > 0. Since s( -A) < w we see that

s( -(2w + A(t))) = -2w + s( -A(t)) :::; -w , t E J.

Let B := 2w + A. Since e-tB(s) = e-2wte-tA(s) we infer from (6.3.8) and the fact
°
that X is a cone that e-tB(s) (X) C X for t ~ and s E J. Thus, thanks to
Theorem 6.3.2,

(1 +c:B(t))-l(X) eX, tEJ, c:>o,


and Theorem 6.3.3 guarantees that

UB(t, s)(X) eX, (t,s) E h .


Now the assertion follows from Remark 2.1.2(d) and from 1R+ Xc X .•

Of course, thanks to Theorem 6.3.2, condition (6.3.6) can also be replaced


by the equivalent requirement (6.3.8).

6.4 Orderings and Positivity


Let E be a real vector space and let P be a convex cone in E. Then we define a
reflexive and transitive binary relation in E, a preorder :::;, by putting

iff y-XEP.

It is obvious that this preorder is linear, that is, x :::; y implies x + z :::; y + z and
ax :::; ay for z E E and a E 1R+. Conversely, given a linear preorder :::; in E, we
put P:= {x E E ; x ~ o}. Then P is a convex cone in E, the positive cone of
the preordered vector space E. The preorder is an order, that is, x :::; y and y :::; x
imply x = y iff its positive cone is proper. By an ordered vector space (E,:::;) we
mean a real vector space E together with a linear order :::;. Instead of (E,:::;) we
often write (E, P), where P is the positive proper cone. Of course, x ~ y means
that y :::; x, and we write x < y iff x:::; y and x ~ y.
If E := (E, P) is an ordered vector space which, in addition, is a TVS then
E is said to be an ordered TVS (OTVS) if its positive cone is closed. Of course, if
E is an OTVS such that it is a LCS (resp. Frechet space, resp. Banach space, etc.)
with its topology then it is said to be an ordered locally convex space (OLCS)
(resp. ordered Frechet space, resp. ordered Banach space (OBS), etc.). Note
that IR is an OBS with its natural order induced by the positive cone 1R+.
Let (E, P) and (F, Q) be preordered vector spaces. Then a linear map T
°
from E to F is positive, in symbols, T ~ 0, if T(P) C Q, that is, Tx ~ whenever
1I.6 Invariance and Positivity 85

x::::: 0 (where we usually use the same symbol for the preorder in either of the
spaces, if no confusion seems likely). Now let (E, P) and (F, Q) be OTVSs. Then

£+(E,F):= {T E £(E,F); T::::: O}

is easily seen to be a closed cone in £(E, F) which is said to induce the natu-
ral preorder in £(E, F). It is a proper cone iff P is total, that is, P - P = E
(e.g., [Sch71, Proposition V.5.I]). Of course, £+(E) := £+(E, E).
Let E be an OLCS. Then pi := £+(E, JR) is the dual cone of E'. Thus
(E', Pi) is an OLCS iff P is total.

Now suppose that E is a Banach space, preordered by a closed convex cone,


and A E Q(E). Then the semigroup {e- tA ; t::::: O} is positive if e- tA 2: 0 for
t 2: o. A linear operator A in E is resolvent positive if (w, CXl) C p( - A) for some
w 2: 0 and
(1 + €A)-l 2: 0 , 0< € < I/w .
If A E Q(E), these two concepts are equivalent, as is seen from the next assertion:

6.4.1 Theorem Let E be a Banach space, preordered by a closed convex cone,


and suppose that A E Q(E). Then the semigroup {e- tA ; t::::: O} is positive iff A is
resolvent positive.

Proof This is an immediate consequence of Theorem 6.3.2 .•

Now suppose that J is a perfect interval and A: J ~ C(E). Also suppose


that U is a parabolic evolution operator for A (with some regularity subspace).
Then U is said to be positive if U(t, s) 2: 0 for (t, s) E h.

6.4.2 Theorem Let (Eo, Ed be a densely injected Banach couple, let J be a


perfect subinterval of jR+ containing 0, and suppose that

for some p E (0,1). Also suppose that Eo is preordered by a closed convex cone.
Then the unique parabolic evolution operator UA for A is positive, provided each
one of the semigroups {e-tA(s) ; t 2: O}, s E J, is positive.

Proof This follows from Theorem 6.3.4 .•


Chapter III

Maximal Regularity

In Chapter II we studied the solvability of the Cauchy problem for linear parabolic
evolution equations by taking into consideration, in particular, the smoothing ef-
fects of analytic semigroups. The regularizing properties deduced from these ef-
fects have important implications for the qualitative theory of quasilinear parabolic
equations. However, they have the disadvantage that, in general, the derivative of a
solution is less regular than the right-hand side of the corresponding parabolic evo-
lution equation. This 'loss of regularity' leads to some difficulties in the treatment
of nonlinear evolution equations. For this reason we investigate in this chapter
situations where such a loss of regularity does not occur, that is, cases of 'maximal
regularity' .
In Section 1 we generalize the concept of solutions of linear parabolic evo-
lution equations to admit distributional solutions belonging to suitable Sobolev
spaces. In addition, we introduce the general concept of maximal regularity. The
remaining sections are devoted to the study of concrete cases of maximal regular-
ity. Namely, in Section 2 we show that we can always have maximal regularity if we
require all functions to be Holder continuous with a prescribed singularity of the
HOlder seminorm at zero. If we wish to consider continuous functions only, the class
of admissible Banach spaces has to be restricted considerably - in dependence
of the generator of the analytic semigroup under consideration. This is shown in
Section 3. Finally, in Section 4 we prove that we can have maximal regularity in
the Sobolev space setting provided we consider a restricted class of Banach spaces
and, in addition, restrict the class of generators as well.

1 General Principles
In this section we give a general discussion of the concept of maximal regularity for
linear evolution equations. Concrete realizations are discussed in Sections 2-4. For
88 III Maximal Regularity

the reader's convenience we begin by recalling some known facts about Sobolev
spaces and absolutely continuous functions.

1.1 Sobolev Spaces


Let X be an open subset of]Rn and E a Banach space. Then V(X, E) is the space
of E-valued test functions on X, that is, of all E-valued COO-functions on X with
compact supports. We equip V(X, E) with the usual inductive limit topology, and
V(X) := V(X,lK). Then
V'(X, E) := £(V(X), E)
is the space of E-valued distributions on X, endowed with the topology of uniform
convergence on bounded subsets of V(X). Thus V'(X, E) is a LCS and
V'(X):= V'(X,lK) = V(X)' .
Recall that V(X) and V'(X) are Montel spaces, that is, they are barreled and
bounded subsets are relatively compact. Thus V(X) (consequently, V'(X) as well)
is reflexive.
As usual, given a multiindex 0' E Nn , the distributional derivative aaT of
T E V'(X, E) is defined by
(aaT)(cp) := (_l)la IT(aacp) , cp E V(X) .
It is easily verified that
aa E £(V'(X, E)) .
Given U E L1,loc(X, E), we put

cp E V(X) .

Since Tu : V(X) -t E is obviously linear, and since


K cc X, cp E V(X) , supp(cp) c K ,
where, given any measurable subset Y of X,

11·ll p := 11·l!p,Y,
is the norm in Lp(Y, E), it follows that Tu E V'(X, E). As in the classical case, it
is not difficult to see that the map
L1,loc(X, E) -t V'(X, E), U f-+ Tu (1.1.1)
is linear and injective, Thus we identify L1,loc(X, E) with its image in V'(X, E)
under the map (1.1.1) so that
L1,loc(X, E) <.......> V'(X, E) .
The elements of L1,loc(X, E) are the regular E-valued distributions on X.
IIL1 General Principles 89

Given kEN and p E [1,00]' we define the Sobolev space

to be the subspace of Lp(X, E) consisting of all u for which E),,'u E Lp(X, E) for
lad::; k, and
(L 118a'II~)1/P , l::;p<oo,
11·11 k,p := 11·11 k,p,X := { lal::;k
maxl18a 'lloo ,
lal::;k
p=oo .

Since Lp(X, E) '----+ L1,loc(X, E), this definition is meaningful and it is easily verified
that Wpk(X, E) is a Banach space. Note that ~o = Lp.
We also define the Frechet spaces

~~IOC(X, E) ,

to consist of all u E Lp,loc(X, E) such that cpu E Wpk(X, E) for cp E V(X). Then
{ u t-+ IIcpullk,p ; cp E V(X) } is a generating family of seminorms for the topology
of ~~loc(X, E).
Now let J be a perfect subinterval of R Then we put
k k 0
~ (J,E):= Wp (J,E) ,

Moreover, given kEN andp E [1,00], we write u E Wp~loc(J, E) iff u E ~~loc(J, E)


and 8 j u E Lp,loc(J, E) for 0::; j ::; k. Note that this means that 8 j u E Lp(J, E) for
o
each compact subinterval J of J (and not only of J).
Of course, if E = IK, we omit this symbol everywhere, if no confusion seems
likely, and denote by
(-,.) := C, ')1J
the duality pairing between V'(X) and V(X).

1.2 Absolutely Continuous Functions


Let E be a Banach space. A function u: J ----+ E is locally absolutely continuous
(on J) if there exists v E £l,loc(J, E) such that, given a E J,

u(t) = u(a) + it V(T) dT , t E J, (1.2.1)

(Here and in the following, £p and £p,loc are the vector spaces of all measurable
functions u: J ----+ E such that t t-+ lu(t)IP is integrable, respectively locally inte-
90 III Maximal Regularity

grable, whereas we use Lp and Lp,loc to denote the respective Banach and Frechet
spaces of equivalence classes of these functions.) The function u is absolutely con-
tinuous (on J) ifthere exists v E £1 (J, E) such that (1.2.1) is true. If u is locally
absolutely continuous, it is continuous, possesses a.e. a derivative it, and it = v
a.e., thanks to Lebesgue's differentiation theorem. Moreover, given a, bE J with
a < b and a sequence (Jk ) of pair-wise disjoint open subintervals, Jk := (ak' bk ),
of (a, b), it follows from (1.2.1) that

L
k
IU(bk) - u(ak)1 ~ 1U.h Ivl dT .

Thus Lebesgue's dominated convergence theorem implies that, given c > 0, there
°
exists 15 > such that

L lu(b k ) - u(ak)1 < c (1.2.2)


k

for each pair a, b E J with a < b and each sequence of pair-wise disjoint subintervals
Jk:= (ak,b k ) of (a,b) satisfying

(1.2.3)

If E is reflexive, u: J --+ E is locally absolutely continuous iff (1.2.2), (1.2.3) are


true, thanks to a theorem of Komura [Kom67].

1.2.1 Remark The above definition of absolute continuity differs from the clas-
sical one of the finite-dimensional case which is based on property (1.2.2), (1.2.3).
In general Banach spaces the two definitions are not equivalent (e.g., [Yos65, Sec-
tion V.5]). Since the integral characterization (1.2.1) is the more useful one, we
have decided to take it as definition of an absolutely continuous function. _

In the following, we adopt the usual convention not to distinguish (in no-
tation) between a measurable function J --+ E and its equivalence class modulo
functions vanishing almost everywhere.

1.2.2 Theorem Suppose that 1 ~ p ~ 00. Then

iff u is locally absolutely continuous and

u, it E Lp(J, E) [resp. Lp,loc(J, E)] .

If this is the case, ou = it.

Proof See [Bar76, Section I.2.2], for example. _


IILI General Principles 91

1.3 Generalized Solutions


Let (Eo, E 1 ) be a densely injected Banach couple. Suppose that

A E Loo,loc(J,L(El,Eo)) and (x,f) E Eo x L1,loc(J,Eo) .

Then, given p E [1,00], we say u is a ~~loc-solution (a generalized solution) of


the linear Cauchy problem

u+A(t)u=J(t) , tEJn(s,oo), u(s) = x (1.3.1) (s,x,A,J)

if
u E ~~loc(P, Eo) n Lp,loc(P, E 1 ) (1.3.2)

and
(8+A)u=J, u(s)=x, (1.3.3)

where (Au)(t) := A(t)u(t) for t E JB := J n [s, 00). Note that u(s) E Eo is well-
defined by Theorem 1.2.2. Of course, u is a Vl~}-solution of (1.3.1)(s,x,A,J) if

(1.3.4)

and (1.3.3) is true.

Suppose that there exists a parabolic evolution operator UA for A. The fol-
lowing proposition shows that every ~~loc -solution is then a mild solution of
(1.3.1) (s,x,A,f). Thus, in this case, U;~loc -solutions are unique, if they exist at all.

1.3.1 Proposition Suppose 1 :::; p :::; 00, there exists a parabolic evolution oper-
ator Jor A, and u is a ~~loc-solution oj the Cauchy problem (1.3.1)(s,x,A,f)' Then

u = u(-, s, x, A, f) = (UAX + UA * f)(', s) .

Proof Givcn t E J n (s, 00), put V(T) := UA (t, T)U(T) for s < T < t. Then it is an
easy consequence of (11.2.1.2), (II.2.1.8), (11.2.1.9), and Theorem 1.2.2 that

and that

a.e. T E (s,t) .

Now the assertion follows from (1.2.1) .•


92 III Maximal Regularity

1.4 Trace Spaces


Let J be a perfect subinterval of lR+ containing 0 and let E and lE( J) be Banach
spaces such that
lE(J) '---+ Wl,loc (J, E) .
It follows from Theorem 1.2.2 that the trace map
"/ := "/0 : lE(J) ~ E, U f-> u(O)
is a well-defined continuous linear operator. We put
,,/lE(J) := im("f) ,
equipped with the quotient topology of the Banach space lE( J) I ker( "/). Thus,
,,/lE( J) is a Banach space with the norm

U f-> IluIII'E(J) = inf{ IlvIIE(J) ; v E lE(J), "/V = u} ,


the trace space of lE( J).
Let J I be another subinterval oflR+ with J I :::> J, and let lE(JI ) be a Banach
space such that lE(JI ) '---+ Wl,loc(J1 , E). Then the trace space ,,/lE(h) is also well-
defined. Suppose that lE( J I ) is a regular extension of lE( J), that is, there exist a
continuous retraction r from lE( J I ) onto lE( J) and a corresponding coretraction r C
such that r is naturally induced by the 'point-wise restriction' U f-> uiJ. Observe
that this implies
rUI (0) = UI (0) ,
and
U E lE(J) .
Thus it is reasonable to expect that the trace spaces ,,/lE(J) and ,,/lE(JI ) are the
same since, at least formally, the trace ,,/U depends on the values of U near zero
only. This is made precise in the following:

1.4.1 Proposition If lE(J1 ) is a regular extension of lE(J) then ,,/lE(JI ) ~ ,,/lE(J).


Proof Let x E "/lE(Jd be given. Then there exists UI E lE(Jd with ,,/UI = x. Con-
sequently, rUl E lE(J) and ,,/rUI = x. This shows that ,,/lE(J1 ) C ,,/lE(J) and

Il xIII'E(J) :::; IlrUIIIE(J) :::; IIrlilluIIIE(J,)


Since this is true for every UI E lE(JI ) with ,,/UI = x,

II x IlI'E(J) :::; Il r llllxIlI'E(Jl) ,


that is, "/lE( h) '---+ ,,/lE( J). A similar argument implies

IIxIII'E(J,) :::; Ilrc llll xIII'E(J) ,


that is, "/lE(J) '---+ ,,/lE(J1 ) • •
IILI General Principles 93

We denote by {As ; s ~ O} the semigroup of left translations on EJfI.+,


that is,
Asu(t)=u(s+t) , UEEJfI.+, s,tEJR+.
Then E(JR+) is translation invariant if {As; s ~ O} induces naturally (that is, by
restriction) a uniformly bounded semigroup on E(JR+). The space E(JR+) is contin-
uously translation invariant if {As ; s ~ O} is a strongly continuous uniformly
bounded semigroup on E(JR+).

Using these concepts it is easy to prove the following important embedding


result.

1.4.2 Proposition Let E(JR+) be a translation invariant regular extension


of E(J). Then
E(J) '-t B(J'I'E(J)) . (104.1)
If, in addition, E(JR+) is continuously translation invariant,

E( J) '-t BUe (J, I'E( J)) . (1.4.2)

Proof Note that I'A sU = u(s). Hence

s E JR+ ,

implies E(JR+) '-t B (JR+, I'E(JR+)). If E(JR+) is continuously translation invariant,


we see from

lIu(r) - u(s)ll--YlE(JfI.+) ::::; IlAs(Ar-s -l)ulllE(JfI.+) ::::; c II(A r - s - l)ull lE (JfI.+)

for 0::::; s < r < 00 and u E E(JR+) that E(JR+) '-t BUe(JR+,I'E(JR+)). Observe
that the following diagram
rC
E(J) E(JR+)

I -- r
f
B (JR+, I'E(JR+))

is commutative, where the left vertical arrow represents the natural injection and
where B has to be replaced by BUe if E(JR+) is continuously translation invariant.
Now the assertion follows from Proposition 1.4.1. •

1.4.3 Remark Suppose that E(JR+) is a regular extension of E(J) for each
J C JR+ containing [0, TJ for some T > 0 such that the corresponding retractions
94 III Maximal Regularity

and coretractions are bounded independently of J. Then it is obvious that the


injections (l.4.1) and (l.4.2), respectively, are uniformly bounded with respect
to J. If this is the case, we say lE(I~+) is a regular extension of lE(J) uniformly
with respect to J:J [0, T] .•

1.5 Pairs of Maximal Regularity


Let J be a perfect subinterval of R+ containing 0 and let (Eo, E I ) be a densely
injected Banach couple. Suppose that

is a pair of Banach spaces such that

(l.5.1 )

and
(l.5.2)
respectively. Lastly, let

Then (lEo, lEI) is said to possess the property of maximal regularity with respect
to A, or to be a pair of maximal regularity for A, if

AEA. (l.5.3)

Of course, by a pair of maximal regularity for A we mean one for A:= {A}.
Moreover, each A E £(EI' Eo) is canonically identified with the constant map

again denoted by A.

1.5.1 Lemma Suppose that A c H(EI,Eo). Then the pair (lEO,lEI) has the
property of maximal regularity with respect to A iff

AEA, (l.5.4)

and, given (x, A, J) E ')IlEI X A x lEo, the Cauchy problem

u + Au = f(t) , t EJ , u(O) = x (l.5.5)(x,A,J)

has a wl 10c -solution u with 11 E lEI.


III.1 General Principles 95

Proof Since, trivially, "( E £(lEl' "(lE l ) , we see that (1.5.4) is equivalent to

AEA.

If (lEo, lEd is a pair of maximal regularity for A then, given

(x,A,!) E "(lE l x A x lEo,

it follows from (1.5.2) that u:= (8+ A, ,,()-lU, x) is a W[loc-solutionofthe Cauchy


problem (1.5.5)(x,A,f)' Conversely, the assumptions imply that (8 + A, "() is sur-
jective. Since it is also injective, thanks to Proposition 1.3.1 and Example 11.2.1.1,
the assertion follows from the open mapping theorem. _

Given (x,A,!) E Eo x H(El' Eo) x Ll,loc(J, Eo), put

(1.5.6)

and
(1.5.7)

Note that

(1.5.8)

and that
(1.5.9)

After these preparations we obtain the following characterization of pairs of


maximal regularity.

1.5.2 Theorem Let (Eo, Ed be a densely injected Banach couple, let (lEo, lE l )
be a pair of Banach spaces satisfying (1.5.1) and (1.5.2), respectively, and let A
be a nonempty subset of H(El' Eo). Then (lEo, lE l ) is a pair of maximal regularity
for A iff

and
RAx+KAf
is a W[,loc -solution of (1.5.5)(x,A,f) for (x, A,!) E "(lE l X A x lEo.

Proof This is an immediate consequence of Lemma 1.5.1, Proposition 1.3.1, and


(1.5.9). _
96 III Maximal Regularity

Let E and E be Banach spaces such that E '---t LI,loc( J, E). Given 0: E JR, put

eaE:= {u E L 1 ,loc(J,E) ; e-atu E E} ,


equipped with the norm

where, as usual, e-atu := [t ~ e-<>tu(t)].


The following observation is almost trivial, though useful in applications.

1.5.3 Proposition Let A be a nonempty subset of1i(El, Eo) and let 0: E R Sup-
pose that (Eo, Ed is a pair of maximal regularity for 0: + A. Then ,(e"'E I ) = ,E 1
and (e<>Eo, e<>Ed is a pair of maximal regularity for A. Moreover,

II (8 + A, ,)-III.C(eoEox-yE"eoE,) = II (8 + 0: + A, ,)-II1£(Eox-yE"E,)
for A E A.

Proof It is obvious that the map (u ~ e<>tu) is an isometric isomorphism from EI


onto e<>E 1 . This readily implies ,(e<>E 1 ) = ,E 1 .
Let (1, x) E e"'Eo x ,E I be given. It follows from Remark II.2.1.2(e) and from
Theorem 1.5.2 that
(8 + A, ,)-1(1, x) = u(·, x, A, f) = e<>tu(·, x, 0: + A, e-<>t j)
=e"t(8+0:+A,,)-I(e-<>tf,x) .

Now the assertion is obvious .•

1.6 Stability
For further considerations we take advantage of the following simple lemma that
we give in slightly greater generality than actually needed.

1.6.1 Lemma Let P j be LeSs for j = 0,1,2 and suppose

(1.6.1)

Then
(1.6.2)
and
j = 1,2, j -I- k . (1.6.3)
111.1 General Principles 97

Proof Clearly, (1.6.1) implies (1.6.3). Hence P := [B21 ker(Bdr 1B2 E £(Fo)
and B2 [B21 ker(Bd] -1 = 1F2 . This implies that p 2 = P, that is, P is a continuous
projection onto Fa. Consequently, Fa = im(P) EB ker(P) and, since im(P) = ker(B1)
and ker(P) = ker(B2 ), the assertion follows .•

1.6.2 Corollary Suppose that (lEo, 1E1) is a pair of maximal regularity for
AC1i(E1,Eo). Put

1E1" := kerh), 1E 1,A:= ker(8 + A)


and

for A E A. Then

and

with

Now it is easy to show that the property of being a pair of maximal regularity
for B E 1i(E1, Eo) is stable under small time-dependent perturbations of B.

1.6.3 Proposition Suppose that (lEo, lEd is a pair of maximal regularity for
BE 1i(E1,Eo) and that BE £(lEl,lEo). Then (1E1,IEO) is a pair of maximal regu-
larity for each A E £(1E 1, lEo) satisfying

(1.6.4)

for some A E (0,1). If this is the case,

where c is independent of A, B, and A.

Proof Note that (8 + A, 'Y)u = (I, x) for u E 1E1 and (I, x) E lEo X 'Y1E1 iff

(8 + B)u + (A - B)u = f, 'YU = x . (1.6.5)

Thanks to Corollary 1.6.2 we have the unique decomposition

u = v +W E 1E1" EB 1E 1 ,B .
98 III Maximal Regularity

Hence (1.6.5) is equivalent to

(8 + B)')'v + (A - B)v + (A - B)w = f, rBw = x ,

which, in turn, is equivalent to

[1 + (A - B)KB j(8 + B)')'v = f - (A - B)RBx, W = RBx .

Note that (1.6.4) implies [1 + (A - B)KB ] E £aut(Eo) and

Thus

and

Hence, given (j, x) E Eo X "(El' there exists a unique

u= v +w EEl,')' EB El,B = El

satisfying (8 + A, "()u = (j, x) and

On the other hand, BE £(El' Eo) and (1.5.4) imply 8 E £(El, Eo). From this we
deduce that 8 + A E £(El' Eo) so that (8 + A, "() E .c(El' Eo x "(El)' This proves
the assertion .•

2 Maximal Holder Regularity


In this section we show that maximal regularity occurs in Holder spaces, provided
we allow the HOlder seminorm to become singular at t = 0 at a controlled rate.
We begin by introducing the needed function spaces.

2.1 Singular HOlder Spaces


Let J be a perfect subinterval of lR+ containing 0 and let E be a Banach space.
Given p, E jR+, put

Ilulle" := lIulle,,(J,E) := sUI>(lA t)l-' lIu(t)11


tEJ
III.2 Maximal Holder Regularity 99

for u E EJ. Then

BCI"(J,E) := ({ U E C(j,E) ; Ilulle" < oo}, 1I'lle,,}


is a Banach space. Note that

BCo(J, E) = BC(j, E) .

If Il > 0, put

BUCI"(J,E)
:= { U E C(j, E); [t f--> (1/\ t)I"u(t)] E BUC(j, E), limt~otl" Ilu(t)11 = o} ,
whereas
BUCo(J, E) := BUC(J, E) .
Observe that BUCI"(J, E) is a closed linear subspace of BCI"(J, E) for Il E JR+.
Thus it is a Banach space as well.
Given p E (0,1) and a nonempty subinterval I of JR, let

{llu(s)-u(t)11 }
* * *
[u]p := [u]p,I := [U]ep(I,E) := sup Is _ W ; s, tEl, 0 < t - s :::; 1 .

For Il E JR+ put

[u]p'I" := [u]p,I",J := [u]e~,,, (J,E) := sup (1/\ E )P+I" [u];,[c,2c]


2€EJ

and
1I'lle~,,, := II· lie" + [. ]p,I" .
Then
BC~'I"(J, E) := ({ u E CpU, E) ; 11'lle~,,, < oo}, 11·lle~,,.)
is a Banach space. It is convenient to put

BC~(J,E):= BC:,o(J,E) , [']p:= [']p,o ,

and

Lastly,

and, of course,
BUC~(J, E) := BUC:,o(J, E) .

Clearly, BUC%,I"(J, E) is a closed linear subspace of BC%,I"(J, E), hence a Banach


space.
100 III Maximal Regularity

If p > 0 and J is compact, the spaces BC~,J.' coincide essentially with the
spaces ZJ.',P introduced in [AcT87]. These authors use the weights tJ.' and tJ.'+P
instead of (11\ t)J.' and (11\ t)J.'+P, respectively. Our choice is more convenient if
J = jR+, a situation that has not been studied in [AcT87]. Related spaces also
occur in [Sob64]. The spaces BUC~ for p> 0 correspond to the spaces zP used in
[Lun87] (again for compact intervals). Also see [Lun95].

In the following proposition, in which we collect some embedding properties


of the spaces introduced above, we let BUCP := BUCP(J, E), etc.

2.1.1 Proposition Suppose that

o ::; T ::; a ::; p < 1 and 0::; J-t ::; 1.1 ::; A< 00 .

Then
BUCP '-+ BUC;,J.' '-+ BUC;,v '-+ BC;,v '-+ BC;,).. '-+ Lp,loc
f or O<p<l/A,

and
BUC; '-+ BUC(J, E) n CP(j, E) .
Moreover,
d
BUC'-+ BUCJ.'

In the Banach space cases, the norms of these injections are bounded by 1.

Proof Note that (11\ t)V ::; (11\ t)J.' for t E jR+. Hence

1I·lIev ::; 11·lle" . (2.1.1)

Moreover, if a > 0,

Ilu(s)Is -_ u(t)11
tl U ::;
Is - tIP-u [u ]*p,[e,2e]::; ()P-U[
11\ c:
]*
U p,[e,2e]

for 0 < c: ::; s < t ::; 2c: and t - s ::; 1, so that

(11\ c:)U['];,[e,2e] ::; (11\ c:)P['];,[e,2e] ,

Consequently, Hu,v ::; [']p,J.' for 0 < a ::; p < 1 so that (2.1.1) implies

11·11c.~,v ::; II· lie:,,, '


III.2 Maximal Wilder Regularity 101

On the other hand, since (1/\ c)P+IL ::; 1 for c > 0, it follows that [']P,IL ::; [']p,J for
0< p < 1. Since 11·lle" ::; 11·lIeo by (2.1.1), we see that 11·lle:." ::; 1I·llep • Lastly,

faa Ilu(t)IIP dt ::; Ilull{';" faa C p )., dt , a E j, a::; 1 ,

and the fact that u is bounded and continuous on J n [a, 00) imply BG)., '---+ Lp,loc
for 0 < p < 1/ A. Now everything, except the density assertion, is obvious.
As for the density of BUG in BUGIL , we can assume that /1 > O. Let u E BUGIL
and c > 0 be given. Then there exists T E (0,1] n J such that tIL Ilu(t)11 ::; c/2for
0< t::; T. Define v EBUC by vl[O, T] := U(T) and vl[T, 00) n J:= ul[T, 00) n J, re-
spectively. Then

tILIlu(t) - v(t)11 ::; tIL Ilu(t)11 + tIL Ilu(T)11 ::; c , O<t::;T,


implies Ilu - vile" ::; c. This proves the proposition. _
Let S denote either one of the symbols BG or BUG. Then, given p E [0,1)
and /1 E IR+,

S~-:-/(J, E) := ({ u E S~'IL(J, E) ; 8u E S~'IL(J, E)}, 11'lle~::) ,


where
Il-IIe +
1 p
P,J.L := 11·lle + 118·lle
p
p,f-t
p
p,p,

Of course, S~-:/ (J, E) is a Banach space and

8 E C(S~'!;/(J,E),S~'IL) . (2.1.2)
Note that [) is the classical derivative in this case.
Let J be bounded. If

O<p<l, /1?0,
then u is uniformly continuous near sup(J). Hence it has a unique continuous
extension 'IT over] and it is clear that 'IT belongs to the set above with J replaced
by ] and has the same norm as u. Thus in the case of the spaces above we can
assume without loss of generality that J is closed. The same is true for the spaces
S~1/(J, E) and BUG~(J, E) with 0 < p < 1 and /1 ? O.

2.1.2 Lemma Let To > 0 be fixed. Then, given p E [0,1) and /1 ? 0, the space
S~'IL(IR+,E) [resp. S~jt(IR+,E)]
is a regular extension of

S~'IL(J, E) [resp. S~jt(J, E)] ,


uniformly with respect to any closed interval J containing [0, To].
102 III Maximal Regularity

Proof Let J := [0, T] for some T? To and let S E {Sg,IL' S~~f} and r := (u!---t ul J).
Then it is clear that r E .c(S(JR+,E),S(J,E)). Given u E S(J,E), put

u(t) := { u(t) , t E J =: [0, T] ,


3u(2T - t) - 2u(3T - 2t) , t E [T, (3/2)T] .

Let !.p E COO(JR+) be equal to one on [0, T + To/6] and vanish for t? T + To/3,
and put rCu:= !.pu. It is easily verified that r C E C(S(J, E), S(JR+, E)) and that we
can choose !.p such that r C is bounded independently of J with J::) [0, To]. Since
rr C = id, the assertion follows. _

2.1.3 Lemma The spaces BCt(JR+, E) [resp. BUC~(JR+, E)] are [resp. contin-
uously] translation invariant for j = 0, 1 and {L ? O.

Proof Note that {As; S?' O} is a contraction semigroup on BCIL (JR+, E) and
on BUCIL(JR+, E). It is, trivially, strongly continuous on BUC(JR, E). Since, thanks
to Proposition 2.1.1, BUC(JR, E) is dense in BUCIL(JR+, E), the semigroup of left
translations is strongly continuous on BUCIL(JR+, E). Now the assertion for the
spaces BC~ and BUC~ follows from the fact that translations and differentiation
commmute. _

Let (Eo, E I ) be a densely injected Banach couple and suppose that

p, {L E [0,1) . (2.1.3)

Put
lEo := lEo(J) := BUC~'IL(J, Eo) (2.1.4)
and
lEI :=lEI(J) :=BUC~'IL(J,EI)nBUc~,tP(J,Eo). (2.1.5 )
Note that (2.1.3) and Proposition 2.1.1 imply that

and
lEI '----> LI,loc(J, EI) n Wl,loc(J, Eo} .
Thus we are in the situation of (1.5.1) and (1.5.2), respectively.

2.2 Semigroup Estimates


Now we prove some technical lemmas that will ultimately lead to maximal regu-
larity results for the pair (lEo, lEI), given suitable additional assumptions. For this
III.2 Maximal Holder Regularity 103

we assume that
A C H(E1' Eo) and there are constants M >
and {j E (0, 7r /2) such that ~t9 C p( - A) and
°
IIAllc(E"Eo ) + (1 + 1..\l)l-j II (..\ + A)-lllc(Eo,E
for (..\,A) E ~t9 x A and j = 0, 1.
j ) ::; M
} (2.2.1 )

Note that (2.2.1) is a special case of (I1.4.2.1). Thus (I1.4.2.2) and Lemma II.4.2.1
are valid. However, we need the following improved estimates. As usual, 11·llj de-
notes the norm in E j .

2.2.1 Lemma Let assumption (2.2.1) be satisfied. Then


(x, A) E E1 X A , (2.2.2)
and there exists a constant w > 0 such that
(t, A) E lR+ X A, kEN. (2.2.3)

Proof The first estimate is a restatement of (II.4.2.2). Since (2.2.1) implies


IIA- 1 11.c(Eo,Ej ) ::; M for A E A, it follows from
(..\ + A)-l = A- 1 (1 + ..\A- 1 )-1 for 1..\IIIA- 1 1Ic(Eo) < 1
that [Izl < l/M] c p(A) and
..\ E [Izl ::; r], AEA ,

where r := 1/(2M). From this and from (2.2.1)


we deduce, letting w := r sin {j > 0, that

[Rez ~ w] C p(2w - A)

and

°
-r -w'
for Re..\ ~ wand A E A. Thus we infer from Re-
mark 1.1.2.1(a) that there exists K ~ 1 such that

-2w+AcH(E1,Eo,K,W) .

Consequently, Corollary 1.1.4.3 implies the ex-


istence of constants M' > 0 and {j' E (0, 7r /2)
such that

(2.2.4)
104 III Maximal Regularity

for A E ~'fl' and A E A. Thus we deduce from Lemma II.4.2.1 that

(2.2.5)

for (t, A) E JR+ X A and kEN. Now (2.2.3) follows (with any positive w that is
strictly smaller than the one of (2.2.5» .•

In the following we put

E(} := (Eo, Ed£,exl , 0<8<1,

and denote by II·II(} the norm in E(}. Then we prove the following fundamental
mapping result.

2.2.2 Lemma Let assumption (2.2.1) be satisfied. Then

Proof Choose w > 0 so that (2.2.5) is satisfied. Then

Ilxllo = II(-w + A)-l(-w + A)xllo ~ cII(-w + A)xll o


and (2.2.2) imply

IIxlll ~ M (II (-w + A)xllo + w Ilxllo) ~ ell (-w + A)xll o


for x E El and A E A. Hence
tl-' IIR-w+Ax(t)lll ~ ctl-'-l Ilt( -w + A)e-t(-w+A)xllo
for t E lR+, x E Eo, and A E A. Thus, if Jl > 0, we deduce from (2.2.5) and
(1.2.10.6)-(1.2.10.10) that

x EEl-I-" AEA ,
and that tl-' RAX(t) -+ 0 in El as t -+ 0 for x E E 1 - w If Jl = 0,

ewt IIRAX(t) II 1 ~ M Ile-t(-w+A)Axlio ~ cllxl1 1 , x E E1 , A E A,


and RAx E C(JR+, E 1 ) for x EEl. This shows that

(2.2.6)

By interpolation we obtain from (2.2.2) and (2.2.3) that

II(tA)ke-tAII.c(E1_",Ed ~ c(k, Jl)Cl-'e- wt (2.2.7)


III.2 Maximal Holder Regularity 105

for (t, A) E lR+ X A and kEN. Since

e- tA _ e- sA = - i t Ae- TA dT , O<s<t<oo, (2.2.8)

we deduce from (2.2.7) that

Ile- tA - e-sAII.qE,_,..,E , ) ~ cit T-I-'-le- WT dT ~ c(l /\ c)-I-' log(t/ s)


for 0 < c ~ s < t < 00. Thanks to
log(t/s) = log(l + (t - s)/s) ~ (t - s)/s ~ (t - s)P(l /\ c)-P (2.2.9)

for 0 < c ~ s<t ~ 2c, we see that


(x,A) EEl-I-' x A. (2.2.10)

Given y E dom(A2), we infer from (2.2.2), (2.2.7), and (2.2.8) that

Ile-tAy - e-SAylh ~ cit T-I-'e- WT dT IIAylll-1-' ~ cc-I-'(t - s) IIAYlll-1-'

for 0 < c ~ s<t< 00. Thus


c > O. (2.2.11)

Given x E El == D(A) and 1/ > 0, there exists y E El such that Ily - Axllo < 1//M,
thanks to the density of El in Eo. Thus
z ;= A-1y E dom(A 2)
and
liz - xlll ~ M IIAz - Axllo = M Ily - Axllo < 1/ .
This shows that dom(A2) is dense in E 1. Since El ~ E1-fJ-' it follows that dom(A2)
is dense in E 1- w Thus, given x E E1-fJ-' we infer from (2.2.10) that there exists
y E dom(A2) such that
(2.2.12)
On the other hand, (2.2.11) implies the existence of co E (0,1] such that

cP+fJ-[RAY]CP([c,2e),Ed < 1//2 , 0< c ~ co. (2.2.13)


Hence we see from (2.2.12) and (2.2.13) that
lim cP+I-'[RAX]CP([c ' 2c) , Ed = 0
g-+O

for x E E1-fJ- and A E A. By combining this with (2.2.6) and (2.2.10), the assertion
follows .•
106 III Maximal Regularity

For the following corollary recall that lEI has been defined in (2.1.5).

2.2.3 Corollary Let assumption (2.2.1) be satisfied. Then

uniformly with respect to J, that is, the norm of this map can be bounded indepen-
dently of J.

Proof Given x EEl-I-" it follows that 8(RAX) = -ARAX in Eo. Hence we infer
from (2.2.2) and Lemma 2.2.2 that

118(RAX)llc~.,,(J,Eo) :::; IIARAXllc~,,(IR+,Eo) :::; M IIRAXllc~,,,(IR+,E,) :::; C Ilxlll_1-'


for (x, A) EEl-I-' X A. Now the assertion is obvious .•

2.2.4 Remarks (a) Put

1Fo := BC~'I-'(J, Eo)


and
1Fl := BC~'I-'(J, Ed n BC~,tP(J,Eo)
if 0 < J1 < 1, and let 1Fj := lEj if J1 = O. Then

lEo '----+ 1Fo '----+ LI,loc(J, Eo) (2.2.14)

and
lEl '----+ 1FI '----+ Ll,loc(J,EI ) nWl,loc(J,Eo) , (2.2.15)
respectively. Moreover, given assumption (2.2.1),

uniformly with respect to J.


Proof This follows by replacing in the first part of the proof of Lemma 2.2.2 (up
to (2.2.10)) the space EI-I-' by (Eo, EIh-I-',(X)' •
(b) Since 8(RAX) = -ARAX, it follows from (a) that RAX is a Wl1oc-solution of
(1.5.5)(x,A,O) for x E (Eo, Elh-/L,OO' • '

2.3 Trace Spaces


Thanks to (2.2.15) the trace spaces ')'lE l and ')'1FI are well-defined. The following
theorem gives a characterization of these spaces. It is convenient to put

(Eo, El)~,oo := (Eo, Elh,oo := EI (2,3.1)

to simplify the presentation.


III.2 Maximal Holder Regularity 107

2.3.1 Theorem Let (Eo, Ed be a densely injected Banach couple and suppose
'H.(EI' Eo) -=1= 0. Then

for p, j.J, E [0,1).

Proof Let A C 'H.(El' Eo) satisfy assumption (2.2.1). Note that we can assume
that A -=1= 0. Indeed, A E 'H.(EI' Eo) iff a + A E 'H.(El' Eo) for a E IR. Thus there
exists Ao E 'H.(EI' Eo) with type( -Ao) < O. Now Corollary 1.1.4.3 implies the ex-
istence of a neighborhood A of A in £(EI' Eo) satisfying (2.2.1).
Let x E EI-J.L (resp. x E (Eo, Edl-J.L,oo) and A E A, and put u := RAxIJ·
Then Corollary 2.2.3 (resp. Remark 2.2.4(a» and "(U = x imply

EI-J.L '---* "(lEI (resp. (Eo, EIh-J.L,oo '---* ''(IF\) .


Conversely, given u E IF\, put x := u(O) E Eo. We deduce from (2.2.15) and
Theorem 1.2.2 that

in Eo. Thus, given A E A,

Ae-tAx = Ae-tAu(t) - Ae- tA fat it,(T) dT ,

From this we obtain by means of Lemma 2.2.1 that

tJ.L-I IltAe-tAxllo :::; c(tJ.L Ilu(t)III + tJ.L-I fat 11it,(T)llo dT )e- wt


(2.3.2)
:::; c{ (1/\ t)J.L Ilu(t)III + sup (1/\ T)J.L 11it,(T) 110 }
O<T-:;t

:::; C IIull1F1
for t E f Thus, if j.J, > 0, we infer from (2.2.3) and (1.2.10.6)-(1.2.10.9) that
AEA.
Since this is true for every u E IF I with "(U = x, it follows that

"(lFI '---* (Eo, EIh-J.L,= .


If, in addition, u E lEI, we obtain from (2.3.2) that
lim tJ.L-l IltAe-tAxllo = 0 .
t->O

Hence (I.2.1O.1O) implies "(lEI '---* El - w


108 III Maximal Regularity

Lastly, suppose that 11 = O. Then we deduce from (2.2.2) and (2.3.2) that

Ilxlh ::; M IIAxll o::; M sup Ile-tA Axllo ::; cIlulllE ,


O<t~l 1

thanks to the fact that F 1 = lEI in this case. Thus {'lEI '-' E1 if 11 = O. This proves
the theorem. _

2.3.2 Remark The above proof shows that, given P,11 E [0,1), there exists a
constant a > 0 such that

and
a- 1 11·IIE, _1' ::; 1I·II-ylE, ::; a II·IIE,-" '
respectively, uniformly with respect to J. _

The fact, contained in Remark 2.2.4(a), that RA E £( (Eo, E 1h-/L,oo, F 1) can


be found - even in greater generality - in [AcT87].
If p = 0, Theorem 2.3.1 is essentially due to Da Prato and Grisvard [DaPG79]
(also cf. [Lun85j), if one takes into consideration that it has been shown by Gris-
vard in the appendix to [Lun85] that the spaces C(I1, E 1, Eo), used in [DaPG79]'
can be replaced by lEI ([0,1]) (for p = 0). To be more precise, in [DaPG79] it is
shown that {'lEI is an interpolation space between Eo and E1 of exponent 1 - 11
if 0 < 11 < 1 and p = 0, even if it is not assumed that Ji(E1' Eo) i= 0 and without
the assumption that E1 is dense in Eo. In [DorF87] it has then been shown that
this 'continuous interpolation method' is equivalent to (', ')~,oo' which explains the
name 'continuous interpolation functor'.
It should be remarked that, given A E Ji(E1' Eo), the notations

and
DA(B) := (Eo, D(A))~,oo ~ (Eo, Ed~,oo
are often used in the literature. Note that the second equivalences follow from
Lemma I. 1. 1. 2.

Now it is easy to prove an important embedding theorem for the spaces lEI
and F 1 , respectively. Here we again use the convention (2.3.1).

2.3.3 Theorem Let J be closed in lR+, let (Eo, Ed be a densely injected Banach
couple with Ji(E1' Eo) i= 0, and suppose that P,11 E [0,1). Then

BC~'/L(J, E 1 ) n BC~-:;/(J, Eo) '-' B(J, (Eo, Ed1-/L,oo)


III.2 Maximal Holder Regularity 109

and

Proof It follows from Lemmas 2.1.2 and 2.1.3 that IF\(JR+) and lEI(JR+) are
translation invariant regular extensions of IF 1 (J) and lEI (J), respectively, and that
lEI (JR+) is continuously translation invariant, provided p = O. Hence the asser-
tion follows from Proposition 1.4.2, Theorem 2.3.1, and the embedding results of
Proposition 2.1.1. •

The above theorem generalizes [Sim92, Satz 5.1], where - by a different


method - the second part of the assertion has been proven for a compact interval
and p = O. It should be noted that such a result does not imply the above em-
bedding theorem if J is not bounded. For this it is important that the elements
of BUCI" (J, E) are uniformly continuous near sup( J). For this reason we have
introduced the spaces BUCI"(J, E) and do not work with the simpler spaces
{u E CU,E); [t t--7 tl"u(t)] E B(j,E), limt .....Qil"u(t) = O} ,
that are used in the literature most frequently (e.g., [CIH+87]).

2.4 Estimates for KA


In the following lemma we study the behavior of the linear operator KA in the
spaces S~'I" introduced above.

2.4.1 Lemma Let assumption (2.2.1) be satisfied and suppose that p is posi-
tive. Then
(A t--7 KA) E B(A,L(S~'I"(J,Eo),S~'I"(J,E1))) ,
uniformly with respect to J.
Proof Given A E A and f E BC~'I"(J, Eo), define Vj := VA,j by

KAf(t) = r/ (t . .
io
2 e .... r)A f(T) dT + it
t/2
e .... (t .... r)A [f(T) - f(t)] dT

+ it
(2.4.1 )
e .... (t .... r)A f(t) dT =: V1(t) + V2(t) + V3(t)
t/2
for t E J. Then Lemma 2.2.1 implies

(III t)1" Il v1(t)III ::; c(ll1 t)1" fo t / 2(t - T) .... 1(111 T) .... l"e .... w(t .... r) dT
sup (lI1T)l"llf(T)llo (2.4.2)
O<r'St
::; c Ilfllcl'([O,t],E o) ::; c Ilfllcl'(J,Eo )
110 III Maximal Regularity

for t E j and A E A, uniformly with respect to J.


Note that, given 0 < t/2 ::::; 7 < t with t E J,

Ilf(7) - f(t)llo ::::; (t - 7)P[Jl;,[t/2,tj ::::; c(t - 7)P(1/\ 7)-P-IL[J]C~,,,([0,t],Eo)

if t - 7 ::::; 1, and that

Ilf(7) - f(t)llo ::::; IIf(7)llo + Ilf(t)llo ::::; c(l/\ 7)-1L 1IfIIc,,([o,t],Eo)


otherwise. From this it easily follows that

(1/\ t)1L Ilv2(t)111 ::::; c(l/\ t)1L itt/2 (t - 7)-le- (t-T) Ilf(t) - f(7)llo d7
w
(2.4.3)
::::; c Ilfllc~,,,([O,t],Eo) ::::; C Ilfllc~,,,(J,Eo)
for (t, A) E j x A, uniformly with respect to J.
Lastly, the closedness of A in Eo implies

AV3(t) = itt/2 Ae-(t-T)A f(t) d7 = (1 - e-(t/2)A)f(t) . (2.4.4)

Thus, by Lemma 2.2.1,

(1/\ t)1L II v3(t)11 1 ::::; c Ilfllc,,([O,tJ,Eo) ::::; c Ilfllc,,(J,Eo) (2.4.5)

for (t, A) E j x A, uniformly with respect to J. Since it is easily verified that


Vj E C(j,E 1 ) (or see the estimates below),

(2.4.6)

uniformly with respect to J.


If f E BUC~'IL(J, Eo) and 11 > 0, we see from (2.4.2), (2.4.3), and (2.4.5) that
lim tIL
t->O
Ilvj(t)lIl = 0, 1::::; j::::; 3. (2.4.7)

If f E BUC~(J, Eo), write

KAf = KA(J - f(O») + KAf(O) . (2.4.8)

By replacing in the estimates above f by f - f(O) (and 11 by 0, of course), we


see that
as t--+O.
Moreover,

AKA(J(O»)(t) = (1 - e-tA)f(O) --+ 0 in Eo as t--+O.


This shows that KAf(t) -> 0 in El as t --+ 0, that is, that KAf is continuous at
t = 0 in the topology of El.
III.2 Maximal Holder Regularity 111

In the following, it is always understood that 0 < 10 ::; S < t ::; 210 with 210 E j,
and that (f, A) E BC~,/lJJ, Eo) x A, and it should be noted that the estimates
below hold uniformly with respect to J. Then, using (2.2.8), the closed ness of A,
and the fact that the estimates below justify the following calculations:

Hence, thanks to Lemma 2.2.1,

::;c[(1!\E)~l"log(2-slt)+ 10 r/ [(s-T)-1_(t-T)-1](1!\T)-l"dT]
2

. Ilflle,,([O,E},Eo )
Observe that

log(2 - sit) ::; 1 - sit = (t - s)P(t - s)l- p It::; (1!\ E)-P(t - s)P

and

r/ [(s - T)-l - (t - T)-l]


10
2
(1!\ T)-I" dT ::; c(t - s)s-1(1!\ s)-I"

::; c(l!\ E)-P-I"(t - sy .

Thus

(1!\ Ey+I"[V1l~p([E,2E},E,) ::; c II file" ([0,2E},Eo) ::; c Ilflle,,(J,Eo ) (2.4.9)

Similarly,

AV2(t) - AV2(S) = ( r -1
It/2
8

8/2
)Ae-(t-r)A(J(T) - f(t)) dT

+1 S
A[e-(t-r)A - e-(s-r)Al(J(T) - f(s)) dT

1
s/2
+ 8
Ae-(t-r)A(J(s) - f(t)) dT
s/2
= ( I 8
t
-
It/2 )Ae-<t-r)A(J(T) -
8/2
f(t)) dT

- 1 8 1tA2e-(0"-r)A [f(T) - f(s)] dadT


8/2 8
+ [e-(t-s)A - e-(t-S/2)Al(J(s) - f(t)) .
112 III Maximal Regularity

Note that, given s/2 :s; 7 :s; ~ :s; t with ~ ?: s,

if ~ - 7 :s; 1, whereas

IIf(~) - f(7)llo :s; Ilf(~)llo + Ilf(7)llo :s; 2(1/\ 7)-1" Ilfll c ,,([0,2c],Eo)
otherwise. Hence we easily deduce from Lemma 2.2.1 that

II v2(t) - v2(s)lh

:s; c(l/\ E)-P-I"{ (t - s)P Ilfllc~,,,([0,2c],Eo)


+ 1 8

s!2
[(s - 7)-1 - (t - 7)-1] (s - 7)P d7 [j]c~,,,([0,2E],Eo)}

:s; c(l/\ E)-P-I"(t - sy Ilfllc~,,,([o,2E],Eo)


for t - s :s; 1, thanks to

is s~
[(s - 7)-1 - (t - 7)-1] (s - 7)P d7 = (t - s) is
s~
(s - 7)p-l(t - 7)-1 d7

:s; (t-s)P l°Oe-l(l+~)-ld~,

where we used the substitution ~:= (s - 7)(t - S)-I. Consequently,

(2.4.10)

Lastly, (2.4.4) implies

AV3(t) - AV3(S) = (1- e-(t!2)A) (j(t) - f(s)) + I


s!2
t!2
Ae- rA f(s) d7 .

Hence we obtain from (2.2.9) and Lemma 2.2.1 that

(2.4.11)

Estimates (2.4.9)-(2.4.11) show that

for (f, A) E BGg,1" (J, Eo) x A, uniformly with respect to J. This, together with
(2.4.6), proves the assertion if S = BG.
III.2 Maximal Holder Regularity 113

Now let 1 E BUC%,I-'(J,Eo). If J1 > 0, we infer from (2.4.9)-(2.4.11) that

~!.To c:P+1-' [KA1]c p ([c,2c],E,) = 0 .

The same is true if J1 = 0 and 1(0) = 0, since 1 E BUC(J, Eo) in this case. Thus
suppose that 1 E BUC%(J, Eo) and put x := 1(0) E Eo. Then, thanks to (2.4.8)
and the preceding observation, it suffices to show that

(2.4.12)

Given y E Eo, thanks to (2.2.9) and Lemma 2.2.1,

IIKAY(s) - KAy(t)lh ::; cllltAe-rAYdTllo


(2.4.13)
::; clog(t/s) Ilyli o::; cc:-P(t - s)P IIYllo
for 0 < c: ::; s < t ::; 2c: and c: ::; 1. If y EEl, we obtain instead

(2.4.14)

Since El is dense in Eo, we deduce from (2.4.13) that, given rJ > 0, there exists
y E El such that

c:::;1.

Then (2.4.14) implies the existence of c:o E (0,1) such that

o < c: ::; C:o .

This proves (2.4.12). Note that KA1 E BUC%,I-'(J,El ) implies

KA1 E BUC(Jn [s,oo),El ) , sE j .

Hence we deduce from (2.4.7) for J1 E (0,1) that [t f-+ tl-' KA1(t)] E BUC(J, E l ) so
that KA1 E BUC%,I-'(J,El ) if 1 E BUC%,I-'(J,Eo) for 0::; J1 < 1..

The above proof uses ideas from the proof of [AcT87, Proposition 2.1(v)].
The case J1 = 0 has also been considered in [Lun87]. In either reference El is not
supposed to be dense in Eo, but J is compact.

2.5 Maximal Regularity


Now we shall show that (lEo, lEd and (Fo,lB't) are pairs of maximal regularity for
H(El' Eo), provided p > O. For this we prepare the following technical result.
114 III Maximal Regularity

2.5.1 Lemma Let assumption (2.2.1) be satisfied and suppose that p > O. Then
KAf E Wl1oc(J, Eo) and
8(KAJ) = f - AKAf

for f E BC%,!-'(J, Eo) and A E A.

Proof Let (J, A) E BC%,!-'(J, Eo) x A be fixed and put T:= sup(J). Note that

(2.5.1)

Thus, given t.p E V(J),

(2.5.2)

Note that

-iT 8t.p(t)e~(t~T)A f(r) dt = t.p(r)f(r) -iT t.p(t)Ae~(t~T)A f(r) dt, (2.5.3)

since the last integral exists in Eo, thanks to

iT t.p(t)Ae~(t~T)A f(r) dt = iT [t.p(t) - t.p(r)] Ae~(t~T)A dt f(r)

+ t.p(r) [1 - e~(T~T)Alf(r) .

Hence we obtain from (2.5.2) and (2.5.3) that

8(KAJ)(t.p) = -KAf(8t.p) = loT t.p(t) [f(t) - lot Ae~(t~T)A f(r) dr] dt

= (J - AKAJ)(t.p) ,

since the proof of Lemma 2.4.1 shows that the inner integral in the second to last
term exists, that A can be interchanged with integration, and that

This proves the assertion .•

2.5.2 Corollary Let assumption (2.2.1) be satisfied and suppose that p > O. Then

uniformly with re8pect to J.


III.2 Maximal HOlder Regularity 115

Proof By Lemma 2.5.1,

Since A E £(El' Eo), it follows from Lemma 2.4.1 that

8(KAf) E 1F'0 [resp. lEo] , fElF'o [resp.lEo] ,

and that the norms of these clements can be bounded independently of

(j, A) E 1F'0 x A [resp. lEo x A] and J .

Now the assertion is a consequence of Lemma 2.4.1. •

After these preparations it is easy to prove the following maximal regularity


result.

2.5.3 Theorem Let assumption (2.2.1) be satisfied and let p E (0, 1) and /.k E [0, 1).
Put

(2.5.4)

Then (lEO,lEl) is a pair of maximal regularity for A, and

")'lEl === E 1- et := (Eo, E 1 )L et ,oo ,

where (Eo,El)~,oo:= E 1 . Moreover,

for A E A, uniformly with respect to J.

Proof It follows from (2.1.2) that 8 E £(lEl' lEo). Since A E £(lEl' lEo), we see
that 8 + A E £(lEl, lEo). Theorem 2.3.1 guarantees that ")'lEl === E 1- et , and Corol-
laries 2.2.3 and 2.5.2 imply

uniformly with respect to J. Note that RAX + KAf E lEI and (2.2.15) together
with Remark 2.2.4(b) and Lemma 2.5.1 imply that RAX + KAf is a Wl.loc-solution
of the Cauchy problem (1.5.5)(x,A,f) for (x, A, f) E E 1 - et X A x lEo. Now the as-
sertion follows from (8 + A, "),)-1 = RAX + KAf and Theorem 1.5.2 .•

2.5.4 Remark Observe that, by replacing in the above proof (lEo, lEI) by (1F'0, IF'I)
and E 1 - et by (Eo, EIh-p,co, respectively, it follows that (IF' 0, 1F'1) is a pair of max-
imal regularity for A if p > 0 .•
116 III Maximal Regularity

The above maximal regularity theorem for the pair (lFo,lF\) is (implicitly)
contained in [AcT87] and, without proof, in [Sob64]. The maximal regularity for
the pair
(BUC:(J, Eo), BUC:(J, E I ) n BUC~+P(J, Eo)) ,
that is, for the pair (Eo, Ed with f.L = 0, has been proven in [Lun87] (cf. [Lun95]).

In the following, we are mainly interested in the pair (Eo,E I ). Reasons be-
ing, for example, that (., ·)L/l-,oo is an admissible interpolation functor whereas
(., ·h-/l-,oo is not and, most importantly, the embedding EI '-+ BUC(J, EI-/l-).
Hence we restrict our considerations to this pair and leave it to the interested
reader to extend the following results to the pair (IF 0, lF I) (if they are valid in
that case).

Using results from Section 1.1, it is not difficult to give sufficient conditions
for assumption (2.2.1) to be satisfied. A particularly important case is contained
in the following statement.

2.5.5 Theorem Suppose that (Eo, EI) is a densely injected Banach couple, let
p E (0, 1) and f.L E [0, 1), and define (Eo, E I ) by (2.5.4) with J := 1R+. Also suppose
that Ao is a compact subset of H(El, Eo) and that there exists Wo > such that°
the spectral bounds satisfy

s(-A) :S -Wo , AEAo·


Then, given w E [O,wo), there exists a neighborhood A of Ao in C(El,Eo) such
that (cWEo, e-wEd is a pair of maximal regularity for A, and

,(e-WE I ) ~ EI-/l- := (Eo, EI)L/l-,oo ,

where (Eo, E)~,oo := E I . Moreover,

II (8 + A, ,)-III.c(e-wlEo xEl_/L,e-wlEl) :S c
for A E A.

Proof It is an immediate consequence of Corollary 1.1.4.3 that there exists a


neighborhood A of An in C(E!, Eo) such that -w + A satisfies condition (2.2.1).
Now the assertion follows from Theorem 2.5.3 and Proposition 1.5.3 .•

If we restrict our consideration to bounded intervals, the next theorem shows


that the pair (2.5.4) is a pair of maximal regularity for all of H(E I , Eo).

2.5.6 Theorem Let J be bounded, (Eo, E I ) be a densely injected Banach couple,


p E (0,1) and f.L E [0,1), and
(2.5.5)
III.2 Maximal HOlder Regularity 117

Then
l'lEl ~ El-I-' := (Eo, El)LI-',oo ,
where (Eo,El)~ 00:= E l , and (lEO,lEI) is a pair of maximal regularity for all
of H(EI' Eo). Given K :::: 1 and w > 0,

A E H(EI,Eo,K,W) .

Proof Assumption (2.2.1) is satisfied for w+H(E1,Eo,K,W), thanks to Corol-


lary I. 1.4.3. Hence Theorem 2.5.3 and Proposition 1.5.3 imply that (ewlEo, eWlE 1 )
is a pair of maximal regularity for H(EI' Eo, K, w). Since J is bounded, it is easily
verified that eWlEj ~ lEj for j = 0, 1. Now the assertion is obvious .•

2.6 Nonautonomous Problems


Let F and F j be Banach spaces and let

(2.6.1)

be a multiplication. It is an easy consequence of (II. 1. 1.6) that point-wise multi-


plication induced by (2.6.1) is bilinear and continuous:

(2.6.2)

for p, f-l E [0, 1). From this we infer that, given Banach spaces E j ,

(2.6.3)

via the natural injection


A>---+(UHAu).

After these preparations it is possible to extend Theorem 2.5.6 to the nonau-


tonomous case.

2.6.1 Theorem Let (Eo, Er) be a densely injected Banach couple and, given
p E (0,1) and f-l E [0,1), put

(2.6.4)

Suppose that A E BUC~(J, H(EI' Eo)) and that either J is bounded or there exists
°
T:::: such that s( -A(T)) < and °
(2.6.5)
118 III Maximal Regularity

for some). E (0,1). Then

Proof First assume that J is bounded. Then we can assume that it is compact.

°
Hence A(J) is compact in 7i(E1' Eo). Thus it follows from Corollary 1.1.3.2 that
there exist K, ;?: 1 and w > such that

A(J) c 7i(E1,Eo,K"w). (2.6.6)

Given 8 E J, denote by Bs the constant map t ~ A(8). Theorem 2.5.6 implies

(2.6.7)

and the existence of a constant a such that

8 E J . (2.6.8)

Let As(t) := A(8 + t) for 8 E J and t E J - 8. Observe that, given 8 E J =: [0, T],
there exists 8 (8) E (0, 1) such that

liAs - BsllcP([o 6]n(J-s) L(El Eo)) =


P' "OStSM(T-s)
max IIA(8 + t) - A(s)IIL(El Eo)
,

+ sup E P [A(8+ .)] [e2E]::; 1/(2a)


O<2e<6/\(T-s) P, ,

for 0 < 8 ::; 8 (8). Note that there exists 80 > 0 such that 8 (8) ;?: 80 for 8 E J. From
this and the compactness of J we easily deduce the existence of a number 8 > and
°
of finitely many points =: 80 < 81 ... < 8m < 8 m +! := T such that 8]+2 - 8j ::; 8
°
for j = 0, 1, ... , m - 1 and such that

IIAsj - B SJ Ilc~([0,8]n(J-sJ),L(El,Eo)) ::; 1/(2a) , j = 0, 1, ... ,m. (2.6.9)

Put

o ::; t ::; 8/\ (T - 8j) ,


8 /\ (T - 8j) < t ::; T ,
for j = 0, 1, ... ,m. Then (2.6.9) and (2.6.3) imply

IIAj - BSj IIL(lE " lE O ) ::; IIAj - BSj Ilc~([O,6]n(J-sj),L(El,Eo)) ::; 1/(2a) .

Hence we infer from (2.6.7), (2.6.8), and Proposition 1.6.3 that

(2.6.10)
III.2 Maximal Holder Regularity 119

and that

j = 0, 1, ... ,m.

Let (f,x) E lEo X El-/L be given. It follows from (2.6.10) that the Cauchy problem

it + A(t)u = f(t) , 0 <t :; 82 , u(O) = x

has a unique solution

Put
f (t)·= { f(8l +t), o :; t :; T - 81 ,
1· f(T) , T - 81 :; t :; T ,

and Xl := uO(8d. Then (h, Xl) E lEo X E l , and (2.6.10) guarantees that the Cau-
chy problem

1i,+A(t)u=f(t) , 81 <t.-:;83 ,

has a unique solution

'Ul E BUC~'/L([81,83],El) rIBUC~~P([8l,83l,Eo) .

Since A E CP([to, Tl, H(El' Eo)) for 0 < to < T, it follows from that part of The-
orem II.1.2.1 which has already been proven in Section 11.4.5 that the Cauchy
problem
it + A(t)u = f(t) , to < t :; t1 , u(to) = X (2.6.11)
has for each X E E1 at most one solution on [to, ttl, where 0 < to < t1 :; T. This
easily implies that

Thus, letting
u(t) := { uo(t) , 0:; t ~ 81 ,

U1 (t) , 81 :; t ~ 83 ,

it follows that

and that u is the unique solution of the Cauchy problem (2.6.11) on this interval.
By repeating this argument a finite number of times we see that (fJ + A, "() is sur-
jective. Since (2.6.3) implies (fJ+A,"() E .C(lE1,lEO x E 1 -/L), and since this map is
injective by what we just have shown, the assertion follows from the open mapping
theorem.
120 III Maximal Regularity

Now suppose that J = JR+ and let (f,x) E lEo(JR+) x EI-I" be given. Let

UT E BUC~,J1. ([0, T + 1], E 1) n BUC~-:: ([0, T + 1], Eo)


be the unique solution of

it,+A(t)u=f(t) , O<t:S;T+l, u(O) = x,

whose existence has been proven above. By Theorem 2.5.5

Assumption (2.6.5) and Proposition 1.6.3 guarantee that

Thus the Cauchy problem

v + A(T + t)v = f(T + t), t >0, v(O) = uT(T)

has a unique solution v E lEI(JR+). Letting

u(t) := { UT(t) O:S;t:S;T+l,


v(t - T) T:S;t<oo,

it follows from uniqueness that U E lE1(lR+) and that it is the unique solution
in lEI (JR+) of it, + A(t)u = f(t) on 1R+ satisfying u(O) = x. Now the assertion follows
similarly as above .•

Theorem 2.6.1 seems to be new in this generality. The assumption that


A E BUCg(J, H(El' Eo)) is natural in this setting. In previous maximal regularity
results in singular Holder spaces, the condition A E B UCP ( J, H (El' Eo)) has been
imposed (e.g., [AcT87]' [Lun87]).

Now it is easy to give the missing part of the proof of the main existence theorem
of Subsection 11.1.2.1.

Proof of Theorem 11.1.2.1: Part 2 Thanks to part 1 of the proof of this


theorem we know already that, given

(x, A, f) E Eo x CP(J, H(El' Eo)) x CP(J, E) ,

the Cauchy problem (1I.1.2.2)(s,x,A,f) has a unique solution

u E C(J\{s},E I ) nCI(J\{s},Eo) nC(J,Eo).


III.3 Maximal Continuous Regularity 121

Let S < So < Sl with Sl E J and put h :=: [0, Sl - so]. Define

(xl,A l ,!1) EEl x CP(Jl ,1t(El , Eo)) x CP(Jl , Eo) (2.6.12)

by Xl := u(so) and
(Al,!1)(t):= (A,f)(so +t) , (2.6.13)

respectively. Then (2.6.11), Proposition 2.1.1, and Theorem 2.6.1 imply that the
Cauchy problem

V+Al(t)V=!1(t) , 0<t:S:Sl- S0, v(O) = Xl

has a unique solution

v E BUC~(J1' E1) n BUC~+P(h Eo) .


It is obvious, letting wo(t) := v(t - so) for So :s: t :s: Sl, that

Wo E BUC~ ([so, Sl], E1) ("I BUC~+P ([so, SI], Eo)


and that Wo is a solution of the Cauchy problem

w+ A(t)w = f(t) , So < t :s: Sl , w(so) = u(so) .


Thus w = ul[so, Sl] by uniqueness. Since this is true for every admissible choice of
So and Sl, it follows that

which proves the theorem. _

3 Maximal Continuous Regularity


Let J be a perfect subinterval of lR + containing 0 and let (Eo, E 1) be a densely in-
jected Banach couple. In the preceding section we have proven maximal regularity
results for the spaces BUC~,p, under the assumption that p > O. In this section we
investigate the case p = 0, that is, the spaces BUCw

3.1 Necessary Conditions


First we consider the case J1 = 0, that is, we consider the pair

(3.1.1)
We also assume that J is bounded, for simplicity.
122 III Maximal Regularity

The following proposition shows that maximal regularity can occur in this
case in the 'parabolic case' only.

3.1.1 Proposition Suppose A E 9(Eo) n C(EI' Eo) and (lEo, lEI) is a pair of
maximal regularity for A. Then A E 1t(EI' Eo).

Proof Note that the proof of Theorem 1.5.2 implies KA E C(lEo, lEI). For this it
suffices to observe that the latter proof did not use all the properties of a parabolic
evolution operator but only those that are valid for

{e-(t-s)A ; 0::; s < t < 00 } .

Fix Xo E Eo and put f(t) := e-tAx for t E J. Then

Thus

im(e- tA ) C dom(A) , IltAe-tAII£(Eo )::; c , (3.1.2)

where we used Lemma 1.1.1.2 for the last estimate. Since it is well-known that
(3.1.2) implies that -A generates a strongly continuous analytic semigroup, the
assertion follows .•

The following theorem, due to Baillon, shows that the assumption that (lEo, lEI)
is a pair of maximal regularity implies serious restrictions on the Banach spaces
Eo and E 1 .

3.1.2 Theorem Suppose that (lEo, lEI) is a pair of maximal regularity for some
A E 1t(El, Eo). Then either EI = Eo or Eo contains a closed subspace which is
top linearly isomorphic to the space Co of null sequences.

Proof For this we refer to Baillon's note [Bai80] (also cf. [EbG91]) .•

3.1.3 Remarks (a) Clearly, EI = Eo iff A E C(Eo). Also observe that EI is


toplineady isomorphic to Eo if 1t(EI' Eo) =1= 0. Hence EI contains a copy of Co
if Eo does.
(b) Every closed linear subspace of a reflexive [resp. weakly sequentially complete]
Banach space is reflexive [resp. weakly sequentially complete]. It is well-known that
Co is neither reflexive nor weakly sequentially complete. Thus, if El =1= Eo, the pair
(lEo, lEI) cannot be a pair of maximal regularity for any A E 1t(E1 , Eo) if Eo is
reflexive or weakly sequentially complete (an LI-space, for example) .•
III.3 Maximal Continuous Regularity 123

3.2 Higher Order Interpolation Spaces


Suppose that
A E 'J-l(El' Eo) with s( -A) < 0,
where s denotes the spectral bound. Then, letting

it follows that
E 2 (A) := (dom(A2); 11·II E 2(A))
is a Banach space such that

(3.2.1)

since the density of dom(A2) in El has already been shown in the proof of
Lemma 2.2.2.
Throughout the remainder of this section we put

0<0<1.

Since (', ·)S 00 is an admissible interpolation functor, it is a consequence of (3.2.1)


that (Eo, EHO(A)) is a densely injected Banach couple for 0 ::; () ::; 1, where we put
El(A) := E l . Let Ao be the Eo-realization of A for 0 ::; ()::; 1. Since A commutes
with its resolvent, it is obvious that p(Al) = p(A). Moreover, using on El the
norm IIA·II,

11(.:\ + A)Axllo
IAllIAxll o + IIAxlll '
shows that

Now again we suppose that assumption (2.2.1) is satisfied, that is,

A C 1-l(El' Eo) and there are constants M > 0

}
and 1J E (0, 7l' /2) such that ~1? C p( -A) and
(3.2.2)
IIAII.C(E"Eo) + (1 + IAl)l-j II(A + A)-lll.C(Eo,Ej) ::; M
for (A,A) E ~1? x A and j = 0,1.

Then, similarly as above, we see that


124 III Maximal Regularity

for (A, A) E ~u x A and j = 0, 1. Thus, by interpolation,

(3.2.3)

for 0::; () ::; 1, (A, A) E ~u x A, and j = 0, 1. Now Lemma 2.2.1 guarantees the
existence of positive constants M(()) and w such that

x E El+o(A) , (3.2.4)

and
t>O, kEN, (3.2.5)

for () E [0,1] and A E A. Note that w is independent of (j.

It should be remarked that the above construction of the higher order inter-
polation spaces E 1+o(A), 0 < () < 1, is a very particular case of the much more
general situation discussed in detail in Chapter V. It should also be observed that,
unlike the spaces Eo, 0::; () ::; 1, the higher order spaces E 1+o(A), 0 < () ::; 1,
depend on A E A, in general.

3.3 Estimates for KA


Now we study the behavior of the map KA in the densely injected Banach couples
(Eo, E 1+o(A)) , 0 < () < 1. For this we begin with the following simple technical

3.3.1 Lemma Suppose that 0 ::; Jl < 1 and 0 < () < 1. Then

.+
for s, t E lR .

Proof If t ::; 2,

If t 2': 2,

Ir(s,t) = 1\S+t-T)IJ-2T-J1-dT+ lt/\s+t-T)0-2dT

::; c{ (s + t)O-l-J1- + (s + t/2)0-1} ::; c(s + 1)0-1 .


1II.3 Maximal Continuous Regularity 125

Similarly, if t ?: 2,

12 (8, t) := i t (8 +t - 7)11-2(1/\ 7)-1-' d7 = i t (8 +t - 7)11-2 d7 :S C8 11 - 1 .


t/2 t/2

Ift:S 1,

h(8, t) = i t (8 +t - 7)11-2 7 -1-' d7 :S CC1-'8 11 - 1 .


t/2

Lastly, if 1 < t < 2,

12 (8, t) =i 1
(8 +t - 7)11-2 7 -1-' d7 + i t (8 + t - 7)11-2 d7 :S C8 11 - 1 •
t/2 1

Now the assertion is obvious .•

After these preparations we can derive the mapping properties of K A in the


interpolation spaces introduced above.

3.3.2 Proposition Let assumption (3.2.2) be satisfied and suppose that 0 :S JL < 1
and 0 < e < 1. Then

KAo E £ (BUGI-'(J, Ell), BUGI-'(J,E1+o(A)) n BUG~(J, Eo))


and
AEA,
uniformly with respect to J.

°
Proof If :S Q < (3 :S 1, it is easily verified that A", :::J A,6. Thus, without fearing
confusion, we simply write A for A"" O:S Q :S 1. Note that

Ae- sA KAf(t) = fatAe-(S+t-T)A f(7) d7 , tEJ, 8>0, (3.3.1 )

for f E BUGI-'(J, Eo). Indeed, it follows from (2.2.7) and the closedness of A that
(3.3.1) is true in El and that, thanks to Lemma 3.3.1,

IIAe-sAKAf(t)lll :S clot (8 + t _7)-2+0(1 /\ 7)-1-' d71Iflle,,([0,t],Ee)


:S C8 0 - 1 (1 Ilflle,,([O,t],Eel
/\ t)-I-'
:S CS Il - 1 (1 /\ t)-I-' Ilflle,,(J,Ee) ,

uniformly with respect to J. Consequently,

sup(l /\ t)1-' SUP8- fl llsAe- sA KAf(t)lh :S c Ilflle (J Ee(A))


t>o s>O " ,
126 III Maximal Regularity

for f E BUC/-,(J, Eo), uniformly with respect to J. From this, from (3.2.5), and
from (I.2.1O.6)-(I.2.1O.9) we deduce that

(3.3.2)

and that

IIKAII£(BUC"(J,Ee),BC"U,(E,,E 2 (A))e,oc)) S; C , AEA, (3.3,3)

uniformly with respect to J.


Now suppose that f E BUCP(J, Eo) for some p E (0,1). Then it follows from
Proposition 2.1.1, estimate (3.2.3), and Theorem 2.5.3 that

Since
d d
BUCP(J, Eo) '--+ BUC(J,Eo) '--+ BUC/-,(J,Ee ) , (3,3.4)

where the density of the first inclusion follows by mollifying, for example, and
where the density of the second injection is proven in Proposition 2.1.1, we deduce
from (3.3.2) and (3.3.3) that

KA E £ (BUC/-,(J, Eo), BUC/-,(J, E1+o(A)))

and that
AEA, (3.3.5)

uniformly with respect to J.


If f E BUCP(J,Eo) for some p E (0,1), we deduce from Theorem 2.5.3 that

KAf E BUC~-::(J,Eo) '--+ BUC~(J,Ee)

and that 8(KAf) = f - AKAf. Hence, thanks to (3.2.3), (3.3.4), and (3.3.5),

118(KAf)IIC,,(J,Ee) S; IlfIIC,,(J,Ee) + c IIKAfllc,,(J,E'+B(A)) S; C IlfIIC,,(J,E e) ,

uniformly with respect to (f, A) E BUC/-,(J, Ee) X A and J. Now the assertion
follows from (3.3.4) .•

3.4 Maximal Regularity


After these preparations it is now easy to prove the main result of this section,
namely the following maximal regularity theorem.
III.3 Maximal Continuous Regularity 127

3.4.1 Theorem Let condition (3.2.2) be satisfied and suppose that 0 :::; /L < 1
and 0 < e < 1. Then

(lEe, lEHe(A))
(3.4.1)
:= (BUCj"(J, Ee), BUCj"(J, El+8(A» n BUC~(J, Ee))
is a pair of maximal regularity for Ali := {Ali; A E A}, and

(3.4.2)

Ali EAIi,

uniformly with respect to J.

Proof Since (Ee, El+8(A)) is a densely injected Banach couple and Ae satisfies
(3.2.3), it follows from Corollary 2.2.3 that

uniformly with respect to J. Similarly, it follows from Theorem 2.3.1 that (3.4.2)
is true. Now the assertion is a consequence of Proposition 3.3.2, Theorem 1.5.2,
and the fact that RAeX + KAef is a Wl,loc(J, Ee)-solution of u + Aeu = f with
u(O) = x whenever (x, 1) E ~(lEHe(A) x lEe. _

3.4.2 Remarks (a) Theorem V.1.5.9 and Remark V.1.5.11 imply

Note however that, in general, (Ee,EHe(A»)~_e,oo # El .


(b) It is clear that one can deduce - by arguments similar to the ones used
in Section 2 - from Theorem 3.4.1 results that correspond to Theorems 2.5.5
and 2.5.6, respectively.
(c) Suppose that J is bounded, A: J --> H(El' Eo) such that A(J) is relatively
compact in H(El' Eo) and, given e E (0,1),

uniformly with respect to t E J, and

Ali E BUC(J, H(E l +8(A(O),EIi )) ,


128 III Maximal Regularity

where Ao(t) := [A(t)]o for t E J. Then the proof of Theorem 2.6.1 shows that

(8 + A, ,) E Cis ( BUGp.(J, EHo(A(O))) n BUG!(J, Eo),

BUGp.(J,Eo) x (Eo,EHO(A(O)))~_p.,oo) .

Of course, the analogue to Theorem 2.6.1 is valid for J = lR+ in the present case
as well. _

Theorem 3.4.1 is due to Da Prato and Grisvard [DaPG79, Theoreme 3.1] for
the case J-t = O. That paper also contains the result of Remark 3.4.2(b) (in the case
J-t = 0, of course, and for J bounded, cf. [DaPG79, Theoreme 3.6]). The extension
to the case J-t > 0 has been carried out in [Ang90]. Our proof of it, more precisely:
our proof of Proposition 3.3.2, is closer to [Sim92, Theorem 5.4].

4 Maximal Sobolev Regularity


In this section we investigate the problem of maximal regularity in Sobolev spaces,
that is, the maximal regularity of pairs of the form

for 1 < p < 00 and a densely injected Banach couple (Eo, Ed. We shall see that
we need restrictions for the underlying Banach spaces E j as well as for the class
of admissible generators.

For the proof of maximal Sobolev regularity we need a considerable amount


of preparatory material that is of interest for its own sake. Below we give almost
self-contained treatments of the Banach-space-valued Hilbert transform and UMD
spaces as well as of parts of the theory of fractional powers of linear operators of
positive type.

4.1 Temperate Distributions


First we recall some simple facts about vector-valued distributions and Fourier
transforms. The proofs are straightforward extensions of the corresponding ones
for the scalar case. We assume that the reader has a working knowledge of scalar
distributions and Fourier transforms and we refer for details to [Hor83, vol. 1],
[Hor66], [Pet83], and [Schw66], for example. For later purposes we consider the
general n-dimensional case though, in this section, only distributions on lR are of
interest.
IlI.4 Maximal Sobolev Regularity 129

Let E:= (E, 1·1) be a Banach space. If X is a nonempty open subset of IR n ,


we denote by
kEN := N U {oo} ,
the Frechet spaces of all E-valued functions on X whose derivatives of order -::; k
are continuous, equipped with the topology induced by the family of seminorms

U f-t Pm,K(U) := I~~ Ilaaulioo,K , m < k + 1, K CC X . (4.1.1)

and Ck(X) := Ck(X, lK). We also put £(X, E) := Coo (X, E) and £(X) := £(X, lK).
Moreover,
£'(X, E) := £(£(X), E)
(equipped, as always, with the bounded convergence topology) so that

£'(X) = £(X)' .
Standard truncation and mollification arguments show that
d
D(X, E) '----+ £(X, E) . (4.1.2)

From this it follows easily that

£'(X,E) '----+ D'(X,E) , (4.1.3)

by restriction, of course. Hence £'(X, E) is a space of distributions and the stan-


dard 'scalar proof' applies to show that U E £'(X,E) iff u is a distribution with
compact support, where the support, supp, of u E D'(X, E) is given by

supp(u) := X\ {x EX; there exists a neighborhood U of xEX


such that u( t.p) = 0 for t.p E D(U) } .

Recall that £(X) and £'(X) are Montel spaces. Hence they are reflexive.

We denote by S(JRn , E) the Schwartz space of smooth rapidly decreasing


E-valued functions on IRn. Thus u E S(JR n , E) iff u E £(JR n , E) and

qk,m(U):= sup (1 + Ixl 2 )k laau(x)1 < 00 , k,mEN. (4.1.4)


xEIR n
Io-I::;m

Then S(JRn , E) is a Frechet space with the topology induced by the family of
seminorms (4.1.4), and S(JR n ) := S(JRn , lK). By standard arguments

(4.1.5)

We define the space S' (IRn, E) of E-valued temperate distributions by


130 III Maximal Regularity

and SI(ffin) := SI(JRn,lK). Note that (4.1.5) implies


£1 (JR n , E) '---+ SI (JR n , E) '---+ Vi (JR n , E) . (4.1.6)

Moreover, S(JR n ) and S/(JR n ) are Montel spaces, thus reflexive.

Lastly, we introduce the space OM (JR n , E) of E-valued slowly increasing


smooth functions on JR n . Namely, U E OM (JR n , E) iff U E £ (JR n , E) and, given
Q E N n , there exist rno: E N and Co: > 0 such that

The space OM(JR n , E) is a LCS with respect to the topology induced by the family
of seminorms

It follows that
S(JR n , E) '---+ OM (JR n ,E) '---+ S/(JRn,E) . (4.1.7)
Of course, OM(JR n ) := OM(JRn,lK). It is easily verified that the map
(4.1.8)
is well-defined and bilinear. Moreover, if rn E £(JR n ) then
(4.1.9)

Clearly, the roles of E and lK can be interchanged on the left-hand side of (4.1.8).
Similarly, OM(JR n ) can be replaced by OM(JRn , E) if S(JR n , E) is replaced by S(JRrl ).

4.2 Fourier Transforms and Convolutions


Let E := (E, 1·1) be a Banach space. Given u E L1 (JR n , E),

:Fu(~) := u(~):= ( e~i(~,x)u(x) dx ,


JlRn
is the Fourier transform of u, where
( I: ) . _ ",71 d j
",X .-L..,j=l"x.

The RIEMANN-LEBESGUES LEMMA asserts that

where, in general, given any locally compact metric space M,

Co(M,E):= (Co(M, E), 11·11=)


is the closed subspace of BUC(M, E) consisting of the continuous functions van-
ishing at infinity. Recall that this means that, given any E > 0, there exists a
III.4 Maximal Sobolev Regularity 131

compact subset K of M such that 1'U(x)1 < E for x E M\K. Thus Co(M, E) is a
Banach space.
The FOURIER INVERSION THEOREM guarantees that
(4.2.1)

and
(4.2.2)
where
u(x) :=u(-x) ,
is the reflection of 'U.

The Fourier transform u:= F'U of the temperate distribution 'U E S'(JR n , E)
is defined by
u('P) := 'U(0) ,
Define the reflection of 'U E 1J' (JR n , E) by

u('P) := 'U(0) ,
Then the FOURIER-SCHWARTZ THEOREM guarantees that

FE Laut (S' (JRn , E)) (4.2.3)

and that (4.2.2) is true for 'U E S' (JR n , E). Moreover, if'U E L1 (JR n , E) C S' (JR n , E),
u
the new definition of coincides with the original one.

It is a fundamental property of the Fourier transform that

(DO<'U)~ := DaU = ('u , 'UES'(JRn,E) , aENn , (4.2.4)

where D j := -ioj for j = 1, ... ,n.

Given a E JRn , we denote by Ta'U the (right) translation of 'U E E~n by a,


that is,
Ta'U(X) :='U(x-a) , x E JRTI. ,
Moreover,

It is not difficult to verify that { Ta ; a E JRn }, the group of translations on JR n ,


is strongly continuous on each one of the spaces J" (lRn ,E) for

J" E {1J, 1J', S, S',£, ['} . (4.2.5)

It is easy to see that


FTa = e-i(a'~)F, a E JRn , (4.2.6)
where e- i (a,l:;) stands for multiplication with the function ~ 1-+ e- i (a,l:;) .
132 III Maximal Regularity

Given a E lR+, we define the dilation 0"a U of u E E lRn by a by

O"aU(X) := u(ax) , x E JRn . (4.2.7)

Moreover,

(4.2.8)

Note that (4.2.8) is consistent with (4.2.7) if u is a regular distribution. It is


easily verified that {O"a ; a E IR+ }, the group of dilations, is strongly continuous
on ~(JRn,E), where ~ satisfies (4.2.5). Moreover, it is not difficult to see that
a;;u = a-nO"a-I it for u E S(JR n , E) and a > 0, that is,
:FO"a = a-nO"a-I:F , a> O. (4.2.9)

Given
u E V' (JR n , E) and ip E V(JR n ) ,
or u E t"(JR n , E) and ip E t'(JR n ) ,
or u E S'(JRn ,E) and ip E S(JR n ) ,

respectively, the convolution, u * ip, of u and ip is defined by

u * ip(x) := U(TxtjJ) ,

It follows that

and that
(4.2.10)
In fact, convolution is a bilinear and separately continuous (indeed, hypocontinu-
ous) map:
V'(JR n , E) xV(JR n ) -+ t'(JR n , E) , (4.2.11)
t" (JR n , E) x t' (JR n ) -+ t'(JRn,E) , (4.2.12)
t"(JR n , E) xV(JR n ) -+ V(JR n , E) , (4.2.13)
S'(JR n , E) x S(JR n ) -+ OM (JR n , E) . (4.2.14)

It is not difficult to verify that

(4.2.15)

Note that
IlIA Maximal Sobolev Regularity 133

Hence
cP * u(x) := TxU(cp) = U * cp(x) , x E~n. (4.2.16)
Of course, if u E L1,loc(~n, E) is a regular distribution and cP E V(~n),

u * cp(x) = r u(x - y)cp(y) dy JJRnr u(y)cp(x - y) dy ,


ifR n
= x E ~n, (4.2.17)

(where, as usual, ea := ae for a E ][( and e E E).

Literally as in the classical case (e.g., [FoI84, Section 8.2]), direct estimates of
the integrals in (4.2.17) combined with density arguments show that convolution
is bilinear and continuous:
BC(~n, E) X L1(~n) -+ BC(~n,E) , (4.2.18)
BUC(~n, E) X L1(~n) -+ BUC(~n, E) , (4.2.19)
Lp(~n, E) X L1(~n) -+ Lp(~n,E) , (4.2.20)
L()()(~n, E) X L1(~n) -+ BUC(~n,E) , (4.2.21)
Lq(~n,E) x Lql(~n) -+ Co(~n,E) (4.2.22)

for 1 :::; p :::; 00 and 1 < q < 00. Moreover, the norms of these bilinear maps are
bounded by 1. Observe that in situation (4.2.20), for example, this means that

which is Young's inequality for convolutions. Note that (4.2.16) implies that
E and ][( can be interchanged on the left-hand sides of (4.2.11)-(4.2.14) and
(4.2.18)-( 4.2.22).
Suppose that cp E £1 (~n) and put

x E ~n, C >0.
Also let a := J cp dx. If a = 1 then {CPc ; c > O} is said to be an approximate
identity and if, in addition, cp E v(~n), cp 2: 0, and suppcp = Jan, it is a mollifier.
Classical arguments (e.g., [FoI84, Section 8.2]) show that

CPc * u -+ au as c-+O (4.2.23)

in
(4.2.24)
and in

BUC(~n,E) if u E BuC(~n,E) , (4.2.25)

respectively. If there exist c > 0 and 15 > 0 such that


Icp(x)1 :::; c(l + Ixl)-n-o , x E ~n, (4.2.26)
134 III Maximal Regularity

then the proof of [FoI84, Theorem (8.15)] carries over to the E-valued situation to
show that, given U E Lp(JR n , E) for some p E [1,00],
<Po * u(x) ---+
0-+0
au(x) , a.a. x E JRn , (4.2.27)

in fact: for every x in the Lebesgue set of u.

It follows from (4.l.7) and (4.2.14) that u * <P E 5'(JRn , E) for u E 5'(JRn , E)
and <P E 5(JR n ). Hence the Fourier transform of u * <P is well-defined and the CON-
VOLUTION THEOREM states that

(u * <p)~ = fiip .

Lastly, suppose that E:= (E,CI')) is a Hilbert space. Then L 2 (JR n , E) is


also a Hilbert space with respect to the inner product

(ulvh := (ulvh2(lRn,E):=
JlRn
r (ulv) dx .
Given u, v E 5(JR n , E), PLANCHEREL'S THEOREM guarantees that
(4.2.28)
Since 5(JRn , E) is dense in L 2 (JR n , E), it follows from (4.2.28) and Fourier's inver-
sion theorem that (27r) -n/2 :F is a unitary operator on L2 (JR n , E).

4.2.1 Remark The formulas (4.2.11)-(4.2.14) and (4.2.17) (4.2.22) exhibit a


nonsymmetry due to the fact that we did not define multiplication and convolu-
tions if both factors are vector-valued. For this suppose that E j , 1::::; j ::::; 3, are
Banach spaces and

is a multiplication (cf. (II.l.lA)). Then, given Uj E V'(JRn,Ej ) for j = 1,2, such


that at least one has compact support, there exists a unique
Ul *. U2 E V'(JR n ,E3 ) ,

the convolution of Ul and U2 (with respect to the multiplication.), such that


(Vl Q9 el) *. (V2 Q9 e2) = (Vl * V2) Q9 (el • e2) , Vj E V'(JR n ) , ej E E j ,
provided Vl or V2 has compact support. Of course, * is the 'scalar' convolution and
(vQgej)(<p) = (v,<p)ej for (v,ej) E V'(JRn) x E j and <p E V(JR n ). Moreover, letting

J j E {V, V',£,£',5,5', OM, BG, BUG, Go, Lp ; 1::::; p < oo} ,


it follows that convolution with respect to • is a bilinear map
Jl(JRn,El ) x J 2 (JR n ,E2 ) --+ J 3 (JR n ,E3 ) ,

having the same continuity properties as the maps (4.2.11)-(4.2.14) and (4.2.18)-
(4.2.22), respectively, provided the triple (J l ,J 2 ,J 3 ) has the same meaning as
I1I.4 Maximal Sobolev Regularity 135

in the corresponding formulas (thus (~1'~2'~3) = (V',V,[) in the analogue to


(4.2.11) etc.). This follows from the general theory of vector-valued distributions
for which we refer to [Schw57b] and [Schw57a]. Since we hardly have occasions to
use this general setting we do not go into details .•

4.3 The Hilbert 'fransform


Unless explicitly stated otherwise, we assume from now on that n = 1 and sim-
ply write ~ for ~(lR?), where ~ stands for anyone of the spaces of distributions
introduced so far. Thus V := V(lR?), S':= S'(lR?), etc. Of course, E := (E, 1·1) is a
Banach space.
Given c > 0, define (llt)E E Ll,loc by

T E lR?.

Then
((l/ t )E,<P? = 1 <p(T)dTIT,
ITI::>.':
<p E S .

It is well-known and easily seen that there exists a unique temperate distribu-
tion PV(llt), the principal value of lit, such that

<pES. (4.3.1)

Given U E S(lR?, E), we define the Hilbert transform Hu of u by

Hu := HEu := 1T- 1 PV(l/t) *u ,

and the truncated Hilbert transform, HE, C > 0, by


c > 0,

so that

t E lR?, u E S(lR?, E) .

The following proposition shows that HE is an approximation of H.

4.3.1 Proposition If U E S(lR?, E) then HEu --+ Hu in OM(lR?, E) as c --+ O.

Proof By (4.3.1) we know that (lit).': --+ PV(1/t) in the w*-topology of st. Since
S is a Montel space, (llt)E --+ PV(l/t) in S' as c --+ 0, by the Banach-Steinhaus
theorem. Now the assertion follows from (4.2.14) .•
136 III Maximal Regularity

Since OM (IR, E) '-t [(IR, E), this proposition justifies the notation

Hu(t) := HEU(t) = -1 /00 u(t - r) dr/r = -1 /00 -u(r)


- dr, t E IR ,
1f -00 1f -00 t - r
for the Hilbert transform of u E S(IR, E), where the integral is a principal value
integral.

Now we show that H extends to a bounded linear operator on L2(1R, E),


provided E is a Hilbert space. For this we need the following Fourier transform:

4.3.2 Lemma (PV(l/t))~ = -i1fsign.


Proof Given 0 < E < R < 00, let
(l/t)~(r) := r-L~[c~ITI~RJ(r) , r E IR.

Then (l/t)~ E L1 '-t S', and

'PES.

Thus
(l/t)~ ----> (l/t)c in S~. as R---->oo. (4.3.2)
Note that

for S E R Hence

1
((l/t)~r ----> -2i sign(-) 00 Si~~ d~ ,
cl·1 ..,
R---->oo ,

in L1,loc '-t S' '-t Sw" Since, trivially, F E £(S~.), we deduce from (4.3.2) that

(l/t);:= ((l/t)c)~ = -2i sign(.) 100 si~~ d~


cl·1 ..,
Hence, given 'P E S,

((l/t);, 'Pl----> -i1f(sign, 'P) , E ----> 0,

that is, (l/t); ----> -i1f sign in S:V. as E ----> O. Since (l/t)e ----> PV(l/t) in S' by the
proof of Proposition 4.3.1 and (4.2.3), the assertion follows .•
1II.4 Maximal Sobolev Regularity 137

Suppose that, given p E [1,00),

IIHuIILp(IR,E) ::; c IluIILp(IR,E) , u E S(JR, E) .

Then, since S(JR, E) is dense in Lp(JR,E), there exists a unique H E C(Lp(JR, E»)
extending H. Of course, H is again denoted by H and said to be the Hilbert
transform on Lp(JR, E).

4.3.3 Theorem Let E be a Hilbert space. Then the Hilbert transform is a unitary
skew-adjoint linear operator on L2 (JR, E).

Proof Given u E S(JR, E), the convolution theorem and Lemma 4.3.2 imply

Hu = -i sign(·)u . (4.3.3)

Thus, by Plancherel's theorem,


(HulHvh = (271")-1(HuIHvh = (271")-1(ulvh = (ulvh , u,v E S(JR, E) .
From this and the density of S (JR, E) in L2 (JR, E) we see that H E Caut (L2 (JR, E»)
and H- 1v = -Hv for v E L 2(JR, E). Now the assertion is obvious .•

Now we show that, given any Banach space E, the Hilbert transform is
bounded on Lp (JR, E) for 1 < p < 00, provided this is true for at least one p E (1, 00 ).
For this we need some preparations.

4.3.4 Lemma Let E be a Banach space. Then Hw E L1 ([ It I > 2TJ, E) and

1 Itl>2T
IHwl dt ::; (2/71") II w l1 1

for T > 0 and wELl (JR, E) satisfying Jw dt = 0 and supp(w) C [-T, 1'] .

Proof Let T > 0 and suppose that v E V(JR, E) with Jv dt = 0 and


supp(v) c [-T - 6, T + 6]
for some 6 E (0, T /2). Given € E (0, T /2) it follows that
138 III Maximal Regularity

It follows from Proposition 4.3.1 and (4.1. 7) that

as 10---+0, rpES.

If supp( rp) C [It I ::0: 2T], similarly as above,

Hev(rp) = ~
7r
r
Jtl>2T
1
rp(t) 1=
-ox;
V(T)[(t - T)-1 - c 1] dTdt (4.3.4)

for 0 < 10 < T/2, that is, Hev(rp) is independent of 10 E (0,T/2) for rp E D with
support in the set [It I ::0: 2T]. Hence

Hevl[ It I > 2T] = Hvl[ It I > 2T] . (4.3.5)

This shows that Hv E L1 ([ It I > 2T], E) and that

1 Itl>2T
2T+8
IHvl dt ::; --;y<:
7r - u
IIvl1 1 (4.3.6)

Let {rpe ; 10 > O} be a mollifier. Then rpc * w E D(JR, E) with support in the inter-
val [-T - 10, T + 10], and, by Fubini's theorem,

1rpc * wdt = 11 rpe(t - T)W(T) dTdt = 1 wdT IIrpEI11 = 0.

Since rpe * W ---+ W in L1 (JR, E), it follows from (4.3.4) and (4.3.5) that (4.3.6) is
true with v replaced by w. Since 8 E (0, T /2) was arbitrary, the assertion follows .•

In the next lemma we recall the CALDERON-ZYGMUND DECOMPOSITION THE-


OREM, that we state in greater generality than presently needed.

4.3.5 Lemma Let E be a Banach space and let u E L1 (JR n , E) and a > 0 be
given. Then there exist v, w k E L1 (JR n, E) for kEN such that
(4.3.7)
and such that
(4.3.8)

and
(4.3.9)
Furthermore, there exists a sequence of pair-wise disjoint cubes Qk, with sides
parallel to the coordinate-hyperplanes, such that
(4.3.10)
and
aLAn(Qk) ::; IIul11 , (4.3.11)
where AnU denotes the n-dimensional Lebesgue measure.
III.4 Maximal Sobolev Regularity 139

Proof The 'scalar proof' given in [Har83, Lemma 4.5.5]' for example, applies
literally to the E-valued case (also see [Tri78, Lemma 2.2.2]) .•

After these preparations we can prove the following important theorem con-
cerning the boundedness of the Hilbert transform.

4.3.6 Theorem Let E be a Banach space and suppose that the Hilbert transform
is bounded on Lp(IR,E) for somep E (1,00). Then it is bounded on Lq(IR,E) and
on Lq(IR,E') for each q E (1,00).

Proof (i) First we show that H is of 'weak type' (1,00) in the sense that

(4.3.12)

By assumption there exist p E (1,00) and f.1 > 0 such that


(4.3.13)

Fix u E S(IR, E). Then, given u > 0, we have the decomposition of Lemma 4.3.5
with a:= u/f.1. Hence, thanks to (4.3.8),

so that

(4.3.14)

Let Q'k be the 'double cube' (that is, interval) with the same center as Qk and
twice its length, and let V := U Q'k. Then, thanks to (4.3.11)

(4.3.15)

Note that H is translation invariant. Thus

UA1 (vc n [LIHwkl > u]) :::; Ivc LIHwkl dt


=L r
In(Q~)C
IHwkldt:::;L r
J1tl?:.2Tk
IHwi.:ldt

for suitable Tk > 0, where wi.: E L1 (IR, E), has its support in [- Tk,Tk], and satisfies
J wi. dt = O. Now we deduce from Lemma 4.3.4 that

(4.3.16)
Lastly, observe that by (4.3.7)

[IHul > u] c [IHvl > u/2] U (Ve n [LIHwkl > u/2]) U V .


140 III Maximal Regularity

Thus we deduce from (4.3.14)-(4.3.16) that


O"A1 ([ IHul > 0"])
~ 2{ (0" /2)A1 ([ IHvl > 0"/2]) + (0" /2)A1 (VC n [~]Hwkl > 0"/2]) } + O"A1 (V)
~ 11 2P Ilvil I + (4/lr)L:ll w kIII + 21111 ull I .
Now (4.3.9) and the density of S(JR., E) in LI(JR., E) imply the validity of (4.3.12).
(ii) From (4.3.12), (4.3.13), and the vector-valued version of the interpolation
theorem of Marcinkiewicz ([Tri78, Theorem 1.18.3]) we see that HE E £(Lq(JR., E))
for 1 < q < p.
(iii) Now let E := K Then Theorem 4.3.3 implies Hoc E £(L2)' Consequently,
we deduce from (ii) that Hoc E £(Lq), 1 < q ~ 2. Theorem 4.3.3 also implies
that (Hoc)' = -Hoc E £(L2)' Hence we deduce from (ii) that (Hoc)' E £(Lq,) for
1 < q' ~ 2, since it is trivially verified that the dual Hoc of Hoc on L q , is again
given by -Hoc. Thus, by the reflexity of Lq for 1 < q < 00, Hoc = (Hoc)' E £(Lq)
for 2 ~ q < 00. That is, Hoc E £(Lq) for 1 < q < 00.
(iv) Suppose that 1 < r < 00. Then we claim that
HE E £(Lr(lR, E)) ¢=? HE' E £(Lr,(JR.,E')) . (4.3.17)
For this we reeall first that
Lr' (JR., E') is (identified with) a closed linear subspace
of Lr(lR, E)' such that

(v',Uhr(JR,E) = k \V'(t),u(t);dt
(4.3.18)

for v' E Lr' (JR., E') and u E Lr(lR, E).


Next, given rp E Lr and e E E, put rp @ e := rpe. Moreover,
Lr @ E := { 'L";=orpj @ ej ; rpj E Ln ej E E, mEN} .
Then
(4.3.19)
thanks to the fact that the simple functions are dense in Lr (JR., E) and contained
in L r @ E. Observe that (iii) and the obvious fact that HE(rp @ e) = (Hocrp) @ e
imply
(4.3.20)
Since
\ rp' @ e', HE(rp @ e); Lr(JR,E) = (rp', Hocrp) Lr (e', e) E
for (rp', rp) E Lr' x Lr and (e', e) E E' x E, it follows from (iii) that
(v', HEuhr(JR,E) = -(HE,v', Uhr(JR,E) (4.3.21)
for u E L r @ E and v' E Lr , @ E' .
I1I.4 Maximal Sobolev Regularity 141

Suppose that HE E C(Lr(lR, E)). Then (4.3.19) and (4.3.21) imply

(HE)'::::J -HEII(Lrl 0 E') .


Thus, using (4.3.18)-(4.3.20) and the fact that (HE)' E C(Lr(lR, En, we see that

(4.3.22)

Conversely, let (4.3.22) be satisfied. Then an analogous argument, combined with


the fact that E is a closed linear subspace of E", proves that HE E C(LrOR,E)).
(v) From (ii) and (iv) it follows that HE' is bounded on Lql (JR, E') for
1 < q' :::; p'. Hence (iv) implies HE E C(Lq(JR, E)) for p:::; q < 00. This proves the
theorem .•

The proofs of this section are adaptions of well-known arguments from har-
monic analysis used in the study of the scalar Hilbert transform (e.g., [Ste70a]).

4.4 UMD Spaces and Fourier Multipliers


Let E := (E, 1·1) be a Banach space. Then, in general, the Hilbert transform is
not bounded on Lp(JR, E) for any p E (1,00). This justifies the following definition:
a Banach space E is a UMD space if the Hilbert transform is bounded on Lp(JR, E)
for some p E (1,00).

4.4.1 Theorem Let E be a UMD space. Then the Hilbert transform is bounded
on Lp(JR, E) for 1 < p < 00.
Proof This is a consequence of Theorem 4.3.6 .•

4.4.2 Remarks (a) Let (fl, A, P) be a probability space and let (Ak) be an
increasing sequence of sub-a-fields of A. A sequence (Uk) of E-valued P-integrable
random variables on (fl, A, P) is said to be a martingale on E if

i Uk dP = i Uj dP , AEAk, O:::;k<j<oo,

that is, if
O:::;k<j<oo,
where 1E(-IAk) is the conditional expectation.
The Banach space E is said to possess the property of unconditionality of
martingale differences (UMD) if for some p E (1, 00 )
142 III Maximal Regularity

for all n EN, Ek E {±1}, and all E-valued martingales (Uk), where U-l := O. It
has been shown by Bourgain [Bour83] that E possesses the property UMD if E is
a UMD space. Conversely, Burkholder [Bur83] proved that the validity of property
UMD implies that E is a UMD space. This explains the name.
(b) A Banach space E is (-convex if there exists a symmetric, biconvex func-
tion ( : E x E --+ JR satisfying ((0,0) > 0 and ((x, y) :::; Ilx + yll for x, y E E with
Ilxll = Ilyll = 1. Burkholder [Bur81] proved that a Banach space is (-convex iff it
is a UMD space.
(c) UMD spaces are reflexive. The converse is not true (e.g., [Rub86]) .•

In order to establish some important properties of UMD spaces we need a


few basic facts about Fourier multipliers.
Let m E OM(JR n ) be given. It follows from (4.1.8), the convolution theorem,
(4.2.1), and (4.2.3) that, given a Banach space E,

(4.4.1)

We also put

and say that m(D) is a translation invariant pseudodifferential operator with


symbol m. This is justified by (4.2.4) and (4.2.15).

Now one can ask what conditions have to be imposed on m so that (4.4.1) is
true if S(JR n , E) is replaced by Lp(JR n , E) or some other Banach space, ~(JRn, E),
satisfying
(4.4.2)
If E = ][{, the famous Mikhlin multiplier theorem gives a satisfactory answer for
the Lp-case. It is interesting and of great importance that this theorem essentially
carries over to the E-valued case, provided E is a UMD space.
To be more precise, let ~(JRn, E) be a Banach space satisfying (4.4.2). Then an
element m E LCXl(JRn ) is said to be a Fourier multiplier for ~(JRn, E) (an ~(JRn, E)-
multiplier), if
(4.4.3)
This means that mu E S'(JR n , E) for U E ~(JRn, E) so that F- 1 (mu) is well-defined,
F-1(mu) E ~(JRn,E), and there exists c > 0 such that

(4.4.4 )

for U E ~(JRn, E). If S(JR n , E) is dense in ~(JRn, E), it suffices, of course, to verify
(4.4.4) for all U E S(JR n , E). (Note that F-1m * u is well-defined for m E LCXl(JR n )
and U E S(JR,E). In the general case the middle term in (4.4.3) can be taken to
define F-1m * .)
I1I.4 Maximal Sobolev Regularity 143

The space of all Lp (JRn , E)-multipliers is denoted by

It is a Banach algebra with respect to the norm

We write MM := MM(JRn ) for the vector space of all mE Loo(JRn ) such that
the distributional derivatives of order a, where a :S (1, ... ,1) E N n , are represented
on IR n by functions and such that

(4.4.5)

Then
(4.4.6)

is a Banach space and a continuous multiplication algebra, that is, point-wise


multiplication
MM X MM ---->MM, (ml,m2) f-+ mlm2
is well-defined and continuous.

After these preparations we can formulate the following vector-valued exten-


sion of the MIKHLIN MULTIPLIER THEOREM.

4.4.3 Theorem Let E be a UMD space. Then

l<p<oo,

and this injection is an algebra homomorphism.

Proof For the first assertion we refer to [Zim89], the second one is obvious .•

It should be remarked that an earlier, somewhat weaker result has been


given in [McC84].

4.4.4 Remark If E = te, Hormander's extension of Mikhlin's theorem guarantees


that MM '---+ Mp := Mp(C) for 1 < p < 00, provided the norm (4.4.5) is replaced by

where [s] denotes the largest integer not exceeding s E R •


144 III Maximal Regularity

4.5 Properties of UMD Spaces


As a first application of the generalized Mikhlin multiplier theorem we show that,
given u E Lp (JRn , E), the truncated Hilbert transform Heu converges in Lp(JRn , E),
and almost everywhere, towards H u if E is a UMD space.

4.5.1 Theorem Suppose that E is a UMD space, 1 < p < 00, and u E Lp(JR, E).
Then Heu ----+ Hu in Lp(JR, E) and almost everywhere as € ----+ O.

Proof Let 'IjJ(~) := e-I~I for ~ E lR. Then

cp(x) := F-1'IjJ(X) = ~ roo e(ix-1)~ d~ + ~ e(ix+1)~ d~


27l' Jo 27l'
1 0

-00

1[ 1 1] 1 1
= 27l' 1 - ix + 1 + ix =;;: 1 + x 2
for x E lR. Hence cP E £1 (JR) and J cp dx = <p(0) = 'IjJ(0) = 1. Thus {CPe; € > O} is
an approximate identity and it follows from (4.2.23), (4.2.24), (4.2.26), and (4.2.27)
and the fact that Hu E Lp(JR, E) that

CPe * Hu ----+ Hu in Lp(JR, E) and a.e. (4.5.1)


as O.
€ ----+

Let
me := _ 2i sign(·)
7l'
1 el.1
00
sin~ d~ ,
~
€ > O.
It is easily verified that me E MM and that IlmellMM ::; c for € > O. Hence, since
by the proof of Lemma 4.3.2 and the convolution theorem, Heu = me(D)u for
€ > 0, Theorem 4.4.3 implies

€ > O. (4.5.2)

Since CPe = C 1 ag-lcp, we infer from (4.2.9) that C;;; = ae<p = ae'IjJ. Note that
'IjJ and m:= -i sign(·) belong to MM. Hence Theorem 4.4.3, the convolution the-
orem, and Lemma 4.3.2 imply

CPe * Hu = F- 1((ae'IjJ)m)Fu = ;:-1 ((ae'IjJ)m) *u , € > O.


Observe that (ae'IjJ)m = ae('ljJm) so that (4.2.9) gives
F-1((a e'IjJ)m) =€-l a g- 1 F- 1('ljJm) =Xe,
where

X(x) := ;:-1 ('ljJm)(x) = -i


27l' Jo
roo
e-W-ix) ~ + ~ 0 eW +ix ) d~
27l' -00
1
i [ -1 1] 1 x
= 27l' 1 - ix + 1 + ix =;;: 1 + x 2
IlI.4 Maximal Sobolev Regularity 145

for x E lR. Consequently,

CPE * H u = XE * U , c > o.
Lastly, let
Ixl > 1,
Ixl < 1.
Then k E £1 (lR) and J k dx = 0 since k is odd. Moreover, Ik(x) I :::; c(l + Ixl)2 for
x E lR. Hence kE * U -> 0 in Lp(JR, E) and a.e. as c -> 0, thanks to (4.2.23), (4.2.24),
(4.2.26), and (4.2.27). Thus

in Lp(JR,E) and a.e. as c -> 0 which, together with (4.5.1) and (4.5.2), implies the
assertion. _

Theorem 4.5.1 was first proven by Burkholder [Bur83]. Our proof has been
motivated by the 'classical' proof of M. Riesz [Rie27] based on conjugate functions
(also cf. [Tit48] and [Pru88]).

In the next theorem we collect the most important properties of UMD spaces.
It follows from this result that most of the reflexive Banach spaces of distributions
commonly used in the theory of partial differential equations are UMD spaces.

4.5.2 Theorem
(i) Every Banach space isomorphic to a UMD space is a UMD space.
(ii) Every Hilbert space is a UMD space.
(iii) Every finite-dimensional Banach space is a UMD space.
(iv) Finite products of UMD spaces are UMD spaces.
(v) If E is a UMD space, E' is one as well.
(vi) If E is a UMD space and (X, JL) is a l7-finite measure space, the Lebesgue
space Lp(X, JL; E) is a UMD space for 1 < p < 00.
(vii) If (Eo, E 1 ) is an interpolation couple of UMD spaces, the interpolation spaces
[Eo, Edo and (Eo, E1)o,p are UMD spaces for 0 < e < 1 and 1 < p < 00.
(viii) Closed linear subspaces of UMD spaces ar'e UMD spaces.
(ix) Quotients of UMD spaces modulo closed linear subspaces are UMD spaces.

Proof (i) Let E be a UMD space and let F be a Banach space such that there
exists T E Lis(E, F). Then, by 'point-wise multiplication',
T E Lis(Lp(JR,E), Lp(JR,F)) , l<p<oo,
and HF = THET~l. This proves the assertion.
146 III Maximal Regularity

(ii) follows from Theorem 4.3.3.


(iii) is a consequence of (i) and (ii).
(iv) Let E := IT E j . Then, using obvious notation, HE = IT HE)" Now the
assertion is evident.
(v) i:> implied by Theorem 4.3.6.
(vi) Given c > 0 and v E Lp(JR., Lp(X, f-l; E)), Fubini's theorem implies

IIHLp(X,/L;E),EVIILp(lR,Lp(X,/L;E)) = ~{
1
{ {I (
llR x Arl?E
v(x, t - r) dr/riP df-l dt} lip

=..!:.{{ (If v(x,t-r)dr/rIPdtdf-lfIP


1X llR lITI?E

{i k
7r

::; IIHE,EII.c(Lp(lR,E)) Iv(x, t)IP dtdf-l} lip

= IIHE,E 11.c(Lp(lR,E)) Ilvll Lp(lR,Lp(X,/L;E))


Let (Cj) be a positive null sequence. Since, thanks to (4.5.2),

IIHE,Ej 11.c(Lp(lR,E)) ::; c , JEN,

it follows that, letting F := Lp(X, f-l; E),

j EN. (4.5.3)

Thus, given 'P E S(JR., F), the sequence (HF,Ej'P) is bounded in Lp(JR., F). Note
that 1 < p < 00 and Remark 4.4.2(c) imply that F is reflexive. Hence Lp(JR., F)
is also reflexive. Consequently, (HF,E j 'P) is weakly relatively compact. Thus there
exists W E Lp(JR., F) and a subsequence converging weakly towards w. By the con-
tinuous injection of Lp(JR., F) in S' (JR., F) it also converges weakly in S' (JR., F)
towards w. On the other hand, we infer from Proposition 4.3.1 and (4.1.7) that
the sequence (HF,Ej'P) converges strongly, hence weakly, in S'(JR., F) towards HF'P.
Thus HF'P = wand (4.5.3) implies

'P E S(JR., E) .

Now we infer from the density of S(JR., E) in Lp(JR., E) that

(vii) Recall that

(4.5.4)

and
(4.5.5)
III.4 Maximal Sobolev Regularity 147

for °< e < 1 and 1 < p < 00 (e.g., ['lIi78, Theorem 1.18.4]). Thus, if

1<p < 00, j = 0,1 ,


it follows that

l<p<oo,
where Eo equals either [Eo, Elle or (Eo, E1)e,p.
(viii) Let F be a closed linear subspace of a UMD space E. Given
m
U = L 'Pj 0 ej E Lp 0 F ,
j=O

it follows from (iii) and (vi) that


m
HEu = L(HIK'Pj) 0 ej E Lp 0 F ,
j=O

that is,
HE(Lp 0 F) c Lp @ Fe Lp(JR, F) .
Now the assertion follows from the density of Lp @ F in Lp(JR, F), the closedness
of Lp(JR, F) in Lp(JR, E), and the continuity of HE on Lp(JR, E).
(ix) Let F be a closed linear subspace of a UMD space. Then (E / F)' is
toplinearly isomorphic to Fl. := { x' E E' ; (;Y;', x) = 0, x E F}. Since Fl. is a
closed linear subspace of E', it follows from (viii) that (ElF)' is a UMD space.
Now the assertion is a consequence of (v) and Remark 4.4.2(c) .•

It should be remarked that the proof of the preceding Theorem 4.5.2 follows
essentially [Rub86].

4.6 Fractional Powers


Let E be a Banach space. A linear operator A in E is said to be of positive type K
if it is closed, densely defined, JR+ C p( - A), and

(1 + s) II(s + A)~lll :s; K ., (4.6.1)

where K 2': 1. We denote by


PK:= PK(E)
the set of all operators of positive type K, and A is of positive type if it belongs to

P := P(E):= U PK(E) .
K:2:1
148 III Maximal Regularity

Throughout the remainder of this section we assume that A E PK for some K :::: 1.

Given 8 E JR+ and ,\ E C satisfying

1'\-81::; (1+8)/(2K) , (4.6.2)

it follows from'\ + A = (8 + A)(l + (,\ - 8)(8 + A)-I) that ,\ E p( -A) and

II('\+A)-Ill::; 11[1+('\-8)(8+A)-lr I IIII(8+A)-111 ::;2K(1+8)-1


(4.6.3)
2K 1 + 8 + 1,\ - 81 2K ( 1) 2K + 1
::; 1 + 1,\1 1+ 8 ::; 1 + 1,\1 1 + 2K = 1 + 1,\1 .

From this we deduce that

S(K):= [Iargzl::; arcsin1/(2K)] U [Izl::; 1/(2K)] c p(-A) (4.6.4)

and that
(1 + 1'\1) 11('\ + A)-III::; 2K + 1 , ,\ E S(K) . (4.6.5)

°
We denote by <P the set of holomorphic functions 'P: C\ (-JR+)
that there exists {j > (depending on 'P) with
--+ C such

as 1,\1 --+ 00 , (4.6.6)

uniformly in [I arg,\1 ::; 7f - E] for each E E (0, 7f). It is obvious that <P is a commu-
tative algebra without unit with respect to point-wise multiplication. Given'P E <P,
we put

Jrr 'P( -,\)(,\ + A)-I d'\ = ~ r 'P('\)('\ -


1.
'P(A) := -2 A)-I d'\ , (4.6.7)
7fZ 27fz Lr

where f is any piece-wise smooth simple curve in S(K)\JR+ running from ooCi'/J
to ooei'/J for some {) E (0, arcsin 1/(2K)]. Of course, -f:= [-,\ E fl. It follows
from (4.6.4)-(4.6.6) and Cauchy's theorem that 'P(A) is well-defined in .c(E) and
independent of the particular choice of f. In fact, more is true.

4.6.1 Lemma The map <P --+ .c(E), 'P 1-+ 'P(A) i8 an algebra homomorphism.

Proof The linearity is obvious. Given 'PI, 'P2 E <P, choose admissible contours fl
IIl.4 Maximal Sobolev Regularity 149

and r2 such that r 1 lies to the left of r 2 . Then

The map >. 1-+ (>' - 1-£) -1 'PI ( - >.) is for each 1-£ E r 2 holomorphic on r I and to the
left of it. Hence it follows from (4.6.6) and Cauchy's theorem that the integral in the
first pair of parentheses equals zero. Similarly, Cauchy's integral formula implies
that the integral in the second pair of parentheses equals 'P2( ->.). Consequently,

which proves the assertion .•

Given Z E C, put

where the principal branch of the logarithm is used. Note that 'Pz E <I> for Re z < O.
If kEN then 'P-k is holomorphic in t
and we can deform r to the positively
oriented circle centered at 0 ofradius r, where 0 < r < liliA-III. Since
00

(>. - A)-l = -A-I(l- >'A-I)-I = -l:A-j-l>.j , 1>'1 = r ,


j=O

it follows that

This justifies the following definition of the fractional powers of A:

Rez < 0,

(cf. (1.2.9.4)).
150 III Maximal Regularity

Suppose that 0 < Re z < 1. Then we can contract r to jR+. Hence

that is,
A- z = -sinrrz
rr
-
0
00
1
s -Z( s+ A)-l d s, 0< Rez < 1 . (4.6.9)

Applying formula (4.6.9) to the case E := C and A := I, in particular, it follows that

1 00
s-Z(I+s)-lds=rr/sinrrz, 0< Rez < 1 . (4.6.10)

Hence we deduce from (4.6.9) and from A E PK that

IIA-ZII <:::Klsinrrzl roo s-Rez(l+s)-lds=K .Isinrrzl (4.6.11)


rr Jo sm(rrRez)
for 0 < Re z < 1. Now it is not difficult to prove the following continuity result:

4.6.2 Theorem {AZ; Rez < O} U {AO = IE} is a strongly continuous holo-
morphic semigroup on E.

Proof It is an easy consequence of the theorem on the differentiation of param-


eter integrals that the map z f-+ A Z is holomorphic on [Re z < 0]. Thus, thanks to
Lemma 4.6.1, it remains to show that it is strongly continuous at z = O.
Note that

for s > o. Thus, given x E dom(A) and z with 0 < Rez < I, it follows from (4.6.9)
and (4.6.10) that
sm rrz
A-Zx - x = - - . 1 00
s-Z(s + A)-lXds - sm
-• -
rrz 1° 00
s-Z(1 + s)-lxds

= --
rr
sin rrz
rr
1 -(--)
0

0
00
s-Z
l+s
(s
rr
+ A)-l(l- A)xds.
Consequently,

O<Rez<l.

Since the integral converges towards 1 as Re z -+ 0, we see that A -Zx -+ x as


z -+ 0 in [I arg zl <::: 0] for each 0 E (0, rr /2). Since A-z is uniformly bounded for z
III.4 Maximal Sobolev Regularity 151

in [I arg zl :::; a] n [0 < Re z < 1] for each a E (0, 7r 12), thanks to (4.6.11), it follows
that AZ --+1 in .Ls(E) as z --+ 0 in [Iargzl:::: 7r/2+c] for each c E (0,7r/2). This
proves the theorem. _

Suppose that AZx = 0 for some x E E and z E C with Rez < o. Then it
follows that AZ+wx = AW AZx = 0 for Rew < o. Thus AW x = 0 for Rew < Rez. In
particular, A -kx = 0 for kEN with k > - Re z. Consequently, x = O. This shows
that AZ is injective for Re z < O. Hence we can define the fractional powers for
Rez > 0 by
Rez > o. (4.6.12)
Clearly, AZ E C(E). Given z,w E C with 0 < Rez < Rew and x E dom(AW), it
follows from

that x E dom(AZ), that is,


dom(AW) c dom(A Z) , 0< Rez < Rew . (4.6.13)
Given x E dom(A), put y := Ax. Since dom(A) is dense in E, we can find
for each c > 0 an element u E dom(A) such that Ilu - yll :::; ciliA-III. Thus, let-
ting v:= Au,
IIA- 2 v - xii = IIA-Iu - A-Iyll :::; IIA-Ililiu - yll :::; c .
This shows that dom(A2) => dom(A). Hence dom(A2) => dom(A) = E which guar-
antees that dom(A 2 ) is dense in E. By induction we see that dom(A k ) is dense
in E for each kEN. Hence we infer from (4.6.13) that
dom(Az) = E , Rez > 0 . (4.6.14)
Now suppose that Rez > 0 and Rew > O. Given
x E dom(Az+W) c dom(A W) n dom(A Z) ,
put y := AZ+'wx. Then x = A-(z+w)y = A-W A-Zy implies AW x = A-Zy which, in
turn, shows that y = AZ AW x , that is,
(4.6.15)
If Rez > Rew and x E dom(AW) then
A-z AWx = A-(z-w) A-w AWx = A-(z-w)x = AW-z x .

Moreover, if x E dom(AZ) then, thanks to (4.6.15),

This proves that, given z, wEe with Rez, Rew, Re(z + w) #- 0,


(4.6.16)
where U E {z, w, z + w} with Reu = max{Rez, Rew, Re(z + w)}.
152 III Maximal Regularity

In order to be able to define A Z for Re z = 0 we need the following extension


of (4.6.9). As usual, the empty product is given the value 1.

4.6.3 Proposition Suppose that mEN. Then

.
A-z=SlO'lfZ
'If
'1
m.
(1 - z)(2 - z) ... (m - z) 0
00 sm-z(s + A)-m-1ds (4.6.17)

for 0 < Re z < m + 1.


Proof Suppose z satisfies 0 < Re z < 1. Then we infer from (4.6.9) by integration
by parts that

Now (4.6.17) follows by induction for 0 < Rez < 1. Thanks to (4.6.1) it is easily
verified that the integral in (4.6.17) converges absolutely for 0 < Re z < m + 1 and
that the right-hand side of (4.6.17) is a holomorphic map from [0 < Rez < m + 1]
into £(E). Now the assertion follows from Theorem 4.6.2 and the identity theorem
for holomorphic functions. •

It is a consequence of Theorem 4.6.2 that {A -t ; t ~ O} is a strongly con-


tinuous semigroup on E. We denote its infinitesimal generator by

-logA

which defines the logarithm of A E P(E). Then the intuitive formula

A -t _-e -tiogA , t ~ 0,

is valid. This also shows that the definition is consistent with the one for A E £(E)
obtained by the Dunford calculus.
Recall that an analytic semigroup {e- tB ; t ~ O} is said to be of angle a,
where 0 < a ~ 'If, if there exists a holomorphic function

T: [I argzl < a] --> £(E)

extending {e- tB ; t ~ O}, such that T is strongly continuous on the closed sec-
tor [I argzl ~ a - e] U {O} for each e E (0, a). Then T is a semigroup on E and
we put e- zB := T(z) for z E [I arg zl < a] U {O}. This notation is justified since
8T(z) = -BT(z) for largzl < a.
111.4 Maximal Sobolev Regularity 153

4.6.4 Theorem Suppose A E PdE). Then {A-t ; t ~ O} is an analytic semi-


group of angle 7r /2. Moreover,

O~t~m, mEN,
and A-z = e-zlogA for Rez < O.
Proof Thanks to Theorem 4.6.2 it remains to prove the asserted bound. By
applying Proposition 4.6.3 to E := C and A := 1 we see that

7r(I- t)(2 -~) ... (m - t) = roc sm-t(l + s)-m-l ds > 0


m! sm trt 10
for 0 < s < m + 1 and mEN. Now Proposition 4.6.3 and (4.6.1) imply

-
sin trt
IIA-tll < Km+1__
7r
m'.
(1 - t)(2 - t) ... (m - t)
10
00
sm-t(1 + s)-m-l dt = K m+1

for 0 < t < m + 1 and mEN.•


Now suppose that -1 < Re z < 1. Then we put

x E dom(A) .

Observe that

Aox = 1 00
(s + A)-2 ds Ax = -(8 + A)-l Axl~ = x, x E dom(A) . (4.6.18)

Moreover, if Rez -=f 0, it follows from (4.6.9) and (4.6.17) that

A z x -- Az-1A x -_ sin 7r(1 - z)


7rZ
1 0
00
s Z( s + A)-2Ax ds -- A zX (4.6.19)

for x E dom(A). Note that

A- 1 Az c sin -
B z := - 7rZ
7rZ
1
0
00
8Z (S + A)- 2 ds E £(E) . (4.6.20)

Let (Xj) be a sequence in dom(A) such that Xj ----> 0 and Azxj ----> yin E. Then,
thanks to (4.6.20), Bzxj ----> 0 and Bzxj ----> A-ly, which implies y = O. Hence A z is
closable. Motivated by (4.6.18) and (4.6.19) we put

A Z := closure of Az , Rez = O. (4.6.21)

By these considerations we have already proven most of the following statement.


154 III Maximal Regularity

4.6.5 Theorem Suppose that A E P(E). Then the fractional power AZ is


for each z E C a densely defined closed linear operator in E. If Re z < 0 then
AZ E £(E) and is given by the Dunford integral

AZ = ~
27rZ
r(_,\y(,\ + A)-l d'\ ,
lr
(4.6.22)

where r is any piece-wise smooth simple curve running in C\lR+ from ooe-i'P
to ooei,p for some cp,1j; E (0, 7r) such that 0"(- A) lies strictly to the left of r.
Moreover,
(i) AZ is the 'usual power' of A if z E Z.
Jo
(ii) AZx = si~;z oo SZ (s + A)-2 Ax ds, x E dom(A), -1 < Re z < 1.
(iii) Suppose that either mEN, x E dom(A2m), and Rez V Rew < m, or Rez,
Re w, and Re( z + w) are distinct from zero and x E dom( AU), where u belongs
to {z, w, z + w} and satisfies Re u = max{Re z, Re w, Re(z + w)}. Then

(iv) AZAw=Az+w, Rez,Rew>O.


d d
(v) D(AW) '----> D(AZ) '----> E, 0 < Rez < Rew.
(vi) AZ E £is(D(Az+W),D(AW)) n£is(D(AZ),E), Rez,Rew > O.
(vii) Given mEN, the map

is holomorphic.

Proof The first part of the assertion follows from the investigations preceding
this theorem.
(i) follows from (4.6.8) and (4.6.12).
(ii) If Rez -=F 0, this has been shown in (4.6.19), and it follows from the
definition of A z if Re z = O.
(iii) If Rez, Rew, and Re(z + w) are all distinct from zero, this is a conse-
quence of (4.6.16) and (4.6.13). From (ii) and (4.6.1) we infer that

(z f--> AZ) Eel ([-1 < Re z < 1], £(D(A), E) n £(D(A2), D(A))) . (4.6.23)

Thus suppose that z, w E [-1 < Re z < 1]. Choose sequences (Zj), (Wj) in

[-1 < Rez < l]\[Rez = 0] =: Z

such that Zj + Wj E Z and such that Zj ---> Z and Wj ---> w. Then, by what we
already know,

Hence, letting j ---> 00, we infer from (4.6.23) that (iii) is true if -1 < Re z, Re w < 1.
III.4 Maximal Sobolev Regularity 155

Suppose that Re z = 0 and I Re wi ~ 1. Fix DO E lR with 0 < a - Re w < 1. Then

for x E D(A2m) with mEN and Rew < m, since -1 < Re(w - a) < 0 and a -=I- O.
Lastly, let Re z -s: -1 < 1 -s: Re wand Re(z + w) = O. Write z = r + s with
-1 < Re r < O. Since the real parts of r, w, and r + ware distinct from zero and z,
r, and s have negative real parts, it follows that AZ = AT AS and AS AWx = AS+w x
for x E dom(AW). Thus

So we can assume that -1 < Rcz < O. Then Re(z + w) = 0 implies 0 < Rew < I,
so that we are back to a situation already considered. Consequently, (iii) has been
completely proven.
(iv) By Theorem 4.6.2 and (iii) it suffices to prove that x E dom(AW) and
AW x dom(AZ) imply x E dom(Aw+Z) ifRez > 0 and Rew > o. Let y:= AZ(AWx).
E
Then it follows from (iii) that x = A-W(A-Zy) = A-(w+z)y E dom(Aw+Z).
(v) From (4.6.13) and (iii) we deduce that

x E D(A11I) .

Since x 1-+ IIAuxl1 is an equivalent norm on D(AU) for Reu > 0, thanks to the
boundcdness of A-u, it follows that D(AW) '----+ D(AZ) '----+ E.
Given x E D(AZ), put y := AZx E E. Since D(AW-Z) is dense in E by (4.6.14),
given E > 0, we can find u E D(AW-Z) such that Ilu - yll < E. Hence

This shows that D(AW) is dense in D(AZ) which, together with (4.6.14), implies
the assertion.
(vi) The first assertion follows from (iv), the second one is trivial.
(vii) Thanks to Theorem 4.6.2 and (4.6.23), we can assume that m ~ 2. Since
(v) implies
£(D(A), E) '----+ £(D(Am), E) ,
we infer that

(4.6.24)

If 0 < Rez < m then (iii) implies AZx = Az-mAmx for x E D(Am). Hence Theo-
rem 4.6.2 guarantees that

This, together with (4.6.24), proves the theorem. _


156 III Maximal Regularity

Observe that Q(E, M, -0") C P(E) for 0" > 0, that is, if -A is the infinitesimal
generator of a strongly continuous exponentially decaying semigroup on E then
A is of positive type. In this case we can obtain another useful representation
formula for AZ with Re z > o.

4.6.6 Theorem Suppose that A E Q(E, M, -0") for some M ::::: 1 and 0" > o. Then

A- Z = _1_
f(z) 1 e-
0
00 1
e
-tAdt
, Rez > 0 .

Proof It is an easy consequence of

111= e-1e- tA dtIIC(E) ::; M 1= tRez-le-at dt


and the known properties of the Euler f-function that the map

[Rez > 0] -+ £(E) , z>---> -1-


r(z)
1=
0
e-1e- tA dt

is holomorphic. Thus, thanks to Theorem 4.6.2 and the identity theorem for holo-
morphic functions, it suffices to prove the asserted equality for 0 < z < 1.
Given z E (0,1),

A - z =sin
-7rZ
7r
- 1=
0
8 -z (s+ A)-l d s

by Proposition 4.6.3. On the other hand, we know from semigroup theory that

(8 + A)-l = 1= e-ste- tA dt , 8> O.

Thus, by Fubini's theorem,

A- z = - . 1= 1
SIn -
7rZ -z 00
e -st e -tA dt ds= -
sIn
• -
7rZ
1= 1= e -tA -tss
s -z e d t
d

1
8
7r 0 0 7r 0 0

=-sin-7rZ
f ( l - z)
00
e-Ie- t A dt.
7r 0

Hence the assertion follows from the well-known 'complementing' formula

f(z)f(l- z) = 7r/sin7rz (4.6.25)

(e.g., [Schw65, VIII,1;20]) .•

Now we assume that H is a Hilbert space and A is a positive definite


self-adjoint linear operator in H, that is, A = A* ::::: 0: > 0 for some 0: > O. Let
IlI.4 Maximal Sobolev Regularity 157

{E).. ; ,\ E lR} be the spectral resolution of A. Then, given Z E <C, we can de-
fine AZ by
Z E <C . (4.6.26)

The following theorem shows that this definition coincides with the former one.

4.6.7 Theorem Let H be a Hilbert space and let A be a positive definite self-
adjoint linear operator in H. Then A E P(H) and the fractional powers defined in
(4.6.26) by means of·the spectral resolution coincide with the fractional powers of
Theorem 4.6.5.

Proof First note that O'(-A) C (-00, -a] if A = A* 2: a > O. Moreover,


(s + a) llxl12 ::; (s + A)xlx) ::; II(s + A)xllllxll , x E dom(.1) ,
implies
!I(s + A)-Ill ::; (s + a)-I::; K(l + s)-1 , s 2: 0 .
Hence A E P(H).
Let r be the contour consisting of the two rays -(3 + lR+ e±i<f' for some
(3 E (0, a) and'P E (0, 7r), and oriented so that the imaginary parts increase along r.
Then, given Z E <C with Re Z < 0 and (1 2: 0:, Cauchy's integral formula implies
_1_ [ (-A)Z d)" = j-lz .
27ri ir ).. + j-l
Hence, by Fubini's theorem and the spectral calculus for A,
1 . {( _)..Y()..
-2 + A)-1 d)" = _.1.. [( _,\Y. [oo().. + (1)-1 dEp, d)"
7rt ir 27rt ir io
= roo [~ { (_,\)Z d)"] dEp, = [00 j-lZ dEp,
io 27rt ir Z + j-l io
in £(H), thanks to the fact that the support of the spectral resolution is contf),ined
in [a, (0). This proves the assertion for Re Z < O. Now the theorem follows from the
spectral calculus for self-adjoint linear operators and the definition of the fractional
powers for A E P(H) given above .•

4.6.8 Remarks (a) Observe that s(s + A)-l :J 1 - (s + .1)-1 A, and thus,
s-l(s+A)-l =s-lA- 1 _(s+A)-lA- 1 .

Thus, by splitting the integral in (4.6.9) at s = 1 and using these formulas, we

11
see that

- - - - -1- A-I x+
A z x- -Sin7rZ{X s z+l( s+ A)-lA- 1 x ds

+1
7r Z l+z 0
00 sz-l(s + A)-l Axds}
158 III Maximal Regularity

for x E dom(A) and -1 < Rez < O. By analytic continuation this representation
is also valid in [-1 < Rez < 1].
(b) Let 0 < 0: < 1. Given A E P(E), it is not difficult to see that 1R+ C p( -k') and

( Ct)-1 1 { 1 ( )-1
oX+A =2?riir oX +(-P,)CtP,+A dp"

where r is a contour as in (4.6.22). By contracting r to 1R+ it follows that

(oX + ACt)-1 = sin ?r0: roo s"(s + A)-1 ds (4.6.27)


?r io s2" + 2oXs" cOS?rO: + oX 2
for oX ~ o. Assume that
[I argzl < 19] U {O} C p(-A)
for some 19 E (O,?r) and that, given c: E (0,19), there exists Kg such that

1 argoXl ::; 19 - c:.


Then by means of the above formulas it can be shown that

[I arg zl < (1 - O:)?r + 0:19] U {O} C p( _A")

and that, given c: E (0, (1- O:)?r + 0:19), there exists Mg such that

This implies, in particular, that A" is the negative generator of a strongly contin-
uous analytic semigroup on E whenever A E Q(E, M, -0-) for some (J > o. Details
can be found in [Kre72] and [Tan79J.
(c) Let A E C(E) be densely defined and assume that lR+ C p( -A) and
s > 0, (4.6.28)

for some K ~ 1. Note that we no longer require that A has a bounded inverse.
Then, given 0: E (0,1), we put

A" := [I,,(oX)r 1 - oX , oX> 0 , (4.6.29)


where I,,(oX) denotes the right-hand side of (4.6.27). It can be shown that this
definition is meaningful and that A" is a closed and densely defined linear operator.
Put Ag := c: + A for c: E (0,1) and note that (4.6.28) implies 1R+ C p( -Ag) and

lI(s+AE )-111::; ~< K_1_, s ~ O.


s+c: c: 1+s
Hence AE is of positive type K / c and, consequently, the fractional power A~ is
well-defined. It can be shown that dom(A~) = dom(ACt) and that there exists a
III.4 Maximal Sobolev Regularity 159

constant c depending on K and Ct only such that

IIA~x - Acxxll :::; Cecx Ilxll , x E dom(A"') . (4.6.30)

Using these facts it is not difficult to prove that the fractional powers A'" enjoy
the following properties in this case in which 0 may belong to O'(A):
(i) if 0 < Ct < f3 :::; 1 then D(Af3) '---+ D(A"');
(ii) if Ct, f3 > 0 and Ct + f3 :::; 1 then x E D(A<>+f3) implies x E D(A'" A(1) and the
relation Acx Af3x = A<>+!1x .
For details we again refer to [Kre72] and [Tan79] .•

Let H be a Hilbert space and let A be a positive semidefinite self-adjoint


linear operator in H, that is, A = A* ::::: O. Then definition (4.6.26) is meaningful.
On the other hand, A obviously satisfies (4.6.28) with K = l. Hence Acx can also be
defined by (4.6.29). The following theorem shows that the two definitions coincide.

4.6.9 Theorem Let H be a Hilbert space and let A be a positive semidefinite


self-adjoint linear operator in H. Then the two definitions (4.6.26) and (4.6.29)
coincide for z = Ct E (0,1).

Proof Given e > 0, the operator As := e + A is positive definite and self-adjoint.


Hence, denoting by {E>JAc:) ; A E IR} the spectral resolution of A e , we infer from
Theorem 4.6.7 that

(Aco)" = 1= ACX dE.\(Ae) = 1 cxo


(p, + e)CX dE" , O<Ct<l,

where the last equality follows from the spectral calculus for self-adjoint operators.
Given x E dom(ACX), it follows from (4.6.30) that (As)OOx -+ AOOx in H as e -+ O.
On the other hand,

1 00
(p, + e)'" dE"x -+ 1 00
p,OO dE"x

in H as e -+ 0 by Lebesgue's theorem. This implies the assertion .•

We close this subsection by showing that we can 'raise powers to powers',


a result that is needed in Chapter V. For this we prove first that A E P(E) entails
A'" E P(E) for 0 < Ct < l. Indeed, we derive the following more precise statement:

4.6.10 Proposition Suppose that A E P(E) and Ct E (0,1). Then A'" E P(E). In
fact, if there exists {J E (0,7f) such that (4.6.1) is satisfied for sEC with I arg 81 :::; {J
then [I arg zl < 7f - (7f - {J)Ct] U {O} c p( -A"') and, given {J' E (0, {J),

(l+lsl)ll(s+A CX )-lll:::;c, largsl :::;7f-(7f-{J')Ct.


160 III Maximal Regularity

Proof Fix s E jR+ and define <p:= <Ps E q> by <p(A) := (s + ).',,)-1. Then

defines 'Ij; E q>. Since (A ~ A-,6) E q> for (3 > 0, it follows from Lemma 4.6.1 that

'Ij;(A) = <p(A)Aa- 1 = (1- s<p(A))A- I . (4.6.31)

Consequently, <p(A)(s + Aa)A -1 = A-I, that is,


<p(A)(s + A"')x =x, x E dom(A) . (4.6.32)

Given x E dom(A"'), by Theorem 4.6.5(v) and (vi) we find Xn E dom(A) such that
Xn ~ x and A"'x n ~ Aa x in E as n ~ 00. This shows that (4.6.32) is valid for
x E dom(A"'). On the other hand, by multiplying (4.6.31) by A-a we obtain

<p(A)A- 1 = A- a (1- s<p(A))A- 1 ,

hence <p(A) = A-"'(l- s<p(A)). This gives (s + Aa)<p(A) = 1, which proves that
jR+ c p( _A"') and (s + A",)-1 = <Ps(A) for s E jR+. By contracting the contour r
in (4.6.7) to jR+ it follows that

(4.6.33)

for s E jR+. Observe that the last integral is an analytic function of s in the sector
[I arg zl < 7r(1 - a)]. Hence this sector belongs to p( _Aa) and (s + Aa)-1 is given
by the integral above. Now suppose that s = lsi ei1/J with 1'Ij;1 < 7r - (7r - {j')a and
o < {j' < {). Then we can again deform the path of integration in the last integral
to the ray r ± := {re±il? ; r 2': O}, where we choose the positive [resp. negative]
sign if 'Ij; 2': 0 [resp. 'Ij; < 0], thanks to the estimate II(A + A)-III::; K(l + IAI)-1 for
Iarg AI ::; {j. Letting (3 := 'Ij; + (7r =F {j)a and ,:= 'Ij; - (7r ± {j)a, where we choose the
upper [resp. lower] sign if we integrate over r + [resp. r _], we obtain the estimate

II(s+Aa)-111 < Ksin7ra roc r"'dr .


- 7r io (1 + r) Ira + lsi e,,6llra + lsi e'll
K sin 7ra roo dT
::; j.;f~ io IT + ei ,6IIT + ei-rl .

This proves that (1 + lsI) II(s + A",)-111 ::; c(a,{j') forO < {j' < {j ands in the sector
[I arg zl ::; 7r - (7r - {jl)a] U {a}, hence the assertion. _
lIlA Maximal Sobolev Regularity 161

4.6.11 Corollary Suppose that A E P(E) and there exists iJ E (O,1r) such that
(4.6.1) is satisfied for sEC with 1arg sl $. iJ. If a. E (0,1) satisfies a. < 1r / [2(1r - iJ)]
then - A Cl genemtes a strongly continuous analytic semigroup on E.

4.6.12 Remark Corollary 4.6.11 and (4.6.5) imply that _ACl generates a strongly
continuous analytic semigroup on E whenever A E P(E) and 0 < a. $. 1/2. More-
over, if -A generates a strongly continuous semigroup on E, this is also true
for _A'" for each a. E (0,1) .•

It follows from Proposition 4.6.10 and Theorem 4.6.5 that the fractional pow-
ers (ACl)Z are well-defined for z E C and a. E (0,1). For simplicity, we restrict our-
selves in the following theorem to the case z E lR.

4.6.13 Theorem
for f3 E JR.
Suppose that A E P(E) and °< a. < 1. Then (ACl),6 = ACl,6

Proof Thanks to Proposition 4.6.10 we can find K 2:: 1 such that A and ACl
belong to PK(E). Then it follows from the proof of that proposition that

( + ACl)-l = _1_ { (,\ + A)-l d'\ fJ E S(K) ,


fJ 21ri lr
fJ + (_,\)Cl '

where r is a piece-wise smooth curve running in S(K)\JR+ from ooe- iiJ to ooeii1
for a sufficiently small positive iJ. Thus, by (4.6.22) and Cauchy's integral formula,

for f3 > 0, where r' is a contour with the same properties as r and lying to the right
of r. Moreover, (ACl),6 = [(ACl)-,6r 1 = [A-Cl,6]-l = ACl,6 for f3 > 0. This proves
the theorem .•

The proofs of this subsection follow essentially [Kre72] and [KraZ76]. By


means of Proposition 4.6.3 it is not difficult to derive a representation formula
similar to the one of Theorem 4.6.5(ii) for arbitrary z E C. This is done in [Tri78],
where such a formula is used to define AZ for A E P(E) without reference to the
Dunford integral representation (4.6.22). For a detailed study offractional powers
of operators that are not necessarily invertible, we refer to [Kom66]. The repre-
sentation formula of Theorem 4.6.6 is often used to define the fractional powers of
162 III Maximal Regularity

generators of strongly continuous semigroups (for real exponents) (e.g., [Gols85],


[Hen81]' [Paz83]).
Lemma 4.6.1 shows that one can define a good functional calculus for a class
of holomorphic functions that is larger than the class of powers. By mimicking
the definition of fractional powers for Re z 2': 0 one can also establish a reason-
able functional calculus for suitable unbounded holomorphic functions. This has
been elaborated by McIntosh [McI86] (also d. [deL87], [deL94] for holomorphic
functional calculi for unbounded operators).

4.7 Bounded Imaginary Powers


Let E be a Banach space. A linear operator A in E is said to have bounded
imaginary powers, in symbols,

A E BIP := BIP(E) ,

provided A E P(E) and there exist c > 0 and M 2': 1 such that

-c ::; t ::; c . (4.7.1)

The following theorem shows that this assumption has far-reaching consequences.

4.7.1 Theorem Suppose that A E BIP. Then {A Z ; Re z ::; O} is a strongly


continuous semigroup on £(E), that is, the map

is a continuous representation of the additive semigroup [Re z ::; 0]. Moreover,


{ Ai t ; t E lR} is a strongly continuous group on E whose infinitesimal genera-
tor is i log A.

Proof Suppose that A E P K . If It I E [nc, (n + l)c) for some n E N, it follows


from Theorem 4.6.4, Theorem 4.6.5(iii), and (4.7.1) that

IIA-s+itxll = IIA-S(A i sign(t)E)n Ai sign(t)(ltl-nE)xll


(4.7.2)
::; KmMn+1llxll = KmMeniogM Ilxll::; KmMel:l1t11lxll

for 0 ::; s ::; m and x E dom(A2), where () := c1log M 2': O. Thus, by the density
of dom(A2) in E,

Re z ::; 0 . (4.7.3)

From this and Theorem 4.6.5(v) and (vii) it follows that the map z 1-+ AZ is
strongly continuous on [Re z ::; 0]. Now, using Theorem 4.6.5(iii) and the density
of dom(A2) in E, we see that {AZ ; Re z ::; O} is a strongly continuous semigroup
IlIA Maximal Sobolev Regularity 163

on E. Consequently, {Ait ; t E JR} is a strongly continuous group on E. Let B


denote the infinitesimal generator of this group and recall that

Bx = lim C1[Aitx - x]
t->O

iff x E dom(B). Since

T- 1[A- 8+i (t+T)x _ A- 8+it X] = T-1(A iT _ 1)A- 8+it x = A-8+itT-l[AiT - l]x

for x E E, s:2: 0, and t, T E JR with T =1= 0, we see that

(4.7.4)

for x E dom(B), s:2: 0, and t E R On the other hand, the analyticity of A-z for
Re z > 0 implies

XEE, s>O, tEJR. (4.7.5)

Since

aa A -8+it x = a8A -8 Aitx = (-log A)A --8 Aitx = (-log A)A- 8+it X

for x E E, s > 0, and t E JR, thanks to im(A -8) C dom(log A) by Theorem 4.6.2,
we deduce from (4.7.4) and (4.7.5) that

s>O, tEJR, (4.7.6)

for x E dom(B). Thus

x E dom(B) ,

as s -+ 0+. Since i log A is closed and A -Sx -+ x for s -+ 0+ we find i log A=:> B.
On the other hand, since the argument leading to (4.7.4) implies

xEE, s>O, tEJR,

it follows from (4.7.5) that

iBA- 8x = A- 8(-logA)x -+ (-logA)x, s -+ 0+, x E dom(1ogA) .

Since B is closed and A -8 X -+ x we see that iBx = (-log A)x for x E dom(log A),
that is, B =:> i log A. This proves the theorem. _

4.7.2 Corollary Suppose that A E SIP. Then there exist constants M :2: 1 and
() :2: 0 such that
t E JR . (4.7.7)
164 III Maximal Regularity

Proof This follows from the proof of (4.7.3) by letting 8 = °


in (4.7.2) .•

The last part of the proof of Theorem 4.7.1 follows [HiP57, Theorem 17.9.2],
where 'boundary values' of analytic semigroups of angle 1r /2 have been studied
in detail.

Suppose that M ?:: 1 and fJ ?:: O. Then we write

A E BZP(M, fJ) := BIP(E; M, fJ)

iff A E BIP and estimate (4.7.7) is valid. Moreover,

BIP(fJ) := BIP(E; fJ):= U BIP(M, fJ)


M2:1

for fJ ?:: 0, so that


BIP = U BIP(fJ) .
02:0

In general, BIP is a proper subset of P, that is, there exist operators of


positive type which do not possess bounded imaginary powers. In fact, Venni
[Ven93] has shown that in every Banach space E with basis, thus in every sep-
arable Hilbert space, there exists A E peE) such that -A generates a strongly
continuous analytic semigroup, Ait = 1 for t E 21rZ, whereas Ait is unbounded
for t E 1r(2Z + 1). Earlier examples showing that BIP(E) -I peE), in general, are
contained in [Kom66, part I, section 14] and [BaiC91].
Although a general characterization of the class BIP is not available at
present, it is known that certain families of operators belong to BIP. Next we
collect some known cases:

4.7.3 Examples (a) Let E be a Hilbert space and suppose that A = A * ?:: a > o.
Then A E BIP(l, 0).
Proof This is an immediate consequence of Theorem 4.6.7.•
(b) Let E be a Hilbert space and let A be maximal accretive with 0 E peA).
Then A E BIP(l, 1r /2).
This is proven in [Kat62].
(c) Let (0., /1) be a a-finite positive measure space and let 1 < p < 00. Suppose
that A is the negative infinitesimal generator of a strongly continuous contrac-
tion semigroup of negative type on Lp(n, /1). Also suppose that A is resolvent
positive with respect to the natural order of Lp induced by the positive cone
Lt := Lp(n, /1; JR+). Then there exists a constant M ?:: 1, depending on p only,
but not on A, such that

t E R. .
III.4 Maximal Sobolev Regularity 165

Thus; given any c > 0, there exists M 2:: 1 such that

A E BIP(Lp(n, 1'); M, (11" +c)/2)

for each A satisfying the above conditions.


This is due to Coifman and Weiss [CoiW77j (also cf. [Ste70bJ).
(d) Let the hypotheses of (c) be true for every p E (1, (0) and let A be self-
adjoint in L2(n,p,). Then, given q E (1,00) and 0 > 11" iq-1 - 2- 1 1, there is K 2:: I,
independently of A, such that A E BIP(Lq(n,p,);K,O).
Proof Thanks to (a) we can assume that q of- 2. Fix any p E (1, (0) with

Then fix c E (0, (11"/fJ) -11"). Since L q : [L2,Lp]o (e.g., [Tri78, 1.18.6/2]), it follows
from (a) and (c) that A E BIP(Lq(n, 1'); K, 0), where K:= M O and 20:= 'l9(11" + t).
Since we can choose p such that Ip-1 - 2- 11 is arbitrarily close to 1/2 and c arbi-
trarily close to 0, the assertion follows .•

We also refer to [CIP90, Theorem 5.8] for bounded imaginary powers of oper-
ators on Lp(n, 1'; E) that are 'tensor product' extensions of operators on Lp(n, 1')
for some p E (1,00).

The following lemma is almost trivial but is useful in proving certain deeper
results about bounded imaginary powers.

4.7.4 Lemma Suppose that K, M 2:: 1 and 0 2:: 0, and let A E P.


(i) If there exists c > °such that
-c<Rez<O,

then A E BIP(M,O).
(ii) If A E BIP(M, 0) n PK then

IIAzll :::; KMeellmzl , -1<Rez<O.

Proof (i) Given x E D(A), it follows from Theorem 4.6.5(vii) that AZx -> Aitx
in E as z -> it and Rez < O. Since D(A) is dense in E, by Theorem 4.6.5(v), the
assertion is now an obvious consequence of the assumption.
(ii) is obtained from A-s+it = A-s Ait for.9, t E JR, and from Theorem 4.6.4 .•
166 III Maximal Regularity

For K 2 1 and fJ E [0, 7r), a linear operator A in E is said to be oftype (K, fJ),
in symbols:
A E P(K, fJ) := P(E; K, fJ) ,
if it is densely defined, if

Sf} := [I argzl ::::; fJ] U {O} C p(-A) ,

and if
A E Sf} . (4.7.8)
Put
P( fJ) := P( E; fJ):= U P(K, fJ)
K~i

and note that, trivially

P(K, fJ) C P(L, 0) , (4.7.9)

Observe that P = P(O), that is, every operator of type (K, fJ) is of positive type.

Let
7r-fJ 1
fJ K := fJ + - 2 - 1\ arcsin 2K . (4.7.10)

An obvious modification of (4.6.3) shows that, given A E P(K, fJ),

II(A + A)-ill ::::; 2K(1 + IAol)-l ::::; (2K + 1)/(1 + IAI)


(4.7.11)
for IA - Aol ::::; (1 + IAol)/(2K) and Ao E So.

Hence
Sf}x U [Izl ::::; 1/(2K)] C p( -A) (4.7.12)
and
P(K,fJ) C P(2K + 1,fJ K ) . (4.7.13)
In the following, we denote by

r := r(K, fJ)

the negatively oriented boundary of S,9K U [Izl ::::; 1/(2K)] C p( -A). Then

Re z < 0, A E P(K, fJ) , (4.7.14)

by Theorem 4.6.5.
After these preparations we can prove the following surprising theorem due to
Dore [Dor93a], which shows that, given any A E P(E), its (E,D(A)L.-realization
(that we again denote by A) belongs to B'IP((E,D(A))a), where (·,·)a is any
admissible real interpolation functor of exponent Q E (0,1).
III.4 Maximal Sobolev Regularity 167

4.7.5 Theorem Suppose that 0 < a < 1 and

(., .)'" E { (., .)""p, (., ·)~,oo ; 1::; p < 00 } ,

and that A E P(E; fJ) for some fJ E (0,7r). Then

Proof Put E", := (E, D(A))", and let A", be the E",-realization of A. Then
A", E C(E",). Since D(A2) C dom(A",) and since D(A2) ~ D(A) ~ E", by The-
orem 4.6.5(v) and the admissibility of (., .)"" it follows that A", is densely defined.
These assertions are also true for AI, the D(A)-realization of A. Since A com-
mutes with its resolvent, it follows that Al E P(D(A), fJ). Hence, by interpola-
tion, A", E P(E", , fJ).
It follows from Theorem 4.6.5 that

Rez = 0, x E D(A) .

Note that
A(8 + A)-l(,\ + A)-l = (8 - ,\)-1 [A('\ + A)-l - A(8 + A)-I]
= (8 - ,\)-1 (1 - ,\(,\ + A)-l - [1 - 8(8 + A)-I])
= (8 - ,\)-1 [8(8 + A)-l _ ,\(,\ + A)-I]

for 8 > 0 and ,\ E p( -A). Hence, by Cauchy's theorem,

A(8+A)-lAZX=~[r
27rZ
(_'\y-1 d'\8A(8+A)-lX+
lr s - ,\
r (-,\)ZA(,\+A)-l x d,\]
lr 8 - ,\

=~
27rZ
r(-,\y A('\ + A)-l Xd'\
lr 8 -,\

for 8 > o. In the last integral we can contract r to the negatively oriented boundary
of Sf}. Then

(4.7.15)

for Rez = 0 and x


D(A). E
Note that dom(A) = dom(e±if} A) and IIAxl1 = Ile±if} Axil for x E dom(A).
Thus D(A) ~ D(e±if} A) and, consequently, (E, D(A)) 0,00 ~ (E, D(e±if} A)) a ,ex)
168 III Maximal Regularity

for 0 < a < 1. Now it follows from (1.2.9.12) that

x I--> sup s'" IIA(s + e±il1 A)-lxll, x I--> sup s'" IIA(s + A)-1 x ll
8>0 8>0

are equivalent norms on (E, D(A)) "',00 . Hence we infer from (4.7.15) that

(4.7.16)

for Rez = 0 and x E D(A), where

M:= c(a) 1 00
C"'{II - te il1 r + 11- te- r
l il1 l } dt .

By invoking (1.2.9.12) once more, we see that

IIAzxll(E,D(A»",oo ~ ce(1r-I1)IItn zlllxll(E,D(A))",oo

for Re z = 0 and x E D(A). Now the assertion follows from the density of D(A)
in (E,D(A))~,oo' provided (-")'" = (",)~,oo' If we have ("')'" = ("')""p for some
p E [1,00), we choose 0 < ao < a < al < 1 and apply (1.2.5.2) and the reiteration
theorem (1.2.8.1), (1.2.8.2) to deduce the assertion .•

4.8 Perturbation Theorems


Now we prove some perturbation theorems for the classes P and BIP. They are
useful for proving that Lp-realizations of rather general elliptic systems possess
bounded imaginary powers.
In the following,

where B(t) := 1 for t 2: 0 and B(t) := 0 for t < 0 is the Heaviside function.

4.8.1 Lemma Suppose that A E P(K,iJ).


(i) If B is a linear opemtor in E satisfying dom(B) ::) dom(A) and

IIB(A+A)-lll ~ (3 < 1, A E 811 , (4.8.1)


then A + BE P(1- (3)-1 K, iJ).
(ii) f.L + A E P(Ks(iJ), iJ) for f.L 2: O.

Proof (i) It follows from (4.8.1) that 1 + B(A + A)-l E .L:aut(E) and
IlIA Maximal Sobolev Regularity 169

Thus we deduce from

>. + A +B = [1 + B(>. + A)-l](>. + A) (4.8.2)

that 8'19 c p(-(A+B)) and

11(>' + A + B)-III :::; (1 - (3)-111(>' + A)-III, >. E 8'19 .

Now the assertion is obvious.


(Ii) follows from the fact that>. E 8'19 implies I>' + 111 2:: 1>'1 sin(71" - -0) if -0 lies
in (71"/2,71"), and I>' + 111 2:: I>.j if 0 < -0 :::; 71"/2 .•

In the following, we put

Eo := E, E 1 := (dom(A), ijA-11)
for A E peE), if no confusion seems likely. We also assume that K 2:: 1, -0 E (0,71"),
and () 2:: o.

Now we can prove an important perturbation theorem for the class Peon"
cerning suitable 'subordinate' perturbations.

4.8.2 Theorem Suppose that A E P(K, -0) and that 0 :::; !X < 1. If !X > 0, put
Ea := J a (Eo, E1), where J is an exact interpolation functor of exponent !X. Then,
given B E £(Ea, Eo) and;3 E (0,1),
11 + A + BE P(l - (3)-lKs(-o),-o) ,
provided
11 2:: 110 := [((1 + K) IIBII ;3-1) l/(l-a) - 1] + S(-O) .
Proof Since A(>. + A)-l = 1 - >.(>. + A)-l we see that

11(>' + A)-ll1c(EQ,E') :::; 1 + K , >. E 8'19 .

This and (4.7.8) imply

>. E S'19 • (4.8.3)


Hence

Since I>' + 111 2:: 11 if 0 < fJ :::; 71"/2, and 1>' + 111 2:: 11 sine71" - fJ) if 71"/2 < fJ < 71", we
see that

Now the assertion follows from Lemma 4.8.1. •


170 III Maximal Regularity

Our next almost obvious lemma implies, in particular, that the classes P( 1'J)
and B'IP( 1'J) are invariant under similarity transformations.

4.8.3 Lemma Let F be a Banach space, let A and B be densely defined lin-
ear operators in E and in F, respectively, and let C E £(E, F) and D E £(F, E).
Suppose that Sf) C p( -A) n p( - B) and

(A + B)-l = C(A + A)-l D , (4.8.4)

Also suppose that A E P(E; K, 1'J). Then


(i) BE P(F; K 1, 1'J) with Kl := IICIIIIDII K.
(ii) If A E B'IP(E;M,8) then BE B'IP(F;M1,8) with M 1 := IICIIIIDII KM.
Proof (i) is obvious.
(ii) Thanks to (4.7.14),

BZ 1.
= -2
7fZ
r
ir
(-AYC(A + A)-lDdA = CAzD, Rez < 0 .

Hence, by Lemma 4.7.4(ii),

IISZII.c(F) :s; IICIIIIDII KMellllmzl , Rez < 0 .

Now the assertion follows from Lemma 4.7.4(i) .•

Next we prove a simple 'splitting lemma' that will greatly simplify our proofs
that a given operator of positive type has bounded imaginary powers.

4.8.4 Lemma Suppose that A E P(K, 1'J) and

(A + A)-l = R(A) + S(A) , A E f:= f(K,1'J) , (4.8.5)

and put
AEf, Rez<O.
Also suppose that
Rz,S E L1(f,ds,£(E))
and
-E < Rcz < 0, (4.8.6)

for some E > 0, ds denoting the 'arc-length measure'. Then

A E B'IP(M + IISIIL" (7r - 1'J) V 8) .


IIl.4 Maximal Sobolev Regularity 171

Proof It follows from (4.7.14) and (4.8.5) that

1.
A Z = -2
7rZ ir
rR z (>,) d>' + ~
27rz ir
r(->'yS(>.) d>' , Rez < 0 .

Thus, thanks to (4.8.6),


IIAzl1 :::; (M + (27r)-1(2K)1 RezIIISIILJe«1r-il)V8)IIrnzl , -10 < Rez < 0,
and Lemma 4.7.4(i) implies the assertion. _

As a first application of this splitting lemma we show that B'IP( B) is invariant


under suitable 'lower order perturbations'.

4.8.5 Theorem Suppose that A E P(K, 13) n B'IP(M, B) and 0 :::; fJ < 1. Fix
K1 ~ K and put
R(>') := (>. + A)-I,
Let B be a linear operator in E satisfying
(i) dom(B) =:> dom(A);
(ii) IIBR(>.)II:::; fJ < 1 for>. E r U Sf);
(iii) IIRBRIIL, (r,ds,L:(E)) :::; a < 00.
Then A + BE B'IP(KM + (1 - (3)-l a , (7r - 13) V B).

Proof Lemma 4.8.1 implies A + BE P((l- (3)-1 K, 79). From (4.8.2) we de-
duce that
(>. + A + B)-l = R(>.) + S(.>') , >. E r,
where
S:= -RBR[l + BRr 1 .

Hence (ii) and (iii) imply S E L1 (r, ds, £(E)) and IISIIL 1 :::; (1 - fJ)-l a . Now the
assertion follows from Lemmas 4.7.4(ii) and 4.8.4 and the fact that r(K,13) can
be replaced by r((l- fJ)-IK 1, 13). _

4.8.6 Corollary ruppose that A E P(K, 13) n B'IP(M, B). Then, given 1/ > 0,
there exists N s'uch that

/-1 + A E B'IP(N, (7r - 13) V B) ,

Proof Let /-11 ~ 0 be fixed and put Al := /-11 + A. Then Lemma 4.8.1 implies
Al E P(K1' 13) with Kl := Ks(13). Suppose that

Al E B'IP(M1 , B)
for some Ml ~ 1. Note that this is true if /-11 = o.
172 III Maximal Regularity

Let °< Jl ::; 1/(6Kl ) =: III and put B := JlIE and


Ri(.A) ;= (.>. + At}-l ,

It follows from (4.7.11) that

II Rl (.>.) II ::; 3Kl(1 + 1'>'1)-1 ,


Hence B satisfies (i)-(iii) of Theorem 4.8.5 with {3 := 1/2 and a := 2KIP, where
P is the Ll(fI,ds)-norm of (1 + 1'1)-2. Thus, thanks to the latter theorem,

Jl + Jll +A = Jl + A1 E BIP(M2, (rr - <0) V 0) ,


where M2 := KlMl + 2a. Now the assertion follows by induction starting with
Jll := 0, since II can be reached in finitely many steps of length at most lIi' •

The following perturbation theorem is particularly important in applications


since it allows to discard 'lower order terms'.

4.8.7 Theorem Suppose A E P(K,<O) n BIP(M,O). Also suppose a E [0,1) and


~'" is an exact interpolation functor of exponent a, if a > O. Lastly, assume that

-lr .
BE £(E", , Eo), where E", := ~",(Eo, Et}, and put

JlB:= [(4K II B I1)1/(1-"'l

Then Jl + A + BE BIP( (rr - <0) V 0) for Jl ? JlB.

Proof Put Jl := JlB and let Ao := Jl + A and Ro(.>.) := (.>. + Ao)-l. Since

sin <OK]
.>. E M(K, <0):= [ Izl::; 2K U StJK

implies.>. + Jl E f(K,<O)UStJK' we deduce from (4.7.8) and (4.7.H) that

IlRo('>')1I ::; 1 +~:+ JlI ' .>. E M(K,fJ) .

From this and from ARo('>') = 1 - (.>. + Jl)Ro('>') it follows that

IIARo(.>.) II ::; 4K , .>. E M(K,<O) . (4.8.7)

Thus, by interpolation,

.>. E M(K,I'J) ,

so that
31 -"'4'" K IIBU
lIBRo('>'}11 :s: (1+ I'>' + J.£1)1- a '
A E MCK,<o) . (4.8.6)
I1I.4 Maximal Sobolev Regularity 173

Of course, we can assume that B =f=. 0. Hence

(4.8.9)

Note that IA + ILl ;::: IL 1\ III - rl for r E (0, sin (19 K )/(2K)) and A E [Izl = r] U 8\9.
Thus it follows from (4.8.8) and (4.8.9) that we can find Kl > K and (3 E (0,1)
such that

Lemma 4.8.1(ii) implies Ao E P(Ks(19), 19). Hence we deduce from (4.7.8) and
(4.7.11) that
(1 + IAI) IIRo(A)11 ::; 3Ks(19) , AE~.

Thus we infer from (4.8.8) that

IIRo(A)BRo(A)11 ::; c(l + IAI),,-2 ,


Since we know from Corollary 4.8.6 that Ao E BIP (( 7r - 19) V 0), the assertion is
now a consequence of Theorem 4.8.5 and Corollary 4.8.6 .•

The results of this subsection are specializations of the corresponding as-


sertions in [AmHS94] where the slightly more general situation of the bounded
Hoo-calculus has been considered. The fact that we have given quantitative esti-
mates in some cases will be useful in Chapter VII.
A closely related result to Corollary 4.8.6 has also been proven in [PriiS90,
Theorem 3]. In that paper a functional calculus for operators of positive type pos-

°
sessing bounded imaginary powers has been developed. However, it is not assumed
that E p(A), but only that A satisfies the assumptions of Remark 4.6.8(c) and
that im(A) is dense in E. This generality requires rather more complicated proofs.
We also refer to [DorV90, Theorem 2.2 and Remark 2.3] and to [PriiS93,
Proposition 3.1] for results related to Theorems 4.8.5 and 4.8.7.

4.9 Sums of Closed Operators


Let E be a Banach space. Given linear operators A, B: E --> E, their commutator
is defined by
[A,B]:= AB - BA.
Thus A and B commute iff [A, B] = 0. If A: dom(A) C E --> E and B: E --> E
are linear then A and B commute if AB :> BA. Note that this means that
B(dom(A)) C dom(A) and ABx = BAx for x E dom(A).
174 III Maximal Regularity

For convenience, we collect some well-known facts about commuting linear


operators. Observe that the first part of the following lemma has already been
used in several of the preceding proofs.

4.9.1 Lemma Suppose that A E C(E). Then

(i) A(>' + A)-I ~ (>. + A)-I A, >. E p( -A).

(ii) If B E £(E) then


AB~BA ¢=} [(>'+A)-I,B] =0 for some>. E p(-A)
¢=} [(>' + A)-I, B] = ° for all >. E p(-A).
(iii) If A, B E CJ(E) then
Ae- tB ~ e- tB A, t 20,
¢=} A(J.l + B)-I ~ (J.l + B)-I A, J.l > type( -B),
¢=} [(>' + A)-I, (J.l + B)-I] = 0, >. > type(-A), J.l > type(-B),
¢=} [(>' + A)-l,e- tB ] =0, >'>type(-A), t20,
¢=} [e-sA,e- tB ] =0, s,t20.
Moreover, the assertions involving>. > type(-A) or J.l > type(-B) are already
true if they hold for some>. > type( - A) or at least one J.l > type( - B), respectively.

Proof (i) follows from the obvious relation

A(>' + A)-I = 1 - >.(>. + A)-I ~ (>. + A)-I A


for>. E p(-A).
(ii) If AB ~ BA then (>. + A)B ~ B(>' + A). Thus, if >. E p( -A), it follows
that B ~ (>. + A)-I B(>. + A); hence B(>' + A)-I ~ (>. + A)-I B. Now the asser-
tion is obvious.
(iii) is an easy consequence of (ii) and the representation formulas (11.6.3.3)
and (II.6.3.4) .•

Two closed linear operators A, B in E are said to be resolvent commuting


if there exist>. E p( -A), J.l E p( -B) such that

(4.9.1)

It follows from Lemma 4.9.1(ii) that (4.9.1) then holds for all >. E p(-A) and
all J.l E p( -B).

4.9.2 Lemma (i) Suppose that A E P(E) and BE £(E). If AB ~ BA then


AZ B ~ BAZ for z E C.
IlI.4 Maximal Sobolev Regularity 175

(ii) Suppose that A, B E peE) and are resolvent commuting. Then

for z E CC and Re w < 0 .

If BE BIP(E) then N B W ~ B WAZ for z E C and Rew s: O.


Proof (i) IfRe z < 0, the assertion follows from Lemma 4.9.1(ii) and from (4.6.22).
Then AZ = (A-z)-l for Rez > 0 implies the assertion for Rez > 0 by again invok-
ing Lemma 4.9.1(ii). Suppose that Re z = O. If ;:r E dom(A), the very same lemma
and the representation formula of Theorem 4.6.b(ii) guarantee that Bx E dom(AZ)
and BAzx = AZ Bx. By the definition of AZ as the closure of its restriction to D(A),
given x E dom(AZ), there exists a sequence (Xj) in D(A) such that Xj -> x and
AZxj -> AZx: in E. Thus AZ BXj = BAZxj -> BAzx by the boundedness of B. Since
BXj -> Bx, the closedness of AZ shows that Bx E D(AZ) and AZ Bx = BAzx.
Hence AZ B ~ BAz in this case as well.
(ii) From (i) we infer AZ(>. + B)-l ~ (..\ + B)-l AZ for z E CC and ..\ E p( -B).
Now the closedness of AZ and (4.6.22) imply AZ BW ~ BW AZ for z E C and Rew < O.
Suppose that B E BIP(E) and Rew = O. Then, given x E D(AZ), we see that

c > O.

Thus, letting c -> 0, we deduce from Theorem 4.7.1 and the closedness of AZ that
AZ BW x = BW AZx, which proves the lemma. _

Throughout the remainder of this subsection we presuppose that

A, BE PK(E) and are resolvent commuting and


AEBIP(E;M,BA) , BEBIP(E;M,BB) (4.9.2)
+ BB <?f.
with BA

It follows from Theorem 4.6.2 that the map

[0 < Rez < 1]-> L(E) , A~ 1-+ - . , - - - -


sin ?fZ
is holomorphic. Note that, thanks to (4.9.2) and Lemma 4.7.4(ii),

IIA -z Bz-111 < 4K2 M2e-7rllmzle(OA +OB)IIrnzl (4.9.3)


Isin ?fzl -
for 0 < Rez < 1 and IImzl 2': 1. Thus, given any C E (0,1),

is well-defined and, thanks to Cauchy's theorem, independent of c E (0,1). Below


we show, following [DorV87]' that S = (A + B)-l provided E is a UMD space.
176 III Maximal Regularity

4.9.3 Lemma If x E dom(A + B) then S(A + B)x = x.


Proof It follows from Lemma 4.9.2(ii) that
1
S(A+B)x=--:-
l
C + iCXJ A-zBzx

.
2z c-iCXJ Slll7rZ
dz---:- .
l
1 c- 1+ iCXJ BZA-zx
2z c-l-iCXJ Slll7rZ
dz. (4.9.4)

Define g: [-1 < Rez < 1] --> E by


A-z BZx = A-z B z - 1Bx , O:SRez<l,
{
g(z):= BZA-zx = BZA-z-1Ax, -1 < Rez < O.
Note that, thanks to Lemma 4.9.2(ii), 9 is well-defined. By Theorem 4.6.5(vii)
the function 9 is holomorphic in [-1 < Rez < 1]\ilR and, by Theorem 4.7.1, it is
continuous on [-1 < Re z < 1]. From this and from Cauchy's theorem we easily de-
duce that JRg(z) dz = 0 for every closed rectangle in [-1 < Rez < 1] whose sides
are parallel to the coordinate axes. Hence 9 is holomorphic in [-1 < Re z < 1] by
Morera's theorem (e.g., [Nar85, Theorem 1.2.5]), which is obviously also valid in
the E-valued case. Put f(z) := g(z)/ sin 7rZ for -1 < Rez < 1. Then f is holomor-
phic in [-1 < Rez < 1]\{O} and has a pole at z = 0 with residuum 7r- 1x. Lastly
we obtain, similarly as in (4.9.3), that
Ilf(z)11 :S ce-(7r-&A-&B)I Irnz l

for -1 < Rez < 1 and I Imzl ::::: 1. Now (4.9.4) and the residue theorem imply the
assertion. _

4.9.4 Lemma Suppose that 0 < c < 1 andx E dom(Al-£). Then Sx E dom(A+B)
and (A + B)Sx = x.
Proof Since A1-e-itBe-Hit x = Be-lBitA-itAl-ex by Theorem 4.7.1 and by
Lemma 4.9.2(ii), the integral

is absolutely convergent. Hence Sx E dom(A) and


_ 1 l£+iCXJ A1-z Bz-1x
ASx- - dz. (4.9.5)
2i £-iCXJ sin7rz
On the other hand, the map

is holomorphic in [c -1 < Rez < 1] by Theorem 4.6.2. Moreover, it follows from


Lemma 4.7.4(ii) that
IIW- 1A-zxll :S IIBz-1111IA£-1-zIIIIA1-£xll :S K2 M 2 1IA 1-£xll e(&A+&B)IIrnzl
I1I.4 Maximal Sobolev Regularity 177

for c - 1 ::::: Re z ::::: 1. Thus, by the residue theorem,


1 jE-I+iOO A-zBz-Ix
Sx = -2. . dz + 7rRes(h, 0) , (4.9.6)
Z E-I-ioo sm 7rZ
where h(z) := A-z Bz-Ix/ sin 7rZ. Since Res(h,O) = 7r- 1 B-Ix and since the inte-
gral in (4.9.6) is abolutely convergent, it follows that Sx E dom(B) and
1 jE-I+iOO A-z BZx
2Z.
BSx=x+- . dz=x-ASx,
E-I-ioo sm 7rZ
thanks to (4.9.5) .•

4.9.5 Lemma The operator A +B is closable, its closure A +B is injective,


and S = (A + B) -1.

Proof Let (Xj) be a null sequence in dom(A + B) such that (A + B)xj -+ y.


Then Xj = S(A + B)xj -+ Sy by Lemma 4.9.3, so that Sy = o. From Lemma 4.9.4
we know that SA-Iy E dom(A + B) and (A + B)SA-I y = A-Iy. Moreover, the
definition of S and Lemma 4.9.2 imply SA-IV = A-ISy = O. Hence A-Iy = O.
Consequently, y = 0 and A + B is closable.
Let x E dom(A + B). There exists a sequence (Xj) in dom(A + B) such that
Xj -+ x and (A + B)xj -+ (A + B)x. Since Xj = S(A + B)xj by Lemma 4.9.3, it
follows that x = S(A + B)x. Conversely, given x E E, choose a sequence (Xj)
in dom(A) such that Xj -+ x. Then SXj E dom(A + B) by Lemma 4.9.4, and
SXj -+ Sx and (A + B)Sxj = Xj -+ x. Hence Sx E dom(A + B) and (A + B)Sx = x,
which proves the assertion. •

As a first application of these considerations we prove the following theorem


concerning the spectrum of operators in BIP(E).

4.9.6 Theorem If A E BIP(E; 0) then a(A) c [I arg AI : : : OJ.


t
Proof We can, of course, assume that 0 < 7r. Fix AE with Iarg AI < 7r - Band
put B := AlE. Then B E BIP(E, Iarg AI) and condition (4.9.2) is satisfied. Since
A + B = A + A is closed, we infer from Lemma 4.9.5 that (A + A)-I E C(E), that
is, A E p( -A). Hence [I arg zl > OJ c p(A), which proves the theorem .•

The preceding theorem can be improved to show that (1 + IAI) "( A + A) -1 "
is bounded in [I arg AI : : : 7r - B - cj for each c E (0,7r - 0) if 0 < 7r (cf. [DorV90,
Theorem 2.1]). This shows, in particular, that -A generates a strongly continuous
analytic semigroup on E if A E BIP(E; B) for some B < 7r /2.

Now we can prove the main result of this subsection, namely the following
DORE-VENNI THEOREM.
178 III Maximal Regularity

4.9.7Theorem LetE be a UMDspace, letA E BIP(E;8A) andB E BIP(E;8B)


with 8 A + 8B < 7r, and let A and B be resolvent commuting. Then A + B E C(E)
and 0 E p(A + B).

Proof We shall prove that Sx E dom(A + B) for x E E. Then Lemma 4.9.5 im-
plies dom(A + B) C dom(A + B). Hence A + B = A + B and S = (A + B)-I.
By Theorems 4.6.2 and 4.7.1 the map z 1-+ A- z B z - 1 /sin7rz is holomorphic
on [0<Rez<1J, continuous on [O:::;Rez:::;I]\{O,I}, and has poles at z=o
and z = 1. Thus it follows from the validity of estimate (4.9.3) for the strip
[0 < Rez < 1] that, given E > 0,

where r j,e are for j = 0, 1 the curves running from j - i 00 to j + i 00 and consisting
of the half-lines {j + it; It I ~ E}, and the circular arcs

{j + Ee i ('P+j11') ; -7r/2:::; r.p:::; 7r/2} .

Thus
Sx = Ro,ex + To,ex = R1,eX + T1,eX ,
where
11 A-j-itBj-Hit x
R j eX =- dt
, 2 Itl~e sin 7r(j + it)
and
TJ. eX -_ (-1)j 1)11'+11'/2 Eei<P
.
A-j-eei "'Bj-l+ee i ", d
X r.p.
, 2 )11'-11'/2 sin7r(j+Ee z 'P)

From Theorem 4.7.1 we easily deduce that To,ex --+ 2- 1 B- 1x and T1,eX --+ 2- 1 A-IX
as E --+ O. Hence

as E --+ O. Since A, B E C(E), it follows from Theorem 4.7.1 and Lemma 4.9.2(ii)
that Ro,ex E dom(B) and R1,eX E dom(A) and that

11 A-itBitx
BRa,ex =- .. dt = -AR1,eX .
2 Itl~e SIll 7rZ t

Consequently, it remains to prove that BRo,ex converges as E --+ O. Then

Sx - 2- 1 B- 1x E dom(B) and Sx - Tl A-IX E dom(A)

by the c10sedness of B and A, respectively. Hence Sx E dom(A) n dom(B).


111.4 Maximal Sobolev Regularity 179

Observe that

(4.9.7)

The first integral is independent of c > 0 and the last one is easily seen to converge
in .cs(E) as c ----> O. Given any s E (0,1), the middle integral can be written as

(4.9.8)

provided 0 < c < 1 - s. It is obvious that

for each x E E and that the first integral in the curly brackets equals -iHeu(s),
where He is the truncated Hilbert transform. Since Heu(s) converges in E for a.e.
s E lR by Theorem 4.5.1, it follows from (4.9.8), by choosing s E (0,1) appropri-
ately, that the middle integral in (4.9.7) also converges in .cs(E) as c ----> O.•

4.9.8 Corollary Let the hypotheses of Theorem 4.9.7 be satisfied and suppose that

Then

where c depends upon K, M, OA, and OB, but not upon the individual operators
A and B.

Proof Since (A + B)-l = S, the estimate for (A + B)-l follows easily from
(4.9.3). The above proof shows that

for x E E, where the integral is a principal value integral. Now the assertion is a
consequence of (4.9.7), (4.9.8), and Theorem 4.5.1. •
180 III Maximal Regularity

The Dore-Venni theorem has been extended by PrUss and Sohr [PrUS90]
to the case 0 ¢. p(A) n p(B). In this case A + B is not invertible, in general.
In [PrUS90] it is also shown that A + BE BIP(E, BA V BB) if the hypotheses of
the Dore-Venni theorem are satisfied and BA =I- BB, which extends a slightly weaker
result of [DorV90]
The sum of two resolvent commuting operators of positive type has been
studied earlier by Sobolevskii [Sob75) and Da Prato and Grisvard [DaPG75]. In the
latter paper it is shown that, given an arbitrary Banach space E and A E P( E; {} A),
B E P(E; {}B) with {}A + {}B > 7r, it follows that the (E, D(A) )1J,q-realization of
A + B is closed and has a bounded inverse for any BE (0,1) and 1 :S q :S 00. In
general, A + B is not closed, even if E is a Hilbert space (cf. [BaiC91]).

4.10 Maximal Regularity


Let (Eo, E 1) be a densely injected Banach couple. We denote by J a perfect subin-
terval of JR+ containing 0 and fix p E (1, (0). Then we put

Eo := Eo(J) := Lp(J, Eo)

and
E1 := E 1(J) := Lp(J, Ed n w;,1(J, Eo) .
It is clear that Eo and E1 satisfy (1.5.1) and (1.5.2), respectively.

4.10.1 Lemma (i) Ej(JR+), j = 0, 1, are continuously tmnslation invariant.


(ii) Let To > 0 be fixed. Then Ej (J) is a regular extension of E j ([0, To», j = 0,1,
uniformly with respect to any subinterval J of JR+ containing [0, To).

Proof (i) Observe that {.As; s 2: O} is a contraction semigroup on Ej(JR+)


whose restriction to the subspace consisting of the smooth Ervalued functions
with compact supports in JR+ is strongly continuous. Since the latter subspace is
dense in Ej(JR+), the assertion follows.
(ii) It is well-known and not difficult to see that the smooth functions with
compact supports in] are dense in Ej(JR+) (cf. Chapter VI). Using this fact it is
easily seen that the coretraction r C constructed in the proof of Lemma 2.1.2 extends
to coretractions from Ej(J) onto Ej(JR+) whose norms are bounded independently
of J ::> [0, To) .•

Now we can characterize the trace space of E1 and prove an important em-
bedding theorem.

4.10.2 Theorem "YE1 ~ (Eo, Et}1-1/p,p and E1 '-+ BUC(J, (Eo, E 1h-1/p,p)'
III.4 Maximal Sobolev Regularity 181

Proof The first assertion follows from the characterization of the real interpo-
lation functors by means of the Lions-Peetre trace method (e.g., [BerL76, Corol-
lary 3.12.3] or [Tri78 , Theorem 1.8.2]). Now the second one is a consequence of
Proposition 1.4.2 and Lemma 4.10.1. •

As in Section 2 we assume that

A C 1-l(EI' Eo) and there are constants M > 0

}
and {} E (0,71)2) such that l.;ii C p(-A) and
(4.10.1)
IIAILc(E1,Eo) + (1 + IAI)I-j II (A + A)-lllc(Eo,Ej ) ::; M
for (A, A) E l.;ii X A and j = 0, 1.

Also recall that RA has been defined in (1.5.6).

4.10.3 Proposition Let assumption (4.10.1) be satisfied. Then

uniformly with respect to J.

Proof Lemma 2.2.1 and Remark 11.5.1.2 imply the existence of w > 0 such that
type( -A) ::; -w for A E A. Now the assertion follows from (2.2.2) and (1.2.10.6)-
(1.2.10.9) .•

4.10.4 Remark Given T > 0, there exists a constant a :::: 1 such that

a-I 11·II(Eo,Elh_l/P,P ::; 11'1I"IEt(J) ::; a 11'II(Eo,Etll-l/P,P )

uniformly with respect to J:J [0, T].

Proof Letting J o := [0, T], we infer from Theorem 4.10.2 that

Next, Lemma 4.1O.1(ii) and Proposition 1.4.1 imply that we can replace J o in the
last inequality by any J:J Jo. This gives the first of the asserted inequalities.
Note that "IRA = 1. Thus, thanks to the definition of the norm in ')'lEI(J),

x E (Eo, EIh-l/p,p .

Now we obtain the second asserted inequality from Proposition 4.10.3.•

Given A E A, it is obvious that

(4.10.2)
182 III Maximal Regularity

Thus, letting lEi,A := ker( 8+ A), it follows from Theorem 4.10.2, Proposition 4.10.3,
and Remark 4.10.4 that, given T > 0,
(4.10.3)
uniformly with respect to J:::> [O,Tj. Hence Lemma 1.5.1 implies that (lEo,lEi) is
a pair of maximal regularity for A if we can show that

a+ A E Lis(lE i ,1', lEo) , AEA, (4.10.4)


where lE i ,1' := kerT For this we put

WP~I' := {u E Wi(J,E o) ; "(u = O}


and define a linear operator B in lEo by

B : l1j,~ l' C lEo --> lEo, u f-+ au .


The following lemma shows that, given c E (0, IT), there exists M ~ 1, indepen-
dently of J, such that

BE P1+T(IEO) n BIP(lEo; M, (IT + c)/2) , (4.10.5)


provided J is bounded, T := sup J, and Eo is a UMD space.

4.10.5 Lemma Suppose that J is bounded with T := sup J and Eo is a UMD


space. Then BE P1+T(lEO) and there exists a constant M, independently of J,
such that
t E lH. .

Proof Given'x E C, it is easily verified that (,X + B)-i exists and is given by

t E J, v E lEo .

Thus we deduce from Young's inequality for convolutions that


,X> O.
Hence (1 + ,X) II (,X + B)-ill::;; 1 + T for ,X ~ 0, that is, B E P1+T(lEO)'
From (4.6.9) we see that
I1I.4 Maximal Sobolev Regularity 183

for v E lEo, t E J, E 2: 0, and 0 < Re Z < 1, thanks to the complementing for-


mula (4.6.25). Denote by r C E .c (lEo , Lp(~, Eo)) the trivial extension operator, that
is, rCu is the extension of u over ~ by zero, and by r E .c (V' (~, Eo), V' (J, Eo)) the
restriction operator to J. Then

v E lEo ,

in V'(J, Eo), where

t > 0,
t :S 0 ,
for 0 < Re Z < 1 and E 2: o. By the convolution theorem

gz,e *r C -r-l
v = J" -r C
mz,eJ"r v , v E lEo ,

where mz,e := !iz,e' Thus

m (T)'= F(g )(T) = _1_


z,e' z,e f(z) io
roo tz-1e-t(e+iT) dt = (E + ir)-Z
f(z)
r
is
(Z-le-( d(
'

where S := { t(E + iT) ; t 2: O}. Since 0 < Re z < 1, we can deform S to ~+, thanks
to Cauchy's theorem. Hence

E>O, TE~.

Note that gz,e ~ gz,o in S', as E -+ O. Thus, by the continuity, hence the weak
continuity, of the Fourier transform in S', and since

as E-+O,

it follows that mz,o(T) = (iT)-Z for T E ~ and 0 < Rez < 1. This shows that
E2:0,O<Rez<1, (4.10.6)

in V'(J, Eo) for v E lEo. It is not difficult to see that

m z,€ (T)I
1 <
_ E-Rezell"IImzl/2 , TE~, E>O, O<Rez<l.

Since mz,e(T) = -iZ(E + iT)-lmz,e(T), it follows that mz,e E MM and

E> 0, 0 < Rez < 1. (4.10.7)

Now we infer from (4.10.6), (4.10.7), Theorem 4.4.3, and Lemma 4.7.4 that

O<E<l, tE~. (4.10.8)


184 III Maximal Regularity

It is easily verified that mz,e ---+ mi t,e in M M as z ---+ it in [0 < Re z < 1]. Hence
we deduce from (4.10.6) and Theorem 4.4.3 that

O<c<l, tElR, (4.10.9)

in Lp(J, Eo) for v E lEo. It is not hard to see that mit,s ~ mit,O in S' as c ---+ O.
Thus, given v E D(J,Eo), the continuity properties of the Fourier transform and
(4.10.6) and (4.10.9) imply

as c ---+ 0 (4.10.10)

for t E JR. On the other hand, it follows from (4.10.8) and the reflexivity of lEo that
there exists a null sequence (Cj) such that (Cj + B)-itv) converges weakly in lEo
to some wE lEo. Since lEo <-t DI(J, Eo), we see that (Cj + B)itv ~ w in D'(J, Eo).
Hence (4.10.10) implies rF-1mit,oFrCv = wE lEo. Finally, it is an easy conse-
quence of Theorem 4.6.5(ii) that (c + B)itv ---+ Bitv in lEo, which shows that
w = Bitv. Now we infer from (4.10.8) that

tEJR, vED(J,Eo) ,

and the assertion follows from the density of D(J, Eo) in lEo .•

4.10.6 Remark The above theorem shows that the symbol of the pseudo-
differential operator 8 i S equals (i T)i S for s E JR, where T is the Fourier variable
corresponding to t .•

After these preparations we can prove the main result of this section.

4.10.7 Theorem Suppose that Eo is a UMD space. Let assumption (4.10.1) be


satisfied and suppose that there exist constants N 2: 1 and (J E [0, 7r /2) such that
A C BIP(Eo; N, (J). Then

(lEo, lEI) := (Lp( J, Eo), Lp(J, E 1) n Wpl(J, Eo))

is a pair of maximal regularity for A, and

"YlEI ='= E1-1/p,p := (Eo, Edl-l/p,p .

Moreover,
11(8 + A,"Y)-IIIc(lEoXE,_,/p,p,lE,) :::; c , AEA,
uniformly with respect to J.

Proof (i) First suppose that J is bounded. We denote the 'multiplication oper-
ator', induced in Lp(J, E 1) by A via A1L(t) := A(u(t)), again by A. Then

(4.10.11)
IlI.4 Maximal Sobolev Regularity 185

and it follows from (2.2.2) that the graph norm of A, considered as a linear operator
in lEo, is equivalent to the Lp(J, E1)-norm. Hence A E C(lEo) by Lemma 1.1.1.2. It
is an obvious consequence of (4.10.1) and the density of Lp(J, E 1) in Lp(J, Eo) that

(4.10.12)

From Theorem 4.6.5(ii) we easily infer that Ail on lEo is given by the corresponding
multiplication operator. Hence

A c BIP(lEo; N, 0) . (4.10.13)

Fix BB E (7f/2,7f - B). Then Lemma 4.10.5 guarantees that B E BIP(lEo;BB)' Of


course, A and B are resolvent commuting for A E A. Since lEo is a UMD space by
Theorem 4.5.2(vi), the Dore-Venni Theorem 4.9.7 guarantees that A + BE C(lEo)
and 0 E p(A + B). Note that

dom(A + B) = dom(A) n dom(B) = Wp~)' n Lp(J, Ed = lEI,)' .

Hence A +B E £is(lE 1,)', lEo) and, thanks to (2.2.2),

for U E lEI,)'" Now Corollary 4.9.8, (4.10.12), (4.10.13), Lemma 4.10.5, and (4.10.5)
imply
II (A + B)-lll.c(lEo,IE" .,) ::; c(J) , AEA. (4.10.14)

Lastly, thanks to (4.10.14) and Proposition 4.10.3,

and
(4.10.15)

for (f,x) E lEo x E1-1/p,p and A E A. Theorem 4.10.2 and (4.10.2) imply

(8+ A,,),) E £(lEl,lEO x E1-1/p,p)

and it is clear that (8+ A,,),) (RAX + (A+B)-lf) = (f,x). Hence

(8 + A, ,),)-l(f, x) = RAX + (A + B)-l f .

Now the assertion follows from (4.10.13) and Lemma 1.5.1.


(ii) Suppose that J = jR+ and recall from (1.5.7) the definition of K A . From
(i) it follows, of course, that RAX + KAf is a WI,loc-solution of (1.5.5)(x,A,f) for
186 III Maximal Regularity

each f E lEo· Hence, thanks to Theorem 1.5.2, Proposition 4.10.3, and (4.10.2), it
suffices to prove that
(4.10.16)

Given v E lEo, put Vj := VX[j,j+1) for j E N. Then

j+1
L1
00

I KAv lli p (IR+,E,) = IIKAV(t)llf dt


j=O J

= 121IKAV(t)llf dt

+ ~1]+1111j-1 e-(t-r)Av(r)dr+ 1~1 e-Ct-r)Av(r)drll: dt


(4.10.17)
:S 121IKA(VO + v1)(t)llf dt

+ 2P - 1 f 1]+1 {111
j=2 J 0
j
-
1e-(t-r)Av(r) drll:

+ 111~1 e-(t-r)A [Vj-1(r) + Vj(r)] drll:} dt .

Part (i) of this proof guarantees that

12I1KA(Vo + vdllf dt :S cIlvo + vlllt((0,2),Eo ) (4.10.18)

From (4.10.1) and Lemma 2.2.1 we see that

~ 1]+1 11 1 j - 1e-(t-r)Av(r) drll: dt

:S cf 1]+1 (l j 1
- (t - r)-l e -w(t-r) Ilv(r)llo dr r
dt

r
(4.10.19)

1 (1
j=2 J 0

:S c 00 t
-
1
e-w(t-r) Ilv(r)llo dr dt

:S w-Pc1 Ilv(t)ll~ 00
dt :S cIlvllj,p(IR+,Eo ) ,

where for the second to the last estimate we used Young's inequality for convolu-
tions. Lastly, using again (i) for the interval [0,2]' we can estimate the integrals
111.4 Maximal Sobolev Regularity 187

from above by

la 2 11la t
e-(t-T)A h-1(T + j -1) + Vj(T + j -1)] dTII: dt
(4.10.20)
::::; c Ilvj-1 + Vj II~p(lR+ ,Eo)
Since the functions Vj have disjoint supports it follows that
00 00

IIvo + v11Iip«0,2),Eo) + L Ilvj-1 + VjIlL(lR+,Eo) ::::; 2 L IlvjIlL(lR+,Eo)


j=2 j=O

= 21Ivll~p(lR+,Eo)
Thus (4.10.17)-(4.10.20) imply
(A ~ KA) E B(A,.c(E o,Lp (lR+,E 1 ))) . (4.10.21)
Now suppose that v E BUCP (lR+, Eo) n Eo for some p E (0,1). Then we know
from Proposition 2.1.1 and Theorem 2.5.3 that w := KAV E C 1(lR+, Eo) and that
wet) = vet) - AKAv(t) for t > o. Thus we infer from (2.2.2) and (4.10.21) that
II(KAV)"IILp(lR+,Eo) ::::; c IlvlllEo (4.10.22)
for A E A and v E BUCP (lR+, Eo) n Eo. Since the latter set is dense in Eo, esti-
mate (4.10.22) is true for all v E Eo and A E A, and (4.10.16) follows from this
and from (4.10.21).
(iii) Lastly, suppose again that J is bounded. Then, by an obvious extension
and restriction argument, we obtain from Lemma 4.1O.2(ii) and step (ii) that
(8 + A, 'Y)-1 can be estimated uniformly with respect to J .•

The next theorem shows that we can weaken assumption (4.10.1) if we restrict
our considerations to bounded intervals.

4.10.8 Theorem Suppose that J is bounded and Eo is a UMD space. Also


suppose that K, N 2:: 1, w > 0, and 8 E [0, 7r /2), and that
A C H(E1, Eo, K, w) satisfies w + A C BTP(Eo; N, 8) . (4.10.23)
Then
(E o,E 1 ) := (Lp(J, Eo), Lp(J, E 1) n Wp1(J, Eo))
is a pair of maximal regularity for A, and

'YE1 == E 1- 1/p,p := (Eo, E1h-l/p,p . (4.10.24)


Moreover,
AEA. (4.10.25)
188 III Maximal Regularity

Proof Thanks to Corollary I.1.4.3 the set w + A satisfies assumption (4.10.1).


Hence Theorem 4.10.7 guarantees that (Eo, E 1 ) is a pair of maximal regularity for
w + A and that (4.10.24) and (4.10.25) are satisfied (with A replaced by w + A in
(4.10.25)). Now the assertion follows from Proposition 1.5.3 and the obvious fact
that eWEj ~ Ej for j = 0, I, thanks to the boundedness of J .•

4.10.9 Remarks (a) It is clear that a result which is analogous to Theorem 2.5.5
holds in this case as well.
(b) By a result of Dore [Dor93b, Theorems 2.1 and 2.2] and Lemma I.1.1.2 the
assumption:

}
there exists p E (1,00) such that
(Lp(J, Eo), Lp(J, E 1) n w;,1(J,Eo») is a pair of (4.10.26)
maximal regularity for A E £(El' Eo) n C(Eo),

implies that A E 1t(El' Eo). Moreover, if J = R+ then type( -A) < o.


(c) It may be worthwhile to reformulate the assertion of Theorem 4.10.7 as:
Given any (x, A, f) E E 1- 1/ P x A x Lp(J, Eo), the Cauchy problem

it + Au = f(t) , t E j , u(o) = x (4.10.27)


has a unique solution u E Lp(J, El) n Wpl(J, Eo), and

l llu(t)llb dt + lllit(t)lIb dt + lIIAu(t)llb dt


(4.10.28)
~ c(ll1f(t)llb dt + IlxIILl/P) ,
where c is independent of (x, A, J) E E 1- 1/p x A x Lp(J, Eo). Moreover, Theo-
rem 4.10.2 implies that the left-hand side of (4.10.28) can be estimated below by

Co sup
tEJ
Ilu(t) II~ I-lip

for a suitable Co > o.•


Now we extend the above maximal regularity results to the nonautonomous case.
The proof of the following theorem is a modification of the one of Theorem 2.6.1.

4.10.10 Theorem Let Eo be a UMD space and let there exist K ~ 1 and w > 0
such that
A E BUC(J,1t(E1 ,Eo,K,W») . (4.10.29)
Suppose that there are N 2: 1 and (J E [0, 7r /2) such that

w + A(t) E BIP(Eo; N, (J) , tE J .


1Il.4 Maximal Sobolev Regularity 189

If J = jR+, also assume that there exists T ::::: ° such that s ( - A(T)) < °
and

sup IIA(t) - A(T) II.C(E1 ,Eo) :::; AjIIKA(T) 11c(lEo,lE,) (4.10.30)


t?T

for some A E (0,1). Then

(8+ A,,),) E Cis(lE1,lEO x E 1 - 1/ p ) ,

where
(lEo, lEI) := (Lp(J, Eo), Lp(J, E 1) n w;,1(J, Eo»)
and E 1- 1/ p := (Eo, E 1 h-1/p,p'

Proof It follows from (4.10.29) and Theorem 4.10.2 that

Thus, by Banach's homomorphism theorem, it remains to show that (8 + A, ')') is


bijective.
(i) First suppose that J is bounded. Then, thanks to (4.10.29), we can assume
that J = [0, T] for some T > 0. Theorem 4.10.8 implies

8 E J , (4.10.31)

and the existence of a constant a such that

8 E J.

By the compactness of J and by (4.10.29) we can find points

° =: 80 < 81 ... < 8 m +2 := T


such that

j=O, ... ,m.

Put
t E Jj := [0,8J+2 - 8j], j = 0, ... , m .
It is an easy consequence of (4.10.29), Lemma 4.10.1, and Proposition 1.6.3 that

j =O, ... ,m. (4.10.32)

Let (j,x) E lEo x E 1 - 1/ p be given and put

fJ(t) := f(8j + t) , t E Jj, j = 0, ... , m .


190 III Maximal Regularity

Then (4.10.32) guarantees that the Cauchy problem

u + Ao(t)u = Jo(t) , t E jo , u(O) = x

has a unique solution Vo E lEI(Jo). From Lemma 4.10.1 and Theorem 4.10.2 we
infer that Vo E BUC(Jo , E I - I / p ), so vO(8d E E 1 - 1 / p . Hence, by again invoking
(4.10.32), we see that the Cauchy problem

has a unique solution VI E lEI (Jl). It is clear that

o :<:; t :<:; 81 ,

81 :<:; t :<:; 83 ,

is the unique solution of the Cauchy problem

u(O) =x .

Now the assertion follows by an obvious finite induction argument.


(ii) Suppose that J = IR+ and fix (j, x) E lEo x E I - I/p. Put Jr := [0, T + 1]
and let UT be the solution of the Cauchy problem

u + A(t)u = J(t) , t E jT , u(O) = x,


whose existence and uniqueness are guaranteed by the first part of this proof. Since
8( - A(T)) < 0 we infer from Corollary 1.1.4.3 that A(T) satisfies condition (4.10.1).
Hence Theorem 4.10.7 shows that

Assumption (4.10.30) and Proposition 1.6.3 imply

Thus the Cauchy problem

u + A(T + t)u = J(T + t), t >0, u(O) = uT(T)


has a unique solution v E lEI. Then

u(t) := { UT(t) , 0:<:; t:<:; T,


vet - T) , t > T,

is the unique solution in lEI of u + A(t)u = J satisfying u(O) = x. Hence the asser-
tion follows .•
IlIA Maximal Sobolev Regularity 191

4.10.11 Remarks (a) The above proof shows that, in the case where J = JR+,
assumption (4.10.29) can be replaced by

A E BC(JR+,H(E1,Eo,K,w)) .
The uniform continuity in (4.lO.29) is only needed if J is bounded and not closed,
in order to extend A over J.
(b) Of course, Theorem 4.10.lO implies an estimate of the form (4.lO.28) .•

Theorem 4.lO.7 is due to Dore and Venni [DorV87, Theorem 3.2] in the case
of a bounded interval (and for vanishing initial values). The extension to JR+ given
in part (ii) of the above prooffollows [Dor93b, Theorem 2.4], where it is attributed
to T. Kato. Other extensions of [DorV87, Theorem 3.2] are due to Pruss and Sohr
[PruS90] and Giga and Sohr [GiS91]. These authors admit operators A that are
not required to have a bounded inverse. Thus the semigroup {e~tA ; t 2:: O} is
only bounded but not exponentially decaying. Hence, thanks to Remark 4.lO.9(b),
the pair (4.lO.26) cannot be a pair of maximal regularity for A in this case. In
fact, in [PruS90] and [GiS91] it is only shown that, given f E Lp(JR+, Eo) for some
p E (1,00), the Cauchy problem

u + Au = f(t) , t >0 , u(O) = 0


has a unique solution u E Wp~loc(JR+, Eo) such that u, Au E Lp(JR+, Eo) and
(4.lO.33)

(In [GiS91] also nonzero initial values are admitted.) Note that the left-hand side
of (4.lO.33) is not an equivalent norm for Lp(JR+, E 1) n Wi (JR+, Eo), since the map
u J--.> IIAul1 is not an equivalent norm on El if 0 E O"(A).
Of course, the assumptions that Eo is a UMD space and A E HIP are suf-
ficient conditions only for (4.lO.26) to be a pair of maximal regularity for A. In
fact, suppose that

with type(-A) <0. (4.10.34)

Then a result of Benedek, Calderon, and Panzone [BCP62] implies that (4.lO.26)
is a pair of maximal regularity for A and any p E (1, 00 ), provided this is true for
some p E (1, (0). In the case that (Eo, E 1) is a densely injected Hilbert couple, by
means of Fourier transforms and Plancherel's theorem it is not difficult to see that
(4.10.26) is a pair of maximal regularity for every A satisfying (4.lO.34), provided
p = 2. By combining these results one obtains a maximal Sobolev regularity result
in the Lp (J, Eo )-setting if Eo is a Hilbert space, as has first been proven in [DeSi64].
In a more concrete situation these observations have been applied in [vW82], for
example, to derive 'Lp-Lq-estimates' for certain classes of parabolic differential
equations. For further 'maximal Sobolev regularity' results we refer to [Sob64],
[DaPG75], [CoL87], [Lam87]' [Mie87], and [GiGS91].
Chapter IV

Variable Domains

So far we have studied the case of 'constant domains', that is, we have always
assumed that A(t) E C(Et , Eo) for t E J, where (Eo, E t ) is a fixed densely injected
Banach couple. In this case the general theory is relatively easy and flexible. In
addition, only mild continuity assumptions for the function t ~ A(t) are needed.

We shall see in later chapters that a great many concrete situations lead to
constant domain problems, either directly or via extrapolation techniques studied
in detail in Chapter V. However, in order to develop a powerful theory of parabolic
evolution equations we cannot restrict ourselves to the constant domain case for
the following two main reasons. First, problems involving 'variable domains' occur
naturally in many concrete applications. Second, if the extrapolation technique can
be applied to transform a variable domain problem into a constant domain case,
we arrive at a 'weak' formulation of the original setting. In order to prove that the
solution of the weak constant domain problem is in fact a 'classical' solution of
the original variable domain problem, we shall rely on bootstrapping arguments
based on solvability results for a variable domain situation.

In Section 1 we prove regularity results for constant domain problems by lift-


ing them into suitable interpolation spaces between Eo and Et. The lifted problem
is then a variable domain problem, in general. In Section 2 we consider situations
where the domain may be variable but a suitable interpolation space between Eo
and D(A(t») is independent oft E J. The results of this section apply in particular
to the constant domain case to give new proofs for some of the results of Chapter II.
In Section 3 we briefly consider cases where no constancy assumption whatsoever
is made. Here we rely on maximal regularity results and study abstract versions of
initial-boundary value problems. The results of this last section provide a basis for
more detailed studies in concrete situations that are carried out in Chapter VIII.
194 IV Variable Domains

1 Higher Regularity
It is the main purpose of this section to prove Theorem 11.1.2.2, that will be
achieved in Subsection 1.5. For this we need some preparation. In Subsection 1.1
we collect a few simple, but useful, results about differentiable functions of one real
variable. Then, in Theorem 1.2.1, we give a general sufficient condition for a mild
solution to be, in fact, a solution. In addition, we prove quantitative estimates for
the restriction of a parabolic evolution operator to suitable intermediate spaces.

1.1 Properties of Differentiable Functions


Let E be a Banach space. Given -00 < a < b < 00, we denote the right derivative
of a function u: [a, b) -+ E by fhu. If E = JR.,

·-1· . f f(x
d+ f( x ) .- ImJil
+ h)h - f(x)
h~O+

is said to be the lower right Dini derivative of f at x. Observe that d+f(x)


exists in Jit

1.1.1 Lemma Suppose that f E C([a,b),JR.) and d+f(x)::; 0 for x E [a,b). Then
f is decreasing.

Proof Given E > 0, put fE(X) := f(x) - EX for x E [a, b). Then d+fE ::; -E. Sup-
pose that fE is not decreasing. Then there exist two points a ::; Xl < x2 < b so
that fE(X2) > fE(xd. Let Xo := max{ X E [Xl, X2J ; fE(X) = fe(XI) }. The continu-
ity of f implies that Xo is well-defined, Xo < X2, and fe(x) > fe(xo) for Xo < X < X2·
Consequently, d+fe(xo) ~ 0, which is impossible. This shows that

f(x) - EX::; f(y) - EY ,

Now the assertion follows by letting E -+ O.•

Suppose that J c lR is a perfect interval and that f E E J is right differ-


entiable in J\ {sup J} (where J\ {(X)} := J). If J is bounded above and closed
on the right, fhf E C(J, E) means that there exists a unique g E C(J, E) such
that 9 ::) 8+f·

1.1.2 Proposition Suppose that f E C(J, E) is differentiable from the right and
8+f E C(J,E). Then f E CI(J, E) and 8f = 8+f·

Proof Let Xo E J be fixed and put

g(x) := f(xo) + l
Xo
x
8+f(t) dt , x E J.
IV.l Higher Regularity 195

Then h := f - 9 E C(J, E) and h(xo) = O. Moreover, h is differentiable from the


right and o+h = O. Let e' E E' be arbitrary and put 'P(x):= Re(e',h(x)), x E J.
Then'P E C(J, lR) and 'P is differentiable from the right with o+'P = Re(e', o+h) = O.
Consequently, 'P(x) = 'P(xo) for x E J, thanks to Lemma 1.1.1. This shows that
Re( e', h(x)) = 0 for x E J and e' E E'. If lK =, C, replace e' by ie' to find that
Im(e',h(x)) = 0 for x E J and e' E E'. Thus (e',h(x) = 0 for x E J and e' E E',
which gives h = O. Hence f = 9 and, since 9 E C 1 (J, E), the assertion follows .•

Next we prove a simple continuity result for differentiable functions in situ-


ations where two topologies are involved. It is an elementary regularity theorem.

1.1.3 Proposition Let f: J --+ E be differentiable, let F be a Banach space


with F '--+ E, and suppose that of E C(J, F) and f(s) E F for some s E J. Then
f E C 1 (J, F) and the derivatives in E and in F coincide.

Proof It follows from F '--+ E that of E C (J, E). Hence f E C 1 (J, E). Thus

f(t) = f(s) + it Of(T) dT , t E J , (1.1.1)

in E. Since of E C(J, F), the integral in (1.1.1) converges in F. Now the assertion
follows by differentiating the integral with respect to its upper limit .•

1.2 General Solvability Results for Cauchy Problems


Let E be a Banach space, let J be a perfect subinterval of lR, and suppose that

(A, J) : J --+ C(E) x E .

Given s E J and x E E, consider the linear Cauchy problem

u+A(t)u=f(t) , tEJn(s,oo), u(s)=x. (1.2.1 )(s,x,A,f)

Below we give sufficient conditions for this Cauchy problem to possess a solution.
Observe that we do not require that dom(A(t)) be independent of t E J.

1.2.1 Theorem Let U be a parabolic evolution operator for A with regularity


subspace F. Suppose that F '--+ Fo '--+ E,

f E C(J,Fo) , (1.2.2)

and there ar'e 0: E [0,1) and c > 0 such that

IIU(t, s)llc(Fo,F) + IIAU(t, s)llc(Fo,E) <::: c(t -- S)-a , (t,s) E J'b. . (1.2.3)
196 IV Variable Domains

Then (1.2.1)(s,x,A,f) has for each (s, x) E J x E a unique solution u(·, s, x) and

u(·,·, x) E C(Ji:,., F) .

IfU E C(h,£s(F)) and x E F then

u(·,·, x) E C(h, F) .

Proof It follows from (II.2.1.2), (II.2.1.6), and (11.2.1.7) that U(·, s)x is a solution
of (1.2.1)(s,x,A,O) such that

U(·,·)x E C(Ji:,.,F).

It is an easy consequence of (11.2.1.2), (1.2.2), and (1.2.3) that

v:= U * f E C(h,F) . (1.2.4)

Hence we infer from Remark II.2.1.2(b) that u := U(·, s)x + v(·, s) has the desired
properties, provided we show that v(·,s) is a solution of (1.2.1)(s,o,A,f).
Let (t,s) E Ji:,. and h > 0 with t + h E J. Then

h- 1 [v(t + h, s) - v(t, s)]

= h- 1 it [U(t + h, r) - U(t, r)] f(r) dr + h- 1 It+h U(t + h, r)f(r) dr


=: h+Ih.
Since (1.2.2) and (II.2.1.2) imply Uf E C(h, E), it follows that

Ih ~ f(t) (1.2.5)

Observe that, thanks to (II.2.1.4)-(II.2.1.7),

U(t + h, r) - U(t, r) = - I t+h AU(O", r) dO" ,


t (t,r) E Ji:,. ,

in £(E). Hence (1.2.3) implies

IIU(t + h, r) - U(t, r)II.c(Fo,E) ::; c(t - r)-Oh .

Now we deduce from Lebesgue's theorem, (1.2.3), and (1.2.4) that

h ~ -it AU(t, r)f(r) dr = -A(t)v(t, s) , h~O+ , (1.2.6)


IV.l Higher Regularity 197

where the last equality follows from the closed ness of A. Thus (1.2.5) and (1.2.6)
show that the right derivative Ol,+V(t, s) exists on J n (s, 00) and satisfies

Ol,+V(t, s) = -A(t)v(t, s) + f(t) , tEJn(s,oo) . (1.2.7)

Moreover, (1.2.2)-(1.2.4) and (1.2.7) imply that

Ol,+V(·,S) E C(Jn [s, 00), E) .

Now the first assertion follows from Proposition 1.1.2.

If U E C(J.c..,.cs(F)) and x E F then U(·, ·)x E C(J.c.., F). Hence (1.2.4) im-
plies the second assertion. _

The next theorem gives a sufficient condition for a solution of (1.2.1)(s,x,A,j)


to be strict.

1.2.2 Theorem Let the hypotheses of Theorem 1.2.1 be satisfied. Also assume
that 0 E p(A(t)) fort E J and

AUA- 1 E C(J.c.., .cs(E)) (1.2.8)

If s E J and x E dom( A(s)) then u(., s, x) is a strict solution of the Cauchy prob-
lem (1.2.1)(s,x,A,J).

Proof Observe that

Au(·, s, x) = AU(-, s)x + (AU * 1)(., s) (1.2.9)

by (1.2.3) and the closedness of A(t). Estimate (1.2.3) easily implies

(1.2.10)

Since AU(t, s)x = AUA-l(t, slY, where we have set y:= A(s)x E E, it follows from
(1.2.8)-(1.2.10) that the assertion is true. _

The above results are due to the author and were first published in [Ama88c].
198 IV Variable Domains

1.3 Estimates for Evolution Operators


Let (Eo, E 1 ) be a densely injected Banach couple, J a perfect subinterval of jR+
containing 0, and p E (0,1). Suppose that

A c CP(J, H(El' Eo)) and there are constants M,7] E jR+ and
(0, 7r /2) such that
() E

[Ajp,J :::; 7] , AEA,

that I;,J c p(-A(s)), and that (1.3.1)

IIA(s)II.c(E"Eo) + (1 + IAI)l~j I (A + A(s)r111.c(Eo,E j ) :::; M

for (s, A, A) E J X I;,J x A and j = 0, 1.

Note that this assumption coincides with (1I.4.2.1). Hence Theorem 11.4.4.1 guar-
antees the existence of a unique parabolic evolution operator UA for A E A pos-
sessing El as regularity subspace. In the following, we fix p as in (11.4.4.2) and
suppose that
(1.3.2)
and that (-, .), is an admissible interpolation functor. Then we put E, := (Eo, Ed,
and derive estimates for the restriction of A to E,.

1.3.1 Lemma Given A E A,

AUA E Si(E, , 1) n Si(E" Eo, 1 - ')') .

More precisely,

fore> 0, (t, s) E J1, and A E A.

Proof Throughout this proof we use the notations of Section 1I.4 and omit the
index A.
First we deduce from (1I.4.2.3) by interpolation that

(1.3.3)

and
(1.3.4)
and that

(t - s)'+j II [A(TWe~(t~s)A(T) 11.c(Eo,E-y) + (t - s) Ilola(t, s)II.c(Kr) :::; c (1.3.5)


IV.1 Higher Regularity 199

for (t, s) E J~, 1" E J, and A E A. From (1I.4.2.2) and (11.4.2.3) it also follows that

IIA(s)e--tA(S)xllo = Ile-tA(s) A(s)xllo ::; eM IIxl1 1 , x E E1, A EA.


Thus, by using (11.4.2.2) again,

1181 a(t, S)II.CCE1,Eo) + (t - s) 1181 a(t, s)llc(Eo) ::; e

for (t, s) E J~ and A E A. Hence, by interpolation,

and

(t - s)l-')' 1181 a(t, S)IIC(E-y, E)


a ::; e , AEA, (t,s)EJ~. (1.3.6)

Also by interpolating we obtain from Lemma 11.4.3.1 that

wE fi.(E')', Eo, 1 - , - p) (1.3.7)

and that
(t - s)l-')'-p IIw(t, s)llqE-y,Eo) ::; e(E)e(J-£+c)(t-s) (1.3.8)
for E > 0, (t, s) E J~, and A E A.
Put
ft(t,s) := e-(t-s)A(t)w(t,s) .
Then we deduce from (1I.4.2.2), (11.4.2.3), (1.3.3), (1.3.5), and (1.3.7), (1.3.8) that

ft E fi.(E')', 1 - p) n fi.(E')', Eo, 1 - , - p) (1.3.9)

and

(t-S)l- p 11ft (t, s) Ilc(E-y) +(t-S)l-')'-p 11ft (t, s )llc(E-y,Eo) ::; e(E)e(J-£+c)Ct-s) (1.3.10)

for E > 0, (t,s) E J~, and A E A.


By interpolation we obtain from (1.3.1) that

(s, A, A) E J X ~1? X A . (1.3.11)

Using this estimate, an obvious modification of the proof of Lemma 11.4.3.2 gives

e E fi.(Eo, E')', 1 +, - p) (1.3.12)


and
Ile(t, s)IICCEo,E-y) ::; e(t - s)p-')'-l (1.3.13)
for (t, s) E J~ and A E A. Now we obtain from (1.3.7), (1.3.8), (1.3.12), (1.3.13),
and (11.3.1.8), (11.3.1.9) on the one hand, and from (1.3.7), (1.3.8), Lemma 11.4.3.2,
200 IV Variable Domains

(11.3.1.8), (11.3.1.9) on the other hand, together with (11.3.1.1), (1I.3.1.2), that

(1.3.14)

and

for c > 0, (t,s) E J't:", and A E A.


By interpolation we infer from (11.4.3.19) that

(T - S)l-"'/ //k(t, s) - k(T, S)/IC(E'Y,Eo ) ::; c(t - T)P

for 0 ::; s < T < t. Thus it is a consequence of (11.4.3.22), (11.4.3.21), (1.3.8), and
(11.4.3.4) that

/Iw(t, s) - W(T, S)/I.c(E'Y,Eo)


::; c(c,(3) [(t - T)P(T - s)"'l-l + IT (t - T)i3(T _ay-i3- 1(O' - s)"'l+p-1 dO'

+ it (t - O')p-1(O' - s)"'l+p-1 dO']e(fL+£)(t-S)

for c > 0, 0 < (3 < p, and 0 ::; s < T < t. Hence

/IW(t,s) - W(T, S)/IC(E'Y,Eo)


(1.3.16)
::; c(c, (3) [(t - T)i3(T - s)"'l-l + J(t, T, s)] e(fL+£)(t-s)
for c > 0, 0 < (3 < p, and 0 ::; s < T < t, where

if 'Y + p - 1 2:: 0 ,
if 'Y + p - 1 < 0 .

Fix (3 E (,)" p) and put

h(t, s):= It A(t)e-(t-T)A(t) [w(t, s) - W(T, s)] dT .

Then we infer from (1.3.3) and (1.3.16) that

/Ih(t, S)/IC(E'Y) ::; c(c) It (t - T)-l-",/ [(t - T)i3(T - s)"'l-l + J(t, T, s)] dre(fL+E)(t-s)

::; c(c) [(t - S )13- 1 + (t - s ?p-1] e(fL+£)(t-s)


IV.1 Higher Regularity 201

for € > 0 and (t, s) E J~. Similarly, (II.4.2.3) and (1.3.16) imply

Ilh(t, S)II.c(E.."Eo) ::::; c(€) [(t - s)1'+,8-1 + (t - s)1'+2 p -1]e(/L+e)(t-s)

for € > 0 and (t, s) E J~. Hence

(1.3.17)

and

(t - S)1-')' IIh(t, s) 11.c(E",) + (t - S)1-2')' Ilh(t, s) 11.c(E.."Eo) ::::; c(€)e(/L+e)(t-s) (1.3.18)

for € > 0 and (t,s) E J~.


Note that Lemma II.4.3.4 and (II.4.3.26) and (II.4.3.30) imply

(1.3.19)

Hence the assertion follows from (1.3.4)-(1.3.6), (1.3.9), (1.3.10), (1.3.14), (1.3.15),
and (1.3.17), (1.3.18) .•

1.3.2 Lemma Given A E A,

Moreover,
IIAUAA- 1(t, S)II.c(E",) ::::; c(€)e(/LP!(P-"')+e)(t-s)

for € > 0, (t, s) E J~, and A E A.

Proof Put

bACt, s) := e-(t-s)A(t) , hA(t, s) := bACt, s) [A(t) - A(s)]

for (t, s) E Ji:!,. and A E A. In the following we again omit the index A. Then
(11.4.2.3) and (1.3.3) imply
hE~(E1,1-p) (1.3.20)
and
Ah E ~(El, E,)" 1 + 'Y - p) , (1.3.21)
respectively, together with the estimate

(1.3.22)

for (t,s) E J~ and A E A.


202 IV Variable Domains

Note that, thanks to (1I.2.2.3),

U=b+h*U. (1.3.23)

From (1.3.1) and (1I.4.2.3) it follows that

AbA- l E R(Eo, 0) (1.3.24)

and
(t,s)EJi:,., AEA. (1.3.25)
Using (1.3.21), (1.3.23), and (1.3.24) we find

AUA- l = AbA- l + A(h * U)A- l = AbA- l + (AhA- l ) * (AUA- l ) , (1.3.26)

where the last 'convolution' is meaningful since AhA- l E R(Eo, 1- p) by (1.3.1)


and (1.3.20).
Observe that
AbA- l = a + (Ab - aA)A- l (1.3.27)
and (cf. the proof of Lemma II.4.3.2)

(Ab - aA)A- l
1.
= -2
7r2
r AeW -
lr
s) (A + A(t)r l [A(t) -A(S)](A+A(s)rlA-l(s)dA. (1.3.28)

From (1.3.1) and (II.4.2.2) we deduce that

I (A + A(s)) -lxll l ::; M IIA(s)(A + A(s)) -lxll o= M II (A + A(s)) -1 A(s)xllo


::; (1 + IAI)-l M211xlll

for x E E1 and A E ~{I. Thus

s E J, AE ~{I, A E .A. (1.3.29)

By interpolation it follows from (1.3.1) and (1.3.29) that

for s E J, A E ~{I, and A E A. By means of (1.3.11) and (1.3.30) we can estimate


the integrand in (1.3.28) in the norm of £(E'Y) by

c(t - sY IAI'Y- 1 eRe.\(t-s) .

Hence Lemma II.4.1.1, (II.3.U), and (II.3.1.5) imply

(Ab - aA)A -1 E Jt(E'Y' 'Y - p) '-+ C( J6., £(E'Y)) n ft(E'Y' 0) (1.3.31 )


IV.l Higher Regularity 203

and
(t - s)'Y-P II(Ab - aA)A- 1 (t, s)II.c(E-y) ::;; c (1.3.32)
for (t, s) E J1. and A E A. Lemma II.4.2.1 implies, by interpolation, that

(1.3.33)

and
Ila(t, s)II.c(E-yl ::;; C , (t,s)Eh, AEA. (1.3.34)
Thus we obtain from (1.3.27) and (1.3.32), (1.3.34) that

AbA- 1 E C(h,£s(KJ) nJt(E'1'O) (1.3.35)

and
IIAbA- 1 (t, s)II.c(E-y) ::;; c(l V (t - s)P-'1) (1.3.36)
for (t, s) E h and A E A. Thanks to (1.3.21), (1.3.22), and (1.3.30)

AhA- 1 E Jt(E'1' 1 + 'Y - p) . (1.3.37)

and

(t, s) E J1, A EA. (1.3.38)

Using (1.3.26), (1.3.35)-(1.3.38), and Theorem II.3.2.2 we find

AUA- 1 E Jt(E'1'O) (1.3.39)

and
(1.3.40)
where l' E Jt(E'1' 1 + 'Y - p). Note that (1.3.38) and Lemma II.3.2.1 imply

(t - s)1+'1-P IIT(t, s)II.c(E-y) ::;; cefLP/(p--Y)(t-s)

for (t, s) E J1. and A E A, provided we increase the constant c(p) in (II.4.4.2)
appropriately. Now it follows from (11.3.1.9) and (1.3.35), (1.3.36) that

(1.3.41 )

and that

(t - S )'1-p Ill' * (AbA -1 )(t, s) 11.c(E-y) ::;; cefLP/(P--Y)(t-s) (1.3.42)

for (t, s) E J1. and A E A. Hence the assertion follows from (1.3.40), (1.3.35),
(1.3.36), and (1.3.41), (1.3.42) .•
204 IV Variable Domains

1.4 Evolution Operators on Interpolation Spaces


Let (Eo, E 1 ) be a densely injected Banach couple, J a perfect subinterval of jR+
containing 0, and p E (0,1). Suppose that

O<,<p (1.4.1 )

and (', .), is an admissible interpolation functor, and let E, := (Eo, Ed,. Given a
linear operator B in Eo, denote by B, its E,-realization.
Now it is easy to prove the following basic theorem which is the main result
of this section.

1.4.1 Theorem Let assumptions (1.3.1) and (1.4.1) be satisfied. Then, given
A E A, the E,-realization, UA", of the parabolic evolution operator UA for A is
the parabolic evolution operator for the E,-realization A, of A. It possesses E1 as
regularity subspace and satisfies

(1.4.2)

Furthermore,

(t - s)i IIA:;,UA,,(t, S)II.C(E-y) + (t - S)l-, IIUA,,(t, s)IIC(E-y,E, )


(1.4.3)
~ c(c)e(Jl+c)(t-s)

and
(1.4.4)
for c > 0, (t, s) E Jb., A E A, and j = 0, 1, where we have put J-t := CoT/lip for a
suitable positive constant co.

Proof Note that the E,-realization of UA is simply the restriction of UA to E"


considered as linear operator in E,. Moreover, x E dom(A,) iff x E El and Ax E E,.
Using these facts, (1.4.2)-(1.4.4) with j = 1 follow immediately from Lemmas 1.3.1

°
and 1.3.2, from (II.4.2.2), and from im(UA(t, s») eEl for (t, s) E Jb.. By interpola-
tion we deduce from (II.4.4.3) that (1.4.3) is true for j = as well. From (II.4.4.1)
and (1.4.3) we easily obtain - thanks to the density of E1 in E, - that

(1.4.5)

which is condition (II.2.1.2) with E := E, and F := E 1.


By restricting 8 1UA(t, s) = -AUA(t, s) for (t, s) E Jb. to E, we see that

(t,s) E Jb.. (1.4.6)


Note that

UA(t, s) = UA(T, s) + it 8 1 UA (0", s) dO" , ° ~ s < T < t, t E J , (1.4.7)


IV.l Higher Regularity 205

in L(E). Thus, by restricting (1.4.7) to E"( and using (1.4.6), we infer

UA,"(t,8) = UA,,,(T, 8) -it APA,"(O',S)dO' , 0:58<r<t, tEJ,

where, thanks to A"(UA,"( E Jl.(E"(, 1), the integral converges in L(E"(). This implies

UA,"(·,8) E al (J n (8,00), L(E,,()) and 8l UA,"(', s) = -A"(UA,"(', 8)

for 8 E J, that is, conditions (11.2.1.6) and (11.2.1.7), respectively.


Recall that UA(t,·) E al([O,t),Ls(EbEO)) and that

(1.4.8)

This shows that

t E J ,

thanks to the continuity properties of A and UA. Since, given x E El,

UA(t, r)x = UA(t, O')x + iT lhUA(t, r)x dr, °:5 0' < T < t, t E j , (1.4.9)

we infer - by restricting (1.4.9) to E"( - from (1.4.5), El ~ E,,(, and (1.4.8) that

for t E j, which are conditions (11.2.1.8) and (11.2.1.9), respectively. Conditions


(11.2.1.3) and (11.2.1.4) are obvious, and (11.2.1.5) is a consequence of Lemma 1.3.1
as has been observed in the beginning of this proof. Hence UA,"( is a parabolic
evolution operator for AI' with regularity subspace El, and the theorem has been
proven.•

1.4.2 Corollary Suppose that 0 < 'Y < P < 1 and

Then the E"(-realization of the parabolic evolution operator for A is the parabolic
evolution operator for the EI'-realization of A. It has El as regularity subspace.
Proof This follows - similarly as in the proof of Corollary II.4.4.2 - by applying
Theorem 1.4.1 to an arbitrary compact subinterval I of J containing 0 and to 0' + A
for a suitable 0' E IR (depending on I and A, of course) .•

1.4.3 Remark Given the hypotheses of Theorem 1.4.1, the EI'-realization A"(t)
of A(t) is densely defined for t E J.
206 IV Variable Domains

Indeed, let E 2(t) := (dom(A2(t)), IIA2(t)·II). Then E2(t) is a Banach space


such that E 2 (t)!:.. El !:.. Eo. Let Al(t) be the Errealization of A(t). It follows
that Al(t) E ,Cis(E2(t),E l ) and

Al(t) C A(t) E ,Cis(El , Eo) .

From this we deduce by interpolation that A,(t) E ,Cis(EH,(t),E,), where


d d
EH,(t) := (El' E 2(t)), '-+ El '-+ E, .

Hence dom(A,(t)) = EH,(t) is dense in E, (cf. Section V.1.5 for more details). _

Now we give a variant of Theorem 1.4.1 in which the interpolation space E,


is replaced by an intermediate space X between Eo and El that is not supposed to
be an interpolation space. Of course, X has to satisfy some additional assumptions.

1.4.4 Theorem Let assumption (1.3.1) be satisfied. Suppose that

°< 'Y- ::; 'Y+ <p, (1.4.10)

that (',' ),_ and (',' ),+ are admissible interpolation functors, and that X is a
Banach space such that
(1.4.11)
where E,± := (Eo, E l ),±. Let Ax(t) be the X -realization of A(t) for A E A and
t E J and suppose that Ax(t) is densely defined and

(1.4.12)

for A E A, t E J and>' E ~t9. Then UA,X, the X -realization of UA, is a parabolic


evolution operator for Ax. It possesses El as regularity subspace and satisfies

(1.4.13)

and
IIAxUA,X A Xl (t, s) IIL(X) ::; c(E)e(/LP/(P-'Y+)+c)(t-s) (1.4.14)
for E > 0, (t, s) E Ji:" and A E A. Moreover,

(t - 8)j IIA~UA,X(t, 8)llc(x) + (t - 8)1-,- IIUA,X(t, 8)11


(1.4.15)
::; c(E)e(/L+c)(t-s)

forc>o, (t,8)EJi:" AEA,andj=O,l.

Proof This follows by obvious modifications of the proof of Theorem 1.4.1 which
we leave to the reader (cf. [Ama88c, Proof of Theorem 2.2] and [Ama90a, Ap-
pendix AD .•
IV.1 Higher Regularity 207

Theorem 1.4.1 - and, of course, Theorem 1.4.4 as well - guarantees the


existence of a parabolic evolution operator for AI' on E,)" which is unique by Re-
mark II.2.1.2(c). It is important to note that dom(A')'(t)) is, in general, not inde-
pendent of t. Thus the situation is different from the one considered in Chapter II.
The above theorems show that the regularizing properties of analytic semi-
groups carry over, to some extent, to parabolic evolution operators. The amount
of regularization is restricted by the smoothness with respect to t as follows from
assumption (1.4.1).
The proof of Theorem 1.4.1 follows the proof of [Ama88c, Theorem 2.2] (also
cf. [AmagOa, Appendix AD.

1.5 The Cauchy Problem


Let (Eo, Ed be a densely injected Banach couple, let J be a perfect subinterval
of lR containing its left endpoint s := min J, and suppose that

(1.5.1)

for some p E (0,1). Also suppose that (., .)1) is for each () E (0,1) an admissible
interpolation functor, and put EI) := (Eo, E 1 )e. Lastly, suppose that 0 < I < p
and 0 < c < 1 - I and

f E Ce(J,E')') + C(J,E,),+e) . (1.5.2)

Consider the Cauchy problem

u+A(t)u=f(t) , tEJ\{s}, u(s) = x, (1.5.3)

where x E Eo.
Now we are ready to prove Theorem 11.1.2.2 which, for the reader's conve-
nience, we recall here:

1.5.1 Theorem If f satisfies (1.5.2), the Cauchy problem (1.5.3) has a unique
solution
u E C(J, Eo) n C(J\{s}, E 1 ) n C 1 (J\{s}, E')') .
It is also a solution of the parabolic equation

1; + A')'(t)v = f(t) , tEJ\{s} , (1.5.4)

in E,),' where AI' is the E,),-realization of A. If x E E')' then u E C(J, E,),)' and if
x E dom(A')' (s)) , it is a strict solution of (1.5.4).

Proof By an obvious translation argument we can assume that s = O. Moreover,


it suffices to consider the case that J is compact.
208 IV Variable Domains

Let U be the parabolic evolution operator for A, whose existence is guaran-


teed, thanks to (1.5.1), by Corollary II.4.4.2. Since J is compact, Proposition 1.1.4.2
guarantees the existence of constants w > 0, K, 2: 1, and a E lR such that

(1.5.5)

Since J is compact, A is uniformly p-HOlder continuous. Hence condition (11.5.0.1)


is satisfied. Thus we deduce from Lemma 11.5.1.3 that

(t - s )1-')' IIU(t, s) 11.c(E., ,Ed:::; C , (t,s) E Jf:,'

This implies, thanks to (1.5.1), that

IIU(t, s)II.c(E."E1 ) + IIAU(t, s)II.c(E."Eo) :::; c(t - s)1'-l , (t,s) E Jf:, .


Since f E C(J, E')') by (1.5.2), Theorem 1.2.1 guarantees that (1.5.3) has a unique
solution
U E C(J, Eo) n c 1 (i, Eo) n c(i, E')') .

As U E c(i,Ed by (II.1.2.3), it remains to show that U E C 1 (j,E,),) and that


U is a solution of (1.5.4) which satisfies U E C( J, E')') if x E E')' and which is strict
if x E dom(A')'(O)).
Thanks to (1.5.5), Corollary 1.1.4.3, and the compactness of J there is a con-
stant Wo such that A:= {wo + A} satisfies condition (1.3.1). By Remark II.2.1.2(e)
we can replace A by Wo + A in the remainder of this proof, that is, we can assume
that A := {A} satisfies condition (1.3.1).
By (1.5.2) there exists a decomposition f = It + h with It E C"(J, E')') and
hE C(J,E,),+,,), It follows from Remark II.2.1.2(b) that
U = Vo + V1 + V2 , (1.5.6)

where
Vo := U(., O)x and Vj:= U * fJe 0) , j = 1,2.
Note that Va(T) EEl '---+ E')' and that

O<T<t, tEJ,
where U')' denotes the E')'-realization of U. From this and from Theorem 1.4.1 it
follows that Va is a solution of the parabolic equation

which belongs to CU, E')') if x E ET From Theorems 1.2.2 and 1.4.1 we also deduce
that Va is a strict solution of the Cauchy problem

v+A')'(t)v=O, t E j , v(O) = x,

provided x E dom(A')'(O)).
IV.1 Higher Regularity 209

Observe that Vj is the solution of the Cauchy problem

v + A(t)v = fj(t), t E j , V(O) = 0 (1.5.7)j

for j = 1,2. Since h E G"(J,Ey) '---> G"(J,Eo), it follows from Theorem 11.1.2.1
that VI is a strict solution of (1.5. 7h. Hence

tEJ . (1.5.8)

From Lemma 1.3.1 and h E G"(J, E'Y) we deduce that

IIA(t)U(t,r) [h(t) - her)] liE s:: e(t -. r),,-l , (t, r) E Ji:.. ,


"Y

and
(t,r) ...... A(t)U(t,r)[h(t) - her)]) E G(Ji:..,E'Y) .
This implies that the first term on the right-hand side of

A(t)V1(t) = fat A(t)U(t, r) [her) - h(t)] dr + A(t) fat U(t, r)h(t) dr (1.5.9)

is a continuous function of t E J with values in ET Let t E j be fixed. Since


U E st(E'Y' E 1 , 1- 'Y) by Lemma 1.3.1 and (1.3.1), we see that

t-f, ~
9i;(t) := Jo U(t, r)h(t) dr g(t):= fat U(t, r)h (t) dr

in E1 as E-+ 0 in (0, t). Hence A(t) E £(E 1, Eo) implies A(t)gf,(t) -+ A(t)g(t) in Eo.
Note that (1.3.9), (1.3.12), (1.3.14), (1.3.17), and (1.3.19) imply

AU(t, r) = A(t)e-(t-T)A(t) + B(t, r) = 8Te-(t-T)A(t) + B(t, r) , (t,r) E Ji:.. ,

where
(1.5.10)
Hence

A(t)gl;(t) = Jo
t- E AU(t, r)!I(t) dr
(t-I;
= [e-f,A(t) - e-tA(t)l!I (t) + Jo B(t, r)!I (t) dr ,

which shows that

A(t)gf.(t) -> [1 - e-tA(t)lh (t) + .f B(t, r)h (t) dr


210 IV Variable Domains

in Eo. Thus

(1.5.11)

for t E f From Lemma 11.4.2.1 and the density of El in E"( we easily deduce that
[t ....... e-tA(t)] E C(J,.cs(E"()). Now we infer from (1.5.10) and (1.5.11) that

(t ....... A(t) lot U(t, 7)ft(t) d7) EC(J,E"().

Hence (1.5.8) and (1.5.9) imply VI E C(J, E"(). By Theorem 1.4.1,

U"( E C(h, .cs(E"()) ,

which trivially gives Vj E C(J,E,,(), j = 1,2. Thus Proposition 1.1.3 guarantees


that VI E Cl(J, E"().
We use the notations of the proof of Lemma 1.3.1. Then it follows from
(1.3.19) that
AU = -ala - I, (1.5.12)
where 1:= It + 12 + h Thanks to (1.3.9), (1.3.14), (1.3.17), and (II.3.1.2),

(1.5.13)

Moreover, from Lemma II.4.2.1 we easily deduce by interpolation that

(1.5.14)

Thus, by (1.5.12)-(1.5.14), (II.3.1.2), and E,,(+€ '-+ E,,(,

AU E Jt(E,,(+€, E,,(, 1 - 'Y /\ f) .


Consequently,

A(t)V2(t) = fat AU(t, 7)12(7) d7 , t E J ,

and AV2 E C(J,E"(). Hence 12 - AV2 E C(J,E"() and V2(t) = f2(t) - A(t)V2(t) for
t E j, together with Proposition 1.1.3, guarantee that V2 E Cl(J, E"(). Now (1.5.6)
and the above considerations imply that u E C l (i, E"() and that u E C l (J, E"() if
x E dom(A"((O)). This and

show that u(t) E dom(A"((t)) for t E j, so that u is a solution of (1.5.4), which is


strict if x E dom(A"((O)) .•
IV.2 Constant Interpolation Spaces 211

Theorem 1.5.1, which is due to the author (cf. [Ama87, Section 6]), is an
important regularity theorem. It implies that the solution u of the evolution equa-
tion (1.5.3), that a priori belongs to the class c(i, Ed n C 1 (j, Eo), in fact sat-
isfies U E C(j, KJ and Ayu = f - it E C(j, E')'). Since, in general, E')' and, con-
sequently, dom(A')'(t)) are smaller spaces than Eo and E l , respectively, we see
that, given the hypotheses of Theorem 1.4.1, the solution U of (1.5.3) possesses
better 'spacial regularity properties' than the ones we could deduce in Chapter II.
This regularizing effect will be the basis for 'abstract bootstrapping arguments'
for quasilinear parabolic evolution equations showing that suitable 'weak solutions'
are already classical ones.

2 Constant Interpolation Spaces


In the preceding section we proved the existence of a parabolic evolution operator
and a solvability theorem for parabolic Cauchy problems in the case where the
domains of the generators A(t) may depend on t. This was done by 'lifting' a 'con-
stant domain case' to a suitable interpolation space. These results are important
for their regularity aspects, but they depend on the fact that we can 'start' with
a 'constant domain case'. It is the purpose of this section to study linear parabolic
Cauchy problems in the case where the domains of A(t) may vary with t but a
suitable interpolation space is constant.

2.1 Semigroup and Convergence Estimates


Let E be a Banach space. Given M > 0, we denote by

the set of all densely defined linear operators A in E such that [Re z 2:: 0] C p( - A)
and
(1 + 1>.1) 11(>' + A)-lIIC(E) :::; M , ReA 2:: °. (2.1.1)
Note that C C(E).
°
Q(M

Let M > be arbitrarily fixed. It follows from Corollary II.6.1.3 that there
exist 1J:= 1J(M) E (0,n/2) and K:= K(M) such that ~1? C p(-A) and

(2.1.2)

Each A E Q( M is the negative infinitesimal generator of a strongly continuous an-


alytic semigroup {e- tA ; t 2:: O} on E, and

t>o, kEN, (2.1.3)


212 IV Variable Domains

r being a piece-wise smooth simple curve running in t{! from ooe- i (cp+1r/2) to
ooe i (cp+7r/2) for some 'P E (O,19J. Note that 21M C PM(E), so that the fractional
powers A Z are well-defined for A E 21 M and z E C.

The following lemma gives an extension of (2.l.3) to fractional exponents.

2.1.1 Lemma Given 0: E 1R+,

A"'e- tA = ~
21fZ
r(-A)"e ),(>. + A)-I dA ,
Jr
t t > 0, A E 21M . (2.l.4)

Proof Fix t > 0 and kEN with 0: < k, and define holomorphic functions 'Pj on
C\ (-IR+) by 'PI (z) := z,,-k and 'P2 (z) := zke- tz , respectively. Given 8 E (0, k - 0:),

Izlo 'Pj (z) -+ 0 as Izl -+ 00 in [I arg zl ::; 1f /2 - EJ


for each E E (0, 1f /2). Thus

j = 1,2,

are well-defined in £( E) for A E 21 M, and the proof of Lemma 111.4.6.1 shows


that 'PI (A)'P2(A) = 'PI'P2(A). Also note that 'PI(A) = A",-k by Theorem 1II.4.6.5,
and 'P2(A) = Ake- tA by (2.l.3). Now the assertion follows since A",-k Ak = A'" by
Theorem 111.4.6.5 .•

2.1.2 Corollary Suppose that 0: E 1R+. Then


(t I-> A"'e- tA ) E C(lR+,L(E))
and
t > 0, A E 21M .

Proof This is an obvious consequence of (2.1.2), (2.1.4), and Lemma II.4.l.l. •

Given A E 21M, let {Ae ; E > O} be the Yosida approximation of A. Recall


from Lemma 1I.6.l.2 that
(2.l.5 )
From this, (2.l.2), and Corollary 11.6.l.3 we deduce that there exist 19 E (0, 1f /2)
and c > 0 such that I;{! C p( -A) n p( -Ae) and
(2.1.6)
for A E I;{!, E > 0, and A E 21 M . Using these facts we now prove some simple but
important estimates and convergence results.
IV.2 Constant Interpolation Spaces 213

2.1.3 Lemma Given A E Q(M and A E E~,

in .L(E)

as c -+ O. The convergence is uniform with respect to A E Q(M and A E E~.

Proof Note that (cf. (11.6.1.12))

(2.1. 7)

for A E E~, c > 0, and A E Q(M. Hence

(A + A€)~l - (A + A)-l = 1: cA + ((1=t-1cA)2 - 1) (1: cA + A)-l


+ (A _ _ A__ ) (_A_ + A)-\A + A)-l
1 + cA 1 + cA

and, consequently, thanks to (2.1.6),

Since 11 + cAl:::: cos 19 > 0 for A E E~, the assertion follows .•


By means of this lemma it is now easy to prove the following convergence result:

2.1.4 Lemma Given a E lR+ and A E Q(M,

in .L(E)

as c -+ 0, locally uniformly with respect to t E lR+.


Proof From (2.1.5) and Lemma 2.1.3 we infer the existence of co °
> such that

0< c < cO . (2.1.8)

Thus Lemma 2.1.1 implies

t>o, c>O.

Hence the assertion follows from Lemma 2.1.:3 and the dominated convergence
theorem. _
214 IV Variable Domains

2.2 Assumptions and Consequences


Throughout the remainder of this section we fix a Banach space Eo := (Eo, 11·110),
a perfect subinterval] of 1Ri.+ containing 0, and a positive constant M.

Given A: ] ----+ I.2lM(Eo), we put

El(A(s)) := (dom(A(s)), IIA(s)·llo) , s E] . (2.2.1)

Note that (Eo, El(A(s))) is a densely injected Banach couple and that

A(s) E 1i(El(A(s)), Eo) , s E] . (2.2.2)

We assume that
(i) A is a nonempty set of maps A: ] ----+ I.2lM(Eo).
(ii) There exist constants e E (0,1) and K, 2: 1,
an admissible interpolation functor (', ')0,
and a Banach space Eo:= (Eo, 11·110) such that

(2.2.3)
and

xE Eo,

for s E ] and A E A.
(iii) °
There exist p E (1 - e, 1) and T) 2: such that
A- 1 E CP(J,£(Eo,Eo)) and [A-1]p,J':::: T) for A E A.

Observe that (2.2.3(ii)) guarantees that the interpolation space Eo (A(s)) between
Eo and D (A( s)) is independent of s E ] and A E A, except for (uniformly) equiv-
alent norms.
Now we deduce some consequences of assumption (2.2.3). First we recall from
(2.1.6) that there exists {) E (0, 7f /2) such that

~1'i C p( -A(s)) n p( -A€(s)) , sE], c:>0, AEA,

and estimate (2.1.6) is valid for A(s), uniformly with respect to s E ] and A E A.

2.2.1 Lemma Given A E A,

(,\+A(S))-lE£(Eo,Eo) and A(s)(,\+A(s))-lE£(Eo,Eo) (2.2.4)

for,\ E ~1'i and s E J. Moreover,


IV.2 Constant Interpolation Spaces 215

for,\ E ~iI, S E J, and A E A. In addition,

for,\ E ~iI, r,s E J, and A E A.

Proof Since
A(s)(,\+A(s)r 1 = 1_,\(,\+A(s))-1, (2.2.7)
we deduce from (2.1.6) that

and

for ,\ E ~iI) S E J, and A E A. Hence (2.2.4) and (2.2.5) follow by interpolation,


thanks to (2.2.3(ii)).
Note that

(,\ + A(r)r 1 _ (,\ + A(s))-l


= (,\ + A(r)) -1 A( s) (,\ + A(s)) -1 _ A(r) (,\ + A(r)) -1 (,\ + A( S))-l
= A(r)('\ + A(r))-1 [A- 1 (r) _ A- 1(s)JA(s)(,\ + A(s)) -1 .

Thus (2.2.6) is a consequence of (2.2.5), assumption (2.2.3(iii)), and (2.2.9) .•

Given s E J and A E A, the fractional powers k'(s) of A(s) are well-defined


for a E IR, and

O<a<I,

by (1.2.9.6). Hence (1.2.11.4), (1.2.5.2), and (2.2.3(ii)) imply Eo '--+ D(AQ(s)) for
a < B. This shows that AQ(s) E £(Eo, Eo). The following lemma gives more precise
information on these operators.

2.2.2 Lemma Suppose that 0 < a < B. Then

(A f-> AQ) E B(A,B(J,£(Eo, Eo))) . (2.2.10)

Moreover, given A E A,

and
(2.2.11)
as c -+ 0, uniformly with respect to s E J.
216 IV Variable Domains

Proof By Lemma 2.2.1 and (2.2.9),

lit" (t + A(s») -2 A(S)Xllo = t" I (t + A(s») -1 A(s)(t + A(s») -lXllo (2.2.12)


~ ct,,-(J-l Ilxll(J
for x E El(A(s»), s E J, t > 0, and A E A. By Theorem III.4.6.5(ii)

sina7l'l°O -2
A"(s)x = - - t"(t + A(s») A(s)xdt, x E El (A(s»). (2.2.13)
a7l' 0

Thus we infer from (2.2.12) and IIA- 1 (s)II.c(Eo ) ~ c that A"(s) E C(E(J, Eo), that
(2.2.13) is true for x E E(J, and, consequently, that

IIA"(s)II.c(Eo,E o) ~ c , SEJ, AEA. (2.2.14)

This proves (2.2.10).


It follows from (2.2.7) that, given A E A,

A e(s)(t+A e(s»)-l_ A (S)(t+A(s»)-l = t[(t+A(S»)-l - (t+Ae(s»)-l] .

Thus, thanks to (2.1.6),

I (t + Ae(s») -2 Ae(s) - (t + Ae(s») -1 A(s)( t + A(s») -111.c(Eo )


~ cll(t+A(s)r 1 - (t+Ae(s»)-lll.c(E o)
for s E J and t ~ O. Similarly, by (2.2.9),

11[(t+AE(s»)-l - (t+A(s»)-l]A(s)(t+A(s»)-III.c(Eo)

~ c II (t + Ac(S») -1 - (t + A(s) r 1 11.c(E )


o

for s E J and t ~ O. Hence we deduce from Lemma 2.1.3 that, given x E El (A(s») ,

as c ~ 0, uniformly with respect to s E J and t in bounded subintervals of ~+.


Since, thanks to (2.1.8), estimate (2.2.12) holds for Ae also, uniformly with respect
to 0 < c < co, we infer from (2.2.13) and the corresponding formula for A~(s)x that
(2.2.11) is true.
By (2.1.7),
A;l = c + A-I. (2.2.15)
From this, E(J<-+Eo, and (2.2.3(iii» it follows that A;l ECP(J,C(Eo»). Thus
Ae E CP ( J, C( Eo») by the analyticity of the inversion map B 1-4 B-1 from Caut( Eo)
IV.2 Constant Interpolation Spaces 217

into itself. This implies A~ E C (J, £( Eo)); hence A~ E C (J, £( Eo, Eo)). Thus the
uniformity of the convergence in (2.2.11) ensures A'" E C(J, £s(E(J, Eo)) . •

2.2.3 Corollary Suppose that 0 < 0: < B. Then


[(r, s) f-7 A<>(r)(A-l(r) - A- 1 (s))] E C(J x J, £(Eo))

and

r,s E J, A EA.

Moreover, given A E A,

as E -> 0, uniformly with respect to r, s E J.

Proof These assertions are immediate consequences of (2.2.3(iii)), Lemma 2.2.2,


and (2.2.15) .•

Next we derive continuity properties of A<>(s)e-tA(s) as a function of s E J


and t E JR+.

2.2.4 Lemma Given 0: E JR+,

(2.2.16)

for r, s E J, t > 0, and A E A. Moreover,

Proof By Lemma 2.1.1

A"'(r)e-tA(r) _A"'(s)e-tA(s) = ~
27ft
1(
r
_A)"'e tA [(A+A(r)r 1 - (A+A(s))-l] dA .

Hence the first assertion follows from Lemmas 2.2.1 and II.4.1.1.
Given r, s E J and 0 < T < t < 00,
A'-"(r)e-·tA(r) _ A<>(s)e-TA(S)
(2.2.17)
= A<>(r)e-tA(r) _ A"'(s)e-tA(s) + A<>(s)[e-tA(s) _ e-TA(S)] .

Since
(2.2.18)
218 IV Variable Domains

we infer from Corollary 2.1.2 that

IIAa(s)[e-tA(s) - e-TA(S)lll.c(EO ) ::; C(T- a - C a ) (2.2.19)

for s E J and 0 < T <t< 00. Now the second assertion is a consequence of (2.2.16)-
(2.2.19) .•

Lastly, we prove the strong continuity of (s, t) f-+ e-tA(s) on J x jR+.

2.2.5 Lemma Given A E A,

Proof Thanks to Lemma 2.2.4 we know that the assertion is true if we replace
J x jR+ by J X lR+. Thus it remains to show that, given So E J,
e-tA(s) -+ 1 in Cs(Eo) (2.2.20)

as (s, t) f-+ (so, 0). Fix a E (0,0) and note that (2.2.18) and Lemma 2.2.2 imply

e-tA(s) _ e-TA(S) = - i t A1-"'(s)e-rrA(s) A"(s) d(J in C(Eo,Eo)

for s E J and 0 < T < t < 00. From this, Corollary 2.1.2, and the strong continuity
of the semigroup {e-tA(s) ; t ~ O} on Eo we infer that

Ile-tA(s)x - xllo ::; ct a Ilxll li , x E Eo, s E J, t >0 .

Now the uniform boundedness of the semigroup {e-tA(s) ; t ~ O}, s E J, on Eo


and the density of Eo in Eo imply (2.2.20) .•

2.3 Construction of Evolution Operators


Suppose, for the moment, that A E CP(J, C(Eo)) for some p E (0,1). Then Re-
mark I.1.2.1(b) and Corollary 1I.4.4.2 imply the existence of a unique parabolic
evolution operator UA for A, and UA solves the integral equations (II.2.2.2) and
(1I.2.2.3). Also assume that type(-A(s)) < 0 for s E J. Then

-[A(t) - A(s)] = A(t) [A-1(t) - A-l(s)]A(s)


(2.3.1 )
= A1-"(t)A"(t) [A-1(t) - A-l(s)]A(s) .

Thus, by applying A1-O(s) from the right to (II.2.2.2) and putting


aA(t, s) := A1-o(s)e-(t-s)A(s)
IV.2 Constant Interpolation Spaces 219

and
kA(t, s) := AD:(t) [A-1(t) - A- 1(s)]A 2 -D:(s)e-(t-s)A(s)
for (t, s) E Jt", we see that VA := UAA1-D: solves the linear Volterra equation

(2.3.2)

Similarly, let

and
hA(t, s) := _A 2 -D:(t)e-(t-s)A(t) AD:(t) [A-I(t) - A-l(s)]

for (t, s) E J. Then, by applying A(t) from the left to (11.2.2.3) and using (2.3.1),
it is easily verified that

WA(t,S) := A(t)UA(t, s) - A(t)e-(t-s)A(t) (2.3.3)

solves the Volterra equation

(2.3.4)

Now suppose that condition (2.2.3) is satisfied and fix a such that

I-p<a</l. (2.3.5)

We shall prove that, given A E A, equations (2.a.2) and (2.3.4) have unique solu-
tions VA and WA, respectively. Then we put

(2.3.6)

and show that WA, indeed, equals the right-hand side of (2.3.3) and that UA is a
parabolic evolution operator for A.

It follows from Corollary 2.1.2 and Lemma 2.2.4 that

(2.3.7)

Similarly, Corollaries 2.1.2 and 2.2.3 and Lemma 2.2.4 imply

(2.3.8)

and that the norm of this map is bounded by cry. Observe that 2 - a - p < 1 by
(2.3.5). Thus Theorem 11.3.2.2 guarantees the existence of a unique solution

(2.3.9)
220 IV Variable Domains

of (2.3.2), and
(2.3.10)
with WA E st(Eo, 2 - Q - p). Note that (11.3.2.2) implies

(2.3.11)

Let

and
W~ := wAA a - 1 . (2.3.12)
Then, thanks to (2.2.3) and Corollaries 2.1.2 and 2.2.3,

(2.3.13)

From (2.3.8) and Lemma 11.3.2.1 we infer that

(t - s)2-a- p IIwA(t, s)llc(Eo) ::; c(€)e(V+E)(t-S) (2.3.14)

for (t, s) E J'A, € > 0, and A E A, where

for some positive constant co. Note that (2.3.11)-(2.3.13) imply

w~ = k~ + WA * k~ . (2.3.15)

Hence we deduce from (2.3.13), (2.3.14), and the results of Subsection 11.3.1 that

W~ E st(Eo, 1 - p) (2.3.16)

and

(t - s)l-p Ilw~(t, s)llc(Eo) ::; c(c)e(V+E)(t-S) , (t,s) E J'A, A EA. (2.3.17)

Put
(t, s) E h . (2.3.18)
We infer from (2.3.10), (2.3.12), and (2.3.16) that UA, defined by (2.3.6), is given by

(2.3.19)

Moreover, (2.3.7), (2.3.16), (2.3.17), and Subsection 11.3.1 imply

aA * w~ E st(Eo, 1 - Q - p) '-t C(h, Eo)


IV.2 Constant Interpolation Spaces 221

and
(t - s)l-a- p lIaA *w~(t,s)IIC(Eo) ~ c(E)e(v+c)(t-s)
for (t, s) E JA, E > 0, and A E A. Consequently, thanks to (2.3.18), (2.3.19), Corol-
lary 2.1.2, and Lemma 2.2.5, we see that

(2.3.20)

and

IIUA(t,s)II.c(E o) :s; c(E)e(v+c)(t-s) , (t,s) E Jt:." E > 0, A EA. (2.3.21)

Next we consider the Volterra equation (2.3.4). For this we need the following
technical result:

2.3.1 Lemma Given A E A,

bA E Jt(Eo, 2 - 0: - p) .
Moreover,

(t,s)EJX, E>O, AEA.

Proof Given (t, s) E JX, we put r := (s + t)/2 and decompose bA as

where

IA(t, s):= it A 2 (t)e-(t-T)A(t) [A- 1 (r) - A- 1 (t)]A(r)e-(T-s)A(T) dr,

IIA(t, s):= l r
A 2 (t)e-(t-T)A(t) [e-(T-s)A(T) - e-(T-s)A(s)] dr,

IIIA(t, s):= l r
A(t)e-(t-T)A(t) [A(s)e-(T-S)A(s) - A(r)e-(T-S)A(T)] dr,

and

We rewrite IA as
222 IV Variable Domains

Then we deduce from Corollaries 2.1.2 and 2.2.3 that

IIIA(t, s)II.c(Eo) :::; c(r - s)-l i t (t - r)a+ p-2 dr = c(r - S)-l(t _ r)a+ p-1
= c(t - s)a+ p-2

for (t, s) E J~ and A E A. Corollary 2.1.2 and Lemma 2.2.4 imply

r
IIIIA(t, s)II.c(Eo) :::; c(t - r)-21 (r - s)Hp-l dr = c(t - s)Hp-2

for (t, s) E J~ and A E A. Similarly, we see that

IIIIIA(t, s)II.c(Eo) :::; c(t - r)-l l r


(r - s)Hp-2 dr = c(t - s)Hp-2

for (t, s) E J~ and A E A. Lastly, define IVl for 0 < E < r - s by replacing the
lower limit s in the integral occurring in the definition of IVA by s + E. Then, by
integrating the first summand of IVl by parts,
IV1(t, s) = A(t)e-(t-r)A(t) [e-(r-s)A(s) _ e-(r-s-e)A(t)e-eA(s)] .

Since IV1(t,s) ~ IVA(t,S) in .cs(Eo) as E ~ 0, we see that


IVA(t, s) = A(t)e-(t-r)A(t) [e-(r-s)A(s) _ e-(r-s)A(t)] .

Hence, thanks to Corollary 2.1.2 and Lemma 2.2.4,

IIIVA(t, s)II.c(Eo) :::; cit - slHp-2

for (t, s) E J~ and A E A. Now the assertion is obvious .•

Finally, we are ready for the proof of the first main result of this section;

2.3.2 Theorem Let assumption (2.2.3) be satisfied. Then, given A E A, there


exists a unique parabolic evolution operator UA for A. It possesses Eo as regularity
subspace. Moreover, the estimates

IIUA(t,s)II.c(E o ) + (t - s) IIA(t)UA(t,s)II.c(Eo ) :::; c(E)e(v+e)(t-s) (2.3.22)

are valid for (t, s) E J~, c > 0, and A E A, where

1/ ;= 1/(0:, "1) ;= COTf1!(a+p-l)

for some 0: E (1 - p, ()) and Co is a positive constant.

Proof We shall show that the function UA, defined by (2.3.6), has the desired
properties.
IV.2 Constant Interpolation Spaces 223

First we note that Corollary 2.1.2 and Lemma 2.2.4 imply

(2.3.23)

Hence, by Lemma 2.3.1 and Theorem 1I.3.2.2, the Volterra equation (2.3.4) pos-
sesses a unique solution WA E Jt(Eo, 2 - 0: - p) and it is given by

(2.3.24)

where
00

rA:= :~:::>A * ... *hA E Jt(Eo, 2 - 0: - p) .


. 1 '---v--"
J= j

It follows from (2.3.23) and Lemma 1I.3.2.1 that

(t - s)2-0I.-p IIrA(t, s)IIC(Eo ) $ c(c:)e(v+g)(t-s) (2.3.25)

for (t, s) E JJ:., c: > 0, and A E A. This estimate, Lemma 2.3.1, formula (2.3.24),
and the results of Section 1I.3.1 imply that WA E Jt(Eo, 2 - 0: - p), and

(t - S)2-0I.-p IIWA(t, s)llc(Eo) $ c(c:)e(v+g)(t-s) (2.3.26)

for (t,s) E JJ:., c: > 0, and A E A.


Now let A E A be fixed, and let {Ag ; c: > o} be the Yosida approximation of
A E A. Then A;l E CP (J, .c(Eo)) by (2.2.15), and, consequently, Ag E CP (J, .c(Eo))
by (2.3.1). Hence all the considerations at the beginning of this subsection apply
to Ag. In particular (cf. (2.3.19)), the unique parabolic evolution operator Ue of Ag
is given by
Ue = a~ + ae * w~ , (2.3.27)
where we write a~ for a~e' etc., for simplicity.
From Corollaries 2.1.2 and 2.2.3 we infer that

(2.3.28)

Moreover, given (t, s) E JJ:., Lemma 2.1.4 and Corollary 2.2.3 imply

as c: ~ O. Now Lebesgue's dominated convergence theorem and an induction ar-


gument guarantee that, given (t, s) E JJ:. and j ~ 2,

j j

as c: ~ O. Moreover, from (2.3.28) we deduce that kg * ... * kg(t, s) satisfies an


estimate of the form (1I.3.2.1), independently of c: > o. Thus, by applying the
224 IV Variable Domains

dominated convergence theorem (in £l(N,L(Eo») to the series (I1.3.2.2), we infer


that, given (t, s) E J a,
(2.3.29)

as c ...... O. Since w~ = k~ + w" * k~ by (2.3.15), we see that w~(t, s) ...... w~(t, s) in


Ls(Eo) for (t, s) E J6.. Now Lemma 2.1.4 and (2.3.13) imply that, given (t, s) E J6.,

(2.3.30)

Put bE := bA, and hE := hA" respectively, and let VVc: be the unique solution
of the Volterra equation W = bE + hE * W for c > o. Then, by similar arguments
as above, given (t, s) E J a,
(2.3.31)

as c ...... O. Since, by (2.3.3),

VVc:(t, s) = A"Uc(t, s) - AE(t)e-(t-s)A,(t) ,

it follows from (2.3.31) and Lemma 2.1.4 that, given (t,s) E J6.,

A"U,,(t,s) = A(t)(1 +cA(t)r1UC(t,s) ...... WA(t,S) + A(t)e-(t-s)A(t) (2.3.32)

in Ls(Eo) as c ...... o. Note that, thanks to Lemma II.6.1.1 and (2.3.30),

as c ...... O. Thus we infer from (2.3.32) and the closedness of A(t) that

im(UA(t,S» C dom(A(t» , (t,s) E J6. , (2.3.33)

and that

AUA(t, s) = WA(t, s) + A(t)e-(t-s)A(t) , (t,s)EJ6.. (2.3.34)

Hence Corollary 2.1.2 and (2.3.26) imply

AUA E fi(Eo, 1) (2.3.35)

and
(t - s) IIAUA(t, S)II.c(Eo ) ~ c(c)e(V+E)(t-S) (2.3.36)

for (t,s) E J6., € > 0, and A E A.


IV.2 Constant Interpolation Spaces 225

Observe that

Ue(t,s) - Ue(T,S) = - it AeUe(a,s)da in C(Eo)

for 0 ~ s < T < t and t E J. From this, (2.3.30), (2.3.32), and (2.3.36) we infer that

UA(t, s) - UA(T, s) = - it AUA(a, s) da

for 0 ~ s < T < t and t E J. Hence

UA(t + h, s) - UA(t, s) + hAUA(t, s) = I t


t +h
[AUA(t, s) - AUA(T, s)] dT

for (t, s) E JX and h > 0 with t + h E J, where the integral converges in C( Eo)
thanks to (2.3.35). Now we deduce from Proposition 1.1.2 that

s E J, (2.3.37)

and that
(t,s) E JX . (2.3.38)
Since Ue is the parabolic evolution operator for Ae and since Ue has the
regularity subspace Eo, by Corollary II.4.4.2, we know that, given x E Eo,

Ue(t, s + h)x - Ue(t, s)x = 1s


8 +h
Ue(t, T)Ae(T)X dT

= 1s
8 +h
Uc(t, T)j1!-<>(T)A~(T)X dT
(2.3.39)

for (t, s) E JX and h > 0 with s + h < t. Note that UeA!-<> = Ve, where, thanks
to (2.3.10),
e> O.
From Lemma 2.1.4 and (2.3.29) we easily infer that, given (t,s) E JA,

as e --+ O. Thus (2.3.6), (2.3.39), and Lemma 2.2.2 imply

UA(t, s + h)x - UA(t, s)x = l s


s +h
VA(t, T)A<> (T)X dT , xEEIJ ,

for (t, s) E JX and h > 0 with s + h < t. Since the integrand is a continuous func-
tion of T, by (2.3.9) and (2.2.10), we see, by again invoking Proposition 1.1.2, that

(2.3.40)
226 IV Variable Domains

and that

(t, s) E J~, x E Eo . (2.3.41 )

If x E dom(A(s)) then (2.3.6), (2.3.41), and Theorem III.4.6.5(iii) imply

VA(t, s)A<>(s)x = UA(t, s)A(s)x .


Hence
(2.3.42)
Next, from (2.3.20), (2.3.35), and (2.3.36) we infer by interpolation that

(2.3.43)

and that
(t - s)o IIUA(t, s)llc(Eo,Eo) :S c(E)e(v+c)(t-s) (2.3.44)
for (t,s) E J~, E > 0, and A E A.
Finally, note that (2.3.30) and the fact that UfO is an evolution operator for
E > 0 imply the validity of (II.2.1.3). Thus, thanks to (2.3.20), (2.3.21), (2.3.33),
(2.3.35)-(2.3.38), and (2.3.40)-(2.3.43), the theorem has been completely proven .•

Parabolic evolution equations with 'nonconstant domains' have been studied


by several authors under a variety of different hypotheses. In several papers it
has been assumed that an intermediate space between Eo and D(A(t)) is inde-
pendent of t. The earliest investigations of this type are due to Kato [Kat61] and
Sobolevskii [Sob61]. These authors assume that the domain of a suitable fractional
power of A(t) is constant. Such an assumption restricts the applicability of these
results considerably since, in general - in particular, if A(t) is induced by elliptic
boundary value problems with coefficients having little regularity - it is rather
difficult (or even not known how) to characterize these domains.
In [AcT86a] it is assumed that the interpolation spaces (Eo,D(A(t)))(I,oo are
independent of t for some e E (0,1), except for equivalent norms. In that paper the
authors do not construct an evolution operator but study the solvability of the Cau-
chy problem directly by means of suitable representation formulas (cf., however,
[AcT86b], [Acq88]). This technique has already been used by the same authors in
the case of constant domains in [AcT85]. In [AcT87] they succeeded to prove by
means of their method the solvability of the Cauchy problem under hypotheses
encompassing and generalizing most of the previously known results.
The theorems of [AcT87] have been further generalized by Yagi [Yag90], who
weakened the assumption that -A(t) should generate an analytic semigroup. In
addition, he proposed a method of proof that is different from the one of Ac-
quistapace and Terreni and closer to the one of Kato in [Kat61]' also used in
earlier papers by Yagi (e.g., [Yag88], [Yag89]). Our proof of Theorem 2.3.2 follows
IV.2 Constant Interpolation Spaces 227

[Yag90] inasmuch as we employ Volterra equations (2.3.2) and (2.304) and the rep-
resentation of bA in, the proof of Lemma 2.3.1. It should be mentioned, however,
that Theorem 2.3.2 seems to be new in this form and the given precision. It has
the advantage that we do not make any assumption on the interpolation space Eo,
except that it is obtained by means of an admissible interpolation functor. This
greatly facilitates its application to concrete problems.
It should be noted that Theorem 2.3.2 contains Theorem IIo404.1 as a spe-
cial case (except for slightly different exponential bounds). In fact, if assump-
tion (IIo4.2.1) is satisfied, it is easily verified that (2.3.1) implies (2.2.3) for any
e E (1 - p, 1). Thus the above considerations provide a new (and more compli-
cated) proof of Theorem IIo404.1.
Theorem 2.3.2 is closely related to Theorem 1.4.1. In fact - using the no-
tations of the latter theorem - we see that El is a constant intermediate space
between E, and D(A,(t)), that is, in general however, not an interpolation space.
Thus, in general, Theorem 1.4.1 cannot be deduced from Theorem 2.3.2. Of course,
conversely, we cannot obtain Theorem 2.3.2 from Theorem 1.4.1 since, usually, we
do not have operators with constant domains to start with. In later chapters we
shall see that each one of these theorems has its own range of applications.
The hypotheses in [AcT87] and [Yag90] are flexible enough to admit cases
where no intermediate space between Eo and D(A(t)) is constant. In fact, those
authors basically assume (2.2.3(i)) and

(2.3045)

for s, t E J and), E ~{I, where p, (J E (0,1] satisfy (J + P > 1. It is easily verified


that (2.2.3(ii)) and (2.2.3(iii)) imply (2.3.45) with (J = e (cf. (2.2.5)). On the other
hand, if we do not assume that there exists a constant interpolation (or fractional
power) space between Eo and D(A(t)), it can be shown that, given (2.2.3(i)), con-
dition (2.3.45) is satisfied if A-I is suitably continuously differentiable (cf. [AcT87,
Section 7]). Moreover, given appropriate differentiability assumptions for A-I,
the existence of an evolution operator has first been shown by Kato and Tanabe
[KaT62] (also cf. [Tan79], [Fat83] for expositions of these results, and [Yag76],
[Yag77] for related investigations). It turns out, however, that differentiability as-
sumptions for A-I are too restrictive to be useful for dealing with quasilinear
parabolic problems that - after all - are our main concern. For this reason we
do not pursue this matter further.

2.4 Estimates for Evolution Operators


We begin by proving an additional continuity result for the parabolic evolution
operators UA .
228 IV Variable Domains

2.4.1 Theorem Let assumption (2.2.3) be satisfied. Then

AEA,

and, letting v have the same meaning as in Theorem 2.3.2,


IIAUAA-1(t, S)IIC(Eo) ::; c(c)e(v+c)(t-s) (2.4.1)

for (t,s) E Jt:,., 10 > 0, and A E A.


Proof Thanks to (2.3.34),

AUAA- 1 = WAA- 1 + AaAA- 1 , (2.4.2)

where aA(t,s):= e-(t-s)A(t). Fixing a E (1- p,e),

Hence Corollaries 2.1.2 and 2.2.3 imply

(t,s) E h, A EA. (2.4.4)

Moreover,
A(t)aA(t,s)A-I(S) - A(r)aA(r,a)A-l(a)
= [A(t)e-(t-s)A(t) _ A(r)e-(T-cr)A(T)]A-1(s)
+ A(r)e-(T-cr)A(T) [A-l(S) _ A-l(a)] ,
assumption (2.2.3(iii», Eo ~ Eo, and Lemma 2.2.4 guarantee that

(2.4.5)

From (2.4.3), Corollaries 2.1.2 and 2.2.3, and Lemma 2.2.5 we see that, given So E J,

AaAA-l(t,s) -+ 1

as (t,s) -+ (so,so) in Jt:,.. Thus

(2.4.6)

Let
CA,j(t, s) := Aj-"(t)e-(t-s)A(t) A"(t) [A-l(S) - A-l(t)]
for j = 1,2, (t, s) E J'A, and A E A. Then

j = 1,2, (2.4.7)

by Corollaries 2.1.2 and 2.2.3. It is easily verified that

AEA.
IV.2 Constant Interpolation Spaces 229

Hence we deduce from (2.4.7), Corollary 2.1.2, and Section II.3.1 that
bA A- 1 E ii(Eo, 1- a - p) ~ C(h, Eo)
and
(t - S)1-o:- p Ilb AA- 1 (t,s)II.c(Eo) ~ c(c)ec(t-s) , (t, s) E J'S., A EA.
Thus (2.3.24) and (2.3.25) imply
WAA- 1 = bAA- 1 + (1' A * bA)A- 1 = bAA- 1 + l' A * (bAA- 1 )
and, consequently,
(2.4.8)
and
(2.4.9)
for (t, s) E Ib., c > 0, and A E A. Now the assertion follows from (2.4.2)-(2.4.6),
(2.4.8), and (2.4.9) .•

2.4.2 Corollary Let assumption (2.2.3) be satisfied. Given A E A,


UA E ii(Eo, E e , 8) n ii(Ee, 0) (2.4.10)
and
AUA E ii(Ee, Eo, 1 - 8) . (2.4.11)
Moreover,

IIUA(t, s)II.c(Ee) + (t - sl'IIUA(t, s)II.c(Eo,E9)


(2.4.12)
+ (t - S)1-11 IIAUA(t, s)ll qE9 ,Eo) ~ c(c)e(v+c)(t-s)

for (t, s) E J'S,., c > 0, and A E A, where l/ has the same meaning as in Theo-
rem 2.3.2. In addition,
sEJ, AEA, (2.4.13)
for each admissible interpolation functor (',' k

Proof Observe that


IIUA(t, s)II.c(E1(A(s)),E1(A(t))) = IIAUA(t, s)II.c(E, (A(8)),Eo)
(2.4.14)
= IIAUAA- 1 (t, s)II.c(Eo)
and that

IIUA(t, s)II.c(Eo,E,(A(t)) = IIA(t)U:'1(t, s)II.c(Eo)


for (t, s) E J'S,. and A E A. Now we obtain (2.4.10)-(2.4.12) easily from Theorems
2.3.2 and 2.4.1 by interpolation (also cf. (2.3.43) and (2.3.44)).
230 IV Variable Domains

To prove (2.4.13) we can assume that J is compact. Then, given s E J


and x E Ee,
IIUA(t, s)x - UA(T, s)xll(Eo,Ee),;
:S cIIUA(t,S)X - UA(T,S)xll~ IIUA(t,S)X - UA(T,S)xIl6-~
:S cllxll eIIUA(t,S)X - UA(T,S)xI16-~
for t, T E J n [s, (0), as follows from (2.4.10). From (2.3.22) and (2.4.10) we deduce
that UA E Jl((Eo,Ee)~,O). Now (2.4.13) is a consequence of UA E C(h,f's(Eo))
and the density of Ef) in (Eo, Eo k _

Lastly, we turn to the case of ordered Banach spaces and study the question
of positivity for evolution operators.

2.4.3 Theorem Let Eo be an ordered Banach space and let assumption (2.2.3)
be satisfied. Given A E A, suppose that A(s) is resolvent positive for s E J. Then
UA is positive.

Proof Let {Ao ; E: > O} be the Yosida approximation for A and let UE be the
evolution operator for A E • Then, letting X be the positive cone of Eo, the proof
of Theorem 11.6.3.3 shows that UE is positive for E: > O. Now the assertion follows
from (2.3.30) and the closed ness of the positive cone. _

Recall from Theorem II.6.3.2 that A( s) is resolvent positive iff the semigroup
{e-tA(s) ; t::::: O} is positive.

2.5 The Cauchy Problem


Let assumption (2.2.3) be satisfied. Let (', .), be an admissible interpolation func-
tor for some "( E (0,1) and put

F := (Eo, Ee), .

Given (x, A) E Eo x A and f: J -> Eo, consider the linear Cauchy problem
it + A(t)u = f(t) , t E j , u(O) = x . (2.5.1)(x,A,J)
The following solvability theorem is the second main result of this section.

2.5.1 Theorem Given (x, A) E Eo x A and


f E CCT(J, Eo) + C(J, F) (2.5.2)
for some (J E (0, I), the Cauchy problem (2.5.1)(x,A,J) has a unique solution
1 .
u E C(J, Eo) n C (J, Eo) .
IV.2 Constant Interpoiati{)n Spaces 231

If (., .)( is an admissible interpolation functor and x E (Eo, Ee)t, then

and U is a strict solution if x E dom( A(O)).

Proof Thanks to Theorem 2.3.2 and Remark II.2.1.2(e) we have to show that
the mild solution
U:= U(·,O)x + U * f("O) ,
where U := UA , is a solution with the asserted regularity properties.
Put Uo := U(·, O)x and

j = 1,2,
where f = h + h with h E Ca(J, Eo) and h E C(J, F). Then Uo is a solution of
(2.5.1)(x,A,O) by Theorem 2.3.2 and the properties of parabolic evolution operators.
Theorem 2.4.1 implies that Uo is strict if x E dom(A(O)). If x E (Eo, Eob it follows
from (2.4.13) that Uo E C(J, (Eo, Ee)().
Observe that Uj is a mild solution of (2.5.1)(o,A,jj) for j = 1,2. If we show
that Uj is, in fact, a solution with the asserted continuity properties, the assertion
follows from U = Uo + Ul + U2.
(i) Let {AE ; C > O} be the Yosida approximation for A, and let Uf: be the
evolution operator for Ac. Then V E := UE * h (., 0) is a solution of (2.5.1)(O,A"J,)'
From (2.1.5), (2.3.21), and (2.3.30) it easily follows that, given t E J,

in Eo (2.5.3)

as c -+ o. Moreover, (2.3.3) implies that

AE(t)VE(t) = fat AcUc(t,T)[h(T) - h(t)] dT+ fat AEUE(t,T)h(t)dT


= rt A,/fc(t,T) [h(T) _ h(t)] dT+ hrt Wc(t,T)h(t)dT
h
(2.5.4)

+ [1- e-tA,,(t)lh(t)
for t E J and c > o. Observe that
(t,T) E Ii:,. .
Hence (2.3.31), (2.3.34), and Lemma 2.1.4 imply that, given (t,T) E J't,.,
in .Ls(Eo)
232 IV Variable Domains

as € -> o. By (2.1.5), (2.3.36), and ft E CU(J, Eo) we see that, given t E j,

From this, (2.5.4), (2.3.31), and Lemma 2.1.4 we easily deduce that, given t E J,

in Eo (2.5.5)

as € -> 0, where

It is clear that wE C(J, Eo). Moreover, (2.5.3), (2.5.5), Lemma 11.6.1.1, and the
closedness of A(t) guarantee that Ul(t) E dom(A(t)) and A(t)Ul(t) = wet). From
(2.5.1)(O,A.,h) we derive

(t, s) E J6,. .

Thus, letting € -> 0, it is easily verified that

(t,s) E h .

Hence Ul E C 1 (J, Eo) and

so that Ul is a strict solution of (2.5.1ho,A,fd. Lastly, since U E f{(Eo, Eo, (J) by


Corollary 2.4.2, it follows that

Ul = U * ft(·, 0) E C(J, Eo) ~ C(J, (Eo, Eo)e) .

(ii) We can assume that J is compact. Then, by interpolating estimates


(2.3.22) and (2.4.12),

(t - s)(1-"()O IIU(t, s)II.c(F,Ee) + (t - S)l-"(O IIAU(t, s)II.c(F,Eo) ::; C

for (t, s) E J6,.. Hence


IIU(t, s)II.c(F,Ee) + IIAU(t, s)II.c(F,Eo) ::; c(t - s)("(-l)O , (t, s) E J6,. .
Now Theorems 1.2.1 and 1.2.2 guarantee that U2 is a strict solution of the Cauchy
problem (2.5.1)(O,A,h). Similarly as for Ut, we see that U2 E C(J, Eo). This proves
the theorem. _
IV.2 Constant Interpolation Spaces 233

The fact that Holder continuity of f implies the solvability of the Cauchy
problem (2.5.1)(x,A,f) is well-known in the literature (e.g., [AcT87J, [Yag89]). Re-
sults which are related to our assumption, namely, that f is continuous with values
in the interpolation space F, can be found in [AcT87J, where also singularities of f
at t = 0 are admitted.
Of course, by using the estimates of Theorem 2.3.2 and Subsection 2.4, it is
not difficult to give bounds for the solution of (2.5.1)(x,A,f) that are uniform with
respect to A E A.

2.6 Abstract Boundary Value Problems


In this subsection we introduce a relatively simple class of hypotheses which imply
assumption (2.2.3). These new conditions are motivated by the theory of elliptic
boundary value problems that are discussed in Chapter VII. There it is shown
how the present abstract hypotheses are met by suitable realizations of elliptic
systems under rather general boundary conditions. In this subsection we illustrate
the abstract theory by means of a simple second-order elliptic boundary value
problem.

Let (Eo, E 1 ) be a densely injected Banach couple and let F be a Banach space.
Given K, 2: I, w > 0, and an increasing function 13: IR+ -> IR+, we denote by

!B(E1 , Eo, F, K" w, 13)

the set of all


(A, 8) E £(E}, Eo x F) (2.6.1)
such that
(w + A, 8) E Lis(EI, Eo x F) , (2.6.2)
that
(2.6.3)
and that
(2.6.4)
for x E E1 and Re), 2: w. We also put

!B(E}, Eo, F, K"w) := U!B(E}, Eo, F, K"W, 13)


f3

and
!B(E},Eo,F;f3):= U !B(E},E ,F,K"w,f3) ,
Q
,,>}
w>o
234 IV Variable Domains

as well as
~(El,Eo,F):= U~(E!'Eo,F;,8) .
{3

Observe that

~(El' Eo, F, "',w,,8) C ~(El, Eo, F, "',w) C ~(El' Eo, F) C .L(El' Eo x F) ,

so that the Banach space .L(E!, Eo x F) induces the natural topologies on these
subsets.
For definiteness we use i1-norms in product Banach spaces. If it is convenient,
we also identify Eo with Eo x {O} etc.

2.6.1 Remarks (a) If F = {OJ (and, consequently, B = 0) then

(b) Suppose that (A, B) satisfies (2.6.1), (2.6.2), and

x E E1 , Re A 2: w .

Then

for x E El and ReA 2: w, where

Proof This is an easy consequence of the identity

x=(w+A,B)-l(A+A+w-A,B)x, x E E1 , A E C,

and the fact that IAI 2: w for Re A 2: w. _


(c) If (A,B) E ~(E!,Eo,F,,,,,w,,8) then

(A + A, B) E Lis(El' Eo x F) , ReA2:w.

Proof Given any r > w, it follows from (2.6.3), thanks to the monotonicity
of ,8, that

XEE 1 , ReA2:w, IAI:::;r,


where c:= 1/(",(1 V ,8(r))). Hence (2.6.2), Proposition 1.1.1.1, and the arbitrari-
ness of r > w imply the assertion. _
IV.2 Constant Interpolation Spaces 235

(d) If (A, B) E I23(El' Eo, F, K, w, (3) then

II (>.. + A, B) -lllc(Eo xF,Ej ) ~ 1).,l i - l K( 1 V (3(1>"1))


for Re>.. 2:: wand j = 0, 1.
Proof This is a consequence of (2.6.3) and (c) .•

Suppose that (A,B) E 123 (El' Eo, F). Then

El,s:= kerB
is a closed linear subspace of E l . We also put

A := As := AIEl,s. (2.6.5)

It follows from Remark 2.6.1(c) that

>.. + A E Lis(El,s, Eo) , Re>.. 2:: w . (2.6.6)

Moreover,
A E 1t(El,s,Eo,K,w) , (2.6.7)
provided El,s is dense in Eo. Lastly, we set

IlxIIE1 • B := II(w + A)xllo , x EEl,s. (2.6.8)

Then we infer from (2.6.3) and (2.6.4) that

Kll Ilxlll·~ IlxiIEl.B ~ Kl Ilxll l ' x EEl,s, (2.6.9)

where Kl := K + W l\illc(E1, E)
0
and i: El '----+ Eo is the injection map.
We also put

and assume that


e
there exist E (0,1) and an admissible }
interpolation functor (', .) e such that (2.6.10)
(Eo, EI,s)e = (Eo, El)e .

The following lemma shows that this assumption implies that the interpola-
tion spaces
(2.6.11)
and
Ee:= (Ee, 11·lle):= (Eo,Ede (2.6.12)
coincide, except for equivalent norms, that is, Ee,s is independent of B, except for
equivalent norms.
236 IV Variable Domains

2.6.2 Lemma Given I'> 2: 1 and w > 0, there exists a constant 1'>0 2:: 1 such that

1'>;1 Ilxllo ::; IIxIIEB,B ::; 1'>01lXll0 , xE Eo,

Jor (A,B) E r.B(E1 ,Eo,F,I'>,w).


Proof From (2.6.9) we infer that

id E £is(E1 ,B, Ei,B) n Cis(Eo)


and that the norms of id E £(E1 ,B, Ei,B) and id E £(Et,B' E 1 ,B) are bounded
by 1'>1. Now the assertion follows by interpolation from (2.6.10) .•

In order to prove the main result of this subsection we introduce the following
assumptions:

(i) (Eo, Ed is a densely injected Banach couple, F is a Banach


space, and J is a perfect subinterval of JR+ containing O.
(ii) I'> 2: 1, w, "10 E 1R+, P E (0,1), and

with (2.6.13)
[(A, B)] p, J ::; "10 .

(iii) E1(s):= E 1,B(s) = kerB(s) is dense in Eo for s E J,


(iv) 1 - P < () < 1 and

s E J.

Then we put
A(s) := A(s)IE1(S) , s E J , (2.6,14)
and
A: J ----> C(Eo) , s t--+ A(s) (2.6.15)
and consider the linear Cauchy problem

it + A(t)u = J(t) , t E j , u(O) = x, (2.6.16)

where x E Eo and J: J ----> Eo. Note that this means that we study the evolution
equation
it + A(t)u = J(t) , t E j , u(O) = x, (2.6,17)
subject to the side condition

B(t)u(t) = 0 , (2.6.18)
IV.2 Constant Interpolation Spaces 237

Since in practical applications the side condition is most often induced by suit-
able boundary conditions, (A, B) is said to be an abstract boundary value prob-
lem (BVP) whenever (A, B) E P,(El,Eo, F). Then (2.6.17), (2.6.18) is an abstract
initial-boundary value problem (IBVP) with homogeneous boundary conditions.

It may be worthwhile to illustrate assumption (2.6.13) by means of a proto-


type problem already now, although we have to rely on results that will be proven
in later chapters only. Thus in the following example we use, without further ex-
planations, facts from the theory of elliptic BVPs and the theory of function spaces
and refer the reader to Chapters VI and VII for notations, definitions, and proofs.

2.6.3 Example: Second Order Elliptic BVPs We denote by X a bounded


domain in ]Rn of class C 2 and fix 8 E c( ax, {O, I}). Then we put

fj:= 8- l (j) = {x E an; 8(x) =oj} , j = 0, 1.


Note that f j is a union of components of ax, hence an open and closed subman-
ifold of ax. We fix p E (1, (X)) and observe that

is a densely injected Banach couple. We also put

with obvious modifications if fo [resp. fll is empty, in which case fl [resp. fol
equals ax.
Set

Given
((ajk), (aj), ao, (bj), bo, co) E JffiW(X) ,
we define a second order (scalar) differential operator A and a first order (scalar)
boundary operator B on X by

and
Bu := 8(bj'"Y88jU + b01'8U)+ (1 - 8)Co1'8U ,
respectively,where 1'8 is the trace operator for ax. Here and in the following,
we
use the summation convention, j and k running from 1 to n. Then (A, B) is a
linear BVP on X of order at most two. Observe that B reduces to the Dirichlet
boundary operator on fo, whereas it is a Neumann type boundary operator on fl.
238 IV Variable Domains

We topologize the set of these BVPs by identifying them with the Banach space
lffiVlP'(X) by means of the identification

We denote by v := (VI, ... ,vn) the outer unit normal vector field on oX.
Given positive constants Q and M, we write

lElffiVlP'(X, Q, M)

for the subset of lffiVlP'(X) consisting of all regularly elliptic BVPs on X of class
(Q, M). By definition, (A, 8) E lffiVlP'(X) is a regularly elliptic BVP of class
(Q, M) if

if

and if
II (A, 8) IhllVJl'(x) ::; M .
It follows from the results proven in Chapter VII that, given positive constants Q
and M, there exist K ? 1 and w > 0 such that

(2.6.19)

where
t(1-I/p)/2 O<t<l,
{
(3(t) := t(2-I/ P)/2:
l::;t<oo.
Now we put
(2.6.20)
where
WP~(1-6ha (X) := { U E Wp2(X) ; (1 - t5h&u = 0 }
is a closed linear subspace of ~2(X). Then (Eo, Ed is a densely injected Banach
couple. We also put
F := W/-1/P(fd .

Then it follows from (2.6.19) that, given Q, ME rR+, there exist K ? 1 and w >0
such that
ElffiVlP'(X, Q, M) '----+ '13 (El' Eo, F, K, W, (3)
(by a slight abuse of notation). Hence assumption (2.6.13(ii)) is satisfied if there
exist p E (0,1) and "7 E lR+ such that

(A,8) E CP (.1, lElffiVlP'(X, Q, M)) (2.6.21 )


IV.2 Constant Interpolation Spaces 239

with
[(A, B)]cP(J,lBVlP'(X» ::; 17 .
Note that
VeX) '--+ Wp~B(X) := { U E W/(X) ; Bu = o}
and that
(A, B) E lBl'VlP'(X) . (2.6.22)
Hence E 1 ,B is dense in Eo for (A, B) E lBl'VlP'(X). Lastly, it follows from the results
of Chapter VI that

o < 2(J < 1 + lip, 2(J =I- lip,


provided (., ·)0 := (., ·)o,p if (J =I- 1/2, and (-, ·h/2 := [., ·]1/2. Thus we see from
(2.6.20) and (2.6.22) that assumption (2.6.13(iv)) is satisfied if (2.6.21) is true
with 2p > 1- lip and if we fix 2(J E (2 - 2p, 1 + lip) with 2(J =I- lip .•

Now it is easy to prove the main result of this subsection.

2.6.4 Theorem Let assumption (2.6.13) be satisfied. Then there exists a unique
parabolic evolution operator U := UA for A. It has Eo as regularity subspace and
satisfies
(2.6.23)

for (t, s) E Je:" E > 0, and j = 0, 1, where v := c017~/(a+p-1) for suitable constants
Co and a E (1 - p, (J). Given an increasing (3: IR+ -+ IR+, estimate (2.6.23) is uni-
form with respect to

(A, B) E CP(J,SJ)(E 1,Eo,F, "",W, (3)) with [(A,B)]p,J::; 170·


If
f E CCT ( J, Eo) + C( J, E (3 )
for some a E (0,1) and (3 E (0, (J), the Cauchy problem (2.6.16) has a unique
solution
UEC(J,Eo)nc 1 (i,Eo) .
It is strict if x E E 1 (0), and U E C(J, Eo) if x E Eo.

Proof It follows from assumption (2.6.13), in particular from (2.6.3) and Re-
mark 2.6.1(c), that

Re >. ~:: W, t E J, j = 0, 1 . (2.6.24)


Thus, since
M:= sup
+ 1>'1 < 00
-I1>.--I ,
Re),:;o.O +W
240 IV Variable Domains

we infer that w + A: J --+ 2l,.M(Eo). Hence assumption (2.2.3(i)) is satisfied for


the set
{w + A ; (A,8) E CP(J, IJ3(E1' Eo, F, K"w)) } . (2.6.25)
Lemma 2.6.2 implies assumption (2.2.3(ii)).
Observe that, given (A, 8), (..4,8) E IJ3 (El, Eo, F), the 'generalized resolvent
equation'

(2.6.26)

is valid for ReA ~ w. From this, from (2.6.24), and from Remark 2.6.1(d) we deduce
by interpolation that

II (w + A(s)) -1 - (w + A(t)) -111.c(Eo,E8)


-::; K, 2 W O- 1 (1 V {3(w)) II (A, 8)(s) - (A,8)(t)II.c(E 1 ,Eo XF)

for s, t E J. Thus, letting 'fl := K,2 W O- 1 (1 V {3(w) ) 'flo , assumption (2.2.3(iii)) is also
satisfied. Now the assertions, but the last one, follow from Theorem 2.3.2, Re-
mark II.2.1.2(d), estimates (2.6.9), and Theorem 2.5.1, since the almost reiteration
property of Remark I.2.1l.2(a) guarantees that

where F:= (Eo, Eo),y, provided 0 < 'Y < {31(J.


To derive the last assertion, the proof of Theorem 2.5.1 shows that it suffices
to verify that U(·,s) E C(Jn [s,oo),.cs(Eo)) for s E J. For this we can assume
that J is compact. Then, given x E E1(S),

IIU(t, s)x - U(T, s)xllo -::; cIIU(t, s)x - U(T, s)xll~ IIU(t, s)x - U(T, S)XIl6- o
for t,T E In [s,oo). Thus assumption (2.6.13(iv)), (2.6.9), formula (2.4.14), and
Theorem 2.4.1 imply

IIU(t, s)x - U(T, s)xllo -::; c Ilxll~ IIU(t, s)x - U(T, s)xIl6- o ,
which shows that U(·,s)x E C(Jn [s,oo),Eo) for x E E1(S). Now the assertion
follows from (2.4.10) and the density of E 1(s) in Eo .•

We close this section by an illustration of this abstract theorem in a prototype


situation building on Example 2.6.3.

2.6.5 Example: Second Order Parabolic BVPs We use the notations of Ex-
ample 2.6.3 and suppose that there exist constants Q, M E lR+ and 2p E (1 - lip, 2)
such that
(A,8) E CP(J,EBVlP'(X,Q,M)) .
IV.3 Maximal Regularity 241

Then, given u O E Lp(X) and

for some a E (0, lip), the parabolic IBVP

atu + A(t)u = f(t) in Xxj,


B(t)u = 0 on ax x j, (2.6.27)
u(·,O)=uo on X,

has a unique solution

u E C(J,Lp(X)) nC 1 (i,Lp(X)) .

If U O E W;(X) for some 8 E (0,1 + IIp)\{llp} and

if lip < 8 < 1 + lip,


then
u E C(J, W;(X)) .
If u O E W;'S(O) (X) then

that is, u is a strict solution of (2.6.27).


This follows from Example 2.6.3, Theorem 2.6.4, and the fact, proven in
Chapter VI, that

for 0 < 2(} < lip.•

3 Maximal Regularity
In this section we are interested in situations in which it is not assumed that an
interpolation space between Eo and D(A(t)) is independent of t. Moreov~r, we
want to impose mild regularity assumptions for the t-dependence only. For this we
choose a particular setting that is an abstract version of parabolic initial-boundary
value problems and employ the concept of maximal regularity of Chapter III in
this case of 'nonconstant domains'. However, we consider in this chapter only
the case where A(t) is a suitably small perturbation of some constant operator
Ao E 1t ( D (Ao ) , Eo). In order to avoid excessive lists of hypotheses we relegate the
case of 'strongly varying' functions t I-t A(t) to Chapter VIII, where we will study
more specific settings.
242 IV Variable Domains

3.1 Abstract Initial Boundary Value Problems


Throughout the following, J denotes a perfect subinterval of IR+ containing o. We
assume that
(Eo, E 1 ) is a densely injected Banach couple
and F is a Banach space. } (3.1.1)

We also suppose that lEj := lEj(J) and IF := IF(J) are Banach spaces such that

lEl L1,loc(J, Ed n WI,loc(J, Eo), lEo


'--+

and IF '--+ Ll,loc(J, F).


'--+ L1,loc(J, Eo),
} (3.1.2)

Recall from Subsection III.l.4 that the trace map

, := ,0 :lEl ~ Eo, u f-+ u(O)


is well-defined, and that ,lEl is the trace space of lE 1 .
Next we assume that

(A, B) E C(lEl' lEo x IF) .


Then, given (J, g, uO) E lEo x IF X ,lEl' we consider the abstract IBVP

8t u + A(t)u = f(t) ,
B(t)u = g(t) , (3.1.3)
u(O) = UO .

This means, of course, that we consider the (nonautonomous) initial value problem

it + A(t)u = f(t) , t E j , u(O) = UO ,

subject to the (nonautonomous) side condition

B(t)u(t) = g(t) , (3.1.4)

Suppose that (3.1.3) has a solution

UEC(J,Ed (3.1.5)

and that
(t f-+ (B,g)(t)) E C(J,C(E1,F) x F) . (3.1.6)
Then it follows from (3.1.3) that the 'compatibility condition'

B(O)uO = g(O)
has to be satisfied. In general, however, u may not possess the regularity (3.1.5), so
that u(O) = uO may not belong to El but to the superspace ,lE 1 . Then B(O)nO may
IV.3 Maximal Regularity 243

not be defined. However, there may exist a linear operator

suitably derived from 8(0), where Fo is a Banach space with F'""--+ Fo, such that
the modified compatibility condition

f38(0)u o = pg

is true, where pg may be a 'partial trace' of g at t = 0. This is illustrated by the


following example where we use again results that will be proven in later chapters.

3.1.1 Example: Second Order Parabolic IBVPs We denote by X a bounded


subdomain of ]Rn of class C 2 and use the notations and conventions of Exam-
ple 2.6.3. We fix p E (1,00), define a densely injected Banach couple by

and put

It follows that
lEI = u.;,2(1-I/ p ) (X) .

We choose 8 E C(8X, {O, I}) and recall that rj := 8- 1 (j) for j = 0,1. Then we let

F-J ''=p
W 2- j - 1 / p (r·)
J , j = 0, 1,
and put

We also put

(J' F) n W(2-
F·J ''= L P P
j -l/P )/2(J L (r .))
, P J , j = 0, 1,

and

Furthermore,

and
B(())(8X) := B(())(r o) EB B(())(r 1 ) ,

Then, given Co E B(())(ro) and ((b j ), bo) E B(())(r 1), we define boundary operators

j = 0,1,
244 IV Variable Domains

by
Bo := (1 - 8)CO,8, B 1 := 8(bj/88j + bo,a) ,
respectively. Then
B:= Bo + Bl E £(El' F) .
We identify the set of all boundary operators of type Bj with lffi!OJ(f j) by means of
the identifications

Now we assume that

(t f-+ (A, B)(t)) E BUC( J, lffiW(X))


such that
(t f-+ Bj(t)) E BUCPj (J,lffi!OJ(fj)) , j = 0, 1,
with 2pj > 1 + j - l/p. It follows that
(A, B) E £(E1,Eo x IF) and (t f-+ B(t)) E C(J,£(El,Lp(8X))) .
Next we define

by
if 3<p<oo,
if 3/2 < p < 3,
if 1 < p < 3/2,
and

by

p~ ~(1-6)
if 3<p<oo,
{ if 3/2 < p < 3,
if 1 < p < 3/2,
where, denotes again the trace at t = 0, that is, ,v := v(O) for v E IF, whenever
this makes sense. Then the compatibility condition
(JB(Ohu = pBu ,
is valid. Note that

IF(B) := { (g, u O) E IF X ,El ; {JB(O)u o = pg }


is a closed linear subspace of IF x ,E 1 . It is shown in Chapter VI that there exists
R E £(IF(B), Ed satisfying
(B, ,)R = 111'(8) ,

provided there exists Q > 0 such that co(x) ~ Q for x E fo and bj(x)vj(x) ~ Q for
x E fl. Thus, in this case, (B,,) is a retraction ofE I onto IF(B) .•
IV.3 Maximal Regularity 245

3.2 Isomorphism Theorems


We presuppose assumptions (3.1.1) and (3.1.2) and assume that

(A, E) E .c(EI' Eo x F) . (3.2.1)

Denoting the constant map (t 1---4 (A, E)) again by (A, E), we also suppose that

(8 + A, E) E .c(lEI' lEo x IF) . (3.2.2)

Next we introduce an abstract 'compatibility condition' by assuming that

IF(E) is a Banach space such that

and (E, "!) is a continuous retraction from lEI onto IF(E).


} (3.2.3)

We denote by EI,B and lE1,B the kernels of E E .c(El' F) and E E .c(lEI,lF),


respectively, and put
A:= AIEI,B.
Then we assume that
(Eo, EI,B) is a densely injected Banach couple,

and (lEo, lEI,B) is a pair of maximal regularity for A.


} (3.2.4)

Recalling definition (III.1.5.7) of K A , we can easily prove the following isomor-


phism result:

3.2.1 Theorem Let assumptions (3.1.1), (3.1.2), and (3.2.1)-(3.2.4) be satisfied.


Then
(8+A, (E,,,!)) E £is (lEI , lEo x IF(E))
and
(8+A,(E,,))-1 = [K A ,(I-KA (8+A))(E,"!)c] , (3.2.5)
where (E, 1Y is a coretraction for (E,,).

Proof It follows from (3.2.2) and (3.2.3) that

(8+A,(E,,)) E .c(lE1,lEO x IF(E)) . (3.2.6)

Moreover, (3.2.4) implies

(3.2.7)
246 IV Variable Domains

and, thanks to Corollary III.1.6.2,

(3.2.8)

where (0 + A)" := (0 + A, ')') I kerb). From (3.2.7) we infer that (0 + A, (B, ')')) is
injective. Given (j, (g, un)) E lEo x F(B), let

Then u E lEI and, thanks to (3.2.8),

Moreover, (B,,),)K A = 0 implies

Hence the map (3.2.6) is surjective and its inverse is given by (3.2.5). The latter
representation (or Banach's homomorphism theorem) also implies the continuity
of the inverse. _

For later purposes we close this section by a qualitative version of a simple


time-dependent perturbation result.

3.2.2 Proposition Let assumptions (3.1.1), (3.1.2), and (3.2.1)-(3.2.4) be sat-


isfied. Also suppose that

(3.2.9)

and that
II (A, (B, ')')) - (A, (B, ')')) IIL(lE l ,lEo xlF(B))
(3.2.10)
::; A II (0 + A, (B, ')')) -111~tlEo xlF(B),IE,)
for some A E (0,1). Then

and
II (0 + A, (8, ')')) -IIIL(lEo xlF(B),IE,)
::; (1- A)-III(o+A,(B,,),))-IIIL(lEoxlF(B),IE') .

Proof From A E L(lEl' lEo) and (3.2.2) we infer 0 E L(lEl' lEo). Hence, thanks
to (3.2.9),
IV.3 Maximal Regularity 247

Put C := (0 + A, (.8, ,)) and note that C E Lis (lEI , lEo x JF(.8)) by Theorem 3.2.1.
Since, by (3.2.10),
1
:s >. ,
~

IIC - ClI.c(JE"JEoXlF(B)) Ilc- IIC(JEoXlF(B),JEI)

the relation C= [1 + (C - C)C- l ] C implies the assertion. _

Given the assumptions of Proposition 3.2.2, it follows that the IBVP

OtU + A(t)u = f(t) ,


B(t)u = g(t) , (3.2.11)
u(O) = uO

ha.<; for each (f, g, nO) E lEo x JF X ,lEI a unique solution n E lEI, provided g and uO
satisfy the compatibility condition (g, nO) E JF(.8). This result is the basis for prov-
ing, in Chapter VIII, solvability results for (3.2.11) without the smallness assump-
tion (3.2.10).
Chapter V

Scales of Banach Spaces

So far we have studied linear parabolic evolution equations in a fixed Banach


space. However, we have already seen, in connection with considerations of higher
regularity and problems with variable domains, that it is useful and necessary to
consider induced equations in interpolation spaces as well.
The theory of linear and nonlinear partial differential equations relies to a
large extent on the concept of weak solutions. As a rule, weak solutions belong to
some space (of distributions) that is larger than the space in which the problem
has originally been set up. These larger spaces are often characterized by dual-
ity arguments and the corresponding extended, or 'extrapolated', operators are
implicitly given via suitable bilinear forms (Dirichlet forms, for example).
In this chapter we develop an abstract framework for these situations. It
allows us to handle rather easily evolution equations in continuous scales of Banach
spaces. These scales are - in most practical cases - constructed with the help
of the generators of semigroups involved and are very well adapted to the given
problems. On the one hand, the scales are being used as precise measures for
the regularity of the solutions. On the other hand, the 'extrapolated' parts of the
scales provide the appropriate setting for an abstract theory of weak solutions.
In concrete applications, some of which are given in the second volume of this
treatise, it turns out that these extrapolation spaces often allow the reduction of
problems with variable domains to fixed domain problems. This is of particular
importance for the study of quasilinear equations.
Later chapters show that the general abstract theory of this chapter is very
flexible and applies to a great variety of concrete situations. In particular, the
theory of interpolation-extrapolation scales provides the right framework for 'con-
tinuous bootstrapping' arguments that can be used to prove that solutions of the
'extrapolated weak equations' are, in fact, solutions of the original problem. It also
allows for a priori estimates in 'weak norms' that can be used in proving global
existence results. For this we refer to the third volume. Lastly, it has the advan-
250 V Scales of Banach Spaces

tage - and this is most important for applications to systems of equations coming
from physics, for example - that it does not depend on coercivity conditions.
In Section 1 we develop the general theory of Banach scales. Section 2 is
devoted to the investigation of the behavior of semigroups and evolution equations
in Banach scales, particularly adapted to a given problem.

1 Banach Scales
In this section we study, in some detail, the concept of Banach scales. These scales
differ from the usual concept of 'scales of Banach spaces', that are simply one-
parameter families of suitably injected Banach spaces, by an additional structure,
expressed in terms of linear operators. In particular, it is shown how Banach scales
can be constructed starting with one Banach space E and a closed and densely
defined linear operator in it, having a nonempty resolvent set. It is most important
that we are able to construct supers paces of E and that we can characterize these
superspaces by a simple duality theorem. These 'extrapolated' spaces and the
techniques presented in this section constitute the abstract basis for our approach
to quasilinear parabolic problems in a 'weak setting'.

1.1 General Concepts


Before introducing the concept of Banach scales we prove a simple technical re-
sult concerning maximal restrictions of linear operators. Recall that B F is the
F-realization of the linear operator B in G if F '---+ G.

1.1.1 Lemma Let F and G be Banach spaces with F'---+ G and let B E C(G).
Suppose that (/1 + B)-l (F) c F for some /1 E p( -B). Then /1 E p( - BF) and

(l.1.1)

If also dom(B) C F then p(B) = p(BF) and (l.1.1) holds for each /1 E p(-B).

Proof Since BF E C(F), the first assertion is obvious. Suppose that dom(B) C F.
Then (A + B)-l (G) c F for each A E p( -B). Hence p(B) C P(BF) and

A E p(-B) .

Given A E p( -B F ) and y E G, put

Then x E dom(B) and (A + B)x = y. Hence A + B is surjective. If x E dom(B) and


(A + B)x = 0, dom(B) C F implies x E dom(BF). Consequently, (A + BF)X = 0,
V.l Banach Scales 251

so that x = O. Thus A + B is bijective. Now A E p( - B) follows from the closed


graph theorem .•

It is easily seen that dome B) C F iff (Ji + B)-l (G) c F for some Ji E p( - B).
Moreover, by iterating the resolvent equation it is easy to generalize the preceding
lemma to the case (Ji + B)-k(G) c F for some kEN (cf. [Are94, Proposition 1.1]).

Denote by A anyone of the sets N, Z, ~+, or lR. Let, for each a E A, be given
a Banach space Ea := (Ea, 11·lla) and an operator AaE £is(Eo+l' Eo.) such that,
for each pair a, (3 E A with a > (3,

and the diagram

(1.1.2)

is commutative. Then
[(Ea , Ac,) ; a E A]
is a Banach scale over the index set A. It is said to be one-sided if A E {N, ~+}
and two-sided otherwise. It is discrete if A E {N, Z} and continuous otherwise.
Of course, the diagram (1.1.2) ends on the right at (3 = 0 if the scale is one-sided.
d
The Banach scale [(Ea, Ac,) ; a E A] is densely injected if Ea '-> E(3 for a > (3.

Let [(Ea, Ac,) ; a E A] be a Banach scale over some A E {N, Z, ~+, ~}. Then
we have the following easy but useful consequences of the above definitions.

1.1.2 Remarks (a) Ac, E C(Ea) for a E A.


Proof Since Au is a toplinear isomorphism, the graph norm of Aa is an equiva-
lent norm for E a + 1 . Hence the assertion follows from Lemma 1.1.1.2 .•
(b) A(3 =:J Au for a > (3. Thus A o , considered as a linear operator in Eo, is closable
in E(3 and its closure equals the restriction of A(3 to

Ea+1,(3+1 := closure of En+l in E(3+1 ,

considered as a linear operator in EfJ.


252 V Scales of Banach Spaces

Proof This is a consequence of (a) and (1.1.2) .•


(c) Ao is the Eo-realization of AJ3 if a> {3.
(d) If the scale is densely injected, the operators Ao are completely determined
by AD (or by Aoo for any fixed 00 E A).
Proof This follows from (b) and (c). •
(e) p(Ao) = p{AJ3) for a,{3 E A.
Proof This is a consequence of Lemma 1.1.1 and (a)-(c), and p(Ao) =1= 0.•
(f) (Ao)j E Cis(Eo+j , Eo) for a E A and j E N, where Ao is considered as a linear
operator in Eo with domain Eo+!.
d
(g) If the scale is discrete, it is densely injected iff Eo+! ~ Eo for a E A.•

n
We put
Eoo:= Eo,
oEA
equipped with its natural projective limit toplogy, that is, the coarsest locally
convex topology for which all the natural injections jo: Eoo ....... Eo, x f-+ x are
continuous. We denote by Aoo the Eoo-realization of AD.

1.1.3 Proposition Eoo is a Prechet space such that Eoo ~ Eo for a E A, and
{11·llj ; j
EN} is a generating family of (semi) norms for Eoo. Moreover,

and Aoo is the Eoo-realization of Ao for each a E A.

Proof Since N is cofinal in A, it is clear that the family { II· IIj ; j EN} generates
the topology. Thus Eoo has a countable neighborhood basis of zero and is obviously
a Hausdorff space. Let (Xk) be a Cauchy sequence in Eoo. Then it is a Cauchy
sequence in E j for each j E N. Now the completeness of E j and an obvious diagonal
sequence argument show that (Xk) converges in Eoo. Thus Eoo is a Frechet space,
and Eoo ~ Eo for each a E A by definition of the topology of Eoo. The last part
of the assertion is an easy consequence of Ao E Cis(Eo+!, Eo) for each a E A, and
of Remark 1.1.2(c) .•

1.1.4 Proposition A one-sided discrete Banach scale is densely injected iff


Eoo is dense in Eo.

Proof Suppose that Eoo is dense in Eo. Let kEN, x E Ek, and c > 0 be given.
By Remark 1.1.2(f) we know that (AO)k E Cis(Ek, Eo). Since Eoo is dense in Eo,
V.I Banach Scales 253

there exists y E E= such that

Proposition 1.1.3 implies z := (Ao)-k y E E=. Hence

This shows that E= is dense in Ek. Hence Eoo C Ek+1 '--+ Ek guarantees that
Ek+1 :!.. Ek. Thus the scale is densely injected by Remark 1.1.2(g).
Conversely, suppose that the scale is densely injected. Define a new norm 1·1 k
on Ek by
k
Ixlk := L Ilxllj ,
j=O
Then, denoting by ik E £(Ek+1' E k ) the natural injection,
k-1k-1
Ilxllk :.:; Ixlk :.:; (1 + L II Ili£IIC(EH1,Et») Ilxll k '
j=O £=j
Hence 1·lk is an equivalent norm for Ek, and

x E Ej, k:.:; j . (1.1.3)

Now let Xo E Eo and c > 0 be given. Since Ek.t-l :!.. Ek for kEN, we find induc-
tively Xk E Ek such that IXk+1 - Xk Ik :.:; c2- k- 1 for kEN. Hence, given £ > m ;::: k,
we infer from (1.1.3) that
£-1 £-1
L L L
00

Ixe - Xmlk :.:; IXj+1 - Xjlk :.:; IXj+1 - Xjli :.:; c Tj-1 = c2- m .
j=m j=m j=m
Hence (Xj)r:=?k is a Cauchy sequence in Ek for each kEN. Thus, given any kEN,
the sequence (Xj)j~k converges in Ek to some Yk. Since E£ '--+ Ek for £ > k,
it follows that Y£ = Yk for 0 :.:; k < £ < 00. This shows that y:= Yo E Eoo and
Ily - xollo = Iy - xolo :.:; c. Thus Eoo is dense in Eo· •
The proof of the sufficiency part of Proposition 1.1.4 follows [Est84, Corol-
lary 2.2J. There a corresponding result, a so-called abstract Mittag-Leffler theorem,
is proven for a projective limit of complete metric spaces (also cf. [Bou53]). For an
application of the abstract Mittag-Leffler theorem to a situation closely related to
our case we refer to [ArEK94, Proposition 6.2J.

1.1.5 Corollary Let [(Ea, Aa) ; a E A] be a densely injected Banach scale.


d
Then Eoo '--+ Ea for a E A.
254 V Scales of Banach Spaces

Proof Let kEN with k > C\' be given. Then Eoo ' - t Ek !:... Ea. By applying
Proposition 1.1.4 to the densely injected discrete scale [(Ek+j,A,,+j) ; j EN] it
d
follows that Eoo '-t E". This proves the assertion. _

A Banach scale is said to be compactly injected if Ea '"'-+ E{3 for C\' > (3.

1.1.6 Proposition Let [(Ea, Aa) ; C\' E A] be a compactly injected Banach scale.
Then Eoo is a Mantel space and Eoo '"'-+ Ea for C\' EA.

Proof Since Eoo is a Frcchet space, it is barreled and metrizable. Hence it is a


Montel space if bounded subsets are relatively sequentially compact. Let B c Eoo
be bounded. Then B is bounded in E j for each j E N. Since Ej+ 1 '"'-+ E j , it follows
that B is relatively sequentially compact in E j for each j E N. Now an obvious
diagonal sequence argument shows that B is relatively sequentially compact in Eoo.
Hence the assertion follows from the factorization Eoo ' - t Ea+1 '"'-+ Ea for C\' E A. _

1.1.7 Proposition Let [(Ea,Aa) ; C\' E A] be a discrete Banach scale. Then it


is compactly injected iff (Ao) -1 E K( Eo) .

Proof Denote by io E £(El' Eo) the natural injection of El into Eo. Then we
deduce from the factorization
A-I
0
Eo El
'"
AOl~ ~o
Eo

that El '"'-+ Eo iff (AO)-1 E K(Eo). Thus we have to show that El '"'-+ Eo guaran-
tees Ea '"'-+ E{3 for C\' > (3. Since A c Z, this follows if we show that Ej+l '"'-+ E j
for j EA.
If j ~ 1, the commutativity of the diagram (see Remarks 1.1.2(b) and (f))

Eo

implies the desired result. If j ::; -1, we infer the assertion from the diagram
which is obtained from the preceding one by replacing Al by Aj+l and Ao by A j ,
respectively. _
V.1 Banach Scales 255

Two Banach scales [(EQ" A.) ; DC E A] and [(Fa, Boo) ; DC E A] over the same
index set A are isomorphic if there exist 100 E Lis (Eo. , Fa), DC E A, such that, given
DC, (3 E A with DC > (3, the diagram

co
I "

'"\
"
(l.1.4)

is commutative. If each 10. is also isometric, the two scales are isometrically iso-
morphic. Two Banach scales are equivalent if they are isomorphic and if E", == Fa
for DC E A.
A Banach scale is said to be reflexive if each Eo. is a reflexive Banach space.
It is a Hilbert scale if each Eo. is a Hilbert space. Lastly, a Banach scale is said
to satisfy interpolation inequalities if

x E Eo.,

provided DC, (3, 'Y E A with 'Y < (3 < DC.


The concepts and results of this subsection are new, except for Lemma 1.1.1,
whose proof follows [Are94]. A different notion of Banach scales has been intro-
duced and studied by KreIn [Kre60] and KreIn and Petunin [KrP66] (also see
[KrPS82, Chapter III]).

1.2 Power Scales


In this subsection we exhibit a simple general technique for constructing one-sided
Banach scales. For this we assume that

E is a Banach space, A E C(E), and 0 E peA) . (l.2.1)

Then it is known that


Aj E C(E) , j EN, (l.2.2)
where, of course, AD := IE (e.g., [HiP57, Theorem 2.16.4]).
256 V Scales of Banach Spaces

We put

j EN, (1.2.3)

so that Eo = E. Since 0 E p(Aj), it follows that 1I·lIi is equivalent to the graph


norm of Aj, that is,
Ei ~ D(Ai) . (1.2.4)
Hence E j is a Banach space by (1.2.2), and

IIxlik ~ IIAk-ill.qE) IIxll j , x E Ej , j > k,


shows that Ej ~ Ek if j > k. Thus
Aj := Ej-realization of A (1.2.5)

is well-defined.

1.2.1 Theorem Let assumption (1.2.1) be satisfied and define Ej and Aj for
j E N by (1.2.3) and (1.2.5), respectively. Then

[(Ej,Aj) ; j EN]
is a Banach scale, the power scale generated by (E, A). It is densely injected iff
A is densely defined, and it is compactly injected iff A-I E K(E).
Proof The first assertion is easily verified. If A is densely defined, that is,
EI ~ E, it follows from A-j E .cis(E,Ej) n.cis(EI ,Ej+1) that Ei+1 ~ Ej for
j ~ 1. Now the density assertion is a consequence of Remaxk 1.1.2(g). The last
part of the assertion follows from Proposition 1.1.7.•

1.2.2 Remarks (a) Every one-sided discrete Banach scale is essentially a power
scale. More precisely: if [(Ej, Aj) ; j E N] is a Banach scale then it is isomorphic
to the power scale generated by (Eo, Ao).
Proof This is a consequence of (1.1.2) and Remarks 1.1.2(a) and (b) .•
(b) If no confusion seems likely, we simply say: a power scale is generated by A,
instead of: by (E, A).
(c) Suppose that BE C(E) with 0 E p(B). Then [(Ej(B),Bj ) ; j EN], the power
scale generated by (E, B), is well-defined. Suppose that

1~k~n,

for some n E N, where Ek := Ek(A). Then Ek ~ Ek(B) and

IIAk B-kll.c(E) IIxlik ~ IlxlIEk(B) ~ IIBk A-kll.c(E) Ilxllk '


for 1 ~ k ~ n.
V.l Banach Scales 257

Proof It follows from (1.2.2) and the closed graph theorem that A kB-k and
Bk A-k belong to £(E). Hence the assertion is obvious .•
(d) Given A E p( -A), we have En(>\ + A) ~ En and

x E En,

for n E N.
Proof This follows from An(A+A)-n = [A(A+A)-If= [1-A(A+A)-lfand
(A + A)nA-n = [(,\,+ A)A-Ir = [,\,A-I + l]n, respectively, and from (c) .•

(e) Fix mEN and put (Fj,Bj) := (Ej+m' Aj +7n ) for j EN. Then [(Fj,Bj) ; j EN]
is the power scale generated by (Em' Am). It is isometrically isomorphic to the
power scale generated by (E, A).
Proof The first assertion is obvious from Remarks 1.1.2( c) and (e). Then, putting
101 := (Aa)-m for Q EN, Remark 1.1.2(f) implies that Ia E Cis(Ea, Fa) and that
the diagram (1.1.4) is commutative. It is easily verified that Ia is an isometric
isomorphism.•

Now we turn to the situation that A is of positive type, that is, A E P(E).
Recall that this means that A E PK(E) for some K 21 which, in turn, means
that A E C(E) with dense domain, JR.+ C p( -A), and (1 + s) lI(s + A)-III S; K for
s E JR.+. Thus, in particular, the power scale generated by A is well-defined. We
also recall the definition of the classes P(E; K, 19) of Subsection III. 4. 7.
For later purposes we prepare the following simple technical lemma, including
operators having bounded imaginary powers. Of course, [(Ej,Aj ) ; j EN] is the
power scale generated by (E, A).

1.2.3 Lemma (i) Suppose that A E P(E; K, 19) for some K 2 1 and iJ E [0,11").
Then Aj E P(Ej;K,iJ) for j EN.
(ii) If A E BIP(E; M, 8) for some M 2 1 and 820 then Aj E BIP(Ej ; M, 8)
for j EN.

Proof (i) From Lemma III.4.9.1(i) and Remarks 1.1.2 we infer that

for j EN, x E E j , and A E p( -A). This implies the assertion.


(ii) From (i) and Theorem III.4.6.5 it follows that (AjY is well-defined for
j EN and z E C If -1 < Rez < 1 and x E Ej+l then (AjYx is given by replacing
in Theorem III.4.6.5(ii) the operator A by Aj everywhere. Now we infer from
Remarks 1.1.2 that

j EN, x E Ej+1' -1 < Rez < 1 .


258 V Scales of Banach Spaces

Thus Lemma III.4.9.2(ii) implies

j EN, x E Ej+1' Rez = 0 .

Now the assertion follows from the density of Ej+1 in E j .•

Suppose that A E P(E). Then the fractional powers, k", of A are well-defined
for Q E JR, and we put

Q E JR+ . (1.2.6)

Observe that this is consistent with (1.2.3) if Q E N.

1.2.4 Theorem Let E be a Banach space and let A be a linear operator of


positive type in E. Define En by (1.2.6) and let An be the En-realization of A for
Q E JR+. Then

is a densely injected Banach scale, the one-sided fractional power scale generated
by (E, A). It satisfies interpolation inequalities. Moreover, it is compactly injected
iff A-I EK.(E).

Proof The fact that the fractional power scale is a densely injected Banach scale
is an easy consequence of Theorem III.4.6.5(v) and (vi).
Suppose that 0 < (3 < Q and that mEN satisfies m::::: Q < m + 1. Then,
thanks to Proposition III.4.6.3,

for 0 < t < 00. Hence

xEE, (1.2.7)

where
(1.2.8)

l'x)
and
'lj;(t) := sm-(3II(s + A)-lllm+I ds
for 0 < t < 00. From (III.4.6.l) it follows that

O<s<oo.
V.1 Banach Scales 259

Thus

1jJ(t) ::; K m+1 1 00


s-l-i3 ds = (3-1 Km+lCi3 , O<t<oo. (1.2.9)

Since A(s + A)-l = 1 - s(s + A)-I, we see from (III.4.6.1) that

O<s<oo. (1.2.10)

Moreover, putting z := m +1- 0:, we infer from Theorem III.4.6.5 and the resol-
vent equation that

Aa-m(s+A)-l = A-ZA(s+A)-l = ~ {(_>')-Z(>' + A)-lA(s + A)-ld>'


211'2 ir
= ~[ { (_>.)-Z A(>' + A)-l d>' _ { (_>.)-Z d>' A(s + A)-I] .
211'2 ir s - >. ir s - >.
Thanks to Cauchy's theorem and the choice of r, the last integral is zero. In the
second to the last integral we can replace r by the curve consisting of the two
half-lines {re±i'P ; r ~ O} for a suitable <p E (0,11'/2) and running from ooe-i'P
to ooei'P. Then using (1.2.10) with m = 1, we obtain the estimate

IIAa-m(s+A)-lll::; I+K roo r-Z[ls-rei'Pl-l+ls-re~i'Pl-l]dr


211' io
= s-z 1 + K roo r- Z [ 11- rei'PI- 1 + 11 _ re-i'PI- 1 ] dr = cs a - m- 1 .
211' io
Consequently, (1.2.8) and (1.2.10) imply

<p(t) ::; c l o
t
sa-i3- 1 ds
C
= __ t a - i3
0:-(3
, O<t<oo.

Now (1.2.7) gives

O<t<oo .

By minimizing the right-hand side with respect to t, it follows that

x E E.
We apply this inequality to Aa x for x E Ea. Then Theorem 1II.4.6.5(iii) implies

x E Ea.
Thus, putting (3' := 0: - (3 and omitting the dash afterwards, we infer from the
latter estimate that, given 0 < (3 < 0: < 00,

x E Ea. (1.2.11)
260 V Scales of Banach Spaces

Suppose that -00 < , < f3 < a < 00. Then (1.2.11) implies

By replacing in this estimate x by A'Yx and using Theorem III.4.6.5(i) and (iii) we
see that

(1.2.12)

for x E dom(Aa), provided -00 < , < f3 < a < 00. This proves, in particular, that
the fractional power scale satisfies interpolation inequalities.
If the fractional power scale is compactly injected, it follows from Propo-
sition 1.1.7 that A-I E K(E). Thus assume A-I E K(E). Then A-k E K(E) for
kEN. Suppose that a > 0 and fix kEN with k > a. Then (1.2.12) implies

x E E. (1.2.13)

Let (Xj) be a bounded sequence in E. Since A-k E K(E), there exists a subse-
quence, again denoted by (Xj), such that (A-k Xj ) is a Cauchy sequence in E.
Then we deduce from (1.2.13) that (A -a Xj ) is a Cauchy sequence in E, which
proves that A-a E K(E). Thus

O<a<oo. (1.2.14)

Suppose that 0 < f3 < a < 00. Let (Xj) be a bounded sequence in Ea. Then, thanks
to (1.2.14), there exists a subsequence, again denoted by (Xj), such that (Xj) is a
Cauchy sequence in Eo. Hence

implies that (Xj) is a Cauchy sequence in E{3. This proves that Ea ~ E{3 which,
together with (1.2.14), guarantees that the fractional power scale is compactly
injected. _

1.2.5 Corollary Let A be an operator of positive type in some Banach space E.


Then A-I E K(E) iff A-a E K(E) for some a > o.

We close this subsection with the following useful technical observation.

1.2.6Proposition SupposethatAEP(E) and let [(Ea,Aa); aEJR+] be the


fractional power scale generated by (E, A). Fix f3 E (0,1] and, E JR+ and put
(Fa, Ba) := (Ea{3+'Y' A a{3+'Y) for a E JR+. Then [(Fa, Ba) ; a E JR+] is the frac-
tional power scale generated by (E'Y' (A.,,){3). If f3 = 1, it is isometrically isomor-
phic to the fractional power scale generated by (E, A).
V.l Banach Scales 261

Proof By an obvious modification of the proof of Lemma 1.2.3(i) we infer that


A")' E P(E")'). Hence (A")').8 E P(E")') by Proposition III.4.6.10, so that the fractional
power scale generated by (E")', (A")').8) is well-defined. Theorem 111.4.6.13 and Re-
mark 1.1.2(b) imply [(A")').8]"'x = Aa.8x for 0: 2: 0 and x E E a.8+")' = Fa. Thus
IiBaxllF = II [(A")').8]'" x II")' = IIA")'Aa.8x ll
(1.2.15)
= II A a.8+")'x II = Ilxlia{H")' = IixliFa
for 0: E 1R+ and x E Ea.8+"Y' This proves the first assertion. If (3 = 1, it is now
clear that
o:EIR+ ,
these isomorphisms are isometric, and the diagrams (1.1.4) are commutative.•

The principal results of this subsection are well-known, at least implicitly and
if -A generates a strongly continuous analytic semigroup (e.g., [Sob66], [Fri69 ,
Section ILl4], [Hen81, Section 1.4], [Paz83, Section 2.6]). Under this assumption,
to the best of our knowledge, the first proof of Theorem 1.2.4 is in [Ama78,
Lemma 2.1]. The above derivation of the moment inequality (1.2.12) follows
[Kre72, Section I.5.3].

1.3 Extrapolation Spaces


Let F be a normed vector space. Recall that a completion (F,J) of F consists
of a Banach space F and an isometric linear isomorphism J from F onto a dense
linear subspace of F.
A particular completion can be obtained by taking for F the Banach space
whose elements are equivalence classes (Xj) of Cauchy sequences in F, where two
Cauchy sequences (Xj) and (Yj) are said to be equivalent if IIxj - Yjli---+ O. This
space is given the natural linear structure, and its norm is defined by

Then J(x) := (x), where (x) is the constant sequence (x,x, ... ) for x E F (e.g.,
[Yos65, Section LlO]).
Let (F1,)1) be another completion of F. Then there exists an isometric
hE £is(F, F1 ) such that the diagram
F

h
F
262 V Scales of Banach Spaces

is commutative. In this sense completions are unique, except for isometric lin-
ear isomorphisms. If no confusion seems likely, we fix a particular completion
F:= (F,J), the completion of F, and identify F with the subspace J(F) of F.
Throughout the remainder ofthis subsection we presuppose that E := (E, 11·11)
is a Banach space and A E C(E) with 0 E p(A). Then [(Ej, Aj) ; j EN] is the dis-
crete power scale generated by (E, A).
Observe that IIA-l·1I is a norm on E which is weaker than 11·11. Thus the
completion
(1.3.1)
of the normed linear space (E, II A -1 '11) is a well-defined Banach space and

(1.3.2)

It is said to be the extrapolation space of E generated by A. Note that

IIAxll-l = Ilxll , x E El , (1.3.3)

and that
yEE. (1.3.4)
Thus let
E1,o := closure of El in E . (1.3.5)
Then it follows from (1.3.3) and (1.3.4) that there exists a unique continuous
extension A_l of A such that A-l is an isometric isomorphism from E 1,o onto E-l
and (A_l)-l :J A-l. In the following, A_l is said to be the E_l-realization of A.

1.3.1 Proposition (i) A_l E Cis(El,o, E_ l ) n C(E-l) and this isomorphism is


isometric.
(ii) A is closable in E-l and its closure equals A_ l .
(iii) p(A-d = p(A).
Proof (i) Since A_l E Cis(E 1,o, E_ l ), the graph norm of A_l is an equivalent
norm on E 1,o. Thus A_1 E C(E-d by Lemma 1.1.1.2. (ii) is now obvious. Lastly,
(iii) is an easy consequence of (i), (ii), and Lemma 1.1.1. •

Since A_l E C(E_ l ) and 0 E P(A-l)' it follows that

E-2 := E_ 1(A_ l ), the extrapolation space of E-l generated by A_ l ,


(1.3.6)
and A_ 2 , the E_ 2 -realization of A_I,

are well-defined. It follows that

(1.3.7)
V.l Banach Scales 263

that
(1.3.8)
that this isomorphism is isometric, and that A-1 is closable in E-2 and A-2 is
its closure.
However, relations (1.3.7) and (1.3.8) have to be interpreted with care. To
be more precise, let (E- 1,j-1) and (E_ 2,j-2) be the completions of (E, \\A- 1'\1)
and (E_l, \\(A_d- 1 '\I), respectively. Then

E j-I E j-2 E
------+ -1 ------+ - 2

and j-1 and j-2 are linear isometries onto dense linear subspaces of E-1 and E-2'
respectively. Thus if we identify, as above, E-1 with its image j_2(E-d in E-2' we
also identify j-1 (E) with j-2 (J-1 (E)). Therefore, by constructing E-2 we have to
replace the already constructed extrapolation space E-1 by an isomorphic image
so that (1.3.7) is true.

1.3.2 Theorem Suppose that A E C(E) is densely defined with 0 E p(A) and fix
mEN. Define E-k and A-k for 1 ::; k ::; m inductively by

E_k := (E_k, \\'\\-k):= (E-k+1' \\(A_k+1)-1·lIt


and
A-k := closure of A-H1 in E-k .
Put
JEN.
Then [(Pj,Bj); j EN] is the power scale generated by (E-m,A- m). It is iso-
metrically isomorphic to the power scale generated by (E, A).

Proof The above considerations imply that the pairs (E-k, A_ k ) are well-defined
for 1 ::; k ::; m. Since E 1,o = Eo = E, thanks to the fact that A is densely defined,
it is an easy consequence of (1.3.3)-(1.3.5) that [(Fj, B j ) ; j E N] is indeed the
power scale generated by (E-m, A-m). The last part of the theorem follows now
from Remark 1.2.2(e) .•

Suppose that A is densely defined and mEN. Then [(Ej - m , A j - m ) ; j EN]


is said to be the extrapolated power scale of order m, generated by (E, A), and
we also denote it by
[(Ej, Aj) ; j E Z n [-m, 00)] .
Thanks to Theorem 1.2.1, it is a densely injected, one-sided, discrete Banach scale.

1.3.3 Remark Given k E Zn [-m, 00), it is clear that [(Ej+k,Aj+k) ; j EN]


is the power scale generated by (E k, A k ) and that [( Ej , Aj) ; j E Z n [-m, 00 )] is
the extrapolated power scale of order m + k, generated by (Ek, Ak)' •
264 V Scales of Banach Spaces

1.3.4 Proposition Suppose that A is densely defined and mEN. Then

[(Ej,Aj) ; j E zn [-m, (0)]


is compactly injected iff (Ak)-l E K(Ek) for some k E Z with k 2: -m. If this is
the case, (Ak)-l E IC(Ek) for each k E Z with k 2: -m.

Proof If the scale


[(Ej,A j ) ; j E Zn [-m, (0)] (1.3.9)
is compactly injected, it follows that

(1.3.10)

is compactly injected for each k E Z with k 2: -m. Hence (Ak)-l E IC(Ek) by


Proposition 1.1.7. Conversely, suppose that (Ak)-l E IC(Ek) for some k E Z with
k 2: -m. Then the scale (1.3.10) is compactly injected by Proposition 1.1.7. Due
to Theorem 1.3.2 and Remark 1.3.3 the full scale (1.3.9) is isomorphic to (1.3.10).
Hence it is compactly injected as well. •

It is tempting to put
E- oo := UE j .
jEZ
Note, however, that this is not meaningful, in general, since we have to construct
the negative spaces inductively and, by constructing E_ m - 1 from E_ m (and A_ m ),
we have to replace the spaces E-l' ... ,E-m already obtained by isomorphic im-
ages. On the other hand, if we do not use the above identifications but work with
the continuous injections j-k : E-k+l --+ E-k provided by the completion process,
we can construct a 'superspace' E- oo as the inductive limit of the spaces E_ j ,
j EN, with respect to these injections (e.g., [Sch71 , Section II.6.3]). Since these
constructions are rather cumbersome and we do not need them, we refrain from
giving details. Moreover, in Subsection 1.4 we give a simple sufficient condition
guaranteeing the existence of a universal superspace E- oo such that E j ~ E- oo
for each j E Z.

Suppose that B E C(E) is densely defined with 0 E p(B). Then the extrapo-
lated power scale [(Ej(B), B j ) ; j E Z n [-m, (0)] of order m, generated by (E, B),
is well-defined. In the following proposition we give a sufficient condition for
E-k(B) ~ E-k := E-k(A) for 1 ::; k ::; m. Recall that the dual of a closed and
densely defined linear operator is well-defined and that it has the same resolvent
set as its predual. Hence the following assumptions are meaningful.

1.3.5 Proposition Suppose that A and B are closed and densely defined in E
with 0 E p(A) n p(B), and that

(1.3.11)
V.1 Banach Scales 265

Then E_k(B) == E-k and

II(B')k(A')-kllctEI) IIxll_ k :::; IlxIlE_k(B) ~; II (A')k(B')-kll.C(E') Ilxll- k


for x E E-k and 1 :::; k :::; m.
Proof Since (A')k and (B')k are closed in E' by (1.2.2), it follows from (1.3.11)
and the closed graph theorem that (B')k(A')-k and (A')k(B')-k belong to C(E').
Hence we deduce from

x E Ek(B) , x' E E' ,

and the density of Ek(B) in E that A-kBk has an extension (A-kB k )- E C(E)
and that

Thus, given x E E,

Similarly, there exists a unique extension (B- k Ak)- E C(E) of B- kAk satisfying

Thus, given x E E,

Now the assertion follows from the density of E in E-k .•

1.3.6 Corollary Suppose that A E C(E) is densely defined with 0 E p(A). Given
mEN and>' E p( -A), it follows that E-k(>' + A) == E-k := E-k(A) with

(1 + 1>'IIIA-llI)-k Ilxll-k :::; IIxIlE_k(>-+A) :::; (1 + 1>'111(>' + A)-lll)k Ilxll-k


for x E E-k and 1 :::; k :::; m.
Proof This follows from

and
[(>. + A),]k(A')-k = ([(>. + A)A-1]k)' = [(AA- l + l)k]'
and Proposition 1.3.5 .•

Next we turn again to the case that A is of positive type. First we prove the
following simple extension of Lemma 1.2.3.
266 V Scales of Banach Spaces

1.3.7 Lemma (i) Suppose that A E P(E; K, '!9) for some K ?: 1 and '!9 E [0, 1l").
Then Aj E P(Ej ; K, '!9) for j E Z n [-m, 00).
(ii) If A E BIP(E; M, 0) for some M?: 1 and 0 ?: 0 then Aj E BIP(E j ; M, 0)
for j E Z n [-m, 00 ) .

Proof Thanks to Theorem 1.3.2 and Lemma 1.2.3 it suffices to prove the asser-
tions for j = -m.
(i) Since

for x E E and A E p( -A-m) = p( -A), the first assertion follows using the density
of E in E_ m .
(ii) From (i) and Theorem III.4.6.5 we infer that (A-mY is well-defined for
z E C. Similarly as in the proof of Lemma 1.2.3 we see that (A_mYx = AZx for
x E EI and -1 < Re z < 1. Thus

for x E EI and Re z = 0, and the desired conclusion follows from the density
of EI in E_ m . •

Suppose that A E P(E) and mEN. It follows from Lemma 1.3.7 by induction
that A-m E P(E). Thus, letting (F, B) := (E-m' A-m), Theorem 1.2.4 guarantees
that the one-sided fractional power scale [(Fa, Bo.) ; 0: E JR+] is well-defined. In
the following, we put

-m :S 0: < 00 , (1.3.12)

and call

the extrapolated fractional power scale of order m, generated by (E, A). Note
that (1.3.12) is consistent with the earlier definitions of (Ec" Acl') for 0: E JR+
and 0: E {-I, -2, ... , -m}.

1.3.8 Theorem Suppose that A E P(E) and mEN. Then the extrapolated frac-
tional power scale of order m, generated by (E,A), is a densely injected Banach
scale satisfying interpolation inequalities. It is isometrically isomorphic to the frac-
tional power scale
[(Ea,Aa) ; 0: E JR+] ,
generated by (E, A). It is compactly injected iff A-I E K.(E). If 0 < 0: :S m then
E- cx is a completion of (E, IIA-a·ll) and A_a is the closure of A in E_ a .
V.I Banach Scales 267

Proof The first assertion is an immediate consequence of Theorem 1.2.4. The


second One follows from Proposition 1.2.6. Now the assertion concerning compact
injections is a consequence of Proposition 1.3.4 and Theorem 1.2.4.
Suppose that 0 < ex < m. Then, thanks to Theorem III.4.6.5(iii),

x E Em-a.

Since Em-a:!.. E, it follows that IIA-axll = IIxll- a for x E E. Thus we deduce


from E:!.. E_ a that E_ a is a completion of (E, IIA-a·II). Since

and A c A-a, we see that A is closable in E_ n and its closure equals A_c<, •

1.3.9 Corollary Suppose that A E P(E). Let [(Ea,Aa) ; ex E [-m,oo)] be the


extrapolated fractional power scale of order m, generated by (E, A). Then Ea ~ Ej3
for -m::; ex < f3 < 00, and (A a )a-j3 is an isometric isomorphism from Ea onto Ej3.

In order to have a unified language we call one-sided [fractional] power scales


extrapolated [fractional] power scales of order zero.

The above construction of E-1 and A-1 occurs perhaps for the first time
in [Nag83] (also cf. [WaI86]). Without being aware of this, it has been reintroduced
in [Ama87] (that paper had already been submitted in 1984; also cf. [Ama86]) and,
again independently, in [Liu89]. Around the same time Da Prato and Grisvard
[DaPG84] proposed a different construction of an extrapolation space E-1 and
an extension iLl of A. In [Ama88c, Remark 6.2] it has been shown that the
extrapolation pairs (E_ 1, A_ 1) and (E-1' A-d are isomorphic (cf. [DaP84] for a
corresponding statement; also see [vN92]).
In [Nag83] and [WaI86] only extrapolated discrete scales have been stud-
ied. Extrapolated fractional power scales have been investigated in [Ama87] and
in [Liu89]. The latter author also introduces at each point ex of the scale induc-
tive limit spaces Up>a Ej3 and projective limit spaces nj3<a Ea. All the papers
cited above have beEll motivated by the theory of semigroups and, as a rule, it is
assumed that A E 9 (E).
Lastly, it should be mentioned that more recently there has been developed
an extrapolation theory by B. Jawerth and M. Milman (e.g., [JM91]) that is of an
entirely different nature than the one considered here (also see [MiI94]).

1.4 Dual Scales


Throughout this subsection we suppose that E is a reflexive Banach space and
A E C(E) with dense domain and 0 E p(A), unless explicitly stated otherwise. Re-
268 V Scales of Banach Spaces

call that A' E C(E') is densely defined and peA') = peA). In order to avoid confu-
sion with too many dashes we put

EH := E', AH:= A' .

Then, given mEN, the extrapolated power scales of order m

(1.4.1 )

and
[(EJ,A!) ; j E Zn [-m,oo)] , (1.4.2)

generated by (E, A) and (EH, AH), respectively, are well-defined. The scale (1.4.2)
is said to be the dual scale of the Banach scale (1.4.1).
Now we establish an important duality theorem for which we need some
preparation. Here and hereafter we denote by (., .) := (., .) E the duality pairing
between E' and E.

1.4.1 Lemma Given k E 1':1 with 1:::; k :::; m, put

Then

is a separating continuous bilinear form of norm 1.

Proof We infer Ak E £is(Ek' E) as well as (A~k)-k E £is(E~k' EH) from Re-


mark 1.1.2(f). It is clear that these isomorphisms are isometric. Now the assertion
is obvious .•

The following simple observation is rather useful and deserves to be singled out.

1.4.2 Lemma Let X and Y be vector spaces and let Band C be linear operators
with domains in X and ranges in Y such that B ::J C. If B is injective and C is
surjective then B = C.

Proof Suppose that x E dom(B). Since C is surjective, there exists y E dom(C)


such that Cy = Bx. Thus B ::J C implies Bx = By, that is, x - Y E ker(B) = {O}.
Hence x = y E dom(C) so that dom(B) C dom(C), which proves the assertion. _

After these preparations we can prove the following duality result. Here and in
similar situations (Ak)' denotes the dual of Ak E C(Ek), that is, as a linear (in gen-
eral: unbounded) operator in Ek and not as a bounded linear map from Ek+l to E k.
V.I Banach Scales 269

1.4.3 Proposition Suppose that kEN with 1 ::;; k ::;; m. Then

(1.4.3)

and
(1.4.4)

with respect to the duality pairing (" .) k'

Proof It follows from Lemma 1.4.1 that x~ f--t (x~, ')k defines a linear isometry
from E~k into (E k),.
Given f E (E k)" we deduce from Ak E Cis(Ek , E) the existence of y~ E E~
such that the diagram

/
f~ / y~

is commutative, that is,

(1.4.5)

Since (A~k)k E Cis(E~, E~k)' the element x~ ;=(A~k)ky~ E E~k is well-defined,


and (1.4.5) implies f(x) = (X~,X)k for x E Ek. Hence (Ek)' = E~k'
Since dom(Ak) = Ek+l is dense in Ek, the dual (Ak)' of Ak E C(Ek) with
respect to the duality pairing (-, . h is well-defined, and (Ak)' E C(E~k) by (1.4.3).
As Ak is the Ek-realization of A and A~k ::) A", it follows that

for x~ E E~k+l and x E Ek+l. Hence (Ak)' ::) A~k' Thanks to p(Ak) = p( (Ak)'),
Remark 1.1.2(e), and Theorem 1.3.2, 0 E p((Ak)') n P(A~k)' Thus (Ak)' = A~k
by Lemma 1.4.2 .•

1.4.4 Corollary If 0 ::;; £ < k ::;; m,


270 V Scales of Banach Spaces

Proof Suppose that x~ E E" and x E E k . Then, thanks to A~k :J A~e :J A~


and A :J A e,
(X",X)k = «A~)-kx",Akx) = «A~)£-k(A~)-£x~,Ak-£Aex)

= «A£-k)'(A")-ex~,Ak-eA£x) = «A~)-£x~,A£x) = (x~,X!£.

Now the assertion follows from E~ !:... E~£ .•


It is a consequence of Corollary 1.4.4 that

O<k~Tn.

Since E" is dense in E~k' we see that (., . h is uniquely determined by (., .), that
is, (., .) k is naturally induced by (., .!. Hereafter, in abuse of notation, we simply
write (., .) for (., .) k' kEN with 0 ~ k ~ Tn, if no confusion seems possible.

1.4.5 Remarks (a) It should be observed that everything said above remains
valid if E is not reflexive but A' is densely defined.
(b) If E is not reflexive and A' is not densely defined, it is still true that
A' E C(E') and 0 E peA'). Thus, letting E~ := E' and A~ := A', the dual power
scale [( EJ, A~) ; j E N] is still well-defined .•

As an easy consequence of Proposition 1.4.3 we prove now the following


fundamental theorem whose corollary contains a most important characterization
of the extrapolation spaces E-k for 1 ~ k ~ Tn.

1.4.6 Theorem Ek and E~ are reflexive for k E Z n [-Tn, 00). Moreover, with
respect to the duality pairing (., . ) ,

-Tn ~ k ~ Tn . (1.4.6)

Proof Since Ek [resp. E~l is toplinearly isomorphic to E [resp. E~l and since E,
hence E~, is reflexive, the first assertion follows. From Proposition 1.4.3 we know
the validity of (1.4.6) for 0 ~ k ~ Tn. Thus

(1.4.7)

Put (Fk,Bk):= (E~,AD for k E zn [-Tn,oo). Then, using (1.4.7) for k = 0, we


infer that the dual scale of [(Fk' B k ) ; k E Z n [-Tn, 00)] equals the scale (1.4.1).
Thus, by applying (1.4.7) to Fk and Bk, we see that

O~k~Tn,

with respect to (., .). This proves the assertion .•


V.1 Banach Scales 271

1.4.7 Corollary If 1:::; k :::; m then E-k = (E~)' and A-k = (A~)' with respect
to the duality pairing (', . >.

On the basis of Theorem 1.4.6 we can now extend our extrapolation scales
to get two-sided scales. For this we need some preparation.
Let X and Y be LCSs such that i: X!:.. Y. Then i' E L(Y',X') and i' is
injective by the density of X in Y. Observe that

i' (y') = y' 0 i = y'IX E X~ , y' E Y' ,

where

X~ := {x' EX' ; x' is continuous with respect to thc Y-topology of X} .

Conversely, given x' E X~, the density of X in Y implies the existence of a unique
y' E Y' with y' =:> x'. Thus i'(Y') = X~. In the following, we identify y' and X~
by means of the injection i'.
Recall that a LC S X is semireftexive if the canonical injection

K:X--->X", Xf-7(',x>x

is surjective. (It is reflexive if K is a toplinear isomorphism.)

1.4.8 Proposition Let X and Y be LCSs. Then

X !:... Y implies Y' '--+ X'

and
(y',x>X = (y',x>y , y' E Y', x EX.

If X is semirefiexive,

X !:... Y implies Y'!:... X' .

Proof The first assertion follows from the considerations above. Thus suppose
d
X Y and X is semireflexive. Let x" E X" satisfy (x", y'> x'
'--+ =0 for y' E Y'.
Then, putting x := K- 1 (x") EX,

(y',x>X = (y',x>y = 0, y' E Y' ,

so that x = 0, hence x" = O. Now the density of Y' in X' follows from the Hahn-
Banach theorem. _
272 V Scales of Banach Spaces

Recall from Proposition 1.1.4 that

O<k<j<oo,

and that Eoo is the projective limit of the sequence of Banach spaces (Ej ) with
respect to the continuous injections Eoo "---t E j . Since E j is reflexive for j EN,
it follows that Eoo is semireflexive (e.g., [Sch71 , Section IV.5.8]). Hence Proposi-
tion 1.4.8 implies

and

(X',X/E= = (X',X/E j = (X',X/Ek = (x',x/ , x' E E', x E Eoo ,

for 0 < k < j < 00. By density, the Ej-Erduality pairing is for each j EN com-
pletely determined by the E'- E-duality pairing (.,. /. Thus (.,. / Ek = (., ./ k for
o :S k :S m, and, letting
E~oo := (Eoo)' ,
it follows from Theorem 1.4.6 that

O:Sk<j:Sm. (1.4.8)

Note that this is true for every mEN. FUrthermore, if we replace in (1.4.8) the
integer m by some larger integer in, we obtain the extrapolation scale

[(EJ, A~) ; j E Z n [-in, (0)]


of order in without changing the previously constructed extrapolation scale of
order -m, generated by An. Thus, proceeding inductively, we can now indeed
construct the two-sided power scale generated by (En, An),

By replacing in the above arguments the power scale [(Ej , Aj) ; j E N] by the
scale [(EJ,A~) ; kEN] and using Corollary 1.4.7 we obtain the two-sided power
scale generated by (E, A),

In the following theorem we collect these observations:

1.4.9 Theorem Let E be a reflexive Banach space and suppose that A E C(E)
with dense domain and 0 E p(A). Then the two-sided power scales
V.1 Banach Scales 273

generated by (E, A) and (E~, A~), where E~ := E' and A~ := A', respectively, are
well-defined densely injected reflexive Banach scales. Moreover,

kEZ,

with respect to the duality pairing naturally induced by the E' -E pairing (., .).

1.4.10 Remark Since, by definition, E- oo is the dual of the projective limit


space E~ = lim<-- EJ, which is a Frechet space, and since E~ ~ EJ for j E N
by Proposition 1.1.4, it follows that E- oo is the inductive limit of the Banach
d
spaces E j , j E Z, with respect to the natural injections E j '---+ Ek for j > k, that
is, E- oo = lim---> E j (cf. [Sch71 , Sections IV.3.4, IV.4.4, and Corollary 2 to Theo-
rem IV.4.3j. Thus

j >k .

Moreover, Proposition 1.1.3 implies

and it is clear that Aj is the Errealization of A-oo for j E Z. Similarly,

j > k,

and
A~oo := (Aoo)' E L:aut(E~oo) .
If A -1 E K(E) then (A~)-l E K(E") by Schauder's theorem. Therefore Eoo and E~
are Montel spaces, hence reflexive, by Proposition 1.1.6. Thus E- oo and E~oo are
Montel spaces, consequently also reflexive (e.g., [Sch71 , Section IV.5.9]) .•

Lastly, we extend the above results to the case of reflexive fractional power
scales provided A is of positive type. For this and for later use we prepare the
following technical observation:

1.4.11 Lemma Let E be a reflexive Banach 8pace and let A E P(E). Then

Rezi-O. (1.4.9)

If A E BIP(E; M, 0) for some M 2: 1 and 02:0 then A' E BIP(E'; M, 0) and


(1.4.9) holds for Rez = 0 as well.

Proof If Rez < 0, relation (1.4.9) is an easy consequence of the integral repre-
sentation (111.4.6.22). Then we obtain (1.4.9) for Rez > 0 from (III.4.6.12). Thus
suppose that A E BIP(E; M, 0). Then IIAzl1 ::; KMe(J11mzl for -1 < Rez < 0 and
274 V Scales of Banach Spaces

a suitable constant K 2:: 1 by Lemma III.4.7.4(ii). Hence we obtain from (1.4.9)


that II (A'Y II ::; K M eO l1m zl for -1 < Re z < 0 which, thanks to Lemma 1II.4. 7.4(i),
entails A' E BIP(E';KM,e). By (1.4.9),
((A')Zx',x) = (x, AZxI , x' E E', x E E, Rez < o.
Thus, letting z tend to it for some t E~, we infer from Theorem III.4.7.1 that
x' E E', x E E, t E ~ ,

which implies that (1.4.9) is also true for Re z = o. Lastly,


t E ~,

shows that A' E BIP(E';M,e) .•

Now it is easy to prove the following theorem on fractional power scales. For
simplicity we restrict ourselves to the reflexive case, the most important one for
applications.

1.4.12 Theorem Let E be a reflexive Banach space and suppose that A E P(E).
Then the two-sided fractional power scale generated by (E, A),
[(Em Ax) ; Q E~] , (1.4.10)
and its dual scale,
(1.4.11)
generated by (E", A"), are well-defined densely injected reflexive Banach scales
satisfying interpolation inequalities. Moreover,
QE~, (1.4.12)
with respect to the duality pairing naturally induced by (., . I.

Proof Theorem 1.3.8 and the considerations leading to Theorem 1.4.9 imply
that the scales (1.4.10) and (1.4.11) are well-defined and satisfy interpolation in-
equalities. Given Q > 0, put

Note that this definition coincides with the one of Lemma 1.4.1 if Q E N. From
Theorem 1.3.8 we also infer that

Hence
(., ·Ia: E~a x Ea -4 IK
is a separating continuous bilinear form of norm 1. From this and Lemma 1.4.11
we infer, by replacing in the proof of Proposition 1.4.3 everywhere k by Q, that
(1.4.12) is true for Q > 0 with respect to the duality pairing (., ·I a .
V.l Banach Scales 275

Suppose that 0 < f3 < a. Then, by replacing k by a and £ by f3 in the proof


of Corollary 1.4.4 and by using (1.4.9), we see that

x~ E E~{3' x E Ee> .
This shows that (', .) e> coincides with the duality pairing naturally induced by (', .).
Lastly, since En = (dom(AOO), IIA" ·11) and E_Q is a completion of (E, IIA '11) -Q

for a> 0, it follows that AQ E Lis(EQ, E) for a E R Hence E Q and, consequently,


E~ also, are reflexive for a E R Now the proof of Theorem 1.4.6 applies to give
(1.4.12) for each a E R •

The preceding duality theorem for reflexive fractional power scales has first
been proven in [Ama87] under the assumption that A E Q(E). It has then been
rediscovered in [Liu89]. To be precise, the results in [Ama87] and [Liu89] are only
valid for extrapolated fractional power scales of order m for some mEN, since
in neither paper - and also not in the papers by [Nag83] and [WaI86], where
two-sided discrete power scales are being studied - the existence of universal
superspaces E_oc> and E~CXJ has been established. The existence of these spaces,
hence the fact that the two-sided [fractional] power scales are well-defined indeed,
is proven here for the first time.

1.5 Interpolation-Extrapolation Scales


Let E be a Banach space and suppose that A E C(E) is densely defined with
o E p(A).
Then, given mEN, the extrapolated discrete power scale

of order m, generated by (E, A), is a well-defined densely injected Banach scale.


e
We specify for each E (0,1) an admissible interpolation functor (', ')0 and put

(1.5.1)

for j < a < j + 1 and j E Z n [-m, 00 ).

1.5.1 Theorem [(Eoo , A,) ; a E [-m,oo)] is a densely injected Banach scale, the
interpolation-extrapolation scale of order m, generated by (E, A) and (', ')0,
o< e< 1. It is compactly injected iff A-I E J( (E). If this is the case,

a E [-m, 00) .

Proof Thanks to the admissibility of the interpolation functors (', ')0 we ob-
tain the first assertion by interpolation from Theorems 1.2.1 and 1.3.2 and Re-
mark 1.1.2(c). The second assertion is an easy consequence of Proposition 1.3.4
276 V Scales of Banach Spaces

and Theorem I.2.11.1. The last assertion is implied by the commutativity of


the diagram

where io: : Eo:+1 ~ Eo: is the natural injection. _

If we assume that A is of positive type, we can exploit the relations between


the domains of fractional powers and suitable real interpolation spaces to obtain
further information.

1.5.2 Theorem Suppose that A E P(E). Then the interpolation-extrapolation


scale of order m, generated by (E, A) and (-, ')0, 0 < e < 1, satisfies interpolation
inequalities.

Proof Suppose that -m ::; k ::; j < £ < 00, where k, j, and £ are integers. Then,
letting F := Ek and B := A k , it follows from Theorem 1.3.2 that

i E {k,j,£} .

Hence (I.2.9.6) implies

(1.5.2)

with the obvious interpretation if j = k. Moreover, we infer from (I.2.11.4) that

j<oo<j+l.

From (1.5.2) and the reiteration property (I.2.8.2) for the real method we de-
duce that

provided j < a < j + 1. Thus it follows that

() d d 0
E k , Ec (o:-k)/(e-k),l '---+ Eo: '---+ (Ek, EC)(o:-k)/(C-k),oo (1.5.3)

for k < a < i!.


Finally, suppose that -m ::; 'Y < j3 < a <00 and fix k, i! E Z n [-m, 00) such
that k < 'Y < j3 < a < i! if 'Y > -m, and put k := 'Y otherwise. Then (1.5.3) and
V.I Banach Scales 277

the reiteration property (1.2.8.2) imply

(E"{, EoJTJ,l '-+ (Ek , ER)((3-k)/(R-k),l '-+ E(3 , (1.5.4)

where "7 := ((3 - ,)/(0: - ,). Hence

by Proposition 1.2.2.1 and Example 1.2.4.1. This proves the assertion. _

It should be noted that, thanks to Theorem 1.2.11.1, interpolation inequal-


ities are always satisfied if j ::; I < (3 < 0: ::; j + 1 for some j E Z, without the
assumption that A be of positive type. The latter assumption has only been used
to interpolate 'across integers'.
Now it is easy to show that the interpolation-extrapolation scale enjoys an
almost reiteration property similar to the one of Remark 1.2.11.2(a). Note that
the latter remark is not applicable since, in general, EO'. ~ (Ej , Ek)(a-j)/(k-j) if
j < 0: < k with j, k E Z and k - j ::::: 2.

1.5.3 Theorem Suppose that A E P(E). Then the interpolation-extrapolation


scale of orderm, generated by (E, A) and (', ')e, 0 < e < 1, possesses the following
almost reiteration property: if -m ::; 0: < (3 < 00 and 0 < "7- < "7 < "7+ < 1 then

Proof From (1.2.11.4), (1.2.5.2), and (1.5.4) we easily deduce that

Similarly, we see that

Since the density assertion is obvious, the theorem follows. _

Suppose that (', ')0 is one of the real interpolation functors (', ')e,q, 1::; q < 00,
or (', ')e = (', ·)~,oo for each e E (0,1), or (., ')e = [', 'le for each e E (0,1). Then it
follows from Remark 1.2.11.2(b) that the reiteration property

0<"7 < 1 , (1.5.5)

holds, provided j ::; 0: < (3 ::; j + 1 for some j E Z. In general, however, this is
not true if we interpolate 'across integers'. For example, E j ~ (Ej-l, E j +1)1/2, in
general (see Chapter VI). In the following, we give a sufficient condition for the
278 V Scales of Banach Spaces

reiteration property (1.5.5) to hold unrestrictedly. Recall that [., ·]e denotes the
complex interpolation functor of exponent e.

1.5.4 Theorem Suppose that A E BIP(E). Then the interpolation-extrapolation


scale [(Ea, Aa) ; 00 E [-m, (0)], generated by (E, A) and [., ·]e, 0 < e < I, is
equivalent to the fractional power scale of order m, generated by (E, A), that is,

...:.. { (dom(Aa),IIAn·ll) o:soo<oo,


En - (E, IIAa.II)~ -m:S a < 0 .

It possesses the reiteration property

-m :S a < {3 < 00, 0< 7] < 1.

Moreover, (Aa)a- i3 E Lis (Eco E i3 ) for -m :S a < {3 < 00.

Proof Fix a, (3 E [-m, (0). By extending the scale, if necessary, we can ass-
sumc that -m < a < (3 < m. Put F := E_ m and B := A_ m. Then B E BIP(F)
by Lemma 1.3.7, and E i - m ~ D(Bi) for i E N by Theorem 1.3.2. Hence we infer
from (1.2.9.8) that

i E N n [0, 2m] . (1.5.6)

Moreover, if -m :S j :S "( < j + 1 :S m with j E Z, we see from E-y = [Ej, E j+1]-y_j,


from (1.5.6), and the reiteration property (1.2.8.4) that

E-y = [[E-m' Em] (j+m)/(2m) , [E-m' E m ](j+1+m)/(2m)]-y_j


~ [E-m, E mky+m)/(2m)

Thus, again by reiteration,

[Ea, E i3 ]1J ~ [[E-m, Em] (a+m)/(2m), [E-m, Em](i3+m)/(2m)]1J


~ [E-m, Em] [(1-1J)a+1Ji3+mJ/(2m) ~ E(1-1J)a+1Ji3 .

Note that

En ~ [E-m' E m](n+m)/(2m) ~ [F, D(B2m)] (n+m)/(2m)


~ D(Ba+m) = D((A_m)n+m)

by (1.2.9.9) and thanks to the fact that A-m E BIP(E_ m ) by Lemma 1.3.7. Now
the assertions follow from Theorem 1.3.8 and Corollary 1.3.9 .•

The following technical proposition extends Lemma 1.3.7 and is of indepen-


dent interest.
V.1 Banach Scales 279

1.5.5 Proposition Let A E P(E) and let [(Ea,Ao,) ; a E [-m,oo)] be the


interpolation-extrapolation scale generated by (E,A) and (., ')0, 0 < 8 < 1, where
each (., ')0 is exact.
(i) If A E P(E; K, iJ) then Aa E P(Ea; K, iJ).
(ii) If A E BIP(E;M,a) then Aa E BIP(Ea;M,a).
Proof This follows from Lemma 1.3.7 by interpolation.•

Of course, if we drop the assumption that the interpolation functors be exact,


Proposition 1.5.5 remains valid if we replace in the conclusions K and M by c(a)K
and c(a)M, respectively, for a E lR\Z, where c(a) is a suitable constant determined
by the interpolation functor.
Suppose that E is reflexive and A E P(E). Then the two-sided interpolation-
extrapolation scale [(Ea, Aa) ; a E lR] generated by (E, A) and (., ')0, 0 < 8 < 1,
is well-defined. Fix "/ E lR\Z and put (Fj,Bj):= (Ej+'Y,Aj+'Y) for j E Z. Since
Bo E P(Fo) by Proposition 1.5.5, it is obvious that [(Fj , B j ) ; j E Z] is the dis-
crete power scale generated by (F, B) := (E'Y' AI' ). (Note that in this case we do
not need to restrict ourselves to extrapolation scales of finite order since the
spaces F_ j are already given for each j E N and since F_ j is obviously a com-
pletion of (F, IIB-j '11) for j EN.) It should be observed that this scale need not
be reflexive.
Determine for each 8 E (0,1) an admissible interpolation functor {-,. }e. Then
the interpolation-extrapolation scale [(Fa,Ba) ; a E lR], generated by (F,B) and
{" '}o, 0 < 8 < 1, is well-defined.

1.5.6 Remark In general, (Fa, Ba) f= (Ea+'Y' Aa+'Y)' as will be seen in the
next chapter. Suppose, however, that (" ')0 = {-, '}e = [', ']0 for 0 < 8 < 1 and let
A E BIP(E). Then it is a consequence of Theorem 1.5.4 and Lemma 1.5.5 that
(Fa, Ba) = (Ea+'Y' Ac"+'Y) for a E lR, except for equivalent norms .•

Now we give a sufficient condition for the scale [(Fa, Ba) ; a E lR] to possess
the reiteration property. To allow for a simple formulation, we put (', .)g,q := (., ')o,q
for 0 < 8 < 1 and 1:::; q < 00.

1.5.7 Theorem Suppose that E is reflexive, A E P(E), and [(Ea,Aa) ; a E lR]


is the interpolation-extrapolation scale generated by (E, A) and (', ')0, 0 < 8 < l.
Fix "/ E lR\Z and denote by [(Fa,Ba); a E lR] the interpolation-extrapolation
scale generated by (F,B):= (E'Y' AI') and {·,'}e, 0 < 8 < 1. Suppose there ex-
ists q E [1,00] such that {-, -}o = (., .)g,q for 0 < 8 < 1 and (', ')'Y-j = (', ')~_j,q,
where j E Z satisfies j < "/ < j + 1. Then
-00 < a < (J < 00, 0 < T) <1.
280 V Scales of Banach Spaces

Proof Fix k, e E IE such that k :s: 0: < {3 < e. Then it follows from (1.5.2) (applied
to the F -scale) that

(1.5.7)

for j E IE and k < j < e. If k :s: j < ~ < j + 1 :s: e then Fr; = {Fj, Fj+1 h-j by def-
inition. Thus we obtain from (1.5.7) and the reiteration property (I.2.8.2) that

~ E (k,e)\IE. (1.5.8)

Now we deduce from (1.5.7) and (1.5.8), thanks to (I.2.5.2) and (I.2.8.2), that

0<17<1. (1.5.9)

Hence (1.5.8) and (1.5.9) guarantee that {Fa, F,a}ry ~ F(l-1)o+1),a for 17 E (0,1)
with (1 - 17)0: + 17{3 rt. IE.
Thus suppose that m := (1 -17)0: + 17{3 E IE. There are i E IE and (J" E (0,1)
such that 'Y = i + (J". Hence Fj = EH-y = E HHa = {EHJ' Ei+Hda for j E IE, since
{-, ·}a = (., ·)a by assumption. So we deduce from (1.5.2) by reiteration that
(1.5.10)

for n E IE n [k, e]. By means of this and by using (I.2.5.2) and (I.2.8.2) we find that

{Fk' Fe}(m-k)/(e-k)
~ { {Ei+k' EiH+da/(Hl-k), {Ei+k' EiH+d(e-k+a)/(Hl-k)} (m-k)/(e-k) (1.5.11)
~ {EHk' EHf+d(m-k+a)/(Hl-k) ~ Fm ,

which, together with (1.5.9), proves the theorem. _

1.5.8 Corollary Suppose that ~ E lR with 'Y + ~ rt. IE. Also suppose (.,.)1) = {-, .}ry
for 17:= 'Y + ~ - b + ~]. Then Fr; ~ E-y+t..

Proof Fix k, e E IE with k < ~ < e. Then (1.5.8) and (1.5.10) imply

Fr; ~ {EHko E iH + 1 }cr;-k+a)/(f+l-k) (1.5.12)

by the same arguments that led to (1.5.11). Note that

Hence, by replacing in (1.5.8) the spaces Fk and Fe by Ei+k and EiH+1' respec-
tively, and using (., .)1) = {., .}1) we find that

{EHk' EiH+1}cr;-k+a)/(f+l-k) ~ Ei+t.+a = E-y+t. .


Now the assertion follows from (1.5.12). _
V.1 Banach Scales 281

The above results allow a considerable strengthening of the almost reiteration


property of Theorem 1.5.3, provided we restrict the interpolation functors suitably.

1.5.9 Theorem Suppose that E is reflexive, A E P(E), and q E [1,00]' and


e
put (', ')0 := (" ')~,q for 0 < < 1. Let [(Ea, Aa) ; a E lR] be the interpolation-
extrapolation scale generated by (E, A) and (', ')0, 0 < < 1. Pute
if a ElR\Z,
(1.5.13)
ifaEZ.

Then (E~, E~)"., ~ E(l-'l)a+'l(3 for -00 < a < (3 < 00 and 0 < 1) < 1.
Proof This follows from Theorem 1.5.7 and Corollary 1.5.8 by letting I := 1/2.•

Of course, we put
if a ElR\Z,
Ae := { Aa (1.5.14)
a E~-realization of A a - 1/ 2 ifaEZ.

Then (E~, A~) = (Fa - 1 / 2, B a- 1 / 2) for a E JR, where [(Fa, Ba) ; a E lR] is the
interpolation-extrapolation scale generated by (F, B) := (El/2' A 1/ 2) and (', ')0,
e
o < < 1. The following theorem shows that this scale is, in fact, a fractional
power scale, provided we restrict the admitted interpolation functors slightly.

1.5.10 Theorem Suppose that E is reflexive, A E P(E, '!9) for some '!9 E [0,7r),
that 1 :s; q < 00, and (', ')0 = (', ')e,q for 1 :s; q < 00. Let [(En' An) ; a E lR] be the
interpolation-extrapolation scale generated by (E, A) and (', ')0, 0 < e < 1, and
define (E~, A~) for a E lR by (1.5.13) and (1.5.14). Then A~ E BIP(E~; 7r - '!9).
Moreover, [( E~, A~) ; a E lR] is equivalent to the fractional power scale generated
by (Eo,Ao), and (A~)a-(3 E £is(E~,E~) for -00 < a < (3 < 00.

Proof Theorem III.4.7.5 guarantees that A 1 / 2 E BIP(E1 / 2; 7r - '!9). Thus, let-


ting (F,B):= (E1/ 2,A 1/ 2) and denoting by [(Fa,B o,) ; a E lR] the interpolation-
extrapolation scale generated by (F, B) and (', ')0, 0 < e < 1, it follows from
Proposition 1.5.5 that Ba E BIP(Fa ;7r - '!9) for a E R Consequently, using the
fact that (Fa, Ba) ~ (E~+l/2' A~+1/2)' we get A~ E BIP(E~; 7r - '!9) for a E R
Let [(G a , e a ) ; a E lR] be the fractional power scale generated by (F, B).
Then Theorem 1.5.4 implies that it is equivalent to the interpolation-extrapolation
e
scale generated by (F,B) and [', ']e, 0 < < 1. Let a E lR be given and let k E Z
satisfy k - 1 < a < k + 1. Then, since G j = F j for j E Z, we deduce from the
reiteration properties of Theorems 1.5.4 and 1.5.7 and from (1.2.8.6)

G n ~ [Gk- 1 , Gk+l](a-k+l)/2 = [Fk-l, Fk+l](a-k+l)/2


~ [(Fk-2, Fk+2h/4, (Fk-2, Fk+2h/4] (a-k+l)/2
~ (Fk-2, Fk+2)(a-k+2)/4 ~ }'a .
282 V Scales of Banach Spaces

This proves that the interpolation-extrapolation scale [(Fa, Ba) ; a E lR] is equiv-
alent to the fractional power scale generated by (E~/2' A~/2)' Now the assertion
follows from Proposition l.2.6 and Corollary l.3.9 .•

1.5.11 Remark The a'lsumption that E be reflexive has only been used to
be able to deal with two-sided scales which simplifies the presentation. It is clear
that everything proven above remains valid - with obvious adaptations - if we
do not presuppose the reflexivity of E but consider extrapolation scales of finite
order only .•

Now we suppose that

E is a reflexive Banach space and


(-, ')0 E {[" ']0,(-' ')o,q, (', ·)~,oc ; 1 < q < oo}, 0 < 0 < l. } (l.5.15)

Then we define the dual interpolation functor (', .)~ by

[', ']0 if (., ')0 = [', ']0 ,


(-, .)~ := { (', ')0,1 if (-, ')0 = (" ·)~,oo ,
(', . )o,q' if (', ')0 = (', ')o,q, 1 < q < 00 .

Note that (', .)~ is an admissible interpolation functor for 0 < 0 < l. Thus the dual
interpolation-extrapolation scale

that is, the interpolation-extrapolation scale generated by (E~, A~) and (-,.)~,
o < 0 < 1, is a well-defined densely injected Banach scale. The next theorem ex-
tends Theorem 1.4.9 to interpolation-extrapolation-scales.

1.5.12 Theorem Let condition (1.5.15) be satisfied and suppose that A E C(E)
is densely defined with 0 E p(A). Then (Ea)' ~ E~a and (Aa)' ~ A~a for a E lR
with respect to the duality pairing naturally induced by the E' -E-duality pair-
ing (', .). If

(', ')0 E {[" ']0' (', ')o,q ; 1 < q < oo} , 0<0<1,
then Ea is reflexive for a E lR.

Proof Thanks to Theorem 1.4.9 it remains to consider the case a E lR\Z. Thus
j< a < j + 1 for some j E Z. Then it follows from (1.2.6.1)-(1.2.6.3), (1.2.5.4), and
Theorem l.4.9 that

(Ea)' = (Ej, EHl)~_j ~ (Ej, Ej+1)~_j = (E~j' E~j_l)~_j


= (E~.
-J-1' E~. )~
-J 1-a+J
. = E~- a .
V.I Banach Scales 283

Suppose that x E Ea+1 '----+ Ej+1 and xU E E~j+1 '----+ E~a+1' Then, using Aj ~ A"
and A~", ~ A~ j together with Theorem 1.4.9, it follows that

IXU , A
\ 0:
x) = Ix U A·x)
\, J = IAUxU
\ - ) ' x) = IAU
\ xU x)
-0: , •

x UE E U
-n+1'

Hence (An)' ~ A~Q' Now 0 E p((AoJ') = p(A",) and A~Q E £is(E~a+1,E~a)' to-
gether with Lemma 1.4.2, guarantee that (Aa)' = A~n'
By reflexivity, (EU)U =E and (AU)U =A. If (',')0 rt. {(-,')O,l'("')~,oo} for
0< e < 1 then (C, .)~)U = (-,.)0. Hence [(Ea,Aa) ; Q E JR.] is the dual scale of
the scale [(E~, AU ; Q E JR.] generated by (EU, AU) and (-, .)~, 0 < e < 1. Now, by
what has already been shown,

Q E JR..
This implies the asserted reflexivity .•

1.5.13 Corollary Suppose that A E BIP(E) and let [(Ea,A,) ; Q E JR.] be the
interpolation-extrapolation scale generated by (E,A) and [', ']0' 0 < e < 1. Then
[(Ea, An) ; Q E JR.] is equivalent to the fractional power scale generated by (E, A),
and its d1wl scale [(E~, A~) ; Q E JR.] is equivalent to the fractional power scale
generated by (E', A').

Proof This is a consequence of Theorem 1.5.4, Lemma 1.4.11, and the fore-
gone theorem .•

Let n E N and suppose that E 1 , ... , En and F are LeSs over the same
field. Then

is the vector space of all continuous n-linear maps from E1 x ... x En to F, and

C(E, F) := £(E, ... ,E; F) . (1.5.16)


'-v--'
n

Note that £l(E, F) = £(E, F). If E 1 , ... , En and Fare normed vector spaces,

(1.5.17)

defines a norm on £(E 1 , ... , En; F), and the latter space is a Banach space if
F is one.
284 V Scales of Banach Spaces

Let b E £(EI,E2;K) be given. Then there exist unique

(1.5.18)

satisfying

(1.5.19)

the form operators induced by b. Note that

(1.5.20)

if EI and E2 are normed vector spaces.

The following observation is almost trivial. However, in concrete situations


it plays an important rOle since it allows to identify extrapolated operators A-a
for certain Q > 0 explicitly.

1.5.14 Proposition Let condition (1.5.15) be satisfied. Suppose that 0 :::; Q :::; 1
and that bE £(ELa' Eo:; K), and let B j , j = 1,2, be the form operators induced
by b. If
b(x~,x) = (xU, Ax) , (x~ , x) E E~ X EI , (1.5.21 )
then BI = A~a and B2 = A a - I .

Proof From (1.5.18), (1.5.19), and Theorem 1.5.12 we infer BI E £(ELo:' E~o:)
and B2 E £(Ea, E a - I ), and (1.5.21) implies B2 ::> A. Since A a - I E £(Ea, Ea-I)
and A a - I ::> A, we obtain A a - I = B2 by the density of EI in Ea. A similar argu-
ment shows that A~a = B I . •

Lastly, we consider the case that E is a Hilbert space and A is self-adjoint


and positive definite, that is, A = A* ;::: a for some a > O.

1.5.15 Theorem Let E := (E, ('1')) be a Hilbert space and let A = A * ;::: a
for some a> O. Then the fractional power scale [(Ea , Aa) ; Q E JR.] generated
by (E, A) is a Hilbert scale. More precisely, Ea = (Ea, (·I·)a), where

x, Y E E(a+I)VI ,

and Ao: = (Aa)* ;::: a for Q E R It is equivalent to the interpolation-extrapolation


scale generated by (E, A) and [".J 0' 0 < B < 1, which, in turn, is equivalent to
the interpolation-extrapolation scale generated by (E, A) and (', ')0,2, 0 < B < l.

Proof Thanks to Theorem III.4.6. 7 and Lemma 104.11, the first assertion is easily
verified. The second one follows from Theorem 1.5.4 and Example III.4.7.3(a). The
last assertion is a consequence of (1.2.9.11) .•
V.I Banach Scales 285

1.5.16 Remarks (a) Let E := (E, (·I·)E) be a Hilbert space. The Riesz duality
mapping '13 := '13 E is the conjugate linear map from E to E' defined by

('I3x,y)E = (yIX)E , x,yE E. (1.5.22)

Thanks to the Riesz representation theorem, '13 is a well-defined bijective isometry.


Defining the dual inner product (·I·)E' by

(X 'I Y') E':= (_0-1


'U
, 'U
y, _0-1 X ') E, x',y' E E' , (1.5.23)

it follows that (E', (·1·) E') is a Hilbert space as well.


Let (F, (,1,) F) be a second Hilbert space and let
B E C( E, F) be densely de-
fined. Then the dual B' E C(F', E') and the adjoint B* E C(F, E) are well-defined,
and it is easily verified that the diagram

B*
F :::> dom(B*) -----+- E

~F 1~ ~F 1~ ~ 1#E (1.5.24)

B'
F' :::> dom(B') E'
is commutative, where, in abuse of notation, we use ~ to denote conjugate-linear
bijections. Thus
(1.5.25)

Given x E dom(B) and y' E dom(B'), it follows from (1.5.22) and (1.5.23) that

(y', BX)F = (B'y',X)E =:= (xW;;/B'y')E = (B'y'I'I3EX)E'


= (y'I(B')*'I3 Ex)F' = ('I3F1(B')*'I3EXI'I3FlY')F
= (Y','I3 F1 (B')*'I3 Ex) .
Consequently, by the analogue to (1.5.24),

(1.5.26)

Now suppose that E = F and that B = B*. Then we infer from (1.5.25) and
(1.5.26) that B' E C(E') is self-adjoint also. Moreover, if B = B* 2: b for some
b E JR, it follows from (1.5.25) and (1.5.23) that

(B'x'lx')E' = ('I3B'I3- 1 x'lx')E' = (B'I3- 1x'I'I3- 1x')E 2: b 11'I3-1X'II~ = b Ilx'II~,

for x' E dom(B'), which shows that B' 2: b.


286 V Scales of Banach Spaces

(b) Let (X, J-t) be a u-finite positive measure space and let E := L2(X, J-t; OCN ) for
some N E N. Then

(u,v) = i(u,V)dJ-t and (ulv) = (u,v) , u,v E E,

where
N

(~,TJ):= 'LeTJ j ,
c ._ (C1
<".- <" , •.. , <"CN)._
,TJ·- ( TJ 1 , ... , TJ N) E uN
1& •

j=l

In this case the duality mapping is given by complex conjugation, that is, 1'Ju = U
for u E E.
(c) Let E:= (E, (-I,)) be a Hilbert space and let A E C(E) be positive definite
and self-adjoint. Then it follows from (a) and Theorems 1.5.12 and 1.5.15 that the
scale [(EtA~); QElRj, dual to the fractional power scale [(Ec"A",); QElRj
generated by (E, A), is a Hilbert scale. _

Most of the theorems of this subsection are generalizations and refinements


of earlier results of the author (cf. [Ama86], [Ama87], [Ama88b], and [Ama88cD.
For Theorem 1.5.4 we also refer to [Tri78, Section 1.5.3]. Theorem 1.5.7 is due to
Simonett [Sim92]. Theorem 1.5.10 is new. In [DaPG84] Da Prato and Grisvard
introduced the extrapolation spaces (E-1' Eo)~,oo' 0 < () < 1.

2 Evolution Equations in Banach Scales


Having established a powerful theory of interpolation-extrapolation scales, we now
study the behavior of semigroups and evolution equations in these scales. We show
that, roughly and imprecisely speaking, we can 'shift' a given evolution equation
'along a scale' to obtain a continuum of related problems. Of particular importance
is the fact that we can develop a good duality theory. This leads to interpretations
of solutions to 'extrapolated' Cauchy problems as naturally defined weak solutions.
We also investigate the case where the pair (E, A), generating the scale, consists
of an ordered Banach space and a resolvent positive operator of positive type. In
the last subsection, for completeness we briefly discuss some implications of our
general results to evolution equations that are not of parabolic type and lie, in
principle, outside the range of our interest.

2.1 Semigroups in Interpolation-Extrapolation Scales


Interpolation-extrapolation scales are most useful for getting precise information
on regularity and smoothing properties of linear operators. In this subsection
V.2 Evolution Equations in Banach Scales 287

we study, in particular, the behavior of semigroups in interpolation-extrapolation


scales. These investigations are the abstract basis for the 'weak theory' of linear
and quasilinear parabolic systems to be developed in later chapters.
The following simple lemma is of general interest and is given in a form that
is more general than presently needed.

2.1.1 Lemma Let E and F be Prechet spaces and suppose that there exists a
LGS G such that E '----> G and F '----> G. If E is a vector subspace of F then E '----> F.

Proof Let i: E -+ F be the natural injection. Let (x j) be a sequence in E such


that Xj -+ x in E and i(xj) -+ y in F. Since E '----> G and F '----> G, we see that
Xj -+ x in G and Xj = i(xj) -+ y in G. Hence x = y, that is, i(x) = y, so that
the graph of i is closed in E x F. Now the assertion follows from the closed
graph theorem .•

2.1.2 Lemma Let F and G be Banach spaces with 1" '----> G. Suppose that
d
B E g( G) with dom(B) c F, that the semigmup {e- tB ; t 2:: O} leaves F invari-
ant, and that Ile-tBIIL(F) :::; c for 0:::; t :::; 1. Then BF E g(l") and e- tBF = e-tBlF
for t 2:: o.

Proof Fixw E p( -B) and let G l := dom(B), endowed with the norm II(w + B)·II.
Denote by Bl the G1-realization of B. Then (w + B)e- tB =) e-tB(w + B), as fol-
lows from Lemma III.4.9.1. This implies the commutativity of the diagram
e- tB
G G

W+llj'" "'jW+B
Ul(t) := e-tBIG l
Gl Gl

From this we see that {Ul(t) ; t 2:: O} is a strongly continuous semigroup on G l .


Let UF(t) := e- tB IF for t 2:: O. Then {UF(t) ; t 2:: O} is a semigroup on F and
UF(t) :J Ul(t) for t 2:: O. Since G l C 1" and G l '----> G, it follows from Lemma 2.1.1
that G l '----> F. Thus, given x E G l , we see that UF(t)X = Ul(t)x -+ x in G l , hence
in F, as t -+ O. Since IIUF(t)IIL(F) :::; c for 0:::; t :::; 1 and since G l is dense in F by
assumption, it follows that {UF(t) ; t 2:: O} is a strongly continuous semigroup
on F. Let A be the negative infinitesimal generator of this semigroup. Given
,,\ > type( - A) V type( - B), we infer from the semigroup representation of the re-
solvent (d. (II.6.3.3)) that

(,,\ + A)-lX = lCXJ e->-.tUp(t)x dt = 1 00


e->-'te-tB x dt = (,,\ + B)-lx , x EF.

Thus, thanks to Lemma 1.1.1, (,,\ + A)-l = (,,\ + BF )-1, which implies A = BF .•
288 V Scales of Banach Spaces

Now we suppose that

E is a Banach space,
B E C(E) and is densely defined with 0 E p(B), mEN, } (2.1.1)
and (', .)(/, 0 < () < 1, are exact admissible interpolation functors.

We denote by
[(Ea , Ba) ; Q E [-m, (0)] (2.1.2)
the interpolation-extrapolation scale of order m, generated by (E, B) and (', .)(/,
o < () < 1, and by
(2.1.3)
the dual power scale generated by (E", B") := (E', B') (recall Remark 1.4.5(b)).
Given any densely defined A E C(E), we put A" := A' and denote by Aa the
Ea-realization of A for Q E lR, which is well-defined. Moreover, if A is closable
in E_ a for some Q E (0, m], we put

A_a := closure of A in E_ a (2.1.4 )

and call A_a the E_a-realization of A. Note that Aa E C(Ea) , whenever it is


well-defined.
Suppose that n E N and A E C(E) satisfy p(A) "10 and
0:::; k:::;n , kEN. (2.1.5)

Then, given I1Ep(-A) and using the fact that dom((I1+A)k) =dom(Ak) for
kEN (d. [HiP57, Theorem 2.16.4]), it follows from the closed graph theorem that
(11 + A)k B- k and Bk(11 + A)-k belong to C(E) for 0:::; k :::; n, or, equivalently, that

(2.1.6)

for 0 :::; k :::; n. Similarly, if

dom((A")k) = E! ' (2.1. 7)

it follows from p(A) = p(A") that


(11 + A")k E C(E!, E"), (11 + A")-k E C(E", E!) (2.1.8)

for 0 :::; k :::; m. Thus

N(A, 11, m, n) := maximum of the norms of all the operators


occurring in (2.1.6) and (2.1.8) } (2.1.9)

is well-defined if (2.1.5) and (2.1.7) are satisfied.


V.2 Evolution Equations in Banach Scales 289

The following fundamental theorem is the main result of this subsection.


In fact, it is one of the main results in the theory of interpolation-extrapolation
scales - at least for our purposes. It shows that, if A E Q(E) and satisfies (2.1.5)
and (2.1.7) then An exists and belongs to Q(E",,) for -m ::; a::; n, and the semi-
groups {e- tA ", ; t::::: O} are obtained naturally from { e- tA ; t::::: O} by restriction
and extension, respectively (cf. diagram (2.1.12)). Moreover, if -A generates an
analytic semigroup, the same is true for -An. In addition, the theorem contains
useful quantitative information. This theorem is the abstract basis for weak for-
mulations of rather general initial-boundary value problems for a wide variety of
concrete differential equations. It is also the foundation of abstract 'bootstrapping'
arguments leading to rather precise regularity results. This will be clear from the
results developed in the subsequent sections (of this and the following volume).

2.1.3 Theorem Let assumption (2.1.1) be satisfied, let [(Ea , Ba) ; a E [-m, 00)]
be the interpolation-extrapolation scale of order m, generated by (E, B) and (',')0,
0< e < 1, and let ((EJ, BJ) ; j EN] be its dual power scale.
Suppose that A E Q(E) and there exists n E N such that
dom(A k ) = Ek , 0::; k ::; n , (2.1.10)
and
O::;k::;m. (2.1.11)
Then Ac, E Q(Ea) and p(Aa) = p(A) for -m :::; a ::; n. Moreover, the diagrams
e- tA",
Ea

f e- tAp
[ (2.1.12)

E{3 E{3

are commutative for -m ::; (3 ::; a ::; nand t ::::: o.


More precisely, suppose that A E Q(EjM,"-.J) for some M::::: 1 and wE lR, fix
/-l E p( -A), and put N ;= N(A, /-l, m, n). Then

-m ::; a ::; n , (2.1.13)


and, ifn::::: 1,
>. + Aa E Cis(Ea + 1 , Ea) , >. E p( -A), -m::; a ::; n - 1 . (2.1.14)
Moreover,
(2.1.15)
and
11(>' + Aet)-lllc(E",E,,+j) ::; N 2 11(>' + A)-lllc(E,Ej ) (2.1.16)
for>'Ep(-A), j=O,1, and-m::;a::;n-l.
290 V Scales of Banach Spaces

If, in addition, A E H(El' Eo), so that n ~ 1, then Aa E H(Ea+l,Ea) for


-m :s;Ct :s; n - 1. Indeed, if A E H(El' Eo, I\:,wo) for some I\: ~ 1 and Wo > 0 then

-m :s; Ct :s; n - 1 ,

where 1\:0 := N 2 (1 + I\: + I\:N 2 1IAII.c(E "


E))'
0
Given (7 > type( -A),

Ile-tAiJll.c(EiJ,Ea ) :s; ctfJ-ae"t , t > 0, (2.1.17)

for -m :s; (3 :s; Ct :s; n, where c depends on Ct - (3.

Proof (i) Suppose that A E Q(E,M,w) and J1 E p(-A). First we consider the
case B := B(J1) := J1 + A. Then (2.1.10) and (2.1.11) are satisfied for each n E N
and N := N(A, J1, m, n) = 1, thanks to the fact that B = fL + A restricts or ex-
tends, respectively, to an isometric isomorphism from Ek+1 to Ek for k E Z.
Lemma I1I.4.9.1 implies

t~O, jEZ. (2.1.18)

Hence, in particular, El is invariant under { e- tA ; t ~ O} and

t ~ 0, x E El .

Now we deduce from Lemma 2.1.2 that Al E Q(El,A1,w) with e- tA, = e- tA IE l ,


and Lemma 1.1.1 gives p(Ad = p(A). By Lemma I1I.4.9.1 we also see that

for" E p(-A), k E {O, I}, and x EEl, which proves (2.1.16) for Ct = 1.
It is clear that Al and B1 are resolvent commuting. Thus, by applying the
above arguments inductively, we obtain the corresponding assertions for each Ct E N.
Letting j = -1 in (2.1.18), we infer that e- tA extends to an element U-l(t)
of £(E-l)' Since B-1 E Lis(E, E-d, it follows that

t ~ O. (2.1.19)

Hence { U- l (t) ; t ~ O} is a strongly continuous semigroup on E-l satisfying

t ~ O.

Let C be its negative infinitesimal generator. Then we deduce from (2.1.19) that
C = B_ 1 AB- 1 with dom(C) = E. Lemma III.4.9.1 implies C :J A. Hence A is
closable in E-1 and C:J A-I. Since e- tA, C e- tA C U_ 1(t), we infer that El is
invariant under {U_ 1(t) ; t ~ O}. Thus, thanks to the density of El in E_ l , the
core theorem of semigroup theory guarantees that El is a core for C, that is, Cis
the closure of its restriction to El (e.g., [Dav80, Theorem 1.9]). Consequently,
V.2 Evolution Equations in Banach Scales 291

C = A-I. This proves that A-I is well-defined, that A-I E Q(E-l' M, w), and
e- tA C e- tA - 1 • Lemma 1.1.1 implies p(A-d = p(A), and (2.1.16) for a = -1 fol-
lows easily from Lemma III.4.9.1 and the definition of E_ l . It is also clear that
A-l and B-1 are resolvent commuting. Hence we can extend the results already
proven to each a E Z n [-m, 00 ).
Now let j < a < j + 1 for some j E Z n [--m, 00). Then we deduce by inter-
polation from what is already proven that Eoo is invariant under {e- tAj ; t:::: O}
and that
t:::: 0 .
Hence Lemma 2.1.2 guarantees that Coo := (Aj )B", E Q(Ea , M, w) and that e- tCa is
the restriction of e- tAj to Ea for t :::: O. If j EN, this gives Coo = Aa. If j < 0, we
infer that El is invariant under {e- tCa ; t:::: O} and e- tC", lEI = e- tA, for t :::: O.
Thus, similarly as above, the core theorem guarantees that Aoo is well-defined and
that Aoo = Ca. The remaining assertions of the theorem, pertaining to the case
A E Q(E,M,w), except estimates (2.1.15), are now obvious.
Note that Lemma III.4.9.1 implies AB-j ~ B-j A for j EN. Hence, given
j E N and x E E Hl ,

IIAjxllj = IIAB-j Bjxllj = IIB-j ABjxllj = IIABjxl1


~ IIAII.c(B"Bo) IIB j xll1 == IIAllqB"Bo) IlxllHl
If j E N n [0, m],

IIA_jxll-j = IIB-j Axil = IIAB-jxll ~ IIAII.cCB"Bo) II X lll-j


for x EEl. Thus by density we obtain the validity of the last estimate for each
x E E l - j . Now interpolation proves (2.1.15).
Suppose that A E 'H(E l , Eo, K, wo). Then (I.l.2.3) and (2.1.16) entail

II(A+Aa)-lll.c(Ba,Ba+,) ~ II(A+A)-lll.c(Bo,B,) ~ K

for Re). :::: Wo and a E [-m, 00). Thus A, E 'H(Ea+l' E u , KO, wo) follows from Re-
mark I.1.2.1(a) and (2.1.15).
Fix a > type(-A). Corollary I.l.4.3 guarantees the existence of 19 E (0,w/2)
and Mo :::: 1 such that a + ~'!? C p( -A) and

Iia + AII.c(B"Bo) + (1 + I).I?-k II(). + a + A)-lll.c(Bo,Bk) ~ Mo


for)' E ~'!? and k = 0, 1. Hence p(Aa) = p(A) and (2.1.15) and (2.1.16), applied to
a + A and /1 - a instead of A and /1, imply

a+ ~'!? C p(-Aoo) (2.1.20)

and
292 V Scales of Banach Spaces

for A E Et9, k = 0, 1, and -m ::; a < 00. From this and Lemma II.4.2.1 we infer,
in particular, that, given fEN,

(2.1.22)

for j E Z n [-m, 00) and t ~ O. Since (2.1.16) implies

II(a + Aj)-lILC(Ej,Ew) ::; (lI(a + A)-lIIC(E,Edr~ ,


estimate (2.1.17) follows from (2.1.22), provided a,f3 E Z n [-m, 00).
We denote by [(Ea,q, Ba,q) ; a E [-m,oo)] the interpolation-extrapolation
scales generated by (E, B) and (', ')~,q' 0 < () < 1, where 1::; q::; 00. Then we
deduce by interpolation from what has just been proven that

t a - j Ile-tAj IIC(Ej,E""q) + tl+ j - a lIe-tAj IIC(E""q,Ej+l) + lIe-tAj IIC(E""q)


(2.1.23)
::; ceut

for t > 0, j E Z n [-m,oo), cr E (j,j + 1), and 1::; q ::; 00. Hence, by interpolating
again and using the reiteration property (1.2.8.2), we see that

t ~ 0, (2.1.24)

for j E Z n [-m, 00) and j < 13 < a < j + 1. Now we infer from (2.1.23), (2.1.24),
and (1.2.11.4) the validity of (2.1.17), provided j ::; 13 ::; a ::; j + 1 for some j E Z
with j ~ -m.
Lastly, suppose that -m ::; 13 < a < 00 and (cr, 13) n Z i= 0. Then there are
integers j and k such that -m ::; j ::; 13 < j + 1 ::; k ::; cr < k + 1. Letting 8 := t/3,
the semigroup property implies

for t > 0, and (2.1.17) follows.


(ii) Now we consider the general case. We denote by

the interpolation-extrapolation scale of order m, generated by (E, B(J1)) and (., ')9,
0< () < 1. It follows from (2.1.10) that dom(Bk(J1)) = Ek for 0::; k::; n, where
B k(J1) := [B(J1)( Hence Remark 1.1.2(f) implies Ek == E k(J1) and

IIB kB- k(J1)II.ctE) IIxll k ::; IlxIIEk(JL) ::; IIB k(J1)B- k llc(E) Ilxll k

for x E Ek and 0 ::; k ::; n. Similarly, dom ( [(B (J1) )'] k) = dom ([ B'l k) for 0 ::; k ::; m
follows from (2.1.11). Consequently, by Proposition 1.3.5, E-k == E-k(J1) and
V.2 Evolution Equations in Banach Scales 293

for x E E-k and 0:::; k:::; m, where B'(p,):= (B(p,))'. Since Bk [resp. (B')k] is an
isometric isomorphism from Ek onto E [resp. E! onto E~], we see that

and

for 0 :::; k :::; n, and

and
II(B')j (B'(p,))-jIlL(Ef) = 11(p, + AU)-jIIL(E~,E;)
for 0 :::; j :::; m. Thus it follows that Ek ='= Ek(p,) and

for -m :::; k :::; n. From this we obtain by interpolation that


-m:::;a:::;n, (2.1.25)
and
x E Eo:, -m:::; a :::; n . (2.1.26)
Now the assertion is an obvious consequence of (2.1.25), (2.1.26), and part (i) of
this proof. •

2.1.4 Corollary Suppose that A E Q(E) and B = p, + A for some p, E p( -A).


Then (2.1.10) is true for every n E Nand (2.1.11) is also satisfied. Hence the asser-
tions of Theorem 2.1.3 hold for -m :::; f3 :::; a < 00. Moreover, N = 1 in this case.

2.1.5 Remarks (a) Let A E Q(E). Given p, E p(-A), denote by

[(Eo: (p,), Bo:(p,)) ; a E [-m, (0)]


the interpolation-extrapolation scale generated by p, + A and (., ·)0, 0 < () < 1.
Then
p"V E p(-A) .

More precisely,
a;:::: 0,
and

IlxIIE_,,(v) :::; (1 + Ip, - Vi II (v + A)-lIIL(E))" IlxIIE_,,(JL) ,


for p"V E p(-A).
294 V Scales of Banach Spaces

Proof This follows from Remark 1.2.2(d) and Corollary 1.3.6 by interpolation .•
(b) It is obvious that Theorem 2.1.3 remains valid if we drop the assumption
that the interpolation functors be exact, provided N is replaced by c(a)N, where
c: lR --> rR + is a I-periodic function satisfying c( k) = 1 for k E Z and depending
on the interpolation functors (., ·)0, 0 < e < 1, only .•

Theorem 2.1.3 gives extensions and sharpenings of earlier results due to the
author (cf. [Ama86], [Ama87]).

2.2 Parabolic Evolution Equations in Banach Scales


Let E:= (E, 11·11) be a Banach space, mEN, and (-,·)0 exact admissible interpo-
lation functors for 0 < e < 1.
Denote by 2t. a subset of geE, M, w) for some M;:::: 1 and wE lR such that
o E p( A) for each A E 2t.. Then, given A E 2t., the interpolation-extrapolation scale
[(E", (A), A) ; a E [-m, (0)] generated by (E, A) and (., ·)0, 0 < e < 1, is well-
defined. It follows from Corollary 2.1.4 that

An E g(E", , M,w) , (2.2.1)

that
(2.2.2)
and

j = 0, 1, (2.2.3)

for A E p( -A) = p( -A.) and -m :::; a < 00, and that the diagrams
e- tA ",
E",(A) E",(A)

r r
(2.2.4)

e- tA {3
Ei3(A) Ei3(A)

are commutative for -m :::; (3 < a < 00 and t ;:::: O. Moreoever, if -A generates an
analytic semigroup, this is true for -A", and, given (J > type( -A),

t > 0, (2.2.5)

for -m :::; (3 < a < 00, where c is a function of a - (3.


Let J be a perfect subinterval of lR+ containing 0 and suppose that

A: J-->2t..
V.2 Evolution Equations in Banach Scales 295

Put

EoJt) := Ea(A(t)) , t E J , Ea := E",(O) , -m::; a < 00 ,


and denote the norm of Ea by 11·lla' In general, the Banach spaces Ea(s) and E",(t)
are distinct, even as vector spaces, if a =1= 0 and s =1= t. Suppose, however, that
there exists a E (0,1) such that Ea-j(t) ~ E a - j for j E {O, I} and t E J. Then
the Cauchy problem

u + A(t)u = f(t) , tE j , 'u(O) = x (2.2.6)

in E and the extrapolated Cauchy problem (of order a-I)

u + Aa-1(t)U = f(t) , t E j , u(O) = x (2.2.7)a-1

in E a - 1 are well-defined, provided, of course, f: J - t E and x E E in case (2.2.6)


and f: J - t E",-l and x E E OI - 1 in case (2.2.7)01-1. Whereas the 'original' Cauchy
problem (2.2.6) can be highly irregular, due to the fact that the domains E1 (t)
of A(t) vary with t E J, the operators A a - 1(t) of the extrapolated Cauchy prob-
lem (2.2.7)01-1 have constant domain, namely Ea. Thus we can apply the general
results developed in Chapters II and III to the extrapolated problem (2.2.7)a-1
and use the results on variable domain problems exposed in Chapter IV to get
solvability and regularity results for the original Cauchy problem (2.2.6). This is
done in the following fundamental existence and regularity theorem, where we use
the notations introduced in Subsection IV.2.1. Also recall that (-, ')~,q := (', ')o,q if
o < B < 1 and 1 ::; q < 00.
2.2.1 Theorem Suppose A: J - t Q(M(E) for some M > 0, 0 < a < f3 < 'Y < 1,
and
Er;(t) ~ Er; , ~ E {a, a-I, f3 - 1, 'Y - I}, t E J .

Also suppose that there exists q E [1,00] such that

(', ')(3 = (', ')~,q, that f3 - a < p < 1 , (2.2.8)

and that Aa-1 E CP(J,£(Eo. , E o. - 1)) and

f E C(J, E.y-d . (2.2.9)

Then, given x E E a - 1, the extrapolated parabolic Cauchy problem

u + Aa-1(t)U = f(t) , t E j , u(O) = x (2.2.10)0.-1

possesses a unique solution u(-, x) and

(2.2.11)
296 V Scales of Banach Spaces

It is also a solution of the extrapolated evolution equation

v + A/3-1(t)V = J(t), t E j , (2.2.12)/3-1

in E/3-1. If there exists a constant /'i, 2: 1 such that

x E Ea, t E J , (2.2.13)

if P > 1 - a, and if
(2.2.14)

for some (j E (0,1) then u(·, x) E C(J, E) n C 1(j, E) and it is a solution of the
'original' evolution equation

u+ A(t)u = J(t) , t E j ,

in E, which is strict if x E E 1.

Proof Put (F1,Fo):= (Ea, E a - 1) and B:= A a - 1. Since A(t) E 1i(E1(t), E), it
follows from B E CP(J, L(F1' Fo)) and Theorem 2.1.3 that

Fix Jl and 1/ such that f3 - a < Jl < 1/ < 'Y - a and Jl < p. Put {., ·}e := (., .)~,q,
0< B < 1, and let Fe := {Fo, F 1}e. Then, by the almost reiteration Theorem 1.5.3,

(2.2.15)

Consequently, J E C(J, F,,) by (2.2.9). Since x E Fo and p> Jl, Theorem II.1.2.2
guarantees that the extrapolated Cauchy problem (2.2.1O)a-1 possesses a unique
solution u(·, x) and

Hence the validity of (2.2.11) follows from (2.2.15). By Theorem II.1.2.2 we also
know that u(·, x) is a solution of the parabolic equation

v + BJl(t)v = J(t) , (2.2.16)

in FJl , where BJl is the FJl-realization of B. Given t E J, let [( Fdt), B«(t)) ; (E jR+]
be the interpolation-extrapolation scale of order zero generated by (F, B (t)) and
the interpolation functors {., ·}e, 0 < B < 1. Then we infer from (2.2.8), Corol-
lary 1.5.8, and Remark 1.5.11 (letting ~ := f3 - a + 1 since 'Y of Corollary 1.5.8
corresponds to a-I in our situation) that F/3-a+1(t) ~ E/3(t) for t E J. Conse-
quently, B/3-a(t) = A/3-1(t) for t E J, and since BJl(t) C B/3-a(t), we deduce from
(2.2.16) that u(·,x) is a solution of (2.2.12)/3-1.
V.2 Evolution Equations in Banach Scales 297

Now let assumptions (2.2.13) through (2.2.14) be satisfied. Since it suffices


to prove the corresponding assertion for each compact subinterval of J, we can
assume that J is compact. As the inversion map

is analytic, our assumptions imply that (A,_lt l E BUCP (J, £(Ea.-I, Ea.)). Thus,
using A a - I :) A and E'-+ Ea.-I,

where p > 1 - O!. Now the last part of the assertion follows from Theorem IV.2.5.1,
the fact that Aa.-I :) A, and by uniqueness .•

2.3 Duality
In this subsection we derive important and useful results about semigroups in
reflexive Banach spaces for which the duality Theorem 1.5.12 is valid. In particular,
we show that the dual operators -A~ generate strongly continuous semigroups as
well that can be obtained naturally by dualizing. We restrict the class of admissible
Banach spaces and interpolation functors by assuming that

E is a reflexive Banach space and


(., ·)e E {[., ·]e, (-, ·)e,q, (., ·)£,00 ; 1 < q < oo}, 0 < f) < 1. } (2.3.1 )

We also suppose that

A E Q(E,M,w) for some M 2:: 1 and wEIR. (2.3.2)

We fix p, > type( -A) and put B := p, + A, so that B E P(E). Then we denote by

the interpolation-extrapolation scale generated by (E, B) and (., ·)e, 0 < () < 1,
and by

its dual scale. It follows from Theorem 1.4.12 that these scales are well-defined. For
each O! E IR we denote by Aa. the Ea.-realization of A and by A~ the E~-rea1iztion
of AU = A'.

In the following theorem we collect some of the basic properties of the op-
erators A~. In particular, we show that -A~, generates a strongly continuous
semigroup on E~.
298 V Scales of Banach Spaces

2.3.1 Theorem Let (2.3.1) and (2.3.2) be satisfied. ThenA~ E 9(EtM,w) and
p(A~) = p(A~) = p(A) for a E lR. The diagrams

r
E~
e ~tA~(3
r
(3

are commutative for -00 < (3 :s: a < 00 and t ;:::: 0, and
A+A;E£is(E;+l,E~), AEp(A) , aEJR.

If A E H(El' Eo, /'i"wo) for some /'i, ;:::: 1 andwo ;:::: 0 then A~ E H(E;+1' E~, /'i,o, wo),
where /'i,o := 1 + /'i, + /'i, IIA~IIL(E~,E~r

Proof Recall from semigroup theory that A~ E g(E~,M,w), thanks to (2.3.1)


and (2.3.2). From [HiP57, Theorem 2.16.4] we infer that dom(Ak) = Ek as well as
dom( (A~)k) = E! for kEN. The reflexivity of E implies (A~)~ = A and (B~)~ = B.
Thus

for kEN. This shows that the assumptions of Theorem 2.1.3 are satisfied for
A~ E g(E~, M, w). Now that theorem implies the assertions. _

In the following theorem we prove important duality results. For this we recall
from Theorem 1.5.12 that
a E JR, (2.3.3)
with respect to the duality pairings naturally induced by the E'-E-duality pair-
ing (', . ). We use these duality pairings without further mention.

2.3.2 Theorem Let assumptions (2.3.1) and (2.3.2) be satisfied. Then


A; = (A~QJ , a E JR ) (2.3.4)
as unbounded operators, and
e~tA~ = (e~tA-Q)' , t;::::O, aEJR, (2.3.5)

that is, {e~tA~ ; t;:::: O} is the dual semigroup to {e~tA-Q ; t;:::: O}. Furthermore,

(2.3.6)

for a E JR, that is, A~ E £(E!+1) E~) is the dual of the bounded linear operator
A~a~l E £(E~a,E~a~l)'
V.2 Evolution Equations in Banach Scales 299

Proof Given a: E JR and .\ E lK, the definition of the Ea-realization implies

x E E(aH)Vl .

Thus we infer from Remark l.l.2(b) and the density of E(aH)Vl in EaH that

(2.3.7)

Similarly,
(.\ + A~)a = .\ + A~ , a:EJR, .\ElK. (2.3.8)
From this and (2.3.3) we deduce that

for a: E JR, which proves (2.3.4). Assertion (2.3.b) is now a well-known consequence
of (2.3.4).
From (2.3.4), dom(A~) = E!+l, and dom(A_a) = E l- a it follows that

(2.3.9)

Since A_ a- l ::) A-a and A_ a- l E £(E-a, E-o<-l), we obtain (2.3.6) by continuity


from (2.3.9), the density of El-o< in E_o<, the fact that A~x~ E E~ for x~ E E!+l,
and from the first part of (2.3.3) .•

The following technical proposition gives useful estimates for the norms of
the operators A~ and (.\ + A~) -1 .

2.3.3 Proposition Let assumptions (2.3.1) and (2.3.2) be satisfied. Then

IIA~ 11.c(E!+l'E~) ::; IIAII.c(E1,E) /\ IIArt 11.c(Ef ,E~)


and

II (.\ + A~)-lll.c EtE!+I) ::; II (.\ + A)-lll.c(E,E I ) /\ II (.\ + Art)-lll.c(Eu,Ef)

for.\ E p(-A) and a: E R.

Proof Thanks to the proof of Theorem 2.3.1 we obtain from Theorem 2.l.3 that

(2.3.10)

and
11(.\ + A~)-lll.c(E~,E!+I) ::; 11(.\ + A~)-lll.c(E~,Ef) (2.3.11)

for.\ E p(-A) and a: E R.


300 V Scales of Banach Spaces

From (2.3.3) and (2.3.6) it follows that

IIA~ILc(E~0+1 E~)
J 0:
= IIA-a-lIlC(E_o,E_o_Il ' a E lR.

Thus we infer from (2.1.15) that

IIA~IIC(E!+1'E~) :::; IIAllc(E1,Eo) ,

which, together with (2.3.10), proves the first assertion.


Recall from Theorems 2.1.3 and 2.3.1 that). + A_ a - 1 E Lis(E_ a , E_ a - 1 )
and), + A~ E Lis (E!+1 ,Ei) for)' E p( -A). Hence we deduce from (2.3.6) that

This implies

Now the second assertion is a consequence of (2.1.16) and (2.3.11) .•

2.4 Approximation Theorems


At this stage we make a digression and prove some useful approximation theorems
that are of general interest.
Let X be a nonempty set, V a vector space, and <I>(X) a vector subspace
of OC X . We define a bilinear map

<I>(X)xV-+VX, (rp,v) ....... rp0v

by rp 0 v(x) := rp(x)v. Then rp 0 v is the tensor product of rp and v and

<I>(X) 0 V := span{ rp 0 v ; rp E <I>(X), v E V} ,

the tensor product of <I>(X) and V, is a well-defined linear subspace of V X .


(Recall that in the proof of Theorem III.4.3.6 we have already used the tensor
product Lr(X, It) 0 E.)

Throughout the remainder of this subsection X denotes a nonempty open


subset of lR n and E := (E, 1·1) a Banach space.

2.4.1 Proposition V(X) 0 E is sequentially dense in V(X, E).

Proof Let u E V(X, E) and an open set K with supp(u) cc K cc X be given.


The family {lJll( e, c) ; e E E} is, for each c > 0, an open covering of the compact
set u(X) =u(K) in E. Thus there are finite subcoverings {lJll(ej,c) ; l:::;j:::; m(e:) }.
V.2 Evolution Equations in Banach Scales 301

FiXe> oand put Uo:= supp(u)C and [,j := u-l(lE(ej,e)) nK for 1-:5. j -:5. m, where
m := m(e). Then {[,j ; 0 -:5. j -:5. m} is an open covering of X. Choose a smooth
°
partition of unity {'l/Jj ; 0 -:5. j -:5. m} on X subordinate to {[,j ; -:5. j -:5. m } and,
letting eo := 0, put v := 2:;~0 'l/Jj Q9 ej. Then v E, D(X) Q9 E with supp(v) c K and
m

lu(x) - v(x)1 -:5. L 'l/Jj(x) lu(x) - ejl -:5. e , xEX.


j=o
This shows that we can find a sequence (Vk) in D(X) 18: E such that SUPP(Vk) cK
and Vk --+ U in BUc(~n, E).
Let {'Pc; e > o} be a mollifier. Then 'PI/j * Vk E D(X) Q9 E provided j 2: jo,
where jo is independent of k, thanks to the fact that the support of 'PI/j * Vk is
contained in the (l/j)-neighborhood of K. Now we infer from (III.4.2.1O) and
(III.4.2.19) that 'PI/j * Vk --+ 'PI/j * U in £(~n, E) as k --+ 00. By reapplying the
last argument and using (III.4.2.23) we find that 'PI/j * U --+ U in £(~n, E). Since
the supports of 'PI/j * v are contained in a fixed compact subset of X, standard
properties of LF-spaces guarantee that the convergence of the above sequences
takes place in D(X, E). _

2.4.2 Corollary Suppose that (Eo, E I ) is a densely injected Banach couple.


d
Then D(X, E I ) '---+ D(X, Eo). In fact, D(X, E 1 ) is sequentially dense in D(X, Eo).
Proof The first assertion is trivial. It is obvious that D(X) Q9 EI is sequentially
dense in D(X) Q9 Eo with respect to the topology induced by D(X, Eo). Hence the
sequential density of D(X) Q9 Eo in D(X, Eo) implies the assertion. _

Next we turn to Sobolev spaces and prove the following important approxi-
mation result.

2.4.3 Theorem Suppose that (Eo, E I ) is a densely injected Banach couple. Then
d .
D(~n, EI ) '---+ Wpm(~n, Eo) for 1 -:5. p < 00 and rn E N.

Proof Thanks to Corollary 2.4.2 we can assume that Eo = E I . Then, letting


D := D(~n, Eo) and Vfj,m := Vfj,m(~n, Eo), it follows from
Ilullm,p -:5. c(K)Pm,K(U) , u ED, supp(u) c K cc ~n ,
that D '---+ W;n.
To prove the asserted density, let U E Vfj,m be given. Let 'I/J E Coo (~n, [0, 1])
satisfy 'l/JllEn = 1 and supp( 'I/J) C 2lBln. Recall that aa denotes dilation and put
'l/Jj := aI/j'I/J for j E N. Then

aa('l/JjU) = L (~)rl{3lal/j(a{3'I/J)aa-{3u , lal-:5.m,


{3~a
302 V Scales of Banach Spaces

by Leibniz' rule. Observe that II(}ljj(aj3~)lloo :s; c and SUPp((}ljj(aj3~)) C 2jIffi n for
j E Nand 1;31 :s; m. Hence ~jU E Wpm and

118"'(~ju - u)llp :s; 21Iullm,p,(jlffin)C + crl Ilulim,p


for j E Nand lal :s; m. Thus, thanks to Lebesgue's theorem, ~jU ----+ U in Wpm as
j ----+ 00. Hence we can assume that u has compact support. Let {'Pe ; E > O} be
a mollifier. Then 'Po * u E V and aD('Pe * u) = 'Pe * a""u for lal :s; m by (III.4.2.13)
and (III.4.2.10), respectively. Thus (III.4.2.23) implies 'Pc * u ----+ u in Wpm as E ----+ 0,
which proves the assertion .•

Given u E V' (lR" , E), define its restriction, TXU, to X by


'P E V(X) .
Then TXU E V'(X, E), and TX := (u f---+ TXU), the restriction operator for X, is
linear and continuous:
rx E .c (V' (lRn, E), V'(X, E)) . (2.4.1)
Observe that rx is independent of E.
It is obvious that rx commutes with differentiation,

aENn , (2.4.2)
and that
Txu=uIX, (2.4.3)
where ulX is the usual point-wise restriction. It is also clear that
(2.4.4)
with norm at most one.
It follows from (2.4.2)-(2.4.4) that
TX E .c(Wpm(lR", E), Wpm(x,E)) , (2.4.5)
and that r x has norm at most one.

The set X is said to possess the extension property if there exists a map

an extension operator for X, such that


l:S;p<oo, mEN,
and
rxex = id ,
independently of E.
V.2 Evolution Equations in Banach Scales 303

2.4.4 Remark Let X have the extension property. Then rx is a retraction


from Wpm (lRn, E) onto Wpm (X, E) for mEN and 1 ~p < 00, and ex is a corre-
sponding coretraction. _

Following Stein [Ste70a, Section VI. 3] , X is said to be a special Lipschitz


domain if it is obtained by rotation from a set of the form

{(y,t) c lRn - 1 x lR; t > 'P(y)} ,


where 'P E BUC 1-(lR n - 1 ). Then 11'Pllcl- is the bound of X. Moreover, X has a
minimally smooth boundary if there exist E > 0, N EN, M::::: 0, and a sequence
Uo, U 1 , ... of open subsets of lR n such that:
(i) if x E aX then lffi(x, E) C Uj for some j;
(ii) no point of lR n is contained in more than N of the f1js;
(iii) for each j there exists a special Lipschitz domain X j ' whose bound does not
exceed M, such that f1j n X = Uj n X j .
Of course, lR n and each half-space in lR n have minimally smooth boundaries. More
interestingly, X has a minimally smooth boundary if X is an n-dimensional C 1-_
submanifold of lR n with compact boundary. For further examples of open sets with
minimally smooth boundaries we refer to [Fra79].

2.4.5 Theorem If X has a minimally smooth boundary, it has the exten-


sion property.

Proof See [Ste70a, Section VI.3] (also cf. [EdE87, Section V.4.4l) and observe
that these proofs extend trivially to the E-valued case. _

Now it is easy to prove the following useful approximation theorem for Sobolev
spaces on X.

2.4.6 Theorem Let (Eo, E 1) be a densely injected Banach couple and let X have
a minimally smooth boundary. Then

is dense in ~m(x, Eo) for 1 ~ P < 00 and mEN.

Proof Given U E Wpm (X, Eo), we deduce from Theorems 2.4.5 and 2.4.3 the
existence of a sequence Uj E D(lR n ,E1) converging in ~m(lRn,Eo) towards exu.
Hence, thanks to (2.4.5), we see that rXuj converges in Wpm (X, Eo) towards u. _

As a first application of the above approximation results we prove a useful


product rule. For this we recall from Proposition 1.4.8 that, given a densely in-
jected Banach couple (Eo,E 1) such that E1 is reflexive if ED of. E 1, (E~,Eb) is a
304 V Scales of Banach Spaces

densely injected Banach couple as well. Moreover, the Eh-Eo-duality pairing deter-
mines the Ei -E1-duality pairing so that we can denote either one by (., .) without
fearing confusion.

2.4.7 Proposition Suppose that X has a minimally smooth boundary. Let


(Eo, E 1) be a densely injected Banach couple such that E1 is reflexive if Eo -I- E 1 .
Also suppose that p E [1,00],

and
v E Lpl (X, E~) n Wp~ (X, E~) .
Then (v,u) E Wl(X) and

j = 1, ... ,n. (2.4.6)

Proof First note that (2.4.6) holds point-wise if u E V(X, E 1) and v E V(X, E£).
Next, HOlder's inequality implies that, given any Banach space F,

[(w',w) f--> (w',w)] E £(Lpl(X,F'),Lp(X,F);L1(X)) . (2.4.7)

Using this, with F equal to Eo or E1 at the appropriate places, the assertion


follows from Theorem 2.4.6, provided 1 < p < 00, thanks to the fact that (Eo, E 1)
and (Ei, Eb) are densely injected Banach couples. If p = 1, we see from (2.4.7)
that (v, u) and the right-hand side of (2.4.6) belong to L1 (X). Let rp E V(X)
be given and fix an open set K such that supp(rp) cc K cc X. Suppose that
u E V(X, E1)' Then (writing again u for rKu etc.), u E L 2 (K, Ed n Wi(K, Eo)
and v E L 2 (K, Eb) n wi
(K, ED. Hence

by what has just been shown. Thus (u, v) E Wl(X) and (2.4.6) is true, provided
u E V(X, E 1) and v E Loo(X, E£) n W~(X, ED. By applying again Theorem 2.4.6
and (2.4.7), we see that the assertion is true if p = 1. Finally, if p = 00, we obtain
the stated result by interchanging the roles of u and v in the last step of the proof. •

2.5 Final Value Problems


In this short subsection we collect some (almost trivial) results about final value
problems. These problems come up naturally in connection with weak solutions
of evolution equations in the next subsection. They are also of importance in the
theory of optimal control for linear and quasilinear parabolic systems.
V.2 Evolution Equations in Banach Scales 305

We fix T E jR+ and put J := [0, T]. We let E and F be Banach spaces with
F'--+ E, assume that

B: J --+ C(E) with dom(B(t)) c F , t E J , (2.5.1)

and that
(x, g) E E x L1,loc(J, E) .
Then we consider the final value problem

- iJ + B(t)v = g(t) , o : : : t < T, v(T) = x . (2.5.2)

Of course, a solution v of (2.5.2) is a function v Eo C 1 ([0, T), E) n C(J, E) such that


v(t) E dom(B(t)) for t E [O,T) and (2.5.2) is point-wise satisfied. Thus a solution
of (2.5.2) is defined hy an ohvious modification of the definition of a solution to
an initial value problem. Clearly, a solution v of (2.5.2) is strict if v E C 1 (J, E)
and v(T) E dom(B(T)). Similarly, the concepts of a Wi,loc-solution and a VY;1-
solution, 1 ::::: p ::::: 00, of (2.5.2) are clear.
With (2.5.2) we associate the initial value problem

it + A(t)u = f(t) , o < t :s T, u(O) =x , (2.5.3)

where A: J --+ C(E) and f E L1,loc(J, E) are defined by

A(t) := B(T - t), f(t):= g(T - t) , t E J, (2.5.4)

respectively.

2.5.1 Lemma Put

u(t) := v(T - t) , 0::::: t::::: T . (2.5.5)

Then v is a [strict] solution (resp. a VY;~loc-solution for some p E [1,00]) of the


final value problem (2.5.2) iff u is a [strict] solution (resp. a Wp~loc -solution) of
the initial value problem (2.5.3).

Proof Obvious .•

A function V: J~ --+ £(E) is said to be a parabolic evolution operator for


the final value problem (2.5.2) (with regularity subspace F) if U: J~ --+ £(E),
defined by
U(t,s) :=V(T-s,T-t) , (t, s) E h , (2.5.6)

is a parabolic evolution operator for A, defined by (2.5.4) (with regularity sub-


space F). Using (2.5.6), it is trivial to deduce from properties (II.2.1.2)-(II.2.1.9)
306 V Scales of Banach Spaces

the properties of V. In particular,

V E C(h,.cs(E)) nC(Ji-".c(E,F)) (2.5.7)

and

V(t, t) = 1, V(t, s) = V(r, s)V(t, r) , (2.5.8)

Moreover,
V(t,·) E C1([0,t),.c(E)) , t E (0, T] , (2.5.9)
and
8 2 V(t,s) = B(s)V(t,s) , O~s<t~T. (2.5.10)
Note that, thanks to Remark I1.2.1.2(c), there exists at most one parabolic evolu-
tion operator V for the final value problem (2.5.2).

2.5.2 Proposition Suppose that V is a parabolic evolution operator for the final
value problem (2.5.2). If v is a solution of (2.5.2),

v(t) = V(T, t)x + iT V(r, t)g(r) dr , O~t~T. (2.5.11)

If 9 = 0 then v := V(T, ·)x is a solution of (2.5.2).

Proof This follows by means of the transformations (2.5.5) and (2.5.6) from the
corresponding results for initial value problems .•

Of course, formula (2.5.11) is meaningful whenever 9 E L1,loc(J, E). Hence


the function v defined by (2.5.11) is said to be a mild solution of the final value
problem (2.5.2). Furthermore, (2.5.11) is the variation-of-constants formula for
the final value problem (2.5.2).

2.5.3 Theorem Suppose that (Eo, E 1) is a densely injected Banach couple and

for some p E (0, 1). Then there exists a unique parabolic evolution operator V
for the final value problem (2.5.2), and it possesses El as regularity subspace. If
f E CP(J, Eo) then (2.5.2) has a unique solution, which is strict if x EEl.

Proof Thanks to Lemma 2.5.1, the assertions are obvious consequences of Corol-
lary 11.4.4.2 and Theorem 11.1.2.1, respectively .•

We leave it to the reader to formulate further existence theorems for (2.5.2)


based upon Theorem 11.1.2.2 or the existence results of Chapter III.
V.2 Evolution Equations in Banach Scales 307

2.6 Weak Solutions and Duality


In this subsection we generalize the concept of solutions of Cauchy problems.
These generalizations are motivated by the notions of weak solutions in the theory
of partial differential equations. Thus they are of particular importance in appli-
cations of the abstract theory to linear and quasilinear parabolic systems given in
later chapters.

Assume that

E is a reflexive Banach space and


(', ')0 E {[" ']0' (', ')o,q, (', ·)~,oo ; 1 < q < (Xl}, 0 < () < 1. } (2.6.1)

Denote by J a perfect subinterval of 1R+ containing 0, and suppose that

A: J --+ geE) and n


tEJ
p(A(t)) =J 0. (2.6.2)

Fix J.L E ntEJP(-A(t)), and denote by

t E J ,

the interpolation-extrapolation scales generated by (E, J.L + A(t)) and the interpo-
lation functors (', ')0, 0 < () < 1, where the latter are independent of t E J. Let

t E J ,

be the respective dual scales. Put

a E 1R. (2.6.3)

Moreover, let Ac,,(t) and A~(t) be the E",(t)- and E~(t)-realizations of A and AU,
respectively, for a E IR and t E J. Lastly, assume that

there exists (3 E [0, 1] such that


E, == E,(t), t E J, 'Y E {{3, {3 - I}. } (2.6.4)

Note that Theorem 1.5.12 implies

t E J, 'Y E {I - {3, -{3} . (2.6.5)

In the following proposition we establish a simple but important 'generalized


Green's formula'.
308 V Scales of Banach Spaces

2.6.1 Proposition Let assumptions (2.6.1), (2.6.2), and (2.6.4) be satisfied and
suppose that
(2.6.6)
Then, given p E [1,00]'

(2.6.7)

and
(2.6.8)
it follows that

fors,t E J with s < t.

Proof From (2.3.6) we know that

(xij,x) E EL,e x E(3, T E J. (2.6.9)

Thus (2.6.6)-(2.6.8) and Holder's inequality imply

(2.6.10)

Now the assertion is an easy consequence of Proposition 2.4.7, Theorem 1.5.12,


and Theorem III.1.2.2 .•

The generalized Green's formula suggests the following definition of 'weak


solutions' of linear Cauchy problems. Let the hypotheses of Proposition 2.6.1 be
satisfied. Suppose that

P E [1,00], (x, J) E E(3-1 x Lp,loc(J, E(3-1) . (2.6.11)

Then u is said to be a weak Lp,loc(J, E(:J)-solution of the Cauchy problem

it + A(t)u = f(t) , t E j , u(O) = x, (2.6.12)

for each uij E Lp,(J, EL(3) n Wp~(J, E~(3) having compact support in J\ {sup(J)}.
Of course, if u is a weak Lp,loc(J, E(3)-solution of (2.6.12) and u E Lp(J, E(3) then
u is a weak Lp(J,E(:J)-solution.
V.2 Evolution Equations in Banach Scales 309

For the next proposition we recall that Mj,~loc -solutions have been defined in
Subsection III.1.3.

2.6.2 Proposition Let assumptions (2.6.1), (2.6.2), and (2.6.4) be satisfied and
suppose that AI'~l and (x, f) satisfy (2.6.6) and (2.6.11), respectively. If u is a
Wp~loc -solution of the extrapolated Cauchy problem

u+AI'~l(t)u=f(t), tEJ, u(O) =x (2.6.13),a~1

then u is a weak Lp,loc(J, EI')-solution of (2.6.12).

Proof This is an immediate consequence of the above definition and Proposi-


tion 2.6.1. •

Let T E j be fixed and supppose that

(x~,f~) E E~,a x Ll.1oc((O,T),E~,a) . (2.6.14)

Then the generalized Green's formula suggests the study of the dual final value
problem of (2.6.12) defined by

(2.6.15f

With (2.6.15f we also associate the extrapolated dual final value problem:

(2.6.16)~,a

It is now easy to give a general sufficient condition for weak Lp,loc (J, EI')-
solutions of (2.6.12) to be unique.

2.6.3 Proposition Let assumptions (2.6.1), (2.6.2), and (2.6.4) be satisfied


and suppose that AI'~l and (x, f) satisfy (2.6.6) and (2.6.11), respectively. Then
o
problem (2.6.12) has at most one weak Lp,loc(J, EI')-solution if, given T E J and
f~ E V(O, T) @ ELI" the extrapolated dual final value problem (2.6.16)~,a with
x" = 0 has a Wp~ -solution.

Proof Suppose that u E Lp,loc(J, E,a) satisfies

1(( -8 + A~I')v", u) dT =0 (2.6.17)

for each
(2.6.18)
having compact support in J\ { sup( J) }. Then we have to show that 1L = O. Given
T E J and f" E V(O, T) @ ELI" let u" be a Mj,~-solution of (2.6.16)~,a for x" = o.
310 V Scales of Banach Spaces

Define v~ by extending u~ by zero over J n (T, 00). Then it follows from Theo-
rem III.1.2.2 that v~ satisfies (2.6.18) and has compact support in J\ {sup(J)}.
Thus, thanks to (2.6.17),

r (f~, u) dT = 0 ,
J.J
f~ E V(O, T) Q9 ELf3 . (2.6.19)

Hence, using ELf3 ~ E~f3'

x ~ E E~1-13'

Since ELf3 is dense in E~f3' the above relation is valid for each x« E E~f3' which

1
shows that
<pudt = 0,

From this we easily infer, by invoking (III.4.2.27), for example, that u = O.•

So far we have required neither restrictive regularity assumptions nor a


parabolicity hypothesis. If we impose conditions of this type, we obtain the fol-
lowing uniqueness and existence result for weak Lp,loc(J, E(3)-solutions of (2.6.12)
and solutions of the extrapolated Cauchy problem (2.6.13)13-1, respectively.

2.6.4 Theorem Let assumptions (2.6.1), (2.6.2), and (2.6.4) be satisfied. Suppose
that -A(t) generates an analytir: semigroup on E for t E J, and

(2.6.20)

for some p E (0,1). Then, given p E [1,00] and

(x,1) E Ef3-1 x Lp,loc(J, E f3 - 1) ,

the Cauchy problem (2.6.12) has at most one weak Lp,loc(J, E(3)-solution. If there
exists c E (0,1) such that

(2.6.21 )

the extrapolated Cauchy problem (2.6.13)13-1 has a unique strict solution, and
(2.6.12) has a unique weak Lp,loc(J, E(3)-solution.

Proof Since A(t) E 1t(E1 (t), E) for t E J by assumption, Theorem 2.3.1 implies

A~f3(t) E 1t(EL f3 , E~(3) , tE J . (2.6.22)

From (2.3.6) we deduce that

IIA~f3(s) - A~f3(t)II£(E~~i3,E~i3) = IIA f3 - 1(s) - Af3-1(t)II£(E!'I,E!'I~1) , s,t E J.


V.2 Evolution Equations in Banach Scales 311

Consequently, thanks to (2.6.20) and (2.6.22),

(2.6.23)

Thus, given T E J and f~ E V(O, T) Q9 EL(3 C CP(J,E~(3)' it follows from Theo-


rem 2.5.3 that the final value problem (2.6.16)~(3 has a solution on [0, TJ, hence a
Wp~-solution. Now the first assertion is a consequence of Proposition 2.6.3. From
A(t) E 7-i(E1(t), E), Corollary 2.1.4, and (2.6.20) we infer that

(2.6.24)

Put (Fa, Fd := (E(3-1, E(3) and Fe := (Fa, Fde for 0 < () < 1. Fix (J" E (0, c). The
almost reiteration property of Theorem 1.5.3 implies E(3-1+E '-+ Fa. Hence (2.6.21)
gives
f E C(J, Fa) + ca(J, Fa) .
Choose a decomposition f = fa + h with fa E Ca(J,Fa) and h E C(J,Fa ) and,
letting B := A(3-1, consider the Cauchy problems

v(O) = Xj, j = 0, 1 ,
where Xa := x and Xl := o. It follows from Theorems 11.1.2.1 and 11.1.2.2, re-
spectively, that these problems have unique strict solutions 7Lo and 7L1, respec-
tively. Hence u := 7Lo + U1 is a strict solution of the extrapolated Cauchy prob-
lem (2.6.13)(3-1; thus a ~~loc-solution of (2.6.13)(3-1. Consequently, thanks to
Proposition 2.6.2, it is a weak Lp,loc(J, E(3)-solution of (2.6.12). Therefore it is
unique by the first part of the assertion. _

We close this subsection by identifying the dual of the evolution opera-


tor U(3-1, where the latter is considered as an element of £(E(3, E(3-1). This iden-
tification is particularly useful in problems of control theory.

2.6.5 Proposition Let assumptions (2.6.1), (2.6.2), and (2.6.4) be satisfied.


Suppose that -A(t) generates an analytic semigroup on E for t E J, and

Let U(3-1 be the parabolic evolution operator of A(3-1 and, given T E J, let V!(3 be
the evolution operator of the extrapolated dual final value problem (2.6.16)~(3. Then

(X~,X) E EL(3 x E(3, 0::::; s::::; t::::; T ,

that is, the dual operator in £(EL(3' E~(3) of U(t, s) E £(E(3, E(3-1) equals V(t, s)
for 0 ::::; s ::::; t ::::; T.
312 V Scales of Banach Spaces

Proof Let s, t E J with s < t be fixed. Since the assumptions of Theorem 2.6.4
are satisfied, (2.6.23) and (2.6.24) are true. Hence Theorem 2.5.3 and Corol-
lary II.4.4.2 guarantee the existence and uniqueness of the evolution operators V~;3
of (2.6.16)~;3 and Uf3 - 1 of A;3-1, respectively. Given (xrt, x) E EL;3 x E;3, put
U(T) := U;3-1(T, s)x and Urt(T) := V~;3(t, T)Xrt for s ~ T ~ t. Then Theorem 11.1.2.1,
Remark II.2.1.2(a), and Theorem 2.5.3 imply

(urt,u) E C([s, t], EL;3 x E;3) nCl([s,t],E~;3 x E;3-1)


and
on [s, t] .
Thus u and u rt satisfy (2.6.7) and (2.6.8), respectively, on the interval [s, t], with
u(s) = x and u(t) = x~. Hence it follows from the generalized Green's formula of
Proposition 2.6.1 that

<X~,U;3-1(t,S)X) = (u~,u)(t) = (u~,u)(s) = <V~;3(t,s)x~,x)


This proves the assertion. _

2.7 Positivity
Let E:= (E, P) be an OES and suppose that A is of positive type, that is,
A E P(E). Fix mEN and, for each 8 E (0, I), an admissible interpolation func-
tor (.,.)0. Denote by [(Ec"Aa); aE [-m,oo)] the interpolation-extrapolation
scale generated by (E, A) and (., ·)e, 0 < 8 < 1. Lastly, assume that A is re-
solvent positive.
Given a E [-m, (0), put

(P) ,
i~l a 2: 0,
Pa := { (2.7.1)
clE" (P) , a < 0,
where io: : Eo: '---7 E is the canonical injection. Then Po: is a closed convex cone
in E(l:' which is proper if a 2: O. Thus Pa induces a preorder in E a , the natural
preorder. Henceforth we put

-m ~ a < 00,

that is, we endow each EQ; with its natural preorder, which we simply denote by ~
without fearing confusion.

2.7.1 Proposition Suppose that A E P(E) and is resolvent positive. Then


Ea is an OES for a 2: 0 with respect to the natural order induced by Po:. Given
d
-m ~ (3 ~ a < 00, the injection Eo: '---7 E;3 is positive, AQ; is resolvent positive, and
POI is dense in Pf3.
V.2 Evolution Equations in Banach Scales 313

Proof The first three assertions are trivial. Observe that -m :::; (3 :::; a < 00 im-
d d
plies Pc; c Pf3. Thus Pc; C Pf3 if a :::; 0 since, by definition, P = Po C Pf3. Hence it
remains to show that Pc; is dense in Pf3 if (32:0. Given (3 2: 0 and x E Pf3, note
that (1 + EA(3)-lX E Pf3+1. Since A is of positive type, Proposition 1.5.5 and the
remarks following it imply Af3 E P(Ef3). Hence it follows from Lemma 11.6.1.1 that
d
(1 + EA(3)-lX -+ x in Ef3 as E -+ O. This shows that PfJ+l C PfJ for (32:0. Thus
d
Pf3 +k C Pf3 for (3 2: 0 and kEN. Consequently, given a 2: (3 2: 0, it follows from
• d
Pf3 + k C Pc; C PfJ, where kEN satisfies (3 + k > a, that Pc; C Pf3 .•

Now we assume that

E is a reflexive Banach space ordered by a total cone P, and


(', ')e E {[" 'le, (', ')e,q, (" .)~ ; 1 < q < oo}, 0 < e < 1. } (2.7.2)

In addition, we retain the assumption that

A E P(E) and is resolvent positive. (2.7.3)

Then the interpolation-extrapolation scale [( E r" Au) ; a E lR1' generated by (E, A)


and Co')e, 0 < e < 1, and its dual scale [( E~, A~) ; a E lR1 are well-defined. We
denote by
P~ the dual cone of E~.

Then E~ = (E~,P~) is an OBS with the so-defined natural (dual) order. Thus,
by replacing P in (2.7.1) by P~, the cones P~ of E~ are well-defined and induce
the natural preorder of E~ for a E R Also note that A~ E P(E") and

which shows that A" is resolvent positive as well. Hence it is a consequence of


Proposition 2.7.1 that P~ is dense in P~ if -00 < (3 < a < 00.

The next duality theorem, the main result of this subsection, shows that Pg is
the dual cone of P-c; for a E R

2.7.2 Theorem Let assumptions (2.7.2) and (2.7.3) be satisfied and let Ec;
and E~ be given their natural preorders, induced by the positive cones Pc; and pt
respectively. Then Eo: and E~ are OBSs for a 2: 0, the injections Eo: '-+ E(3 and
E~ '-+ E~ are positive for -00 < (3 < a < 00, the operators An and A~ are resol-
vent positive, and Pc; [resp. P~l is dense in P(3 [resp. P~l for -00 < (3 < a < 00.
Moreover,
ooElR, (2.7.4)
314 V Scales of Banach Spaces

that is, given a E IR and xU E E~,


x~EP1 iff (xU,x):::::O,

Proof Only (2.7.4) remains to be proven. Thus assume first that a> O. Given
x~ E pl = pU nE~, it follows that (x~,x)::::: 0 for x E P. Hence the density of P
in P- a and Theorem 1.5.12 imply (xU,x) ::::: 0 for x E P- a , that is,
(2.7.5)

Now let a < O. Given xU E Pl, there exists a sequence (x~) in P~ converging in E~
towards x~. Hence (x~,x) --+ (x~,x) as j --+ 00 for x E P- a C P, and (x~,x) ::::: 0
implies that (2.7.5) is true in this case as well.
To prove the inclusion reverse to (2.7.5), let xU E E~ satisfy (x~,x) ::::: 0 for
x E P- a . If a > 0, it follows that (x~, x) ::::: 0 for x E P so that xU E pU. Conse-
quently, xU E pl = P~ n Ea;, which shows that (P- a )' C pl in this case. Thus
let a < 0 and fix kEN with a + k > O. Then, thanks to Theorem 1.5.12 and
Remark 1.1.2(f),
((1 + cA~)-kx~,x) = (x~, (1 + cA_a)-kx ) ::::: 0, x E P- a , (2.7.6)

due to the resolvent positivity of A-a. Since (1 + cA~)-kx~ E E!+k C E~, we de-
duce from (2.7.6) and the density of P- a in P that (1 + cA~)-kxU E P~. Finally,
we infer from Lemma II.6.1.1 that (1 + cA~)-kxU --+ xU in E~. This proves that
(P- a )' C pl in this case as