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 ParisDauphine, France
Associate Editors:
H. Araki, Kyoto University
J. Ball, HeriotWatt 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, MaxPlanckInstitut Bonn
Herbert Amann
Linear
and Quasilinear
PamboHc Problems
Volume I
Abstract Linear Theory
A CIP catalogue record for this book is available from the Library of Congress,
Washington D.C., USA
This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, reuse of illustrations, 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, CH4010 Basel, Switzerland
Printed on acidfree paper produced of chlorinefree 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
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
preparation. For this reason our treatise is divided in three volumes, carrying the
respective titles:
Preface . . . . vii
Introduction xiii
3 Maximal Regularity
3.1 Abstract Initial Boundary Value Problems 242
3.2 Isomorphism Theorems . . . . . . . . . . . 245
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.
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)+(18)u=0 onoX, t > O. (5)
Lastly, in order to determine the timeevolution 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
initialboundary value problem:
OtU + Au = r(·, t, u)
Bu=O
in X
on ax } t > 0,
(7)
u(',O) = u o onX.
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 conormal 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 initialboundary 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,
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 functionalanalytical 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
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)
Thus, if we impose the condition p > n12, it follows from El <t C <t Eo that the
minimal requirement (15) is met.
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 halfplane [Re A ::: AO 1 for some
AO E lR and that there exists a constant K such that
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
(23)
and u satisfies (13) pointwise, 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 variationofconstants 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
(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
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 righthand side of (24)
is welldefined 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.
(28)
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
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)
(31)
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,
we infer from (20) that f satisfies the modified minimal requirement (15)" in this
case as well.
(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
(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.
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 nvectors dik , respectively, and the Nvectors 4
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) ,
(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
reactiondiffusion system (36)~(39) is rewritten as a quasilinear initialboundary
value problem:
satisfying (42) in the obvious 'pointwise' 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 variationofconstants formula
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 righthand 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 vdependence of A and f, re
spectively, for the moment. This Cauchy problem corresponds to a Banach space
formulation of the nonautonomous linear initialboundary 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 tdependent, 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
(51)
and
and that
(53)
Note that this implies the existence of a boundary decomposition ax = aox u a1x
such that
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
ofconstants formula
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:
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 {etA(s) ; t 2:: O}, 8 E J, restricts to a strongly continuous
semigroup on Ea and (32) is replaced by
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 welldefined 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.
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
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
Then, thanks to the divergence structure of (40) and (41), the natural weak for
mulation of the initialboundary value problem (42) takes the form:
1{ (v, u) + a(t, u)( v, u)} dt = 1 (v, F(t, u)) dt + (v(O), uo) (64)
It is easily seen that every solution of the initialboundary value problem (42) is a
weak solution in the sense of (64). Furthermore, (64) is equivalent to the abstract
quasilinear Cauchy problem in p, WJ:
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)
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 Flrealization
xxx Introduction
1<,6<0<1.
Furthermore, denoting the Fa_Irealization of B1 by B a  1 for 0 E [0,1], it follows
that
B a  1 E H(Fa , Fat) ,
These extensions are 'natural' in the sense that etB"'l is the restriction to F a  1
of e tBj3  1 for 0 :S f3 < 0 :S l.
This 'interpolationextrapolation' 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, interpolationextrapolation 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 interpolationextrapolation 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 interpolationextrapolation 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
Aal (t, u) E H(E", , E al) , 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 interpolationextrapolation technique outlined above enables us to solve the
initialboundary 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
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 Al/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
is wellposed in E,6l for some {3 E (a,l], its solution coincides with un. Note
that a solution u of (72) satisfies u(t) E dom(ABl(t,Uo(t))) for t E jo. Hence it
follows that
uo(t) E dom(A,6l(t,UO(t))) C Wp2,62 (73)
for t E jo, provided (72) is wellposed.
iJ + A,6l(t)V = F(t) , t E j ,
in E,6l, where we are particularly interested in the case 2{3 > 1 + l!p, so that
dom(A,6l(t)) is no longer constant. Since Aal(t) :J A,6l(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,6l '+ E a  1 is an evolution operator for { A iJ  l (t) ; t E J} on E,6l.
This is true in fact, provided
p>{3a, (74)
(75)
if (J > 20:  1. From (71) we infer, using suitable embedding theorems, that
provided 2p < 20:  (J  nip. By combining (75) and (76) we see that
(77)
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 initialboundary value problem (42).
This problem is also studied in detail in the second volume by making use of the
interpolationextrapolation 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 reactiondiffusion 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 infinitedimensional 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,
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:
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
interpolationextrapolation 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.
By now the reader surely feels that our approach to the quasi linear initialboundary
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
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,
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
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
AcCB,
_ 0
if A is compact and contained in R.
!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
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 the set of all topological linear (toplinear) isomorphisms from X onto Y and
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
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
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
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
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 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
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
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
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
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
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
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,
and
T E K(X, Y) ¢=> Tc E K(Xc, Yc) .
Moreover, (3.3) easily implies
and it follows from (3.1) that the duality pairing (., ·)xc : X~ x Xc ~ C is given by
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
3.1 Remark Let (F, 11·11 F) be a complex Banach space of complexvalued func
tions and let (X, 11·11) be the real subspace of IRvalued 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
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.
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 Xrealization, Ax, (the part of A in X, or the maximal restriction of A
to X) is the linear operator in X, defined by
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.
BEC(X,L(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 := B1(xj)f, it follows that
which shows that (ej) is a Cauchy sequence in E. Thus ej > e for some e E E and
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
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) .•
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
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
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
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) : : : (IAIIIAII) 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:
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 wellknown 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~)
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
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
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).
1.3.1 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/(IKr),wV!3/r) .
Proof (i) Given B E £(EI' Eo) with IIBII :::; r, we see that
x E EI , 0:::; t :::; 1.
Hence
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).
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. _
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
Then
II(A + A)xllo ~ 11(>'0 + A)xllo I>'  >'olllxli o
(1.4.2)
~ (",ll>'olIA  >'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>'011>'  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
11(>' + A)xllo ::; (KW + 1>'1 + w) Ilxll o + K IIxl1 1 ::; 5K(I>.llI x ll o + Ilxlh) .
o w6 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. •
Consequently, we deduce from (1.4.8) that [ReA 2': 0] c p( ((5 + A)) and
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
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 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 timedependent linear evolution
equations. However, a first application of Corollary 1.4.3 will lead to the pertur
bation result of Theorem 1.5.1.
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)
Re).~w, (1.5.2)
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.
where we use obvious matrix notation. For definiteness, in the following theorems
we use the l'Inorm on the product Banach spaces.
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
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
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,
(1.6.3)
Put
B:= [~ A~2] (1.6.4)
1.6.2 Remarks (a) Suppose that All and A22 satisfy the hypotheses of Theo
rem 1.6.1 and replace K by
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,
and, letting ~ := 0,
yEH,
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
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 welldefined
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
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
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
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
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).
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
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),
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 . •
j = 0, 1,
2.4.3 Remark Let (E, II· liE) be a complex Banach space of complexvalued func
tions and let (F, 11·11) be the real subspace of realvalued functions. Then we know
from Remark 3.1 that Fe ~ E but II·IIF" III·IIE, in general. This fact has to be
kept in mind when employing the complex interpolation functor in the real case.•
Let (Eo, E l ) be a densely injected Banach couple. Given 8 E (0, I), put
(', ')~,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
(2.5.1)
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)
(2.5.4)
1 1
+=1.
q q'
(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
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. (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
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
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
Hence
0<7)<1. (2.8.7)
). E lE.+ . (2.9.1)
and
(2.9.3)
Let r be any piecewise smooth simple curve in R<p running from ooei<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 welldefined bounded linear operator in E, thanks to esti
mate (2.9.3).
(2.9.5)
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(10)<>+Oi3) (2.9.8)
for 0 ~ Re a < Re;3, ° < e < 1. More precisely, let
E A := (dom(A<», IIA<>·II) , Rea2:0.
°
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].
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].
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
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 pairwise commuting. Put
E:= nn
j=1
D(Aj) (2.10.1)
and
(2.10.2)
for t > 0 and x E E. If (J < 0, there exists a constant c := c(M, (J) ~ 1 such that
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)
x E Eo. (2.10.7)
Then
(Eo, E1)e,q = {x E Eo ; IIXlle,q,A < 00 }
for all A E H(El' Eo) satisfying (2.10.6) with type( A) :::; (J". Finally,
Then, given M 2': I, there exists a constant c := c(M) 2': 1 such that
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].
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.
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
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
we infer from (2.11.4) and the reiteration property (2.8.2) of the real method that
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
and
(Eo,Ed(lO:)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
Proof This follows from an obvious modification of the proof of [Ama90b, Propo
sition 6.4j .•
(1.1.3)
for p, IJ, T E [0,1) U {1} with T :s; IJ :s; p, where BUCo := BUC and CO := C.
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,
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
(1.1.4)
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
pointwise 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).
(1.1.5)
and, if p > 0,
(1.1.6)
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
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)
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
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)
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
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 wellsuited 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
h := { (t, s) E J x J ; sst}
and we put
J!:,. := { (t, s) E h ; s < t} .
If 'P is an operatorvalued function on J!:,. and t/J is such a function on J, we write
and
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
(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 . •
(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
(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
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),
(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
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
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)
In particular,
(3.1.6)
by the identification
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
(3.1.8)
and
(3.1.9)
w:=L~' (3.2.2)
j=l j
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
(3.2.5)
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. _
if a + (3 < 1.
These generalizations are occasionally useful  in Theorem 3.3.1, for example. _
52 II Cauchy Problems and Evolution Operators
3.3.1 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
1t
and
u(t) ::; AC,6 +B (t  T)aU(T) dT , a.a. t E j , (3.3.2)
a.a. t E j , (3.3.3)
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
U = (a  b) + w * (a  b) .
By Lemma 3.2.1,
t>O, E>O.
Since
1t eV(tTl(t  T)O:T,6 dT::; evt 1t (t  T)"T,6 dT
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(ts) in Lemma 3.2.1 and in Theorem 3.3.1
and its corollary  where J.t:= m1/(1a) in Lemma 3.2.1  can be replaced
by e(/L+c)(ts) . •
4.1.1 Lemma Given {J E (0, 7f 12), let r be an arbitrary piecewise 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.+ .
Then
1f(z,x,t)dz=
r
(1 + Js, + Jr,
r,
f fJf(z,x,t)dz
+ J iI+1I"/2
il1I"/2
f(tIei'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
It should be remarked that estimates of the type given in this lemma are
standard in the theory of analytic semigroups (e.g., [Tan79]).
(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:
Ak(05)etA(s) = (It
27rZ it
r >.ket'\(>. + A(s)r 1 d>' , t>o, kEN,
r being a piecewise smooth curve running in t1'l from ooe i (1'I+71j2) to ooe i (1'I+7r/2).
From the resolvent estimates of (4.2.1) and from Lemma 4.1.1 we deduce that
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> etA(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
Moreover,
etA(s) _ etA(s') = ~
27TZ
r etA [(>. + A(s)) (>. + A(s')r l ] d>'
Jr
1 
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.
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)
II(A+A(t))~l_ (A+A(s))~lll.c(Eo)
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
Ila(t, s)  a(T, s)llc(Ej,E1 ) :::; M it (0"  S)j2 dO" :::; M(t  T)(T  S)j2
II [A(T)  A(s)] [a(t, s)  a(T, s)] Ilc(Ej,Eo) :::; c(t  T)(T  S)p+j2 (4.3.16)
Ilk(t, s)  k(T, S)II.c(EO ) ::; C[(t  S)pI + (T  SyI] ::; C(T  syI (4.3.20)
+ 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
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),
I
Put
t £
d£(t,s):= s a(t, T)W(T,S) dT , O<c<ts,
II.4 Existence of Evolution Operators 61
as E + 0. (4.3.23)
Moreover,
+ i t ce
 (t,r)w(r,s)dr (4.3.25)
+ it e(t,r)w(r,s)dr (4.3.26)
(4.3.27)
and
(t  s)Ijp IldA(t, s)llc(Ej,Eo ) ::; c(E)e(I'+c:)(ts) (4.3.28)
for E > 0, (t, s) E J'6., A E A, and j = 0, 1, and that
aA*wA(',S) ECI(Jn(s,oo),.cs(Eo)), s E J ,
and
(4.3.30)
Consequently,
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)
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)(ts) (4.3.35)
jorc>O, (t,S)EJ'A, AEA, andj,kE{0,1}.
Proof Put
A , s) .=
UC(t . e(ts)A(s) +d (t s)
E:, , o<c<ts.
Then it is easily verified that
8 1 Ul(t, s) + A(t)Ul(t, s)
= kA(t, s) + ecA(tc)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
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)
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)
and
02a(t, s) = A( s )e~(ts)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
v E Jt(Eo, 0) (4.4.8)
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*mv = 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
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
implies
[Ag]p,J:::; 'TJ, c: > 0 . (4.4.14)
Since, as is wellknown 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
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
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
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 .•
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
Put
Vs(t) := 1tce(tT)A(T) f(T) dT , E>O, tEJ, t?E,
and that
ItEA(T)e(tT)A(T) f(T) dT
=  I t  Ee (t,T)f(T)dT+ I t  cA (t)e(tT)A(t)f(T)dT.
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
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
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.
5.1.1 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
°
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
(5.1.2)
such that
for (t, s) E J'l:,., A E A, and j = 0,1, where Il:= Co (p)1]l/p . Since, by Remark 2.1.2(d),
UA(t , s) .=
.  A(t , s)
e(o+w)(ts)Ua+w+ , (t, s) E h , (5.1.3)
5.1.2 Remark Suppose that A E 1i(E1 , Eo) and II> s( A). Then
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
HilleYosida theorem implies the converse estimate. _
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(ts)
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
and
(t  S)(lI;)" IIUA(t, s)11 ( ) :::; cel/(ts)
£ 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). _
Proof By replacing in the argument leading to the integral equation (2.2.2) the
function e(rs)A(s) by UB(r, s), it follows that
(t, s) E h .
Recall from Remark 2.1.2(e) that the mild solution of (5.2.1)(x,A,f) is given by
5.2.1 Theorem Suppose that 0 ::; f3 ::; n ::; 1 with n > 0 and f3 < 1 and 0 < 'Y ::; 1.
Then
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 Xspaces 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
X~ '+ Eo UA
+ El '+
X~
that
~ E {a,,8,,} .
Similarly,
and
(t  s) f3 IIUA(t,s)IIL:(x.."x/3) :S cev(ts)
for (t, s) E J6.. Hence we deduce from the representation (5.2.2) that
u(·,x,A,f) E C(j,Xf3)
(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
with
01V(t, s) = A(t)UA(t, s)x .
Hence it follows from (5.3.4) that
and
Ilv(t, s)  v(r, s)llo ::; c(t  r)"'e(ve)(ts) 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)(ts) 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 (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
for T ::; 05 ::; t. This estimate shows that the second integral in (5.3.9) is bounded
in the E;3norm 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 .•
5.4.1 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~
5.4.2 Corollary Let the assumptions of Theorem 5.4.1 be satisfied and suppose
v < o. Then
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
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
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 welldefined for 0 < E < l/w. Since
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/(IEw)] 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
+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),
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. _
as c ~ O.
Let
B:=w+A
and note that (6.1.2) implies
and
6.2.1 Lemma There exists c > 0, depending on K and w only, such that
Proof Corollary 6.1.3 and (6.2.3) imply the existence of {) E (0, 7r /2) and of a
constant c with
s E J.
sEJ, 1';>0,
I'; > 0.
II.6 Invariance and Positivity 79
Consequently,
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(ts) , (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)
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",(ts) , (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:
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
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 .•
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,
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 halfspaces [Re(e', x) ::; a]
containing X .•
t ~ 0, (6.3.1)
iff
(1 + SA)l(X) eX, (6.3.2)
(1 + SA)lx = 1 CXJ
etecAtxdt , x E E, 0 < s < l/a+ . (6.3.3)
Thus, letting JL := etdt on Sl := lR+, Lemma 6.3.1 and (6.3.1) imply (6.3.2). On
the other hand,
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
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
(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
Let v(t) := e(tS)/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
f(w):=v+Kw, wE C(I,Eo) ,
that u(·, x) = lim Uk in C(I, Eo), where
Note that
f(w) = ax + (1  a)Mw ,
where a(t) := a(t, s) := e(tS)/E and
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
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
Let B := 2w + A. Since etB(s) = e2wtetA(s) we infer from (6.3.8) and the fact
°
that X is a cone that etB(s) (X) C X for t ~ and s E J. Thus, thanks to
Theorem 6.3.2,
iff yXEP.
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
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.
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 {etA(s) ; t 2: O}, s E J, is positive.
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 righthand 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 24. For
88 III Maximal Regularity
the reader's convenience we begin by recalling some known facts about Sobolev
spaces and absolutely continuous functions.
cp E V(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 Evalued distributions on X.
IIL1 General Principles 89
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) ,
(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 pairwise 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
for each pair a, b E J with a < b and each sequence of pairwise disjoint subintervals
Jk:= (ak,b k ) of (a,b) satisfying
(1.2.3)
1.2.1 Remark The above definition of absolute continuity differs from the clas
sical one of the finitedimensional 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.
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)
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 ~~locsolution oj the Cauchy problem (1.3.1)(s,x,A,f)' Then
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) .
s E JR+ ,
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
(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)
Proof Since, trivially, "( E £(lEl' "(lE l ) , we see that (1.5.4) is equivalent to
AEA.
(1.5.6)
and
(1.5.7)
Note that
(1.5.8)
and that
(1.5.9)
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.
Let E and E be Banach spaces such that E 't LI,loc( J, E). Given 0: E JR, put
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(eoEoxyE"eoE,) = II (8 + 0: + A, ,)II1£(EoxyE"E,)
for A E A.
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)
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
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 timedependent 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)
Proof Note that (8 + A, 'Y)u = (I, x) for u E 1E1 and (I, x) E lEo X 'Y1E1 iff
u = v +W E 1E1" EB 1E 1 ,B .
98 III Maximal Regularity
Thus
and
u= v +w EEl,')' EB El,B = El
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 .•
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 .
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
and
Lastly,
and, of course,
BUC~(J, E) := BUC:,o(J, E) .
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].
o ::; T ::; a ::; p < 1 and 0::; Jt ::; 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
Moreover, if a > 0,
Ilu(s)Is _ u(t)11
tl U ::;
Is  tIPu [u ]*p,[e,2e]::; ()PU[
11\ c:
]*
U p,[e,2e]
Consequently, Hu,v ::; [']p,J.' for 0 < a ::; p < 1 so that (2.1.1) implies
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,
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
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
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
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. _
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.
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)lj 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 .
[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) .
(2.2.4)
104 III Maximal Regularity
(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» .•
and denote by II·II(} the norm in E(}. Then we prove the following fundamental
mapping result.
x EElI" AEA ,
and that tl' RAX(t) + 0 in El as t + 0 for x E E 1  w If Jl = 0,
(2.2.6)
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 ;= A1y 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 ~ E1fJ' it follows that dom(A2)
is dense in E 1 w Thus, given x E E1fJ' 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
for x E E1fJ 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).
uniformly with respect to J, that is, the norm of this map can be bounded indepen
dently of J.
Proof Given x EElI" it follows that 8(RAX) = ARAX in Eo. Hence we infer
from (2.2.2) and Lemma 2.2.2 that
and
lEl '+ 1FI '+ Ll,loc(J,EI ) nWl,loc(J,Eo) , (2.2.15)
respectively. Moreover, given assumption (2.2.1),
2.3.1 Theorem Let (Eo, Ed be a densely injected Banach couple and suppose
'H.(EI' Eo) =1= 0. Then
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 EIJ.L (resp. x E (Eo, EdlJ.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
:::; 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
Lastly, suppose that 11 = O. Then we deduce from (2.2.2) and (2.3.2) that
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·IIylE, ::; a II·IIE," '
respectively, uniformly with respect to J. _
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
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. •
2.4.1 Lemma Let assumption (2.2.1) be satisfied and suppose that p is posi
tive. Then
(A t7 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
(1/\ t)1L Ilv2(t)111 ::::; c(l/\ t)1L itt/2 (t  7)le (tT) 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
(2.4.6)
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:
::;c[(1!\E)~l"log(2slt)+ 10 r/ [(sT)1_(tT)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
Thus
Similarly,
AV2(t)  AV2(S) = ( r 1
It/2
8
8/2
)Ae(tr)A(J(T)  f(t)) dT
+1 S
A[e(tr)A  e(sr)Al(J(T)  f(s)) dT
1
s/2
+ 8
Ae(tr)A(J(s)  f(t)) dT
s/2
= ( I 8
t

It/2 )Ae<tr)A(J(T) 
8/2
f(t)) dT
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!2
[(s  7)1  (t  7)1] (s  7)P d7 [j]c~,,,([0,2E],Eo)}
is s~
[(s  7)1  (t  7)1] (s  7)P d7 = (t  s) is
s~
(s  7)pl(t  7)1 d7
(2.4.10)
(2.4.11)
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
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)
(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.
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.
2.5.1 Lemma Let assumption (2.2.1) be satisfied and suppose that p > O. Then
KAf E Wl1oc(J, Eo) and
8(KAJ) = f  AKAf
Proof Let (J, A) E BC%,!'(J, Eo) x A be fixed and put T:= sup(J). Note that
(2.5.1)
(2.5.2)
Note that
+ t.p(r) [1  e~(T~T)Alf(r) .
= (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
2.5.2 Corollary Let assumption (2.2.1) be satisfied and suppose that p > O. Then
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)
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.locsolution
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, EIhp,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
II (8 + A, ,)III.c(ewlEo xEl_/L,ewlEl) :S c
for A E A.
Then
l'lEl ~ ElI' := (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) .
(2.6.1)
(2.6.2)
for p, fl E [0, 1). From this we infer that, given Banach spaces E j ,
(2.6.3)
2.6.1 Theorem Let (Eo, Er) be a densely injected Banach couple and, given
p E (0,1) and fl 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
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
(2.6.7)
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
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
Put
IIAj  BSj IIL(lE " lE O ) ::; IIAj  BSj Ilc~([O,6]n(Jsj),L(El,Eo)) ::; 1/(2a) .
(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
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 ,
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 EII" be given. Let
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 .•
Now it is easy to give the missing part of the proof of the main existence theorem
of Subsection 11.1.2.1.
Let S < So < Sl with Sl E J and put h :=: [0, Sl  so]. Define
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
(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
Thus
where we used Lemma 1.1.1.2 for the last estimate. Since it is wellknown 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]) .•
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.
11(.:\ + A)Axllo
IAllIAxll o + IIAxlll '
shows that
}
and 1J E (0, 7l' /2) such that ~1? C p( A) and
(3.2.2)
IIAII.C(E"Eo) + (1 + IAl)lj II(A + A)lll.C(Eo,Ej) ::; M
for (A,A) E ~1? x A and j = 0,1.
(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)
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.
3.3.1 Lemma Suppose that 0 ::; Jl < 1 and 0 < () < 1. Then
.+
for s, t E lR .
Proof If t ::; 2,
If t 2': 2,
Similarly, if t ?: 2,
Ift:S 1,
12 (8, t) =i 1
(8 +t  7)112 7 1' d7 + i t (8 + t  7)112 d7 :S C8 11  1 •
t/2 1
3.3.2 Proposition Let assumption (3.2.2) be satisfied and suppose that 0 :S JL < 1
and 0 < e < 1. Then
°
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
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,
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
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
and that
AEA, (3.3.5)
and that 8(KAf) = f  AKAf. Hence, thanks to (3.2.3), (3.3.4), and (3.3.5),
uniformly with respect to (f, A) E BUC/,(J, Ee) X A and J. Now the assertion
follows from (3.3.4) .•
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,
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. _
where Ao(t) := [A(t)]o for t E J. Then the proof of Theorem 2.6.1 shows that
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 Jt = O. That paper also contains the result of Remark 3.4.2(b) (in the case
Jt = 0, of course, and for J bounded, cf. [DaPG79, Theoreme 3.6]). The extension
to the case Jt > 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].
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.
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)
Recall that £(X) and £'(X) are Montel spaces. Hence they are reflexive.
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)
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 welldefined and bilinear. Moreover, if rn E £(JR n ) then
(4.1.9)
Clearly, the roles of E and lK can be interchanged on the lefthand 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 ).
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 FOURIERSCHWARTZ THEOREM guarantees that
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.
Moreover,
(4.2.8)
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 ) ,
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)
(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),
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 lefthand 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
in
(4.2.24)
and in
then the proof of [FoI84, Theorem (8.15)] carries over to the Evalued 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)
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 welldefined and the CON
VOLUTION THEOREM states that
(u * <p)~ = fiip .
(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).
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
T E lR?.
Then
((l/ t )E,<P? = 1 <p(T)dTIT,
ITI::>.':
<p E S .
It is wellknown and easily seen that there exists a unique temperate distribu
tion PV(llt), the principal value of lit, such that
<pES. (4.3.1)
so that
t E lR?, u E S(lR?, E) .
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 BanachSteinhaus
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
'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
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
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
skewadjoint 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)
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.
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'] .
as 10+0, rpES.
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
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,
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 .•
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 pairwise disjoint cubes Qk, with sides
parallel to the coordinatehyperplanes, such that
(4.3.10)
and
aLAn(Qk) ::; IIul11 , (4.3.11)
where AnU denotes the ndimensional 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 Evalued 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)
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)
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)
(v',Uhr(JR,E) = k \V'(t),u(t);dt
(4.3.18)
(4.3.22)
The proofs of this section are adaptions of wellknown arguments from har
monic analysis used in the study of the scalar Hilbert transform (e.g., [Ste70a]).
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 subafields of A. A sequence (Uk) of Evalued Pintegrable
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 Evalued martingales (Uk), where Ul := 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]) .•
(4.4.1)
We also put
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 Lpcase. It is interesting and of great importance that this theorem essentially
carries over to the Evalued 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 welldefined,
F1(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 F1m * u is welldefined 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 F1m * .)
I1I.4 Maximal Sobolev Regularity 143
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)
l<p<oo,
Proof For the first assertion we refer to [Zim89], the second one is obvious .•
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.
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
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 aglcp, 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 * 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 finitedimensional 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 l7finite 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 'pointwise 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
IIHLp(X,/L;E),EVIILp(lR,Lp(X,/L;E)) = ~{
1
{ {I (
llR x Arl?E
v(x, t  r) dr/riP dfl dt} lip
{i k
7r
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) .
(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
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
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].
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.
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
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 pointwise multiplication. Given'P E <P,
we put
where f is any piecewise 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 welldefined 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
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,
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 'Pk is holomorphic in t
and we can deform r to the positively
oriented circle centered at 0 ofradius r, where 0 < r < liliAIII. Since
00
it follows that
Rez < 0,
(cf. (1.2.9.4)).
150 III Maximal Regularity
that is,
A z = sinrrz
rr

0
00
1
s Z( s+ A)l d s, 0< Rez < 1 . (4.6.9)
1 00
sZ(I+s)lds=rr/sinrrz, 0< Rez < 1 . (4.6.10)
4.6.2 Theorem {AZ; Rez < O} U {AO = IE} is a strongly continuous holo
morphic semigroup on E.
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
AZx  x =   . 1 00
sZ(s + A)lXds  sm
• 
rrz 1° 00
sZ(1 + s)lxds
= 
rr
sin rrz
rr
1 ()
0
0
00
sZ
l+s
(s
rr
+ A)l(l A)xds.
Consequently,
O<Rez<l.
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
.
Az=SlO'lfZ
'If
'1
m.
(1  z)(2  z) ... (m  z) 0
00 smz(s + A)m1ds (4.6.17)
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 righthand 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. •
logA
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
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
O~t~m, mEN,
and Az = ezlogA 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

sin trt
IIAtll < Km+1__
7r
m'.
(1  t)(2  t) ... (m  t)
10
00
smt(1 + s)ml dt = K m+1
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)
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 > Aly, which implies y = O. Hence A z is
closable. Motivated by (4.6.18) and (4.6.19) we put
AZ = ~
27rZ
r(_,\y(,\ + A)l d'\ ,
lr
(4.6.22)
where r is any piecewise smooth simple curve running in C\lR+ from ooei'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
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
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
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 = AW(AZy) = 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 Au, it follows that D(AW) '+ D(AZ) '+ E.
Given x E D(AZ), put y := AZx E E. Since D(AWZ) is dense in E by (4.6.14),
given E > 0, we can find u E D(AWZ) 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 = AzmAmx for x E D(Am). Hence Theo
rem 4.6.2 guarantees that
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 .
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
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
=sin7rZ
f ( l  z)
00
eIe t A dt.
7r 0
{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.
4.6.8 Remarks (a) Observe that s(s + A)l :J 1  (s + .1)1 A, and thus,
sl(s+A)l =slA 1 _(s+A)lA 1 .
11
see that
    1 AI x+
A z x Sin7rZ{X s z+l( s+ A)lA 1 x ds
+1
7r Z l+z 0
00 szl(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"
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
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] .•
where the last equality follows from the spectral calculus for selfadjoint 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
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),
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
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 Aa we obtain
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
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 welldefined 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
where r is a piecewise 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 = [ACl,6]l = ACl,6 for f3 > 0. This proves
the theorem .•
A E BIP := BIP(E) ,
provided A E P(E) and there exist c > 0 and M 2': 1 such that
The following theorem shows that this assumption has farreaching consequences.
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
Bx = lim C1[Aitx  x]
t>O
(4.7.4)
for x E dom(B), s:2: 0, and t E R On the other hand, the analyticity of Az for
Re z > 0 implies
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
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
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
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.
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 afinite 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
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 Ip1  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.
then A E BIP(M,O).
(ii) If A E BIP(M, 0) n PK then
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 As+it = As 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
and if
A E Sf} . (4.7.8)
Put
P( fJ) := P( E; fJ):= U P(K, fJ)
K~i
Observe that P = P(O), that is, every operator of type (K, fJ) is of positive type.
Let
7rfJ 1
fJ K := fJ +  2  1\ arcsin 2K . (4.7.10)
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
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
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 _ ,\(,\ + A)I]
A(8+A)lAZX=~[r
27rZ
(_'\y1 d'\8A(8+A)lX+
lr s  ,\
r (,\)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)
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)
M:= c(a) 1 00
C"'{II  te il1 r + 11 te r
l il1 l } dt .
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 .•
where B(t) := 1 for t 2: 0 and B(t) := 0 for t < 0 is the Heaviside function.
Proof (i) It follows from (4.8.1) that 1 + B(A + A)l E .L:aut(E) and
IlIA Maximal Sobolev Regularity 169
Eo := E, E 1 := (dom(A), ijA11)
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 ;31) l/(la)  1] + S(O) .
Proof Since A(>. + A)l = 1  >.(>. + A)l we see that
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
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
BZ 1.
= 2
7fZ
r
ir
(AYC(A + A)lDdA = CAzD, Rez < 0 .
Next we prove a simple 'splitting lemma' that will greatly simplify our proofs
that a given operator of positive type has bounded imaginary powers.
and put
AEf, Rez<O.
Also suppose that
Rz,S E L1(f,ds,£(E))
and
E < Rcz < 0, (4.8.6)
1.
A Z = 2
7rZ ir
rR z (>,) d>' + ~
27rz ir
r(>'yS(>.) d>' , Rez < 0 .
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
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
lr .
BE £(E", , Eo), where E", := ~",(Eo, Et}, and put
Proof Put Jl := JlB and let Ao := Jl + A and Ro(.>.) := (.>. + Ao)l. Since
sin <OK]
.>. E M(K, <0):= [ Izl::; 2K U StJK
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
(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~.
°
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.
(4.9.1)
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).
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. _
.
2z ciCXJ Slll7rZ
dz: .
l
1 c 1+ iCXJ BZAzx
2z cliCXJ Slll7rZ
dz. (4.9.4)
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 A1eitBeHit x = BelBitAitAlex by Theorem 4.7.1 and by
Lemma 4.9.2(ii), the integral
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
DOREVENNI THEOREM.
178 III Maximal Regularity
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 halflines {j + it; It I ~ E}, and the circular arcs
Thus
Sx = Ro,ex + To,ex = R1,eX + T1,eX ,
where
11 AjitBjHit x
R j eX = dt
, 2 Itl~e sin 7r(j + it)
and
TJ. eX _ (1)j 1)11'+11'/2 Eei<P
.
Ajeei "'Bjl+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 AIX
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 AitBitx
BRa,ex = .. dt = AR1,eX .
2 Itl~e SIll 7rZ t
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)
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 DoreVenni 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 DoreVenni 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,qrealization 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]).
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.
Now we can characterize the trace space of E1 and prove an important em
bedding theorem.
4.10.2 Theorem "YE1 ~ (Eo, Et}11/p,p and E1 '+ BUC(J, (Eo, E 1h1/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 LionsPeetre 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. •
}
and {} E (0,71)2) such that l.;ii C p(A) and
(4.10.1)
IIAILc(E1,Eo) + (1 + IAI)Ij II (A + A)lllc(Eo,Ej ) ::; M
for (A, A) E l.;ii X A and j = 0, 1.
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
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, EIhl/p,p .
(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
Proof Given'x E C, it is easily verified that (,X + B)i exists and is given by
t E J, v E lEo .
v E lEo ,
t > 0,
t :S 0 ,
for 0 < Re Z < 1 and E 2: o. By the convolution theorem
gz,e *r C rl
v = J" r C
mz,eJ"r v , v E lEo ,
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)
m z,€ (T)I
1 <
_ ERezell"IImzl/2 , TE~, E>O, O<Rez<l.
Now we infer from (4.10.6), (4.10.7), Theorem 4.4.3, and Lemma 4.7.4 that
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
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 rF1mit,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.
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)
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)
and
(4.10.15)
for (f,x) E lEo x E11/p,p and A E A. Theorem 4.10.2 and (4.10.2) imply
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)
j+1
L1
00
= 121IKAV(t)llf dt
+ 2P  1 f 1]+1 {111
j=2 J 0
j

1e(tr)Av(r) drll:
:S cf 1]+1 (l j 1
 (t  r)l e w(tr) Ilv(r)llo dr r
dt
r
(4.10.19)
1 (1
j=2 J 0
:S c 00 t

1
ew(tr) Ilv(r)llo dr dt
:S wPc1 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(tT)A h1(T + j 1) + Vj(T + j 1)] dTII: dt
(4.10.20)
::::; c Ilvj1 + Vj II~p(lR+ ,Eo)
Since the functions Vj have disjoint supports it follows that
00 00
= 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.
4.10.9 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),
Co sup
tEJ
Ilu(t) II~ Ilip
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
If J = jR+, also assume that there exists T ::::: ° such that s (  A(T)) < °
and
where
(lEo, lEI) := (Lp(J, Eo), Lp(J, E 1) n w;,1(J, Eo»)
and E 1 1/ p := (Eo, E 1 h1/p,p'
8 E J , (4.10.31)
8 E J.
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
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
o :<:; t :<:; 81 ,
81 :<:; t :<:; 83 ,
u(O) =x .
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
(In [GiS91] also nonzero initial values are admitted.) Note that the lefthand 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
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 'LpLqestimates' 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.
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· . f f(x
d+ f( x ) . ImJil
+ h)h  f(x)
h~O+
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 welldefined, Xo < X2, and fe(x) > fe(xo) for Xo < X < X2·
Consequently, d+fe(xo) ~ 0, which is impossible. This shows that
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·
g(x) := f(xo) + l
Xo
x
8+f(t) dt , x E J.
IV.l Higher Regularity 195
Proof It follows from F '+ E that of E C (J, E). Hence f E C 1 (J, E). Thus
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 .•
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.
f E C(J,Fo) , (1.2.2)
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) .
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).
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
Ih ~ f(t) (1.2.5)
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
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. _
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
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).
(1.2.10)
Since AU(t, s)x = AUAl(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
A c CP(J, H(El' Eo)) and there are constants M,7] E jR+ and
(0, 7r /2) such that
() E
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,.
More precisely,
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
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
and
and that
(t  s)l')'p IIw(t, s)llqEy,Eo) ::; e(E)e(J£+c)(ts) (1.3.8)
for E > 0, (t, s) E J~, and A E A.
Put
ft(t,s) := e(ts)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
and
(tS)l p 11ft (t, s) Ilc(Ey) +(tS)l')'p 11ft (t, s )llc(Ey,Eo) ::; e(E)e(J£+c)Cts) (1.3.10)
Using this estimate, an obvious modification of the proof of Lemma 11.4.3.2 gives
(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 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
if 'Y + p  1 2:: 0 ,
if 'Y + p  1 < 0 .
/Ih(t, S)/IC(E'Y) ::; c(c) It (t  T)l",/ [(t  T)i3(T  s)"'ll + J(t, T, s)] dre(fL+E)(ts)
for € > 0 and (t, s) E J~. Similarly, (II.4.2.3) and (1.3.16) imply
(1.3.17)
and
(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) .•
Moreover,
IIAUAA 1(t, S)II.c(E",) ::::; c(€)e(/LP!(P"')+e)(ts)
Proof Put
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~(E1,1p) (1.3.20)
and
Ah E ~(El, E,)" 1 + 'Y  p) , (1.3.21)
respectively, together with the estimate
(1.3.22)
U=b+h*U. (1.3.23)
and
(t,s)EJi:,., AEA. (1.3.25)
Using (1.3.21), (1.3.23), and (1.3.24) we find
(Ab  aA)A l
1.
= 2
7r2
r AeW 
lr
s) (A + A(t)r l [A(t) A(S)](A+A(s)rlAl(s)dA. (1.3.28)
and
(t  s)'YP II(Ab  aA)A 1 (t, s)II.c(Ey) ::;; 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(Eyl ::;; C , (t,s)Eh, AEA. (1.3.34)
Thus we obtain from (1.3.27) and (1.3.32), (1.3.34) that
and
IIAbA 1 (t, s)II.c(Ey) ::;; 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)
and
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
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
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
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,
and
(1.4.4)
for c > 0, (t, s) E Jb., A E A, and j = 0, 1, where we have put Jt := CoT/lip for a
suitable positive constant co.
°
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)
where, thanks to A"(UA,"( E Jl.(E"(, 1), the integral converges in L(E"(). This implies
(1.4.8)
t E J ,
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
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
Hence dom(A,(t)) = EH,(t) is dense in E, (cf. Section V.1.5 for more details). _
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)
(1.4.13)
and
IIAxUA,X A Xl (t, s) IIL(X) ::; c(E)e(/LP/(P'Y+)+c)(ts) (1.4.14)
for E > 0, (t, s) E Ji:" and A E A. Moreover,
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
(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
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
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).
(1.5.5)
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
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)
and
(t,r) ...... A(t)U(t,r)[h(t)  her)]) E G(Ji:..,E'Y) .
This implies that the first term on the righthand 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)
tf, ~
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
where
(1.5.10)
Hence
A(t)gl;(t) = Jo
t E AU(t, r)!I(t) dr
(tI;
= [ef,A(t)  etA(t)l!I (t) + Jo B(t, r)!I (t) dr ,
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 ....... etA(t)] E C(J,.cs(E"()). Now we infer from (1.5.10) and (1.5.11) that
(1.5.13)
(1.5.14)
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
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.
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)
r being a piecewise 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 welldefined for A E 21 M and z E C.
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:),
j = 1,2,
in .L(E)
(2.1. 7)
in .L(E)
t>o, c>O.
Hence the assertion follows from Lemma 2.1.:3 and the dominated convergence
theorem. _
214 IV Variable Domains
Note that (Eo, El(A(s))) is a densely injected Banach couple and that
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 [A1]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
and estimate (2.1.6) is valid for A(s), uniformly with respect to s E ] 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
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.
Moreover, given A E A,
and
(2.2.11)
as c + 0, uniformly with respect to s E J.
216 IV Variable Domains
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
11[(t+AE(s»)l  (t+A(s»)l]A(s)(t+A(s»)III.c(Eo)
for s E J and t ~ O. Hence we deduce from Lemma 2.1.3 that, given x E El (A(s») ,
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)) . •
and
r,s E J, A EA.
Moreover, given A E A,
(2.2.16)
A"'(r)etA(r) _A"'(s)etA(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)eTA(S)
(2.2.17)
= A<>(r)etA(r) _ A"'(s)etA(s) + A<>(s)[etA(s) _ eTA(S)] .
Since
(2.2.18)
218 IV Variable Domains
for s E J and 0 < T <t< 00. Now the second assertion is a consequence of (2.2.16)
(2.2.19) .•
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,
etA(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
for s E J and 0 < T < t < 00. From this, Corollary 2.1.2, and the strong continuity
of the semigroup {etA(s) ; t ~ O} on Eo we infer that
and
kA(t, s) := AD:(t) [A1(t)  A 1(s)]A 2 D:(s)e(ts)A(s)
for (t, s) E Jt", we see that VA := UAA1D: solves the linear Volterra equation
(2.3.2)
Similarly, let
and
hA(t, s) := _A 2 D:(t)e(ts)A(t) AD:(t) [AI(t)  Al(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
(2.3.4)
Now suppose that condition (2.2.3) is satisfied and fix a such that
Ip<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 righthand side of (2.3.3) and that UA is a
parabolic evolution operator for A.
(2.3.7)
(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)
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
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)
and
(t  s)la p lIaA *w~(t,s)IIC(Eo) ~ c(E)e(v+c)(ts)
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
Next we consider the Volterra equation (2.3.4). For this we need the following
technical result:
bA E Jt(Eo, 2  0:  p) .
Moreover,
where
IIA(t, s):= l r
A 2 (t)e(tT)A(t) [e(Ts)A(T)  e(Ts)A(s)] dr,
IIIA(t, s):= l r
A(t)e(tT)A(t) [A(s)e(TS)A(s)  A(r)e(TS)A(T)] dr,
and
We rewrite IA as
222 IV Variable Domains
IIIA(t, s)II.c(Eo) :::; c(r  s)l i t (t  r)a+ p2 dr = c(r  S)l(t _ r)a+ p1
= c(t  s)a+ p2
r
IIIIA(t, s)II.c(Eo) :::; c(t  r)21 (r  s)Hpl dr = c(t  s)Hp2
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(tr)A(t) [e(rs)A(s) _ e(rse)A(t)eeA(s)] .
Finally, we are ready for the proof of the first main result of this section;
Proof We shall show that the function UA, defined by (2.3.6), has the desired
properties.
IV.2 Constant Interpolation Spaces 223
(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
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
(2.3.28)
Moreover, given (t, s) E JJ:., Lemma 2.1.4 and Corollary 2.2.3 imply
j j
(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)
it follows from (2.3.31) and Lemma 2.1.4 that, given (t,s) E J6.,
as c ...... O. Thus we infer from (2.3.32) and the closedness of A(t) that
and that
and
(t  s) IIAUA(t, S)II.c(Eo ) ~ c(c)e(V+E)(tS) (2.3.36)
Observe that
for 0 ~ s < T < t and t E J. From this, (2.3.30), (2.3.32), and (2.3.36) we infer that
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,
= 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,
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
(2.3.43)
and that
(t  s)o IIUA(t, s)llc(Eo,Eo) :S c(E)e(v+c)(ts) (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 .•
[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)
AEA,
Moreover,
A(t)aA(t,s)AI(S)  A(r)aA(r,a)Al(a)
= [A(t)e(ts)A(t) _ A(r)e(Tcr)A(T)]A1(s)
+ A(r)e(Tcr)A(T) [Al(S) _ Al(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,
AaAAl(t,s) + 1
(2.4.6)
Let
CA,j(t, s) := Aj"(t)e(ts)A(t) A"(t) [Al(S)  Al(t)]
for j = 1,2, (t, s) E J'A, and A E A. Then
j = 1,2, (2.4.7)
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)1o: p Ilb AA 1 (t,s)II.c(Eo) ~ c(c)ec(ts) , (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) .•
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
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
{etA(s) ; t::::: O} is positive.
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.
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)
+ [1 etA,,(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
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,. .
(t,s) E h .
The fact that Holder continuity of f implies the solvability of the Cauchy
problem (2.5.1)(x,A,f) is wellknown 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.
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
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
so that the Banach space .L(E!, Eo x F) induces the natural topologies on these
subsets.
For definiteness we use i1norms in product Banach spaces. If it is convenient,
we also identify Eo with Eo x {O} etc.
x E E1 , Re A 2: w .
Then
x=(w+A,B)l(A+A+wA,B)x, x E E1 , A E C,
(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
El,s:= kerB
is a closed linear subspace of E l . We also put
A := As := AIEl,s. (2.6.5)
Moreover,
A E 1t(El,s,Eo,K,w) , (2.6.7)
provided El,s is dense in Eo. Lastly, we set
where Kl := K + W l\illc(E1, E)
0
and i: El '+ Eo is the injection map.
We also put
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
In order to prove the main result of this subsection we introduce the following
assumptions:
with (2.6.13)
[(A, B)] p, J ::; "10 .
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
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
initialboundary value problem (IBVP) with homogeneous boundary conditions.
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(1I/p)/2 O<t<l,
{
(3(t) := t(2I/ P)/2:
l::;t<oo.
Now we put
(2.6.20)
where
WP~(16ha (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
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
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+p1) 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
Proof It follows from assumption (2.6.13), in particular from (2.6.3) and Re
mark 2.6.1(c), that
(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
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
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 .•
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
u E C(J,Lp(X)) nC 1 (i,Lp(X)) .
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 tdependence only. For this we
choose a particular setting that is an abstract version of parabolic initialboundary
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 It A(t) to Chapter VIII, where we will study
more specific settings.
242 IV Variable Domains
We also suppose that lEj := lEj(J) and IF := IF(J) are Banach spaces such that
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
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
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
and put
It follows that
lEI = u.;,2(1I/ p ) (X) .
We choose 8 E C(8X, {O, I}) and recall that rj := 8 1 (j) for j = 0,1. Then we let
FJ ''=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
by
if 3<p<oo,
if 3/2 < p < 3,
if 1 < p < 3/2,
and
by
p~ ~(16)
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
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
Denoting the constant map (t 14 (A, E)) again by (A, E), we also suppose that
(3.2.7)
246 IV Variable Domains
(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
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. _
(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 >. ,
~
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
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'.
1.1.1 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) .
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,
(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 onesided if A E {N, ~+}
and twosided 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 onesided.
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.
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 Eoorealization 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,
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) .•
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
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,
k1k1
Ilxllk :.:; Ixlk :.:; (1 + L II Ili£IIC(EH1,Et») Ilxll k '
j=O £=j
Hence 1·lk is an equivalent norm for Ek, and
Now let Xo E Eo and c > 0 be given. Since Ek.tl :!.. 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 Tj1 = 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 socalled abstract MittagLeffler theorem,
is proven for a projective limit of complete metric spaces (also cf. [Bou53]). For an
application of the abstract MittagLeffler theorem to a situation closely related to
our case we refer to [ArEK94, Proposition 6.2J.
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 Denote by io E £(El' Eo) the natural injection of El into Eo. Then we
deduce from the factorization
AI
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.,
We put
j EN, (1.2.3)
is welldefined.
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 AI E K(E).
Proof The first assertion is easily verified. If A is densely defined, that is,
EI ~ E, it follows from Aj 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 onesided 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 welldefined. Suppose that
1~k~n,
Proof It follows from (1.2.2) and the closed graph theorem that A kBk and
Bk Ak 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= [1A(A+A)lfand
(A + A)nAn = [(,\,+ A)AIr = [,\,AI + 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 welldefined. 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
Suppose that A E P(E). Then the fractional powers, k", of A are welldefined
for Q E JR, and we put
Q E JR+ . (1.2.6)
is a densely injected Banach scale, the onesided fractional power scale generated
by (E, A). It satisfies interpolation inequalities. Moreover, it is compactly injected
iff AI 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,
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
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
<p(t) ::; c l o
t
sai3 1 ds
C
= __ t a  i3
0:(3
, O<t<oo.
O<t<oo .
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 AI E K(E). Thus assume AI E K(E). Then Ak 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 Ak E K(E), there exists a subse
quence, again denoted by (Xj), such that (Ak 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 Aa 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. _
The principal results of this subsection are wellknown, 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].
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 IIAl·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 welldefined Banach space and
(1.3.2)
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 Al is an isometric isomorphism from E 1,o onto El
and (A_l)l :J Al. In the following, A_l is said to be the E_lrealization of A.
(1.3.7)
V.l Banach Scales 263
that
(1.3.8)
that this isomorphism is isometric, and that A1 is closable in E2 and A2 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,j1) and (E_ 2,j2) be the completions of (E, \\A 1'\1)
and (E_l, \\(A_d 1 '\I), respectively. Then
E jI E j2 E
+ 1 +  2
and j1 and j2 are linear isometries onto dense linear subspaces of E1 and E2'
respectively. Thus if we identify, as above, E1 with its image j_2(Ed in E2' we
also identify j1 (E) with j2 (J1 (E)). Therefore, by constructing E2 we have to
replace the already constructed extrapolation space E1 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 Ek and Ak for 1 ::; k ::; m inductively by
Proof The above considerations imply that the pairs (Ek, A_ k ) are welldefined
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 (Em, Am). The last part of the theorem follows now
from Remark 1.2.2(e) .•
(1.3.10)
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 El' ... ,Em already obtained by isomorphic im
ages. On the other hand, if we do not use the above identifications but work with
the continuous injections jk : Ek+l + Ek 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 welldefined. In the following proposition we give a sufficient condition for
Ek(B) ~ Ek := Ek(A) for 1 ::; k ::; m. Recall that the dual of a closed and
densely defined linear operator is welldefined 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
and the density of Ek(B) in E that AkBk has an extension (AkB 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,
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 Ek(>' + A) == Ek := Ek(A) with
and
[(>. + A),]k(A')k = ([(>. + A)A1]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( Am) = 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 (AmY is welldefined 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 Am E P(E). Thus, letting (F, B) := (Em' Am), Theorem 1.2.4 guarantees
that the onesided fractional power scale [(Fa, Bo.) ; 0: E JR+] is welldefined. 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 AI E K.(E). If 0 < 0: :S m then
E cx is a completion of (E, IIAa·ll) and A_a is the closure of A in E_ a .
V.I Banach Scales 267
x E Ema.
and A c Aa, we see that A is closable in E_ n and its closure equals A_c<, •
The above construction of E1 and A1 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 E1 and
an extension iLl of A. In [Ama88c, Remark 6.2] it has been shown that the
extrapolation pairs (E_ 1, A_ 1) and (E1' Ad 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]).
call that A' E C(E') is densely defined and peA') = peA). In order to avoid confu
sion with too many dashes we put
(1.4.1 )
and
[(EJ,A!) ; j E Zn [m,oo)] , (1.4.2)
generated by (E, A) and (EH, AH), respectively, are welldefined. 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.
Then
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.
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)
and
(1.4.4)
Proof It follows from Lemma 1.4.1 that x~ ft (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~
(1.4.5)
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 .•
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 welldefined .•
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)
O~k~Tn,
1.4.7 Corollary If 1:::; k :::; m then Ek = (E~)' and Ak = (A~)' with respect
to the duality pairing (', . >.
On the basis of Theorem 1.4.6 we can now extend our extrapolation scales
to get twosided 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
where
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", Xf7(',x>x
and
(y',x>X = (y',x>y , y' E Y', x EX.
If X is semirefiexive,
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,
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
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
for 0 < k < j < 00. By density, the EjErduality pairing is for each j EN com
pletely determined by the E' Eduality 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
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 twosided power
scale generated by (E, A),
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 twosided power scales
V.1 Banach Scales 273
generated by (E, A) and (E~, A~), where E~ := E' and A~ := A', respectively, are
welldefined densely injected reflexive Banach scales. Moreover,
kEZ,
with respect to the duality pairing naturally induced by the E' E pairing (., .).
j >k .
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
ReziO. (1.4.9)
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
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 twosided 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 welldefined 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 welldefined 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
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
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
twosided 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 twosided [fractional] power scales are welldefined indeed,
is proven here for the first time.
(1.5.1)
1.5.1 Theorem [(Eoo , A,) ; a E [m,oo)] is a densely injected Banach scale, the
interpolationextrapolation scale of order m, generated by (E, A) and (', ')0,
o< e< 1. It is compactly injected iff AI 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
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,£} .
(1.5.2)
j<oo<j+l.
From (1.5.2) and the reiteration property (I.2.8.2) for the real method we de
duce that
() d d 0
E k , Ec (o:k)/(ek),l '+ Eo: '+ (Ek, EC)(o:k)/(Ck),oo (1.5.3)
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
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 ~ (Ejl, 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.
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
Note that
by (1.2.9.9) and thanks to the fact that Am E BIP(E_ m ) by Lemma 1.3.7. Now
the assertions follow from Theorem 1.3.8 and Corollary 1.3.9 .•
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.
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 hj 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(l1)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 = EHy = 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}(mk)/(ek)
~ { {Ei+k' EiH+da/(Hlk), {Ei+k' EiH+d(ek+a)/(Hlk)} (mk)/(ek) (1.5.11)
~ {EHk' EHf+d(mk+a)/(Hlk) ~ Fm ,
1.5.8 Corollary Suppose that ~ E lR with 'Y + ~ rt. IE. Also suppose (.,.)1) = {, .}ry
for 17:= 'Y + ~  b + ~]. Then Fr; ~ Ey+t..
Proof Fix k, e E IE with k < ~ < e. Then (1.5.8) and (1.5.10) imply
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
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
interpolationextrapolation 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
interpolationextrapolation 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.
This proves that the interpolationextrapolation 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 twosided 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 .•
Note that (', .)~ is an admissible interpolation functor for 0 < 0 < l. Thus the dual
interpolationextrapolation scale
that is, the interpolationextrapolation scale generated by (E~, A~) and (,.)~,
o < 0 < 1, is a welldefined densely injected Banach scale. The next theorem ex
tends Theorem 1.4.9 to interpolationextrapolationscales.
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' Eduality 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
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
interpolationextrapolation 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 nlinear maps from E1 x ... x En to F, and
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
(1.5.18)
satisfying
(1.5.19)
(1.5.20)
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, EaI)
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 . •
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 ,
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
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 conjugatelinear
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
(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 selfadjoint also. Moreover, if B = B* 2: b for some
b E JR, it follows from (1.5.25) and (1.5.23) that
(b) Let (X, Jt) be a ufinite positive measure space and let E := L2(X, Jt; OCN ) for
some N E N. Then
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 selfadjoint. 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. _
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.
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 IletBIIL(F) :::; c for 0:::; t :::; 1. Then BF E g(l") and e tBF = etBlF
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 G1realization of B. Then (w + B)e tB =) etB(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) := etBIG l
Gl Gl
Thus, thanks to Lemma 1.1.1, (,,\ + A)l = (,,\ + BF )1, which implies A = BF .•
288 V Scales of Banach Spaces
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 interpolationextrapolation 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
Earealization of A for Q E lR, which is welldefined. Moreover, if A is closable
in E_ a for some Q E (0, m], we put
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)
2.1.3 Theorem Let assumption (2.1.1) be satisfied, let [(Ea , Ba) ; a E [m, 00)]
be the interpolationextrapolation 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
m :s; Ct :s; n  1 ,
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 ~ 0, x E El .
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 Ul(t)
of £(El)' Since B1 E Lis(E, Ed, it follows that
t ~ O. (2.1.19)
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 E1 and C:J AI. 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 = AI. This proves that AI is welldefined, that AI E Q(El' M, w), and
e tA C e tA  1 • Lemma 1.1.1 implies p(Ad = 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
Al and B1 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 welldefined 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 ABj ~ Bj A for j EN. Hence, given
j E N and x E E Hl ,
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
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 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
the interpolationextrapolation 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, Ek == Ek(J1) and
V.2 Evolution Equations in Banach Scales 293
for x E Ek 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
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
More precisely,
a;:::: 0,
and
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 Iperiodic 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]).
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)
A: J>2t..
V.2 Evolution Equations in Banach Scales 295
Put
(2.2.11)
296 V Scales of Banach Spaces
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)a1 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
in FJl , where BJl is the FJlrealization of B. Given t E J, let [( Fdt), B«(t)) ; (E jR+]
be the interpolationextrapolation 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 aI in our situation) that F/3a+1(t) ~ E/3(t) for t E J. Conse
quently, B/3a(t) = A/31(t) for t E J, and since BJl(t) C B/3a(t), we deduce from
(2.2.16) that u(·,x) is a solution of (2.2.12)/31.
V.2 Evolution Equations in Banach Scales 297
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
We fix p, > type( A) and put B := p, + A, so that B E P(E). Then we denote by
the interpolationextrapolation 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 welldefined. 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
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'Eduality pair
ing (', . ). We use these duality pairings without further mention.
that is, {e~tA~ ; t;:::: O} is the dual semigroup to {e~tAQ ; 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
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 wellknown 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)
The following technical proposition gives useful estimates for the norms of
the operators A~ and (.\ + A~) 1 .
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)
IIA~ILc(E~0+1 E~)
J 0:
= IIAalIlC(E_o,E_o_Il ' a E lR.
This implies
FiXe> oand put Uo:= supp(u)C and [,j := ul(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
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.
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
aENn , (2.4.2)
and that
Txu=uIX, (2.4.3)
where ulX is the usual pointwise 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
Proof See [Ste70a, Section VI.3] (also cf. [EdE87, Section V.4.4l) and observe
that these proofs extend trivially to the Evalued 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
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. _
densely injected Banach couple as well. Moreover, the EhEoduality pairing deter
mines the Ei E1duality pairing so that we can denote either one by (., .) without
fearing confusion.
and
v E Lpl (X, E~) n Wp~ (X, E~) .
Then (v,u) E Wl(X) and
Proof First note that (2.4.6) holds pointwise if u E V(X, E 1) and v E V(X, E£).
Next, HOlder's inequality implies that, given any Banach space F,
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. •
We fix T E jR+ and put J := [0, T]. We let E and F be Banach spaces with
F'+ E, assume that
and that
(x, g) E E x L1,loc(J, E) .
Then we consider the final value problem
respectively.
Proof Obvious .•
and
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),
Proof This follows by means of the transformations (2.5.5) and (2.5.6) from the
corresponding results for initial value problems .•
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 .•
Assume that
t E J ,
the interpolationextrapolation 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 ,
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
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
(2.6.10)
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
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.
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)
x ~ E E~113'
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.•
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)
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)131 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
(2.6.23)
(2.6.24)
Put (Fa, Fd := (E(31, 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(31+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(31, 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)(31; thus a ~~locsolution of (2.6.13)(31. 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. _
Let U(31 be the parabolic evolution operator of A(31 and, given T E J, let V!(3 be
the evolution operator of the extrapolated dual final value problem (2.6.16)~(3. Then
that is, the dual operator in £(EL(3' E~(3) of U(t, s) E £(E(3, E(31) 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;31, respectively. Given (xrt, x) E EL;3 x E;3, put
U(T) := U;31(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
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 interpolationextrapolation
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.
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 .•
Then E~ = (E~,P~) is an OBS with the sodefined natural (dual) order. Thus,
by replacing P in (2.7.1) by P~, the cones P~ of E~ are welldefined and induce
the natural preorder of E~ for a E R Also note that A~ E P(E") and
The next duality theorem, the main result of this subsection, shows that Pg is
the dual cone of Pc; 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
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 Aa. 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