An Introduction To Interpolation Theory, 2007 (Lunardi - A.)

An Introduction to Interpolation Theory
Dottorato di Ricerca in Matematica, consorzio

Milano-Insubria-Parma-Trieste
Alessandra Lunardi
February 2007
2
Contents
1 Real interpolation 7
1.1 The K-method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.1.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 The trace method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3 Intermediate spaces and reiteration . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.3.2 Applications. The theorems of Marcinkiewicz and Stampacchia . . . 31
1.3.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2 Complex interpolation 35
2.1 Denitions and properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.1.2 The theorems of HausdorYoung, RieszThorin, Stein . . . . . . . 45
2.1.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3 Interpolation and domains of operators 49
3.1 Operators with rays of minimal growth . . . . . . . . . . . . . . . . . . . . . 49
3.1.1 Two or more operators . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.1.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2 The case where A generates a semigroup . . . . . . . . . . . . . . . . . . . . 58
3.2.1 Examples and applications. Schauder type theorems . . . . . . . . . 62
4 Powers of positive operators 67
4.1 Denitions and general properties . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.1 Powers of nonnegative operators . . . . . . . . . . . . . . . . . . . . 75
4.2 Operators with bounded imaginary powers . . . . . . . . . . . . . . . . . . . 76
4.2.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.2.2 The sum of two operators with bounded imaginary powers . . . . . 82
4.2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.3 M-accretive operators in Hilbert spaces . . . . . . . . . . . . . . . . . . . . . 88
4.3.1 Self-adjoint operators in Hilbert spaces . . . . . . . . . . . . . . . . . 93
5 Analytic semigroups and interpolation 99
5.1 Characterization of real interpolation spaces . . . . . . . . . . . . . . . . . . 100
5.2 Generation of analytic semigroups by interpolation . . . . . . . . . . . . . . 102
5.3 Regularity in abstract parabolic equations . . . . . . . . . . . . . . . . . . . 103
3
4
5.4 Applications to regularity in parabolic PDEs . . . . . . . . . . . . . . . . . 110
A The Bochner integral 113
A.1 Integrals over measurable real sets . . . . . . . . . . . . . . . . . . . . . . . 113
A.2 L
p
and Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
A.3 Weighted L
p
spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
B Vector-valued holomorphic functions 119
Index 125
Introduction
Let X, Y be two real or complex Banach spaces. By X = Y we mean that X and Y
have the same elements and equivalent norms. By Y X we mean that Y is continuously
embedded in X.
The couple of Banach spaces (X, Y ) is said to be an interpolation couple if both X and
Y are continuously embedded in a Hausdor topological vector space 1. In this case the
intersection X Y is a linear subspace of 1, and it is a Banach space under the norm
|v|
XY
= max|v|
X
, |v|
Y
.
Also the sum X + Y = x + y : x X, y Y is a linear subspace of 1. It is endowed
with the norm
|v|
X+Y
= inf
xX, yY, x+y=v
|x|
X
+|y|
Y
.
As easily seen, X+Y is isometric to the quotient space (XY )/D, where D = (x, x) :
x XY . Since 1 is a Hausdor space, then D is closed, and X +Y is a Banach space.
Moreover, |x|
X
|x|
X+Y
and |y|
Y
|y|
X+Y
for all x X, y Y , so that both X
and Y are continuously embedded in X +Y .
The space 1 is used only to guarantee that X +Y is a Banach space. It will disappear
from the general theory.
If (X, Y ) is an interpolation couple, an intermediate space is any Banach space E such
that
X Y E X +Y.
An interpolation space between X and Y is any intermediate space such that for every
T L(X) L(Y ) (that is, for every T L(X + Y ) whose restriction to X belongs to
L(X) and whose restriction to Y belongs to L(Y )), the restriction of T to E belongs
to L(E). We could also require that there is a constant independent of T such that
|T|
L(E)
C(|T|
L(X)
+|T|
L(Y )
), but often this property is neglected.
The general interpolation theory is not devoted to characterize all the interpolation
spaces between X and Y but rather to construct suitable families of interpolation spaces
and to study their properties. The most known and useful families of interpolation spaces
are the real interpolation spaces which will be treated in chapter 1, and the complex
interpolation spaces which will be treated in chapter 2.
Interpolation theory has a wide range of applications. We shall emphasize applications
to partial dierential operators and partial dierential equations, referring to [36], [12] for
applications to other elds. In particular we shall give self-contained proofs of optimal
regularity results in Holder and in fractional Sobolev spaces for linear elliptic and parabolic
dierential equations.
The domains of powers of positive operators in Banach spaces are not interpolation
spaces in general. However in some interesting cases they coincide with suitable complex
5
6 Introduction
interpolation spaces. In any case the theory of powers of positive operators is very close
to interpolation theory, and there are important connections between them. Therefore
in chapter 4 we give an elementary treatment of the powers of positive operators, with
particular attention to the imaginary powers.
Chapter 1
Real interpolation
Let (X, Y ) be a real or complex interpolation couple.
If I is any interval contained in (0, +), L
p
(I) is the Lebesgue space L

p
with respect
to the measure dt/t in I. In particular, L
(I) = L
(I). See Appendix, 2.

1.1 The K-method
Denition 1.1.1 For every x X +Y and t > 0, set
K(t, x, X, Y ) = inf
x=a+b, aX, bY
(|a|
X
+t|b|
Y
) . (1.1)
If there is no danger of confusion, we shall write K(t, x) K(t, x, X, Y ) instead of K(t, x, X,
Y ).
Note that K(1, x) = |x|
X+Y
, and for every t > 0 K(t, ) is a norm in X+Y , equivalent
to the norm of X +Y . Now we dene a family of Banach spaces by means of the function
K.
Denition 1.1.2 Let 0 < < 1, 1 p , and set
_
_
_
(X, Y )
,p
= x X +Y : t t
K(t, x, X, Y ) L
p
(0, +),
|x|
(X,Y )
,p
= |t
K(t, x, X, Y )|
L
p
(0,)
;
(1.2)
(X, Y )
= x X +Y : lim
t0
+ t
K(t, x, X, Y )
= lim
t+
t
K(t, x, X, Y ) = 0.
(1.3)
Such spaces are called real interpolation spaces.
Since t K(t, x) is continuous in (0, ) for each x X+Y , then (X, Y )
(X, Y )
,
.
The spaces (X, Y )
are also called continuous interpolation spaces.

The mapping x |x|
(X,Y )
,p
is easily seen to be a norm in (X, Y )
,p
. If no confusion
may arise, we shall write |x|
,p
instead of |x|
(X,Y )
,p
.
Note that K(t, x, X, Y ) = tK(t
1
, x, Y, X) for each t > 0. By the transformation
= t
1
, which preserves L
p
(0, ), we get
(X, Y )
,p
= (Y, X)
1,p
, 0 < < 1, 1 p , (1.4)
7
8 Chapter 1
and
(X, Y )
= (Y, X)
1
. (1.5)
So, pay attention to the order!
Let us consider some particular cases.
(a) Let X = Y . Then X +Y = X, and K(t, x) mint, 1|x|. Therefore
X (X, X)
,p
, 0 < < 1, 1 p .
In the next proposition we will see that for any interpolation couple we have (X, Y )
,p
X +Y . So, if X = Y then (X, X)
,p
= X.
(b) If X Y = 0, then for each x X + Y there are a unique a X and a unique
b Y such that x = a + b, hence K(t, x) = |a|
X
+ t|b|
Y
and t t
K(t, x) does
not belong to any L
p
(0, ), unless x = 0. Therefore, (X, Y )

,p
= (X, Y )
= 0 for
every p [1, ], (0, 1).
(c) In the important case where Y X we have K(t, x) |x|
X
for every x X, so that
t t
K(t, x) L
p
(a, ) for all a > 0, and lim

t+
t
K(t, x) = 0. Therefore,
only the behavior near t = 0 of t
K(t, x) plays a role in the denition of (X, Y )

,p
and of (X, Y )
. Indeed, one could replace the haline (0, +) by any interval (0, a)
in denition 1.1.2, obtaining equivalent norms.
The inclusion properties of the real interpolation spaces are stated below.
Proposition 1.1.3 For 0 < < 1, 1 p
1
p
2
we have
X Y (X, Y )
,p
1
(X, Y )
,p
2
(X, Y )
(X, Y )
,
X +Y. (1.6)
Moreover,
(X, Y )
,
X Y ,
where X, Y are the closures of X, Y in X +Y .
Proof. Let us show that (X, Y )
,
is contained in XY and it is continuously embedded
in X +Y . For x (X, Y )
,
we have
K(t, x) = inf
x=a+b
|a|
X
+t|b|
Y
t
|x|
,
, t > 0,
so that for every n N (taking t = 1/n) there are a
n
X, b
n
Y such that x = a
n
+b
n
,
and
|a
n
|
X
+
1
n
|b
n
|
Y

2
n
|x|
,
.
In particular, |x b
n
|
X+Y
= |a
n
|
X+Y
|a
n
|
X
2|x|
,
n
, so that the sequence

b
n
goes to x in X + Y as n . This implies that (X, Y )
,
is contained in Y .
Arguing similarly (i.e., replacing 1/n by n and letting n ), or else recalling that
(X, Y )
,
= (Y, X)
1,
) we see that (X, Y )
,
is contained also in X. Moreover, by
denition |x|
X+Y
= K(1, x). Therefore
|x|
X+Y
= K(1, x) |x|
,
, x (X, Y )
,
,
Real interpolation 9
so that (X, Y )
,
is continuously embedded in X +Y .
The inclusion (X, Y )
(X, Y )
,
is obvious because K(, x) is continuous (see exercise
1, 1.1.1) so that t
K(t, x) is bounded in every interval [a, b] with 0 < a < b.

Let us show that (X, Y )
,p
is contained in (X, Y )
and it is continuously embedded in

(X, Y )
,
for p < . For each x (X, Y )
,p
and t > 0, recalling that K(, x) is increasing
we get
t
K(t, x) = (p)
1/p
__
+
t
s
p1
ds
_
1/p
K(t, x)
(p)
1/p
__
+
t
s
p1
K(s, x)
p
ds
_
1/p
, t > 0.
(1.7)
The right hand side is bounded by (p)
1/p
|x|
,p
. Therefore x (X, Y )
,
, and |x|
,

(p)
1/p
|x|
,p
. Changing with 1 and X with Y we obtain |x|
,
((1)p)
1/p
|x|
,p
.
Therefore,
|x|
,
[min, 1 p]
1/p
|x|
,p
. (1.8)
Moreover letting t we get lim
t
t
K(t, x) = 0. To prove that x (X, Y )
we need
also that lim
t0
t
K(t, x) = 0. This can be seen as follows: since (X, Y )

,p
= (Y, X)
1,p
then
0 = lim
t+
t
(1)
K(t, x, Y, X) = lim
t+
t
K(t
1
, x, X, Y ) = lim
0
+
K(, x, X, Y ).
Let us prove that (X, Y )
,p
1
(X, Y )
,p
2
for p
1
< p
2
< +. For x (X, Y )
,p
1
we
have
|x|
,p
2
=
__
+
0
t
p
2
K(t, x)
p
2
dt
t
_
1/p
2
=
__
+
0
t
p
1
K(t, x)
p
1
(t
K(t, x))
p
2
p
1
dt
t
_
1/p
2
__
+
0
t
p
1
K(t, x)
p
1
dt
t
_
1/p
2
_
sup
t>0
t
K(t, x)
_
(p
2
p
1
)/p
2
= (|x|
,p
1
)
p
1
/p
2
(|x|
,
)
1p
1
/p
2
,
and using (1.8) we nd
|x|
,p
2
[min, 1 p
1
]
1/p
1
1/p
2
|x|
,p
1
. (1.9)
Finally, from the inequality K(t, x) min1, t |x|
XY
for every x XY it follows
immediately that X Y is continuously embedded in (X, Y )
,p
for 0 < < 1, 1 p .
The statement is so completely proved.
The rst part of the proof of proposition 1.1.3 shows the connection between inter-
polation theory and approximation theory. Indeed, the sequence b
n
in the proof consists
of elements of Y and converges to x in X + Y . The rate of convergence of b
n
and the
rate of blowing up of |b
n
|
Y
are described precisely by the fact that x (X, Y )
,
(or
x (X, Y )
,p
(X, Y )
,
). In particular, if x (X, Y )
,
then |x b
n
|
X+Y
const.
n
, and |b
n
|
Y
const. n
1
.
If Y X other embeddings hold.
10 Chapter 1
Proposition 1.1.4 If Y X, for 0 <
1
<
2
< 1 we have
(X, Y )
2
,
(X, Y )
1
,1
. (1.10)
Therefore, (X, Y )
2
,p
(X, Y )
1
,q
for every p, q [1, ].
Proof. For x (X, Y )
2
,
we have, using the inequalities K(t, x) |x|
X
for t 1 and
K(t, x) t
2
|x|
2
,
for 0 < t 1,
|x|
1
,1
=
_
1
0
t
1
1
K(t, x)dt +
_
+
1
t
1
1
K(t, x)dt
_
1
0
t
1
1
|x|
2
,
t
2
dt +
_
+
1
t
1
1
|x|
X
dt
1
|x|
2
,
+
1
1
|x|
X
,
(1.11)
and the statement follows since (X, Y )
2
,
X +Y = X because Y X.
Note that (1.10) is not true in general. See next example 1.1.10.
Proposition 1.1.5 For all (0, 1), p [1, ], (X, Y )
,p
is a Banach space. For all
(0, 1), (X, Y )
is a Banach space, endowed with the norm of (X, Y )

,
.
Proof. Let x
n
nN
be a Cauchy sequence in (X, Y )
,p
. By the continuous embedding
(X, Y )
,p
X + Y , x
n
nN
is a Cauchy sequence in X + Y too, so that it converges to
an element x X +Y .
Let us estimate |x
n
x|
,p
. Fix > 0, and let |x
n
x
m
|
,p
for n, m n
. Since
y K(t, y) is a norm in X + Y , for every n, m N and t > 0 we have K(t, x
n
x)
K(t, x
n
x
m
) +K(t, x
m
x), so that
t
K(t, x
n
x) t
K(t, x
n
x
m
) +t
maxt, 1|x
m
x|
X+Y
. (1.12)
Let p = . Then for every t > 0 and n, m n
K(t, x
n
x) +t
maxt, 1|x
m
x|
X+Y
.
Letting m +, we nd t
K(t, x
n
x) for every t > 0. This implies that
x (X, Y )
,
and that x
n
x in (X, Y )
,
. Therefore (X, Y )
,
is complete.
It is easy to see that (X, Y )
is a closed subspace of (X, Y )

,
. Since (X, Y )
,
is
complete, then also (X, Y )
is complete.
Let now p < . Then
|x
n
x|
,p
= lim
0
_
_
1/
t
p1
K(t, x
n
x)
p
dt
_
1/p
.
Due again to (1.12), for every (0, 1) we get, for n, m n
,
_
_
1/
t
p1
K(x
n
x)
p
dt
_
1/p
|x
n
x
m
|
,p
+|x
m
x|
X+Y
_
_
1/
t
p1
maxt, 1dt
_
1/p
+C(, p)|x
m
x|
X+Y
.
Letting rst m and then 0 we get x (X, Y )
,p
and x
n
x in (X, Y )
,p
. So,
(X, Y )
,p
is complete.
The spaces (X, Y )
,p
and (X, Y )
are interpolation spaces, as a consequence of the

following important theorem.
Theorem 1.1.6 Let (X
1
, Y
1
), (X
2
, Y
2
) be interpolation couples. If T L(X
1
, X
2
)
L(Y
1
, Y
2
), then T L((X
1
, Y
1
)
,p
, (X
2
, Y
2
)
,p
) L((X
1
, Y
1
)
, (X
2
, Y
2
)
) for every (0, 1)

and p [1, ]. Moreover,
|T|
L((X
1
,Y
1
)
,p
,(X
2
,Y
2
)
,p
)
(|T|
L(X
1
,X
2
)
)
1
(|T|
L(Y
1
,Y
2
)
)
. (1.13)
Proof. If T is the null operator, there is nothing to prove. If T ,= 0, either |T|
L(X
1
,X
2
)
,= 0
or |T|
L(Y
1
,Y
2
)
,= 0. Assume that |T|
L(X
1
,X
2
)
,= 0. Let x (X
1
, Y
1
)
,p
: then for every
a X
1
, b Y
1
such that x = a +b and for every t > 0 we have
|Ta|
X
2
+t|Tb|
Y
2
|T|
L(X
1
,X
2
)
_
|a|
X
1
+t
|T|
L(Y
1
,Y
2
)
|T|
L(X
1
,X
2
)
|b|
Y
1
_
,
so that, taking the inmum over all a, b as above, we get
K(t, Tx, X
2
, Y
2
) |T|
L(X
1
,X
2
)
K
_
t
|T|
L(Y
1
,Y
2
)
|T|
L(X
1
,X
2
)
, x, X
1
, Y
1
_
. (1.14)
Setting s = t
|T|
L(Y
1
,Y
2
)
|T|
L(X
1
,X
2
)
we get Tx (X
2
, Y
2
)
,p
, and
|Tx|
(X
2
,Y
2
)
,p
|T|
L(X
1
,X
2
)
_
|T|
L(Y
1
,Y
2
)
|T|
L(X
1
,X
2
)
_
|x|
(X
1
,Y
1
)
,p
,
and (1.13) follows. From (1.14) it follows also that
lim
t0
t
K(t, x, X
1
, Y
1
) = lim
t
t
K(t, x, X
1
, Y
1
) = 0 =
= lim
t0
t
K(t, Tx, X
2
, Y
2
) = lim
t
t
K(t, Tx, X
2
, Y
2
) = 0,
that is, T maps (X
1
, Y
1
)
into (X
2
, Y
2
)
.
In the case where |T|
L(X
1
,X
2
)
= 0 we get the result either replacing everywhere
|T|
L(X
1
,X
2
)
by > 0 and then letting 0, or else replacing X
i
by Y
i
for i = 1, 2
and by 1 (see (1.4) and (1.5)).
Taking X
1
= X
2
= X, Y
1
= Y
2
= Y , it follows that (X, Y )
,p
and (X, Y )
are
interpolation spaces. Another important consequence is the next corollary.
Corollary 1.1.7 Let (X, Y ) be an interpolation couple. For 0 < < 1, 1 p there
is c(, p) such that
|y|
(X,Y )
,p
c(, p)|y|
1
X
|y|
Y
y X Y. (1.15)
12 Chapter 1
Proof. Set K = R or K = C, according to the fact that X, Y are real or complex
Banach spaces. Let y X Y , and dene T by T() = y for each K. Then
|T|
L(K,X)
= |y|
X
, |T|
L(K,Y )
= |y|
Y
, and |T|
L(K,(X,Y )
,p
)
= |y|
(X,Y )
,p
. The statement
follows now taking X
1
= Y
1
= K and X
2
= X, Y
2
= Y in theorem 1.1.6, and recalling
that (K, K)
,p
= K.
Another more direct proof is the following: for y X Y 0, we have K(t, y)
min|y|
X
, t|y|
Y
, so that
t
|y|
X
|y|
Y
= K(t, y) t|y|
Y
= t
K(t, y) t
1
|y|
Y
|y|
1
X
|y|
Y
,
and
t
|y|
X
|y|
Y
= K(t, y) |y|
X
= t
K(t, y)
_
|y|
Y
|y|
X
_
|y|
X
= |y|
1
X
|y|
Y
.
Therefore |y|
(X,Y )
,
= sup
t>0
t
K(t, y) |y|
1
X
|y|
Y
, and the statement follows for
p = + with constant c(, ) = 1. For p < + we already know that (X, Y )
,p
is
continuously embedded in (X, Y )
,
, and the proof is complete.
1.1.1 Examples
Let us see some basic examples. C
b
(R
n
) is the space of the bounded continuous functions
in R
n
, endowed with the sup norm ||
; C
1
b
(R
n
) is the subset of the continuously dieren-
tiable functions with bounded derivatives, endowed with the norm |f|
n
i=1
|D
i
f|
.
For (0, 1), C
b
(R
n
) is the set of the bounded and uniformly Holder continuous functions,
endowed with the norm
|f|
C
b
= |f|
+ [f]
C
= |f|
+ sup
x,=y
[f(x) f(y)[
[x y[
.
For (0, 1), p [1, ), W
,p
(R
n
) is the space of all f L
p
(R
n
) such that
[f]
W
,p =
__
R
n
R
n
[f(x) f(y)[
p
[x y[
p+n
dxdy
_
1/p
< .
It is endowed with the norm | |
L
p + [ ]
W
,p.
Example 1.1.8 For 0 < < 1, 1 p < we have
(C
b
(R
n
), C
1
b
(R
n
))
,
= C
b
(R
n
), (1.16)
(L
p
(R
n
), W
1,p
(R
n
))
,p
= W
,p
(R
n
), (1.17)
with equivalence of the respective norms.
Proof. Let us prove the rst statement. Let f (C
b
(R
n
), C
1
b
(R
n
))
,
. Since (X, Y )
,

X + Y , we have |f|
const. |f|
,
. In our case the constant is 1, because for every
decomposition f = a +b, with a C
b
(R
n
), b C
1
b
(R
n
) we have |f|
|a|
+|b|
, so
that
|f|
K(1, f, C
b
(R
n
), C
1
b
(R
n
)) |f|
,
.
Moreover for x ,= y and again for every decomposition f = a + b, with a C
b
(R
n
),
b C
1
b
(R
n
), we have
[f(x) f(y)[ [a(x) a(y)[ +[b(x) b(y)[ 2|a|
+|b|
C
1[x y[,
so that, taking the inmum over all the decompositions,
[f(x) f(y)[ 2K([x y[, f, C
b
(R
n
), C
1
b
(R
n
)) 2[x y[
|f|
,
.
Therefore f is -Holder continuous and |f|
C
= |f|
+ [f]
C
3|f|
,
.
Conversely, let f C
b
(R
n
). Let C
(R
n
) be a nonnegative function with support
in the unit ball, such that
_
R
n
(x)dx = 1. For every t > 0 set
b
t
(x) =
1
t
n
_
R
n
f(y)
_
x y
t
_
dy, a
t
(x) = f(x) b
t
(x), x R
n
. (1.18)
Then
a
t
(x) =
1
t
n
_
R
n
(f(x) f(x y))
_
y
t
_
dy
so that
|a
t
|
[f]
C
1
t
n
_
R
n
[y[
(y/t)dy = t
[f]
C
_
R
n
[w[
(w)dw.
Moreover, |b
t
|
|f|
, and
D
i
b
t
(x) =
1
t
n+1
_
R
n
f(y)D
i
_
x y
t
_
dy.
Since
_
R
n
D
i
((x y)/t)dy = 0, we get
D
i
b
t
(x) =
1
t
n+1
_
R
n
(f(x y) f(x))D
i
_
y
t
_
dy, (1.19)
which implies
|D
i
b
t
|
t
1
[f]
C
_
R
n
[w[
[D
i
(w)[dw.
Therefore,
t
K(t, f) t
(|a
t
|
+t|b
t
|
C
1) C|f|
C
, 0 < t 1.
For t 1 (see remark (c) after denition 1.1.1), we can take a
t
= f, b
t
= 0 which implies
t
K(t, f) t
|f|
|f|
, t 1.
The embedding C
b
(R
n
) (C
b
(R
n
), C
1
b
(R
n
))
,
follows.
The proof of the second statement is similar. We recall that for every b W
1,p
(R
n
)
and h R
n
0 we have (see e.g. [8])
__
R
n
_
[b(x +h) b(x)[
[h[
_
p
dx
_
1/p
|Db|
L
p.
For every f (L
p
(R
n
), W
1,p
(R
n
))
,p
and h R
n
let a = a(h) L
p
(R
n
), b = b(h)
W
1,p
(R
n
) be such that f = a +b, and
|a|
L
p +[h[ |b|
W
1,p 2K([h[, f).
14 Chapter 1
Then
[f(x +h) f(x)[
p
[h[
p+n
2
p1
_
[a(x +h) a(x)[
p
[h[
p+n
+
[b(x +h) b(x)[
p
[h[
p+n
so that
_
R
n
[f(x +h) f(x)[
p
[h[
p+n
dx
2
p1
_
R
n
_
[a(x +h) a(x)[
p
[h[
p+n
+
[b(x +h) b(x)[
p
[h[
p+n
_
dx
2
2p2
|a|
p
L
p
[h[
p+n
+ 2
p1
[h[
p
| [Db[ |
p
L
p
[h[
p+n
C
p
[h[
pn
(|a|
L
p +[h[ |b|
W
1,p)
p
C
p
[h[
pn
K([h[, f).
It follows that
_
R
n
R
n
[f(x +h) f(x)[
p
[h[
p+n
dxdh
C
p
_
R
n
[h[
pn
K([h[, f)
p
dh
= C
p
_

0
K(r, f)
p
r
p+1
dr
_
B(0,1)
d
n1
= C
p,n
|f|
p
,p
.
Therefore, (L
p
(R
n
), W
1,p
(R
n
))
,p
is continuously embedded in W
,p
(R
n
). To be precise, we
have estimated so far only the seminorm [f]
,p
. But we already know that each (X, Y )
,p
is continuously embedded in X + Y ; in our case X + Y = X = L
p
(R
n
) so that we have
also |f|
L
p C|f|
,p
.
To prove the other embedding, for each f W
,p
dene a
t
and b
t
by (1.18). Then
|a
t
|
p
L
p
=
_
R
n
__
R
n
[f(y) f(x)[
1
t
n

_
x y
t
_
dy
_
p
dx
_
R
n
_
R
n
[f(y) f(x)[
p
1
t
n

_
x y
t
_
dydx.
were for p > 1 we applied the Holder inequality to the product
[f(x) f(y)[(t
n
((x y)/t))
1/p
(t
n
((x y)/t))
11/p
.
So we get
_

0
t
p
|a
t
|
p
L
p
dt
t

_

0
__
R
n
R
n
[f(y) f(x)[
p
1
t
n
_
x y
t
_
dy dx
_
dt
t
=
_
R
n
R
n
[f(y) f(x)[
p
__

0
t
p
1
t
n
_
x y
t
_
dt
t
_
dy dx
=
_
R
n
R
n
[f(y) f(x)[
p
__

[xy[
t
p
1
t
n
_
x y
t
_
dt
t
_
dy dx
||
p +n
_
R
n
R
n
[f(y) f(x)[
p
[y x[
p+n
dxdy = C[f]
p
W
,p
.
Using (1.19) and arguing similarly, we get also
_

0
t
(1)p
|D
i
b
t
|
p
L
p
dt
t

C
p1
i
|D
i
|
p +n
[f]
p
W
,p
,
with C
i
=
_
R
n
[D
i
(y)[dy, while |b
t
|
L
p |f|
L
p||
L
1 = |f|
L
p.
Therefore, t
K(t, f, L
p
, W
1,p
) t
|a
t
|
L
p +t
1
|b
t
|
W
1,p L
p
(0, 1), with norm esti-

mated by C|f|
W
,p, and the second part of the statement follows.
Note that the proof of (1.16) yields also
(L
(R
n
), Lip(R
n
))
,
= (BUC(R
n
), BUC
1
(R
n
))
,
= C
b
(R
n
).
We shall see later (3.2) another method to prove (1.16) and (1.17).
Example 1.1.9 Let R
n
be an open set with the following property: there exists an
extension operator E such that E L(C
b
(), C
b
(R
n
)) L(C
1
b
(), C
1
b
(R
n
)), and also, for
some (0, 1), E L(C
b
(), C
b
(R
n
)) (by extension operator we mean that Ef
[
= f(x),
for all f C
b
()). Then
(C
b
(), C
1
b
())
,
= C
b
().
Proof. Theorem 1.1.6 implies that
E L((C
b
(), C
1
b
())
,
, (C
b
(R
n
), C
1
b
(R
n
))
,
).
We know already that (C
b
(R
n
), C
1
b
(R
n
))
,
= C
b
(R
n
). So, for every f (C
b
(), C
1
b
())
,
the extension Ef is in C
b
(R
n
) and |Ef|
C
b
(R
n
)
C|f|
(C
b
(),C
1
b
())
,
. Since f = Ef
[
,
then f C
b
() and |f|
C
b
()
C|f|
(C
b
(),C
1
b
())
,
.
Conversely, if f C
b
() then Ef C
b
(R
n
) = (C
b
(R
n
), C
1
b
(R
n
))
,
. The retraction
operator Rg = g
[
belongs obviously to L(C
b
(R
n
), C
b
()) L(C
1
b
(R
n
), C
1
b
()). Again by
theorem 1.1.6, f = R(Ef) (C
b
(), C
1
b
())
,
, with norm not exceeding C|Ef|
C
b
(R
n
)

C
t
|f|
C
b
()
.
Such a good extension operator exists if is an open set with uniformly C
1
boundary.
is said to be uniformly C
1
if there are N N and a (at most) countable set of balls
B
k
whose interior parts cover , such that the intersection of more than N of these balls
is empty, and dieomorphisms
k
: B
k
B(0, 1) R
n
such that
k
(B
k
) = y
B(0, 1) : y
n
0, and |
k
|
C
1 + |
1
k
|
C
1 are bounded by a constant independent of k.
(In particular, each bounded with C
1
boundary has uniformly C
1
boundary).
It is sucient to construct E when = R
n
+
. The construction of E for any open set
with uniformly C
1
boundary will follow by the usual method of local straightening the
boundary.
If = R
n
+
= x = (x
t
, x
n
) R
n
: x
n
> 0 we may use the reection method: we set
Ef(x) =
_
_
_
f(x), x
n
0,
1
f(x
t
, x
n
) +
2
f(x
t
, 2x
n
), x
n
< 0,
where
1
,
2
satisfy the continuity condition
1
+
2
= 1 and the dierentiability condition
1
2
2
= 1, that is
1
= 3,
2
= 2.
16 Chapter 1
Then E L(C(R
n
+
), C
b
(R
n
)) L(C
(R
n
+
), C
b
(R
n
)) L(C
1
(R
n
+
), C
1
b
(R
n
)), for every
(0, 1).
Let now (, ) be a -nite measure space. To dene the Lorentz spaces L
p,q
() we
introduce the rearrangements as follows. For every measurable f : R or f : C
set
m(, f) = x : [f(x)[ > , 0,
and
f
(t) = inf : m(, f) t, t 0.

Both m(, f) and f
are nonnegative, decreasing (i.e. nonincreasing), right continuous,

and f
, [f[ are equi-measurable, that is for each

0
> 0 we have
[t > 0 : f
(t) >
0
[ = m(
0
, f) = x : [f(x)[ >
0
,
and consequently [t > 0 : f
(t) [
1
,
2
][ = x : [f(x)[ [
1
,
2
], etc. Therefore,
for each p 1,
_
[f(x)[
p
(dx) =
_

0
(f
(t))
p
dt; sup ess [f(x)[ = f
(0) = sup ess f
(t), (1.20)
and for each measurable set E ,
_
E
[f(x)[(dx) =
_
(E)
0
f
(t)dt.
f
is called the nonincreasing rearrangement of f onto (0, ).

The Lorentz spaces L
p,q
() (1 p , 1 q ) are dened by
L
p,q
() =
_
f L
1
() +L
() : |f|
L
p,q =
__

0
(t
1/p
f
(t))
q
dt
t
_
1/q
<
_
,
for q < , and
L
p,
() = f L
1
() +L
() : |f|
L
p, = sup
t>0
t
1/p
f
(t) < .
(For p = we set as usual 1/ = 0).
Note that in general | |
L
p,q is not a norm but only a quasi-norm, i.e. the triangle
inequality is replaced by |f +g| C(|f| +|g|). Moreover, due to (1.20),
L
p,p
() = L
p
(), 1 p .
Example 1.1.10 Let (, ) be a -nite measure space. Then (L
1
(), L
()) is an
interpolation couple. For 0 < < 1, 1 q we have
(L
1
(), L
())
,q
= L
1
1
,q
(). (1.21)
Proof. Let 1 be the space of all measurable, a.e. nitely valued (real or complex) functions
dened in . 1 is a linear topological Hausdor space under convergence in measure on
each measurable E with nite measure (E). Both L
1
() and L
() are continuously
embedded in 1. Therefore (L
1
(), L
()) is an interpolation couple.

The proof of (1.21) is based on the equality
K(t, f, L
1
(), L
()) =
_
t
0
f
(s)ds, t > 0. (1.22)

Once (1.22) is established, (1.21) follows easily. Indeed, since f
is decreasing then
K(t, f) tf
(t), so that for q <

|t
K(t, f)|
q
L
q
=
_

0
t
q
K(t, f)
q
dt
t

_

0
t
q+q
f
(t)
q
dt
t
= |f|
q
L
1/(1),q
()
,
and similarly, for q =
sup
t>0
|t
K(t, f)|
L
sup
t>0
|t
1
f
(t)|
L
= |f|
L
1/(1),
()
.
The opposite inequality follows from the Hardy-Young inequality (A.10) (i) for q < :
|t
K(t, f)|
q
L
q
=
_

0
t
q
__
t
0
sf
(s)
ds
s
_
q
dt
t
q
_

0
s
(1)q
(f
(s))
q
ds
s
= |f|
q
L
1/(1),q
()
,
and from the obvious inequality
|t
K(t, f)|
L
t
_
t
0
ds
s
1
|s
1
f
(s)|
L
=
1
|f|
L
1/(1),
()
for q = .
Let us prove (1.22). To prove , for every f L
1
() +L
() and t > 0, x we set

a(x)
_
_
= f(x) f
(t)
f(x)
[f(x)[
, if [f(x)[ > f
(t),
= 0 otherwise,
b(x) = f(x) a(x).
The function a is dened in such a way that [a(x)[ = [f(x)[ f
(t) if [f(x)[ > f
(t),
[a(x)[ = 0 if [f(x)[ f
(t). Then
|a|
L
1 =
_
E
([f(x)[ f
(t))(dx),
where E = x : [f(x)[ > f
(t) has measure (E) = [s > 0 :> f
(s) > f
(t)[
(because [f[ and f
are equi-measurable) = m(f
(t), f) t, and f
is constant in [(E), t].

Therefore,
|a|
L
1 =
_
(E)
0
(f
(s) f
(t))ds
_
t
0
(f
(s) f
(t))ds.
18 Chapter 1
Moreover,
[b(x)[
_
_
_
= [f(x)[ if [f(x)[ f
(t),
= f
(t) if [f(x)[ > f
(t),
so that
[b(x)[ f
(t) =
1
t
_
t
0
f
(t)ds, x .
Therefore,
K(t, f, L
1
, L
) |a|
L
1 +t|b|
L

_
t
0
f
(s)ds.
To prove the opposite inequality we use the fact that for every decomposition f = a+b
we have (see exercise 7, 1.1.2)
f
(s) a
((1 )s) +b
(s), s 0, 0 < < 1.

Then, if a L
1
(), b L
(), and a +b = f we have

_
t
0
f
(s)ds
_
t
0
a
((1 )s)ds +
_
t
0
b
(s)ds
1
1
_

0
a
()d +tb
(0) =
1
1
_
[a(x)[(dx) +t sup ess [b(x)[.

Letting 0 we get
_
t
0
f
(s)ds |a|
L
1 +t|b|
L
,
so that
K(t, f, L
1
(), L
())
_
t
0
f
(s)ds,
and the statement follows.
1.1.2 Exercises
1) Prove that for every x X + Y , t K(t, x) is concave (and hence continuous) in
(0, ).
2) Prove that x |x|
,p
is a norm in (X, Y )
,p
, for each (0, 1), p [1, ].
3) Take = 0 in denition 1.1.1 and show that (X, Y )
0,p
= (X, Y )
0
= 0, for all
p [1, ). Show that X (X, Y )
0,
.
Take = 1 in denition 1.1.1 and show that (X, Y )
1,p
= (X, Y )
1
= 0, for all
p [1, ). Show that Y (X, Y )
1,
.
4) Following the method of example 1.1.8 show that (C
b
(R
n
), C
1
b
(R
n
))
1,
= Lip(R
n
), and
that for 1 < p < , (L
p
(R
n
), W
1,p
(R
n
))
1,
= W
1,p
(R
n
).
5) Show that for 0 < < 1, C
1
b
(R
n
) is not dense in C
b
(R
n
). Show that (C
b
(R
n
), C
1
b
(R
n
))
is the space of the little Holder continuous functions h
(R
n
), consisting of those bounded
functions f such that
lim
h0
sup
xR
n
[f(x +h) f(x)[
[h[
= 0.
6) Following the method of example 1.1.8 show that
(L
p
(R
n
), W
1,p
(R
n
))
,q
= B
p,q
(R
n
),
dened by B
p,q
(R
n
) = f L
p
(R
n
) : [f]
B
p,q
< , where
[f]
B
p,q
=
__
R
n
__
R
n
[f(x) f(x +h)[
p
dx
_
q/p
1
[h[
q+n
dh
_
1/q
,
and |f|
B
p,q
= |f|
L
p + [f]
B
p,q
.
7) Let be an open set in R
n
such that there exists an extension operator E belonging to
L(L
p
(), L
p
(R
n
)) L(W
,p
(), W
,p
(R
n
)) L(W
1,p
(), W
1,p
(R
n
)), for some p [1, )
and (0, 1). Show that
(L
p
(), W
1,p
())
,p
= W
,p
().
Show that if has uniformly C
1
boundary such extension operator E does exist (see the
remarks after example 1.1.9).
The space W
,p
() is usually dened as the set of the functions f L
p
() such that
[f]
W
,p =
__
[f(x) f(y)[
p
[x y[
p+n
dxdy
_
1/p
< .
8) Let (, ) be any measure space. Prove that for each a, b L
1
() +L
() we have
(a +b)
(s) a
((1 +)s) +b
(s), s 0, 0 < < 1.

(This is used in example 1.1.10). Hint: show preliminarly that
m(
0
+
1
, a +b) m(
0
, a) +m(
1
, b),
0
,
1
0.
9) Prove that for each x X +Y the function K(, x) satises
K(t, x)
t
s
K(s, x) 0 < s < t.
1.2 The trace method
In this section we describe another construction of the real interpolation spaces, which
will be useful for proving other properties, and will let us see the connection between
interpolation theory and trace theory.
We shall use L
p
and Sobolev spaces of functions with values in Banach spaces, whose
denitions and elementary properties are in Appendix A.
Denition 1.2.1 For 0 < < 1 and 1 p dene V (p, , Y, X) as the set of all
functions u : R
+
X +Y such that u W
1,p
(a, b; X +Y ) for every 0 < a < b < , and
t u
(t) = t
u(t) L
p
(0, +; Y ),
t v
(t) = t
u
t
(t) L
p
(0, +; X),
20 Chapter 1
with norm
|u|
V (p,,Y,X)
= |u
|
L
p
(0,+;Y )
+|v
|
L
p
(0,+;X)
.
Moreover, for p = + dene a subspace of V (, , Y, X), by
V
0
(, , Y, X) = u V (, , Y, X) : lim
t0
|t
u(t)|
X
= lim
t0
|t
u
t
(t)|
Y
= 0.
It is not dicult to see that V (p, , Y, X) is a Banach space endowed with the norm
| |
V (p,,Y,X)
, and that V
0
(, , Y, X) is a closed subspace of V (, , Y, X). Moreover any
function belonging to V (p, , Y, X) has a X-valued continuous extension at t = 0. Indeed,
for 0 < s < t from the equality u(t) u(s) =
_
t
s
u
t
()d it follows, for 1 < p < ,
|u(t) u(s)|
X

__
t
s
|
1/p
u
t
()|
p
X
d
_
1/p
__
t
s
(1/p)p
d
_
1/p
|u|
V (p,,Y,X)
[p
t
(1 )]
1/p
(t
p
(1)
s
p
(1)
)
1/p
,
with p
t
= p/(p 1). Arguing similarly, one sees that also if p = 1 or p = , then u is
uniformly continuous near t = 0.
With the aid of corollary A.3.1 we are able to characterize the real interpolation spaces
as trace spaces.
Proposition 1.2.2 For (, p) (0, 1) [1, +], (X, Y )
,p
is the set of the traces at t = 0
of the functions in V (p, 1 , Y, X), and the norm
|x|
Tr
,p
= inf|u|
V (p,1,Y,X)
: x = u(0), u V (p, 1 , Y, X)
is an equivalent norm in (X, Y )
,p
. Moreover, for 0 < < 1, (X, Y )
is the set of the

traces at t = 0 of the functions in V
0
(, 1 , Y, X).
Proof. Let x (X, Y )
,p
. We need to dene a function u V (p, 1 , Y, X) such that
u(0) = x.
For every t > 0 there are a
t
X, b
t
Y such that |a
t
|
X
+t|b
t
|
Y
2K(t, x). It holds
t
1
|b
t
|
Y
2t
K(t, x), and the function t t
K(t, x) is in L
p
(0, +). Moreover, we

already know (see the proof of proposition 1.1.3) that lim
t0
b
t
= x in X + Y . So, the
function t b
t
looks a good candidate for u. But in general it is not measurable with
values in Y , and it is not in W
1,p
loc
(0, ) with values in X. So we have to modify it, and
we proceed as follows.
For every n N let a
n
X, b
n
Y be such that a
n
+b
n
= x, and
|a
n
|
X
+
1
n
|b
n
|
Y
2K(1/n, x).
For t > 0 set
u(t) =
n=1
b
n+1
(
1
n+1
,
1
n
]
(t) =
n=1
(x a
n+1
)
(
1
n+1
,
1
n
]
(t),
where
I
is the characteristic function of the interval I, and
v(t) =
1
t
_
t
0
u(s)ds.
Since (X, Y )
,p
is contained in (X, Y )
,
then t
K(t, x) is bounded, so that lim

t0
K(t, x)
= 0. Therefore, lim
n
|a
n
|
X
= 0, so that |x b
n
|
X+Y
|a
n
|
X
0 as n , and
x = lim
t0
u(t) = lim
t0
v(t) in X +Y . Moreover,
|t
1
u(t)|
Y
t
1
n=1
(
1
n+1
,
1
n
]
(t)2(n + 1)K(1/(n + 1), x) 4t
K(t, x), (1.23)

so that t t
1
u(t) L
p
(0, +; Y ). By Corollary A.3.1, t t

1
v(t) belongs to
L
p
(0, +; Y ), and
|t
1
v|
L
p
(0,+;Y )
4
1
|x|
,p
.
On the other hand,
v(t) = x
1
t
_
t
0
n=1
(
1
n+1
,
1
n
]
(s)a
n+1
ds,
so that v is dierentiable almost everywhere with values in X, and
v
t
(t) =
1
t
2
_
t
0
g(s)ds
1
t
g(t),
where g(t) =
n=1
(
1
n+1
,
1
n
]
(t)a
n+1
is such that
|g(t)|
X

n=1
(
1
n+1
,
1
n
]
(t)2K(1/(n + 1), x) 2K(t, x).
It follows that
|t
1
v
t
(t)| t
sup
0<s<t
|g(s)| +|t
g(t)| 4t
K(t, x). (1.24)

Then t t
1
v
t
(t) belongs to L
p
(0, +; X), and

|t
1
v
t
|
L
p
(0,+;X)
4|x|
,p
.
Therefore, x is the trace at t = 0 of a function v V (p, 1 , Y, X), and
|x|
Tr
,p
2(2 + 1/)|x|
,p
.
If x (X, Y )
, then, by (1.23), lim

t0
t
1
|u(t)|
Y
= 0, so that lim
t0
t
1
|v(t)|
Y
= 0.
By (1.24), lim
t0
t
|g(t)|
X
= 0, so lim
t0
t
1
|v
t
(t)|
X
= 0. Then v V
0
(, 1, Y, X).
Conversely, let x be the trace at t = 0 of a function u V (p, 1 , Y, X). Then
x = x u(t) +u(t) =
_
t
0
u
t
(s)ds +u(t) t > 0,
so that
t
K(t, x) t
1
_
_
_
_
1
t
_
t
0
u
t
(s)ds
_
_
_
_
X
+t
1
|u(t)|
Y
. (1.25)
Corollary A.3.1 implies now that t t
K(t, x) belongs to L
p
(0, +), so that x

(X, Y )
,p
, and
|x|
,p

1
|x|
Tr
,p
.
If x is the trace of a function u V
0
(, 1, Y, X), we may assume without loss of gen-
erality that u vanishes for t large. Then, by (1.25), lim
t0
t
K(t, x) = lim
t
t
K(t, x)
= 0, so that x (X, Y )
.
22 Chapter 1
Example 1.2.3 Choosing X = L
p
(R
n
), Y = W
1,p
(R
n
), and = 1 1/p, 1 < p < ,
we get the following well known characterization of W
11/p,p
(R
n
): W
11/p,p
(R
n
) is the
space of the traces at (x, 0) of the functions (x, t) v(x, t) W
1,p
(R
n+1
+
). Indeed, we
already know that W
11/p,p
(R
n
) = (L
p
(R
n
), W
1,p
(R
n
))
11/p,p
, thanks to to example 1.1.8.
By proposition 1.2.2, W
11/p,p
(R
n
) is the space of the traces at t = 0 of the functions
v V (1/p, p, W
1,p
(R
n
), L
p
(R
n
)). But v V (1/p, p, W
1,p
(R
n
), L
p
(R
n
)) if and only if
the function v(x, t) = v(t)(x) is in W
1,p
(R
n+1
+
). Indeed, concerning measurability, it is
possible to see that a function w : (0, +) L
p
(R
n
)) (resp., w : (0, +) W
1,p
(R
n
))
is measurable in the sense of denition A.1.1 i (t, x) w(t, x) is measurable (resp.,
(t, x) w(t, x) and (t, x) D
i
w(t, x) are measurable for all i = 1, . . . n). Concerning
estimates, the condition t t
1/p
v(t) L
p
((0, +), W
1,p
(R
n
)) means that
_
+
0
_
R
n
_
[v(x, t)[
p
+
n
i=1
[v
x
i
(x, t)[
p
_
dxdt < ,
and the condition t t
1/p
v
t
(t) L
p
((0, +), L
p
(R
n
)) means that v
t
is measurable with
values in L
p
(R
n
)) and
_
+
0
_
R
n
[v
t
(x, t)[
p
dxdt < .
In particular, choosing p = 2 we get that H
1/2
(R
n
) is the space of the traces at (x, 0)
of the functions (x, t) v(x, t) H
1
(R
n+1
+
).
This example shows an important connection between interpolation theory and trace
theory.
Remark 1.2.4 (important) 1. By Proposition 1.2.2, if x (X, Y )
,p
or x (X, Y )
,
then x is the trace at t = 0 of a function u belonging to L
p
(a, b; Y ) W
1,p
(a, b; X) for
0 < a < b. But it is possible to nd a more regular function v V (p, 1 , Y, X) (or
v V
0
(, 1 , Y, X)) such that v(0) = x. For instance we may take
v(t) =
1
t
_
t
0
u(s)ds, t 0.
Then v W
1,p
(a, b; Y ) W
2,p
(a, b; X) for 0 < a < b, v(0) = x, and moreover t t
1
v(t)
belongs to L
p
(0, +; Y ), t t
2
v
t
(t) belongs to L
p
(0, +; Y ), and t t
1
v
t
(t) belongs
to L
p
(0, +; X), with norms estimated by const. |u|

V (p,1,Y,X)
.
Even better, choose any smooth nonnegative function : R
+
R, with compact
support and
_
0
s
1
(s)ds = 1, and set
v(t) =
_

0
_
t
_
u()
d
=
_

0
(s)u
_
t
s
_
ds
s
.
Then v C
(R
+
; X Y ), v(0) = x, and
t t
n
v
(n)
(t) L
p
(0, +; X), n N,
t t
n+1
v
(n)
(t) L
p
(0, +; Y ), n N 0,
with norms estimated by c(n)|u|
V (p,1,Y,X)
. If in addition p = and x (X, Y )
then
lim
t0
t
n
|v
(n)
(t)|
X
= 0, n N,
lim
t0
t
n+1
|v
(n)
(t)|
Y
= 0. n N 0,
2. Let x X + Y be the trace at t = 0 of a function v V (p, 1 , Y, X). Fix any
C
0
([0, +)) such that 1 in a right neighborhood of 0, say in (0, 1]. The function
t (t)v(t) is in V (p, 1 , Y, X), its norm does not exceed C|v|
V (p,1,Y,X)
, with C
depending only on , and its trace at t = 0 is still x. Moreover, it has compact support in
[0, +). This shows that in the dention of the trace spaces one could consider just the
subset of V (p, 1 , Y, X) consisting of the functions with compact support, obtaining an
equivalent trace space (i.e., the same space with an equivalent norm).
By means of the trace method it is easy to prove some important density properties.
Proposition 1.2.5 Let 0 < < 1. For 1 p < , X Y is dense in (X, Y )
,p
. For
p = , (X, Y )
is the closure of X Y in (X, Y )

,
.
Proof. Let p < , and let x (X, Y )
,p
. By Remark 1.2.4, x = v(0), where v
C
((0, ); X Y ) V (p, 1 , Y, X), and moreover t t

2
v
t
L
p
(0, +; Y ). Set
x
= v(), > 0.
Then x
X Y , and we shall show that x
x in (X, Y )
,p
.
We have x
x = z
(0), where
z
(t) = (v() v(t))

[0,]
(t).
It is not hard to check that z
W
1,p
(a, b; X) for 0 < a < b < , and that z
t
(t) =
v
t
(t)
(0,)
(t). It follows that
lim
0
|t
1
z
t
(t)|
L
p
(0,+;X)
= 0.
Moreover, due to the equality
z
(t) =
_
+
t
(0,)
(s)v
t
(s)ds,
we get, using the Hardy-Young inequality (A.10)(ii),
|t
1
z
(t)|
L
p
(0,+;Y )

__
+
0
t
(1)p
__
+
t
(0,)
(s)s|v
t
(s)|
Y
ds
s
_
p
dt
t
_
1/p
1
1
__
+
0
(0,)
(s)s
(2)p
|v
t
(s)|
Y
ds
s
_
1/p
,
so that t t
1
z
(t) L
p
(0, +; Y ) for every , and

lim
0
|t
1
z
(t)|
L
p
(0,+;Y )
= 0.
Therefore, z
0 in V (p, 1 , Y, X) as 0, which means that |x
x|
T
,p
0 as
0. From Proposition 1.2.2 we get lim
0
|x
x|
,p
= 0.
Let now x (X, Y )
. Due again to remark 1.2.4, x is the trace at t = 0 of a function

v V
0
(, 1 , Y, X), such that t t
2
v
t
(t) L
(0, +; Y ) and lim

t0
t
2
|v
t
(t)|
Y
= 0. Let x
, z
be dened as above. Then lim

t0
t
1
|v
t
(t)|
X
= 0, so that
sup
t>0
t
1
|z
t
(t)|
X
= sup
0<t
t
1
|v
t
(t)|
X
0, as 0,
24 Chapter 1
and
sup
t>0
t
1
|z
(t)|
Y
= sup
0<t
t
1
|v() v(t)|
Y
2 sup
0<s
s
1
|v(s)|
Y
0,
as 0. Arguing as before, it follows that |x
x|
,
0 as 0.
1.2.1 Exercises
1) Prove the statements of remark 1.2.4.
2) Prove that (X, Y )
,p
is the set of the elements x X + Y such that x = u(t) + v(t)
for almost all t > 0, with t t
u(t) L
p
(0, ; X), t t
1
u(t) L
p
(0, ; Y ), and the

norm
x inf
x=u(t)+v(t)
|t
u(t)|
L
p
(0,;X)
+|t
1
v(t)|
L
p
(0,;Y )
is equivalent to the norm of (X, Y )
,p
.
3) Prove that (X, Y )
,p
is the set of the elements x X + Y such that x =
_
0
u(t)dt/t,
with t t
u(t) L
p
(0, ; X), t t
1
u(t) L
p
(0, ; Y ), and the norm

x inf
x=
R

0
u(t)dt/t
|t
u(t)|
L
p
(0,;X)
+|t
1
v(t)|
L
p
(0,;Y )
is equivalent to the norm of (X, Y )
,p
.
Hint: use remark 1.2.4 to write any x (X, Y )
,p
as x =
_
0
v(t)dt, as in the proof
of next proposition 1.3.2.
4) Let R
n
be a bounded open set with C
1
boundary. Prove that for 1 < p < , the
space W
11/p,p
() is the space of the traces on of the functions in W
1,p
().
1.3 Intermediate spaces and reiteration
Let us introduce two classes of intermediate spaces for the interpolation couple (X, Y ).
Denition 1.3.1 Let 0 1, and let E be a Banach space such that X Y E
X +Y .
(i) E is said to belong to the class J
between X and Y if there is a constant c such that

|x|
E
c|x|
1
X
|x|
Y
, x X Y.
In this case we write E J
(X, Y ).
(ii) E is said to belong to the class K
between X and Y if there is k > 0 such that

K(t, x) kt
|x|
E
, x E, t > 0.
In this case we write E K
(X, Y ).
If (0, 1) this means that E is continuously embedded in (X, Y )
,
.
A useful characterization of J
(X, Y ) is the following one.

Proposition 1.3.2 Let 0 < < 1, and let E be a Banach space such that X Y E
X +Y . The following statements are equivalent:
(i) E belongs to the class J
between X and Y ,
(ii) (X, Y )
,1
E.
Proof. The implication (ii) (i) is a straightforward consequence of Corollary 1.1.7, with
p = 1. Let us show that (i) (ii). For every x (X, Y )
,1
, let u V (1, 1 , Y, X) be
such that u(t) vanishes for large t, u(0) = x, and set
v(t) =
1
t
_
t
0
u(s)ds.
By remark 1.2.4, t t
2
v
t
(t) belongs to L
1
(0, +; Y ), and t t
1
v
t
(t) belongs to
L
1
(0, +; X). Moreover v(0) = x, v(+) = 0, so that

x =
_
+
0
v
t
(t)dt.
Let c be such that |y|
E
c|y|
Y
|y|
1
X
for every y X Y . Then
|v
t
(t)|
E
c|v
t
(t)|
Y
|v
t
(t)|
1
X
= ct
1
|t
2
v
t
(t)|
Y
|t
1
v
t
(t)|
1
X
.
Since the function t |t
2
v
t
(t)|
Y
belongs to L
1/
(0, +) and t |t
1
v
t
(t)|
1
X
belongs
to L
1/(1)
(0, +), by the Holder inequality v

t
L
1
(0, +; E), and
|x|
E
c(|t
2
v
t
(t)|
L
1
(0,;Y )
)
(|t
1
v
t
(t)|
L
1
(0,;X)
)
1
const. |x|
,1
.
By Denition 1.3.1 and Proposition 1.2.5, if 0 < < 1 a space E belongs to K
(X, Y )
J
(X, Y ) if and only if

(X, Y )
,1
E (X, Y )
,
.
In particular, (X, Y )
,p
and (X, Y )
are in K
(X, Y ) J
(X, Y ), for every p [1, ]. We

shall see in chapter 2 that also the complex interpolation spaces [X, Y ]
are in K
(X, Y )
J
(X, Y ).
But there are also intermediate spaces belonging to K
(X, Y ) J
(X, Y ) which are

not interpolation spaces.
Example 1.3.3 C
1
b
(R
n
) J
1/2
(C
b
(R
n
), C
2
b
(R
n
)) K
1/2
(C
b
(R
n
), C
2
b
(R
n
)). But C
1
b
(R
n
)
is not an interpolation space between C
b
(R
n
) and C
2
b
(R
n
).
Proof. From the inequalities (i = 1, . . . , n)
[f(x +he
i
) f(x) D
i
f(x)h[
1
2
|D
ii
f|
h
2
, x R
n
, h > 0,
we get
[D
i
f(x)[
[f(x +he
i
) f(x)[
h
+
1
2
|D
ii
f|
h, x R
n
, h > 0,
26 Chapter 1
so that
|D
i
f|

2|f|
h
+
1
2
|D
ii
f|
h, h > 0.
Taking the minimum on h over (0, +) we get
|D
i
f|
2(|f|
)
1/2
(|D
ii
f|
)
1/2
, f C
2
b
(R
n
)
so that
|f|
C
1
b
(|f|
)
1/2
_
(|f|
)
1/2
+ 2
n
i=1
(|D
ii
f|
)
1/2
_
C(|f|
)
1/2
(|f|
C
2
b
)
1/2
.
This implies that C
1
b
(R
n
) belongs to J
1/2
(C
b
(R
n
), C
2
b
(R
n
)). To prove that it belongs also to
K
1/2
(C
b
(R
n
), C
2
b
(R
n
)), namely that it is continuously embedded in (C
b
(R
n
), C
2
b
(R
n
))
1/2,
,
we argue as in example 1.1.8: for every f C
1
b
(R
n
) the functions a
t
, b
t
dened in (1.18)
are easily seen to satisfy
|a
t
|
Ct[f]
Lip
, |b
t
|
C
1
b
C|f|
C
1
b
, |D
ij
b
t
|
Ct
1
[f]
Lip
.
Therefore, K(t, f, C
b
(R
n
), C
2
b
(R
n
)) |a
t
1/2|
+t|b
t
1/2|
C
2
b
Ct
1/2
|f|
C
1
b
so that C
1
b
(R
n
)
is in K
1/2
(C
b
(R
n
), C
2
b
(R
n
)).
But C
1
b
(R
n
) is not a real interpolation space between C
b
(R
n
) and C
2
b
(R
n
), even for
n = 1. More precisely, there does not exist any constant C such that |T|
L(C
1
b
(R))

C(|T|
L(C
2
b
(R))
)
1/2
(|T|
L(C
b
(R))
)
1/2
for all T L(C
2
b
(R)) L(C
b
(R)).
Indeed, consider the family of operators
(T
f)(x) =
_
1
1
x
x
2
+y
2
+
2
(f(y) f(0)) dy, x R.
It is easy to see that |T
|
L(C
b
(R))
and |T
|
L(C
2
b
(R))
are bounded by a constant independent
of . Indeed, for every continuous and bounded f,
[(T
f)(x)[ 2
_
1
1
[x[
x
2
+y
2
+
2
|f|
dy 2|f|
,
(T
f)
t
(x) =
_
1
1
x
2
+y
2
+
2
(x
2
+y
2
+
2
)
2
(f(y) f(0))dy,
and for every f C
1
b
(R),
(T
f)
tt
(x) =
_
1
1
2x(x
2
+ 3y
2
+ 3
2
)
(x
2
+y
2
+
2
)
3
_
y
0
f
t
(s)ds dy
=
_
1
1
2x(x
2
+ 3y
2
+ 3
2
)
(x
2
+y
2
+
2
)
3
_
y
0
(f
t
(s) f
t
(0))ds dy,
so that, if f C
2
b
(R),
[(T
f)
tt
(x)[ [x[
_
1
1
x
2
+ 3y
2
+ 3
2
(x
2
+y
2
+
2
)
2
dy |f
tt
|
3|f
tt
|
.
On the contrary, choosing f
(x) = (x
2
+
2
)
1/2
(x), with C
0
(R), 1 in [1, 1], we
get
(T
)
t
(0) =
_
1
1
(y
2
+
2
)
1/2
y
2
+
2
dy =
1
_
1/
1/
(s
2
+ 1)
1/2
1
s
2
+ 1
ds
so that lim
0
(T
)
t
(0) = +, while the C
1
b
norm of f
is bounded by a constant inde-

pendent of . Therefore |T
|
L(C
1
b
(R))
blows up as 0. By theorem 1.1.6, C
1
b
(R) cannot
be a real interpolation space between C
b
(R) and C
2
b
(R).
This counterexample is due to Mitjagin and Semenov, it shows also that C
1
([1, 1]) is
not a real interpolation space between C([1, 1]) and C
2
([1, 1]), and it may be obviously
adapted to show that for any dimension n, C
1
b
(R
n
) is not a real interpolation space between
C
b
(R
n
) and C
2
b
(R
n
).
Remark 1.3.4 Arguing similarly one sees that C
k
b
(R
n
) is in J
1/2
(C
k1
b
(R
n
), C
k+1
b
(R
n
))
K
1/2
(C
k1
b
(R
n
), C
k+1
b
(R
n
)), for every k N. It follows easily that for m
1
< k < m
2

N, C
k
b
(R
n
) belongs to J
(km
1
)/(m
2
m
1
)
(C
m
1
b
(R
n
), C
m
2
b
(R
n
)). For instance, knowing that
C
1
b
(R
n
) belongs to J
1/2
(C
b
(R
n
), C
2
b
(R
n
)) and C
2
b
(R
n
) belongs to J
1/2
(C
1
b
(R
n
), C
3
b
(R
n
)) one
gets, for every f C
3
b
(R
n
),
|f|
C
1
b
C|f|
1/2
|f|
1/2
C
2
b
C
t
|f|
1/2
(|f|
1/2
C
1
b
|f|
1/2
C
3
b
)
1/2
so that |f|
3/4
C
1
b
C
t
|f|
1/2
|f|
1/4
C
3
b
, which implies
|f|
C
1
b
C
tt
|f|
2/3
|f|
1/3
C
3
b
that is, C
1
b
(R
n
) belongs to J
1/3
(C
b
(R
n
), C
3
b
(R
n
)).
Now we are able to state the Reiteration Theorem. It is one of the main tools of
general interpolation theory.
Theorem 1.3.5 Let 0
0
,
1
1,
0
,=
1
. Fix (0, 1) and set = (1 )
0
+
1
.
The following statements hold true.
(i) If E
i
belong to the class K
i
(i = 0, 1) between X and Y , then
(E
0
, E
1
)
,p
(X, Y )
,p
, p [1, ], (E
0
, E
1
)
(X, Y )
.
(ii) If E
i
belong to the class J
i
(i = 0, 1) between X and Y , then
(X, Y )
,p
(E
0
, E
1
)
,p
, p [1, ], (X, Y )
(E
0
, E
1
)
.
Consequently, if E
i
belong to K
i
(X, Y ) J
i
(X, Y ), then
(E
0
, E
1
)
,p
= (X, Y )
,p
, p [1, ], (E
0
, E
1
)
= (X, Y )
,
28 Chapter 1
Proof. Without loss of generality (recalling that (E
1
, E
0
)
,p
= (E
0
, E
1
)
1,p
and (E
1
, E
0
)
=
(E
0
, E
1
)
1
) we may assume that
0
<
1
.
Let us prove statement (i). Let k
i
be such that K(t, x) k
i
t
i
|x|
E
i
for every x E
i
,
i = 0, 1. For each x (E
0
, E
1
)
,p
, let a E
0
, b E
1
be such that x = a +b. Then
K(t, x, X, Y ) K(t, a, X, Y ) +K(t, b, X, Y ) k
0
t
0
|a|
E
0
+k
1
t
1
|b|
E
1
.
Since a and b are arbitrary, it follows that
K(t, x, X, Y ) maxk
0
, k
1
t
0
K(t
0
, x, E
0
, E
1
).
Consequently,
t
K(t, x, X, Y ) maxk
0
, k
1
t
(
1
0
)
K(t
0
, x, E
0
, E
1
). (1.26)
By the change of variable s = t
0
we see that t t
K(t, x, X, Y ) is in L
p
(0, +),
which means that x belongs to (X, Y )
,p
, and
_
_
_
|x|
(X,Y ),p
maxk
0
, k
1
(
1
0
)
1/p
|x|
(E
0
,E
1
)
,p
, if p < ,
|x|
(X,Y ),
maxk
0
, k
1
|x|
(E
0
,E
1
)
,p
, if p = .
If x (E
0
, E
1
)
, by (1.26) we get
lim
t0
t
K(t, x, X, Y ) maxk
0
, k
1
lim
s0
s
K(s, x, E
0
, E
1
) = 0,
and
lim
t+
t
K(t, x, X, Y ) maxk
0
, k
1
lim
s+
s
K(s, x, E
0
, E
1
) = 0,
so that x (X, Y )
.
Let us prove statement (ii). By proposition 1.2.2 and remark 1.2.4, every x (X, Y )
,p
is the trace at t = 0 of a regular function v : R
+
X Y such that v(+) = 0,
t t
1
v
t
(t) belongs to L
p
(0, +, X), t t
2
v
t
(t) belongs to L
p
(0, +, Y ), and
|t
1
v
t
(t)|
L
p
(0,+,X)
+|t
2
v
t
(t)|
L
p
(0,+,Y )
k|x|
Tr
(X,Y ),p
,
with k independent of x and v. We shall show that the function
g(t) = v(t
1/(
1
0
)
), t > 0,
belongs to V (p, 1 , E
0
, E
1
): since g(0) = x, this will imply, through proposition 1.2.2,
that x (E
0
, E
1
)
,p
.
To this aim we preliminarly estimate |v
t
(t)|
E
i
, i = 0, 1. Let c
i
be such that
|y|
E
i
c
i
|y|
1
i
X
|y|
i
Y
y Y, i = 0, 1.
Then
|v
t
(s)|
E
i

c
i
s
i
+1
|s
1
v
t
(s)|
1
i
X
|s
2
v
t
(s)|
i
Y
, i = 0, 1,
so that from the equalities
0
+ 1 = 1 (
1
0
),
1
+ 1 = 1 + (1 )(
1
0
),
we get
_
_
(i) |s
1(
1
0
)
v
t
(s)|
L
p
(0,+;E
0
)
c
0
k |x|
Tr
(X,Y ),p
,
(ii) |s
1+(1)(
1
0
)
v
t
(s)|
L
p
(0,+;E
1
)
c
0
k |x|
Tr
(X,Y ),p
.
(1.27)
From the equality v(t) =
_
t
v
t
(s)ds and 1.27(ii), using the Hardy-Young inequality
(A.10)(ii) if p < , we get
|t
(1)(
1
0
)
v(t)|
L
p
(0,+;E
1
)

c
0
k
(1 )(
1
0
)
|x|
Tr
(X,Y ),p
.
It follows that t t
1
g(t) L
p
(0, +; E
1
), and
|t
1
g(t)|
L
p
(0,+;E
1
)
(
1
0
)
1/p
|t
(1)(
1
0
)
v(t)|
L
p
(0,+;E
1
)
.
Moreover g
t
(t) = (
1
0
)
1
t
1+1/(
1
0
)
v
t
(t
1/(
1
0
)
), so that, by (1.27)(i), t t
1
g
t
(t)
= (
1
0
)
1
t
(1(
1
0
))/(
1
0
)
v
t
(t
1/(
1
0
)
) L
p
(0, +; E
0
), and
|t
1
g
t
(t)|
L
p
(0,+;E
0
)
(
1
0
)
11/p
|t
1(
1
0
)
v
t
(t)|
L
p
(0,+;E
0
)
.
Therefore, g V (p, 1 , E
0
, E
1
), so that x = g(0) belongs to (E
0
, E
1
)
,p
, and
|x|
T
(E
0
,E
1
)
,p
(
1
0
)
11/p
k|x|
T
(X,Y ),p
.
If x (X, Y )
, then (1.27)(i) has to be replaced by

lim
s0
s
1(
1
0
)
|v
t
(s)|
E
0
= 0,
so that
lim
t0
t
1
|g
t
(t)|
E
0
= lim
t0
t
+1/(
1
0
)
0
|v
t
(t
1/(
1
0
)
)|
E
0
= 0.
Similarly, (1.27)(ii) has to be replaced by
lim
s0
s
1+(1)(
1
0
)
|v
t
(s)|
E
1
= 0.
Using the equality
t
1
g(t) = t
1
_

1/(
1
0
)
t
1/(
1
0
)
v
t
(s)ds +
t
1
1
_
1
_
+
1/(
1
0
)
v
t
(s)ds
_
,
which holds for 0 < t < , one deduces that lim
t0
t
1
|g(t)|
E
1
= 0.
Remark 1.3.6 The assumption
0
,=
1
is not removable. Consider for instance the case
E
0
= E
1
= (X, Y )
0
,
, which is in K
0
(X, Y ). If statement (i) of the theorem would be
true for
0
=
1
then (X, Y )
0
,
(X, Y )
0
,p
for each p < , which is false in general.
Remark 1.3.7 By proposition 1.2.3, (X, Y )
,p
and (X, Y )
are in K
(X, Y ) J
(X, Y )
for 0 < < 1 and 1 p . The Reiteration Theorem yields
((X, Y )
0
,q
0
, (X, Y )
1
,q
1
)
,p
= (X, Y )
(1)
0
+
1
,p
,
((X, Y )
0
, (X, Y )
1
,q
)
,p
= (X, Y )
(1)
0
+
1
,p
,
((X, Y )
0
,q
, (X, Y )
1
)
,p
= (X, Y )
(1)
0
+
1
,p
,
30 Chapter 1
for 0 <
0
,
1
< 1, 1 p, q . Moreover, since X belongs to K
0
(X, Y ) J
0
(X, Y ), and
Y belongs to K
1
(X, Y ) J
1
(X, Y ) between X and Y , then
((X, Y )
0
,q
, Y )
,p
= (X, Y )
(1)
0
+,p
, ((X, Y )
0
, Y )
= (X, Y )
(1)
0
+
,
and
(X, (X, Y )
1
,q
)
,p
= (X, Y )
1
,p
, (X, (X, Y )
1
)
= (X, Y )
,
for 0 <
0
,
1
< 1, 1 p, q .
1.3.1 Examples
The following examples are immediate consequences of examples 1.1.8, 1.1.10, 1.1.9 and
remark 1.3.7.
Example 1.3.8 Let 0
1
<
2
1, 0 < < 1. Then
(C
1
b
(R
n
), C
2
b
(R
n
))
,
= C
(1)
1
+
2
b
(R
n
).
If is an open set in R
n
with uniformly C
1
boundary, then
(C
1
b
(), C
2
b
())
,
= C
(1)
1
+
2
b
().
Example 1.3.9 Let 0
1
<
2
1, 0 < < 1, 1 p < . Then
(W
1
,p
(R
n
), W
2
,p
(R
n
))
,p
= W
(1)
1
+
2
,p
(R
n
).
If is an open set in R
n
with uniformly C
1
boundary, then
(W
1
,p
(), W
2
,p
())
,
= W
(1)
1
+
2
,p
().
Example 1.3.10 Let (, ) be a -nite measure space, and x 1 p
0
, p
1
. Then for
each q , 0 < < 1 we have
(L
p
0
,q
0
(), L
p
1
,q
1
())
,q
= L
p,q
(),
with
1
p
=
1
p
0
+

p
1
.
Recalling that L
p,p
() = L
p
(), and taking p
0
= q
0
, p
1
= q
1
, we get
(L
p
0
(), L
p
1
())
,q
= L
p,q
(),
1
p
=
1
p
0
+

p
1
.
In particular,
(L
p
0
(), L
p
1
())
,q
= L
q,q
() = L
q
() for =
_
1
p
0
q
__
1
p
0
p
1
_
1
.
Another generalization of example 1.1.8 is the following.
Example 1.3.11 For 0 < < 1, 1 p, q < , m N,
(L
p
(R
n
), W
m,p
(R
n
))
,q
= B
m
p,q
(R
n
).
Here B
s
p,q
(R
n
) is the Besov space dened as follows: if s is not an integer, let [s] and
s be the integral and the fractional parts of s, respectively. Then B
s
p,q
(R
n
) consists of
the functions f W
[s],p
(R
n
) such that
[f]
B
s
p,q
=
[[=[s]
__
R
n
dh
[h[
n+sq
__
R
n
[D
f(x +h) D
f(x)[
p
dx
_
q/p
_
1/q
is nite. In particular, for p = q we have B
s
p,p
(R
n
) = W
s,p
(R
n
).
If s = k N, then B
k
p,q
(R
n
) consists of the functions f W
k1,p
(R
n
) such that
[f]
B
k
p,q
=
[[=[s]1
__
R
n
dh
[h[
n+q
__
R
n
[D
f(x + 2h) 2D
f(x +h) +D
f(x)[
p
dx
_
q/p
_
1/q
is nite. For m = 1 see exercise 5, 1.1.2. For the complete proof see [36, 2.3, 2.4].
1.3.2 Applications. The theorems of Marcinkiewicz and Stampacchia
Let (, ), (, ) be two -nite measure spaces. Traditionally, a linear operator T :
L
1
() +L
() L
1
() +L
() is said to be of weak type (p, q) if there is M > 0 such

that
sup
>0
(y : [Tf(y)[ > )
1/q
M|f|
L
p
()
,
for all f L
p
(). This is equivalent to say that the restriction of T to L
p
() is a
bounded operator from L
p
() to L
q,
(). Indeed, by the properties of the nonincreasing
rearrangements,
sup
>0
(y : [g(y)[ > )
1/q
= sup
t>0
t
1/q
g
(t) = |g|
L
q,. (1.28)
T is said to be of strong type (p, q) if its restriction to L
p
() is a bounded operator
from L
p
() to L
q
().
Since L
q
() = L
q,q
() L
q,
(), then any operator of strong type (p, q) is also of
weak type (p, q).
Theorem 1.3.12 Let T : L
1
() +L
() L
1
() +L
() be of weak type (p
0
, q
0
) and
(p
1
, q
1
), with constants M
0
, M
1
respectively, and
1 p
0
, p
1
, 1 < q
0
, q
1
,
q
0
,= q
1
, p
0
q
0
, p
1
q
1
.
For every (0, 1) dene p and q by
1
p
=
1
p
0
+

p
1
,
1
q
=
1
q
0
+

q
1
.
Then T is of strong type (p, q), and there is C independent of such that
|Tf|
L
q
()
CM
1
0
M
1
|f|
L
p
()
, f L
p
().
32 Chapter 1
Proof. For i = 1, 2, T is bounded from L
p
i
() to L
q
i
,
(), with norm not exceed-
ing CM
i
. By the interpolation theorem 1.1.6, T is bounded from (L
p
0
(), L
p
1
())
,p
to
(L
q
0
,
(), L
q
1
,
())
,p
, and
|T|
(L
p
0(),L
p
1())
,p
,(L
q
0
,
(),L
q
1
,
())
,p
)
CM
1
0
M
1
.
On the other hand, we know from example 1.3.10 that
(L
p
0
(), L
p
1
())
,p
= L
p,p
() = L
p
(),
and
L
q
i
,
() = (L
1
(), L
())
11/q
i
,
, i = 1, 2
(it is here that we need q
i
> 1: L
1,
() is not a real interpolation space between L
1
()
and L
()), so that by the Reiteration Theorem

(L
q
0
,
(), L
q
1
,
())
,p
= (L
1
(), L
())
(1)(11/q
0
)+(11/q
1
),p
= (L
1
(), L
())
11/q,p
.
The last space is nothing but L
q,p
(), again by example 1.1.10. Since p
0
q
0
and p
1
q
1
then p q, so that L
q,p
() L
q,q
() = L
q
(). It follows that T is bounded from L
p
()
to L
q
(), with norm not exceeding C
t
M
1
0
M
1
.
Theorem 1.3.12 is slightly less general than the complete Marcinkiewicz Theorem,
which holds also for q
0
or q
1
= 1.
Since every T of strong type (p, q) is also of weak type (p, q) we may recover a part of
the RieszThorin Theorem from theorem 1.3.12: we get that if T is of strong type (p
i
, q
i
),
i = 0, 1, with p
i
, q
i
subject to the restrictions in theorem 1.3.12, then T is of strong type
(p, q). The full RieszThorin Theorem will be proved in Chapter 2.
Let f be a locally integrable function. For each measurable subset A with positive
measure (A) we dene the mean value of f on A by
f
A
=
1
(A)
_
A
f(x)(dx).
The space BMO() (BMO stands for bounded mean oscillation) consists of those locally
integrable functions f such that
sup
0<(A)<
1
(A)
_
A
[f(x) f
A
[(dx) < .
If () < we see immediately that L
() BMO() L
1
().
It is possible to show that if R
n
is a Lipschitz bounded domain, then BMO() is
contained in each L
p
(), 1 p < , and that the norms
|f|
L
p + [f]
BMO,p
= |f|
L
p + sup
A
_
1
(A)
_
A
[f(x) f
A
[
p
dx
_
1/p
, 1 p <
are equivalent in BMO(). This was proved by John and Nirenberg in the well known
paper [24] in the case where is a cube. For such s we may extend the result of example
1.1.10 to the interpolation couple (L
1
(), BMO()).
Example 1.3.13 Let be a bounded domain in R
n
with Lipschitz continuous boundary.
Then for 0 < < 1, 1 q ,
(L
1
(), BMO())
,q
= L
1
1
,q
().
As a consequence we get the main part of the Stampacchia interpolation theorem.
Theorem 1.3.14 Let be a bounded domain in R
n
with Lipschitz continuous boundary,
let 1 r < and let T L(L
r
()) L(L
(), BMO()), or else T L(L

r
())
L(BMO()). Then T L(L
p
()) for every p (r, ), and
|T|
L(L
p
)
C|T|
r/p
L(L
r
)
|T|
1r/p
L(L
,BMO)
in the rst case,
|T|
L(L
p
)
C|T|
r/p
L(L
r
)
|T|
1r/p
L(BMO)
in the second case.
Proof. In the rst case, the interpolation theorem 1.1.6 implies that T L((L
r
, L
)
,p
,
(L
r
, BMO)
,p
) for every (0, 1) and p [1, ], and
|T|
L((L
r
,L
)
,p
,(L
r
,BMO)
,p
)
|T|
1
L(L
r
)
|T|
L(L
,BMO)
.
By example 1.3.10, (L
r
(), L
())
,p
= L
r
1
,p
(). By example 1.3.13 and reiteration,
we still have
(L
r
(), BMO())
,p
= L
r
1
,p
().
Taking = 1 r/p (so that r/(1 ) = p) gives the rst statement through the equality
L
p,p
() = L
p
(). The proof of the second statement is similar.
Campanato and Stampacchia ([15]) used the above interpolation theorem to prove
optimal regularity results for elliptic boundary value problems, as follows.
Let
/ =
n
i,j=1
D
i
(a
ij
(x)D
j
)
be an elliptic operator with L
coecients in a bounded open set R

n
with regular
boundary. If f
ji
, j = 0, . . . , n are in L
2
(), a weak solution of the Dirichlet problem
_
_
_
/u = f in ,
u = 0 in
is any u H
1
0
() such that for every C
0
() it holds
_
i,j=1
a
ij
(x)D
j
u(x)D
i
(x)dx =
_
f
0
(x)(x)dx +
_
j=1
f
j
(x)D
j
(x)dx.
Using the Lax-Milgram theorem, it is not hard to see that the Dirichlet problem has a
unique weak solution u, and |u|
H
1 C
n
j=0
|f
j
|
L
2. This is a rst step in several basic
courses in PDEs; see e.g. [8, Ch. IX] or [1]. The second step is a regularity theorem: if
f
j
= 0 for each j and the coecients a
ij
C
1
(), then u H
2
() and |u|
H
2 C|f
0
|
L
2.
See again [8] or [1].
Moreover Campanato in [14, Thm. 16.II] was able to prove the following:
34 Chapter 1
(i) if the coecients a
ij
are in C
() for some (0, 1) and the functions f

j
are
in BMO(), then each derivative D
i
u belongs to BMO(), and |D
i
u|
BMO,2

C
n
j=0
|f
j
|
BMO,2
;
(ii) if the coecients a
ij
are in C
1+
() for some (0, 1), f
j
= 0 and f
0
BMO(),
then u H
2
() satises the equation a.e., each second order derivative D
ij
u belongs
to BMO(), and |D
ij
u|
BMO,2
C|f
0
|
BMO,2
.
In case (i), applying theorem 1.3.14 with r = 2 to the operators T
i
, i = 1, . . . , n,
dened by T
i
(f
0
, . . . , f
n
) = D
i
u, u being the solution of the Dirichlet problem, we get
that if the f
j
s are in L
p
(), 2 < p < , then each derivative D
i
u belongs to L
p
(), and
|D
i
u|
L
p C
n
j=0
|f
j
|
L
p.
In case (ii), applying again theorem 1.3.14 with r = 2 to the operators T
ij
, i, j =
1, . . . , n, dened by T
ij
f
0
= D
i
u, we get that if f
0
L
p
(), 2 < p < , then each
derivative D
ij
u belongs to L
p
(), and |D
ij
u|
L
p C|f
0
|
L
p.
1.3.3 Exercises
1) Show that for 0 < < 1, ,= 1/2,
(C
b
(R
n
), C
2
b
(R
n
))
,
= C
2
b
(R
n
).
Hint: prove that (C
1
b
(R
n
), C
2
b
(R
n
))
,
= C
1+
b
(R
n
) using example 1.1.8, then use the
Reiteration Theorem with E = C
1
b
(R
n
).
2) (extreme cases in reiteration)
(a) Using the examples of 1.1.1 nd some interpolation couple (X, Y ) and intermediate
spaces in the classes J
0
, J
1
, K
1
between X and Y that do not coincide with X or Y .
(b) Give a direct proof of statement (ii) of the Reiteration Theorem in the case (
0
,
1
) =
(0, 1).
Chapter 2
Complex interpolation
The complex interpolation method is due to Calderon [13]. It works in complex interpola-
tion couples. It may sound articial compared to the more natural real interpolation
method of chapter 1, see next denitions 2.1.1 and 2.1.3. It is in fact an abstraction and a
generalization of the method used in the proof of the RieszThorin interpolation theorem,
which we show below.
The theorem of RieszThorin. Let (, ), (, ) be -nite measure spaces. Let
1 p
0
, p
1
, q
0
, q
1
and let T : L
p
0
() +L
p
1
() L
q
0
() +L
q
1
() be a linear operator
such that
T L(L
p
0
(), L
q
0
()) L(L
p
1
(), L
q
1
()).
Then
T L(L
p
(), L
q
()), 0 < < 1,

with
1
p
=
1
p
0
+

p
1
,
1
q
=
1
q
0
+

q
1
, (2.1)
and setting M
i
= |T|
L(L
p
i (),L
q
i ())
, i = 0, 1, then
|T|
L(L
p
(),L
q
())
M
1
0
M
1
.
In the case that some of the p
i
s or the q
i
s is the statement still holds if we set as usual
1/ = 0.
Proof We recall that the set of the simple functions (= nite linear combinations of
characteristic functions of measurable sets with nite measure) a : C is dense in
L
p
() for every p [1, +), and the set of the simple functions b : C is dense in
L
q
(), for every q [1, +). Moreover, for each measurable function f : C we have
|f|
L
q
()
= sup
1
|b|
L
q
()
f(x)b(x) (dx)
where b , 0 varies in the set of the simple functions from to C.

Let a : C be a simple function. Then a L
p
() for each p [1, +]; in particular
a L
p
(). To estimate |Ta|L

q
() we use the above characterization of the L

q
-norm,
i.e. we estimate the integral
(Ta)(x)b(x) (dx)
35
36 Chapter 2
where b : C is any simple function. To this aim, for every z S = z = x +iy C :
0 x 1, we dene
f(z)(x) =
_
_
[a(x)[
p
_
1z
p
0
+
z
p
1
_
a(x)
[a(x)[
, if x , a(x) ,= 0,
0, if x , a(x) = 0,
g(z)(x) =
_
_
[b(x)[
q
_
1z
q
0
+
z
q
1
_
b(x)
[b(x)[
, if x , b(x) ,= 0,
0, if x , b(x) = 0.
Then f() = a, g() = b and for each z S, f(z) L
p
() for every p, g(z) L
q
() for
every q. In particular, f(z) L
p
0
() L
p
1
() so that Tf(z) L
q
0
() L
q
1
(), and the
function
F : S C, F(z) =
_
Tf(z) g(z) (dx)

is well dened, holomorphic in the interior of S, continuous and bounded in S, and F() =
_
(Ta)(x)b(x) (dx) is the integral that we want to estimate. For every t R we have
[F(it)[ |Tf(it)|
L
q
0()
|g(it)|
L
q
0()
|T|
L(L
p
0(),L
q
0())
|a|
p
/p
0
L
p
()
|b|
q
/q
0
L
q
()
,
[F(1 +it)[ |Tf(1 +it)|
L
q
1()
|g(1 +it)|
L
q
1()
|T|
L(L
p
1(),L
q
1())
|a|
p
/p
1
L
p
()
|b|
q
/q
1
L
q
()
.
By the three lines theorem (see exercise 2, 2.1.3) we get
[F()[ =
_

Ta b (dx)
(sup
tR
[F(it)[)
1
(sup
tR
[F(1 +it)[)
|T|
1
L(L
p
0(),L
q
0())
|T|
L(L
p
1(),L
q
1())
|a|
L
p
()
|b|
L
q
()
.
Since |Ta|
L
q
()
is the supremum of [F()[/|b|
L
q
()
when b , 0 runs in the set of the
simple functions on , we get
|Ta|
L
q
()
|T|
1
L(L
p
0(),L
q
0())
|T|
L(L
p
1(),L
q
1())
|a|
L
p
()
,
for every simple function a : C. Since the set of such as is dense in L
p
() the
statement follows.
Taking in particular p
i
= q
i
, we get that if T L(L
p
0
(), L
p
0
()) L(L
p
1
(), L
p
1
())
then T L(L
r
(), L
r
()) for every r [p
0
, p
1
].
The crucial part of the proof is the use of the three lines theorem for the function F.
The explicit expression of F is not important; what is important is that F is holomorphic
in the interior of S, continuous and bounded in S, that F() leads to the norm |Ta|
L
q
()
,
and that the behavior of F in iR and in 1+iR is controlled. Banach space valued functions
of this type are precisely those used in the construction of the complex interpolation spaces.
Complex interpolation 37
2.1 Denitions and properties
Throughout the section we shall use the maximum principle for holomorphic functions with
values in a complex Banach space X: if is a bounded open subset of C and f : X
is holomorphic in and continuous in , then |f()|
X
max|f(z)|
X
: z , for
every . This is well known if X = C, and may be recovered for general X by the
following argument. For every let x
t
X
t
be such that |f()|
X
= f(), x
t
) and
|x
t
|
X
= 1. Applying the maximum principle to the complex function z f(z), x
t
) we
get
|f()|
X
= [f(), x
t
)[ max[f(z), x
t
)[ : z
max|f(z)|
X
: z .
The maximum principle holds also for functions dened in strips. Dealing with complex
interpolation, we shall consider the strip
S = z = x +iy C : 0 x 1.
If f : S X is holomorphic in the interior of S, continuous and bounded in S, then for
each S
|f()|
X
maxsup
tR
|f(it)|
X
, sup
tR
|f(1 +it)|
X
.
See exercise 1, 2.1.3.
Let (X, Y ) be an interpolation couple of complex Banach spaces.
Denition 2.1.1 Let S be the strip z = x + iy C : 0 x 1. T(X, Y ) is the space
of all functions f : S X +Y such that
(i) f is holomorphic in the interior of the strip and continuous and bounded up to its
boundary, with values in X +Y ;
(ii) t f(it) C
b
(R; X), t f(1 +it) C
b
(R; Y ), and
|f|
T(X,Y )
= maxsup
tR
|f(it)|
X
, sup
tR
|f(1 +it)|
Y
< .
T
0
(X, Y ) is the subspace of T(X, Y ) consisting of the functions f : S X +Y such that
lim
[t[
|f(it)|
X
= 0, lim
[t[
|f(1 +it)|
Y
= 0.
It is not hard to see that T(X, Y ) and T
0
(X, Y ) are Banach spaces. Indeed, if f
n
is a
Cauchy sequence, the maximum principle gives, for all z S,
|f
n
(z) f
m
(z)|
X+Y
maxsup
tR
|f
n
(it) f
m
(it)|
X+Y
, sup
tR
|f
n
(1 +it) f
m
(1 +it)|
X+Y
maxsup
tR
|f
n
(it) f
m
(it)|
X
, sup
tR
|f
n
(1 +it) f
m
(1 +it)|
Y
.
Therefore for every z S there exists f(z) = lim
n
f
n
(z) in X +Y , and it is easy to see
that f T(X, Y ). Since t f
n
(it) converges in C
b
(R; X) and t f
n
(1 + it) converges
38 Chapter 2
in C
b
(R; Y ), then f
n
converges to f in T(X, Y ). Moreover, since T
0
(X, Y ) is closed in
T(X, Y ), then T
0
(X, Y ) is a Banach space too.
An important technical lemma about the space T
0
(X, Y ) is the following one. Its
intricate proof is due to Calderon [13, p. 132-133]. A very detailed proof is in the book of
KreinPetuninSemenov [28, p. 217-220].
Lemma 2.1.2 The linear hull of the functions e
z
2
+z
a, > 0, R, a X Y , is
dense in T
0
(X, Y ).
The complex interpolation spaces [X, Y ]
are dened through the traces of the functions

in T(X, Y ).
Denition 2.1.3 For every [0, 1] set
[X, Y ]
= f() : f T(X, Y ), |a|

[X,Y ]
= inf
fT(X,Y ), f()=a
|f|
T(X,Y )
.
[X, Y ]
is isomorphic to the quotient space T(X, Y )/^
, where ^
is the subset of
T(X, Y ) consisting of the functions which vanish at z = . Since ^
is closed, the quotient

space is a Banach space and so is [X, Y ]
.
Some immediate consequences of the denition are listed below.
(i) For every (0, 1),
[X, Y ]
= [Y, X]
1
.
(ii) We get an equivalent denition of [X, Y ]
replacing the space T(X, Y ) by the space

T
0
(X, Y ). Indeed, for each f T(X, Y ) and > 0 the function f
(z) = e
(z)
2
f(z)
is in T
0
(X, Y ), f
() = f(), and |f
|
T(X,Y )
maxe
2
, e
(1)
2
|f|
T(X,Y )
. Let-
ting 0 we obtain also
inf
fT(X,Y ), f()=a
|f|
T(X,Y )
= inf
fT
0
(X,Y ), f()=a
|f|
T(X,Y )
.
(iii) If Y = X then [X, X]
= X, with identical norms (see exercise 1, 2.1.3).

(iv) For every t R and for every f T(X, Y ), then f( + it) [X, Y ]
for each
(0, 1), and |f( + it)|
[X,Y ]
= |f()|
[X,Y ]
. (this is easily seen replacing the

function f by g(z) = f(z +it))
Finally, from lemma 2.1.2 it follows that X Y is dense in [X, Y ]
for every (0, 1).

In the present chapter, this fact will be used only in example 2.1.11.
Proposition 2.1.4 Let 0 < < 1. Then
X Y [X, Y ]
X +Y.
Proof Let a X Y . The constant function f(z) = a belongs to T(X, Y ), and
|f|
T(X,Y )
max|a|
X
, |a|
Y
.
Therefore, a = f() [X, Y ]
and |a|
[X,Y ]
|a|
XY
.
The embedding [X, Y ]
X+Y follows again from the maximum principle: if a = f()

with f T(X, Y ) then
|a|
X+Y
maxsup
tR
|f(it)|
X+Y
, sup
tR
|f(1 +it)|
X+Y
maxsup
tR
|f(it)|
X
, sup
tR
|f(1 +it)|
Y
= |f|
T(X,Y )
so that |a|
X+Y
|a|
[X,Y ]
.
Remark 2.1.5 Let 1(X, Y ) be the linear hull of the functions of the type (z)x, with
T
0
(C, C) and x X Y . If a X Y , its [X, Y ]
-norm may be obtained also as

|a|
[X,Y ]
= inf
f1(X,Y ), f()=a
|f|
T(X,Y )
. (2.2)
Proof For each > 0 let f
0
T
0
(X, Y ) be such that f
0
() = a and |f
0
|
T(X,Y )

|a|
[X,Y ]
+. Let z r(z) be a function continuous in S and holomorphic in the interior

of S with values in the unit disk, and such that r() = 0, r
t
() ,= 0, r(z) ,= 0 for z ,= .
For instance, we may take
r(z) =
z
z +
, z S.
Set
f
1
(z) =
f
0
(z) e
(z)
2
a
r(z)
, z S.
Then f
1
T
0
(X, Y ), so that by lemma 2.1.2 there exists a function
f
2
(z) =
n
k=1
exp(
k
z
2
+
k
z)x
k
,
with
k
> 0,
k
R, x
k
X Y , such that |f
1
f
2
|
T(X,Y )
. Set
f(z) = e
(z)
2
a +r(z)f
2
(z), z S.
Then f 1(X, Y ) and
|f|
T(X,Y )
|f
0
|
T(X,Y )
+|f f
0
|
T(X,Y )
|a|
[X,Y ]
+ 2.
Remark 2.1.5 will be used in theorem 2.1.7 and in theorem 4.2.6.
We prove now that the spaces [X, Y ]
are interpolation spaces.

1
, Y
1
), (X
2
, Y
2
) be complex interpolation couples. If a linear oper-
ator T belongs to L(X
1
, X
2
) L(Y
1
, Y
2
), then the restriction of T to [X
1
, Y
1
]
belongs to
L([X
1
, Y
1
]
, [X
2
, Y
2
]
) for every (0, 1). Moreover,

|T|
L([X
1
,Y
1
]
,[X
2
,Y
2
]
)
(|T|
L[X
1
,X
2
]
)
1
(|T|
L[Y
1
,Y
2
]
)
. (2.3)
40 Chapter 2
Proof First let |T|
L(X
1
,X
2
)
,= 0 and |T|
L(Y
1
,Y
2
)
,= 0. If a [X
1
, Y
1
]
, let f T(X
1
, Y
1
)
be such that f() = a. Set
g(z) =
_
|T|
L(X
1
,X
2
)
|T|
L(Y
1
,Y
2
)
_
z
Tf(z), z S.
Then g T(X
2
, Y
2
), and
|g(it)|
X
2
(|T|
L(X
1
,X
2
)
)
1
(|T|
L(Y
1
,Y
2
)
)
|f(it)|
X
1
,
|g(1 +it)|
Y
2
(|T|
L(X
1
,X
2
)
)
1
(|T|
L(Y
1
,Y
2
)
)
|f(1 +it)|
Y
1
,
so that |g|
T(X
2
,Y
2
)
(|T|
L(X
1
,X
2
)
)
1
(|T|
L(Y
1
,Y
2
)
)
|f|
T(X
1
,Y
1
)
. Therefore Ta = g()
[X
2
, Y
2
]
, and
|Ta|
[X
2
,Y
2
]
|g|
T(X
2
,Y
2
)
(|T|
L(X
1
,X
2
)
)
1
(|T|
L(Y
1
,Y
2
)
)
|f|
T(X
1
,Y
1
)
.
Taking the inmum over all f T(X
1
, Y
1
) we get
|Ta|
[X
2
,Y
2
]
(|T|
L[X
1
,X
2
]
)
1
(|T|
L[Y
1
,Y
2
]
)
|a|
[X
1
,Y
1
]
.
If either |T|
L(X
1
,X
2
)
or |T|
L(Y
1
,Y
2
)
vanishes, replace it by > 0 in the denition of g
and then let 0 to get the statement.
If both |T|
L(X
1
,X
2
)
and |T|
L(Y
1
,Y
2
)
vanish, set g(z) = Tf(z).
Theorem 2.1.6 has an interesting extension to linear operators depending on z S.
1
, Y
1
), (X
2
, Y
2
) be complex interpolation couples.
For every z S let T
z
L(X
1
Y
1
, X
2
+ Y
2
) be such that z T
z
x is holomorphic
in S and continuous and bounded in S for every x X
1
Y
1
, with values in X
2
+ Y
2
.
Moreover assume that t T
it
x C(R; L(X
1
, X
2
)), t T
1+it
x C(R; L(Y
1
, Y
2
)) and that
|T
it
|
L(X
1
,X
2
)
, |T
1+it
|
L(Y
1
,Y
2
)
are bounded by a constant independent of t.
Then, setting
M
0
= sup
tR
|T
it
|
L(X
1
,X
2
)
, M
1
= sup
tR
|T
1+it
|
L(Y
1
,Y
2
)
for every (0, 1) we have
|T
x|
[X
2
,Y
2
]
M
1
0
M
1
|x|
[X
1
,Y
1
]
so that T
has an extension belonging to L([X

1
, Y
1
]
, [X
2
, Y
2
]
) (which we still call T
)
satisfying
|T
|
L([X
1
,Y
1
]
,[X
2
,Y
2
]
)
M
1
0
M
1
. (2.4)
Proof The proof is just a modication of the proof of theorem 2.1.6.
Assume rst that M
0
and M
1
are positive. For every a X
1
Y
1
let f 1(X
1
, Y
1
) be
such that f() = a (see remark 2.1.5), and set
g(z) =
_
M
0
M
1
_
z
T
z
f(z), z S.
Then g T(X
2
, Y
2
) and
|g(it)|
X
2
M
1
0
M
1
|f(it)|
X
1
,
|g(1 +it)|
Y
2
M
1
0
M
1
|f(1 +it)|
Y
1
,
so that |g|
T(X
2
,Y
2
)
M
1
0
M
1
|f|
T(X
1
,Y
1
)
. Therefore T
a = g() [X
2
, Y
2
]
, and
|T
a|
[X
2
,Y
2
]
M
1
0
M
1
inf
f1(X
1
,Y
1
), f()=a
|f|
T(X
1
,Y
1
)
,
but by remark 2.1.5,
|a|
[X
1
,Y
1
]
= inf
f1(X
1
,Y
1
), f()=a
|f|
T(X
1
,Y
1
)
,
If either M
0
or M
1
vanishes, replace it by in the denition of g, and then let 0.
If both M
0
and M
1
vanish, dene g by g(z) = T
z
f(z) and follow the above arguments.
Let us come back to theorem 2.1.6 and to its consequences. The same proof of corollary
1.1.7 (through the equality [C, C]
= C with the same norm) yields

Corollary 2.1.8 For every (0, 1) we have
|y|
[X,Y ]
|y|
1
X
|y|
Y
, y X Y. (2.5)
Therefore, [X, Y ]
(X, Y ). This means that (X, Y )

,1
[X, Y ]
, thanks to propo-
sition 1.3.2. It is also true that [X, Y ]
(X, Y )
,
; to prove it we need the following
lemma, which gives a Poisson formula for holomorphic functions in a strip with values in
Banach spaces.
Lemma 2.1.9 For every bounded f : S X which is continuous in S and holomorphic
in S we have
f(z) = f
0
(z) +f
1
(z), z = x +iy S,
where
f
0
(z) =
_
R
e
(yt)
sin(x)
f(it)
sin
2
(x) + (cos(x) exp((y t)))
2
dt,
f
1
(z) =
_
R
e
(yt)
sin(x)
f(1 +it)
sin
2
(x) + (cos(x) + exp((y t)))
2
dt.
(2.6)
Sketch of the proof Let rst X = C. Then (2.6) may be obtained using the Poisson
formula for the unit circle,
f() =
1
2
_
[[=1
f()
1 [[
2
[ [
2
d, [[ < 1,
(which holds for every

f which is holomorphic in the interior and continuous up to the
boundary), and the conformal mapping
(z) =
e
iz
i
e
iz
+i
, z S,
42 Chapter 2
which transforms S into the unit circle. If X is a general Banach space (2.6) follows as
usual, considering the complex functions z f(z), x
t
) for every x
t
X
t
, and applying
(2.6) to each of them.
Proposition 2.1.10 For every (0, 1), [X, Y ]
(X, Y ), that is [X, Y ]
is continu-
ously embedded in (X, Y )
,
.
Proof Let a [X, Y ]
. For every f T(X, Y ) such that f() = a split a = f

0
()+f
1
()
according to (2.6).
Note that for z = x +iy S we have
0 <
_
R
e
(yt)
sin(x)
1
sin
2
(x) + (cos(x) exp((y t)))
2
dt < 1,
0 <
_
R
e
(yt)
sin(x)
1
sin
2
(x) + (cos(x) + exp((y t)))
2
dt < 1.
Indeed, both kernels are positive so that both integrals are positive; moreover if f is
holomorphic in S, continuous and bounded in S and f 1 on iR and on 1+iR then f 1
in S, so that the sum of the integrals is 1.
Therefore for every z S, f
0
(z) X, |f
0
(z)|
X
sup
R
|f(i)|
X
, and f
1
(z) Y ,
|f
1
(z)|
Y
sup
R
|f(1 +i)|
Y
. Then for each t > 0 we get
K(t, a) |f
0
()|
X
+t|f
1
()|
Y
sup
R
|f(i)|
X
+t sup
R
|f(1 +i)|
Y
.
The function g(z) = t
z
f(z) is in T(X, Y ), and g() = a. Applying the above estimate
to g we get
K(t, a) sup
R
|f(i)|
X
t
+t sup
R
|f(1 +i)|
Y
t
1
2t
|f|
T(X,Y )
.
Since f is arbitrary,
K(t, a) 2t
|a|
[X,Y ]
,
By corollary 2.1.8 and proposition 1.1.3, [X, Y ]
(X, Y ) K
(X, Y ).
This implies, through proposition 1.1.4, that [X, Y ]
2
[X, Y ]
1
for
1
<
2
whenever
Y X.
This also allows to use the Reiteration Theorem to characterize the real interpolation
spaces between complex interpolation spaces. We get, for 0 <
1
<
2
< 1, 0 < < 1,
1 p ,
([X, Y ]
1
, [X, Y ]
2
)
,p
= (X, Y )
(1)
1
+
2
,p
.
Further reiteration properties are the following. Calderon ([13]) showed that if one of the
spaces X, Y is continuously embedded in the other one, or if X, Y are reexive and XY
is dense both in X and in Y , then
[[X, Y ]
1
, [X, Y ]
2
]
= [X, Y ]
(1)
1
+
2
,
Lions ([30]) proved that if X and Y are reexive, then for 0 <
1
<
2
< 1, 0 < < 1,
1 < p < ,
[(X, Y )
1
,p
, (X, Y )
2
,p
]
= (X, Y )
(1)
1
+
2
,p
.
The question whether [X, Y ]
coincides with some (X, Y )

,p
has no general answer.
We will see in the next chapter (sect. 3.4) that if X and Y are Hilbert spaces then
[X, Y ]
= (X, Y )
,2
, 0 < < 1,
but in the non hilbertian case there are no general rules. See next examples 2.1.11 and
2.1.12.
2.1.1 Examples
Example 2.1.11 Let (, ) be a measure space with -nite measure, and let 1 p
0
, p
1

, 0 < < 1. Then
[L
p
0
(), L
p
1
()]
= L
p
(),
1
p
=
1
p
0
+

p
1
,
and their norms coincide. (In the case p
0
= or p
1
= the statement is correct if we
set as usual 1/ = 0).
Proof The proof follows the proof of the RieszThorin theorem at the beginning of the
chapter.
We recall that L
p
0
() L
p
1
() is dense both in L
p
() and in [L
p
0
(), L
p
1
()]
.
Let a L
p
0
() L
p
1
(). We may assume without loss of generality that |a|
L
p = 1.
For z S set
f(z)(x) = [a(x)[
p
_
1z
p
0
+
z
p
1
_
a(x)
[a(x)[
, if x , a(x) ,= 0,
f(z)(x) = 0, if x , a(x) = 0.
Then f is continuous in S and holomorphic in the interior of S with values in L
p
0
() +
L
p
1
(), and for each t R
[f(it)(x)[ = [a(x)[
p
p
0
, |f(it)|
L
p
0 |a|
p/p
0
L
p
= 1,
[f(1 +it)(x)[ = [a(x)[
p
p
1
, |f(1 +it)|
L
p
1 |a|
p/p
1
L
p
= 1.
Moreover, t f(it) is continuous with values in L
p
0
() and t f(1 + it) is continuous
with values in L
p
1
(). Therefore, f T(L
p
0
(), L
p
1
()) and |f|
T(L
p
0,L
p
1)
1. Since
f() = a, then
|a|
[L
p
0,L
p
1]
1 = |a|
L
p.
To prove the opposite inequality we remark that
1 = |a|
L
p = sup
_
a(x)b(x)(dx)
: b L
p
0
L
p
1
, |b|
L
p
= 1
_
.
For every b L
p
0
L
p
1
with |b|
L
p
= 1 set, as before
g(z)(x) = [b(x)[
p
_
1z
p
0
+
z
p
1
_
b(x)
[b(x)[
, if x , b(x) ,= 0,
g(z)(x) = 0, if x , b(x) = 0,
44 Chapter 2
and dene, for every f T(L
p
0
, L
p
1
) such that f() = a,
F(z) =
_
f(z)(x)g(z)(x)dx, z S.
Then F is holomorphic in the interior of S and continuous in S, so that the maximum
principle (see exercise 2, 2.1.3) implies that for every z S it holds
[F(z)[ maxsup
tR
[F(it)[, sup
tR
[F(1 +it)[.
But [F(it)[ and [F(1 +it)[ may be easily estimated:
[F(it)[ |f(it)|
L
p
0 |g(it)|
L
p
0
= |f(it)|
L
p
0 |b|
p
/p
0
L
p
= |f(it)|
L
p
0 ,
[F(1 +it)[ |f(1 +it)|
L
p
1 |g(1 +it)|
L
p
1
= |f(1 +it)|
L
p
1 |b|
p
/p
1
L
p
= |f(1 +it)|
L
p
1 ,
so that
[F(z)[ maxsup
tR
|f(it)|
L
p
0 , sup
tR
|f(1 +it)|
L
p
1
|f|
T(L
p
0,L
p
1)
, z S.
Therefore,

a(x)b(x)dx
= [F()[ |f|
T(L
p
0,L
p
1)
.
Since b is arbitrary,
|a|
L
p |f|
T(L
p
0,L
p
1)
.
Since f is arbitrary,
|a|
L
p |a|
[L
p
0,L
p
1]
.
Therefore the identity is an isometry between L
p
0
L
p
1
with the L
p
norm and L
p
0
L
p
1
with the [L
p
0
, L
p
1
]
norm. Since L
p
0
L
p
1
is dense respectively in L
p
and in [L
p
0
, L
p
1
]
,
the statement follows.
Example 2.1.12 For 0 < < 1, 1 < p < , m N,
[L
p
(R
n
), W
m,p
(R
n
)]
= H
m,p
(R
n
).
The proof is in [36, 2.4.2].
We recall that for s > 0
H
s,p
(R
n
) = f L
p
(R
n
) : |f|
H
s,p = |T
1
(1 +[x[
2
)
s/2
Tf|
L
p < ,
where T is the Fourier transform. It is known that if s = k is integer then
H
k,p
(R
n
) = W
k,p
(R
n
), k N.
Moreover it is known that
B
s
p,p
(R
n
) H
s,p
(R
n
) B
s
p,2
(R
n
), 1 < p 2,
B
s
p,2
(R
n
) H
s,p
(R
n
) B
s
p,p
(R
n
), 2 p < ,
and the inclusions are strict if p ,= 2. See [36, 2.3.3]. We recall (example 1.3.11) that
(L
p
(R
n
), W
m,p
(R
n
))
,p
= B
m
p,p
(R
n
). Therefore
[L
p
(R
n
), W
m,p
(R
n
)]
,= (L
p
(R
n
), W
m,p
(R
n
))
,p
unless p = 2.
2.1.2 The theorems of HausdorYoung, RieszThorin, Stein
Applying theorem 2.1.6 to the spaces X
1
= L
p
0
(), X
2
= L
q
0
(), Y
1
= L
p
1
(), Y
2
=
L
q
1
() and recalling example 2.1.11 we get the RieszThorin theorem, as stated at the
beginning of the chapter. However, the proof of example 2.1.11 is modeled on the proof
of the RieszThorin theorem, so that this has not to be considered an alternative proof.
An important application of the RieszThorin theorem (or, equivalently, of theorem
2.1.6 and example 2.1.11) is the theorem of Hausdor and Young on the Fourier transform
in L
p
(R
n
). We set, for every f L
1
(R
n
),
(Tf)(k) =
1
(2)
n/2
_
R
n
e
ix,k)
f(x)dx, k R
n
.
As easily seen, |Tf|
L
2 = |f|
L
2 for every f C
0
(R
n
), so that T is canonically extended
to an isometry (still denoted by T) to L
2
(R
n
).
Theorem 2.1.13 If 1 < p 2, T is a bounded operator from L
p
(R
n
) to L
p
(R
n
), p
t
=
p/(p 1), and
|T|
L(L
p
,L
p
)

1
(2)
n(1/p1/2)
.
Proof Since |T|
L(L
1
,L
)
(2)
n/2
and |T|
L(L
2
)
= 1, by the RieszThorin theorem
T L(L
p
, L
q
) for every (0, 1), p
and q
being dened by (2.1): p
= 2/(1 + ),
q
= 2/(1 ) = p
t
. Moreover,
|T|
L(L
p
,L
p
)
|T|
L(L
1
,L
)
|T|
1
L(L
2
)

_
1
(2)
n/2
_
2/p1
. (2.7)
The use of the RieszThorin theorem may be avoided by using directly the results of
theorem 2.1.6 and of example 2.1.11: by theorem 2.1.6 T is a bounded operator from
[L
1
, L
2
]
to [L
, L
2
]
for every (0, 1); by example 2.1.11, [L

1
, L
2
]
= L
p
, p
= 2/(1+),
and [L
, L
2
]
= L
q
, q
= 2/(1 ), with identical norms. Therefore, (2.7) holds.

When runs in (0, 1), p
= 2/(1 +) runs in (1, 2), and the statement is proved.

A useful generalization of the Riesz-Thorin theorem is the Stein interpolation theorem.
It is obtained by applying theorem 2.1.7 to the interpolation couples (L
p
0
(), L
p
1
()),
(L
q
0
(), L
q
1
()), using the characterization of example 2.1.11.
Theorem 2.1.14 Let (, ), (, ) be -nite measure spaces. Let p
0
, p
1
, q
0
, q
1
[1, ]
and dene as usual p
, q
by
1
p
=
1
p
0
+

p
1
,
1
q
=
1
q
0
+

q
1
, 0 < < 1.
Assume that for every z S, T
z
: L
p
0
() L
p
1
() L
q
0
() +L
q
1
() is a linear operator
such that
(i) for each f L
p
0
() L
p
1
(), z T
z
f is holomorphic in the interior of S and
continuous and bounded in S with values in L
q
0
() +L
q
1
();
(ii) for each f L
p
0
() L
p
1
(), t T
it
f is continuous and bounded in R with values
in L
q
0
(), t T
1+it
f is continuous and bounded in R with values in L
q
1
();
46 Chapter 2
(iii) there are M
0
, M
1
> 0 such that for each f L
p
0
() L
p
1
(),
sup
tR
|T(t)f|
L
q
0()
M
0
|f|
L
p
0()
, sup
tR
|T(t)f|
L
q
1()
M
1
|f|
L
p
1()
.
Then for each (0, 1) and for each f L
p
0
() L
p
1
() we have
|T
f|
L
q
()
M
1
0
M
1
|f|
L
p
()
.
Therefore, T
may be extended to a bounded operator (which we still call T
) from L
p
()
to L
q
(), with p
and q
dened in (2.1), and

|T
|
L(L
p
(),L
q
())
M
1
0
M
1
.
Theorem 2.1.14 has a slightly sharper version, stated below, obtained modifying the
direct proof of the RieszThorin theorem. We recall that if (, ) is a measure space, a
simple function is a (nite) linear combination of characteristic functions of measurable
sets with nite measure.
Theorem 2.1.15 Let (, ), (, ) be -nite measure spaces. Assume that for every
z S, T
z
is a linear operator dened in the set of the simple functions on , with values
into measurable functions on , such that for every couple of simple functions a : C
and b : C, the product T
z
a b is integrable on and
z
_
(T
z
f)(x)g(x)(dx),
is continuous and bounded in S, holomorphic in the interior of S.
Assume moreover that for some p
j
, q
j
[1, +], j = 0, 1, we have
|T
it
a|
L
q
0()
M
0
|a|
L
p
0()
, |T
1+it
a|
L
q
1()
M
0
|a|
L
p
1()
, t R,
for every simple function a. Then for each (0, 1), T
may be extended to a bounded

operator (which we still call T
) from L
p
() to L
q
(), with p
and q
dened in (2.1),
and
|T
|
L(L
p
(),L
q
())
M
1
0
M
1
.
Proof The proof is just a modication of the proof of the RieszThorin theorem. For
every couple of simple functions a C, b : C, we apply the three lines theorem to
the function
F(z) =
_
T
z
f(z) g(z) (dx)
where f and g are dened as in the proof of the RieszThorin theorem, i.e.
f(z)(x) = [a(x)[
p
_
1z
p
0
+
z
p
1
_
a(x)
[a(x)[
, if x , a(x) ,= 0;
f(z)(x) = 0, if x , a(x) = 0.
g(z)(x) = [b(x)[
p
_
1z
p
0
+
z
p
1
_
b(x)
[b(x)[
, if x , b(x) ,= 0,
g(z)(x) = 0, if x , b(x) = 0.
We get
[F()[ =
(T
a)(x)b(x) (dx)
M
1
0
M
1
|a|
L
p
()
|b|
L
q
()
,
so that
|Ta|
L
q
()
M
1
0
M
1
|a|
L
p
()
,
for every simple a dened in . Since the set of such as is dense in L
p
() the statement
follows.
2.1.3 Exercises
1) The maximum principle for functions dened on a strip. Let f : S X be holomorphic
in the interior of S, continuous and bounded in S. Prove that for each S
|f()| maxsup
tR
|f(it)|, sup
tR
|f(1 +it)|.
(Hint: for each (0, 1) let z
0
be such that |f(z
0
)| |f|
(1); consider the functions

f
(z) = exp((z z
0
)
2
)f(z), and apply the maximum principle in the rectangle [0, 1]
[M, M] with M large).
2) The three lines theorem. Let f : S X be holomorphic the interior of S, continuous
and bounded in S. Show that
|f()|
X
(sup
tiR
|f(it)|
X
)
1
(sup
tiR
|f(1 +it)|
X
)
, 0 < < 1.
This estimate implies that if f vanishes in iR or in 1 +iR then f vanishes in S.
(Hint: apply the maximum principle of exercise 1 to (z) = e
z
f(z) and then choose > 0
properly).
3) Show that [X, X]
= X, with identical norms.

4) Using Lemma 2.1.2 prove that X Y is dense in [X, Y ]
for every (0, 1).

48 Chapter 2
Chapter 3
Interpolation and domains of
operators
3.1 Operators with rays of minimal growth
Let X be a real or complex Banach space with norm | |. In this section we consider a
linear operator A : D(A) X X such that
(A) (0, ), M : |R(, A)|
L(X)
M, > 0. (3.1)
Since (A) is not empty, then A is a closed operator, so that D(A) is a Banach space with
the graph norm |x|
D(A)
= |x| + |Ax|. Moreover for every m N also A
m
is a closed
operator (see exercise 1, 3.2.1).
This section is devoted to the study of the real interpolation spaces (X, D(A))
,p
, and,
more generally, (X, D(A
m
))
,p
. Since for every t, R the graph norm of D(A
m
) is
equivalent to the graph norm of D(B
m
) with B = e
it
(A+I), the case of an operator B
satisfying
(B) e
i
: >
0
, M : |R(e
i
, B)|
L(X)
M, >
0
for some [0, 2),
0
0, may be easily reduced to this one. The haline r = e
i
:
>
0
is said to be a ray of minimal growth of the resolvent of B. See [2, Def. 2.1].
Proposition 3.1.1 Let A satisfy (3.1). Then
(X, D(A))
,p
= x X : () =
|AR(, A)x| L
p
(0, +)
and the norms |x|
,p
and
|x|
,p
= |x| +||
L
p
(0,+)
are equivalent.
Proof. Let x (X, D(A))
,p
. Then if x = a + b with a X, b D(A), for every > 0
we have
|AR(, A)x|
|AR(, A)a| +
|R(, A)Ab|
(M + 1)
|a| +M
1
|Ab|
(M + 1)
(|a| +
1
|b|
D(A)
),
49
50 Chapter 3
so that
|AR(, A)x| (M + 1)
K(
1
, x).
With the changement of variables
1
we see that the right hand side belongs to
L
p
(0, ), with norm equal to (M + 1)|x|

,p
. Therefore () =
|AR(, A)x| is in
L
p
(0, ), and
|x|
,p
(M + 1)|x|
,p
.
Conversely, if L
p
(0, ), set for every 1

x = a
+b
= AR(, A)x +R(, A)x,

so that
K(
1
, x)
(|AR(, A)x| +
1
|R(, A)x|
D(A)
)
=
(2|AR(, A)x| +|R(, A)x|).

The right hand side belongs to L
p
(1, ), with norm estimated by

2|x|
,p
+M
_
1
(1 )p
_
1/p
|x|, p < ,
2|x|
,
+M|x|, p = .
It follows that t t
K(t, x) L
p
(0, 1), and hence x (X, D(A))

,p
and
|x|
,p
C
p
(|x|
,p
+|x|).
The following notation is widely used.

Denition 3.1.2 For 0 < < 1, 1 p we set
D
A
(, p) = (X, D(A))
,p
.
For 0 < < 1, k N, 1 p we set
D
A
( +k, p) = x D(A
k
) : A
k
x D
A
(, p),
|x|
D
A
(+k,p)
= |x| +|A
k
x|
D
A
(,p)
,
that is, D
A
( +k, p) is the domain of the part of A
k
in D
A
(, p).
From the denition we get easily
Lemma 3.1.3 For 0 < < 1, 1 p ,
D
A
( + 1, p) = (D(A), D(A
2
))
,p
and, more generally,
D
A
( +k, p) = (D(A
k
), D(A
k+1
))
,p
.
Interpolation and domains of operators 51
Proof. It is sucient to remark that (I A)
k
is an isomorphism from D(A
k
) to X, and
also from D(A
k+1
) to D(A). By the interpolation theorem 1.1.6, it is an isomorphism
between (D(A
k
), D(A
k+1
))
,p
and (X, D(A))
,p
, and the statement follows.
It is also important to characterize the real interpolation spaces between X and D(A
2
),
or more generally, between X and D(A
m
). The following proposition is useful.
Proposition 3.1.4 Let A satisfy (3.1). Then D(A) J
1/2
(X, D(A
2
)) K
1/2
(X, D(A
2
)).
Proof. Let us prove that D(A) J
1/2
(X, D(A
2
)). For every x D(A) it holds
lim
R(, A)x = lim
R(, A)Ax +x = x.
Setting f() = R(, A)x for > 0, we have
f
t
() = R(, A)x R(, A)
2
x = R(, A)(I R(, A))x = R(, A)
2
Ax
and f(+) = x, so that
x R(, A)x =
_

R(, A)
2
Axd, > 0,
and if x D(A
2
),
Ax = AR(, A)x
_

R(, A)
2
A
2
xd, > 0.
Therefore,
|Ax| (M + 1)|x| +
M
2
|A
2
x|, > 0.
Taking the inmum for (0, ) we get
|Ax| 2M(M + 1)
1/2
|x|
1/2
|A
2
x|
1/2
, x D(A
2
), (3.2)
so that
|x|
D(A)
C|x|
1/2
|x|
1/2
D(A
2
)
, x D(A
2
),
that is, D(A) J
1/2
(X, D(A
2
)).
Let us prove that D(A) K
1/2
(X, D(A
2
)). For every x D(A) split x as
x = R(, A)Ax +R(, A)x, > 0,
where
|R(, A)Ax|
M
|x|
D(A)
,
|R(, A)x|
D(A
2
)
= |R(, A)x| +|AR(, A)Ax|
M|x| +(M + 1)|Ax|
so that setting t =
2
K(t, x, X, D(A
2
)) |R(t
1/2
, A)Ax| +t|t
1/2
R(t
1/2
, A)x|
D(A
2
)
Mt
1/2
|x|
D(A)
+Mt|x| + (M + 1)t
1/2
|Ax|, t > 0
52 Chapter 3
which implies that t K(t, x, X, D(A
2
)) is bounded in (0, 1] by (2M + 1)|x|
D(A)
. Since
it is bounded by |x| in (1, ), then x (X, D(A
2
))
1/2,
and
|x|
(X,D(A
2
))
1/2,
(2M + 1)|x|
D(A)
.
But in general D(A) is not an interpolation space between X and D(A

2
). As a coun-
terexample we may take X = C
b
(R), A = realization of /x in X. See example 1.3.3.
As a corollary of proposition 3.1.4 we get a useful characterization of (X, D(A
2
))
,p
.
Proposition 3.1.5 Let A satisfy (3.1). Then for ,= 1/2
(X, D(A
2
))
,p
= D
A
(2, p).
Proof. Taking into account that X belongs to J
0
(X, D(A
2
)) K
0
(X, D(A
2
)) and D(A)
belongs to J
1/2
(X, D(A
2
)) K
1/2
(X, D(A
2
)), and applying the Reiteration Theorem with
E
0
= X, E
1
= D(A) we get
D
A
(, p) = (X, D(A))
,p
= (X, D(A
2
))
/2,p
, 0 < < 1,
and setting = 2 the statement follows for 0 < < 1/2. Taking into account that
D(A
2
) belongs to J
1
(X, D(A
2
)) K
1
(X, D(A
2
)) and D(A) belongs to J
1/2
(X, D(A
2
))
K
1/2
(X, D(A
2
)), and applying the Reiteration Theorem with E
0
= D(A), E
1
= D(A
2
) we
get
D
A
( + 1, p) = (D(A), D(A
2
))
,p
= (X, D(A
2
))
(+1)/2,p
, 0 < < 1,
and setting + 1 = 2 the statement follows for 1/2 < < 1.
Another characterization, which holds also for = 1/2, is the following one.
Proposition 3.1.6 Let A satisfy (3.1). Then for 0 < < 1, 1 p
(X, D(A
2
))
,p
= x X : () =
2
|(AR(, A))
2
x| L
p
(0, ),
and the norms |x|
(X,D(A
2
))
,p
and
|x|
e
,p
= |x| +| |
L
p
(0,)
are equivalent.
Proof. The proof is very close to the proof of proposition 3.1.1. Let x (X, D(A
2
))
,p
.
Then if x = a +b with a X, b D(A
2
), for every > 0 we have
2
|(AR(, A))
2
x|
2
|(AR(, A))
2
a| +
2
|R(, A)
2
A
2
b|
(M + 1)
2
2
|a| +M
2
22
|A
2
b| (M + 1)
2
2
(|a| +
2
|b|
D(A
2
)
),
so that
2
|(AR(, A))
2
x| (M + 1)
2
2
K(
2
, x).
We know that
K(, x) L
p
(0, ). With the change of variable =

2
we get
that
2
K(
2
, x) L
p
(0, ), with norm equal to 2

1/p
|x|
,p
. Therefore () =
2
|(AR(, A))
2
x| is in L
p
(0, ), and
|x|
e
,p
2
1/p
(M + 1)
2
|x|
,p
.
(The formula is true also for p = if we set 1/ = 0).
Conversely, if L
p
(0, ), from the obvious identity

x =
2
R(, A)
2
x 2AR(, A)
2
x +A
2
R(, A)
2
x,
where
AR(, A)
2
x = ( A)AR(, A)
3
x
= AR(, A)
2
R(, A)
2
x R(, A)A
2
R(, A)
2
x
we get
x = (I 2AR(, A))
2
R(, A)
2
x + (2R(, A) +I)A
2
R(, A)
2
x, 1,
where
|(I 2AR(, A))
2
R(, A)
2
x|
D(A
2
)
= |(I 2AR(, A))
2
R(, A)
2
x| +|(I 2AR(, A))
2
A
2
R(, A)
2
x|
(2M + 3)M
2
|x| + (2M + 3)
2
|A
2
R(, A)
2
x|
and
|(2R(, A) +I)A
2
R(, A)
2
x| (2M + 1)|A
2
R(, A)
2
x|.
Therefore,
2
K(
2
, x, X, D(A
2
))

2
(|(2R(, A) +I)A
2
R(, A)
2
x|
+
2
|(I 2AR(, A))
2
R(, A)
2
x|
D(A
2
)
)
(4M + 4)
2
|A
2
R(, A)
2
x| + (2M + 3)M
2
22
|x|.
The right hand side belongs to L
p
(1, ), with norm estimated by

(4M + 3)|x|
e
,p
+ (2M + 2)M
2
_
1
(2 2)p
_
1/p
|x|,
which is true also for p = with the convention (1/)
1/
= 1. It follows that t
t
K(t, x, X, D(A
2
)) L
p
(0, 1), and hence x (X, D(A

2
))
,p
and
|x|
(X,D(A
2
))
,p
C
p
(|x|
e
,p
+|x|).
Propositions 3.1.4 and 3.1.5 may be generalized as follows.

54 Chapter 3
Proposition 3.1.7 Let A satisfy (3.1), and let r, m N, r > m. Then D(A
r
)
J
r/m
(X, D(A
m
)) K
r/m
(X, D(A
m
)).
Proposition 3.1.8 Let A satisfy (3.1), and let m N. Then for (0, 1) such that
m / N, and for 1 p
(X, D(A
m
))
,p
= D
A
(m, p).
3.1.1 Two or more operators
Let us consider now two operators A : D(A) X, B : D(B) X, both satisfying (3.1).
Throughout the section we shall assume that A and B commute, in the sense that
R(, A)R(, B) = R(, B)R(, A), > 0.
It follows that D(A
k
B
h
) = D(B
h
A
k
) for all natural numbers h, k, and that A
k
B
h
x =
B
h
A
k
x for every x in D(A
k
B
h
).
Denition 3.1.9 For every m N set
K
m
=
m
j=0
D(A
j
B
mj
), |x|
K
m = |x| +
m
j=0
|A
j
B
mj
x|.
The main result of the section is the following.
Theorem 3.1.10 Let m N, p [1, ] and (0, 1) be such that m is not integer,
and set k = [m], = m. Then we have
(X, K
m
)
,p
= x K
k
: A
j
B
kj
x D
A
(, p) D
B
(, p), j = 0, . . . , k,
and the norms
x |x|
(X,K
m
)
,p
,
x |x| +
k
j=0
(|A
j
B
kj
x|
D
A
(,p)
+|A
j
B
kj
x|
D
A
(,p)
)
are equivalent.
The theorem will be proved in several steps. The rst one is the case m = 1.
Proposition 3.1.11 For every p [1, ] and (0, 1) we have
(X, K
1
)
,p
= D
A
(, p) D
B
(, p),
and the norms
|x|
(X,K
1
)
,p
, |x|
D
A
(,p)
+|x|
D
B
(,p)
are equivalent.
Proof. The embedding (X, K
1
)
,p
D
A
(, p) D
B
(, p) is obvious, since K
1
= D(A)
D(B) is continuously embedded both in D(A) and in D(B).
Let x D
A
(, p) D
B
(, p). We recall (see proposition 3.1.1) that the functions

|AR(, A)x|,
|BR(, B)x|, > 0,

belong to L
p
(0, ) and their norms are less than C|x|

D
A
(,p)
, C|x|
D
B
(,p)
, respectively.
For every > 0 set
v() =
2
R(, A)R(, B)x, > 0, (3.3)
and split x = x v() +v(). It holds
|v() x| |R(, A)(R(, B)x x)| +|R(, A)x x|
M|BR(, B)x| +|AR(, A)x|,
and
|v()|
K
1 = |v()| +|Av()| +|Bv()|
M
2
|x| +M|AR(, A)x| +M|BR(, B)x|.
Therefore, for 1
K(
1
, x, X, K
1
) 2M
(|AR(, A)x| +|BR(, B)x|) +M

2
1
|x|,
so that
K(
1
, x, X, K
1
) L
p
(1, ), with norm estimated by const. (|x|

+|x|
D
A
(,p)
+|x|
D
B
(,p)
). Then
K(, x, X, K
1
) L
p
(0, 1), with the same norm,

As a second step we show that
Proposition 3.1.12 For every p [1, ] and (0, 1) we have
(K
1
, K
2
)
,p
= x K
1
: Ax, Bx D
A
(, p) D
B
(, p)
= D
A
( + 1, p) D
B
( + 1, p),
and the norms
x |x|
(K
1
,K
2
)
,p
,
x |x| +|Ax|
D
A
(,p)
+|Ax|
D
B
(,p)
+|Bx|
D
A
(,p)
+|Bx|
D
B
(,p)
,
x |x|
D
A
(+1,p)
+|x|
D
B
(+1,p)
are equivalent.
Proof. Let us prove the embeddings . Since K
1
D(A) and K
2
D(A
2
) then
(K
1
, K
2
)
,p
(D(A), D(A
2
))
,p
= D
A
( + 1, p). Similarly, (K
1
, K
2
)
,p
D
B
( + 1, p). It
remains to show that each x (K
1
, K
2
)
,p
is such that Ax D
B
(, p) and Bx D
A
(, p).
For every a K
1
, b K
2
such that x = a +b we have
|BR(, B)Ax|
|BR(, B)Aa| +
|BR(, B)Ab|

(M + 1)|Aa| +M
1
|BAb| (M + 1)
(|a|
K
1 +
1
|b|
K
2)
56 Chapter 3
so that
|BR(, B)Ax| (M + 1)
K(
1
, x, K
1
, K
2
), > 0.
It follows that
|BR(, B)Ax| L
p
(0, ) with norm not exceeding (M+1)|x|

(K
1
,K
2
)
,p
,
and the embedding is proved.
The proof of the embedding x K
1
: Ax, Bx D
A
(, p) D
B
(, p) (K
1
, K
2
)
,p
is similar to the corresponding proof in proposition 3.1.11, and is omitted.
Let us prove that D
A
(+1, p)D
B
(+1, p) x K
1
: Ax, Bx D
A
(, p)D
B
(, p).
We have only to show that if x D
A
( + 1, p) D
B
( + 1, p) then Ax D
B
(, p) and
Bx D
A
(, p). Indeed, for each > 0 we have
|B
2
R(, B)
2
Ax|

+1
|B
2
R(, B)
2
R(, A)Ax| +
|B
2
R(, B)
2
AR(, A)Ax|

M(M + 1)|BR(, B)Bx| +
(M + 1)
2
|AR(, A)Ax|
so that
|B
2
R(, B)
2
Ax| L
p
(0, ) with norm not exceeding

M(M + 1)|Bx|
D
B
(,p)
+ (M + 1)
2
|Ax|
D
A
(,p)
.
Thanks to proposition 3.1.6, Ax D
B
2(/2, p), which coincides with D
B
(, p) thanks to
proposition 3.1.5. So, Ax D
B
(, p) and |Ax|
D
B
(,p)
C(|x|
D
A
(+1,p)
+ |x|
D
B
(+1,p)
.
Similarly, Bx D
A
(, p) and |Bx|
D
A
(,p)
C(|x|
D
B
(+1,p)
+|x|
D
A
(+1,p)
).
In the last part of the proof of proposition 3.1.12 we have shown that if x D
A
( +
1, p) D
B
( + 1, p) then Ax D
B
(, p) and Bx D
A
(, p), a sort of mixed regularity
result. However it is not true in general that D(A
2
) D(B
2
) D(AB).
For instance, let A be the realization of /x and let B be the realization of /y
in X = C(R
2
). Then D
A
( + 1, ) consists of the functions f X such that x
f(x, y) C
+1
(R), uniformly with respect to y R, and similarly D
B
( + 1, ) consists
of the functions f X such that y f(x, y) C
+1
(R), uniformly with respect to x R.
Proposition 3.1.12 states that if f D
A
(+1, )D
B
(+1, ) then f/x D
B
(, ),
that is it is Holder continuous also with respect to y, and f/y D
A
(, ), that is it
is Holder continuous also with respect to x. On the other hand, it is known that in this
example D(A
2
) D(B
2
) is not embedded in D(AB).
A similar proof yields
Proposition 3.1.13 For every k N, p [1, ] and (0, 1) we have
(K
k
, K
k+1
)
,p
= x K
k
: A
j
B
kj
x D
A
(, p) D
B
(, p), j = 0, . . . , k
and the norms
|x|
(K
k
,K
k+1
)
,p
, |x| +
k
j=0
(|A
j
B
kj
x|
D
A
(,p)
+|A
j
B
kj
x|
D
B
(,p)
)
are equivalent.
Next step consists in proving that K
1
belongs to J
1/2
(X, K
2
) and to K
1/2
(X, K
2
).
Proposition 3.1.14
K
1
J
1/2
(X, K
2
) K
1/2
(X, K
2
).
Proof. We already know that D(A) J
1/2
(X, D(A
2
)) and that D(B) J
1/2
(X, D(B
2
)).
Therefore there is C > 0 such that
|x|
K
1 |x|
D(A)
+|x|
D(B)
C|x|
1/2
(|x|
1/2
D(A
2
)
+|x|
1/2
D(B
2
)
) C
t
|x|
1/2
|x|
1/2
K
2
,
which means that K
1
J
1/2
(X, K
2
).
To prove that K
1
K
1/2
(X, K
2
), for every x K
1
we split again x = x v() +v()
for every > 0, where v is the function dened in (3.3). Then
|v() x| |R(, A)(R(, B)x x)| +|R(, A)x x|
= |R(, A)R(, B)Bx| +|R(, A)Ax|

1
(M(M + 1)|Bx| +M|Ax|),
and
|v()|
K
2 = |v()| +|A
2
v()| +|ABv()| +|B
2
v()|
M
2
|x| + 2M(M + 1)(|Ax| +|Bx|).
Setting = t
1/2
we deduce that
t
1/2
K(t, x, X, K
2
) t
1/2
(|x v(t
1/2
)| +t|v(t
1/2
)|
K
2)
C(|x|
K
1 +t
1/2
|x|),
is bounded in (0, 1). We know already that t t
1/2
K(t, x, X, K
2
) is bounded in [1, ).
Therefore K
1
is in the class K
1/2
between X and K
2
.
Arguing similarly one shows that
Proposition 3.1.15 For every k N 0, K
k+1
J
1/2
(K
k
, K
k+2
) K
1/2
(K
k
, K
k+2
).
More generally, K
k+1
J
1/s
(K
k
, K
k+s
) K
1/s
(K
k
, K
k+s
).
The Reiteration Theorem and proposition 3.1.14 yield now
Proposition 3.1.16 Let p [1, ], (0, 1), ,= 1/2. Then
(X, K
2
)
,p
= D
A
(2, p) D
B
(2, p),
and for > 1/2 we have also
(X, K
2
)
,p
= x K
1
: Ax, Bx D
A
(2 1, p) D
B
(2 1, p),
Proof. For < 1/2 we apply the Reiteration Theorem with Y = K
2
, E
0
= X, E
1
= K
1
,
and the statement follows from proposition 3.1.11. For > 1/2 we apply the Reiteration
Theorem with Y = K
2
, E
0
= K
1
, E
1
= K
2
, and the statement follows from proposition
3.1.12.
The above proposition is a special case of theorem 3.1.10, with m = 2. Theorem 3.1.10
in its full generality may be proved by recurrence, arguing similarly. See the exercises of
3.1.2.
The results and the procedures of this section are easily extended to the case of a nite
number of operators.
58 Chapter 3
3.1.2 Exercises
1) Let A : D(A) X X satisfy (3.1). Prove that for every m N, A
m
is a closed
operator. (Hint: use estimate (3.2)).
2) Prove proposition 3.1.7. Hint: to show that D(A
r
) K
r/m
(X, D(A
m
)) prove prelimi-
narly that D(A
r
) K
1/(mr)
(D(A
r1
), D(A
m
)), using a procedure similar to the one of
proposition 3.1.4, and then argue by reiteration.
3) Prove proposition 3.1.8.
4) Prove proposition 3.1.13.
5) Prove proposition 3.1.15. Hint: for the rst statement, follow step by step the proof of
3.1.14; for the second statement replace v() by w() =
2s
R(, A)
s
R(, B)
s
x.
6) Prove theorem 3.1.10 by recurrence on m, using the procedure of proposition 3.1.16
and the results of propositions 3.1.13 and 3.1.15.
7) Prove that (0, +) is a ray of minimal growth for the following operators:
(a) A : D(A) = C
1
b
(R) C
b
(R) (resp. A : D(A) = W
1,p
(R) L
p
(R), 1 p < ),
Af = f
t
(b) A : D(A) = C
2
b
(R) C
b
(R) (resp. A : D(A) = W
2,p
(R) L
p
(R), 1 p < ),
Af = f
tt
(c) A : D(A) = f C
2
([0, ]) : f(0) = f() = 0 C([0, ]) (resp. A : D(A) =
W
2,p
(0, ) W
1,p
0
(0, ) L
p
(0, ), 1 p), Af = f
tt
3.2 The case where A generates a semigroup
Let A : D(A) X X satisfy (3.1).
Due to the Hille-Yosida Theorem, if in addition D(A) is dense in X and for every n N
|(R(, A))
n
|
L(X)
M, then A is the innitesimal generator of a strongly continuous
semigroup T(t), and the following representation formula holds.
R(, A) =
_

0
e
t
T(t)dt, > 0. (3.4)
Since AR(, A) = R(, A) I, then
AR(, A) =
_

0
e
t
(T(t) I)dt, > 0. (3.5)
Proposition 3.2.1 Let A generate a semigroup T(t). Then
(X, D(A))
,p
= x X : t (t) = t
|T(t)x x| L
p
(0, )
and the norms |x|
,p
and
|x|
,p
= |x| +||
L
p
(0,)
are equivalent.
Proof. Recall that for every b D(A) we have
T(t)b b =
_
t
0
AT(s)b ds =
_
t
0
T(s)Ab ds, t > 0.
Let x (X, D(A))
,p
. Then if x = a+b with a X, b D(A), for every t > 0 we have
t
|T(t)x x| t
(|T(t)a a| +|T(t)b b|)

t
((M + 1)|a| +tM|Ab|) (M + 1)t
K(t, x).
Therefore (t) = t
|T(t)x x| L
p
(0, ) and
|x|
,p
(M + 1)|x|
,p
.
Conversely, if L
p
(0, ) let us use (3.5) to get
|AR(, A)x|
_

0
+1
t
+1
e
t
|T(t)x x|
t
dt
t
,
that is, is the multiplicative convolution between the functions f(t) = t
+1
e
t
and
(t) = t
|T(t)x x|. Since f L

1
(0, ) and L
p
(0, ), then L
p
(0, ) and
||
L
p
(0,)
|f|
L
1
(0,)
||
L
p
(0,)
, so that
|x|
,p
( + 1)|x|
,p
,
Proposition 3.2.2 Under the assumptions of proposition 3.2.1, for every (0, 1) and
p [1, ] we have
(X, D(A
2
))
,p
= x X : t

(t) = t
2
|(T(t) I)
2
x| L
p
(0, )
and the norms |x|
,p
and
|x|
ee
,p
= |x| +|
|
L
p
(0,)
are equivalent.
Proof. Recall that for every b D(A
2
) we have
(T(t) I)
2
b = (T(t) I)
_
t
0
T()Ab d =
_
t
0
_
t
0
T(s +)A
2
b ds d, t > 0,
so that
|(T(t) I)
2
b| t
2
M|A
2
b|.
Let x (X, D(A
2
))
,p
. Then if x = a +b with a X, b D(A
2
), for every t > 0 we have
t
2
|(T(t) I)
2
x| t
2
(|(T(t) I)
2
a| +|(T(t) I)
2
b|)
t
2
((M + 1)
2
|a| +t
2
M
2
|A
2
b|)
so that
t
2
|(T(t) I)
2
x| (M + 1)
2
t
2
K(t
2
, x).
60 Chapter 3
Therefore

(t) = t
2
|(T(t) I)
2
x| L
p
(0, ) and
|x|
e e
,p
2
1/p
(M + 1)
2
|x|
,p
.
Conversely, let x be such that

(t) L
p
(0, ). Then from (3.5) it follows that

(AR(, A))
2
x =
2
_

0
_

0
e
(t+s)
(T(t +s) T(t) T(s) +I)xds dt
= 2
2
_

0
e
2u
du
_
2u
0
(T(2u) T(t) T(2u t) +I)xdt
= 2
2
_

0
e
2u
(T(2u) 2T(u) +I)x
_
2u
0
dt du
+2
2
_

0
e
2u
du
_
2u
0
(2T(u) T(t) T(2u t))xdt.
The rst integral is nothing but
4
2
_

0
ue
2u
(T(u) I)
2
xdu.
To rewrite the second one we note that
_
2u
0
(2T(u) T(t) T(2u t))dt =
_
2u
0
(2T(u) 2T(t))dt
=
__
u
0
+
_
2u
u
_
(2T(u) 2T(t))dt
=
_
u
0
2(T(u t) I)T(t)dt +
_
2u
u
2T(u)(I T(t u))dt
=
_
u
0
2(T(s) I)T(u s)ds +
_
u
0
2T(u)(I T(s))ds
= 2
_
u
0
(T(s) I)(T(u s) T(u))ds = 2
_
u
0
(T(s) I)
2
T(u s)ds.
Therefore,
|
2
(AR(, A))
2
x| 4[(f

)()[ + 2[(f

1
)()[,
where stands for the multiplicative convolution and
f(u) = u
2+2
e
2u
,

1
(u) =
M
u
1+2
_
u
0
|(T(s) I)
2
x|ds.
Let us remark now that the Hardy-Young inequality (A.10)(i) implies that if a function
z is such that t t
z(t) L
p
(0, ) the same is true for its mean v(t) = t

1
_
t
0
z(s)ds,
with
|t t
v(t)|
L
p
(0,)

1
( + 1)
|t t
z(t)|
L
p
(0,)
,
and this is easily seen to be true also for p = . Therefore,

1
L
p
(0, ) and
|
1
|
L
p
(0,)
(2 + 1)
1
|
|
L
p
(0,)
. It follows that () = |
2
(AR(, A))
2
x|
L
p
(0, ), and
||
L
p
(0,)
= |x|
,p
4|f|
L
1
(0,)
(|
|
L
p
(0,)
+|
1
|
L
p
(0,)
) C
p
|x|
,p
,
Remark 3.2.3 In the proof of proposition 3.2.1 we have not used the fact that T(t) is
strongly continuous or that the domain of A is dense. The only essential assumption is
that T(t) is a semigroup such that |T(t)|
L(X)
M and for > 0 the operators
R() =
_

0
e
t
T(t)dt
are well dened and invertible. Indeed, in that case due to the semigroup property R()
satises the resolvent identity R() R() = ( )R()R(), for , > 0. From the
general spectral theory it follows that there exists a unique closed operator A such that
(A) (0, ) and R() = R(, A), for every > 0. The results of propositions 3.2.1 and
3.2.2 hold also for such semigroups.
The operator A may still be called generator of T(t), even if it is the innitesimal
generator in the usual sense if and only if T(t) is strongly continuous.
This is the case of the translations semigroups (T
i
(t)f)(x) = f(x+te
i
) in X = C
b
(R
n
),
of the Gauss-Weierstrass semigroup
P(t)f(x) =
1
(4t)
n/2
_
R
n
e
|y|
2
4t
f(x y)dy,
again in C
b
(R
n
), of the Ornstein-Uhlenbeck semigroup
T(t)f(x) =
1
(4t)
n/2
(det K
t
)
1/2
_
R
n
e
|K
1/2
t
y|
2
4t
f(e
tB
x y)dy,
with Q 0, B ,= 0 arbitrary n n matrices,
K
t
=
1
t
_
t
0
e
sB
Qe
sB
ds,
both in C
b
(R
n
) and in BUC(R
n
), etc. None of these semigroups is strongly continuous.
A useful embedding result in applications to PDEs is the following.
Theorem 3.2.4 Let T(t) be a semigroup in X. Assume moreover that there exists a
Banach space E X and m N, 0 < < 1, C > 0 such that
|T(t)|
L(X,E)

C
t
m
, t > 0,
and that t T(t)x is measurable with values in E, for each x X. Then E
J
(X, D(A
m
)), so that (X, D(A
m
))
,p
(X, E)
,p
, for every (0, 1), p [1, ].
62 Chapter 3
Proof. Let x D(A
m
), > 0 and set (I A)
m
x = y. Then x = (R(, A))
m
y so that
x =
(1)
m1
(m1)!
d
m1
d
m1
R(, A)y =
1
(m1)!
_

0
e
s
s
m1
T(s)y ds,
so that for every > 0
|x|
E

C
(m1)!
_

0
e
s
s
m(1)1
ds|y| =
C(m(1 ))
(m1)!

mm
|y|
=
C(m(1 ))
(m1)!

mm
_
_
_
_
m
r=0
_
m
r
_
mr
(1)
r
A
r
x
_
_
_
_
C
t
m
r=0
mr
|A
r
u|.
Let us recall that D(A
r
) belongs to J
m/r
(X, D(A
m
)) so that there is C such that
|x|
D(A
r
)
C|x|
r/m
D(A
m
)
|x|
1r/m
X
. Using such inequalities and then ab C(a
p
+ b
p
) with
p = n/r, p
t
= r/(n r) we get
|x|
E
C
m
(
m
|u|
D(A
m
)
+|u|), > 0,
so that taking the minimum for > 0
|x|
E
C|u|
1
|u|
D(A
m
)
and the statement holds.
3.2.1 Examples and applications. Schauder type theorems
Example 3.2.5 Let us apply propositions 3.2.1, 3.2.2 to the case X = L
p
(R), 1 p < ,
A : D(A) = W
1,p
(R) L
p
(R), Af = f
t
. Then T(t) is the translations semigroup,
T(t)f(x) = f(x +t). Applying proposition 3.1.1 we get for 0 < < 1
(L
p
(R), W
1,p
(R))
,p
= D
A
(, p)
= f L
p
: t t
|f( +t) f|
L
p
L
p
(0, ) = W
,p
(R),
which we knew already (example 1.1.8), but this is an alternative proof. Applying propo-
sition 3.1.5 and recalling that D(A
2
) = W
2,p
(R) we get for ,= 1/2
(L
p
(R), W
2,p
(R))
,p
= D
A
(2, p)
so that for < 1/2
(L
p
(R), W
2,p
(R))
,p
= W
2,p
(R),
and for > 1/2, by proposition 3.1.5,
(L
p
(R), W
2,p
(R))
,p
= f W
1,p
: f
t
D
A
(2 1, p) = W
2,p
(R).
For = 1/2 we need proposition 3.1.6: we get
(L
p
(R), W
2,p
(R))
1/2,p
=
= f L
p
: t t
1
|f( + 2t) 2f( +t) +f|
L
p L
p
(0, ) = B
1
p,p
(R),
which coincides with W
1,p
(R) only for p = 2.
Choosing X = C
b
(R), A : D(A) = C
1
b
(R) C
b
(R), Af = f
t
, we get, recalling remark
3.2.3,
(C
b
(R), C
2
b
(R))
,
= C
2
b
(R), ,= 1/2,
(C
b
(R), C
2
b
(R))
1/2,
=
=
_
f C
b
(R) : sup
t,=0, xR
[f(x + 2t) 2f(x +t) +f(x)[
t
<
_
,
which Zygmund called
1
(R). It is easy to see that Lip(R)
1
(R), but the converse is
not true.
Example 3.2.6 Let A
i
, i = 1, . . . , n be the realization of the partial derivative /x
i
in
C
b
(R
n
), or in BUC(R
n
), or in L
p
(R
n
), 1 p . Each A
i
satises (3.1), with
(R(, A
i
)f)(x) =
_
+
x
i
e
(x
i
s)
f(x
1
, . . . , x
i1
, s, x
i+1
, . . . , x
n
)ds,
for > 0, f X, x R
n
, so that M = 1 for every i, and R(, A
i
)R(, A
j
) =
R(, A
j
)R(, A
i
) for every i, j. We apply theorem 3.1.10 for those such that m is
not integer, m = k +, k = [m], 0 < < 1.
If X = C
b
(R
n
) (resp., X = BUC(R
n
)) then K
m
= C
m
b
(R
n
) (resp., K
m
= BUC
m
(R
n
)).
From the second part of proposition 3.1.1, or else from example 1.1.8 we know that
D
A
i
(, ) = f X : s f(x
1
, . . . , x
i1
, s, x
i+1
, . . . , x
n
) C
b
(R), so that
n
i=1
D
A
i
(, ) = C
b
(R
n
).
From theorem 3.1.10 we get
(X, K
m
)
,
= f K
k
: D
f C
(R
n
), [[ = k = C
m
b
(R
n
).
Let now X = L
p
(R
n
), 1 p . From the second part of proposition 3.1.1 and from
example 1.1.8 we know that D
A
i
(, p) = f X : s f(x
1
, . . . , x
i1
, s, x
i+1
, . . . , x
n
)
W
,p
(R), so that
n
i=1
D
A
i
(, p) = W
,p
(R
n
).
From theorem 3.1.10 we get
(X, K
m
)
,p
= (L
p
, W
k,p
)
,p
= f W
k,p
(R
n
: D
f W
,p
(R
n
), [[ = k = W
m,p
(R
n
).
After such characterizations we are able to characterize other important interpolation
spaces by means of theorem 3.2.4.
Example 3.2.7 Let A be the realization of the Laplace operator in X = L
p
(R
n
), 1
p < . Then for 0 < < 1
W
,p
(R
n
) = D
A
(/2, p), W
+2,p
(R
n
) = D
A
(/2 + 1, p).
If A is the realization of in X = BUC(R
n
), in X = C
b
(R
n
) or in X = L
(R
n
), then
C
b
(R
n
) = D
A
(/2, ), C
+2
b
(R
n
) = D
A
(/2 + 1, ).
64 Chapter 3
Proof. The embeddings are easy consequences of example 3.2.6. Indeed, let X =
L
p
(R
n
), 1 p < . Example 3.2.6 yields W
,p
(R
n
) = (L
p
(R
n
), W
2,p
(R
n
))
,p
. Since
W
2,p
D(A), then we have
W
,p
(R
n
) = (L
p
(R
n
), W
2,p
(R
n
))
/2,p
D
A
(/2, p).
Similarly, from example 3.2.6 we know that W
+2,p
(R
n
) = (L
p
(R
n
), W
4,p
(R
n
))
(+2)/4,p
.
Since W
4,p
(R
n
) D(A
2
), then we have
W
+2,p
(R
n
) = (L
p
(R
n
), W
4,p
(R
n
))
(+2)/4,p
(X, D(A
2
))
(+2)/4,p
= D
A
(/2 + 1, p),
where the last equality follows from proposition 3.1.5.
The same proof works in the case X = BUC(R
n
).
To prove the opposite inclusions we introduce the Gauss-Weierstrass semigroup,
P(t)f(x) =
1
(4t)
n/2
_
R
n
e
|xy|
2
4t
f(y)dy, t > 0, x R
n
. (3.6)
P(t) may be seen as a (strongly continuous) semigroup in X = L
p
(R
n
), 1 p < or in
X = BUC(R
n
). Its innitesimal generator is the realization of the Laplace operator in X.
It is easy to see that if 1 p , P(t)f C
(R
n
) for every f L
p
(R
n
), and that
|D
P(t)f|
L
p
C
,p
t
[[/2
|f|
L
p, t > 0.
In particular,
_
_
|P(t)|
L(L
p
,W
1,p
)
C
_
1 +
1
t
1/2
_
, t > 0,
|P(t)|
L(L
p
,W
3,p
)
C
_
1 +
1
t
1/2
+
1
t
+
1
t
3/2
_
, t > 0,
(3.7)
and similarly
_
_
|P(t)|
L(BUC(R
n
),BUC
1
(R
n
))
C
_
1 +
1
t
1/2
_
, t > 0,
|P(t)|
L(BUC(R
n
),BUC
3
(R
n
))
C
_
1 +
1
t
1/2
+
1
t
+
1
t
3/2
_
, t > 0.
(3.8)
Replacing P(t) by T(t) = P(t)e
t
(the semigroup generated by AI) we get
_
_
|T(t)|
L(L
p
,W
1,p
)

C
t
1/2
, t > 0,
|T(t)|
L(L
p
,W
3,p
)

C
t
3/2
, t > 0,
and
_
_
|T(t)|
L(BUC(R
n
),BUC
1
(R
n
))

C
t
1/2
, t > 0,
|T(t)|
L(BUC(R
n
),BUC
3
(R
n
))

C
t
3/2
, t > 0.
Let X = L
p
(R
n
), 1 p < . Since D(A) = D(A I) then D
A
(/2, p) = D
AI
(/2, p).
Using theorem 3.2.4, with E = W
1,p
(R
n
), m = 1, = 1/2, we get
D
AI
(/2, p) = (X, D(AI))
/2,p
(L
p
(R
n
), W
1,p
(R
n
))
,p
= W
,p
(R
n
).
Therefore,
W
,p
(R
n
) D
A
(/2, p).
Moreover, since D(A
2
) = D((A I)
2
) then D
A
(/2 + 1, p) = D
AI
(/2 + 1, p). Using
again theorem 3.2.4, with E = W
3,p
(R
n
), m = 2, = 3/4, we get
D
AI
(/2 + 1, p) = (X, D((AI)
2
))
(+2)/4,p
(L
p
(R
n
), W
3,p
(R
n
))
(+2)/3,p
= W
+2,p
(R
n
),
the last equality following from example 3.2.6. Therefore,
W
+2,p
(R
n
) D
A
(/2 + 1, p),
and the rst part of the statement is proved. The same procedure works in the case
X = BUC(R
n
), X = C(R
n
), X = L
(R
n
).
Remark 3.2.8 Note that the embeddings hold for every operator A : D(A) L
p
(R
n
)
L
p
(R
n
) such that D(A) W
2,p
(R
n
) and D(A
2
) W
4,p
(R
n
) (respectively, A : D(A)
BUC(R
n
) BUC(R
n
) such that D(A) BUC
2
(R
n
) and D(A
2
) BUC
4
(R
n
)), whereas
the embeddings hold for every operator A : D(A) L
p
(R
n
) L
p
(R
n
) (respectively,
A : D(A) BUC(R
n
) BUC(R
n
)) which generates a semigroup P(t) satisfying es-
timates (3.7) (respectively, (3.8)). For 1 < p < one could prove the statement also
using the known characterizations D(A) = W
2,p
(R
n
), D(A
2
) = W
4,p
(R
n
). However, such
characterizations are not true for p = 1; similarly, it is not true that if X = BUC(R
n
)
then D(A) = BUC
2
(R
n
) and D(A
2
) = BUC
4
(R
n
).
An important consequence of example 3.2.7 are the optimal regularity theorems for
the Laplace equation in Holder and in fractional Sobolev spaces.
Corollary 3.2.9 (i) (Schauder Theorem) Let u C
2
b
(R
n
) be such that u C
b
(R
n
) with
0 < < 1. Then u C
+2
b
(R
n
), and
|u|
C
+2
b
(R
n
)
C(|u|
+|u|
C
b
(R
n
)
).
(ii) Let u W
2,p
(R
n
) be such that u W
,p
(R
n
) with 0 < < 1, 1 p < . Then
u W
+2,p
(R
n
), and
|u|
W
+2,p
(R
n
)
C(|u|
L
p +|u|
W
,p
(R
n
)
).
66 Chapter 3
Chapter 4
Powers of positive operators
The powers (with real or complex exponents) of positive operators are important tools in
the study of partial dierential equations. The theory of powers of operators is very close
to interpolation theory, even if in general the domain of a power of a positive operator is
not an interpolation space.
Through the whole chapter X is a complex Banach space.
4.1 Denitions and general properties
Denition 4.1.1 A linear operator A : D(A) X X is said to be a positive operator
if the resolvent set of A contains (, 0] and there is M > 0 such that
|R(, A)|
L(X)

M
1 +[[
, 0. (4.1)
Note that A is a positive operator i (, 0) is a ray of minimal growth for the
resolvent R(, A) and 0 (A). So, if A is a positive operator, then A satises (3.1) so
that all the results of 3.1 are applicable.
Examples of unbounded positive operators are readily given: for instance, the realiza-
tion of the rst order derivative with Dirichlet boundary condition at x = 0 in C([0, 1])
or in L
p
(0, 1), 1 p is positive. More generally, if A is the generator of a strongly
continuous or analytic semigroup T(t) such that |T(t)| Me
t
for some > 0, then
A is a positive operator. This can be easily seen from the already mentioned resolvent
formula
R(, A) = R(, A) =
_

0
e
t
T(t)dt, > .
This section is devoted to the construction and to the main properties of the powers
A
z
, where z is an arbitrary complex number.
If A : X X is a bounded positive operator the powers A
z
are readily dened by
A
z
=
1
2i
_
z
R(, A)d,
where is any piecewise smooth curve surrounding (A), avoiding (, 0], with index
1 with respect to every element of (A). Several properties of A
z
follow easily from the
denition: for instance, z A
z
is holomorphic with values in L(X); if z = k Z then
67
68 Chapter 4
A
z
dened above coincides with A
k
; for each z, w C we have A
z
A
w
= A
w
A
z
= A
z+w
;
(A
1
)
z
= A
z
, etc.
In the case where A is unbounded the theory is much more complicated. To dene A
z
we shall use an elementary but important spectral property, stated in the next lemma.
Lemma 4.1.2 Let A be a positive operator. Then the resolvent set of A contains the set
= C : Re 0, [Im [ < ([Re [ + 1)/M C : [[ < 1/M,
where M is the number in formula (4.1), and for every
0
(0, arctan1/M), r
0
(0, 1/M)
there is M
0
> 0 such that
|R(, A)|
M
0
1 +[[
for all C with [[ r
0
, and for all C with Re < 0 and [Im [/[Re [ tan
0
.
Proof. It is sucient to recall that for every
0
(A) the resolvent set (A) contains the
open ball centered at
0
with radius 1/|R(
0
, A)|, and that for [
0
[ < 1/|R(
0
, A)|
it holds
R(, A) =
n=0
(1)
n
(
0
)
n
R(
0
, A)
n+1
.
The union of the balls centered at
0
(, 0] with radius 1/|R(
0
, A)| contains the set
, and the estimate follows easily.
For (/2, ), r > 0, let
r,
be the curve dened by
r,
=
(1)
r,

(2)
r,
+
(3)
r,
,
where
(1)
r,
,
(3)
r,
are the half lines parametrized respectively by z = e
i
, z = e
i
, r,
and
(2)
r,
is the arc of circle parametrized by z = re
i
, . See the gure.
0
r,
r
Fig. 1. The curve
r,
.
Now we are ready to dene A
z
for Re z < 0 through a Dunford integral.
Denition 4.1.3 Fix any r (0, 1/M), ( arctan1/M, ). For Re < 0 set
A
=
1
2i
_
r,
R(, A)d. (4.2)

Powers of positive operators 69
Since
R(, A) is holomorphic in (, 0] with values in L(X), the integral

is an element of L(X) independent of r and . Writing down the integral we get
A
=
1
2i
_

r
(e
i(+1)
R(e
i
, A) +e
i(+1)
R(e
i
, A))d
r
+1
2
_

e
i(+1)
R(re
i
, A)d
(4.3)
for every r (0, 1/M), ( arctan1/M, ).
Of course formula (4.3) may be reworked to get simpler expressions for A
. For
instance, if 1 < Re < 0 we may let r 0, to get
A
x =
sin()
_

0
(I +A)
1
xd. (4.4)
Note that for every a > 0 and (1, 0) we have
a
=
sin()
_

0
+a
d, (4.5)
which agrees with (4.4) of course, and will be used later.
From the denition it follows immediately that the function z A
z
is holomorphic
in the half plane Re z < 0, with values in L(X). Its behavior near the imaginary axis is
not obvious, but it is of great importance in the developements of the theory and will be
discussed in the next section.
Let us see some basic properties of the operators A
.
Proposition 4.1.4 The following statements hold true.
(i) For = n, n N, the operator dened in (4.2) coincides with A
n
= n-th power
of the inverse of A.
(ii) For Re z < k, k N, the range of A
z
is contained in the domain D(A
k
), and
A
k
A
z
x = A
k+z
x, x X.
(iii) For Re z < 0 and x D(A
k
), k N, A
z
x D(A
k
), and
A
z
A
k
x = A
k
A
z
x.
(iv) For Re z
1
, Re z
2
< 0 we have
A
z
1
A
z
2
= A
z
1
+z
2
.
Proof. (i) Let = n. It is easy to see that
1
2i
_
r,
n
R(, A)d = lim
k
1
2i
_
n
R(, A)d,
with
k
as in gure.
70 Chapter 4
r
k
Fig. 2. The curve
k
.
For every k N the function R(, A) is holomorphic in the bounded region
surrounded by
k
. For every k N we have
1
2i
_
n
R(, A)d =
1
(n 1)!
d
n1
d
n1
R(, A)
=0
= A
n
,
and letting k ,
1
2i
_
r,
n
R(, A)d = A
n
.
(ii) Let k = 1, Re z < 1. Then, since
|
z
AR(, A)| = |
z
(R(, A) I)| (M
0
+ 1)[[
Re z
,
the integral dening A
z
is in fact an element of L(X, D(A)), and
A
1
2i
_
r,
z
R(, A)d =
1
2i
_
r,
z+1
R(, A)d
1
2i
_
r,
z
dI.
But the last integral vanishes, so that A A
z
= A
1+z
, and the statement is proved for
k = 1. The statement for any k follows arguing by recurrence.
Statement (iii) is obvious because A
k
commutes with R(, A) on D(A
k
), and this
implies that A
k
commutes with A
z
on D(A
k
).
(iv) Let
1
<
2
< , 1/M > r
1
> r
2
> 0, so that
r
1
,
1
is on the right hand side of
r
2
,
2
. Then
A
z
1
A
z
2
=
1
(2i)
2
_
r
1
,
1
z
1
R(, A)d
_
r
2
,
2
w
z
2
R(w, A)dw
=
1
(2i)
2
_
r
1
,
1
r
2
,
2
z
1
w
z
2
R(, A) R(w, A)
w
ddw
=
1
(2i)
2
_
r
1
,
1
z
1
R(, A)d
_
r
2
,
2
w
z
2
w
dw
1
(2i)
2
_
r
2
,
2
w
z
2
R(w, A)dw
_
r
1
,
1
z
1
w
d
=
1
2i
_
r
1
,
1
z
1
+z
2
R(, A)d = A
z
1
+z
2
.
Statement (iv) of the proposition implies immediately that A

z
is one to one. Indeed, if
A
z
x = 0 and n N is such that n < Re z, then A
n
x = A
nz
A
z
x = 0, so that x = 0.
Therefore it is possible to dene A
if Re > 0 as the inverse of A
. But in this way

the powers A
it
, t R, remain undened. So we give a unied denition for Re 0.
Denition 4.1.5 Let 0 Re < n, n N. We set
D(A
) = x X : A
n
x D(A
n
), A
x = A
n
A
n
x.
From proposition 4.1.4 it follows that the operator A
is independent of n: indeed,
if n, m > Re , then A
m
x = A
nm
A
n
x both for n < m (by proposition 4.1.4(iv))
and for n > m (by proposition 4.1.4(ii), taking z = n and k = n m), so that
A
m
x D(A
m
) i A
nm
A
n
x D(A
m
) i.e. A
n
x D(A
n
).
For = 0 we get immediately A
0
= I. Moreover for Re > 0 we get
D(A
) = A
(X); A
= (A
)
1
.
Indeed, A
n
x D(A
n
) i there is y X with A
n
x = A
n
y. Such a y is obviously
unique, and A
x = y by denition. Moreover A
n
A
y = A
A
n
y = A
A
n
x =
A
n
x so that x = A
y is in the range of A
and A
= (A
)
1
.
Since A
has a bounded inverse, then it is a closed operator, so that D(A
) is a Banach
space endowed with the graph norm. Again, since A
has a bounded inverse, its graph

norm is equivalent to
x |A
x|,
which is usually considered the canonical norm of D(A
).
If Re = 0, = it with t R, A
it
is the inverse of A
it
in the sense that for each
x D(A
it
), A
it
x D(A
it
) and A
it
A
it
x = x. Indeed, if x D(A
it
) then A
it1
x D(A),
and A
it
x = A(A
it1
x) by denition. Therefore A
1it
A
it
x = A
1it
A A
it1
x = A
A
1it
A
it1
x = A A
2
x D(A), which implies that A
it
x D(A
it
) and A
it
A
it
x = x.
But in general the operators A
it
are not bounded, see next example 4.2.1. However,
they are closed operators, because A
1+it
is bounded and A is closed (see next exercise 6,
4.2.1). Therefore also D(A
it
) is a Banach space under the graph norm.
From the denition it follows easily that for 0 Re < n N, the domain D(A
n
)
is continuously embedded in D(A
): indeed for each x D(A

n
), A
n
x D(A
n
) by
proposition 4.1.4(iii), and A
x = A
n
A
n
x = A
n
A
n
x so that |A
x| |A
n
| |A
n
x|.
This property is generalized in the next theorem.
Theorem 4.1.6 Let , C be such that Re < Re . Then D(A
) D(A
), and
for every x D(A
),
A
x = A
x.
Moreover for each x D(A
), A
x D(A
) and
A
x = A
x.
Conversely, if x D(A
) and A
x D(A
), then x D(A
) and again A
x =
A
x.
72 Chapter 4
Proof. The embedding D(A
) D(A
) is obvious if Re < 0; it has to be proved for

Re 0.
If x D(A
), A
n+
x D(A
n
) for n > Re . Therefore A
n+
x = A
A
n+
x
D(A
n
), thanks to proposition 4.1.4(iii), so that x D(A
), and A
x = A
n
A
A
n+
x =
A
x. Since A
is a bounded operator, |A
x| |A
|
L(X)
|A
x|, and D(A
)
).
Let again x D(A
), and let n > maxRe , Re ( ). Then

A
n+
A
x = A
n+
A
x = A
n
A
x D(A
n
),
so that A
x D(A
) and A
x = A
x.
Let now x D(A
) be such that A
x D(A
), and x n > max Re , Re .

Then
A
2n
x = A
n
A
n+
x = A
n
A
n
A
x = A
n
A
n
A
x
is in D(A
2n
), so that x D(A
) and A
x = A
2n
A
2n
x = A
x.
The condition Re < Re is essential in the above theorem when Re > 0. In fact
for every > 0, t R we have D(A
) = D(A
+it
) if and only if A
it
is bounded. See
exercise 2, 4.2.1.
Now we give some representation formulas for A
x when x D(A
). We consider rst
the case where 0 < Re < 1. Taking n = 1 in the denition, we see that x D(A
) if
and only if A
1
x D(A). Letting r 0 and in the representation formula (4.3)
for A
1
x (i.e., using formula (4.4) with replaced by 1) we get
A
1
x =
sin()
_

0
1
(I +A)
1
xd. (4.6)
Therefore x D(A
) if and only if the integral

_
0

1
(I + A)
1
xd is in the domain
of A, and in this case
A
x =
sin()
A
_

0
1
(I +A)
1
xd
=
1
()(1 )
A
_

0
1
(I +A)
1
xd,
(4.7)
which is the well-known Balakrishnan formula.
Another important representation formula holds for 1 < Re < 1. The starting
point is again formula (4.3) for A
1
x. We let and then we integrate by parts in
the integrals between r and , getting
A
1
x =
sin()
_

r
(I +A)
2
xd r
sin()
(rI +A)
1
x
2
_

e
i
(re
i
I +A)
1
xd
(with (sin())/() replaced by 1 if = 0) and letting r 0 we get (both for Re
(0, 1) and for Re (1, 0])
A
1
x =
1
(1 )(1 +)
_

0
(I +A)
2
xd. (4.8)
Therefore x D(A
) if and only if the integral

_
(I + A)
2
xd is in the domain of
A, and in this case
A
x =
1
(1 +)(1 )
A
_

0
(I +A)
2
xd. (4.9)
The most general formula of this type may be found as usual in the book of Triebel:
for n N 0, m N, n < Re < mn we have
A
x =
(m)
( +n)(mn )
A
mn
_

0
t
+n1
(tI +A)
m
xdt
for every x D(A
). See [36, 1.5.1].

We already know that the domain D(A) is continuously embedded in D(A
) for Re
[0, 1). With the aid of the representation formulas (4.7) and (4.9) we are able to prove
more precise embedding properties of D(A
).
Proposition 4.1.7 For 0 < Re < 1, D(A
) J
Re
(X, D(A)) K
Re
(X, D(A)), i.e.
(X, D(A))
Re ,1
D(A
) (X, D(A))
Re ,
.
Proof. The embedding (X, D(A))
Re ,1
D(A
) is easy, because for > 0

|A
1
(I +A)
1
x| =
Re 1
|A(I +A)
1
x|
and for every x (X, D(A))
Re ,1
the function
Re
|A(A + I)
1
x| is in L
1
(0, )
thanks to proposition 3.1.1. Using the representation formula (4.6) for A
1
x, we get
A
1
x D(A), i.e. x D(A
) and by (4.7)
|A
x|
1
[()(1 )[
_

0
Re
|A(A+I)
1
x|
d
C|x|
(X,D(A))
Re ,1
.
Let now x D(A
). Then x = A
y, with y = A
x, so that x = A A
1
y. We use
the representation formula (4.8) for A
1
y, that gives
x =
A
(1 )(1 +)
_

0
t
(A+tI)
2
y dt.
On the other hand, by proposition 3.1.1 we have
|x|
(X,D(A))
Re ,
C() sup
>0
|
Re
A(A+I)
1
x|,
so that
|x|
(X,D(A))
Re ,
C() sup
>0
_
_
_
_
Re
A
2
(A+I)
1
(1 )(1 +)
_

0
t
(A+tI)
2
y dt
_
_
_
_
.
For every > 0 we have
Re
_
_
_
_
A
2
(A+I)
1
_

0
t
(A+tI)
2
y dt
_
_
_
_

Re
M
1 +
_

0
t
Re
(M + 1)
2
|y| dt
+
Re
(M + 1)
_

t
Re
M(M + 1)
1 +t
|y|dt
C|y|
74 Chapter 4
so that x (X, D(A))
Re ,
and
|x|
(X,D(A))
Re ,
C
t
|y| = C
t
|A
x|,
which implies that D(A
) (X, D(A))
Re ,
.
Remark 4.1.8 Arguing similarly (using formula (4.9) instead of (4.7)) we see easily that
for every (0, 1) and t R, (X, D(A))
,1
is contained in D(A
it
). Indeed the function
|
it
A(I +A)
2
x| M(1+)
1
|A(I +A)
1
x| is in L
1
(0, ) for every x in (X, D(A))
,1
,
so that the integral
_
0

it
(I +A)
2
xd belongs to the domain of A. Therefore, for every
(0, 1) and p [1, ], t R, (X, D(A))
,p
it
) (because
it is continuously embedded in (X, D(A))
/2,1
).
Remark 4.1.9 Let 0 < < 1. It is possible to show that in its turn A
is a positive
operator, and that
R(, A
) =
1
2i
_
r,
R(z, A)
z
dz, 0. (4.10)
(see exercise 5, 4.2.1). Using the above formula for the resolvent, one shows that A
is a
sectorial operator. This may be surprising, since A is not necessarily sectorial. This also
may help in avoiding mistakes driven by intuition. Consider for instance the case where
X = L
2
(0, ) and A is the realization of d
2
/dx
2
with Dirichlet boundary condition, i.e.
A : D(A) = H
2
(0, ) H
1
0
(0, ), Au = u
tt
. One could think that A
1/2
is a realization of
i d/dx with some boundary condition, but this cannot be true because such operators are
not sectorial. See next example 4.3.10.
4.1.1 Powers of nonnegative operators
A part of the theory of powers of positive operators may be extended to nonnegative
operators.
Denition 4.1.10 A linear operator A : D(A) X X is said to be nonnegative if the
resolvent set of A contains (, 0) and there is M > 0 such that
|(I +A)
1
|
M
, > 0.
In other words, A is a nonnegative operator i (, 0) is a ray of minimal growth for the
resolvent of A.
An important example of nonnegative operator is the realization A of (the Laplace
operator) in L
p
(R
n
), 1 p . But A is not positive because 0 (A). However if
p < then A is one to one. See exercise 13, 4.2.1.
If 0 (A) but A is one to one, it is still possible to dene A
z
for 1 < Re z < 1.
Let 1 < Re z < 1, z ,= 0, and dene an operator B
z
on D(A) R(A) by
B
z
x =
sin(z)
_
x
z

A
1
x
1 +z
+
+
_
1
0
z+1
(I +A)
1
A
1
xd +
_

1
z1
(I +A)
1
Axd
_
(4.11)
for each x D(A) R(A) (note that in the case where 0 (A), B
z
x coincides with A
z
x
since formula (4.11) is obtained easily from (4.9)). Then one checks that B
z
: D(A)
R(A) H is closable, and denes A
z
as the closure of B
z
.
Another way to dene A
for 0 < < 1, even if A is not one to one, is the following:

for > 0 one denes (I +A
)
1
by
R
=
sin()
_

0
2
+ 2
cos() +
2
(I +A)
1
d, (4.12)
then one checks that R
is invertible for every > 0 and R
= (R
)/( ).
Therefore there exists a unique closed operator B such that R
= (I+B)
1
for > 0, and
we set A
= B. (Note that in the case where 0 (A) the above formula for (I +A
)
1
is correct because it is obtained from (4.10) letting r 0 and ).
From the representation formula (4.12) it follows that
lim
0
(I + (I +A)
)
1
= (I +A
)
1
, in L(H),
which will be used in the proof of next lemma. In its turn, lemma 4.1.11 will be used in
the proof of theorem 4.3.4.
Lemma 4.1.11 Let A : D(A) X X be any nonnegative densely dened operator.
Then D(A+I)
= D(A
) for each [0, 1], and there is C independent of such that

|(A+I)
x A
x| C
|x|, x D(A
).
Proof. For 0 < < , A+I and A+I are positive operators. Using the Balakrishnan
formula for 0 < < 1, we get
(A+I)
x (A+I)
x =
=
sin()
_
( )
_

(( +)I +A)
1
(( +)I +A)
1
xd
+
_

0
1
((A+I)(( +)I +A)
1
(A+I)(( +)I +A)
1
)xd
_
for every x D(A) and > 0. Therefore,
|(A+I)
x (A+I)
x|
sin()
_
( )M
2
_

2
d + 2(1 +M)
_

0
1
d
_
=
sin()
_

1
M
2
1
+
2(1 +M)
_
|x|
for every x D(A) and > 0. Taking = ( ) we get
|(A+I)
x (A+I)
x| C()( )
|x|.
76 Chapter 4
Therefore, for every x D(A) the function (A + I)
x is uniformly continuous, so
that there exists the limit lim
0
(A + I)
x = Bx. Letting 0 in the above estimate

we nd
|(A+I)
x Bx| C()
|x|. (4.13)
for each > 0, for each x D(A). But D(A) is a core of (A+I)
, that is the closure of

the restriction of (A+I)
to D(A) is (A+I)
itself: indeed, for every y D((A+I)
)
the sequence y
n
= n(nI +A)
1
y is in D(A), y
n
y and (A+I)
y
n
= n(nI +A)
1
(A+
I)
y (A + I)
y as n , because D(A) is dense. This implies that B is closable,

its closure B has domain D((A + I)
), and inequality (4.13) holds also for B. So,

(A+I)
x Bx uniformly for x D(B), |x| 1, and this implies that (, 0) (B),

and (I + (A + I)
)
1
(I + B)
1
as 0 in L(X). Since (I + (A + I)
)
1
(I +A
)
1
, then B = A
.
4.2 Operators with bounded imaginary powers
Let again A be a positive operator in a complex Banach space X. We know that the
operators A
z
are bounded if Re z < 0 and unbounded in general if Re z > 0 (this is
because D(A
z
) (X, D(A))
Re z,
by proposition 4.1.7). Moreover remark 4.1.8 tells us
that for every t R, the domain D(A
it
) is not very far from the underlying space X
because all the interpolation spaces (X, D(A))
,p
are continuously embedded in D(A
it
).
So, it is natural to ask whether D(A
it
) = X and A
it
is a bounded operator also for t ,= 0.
The general answer is no, as the following example shows.
Example 4.2.1 Let S be the shift operator, S(
1
,
2
,
3
, . . .) = (0,
1
,
2
, . . .), in X = c
0
=
the space of all complex valued sequences
n
such that lim
n
n
= 0, endowed with the
sup norm, and set A = (I S)
1
. Then A is a positive operator, and for every t ,= 0, A
it
is unbounded.
Proof. It is easily seen that the domain of A is the subset of c
0
consisting of the sequences
=
n
such that

n=1
n
= 0 (which is dense in c
0
), and
A(
1
,
2
,
3
, . . .) = (
1
,
1
+
2
,
1
+
2
+
3
, . . .),
An easy computation shows that for > 0
(I +A)
1
=
I
+ 1

1
( + 1)
2
k=1
_

+ 1
_
k1
S
k
,
so that
|(I +A)
1
|
1
+ 1
_
1 +
1
+ 1
k=1
_

+ 1
_
k1
_
2
+ 1
,
which implies that A is a positive operator. Replacing in (4.2), for Re < 0, and then
also for Re = 0, we get
A
= I +S +
( + 1)
2
S
2
+
( + 1)( + 2)
3!
S
3
+. . .
So, for = it the n-th component of A
it
is
n
+it
n1
+
it(it + 1)
2

n2
+. . . +
it(it + 1) . . . (it +n 2)
(n 1)!

1
.
Fix any n N and dene a sequence D(A) as follows:
n
= 0,
n1
=
1
i
,
n2
=
1
i(it + 1)
,
n3
=
2
i(it + 1)(it + 2)
, . . . ,
1
=
(n 2)!
i(it + 1) . . . (it +n 2)
,
while for k > n
k
is arbitrary, subject only to [
k
[ 1 and

k=1
k
= 0. Then we get
|A
it
| [(A
it
)
n
[ = [t[
_
1 +
1
2
+
1
3
+. . . +
1
n 2
_
while the norm of is 1. Letting n we obtain sup
D(A), ||=1
|A
it
| = +, so that
A
it
is unbounded.
Let us discuss the behavior of A
z
for Re z < 0, z close to the imaginary axis. From
the representation formula (4.4) we get easily, using (4.5),
|A
z
|
M
[ sin(z)[
_

0
Re z
+ 1
d M
[ sin(z)[
[ sin(Re z)[
, Re z (1, 0). (4.14)
In particular, for real z = , with 0 < < 1 we get |A
| M, so that |A
| is
bounded on the real interval (1, 0), and hence it is bounded on any real interval (a, 0),
a > 0. But in general |A
z
| may be unbounded in other subsets of the left halfplane
such that iR ,= . However, if the operator A
it
is bounded for t I R and |A
it
| C
for every t I, then for z = +it we have
|A
+it
| |A
| |A
it
| MC, 0 < < 1/2, t I.
A sort of converse of the above considerations is in the next lemma. It gives a simple
(but hard to be checked) sucient condition for A
it
to be bounded.
Lemma 4.2.2 Let A be a positive operator with dense domain D(A). Assume that there
are a set z C : Re z < 0 and a constant C > 0 such that iR ,= and
|A
z
| C for z . Then for every t R such that it , A
it
is a bounded operator
and |A
it
| C.
Proof. For every x D(A) the function z A
z
x is continuous for Re z < 1, so that
lim
zit, z
A
z
x = A
it
x. Since D(A) is dense and |A
z
| C for z , it follows that for
every x X there exists the limit lim
zit, z
A
z
x. Denoting such a limit by Tx, we get
|Tx| C|x|. Then A
it
is a closed operator which coincides with the bounded operator
T on a dense subset. This implies that A
it
= T so that A
it
is bounded.
The most popular examples of positive unbounded operators with bounded imaginary
powers are m-accretive positive operators in Hilbert spaces, which will be discussed in
the next section. Self-adjoint positive operators belong to this class. Another interesting
example is the following.
78 Chapter 4
Example 4.2.3 Let A : D(A) X be any positive operator, and let 0 < < 1, 1
p . Then the parts of A in (X, D(A))
,p
and in (X, D(A))
have bounded imaginary

powers.
Proof. It is not hard to check (see exercise 1, 4.2.1) that the part of A in any of the
above spaces is still a positive operator.
First we consider the case p = . Let A
: D
A
( +1, ) D
A
(, ) be the part of A
in D
A
(, ). We already know (remark 4.1.8) that D
A
(, ) = (X, D(A))
,
is contained
in D(A
it
). Therefore for every x D
A
(, ), A
it1
x belongs to the domain D(A). To
obtain an estimate for |A
it
x| = |A(A
it1
x)| we use the representation formula (4.8) for
A
it1
x,
A
it1
x =
1
(1 it)(1 +it)
_

0
it
(A+I)
2
xd
=
sin(it)
it
_

0
it
(A+I)
2
xd,
which gives
|A
it
x|
e
t
e
t
2t
_

0
M
(1 +)
|
A(A+I)
1
x| d
C()
e
t
e
t
t
|x|
(X,D(A))
,
.
We prove now that in fact A
it1
x belongs to D
A
(+1, ). We use again the representation
formula (4.8) for A
it1
x, which implies that for every > 0
|
A(I +A)
1
A(A
it1
x)|
e
t
e
t
2t
_
_
_
_
(I +A)
1
_

0
it
A(A+I)
1
A(A+I)
1
xd
_
_
_
_
+
e
t
e
t
2t
_
_
_
_
A(I +A)
1
_

it
(A+I)
1
A(A+I)
1
xd
_
_
_
_
e
t
e
t
2t
_
M
+ 1
(M + 1)
1
1
+ (M + 1)
_
|x|
(X,D(A))
,
C
t
e
t
e
t
t
|x|
(X,D(A))
,
.
Therefore, A
it1
x belongs to D
A
( + 1, ), which is the domain of A
in D
A
(, ). It
follows that x is in the domain of A
it
, and
|A
it
x|
(X,D(A))
,
C
tt
e
t
e
t
t
|x|
(X,D(A))
,
.
The rest of the statement follows by interpolation: knowing that for 0 <
1
<
2
< 1
the part of A
it
in (X, D(A))
1
,
and in (X, D(A))
2
,
is a bounded operator, from theorem
1.1.6 it follows that for every (
1
,
2
) the part of A
it
in (X, D(A))
,p
and in (X, D(A))
is a bounded operator.
Another important example follows from the so called transference principle. See
CoifmanWeiss [16].
Theorem 4.2.4 Let (, ) be a -nite measure space, and let 1 < p < . If A is a
positive operator in L
p
(, ) such that |(I + A)
1
| 1/ and (I + A)
1
is positivity
preserving for > 0 (i.e. f(x) 0 a.e. implies ((I + A)
1
f)(x) 0 a.e), then the
operators A
it
are bounded in L
p
(, ), and there is C > 0 such that
|A
it
| C(1 +t
2
)e
[t[/2
, t R.
Let us come back to the general theory. Due to theorem 4.1.6, if the operators A
it
are bounded for any t in a small neighborhood of 0, then they are bounded for every
t R. Moreover, if |A
it
| C for t , then there exists C
t
, > 0 such that
|A
it
| C
t
e
[t[
for every t R.
Lemma 4.2.5 Let A be a positive operator such that A
it
L(X) for every t R, and
t |A
it
| is locally bounded. Then for every x D(A) the function z A
z
x is continuous
in the closed halfplane Re z 0.
Proof. If x D(A) then z A
z
x is holomorphic for Re z < 1, so that it is obviously
continuous for Re z 0. We have already remarked that (4.14) implies |A
| M for
1/2 < < 0, so that for z = +i, 1/2 < < 0, we get
|A
z
| M|A
i
|.
In particular, for every t R and r > 0 small enough the norm |A
z
A
it
| is bounded in
the half circle z : [z it[ r, Re z 0, by a constant independent of z. It follows that
for every x D(A), lim
zit
A
z
x = A
it
x.
Note that if x / D(A) the function z A
z
x cannot be continuous in the halfplane
Re z 0. Indeed by proposition 4.1.7, A
z
x (X, D(A))
Re z,
for 1 < Re z < 0, and
(X, D(A))
Re z,
is contained in D(A).
The family of operators A
it
: t R plays an important role also in the interpolation
properties of the domains D(A
z
).
Theorem 4.2.6 Let A be a positive operator with dense domain such that for every t R
A
it
L(X), and there are C, > 0 such that
|A
it
| Ce
[t[
, t R.
Then for 0 Re < Re
[D(A
), D(A
)]
= D(A
(1)+
).
Proof. Thanks to theorem 4.1.6 we may assume that = 0 without loss of generality.
Moreover since A
it
is bounded for every t R, then D(A
) = D(A
Re
) for Re > 0. See
exercise 2, 4.2.1. So we may also assume that (0, ).
Let x D(A
), and set
f(z) = e
(z)
2
A
(z)
x, 0 Re z 1.
80 Chapter 4
Let us prove that f T(X, D(A
)). f is obviously holomorphic in the strip Re z (0, 1)

and continuous up to Re z = 1 with values in X. Since D(A) is dense in X, f is also
continuous up to Re z = 0 with values in X. Indeed, A
(z)
x = A
z
A
x, and we
know from lemma 4.2.5 that w A
w
y is continuous with values in X for Re w 0 for
every y D(A) = X. Similarly, t f(1 + it) is continuous with values in D(A
). f is
also bounded, since
|A
(z)
x| = |A
Imz
A
Re z
A
x| |A
Re z
| Ce
[Imz[
|A
z|
Therefore, f T(X, D(A
)). Since f() = x, then x [X, D(A
)]
and
|x|
[X,D(A
)]
maxsup
tR
|e
t
2
+
2
A
(it)
x|,
sup
tR
|e
t
2
+(1)
2
A
(1+it)
x|
D(A
C
t
|A
x|.
It follows that D(A
) is continuously embedded in [X, D(A
)]
.
Conversely, let x D(A
), and let f T(X, D(A
)) be such that f() = x.

The function
F(z) = e
(z)
2
A
z
f(z),
is continuous with values in X both for Re z = 0 and for Re z = 1, and we have
sup
tR
|F(it)| sup
tR
e
t
2
+
2
Ce
[t[
sup
tR
|f(it)| C
t
|f|
T(X,D(A
))
, (4.15)
sup
tR
|F(1 +it)| sup
tR
e
t
2
+(1)
2
Ce
[t[
sup
tR
|A
f(1 +it)| C
t
|f|
T(X,D(A
))
, (4.16)
so that F is bounded with values in X for Re z = 0 and for Re z = 1. If F would
be holomorphic in the interior of S and continuous in S, we could apply the maximum
principle to get |A
x| C
t
|f|
T(X,D(A
))
. But in general F is not even dened in the
interior of S, because f has values in X and not in the domain of some power of A. So
we have to modify this procedure.
By remark 2.1.5,
|x|
[X,D(A
)]
= inf|f|
T(X,D(A
))
: f 1(X, D(A
)), f() = x.
So, let f 1(X, D(A
)) be such that f() = x. The function

F(z) = e
(z)
2
A
z
f(z), 0 Re z 1,
is now well dened and holomorphic for Re z (0, 1), continuous with values in X up
to Re z = 0, Re z = 1, and bounded with values in X. By the maximum principle (see
exercise 1, 2.1.1),
|A
x| = |f()| maxsup
tR
|F(it)|, sup
tR
|F(1 +it)|
C
t
|f|
T(X,D(A
))
,
where the last inequality follows from estimates (4.15) and (4.16).
Taking the inmum over all the f 1(X, D(A
)) such that f() = x we get

|A
x| C
t
|x|
[X,D(A
)]
, x D(A
).
Since D(A
) is dense in [X, D(A
)]
it follows that [X, D(A
)]
is continuously embedded
in D(A
).
4.2.1 Exercises
In exercises 19, A is a positive operator in general Banach space X.
1) Let Y be any interpolation space between X and D(A). Prove that the part of A in Y
is a positive operator.
2) Let Re > 0, t R. Show that D(A
) = D(A
+it
) with equivalence of the norms if
and only if A
it
is bounded.
3) Prove that (4.8) holds for 1 < Re < 1.
4) Show that for Re > 1, Re / N, D(A
) belongs to J
Re
(D(A
[Re ]
), D(A
[Re ]+1
))
and to K
Re
(D(A
[Re ]
), D(A
[Re ]+1
)), where Re and [Re ] are the fractional part
and the integral part of Re , respectively.
5) Prove that for every > 0, A
is a positive operator, and that

R(, A
) =
1
2i
_
r,
R(z, A)
z
dz, 0.
6) Prove that if : D() X X is a linear closed operator, and B : X X is a linear
bounded operator, then C : D(C) = x X : Bx D() X, Cx = Bx, is a closed
operator. (This is used to check that A
it
is a closed operator for each t R).
7) Prove that if t, s R and x D(A
is
) D(A
i(t+s)
) then A
is
x D(A
it
) and A
it
A
is
x =
A
i(t+s)
x.
8) Improve the estimate of proposition 4.2.3 showing that for each < arctan1/M there
is C such that
|A
it
| Ce
()[t[
, t R.
Hint: instead of using formula (4.9), modify formula (4.3) for A
it1
x letting only r 0
and leaving < arctan1/M xed.
9) Prove that A
1
is a nonnegative operator, and that for 0 < Re < 1
A
= (A
1
)
.
10) Let A be a densely dened nonnegative operator such that R(A) is dense in X. Show
that R(A) D(A) is dense in X, and that the operators B
z
: D(A) R(A) X dened
in (4.11) are closable.
11) Let A be a nonnegative one to one operator, and set A
= (I + A)(A + I)
1
for
> 0. Show that (i) (A
) (, 0], (ii) (I +A
)
1
(I +A)
1
for > 0, A
x x
for x D(A), A
1
x A
1
x for x R(A) as 0, (iii) A
z
x A
z
x for x D(A)R(A)
as 0.
12) Let A be a positive operator in a Hilbert space H. Show that for each C,
(A
= (A
, so that if is real, then (A
= (A
.
13) Prove that the realization A of in L
p
(R
n
), 1 p , is a nonnegative operator.
Prove that if p < A is one to one, and that if p = the kernel of A consists of the
constant functions.
Hint: denoting by T(t) the Gauss-Weierstrass semigroup, use |D
i
T(t)|
L(L
p
)
C/
t
to show that if f = 0 then D
i
f = D
i
T(t)f vanishes for every i = 1, . . . , n, so that f is
constant.
82 Chapter 4
4.2.2 The sum of two operators with bounded imaginary powers
We consider now two positive operators, A : D(A) X X, B : D(B) X X,
having bounded imaginary powers and such that
|A
it
| Me
A
[t[
, |B
it
| Me
B
[t[
, t R. (4.17)
We also assume that A and B commute in the resolvent sense,
R(, A)R(, B) = R(, B)R(, A), , 0. (4.18)
Our assumptions imply immediately (through formula (4.2)) that A
z
B
w
= B
w
A
z
for Re
w, Re z < 0 and also (less immediately but easily) for Re w, Re z 0.
We shall study the closability and the invertibility of the operator A + B, following
the approach of Dore and Venni [18].
Proposition 4.2.7 Let A, B be positive operators satisfying (4.17) and (4.18). Assume in
addition that
A
+
B
< and that D(A) or D(B) is dense in X. Then A+B : D(A)D(B)
is closable, and its closure A+B is invertible with inverse S given by
S =
1
2i
_
a+i
ai
A
z
B
z1
sin(z)
dz =
1
2i
_
a+i
ai
B
z1
A
z
sin(z)
dz (4.19)
with any a (0, 1). Moreover, S is a left inverse of A+B.
Proof. Since
A
+
B
< the norm of the operator A
z
B
z1
/ sin(z) in the integral
decays exponentially as [Im z[ . Therefore S is a bounded operator, and since
z A
z
B
z1
/ sin(z) is holomorphic in the strip Re z (0, 1), S is independent of
a (0, 1).
The proof is in three steps: (i) we show that S is a left inverse of A+B; (ii) we show that
if D(A) (resp. D(B)) is dense, then for every (0, 1) S maps D(A
1
) (resp. D(B
1
))
to D(A)D(B) and (A+B)Sx = x for each x D(A
1
) (resp. x D(B
1
)); (iii) using
steps (i) and (ii) we show in a standard way that A+B is closable and S = A+B
1
.
(i) For every x D(A) D(B) it holds
S(Ax +Bx) =
1
2i
_
a+i
ai
_
B
z1
A
1z
x
sin(z)
+
A
z
B
z
x
sin(z)
_
dz
=
1
2i
_
a1+i
a1i
B
z
A
z
x
sin(z)
dz +
1
2i
_
a+i
ai
A
z
B
z
x
sin(z)
dz.
The function g : = z C : 1 < Re z < 1 X dened by
g(z) =
_
_
_
B
z
A
z
x, 1 < Re z 0,
A
z
B
z
x, 0 Re z < 1,
is holomorphic in the strips 1 < Re z < 0, 0 < Re z < 1 and continuous up to the
imaginary axis (see exercise 2, 4.2.3). Therefore it is holomorphic in the whole strip .
It follows that z g(z)/ sin(z) is holomorphic in 0. Moreover it has a simple pole
at z = 0, with residue x/, and it decays exponentially, uniformly for 1 < Re z < 1, as
[ Im z[ . It follows that
S(A+B)x = Res
_
g(z)
sin(z)
, 0
_
= x.
(ii) Assume for instance that that D(A) is dense. Fix x D(A
1
), with 0 < < 1 and
choose a = in the denition of S. The function z A
1z
B
z1
x/ sin(z) is well dened
because B
z1
maps D(A
1
) = D(A
1z
) into itself, and it is integrable over +iR, since
|A
1z
B
z1
x| = |A
z
B
z1
A
1
x| Ce
(
A
+
B
)[Imz[
|A
1
x|.
Therefore, Sx D(A) and
ASx =
1
2i
_
a+i
ai
A
1z
B
z1
x
sin(z)
dz =
1
2i
_
a+i
ai
B
z1
A
1z
x
sin(z)
dz.
To show that Sx belongs to D(B), we remark that the function z B
z1
A
z
x/ sin(z)
is holomorphic for 1 < Re z < 1, z ,= 0, continuous up to Re z = 1 (because D(A)
is dense, see lemma 4.2.5), and it decays exponentially as [Imz[ . Therefore, we may
shift the vertical line Re z = a to Re z = 1 in the denition of S, to get
Sx =
1
2i
_
1+i
1i
B
z1
A
z
x
sin(z)
dz +Res
_
B
z1
A
z
x
sin(z)
, 0
_
=
1
2i
_
1+i
1i
B
z1
A
z
x
sin(z)
dz +B
1
x.
As easily seen, the integral denes an element of D(B). Therefore Sx D(B) and
BSx =
1
2i
_
1+i
1i
B
z
A
z
x
sin(z)
dz +x
=
1
2i
_
+i
i
B
z1
A
1z
x
sin(z)
dz +x = ASx +x.
(iii) Let us show that A + B is closable. Let x
n
D(A) D(B) be such that x
n
0,
(A+B)x
n
y as n . Since S is a left inverse of A+B then
0 = lim
n
x
n
= lim
n
S(A+B)x
n
= Sy.
Since A
1
y D(A) D(A
1
) for each > 0, then by step (ii) SA
1
y D(A) D(B)
and
A
1
y = (A+B)SA
1
y = (A+B)A
1
Sy = 0,
so that y = 0 and A+B is closable.
As a last step we prove that A+B is invertible and its inverse is S. If x D(A+B)
there is a sequence x
n
D(A) D(B) such that x
n
x and (A + B)x
n
A+Bx as
n . Then
x = lim
n
x
n
= lim
n
S(A+B)x
n
= SA+Bx,
84 Chapter 4
which means that S is a left inverse of A+B. Now for x X let x
n
D(A) be such
that x
n
x as n . By step (ii) Sx
n
D(A) D(B), lim
n
Sx
n
= Sx and
lim
n
(A+B)Sx
n
= lim
n
x
n
= x. This implies that Sx D(A+B) and A+BSx =
x, so that S is also a right inverse of A+B.
Under the assumptions of proposition 4.2.7 the operator A+B is not closed in general.
If we knew that A + B is closed, it would follow that A + B is invertible with bounded
inverse. In the next proposition we use this fact to get information on the resolvent set of
a positive operator with bounded imaginary powers.
Proposition 4.2.8 Let A be a positive operator with bounded imaginary powers, satisfying
|A
it
| Ce
A
[t[
, t R.
If 0
A
< , the resolvent set of A contains the sector C : [arg ( )[ <
A
.
Proof. We apply proposition 4.2.7 to the operators A, B = I for every in the sector.
If = e
i
with > 0, (, ), then |B
it
| = |()
it
I| = e
t
, so that
B
= =
arg (). Since D(B) = X and A + B = A I is closed, the statement follows from
proposition 4.2.7.
Under a further suitable assumption on the space X, the conditions of proposition 4.2.7
are sucient for A+B to be closed. Such assumption has several equivalent formulations.
The geometric formulation is the following: there exists a symmetric function :
XX R which is convex in each variable, such that (0, 0) > 0, and (x+y) |x+y|
whenever |x| 1 |y|. In this case the space X is said to be -convex.
The probabilistic formulation is the following: there are p (1, ), C > 0 such that
for every probability space (, T, P) and for every martingale u
k
: X with respect
to any ltration T
k
, for each choice of
k
1, 1, and for each n N we have
|
n
k=0
k
(u
k
u
k1
)|
L
p
(;X)
C|
n
k=0
(u
k
u
k1
)|
L
p
(;X)
(where we set u
1
= 0). In this case X is said to be a UMD space, or to have the property
of unconditionality of martingale dierences.
The formulation which is useful here is the following: for some p (1, ) and > 0,
the truncated Hilbert transform
(H
f)(t) =
1
_
[y[
f(t y)
y
dy, t R
is a bounded operator from L
p
(R; X) to itself. Then it is possible to show that this is true
for every p (1, ) and > 0. Moreover for each f L
p
(R; X) there exists the limit
lim
0
H
f
in L
p
(R; X) and a.e. pointwise. Such a limit is denoted by Hf and it is called the Hilbert
transform of f.
Equivalence of the above properties is not trivial. Bourgain [6] showed that any UMD
space has the Hilbert transform property. The converse was proved by Burkholder [11],
who also proved in [10] that X is -convex i it is a UMD space.
Theorem 4.2.9 Let X be a -convex space, and let A, B be densely dened
(1)
positive
operators in X satisfying the assumptions of proposition 4.2.7. Then A + B : D(A)
D(B) X is closed, and 0 (A+B).
Proof. After proposition 4.2.7, we have only to show that S maps X into D(A) D(B).
This obviously implies that S is a right inverse of A+B, since by proposition 4.2.7 S is a
right inverse of A+B. Again by proposition 4.2.7, S is a left inverse of A+B. Therefore
S is the inverse of A+B and 0 (A+B).
Let us show that S maps X into D(B). For 0 < < 1/2 we have
Sx =
1
2i
_
A
z
B
z1
x
sin(z)
dz
where
is the curve is : [s[ z C : [z[ = , Re z 0, oriented with increasing

imaginary part. So,
Sx =
1
2
_
[s[
A
is
B
is1
x
sin(is)
ds +
1
2
_
/2
/2
e
i
sin(e
i
)
A
e
i
B
e
i
1
xd
= I
1,
+I
2,
.
Since D(A) is dense, I
2,
goes to B
1
x/2 as 0. Moreover I
1,
is in D(B) for every .
We reach our goal if we show that I
1,
converges in D(B) as 0, i.e. if BI
1,
converges
in X as 0. Indeed, in that case we have
Sx = B
1
x/2 + lim
0
I
1,
D(B).
Let us split BI
1,
into the sum
BI
1,
=
1
2
_
[s[1
A
is
B
is
x
sin(is)
ds +
1
2
_
[s[1
A
is
B
is
x
is
ds
+
1
2
_
[s[1
A
is
B
is
x
_
1
sin(is)

1
is
_
ds.
The rst term is independent of , the third one is easily seen to converge as 0 because
1/(sin(is)) 1/(is) is bounded. The second term is equal to 1/(2i)H
f(0) where
f(s) =
(1,1)
(s)A
is
B
is
x, s R.
Since f L
p
(R; X) for each p > 1 and X is -convex, the truncated Hilbert transform of
f converges in X for almost every t R. Let us prove that 0 is one of such ts. Fix once
and for all t (0, 1) such that H
f(t) converges. Then

(H
f)(0) = A
it
B
it
_
[s[1
A
i(ts)
B
i(st)
x
s
ds = A
it
B
it
__
[s[
f(t s)
s
ds +
_
t1
1
A
i(ts)
B
i(st)
x
s
ds
_
t+1
1
A
i(ts)
B
i(st)
x
s
ds
_
1
Every -convex space X is reexive, and therefore by [27] all positive operators in X are densely
dened.
86 Chapter 4
converges as 0. This concludes the proof that Sx D(B). In the same way one
shows that Sx D(A), and the statement follows.
The main application at least, from our point of view of the theorem of Dore and
Venni is to evolution equations in a -convex space X,
_
_
_
u(t) +Au(t) = f(t), 0 < t < T,
u(0) = 0,
(4.20)
with f L
p
(0, T; X) for some p (1, ), T > 0. Here we assume that A : D(A) X X
is a positive operator having bounded imaginary powers, and
|A
is
| Ce
A
[s[
, s R.
Then it is not hard to see that the operator / dened by
/ : L
p
(0, T; D(A)) A = L
p
(0, T; X), (/f)(t) = Af(t)
is positive and has bounded imaginary powers, given of course by
(/
is
f)(t) = A
is
f(t), s R, f A,
and
|/
is
| Ce
A
[s[
, s R.
It is also possible to show that the operator
B : D(B) = f W
1,p
(0, T; X) : f(0) = 0 A, (Bf)(t) = f(t),
is positive and has bounded imaginary powers, satisfying the estimate
|B
is
| C(1 +[s[
2
)e
[s[/2
, s R.
See [18]. Therefore, if
A
< /2 it is possible to apply theorem 4.2.9 to equation (4.20),
seen as the equation in A
/u +Bu = f,
getting that for every f L
p
(0, T; X) problem (4.20) has a unique solution u W
1,p
(0, T;
X) L
p
(0, T; D(A)), which depends continuously on f.
For instance, if A is the realization of in L
p
(), 1 < p < , with Dirichlet
boundary condition:
D(A) = W
2,p
() W
1,p
0
(), Au = u,
being an open bounded set in R
n
with regular boundary, then A is a positive operator
with bounded imaginary powers, and
A
, /2 due to [34]. We get that for each f
L
p
((0, T) ) the problem
_
_
u
t
(t, x) = u(t, x) +f(t, x), 0 < t < T, x ,
u(t, x) = 0, 0 < t < T, x ,
u(0, x) = 0, x ,
has a unique solution u W
1,2
p
((0, T) ), i.e. u, u
t
, D
ij
u L
p
((0, T) ).
4.2.3 Exercises
1) Let A, B be two positive operators with bounded imaginary powers, satisfying (4.18).
Show that A
z
B
w
= B
w
A
z
for Re z, Re w 0. Show that for Re z 0, A
z
maps D(B)
into itself, and A
z
Bx = BA
z
x for each x D(B). Show that for Re z 0, A
z
maps D(B)
into itself.
2) Let A, B be two positive operators with bounded imaginary powers, satisfying (4.17)
and (4.18). Show that for every x D(A) D(B) the functions z B
z
A
z
x, 1 <
Re z 0, and z A
z
B
z
x, 0 Re z < 1, are continuous (this is used in the proof of
proposition 4.2.7).
4.3 M-accretive operators in Hilbert spaces
Throughout this section H is a complex Hilbert space, and A : D(A) H H is a linear
operator satisfying
D(A) = H, (A) (, 0), |(I +A)
1
| 1/, > 0. (4.21)
Therefore, A is a nonnegative operator. Moreover it satises the resolvent estimate with
constant M = 1, and this is not a mere notational simplication but it is a crucial as-
sumption.
Due to the Hille-Yosida theorem, assumption (4.21) is equivalent to the hypothesis
that A be the innitesimal generator of a contraction semigroup e
tA
. Therefore, for
each x D(A),
Re Ax, x) = lim
t0
Re
e
tA
x x
t
, x) 0.
Any operator B : D(B) H H satisng
Re Bx, x) 0, x D(B)
is called accretive. Therefore, A is accretive.
It is possible to show that A is m-accretive (maximal accretive), in the sense that it
has no proper accretive extension, and conversely, any closed m-accretive operator satises
(4.21). So, operators satisfying (4.21) are often referred in the literature as m-accretive
operators.
It is easy to see that A satises (4.21) i A
does. Moreover, if A satises (4.21) and

it is one to one, then the range of A is dense in H.
We shall follow the approach of Kato [25, 26] to study the imaginary powers of A and
the relationship between the domains of A
and (A
.
First we consider m-accretive bounded operators satisfying in addition
Re Ax, x) |x|
2
, x H. (4.22)
for some > 0. The case of unbounded operators will be reduced to this one, through the
use of the Yosida approximations nA(nI +A)
1
.
Assumption (4.22) implies that the resolvent set of A contains (, ), and |(I +
A)
1
| ( + )
1
for every 0. Therefore A is a positive bounded operator, so that
for each z C the complex powers A
z
are dened by
A
z
=
1
2i
_
z
R(, A)d, (4.23)
88 Chapter 4
where is any regular curve surrounding (A) with index 1 with respect to every point
of (A), and avoiding (, 0]. Moreover we have
(A
= (A
, C.
See exercise 12, 4.2.1.
We will need the following lemma.
Lemma 4.3.1 Let A be a bounded m-accretive operator satisfying (4.22). Then for real
[1, 1], A
satises
Re A
x, x)
|x|
2
, 0 1, x H, (4.24)
Re A
x, x) (|A|
2
)
|x|
2
, 1 0, x H. (4.25)
Proof. For 0 < < 1 we use the Balakrishnan formula (4.7), which implies
A
x, x) =
sin()
_

0
1
A(I +A)
1
x, x)d.
Recalling that
ReA(I +A)
1
x, x) = Re(I (I +A)
1
)x, x)
|x|
2
[(I +A)
1
x, x)[ |x|
2

+
|x|
2
=

+
|x|
2
,
we obtain, through (4.5),
A
x, x)
sin()
_

0
1
+
|x|
2
d =
|x|
2
,
i.e. (4.24) holds. Moreover,
ReA
1
x, x) = ReA
1
x, AA
1
x) |A
1
x|
2
|A|
2
|x|
2
,
so that (4.25) holds for = 1. But using again the representation formula (4.23) we see
easily that
A
= (A
1
)
, C,
so that (4.25) holds for every [1, 0].
Theorem 4.3.2 Let A : H H be a bounded operator. Assume that there exists > 0
such that (4.22) holds. Then for each [0, 1/2) and for each x H
|(A
x| c
|A
x|, |A
x| c
|(A
x|,
with c
= tan(1 + 2)/4. Moreover,

|A
it
| e
[t[/2
, t R.
Proof. Let us introduce the real part and the imaginary part of A
dened by
H
=
A
+ (A
2
, K
=
A
(A
2i
,
i.e.
A
= H
+iK
, (A
= H
iK
.
Then for every x H
|H
x|
2
|K
x|
2
= Re A
x, (A
x) = Re A
+
x, x).
If 1/2 Re 1/2, then A
+
still satises (4.22) with replaced by the constant
= min
, (|A|
2
)
thanks to lemma 4.3.1. Therefore,

|H
x|
2
|K
x|
2
+
|x|
2
, x H. (4.26)
This implies that H
is one to one. Since H
= (A
+ (A
)/2, and A
satises the
assumptions of the theorem, also H
is one to one. Therefore H
is invertible, and since

|K
y| |H
y| for every y because of (4.26), then |K
H
1
x| |H
H
1
x| = |x| for
each x, so that
|K
H
1
| 1, 1/2 Re 1/2.
Now we improve this estimate applying the maximum principle to the function
() =
K
H
1
tan(/2)
, 1/2 Re 1/2.
Such a function is holomorphic in the whole strip (even at = 0, because K
0
= 0)
and bounded with values in L(H); moreover [ tan(/2)[ = 1 if Re = 1/2 so that
|()| 1 on the boundary of the strip. Therefore |()| 1 for each in the strip,
i.e.
|K
H
1
| tan(/2), 1/2 Re 1/2. (4.27)

This is an important improvement of |K
H
1
| 1, because [ tan(/2)[ < 1 for [Re

[ < 1/2, so that I iK
H
1
is invertible with bounded inverse.

Since A
= H
+iK
, (A
= H
iK
, we get for [Re [ < 1/2

|(A
| = |(I iK
H
1
)(I +iK
H
1
)
1
|
1 +[ tan(/2)[
1 [ tan(/2)[
. (4.28)
Therefore for real [0, 1/2)
|(A
x|
1 + tan(/2)
1 tan(/2)
|A
x| = tan
(1 + 2)
4
|A
x|, x H,
and we get a similar inequality exchanging A with A
. So, the rst statement is proved.

Moreover, taking = it with real t, in (4.28) we get
|A
it
x|
2
= (A
)
it
A
it
x, x)
1 +[ tan(it/2)[
1 [ tan(it/2)[
|x|
2
= e
[t[
|x|
2
,
and the last statement follows.
Theorem 4.3.2 is the starting point to show several properties of the powers of general
m-accretive operators. The rst property concerns the equivalence of the domains D(A
),
D((A
) for 0 < 1/2. But we need a preliminary lemma.

90 Chapter 4
Lemma 4.3.3 Let A : D(A) H H satisfy (4.21). For every n N set
J
n
=
_
I +
A
n
_
1
= n(nI +A)
1
.
Then J
n
is a bounded nonnegative operator, and for every [0, 1]
|J
n
| 1,
lim
n
J
n
x = x, x H.
Proof. As easily seen, for every > 0 I +J
n
is invertible with (bounded) inverse
(I +J
n
)
1
= (nI +A)(n( + 1) +A)
1
.
For every > 0 the operator I +A is positive. The representation formula (4.6) gives
(I +A/n)
=
sin()
_

0
(I +I +A/n)
1
d,
where |(I +I +A/n)
1
| 1/( + 1), so that due to (4.5)
|(I +A)
|
sin()
_

0
1
( + 1)
d = 1.
It follows (see exercise 9, 4.2.1)
|(I +A/n)
| = |((I +A/n)
1
)
| = |J
n
| 1,
and by the dominated convergence theorem
lim
n
J
n
x = lim
n
(I +A/n)
x = lim
n
sin()
_

0
+ 1
xd = x, x H.
Theorem 4.3.4 Let A : D(A) H H be any m-accretive operator. Then for every
[0, 1/2), D(A
) = D((A
) and for each x in the common domain

|(A
x| tan
(1 + 2)
4
|A
x|, |A
x| tan
(1 + 2)
4
|(A
x|.
Proof. We know already (theorem 4.3.2) that the statement is true if A is bounded and
satises (4.22) for some > 0. As a second step, we prove that the statement is true if A
satises (4.21) and 0 (A). Then it will follow easily that the statement is true in the
general case.
Let 0 (A). Let us consider the Yosida approximations,
A
n
= AJ
n
= nA(nI +A)
1
, n N.
It is not hard to see that the operators A
n
are bounded (with |A
n
| n), m-accretive
(with (I +A
n
)
1
= n
2
/(n+)
2
(n/(n+)I +A)
1
+I/(n+)), and satisfy (4.22) since
A
n
x, x) = AJ
n
x,
1
n
(nI +A)J
n
x) = AJ
n
x, J
n
x) +
1
n
|A
n
x|
2
and A
n
is invertible, with A
1
n
= A
1
+I/n. Therefore by theorem 4.3.2 we have
_
_
|(A
n
)
| tan
(1 + 2)
4
|A
n
|,
|A
n
| tan
(1 + 2)
4
|(A
n
)
|,
(4.29)
for every n N.
Note that since I + A/n is a positive operator, then (I + A/n)
is well dened for

every 0, and by exercise 9, 4.2.1, (I +A/n)
= J
n
for 0 1. Therefore,
J
n
= (A
1
A
n
)
= A
n
A
= A
n
,
as it is easy to check. Therefore for every x D(A
), J
n
x D(A
), and we have
A
n
x = A
n
x = J
n
A
x, 0 1.
Now we use lemma 4.3.3, which states that |J
n
| 1, and lim
n
J
n
x = x for each x.
We get
(i) |A
n
x| |A
x|, (ii) lim

n
A
n
x = A
x, x D(A
). (4.30)
Using (4.29) for 0 < 1/2 and then (4.30)(i) we get
|(A
n
)
x| tan
(1 + 2)
4
|A
n
x| tan
(1 + 2)
4
|A
x|, (4.31)
so that the sequence (A
n
)
x is bounded for each x D(A
). Moreover, for every y

D(A
) we have, due to (4.30)(ii),

(A
n
)
x, y) = x, A
n
y) x, A
y), n .
Since D(A
) is dense in H (because D(A) is dense and D(A
) D(A)), and |(A
n
)
x| is
bounded thanks to (4.31), then (A
n
)
x converges weakly to some w H. Such a w satises

w, y) = x, A
y) for every y D(A
), and this implies that x D((A
) = D((A
)
and w = (A
x. Therefore, the domain of A
is contained in the domain of (A
.
Exchanging the roles of A and A
we get that D(A
) = D((A
), and lim
n
(A
n
)
x =
(A
x for each x in the common domain. Letting n in (4.31), and in the similar
estimate with A
in the place of A, we get the claimed estimates.

Now we consider the case where 0 (A). For each > 0 the operator A+I satises
(4.21) and 0 belongs to its resolvent set, so D((A+I)
) = D((A
+I)
) for 0 < 1/2,

and for every x in the common domain we have
|(A
+I)
x| tan
(1 + 2)
4
|(A+I)
x|,
|(A+I)
x| tan
(1 + 2)
4
|(A
+I)
x|.
By lemma 4.1.11, D((A + I)
) = D(A
) and (A + I)
x A
x for each x D(A
),
and the same holds with A replaced by A
. Letting 0 the statement follows.

Let us consider now the imaginary powers A
it
, t R.
92 Chapter 4
Theorem 4.3.5 Assume that A : D(A) H H is m-accretive and that 0 (A).
Then the imaginary powers A
it
, t R, are bounded operators, and
|A
it
| e
[t[/2
, t R.
Proof. Theorem 4.3.2 implies that the statement is true if A is bounded and satises
(4.22).
Next step is to show that the statement is true if A is bounded, m-accretive and one
to one. This is done considering the operators A+I with > 0, which satisfy (4.22) with
= , and letting 0. Indeed, since (A+I)x x and (I +A+I)
1
x (A+I)
1
x
for each x H, and (A +I)
1
x A
1
x for each x R(A) as 0, we may let 0
in formula (4.11) getting
|A
it
x| = lim
0
|(A+I)
it
x| e
[t[/2
|x|, t R.
Since A
it
is closed, we may conclude that D(A
it
) = H and |A
it
| e
[t[/2
for every t R.
The fact that (A + I)
1
x A
1
x for each x R(A) as 0 may be proved as
follows: from the equality
A((A+I)
1
y A
1
y) +((A+I)
1
y A
1
y) = y,
which holds for each y H, we obtain
|A((A+I)
1
y A
1
y)|
2
+
2
|(A+I)
1
y A
1
y|
2
|y|
2
,
so that 0 = lim
0
A((A+I)
1
yA
1
y) = lim
0
(A+I)
1
AyA
1
Ay, i.e. lim
0
(A+
I)
1
x = A
1
x for every x = Ay in the range of A.
In the nal step we consider a general m-accretive operator A with 0 (A). Therefore
A
1
is bounded, m-accretive and one to one. It is not hard to see that
A
it
= (A
1
)
it
, t R.
But we already know that (A
1
)
it
is bounded, with norm not exceeding e
[t[/2
. The
statement follows.
Theorems 4.3.5 and 4.2.6 yield the next corollary.
Corollary 4.3.6 If A : D(A) H H is m-accretive and 0 (A) then for 0 Re <
Re we have
[D(A
), D(A
)]
= D(A
(1)+
).
4.3.1 Self-adjoint operators in Hilbert spaces
Here H is again a complex Hilbert space, and A : D(A) H H is a positive denite
self-adjoint operator. A self-adjoint operator is said to be positive if there is > 0 such
that
Ax, x) |x|
2
, x D(A). (4.32)
Lemma 4.3.7 If A is a self-adjoint operator satifying (4.32), then (A) contains (, )
and
|R(, A)|
1

, < . (4.33)
Proof. Let < , x D(A). Then
|(I A)x|
2
= |( )x + (I A)x|
2
= ( )
2
|x|
2
+ 2( )x, x Ax) +|(I A)x|
2
,
so that
|(I A)x|
2
( )
2
|x|
2
, (4.34)
and therefore I A is one to one. We prove that it is also onto, showing that its range
is both closed and dense in H. Let x
n
D(A) be such that x
n
Ax
n
converges. From
the inequality (4.34) we get
|(I A)(x
n
x
m
)|
2
( )
2
|x
n
x
m
|
2
, n, m N,
so that x
n
is a Cauchy sequence, x
n
is a Cauchy sequence, and Ax
n
is a Cauchy sequence.
Then x
n
and Ax
n
converge; let x, y be their limits. Since A is self-adjoint, then it is closed
(we recall that each adjoint operator is closed), so that x D(A), Ax = y, and x
n
Ax
n
converges to x Ax Range (I A). Therefore, the range of I A is closed.
Let y be orthogonal to the range of (I A). Then for each x D(A) we have
y, x Ax) = 0, so that y D(A
) = D(A) and y A
y = y Ay = 0. Since I A
is one to one, it follows y = 0. Therefore, the range of (I A) is dense.
Let us estimate |R(, A)|. For x H, let u = R(, A)x. From the equality x, u) =
u Au, u) it follows x, u) ( )|u|
2
0, so that
( )|u|
2
[x, u)[ |x| |u|
which implies |R(, A)| ( )
1
.
In this case it is possible to dene the powers A
z
for z C through the spectral
decomposition of A.
Theorem 4.3.8 Let A : D(A) H be a self-adjoint operator. There exists a unique
family of projections E
L(H), R, with the following properties:

(i) E
= E
= E
, for < ,
(ii) lim
t0
+ E
+t
x = E
x, R, x H,
(iii) lim
x = 0, lim
+
E
x = x, x H,
such that
D(A) = x H :
_
+
2
d|E
x|
2
< , Ax =
_
+
dE
x, x D(A).
The above integrals are meant as improper integrals of Stieltjes integrals (note that
|E
x|
2
= E
x, x) is nonnegative and nondecreasing). The family E
: R is
called the spectral resolution or spectral decomposition of A. See e.g. [33, Ch. VIII, sect.
120121].
If f : R C is continuous, the operator f(A) is dened by
D(f(A)) = x H :
_
+
[f()[
2
d|E
x|
2
< ,
94 Chapter 4
f(A)x, y) =
_
+
k
dE
x, y), x D(f(A)), y H.
It is possible to show (see [33, Ch. IX, sect. 126128]) that for every (A) we have
R(, A)x, y) =
_
+
infty
1

dE
x, y), y H,
and that this denition agrees with the usual denition in the case where f is a power
with integer exponent, i.e.
D(A
k
) = x H :
_
+
2k
d|E
x|
2
< , k Z,
A
k
x, y) =
_
+
k
dE
x, y), k Z, x D(A
k
), y H.
More generally, this denition agrees with the denition of the powers A
z
for any
z C.
Theorem 4.3.9 For every z C we have
D(A
z
) = x H : |A
z
x|
2
=
_
+
0
2Rez
d|E
x|
2
< , (4.35)
and
A
z
x =
_
+
0
z
dE
x, x D(A
z
). (4.36)
Proof. Let Re z < 0. Then
A
z
=
1
2i
_
r,
z
R(, A)d
1
2i
_
r,
z
_

0
1

dE
d =
_

0
1
2i
_
r,
z

d dE
=
_

0
z
dE
,
and the statement holds for Re z < 0. If Re z 0 let n > Re z, n N. By denition,
x D(A
z
) if and only if
A
zn
x =
_
+
0
zn
dE
x D(A
n
).
For each y H, y D(A
n
) i
_
+
0

2n
d|E
y|
2
< . Therefore, x D(A
z
) i
_
+
0
2n
d|E
A
zn
x|
2
=
_
+
0
2n
d|E
_

0
zn
dE
x|
2
_
+
0
2n
d|
_

0
zn
E
x|
2
=
_
+
0
2Re z
d|E
x|
2
<
and in this case
A
z
x, y) =
_
+
0
n
d
_
+
0
zn
dE
x, E
y)
=
_
+
0
n
d
_

0
zn
d
x, y) =
_
+
0
z
d
x, y).
The function E
is constant on each interval contained in the resolvent set of A. In

the case of a positive operator the spectrum of A is contained in (0, ), so that all the
above integrals are in fact integrals over (0, ).
It follows from the denition that if f is bounded, then f(A) is a bounded operator,
with norm not exceeding |f|
. If A satises (4.32), then (A) [, +), so that dening

f() =
it
for and extending f to a continuous bounded function to the whole R,
we obtain that A
it
is a bounded operator, and |A
it
| 1, for every real t.
Example 4.3.10 Let be a bounded open set with C
2
boundary, and let H = L
2
(),
A : D(A) = H
2
() H
1
0
(), Au = u. It is well known that (A) consists of a sequence
of positive eigenvalues, each of them with nite dimensional eigenspace, and there exists
an orthonormal basis of H consisting of eigenfunctions of A. Denoting such a basis by
e
n
nN
and denoting by
n
the eigenvalue with eigenfunction e
n
, we have
E
u =
n: n
u, e
n
)e
n
.
For every with positive real part, the domain of A
consists of those u L
2
() such
that
n=1
2Re
n
[u
n
[
2
< ,
where u
n
= u, e
n
), and
A
u =
n=1
n
u
n
e
n
, u D(A
).
Moreover, e
n
/
n
: n N is an orthonormal basis of H
1
0
(), with respect to the
scalar product Du, Dv) =
_
u(x)v(x)dx (see e.g. [8, IX.8]).

In the particular case = (0, ) it is easy to see that
e
n
(x) =
_
2/ sin(nx),
n
= n
2
.
Therefore, u D(A) = H
2
(0, )H
1
0
(0, ) i
n=1
n
4
[u
n
[
2
< , and Au =
n=1
n
2
u
n
e
n
.
Taking = 1/2, we get u D(A
1/2
) i

n=1
n
2
[u
n
[
2
< , that is i u H
1
0
(0, ), and
A
1/2
u =
n=1
nu
n
e
n
.
Let us come back to the general theory. If A satises (4.32), (4.35)-(4.36) could be
taken as a denition of A
z
, and all the properties of A
z
could be deduced, without invoking
any result of the previous sections. For instance, it follows immediately that for every
x D(A
) the function z A
z
x is holomorphic in the halfplane Re z < Re , that
D(A
z
) depends only on Re z, and if Re z
1
< Re z
2
then D(A
z
1
) D(A
z
2
).
96 Chapter 4
Most important, it is clear from (4.35) that for every t R, A
it
is a bounded op-
erator, and |A
it
| 1. Thanks to theorem 4.2.6, this implies that [D(A
), D(A
)]
=
D(A
(1)+
) for 0 Re < Re . Moreover, these spaces coincide also with the real
interpolation spaces (D(A
), D(A
))
,2
, as the next theorem 4.3.12 shows. For its proof
we need a lemma.
Lemma 4.3.11 Let L : D(L) H be a self-adjoint positive operator. Then iL generates
a strongly continuous group of contractions e
itL
in H.
Proof. Let us prove that the resolvent set of iL contains R 0.
Since L is self-adjoint, then Lx, x) R for each x D(L), so that Re iLx, x) = 0 for
each x D(L) = D(iL). This implies that for every > 0, I iL and I iL are one
to one, and |(I iL)
1
x| |x|/, for each x (I iL)(H), |(I +iL)
1
x| |x|/,
for each x (I +iL)(H). Indeed for > 0, x X, setting y = R(, iL)x we have
|y|
2
Re |y|
2
Re iLy, y) = Re x, y) |x| |y|
so that |y| = |R(, iL)x| |x|/.
To prove that I iL and I iL are onto it is enough to check that their ranges
are dense in H. Let y be orthogonal to the range of I iL, then x iLx, y) = 0 for
each x D(L), so that
y D(I (iL)
), y (iL)
y = y +iLy = 0,
so that y = 0. It follows that the range of I iL is dense in H. Similarly one shows that
the range of I +iL is dense in H. Consequently, R 0 (iL) and |R(, iL)| 1/,
and this implies the statement.
Theorem 4.3.12 Let , C with Re 0, Re 0. Then for every (0, 1)
[D(A
), D(A
)]
= (D(A
), D(A
))
,2
= D(A
(1)+
).
Proof. It is sucient to prove that for real > 0
(H, D(A
))
,2
= D(A
)
and the general statement will follow by interpolation and reiteration.
Since A
is in its turn a positive self-adjoint operator, iA
generates a strongly con-

tinuous group of operators
T(t) = e
itA
, t R.
By proposition 3.2.1 (H, D(A
))
,2
consists of the elements x such that t (t) =
t
|T(t)x x| L
2
(0, ). We have
||
2
L
2
(0,)
=
_

0
t
2
|T(t)x x|
2
dt
t
=
_

0
t
2
_

0
[e
it
1[
2
d|E
x|
2
dt
t
=
_

0
__

0
[e
it
1[
2
t
2
dt
t
_
d|E
x|
2
=
_

0
[e
i
1[
2
2+1
d
_

0
2
d|E
x|
2
= C|A
x|
2
,
so that (H, D(A
))
,2
= D(A
), with equivalence of the norms.

An important corollary about interpolation in Hilbert spaces follows.
Corollary 4.3.13 Let H
1
, H
2
be Hilbert spaces, with H
2
H
1
, H
2
dense in H
1
. Then
for every (0, 1),
[H
1
, H
2
]
= (H
1
, H
2
)
,2
.
Proof. It is known (see e.g. [33, Ch. VIII, sect. 124]) that there exists a self-adjoint
positive operator A in H
1
such that D(A) = H
2
. The statement is now a consequence of
theorem 4.3.12.
98 Chapter 4
Chapter 5
Analytic semigroups and
interpolation
Throughout the chapter X is a complex Banach space, and A : D(A) X X is a
sectorial operator, that is there are constants R, (/2, ), M > 0 such that
_
_
(i) (A) S
,
= C : ,= , [arg( )[ < ,
(ii) |R(, A)|
L(X)

M
[ [
S
,
.
(5.1)
(5.1) allows us to dene a semigroup e
tA
in X, by means of the Dunford integral
e
tA
=
1
2i
_
+r,
e
t
R(, A)d, t > 0,
where r > 0, (/2, ), and
r,
is the curve C : [arg[ = , [[ r
C : [arg[ , [[ = r, oriented counterclockwise. We also set e
0A
= I.
It is possible to show that the function t e
tA
is analytic in (0, +) with values in
L(X) (in fact, with values in L(X, D(A
m
)) for every m), so that e
tA
is called analytic
semigroup generated by A. One sees easily that for x X there exists lim
t0
e
tA
x if and
only if x D(A) (and in this case the limit is x). Therefore e
tA
is strongly continuous
if and only if D(A) is dense in X. In any case e
tA
L(X, D(A
m
)) for every t > 0 and
m N, and d
m
/dt
m
e
tA
= A
m
e
tA
for t > 0. Moreover there are constants M
k
> 0 such
that
_
_
_
(a) |e
tA
|
L(X)
M
0
e
t
, t > 0,
(b) |t
k
(AI)
k
e
tA
|
L(X)
M
k
e
t
, t > 0.
(5.2)
Note that for every x X and 0 < s < t, the function e
A
x is in C
1
([s, t]; X)
C([s, t]; D(A)), so that
e
tA
x e
sA
x =
_
t
s
Ae
A
xd = A
_
t
s
e
A
xd.
The same is true also for s = 0, in the sense specied by the following lemma.
Lemma 5.0.14 For every x X and t 0, the integral
_
t
0
e
sA
xds belongs to D(A), and
A
_
t
0
e
sA
xds = e
tA
x x.
99
100 Chapter 5
If in addition the function s Ae
sA
x belongs to L
1
(0, t; X), then
e
tA
x x =
_
t
0
Ae
sA
xds.
Other properties of sectorial operators and analytic semigroups may be found in [32,
Ch. 2].
5.1 Characterization of real interpolation spaces
Throughout the section we denote by M
0
, M
1
two constants such that |e
tA
|
L(X)
M
0
,
|tAe
tA
|
L(X)
M
1
, for every t (0, 1].
Since A is sectorial, the operator A I satises the assumptions of chapter 3, and
we may use the results of 3.1 to characterize the interpolation spaces (X, D(A
m
))
,p
.
Another characterization, which is very useful in abstract parabolic problems, was
found by Butzer and Berens (see e.g. [12]) for m = 1.
Proposition 5.1.1 For 0 < < 1, 1 p we have
(X, D(A))
,p
= x X : (t) = t
1
|Ae
tA
x| L
p
(0, 1),
and the norms | |
(X,D(A))
,p
and
x |x| +||
L
p
(0,1)
are equivalent.
Proof. For every x (X, D(A))
,p
let a X, b D(A) be such that x = a +b. Then
t
1
|Ae
tA
x| t
1
|Ae
tA
a| +t
1
|Ae
tA
b|
t
M
1
|a| +t
1
M
0
|Ab| maxM
0
, M
1
t
(|a| +t|b|
D(A)
)
so that
t
1
|Ae
tA
x| maxM
0
, M
1
t
K(t, x, X, D(A)),
which implies that L
p
(0, 1) with norm not exceeding M

0
, M
1
|x|
(X,D(A))
,p
max
M
0
, M
1
.
Conversely, if L
p
(0, 1) write x as
x = (x e
tA
x) +e
tA
x =
_
t
0
Ae
sA
xds +e
tA
x, 0 < t < 1.
Since t
1
|Ae
tA
x| L
p
(0, 1) the same is true for t

1
v(t), where v is the mean value
v(t) = t
1
_
t
0
|Ae
sA
x|ds, thanks to corollary A.3.1, and
|t t
1
v(t)|
L
p
(0,1)
=
_
_
_
_
t t
_
t
0
|Ae
sA
x|ds
_
_
_
_
L
p
(0,1)
||
L
p
(0,1)
.
Moreover,
|e
tA
x|
D(A)
= |e
tA
x| +|Ae
tA
x| M
0
|x| +t
1+
(t).
Analytic semigroups and interpolation 101
Therefore
t
K(t, x, X, D(A)) t
_
t
0
|Ae
sA
x|ds +t
1
M
0
|x| +(t),
so that t
K(t, x, X, D(A)) L
p
(0, 1), with norm not exceeding C(|x| + ||

L
p
(0,1)
).
We recall that it belongs to L
p
(1, ), with norm not exceeding C|x|. Therefore, x

(X, D(A))
,p
, and the statement follows.
Proposition 5.1.1 has the following generalization.
Proposition 5.1.2 For 0 < < 1, 1 p we have
(X, D(A
m
))
,p
= x X :
m
(t) = t
m(1)
|A
m
e
tA
x| L
p
(0, 1),
and the norms | |
(X,D(A
m
))
,p
and
x |x| +|
m
|
L
p
(0,1)
are equivalent.
Proof. (Sketch) Let m = 2. The embedding may be proved as in proposition 5.1.1,
splitting
t
m(1)
A
m
e
tA
x = t
m(1)
A
m
e
tA
a +t
m(1)
A
m
e
tA
b
if x = a +b. Also the idea of the proof of the other embedding is similar, but now instead
of the kernel Ae
sA
x we have to use sA
2
e
sA
x: we have
_
t
0
sA
2
e
sA
xds =
_
t
0
s
d
ds
Ae
sA
xds
=
_
t
0
Ae
sA
xds +tAe
tA
x = x e
tA
x +tAe
tA
x.
For every t (0, 1) we split x as
x =
_
t
0
sA
2
e
sA
xds +e
tA
x tAe
tA
x.
Since t
12
|tA
2
e
tA
x| L
p
(0, 1) the same is true for t

12
v(t), where v is the mean value
v(t) = t
1
_
t
0
|sA
2
e
sA
x|ds. Therefore
t g(t) = t
2
_
t
0
|sA
2
e
sA
x|ds L
p
(0, 1).
So,
t
2
K(t
2
, x, X, D(A
2
)) t
2
__
t
0
|sA
2
e
sA
x|ds +t
2
|e
tA
x tAe
tA
x|
D(A
2
)
_
g(t) +t
22
((M
0
+M
1
)|x| +|A
2
e
tA
x| + 2M
1
|A
2
e
tA/2
x|)
and t
2
K(t
2
, x, X, D(A
2
)) L
p
(0, 1), with norm less than C(|

2
|
L
p
(0,1)
+ |x|). Then
t
K(t, x, X, D(A
2
)) L
p
(0, 1), again with norm less than C

t
(|
2
|
L
p
(0,1)
+|x|), and the
statement follows for m = 2.
If m is arbitrary the procedure is similar, with the kernel s
m1
A
m
e
sA
x replacing
sA
2
e
sA
x.
In the case where m is not integer, proposition 5.1.2 may be deduced from proposition
3.1.8 and proposition 5.1.1.
102 Chapter 5
5.2 Generation of analytic semigroups by interpolation
In this section we shall use interpolation to check that certain operators are sectorial in
suitable functional spaces.
Theorem 5.2.1 Let Y be any interpolation space between X and D(A). Then the part
of A in Y , that is the operator
A
Y
: D(A
Y
) Y, D(A
Y
) = y D(A) : Ay Y , A
Y
y = Ay
is sectorial in Y .
In particular, for every (0, 1), 1 p . the parts of A in D
A
(, p), in D
A
(),
and in [X, D(A)]
are sectorial operators. Similarly, for every k N the part of A in

D
A
( +k, p) is sectorial in D
A
( +k, p).
Proof. Let S
,
. Since R(, A) commutes with A on D(A), then |R(, A)|
L(D(A))

M/[[. By interpolation it follows that R(, A) L(Y ) and |R(, A)|
L(Y )
M/[[,
In the next theorem we use the Stein interpolation theorem to prove generation of
analytic semigroups in L
p
spaces.
Theorem 5.2.2 Let (, ) be a -nite measure space, and let T(t) : L
2
() +L
()
L
2
() + L
() be a semigroup such that its restriction to L

2
() is a bounded analytic
semigroup in L
2
() and its restriction to L
() is a bounded semigroup in L
(). Then
the restriction of T(t) to L
p
() is a bounded analytic semigroup in L
p
(), for every p
(2, ).
Proof. Let
0
(0, /2) and M > 0 be such that T(t) has an analytic extension to the
sector
2
= z C : z ,= 0, [arg z[ <
0
, and |T(z)|
L(L
2
)
M for every z
2
,
|T(t)|
L(L
)
M for every t > 0.
Let S be the strip z C : Re z [0, 1]. For every r > 0 and (
0
,
0
) dene a
function h : S
2
by
h(z) = re
i(1z)
, z S,
and dene a family of operators
z
L(L
2
) by
z
= T(h(z)), z S.
Then z
z
is continuous and bounded in S, holomorphic in the interior of S, with
values in L(L
2
). Consequently, if a is a simple function on and b is a simple function
on , the function
z
_
(
z
a)(x)b(x)(dx)
is continuous and bounded in S, holomorphic in the interior of S. If Re z = 1, h(z) =
re
Imz
is a positive real number, so that
z
L(L
). Moreover
sup
tR
|
it
|
L(L
2
)
M, sup
tR
|
1+it
|
L(L
)
M.
By the Stein interpolation theorem 2.1.15, applied with p
0
= q
0
= 2, p
1
= q
1
= +,
for every s (0, 1) the operator
s
has an extension in L(L
p
) with p = 2/(1 s), and
|
s
|
L(L
p
)
M. Note that since s runs in (0, 1), then p runs in (2, ).
s
is nothing but
T(re
i(1s)
). Since r and are arbitrary, for every z ,= 0 with [arg z[ <
0
(1 s) = 2
0
/p
T(z) L(L
p
), with norm less or equal to M.
Let us show that z T(z) is holomorphic in the sector
p
= z ,= 0 : [arg z[ < 2
0
/p
with values in L(L
p
) (for the moment we know only that it is bounded).
Let f L
p
, g L
p
, and let f
n
L
p
L
2
, g
n
L
p
L
2
, be such that f
n
f in L
p
,
g
n
g in L
p
. Then the functions z T(z)f

n
, g
n
) are holomorphic in
2

p
, and
converge to T(z)f, g) uniformly in
p
because |T(z)|
L(L
p
)
is bounded in
p
. Therefore
z T(z)f, g) is holomorphic in
p
. This implies that z T(z) is holomorphic with
values in L(L
p
).
Theorem 5.2.2 may be easily generalized as follows: if a semigroup T(t) is analytic in
L
p
0
and bounded in L
p
1
then it is analytic in L
p
for every p in the interval with endpoints
p
0
and p
1
. But the most common situation is p
0
= 2, p
1
= . For instance, if
/u =
n
i,j=1
D
i
(a
ij
(x)D
j
u)(x), x R
n
,
where the coecients a
ij
are in H
1
loc
(R
n
) and
n
i,j=1
a
ij
(x)
i
j
0, x, R
n
,
then the realization of / in L
2
(R
n
) generates an analytic bounded semigroup in L
2
(R
n
),
whose restriction to L
(R
n
) is a bounded semigroup in L
(R
n
). The proof may be found
in the book of Davies [17, Ex. 3.2.11].
5.3 Regularity in abstract parabolic equations
Throughout the section we x T > 0 and we set
M
k
= sup
0<tT+1
|t
k
A
k
e
tA
|
L(X)
, k N 0, (5.3)
and, for (0, 1),
M
k,
= sup
0<tT+1
|t
k
A
k
e
tA
|
L(D
A
(,),X)
, k N. (5.4)
Let f : [0, T] X, and u
0
X. Cauchy problems of the type
_
_
_
u
t
(t) = Au(t) +f(t), 0 < t < T,
u(0) = u
0
(5.5)
are called abstract parabolic problems, since the most known examples of sectorial oper-
ators are the realizations of elliptic dierential operators in the usual functional Banach
spaces (L
p
spaces, spaces of continuous or Holder continuous functions in R
n
or in open
sets of R
n
, etc.).
In this section we will see some optimal regularity results for (5.5) involving the inter-
polation spaces D
A
(, p). To begin with, we need some notation and general results about
abstract parabolic problems.
104 Chapter 5
Denition 5.3.1 Let T > 0, let f : [0, T] X be a continuous function, and let u
0
X.
Then:
(i) A function u C
1
([0, T]; X) C([0, T]; D(A)) is said to be a strict solution of (5.5)
in the interval [0, T] if u
t
(t) = Au(t) +f(t) for each t [0, T], and u(0) = u
0
.
(ii) A function u C
1
((0, T]; X) C((0, T]; D(A)) C([0, T]; X) is said to be a classical
solution of (5.5) in the interval [0, T] if u
t
(t) = Au(t) +f(t) for each t (0, T], and
u(0) = u
0
.
From denition 5.3.1 it follows easily that if problem (5.5) has a strict solution then
u
0
D(A), Au
0
+f(0) D(A), (5.6)
and if problem (5.5) has a classical solution, then
u
0
D(A). (5.7)
Moreover, any strict solution is also classical.
Proposition 5.3.2 Let f C([0, T], X), and let u
0
D(A). If u is a classical solution
of (5.5), then it is given by the variation of constants formula
u(t) = e
tA
u
0
+
_
t
0
e
(ts)A
f(s)ds, 0 t T. (5.8)
Proof. Let u be a classical solution of (5.5) in [0, T], and let t (0, T]. Since u
C
1
((0, T]; X) C([0, T]; X) C((0, T]; D(A)), then u(t) belongs to D(A) for 0 t T,
the function
v(s) = e
(ts)A
u(s), 0 s t,
belongs to C([0, t]; X) C
1
((0, t), X), and
v(0) = e
tA
u
0
, v(t) = u(t),
v
t
(s) = Ae
(ts)A
u(s) +e
(ts)A
u
t
(s) = e
(ts)A
f(s), 0 < s < t.
Then, for 0 < 2 < t,
v(t ) v() =
_
t
e
(ts)A
f(s)ds,
so that letting 0 we get
v(t) v(0) =
_
t
0
e
(ts)A
f(s)ds,
Proposition 5.3.2 implies that the classical solution of (5.5) is unique. Therefore the
strict solution is unique. Unfortunately, in general the function dened by (5.8) is not a
classical or a strict solution of (5.5). It is called mild solution of (5.5). The rst term
t e
tA
u
0
is OK: it is a classical solution of w
t
= Aw, w(0) = u
0
if u
0
D(A), it is a
strict solution if u
0
D(A), Au
0
D(A). The diculties come from the second term,
v(t) =
_
t
0
e
(ts)A
f(s)ds, 0 t T.
Its regularity properties are stated in the next proposition.
Proposition 5.3.3 Let f L
(0, T; X). Then, for every (0, 1), v C
([0, T]; X)
C([0, T]; D
A
(, 1)). More precisely, it belongs to C
1
([0, T]; D
A
(, 1)), and there is C
independent of f such that
|v|
C
1
([0,T];D
A
(,1))
C|f|
L
(0,T;X)
. (5.9)
Proof. Due to estimate (5.3), with k = 0, v satises
|v(t)| M
0
_
t
0
|f(s)|ds M
0
T|f|
, 0 t T. (5.10)
Since |e
tA
|
L(X)
and |te
tA
|
L(X,D(A))
are bounded in (0, T], by interpolation there is K
0,
>
0 such that |e
tA
|
L(X,D
A
(,1))
K
0,
t
for 0 < t T. Similarly, since |tAe

tA
|
L(X)
and
|t
2
Ae
tA
|
L(X,D(A))
are bounded in (0, T], by interpolation there is K
1,
> 0 such that
|Ae
tA
|
L(X,D
A
(,1))
K
1,
t
1
for 0 < t T.
Therefore, s |e
(ts)A
|
L(X,D
A
(,1))
belongs to L
1
(0, t) for every t (0, T]. Then v(t)
belongs to D
A
(, 1) for every (0, 1), and
|v(t)|
D
A
(,1)
K
0,
(1 )
1
T
1
|f|
L
(0,T;X)
, (5.11)
Moreover, for 0 s t T,
v(t) v(s) =
_
s
0
_
e
(t)A
e
(s)A
_
f()d +
_
t
s
e
(t)A
f()d
=
_
s
0
d
_
t
s
Ae
A
f()d +
_
t
s
e
(t)A
f()d,
which implies
|v(t) v(s)|
D
A
(,1)
K
1,
_
s
0
d
_
t
s
1
d |f|
+K
0,
_
t
s
(t )
d |f|

_
K
1,
(1 )
+
K
0,
1
_
(t s)
1
|f|
,
(5.12)
so that v is (1)-Holder continuous with values in D
A
(, 1). Estimate (5.9) follows now
from (5.11) and (5.12).
The estimates for v(t) blow up as 1. In fact it is possible to show that in general
v is not Lipschitz continuous with values in X, nor bounded with values in D(A).
Next lemma is useful because it cuts half of the job in showing that a mild solution is
classical or strict.
Lemma 5.3.4 Let f C([0, T]; X), let u
0
D(A), and let u be the mild solution of (5.5).
The following conditions are equivalent.
(a) u C((0, T]; D(A)),
(b) u C
1
((0, T]; X),
106 Chapter 5
(c) u is a classical solution of (5.5).
Moreover the following conditions are equivalent.
(a
t
) u C([0, T]; D(A)),
(b
t
) u C
1
([0, T]; X),
(c
t
) u is a strict solution of (5.5).
Proof. Of course, (c) is stronger than (a) and (b). Let us show that if either (a) or
(b) holds, then u is a classical solution. Since u
0
D(A), then t e
tA
u
0
belongs to
C([0, T]; X). We know already that u belongs to C([0, T]; X). Moreover the integral
_
t
0
u(s)ds belongs to D(A), and
u(t) = u
0
+A
_
t
0
u(s)ds +
_
t
0
f(s)ds, 0 t T. (5.13)
Indeed, for every t [0, T] we have
_
t
0
u(s)ds =
_
t
0
e
sA
u
0
ds +
_
t
0
ds
_
s
0
e
(s)A
f()d
=
_
t
0
e
sA
u
0
ds +
_
t
0
d
_
t
e
(s)A
f()ds.
By lemma 5.0.14, the integral
_
t
e
(s)A
f()ds =
_
t
0
e
A
f()d is in D(A), and
A
_
t
e
(s)A
f()ds = (e
(t)A
I)f() L
1
(0, t).
Therefore the integral
_
t
0
u(s)ds belongs to D(A), and
A
_
t
0
u(s)ds = e
tA
u
0
u
0
+
_
t
0
(e
(t)A
1)f()d, 0 t T,
so that (5.13) holds.
From (5.13) we infer that for every t, h such that t, t +h (0, T],
u(t +h) u(t)
h
=
1
h
A
_
t+h
t
u(s)ds +
1
h
_
t+h
t
f(s)ds. (5.14)
Since f is continuous at t, then
lim
h0
1
h
_
t+h
t
f(s)ds = f(t). (5.15)
Let (a) hold. Then Au is continuous at t, so that
lim
h0
1
h
A
_
t+h
t
u(s)ds = lim
h0
1
h
_
t+h
t
Au(s)ds = Au(t).
By (5.14) and (5.15) we get now that u is dierentiable at the point t, with u
t
(t) =
Au(t) +f(t). Since both Au and f are continuous in (0, T], then u
t
too is continuous, and
u is a classical solution.
Let now (b) hold. Since u is continuous at t, then
lim
h0
1
h
_
t+h
t
u(s)ds = u(t).
On the other hand, by (5.14) and (5.15), there exists the limit
lim
h0
A
_
1
h
_
t+h
t
u(s)ds
_
= u
t
(t) f(t).
Since A is a closed operator, then u(t) belongs to D(A), and Au(t) = u
t
(t) f(t). Since
both u
t
and f are continuous in (0, T], then also Au is continuous in (0, T], so that u is a
classical solution.
The equivalence of (a
t
), (b
t
), (c
t
) may be proved in the same way.
Now we are ready to prove regularity results for (5.5).
Let u be the mild solution of (5.5), and set u = u
1
+u
2
, where
_
_
u
1
(t) =
_
t
0
e
(ts)A
(f(s) f(t))ds, 0 t T,
u
2
(t) = e
tA
u
0
+
_
t
0
e
(ts)A
f(t)ds, 0 t T.
(5.16)
Theorem 5.3.5 Let 0 < < 1, f C
([0, T], X), u

0
X. Then the mild solution u of
(5.5) belongs to C
([, T], D(A)) C

1+
([, T], X) for every (0, T), and
(i) if u
0
D(A), then u is a classical solution of (5.5);
(ii) if u
0
D(A) and Au
0
+f(0) D(A), then u is a strict solution of (5.5), and there
is C such that
|u|
C
1
([0,T],X)
+|u|
C([0,T],D(A))
C(|f|
C
([0,T],X)
+|u
0
|
D(A)
); (5.17)
(iii) if u
0
D(A) and Au
0
+ f(0) D
A
(, ), then u
t
and Au belong to C
([0, T], X),

u
t
belongs to B([0, T]; D
A
(, )), and there is C such that
|u|
C
1+
([0,T],X)
+|Au|
C
([0,T],X)
+|u
t
|
B([0,T],D
A
(,))
C(|f|
C
([0,T],X)
+|u
0
|
D(A)
+|Au
0
+f(0)|
D
A
(,)
).
(5.18)
Proof. Thanks to lemma 5.3.4, to prove statements (i) and (ii) it is sucient to show
that u belongs to C((0, T]; D(A)) in the case where u
0
D(A), and to C([0, T]; D(A)) in
the case where u
0
D(A) and Au
0
+f(0) D(A). We know already by proposition 5.3.3
that u C
([, T]; X) for every (0, T), and that u C([0, T]; X) if u
0
D(A). So we
have to study Au.
108 Chapter 5
Let u
1
and u
2
be dened by (5.16). Then u
1
(t) D(A) for t 0, u
2
(t) D(A) for
t > 0, and
_
_
(i) Au
1
(t) =
_
t
0
Ae
(ts)A
(f(s) f(t))ds, 0 t T,
(ii) Au
2
(t) = Ae
tA
u
0
+ (e
tA
1)f(t), 0 < t T.
(5.19)
If u
0
D(A), then (5.19)(ii) holds also for t = 0.
Let us show that Au
1
is Holder continuous in [0, T]. For 0 s t T
Au
1
(t) Au
1
(s) =
_
s
0
A
_
e
(t)A
e
(s)A
_
(f() f(s))d
+(e
tA
e
(ts)A
)(f(s) f(t)) +
_
t
s
Ae
(t)A
(f() f(t))d,
(5.20)
so that, since A(e
(t)A
e
(s)A
) =
_
t
s
A
2
e
A
d,
|Au
1
(t) Au
1
(s)| M
2
_
s
0
(s )
_
t
s
2
d d [f]
C
+2M
0
(t s)
[f]
C
+M
1
_
t
s
(t )
1
d [f]
C
M
2
_
s
0
d
_
t
s
2
d [f]
C
+ (2M
0
+M
1
1
)(t s)
[f]
C
_
M
2
(1 )
+ 2M
0
+
M
1
_
(t s)
[f]
C
.
(5.21)
Therefore, Au
1
is -Holder continuous in [0, T]. Moreover, Au
2
is obviously -Holder
continuous in [, T]: hence, if u
0
D(A), then u C([0, T], X) and Au C((0, T]; X), so
that, by Lemma 5.3.4, u is a classical solution of (5.5), and statement (i) is proved.
If u
0
D(A) we have
Au
2
(t) = e
tA
(Au
0
+f(0)) +e
tA
(f(t) f(0)) f(t), 0 t T, (5.22)
so that if Au
0
+ f(0) D(A) then Au
2
is continuous also at t = 0, and statement (ii)
follows.
In the case where Ax +f(0) D
A
(, ), from (5.22) we get, for 0 s t T,
|Au
2
(t) Au
2
(s)| |(e
tA
e
sA
)(Au
0
+f(0))|
+|(e
tA
e
sA
)(f(s) f(0))| +|(e
tA
1)(f(t) f(s))|
_
t
s
|Ae
A
|
L(D
A
(,),X)
d |Au
0
+f(0)|
D
A
(,)
+s
_
_
_
_
A
_
t
s
e
A
d
_
_
_
_
L(X)
[f]
C
+ (M
0
+ 1)(t s)
[f]
C
M
1,
|Au
0
+f(0)|
D
A
(,)
(t s)
+
_
M
1
+M
0
+ 1
_
(t s)
[f]
C
,
(5.23)
so that also Au
2
is Holder continuous, and the estimate
|u|
C
1+
([0,T];X)
+|Au|
C
([0,T];X)
C(|f|
C
([0,T],X)
+|u
0
|
D(A)
+|Au
0
+f(0)|
D
A
(,)
)
follows easily.
Let us estimate |u
t
(t)|
D
A
(,)
. For 0 t T we have, by (5.19),
u
t
(t) =
_
t
0
Ae
(ts)A
(f(s) f(t))ds +e
tA
(Au
0
+f(0)) +e
tA
(f(t) f(0)),
so that for 0 < 1
|
1
Ae
A
u
t
(t)|
_
_
_
_
1
_
t
0
A
2
e
(t+s)A
(f(s) f(t))ds
_
_
_
_
+|
1
Ae
(t+)A
(Au
0
+f(0))| +|
1
Ae
(t+)A
(f(t) f(0))|
M
2
1
_
t
0
(t s)
(t + s)
2
ds [f]
C
+M
0
[Au
0
+f(0)]
D
A
(,)
+M
1
1
(t +)
1
t
[f]
C
M
2
_

0
( + 1)
2
d [f]
C
+M
0
[Au
0
+f(0)]
D
A
(,)
+M
1
[f]
C
.
(5.24)
Therefore, |u
t
(t)|
D
A
(,)
is bounded in [0, T], and the proof is complete.
Theorem 5.3.6 Let 0 < < 1, u
0
X, f C([0, T]; X) B([0, T]; D
A
(, )), and
let u be the mild solution of (5.5). Then u C
1
((0, T]; X) C((0, T]; D(A)), and u
B([, T]; D
A
( + 1, )) for every (0, T). Moreover, the following statements hold.
(i) if u
0
D(A), then u is a classical solution;
(ii) if u
0
D(A), Au
0
D(A), then u is a strict solution;
(iii) if u
0
D
A
(+1, ), then u
t
and Au belong to C([0, T]; X) B([0, T]; D
A
(, )),
Au belongs to C
([0, T]; X), and there is C such that

|u
t
|
B([0,T];D
A
(,))
+|Au|
B([0,T];D
A
(,))
+|Au|
C
([0,T];X)
C(|f|
B([0,T];D
A
(,))
+|u
0
|
D
A
(+1,)
).
(5.25)
Proof. Let us consider the function v. We are going to show that it is the strict solution
of
v
t
(t) = Av(t) +f(t), 0 < t T, v(0) = 0, (5.26)
110 Chapter 5
and moreover v
t
and Av belong to B([0, T]; D
A
(, )), Av C
([0, T]; X), and there is

C such that
|v
t
|
B([0,T];D
A
(,))
+|Av|
B([0,T];D
A
(,))
+|Av|
C
([0,T];X)
C|f|
B([0,T];D
A
(,))
.
(5.27)
For 0 t T, v(t) belongs to D(A), and
|Av(t)| M
1,
_
t
0
(t s)
1
ds |f|
B(D
A
(,))
=
T
M
1,
|f|
B(D
A
(,))
. (5.28)
Moreover, for 0 < 1,
|
1
Ae
A
Av(t)| =
1
_
_
_
_
_
t
0
A
2
e
(t+s)A
f(s)ds
_
_
_
_
M
2,
1
_
t
0
(t + s)
2
ds|f|
B([0,T];D
A
(,))

M
2,
1
|f|
B([0,T];D
A
(,))
,
(5.29)
so that Av is bounded with values in D
A
(, ). Let us show that Av is Holder continuous
with values in X: for 0 s t T we have
|Av(t) Av(s)|
_
_
_
_
A
_
s
0
_
e
(t)A
e
(s)A
_
f()d
_
_
_
_
+
_
_
_
_
A
_
t
s
e
(t)A
f()d
_
_
_
_
M
2,
_
s
0
d
_
t
s
2
d |f|
B([0,T];D
A
(,))
+M
1,
_
t
s
(t )
1
d |f|
B([0,T];D
A
(,))
_
M
2,
(1 )
+
M
1,
_
(t s)
|f|
B([0,T];D
A
(,))
,
(5.30)
so that Av is -Holder continuous in [0, T]. Estimate (5.27) follows now from (5.28),
(5.29), (5.30). Moreover, thanks to Lemma 5.3.4, v is a strict solution of (5.26).
Let us consider now the function t e
tA
u
0
.
If u
0
D(A), t e
tA
u
0
is the classical solution of w
t
= Aw, t > 0, w(0) = u
0
.
If u
0
D(A) and Au
0
D(A) it is a strict solution. If x D
A
( + 1, ), it is a strict
solution, moreover it belongs to B([0, T]; D
A
(+1, )) C
([0, T]; D(A)). Summing up,

the statement follows.
5.4 Applications to regularity in parabolic PDEs
Consider the problem
_
_
_
u
t
(t, x) = u(t, x) +f(t, x), 0 t T, x R
n
,
u(0, x) = u
0
(x), x R
n
,
(5.31)
where f and u
0
are continuous and bounded functions. We read it as an abstract Cauchy
problem of the type (5.5) in the space X = C(R
n
) with the sup norm, setting u(t) = u(t, )
and f(t) = f(t, ). A is the realization of the Laplace operator in X. It is the generator
of the Gauss-Weierstrass analytic semigroup dened by (3.6). Note that the domain of A,
D(A) = C(R
n
) : (in the sense of distributions) C(R
n
)
= C(R
n
) W
2,p
loc
(R
n
) p 1 : C(R
n
)
contains properly C
2
(R
n
). However, we already know that for 0 < < 1, ,= 1/2
D
A
(, ) = C
2
(R
n
), D
A
( + 1, ) = C
2+2
(R
n
).
For every f : [0, T] R
n
C set

f(t) = f(t, ), 0 t T. The following statements
are easy to be checked.
(i)

f : [0, T] X is continuous i f is continuous, bounded, and for every t
0
[0, T]
lim
tt
0
sup
xR
n [f(t, x) f(t
0
, x)[ = 0;
(ii) if 0 < < 1, then

f C
([0, T]; X) if and only if f is continuous and bounded, and

moreover sup
s,=t, xR
n [f(t, x) f(s, x)[/[t s[
< ;
(iii) if 0 < < 1, ,= 1/2,

f B([0, T]; D
A
(, )) i f(t, ) C
2
(R
n
) for every t, and
sup
0tT
|f(t, )|
C
2
(R
n
)
< .
So, we may apply theorems 5.3.5 and 5.3.6. Theorem 5.3.5 gives
Theorem 5.4.1 Let 0 < < 1, ,= 1/2, and let f : [0, T] R
n
C be continuous,
bounded, such that f(, x) C
([0, T]) for each x R

n
and sup
xR
n |f(, x)|
C
([0,T])
< .
Let u
0
D(A) be such that u
0
+ f(0, ) C
2
(R
n
). Then problem (5.31) has a unique
solution u such that u, u
t
, u are continuous, bounded, and sup
xR
n |u
t
(, x)|
C
([0,T])
<
, sup
xR
n |u(, x)|
C
([0,T])
< , sup
t[0,T]R
n |u
t
(t, )|
C
2
(R
n
)
< .
Applying theorem 5.3.6 gives
Theorem 5.4.2 Let 0 < < 1, ,= 1/2, and let f : [0, T] R
n
C be uniformly
continuous, bounded, and such that sup
t[0,T]
|f(t, )|
C
2
(R
n
)
< . Let u
0
C
2+2
(R
n
).
Then problem (5.31) has a unique solution u such that u, u
t
, u are uniformly con-
tinuous, bounded, and sup
t[0,T]
|u
t
(t, )|
C
2
(R
n
)
< , sup
t[0,T]
|D
ij
u(t, )|
C
2
(R
n
)
< ,
sup
xR
n |u(, x)|
C
([0,T])
< .
Putting together theorems 5.3.5 and 5.3.6 we get the LadyzhenskajaSolonnikov
Uralceva theorem (see [29] for a completely dierent proof). For simplicity we consider
only the case 0 < < 1/2. It is convenient to adopt the usual notation: we denote by
C
,2
([0, T] R
n
) the space of the bounded functions f such that
sup
t,=s, x,=y
[f(t, x) f(s, y)[
[t s[
+[x y[
2
< ,
and we denote by C
1+,2+2
([0, T] R
n
) the space of the bounded functions f with
bounded f
t
, D
ij
f, such that f
t
, D
ij
f are in C
,2
([0, T] R
n
). It is possible to see that
this implies that the space derivatives D
i
f are (1/2 + )-Holder continuous with respect
to t, with Holder constant independent of x.
112 Chapter 5
Theorem 5.4.3 (LadyzhenskajaSolonnikovUralceva) Let 0 < < 1/2 and let f
C
,2
([0, T] R
n
), u
0
C
2+2
(R
n
). Then problem (5.31) has a unique solution u
C
1+,2+2
([0, T] R
n
).
Proof. Almost all follows from patching together the results of the above two theorems.
It remains to show that the space derivatives D
ij
u are time -Holder continuous. To
this aim we use the fact that C
2
(R
n
) belongs to the class J
1
between C
2
(R
n
) and
C
2+2
(R
n
) (see the exercises of chapter 1). Moreover, since |u
t
(t, )|
C
2
(R
n
)
is bounded
in [0, T], then t u(t, ) is Lipschitz continuous with values in C
2
(R
n
), with Lipschitz
constant sup
0T
|u
t
(, )|
C
2. So, for 0 s t T we have
|u(t, ) u(s, )|
C
2 C|u(t, ) u(s, )|
C
2
|u(t, ) u(s, )|
1
C
2+2
C((t s) sup
0T
|u
t
(, )|
C
2)
(2 sup
[0,T]
|u(, )|
C
2+2)
1
C
t
(t s)

This procedure works also if the Laplacian is replaced by any uniformly elliptic operator
with regular and bounded coecients. Indeed it is known that the realization A of such an
operator in C(R
n
) is sectorial, and that D
A
(, ) = C
2
(R
n
), D
A
(+1, ) = C
2+2
(R
n
)
for ,= 1/2. But the proofs of this properties are not trivial; they rely on the Stewarts
theorem [35] which, in its turn, is based on the AgmonDouglisNirenberg theorem [3].
See [32, Ch. 3].
Appendix A
The Bochner integral
A.1 Integrals over measurable real sets
We recall here the few elements of Bochner integral theory that are used in these notes.
Extended treatments, with proofs, may be found in the books [7], [37].
X is any real or complex Banach space. We consider the usual Lebesgue measure in
R, and we denote by / the -algebra consisting of all Lebesgue measurable subsets of R.
If A R,
A
denotes the characteristic function of the set A.
Denition A.1.1 A function f : R X is said to be simple if there are n N, x
1
, .., x
n

X, A
1
, .., A
n
/, with meas A
i
< and A
i
A
j
= for i ,= j, such that
f =
n
i=1
x
i
A
i
.
If I /, a function f : I X is said to be Bochner measurable if there is a sequence of
simple functions f
n
such that
lim
n
f
n
(t) = f(t) for almost all t I.
It is easy to see that every continuous function is measurable.
If f =
n
i=1
x
i
A
i
is a simple function we set
_
R
f(t)dt =
n
i=1
x
i
measA
i
. (A.1)
Denition A.1.2 Let f : R X. f is said to be Bochner integrable if there is a sequence
of simple functions f
n
converging to f almost everywhere, such that
lim
n
_
R
|f
n
(t) f(t)|dt = 0.
Then n
_
R
f
n
(t)dt is a Cauchy sequence in X. We set
_
R
f(t)dt = lim
n
_
R
f
n
(t)dt. (A.2)
113
114 Appendix A
Arguing as in the case X = R, one sees that
_
R
f(t)dt is independent of the choice of the
sequence f
n
. If f is dened on a measurable set, the above denition can be extended
as follows.
Denition A.1.3 If I / and f : I X, f is said to be integrable over I if the
extension

f dened by
f(t)
_
= f(t), if t I
= 0, if t / I
is integrable. In this case we set
_
I
f(t)dt =
_
R
f(t)dt. (A.3)
If I = (a, b), with a b +, we set as usual
_
(a,b)
f(t)dt =
_
b
a
f(t)dt;
_
a
b
f(t)dt =
_
b
a
f(t)dt
A simple criterion for establishing whether a function is integrable is stated in the
following proposition.
Proposition A.1.4 Let I /, and let f : I X. Then f is integrable if and only if f
is measurable and t |f(t)| is Lebesgue integrable on I. Moreover,
_
_
_
_
_
I
f(t)dt
_
_
_
_
_
I
|f(t)|dt. (A.4)
From the denition it follows easily that if Y is another Banach space and A L(X, Y ),
then for every integrable f : I X the function Af : I Y is integrable, and
_
I
Af(t)dt = A
_
I
f(t)dt.
In particular, if X
t
then for every integrable f : I X the function t f(t), ) is
integrable, and
_
I
f(t)dt, ) =
_
I
f(t), )dt.
It follows that for every couple of integrable functions f, g it holds
_
I
(f(t) +g(t))dt =
_
I
f(t) +
_
I
g(t)dt, , C,
Another important commutativity property is the following one.
Proposition A.1.5 Let X, Y be Banach spaces, and let A : D(A) X Y be a closed
operator. Let I /, let f : I X be an integrable function such that f(t) D(A) for
almost all t I, and Af : I Y is integrable. Then the integral
_
I
f(t)dt belongs to
D(A), and
A
_
I
f(t)dy =
_
I
Af(t)dt.
The Bochner integral 115
A.2 L
p
and Sobolev spaces
On the set of all measurable functions on I we dene the equivalence relation
f g f(t) = g(t) for almost all t I. (A.5)
Denition A.2.1 L
1
(I; X) is the set of all equivalence classes of integrable functions
f : I X, with respect to the equivalence relation (A.5).
Since no confusion will arise, in the sequel we shall identify the equivalence class [f]
L
1
(I; X) with the function f itself. We dene a norm on L
1
(I, X) by setting
|f|
L
1
(I;X)
=
_
I
|f(t)|dt. (A.6)
We dene now the spaces L
p
(I; X) for p > 1.
Denition A.2.2 Let p (1, +], and I /. L
p
(I; X) is the set of all equivalence
classes of measurable functions f : I X,with respect to the equivalence relation (A.5),
such that t |f(t)| belongs to L
p
(I).
L
p
(I; X) is endowed with the norm
|f|
L
p
(I;X)
=
__
I
|f(t)|
p
dt
_
1/p
, if p < (A.7)
|f|
L
(I;X)
= ess sup |f(t)| : t I (A.8)
Arguing as in the case X = R, it is not dicult to see that for 1 p , the space
L
p
(I; X) is complete.
In the following, if there is no danger of confusion, we shall write |f|
p
instead of
|f|
L
p
(I;X)
.
To introduce the Sobolev space W
1,p
(a, b; X) we need a lemma.
Lemma A.2.3 Let p [1, ). Then the operator
L
0
: D(L
0
) = C
1
([a, b]; X) L
p
(a, b; X), L
0
f = f
t
is preclosed in L
p
(a, b; X), that is the closure of its graph is the graph of a closed operator.
Denition A.2.4 Let L : D(L) L
p
(a, b; X) be the closure of the operator L
0
dened
in Lemma (A.2.3). We set
W
1,p
(a, b; X) = D(L)
and we endow it with the graph norm. For every f W
1,p
(a, b; X), Lf is said to be the
strong derivative of f, and we denote it by f
t
.
In other words, f W
1,p
(a, b; X) if and only if there is a sequence f
n
C
1
([a, b]; X)
such that f
n
f in L
p
(a, b; X) and f
t
n
g in L
p
(a, b; X), and in this case g = f
t
.
Moreover we have
|f|
W
1,p
(a,b;X)
= |f|
L
p
(a,b;X)
+|f
t
|
L
p
(a,b;X)
f W
1,p
(a, b; X).
116 Appendix A
Since L is a closed operator, then W
1,p
(a, b; X) is a Banach space.
Let f W
1,p
(a, b; X), and let f
n
C
1
([a, b]; X) be such that f
n
f and f
t
n
f
t
in L
p
(a, b; X). From the equality
f
n
(t) f
n
(s) =
_
t
s
f
t
n
()d (A.9)
we get, integrating with respect to s in (a, b) and letting n ,
f(t) =
1
b a
__
b
a
f(s)ds +
_
t
a
( a)f
t
()d
_
, a.e.in (a, b).
Therefore, W
1,p
(a, b; X) is continuously embedded in C([a, b]; X). Letting n in (A.9)
we get also
f(t) f(s) =
_
t
s
f
t
()d.
Sometimes it is easier to deal with weak (or distributional) derivatives, dened as
follows.
Denition A.2.5 Let f L
p
(a, b; X). A function g L
1
(a, b; X) is said to be the weak
derivative of f in (a, b) if
_
b
a
f(t)
t
(t)dt =
_
b
a
g(t)(t)dt, C
0
(a, b).
It can be shown that weak and strong derivatives do coincide. More precisely, the
following proposition holds.
Proposition A.2.6 Let f W
1,p
(a, b; X). Then f is weakly dierentiable, and f
t
is the
weak derivative of f.
Conversely, if f L
p
(a, b; X) admits a weak derivative g L
p
(a, b; X), then f
W
1,p
(a, b; X), and g = f
t
.
A.3 Weighted L
p
spaces
Let I be an interval contained in (0, +). For 1 p we denote by L
p
(I) the space

of the L
p
functions in I with respect to the measure dt/t, endowed with its natural norm
|f|
L
p
(I)
=
__
+
0
[f(t)[
p
dt
t
_
1/p
, if p < ,
|f|
L
(I)
= ess sup
tI
[f(t)[.
Dealing with L
p
spaces, the Hardy-Young inequalities are often more useful than the
Holder inequality. They hold for every positive measurable function : (0, a) R,
0 < a , and every > 0, p 1. See [23, p.245-246].
_
_
(i)
_
a
0
t
p
__
t
0
(s)
ds
s
_
p
dt
t

1
p
_
a
0
s
p
(s)
p
ds
s
,
(ii)
_
a
0
t
p
__
a
t
(s)
ds
s
_
p
dt
t

1
p
_
a
0
s
p
(s)
p
ds
s
(A.10)
Vector-valued holomorphic functions 117
The measure m(dt) = dt/t is the Haar measure of the multiplicative group R
+
. So,
it is invariant under multiplication: m(A) = m(A), for every measurable A R
+
and
> 0, and ||
L
p
(0,)
= |()|
L
p
(0,)
.
For every ,= 0 the space L
p
(0, ) is invariant under the change of variable t t
,
in the sense that L
p
(0, ) i t (t
) L
p
(0, ), and
||
L
p
(0,)
= [[
1/p
|t (t
)|
L
p
(0,)
.
(This is obviously true also for p = , with the usual convention 1/ = 0). In particular,
for = 1 we get an isometry:
||
L
p
(0,)
= |t (t
1
)|
L
p
(0,)
.
Moreover, the change of variable t t
1
is an isometry also between L
p
(1, ) and L
p
(0, 1).
If X is any Banach space and 1 p the space L
p
(I; X) is the set of all Bochner

measurable functions f : I X, such that t |f(t)|
X
is in L
p
(I). It is endowed with

the norm
|f|
L
p
(I;X)
= |t |f(t)|
X
|
L
p
(I)
.
In chapters 1 and 3 we have used the following consequence of inequality (A.10)(i).
Corollary A.3.1 Let u be a function such that t u
(t) = t
u(t) belongs to L
p
(0, a; X),
with 0 < a , 0 < < 1 and 1 p . Then also the mean value
v(t) =
1
t
_
t
0
u(s)ds, t > 0 (A.11)
has the same property, and setting v
(t) = t
v(t) we have
|v
|
L
p
(0,a;X)

1
1
|u
|
L
p
(0,a;X)
(A.12)
118 Appendix B
Appendix B
Vector-valued holomorphic
functions
Let X be a complex Banach space, let be an open subset of C, f : X be a
continuous function and : [a, b] be a piecewise C
1
-curve. The integral of f along
is dened by
_
f(z)dz =
_
b
a
f((t))
t
(t)dt.
As usual, we denote by X
t
the dual space of X consisting of all linear bounded operators
from X to C. For each x X, x
t
X
t
we set x
t
(x) = x, x
t
).
Denition B.0.2 f is holomorphic in if for each z
0
the limit
lim
zz
0
f(z) f(z
0
)
z z
0
:= f
t
(z
0
)
exists in X. f is weakly holomorphic in if it is continuous in and the complex-valued
function z f(z), x
t
) is holomorphic in for every x
t
X
t
.
Clearly, any holomorphic function is weakly holomorphic; actually, the converse is also
true, as the following theorem shows.
Theorem B.0.3 Let f : X be a weakly holomorphic function. Then f is holomor-
phic.
Proof. Let B(z
0
, r) be a closed ball contained in ; we prove that for all z B(z
0
, r) the
following Cauchy integral formula holds:
f(z) =
1
2i
_
B(z
0
,r)
f()
z
d. (B.1)
First of all, we observe that the right hand side of (B.1) is well dened because f is con-
tinuous. Since f is weakly holomorphic in , the complex-valued function z f(z), x
t
)
is holomorphic in for all x
t
X
t
, and hence the ordinary Cauchy integral formula in
B(z
0
, r) holds, i.e.,
f(z), x
t
) =
1
2i
_
B(z
0
,r)
f(), x
t
)
z
d =
_
1
2i
_
B(z
0
,r)
f()
z
d, x
t
_
.
119
120 Appendix B
Since x
t
X
t
is arbitrary, we obtain (B.1). We can dierentiatee with respect to z under
the integral, so that f is holomorphic and
f
(n)
(z) =
n!
2i
_
B(z
0
,r)
f()
( z)
n+1
d
for all z B(z
0
, r) and n N.
Denition B.0.4 Let f : X. We say that f admits a power series expansion around
a point z
0
if there exist a X-valued sequence (a
n
) and r > 0 such that B(z
0
, r)
and
f(z) =
+
n=0
a
n
(z z
0
)
n
in B(z
0
, r).
Theorem B.0.5 Let f : X be a continuous function; then f is holomorphic if and
only if f has a power series expansion around every point of .
Proof. Assume that f is holomorphic in . Then, if z
0
and B(z
0
, r) , the Cauchy
integral formula (B.1) holds for every z B(z
0
, r).
Fix z B(z
0
, r) and observe that the series
+
n=0
(z z
0
)
n
( z
0
)
n+1
=
1
z
converges uniformly for in B(z
0
, r), since
(zz
0
)/(z
0
)[ = r
1
[zz
0
[. Consequently,
by (B.1) we obtain
f(z) =
1
2i
_
B(z
0
,r)
f()
+
n=0
(z z
0
)
n
( z
0
)
n+1
d
=
+
n=0
_
1
2i
_
B(z
0
,r)
f()
( z
0
)
n+1
d
_
(z z
0
)
n
,
the series being convergent in X.
Conversely, suppose that
f(z) =
+
n=0
a
n
(z z
0
)
n
, z B(z
0
, r),
where (a
n
) is a sequence with values in X. Then f is continuous, and for each x
t
X
t
,
f(z), x
t
) =
+
n=0
a
n
, x
t
)(z z
0
)
n
, z B(z
0
, r).
This implies that the complex-valued function z f(z), x
t
) is holomorphic in B(z
0
, r)
for all x
t
X
t
and hence f is holomorphic by Theorem B.0.3.
Now we extend some classical theorems of complex analysis to the case of vector-valued
holomorphic functions.
Vector-valued holomorphic functions 121
Theorem B.0.6 (Cauchy) Let f : X be holomorphic in and let D be a regular
domain contained in . Then _
D
f(z)dz = 0.
Proof. For each x
t
X
t
the complex-valued function z f(z), x
t
) is holomorphic in
and hence
0 =
_
D
f(z), x
t
)dz =
_
D
f(z)dz, x
t
).
Remark B.0.7 [improper complex integrals] As in the case of vector-valued func-

tions dened on a real interval, it is possible to dene improper complex integrals in an
obvious way. Let f : X be holomorphic, with C possibly unbounded. If I = (a, b)
is a (possibly unbounded) interval and : I C is a piecewise C
1
curve in , then we set
_
f(z)dz := lim
sa
+
, tb
_
t
s
f(())
t
()d,
provided that the limit exists in X.
Theorem B.0.8 (Laurent expansion) Let f : D := z C : r < [z z
0
[ < R X
be holomorphic. Then, for every z D
f(z) =
+
n=
a
n
(z z
0
)
n
,
where
a
n
=
1
2i
_
B(z
0
,)
f(z)
(z z
0
)
n+1
dz, n Z,
and r < < R.
Proof. Since for each x
t
X
t
the function z f(z), x
t
) is holomorphic in D the usual
Laurent expansion holds, that is
f(z), x
t
) =
+
n=
a
n
(x
t
)(z z
0
)
n
where the coecients a
n
(x
t
) are given by
a
n
(x
t
) =
1
2i
_
B(z
0
,)
f(z), x
t
)
(z z
0
)
n+1
dz, n Z.
It follows that
a
n
(x
t
) = a
n
, x
t
), n Z,
where the a
n
are those indicated in the statement.
122 Appendix B
Bibliography
[1] S. Agmon: Lectures on elliptic boundary value problems, Van Nostrand Math. Studies,
Princeton, NJ (1965).
[2] S. Agmon: On the eigenfunctions and the eigenvalues of general elliptic boundary
value problems, Comm. Pure Appl. Math. 15 (1962), 119-147.
[3] S. Agmon, A. Douglis, L. Nirenberg: Estimates near the boundary for solutions
of elliptic partial dierential equations satisfying general boundary conditions, Comm.
Pure Appl. Math. 12 (1959), 623-727.
[4] B. Beauzamy: Espaces dinterpolation reels: topologie et geometrie, Springer Verlag,
Berlin (1978).
[5] J. Bergh, J. L ofstr om: Interpolation Spaces. An Introduction, Springer Verlag,
Berlin (1976).
[6] J. Bourgain: Some remarks on Banach spaces in which the martingale dierences
are unconditional, Arkiv Mat. 21 (1983), 163168.
[7] H. Br ezis: Operateurs maximaux monotones et semi-groupes de contractions dans
les espaces de Hilbert, North-Holland Mathematics Studies 5, Amsterdam (1973).
[8] H. Br ezis: Analyse Fonctionnelle, Masson, Paris (1983).
[9] Yu. Brudnyi, N. Krugljak: Interpolation functors and interpolation spaces, North-
Holland, Amsterdam (1991).
[10] D.L. Burkholder: A geometrical characterization of Banach spaces in which mar-
tingale dierence sequences are unconditional, Ann. Probability 9 (1981), 9971011.
[11] D.L. Burkholder: A geometric condition that implies the existence of certain sin-
gular integrals of Banach space valued functions, in: Conference on Harmonic Anal-
ysis in Honour of Antoni Zygmund, Chicago 1981, W. Beckner, A.P. Calderon, R.
Feerman, P.W. Jones Eds., 270286, Belmont, Cal. 1983, Wadsworth.
[12] P.L. Butzer, H. Berens: Semigroups of Operators and Approximation, Springer
Verlag, Berlin (1967).
[13] A.P. Calder on: Intermediate spaces and interpolation, the complex method, Studia
Math. 24 (1964), 113-190.
[14] S. Campanato: Equazioni ellittiche del II
o
ordine e spazi L
2,
, Ann. Mat. Pura
Appl. 69 (1965), 321381.
123
124 References
[15] S. Campanato, G. Stampacchia: Sulle maggiorazioni in L
p
nella teoria delle
equazioni ellittiche, Boll. U.M.I. 20 (1965), 393-399.
[16] R.R. Coifman, G. Weiss: Transference methods in analysis, Conference Board
of the Mathematical Sciences, Regional Conference Series in Mathematics 31, AMS,
Providence (1977).
[17] E.B. Davies: Heat kernels and spectral theory, Cambridge Univ. Press (1990).
[18] G. Dore, A. Venni: On the closedness of the sum of two closed operators, Math.
Z. 186 (1987), 189-201.
[19] N. Dunford, J. Schwartz: Linear Operators, Wiley-Interscience, New York
(1958).
[20] P. Grisvard: Commutativite de deux functeurs dinterpolation et applications, J.
Maths. Pures Appl. 45 (1966), 143-290.
[21] P. Grisvard:

Equations dierentielles abstraites, Ann. Scient.

Ec. Norm. Sup., 4
e
serie 2 (1969), 311-395.
[22] P. Grisvard: Caracterisation de quelques espaces dinterpolation, Arch. Rat. Mech.
Anal. 25 (1967), 40-63.
[23] G.H. Hardy, J.E. Littlewood, G. P` olya: Inequalities, Cambridge Univ. Press,
Cambridge (1934).
[24] F. John, L. Nirenberg: On functions of bounded mean oscillation, Comm. Pure
Appl. Math. 14 (1961), 415426.
[25] T. Kato: Fractional powers of dissipative operators, Proc. Math. Soc. Japan 13
(1961), 246274.
[26] T. Kato: Fractional powers of dissipative operators, II, Proc. Math. Soc. Japan 14
(1962), 242248.
[27] T. Kato: remarks on pseudo-resolvents and innitesimal generators of semigroups,
Proc. Japan Acad. 35 (1959), 467468.
[28] S.G. Krein, Yu. Petunin, E.M. Semenov: Interpolation of linear operators,
Amer. Math. Soc., Providence (1982).
[29] O.A. Ladyzhenskaja, V.A. Solonnikov, N.N. Uralceva: Linear and quasilin-
ear equations of parabolic type, Nauka, Moskow 1967 (Russian). English transl.: Transl.
Math. Monographs, AMS, Providence (1968).
[30] J.L. Lions: Theoremes de traces et dinterpolation, (I),. . .,(V); (I), (II) Ann. Sc.
Norm. Sup. Pisa 13 (1959), 389-403; 14 (1960), 317-331; (III) J. Maths. Pures Appl. 42
(1963), 196-203; (IV) Math. Annalen 151 (1963), 42-56; (V) Anais de Acad. Brasileira
de Ciencias 35 (1963), 1-110.
[31] J.L. Lions, J. Peetre: Sur une classe despaces dinterpolation, Publ. I.H.E.S. 19
(1964), 5-68.
References 125
[32] A. Lunardi: Analytic semigroups and optimal regularity in parabolic problems,
Birkhauser, Basel (1995).
[33] F. Riesz, B. Sz. Nagy: Functional Analysis, Dover (1990).
[34] R. Seeley: Norms and domains of the complex powers A
z
B
, Amer. J. Math. 93
(1971), 299309.
[35] H.B. Stewart: Generation of analytic semigroups by strongly elliptic operators,
Trans. Amer. Math. Soc. 199 (1974), 141-162.
[36] H. Triebel: Interpolation Theory, Function Spaces, Dierential Operators, North-
Holland, Amsterdam (1978).
[37] K. Yosida: Functional Analysis, Springer Verlag, Berlin (1965).

An Introduction To Interpolation Theory, 2007 (Lunardi - A.)

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

An Introduction To Interpolation Theory, 2007 (Lunardi - A.)

Încărcat de

Drepturi de autor:

Formate disponibile

An Introduction to Interpolation Theory

Dottorato di Ricerca in Matematica, consorzio

(I) is the Lebesgue space L

(I). See Appendix, 2.

are also called continuous interpolation spaces.

(0, ), unless x = 0. Therefore, (X, Y )

(a, ) for all a > 0, and lim

K(t, x) plays a role in the denition of (X, Y )

, so that the sequence

K(t, x) is bounded in every interval [a, b] with 0 < a < b.

and it is continuously embedded in

K(t, x) = 0. To prove that x (X, Y )

K(t, x) = 0. This can be seen as follows: since (X, Y )

is a Banach space, endowed with the norm of (X, Y )

is a closed subspace of (X, Y )

are interpolation spaces, as a consequence of the

) for every (0, 1)

(0, 1), with norm esti-

(t) = inf : m(, f) t, t 0.

are nonnegative, decreasing (i.e. nonincreasing), right continuous,

, [f[ are equi-measurable, that is for each

(0) = sup ess f

is called the nonincreasing rearrangement of f onto (0, ).

()) is an interpolation couple.

(s)ds, t > 0. (1.22)

(t), so that for q <

() and t > 0, x we set

(t) if [f(x)[ > f

(t) has measure (E) = [s > 0 :> f

are equi-measurable) = m(f

is constant in [(E), t].

(t) if [f(x)[ > f

(s), s 0, 0 < < 1.

(), and a +b = f we have

[a(x)[(dx) +t sup ess [b(x)[.

is the space of the little Holder continuous functions h

(s), s 0, 0 < < 1.

is the set of the

K(t, x), and the function t t

(0, +). Moreover, we

K(t, x) is bounded, so that lim

K(t, x), (1.23)

(0, +; Y ). By Corollary A.3.1, t t

K(t, x). (1.24)

(0, +; X), and

, then, by (1.23), lim

(0, +), so that x

(0, +; X), with norms estimated by const. |u|

is the closure of X Y in (X, Y )

((0, ); X Y ) V (p, 1 , Y, X), and moreover t t

X Y , and we shall show that x

(t) = (v() v(t))

(0, +; Y ) for every , and

0 in V (p, 1 , Y, X) as 0, which means that |x

. Due again to remark 1.2.4, x is the trace at t = 0 of a function

(0, +; Y ) and lim

be dened as above. Then lim

(0, ; Y ), and the

(0, ; Y ), and the norm

between X and Y if there is a constant c such that

between X and Y if there is k > 0 such that

(X, Y ) is the following one.

(0, +; X). Moreover v(0) = x, v(+) = 0, so that