Anda di halaman 1dari 46

Evolution equations

Georg Prokert, with additions by Mark Peletier


May 15, 2015

Prerequisites

In these notes we will assume that the reader is familiar with a number of concepts and tools from
Functional Analysis:
Real and complex Banach and Hilbert spaces, bounded linear operators in such spaces,
topological duals, the Hahn-Banach theorem;
Banach spaces of continuous and continuously differentiable functions (Cb and Cbk );
Weak differentiation, weak derivatives;
Lebesgue integration theory, and the spaces Lp , W k,p , and H k ;
Solvability of elliptic equations in W 2,2 in bounded domains.
Good references for (some of) this material include [Bre11, Eva02, Kre89, RR04].

Unbounded linear operators

Let X, Y be normed spaces. The set of bounded linear operators is noted as L(X, Y ).
Let now D = D(A) X be a linear subspace, and A : D Y a linear (not necessarily
bounded!) operator.
Notation: (A, D(A)) : X Y
Definition 1.1. G(A) := {(x, Ax) | x D} is called the graph of A. The linear operator A is
called closed if G(A) is closed in X Y . The linear operator A is called closable if G(A) = G(A)
for some linear operator (A, D(A)) : X Y . In this case A is called the closure of A.
G(A) is a linear subspace of X Y . If a closure exists, it is obviously unique. Clearly, if A is
closed then A is closable and A = A.
Lemma 1.2. Let (A, D(A)) : X Y be a linear operator. The following statements are
equivalent:
(i) A is closed.
X

(ii) For any sequence (xn ) in D(A) that satisfies xn x, Axn y for some x X, y Y , one
has x D(A), Ax = y.
Proof: Exercise!
Lemma 1.3. Let (A, D(A)) : X Y be a linear operator. The following statements are
equivalent:
(i) A is closable.
1

(ii) For any sequence (xn ) in D(A) that satisfies xn 0 and (Axn ) converges, we have Axn 0.
Proof: Exercise!
Examples:
Bounded linear operators (A, D(A)) : X Y are closable with D(A) = D(A). In particular, bounded operators are closed if and only if their domain of definition is closed. In
particular, operators A L(X, Y ) are closed.
The following example shows that differential operators with smooth coefficients are closable.
(This fact is not restricted to the L2 -context in which it is formulated here.)
Let Rd be a domain, X = Y = L2 (). Define (A, D(A)) by
X
D(A) = Cc (),
Au =
a u, a C (),

u D(A).

Then A is closable. To see this, let H m () := (H0m ())0 and define A L(L2 (), H m ())
by
Z
X
||
m
(1)
(a ) dx.
H m hA, iH0 :=
m

D(A) = A, if functions are


(Check that indeed A L(L2 (), H m ()) !) Furthermore, A|
identified with (anti)linear forms in the usual way. Note that with this identification we have
L2 () , H m (). Assume now that for some sequence (xn ) in D(A) we have xn 0,
n for all n and Ax
n 0 in H m (), so y = 0.
Axn y in L2 (). Then Axn = Ax
Therefore, A is closable by Lemma 1.3.
Let X = L2 (R), Y = K, and the operators (A, D(A)), (B, D(B)) given by
Z
D(A) = D(B) = C0 (R), Au = u(0), Bu =
u dx.
R

(As usual, A has to be interpreted in the sense of the continuous representative.) Then
neither A nor B are closable. (Exercise!)
Let X = Y = L2 (1, 1) (over R) and (A, D(A)) given by
D(A) = Cb1 (1, 1),

Au = u0 .

Then A is not closed. Indeed, let u, un X be given by


u(t) = |t|,

un (t) = p

t
t2

+ 1/n

Then un D(A), un u in X but u0n f in X, where



1 (t < 0)
f (t) =
1
(t > 0).
(Check!) So f
/ D(A) and A is not closed by Lemma 1.2.
However, A has the closure (A, D(A)) given by
D(A) = H 1 (1, 1),

Au = u0

(in the sense of weak derivatives). To prove this, we have to show that
{u, u0 ) | u Cb1 (1, 1)} = {(u, u0 ) | u H 1 (1, 1)},
2

where the closure on the left is in the norm of L2 (1, 1) L2 (1, 1).
I. : Let (un ) be a sequence in D(A) such that un u, u0n f in L2 (1, 1). By
the properties of weak derivatives, this implies that u is weakly differentiable and u0 = f
(recall!). Therefore u H 1 (0, 1).
II. : Fix u H 1 (1, 1) and define un := n u where n is a standard mollifier. (Extend
u by zero and restrict the convolution to (1, 1).) Then un u and u0n = n u0 u0 in
L2 (1, 1) (check!).
(Closure of elliptic operators)
Let Rd be a bounded smooth domain. Let X = Y = L2 () and (A, D(A))) be given by
D(A) = Cc (),

Au = u.

Its closure (A, D(A)) is given by


D(A) = H02 (),

Au = u

(again in the sense of weak derivatives.) To prove this, we have to show that
{(u, u) | u Cc ()} = {(u, u) | u H02 ()}.
I. : Let (un ) be a sequence in D(A) such that (un , Aun ) (u, f ) in L2 () L2 ().
Using the a priori estimate

v Cc ()
(1.1)
kvkH 2 () C kvkL2 () + kvkL2 ()
we get

kun um kH 2 () C kun um kL2 () + kun um kL2 () ,
i.e. (un ) is a Cauchy sequence in H 2 (). Therefore, u H02 () and u = f by continuity.
II. : Fix u H02 (). By definition of this space, there is a sequence (un ) in Cc () such
that un u in H 2 (). By continuity, un u in L2 ().
Remark: Analogous results hold for more general elliptic operators of order 2m, given by
X
Au =
a u
||2m

satisfying the ellipticity condition


X

a c||2m ,

Rd

||=2m

for some fixed c > 0. The essential point is that these assumptions imply an H 2m -estimate
parallel to (1.1).
Remark: The choice of D(A) is crucial. For example, if we define (A, D(A)) by
D(A) = {u Cb () | u| = 0},

Au = u,

we get
D(A) = H 2 H01 ().
(Closure of nonelliptic operators)
If a differential operator is not elliptic, the closure will in general be larger than the Sobolev
space corresponding to the order of the operator.

Consider, for example, X = Y = L2 ((0, 1)2 ) and the operator (A, D(A)) given by
D(A) = Cb1 ((0, 1)2 ),

Au = 1 u + 2 u.

Let A denote the characteristic function of the set A. Let (n ) be a sequence of smooth
functions on (1, 1) such that n (0,1) . Define un Cb1 ((0, 1)2 ) by
un (x, y) = n (x y).
Then Aun = 0 for all n and un {x>y} in L2 ((0, 1)2 ) (check!). Therefore, the characteristic
function {x>y} belongs to D(A), but obviously not to H 1 ((0, 1)2 ).
Definition 1.4. Let (A, D(A)) : X Y be a linear operator. The norm k kD(A) on D(A) given
by
kxkD(A) := kxkX + kAxkY
is called the graph norm of D(A).
Theorem 1.5. Let X and Y be Banach spaces. Then (D(A), k kD(A) ) is a Banach space if and
only if A is closed.
Proof: The map x 7 (x, Ax) induces an isomorphism from the normed space (D(A), kkD(A) )
to the normed space (G(A), k kXY ). As X Y is a Banach space, the latter (and therefore also
the former) is a Banach space if and only if it is closed, i.e. if A is closed.

Some basic theorems


of linear Functional Analysis

In this section, BX (x, r) will denote the open ball in the normed space X with center x X and
radius r > 0. If no confusion is likely, the space will not be indicated.
S
Theorem 2.1. (Baire) Let X be a Banach space. Assume X = nN Mn where all Mn are
closed subsets of X. Then one of the Mn contains an interior point.
Proof: Assume the opposite:
x X, > 0, n N :

B(x, ) \ Mn 6= .

(2.1)

Fix x0 X, 0 > 0 arbitrary. The set B(x0 , 0 ) \ M0 is nonempty by (2.1) and open, so there
exist x1 X, 1 (0, 0 /2) such that B(x1 , 21 ) B(0 ) \ M0 . Proceeding by induction and
using (2.1) repeatedly, we construct sequences (xn ) in X and (n ) in R+ such that
n+1 (0, n /2),

B(xn+1 , 2n+1 ) B(xn , n ) \ Mn .

Obviously, n 0. Moreover, (xn ) is Cauchy. Indeed, for given > 0 let n0 large enough
to ensure 2n0 < and let n, m n0 . Then by construction xn , xm B(xn0 , n0 ) and thus
kxn xm k < 2n0 < . So xn x in X. For any N N and any n N + 1, we have
xn B(xN +1 , N +1 ) and therefore
x B(xN +1 , N +1 ) B(xN +1 , 2N +1 ) B(xN , N ) \ MN .
S
In particular, x
/ MN for all N N. This is in contradiction to X = nN Mn .
Remark: It is clear from the proof that Theorem 2.1 holds more generally if X is just a complete metric space. However, completeness is necessary, as simple counterexamples show. (See
exercises.)

Definition 2.2. Definition: A mapping T : X Y is called open if for any open set U X,
the image T (U ) is open in Y .
Lemma 2.3. Let T : X Y be linear. Then T is open if and only if 0 in an interior point of
T (BX (0, 1)), i.e.
> 0 : BY (0, ) T (BX (0, 1)).
Proof: : trivial.
Fix U X open and y T (U ). Fix x U such that y = T x. As U is open, there is
> 0 such that B(x, ) U . Then, by linearity of T ,
BY (y, ) = y + BY (0, ) T (x + BX (0, 1)) = T (BX (x, )) T (U ).
Theorem 2.4. (Open Mapping theorem) Let X and Y be Banach spaces, T L(X, Y ). Then
T is surjective if and only if T is open.
Proof: : straightforward. (Check!)
: If T is surjective then
[
Y =
T (BX (0, n)).
nN+

By Baires theorem 2.1, there exist n0 N+ , 0 > 0 and y0 Y such that


BY (y0 , 0 ) T (BX (0, n0 )).

(2.2)

BY (0, ) T (BX (0, 1))

(2.3)

Fix x0 X such that T x0 = y0 .


1. We are going to show that

with =

0
n0 +kx0 k .

Indeed, by shifting we get from (2.2)

BY (0, 0 ) = BY (y0 , 0 ) y0 T (BX (0, n0 )) T x0 = T (B(x0 , n0 ))


T (B(0, n0 + kx0 k))
and upon scaling with the factor (n0 + kx0 k)1 this yields (2.3).
2. For as defined above, we are going to show
BY (0, ) T (B(0, 3)),

(2.4)

which implies that T is open by Lemma 2.3.


We fix y BY (0, ) and use (2.3) to successively construct approximate solutions to the
equation T x = y.
By (2.3),

x0 BX (0, 1) : kT x0 yk < , i.e. 2(T x0 y) BY (0, ),


2
so

x1 BX (0, 1) : kT x1 2(T x0 y)k < , i.e. 4(T x0 + 12 T x1 y) BY (0, ).


2
Continuing in this way, we obtain a sequence (xj ) in BX (0, 1) such that




k
X


j
j1
y T

2
x
.
j 2



j=0
Y
P j
As X is a Banach space, we can define x = j=0 2 xj and by continuity of T we get T x = y.
Moreover,

X
kxk
2j = 2 < 3.
j=0

This shows (2.4).


5

Corollary 2.5. (Inverse Mapping Theorem) Let X and Y be Banach spaces and T L(X, Y )
be bijective. Then T 1 L(Y, X).
Proof: As T is surjective, T is open by Theorem 2.4. This is equivalent to the continuity of
T 1 .
Corollary 2.6. (Equivalence of norms on Banach spaces) Let X be a vector space equipped
with two norms k k1 and k k2 , such that X is complete under both norms, and there is a C > 0
such that
kxk1 Ckxk2 ,
x X.
Then there is also a constant C 0 such that
kxk2 C 0 kxk1 ,

x X,

i.e. both norms are equivalent.


Proof: Apply Corollary 2.5 to the identity i : (X, k k2 ) (X, k k1 ).
Corollary 2.7. (Closed-Graph theorem)
Let X and Y be Banach spaces and A : X Y a linear operator with D(A) = X. Then A
is bounded iff A is closed.
Proof: If A is bounded, then A is continuous, and therefore the graph G(A) is closed. On the
other hand, if G(A) is closed, then the space X = D(A) can be given the graph norm k kD(A) ,
and the space (X, k kD(A) ) is a Banach space by Theorem 1.5. Moreover, kxkX kxkA for all
x X. Then the norms k kX and k kD(A) are equivalent by Corollary 2.6. In particular, there
is a constant C > 0 such that kAxkY CkxkX for all x X.
Definition 2.8. Let T L(X, Y ) be a family of bounded linear operators. T is called pointwise
bounded if for all x X the set {T xk | T T } is bounded in Y .
Theorem 2.9. (Uniform boundedness principle, Banach-Steinhaus theorem)
Let X be a Banach space. A family T L(X, Y ) is pointwise bounded if and only if it is
bounded in the norm of L(X, Y ), i.e. if
C > 0 : x X, T T :

kT xkY CkxkX .

Proof: : trivial.
: For k N+ we define the closed sets
\
Ak :=
{x X | kT xk k}.
T T

If T is pointwise bounded, then X =


contains an interior point, i.e.
x0 X, k0 N+ , > 0 :

kN+

z X :

Fix x X \ {0} and define z = z(x) =

Ak , and by Baires theorem 2.1 one of the sets Ak


kz x0 k (T T : kT zk k0 ).

kxk
x
+ x0 . Then x =
(z x0 ), kz x0 k = , and
kxk

therefore for any T T


kT xk =

kxk
2k0
kxk
kT z T x0 k
(kT zk + kT x0 k)
kxk.

By contraposition, we get:
Corollary 2.10. (Resonance theorem)
Let X be a Banach space and let T L(X, Y ) be unbounded. Then there exist an x X and
a sequence (Tn ) in T such that kTn xk .
6

Basic spectral theory

Definition 3.1. A linear operator (T, D(T )) : X Y is called invertible if T is bijective


(between D(T ) and Y ) and its inverse is in L(Y, X). (Note that the range of the inverse is D(T ),
but in general not X.)
In the remainder of this section, let X be a complex Banach space and (A, D(A)) : X X
an in general unbounded linear operator.
Definition 3.2. The set
(A) := { C | I A is invertible}
is called the resolvent set of A. The function R(, A) : (A) L(X) given by
R(, A) := (I A)1 ,

(A),

is called the resolvent of A.


The set
(A) := C \ (A)
is called the spectrum of A.
The set
p (A) := { C | N (I A) 6= {0}}
is called the point spectrum of A. Its elements are the eigenvalues of A. Obviously, p (A)
(A).
Lemma 3.3. If (A) is nonempty then A is closed.
Proof: Exercise!
This shows that spectrum and resolvent are interesting only for closed operators.
Lemma 3.4. (Properties of the resolvent)
(i) (A) is an open subset of C.
(ii) The resolvent R(, A) is an analytic L(X)-valued function on (A).
(iii) The resolvent has no analytic extension to any subset of C larger than (A).
Proof: Fix 0 (A), let C such that
| 0 | < kR(0 , A)k1
L(X) .

(3.1)

then, by using
I A = 0 I A + ( 0 )I
and applying a Neumann series for the calculation of the inverse, we get (A) and
R(, A) =

(1)n ( 0 )n R(0 , A)n+1 ,

n=0

where the operator power series on the right converges absolutely in L(X). (Exercise!) This shows
(i) and (ii).
More precisely, as (3.1) implies (A) we have for all (A)
| 0 | kR(0 , A)k1
L(X)
and hence
kR(0 , A)k

1
.
| 0 |

Taking the supremum over these yields


kR(0 , A)k

1
dist(0 , (A))

which implies (iii).


7

(3.2)

Operator semigroups and generators

Let X be a Banach space.


Definition 4.1. A family {T (t) | t 0} L(X) of bounded linear operators on X such that
T (0) = I,
s, t 0

T (s + t) = T (s)T (t)

is called a semigroup of (linear) operators on X. This semigroup is called


- uniformly continuous if the map t 7 T (t) is continuous from [0, ) to L(X),
- strongly continuous or C0 -semigroup if for all u X the map t 7 T (t)u is continuous
from [0, ) to X.
We will also use the terms strongly continuous and uniformly continuous (and obvious modifications of it) for the operator valued mappings themselves.
Clearly, any uniformly continuous semigroup is strongly continuous.
Examples:
Let X = L (R) and define for t 0
(T (t)u)(x) = u(x + t),

u L (R).

Then {T (t)} is a semigroup of linear operators on X which is not strongly continuous.


(Exercise!)
Let X = BU C(R) and define for t 0
(T (t)u)(x) = u(x + t),

u BU C(R).

(4.1)

Then {T (t)} is a semigroup of linear operators on X which is strongly but not uniformly
continuous. (Exercise!)
For any complex Banach space X and any C, the semigroup {T (t)} of operators on X
given by
T (t) = et I
(4.2)
is uniformly continuous.
Definition 4.2. Let {T (t)} be a strongly continuous semigroup of operators on X. The linear
(possibly unbounded) operator (A, D(A)) : X X given by


T (t)u u
exists in X ,
D(A) := u X | lim
t0
t
T (t)u u
Au := lim
, u D(A),
t0
t
is called the generator of the semigroup {T (t)}.
Examples:
Let X = BU C(R) and let {T (t)} be given by (4.1). The generator (A, D(A)) is an unbounded
operator that satisfies
Au = u0 ,

BU C 1 (R) D(A) 6= X,

(Check! Later we will see that D(A) = BU C 1 (R).)


8

u BU C 1 (R).

Let X be an arbitrary complex Banach space and let {T (t)} be given by (4.2). The generator
(A, D(A)) is the bounded operator given by
D(A) = X,

A = I.

(Check!)
The above definitions give rise to a number of questions:
Which (bounded or unbounded) linear operators on a Banach space generate a (stronglycontinuous) semigroup of operators? Is this semigroup unique?
What properties of the semigroup are implied by properties of its generator, and vice versa?
Answering these questions is a core subject of the theory of operator semigroups.

Uniformly continuous semigroups

A particularly simple (in some sense, nearly trivial) case in point is given by the property of
uniform continuity. This property is in one-to-one correspondence with boundedness of the
generator.
Theorem 5.1. (Uniformly continuous semigroups)
Let X be a Banach space.
(i) Any bounded linear operator A L(X) generates precisely one uniformly continuous semigroup of operators on X.
(ii) The generator of any uniformly continuous semigroup of operators on X is bounded and
satisfies D(A) = X.
Proof: 1. Fix A L(X) and define for t 0
T (t) :=

X
1 k k
t A .
k!

(5.1)

k=0

Then clearly T (0) = I, and the series converges absolutely in L(X), uniformly as t varies in
compact subsets of R. This implies T (t + s) = T (t)T (s), t 7 T (t) is continuous from [0, ) to
L(X), and
T (t) I
lim
= A in L(X)
t0
t
Therefore, {T (t)} is a uniformly continuous semigroup on X generated by A.
2. Let {T (t)} Rbe a uniformly continuous semigroup on X. Then, forR sufficiently small > 0,

the mean value 1 0 T (s) ds is close to I and therefore invertible, hence 0 T (s) ds is invertible as
well. So, for any such , the operator
Z
A := (T () I)

1
T (s) ds

is a well-defined element of L(X). We are going to show that


A = lim
t0

T (t) I
,
t

so A generates {T (t)} (and, in particular, is independent of .)

(5.2)

Indeed, for t (0, ) we have



Z
Z
Z
1
1
T (s) ds
T (s) ds =
T (s + t) ds
(T (t) I)
t
t
0
0
0


Z +t
Z +t
Z
Z t
1
1
=
T (s) ds
T (s) ds =
T (s) ds
T (s) ds ,
t
t
t
0

0
so
1
1
(T (t) I) =
t
t

Z

+t

T (s) ds

 Z
T (s) ds

1
t0
T (s) ds
A.

This proves (5.2).


3. Let {T (t)}, {S(t)} be uniformly continuous semigroups on X, both generated by A L(X).
Then, by 2.,
S(t) I
T (t) I
= lim
= A.
(5.3)
lim
t0
t0
t
t
Fix > 0, > 0 arbitrary. By continuity and compactness, there is a C > 0 such that
kT (t)k, kS(t)k C,

t [0, ],

and by (5.3) there is a > 0 such that for all h (0, )

kT (h) S(h)k
< 2 .
h
C
Fix t [0, ]. Choose n N large enough such that t/n < . Then
kT (t) S(t)k = kT (t/n)n S(t/n)n k

n1
X
k=0

T (t/n)nk1 (T (t/n) S(t/n)) S(t/n)k k


|
{z
}
| {z }
T ((n k 1)t/n)
S(kt/n)

nC 2 kT (t/n) S(t/n)k < nC 2

t/n
.
C 2

Therefore T (t) = S(t) for t [0, ] and {T (t)} = {S(t)} as was arbitrary.
Remarks:
For A L(X), the semigroup given by (5.1) is called the (operator) exponential function
of A and is also denoted by etA .
This exponential function extends analytically (with values in L(X)) to the whole complex
plane. In particular, the map t 7 T (t) is analytic from R to L(X) and has the property
T (s + t) = T (t)T (s) for all s, t R, so the family {T (t) | t R} is called an (analytic) group
of operators on X. It satisfies
AT (t) = T (t)A,

dn
T (t) = An T (t),
dtn

and the bound


kT (t)k

t R, n N,

X
1 k
|t| kAkk = ekAk|t| .
k!

k=0

Theorem (5.1) can be straightforwardly generalized to the case where L(X) is replaced by
an arbitrary Banach algebra.
The uniqueness of the semigroup generated by A will later be reproved in the more general context of strongly continuous semigroups. The proof given above is included for selfcontainedness and because the proof for strongly continuous semigroups does not generalize
to the Banach-algebra case.
10

Basic properties of strongly continuous semigroups

Let X again be a Banach space. We return to the general situation and recall the definition of
strong continuity from Section 4. Let {T (t)} be a semigroup of operators on X. We will investigate
the following properties:
(a) t 7 T (t) is strongly continuous,
(b) t 7 T (t) is strongly right continuous,
(b) t 7 T (t) is strongly right continuous at t = 0,
(c1) for all > 0 there is an M > 0 such that kT (t)k M for all t [0, ],
(c1) there are > 0, M > 0 such that kT (t)k M for all t [0, ],
t0

(c2) there is a dense subset D X such that T (t)x x for all x D.


Lemma 6.1. (Criteria for strong continuity)
For any operator semigroup {T (t)} we have
(c1) (c1)

and

(a) (b) (b) ((c1)(c2)).

Proof: (a)(b)(b)(c2): trivial.


(c1)(c1): trivial.
(c1)(c1): see exercises!
(b)(b): Fix t0 > 0, x X, and let h > 0. Then
h0

T (t0 + h)x = T (h)T (t0 )x T (t0 )x.


(b)(c1): Assume (c1) does not hold: Then there is a sequence (tn ) of positive numbers such
that tn 0 and kT (tn )k . By Corollary 2.10, this implies that there is an x X such that
the set {T (tn )x | n N} is unbounded. This is in contradiction to (b).
((c1)(c2))(b): Fix x X, > 0. Choose x
D such that k
x xk < /(2(M + 1)). Choose
(0, ) such that kT (t)
xx
k < /2 for all t [0, ]. For t [0, ] we have then
kT (t)x xk kT (t)(x x
)k + kT (t)
x xk < .
xx
k + k
|
{z
} |
{z
}
| {z }
< 2
< 2(M+1)
< M
2(M +1)

This implies (b).


((b)(c1))(a): We have to show that t 7 T (t) is strongly left continuous on (0, ). Fix
t0 > 0, x X, and let h (0, t0 ). Then
kT (t0 h)x T (t0 )xk kT (t0 h)k kx T (h)xk 0
{z
}|
{z
}
|
0
Mt0

as h 0.

Gathering the above implications, we finally get


((c1)(c2)) ((b)(c1)) ((b)(c1)) (a).
This completes the proof.
Theorem 6.2. (Basic properties)
Let {T (t)} be a strongly continuous semigroup of operators on X with generator (A, D(A)).
Then:
(i) There are M, > 0 such that
kT (t)k M et ,
11

t 0.

(ii) For all x X, t > 0 we have


Z t
T (s)x ds D(A),

T (s)x ds = T (t)x x.

A
0

(iii) For all x D(A), t 0 we have


T (t)x D(A),

AT (t)x = T (t)Ax.

(6.1)

Moreover, the mapping t 7 T (t)x is continuously differentiable from [0, ) to X, and


d
T (t)x = AT (t)x.
dt

(6.2)

Caution: In general, (6.1) and (6.2) are not true for all x X!
Proof of Theorem 6.2: (i): See exercise!
(ii): Fix x X, t > 0, and let h > 0. Then
Z
Z
T (h) I t
1 t
T (s)x ds =
(T (s + h)x T (s)x) ds
h
h 0
0
Z
Z
1 t+h
1 h
h0
=
T (s)x ds
T (s)x ds T (t)x x
h t
h 0
since the integrand is continuous with values in X. By the definition of D(A), therefore
D(A), and its image under A is T (t)x x.
(iii): Fix x D(A), t 0, and let h > 0. Then

Rt
0

T (s)x dx

T (t + h)x T (t)x
T (h)x x h0
T (h) I
T (t)x =
= T (t)
T (t)Ax
h
h
h
by continuity of T (t). This implies (6.1) and the differentiability of t 7 T (t)x from the right. To
show left differentiability and (6.2), fix t > 0 and let h (0, t). Then
T (t)x T (t h)x
T (t)Ax
h


T (h)x x
h0
=T (t h)
Ax + T (t h)Ax T (t)Ax 0,
|
{z
}
h
|
{z
}
h0
0
h0
0
where we used that the operators T (t h) are bounded uniformly in h by (i) and that the map
t 7 T (t)Ax is (right) continuous at t. Finally, by (6.1) and (6.2) we have
d
T (t)x = T (t)Ax,
dt
so the derivative is continuous in t.
Theorem 6.3. (Properties of generators)
Let (A, D(A)) be the generator of a strongly continuous semigroup {T (t)} on X. Then D(A)
is dense in X, and A is a closed operator on X.
Proof: For any x X we have
1
t
and

Rt
0

t0

T (s)x ds x
0

T (s)x ds D(A) for all t > 0 by Theorem 6.2 (ii). This proves that D(A) is dense in X.
12

To prove closedness, fix a sequence (xn ) in D(A) such that xn x, Axn y in X. Let t > 0.
By Theorem 6.2 (iii), the mapping
s 7 (s) := T (s)xn
is continuously differentiable on [0, t], and 0 (s) = T (s)Axn . Integration from 0 to t yields
Z t
T (t)xn xn =
T (s)Axn ds.

(6.3)

As the operators T (s), s [0, t] are uniformly bounded in s, we have T (s)Axn T (s)y uniformly
in s. Therefore, taking n in (6.3),
Z t
T (t)x x =
T (s)y ds
0

and hence
1
T (t)x x
=
t
t

t0

T (s)y ds y.
0

Consequently, x D(A) and Ax = y. This proves that A is closed.


Theorem 6.4. (Uniqueness)
Let {T (t)}, {S(t)} be two strongly continuous semigroups of operators on X having the same
generator. Then T (t) = S(t) for all t 0.
Proof: See exercises!
Definition 6.5. For a given linear operator (A, D(A)) : X X we define inductively its powers
(An , D(An )) : X X by
D(An ) := {x D(An1 ) | Ax D(An1 )},

An x = AAn1 x,

x D(An )

(n 2). Moreover, let

D(A ) :=

D(An ).

(6.4)

n=1

(Note that there is no definition of any operator named A !)


Lemma 6.6. Let (A, D(A)) be the generator of a strongly continuous semigroup of operators
{T (t)}. Then even D(A ) is dense in X.
Proof: For x X and Cc ((0, )), let
Z
x() :=
(s)T (s)x ds
0

and
Y := {x() | x X, Cc ((0, ))}.
Obviously, Y is a linear subspace of X.
1. We show Y D(A ). Fix x X, Cc ((0, )), and let h > 0 be small. Then
Z
T (h) I
1
x() =
(s)(T (s + h) T (s))x ds
h
h 0
Z
Z
(s) (s h)
h0
=
T (s)x ds
0 (s)T (s)x ds,
h
0
0
where we used that the support of is away from zero and
s (0, ) by compactness of the support of .
13

(s)(sh)
h

0 (s) uniformly for

Hence x() D(A) with Ax() = x(0 ). Induction shows then x D(An ) and An x() =
(1)n x((n) ).
2. We show that Y is dense in X. Suppose the opposite. Then Y 6= X, and the Hahn-Banach
theorem1 implies the existence of a nonzero bounded linear functional x X \ {0} such that
Z
hx , x()i =
(s)hx , T (s)xi ds
0

for all

Cc ((0, ))

and x X. This implies, by continuity of s 7 hx , T (s)xi that


hx , T (s)xi = 0

for all s 0

and in particular hx , xi = 0 for all x X, in contradiction to x 6= 0.

The Hille-Yosida theorem for contractions

Let X be a complex Banach space. Here and in the sequel we will write R+ = (0, ).
Definition 7.1. Definition: A semigroup {T (t)} of operators on X is called a contraction
semigroup2 if kT (t)k 1 for all t 0.
Theorem 7.2. (Hille-Yosida)
A linear operator (A, D(A)) on X is the generator of a strongly continuous contraction semigroup {T (t)} if and only if
(i) A is closed and densely defined,
(ii) R+ (A), and
kR(, A)k

1
,

> 0.

Proof: =: Statement (i) follows by Theorem 6.3. To show (ii), we fix > 0 and define
B L(X) by
Z
Bx :=
et T (t)x dt.
0

The assumption that {T (t)} is a strongly continuous contraction semigroup implies that B is well
defined and that
Z
1
kBk
et dt = .

0
We will show that B is in fact the resolvent operator R(, A). For this, we have to show that
x X :

Bx D(A), ABx = Bx x,

(7.1)

and
x D(A) :

BAx = ABx.

(7.2)

(Check!)
To show (7.1), fix x X and let h > 0. Then
Z
T (h) I
1 t
Bx =
e (T (t + h) T (t))x dt
h
h 0
Z
Z
eh t
1 t
=
e T (t)x dt
e T (t)x dt
h h
h 0
Z
Z
eh 1 t
1 h t
h0
=
e T (t)x dt
e T (t)x dt B x.
h
h
| {z } | h
{z
} | 0
{z
}
h0
h0
h0

B
x
1 A proof of this theorem is not part of these lectures. The reader is encouraged to consult any book on basic
linear functional analysis, eg. [Bre11, Kre89].
2 Note the different usage of the term contraction in contrast to the Banach contraction principle!

14

To show (7.2), fix x D(A). Then


Z
Z
et AT (t)x dt
et T (t)Ax dt =
BAx =
0
0
Z
et T (t)x dt = ABx.
=A
0

Observe that the closedness of A is used in the third equality.


=: To shorten notation, we write R() instead of R(, A).
1. We show
x X :
lim R()x = x.

(7.3)

Assume first x D(A). Then


kR()x xk = kAR()xk = kR()Axk

kAxk 0.

In the general case, fix x X and > 0. As D(A) is dense, there is a x


D(A) such that
kx x
k < /3. For large we have kR()
xx
k < /3 and thus
kR()x xk kR()(x x
)k + kx x
k + kR()
xx
k < .
2. For > 0 we define the so-called Yosida approximation
A := AR() = (R() I).
Note that A L(X) for all > 0, and by (7.3), for all x D(A),

A x = AR()x = R()Ax Ax.

(7.4)

The semigroups {etA } generate by A are uniformly continuous, and they are contraction semigroups because

2

tA
2
2



e =
(7.5)
et( R()I) et et R() et et kR()k 1.
Furthermore, we can estimate the dependence of etA x on by
Z

tA
1 d h tsA t(1s)A i

e x etA x =
e
e
x ds


0 ds
Z 1

tsA t(1s)A

t
e
(A A )x ds tk(A A )xk
e
0

for t 0, x X, , > 0. In particular, for x D(A), by (7.4),


tA

e x etA x t(k(A A)xk + k(A A)xk) ,
0.
3. This implies that the limit
T (t)x := lim etA x

(7.6)

exists for all t 0 and x D(A). (Check!) The operator T (t) defined by this is clearly linear
and bounded with kT (t)k 1 because of (7.5), and therefore it extends by density with the
same estimate to all x X. We straightforwardly have the semigroup properties T (0) = I and
T (s + t) = T (t)T (s). Finally, as the convergence in (7.6) is uniform with respect to t for t varying
in bounded sets, we conclude that t 7 T (t)x is continuous. (Check!)
4. Let (B, D(B)) the generator of the strongly continuous semigroup of contractions {T (t)}.
We have to show B = A. By assumption and from the first part of the proof we get R+
15

(A) (B). Therefore it is sufficient to show A B (see exercise!). To see this, fix x D(A).
Then
Z t
Z t
T (s)Ax ds,
T (t)x x = lim etA x x = lim
esA A x ds =

0
sA

where the last equality follows from the fact that e A x T (s)Ax uniformly in s [0, t].
(Check!) Consequently,
Z
T (t)x x
1 t
t0
T (s)Ax ds Ax,
=
t
t 0
so x D(B) and Bx = Ax.
Remarks:
The first part of the proof can be easily modified to show that if A generates a contraction
semigroup {T (t)}, then the right complex half plane C+ := { C | Re > 0} belongs to
(A), and we have
Z
R(, A)x =
et T (t)x dt,
x X, C+ ,
(7.7)
0

with an estimate

1
.
Re
Note, furthermore, that 7 R(, A)x is in fact the (vector valued) Laplace transform of
t 7 T (t)x.
kR(, A)k

The Hille-Yosida theorem and the above remark can be straightforwardly generalized to give
a characterization of generators of so-called quasicontraction semigroups, i.e. semigroups
that satisfy kT (t)k et for some R. (See exercises.)

Dissipative operators
and the Lumer-Phillips theorem

Let X be a complex Banach space.


Definition 8.1. For x X, we call
F (x) := {x X | hx , xi = kx k2 = kxk2 }
the duality set of x.
Remarks:
The Hahn-Banach theorem implies F (x) 6= for all x X.
If X is a Hilbert space, then F (x) = {i(x)}, where i is the Riesz antiisomorphism. (Check!)
In general, F (x) is not necessarily a singleton. (Example?)
Definition 8.2. A linear operator (A, D(A)) on X is called dissipative if
x D(A) :

x F (x)

Rehx , Axi 0.

In the case of a Hilbert space with inner product (, ) this just means
Re(Ax, x) 0,

x D(A).

This shows the close relationship of dissipativity to energy estimates of the corresponding linear
evolution equation x = Ax.
16

Theorem 8.3. A linear operator (A, D(A)) on X is dissipative if and only if


kx Axk kxk,

for all > 0, x D(A).

Proof: We only show here. Let A be dissipative. Fix > 0, x D(A), and x F (x)
such that Rehx , Axi 0. Then
kx Axkkxk |hx , x Axik Rehx , xi Rehx , Axi kxk2 .
This implies the result.
Corollary 8.4. Let (A, D(A)) be dissipative and assume there is a 0 > 0 such that 0 I A is
surjective. Then R+ (A), and
kR(, A)k

1
,

> 0.

Proof: Let S := R+ (A). Then S 6= by assumption and S is open. By Theorem 8.3,


we have R(, A) 1/ for all S. Therefore (cf. (3.2)) dist(, (A)) for these , which
implies S = R+ .
Theorem 8.5. (Lumer-Phillips)
Let (A, D(A)) be a densely defined linear operator on X.
(i) If A is dissipative and there is a 0 > 0 such that 0 I A is surjective then A generates a
strongly continuous semigroup of contractions.
(ii) If A generates a strongly continuous semigroup of contractions {T (t)} then
x D(A) :

x F (x) :

Rehx , Axi 0.

(8.1)

(Recall that F (x) 6= , hence (8.1) implies that A is dissipative.)


Proof: (i) follows by Corollary 8.4 and the Hille-Yosida theorem.
(ii): Fix x D(A) and x F (x). Let t > 0. Then
Rehx , T (t)x xi = Rehx , T (t)xi kxk2 |hx , T (t)xi| kxk2 0.
The result follows by dividing by t and taking the limit t 0.

The Hille-Yosida theorem


for bounded semigroups

Definition 9.1. A semigroup of operators {T (t)} on the Banach space X is called bounded if
there is an M > 0 such that kT (t)k M for all t 0.
Observe that if A generates a strongly continuous semigroup of operators then there is an
> 0 such that A I generates a bounded strongly continuous semigroup. (Check!) Hence,
up to a spectral shift, a characterization of generators of bounded strongly continuous semigroups
covers the general case. The main result of this section gives such a characterization, although its
direct practical applicability is limited.
The proof rests on the idea of introducing equivalent norms with respect to which given bounded
operators have a norm not exceeding 1.
Lemma 9.2. Let (A, D(A)) be a linear operator on X satisfying R+ (A) and
M > 0 :

n N+ , > 0 :

kn R(, A)n k M.

Then there is an equivalent norm ||| ||| on X such that


|||R(, A)x||| |||x|||,
for all x X, > 0.
17

kxk |||x||| M kxk

Proof: For > 0 we define the norm k k on x by


kxk := sup kn R(, A)n xk .
n0

Then, for all x X, by assumption,


kxk kxk M kxk

(9.1)

kR(, A)xk kxk .

(9.2)

and by definition of k k
Fix (0, ] and x X. Define y := R(, A)x = R(, A)(x + ( )y). Then, by (9.2),


1

1
kyk .
kyk kx + ( )yk kxk + 1

Subtracting the
kyk kxk , or

second

term

on

the

right

kR(, A)k kxk ,

and

multiplication

0 < .

by

yields
(9.3)

Define now
|||x||| := lim kxk .

The lemma follows by taking the upper limit in (9.1) and (9.3).
Theorem 9.3. (Hille-Yosida)
A linear operator (A, D(A)) on X is the generator of a strongly continuous semigroup {T (t)}
that satisfies kT (t)k M for all t 0 if and only if
(i) A is closed and densely defined,
(ii) R+ (A), and
kR(, A)n k

M
,
n

n N, > 0.

Proof: : Define on X the norm ||| ||| by


|||x||| := sup kT (s)xk.
s0

This norm is obviously equivalent to kk, so {T (t)} is a strongly continuous semigroup of operators
on (X, ||| |||), and it is even a semigroup of contractions on this space, as
|||T (t)x||| = sup kT (t + s)xk = sup kT (s)xk |||x|||.
s0

st

The generator of the semigroup and its topological properties (closedness, denseness of D(A),
spectrum) do not depend on the norm. It remains to show the resolvent estimate. We will denote
the operator norm of L(X) corresponding to ||| ||| by that symbol as well. Fix > 0, n N, and
x X. Theorem 7.2 yields that |||R(, A)||| 1, hence |||n R(, A)n ||| 1, and
kn R(, A)n xk |||n R(, A)n x||| |||x||| M kxk.
: Let |||||| be the norm from Lemma 9.2. By Theorem 7.2, A generates a strongly continuous
semigroup of contractions {T (t)} on (X, ||| |||). By equivalence of norms, {T (t)} is also strongly
continuous (and also generated by A) on (X, k k). Moreover, for t 0, x X,
kT (t)xk |||T (t)x||| |||x||| M kxk.

18

Remark: It is not hard to see that the first part of the proof can be generalized to yield (A)
for all complex satisfying Re > 0 and for these ,
kR(, A)k

M
.
Re

(9.4)

Moreover, inspection of the arguments that prove the representation formula (7.7) shows that it
remains valid for (generators of) bounded semigroups.
In view of Theorem 6.2(i) and the remarks after Theorem 7.2 we can give now the following
characterization of generators of strongly continuous semigroups:
Corollary 9.4. A linear operator (A, D(A)) on a Banach space X generates a strongly continuous
semigroup of operators if and only if A is closed and densely defined and there are constants M > 0
and R such that (, ) (A) and
n N, > :

kR(, A)n k

M
.
( )n

In this case, we even have (A) and


kR(, A)n k

n N :

M
(Re )n

for all complex that satisfy Re > .


The semigroup {T (t)} generated by A satisfies
kT (t)k M et ,

t 0.

In the course of the proof of Theorem 9.3 we also showed the following property:
Corollary 9.5. Let T (t) be a bounded strongly continuous semigroup on X. Then there exists an
equivalent norm ||| ||| on X such that T is a contraction with respect to ||| |||.

10

Inverse Laplace transform

As we have seen above, the resolvent of the generator can (in a certain sense) be obtained as
Laplace transform of the semigroup. (See the first part of the proof of Theorem 7.2. The resolvent
representation given there generalizes straightforwardly to bounded semigroups.) Therefore, it is
natural to construct the semigroup by inverting the Laplace transform. This corresponds to a
classical technique of solving inital value problems for linear evolution equations.
We start with an informal discussion of the idea behind the inversion of the Laplace transform.
Let {T (t)} be a strongly continuous semigroup of operators on the Banach space X which we
assume (without loss of generality) to be bounded. For f L1 (R+ ) we define T [f ] L(X) by
Z
(T [f ])x :=
f (s)T (s)x dx,
x X.
0

Is is easy to check that



T L L1 (R+ ), L(X) .

(10.1)

Observe that for x X


R(, A)x = T [exp ]x,
C+ ,
Z t
S(t)x :=
T (s)x ds = T [1[0,t] ]x,
0

19

t 0.

where 1[0,t] , exp L1 (R+ ) are given by



1 (s < t)
1[0,t] (s) :=
,
0 (s > t)

exp (s) = es .

Fix > 0. We are going to approximate 1[0,t] by functions Hn,t given by the complex line
integral
Z +in t
Z +in (ts)
1
e
e
1
Hn,t (s) :=
exp (s) d =
d
2i in
2i in

with n . (The integration is along the straight line from in to +in). For t and s 6= t fixed,
the improper complex integral arising for n can be calculated using the residue theorem,
giving 1 for t > s and 0 for t < s, respectively:
in

in

t>s

t<s

The next lemma provides the precise result we will need:


Lemma 10.1. For fixed > 0 we have
lim Hn,t = 1[0,t]

in L1 (R+ ),

uniformly with respect to t [0, ].


Proof: We split
kHn,t 1[0,t] kL1 (R+ )

Z t
Z +in (ts)
Z
Z +in (ts)


1
1
e
e

ds +

d

1
d ds
=
2i

2i in

0
in
t

Z t
Z +in u
Z
Z +in u
1



e
e

1
=
d 1 du +
d du,
2i

2i in

0
in
0
{z
} |
{z
}
|
J1 (n, t)
J2 (n)
where we substituted u = t s > 0 in the first integral and u = s t > 0 in the second.
n
1. As J1 (n, t) is increasing in t, it is sufficient to show J1 (n, t) 0 for all t [0, ]. In view
Rt
of the structure J1 (n, t) = 0 n (u) du this follows by Lebesgues dominated convergence theorem
n
if we show n (u) 0 for all u (0, t] and n (u) C with C independent of u and n.
0
Let := { + nei | (/2, 3/2)}. The (oriented) union of 0 and the line from in
to + in is a closed, simple, positively oriented curve that encloses the origin if n > . Therefore,
the residue theorem yields
Z u
1
e
d = 1
2i

20

and therefore

Z
1
n (u) =
2i

+in

in



Z 3/2

i
1
eu
exp(u(
+
ne
))

i
ine
d 1 =
d
.
2i /2

+ nei

Using the estimates | + nei | n , | exp(u( + nei ))| = eu(+n cos ) we get
neu
n (u)
2(n )

3/2

un cos

/2

neu
d =
(n )

/2

eun cos d.

By the concaveness of the cosine function, we have cos 1 2 for (0, /2). Using this and
elementary calculations, we get
neu
0 n (u)
(n )

/2

eun(1 ) d =

eu n 1 enu
e
2 n | nu
{z }
n

for n > 2. This shows that J1 (n,Rt) 0 uniformly in t as announced.

2. Similarly, we have J2 (n) = 0 n (u) du, n (u) 0 and will show that n (u) 0 as n

for all u (0, ) and n (u) (u)


with L1 (R+ ). This implies J2 (n) 0 by dominated
convergence.
Let 00 := { + nei | (0, /2)}. Then, by the Cauchy integral theorem,
Z

+in

in

eu
d =

Z
00

eu
d

and analogous estimates as in 1. (with || n) yield



Z +in u
eu 1 enu n
1
e

d
n (u) =
0.
2i in

2
nu
Moreover, as the function x 7

1ex
x

is decreasing on R+ ,

n (u)

eu 1 eu

=: (u)
2
u

with L1 (R+ ). This proves J2 (n) 0.


To simplify notation for we set for > 0
Z

+i

+in

f () d := lim

f () d.
in

For further reference we recall that we have shown in particular



Z +i u
e
1
1 if u > 0,
d =
0 if u < 0.
2i i
Theorem 10.2. (Complex inversion formula)
For > 0, x X, t > 0 we have
Z

T (s)x ds =
0

1
2i

+i

et
R(, A)x d.

The convergence of the improper integral is uniform with respect to t [0, ].

21

(10.2)

Proof: For fixed n N+ , t [0, ], and x X we have

Z
T [Hn,t ]x =

Hn,t (s)T (s)x ds =


0

1
2i

+in

in

e(ts)
T (s)x dds.


(ts)
As the integrand (, s) 7 e T (s)x is in L1 ( + i(n, n)) (0, ), X we get by (the vector
valued version of) Fubinis theorem
1
T [Hn,t ]x =
2i

+in

in

et

1
T (s)x dsd =
2i

+in

in

et
R(, A)x d.

Hence, by (10.1) and Lemma 10.1,


Z

T [1[0,t] ]x = T [ lim Hn,t ]x = lim T [Hn,t ]x

T (s)x ds =

1
2i

+i

et
R(, A)x d.

Note, however, that the integral does not converge absolutely. For a similar reason, the validity
of the following classical inversion formula for Laplace transforms is restricted to x D(A) an
cannot be extended by a density argument to the complete space.
Corollary 10.3. For x D(A), > 0, t > 0 we have
T (t)x =

1
2i

+i

et R(, A)x d.

(10.3)

Proof: We have

+i

et
R(, A)Ax d

i
0
Z +i
Z +i t
1
1
e
et R(, A)x d
x d
2i i
2i i

Z
T (t)x x =

T (s)Ax ds =

1
2i

and the result follows from (10.2).

11

An exponential formula

Another way of approximating the semigroup with generator A powers of the resolvent and is
closely related to solving the initial value problem
u = Au,

u(0) = x,

approximately by the Euler-backward method. For fixed t > 0, n N large, h := t/n one
approximates u(t) := T (t)x by un , where
uk+1 uk
= Auk+1 ,
h
Then, as one easily checks,
un =

n

k = 0, . . . , n 1,

n


, A )n x.

u0 = x.

t
t
The following theorem can be considered as a convergence result for this approximation.

22

Theorem 11.1. Let t 7 T (t) be a strongly continuous semigroup of operators on X generated by


A. Then
 n  n n
T (t)x = lim
R
,A
x
(11.1)
n t
t
uniformly for t (0, ].
Eqn. (11.1) is called an exponential formula or the Post-Widder inversion formula. Note
that it makes sense because large real numbers belong to (A).
Proof: Fix M, 0 large enough to ensure kT (t)k M et , t 0. Then for > 0 we have
Z
R(, A)x =
es T (s)x ds.
0

Taking the n-th order derivative with respect to yields


Z
n
n+1
n
(1) n!R(, A)
x = (1)
sn es T (s)x ds.
0

Substituting s = vt, setting :=


n
t

n
t,

and multiplication by nn+1 gives

Z
n+1
nn+1 v n
,A
x=
(ve ) T (vt)x dv.
t
n! 0

n

Using
nn+1
n!

(vev )n dv = 1

(11.2)

we find
:=

n
t

Z
n+1
nn+1 v n
x T (t)x =
,A
(ve ) (T (vt)x T (t)x) dv,
t
n! 0

n

and, in view of (7.3), it is sufficient to show 0 as n .


Fix > 0. By continuity, there are a < 1, b > 1 such that
kT (vt)x T (t)xk <

for all v (a, b) and all t [0, ].

We split
nn+1
=
n!

Z
... +

Z
... +

...
b

and estimate the terms on the right separately.


1. For v (0, a) we have 0 < vev < aea := < e1 . Therefore
Z a

n+1

nn+1
(vev )n (T (vt)x T (t)x) dv n
n aM (ea + e )kxk


n!
n!
0
nn+1 n n
C
0.
n!
2. In view of (11.2) and (11.3) we directly have
Z

nn+1 Z b
b
nn+1


v n

(vev )n dv < .
(ve ) (T (vt)x T (t)x) dv

n! a
n!
a

23

(11.3)

3. We have := beb < e1 . So



Z

nn+1
v n

(ve ) (T (vt)x T (t)x) dv


n!
b
Z
n+1
n

2M
v n evn e dv kxk
n!
b
Z
0
0
nn+1 bn b 0
v=v 0 +b
(v + b)n ev n ev dv 0
=
2M kxk
e
e
n!
0
Z
n

nn bn 
=v 0 n
=
C
e
+ b e e n d
n!
n
0
n
Z 

/b
nn n
n
1+
=
C

e(1 n ) d 0,
n
|n!{z } 0 |
{z
}
n
0
(1 1 )
b
n
<e
as the integral exists and is bounded independently of n for n large.

12

Analytic Semigroups of Operators

For (0, ) we define the sector


:= {z C \ {0} | | arg z| < }
in the complex plane. Note that is open, (0, ) , and that for > 0, z, w we have
z, z + w .
Let X be a Banach space.
Definition 12.1. A mapping T : L(X) is called an analytic semigroup of operators
(in ) if
the mapping z 7 T (z) is analytic,3
T (z1 + z2 ) = T (z1 )T (z2 ) for all z1 , z2 ,
limz0 T (z)x = x for all x X.
For some purposes it is convenient to define additionally T (0) = I. It is clear that with this
definition, the restriction of T to the real half axis [0, ) is a strongly continuous semigroup of
operators t 7 T (t).
Remark: Observe, however, that while T is analytic (with values in L(X)) on the open set
(0, ), in general it need not even be continuous (with values in L(X)) on the closed interval
[0, ). In fact, continuity on the closed interval would imply that the semigroup is uniformly
continuous. Of course, for a fixed x X, t 7 T (t)x is continuous with values in X, as required
by the definition above.
The analyticity implies that for any x X the map t 7 T (t)x has derivatives of all orders on
(0, ). This implies (recall (6.4) and check!)
T (t)x D(A ),

dn
T (t)x = An T (t)x
dtn

(12.1)

for all x X, t > 0, n N.


3 Recall that analyticity of a map on an open subset of C can be defined either as (complex) differentiability or
by the demand that any point of the domain of definition has a neighborhood in which the function is represented
by a convergent power series. Both definitions are equivalent and can be carried over to maps that take values in
Banach spaces.

24

In applications, analytic semigroups arise from parabolic evolution problems. The property
(12.1) then translates into the smoothing property which is characteristic for these equations.
As we will show, semigroups of this type are generated by so-called sectorial operators (and only
by these). To simplify notation, we define for R, (0, /2) the sector S, := + /2+ .
Definition 12.2. A densely defined linear operator (A, D(A)) on X is called sectorial (or (, )sectorial) if there are R, M, > 0 such that (A) S, and
kR(, A)k

M
,
| |

S, .

(12.2)

S ,

(A)

As A I is (0, )-sectorial if A is (, )-sectorial we can restrict ourselves to the case = 0


without loss of generality.
We introduce the following notation for paths of integration: For r > 0, (0, /2), let
r,

(1)

(2)

(3)

:= r, r, r, ,

(1)

:= {ei(/2) | (, r)},

(2)

:= {rei | (( + /2), + /2)},

(3)

:= {ei(+/2) | (r, )}.

r,

r,

r,

The orientation is with increasing and , respectively.


Let A be a (0, )-sectorial operator, 0 (0, ), (0, 0 ) and r > 0. We construct the
analytic semigroup z 7 T (z) on 0 generated by A by means of a so-called Dunford integral,
given by
Z
1
ez R(, A) d
(12.3)
T (z) :=
2i r,

25

r,

Observe the formal similarity to the complex Laplace inversion formula (10.3). Note, however,
that due to the stronger assumptions on A and and the different integration path, here we will
obtain absolute convergence of the integral, even in L(X). Moreover, the function 7 ez can be
replaced by an arbitrary analytic function 7 f () from a certain class, to obtain a functional
calculus, i.e. a linear map f 7 f (A) that translates products of functions in this class into
compositions of operators in L(X). Here we content ourselves with T (z) = eAz .
Theorem 12.3. Under the assumptions given above, the following holds:
(i) The integral (12.3) exists in L(X) for all z 0 . It is independent of r (0, ) and
(0, 0 ). Moreover, kT (z)k C0 for all z 0 , and the mapping z 7 T (z) is
analytic on 0 .
(ii) We have T (z1 + z2 ) = T (z1 )T (z2 ), z1 , z2 0 .
(iii) For all x X,
lim

0 3z0

T (z)x = x.

(12.4)

(iv) The strongly continuous semigroup t 7 T (t), t [0, ), has generator A.


Remark:
As the theorem holds for all 0 (0, ), T can be extended uniquely and analytically
S
to = 0 < 0 , and the semigroup property is preserved. However, T may be unbounded on
the intersection of any neighborhood of 0 with , and in (12.4), 0 cannot be replaced by .
Therefore, z 7 T (z) is not necessarily an analytic semigroup on in the sense of our definition.
Proof of Theorem 12.3: (i) We will first show the convergence of the integral and the bound
for r = 1/|z| and = (0, 0 ).
(3)
For r, we have
arg + arg z
Re(z)
z

|e |

>

/2 + 0 = /2 + ,

= |z| cos(arg + arg z) |z| cos(/2 + ) = |z| sin ,


e|z| sin ,

and with (12.2)


kez R(, A)k

M |z| sin
e
.
||

Now, parameterizing = ei(/2+) ,


Z
Z
z
i(/2+)
e R(, A) d = e
(3)

r,

26

ez R(, A) d

(12.5)

and the integral converges because from (12.5) we get


kez R(, A)k

M |z| sin
e
.

Additionally, we use r|z| = 1 to get the estimate



Z




z

(3) e R(, A) d

e|z| sin

M
r

r,

=|z|

sin

r|z|

d0
= C0 .
0

(1)

The integral over r, can be discussed analogously.


(2)

For r, we have |z| = 1, |ez | e, kR(, A)k M/r, and therefore


Z


(2)




e R(, A) d 2M e.

z

r,

This proves that the integral (12.3) exists in L(X), and


Z





ez R(, A) d C0

r,

for our choices of and r. However, as the map 7 ez R(, A) is analytic for fixed z, we can
show by deforming the path of integration (and using the exponential decay) that (12.3) is
independent of r (0, ) and (0, 0 ).
The integrand 7 ez R(, A) depends analytically on the parameter z, with derivative 7
z
e R(, A). Similar to what has been shown for (12.3) above, we have absolute convergence of
the integral
Z
ez R(, A) d.
(12.6)
r,

This shows that z 7 T (z) given by (12.3) is complex differentiable and therefore analytic in 0 .
(ii) In the sequel, we fix r and and write := r, . Let c > 0 be such that 0 := + c lies
completely to the right of .

Fix z1 , z2 0 . By shifting the path of integration, applying the resolvent identity, and Fubinis

27

theorem,
T (z1 )T (z2 )

1
(2i)2

Z Z
0

ez1 ez2 R(, A)R(, A) dd


|
{z
}
R(,A)R(,A)

1
(2i)

1
(2i)

"Z

ez2
d

0
{z
}
ez2
#
Z
Z
ez1
1
z2
e R(, A)d
d
(2i) 0

|
{z
}
0
Z
e

z1

1
R(, A) d
(2i)
|

e(z1 +z2 ) R(, A) d = T (z1 + z2 ).

The integrals over the single paths have been calculated by closing the integration contour and
applying the Residue theorem and Cauchys theorem; see sketches:

(The integrand is O(eR ) with some > 0 on the closing arc with radius R.)
(iii) As kT (z)k is bounded uniformly in z 0 , it is sufficient to show the limit relation for
x D(A). In view of
Z z
1
e
d = 1
2i
we have
T (z)x x =

1
2i



Z z
I
1
e
ez R(, A)
x d =
R(, A)Ax d.

2i

Because of the estimate


(
z

e



R(, A)Ax () :=

M
||2 kAxk
M r
||2 e kAxk

(12.7)

(1) (3) ,
(2)

for z 0 , |z| 1 with L1 () we can take the limit in (12.7) by dominated convergence and
obtain
Z
1
R(, A)
lim T (z)x x =
Ax d = 0,
0 3z0
2i

where the last integral is calculated by closing the integration contour by a circular arc of radius
R on the right.
28

(The integrand is O(R2 ) on this arc, while its length is O(R), so that the integral on the arc
vanishes for R . Cauchys theorem yields 0 for the integral over the closed contour.)
(iv) Let B be the generator of the strongly continuous semigroup t 7 T (t), t [0, ). We
will show R(, B) = R(, A) for real > max Re + 1. Fix t0 > 0 and x X. By Fubinis
theorem,
Z t0
Z Z t0
1
t
e T (t)x dt =
e R(, A)x dtd
2i 0
0
Z ()t0
e
1
1
R(, A)x d
=
2i

Z
Z ()t0
R(, A)
e
1
1
=
x d +
R(, A)x d.
2i
2i
|
{z
}
R(, A)x
The first integral can be calculated by closing the contour as in (iii), taking into account the
residuum R(, A) of 7 R(,A)
at = and the clockwise orientation of the contour.

To estimate the second integral we use


()t

0
e

M
t0


C||2 et0 .
R(, A) |||| | e
So this vanishes as t0 , and in this limit we obtain
Z
R(, B)x =
et T (t)x dt = R(, A)x.
0

29

Remark: Sectoriality of an operator may be defined without the demand that D(A) is dense in
X. Then (i) and (ii) remain true while (12.4) holds only for x D(A). Accordingly, A generates
a strongly continuous semigroup of operators on D(A). In this sense, the restriction to densely
defined operators is without loss of generality (upon replacing X by D(A)).
Our next result provides the reverse of the generation result from Theorem 12.3 together with
an equivalent characterization of sectoriality. We define C+ := {z C | Re z > 0}.
Theorem 12.4. Let (A, D(A)) be a densely defined linear operator on a Banach space X, The
following statements are equivalent:
(i) A is a (0, )-sectorial operator for some > 0,
(ii) A generates a strongly continuous semigroup of operators that extends to a bounded analytic
semigroup in a sector 0 for some 0 > 0,
(iii) We have C+ (A) and there is a C > 0 such that
kR(, A)k

kR(, A)k

C
,
Re
C
,
| Im |

C+
C+ \ (0, )

(12.8)
(12.9)

Proof: (i)(ii): This follows (with 0 (0, )) from Theorem 12.3.


(ii)(iii): The inclusion C+ (A) and (12.8) follow from the Hille-Yosida theorem 9.3 and
(9.4). To show (12.9), assume first Im > 0 and fix (, 0), so that ei = a bi, a, b > 0.
As the integrand z 7 exp(z)T (z)x is analytic in 0 and decays exponentially for Re z large in
the part of 0 below the positive real axis we can shift the path of integration in (9.4) to the ray
P := {ei r |, r (0, )} and obtain for x X
Z
Z
z
i
R(, A)x =
e
T (z)x dz = e
exp(rei )T (rei )x dr,
P

and from this the estimate


kR(, A)k

C
C/b
C
=

.
i
Re(e )
a Re + b Im
Im

(iii)(i): We will show the result for (0, arctan(1/C)).

2C
Fix S0, . If Re > 0 then || 2 max(Re , | Im |) and this implies kR(, A)k
.
||
If Re 0 then := i Im 6= 0 and
| | = | Re |

q
||
C

for some q (0, 1). As


C
,
||

kR( + i, A)k
we also have (A) and kR(, A)k

for all > 0,

C
. As
||

dist(, (A))

||
1

kR(, A)k
C

we have (A), and from the Neumann series representation


R(, A) =

( )k R(, A)k+1

k=0

30

(12.10)

and (12.10) we get


kR(, A)k

X
C0
q k ||k C k+1
C 00

.
k
k+1
C ||
||
||

k=0

The following corollary gives another characterization for sectoriality which is easier to apply
in practical applications:
Corollary 12.5. A densely defined operator (A, D(A)) is sectorial if and only if there are constants
M 1, 0 0 such that (A) { 0 } + C+ and
{ 0 } + C+ .

kR(, A)k M,

(12.11)

Proof: : Let A be (, )-sectorial. Choose 0 > max(0, ). Then


||
C1 ,
| |

{ 0 } + C+ ,

and, as these are also in S, ,


C2
M

.
| |
||

kR(, A)k

: Let A satisfy (A) { 0 }+C+ and (12.11). Then, for { 0 }+C+ , 0 (A 0 I)


and
M
kR( 0 , A 0 I)k = kR(, A)k
.
||
Replacing 0 by we conclude (A 0 I) C+ and
kR(, A 0 I)k

M
M

,
0
| + |
||

C+ .

So by Theorem 12.4, A 0 I is (0, )-sectorial for some > 0, and therefore A is ( 0 , )-sectorial.
Recall that T (t)x x as t 0 for x X, but AT (t)x Ax only if x D(A). In general,
kAT (t)xk in this limit, i.e. the time derivative of T (t)x blows up near t = 0. The following
theorem provides a bound for the order of this blowup, while at the same time describing the
decay for large t.
Theorem 12.6. (Asymptotics of analytic semigroups)
Let (A, D(A)) be a (0, )-sectorial operator generating the (bounded, analytic) semigroup t 7
T (t). There is a C > 0 such that
ktAT (t)k C,

t 0.

Proof: Assume t > 0 without loss of generality. Recall the definition (12.3) and the absolute
convergence of this integral and the integral (12.6). Recall further that AR(, A) = R(, A) I
implying the estimate
kAR(, A)k M + 1,
S0, .
This yields
AT (t) =

1
2i

et AR(, A) d

r,

and estimates (cf. the proof of Theorem 12.3 (i))



Z
kAT (t)k C
et sin d + r
r

for r > 0 with C independent of r and t, which imply the result upon taking r 0.
31

Corollary 12.7.

(i) Under the assumptions of Theorem 12.6, there are constants Cn such that
ktn An T (t)k Cn ,

n N+ , t 0.

(ii) Let (A, D(A)) be a (, )-sectorial operator generating the semigroup t 7 T (t), let > 0,
n N+ . There are constants Cn, (independent of t) such that
ktn An T (t)k Cn, e(+)t ,

n N+ , t 0.

Proof: (i) By the semigroup property and commuting,


 n

t
t
n n
n
n
t A T (t) = (tAT (t/n)) = n
AT
n
n
and thus, by Theorem 12.6,
ktn An T (t)k nn C1n := Cn .
(ii) Exercise!

13

Perturbation theorems

Let (A, D(A)) be a generator of a semigroup of operators and (B, D(B)) (its perturbation) a
linear operator on X with D(B) D(A). Perturbation theorems provide sufficient conditions
under which (A + B, D(A)) (the perturbed operator) also generates a semigroup of operators,
and describe the properties of this semigroup. We give two simple, important results, one for
strongly continuous semigroup and another for analytic semigroups.
The following preliminary lemma deals with perturbations of the resolvent and is guided by
the algebraic identity
1
1
1
.
=
b
(a + b)
a 1 a
Lemma 13.1. Let (A, D(A)), (B, D(B)) be two linear operators on a Banach space X with
D(B) D(A). Let (A). If BR(, A) L(X) and
kBR(, A)k < 1
then (A + B) and
kR(, A + B)k

1
R(, A).
1 kBR(, A)k

Proof: Recall that I BR(, A) is invertible and


k(I BR(, A))1 k

1
.
1 kBR(, A)k

Define
R := R(.A)(I BR(, A))1 .
It remains to show that R = R(, A + B). Indeed, we have
(I A B)R

(I A)R BR

(I BR(, A))1 BR(.A)(I BR(, A))1 = I

32

and for x D(A), using the Neumann series representation of (I BR(, A))1 ,
R(I A B)x

k=0

R(, A)[BR(, A)]k (I A)x


R(, A)[BR(, A)]k Bx

k=0

[R(, A)B]k x

k=0

[R(, A)B]k+1 x = x.

k=0

Theorem 13.2. (Perturbation of generators of strongly continuous semigroups)


Let (A, D(A) be the generator of the strongly continuous semigroup of operators t 7 T (t) on
X, satisfying kT (t)k M et . Let B L(X). Then (A + B, D(A)) also generates a strongly
continuous semigroup of operators t 7 S(t) on X, satisfying
kS(t)k M e(+M kBk)t .
Proof: Introducing the equivalent norm ||| ||| on X by
|||x||| := sup kes T (s)xk,

x X,

s0

we find that t 7 T (t) is a semigroup of quasicontractions with respect to this norm, i.e.
|||T (t)||| et ,

t 0,

cf. the proof of Theorem 9.3. Consequently, by the Hille-Yosida theorem for quasicontractions, in
the corresponding operator norm we have (, ) (A) and
|||R(, A)|||

1
,

> .

Fix > + |||B|||. Then (A), and |||BR(, A)||| < 1 because of |||B||| < . So Lemma
13.1 (in the space (X, ||| |||)) yields (A + B), and
|||R(, A + B)|||

1
1

|||B|||

1
1
=
.

( + |||B|||)

Consequently, by the Hille-Yosida theorem 7.2, A + B generates a strongly continuous semigroup


t 7 S(t), and
|||S(t)||| e(+|||B|||)t .
The result follows by translating this back to the original norms.
Theorem 13.3. (Perturbation of generators of analytic semigroups)
Let (A, D(A)) be a sectorial operator. There is an > 0 with the following property: If a linear
operator (B, D(B)) satisfies D(B) D(A) and
kBxk kAxk + CB kxk,

x D(A),

(13.1)

for some CB > 0, then (A + B, D(A)) is sectorial.


Remarks: Note that depends only on A while CB may depend on B. In practical applications, the value of for a concrete sectorial operator A is usually not explicitly known. To
apply Theorem 13.3 to concrete operators A and B one usually shows that for all > 0 there is a
constant C such that
kBxk kAxk + C kxk,
x D(A).
33

This is often done by means of so-called interpolation inequalities. The situation of the theorem
occurs, for example, if A is a suitable differential operator and B is a differential operator of order
lower than A. (See exercises.)
Proof of Theorem 13.3: By Corollary 12.5, there are M 1, r 0 such that (A) {r} + C+
and
M
kR(, A)
,
{r} + C+ .
||
We will show that := 1/(4(M + 1)) has the property announced in the theorem.
Indeed, let an operator B satisfy the assumptions of that property. Let r0 := max(r, 4CB M ).
Then, for {r0 } + C+ , x X, we have (A) and
kBR(, A)xk

k AR(, A)x k + CB kR(, A)xk


| {z }
R(,A)xx

CB M
(M + 1) +
||


kxk

1
kxk.
2

Therefore BR(, A) L(X) and kBR(, A)k 12 . Now Lemma 13.1 yields (A + B) and
kR(, A + B)k 2R(, A)

2M
.
||

Therefore, Corollary 12.5 implies that A + B is sectorial.

14

Homogeneous Cauchy problems

Now we turn to the investigation of (abstract) evolution equations, which is the main aim of the
theory of operator semigroups. The simplest equation of this type is the so-called homogeneous
Cauchy problem which has the following informal formulation: Let (A, D(A)) be a linear operator on a Banach space X and x X. We are looking for a function u on the (time) interval
[0, ] such that

u = Au,
(HCP)
u(0) = x.
To make this precise, we introduce a hierarchy of solution concepts to (HCP).
Definition 14.1.
A function u C 1 ([0, ], X) is called strict solution to (HCP) if u(0) =
x and for all t [0, ] we have u(t) D(A) and u(t)

= Au(t).
A function u C([0, ], X)C 1 ((0, ], X) is called classical solution to (HCP) if u(0) = x
and for all t (0, ] we have u(t) D(A) and u(t)

= Au(t).
A function u C([0, ], X) is called mild solution to (HCP) if u(0) = x and for all t [0, ]
Rt
we have 0 u(s) ds D(A) and
Z t
u(t) = x + A
u(s) ds,
t [0, ].
0

We clearly have (for a given problem (HCP))


u strict solution

u classical solution

u mild solution

whenever A is closed. Moreover, if u is a mild solution to (HCP) then


Z
1 t
x = u(0) = lim
u(s) ds D(A),
t0 t 0
|
{z
}
D(A)
34

(14.1)

but for the sake of generality we will not assume in this section that A is densely defined.
In the case that A generates a strongly continuous semigroup of operators we immediately
have, from previous results,

strict solution for x D(A),
u(t) := T (t)x is a
mild solution for x X.
If A generates an analytic semigroup then
u(t) := T (t)x is a classical solution for x X.
These solutions are unique, as the following results shows:
Lemma 14.2. (Uniqueness of solutions to (HCP))
Assume that A generates a strongly continuous semigroup of operators, and u is a mild solution
to (HCP) with x = 0. Then u 0.
Proof: Exercise!
Note, however, that existence and uniqueness of (even strict) solutions to (HCP) for all x
D(A) does not imply that A generates a strongly continuous semigroup of operators, even if A is
closed. (See the exercises for a counterexample.) Before we can clarify the situation completely
we need the following lemma:
Lemma 14.3. (Cores for A)
Let (A, D(A)) be the generator of the strongly continuous semigroup t 7 T (t). Let F be a
dense subspace of D(A) which satisfies T (t)F F for all t 0. Then F is dense in D(A) even
with respect to the graph norm k kD(A) .
Any subspace F with these properties is called a core for A.
Proof of Lemma 14.3: Fix x D(A) and > 0. As the map s 7 T (s)x is continuous with
respect to k kD(A) , we have






Z t

1

x
<
T
(s)x
ds


t
2

| 0 {z
}


=: z
D(A)
for t > 0 sufficiently small. Fix such a t. Denote by Fb the closure of F with respect to D(A).
X
Choose a sequence (xn ) in F such that xn x. As s 7 T (s)xn is continuous with respect to
k kD(A) and T (s)xn F for s [0, t] we have (as the integral is a limit of Riemann sums)
Z
1 t
T (s)xn ds Fb.
zn :=
t 0
Now
kzn zkD(A)

= kzn zkX + kAzn AzkX



Z t
1


T
(s)(x

x)
ds
=
n
t

0








1 Z t


T (s)(xn x) ds
+ A
t

| 0
{z
}


T (t)(xx )(xx )
n

kx xn kX <
t
2

for n large, hence for such n we have kx zn kD(A) < . So Fb and consequently F are dense in
D(A) with respect to k kD(A) .
We will say that a densely defined linear operator A satisfies property (EU) if for all x D(A),
(HCP) has a unique strict solution. We will denote this solution by u(, x).
35

Theorem 14.4. (Homogeneous Cauchy problems and strongly continuous semigroups)


Let (A, D(A)) be a closed linear operator on X. The following conditions are equivalent:
(i) A is generator of a strongly continuous semigroup of operators t 7 T (t),
(ii) A satisfies property (EU) and (A) 6= .
X

(iii) A satisfies property (EU), is densely defined, and for any sequence (xn ) in D(A) with xn 0
we have u(, xn ) 0 in C[[0, ], X) for all > 0.
Observe that condition (iii) can be interpreted as a well-posedness statement for (HCP) as it
provides, additionally to existence and uniqueness of the solution, continuity of the map x 7 u(, x)
in a precise sense.
Proof of Theorem 14.4: (i)(ii): follows directly from previous results on semigroups.
(ii)(iii): The proof proceeds by constructing mild solutions v(, x) for any x X, showing that
these are unique, and proving the continuous dependence on x for these. By uniqueness, for
x D(A), these solutions are the strict solutions u(, x). By (14.1) we also get D(A) dense in X
once the existence of a weak solution for any x X is established.
I. Construction of the mild solution: Fix (A). For x X, let y := R(, A) D(A) and
u := u(, y) the corresponding strict solution. Then v := v(, x) := (I A)u(, y) is a mild solution
to (HCP). Indeed, v(0) = x, v = u Au = u u C([0, ], X), and, using the closedness of
A, we get from
Z
t

u(t) = y + A

u(s) ds
0

by applying (I A) (which is closed as well)


t

Z
v(t) = x + A

v(s) ds.
0

II. Uniqueness of the mild solution: Let v = v(, 0) be a mild solution to (HCP) with x = 0.
Rt
This means that for all t [0, ], the integral 0 v(s) ds is in D(A) and
t

Z
v(t) = A

v(s) ds.
0

Therefore, u given by
Z

u(t) =

v(s) ds
0

is a strong solution to (HCP) with x = 0 and hence, by assumption, u 0. This implies in turn
v 0.
III. Continuous dependence on the initial value: Define the linear map : X C([0, ], X)
by x = v(, x), the mild solution to (HCP) with v(0) = x. To show that is closed, fix x X,
X
assume xn x, and xn y in C([0, ], X). By integration
Z t
Z t
X
v(s, xn ) ds
y(s) ds,
t [0, ],
0

and, as the v(, xn ) are mild solutions,


Z t
X
A
v(s, xn ) ds = v(t, xn ) x y(t) x.
0

As A is closed, this implies


Z t
y(s) ds D(A)

Z
and

36

y(s) ds = y(t) x,

i.e. y is the mild solution with y(0) = x, so y = x. This shows that is closed, and continuity
follows by the Closed Graph theorem (Corollary 2.7).
(iii)(i): Define T (t)x := u(t, x) for x D(A). It follows directly from our assumptions that
T (t) L(D(A), X) for all t 0, and T (t) extends by density to an operator T (t) L(X).
Moreover, there is a constant M such that kT (t)k M for t [0, 1]. (Otherwise, there would be
X
sequences (tn ) in [0, 1] and (xn ) in D(A) such that xn 0 and
kT (tn )xn k = ku(tn , xn )k 1,
in contradiction to the assumption u(, xn ) 0 uniformly in [0, 1].)
Uniqueness yields, for t, s 0, x D(A),
T (t + s)x = u(t + s, x) = u(t, u(s, x)) = T (t)T (s)x,
and this extends by continuity to all x X. Moreover for x in the dense set D(A) we have
continuity of the map t 7 T (t)x. By Lemma 6.1, this implies that t 7 T (t) is a strongly
continuous semigroup of operators. Let (B, D(B)) be its generator. We clearly have A B. As
D(A) is dense in X, it is also dense in the smaller set D(B), and because T (t)D(A) D(A),
D(A) is a core for B. Now Lemma 14.3 implies that the set {(x, Ax) | x D(A)} is dense in
{(x, Bx) | x D(B)}, and both are closed in X X. Thus A = B.

15

Intermezzo: more general uniqueness

Lemma 14.2 states that mild solutions of (HCP) are unique, if the operator A generates a strongly
continuous semigroup. The same can be proved under weaker conditions on A, if we consider
strict solutions:
Lemma 15.1. Let (A, D(A)) be dissipative. There exists at most one classical solution of (HCP).
This lemma does not require A to be a generator of a strongly continuous semigrouponly to
be dissipative.
Proof: In order to prove uniqueness, it is sufficient to prove that any curve x C([0, ]; X)
C 1 ((0, ]; X) with x = Ax and x(0) = 0 satisfies x 0. We will do this by showing that
the function f (t) := kx(t)kX is non-increasing. Fix t > 0 and choose x F (x(t)) such that
Rehx , Ax(t)i 0 (this is possible since A is dissipativesee Definition 8.2). Set y = x /kx kX ;
since x F (x(t)), we have
kx(t)kX = hy , x(t)i

and X :

kkX hy , i.

Then
lim sup
h0



1
1
f (t) f (t h) = Re lim sup f (t) f (t h)
h
h
h0

1
= Re lim sup kx(t)kX kx(t h)kX
h
h0
1
Re lim sup hy , x(t) x(t h)i
h
h0
= Re hy , x(t)i

0.

The following lemma then implies that f is decreasing in t.


Lemma 15.2. Let f : [0, 1] R be continuous, and assume that f satisfies
t (0, 1] :

lim sup
h0


1
f (t) f (t h) 0.
h

Then f is non-increasing.
37

Proof: Assume that f is not non-increasing; then there exist 0 a < b 1 such that
` :=

f (b) f (a)
> 0.
ba

The continuous function g : [a, b] R, g(t) := f (t) f (a) `(t a) assumes its maximum over
[a, b] at some t [a, b]; since g(a) = g(b) = 0, we can assume that t (a, b].
From the maximality of g at t we then deduce that
0 lim sup
h0



1
1
g(t) g(t h) = lim sup f (t) f (t h) ` ` < 0,
h
h
h0

a contradiction.

16

Inhomogeneous Cauchy problems

Let X be a Banach space and A the generator of the strongly continuous T 7 T (t). Let f :
[0, ] X be a function of which we will demand at least local integrability, further demands will
depend on the situation. The (abstract) inhomogeneous Cauchy problem can be informally
stated as

u = Au + f,
(ICP)
u(0) = x.
The following definitions are essentially analogous to the homogeneous case (we exlude the right
boundary point for technical reasons):
Definition 16.1.
A function u C 1 ([0, ), X) is called strict solution to (ICP) if u(0) = x
and for all t [0, ] we have u(t) D(A) and u(t)

= Au(t) + f (t).
A function u C([0, ), X) C 1 ((0, ), X) is called classical solution to (ICP) if u(0) = x
and for all t (0, ] we have u(t) D(A) and u(t)

= Au(t) + f (t).
Again, strict solutions are classical solutions.
The definition of a mild solution is based on the following important representation formula:
Lemma 16.2. (Variation of constants formula, Duhamels principle)
Assume f L1 ((0, ), X). 4 Let u be a classical solution to (ICP). Then
Z t
T (t s)f (s) ds,
t [0, ).
u(t) = T (t)x +
0

Proof: The representation formula obviously holds for t = 0. Now fix t (0, ] and define
g C([0, t], X) C 1 ((0, t), X) by
g(s) := T (t s)u(s),

s [0, t].

Then
g(s)

= AT (t s)u(s) + T (t s)u(s)

= AT (t s)u(s) + T (t s)(Au(s) + f (s)) = T (t s)f (s),

and by integration
Z
u(t) T (t)x = g(t) g(0) =

T (t s)f (s) ds.


0

4 Readers

unfamiliar with this space can replace this assumption by the demand that f should be regulated on
R
(0, ), and kf kL1 ((0,),X) := 0 kf (s)kX ds should be finite.

38

Corollary 16.3. Classical solutions to (ICP) with f L1 ((0, ), X) are unique.


Observe, however, that Lemma 16.2 does not provide existence of classical solutions. Rather,
it gives rise to a generalized solution concept:
Definition 16.4. For x X, f L1 ((0, ), X), the function u C([0, ], X) given by
Z t
u(t) = T (t)x +
T (t s)f (s) ds,
t [0, ]

(MS)

is called the mild solution to (ICP).


In view of our results on homogeneous Cauchy problems, this definition is consistent with the
earlier one in the case f 0.
Lemma 16.2 shows that (MS) provides the only candidate for a classical solution. A counterexample shows that even for x = 0, f C([0, ], X) is not sufficient for (MS) to provide a
classical solution (see exercise).
The next result provides criteria (in terms of the inhomogeneous part of (MS)) under which
(MS) actually is a classical solution.
Lemma 16.5. (Conditions for classical solutions)
Assume f L1 ((0, ), X) C((0, ], X) and define v C([0, ), X) by
Z t
v(t) =
T (t s)f (s) ds,
t [0, ].

(16.1)

The following statements are equivalent:


(i) v C 1 ((0, ), X),
(ii) v(t) D(A) for all t (0, ) and Av() C((0, ), X),
(iii) (MS) is the classical solution to (IVP) for some x D(A),
(iv) (MS) is the classical solution to (IVP) for all x D(A).
Proof: (iii)(i): Fix x D(A) such that (MS) is the classical solution to (IVP). Note that
then both t 7 T (t)x and u are in C 1 ((0, ), X). Statement (i) follows from v(t) = u(t) T (t)x.
(iii)(ii): With x D(A) as in the previous part of the proof, t (0, ), we have v(t) =
u(t) T (t)x D(A) and
Av(t) = u(t)
f (t) T (t)Ax
for t (0, ]. Hence Av() C((0, ], X].
(i)(iv): Let t (0, ). For small positive h we have (check!)
Z
T (h) I
v(t + h) v(t)
1 t+h
v(t) =

T (t + h s)f (s) ds.


h
h
h t

(16.2)

Taking h 0 (first on the right) yields v(t) D(A), Av(t) = v(t)


f (t). From this and v(0) = 0
we directly conclude that for any x D(A), the classical solution u is given by u(t) = T (t)x + v(t).
(ii)(iv): Let t (0, ). From v(t) D(A) we conclude by (16.2) that v is right differentiable in
t, with
d+
v(t) = Av(t) + f (t),
t (0, ),
dt
and, as the right derivative is continuous, v is differentiable on (0, ) (see exercise). From this
and v(0) = 0 we again conclude that for any x D(A), the classical solution u is given by
u(t) = T (t)x + v(t).
(iv)(iii): trivial.
Remark: An analogous theorem holds for strict solutions when the open intervals (0, ) are
replaced by half-open intervals [0, ). (Check!)
39

Theorem 16.6. Let f C 1 ([0, ], X), x D(A). Then (ICP) has a unique strict solution given
by (MS).
Proof: In the notation of Lemma 16.5,
Z t
Z t
v(t) =
T (t s)f (s) ds =
T (s)f (t s) ds,
0

so v C 1 ([0, ], X), and by this lemma and the remark, (MS) yields a strict solution for all
x D(A).
Due to the smoothing effect of analytic semigroups, the situation is better in that case.
Recall that a function f : [0, ] X is called H
older continuous with exponent (0, 1) if
there is a constant C > 0 such that
kf (t) f (s)k C|t s| ,

t, s [0, ].

We will write C ([0, ], X) for the set of these functions.


Theorem 16.7. Let (0, 1), and let A be sectorial. Let f C ([0, ], X), x X. Then (ICP)
has a unique classical solution given by (MS).
Proof: As t 7 T (t)x provides a classical solution to (HCP), we can restrict ourselves to (ICP)
with x = 0 and are going to apply Lemma 16.5. For t (0, ) we define
Z t
Z t
Z t
v(t) :=
T (t s)f (s) ds =
T (t s)(f (s) f (t)) ds +
T (t s)f (t) ds
0
0
0
{z
} |
{z
}
|
=: v1 (t)
=: v2 (t)
and will show that for i = 1, 2, vi (t) D(A) and t 7 Avi (t) continuous. Then the result will
follow from Lemma 16.5. As
Z t
v2 (t) =
T (s)f (t) ds
0

we immediately have v2 (t) D(A) and Av2 (t) = T (t)f (t) f (t) which is continuous in t.
To get the same results for v1 , we have to discuss interchanging A and an improper integral.
For > 0, define
Z t
Z t
v1, (t) =
T (t s)(f (s) f (t)) ds =
T (s)(f (t s) f (t)) ds.
0

As t 7 T (t) is an analytical semigroup, the map s 7 AT (s)(f (t s) f (t)) is continuous on [t, ].


Consequently, this function is integrable there, and, by closedness of A,
Z t
Av1, (t) =
AT (s)(f (t s) f (t)) ds.

By Theorem 12.6 and the H


older continuity of f ,
kAT (s)(f (t s) f (t))k Cs1

(16.3)

with C independent of s, t (0, ]. Therefore, by dominated convergence,


Z t
Av1, (t)
AT (s)(f (t s) f (t)) ds
0

as 0. Using v1, (t) v1 (t) as 0 and the closedness of A we conclude v1 (t) D(A) and
Z t
Av1 (t) =
AT (s)(f (t s) f (t)) ds.
0

40

Finally, (16.3) implies that for any > 0 there is a > 0 such that

Z




t (0, ].
AT (s)(f (t s) f (t)) ds < ,


0
Now from
Z

Z
AT (s)(f (t s) f (t)) ds +

Av1 (t) =
0

AT (s)(f (t s) f (t)) ds

and the continuity of the second term in t, it follows that t 7 Av1 (t) is continuous (check!).
Our next result will be important when we discuss semilinear parabolic problems and concerns
H
older continuity of the mild solution to (ICP) if f is only continuous.
Lemma 16.8. Let f C([0, T ], X), (0, 1). Then v given by (16.1) is H
older continuous with
exponent on [0, ].
Proof: Pick s, t [0, ], assume without loss of generality s < t. Then
Z s
Z t
v(t) v(s) =
(T (t ) T (s ))f () d +
T (t )f () d
0
s
Z s Z t
Z t
=
AT ( )f () d d +
T (t )f () d.
0

We estimate the integrals separately. Using standard bounds for T (t) we immediately get

Z t


Ckf k (t s) C 0 kf k (t s) .

T
(t

)f
()
d


s

(Check!)
For the first term we find by Theorem 12.6

Z s Z t




AT
(
)f
()
d
d


0
s
Z s Z t
Z s
Z
1

Ckf k
d d Ckf k
(s )
0

s
s

(s )

Ckf k

1 d d

ts

( + s )1 d d

Z ts
(s )
1 d d
0
0
Z s
1

Ckf k (t s)
(s ) d
Z

Ckf k

Ckf k 1 (1 )1 1 (t s) .
This completes the proof.

17

A simple class of semilinear Cauchy problems

Let (A, D(A)) generate a strongly continuous semigroup of operators t 7 T (t) on the Banach
space X. Fix > 0, and let
F : [0, ] X X
be a continuous function for which we additionally demand that it is Lipschitz continuous on
bounded sets in the second argument, i.e.:
41

(A1) For all R > 0 there is an LR > 0 such that


kF (t, x) F (t, y)k LR kx yk,

x, y BX (0, R), t [0, ].

Under these assumptions we consider for x X the (abstract) semilinear Cauchy problem

u(t)

= Au(t) + F (u(t), t),


(SCP)
u(0) = x.
The concepts of strong solution and classical solution are defined as in the linear case. A
function u C([0, ], X) is called a mild solution of (SCP) if
Z

T (t s)F (s, u(s)) ds.

u(t) = T (t)x +
0

(Note that this is this is a nonlinear integral equation defining u implicitly, so existence and
uniqueness are nontrivial issues here!) Again, it is straightforward to see that strong solutions are
classical solutions, and classical solutions are mild solutions.
Theorem 17.1. (Short-time well-posedness of (SCP) for mild solutions)
Let F satisfy (A1). For any x
X, there are r, , K > 0 such that
for all x B(
x, r), (SCP) has a unique mild solution u = u(, x) on the interval [0, ],
we have
ku(t, x1 ) u(t, x2 )k Kkx1 x2 k

x1 , x2 B(
x, r), t [0, ].

(17.1)

Observe that (17.1) provides, in a certain sense, continuous dependence of the mild solution
on the initial data.
Proof: 1. Let M := supt[0,] kT (t)k, R 8M k
xk, r := R/(8M ),
(

1
R
:= min ,
,
2M LR 4M sups[0,] kF (s, 0)k + 1

)
.

Define
Y := YR := {v C([0, ], X) | sup kv(t)k R}.
t[0,]

As Y is a closed ball in the Banach space C([0, ], X) (with respect t o the supremum norm which we
will denote by k k ), Y is a complete metric space. Define further the map : Y C([0, ], X)
by
Z
t

T (t s)F (s, v(s)) ds,

(v)(t) = T (t)x +

t [0, ].

Clearly, a function u Y is a mild solution to (SCP) on [0, ] if and only if u = u. We will show
that [Y ] Y and that is ( 21 -)contracting, hence the existence of a mild solution will follow
from the Banach fixed point theorem.
Indeed, because for v1,2 Y we have
Z
v1 (t) v2 (t) =

T (t s)(F (s, v1 (s)) F (s, v2 (s))) ds M LR kv1 v2 k ,


0

our assumptions on imply


kv1 v2 k

42

1
kv1 v2 k .
2

Furthermore, for v Y we have, due to the contraction property,


kvk

kx 0k + k0k

1
2 kvk

+ M (k
xk + r) + M sup kF (s, 0)k
s[0,]

R R R
+ +
=R
2
4
4

and therefore v Y .
In particular, due to the uniqueness of the fixed point in Y , we have the following intermediate
uniqueness result: For any sufficiently large R, there is a = (R, x) > 0 such that among all
functions in C([0, ], X) that satisfy kuk R, there is precisely one mild solution to (SCP).
2. Fix x X and let u1,2 C([0, ], X) be two mild solutions to (SCP). Let
t0 := sup({0} {t (0, ] | u1 |[0,t] = u2 |[0,t] }).
We will show by contradiction that t0 = . Assume t0 < . Define u
1,2 C([0, t0 ], X) by
u
i (t) = ui (t + t0 ), t [0, t0 ]. Then u
1,2 are mild solutions to the semilinear Cauchy problem

u
(t) = A
u(t) + F (t + t0 , u
(t))
u
(0) = u(t0 )
on [0, t0 ] (exercise!). Choose R max{ku1 k , ku2 k , 8M kxk}. By the uniqueness result of
Step 1, we find that u
1 and u
2 coincide on some interval [0, 0 ] of positive length, contradicting
the definition of t0 .
3. Fix x1,2 B(
x, r) and write ui := u(, xi ) for the corresponding mild solutions of (SCP).
Then
Z
t

u1 (t) u2 (t) = T (t)(x1 x2 ) +

T (t s)(F (s, u1 (s)) F (s, u2 (s))) ds.


0

and, because of kui k R,


ku1 u2 k M kx1 x2 k + M LR ku1 u2 k ,
| {z }
12
and so
ku1 u2 k 2M kx1 x2 k.
Using arguments similar to the ones given in Step 2 of the proof, one can also show existence
and uniqueness of a maximal solution to (SCP), i.e. that there is a mild solution u
on an interval
I [0, ] such that for any mild solution u on some interval I [0, ] we have I I and u = u
|I .
(Exercise!)
To discuss regularity of the mild solutions, we additionally assume
(A2) A is sectorial, i.e. the semigroup t 7 T (t) is analytic,
(A3) F is H
older continuous with respect to t, more precisely, there is an (0, 1) such that for
all R > 0 there is a constant CR such that
kF (t, x) F (s, x)k CR |t s| ,

x B(0, R).

Theorem 17.2. (Classical solutions to (SCP) in the parabolic case)


Let (A1)(A3) be satisfied, x X. Let u be the mild solution to (SCP) on some interval [0, ]
Proof: Observe first thst the map s 7 f (s) := F (s, u(s)) is continuous, so u is the mild solution
the linear problem (ICP) with that function f , and Lemma (16.5) applies. We will show that
u C 1 (, ) for all > 0. By the analyticity of t 7 T (t)x and Lemma 16.8 we find that
43

u C ([, ], X). Define now u


, f by u
:= u(t + ), f(t) := F (t + , u
(t)), t [0, ], and
observe that u
is the mild solution to
u
= A
u + f,

u
(0) = u().

(17.2)

(Check, see exercise.) As f C ([0, ], X) because of (A1), (A3) and the Holder continuity of
u (exercise!) we find by Theorem 16.7 applied to (17.2) that u
C 1 ([0, ], X) which implies
our claim.

18

Global existence and stability: a toy example

Instead of giving general results in an abstract context, we want to demonstrate the application of
the theory of (analytic) semigroups to the stability analysis of an equibrium of a simple semilinear
parabolic initial-boundary value problem in one space variable.
We are seeking functions u = u(x, t), x [0, 1], t 0 such that

ut (x, t) = uxx (x, t) + u(x, t)2 x [0, 1], t 0,


u(0, t) = u(1, t) = 0
t0
(18.1)

u(x, 0) = u0 (x)
x [0, 1],
where u0 is a given continuous function that satisfies the compatibility conditions u0 (0) = u0 (1) =
0.
Obviously u 0 is a trivial stationary solution. We are going to show its stability, i.e.we we
will show that solutions with initial value near zero exist for all positive times and decay towards
zero for large times.
Proposition 18.1. For any (0, 2) there is an > 0 such that (18.1) has a unique classical
solution on (0, ). It satisfies the estimate
sup |u(x, t)| = ku(, t)k 2et .

(18.2)

Proof: We choose the following setting: Let


X = {v C([0, 1]) | v(0) = v(1) = 0},
with the supremum norm k k, let (A, D(A)) be given by
D(A) = {v C 2 ([0, 1]) | v(0) = v(1) = 0},

Av = vxx .

Then (A, D(A)) generates a strongly continuous analytic semigroup of operators t 7 T (t) on X. Its
2
spectrum consists of negative real eigenvalues, the largest being 2 . Moreover, kT (t)k e t .
(Exercises!) Further, let F : [(0, ) X X be given by
F (t, v)(x) = v(x)2 .
It is not hard to check that F satisfies (A1) and (A3). With these choices for X, A, and F , and
x := u0 X, Problem (18.1) belongs tothe class of problems ginven by (SCP) (on an infinite time
iterval).
(Danger: The restriction to one space dimension makes it possible to choose particularly
simple function spaces. To treat the spatially multidimensional analogon to (18.1), different spaces
have to be chosen.)
In the sequel, we identify continuous functions u : [0, 1] [0, ) R that satisfy (18.1)2 with
X-valued functions t 7 u(, t) =: u(t) X without changing the notation.
For (0, 2 ), > 0 we define
M, := {v C([0, ), X) | t 7 et v(t) bounded , kvk 2}
44

where
kvk := sup ket v(t)k.
t0

Note that M, , with the metric induced by k k is a complete metric space of continuous
functions that satisfy the decay estimate (18.2). In view of Theorems 17.1, 17.2, the proposition
is established once we find a mild solution to (18.1) on [0, ) in M, , i.e. a fixed point of the
operator
: M, C([0, ), X)
given by
Z
(v)(t) := T (t)u0 +

T (t s)v(s)2 ds.

To apply Banachs fixed point theorem, we have to show that for sufficiently small, is a
1
2 -contraction and [M, ] M, . Indeed, for v, w M, , t 0, < /8, we have

t
e (v(t) w(t)

kT (t s)kkkv(s)2 w(s)2 k ds

(ts)

kv(s) + w(s)k kv(s) w(s)k ds


|
{z
}
s
4e

( )(ts)
w(s)k ds
{z
} kv(s)
|e
|
{z
}
1
kv wk es

4
0

Z
4

es ds kv wk 21 kv wk ,
| {z }
1
0

and so, taking the supremum over t,


kv wk 21 kv wk .
Moreover, for v M, ,
kvk

kv 0k + k0k 12 kvk + sup et kT (t)u0 k


t

+ sup e

()t

ku0 k = 2.

45

References
[Bre11]

H. Brezis. Functional Analysis, Sobolev Spaces and Partial Differential Equations.


Springer, New York, 2011.

[EN00]

K.-J. Engel and R. Nagel. One-parameter semigroups for linear evolution equations.
Springer, 2000.

[EN06]

K.-J. Engel and R. Nagel. A short course on operator semigroups. Springer, 2006.

[Eva02]

L.C. Evans. Partial Differential Equations, volume 19 of Graduate Studies in Mathematics. American Mathematical Society, first edition, 2002.

[Kre89]

E. Kreyszig. Introductory Functional Analysis with Applications, volume 130. John


Wiley & Sons New York, 1989.

[LLMP04] L. Lorenzi, A. Lunardi, G. Metafune, and D. Pallara. Analytic semigroups and reactiondiffusion problems. In Internet Seminar, 2004.
[Lun12]

A. Lunardi. Analytic semigroups and optimal regularity in parabolic problems. Springer,


2012.

[Paz83]

A. Pazy. Semigroups of linear operators and applications to partial differential equations, volume 44. Springer New York, 1983.

[RR04]

M. Renardy and R. C. Rogers. An Introduction to Partial Differential Equations,


volume 13. Springer Science & Business Media, 2004.

[Sch]

B. Schweizer. Evolutionsgleichungen (german lecture notes, 2009/10).

46

Anda mungkin juga menyukai