Lecture Notes Spectral Theory: Roland Schnaubelt

Download as pdf or txt
Download as pdf or txt
You are on page 1of 90

Lecture Notes

Spectral Theory

Roland Schnaubelt

These lecture notes are based on my course from the summer semester 2010. I
kept the numbering and the contents of the results presented in the lectures (except
for some minor corrections). Typically, the proofs and calculations in the notes are
a bit shorter than those given in the lecture. Moreover, the drawings and many
additional, mostly oral remarks from the lectures are omitted here. On the other
hand, I have added a few lengthy proofs not shown in the lectures.
I like to thank Bernhard Konrad for his support in preparing these notes and
Martin Meyries for pointing out a few mistakes and gaps in Section 4.3.
Karlsruhe, 6. August 2012

Roland Schnaubelt

Contents
1 General spectral theory

1.1

Closed operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

The spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Spectral theory of compact operators

17

2.1

Compact operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.2

The Fredholm alternative . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Sobolev spaces and weak derivatives

29

3.1

Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2

Density and embedding theorems . . . . . . . . . . . . . . . . . . . . . 39

4 Selfadjoint operators

53

4.1

Closed operators and their spectra, revisited . . . . . . . . . . . . . . . 53

4.2

Selfadjoint operators and their spectra . . . . . . . . . . . . . . . . . . 60

4.3

The spectral theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5 Holomorphic functional calculus

76

5.1

The bounded case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5.2

Sectorial operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Bibliography

89

Chapter 1

General spectral theory


General notation. X 6= {0}, Y 6= {0} and Z 6= {0} are Banach spaces over C with
norms kk (or kkX etc.). The space
L(X, Y ) = {T : X Y : T is linear and continuous}
is endowed with the operator norm kT k = supkxk1 kT xk, and we abbreviate L(X) :=
L(X, X).
Let D(A) be a linear subspace of X and A : D(A) Y be linear. Then A, or
(A, D(A)), is called linear operator from X to Y (and on X if X = Y ) with domain
D(A). We denote by
N (A) = {x D(A) : Ax = 0},
and

R(A) = {y Y : there exists some x D(A) with y = Ax}

the kernel and range of A.

1.1

Closed operators

Example 1.1. Let X = C([0, 1]) be endowed with the supremum norm and let
Af = f 0 with D(A) = C 1 ([0, 1]). Then A is linear, but not bounded. Indeed,

consider the functions un D(A) given by un (x) = (1/ n) sin(nx) for n N, which
satisfy kun k 0 and

kAun k |u0n (0)| = n as n .


However, if fn D(A) satisfy fn f and Afn g in C([0, 1]) as n , then
f D(A) = C 1 ([0, 1]) and Af = g (see Analysis 1 & 2).
Definition 1.2. Let A be a linear operator from X to Y . The operator A is called
closed if for all xn D(A), n N, such that there exists x = limn xn in X and
y = limn Axn in Y it holds that x D(A) and Ax = y.
Hence, limn (Axn ) = A(limn xn ) if both (xn ) and (Axn ) converge.

CHAPTER 1. GENERAL SPECTRAL THEORY

Remark. It is clear that every operator A L(X, Y ) is closed (with D(A) = X).
The operator A from Example 1.1 is closed.
Example 1.3. a) Let X = C([0, 1]) and Af = f 0 with
D(A) = {f C 1 ([0, 1]) : f (0) = 0}.
Let fn D(A) and f, g X be such that fn f and Afn = fn0 g in X as
n . Again by Analysis 1 & 2, there exists f C 1 ([0, 1]) such that f 0 = g. Since
0 = fn (0) f (0) as n , we obtain f D(A). This means that A is closed on
X. In the same way we see that A1 f = f 0 with
D(A1 ) = {f C 1 ([0, 1]) : f (0) = f 0 (0) = 0}
is closed. There are many more variants.
b) Let X = C([0, 1]) and Af = f 0 with
D(A) = Cc1 ((0, 1]) = {f C 1 ([0, 1]) : supp f (0, 1]},
where the support supp f of f is the closure of {t [0, 1] : f (t) 6= 0} in R.
This operator is not closed. In fact, consider the functions fn D(A) given by
(
0,
0 t 1/n,
fn (t) =
2
(t 1/n) , 1/n t 1,
for every n N. Then, fn f and fn0 f 0 in X as n , where f (t) = t2 .
However, supp f = [0, 1] and so f
6 D(A).
c) Let X = Lp (Rd ), 1 p , and m : Rd C be measurable. Define Af = mf
with
D(A) = {f X : mf X}.
This is the maximal domain. Then A is closed. Indeed, let fn f and Afn =
mfn g in X as n . Then there is a subsequence such that fnj (x) f (x) and
m(x)fnj (x) g(x) for a.e. x Rd , as j . Hence, mf = g in Lp (Rd ) and we
thus obtain f D(A) and Af = g.
d) Let X = L1 ([0, 1]), Y = C, and Af = f (0) with D(A) = C([0, 1]). Then A is
not closed. In fact, consider the functions fn D(A) given by
(
1 nt, 0 t 1/n,
fn (t) =
1/n t 1,
0,
for every n N. Then kfn k1 = 1/(2n) 0 as n , but Afn = fn (0) = 1.
Definition 1.4. Let A be a linear operator from X to Y . The graph of A is given
by
gr(A) = {(x, Ax) X Y : x D(A)}.
The graph norm of A is defined by kxkA = kxkX + kAxkY . We write [D(A)] if we
equip D(A) with kkA .

1.1. CLOSED OPERATORS

Of course, kkA is equivalent to kkX if A L(X, Y ). We endow X Y with the


norm k(x, y)kXY = kxkX + kykY .
Lemma 1.5. For a linear operator A from X to Y the following assertions hold.
1. gr(A) X Y is a linear subspace.
2. [D(A)] is a normed vector space and A L([D(A)], Y ).
3. A is closed if and only if gr(A) is closed in X Y if and only if [D(A)] is a
Banach space.
4. Let A be injective and put D(A1 ) := R(A). Then, A is closed from X to Y if
and only if A1 is closed from Y to X.
Proof. Assertions 1) and 2) are straightforward to check.
3) The operator A is closed if and only if for all xn D(A), n N, and (x, y)
X Y with (xn , Axn ) (x, y) in X Y as n , we have (x, y) gr(A). This
property is equivalent to the closedness of gr(A). Since k(x, Ax)kXY = kxkX +
kAxkY , a Cauchy sequence or a converging sequence in gr(A) corresponds to a Cauchy
or a converging sequence in [D(A)], respectively. Hence, [D(A)] is complete if and
only if (gr(A), kkXY ) is complete if and only if gr(A) X Y is closed.
4) Assertion 4) follows from 3) since
gr(A1 ) = {(y, A1 y) : y R(A)} = {(Ax, x) : x D(A)}
is closed in Y X if and only if gr(A) is closed in X Y .
Theorem 1.6 (Closed Graph Theorem). Let X and Y be Banach spaces and A be
a closed operator from X to Y . Then A is bounded (i.e., kAxk ckxk for some
c 0 and all x D(A)) if and only if D(A) is closed in X.
In particular, a closed operator with D(A) = X already belongs to L(X, Y ).
Proof. : Let D(A) be closed in X. Then D(A) is a Banach space for kkX
and kkA . Since kxkX kxkA for all x D(A), a corollary to the open mapping
theorem (see e.g. Corollary 3.17 in [FA]) shows that there is some c > 0 such that
kAxkY kxkA ckxkX for all x D(A).
: Let A be bounded and let xn D(A) converge to x X with respect to
kkX . Then kAxn Axm kY ckxn xm kX , and so the sequence (Axn )n is Cauchy
in Y . There thus exists y := limn Axn in Y . The closedness of A shows that
x D(A); i.e., D(A) is closed in X.
Remark. a) Theorem 1.6 is wrong without completeness. Consider for instance the
operator T given by (T f )(t) = tf (t), t R, on Cc (R) with supremum norm. This
linear operator is everywhere defined, unbounded and closed: Take fn , f, g Cc (R)
such that fn (t) f (t) and (T fn )(t) = tfn (t) g(t) uniformly for t R as n .
Then g(t) = tf (t) for all t R and so g = T f .

CHAPTER 1. GENERAL SPECTRAL THEORY

b) Let X be an infinite dimensional Banach space and let B be an algebraic basis


of X (i.e., for all x X there are unique 1 , . . . , n C, b1 , . . . , bn B and n N
such that x = 1 b1 + + n bn ). We may assume that kbk = 1 for all b B. Choose
a countable subset B0 = {bk : k N} of B and set
T bk = kbk for each bk B0 , and T b = 0 for each b B\B0 .
Then T can be extended to a linear operator on X which is unbounded, since kT bk k =
k and kbk k = 1. Thus T is not closed.
Proposition 1.7. Let A be closed from X to Y , T L(X, Y ), and S L(Z, X).
Then the following operators are closed.
1. B = A + T with D(B) = D(A),
2. C = AS with D(C) = {z Z : Sz D(A)}.
Proof. a) Let xn D(B), n N, and x X, y Y such that xn x in X
and Bxn = Axn + T xn y in Y as n . Since T is bounded, there exists
T x = limn T xn and so Axn y T x as n . Since A is closed, we deduce
that x D(A) = D(B) and Ax = y T x; i.e., Bx = Ax + T x = y.
b) Let zn D(C), n N, and z Z, y Y such that zn z in Z and ASzn y
in Y as n . Since S is bounded, xn := Szn converges to Sz. Since Axn y
and A is closed, we obtain Sz D(A) and ASz = y; i.e., z D(C) and Cz = y.
Corollary 1.8. Let A be linear on X and C. Then the following assertions
hold.
1. If I A (or I + A) is closed, then A is closed.
2. If I A is bijective with (I A)1 L(X), then A is closed.
Proof. Assertion 1) is a consequence of Proposition 1.7 since A = ((I A) I).
For 2), Lemma 1.5 shows that I A is closed, and then assertion 1) yields 2).
The following examples show that closedness can be lost when taking sums or
products of closed operators. See the exercises for further related results.
Example 1.9. a) Let E = Cb (R2 ) and Ak = k with
D(Ak ) = {f E : the partial derivative k f exists and belongs to E},
for k = 1, 2. Set B = 1 + 2 on
D(B) := D(A1 ) D(A2 ) = Cb1 (R2 ) = {f C 1 (R2 ) : f, 1 f, 2 f E}.
By an excercise, A1 and A2 are closed.
However, B is not closed: Take n Cb1 (R) converging uniformly to some
Cb (R)\C 1 (R). Set fn (x, y) = n (x y) and f (x, y) = (x y) for (x, y) R2 and

1.2. THE SPECTRUM

n N. We then obtain f E, fn D(B), kfn f k = kn k 0 and


Bfn = 0n 0n = 0 0 as n , but f 6 D(B).
b) Let X = C([0, 1]), Af = f 0 with D(A) = C 1 ([0, 1]) and m C([0, 1]) such
that m = 0 on [0, 1/2]. Define T L(X) by T f = mf for all f X. Then the
operator T A with D(T A) = D(A) is not closed.
To see this, take functions fn D(A) such that fn = 1 on [1/2, 1] and fn f in
X with f 6 C 1 ([0, 1]). Then, T Afn = mfn0 = 0 converges to 0, but f
/ D(A).

1.2

The spectrum

Definition 1.10. Let A be a closed operator on X. The resolvent set of A is given


by
(A) = { C; I A : D(A) X is bijective},
and its spectrum by
(A) = C\(A).
We further define the point spectrum of A by
p (A) = { C : there exists some v D(A)\{0} with v = Av} (A),
where we call p (A) an eigenvalue of A and the corresponding v an eigenvector
or eigenfunction of A. For (A) the operator
R(, A) := (I A)1 : X X
and the set {R(, A) : (A)} are called the resolvent.
Remark 1.11. a) Let A be closed in X and (A). The resolvent R(, A) has
the range D(A). Lemma 1.5 and Corollary 1.8 further show that R(, A) is closed
and thus it belongs to L(X) by Theorem 1.6.
b) Let A be a linear operator such that I A : D(A) X has a bounded inverse
for some C. Then A is closed by Corollary 1.8. In this case, the closedness
assumption in Definition 1.10 is redundant.
We set e (t) = et for C, t J, and any interval J R.
Example 1.12. a) Let X = Cd and T L(X). Then (T ) only consists of the
eigenvalues 1 , . . . , m of T , where m d.
b) Let X = C([0, 1]), and Au = u0 with D(A) = C 1 ([0, 1]). Then e D(A) and
Ae = e for each C. Hence, p (A) and so (A) = p (A) = C.
c) Let X = C([0, 1]), and Au = u0 with D(A) = {u C 1 ([0, 1]) : u(0) = 0}.
Let C and f X. We then have u D(A) and (I A)u = f if and only if
u C 1 ([0, 1]), u0 (t) = u(t) f (t), t [0, 1], and u(0) = 0, which is equivalent to
Z t
u(t) =
e(ts) f (s)ds,
0

CHAPTER 1. GENERAL SPECTRAL THEORY

for all 0 t 1. Hence, (A) = and the resolvent is given by


Z

R(, A)f (t) =

e(ts) f (s)ds,

for all 0 t 1, f X, and C.


Let U C be open. The derivative of a function f : U Y at U is given by
1
(f () f ()) Y,

f 0 () = lim
if the limit exists in Y .

Theorem 1.13. Let A be a closed operator on X and let (A). Then the
following assertions hold.
1. AR(, A) = R(, A) I, AR(, A)x = R(, A)Ax for all x D(A), and
1
(R(, A) R(, A)) = R(, A)R(, A) = R(, A)R(, A)

if (A)\{}. The latter identity is called the resolvent equation.


2. The spectrum (A) is closed, where B(, 1/kR(,A)k) (A) and
R(, A) =

( )n R(, A)n+1 =: R ,

n=0

if | | < 1/kR(,A)k. The series converges absolutely in L(X, [D(A)]), uniformly on B(, /kR(,A)k) for each (0, 1). Moreover,
kR(, A)kL(X,[D(A)])

c()
1

for all B(, /kR(,A)k) and a constant c() depending only on .


3. The function (A) L(X, [D(A)]), 7 R(, A), is infinitely often differentiable with
 n
d
R(, A) = (1)n n! R(, A)n+1
for every n N.
d

4.
kR(, A)k

1
.
d(, (A))

1.2. THE SPECTRUM

Proof. 1) The first assertions follow from x = (I A)R(, A)x = R(, A)(I A)x,
where x X in the first equality and x D(A) in the second one. For (A) we
further have
(R(, A) AR(, A))R(, A) = R(, A),
R(, A)(R(, A) AR(, A)) = R(, A).
The resolvent equation then follows by subtraction and interchanging and .
2) Let | | /kR(,A)k for some (0, 1) and x X with kxk 1. We then
have

n
k( )n R(, A)n+1 xkA
kAR(, A)R(, A)n xk + kR(, A)n+1 xk
n
kR(, A)k
n (kR(, A)k + 1 + kR(, A)k) =: n c(),
where we used 1). So the series in 2) converges absolutely in L(X, [D(A)]) uniformly
on B(, /kR(,A)k) and can be estimated in norm by c()(1 )1 (cf. Proposition
3.12 in [FA]). Using also (I A)R(, A) = ( )R(, A) + I, we obtain
(I A)R =

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

n=0

( )n R(, A)n = I,

n=0

and similarly R (I A)x = x for all x D(A). Hence, (A) and R = R(, A).
3) Since 7 R(, A) L(X, [D(A)]) is locally bounded, due to the estimate
in 2), the resolvent equation implies that the map 7 R(, A) L(X, [D(A)]) is
continuous. The resolvent equation then also yields assertion 3) for n = 1. Assume
that 3) holds for some n N. We then obtain
 n+1

d
d 
R(, A) =
(1)n n! R(, A)n+1 .
d
d
Using the formula
n+1

R(, A)

R(, A)

n+1

n
X

R(, A)nj (R(, A) R(, A))R(, A)j ,

j=0

the continuity of R(, A) and the assertion for n = 1, we then conclude that 3) holds
for n + 1.
4) Assertion 4) follows from 2).
Proposition 1.14. Let Rd , m C(), X = Cb (), and Af = mf with
D(A) = {f X : mf X}. Then A is closed,
(A) = m(),
and R(, A)f =

1
m f

for all (A) and f X.

In particular, for every closed (compact) subset S C there is a closed (bounded)


operator B on a Banach space with (B) = S.

CHAPTER 1. GENERAL SPECTRAL THEORY

Proof. The closedness of A can be shown as in the remark on page 5. Let 6 m()
1
and g Cb (). The function f := m
g then belongs to Cb () and f mf = g so
that mf = f g Cb (). As a result, f D(A) and f is the unique solution in
1
D(A) of the equation f Af = g. This means that (A) and R(, A)g = m
g.
In the case that = m(z) for some z , we obtain
((I A)f )(z) = f (z) m(z)f (z) = 0
for every f D(A). Consequently, I A is not surjective and so (A). We
now conclude that (A) = m() since the spectrum is closed.
The final assertion follows from Example 1.12c) if S = . Otherwise, consider
= S. Define A and X as above. Then, (A) = S and A is bounded if S is compact
(where Cb (S) = C(S)).
A similar result holds on Lp -spaces, see e.g. Example IX.2.6 in [Con90].
Example 1.15. Let X = C0 (R+ ) = {f C(R+ ) : limt f (t) = 0} with the
supremum norm and Af = f 0 with
D(A) = C01 (R+ ) = {f C 1 (R+ ) : f, f 0 X}.
As in Example 1.3 one sees that A is closed. Let C with Re > 0 and f X.
We then have u D(A) and u Au = f if and only if u X C 1 (R+ ) and
u0 (t) = u(t) f (t) for all t 0. This equation is uniquely solved by
Z
u(t) =
e(ts) f (s)ds =: (R f )(t),
t

for t 0.
We have to show that R f X. Let > 0. Then there is an t 0 such that
|f (s)| for all s t . Hence
Z
Z

|R f (t)|
e(Re )(ts) |f (s)|ds
e Re r dr =
,
Re
t
0
for all t t , where we substituted r = s t. As a result, R f C0 (R+ ) and so
(A) with R = R(, A). If Re < 0, then e X and e0 = e X. Hence,
e is an eigenfunction for the eigenvalue and { C : Re < 0} (A). Since
(A) is closed, we deduce that
{ C : Re 0} = (A).
Theorem 1.16. Let T L(X). Then (T ) is a non-empty compact set. The
spectral radius r(T ) := max{|| : (T )} is given by
r(T ) = lim kT n k /n = inf kT n k /n kT k,
1

nN

and for C with || > r(T ) we have


R(, T ) =

n1 T n =: R .

n=0

10

1.2. THE SPECTRUM

Proof. 1) Since kT n+m k kT n k kT m k for all n, m N, an elementary lemma (see


Lemma VI.1.4 in [Wer05]) yields that there exists
lim kT n k /n = inf kT n k /n =: r kT k.
1

nN

If || > r, then
lim supkn T n k /n =
1

1
r
1
lim kT n k /n =
< 1.
n
||
||

Therefore the series R converges absolutely in L(X), and uniformly for in all
compact subsets of C\B(0, r) (proof as in Analysis 1). Moreover,
(I T )R =

n=0

n1 T n+1 = I,

n=0

and similarly R (I T ) = I. Hence, (T ) and R = R(, T ). Due to its


closedness, the spectrum (T ) B(0, r) is compact. Therefore, r(T ) exists as the
maximum of a compact subset of R, and r(T ) r.
2) Take s > r(T ), L(X) , and m N. We define f () = (R(, T )) for
(T ). Note that f : (T ) C is complex differentiable. We set
Z
1
Cm () =
m (R(, T ))d.
2i ||=s
Since f is holomorphic, this integral does not depend on s > r(T ) due to complex
analysis. So we may choose for a moment s > r and use the uniformly convergent
series of step 1) to deduce
Z

X
1
Cm () =
mn1 d (T n )
2i ||=s
n=0
Z 2

X
1
=
(seit )mn1 iseit dt (T n ) = (T m ).
2i 0
n=0

Employing the Hahn-Banach theorem, we now choose a functional m L(X)


with km k = 1 and m (T m ) = kT m k (see e.g. Corollary 4.9 in [FA]). Again for any
s > r(T ), we can then estimate
Z 2
1
m
kT k = Cm (m )
|seit |m km k kR(seit , T )k|seit | dt
2 0
sm s max kR(, T )k =: c(s)sm .
||=s

Thus, kT m k1/m sc(s)1/m and so r s. This means that r(T ) = r.


Finally, suppose that (T ) = . Then the functions f are entire for every
L(X) . Moreover, step 1) yields
|f ()| kk||

X
kT kn
n=0

11

||n

2kk
,
||

CHAPTER 1. GENERAL SPECTRAL THEORY

for all C with || 2kT k. Hence, f is bounded and thus constant by Liouvilles
theorem from complex analysis. The above estimate then shows that (R(, T )) = 0
for all C and L(X ). Using again the Hahn-Banach theorem (see e.g.
Corollary 4.9 in [FA]), we obtain R(, T ) = 0, which is impossible since R(, T ) is
injective and X 6= {0}.
Example 1.17. a) Let X = C([0, 1]) and define the Volterra operator V on X by
Z t
V f (t) =
f (s)ds
0

for t [0, 1] and f X. Then V L(X) and


Z sn1
Z t Z s1
1
n
kf k dsn . . . ds1 kf k ,
...
|V f (t)|
n!
0
0
0
for all n N, t [0, 1], and f X. Hence, kV n k 1/n!. For f =
kV n k kV n 1k = 1/n! and so kV n k = 1/n!. This gives

r(V ) = lim

1
n!

1 we obtain

1/n
=0

and (V ) = {0}.

Observe that p (V ) = since V f = 0 implies that f = (V f )0 = 0. Moreover,


kV k = 1 > r(V ) = 0.
b) Let left shift L given by Lx = (xn+1 ) on X {c0 , `p : 1 p } has the
spectrum
(L) = B(0, 1).
In fact, L L(X) has norm 1 (see e.g. Example 1.57 in [FA]), and so (L) B(0, 1).
Moreover, L(1, 0, ) = 0, and for || < 1 the sequence v = (n )nN belongs to X
and satisfies v = Lv so that B(0, 1) p (L) (L). The closedness of (L) then
yields (L) = B(0, 1). Note that p (L) = B(0, 1) if X = ` , but p (L) = B(0, 1) in
the other cases.
Definition 1.18. Let A be a closed operator on X. Then we call
ap (A) = { C : there exist xn D(A) with kxn k = 1 for all n N,
and xn Axn 0 as n }
the approximate point spectrum of A and
r (A) = { C : (I A)D(A) is not dense in X}
the residual spectrum of A.
Proposition 1.19. For a closed operator A on X the following assertions hold.
1. ap (A) = p (A) { C : (I A)D(A) is not closed in X}.
2. (A) = ap (A) r (A).

12

1.2. THE SPECTRUM

3. (A) ap (A).
(Note that the unions need not be disjoint.)
Proof. 1) We have 6 ap (A) if and only if there is a c > 0 such that k(I A)xk
ckxk for all x D(A). This lower estimate implies that 6 p (A). Moreover, if
yn = xn Axn y in X as n for some xn D(A), then the lower estimate
shows that (xn ) is Cauchy in X, so that xn converges to some x X. Hence,
Axn = xn yn x y and the closedness of A yields x D(A) and x Ax = y.
Consequently, (I A)D(A) is closed.
Conversely, if (I A)D(A) is closed and 6 p (A), then the inverse (I A)1
exists and is closed on its closed domain (I A)D(A). The closed graph theorem
1.6 then yields the boundedness of (I A)1 . Thus,
kxk = k(I A)1 (I A)xk Ck(I A)xk
for all x D(A) and a constant C > 0. This means that 6 ap (A).
2) Assertion 2) follows from 1).
3) Let (A). Then there exist n (A) with n as n . Due
to Theorem 1.13(4), kR(n , A)k as n , and thus there are yn X such
that kyn k = 1 for all n N and an := kR(n , A)yn k as n , where we can
assume that an > 0 for all n N. Set xn = a1n R(n , A)yn D(A). We then have
kxn k = 1 for all n N and xn Axn = ( n )xn + a1n yn 0 as n . As a
result, ap (A).
Proposition 1.20. Let A be closed on X and (A). Then the following assertions hold.
1
1. (R(, A))\{0} = ( (A))1 = {
: (A)}.

2. j (R(, A))\{0} = ( j (A))1 for j = p, ap, r.


3. If x is an eigenvector for the eigenvalue 6= 0 of R(, A), then y = R(, A)x is
an eigenvector for the eigenvalue = 1/ of A. If y D(A) is an eigenvector
for the eigenvalue = 1/ of A with C \ {0}, then x = 1 (y Ay)
is an eigenvector for the eigenvalue of R(, A).
4. r(R(, A)) = 1/d(,(A)).
5. If A is unbounded (i.e., D(A) 6= X), then 0 (R(, A)).
Proof. For C\{0} we have



1
I R(, A) =

I A R(, A).

Since R(, A) : X D(A) is bijective, we obtain assertion 2) for j = p, part 3)


and the equality of the ranges of the operators I R(, A) and ( 1 )I A. Thus
the assertions 1) and 2) for j = ap, r follow from Proposition 1.19. Assertion 4) is a
consequence of 1). Finally, 5) is true because of R(, A)1 = I A.

13

CHAPTER 1. GENERAL SPECTRAL THEORY

Example 1.21. a) Let X = Lp (R), 1 p , and the translation T (t) be given


by (T (t)f )(s) = f (s + t) for s R, f X, and t R. Recall from Example 3.8
in [FA], that T (t) is an isometry on X with inverse (T (t))1 = T (t) for every
t R. By Theorem 1.16 we have (T (t)) B(0, 1). Proposition 1.20 further yields
(T (t))1 = (T (t)1 ) = (T (t)) B(0, 1) so that (T (t)) B(0, 1) for all
t R. Fix t 6= 0. For every iR, the function e belongs to Cb (R) L (R) and
(T (t)e )(s) = e(s+t) = et e (s)
for all s R. Hence, (T (t)) = p (T (t)) = B(0, 1) for p = .
If p [1, ), we use e to construct an approximate eigenfunction. For n N
set fn = n1/p 1[0,n] e . We then have kfn kp = n1/p k1[0,n] kp = 1 and (see above)
kT (t)fn et fn kp = n /p ket (1[t,nt] 1[0,n] )kp = n /p |2t| /p 0,
1

as n . As a result, (T (t)) = B(0, 1) if t 6= 0.


b) Let X = C0 (R) and Au = u0 with D(A) = C01 (R) := {f C 1 (R) : f, f 0
C0 (R)}. As in Example 1.15 one sees that (A) if Re 6= 0 and
Z

e(ts) f (s)ds, if Re > 0,

R(, A)f (t) =


t

R(, A)f (t) =

e(ts) f (s)ds, if Re < 0,

for all t R and f X. Let Re = 0. Choose n Cc1 (R) with k0n k 1/n and
kn k = 1, and set un = n e for all n N. Then, kun k = 1, un D(A) and
Aun = 0n e + n e0 = 0n e + un . Since k0n e k 1/n, we obtain ap (R) and
so (A) = iR.
Definition 1.22. Let A be a linear operator from X to Y with dense domain. We
define its adjoint A from Y to X by setting
D(A ) = {y Y ; z X x D(A) : hAx, y i = hx, z i},
A y = z .
Observe that it holds
hAx, y i = hx, A y i
for all x D(A) and y D(A ).
Remark 1.23. Let A be linear from X to Y with D(A) = X.
a) Since D(A) is dense, there is at most one vector z = A y as in Definition 1.22,
so that A : D(A ) X is a map. It is clear that A is linear. If A L(X, Y ), then
Definition 1.22 coincides with the definition of A in 4.4 of [FA], where D(A ) = Y .
b) The operator A is closed from Y to X .

14

1.2. THE SPECTRUM

Proof. Let yn D(A ), y Y , and z X such that yn y in Y and


zn := A yn z in X as n . Let x D(A). It then holds that
hx, z i = lim hx, zn i = lim hAx, yn i = hAx, y i.
n

As a result, y D(A ) and A y = z .


c) If T L(X, Y ), then the sum A + T with D(A + T ) = D(A) has the adjoint
(A + T ) = A + T with D((A + T ) ) = D(A ).
Proof. Let x D(A) and y Y . We obtain
h(A + T )x, y i = hAx, y i + hx, T y i.
Hence, y D((A + T ) ) if and only if y D(A ), and then (A + T ) y = A y +
T y.
Remark. The operator Af = f 0 with D(A) = {f C 1 ([0, 1]) : f (0) = 0} is not
densely defined on X = C([0, 1]) since D(A) = {f X : f (0) = 0}.
Theorem 1.24. Let A be a closed operator on X with dense domain. Then the
following assertions hold.
1. r (A) = p (A ).
2. (A) = (A ) and R(, A) = R(, A ) for every (A).
Proof. 1) Due to a corollary of the Hahn-Banach theorem (see e.g. Proposition 4.11
in [FA]), the set (I A)D(A) is not dense in X if and only if there is a vector
y X \{0} such that hx Ax, y i = 0 for every x D(A). This fact is equivalent
to the equality hAx, y i = hx, y i for every x D(A), which in turn means that
y D(A )\{0} and A y = y ; i.e., p (A ).
2) Let (A). Take x D(A), x X , and set y = R(, A) x . We then
obtain
h(I A)x, y i = hR(, A)(I A)x, x i = hx, x i.
Thus, y D(A ) and x = (I A) y = (I A )y , where we use Remark 1.23.
This means that I A is surjective. Further, take x D(A ) and x X. We
compute
hx, R(, A) (I A )x i = hR(, A)x, (I A )x i
= h(I A)R(, A)x, x i = hx, x i,
using Definition 1.22 and that R(, A)x belongs to D(A). Hence, R(, A) (I
A )x = x so that I A is also injective. It thus exists R(, A ) = R(, A) .
Conversely, let (A ). Take x D(A). For every x X , we compute as
above
h(I A)x, R(, A )x i = hx, (I A )R(, A )x i = hx, x i.

15

CHAPTER 1. GENERAL SPECTRAL THEORY

Due to a corollary of the Hahn-Banach theorem (see e.g. Corollary 4.9 in [FA]), there
is a y X such that ky k = 1 and hx, y i = kxk. Hence,
kxk = h(I A)x, R(, A )y i kR(, A )kkx Axk.
This estimate implies that 6 ap (A). Further, 6 p (A ) = r (A) by part a), and
so Proposition 1.19 shows that 6 (A).
Example 1.25. Let X = `p , 1 p < , or X = c0 . Let Rx = (0, x1 , x2 , . . . ) be
the right shift on X. Then R = L, where the left shift L acts on `1 if X = c0 and
0
on `p otherwise. Since (L) = B(0, 1) by Example 1.17, Theorem 1.24 yields
(R) = (L) = B(0, 1).
Moreover, r (R) = p (L) = B(0, 1) if X = c0 or X = `p with 1 < p < and
r (R) = p (L) = B(0, 1) if X = `1 . If X = ` , then R = L for L on `1 so that
again (R) = B(0, 1).
Let A be a linear operator. Then a linear operator B from X to Y is called
A-bounded if D(A) D(B) and B L([D(A)], Y ).
Theorem 1.26. Let A be a closed operator on X and (A). Further, let B be
linear on X, A-bounded and satisfy kBR(, A)k < 1. Then A + B with D(A + B) =
D(A) is closed, (A + B),
R(, A + B) = R(, A)

(BR(, A))n = R(, A)(I BR(, A))1 ,

n=0

and
kR(, A + B)k

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

P
n It is known that the Neumann series
Proof.
P Set R = nR(, A) n=0 (BR(, A)) . 1
and that the asserted norm estimate
n=0 (BR(, A)) converges to (IBR(, A))
holds, see e.g. Proposition 3.12 in [FA]. Clearly, R maps into D(A) and it holds
(I A B)R =

(BR(, A))n

n=0

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

n=0

P
n
Since R =
n=0 (R(, A)B) R(, A), we also obtain R (I A B)x = x for all
x D(A). As a result (A + B) and R = R(, A + B).
Observe that the smallness condition in this theorem is sharp in general: Let
X = C, a C
= L(C), a 6= 0, and b = a. Then a is invertible, but a a = 0 is not.
Here we have = 0 and |bR(0, a)| = | aa | = 1.

16

Chapter 2

Spectral theory of compact


operators
2.1

Compact operators

A nonempty subset B X is called relatively compact if its closure is compact in


X. We will often use that this property holds if and only if each sequence in B
has a converging subsequence (with limit in B). In fact, the necessity of the latter
condition is clear. Conversely, assume that each sequence in B has a converging
subsequence. Let (xn ) be a sequence in B. Then, for each n N there exists a
yn B with kxn yn k 1/n. By the assumption, we have a subsequence (ynj )j with
limit y in B. Consequently, xnj converges to y as j .
Definition 2.1. A linear map T : X Y is called compact if T B(0, 1) is relatively
compact in Y . The set of all compact operators is denoted by L0 (X, Y ).
Remark 2.2. a) If T is compact, then T B(0, 1) is bounded and thus T is bounded;
i.e., L0 (X, Y ) L(X, Y ).
b) Let T : X Y be linear. Then the following assertions are equivalent.
(i) T is compact.
(ii) T maps bounded sets of X into relatively compact sets of Y .
(iii) For every bounded sequence (xn )n X there exists a convergent subsequence
(T xnj )j in Y .
Proof. (i) (ii): If T is compact and B X is bounded, then B B(0, r) for some
r 0 and T B T B(0, r) = rT B(0, 1), which is compact. Hence, T B is compact.
The implications (ii) (iii) (i) are clear.
c) The space of operators of finite rank is defined by
L00 (X, Y ) = {T L(X, Y ) : dim T X < }.

17

CHAPTER 2. SPECTRAL THEORY OF COMPACT OPERATORS

The Heine Borel theorem yields that L00 (X, Y ) L0 (X, Y ).


d) The identity I : X X is compact if and only if B(0, 1) is compact if and
only if dim X < (see e.g. Theorem 1.40 in [FA]).
Proposition 2.3. L0 (X, Y ) is a closed linear subspace of L(X, Y ). Let T L(X, Y )
and S L(Y, Z). If one of the operators T or S is compact, then ST is compact.
Proof. Let xk X, k N, satisfy supkN kxk k =: c < .
a) Let T, R L0 (X, Y ) be compact. If C, then T is also compact. There
further exists a converging subsequence (T xkj )j . Since (xkj )j is still bounded, there
is another converging subsequence (Rxkjl )l . Hence, ((T + R)xkjl )l converges and so
T + R L0 (X, Y ).
b) Let Tn L0 (X, Y ) converge in L(X, Y ) to some T L(X, Y ) as n . Since
T1 is compact, there is a converging subsequence (T1 x1 (j) )j . Because kx1 (j) k c
for all j, there exists a subsubsequence 2 1 such that (T2 x2 (j) )j converges. Note
that (T1 x2 (j) )j still converges. Iteratively, we obtain subsequences l l1 such
that (Tn xl (j) )j converges for all l n.
Set um = xm (m) . Then (Tn um )m converges as m for each n N. Let
> 0. Fix N = N N such that kTN T k . Let M m N . We then obtain
kT uM T um k k(T TN )uM k + kTN (uM um )k + k(TN T )um k
2c + kTN (uM um )k.
Therefore (T um ) is a Cauchy sequence. So we have shown that T is compact. Hence,
L0 (X, Y ) is closed.
c) Let S L0 (X, Y ). Since (T xk )k is bounded, there is a converging subsequence
(ST xkj )j , so that ST is compact.
If T L0 (X, Y ), then there is a converging subsequence (T xkj )j and thus ST xkj
converges. Again, ST is compact.
Remark 2.4. Strong limits of compact operators need not be compact. Consider,
e.g., X = `2 and Tn x = (x1 , . . . , xn , 0, 0, . . . ) for all x X and n N. Then
Tn L00 (X) L0 (X) but Tn x x = Ix as n for every x X and I 6 L0 (X).
Example 2.5. a) Let X {C([0, 1]), Lp ([0, 1]), 1 p }, Y = C([0, 1]), and
k C([0, 1]2 ). Setting
Z 1
T f (t) =
k(t, )f ( )d,
0

for f X and t [0, 1], we define the integral operator T : X Y for the kernel
k. Since kT f k kkk kf k1 kkk kf kp (using that ([0, 1]) = 1), we obtain that
T L(X, Y ) and thus T B X (0, 1) is bounded in Y . Moreover, for t, s [0, 1] and
f B X (0, 1) we have
Z
|T f (t) T f (s)|

|k(t, ) k(s, )||f ( )|d


0

18

2.1. COMPACT OPERATORS

sup |k(t, ) k(s, )|kf k1


[0,1]

sup |k(t, ) k(s, )|kf kp .


[0,1]

The right hand side tends to 0 as |t s| 0 uniformly in f B(0, 1), because k is


uniformly continuous. Therefore T B is equicontinuous. The Arzela-Ascoli theorem
(see e.g. Theorem 1.45 in [FA]) then implies that T B X (0, 1) is relatively compact.
Hence, T L0 (X, Y ).
b) Let Z = L2 ([0, 1]) and k L2 ([0, 1]2 ). Define T L(Z) as in a), cf. Example 2.23 in [FA]. There are kn C([0, 1]2 ) converging to k L2 ([0, 1]2 ) (cf. Analysis 3). Let Tn be the corresponding integral operators in L(Z) and L(Z, C([0, 1])) as
in a). H
olders inequality yields
Z

|k(t, ) kn (t, )||f ( )|d

|T f (t) Tn f (t)|
0

Z

|k(t, ) kn (t, )| d

 12
kf k2

for all n N, f Z, and t [0, 1]. Hence kT f Tn f k2 kk kn k2 kf k2 and so


Tn T as n in L(Z). Let (fk ) Z be bounded and n N. Due to a), there is
a subsequence (Tn fkj )j converging for kk , and hence for kk2 (since ([0, 1]) < ).
So, Tn L0 (Z) and the compactness of T then follows from Proposition 2.3.

Proposition 2.6. Let T L0 (X, Y ) and xn x in X as n (i.e., hxn , x i


hxn , x i as n for every x X ). Then T xn converges to T x in Y as n .
Proof. We have hT xn T x, y i = hxn x, T y i 0 as n for each y Y ,

hence T xn T x as n .
Suppose that T xn does not converge to T x. Then there are > 0 and a subsequence such that
kT xnj T xk > 0
for all j N. By compactness, there is a subsubsequence and a y Y such that

T xnjl y as l . On the other hand, T xnjl T x and T xnjl y. Since weak


limits are unique, it follows that y = T x, which is impossible.
Theorem 2.7 (Schauder). An operator T L(X, Y ) is compact if and only if T
L(Y , X ) is compact.
Proof. 1) Let T be compact and yn Y , n N, with supnN kyn k =: c < . The
set K := T B X (0, 1) is a compact metric space for the restriction of the norm of Y .
Set fn := yn |K C(K) for each n N. Putting c1 := maxyK kyk < , we obtain
kfn k = max|hy, yn i| cc1
yK

19

CHAPTER 2. SPECTRAL THEORY OF COMPACT OPERATORS

for every n N. Moreover, (fn )nN is equicontinuous since


|fn (y) fn (z)| kyn k ky zk c ky zk
for all n N and y, z K.
The Arzela-Ascoli theorem then yields a subsequence (fnj )j converging in C(K).
We thus deduce that
kT yn j T yn l kX = sup |hx, T (yn j yn l i|
kxk1

= sup |hT x, yn j yn l i| = kfnj fnl kC(K)


kxk1

tends to 0 as j, l (using that T B(0, 1) is dense in K). This means that (T yn j )j


converges and so T is compact.
2) Let T be compact. By Step 1), the biadjoint T is compact. Let JX : X
be the canonical isometric embedding. Due to e.g. Proposition 4.39 in [FA], it
holds that
T JX = JY T : X Y .
X

Using also Proposition 2.3 we see that JY T is compact. If (xn ) is bounded, we thus
obtain a converging subsequence (JY T xnj )j which is Cauchy. Since JY is isometric,
also (T xnj )j is Cauchy and thus converges; i.e., T is compact.

2.2

The Fredholm alternative

We need some fact from functional analysis. For nonempty sets M X and N X
we define the annihilators
M = {x X : y M hy, x i = 0},

N = {x X : y N hx, y i = 0}.

These sets are equal to X or X if and only if M = {0} or N = {0}, respectively,


see e.g. Remark 4.18 in [FA]. For T L(X) it holds
(2.1)

R(T ) = N (T ),

R(T ) = N (T ) N (T ) = R(T ),

R(T ) N (T ) .

In particular, R(T ) is dense if and only if T is injective; and if R(T ) is dense,


then T is injective. (See e.g. Proposition 4.41 in [FA].)
The following theorem extends fundamental results on matrices known from Linear Algebra.
Theorem 2.8 (Riesz 1918, Schauder 1930). Let K L0 (X) and T = I K. Then
the following assertions hold.
a) R(T ) is closed.

20

2.2. THE FREDHOLM ALTERNATIVE

b) dim N (T ) < and codim R(T ) := dim X/R(T ) < .


c) T is bijective T is surjective T is injective T is bijective T is
surjective T is injective.
More precisely, we have
(2.2)

dim N (T ) = codim R(T ) = dim N (T ) = codim R(T ).

Corollary 2.9 (Fredholm alternative). Let L L0 (X), C\{0}, and y X.


Then one of the following alternatives hold:
A) x = Lx has only the trivial solution x = 0.
Then for every y X there is a unique solution x X of x Lx = y given by
x = R(, L)y.
B) x = Lx has an n-dimensional solution space N (I L) for some n N.
Then there are n linearly independent solutions x1 , . . . , xn X of x = Lx ,
and the equation x Lx = y has a solution x X if and only if hy, xk i = 0 for
every k = 1, . . . , n.
Finally, every z X satisfying z Lz = y is of the form z = x + x0 , where
x Lx = y and x0 N (I L).
Proof of Corollary 2.9. Apply Theorem 2.8 to K := 1 L L0 (X) and note that
xLx = y is equivalent to (IK)x = 1 y. By Theorem 2.8b), either dim N (IK) =
0 (case A) or dim N (I K) = n N (case B).
In the first case, I K is bijective due to Theorem 2.8c) which yields part A.
In the second case, Theorem 2.8a) shows that R(I K) is closed so that R(I
K) = N (I K ), by equation (2.1). This gives the solvability condition from
case B noting that dim N (I K ) = n due to equation (2.2). If x Kx = y and
z Kz = y, then z x belongs to N (I K ), as required in case B.
Rt
Example 2.10. Let X = C([0, 1]) and Kf (t) = 0 f (s)ds for t [0, 1] and f X.
Then
R(K) = {g C 1 ([0, 1]) : g(0) = 0},
which is not closed in X. Moreover, N (K) = {0} by Example 1.17. Also, K is
compact since KB X (0, 1) is contained in the closed unit ball of C 1 ([0, 1]), which is
a compact subset of X due to the Arzela-Ascoli theorem. In this case, Kf = g can
not be solved for all g X. Summing up, the Fredholm alternative fails for = 0.
Proof of Theorem 2.8. 1) The space N := N (T ) is closed in X. For x N we have
Kx = x N , so that K leaves N invariant and its restriction KN to N coincides
with the identity on N . On the other hand, KN is still compact so that dim N <
by Remark 2.2d).
2) Since dim N < , there is a closed subspace C X such that N C = {0}
and N + C = X, i.e., X = N C, see e.g. Proposition 4.15 in [FA].

21

CHAPTER 2. SPECTRAL THEORY OF COMPACT OPERATORS

Let T : C R(T ) be the restriction of T to C and endow C and R(T ) with the
norm of X. (Observe that C is a Banach space, but we do not know yet whether
R(T ) is a Banach space.)
If Tx = 0 for some x C, then x N and so x = 0. Let y R(T ). Then there
is an x X such that T x = y. We can write x = x0 + x1 with x0 N and x1 C.
Hence, Tx1 = T x1 + T x0 = y, and so T is bijective.
A corollary to the open mapping theorem (see e.g. Corollary 3.19 in [FA]) yields
that R(T ) = R(T) is closed if and only if T1 : R(T ) C is bounded. Assume that
T1 were unbounded. Then there would exist yn = Txn R(T ) with xn C such
that yn 0 as n and kxn k = kT1 yn k for some > 0 and all n N. If
nl := kx1n k xnl and note
there exists an unbounded subsequence (xnl ), then we set x
l
that k
xnl k = 1 and that
1
nl =
ynl := Tx
yn
kxnl k l
still converges to 0 as l . So we can assume that (xn ) is bounded.
Since K is compact, there exists a subsequence (xnj )j and a z X such that
Kxnj z as j . Consequently, xnj = ynj + Kxnj z as j and so
kzk > 0. We further deduce that z C from the closedness of C and that z N
from
T z = z Kz = lim Kxnj K lim xnj = 0.
j

Hence, z C N = {0}, which contradicts kzk > 0. So, assertion a) has been
shown.
3) Theorem 2.7 shows that K is also compact so that dim N (I K ) < by
the first step. Using equation (2.1) and Proposition 4.20 in [FA], we further obtain
N (T ) = R(T )
= (X/R(T )) .
Since N (T ) is finite-dimensional, linear algebra yields that
()

dim N (T ) = dim(X/R(T )) = dim X/R(T ) = codim R(T )

and so codim R(T ) < , i.e., assertion b) is true.


with dim N
< and a closed linear
4) Claim A: There is a linear subspace N

subspace R of X such that


R,

X=N

N
,
TN

TR

and

R
is bijective.
T2 := T|R : R

), we have
Suppose for a moment that Claim A is true. Setting T1 := T|N L(N
the following facts.
/R(T1 ) = dim N (T1 ) (by the dimension formula in Cn ).
(i) dim N
and x2 R
we deduce that T x = 0 if
(ii) Writing x = x1 + x2 for x X, x1 N
R
= {0} which is, by the bijectivity of T2 , equivalent
and only if T2 x2 = T1 x1 N
and T1 x = 0, which
to x2 = 0 and T1 x1 = 0. Hence, x N (T ) if and only if x N
means that N (T ) = N (T1 ).

22

2.2. THE FREDHOLM ALTERNATIVE

(iii) We define the map


/R(T1 ) X/R(T ),
:N

x + R(T1 ) 7 x + R(T ),

X. Because of R(T1 ) R(T ), the map is well defined, and it is linear.


for x N
We want to show that is bijective.
, then there is a y X such that x = T y.
If (x + R(T1 )) = 0 for some x N

There are y1 N and y2 R with y = y1 + y2 . Hence, T2 y2 = x T1 y1 belongs to


N
= {0}. Since T2 is injective, we infer that y2 = 0. Thus, is injective.
R
and x2 R
= TR
with x = x1 + x2 . We now
Take x X. There are x1 N
conclude that x x1 = x2 R(T ) and so
(x1 + R(T1 )) = x1 + R(T ) = x + R(T ).
/R(T1 ).
Hence, is bijective, which leads to dim X/R(T ) = dim N
The facts (i)-(iii) imply that
()

/R(T1 ) = dim X/R(T ) = codim R(T ).


dim N (T ) = dim N (T1 ) = dim N

Since also K is compact by Theorem 2.7, we further obtain


dim N (T ) = codim R(T ).

( )

Combining ()-( ), we arrive at assertion c).


5) Proof of Claim A. We set Nk = N (T k ) and Rk = R(T k ) for k N0 . It holds
{0} = N0 N1 N2 . . . ,

X = R0 R1 R2 . . . ,

T Nk Nk1 Nk , and T Rk = Rk+1 Rk ,

(+)

for all k N0 . We further have


k  
X
k
T = (I K) = I
(1)j+1 K j =: I Ck ,
j
k

j=1

where Ck is compact for each k N, due to Proposition 2.3.


Assertions a) and b) thus show that Rk is closed, dim Nk < , and codim Rk <
for every k N. We need four more claims to establish Claim A.
Claim 1: There is a minimal n N0 such that Nn = Nn+j for all j N0 .
Indeed, if it were true that Nj $ Nj+1 for all j N0 , then Riesz Lemma (see
e.g. Lemma 1.42 in [FA]) would give xj Nj with kxj k = 1 and d(xj , Nj1 ) 1/2
for every j N0 . Take l > k 0. Since T xl + xk T xk Nl1 , we deduce that
kKxl Kxk k = kxl (T xl + xk T xk )k 1/2.
As a result, (Kxk )k has no converging subsequence, which contradicts the compactness of K. So there is a minimal n N0 with Nn = Nn+1 . Let x Nn+2 . Then,

23

CHAPTER 2. SPECTRAL THEORY OF COMPACT OPERATORS

T x Nn+1 = Nn so that x Nn+1 . This means that Nn+1 = Nn+2 , and Claim 1
follows by induction.
Claim 2: There is a minimal m N0 such that Rm = Rm+j for all j N0 .
This claim can be shown as Claim 1, see e.g. Lemma VI.2.2 in [Wer05].
Claim 3: Nn Rn = {0} and Nm + Rm = X.
Indeed, for the first part, let x Nn Rn . Then T n x = 0 and there is a y X such
that T n y = x. Hence, T 2n y = 0 and so y N2n = Nn by Claim 1. Consequently,
x = T n y = 0.
For the second part, let x X. By Claim 2 we have T m x Rm = R2m , and thus
there exists a y X with T m x = T 2m y. Therefore x = (xT m y)+T m y Nm +Rm .
Claim 4: It holds n = m.
Indeed, suppose that n > m. Due to Claim 1 and Claim 2, there is an x Nn \Nm
and it holds Rn = Rm . Claim 3 further gives y Nm Nn and z Rm = Rn such
that x = y + z. Therefore, z = x y Nn so that z = 0 by Claim 3. As a result,
x = y Nm , which is impossible. The inequality n < m can be excluded in a similar
way, see e.g. Lemma VI.2.2 in [Wer05].
:= Nn and R
= Rn . The
We can now finish the proof of Claim A, setting N
closedness of Nn and Rn have been established before Claim 1. From Claim 3 and
R.
Moreover, equation (+) yields T N
N
and
4 we thus deduce that X = N
= R.

TR
and y X, then y Nn+1 = Nn by Claim
If T x = 0 for some x = T n y R
1. Therefore, x = 0 and T|R is bijective and thus isomorphic by the open mapping
theorem.
We now reformulate the Riesz-Schauder theorem in terms of spectral theory.
Observe that the Voltera operator in Example 2.10 is compact and has the spectrum
(V ) = {0} with p (V ) = .
Theorem 2.11. Let dim X = and K L(X) be compact. Then the following
assertions hold.
a)
(K) = {0} {j : j J},
where either J = , or J = N, or J = {1, . . . , n} for some n N.
b) Each (K)\{0} is an eigenvalue of K and
dim N (I K) = codim(I K)X < .
c) If J = N, then j 0 as j . (This means that for all > 0 the set
(K)\B(0, ) is finite.)

24

2.2. THE FREDHOLM ALTERNATIVE

Proof. If 0 (K), then K would be invertible. Proposition 2.3 now shows that
I = K 1 K would be compact which contradicts dim X = .
Hence, 0 always belongs to (K). Observe that thus assertion a) follows from c)
by taking = 1/n, n N.
For C\{0} we have I K = (I 1 K). Since 1 K L0 (X), Theorem 2.8
implies either (K) or p (K) with


dim N (I K) = dim N I 1 K = codim I 1 K X = codim(I K)X < .
So we have established part b).
To prove assertion c), we suppose that for some > 0 we have n (K)\B(0, )
with n 6= m for all n 6= m in N and xn X\{0} with Kxn = n xn . In Linear
Algebra it is shown that eigenvectors to different eigenvalues are linearly independent.
Hence, the subspaces
Xn := lin{x1 , . . . , xn }
satisfy Xn $ Xn+1 for every n N. Moreover, KXn Xn and Xn is closed for
each n N (since dim Xn < ). Riesz lemma (see e.g. Lemma 1.42 in [FA]) gives
yn Xn such that kyn k = 1 and d(yn , Xn1 ) 1/2 for each n N. There are
n,j C with yn = n,1 x1 + + n,n xn , and hence the vector
n yn Kyn =

n
X

(n j )n,j xj =

j=1

n1
X

(n j )n,j xj

j=1

belongs to Xn1 . For n > m we thus obtain


kKyn Kym k = |n | kyn

1
|n |

(n yn Kyn + Kym )k
,
n
2
2

since n yn Kyn + Kym Xn1 . This fact contradicts the compactness of K.


Example 2.12 (The Dirichlet problem and boundary integrals). Let D R3 be
open and bounded with D C 2 (see Analysis 3). Let C(D) be given. We
look for u C 2 (D) C(D) satisfying
(
u(x) = 0,
x D,
(2.3)
u(x) = (x), x D.
In the lecture Partial Differential Equations (see e.g. Corollary 2.1.1 in [Jos08]) it
is shown that the problem (2.3) has at most one solution. To find a solution, we
1
consider the Newton potential in R3 given by (x) = 4|x|
for x R3 \{0}. Let
2
(y) be the outer unit normal at y D. We define
k(x, y) =

(x y) (y)
(x y) = ()(x y) (y) =
(y)
4|x y|32

for all x R3 and y D with x 6= y. One defines the double layer potential by
Z
Sg(x) =
k(x, y)g(y)d(y)
D

25

CHAPTER 2. SPECTRAL THEORY OF COMPACT OPERATORS

for all x R3 \D and g C(D). Using results from Analysis 2 or 3, one derives
that Sg C (R3 \D) and Sg = 0 on R3 \D employing that = 0 on R3 \ {0}.
For each C(D) one thus obtains the solution Sg|D of (2.3) if one can find a
g C(D) such that
lim Sg(x) = (z)
xz
xD

for all z D.
To that purpose we recall from Analysis 3 that the surface integral for a measurable function h : D C is given by
Z
m Z
q
X
h(y)d(y) =
j (Fj (t))h(Fj (t)) det F 0 (t)T F 0 (t)dt,
D

j=1

Uj

Pm
if the right hand side exists. Here, 0 j Cc (R3 ) satisfy
j=1 j = 1,
3

j is a
{V1 , . . . , Vm } is an open cover of D in R , Vj = Vj D, j : Vj U
1
1
2
C -diffeomophism such that k j , kl j , k j and kl j have continuous exj , respectively, and Fj = 1 |U has the range Vj , where
tensions to Vj and U
j
j
j (R2 {0}). We identify Uj with a subset of R2 writing t R2 instead of
Uj = U
(t, 0) R3 . In the following we omit the index j {1, . . . , m}.
Recall that for y = F (t) V , the tangent plane of D at y is spanned by 1 F (t)
) yields, for
and 2 F (t), where t U R2 . Taylors formula applied to 1 Cb2 (U
1
2
x = (s, 0) with s U R , that


st
1 0
+ O(|s t|22 ) = y + F 0 (t)(s t) + O(|s t|22 ).
x = y + ( ) (t, 0)
0
Using that (y) is orthogonal to j F (t), we deduce that
(x y) (y) = (y)T F 0 (t)(s t) + O(|s t|22 ) = O(|s t|22 ).
On the other hand, 1 and are globally Lipschitz so that
c|s t|2 |x y|2 C|s t|2
for all x = F (s), y = F (t) V with s, t U and some constants C, c > 0. In the
following we denote by c various, possibly differing constants.
The above facts can be established for all j = 1, . . . , m, and so we obtain
c
c
|k(x, y)|

|x y|2
|s t|2
for all x = F (s) and y = F (t) in D with x 6= y. As a result, the integrands
q
(F (t))k(F (s), F (t))g(F (t)) det F 0 (t)T F 0 (t)
of Sg are bounded by a constant times |s t|1
2 kgk for all x = F (s) D and
y = F (t) D with x 6= y. We now define k(x, x) = 0 for x D and
(
k(x, y),
|x y|2 > 1/n,
kn (x, y) =
3
1
n (4) (x y) (y), |x y|2 1/n,

26

2.2. THE FREDHOLM ALTERNATIVE

1
for n N. Observe that |kn (x, y)| cn3 |s t|22 c|s t|1
2 if |x y|2 /n because
then |s t|2 c/n.
Since kn is continuous on D D, we see that
Z
kn (x, y)g(y)d,
Tn g(x) =
D

for x D and g C(D), defines an operator Tn L(C(D)). As in Example


2.5a), one also shows that Tn is compact by means of the Arzela-Ascoli theorem. For
g C(D) and x D we further estimate
Z
Z
|(k(x, y) kn (x, y))g(y)|d(y) =
|k(x, y) kn (x, y)||g(y)|d(y)
D

1
DB(x, n
)
Z
m
X

j=1

Uj B(s, nc )

Z
ckgk

B(0, nc )

Z
ckgk
0

c
n

kgk
dt
|s t|2

dv
|v|2

rdr
ckgk

,
r
n

using polar coordinates in R2 .


Hence, for each x D the function y 7 k(x, y)g(y) is integrable for the surface
measure d on D. Moreover, Tn g converges uniformly for x D to the function
Z
T g(x) :=
k(x, y)g(y)d(y),
D

as n , for every g C(D). This fact yields that T g C(D). Clearly


T : C(D) C(D) is linear.
We have even shown that Tn converges to T in operator norm so that T is
bounded and, by Proposition 2.3, compact. One can now prove that, for all z D
and g C(D), it holds
(2.4)

1
g(z),
lim
Sg(x)
=
T
g(z)

xz
2
xD

and

1
lim
Sg(x)
=
T
g(z)
+
g(z),
xz
2
3

xR \D

see e.g. Theorem IX in Chapter VI of [Kel67]. As a result, Sg C 2 (D) C(D) solves


(2.3) if and only if there is a g C(D) such that 21 g T g = .
In view of the Fredholm alternative we only have to establish that 12 I T is
injective. So let g C(D) satisfy 12 g0 = T g0 . We have already seen that then Sg0
solves (2.3) with = 0. This problem is also solved by u = 0. The uniqueness of
(2.3) now yields that Sg0 = 0 on D.
Equation (2.4) and T g0 = 21 g0 show that the function
(
Sg0 (x), x R3 \D,
v(x) =
g0 (x),
x D,

27

CHAPTER 2. SPECTRAL THEORY OF COMPACT OPERATORS

is continuous on R3 \D. Fix r0 > 0 such that D B(0, r0 ) =: B(r0 ). For r r0 + 1


we can estimate
c
c
c

2,
(r r0 )2
r
|x y|22
Z
c
c
kg0 k d 2 ,
|Sg0 (x)|
2
r
D r
|k(x, y)|

for all x B(r) and y D. Assume that g0 6= 0. Then we can choose r r0 + 1


such that
kg0 k > max |Sg0 (x)|.
xB(r)

In particular, v is not constant on B(r)\D. Since v = 0 on B(r)\D, the strong


maximum principle says that v attains no maximum in B(r)\D (see e.g. Theorem
2.1.2 in [Jos08]). As the maximum exists on B(r)\D it must be attained at a y0 D
and v(y0 ) > v(x) for all x B(r)\D. Moreover, v(y0 ) exists and is equal to 0 since
Sg0 = 0 on D, see e.g. Theorem X in Chapter VI of [Kel67]. These facts contradict
Hopfs Lemma (see e.g. Lemma 2.1.2 in [Jos08]) so that g0 = 0. As a result, 21 I T
is invertible on C(D) and

1 !
1
u = S
I T

2
|D

solves (2.3). Summing up, we have found a unique u C 2 (D) C(D) solving the
Dirichlet problem (2.3).

28

Chapter 3

Sobolev spaces and weak


derivatives
Throughout, U Rd is open and nonempty.

3.1

Basic properties

We are looking for properties of C 1 function and their derivatives which can be
generalized to a concept of derivatives suited to Lp spaces, which is in particular not
based on pointwise limits. To that purpose, take f C 1 (U ) and Cc (U ). There
is an open set V in Rd such that
supp V V U.
We write x V as x = (x1 , y) for x1 Vy := {t R : (t, y) V } and y Rd1 with
Vy 6= . Observe that Vy R is open and thus a disjoint union of open intervals.
Integrating by parts, we deduce
Z
Z
Z
(1 f )dx =
1 f (x1 , y)(x1 , y)dx1 dy
y:Vy 6=

Vy

=
y:Vy 6=

Vy

f (x1 , y)1 (x1 , y)dx1 + [f (x1 , y)(x1 , y)]|x1 Vy

f (x1 , y)1 (x1 , y)dx1 dy =


y:Vy 6=

dy

Vy

f 1 dx,
U

since vanishes on Vy for each y Rd1 with Vy 6= . Inductively one shows that
Z
(3.1)

||

( f )dx = (1)
U

f dx,

for all f C k (U ), Cc (U ) and Nd0 with || k.

29

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

We set
L1loc (U ) = {f : U C : f is measurable, f|K L1 (K) for all compact K U }.
We extend f L1loc (U ) by 0 to a measurable function f : Rd C without further
notice. Convergence in L1loc (U ) means that (fn )|K converges in L1 (K) for all compact
K U . Observe that Lp (U ) L1loc (U ) for all 1 p .
Definition 3.1. Let f L1loc (U ) and Nd0 . If there is a function g L1loc (U ) such
that
Z
Z
||
f dx,
gdx = (1)
U

Cc (U ),

for all
then g =:
is called weak derivative of f . If f possesses
weak derivates for all || k, then we write f W k (U ). Moreover, one defines the
Sobolev spaces by
Wpk (U ) = {f Lp (U ) W k (U ) : f Lp (U ) for all || k}
for k N and 1 p and endows them with

1/p

k f kpp ,

0||k
kf kk,p =

max k f k ,

1 p < ,

p = ,

0||k

where 0 f := f . We write Wp0 (U ) = Lp (U ).


As usually, the spaces L1loc (U ), W k (U ) and Wpk (U ) are spaces of equivalence
classes modulo the subspace N = {f : U R : f is measurable, f = 0 a.e.}.
Remark 3.2. a) We will see in Lemma 3.5 that f is uniquely defined. It is then
also clear that the map
: W k (U ) L1loc (U )
is linear if || k.
b) Formula (3.1) implies that C k (U ) + N W k (U ) and that weak and classical
derivates coincide for f C k (U ).
c) Let 1 p and k N. Clearly, (Wpk (U ), kkk,p ) is a normed vector space
and
J : Wpk (U ) Lp (U )m , f 7 ( f )||m ,
is an isometry where m = 1 + d + . . . + dk . We see in the proof of the next proposition
that Wpk (U ) is isometrically isomorphic to a closed subspace of Lp (U )m .
d) Let 1 p and k N. Since the p-norm and the 1-norm on Rm are
equivalent, there are constants Ck , ck > 0 such that
X
X
ck
k f kp kf kk,p Ck
k f kp
0||k

0||k

30

3.1. BASIC PROPERTIES

for all f Wpk (U ).


e) Observe that kf kp1,p = kf kpp + k|f |p kpp for all 1 p < .
Proposition 3.3. For all 1 p and k N, Wpk (U ) is a Banach space. It is
separable if p < and reflexive if 1 < p < . Moreover, W2k (U ) is a Hilbert space
endowed with the scalar product
X Z
hf, giW k =
( f ) gdx.
2

||k

Proof. Let (fn )n be a Cauchy sequence in Wpk (U ). Then ( fn )n is a Cauchy sequence in Lp (U ) for every Nd0 with || k and thus fn g in Lp (U ) for
some g Lp (U ) as n , where we set f := g0 .
Let Cc (U ) and || k. Since fn Wpk (U ), we deduce
Z
Z
Z

||
f dx = lim
fn dx = lim (1)
( fn )dx
n
n
U
UZ
U
g dx.
= (1)||
U

This means that g is the weak derivative f so that f Wpk (U ) and fn


f in Wpk (U ). Hence, Wpk (U ) is a Banach space. Using Remark 1.60 in [FA], we
then deduce from Remark 3.2c) that Wpk (U ) is isometrically isomorphic to a closed
subspace of Lp (U ). The remaining assertions now follow by isomorphy from known
results of functional analysis.
Example 3.4. a) Let f Cc (R) be such that f := f|R belong to C 1 (R ). We
then have f W 1 (R) with

 0
f+ on [0, )
=: g.
f := 1 f =
f0 on (, 0)
For f (x) = |x|, we thus obtain f = 1R+ 1(,0) .
Proof. For every Cc (U ), we compute
Z

f dt =
R

f dt +

f+ 0 dt

0
0

f0 dt

f |0

f+0 dt + f+ 0 |
0

Z
=

gdt,
R

since f+ (0) = f (0) by the continuity of f .


b) The function f = 1R+ does not belong to W 1 (R).

31

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

Proof. Assume there would exist g = f L1loc (R). Then we would obtain for every
Cc (R) that
Z
Z
Z
0 (t)dt = (0).
gdt = 1R+ 0 dt =
R

Taking with supp (0, ), we deduce from Lemma 3.5 below that g = 0 on
(0, ). Similarly, it follows that g = 0 on (, 0). Hence, g = 0 and so (0) = 0
for all Cc (R), which is false.
c) Let d 2, U = B(0, 1), 1 p < d, and f (x) = ln|x|2 for x U \{0}. Then we
have f Wp1 (U ) with
xj
j f (x) =
=: gj (x),
|x|22
for x 6= 0. Observe that f is unbounded and has no continuous extension at x = 0.
Proof. Using polar coordinates, we obtain
Z 1
p
kf kp = c
|ln r|p rd1 dr < ,
0

for all p [1, ). Estimating |xj | r, we further compute


Z 1 p
Z 1
r d1
p
kgj kp c
r dr = c
rdp1 dr < ,
2p
0 r
0
and thus gj Lp (U ), for all p [1, d) and j {1, . . . , d}. Take j = 1, (0, 1)
and Cc (U ). We set J = (1, ] [, 1) and write y = (x2 , . . . , xd ). We then
obtain
Z
Z
f 1 dx =
f (x1 , y)1 (x1 , y)dx1 dy
U
[1,1]d

Z
Z  q
2
2
= lim
ln x1 + |y|2 1 (x1 , y)dx1 dy
0 [1,1]d1 J
 q

Z
h Z
x1
2
2
= lim

+ |y|2 (, y)
2
2 (x1 , y)dx1 ln
0 (1,1)d1
J x1 + |y|2
 q

i
2
2
+ ln + |y|2 (, y) dy
Z
=
g1 dx.
U

Here we used the theorem of dominated convergence with majorants k1 k |f | and


kk |g1 | and that
 q



2
2
| ln + |y|2 ((, y) (, y)| 2 |ln | + ln d k1 k
converges to 0 as 0 uniformly in y (1, 1)d1 . One similarly sees that j f = gj
for j = 2, . . . , d.

32

3.1. BASIC PROPERTIES

We next investigate the properties of mollifiers. Besides duality, they are the
basic tool in the study of Sobolev spaces.
1
d
Set (x) = exp 1|x|
2 for |x|2 < 1 and (x) = 0 for |x|2 1, where x R .
2

Observe that C (Rd ). We define


(3.2)

(x) =

1
(x)
kk1

and

1
(x) = d ( x),

for x Rd and > 0. We then have 0 C (Rd ), (x) > 0 if and only if
|x|2 < , = 0 on Rd \B(0, ) and k k1 = 1. For f L1loc (Rd ) and > 0, we now
introduce the mollifier T by
Z
Z
(z)f (x z)dz,
(x y)f (y)dy =
(3.3) T f (x) = ( f ) (x) =
B(0,)

B(x,)

for x Rd , where we have put f (x) := 0 for x Rd \U . From e.g. Example 3.9 and
Proposition 3.10 in [FA], we recall that
T f C (Rd ),

(3.4)
(3.5)

supp T f S := S + B(0, ),

if supp f = S,

(3.6) kT f kLp (U ) kT f kLp (Rd ) k k1 kf kp = kf kp ,


(3.7)

if f Lp (U ), 1 p ,

T f f in Lp (U ) as 0, if f Lp (U ) and 1 p < .

Observe that S is compact if S is compact.


Lemma 3.5. Let K U be compact. Then there is a function Cc (U ) such
that 0 1 on U and = 1 on K. Let f L1loc (U ) satisfy
Z
f dx = 0
U

for all Cc (U ). Then f = 0 a.e.. In particular, weak derivatives are uniquely


defined.
Proof. Assume that f 6= 0 on a Borel set B U with (B) > 0. By Analysis 3,
there is a compact set K B U with (K) > 0. Fix 0 < < 21 dist(K, U ) so
that K2 U . The function := T 1K belongs to Cc (U ) by (3.4) and (3.5), where
supp K2 . Moreover, (3.6) and (3.3) yield that 0 (x) kk k1K k = 1
for all x U and that
Z
(x) =
(x y)1K (y)dy = k k1 = 1
B(x,)

for all x K. This construction shows the first claim.


Since f L1 (U ), the functions T (f ) converge to f in L1 (U ) as 0, due
to (3.7). Hence, there is a nullset N and a subsequence j 0 with j , such

33

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

that (Tj (f ))(x) f (x) 6= 0 as j for each x K\N . For every x K\N
and j N, we also deduce
Z
j (x y)(y)f (y)dy = 0
(Tj (f ))(x) =
U

from the assumption, since the function y 7 j (x y)(y) belongs to Cc (U ). This


is a contradiction.
Recall that H
olders inequality implies that the map
Z
p
p0
f gdx,
(3.8)
L (B) L (B) C, (f, g) 7
B

is continuous for all 1 p and Borel sets B Rd .


Lemma 3.6. a) Let f Lploc (U ) possess the weak derivative f Lploc (U ) for some
Nd0 and p [1, ). Then the functions T f C (U ) converge to f and (T f )
converge to f in Lploc (U ) as 0. We further have (T f )(x) = T ( f )(x)
for all x U and < d(x, U ). Moreover, there is a sequence n 0 such that
Tn f f and (Tn f ) f a.e. on U as n . If f W k (U ), we can take
the same fn for all || k.
b) If f, g L1loc (U ) and there are fn W || (U ) such that fn f and fn g
in L1loc (U ) as n , then g is the weak derivative f . If this convergence holds
in Lp (U ) for some p [1, ] and all with || k, then f Wpk (U ).
Proof. a) Let > 0 and x U . If < d(x, U ), then the function y 7 ,x (y) =
(xy) belongs to Cc (U ). Using a corollary to Lebesgues theorem (see Analysis 3)
and Definition 3.1, we can thus deduce
Z
Z

||
T f (x) =
x (x y)f (y)dy = (1)
( ,x )(y)f (y)dy
U
ZU
=
,x (y)( f )(y)dy = (T f )(x).
U

Choose a compact subset K U and fix > 0 with K U . Take (0, ].


Note that the integrand of (T g)(x) is then supported in K for every x K and
g L1loc (U ), see (3.3). Due to (3.7), the functions

1K (T f ) = 1K T ( f ) = 1K T (1K f )
converge in Lp (K) to 1K 1K f = 1K f as 0. So we have shown the asserted
convergence in Lploc (U ). The sets
Km = {x U : d(x, U )

1
and |x|2 m}
m

S
are compact and mN Km = U . Let k 0. Then, for each m N there is a null
set Nm Km and a subsequence m (j) such that Tm (j) f (x) converges to f (x)

34

3.1. BASIC PROPERTIES

and Tm (j) f (x) converges to f (x) for all x Km \Nm as j . By means of a


diagonal sequence, one obtains a S
sequence n 0 such that S
Tn f (x) f (x)
and Tn f (x) f (x) for x U \( mN Nm ), as n , where mN Nm is a null
set. This procedure can also be done for finitely many f at the same time.
b) Let fn W || (U ) be given such that fn f and fn g in L1loc (U ) as
n . From (3.8) on supp we deduce that
Z
Z
Z
Z

gdx
( fn )dx = (1)
fn dx = (1) lim
f dx = lim
U

for all
obtain f

n U

n U

Cc (U ).
Lp (U )

f .

Hence, g =
In the setting of the last assertion we thus
for all || k, and hence f Wpk (U ).

Proposition 3.7. a) Let f, g W 1 (U ) L (U ). Then, f g W 1 (U ) L (U ) and


(3.9)

j (f g) = (j f )g + f (j g)

holds for all j {1, . . . , d}.


b) Let 1 p , f Wp1 (U ) and g Wp10 (U ). Then, f g W11 (U ) and (3.9)
holds.
Proof. 1) Let f, g W 1 (U ). Set fn = Tn f C (U ) and gn = Tn g C (U ) with
n 0 as in Lemma 3.6a). Fix m N and take Cc (U ) and j {1, . . . , d}.
Choose an open and bounded set V such that supp V V U . Since fn f
and j fn j f on L1 (V ) by Lemma 3.6a), the formulas (3.8) and (3.1) yield
Z
Z
Z
f gm j dx = lim
fn gm j dx = lim
((j fn )gm + fn (j gm )) dx
n V
n V
U
Z
=
((j f )gm + f (j gm )) dx
U

so that the weak derivative j (f gm ) = (j f )gm + f (j gm ) L1loc (U ) exists.


2) Let f, g W 1 (U )L (U ) and gm as in 1). Note that gm g and j gm j g
in L1loc (U ) as m . Since f is bounded, we obtain
Z

Z
Z
Z
f gj dx = lim
f gm j dx = lim
(j f )gm dx +
f (j gm )dx ,
U

m U

R
using Step 1). On the right hand side, the second integral converges to U f j gdx,
again because of f L (U ). For the first integral we use that gm g a.e. by Lemma
3.6a) and that kgm k kgk by (3.6). The theorem of dominated convergence (with
the majorant |j f |kgk kk 1supp ) yields
Z
Z
f gj dx = ((j f )g + f (j g))dx,
U

as required. Note that (j f )g + f (j g) belongs to L1loc (U ) by our assumptions.


3) Let f Wp1 (U ) and g Wp10 (U ). If p (1, ] we show (3.9) as in Step 2), using
0

(3.8) and that gm , j gm converge in Lploc (U ) by Lemma 3.6a). If p = 1, we replace


the roles of f and g. H
olders inequality and (3.9) then yield j (f g) L1 (U ).

35

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

Proposition 3.8. a) Let f W 1 (U ) be real valued and h C 1 (R) with h0 Cb (R).


We then have h f W 1 (U ) and
j (h f ) = (h0 f )j f
for all j {1, . . . , d}.
b) Let f W 1 (U ), V Rd be open and : V U be a diffeomorphism such
that 0 and (1 )0 are bounded. We then have f W 1 (V ) and
j (f ) =

d
X

((m f ) ) j m

m=1

for all j = 1, . . . , d.
In both results we can replace W 1 (U ) by Wp1 (U ) for 1 p , if in a) also
h(0) = 0 holds in the case that (U ) = and p < .
Proof. a) By Lemma 3.6, there are fn C (U ) such that fn f and j fn j f
in L1loc (U ) and a.e. as n , for every j {1, . . . , d}. Since
|h(f (x))| |h(f (x)) h(0)| + |h(0)| kh0 k |f (x)| + |h(0)|
for all x U , the function h f belongs to L1loc (U ) (and to Lp (U ) if f Lp (U ) and
if h(0) = 0 in the case that (U ) = and p 6= ). Let K U be compact. We
obtain that
Z
Z
|h(fn (x)) h(f (x))|dx kh0 k
|fn (x) f (x)|dx 0,
K
K
Z
|h0 (fn (x))j fn (x) h0 (f (x))j f (x)|dx
K
Z
Z
0
kh k
|j fn (x) j f (x)|dx +
|h0 (fn (x)) h0 (f (x))||j f (x)|dx 0
K

as n where we also used Lebesgues theorem and the majorant 2kh0 k |j f |


in the last integral. Since h fn C 1 (U ), (h0 f )j f L1loc (U ) and j (h fn ) =
(h0 fn )j fn , Lemma 3.6b) yields assertion a). If f Wp1 (U ) then (h0 f )j f Lp (U )
and so h f Wp1 (U ).
Assertion b) can be shown similarly using the transformation rule.
Corollary 3.9. Let f W 1 (U ) be real valued. Then f+ , f , |f | W 1 (U ) with
j f = 1{f 0} j f

and

j |f | =

1{f >0} 1{f <0} j f




for all j {1, . . . , d}. Here one can replace W 1 by Wp1 for all 1 p .

Proof. We employ the function h C 1 (R) given by h (t) := t2 + 2 t


for t 0 and h (t) := 0 for t < 0, where > 0. Observe that kh0 k = 1 and

36

3.1. BASIC PROPERTIES

h (t) t for t > 0 and h (t) 0 for t 0, as 0. Proposition 3.8 shows that
h f W 1 (U ) and
Z
Z
Z
f
0
p
h (f )j dx =
h (f )(j f )dx =
(j f )dx
f 2 + 2
U
U
{f >0}
for each Cc (U ). Thanks to the majorants kj k 1B |f | and kk |j f |1B with
B = supp , Lebesgues convergence theorem shows that
Z
Z
Z
f
f+ j dx =
(j f )dx =
1{f >0} (j f )dx.
U
{f >0} |f |
U
There thus exists j f+ = 1{f >0} j f L1loc (U ). The other assertions follow from
f = (f )+ and |f | = f+ + f .
Theorem 3.10. Let J R be an open interval and f L1loc (J). We then have
f W 1 (J) if and only if there is a g L1loc (J) and a continuous representative of f
such that
Z t
(3.10)
f (t) = f (s) +
g( )d
s

for all s, t J. In this case, g = f a.e..


Proof. 1) Let f W 1 (J). Take fn = Tn f C (J) from Lemma 3.6a). Then for
a.e. t J and for a.e. t0 J we have
Z
f (t) f (t0 ) = lim (fn (t) fn (t0 )) = lim
n

n t
0

fn0 ( )d

f ( )d.
t0

Rt
Fixing one such t0 and noting that t 7 t0 f ( )d is continuous, we obtain a continuous representative of f which satisfies (3.10) for s = t0 and g = f . Subtracting
two such equations for any given t, s J and the fixed t0 , we deduce (3.10) with
g = f for all t, s J.
2) If (3.10) holds for some g L1loc (J), take gn C (J) such that gn g in
Rt
L1loc (J) as n . For any s J and n N, the function fn (t) := f (s) + s gn ( )d ,
t J belongs to C (J) with fn0 = gn . Moreover, for [a, b] J with s [a, b], we
estimate
Z bZ t
kfn f kL1 ([a,b])
|gn ( ) g( )|d dt (b a)kgn gkL1 ([a,b]) ,
a

using (3.10), so that fn f in L1loc (J) as n . Lemma 3.6b) then yields


f W 1 (J) and f = g.
Remark 3.11. a) Let J = (a, b) for some a < b in R and f : J C. We then have
f W11 (J) if and only if f is absolutely continuous, i.e.

37

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

> 0 > 0 a < 1 < 1 < < n < n < b, n N with


it holds
n
X
|f (j ) f (j )| .

Pn

i=1 (j j )

j=1

In this case, f is differentiable for a.e. t J and the pointwise derivative f 0 is equal
a.e. to the weak derivative f L1 (J).
(Remark: A Lipschitz continuous function is absolutely continuous; an absolutely
continuous function is uniformly continuous.)
Proof. The implication holds since (3.10) yields that

Z
n
n Z j

X
X


|f (j ) f (j )| =
|f 0 ( )|d =: S,
f 0 ( ) d S

n
j

j=1 (j ,j )
j=1

j=1

where S 0 as (nj=1 (j , j )) 0.
The other implication 00 and the last assertion is shown in Theorem 7.20 of
[Rud87] (combined with our Theorem 3.10).
b) There is a continuous increasing function f : [0, 1] R with f (0) = 0 and
f (1) = 1 such that f 0 (t) = 0 exists for a.e. t [0, 1]. Consequently,
Z
1 = f (1) 6= f (0) +

f 0 ( )d = 0,

and this function is not absolutely continuous, does not belong to W11 ((0, 1)) and
violates (3.10). (See 7.17 in [Rud87].)
1 (U ) is isomorphic to
Proposition 3.12. Let U be convex. Then W

Cb1 (U ) = {f Cb (U ) : f is Lipschitz},
1 (U ) is equivalent to
and the norm of W

kf kC 1 = kf k + [f ]Lip ,
b

where [f ]Lip is the Lipschitz constant of f .


1 (U ). Take 0 from Lemma 3.6a). Let K U be compact.
Proof. Let f W
n
For sufficiently large n N, Lemma 3.6 and (3.6) yield

|j Tn f (z)| = |Tn j f (z)| kj f k kf k1, ,


for all j {1, . . . , d} and z K. Using that Tn f (x) f (x) as n for all
x U \N and a null set N , we thus estimate
|f (x) f (y)| = lim |Tn f (x) Tn f (y)|
n

38

3.2. DENSITY AND EMBEDDING THEOREMS

Z

= lim
n



Tn f (y + (x y)) (x y) d kf k1, |x y|2

for all x, y U \N . Hence, f has a representative with Lipschitz constant kf k1, .


Let f Cb1 (U ). Take Cc (U ), j {1, . . . , d}, and > 0 such that
(supp ) U . For (0, ] the difference quotient 1 ((x + ej ) (x)) converges
uniformly on supp as 0, and hence
Z
Z
1
| f j dx| = lim |
f (x) ((x + ej ) (x))dx|
0 supp

U
Z
1
lim
|f (y ej ) f (y)| |(y)| dy
0 supp
[f ]Lip kk1 .
Taking
into account that Cc (U ) is dense in L1 (U ), we see that the map 7
R
U f j dx has a continuous linear extension Fj : L1 (U ) C. Therefore there is a
function gj L (U ) = L1 (U ) with kgj k = kFj k [f ]Lip such that
Z
Z

f j dx = Fj () =
gj dx
U

for
As

all Cc (U ). This means that f


1 (U ) and kf k
a result, f W
1 (U )
W

has the weak derivative j f = gj L (U ).


kf k + [f ]Lip .

In the above proof convexity is only used to estimate the Lipschitz constant by
kf k . Instead of convexity, it suffices to assume that there exist a > 0 such
that for all x, y U with |x y|2 the line segment from x to y belongs to U .
For such x and y we can argue as above. If |x y|2 and one chooses bounded
representative of f , then one obtains |f (x) f (y)| 2kf k 1 |x y|2 .

3.2

Density and embedding theorems

In this section we prove some of the most important theorems on Sobolev spaces.
Definition 3.13. For k N and 1 p < , the closure of Cc (U ) in Wpk (U ) is
k (U ).
denoted by W
p
Theorem 3.14. Let k N and p [1, ). We then have
pk (Rd ) = Wpk (Rd ).
W
Moreover, the set C (U ) Wpk (U ) is dense in Wpk (U ).
Proof. We prove the theorem only for k = 1, the general case can be treated similarly.
1) Let f Wp1 (Rd ). Take any C (R) such that 0 1, = 1 on [0, 1]
and = 0 on [2, ). Set


1
n (x) =
|x|2
(cut-off function)
n

39

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

for n N and x Rd . We then have n Cc (Rd ), 0 n 1 and kj n k


k0 k n1 for all n N, as well as n (x) 1 for all x Rd as n . Thus
kn f f kp 0 as n by Lebesgues convergence theorem. Further, Proposition
3.7 implies that
kj (n f f )kp = k(n j f j f ) + (j n )f kp
1
kn j f j f kp + k0 k kf kp ,
n
and the right hand side tends to 0 as n for each j {1, . . . , d}. Given > 0,
we can thus fix m N such that km f f k1,p . Due to (3.4) and (3.5), the
functions T 1 (m f ) belong to Cc (Rd ) for all n N . Equation (3.7) and Lemma 3.6
n
further yield that
T 1 (m f ) m f
n

and j T 1 (m f ) = T 1 j (m f ) j (m f )
n

in Lp (Rd ) as n , for each j {1, . . . , d}. So there is an n N such that


kT 1 (m f ) m f k1,p ,
n

and thus
kT 1 (m f ) f k1,p 2.
n

2) For the second assertion, we can assume that U 6= . Let f Wp1 (U ). Set

Un =

1
x U : |x|2 < n and d(x, U ) >
n

S
for all n N. ThenSUn U n Un+1 U , U n is compact and
n=1 Un = U .

Observe that U = n=1 Un+1 \U n1 , where U0 , U1 := .P There are functions


n Cc (U ) such that supp n Un+1 \U n1 , n 0, and
n=1 n (x) = 1 for all
x U (see 2.19 in [Alt06] or Analysis 3).
Fix > 0. As in Step 1), for each n N there is a n > 0 such that gn :=
Tn (n f ) Cc (U ),P
supp gn (supp n f )n Un+1 \U n1 and kgn n f k1,p
n
2 . Define g(x) =
that on each ball B(x, r) U
n=1 gn (x) for all x U . Observe
P
this sum is finite. Hence, g C (U ). Since f =
n=1 n f , we further have
g(x) f (x) =

(gn (x) n (x)f (x)),

n=1

for all x U and n N. Due to kgn n f k1,p 2n , this series converges absolutely
in Wp1 (U ), and

X
kf gk1,p
kgn n f k1,p .
n=1

40

3.2. DENSITY AND EMBEDDING THEOREMS

k (U ) 6= W k (U ), see Lemma 6.67 in


Remark 3.15. a) If U is bounded, then W
p
p
[Ren04].
b) For good U one can replace in C (U ) by C (U ) in Theorem 3.14, see
Corollary 3.22 below.
We now want to study embeddings of Sobolev spaces. We clearly have
Wpk (U ) , Wpj (U )

if k j 0

and
Wpk (U ) , Wqj (U )
Wp0 (U )

if k j 0, 1 q p and (U ) < .

Lp (U )

(Here we put
=
for 1 p .) The embedding X , Y means that
there is an injective map J L(X, Y ). Above it holds Jf = f , and below we also
use Jf = f + N . Writing c = kJk, one obtains kf kY ckf kX if one identifies Jf
with f .
Theorem 3.16 (Sobolev, Morrey). Let k N and 1 p < . We have the
following embeddings.
a) If kp < d, then
p :=

pd
(p, )
d kp

and

Wpk (Rd ) , Lq (Rd )

for all q [p, p ].


b) If kp = d, then
Wpk (Rd ) , Lq (Rd )
for all q [p, ).
c) If kp > d, then there are either j N0 and (0, 1) such that k dp = j +
or k dp N. In the latter case we set j := k dp 1 N0 and take any (0, 1).
Then
Wpk (Rd ) , C0j (Rd ) and | f (x) f (y)| c|x y|2
for all x, y Rd , || j, a constant c > 0, and a representative f C0j (Rd ) of f ,
where
C0j (Rd ) = {u C j (Rd ) : u(x) 0 as |x|2 for all 0 || j}.
Corollary 3.17. Let k N and p [1, ). If there are j N0 and q [p, ) with
k dp = j dq , then
Wpk (Rd ) , Wqj (Rd ).
Observe that Theorem 3.16a) is a special case with j = 0 and q = p .
Proof of Corollary 3.17. By assumption, we have (k j)p = d
rem 3.16a) thus yields
Wpkj (Rd ) , Lq (Rd )

since q =

41

pd
.
d kp + jp

dp
q

(0, d). Theo-

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

Applying this embedding to f Wpkj (Rd ) for all || j and f Wpk (Rd ), we
deduce that
k f kq ck f kkj,p ckf kk,p ,
as asserted.
Example 3.18. There is an unbounded function f Wd1 (Rd ) for d 2, showing
that Theorem 3.16b) is sharp. In fact, for any (0, 1 d1 ) and Cc (Rd ) with
supp B(0, 3/4) and = 1 on B(0, 1/2), we define
(
(x)( ln|x|2 ) , 0 < |x|2 3/4,
f (x) :=
0,
|x|2 > 3/4 or x = 0.
Arguing as in Example 3.4c), one sees that f Lp (Rd ) for all p < , f 6 L (Rd )
and that, for all j {1, . . . , d}, we have
j f (x) = (j (x))( ln|x|2 ) (x)( ln|x|2 )1

xj
|x|22

for 0 < |x|2 < 3/4 and j f (x) = 0 otherwise. Using polar coordinates, we further
estimate
!1/d
Z
1
Z
/d

|j f |d dx

3/4

(ln r)d rd1 dr

ckj k

Rd

Z
+ ckk
0

Z
c+c
0

3/4

3/4

(ln r)(1)d d1
r dr
rd
!1/d

dr
dr
r(ln r)(1)d

!1/d

<

for some constants c > 0, since (1 )d > 1. Hence, f Wd1 (Rd )\L (Rd ).
For the proof of Theorem 3.16 we set x
j = (x1 , . . . , xj1 , xj+1 , . . . , xd ) Rd1 for
d
all x R , j {1, . . . , d} and d 2. We start with a lemma.
Lemma 3.19. Let d 2 and f1 , . . . , fd Ld1 (Rd1 ) C(Rd1 ). Set f (x) =
f1 (
x1 ) . . . fd (
xd ) for x Rd . We then have f L1 (Rd ) and
kf kL1 (Rd ) kf1 kLd1 (Rd1 ) . . . kfd kLd1 (Rd1 ) .
Proof. If d = 2, then Fubinis theorem shows that
Z
Z Z
|f (x)|dx =
|f1 (x2 )||f2 (x1 )|dx1 dx2 = kf1 k1 kf2 k1 ,
R2

as asserted. Assume that the assertion holds for some d N with d 2.


Take f1 , . . . , fd+1 Ld (Rd ) C(Rd ). Write y = (x1 , . . . , xd ) Rd and x =
(y, xd+1 ) Rd+1 . For a.e. xd+1 R, the maps yj 7 |fj (
y j , xd+1 )|d are integrable

42

3.2. DENSITY AND EMBEDDING THEOREMS

on Rd1 for every j {1, . . . , d} due to Fubinis theorem. Fix such a xd+1 R and
write
d
Y

f (y, xd+1 ) :=
fj (
xj ).
j=1

Using H
olders inequality and d0 =
Z

d
d1 ,

|f(y, xd+1 )||fd+1 (y)|dy

|f (y, xd+1 )|dy =


Rd

we obtain

Rd

Z
kfd+1 kLd (Rd )

d0

|f(y, xd+1 )| dy

1/d0
.

Rd

We set gj (
y j ) = |fj (
y j , xd+1 )|d for j {1, . . . , d} and x Rd+1 . Since d0 (d 1) = d,
d1
we have gj L (Rd1 ), and the induction hypothesis yields
Z
Z
d0

|f (y, xd+1 )| dy =
g1 (
y 1 ) . . . gd (
y d )dy kg1 kd1 . . . kgd kd1
Rd

Rd
d Z
Y

1
d1

|fj (
y , xd+1 )| dy

Rd1

j=1

Integrating over xd+1 R, we thus arrive at


Z
|f |dx kfd+1 kd

Z Y
d Z

Rd+1

1 d1
d1 d

|fj (
x )| dy

dxd+1 .

Rd1

R j=1

Applying the d-fold H


older inequality to the xd+1 -integral, we conclude that
Z
|f |dx kfd+1 kd
Rd+1

Z Z
d
Y
j=1

|fj (
xj )|d dy

! d1

 1 d
d

dxd+1

Rd1

= kf1 kd . . . kfd+1 kd .
Recall from Analysis 3 that for f Lp (U ) Lq (U ) and r [p, q] with 1 p <
q , we have
(3.11)

kf kr kf kp kf kq1 kf kp + (1 )kf kq ,

where [0, 1] is given by


Analysis 1 and 2.

1
r

1
q

and we also used Youngs inequality from

Proof of Theorem 3.16. We only prove the case k = 1, the rest can be done by
induction, see e.g. 5.6.3 in [Eva10]. Since Wp1 (Rd ) , Lp (Rd ), estimate (3.11) implies
that for assertion a) it suffices to show

Wp1 (Rd ) , Lp (Rd ).

43

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

d
1) Let f Cc1 (Rd ). Let first p = 1 < d, whence p = d1
. For x Rd and
j {1, . . . , d}, we then obtain
Z
Z xj
j f (x1 , . . . , xj1 , t, xj+1 , . . . , xd )dt| |j f (x)|dxj ,
|f (x)| = |

|f (x)|d

d Z
Y

|j f (x)|dxj .

d
j=1 R

R
1
Setting gj (
xj ) = ( R |j f (x)|dxj ) d1 , we deduce
|f (x)|

d
d1

d
Y

gj (
xj ).

j=1

After integration over x Rd , Lemma 3.19 yields


kf k

d
d1
d
L d1 (Rd )

g1 (
x1 ) . . . gd (
xd )dx

Rd

kf k

d
L d1 (Rd )

d
Y

Z
|j f (x)|dxj d
x

Rd1

j=1

(3.12)

kgj kLd1 (Rd1 )

j=1

d Z
Y

d
Y

1
d1

kj f kLd 1 (Rd ) k|f |1 k1 kf k1,1 .

j=1

pd

2) Next, let p (1, d) and p = dp


. Set t := d1
d p =
d
calculation shows that (t 1)p0 = t d1
= p . Set

g = f |f |t1 = f (f f )

d1
dp p

> 1. An elementary

t1
2

for t > 1. We compute



t1
t1
(f f ) 2 1 (j f )f + f (j f )
2
= j f |f |t1 + (t 1)f |f |t3 Re(f j f ),

j g = j f |f |t1 + f
|g| = |f |t ,

|j g| t|j f ||f |t1 .

Applying (3.12) to g, we estimate


t

kf k

td
d1

Z

d
t d1

|f |

 d1 Z
 d1
d
d
d
dx
=
|g| d1 dx

Rd

d
Y

Rd
1
d

kj gk1

j=1

1
d

Z

Rd

t1

|j f ||f |

1

dx

Rd

j=1

d Z
Y
j=1

d
Y

|j f | dx
| {z }

 1 Z

|f |pp

44

dp

Rd

(t1)p0

|f |


dx

1
p0 d

3.2. DENSITY AND EMBEDDING THEOREMS

d
Y

t1

t1
d
k|f |p kpd kf k(t1)p
0 = tk|f |p kp kf k(t1)p0 ,

j=1

where we used H
olders inequality. For t = t , we use the properties of t stated above
and obtain
kf kp p

(3.13)

d1
d1
k|f |p kp p
kf k1,p .
dp
dp

By density (see Theorem 3.14), this estimate can be extended to all f Wp1 (Rd ).
Then the identity map on Wp1 (Rd ) is the required embedding.
3) Now let f Cc1 (Rd ), p = d, and t > 1. Then p0 =
kf kt

()

1 1

d
d1

t
t t kf k(t1)

d
d1

d
d1 ,


1
k|f |p kdt c kf k(t1)

d
d1

and Step 2) yields


+ k|f |p kd

using Youngs inequality from Analysis 1 and 2. For t = d, this estimate gives
d2

f L d1 (Rd ) and
kf k

d2
d1

ckf k1,d .

d
Here and below the constant c > 0 does not depend on f . For q (d, d d1
),
inequality (3.11) further yields

kf kq c(kf kd + kf k

d2
d1

) ckf k1,d .

Now, we can apply () with t = d + 1 and obtain


kf k d2 +d c(kf k
d1

d2
d1

+ k|f |p kd ) ckf k1,d .

d+1
As above, we see that f Lq (Rd ) for d q d d1
. We can now iterate this
procedure with tn = d + n and obtain

kf kq c(q)kf k1,p
for all q < . Again, b) follows by density.
4) Let f Cc1 (Rd ), p > d, Q(r) = [ 2r , 2r ]d for r > 0, and x0 Q(r). Set
R
M (r) = rd Q(r) f dx. We further put := 1 dp (0, 1). Using |x x0 | r for
x Q(r), the transformation y = t(x x0 ) and Holders inequality, we compute
Z
d
|f (x0 ) M (r)| = |r
(f (x0 ) f (x))dx|
Q(r)

Z
= rd |

Q(r)

rd

1
1

d
f (x0 + t(x x0 ))dtdx|
dt

|f (x0 + t(x x0 )) (x x0 )|dtdx


Q(r)

45

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

r1d

1Z

|f (x0 + t(x x0 ))|1 dxdt


0

=r

1d

Q(r)
1Z

Z
0

1d

|f (x0 + y)|1 dy td dt

t(Q(r)x0 )
p

|f (x0 +
0

cr

!1

1d

t(Q(r)x0 )
1

r p0 t p0

k|f |p kLp (Rd )

y)|p1 dy

vol(t(Q(r) x0 )) p0 td dt

dt

= Cr

1 dp

k|f |p kLp (Rd )

for constants C, c > 0 only depending on d and p, using also that


to p > d. A translation then gives
|f (x0 + z) r

d
p0

d > 1 due

f (y)dy| Cr k|f |p kLp (Rd )

z+Q(r)

for all z Rd . Taking x = z, x0 = 0, r = 1, and using Holders inequality, we thus


obtain
Z
Z
(+) |f (x)| |f (x)
f dy| + |
f dy| Ck|f |p kp + kf kp ckf k1,p
x+Q(1)

x+Q(1)

for all x Rd , where c only depends on d and p. Given x, y Rd , we find a cube Q


of side length |x y| =: r such that x, y Q and Q is parallel to the axes. Hence,
|f (x) f (y)| |f (x) r

Z
f dy| + |r

f dy f (y)| 2Ck|f |p kp |x y|

2Ck|f |p kp |x y|2 .
If f Wp1 (Rd ), then there are fn Cc1 (Rd ) converging to f in Wp1 (Rd ). By (+), fn
is a Cauchy sequence in C0 (Rd ). Hence, f has a representative f C0 (Rd ) such that
fn f uniformly as n . So the above estimates imply that
kfk + sup
x6=y

|f(x) f(y)|
|x y|2

ckf k1,p .

The map f 7 f is the required embedding.


Remark 3.20. The assertions of Theorem 3.16 hold on an open set U instead of Rd
pk . In fact, Theorem 3.16 can be applied to the 0-extension of
if we replace Wpk by W

Cc (U ) and then the result follows by density.


Problem: How to extend the Sobolev embedding to Wpk (U )?

46

3.2. DENSITY AND EMBEDDING THEOREMS

Definition 3.21. Let k N. The open set U Rd has the k-extension property if
for all m {0, 1, . . . , k} and p [1, ), there is an operator
Em,p L(Wpm (U ), Wpm (Rd ))
with Em,p f = f on U and Em,p f = El,q f for all f Wpm (U ) Wql (U ), 1 m, l k
and 1 p, q < . We write EU instead of Em,p .
Observe that E0 f (x) = f (x) for x U and E0 f (x) = 0 for x Rd \U defines an
isometry E0 : Lp (U ) Lp (Rd ). Moreover, PU f = f|U defines a contractive map in
all spaces L(Wpk (Rd ), Wpk (U )).
Corollary 3.22. If U has the m-extension property for some m N, then Theorem
3.16 and Corollary 3.17 hold for k {1, . . . , m} with Rd replaced by U and C0j (Rd )
replaced by
C0j (U ) = {f C j (U ) : f has a continuous extension to U
and f (x) 0 as |x|2 if U is unbounded, for all 0 || j}.
Proof. Consider e.g. Theorem 3.16a). We have the embedding

J : Wpk (Rd ) , Lp (Rd )


given by the identity. Then,

RU JEU : Wpk (U ) Lp (U )
is continuous and injective. The other assertions are proved in the same way.
Corollary 3.23. Let U possess the k-extension property for some k N. Then
Wpk (U ) C (U ) is dense in Wpk (U )
for all 1 p < .
Proof. If f Wpk (U ), then EU f Wpk (Rd ). By Theorem 3.14, there are gn
Cc (Rd ) converging to EU f in Wpk (Rd ). Hence, RU gn Wpk (U ) C (U ) converge
to f = RU EU f in Wpk (U ) as n .
Theorem 3.24. Let U Rd be bounded and open with U C k . Then U has the
k-extension property.
Sketch of the proof for k = 1. (See also Theorem 5.22 in [AdF07].) 1) Let
H = {(y, t) Rd : y Rd1 , t 0}
and f Wp1 (H ) C 1 (H ). Define
(
f (y, t),
E f (y, r) =
4f (y, 2t ) 3f (y, t),

47

(y, t) H ,
(y, t) H+ .

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

Note that E f C 1 (Rd ). One can check that kE f kWp1 (Rd ) ckf kWp1 (H ) for a
constant c > 0.
2) We show that Wp1 (H ) C 1 (H ) is dense in Wp1 (H ), so that E can be
extended to a 1-extension operator on Wp1 (H ). In fact, let f Wp1 (H ) and > 0.
By Theorem 3.14, there is a function g C (H ) Wp1 (H ) with kf gk1,p .
Setting gn (y, t) = g(y, t n1 ) for t 0, y Rd1 and n N, we define the functions
gn C 1 (H ) Wp1 (H ). Observe that
gn = RH Sn E0 g
for 0 || 1, where Sn L(Lp (Rd )) is given by Sn h(y, t) = h(y, t n1 ) for
h Lp (Rd ). One can see that Sn h h in Lp (Rd ) as in Example 3.8 of [FA]. Hence,
gn converges to g in Wp1 (H ) implying the claim.
3) Since U C 1 , there are bounded open sets U0 , U1 , . . . , Um Rd such that
U U0 Um , U 0 U and U U1 Um , as well as a diffeomorphism
0
j : Uj Vj such that 0j and (1
j ) are bounded and j (Uj U ) H and
0j (Uj U ) Rd1 {0}, for each j {1, . . . , m}. Moreover, there are functions
P
0 j Cc (Rd ) with supp j Uj for all j = 0, 1, . . . , m and m
j=0 j (x) = 1 for
all x U (see Analysis 3).
Let j {1, . . . , m}. Set Sj g(y) = g(1
j (y)) for y H Vj and Sj g(y) = 0 for
1
y H \Vj , where g Wp (Uj U ). For h Wp1 (Rd ), set Sj h(x) = h(j (x)) for
x Uj and Sj h(x) = 0 for x Rd \Uj . Take any j Cc (Rd ) with supp j Uj
and j = 1 on supp j (see Lemma 3.5).
Let f Wp1 (U ). Define
EU f = E0 0 f +

m
X



j Sj E Sj R(Uj U ) j R(Uj U ) f .

j=1

Note that supp EU f U0 U1 Um . Using part 2) and Proposition 3.7 and 3.8,
we see that EU L(Wp1 (U ), Wp1 (Rd )). Let x U . If x Uk for some k {1, . . . , m},
we have k (x) H . If x 6 Uj , then j (x) = 0. Thus
X
EU f (x) = 0 (x)f (x) +
j (x)(j f )(1
j (j (x)))
j=1,...,m
xUj

m
X

j (x)f (x) = f (x).

j=0

If x U0 \(U1 Um ), we also have EU f (x) = 0 f (x) = f (x).


We refer to [AdF07] for other extension theorems and a detailed study of the
impact of the regularity of U .
Theorem 3.25 (Rellich-Kondrachov). Let U Rd be bounded and open with U
C k , k N, and 1 p < . Then the following assertions hold.

48

3.2. DENSITY AND EMBEDDING THEOREMS

a) If kp d and 1 q < p =

dp
dkp

(p, ], then the embedding

J : Wpk (U ) , Lq (U )
is compact. (For instance, q = p.)
b) If k

d
p

> j N0 , then the embedding


J : Wpk (U ) , C j (U )

is compact.
Note that a compact embedding J : Y , X means that any bounded sequence
(yn ) in Y has a subsequence such that (Jynj )j converges in X. In Theorem 3.25a),
J is given by the identity, in b) it is given by choosing the representative in C j .
Proof of Theorem 3.25. We prove the result only for k = 1 (and thus j = 0). Part
b) follows from the Arzela-Ascoli theorem since Corollary 3.22 and Theorem 3.16c)
give constants , c > 0 such that |f (x) f (y)| c|x y| and |f (x)| c for all
x, y U and f Wp1 (U ) with kf k1,p 1, where p > d.
In the case p < d, take fn Wp1 (U ) with kfn k1,p M for all n N. Set
gn = EU fn Wp1 (Rd ). From the proof of Theorem 3.24 we know that the support of
Rd containing U . Moreover, kgn k1,p
gn belongs to a fixed open bounded set U
kEU kM =: M1 for all n N. Take q [1, p ) and (0, 1] with 1q = 1 + 1
p .
Inequality (3.11) and Theorem 3.16 yield that
kfn fm kLq (U ) kgn gm kLq (U ) kgn gm kL1 (U ) kgn gm k1

Lp (U
)
1
1
)
+ kgm k1,p
ckgn gm kL1 (U ) (kgn k1,p

2cM11 kgn gm kL1 (U )


for all n, m N. So it suffices to construct a subsequence of gn which converges in
).
L1 (U
, n N and > 0, we compute
For x U
Z
|gn (x) T gn (x)| = |

(x y)(gn (x) gn (y))dy|

RdZ

d
Z
=

1
( (x y))|gn (x) gn (y)|dy

B(x,)
(z)|gn (x) gn (x z)|dz

B(0,1)

Z 0
d
=
(z)|
gn (x tz)dt|dz
B(0,1)
dt
Z
Z

(z)
|gn (x tz) z|dtdz
Z

B(0,1)

49

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

(z)|gn (x tz)|2 dzdt


0

B(0,1)
Z

1
( (x y))|gn (y)|2 td dydt
t
B(x,t)

=
0

kt |gn |2 k1 dt sup kt k1 k|gn |2 k1

0t

ck|gn |p kp cM1 ,
where we have used the transformations z = 1 (x y) and y = x tz, as well as
) , L1 (U
). We thus obtain
Fubinis theorem, Youngs inequality, (3.6) and Lp (U
)M1 =: C
kgn T gn kL1 (U ) c(U

()

for all n N and > 0. On the other hand, the definition of T gn yields
|T gn (x)| k k kgn kL1 (U ) ,

and |T gn (x)| k k kgn kL1 (U )

and n N and each fixed > 0. The Arzela-Ascoli theorem now implies
for all x U
 
for each > 0, and
that the set F := {T gn : n N} is relatively compact in C U
 
) since C U
, L1 (U
). Let > 0 be given and fix = . Then there
thus in L1 (U
2C
are n1 , . . . , nl N such that
F

l
[
j=1

l
[

BL1 (U ) (T gnj , ) =:
Bj .
2
j=1

Hence, given n N, there is a nj such that T gn Bj . The estimates () and (3.6)


then yield
kgn T gnj kL1 (U ) kgn T gn kL1 (U ) + kT (gn gnj )kL1 (U ) C + /2 = .
We have shown that, for each > 0, the set G := {gn : n N} is covered by finitely
). Thus G contains
many open balls Bj of radius /2, i.e., G is totally bounded in L1 (U
1

a subsequence converging in L (U ) (see Satz B.1.7 in [Wer05]). In the case p = d


one simply replaces p by any r (q, ).
Remark 3.26. a) Theorem 3.25 is wrong for unbounded domains. In fact, let
k N, p [1, ) and define the functions fn = f ( n) in Wpk (R) for any function
0 6= f C (R) with supp f (1/2, 1/2). Then kfn kk,p and kfn fm kq > 0 do not
depend on n 6= m in N so that (fn ) is bounded in Wpk (R) and has no subsequence
which converges in Lq with 1 q < p .

b) The embedding Wp1 (U ) , Lp (U ) is never compact, see Example 6.12 in


[AdF07].
2 (U ) or U be bounded with
Proposition 3.27. Let 1 p < and either f W
p
2
2
U C and f Wp (U ). Then there are constants C, 0 > 0 such that

1/p

1/p
d
d
X
X
C
kj f kpp
kij f kpp + kf kp ,

j=1

i,j=1

50

3.2. DENSITY AND EMBEDDING THEOREMS

2 (U ) and for all 0 < 0 if f W 2 (U ).


for all > 0 if f W
p
p
Proof. Let f Cc2 (U ) and extend it to Rd by 0. Take j = 1. Write x = (t, y)
R Rd1 for x Rd . Fix y Rd1 and set g(t) = f (t, y) for t R. Let > 0 and
a, b R with b a = . Take any r (a, a + 3 ) and t (b 3 , b). There there is a
s = s(r, t) (a, b) such that
|g 0 (s)| = |

g(t) g(r)
3
| (|g(t)| + |g(r)|).
tr

For every s (a, b) we thus obtain


0

g 00 ( )d |
Z b
3
|g 00 ( )|d.
(|g(r)| + |g(t)|) +

|g (s)| = |g (s) +

Integrating first over r and then over t, we conclude

Z
Z
0
3 a+ 3
b 00
|g (s)|
|g(r)|dr + |g(t)| +
|g ( )|d,
3
a
3 a
Z a+
Z b
Z
3
2 0
2 b 00
|g ( )|d,
|g (s)|
|g(r)|dr +
|g(t)|dt +
9
9 a
b 3
a
Z
Z b
9 b
|g 0 (s)| 2
|g( )|d +
|g 00 ( )|d
a
a
Z b
Z b
1/p
1/p
1 9 
1 
p
0
0
p 2
|g 00 ( )|p d
|g( )| d
+ p

a
a
 p Z b
1/p
Z b
p1 p1
9
p
00
p
p
|g( )| d +
|g ( )|d
,
2
2
a
a

where we used H
olders inequality first for the integrals and then in R2 . We take
now the p-th power and then integrate over s arriving at

 p Z b
Z b
Z b
9
0
p
p1 p1
p
00
p
|g( )| d +
|g ( )| d .
|g (s)| ds 2
2p a
a
a
Now choose a = ak = k and b = bk = (k + 1) for k Z. Summing the integrals on
[k, (k + 1)) for k Z and then integrating over y Rd1 , it follows that
 p Z

Z
Z
9
0
p
p p1
p
00
p
|g ( )| d 2
|g( )| d + |g ( )| d ,
2p R
R
Z
Z
ZR
36p
p
p
p
(3.14)
|1 f | dx (2)
|11 f | dx +
|f |p dx.
p
(2)
U
U
U
2 (U ). The same result
By approximation, (3.14) can be established for all f W
p
holds for j f and jj f with j {2, . . . , d}. We now replace 2 by , sum over j and

51

CHAPTER 3. SOBOLEV SPACES AND WEAK DERIVATIVES

take the p-th root to arrive at


1/p

1/p
d
d
p
X
X
36
kj f kpp p
kjj f kpp + p kf kpp

j=1

j=1

1/p
d
X
36
kjj f kpp + kf kp ,

(3.15)

j=1

2 (U ), as asserted. If u W 2 (U ) and U is bounded with U C 2 , we


for all f W
p
p
use the extension operator EU L(Wp2 (U ), Wp2 (Rd )) from Theorem 3.24 to deduce
from (3.15) with U = Rd that

d
X
kj f kp p

L (U )

1/p

j=1

1/p
d
X
kj EU f kpLp (Rd )

j=1

1/p
d
X
36
kjj EU f kpLp (Rd ) + kEU f kLp (Rd )

j=1

36
kEU f kWp2 (Rd ) + kEU f kLp (U )

c
ckf kWp2 (U ) + kf kLp (U )

1/p

1/p
d
d
X
X
c
c0
kij f kpp + c1 kj f kpp + kf kp

i,j=1

j=1

where we assume that (0, 1] and the constants c, c0 , c1 do not depend on or f .


Choosing 1 = min{ 2c11 , 1} we arrive at

1/p

1/p
d
d
X
1 X
c
kj f kpp c0
kij f kpp + kf kp
2

j=1

i,j=1

if 0 < 1 . This inequality implies the assertion for U with U C 2 after


replacing by /(2c0 ) and 1 by 0 = c0/c1 .

52

Chapter 4

Selfadjoint operators
4.1

Closed operators and their spectra, revisited

Example 4.1. Let J R be an open interval, X = Lp (J), 1 p < , and


Ak u = k u with D(Ak ) = Wpk (J) for k = 1, 2. By Exercise 11, A1 is closed and
A2 is closed on the domain D2 = {u X : 2 u X}. Exercise 16 further yields
D(A2 ) = D2 .
Example 4.2. Let 1 p < , X = Lp (R) and A = with D(A) = Wp1 (R). For
Re > 0 we set
Z
(R f )(t) =
e(ts) f (s)ds,
t

for t R and f X, see Example 1.21. Let = eRe 1R . Since


Z

|R (f (t))|

eRe (ts) |f (s)|ds = ( |f |)(t)

for all t R, we deduce from Youngs inequality (see e.g. formula (3.1) in [FA]) that
kR f kp k k1 kf kp =

1
kf kp .
Re

Moreover, a slight variant of Theorem 3.10 shows that R f W 1 (R) with R f =


R f f Lp (R). As a result, R f Wp1 (R) = D(A) , C0 (R) and (I A)R = I.
Next, let C and suppose that there would exists some u D(A) and u = Au =
u. Then Theorem 3.10 yields that
1
1
(u(t) u(s)) =
ts
ts

Z
s

1
u( )d =
ts

u( )d u(s)
s

as t s, i.e., u is differentiable with the continuous derivative u. (Observe that


the above argument shows that if u W 1 (J) and u is continuous, then u C 1 (J)
for any open interval J R.) Consequently, u is equal to a multiple of e . Since
e 6 X, we obtain u = 0 and so I A is injective. Summing up, we have shown

53

CHAPTER 4. SELFADJOINT OPERATORS

that p (A) = and (A) with R(, A) = R if Re > 0. In the same way, one
sees that (A) if Re < 0, where the resolvent is given by
t

e(ts) f (s)ds,

R(, A)f (t) =

for t R and f X. Now, let iR. We take n Cc1 (R) with n = n1/p on
[n, n], n (t) = 0 for |t| n + 1 and kn k 2n1/p , for every n N. We then
obtain un = n e Wp1 (R),
Z

1/p

n
p

kun kp

|n (t)e (t)| dt

= n /p (2n) /p = 2 /p ,
1

kun Aun kp = k0n e kp =


2n

1/p

Z
{n|t|n+1}

|0n (t)e (t)|p dt

1/p

1/p

2 ,

for every n N. As a result, ap (A) and so (A) = iR.


Example 4.3. Let X = Lp (R), 1 p < , and B = 2 with D(B) = Wp2 (R). For
6 R there is a C with Re > 0 such that 2 = . Using A from Example
4.2, we obtain
()

(I B)u = (I A)(I + A)u

for u Wp2 (R). Since I A and I + A = (I A) are invertible by Example


4.2, we infer that
u = (I + A)1 (I A)1 (I B)u
for all u D(B). On the other hand, for v Wp1 (Rd ) it holds that
(I + A)1 v = A(I + A)1 v = (I + A)1 v + v Wp1 (R).
This fact implies that (I + A)1 Wp1 (R) Wp2 (R). Therefore we can choose u =
(I + A)1 (I A)1 f for any f X in (), arriving at
(I B)(I + A)1 (I A)1 f = f
for all f X. Hence, (B) and R(, B) = (I + A)1 (I A)1 , so that
(B) R . For 0, we have = 2 for some iR. Then I A is not
surjective (since its range is not closed by Example 4.2). Equation () thus implies
that I B is not surjective. As a result, (B) = R = (A2 ).
Definition 4.4. A closed operator A in X has a compact resolvent if there exists a
(A) such that R(, A) L(X) is compact.

54

4.1. CLOSED OPERATORS AND THEIR SPECTRA, REVISITED

Remark 4.5. a) If A has a compact resolvent R(, A) and (A), then


R(, A) = R(, A) + ( )R(, A)R(, A)
by the resolvent equation, so that R(, A) is also compact due to Proposition 2.3.
b) For a closed operator A with (A) the following assertions are equivalent,
where [D(A)] = (D(A), kkA ).
1) A has a compact resolvent,
2) Each bounded sequence in [D(A)] has a subsequence which converges in X,
3) The inclusion map J : [D(A)] X is compact.
Proof. 1) 2): Let xn D(A) with kxn kA c for all n N. Set yn = xn Axn .
Then kyn k (|| + 1)c for every n N so that xn = R(, A)yn has a subsequence
which converges in X.
2) 3): This implication is clear, see Remark 2.2.
3) 1): Define R L(X, [D(A)]) by R x = R(, A)x for x X. The
operator R(, A) = JR : X X is then compact due to Proposition 2.3.
c) If dim X = , then a closed operator A with compact resolvent R(, A) cannot
be bounded since otherwise I A were bounded and thus I = (I A)R(, A) were
compact by Proposition 2.3.
Theorem 4.6. Let dim X = and A be a closed operator with compact resolvent.
Then (A) is either empty or contains only at most countably many eigenvalues n
with dim N (n I A) < . If A has infinitely many eigenvalues n , then |n |
as n .
Proof. Fix a (A). By Theorem 2.11, the spectrum (R(, A)) only contains 0
and either no or finitely many eigenvalues n or a nullsequence of eigenvalues n .
Moreover, dim N (n I R(, A)) < for all n.
By Proposition 1.20, we have ((A))1 = (R(, A))\{0} and (p (A))1 =
p (R(, A))\{0}. As a result, the spectrum of A consists only of the eigenvalues
n := 1n (if (R(, A))\{0} =
6 ). If there are infinitely many n , then n 0
and hence |n | as n . Since
n I A = (n I R(, A))

1
(I A),
n

linearly independent eigenvectors of n I A give linearly independent eigenvectors


1
n (I A)v of n I R(, A), so that the kernel of n I A is finite dimensional.
Example 4.7. Let U Rd be open and bounded with U C 1 if 1 p < . Let
X = Lp (U ) for p [1, ) and X = C(U ) for p = . Let A be a closed operator
such that [D(A)] , Wp1 (U ) if 1 p < and [D(A)] , C 1 (U ) if p = . Theorem

55

CHAPTER 4. SELFADJOINT OPERATORS

3.25 of Rellich-Kondrachov and the theorem of Arzela-Ascoli (see e.g. Theorem 1.45
in [FA]) show that Wp1 (U ) and C 1 (U ) are compactly embedded in X, respectively,
so that the embedding [D(A)] , X is compact as a product of a bounded and a
compact operator. By Remark 4.5, A has a compact resolvent in X if (A) 6= .
We stress that for Lp (0, 1) and C([0, 1]) the operators Ak = k with domains
Wpk (0, 1) and C k ([0, 1]), respectively, for k = 1, 2 have empty resolvent sets since
e D(Ak ) is an eigenfunction for the eigenvalue if k = 1 and 2 if k = 2, for
every C.
To study boundary values of functions in Wp1 (U ), we need the following result
proved e.g. in 5.5 of [Eva10].
Theorem 4.8 (Trace theorem). Let U Rd be bounded and open with U C 1 ,
and let p [1, ). Then the trace map f 7 f|U from Wp1 (U ) C(U ) to Lp (U, d)
p1 (U ).
has a bounded linear extension T : Wp1 (U ) Lp (U, d) whose kernel is W
Proof. 1) Let u C 1 (U ). By the definition of the surface integral, see Analysis 3,
ej Vej and j C 1 (U
ej ) with 0
there are finitely many diffeomorphisms j : U
c
p
j 1 such that kukLp (U ) is dominated by
c

m Z
X
j=1

Vj

1 p
0
j 1
j |u j | dy

ej and Vej are open subsets of Rd , the sets U


ej cover U , the functions j
where U
ej , Vj := {(y 0 , yd ) Vej : yd = 0}, Vj+ :=
form a partition of unity subordinated to U
ej U ) = Vj , and j (U
ej U ) = Vj+ . We set
{(y 0 , yd ) Vej : yd > 0}, j (U
1
1 e
v = u 1
j and = j j Cc (Vj ) and drop the indices j below. By means of
Fubinis theorem and the fundamental theorem of calculus, we compute
Z
Z
Z
|v(y 0 )|p dy 0 =
d ( |v|p ) dy =
[(d ) |v|p + p|v|p2 Re(vd v)] dy
V
V
V+
Z +
c
[|v|p + |v|p1 |d v|] dy c kvkpp + c kvkp1
kd vkp
p
V+

c (kvkpp + kd vkpp ) c kvkpW 1 (V+ ) c kukpW 1 (U ) .


p

Here we also used H


olders and Youngs inequality, Proposition 3.8 and the transformation rule. As a result, the map T : (C 1 (U ), k k1,p ) Lp (U, d), T u = u|U ,
is continuous. Corollary 3.23 and Theorem 3.24 allow to extend T to an operator in
L(Wp1 (U ), Lp (U, d)). If we start with a u Wp1 (U ) C(U ), then we can define
an approximating sequence of un C 1 (U ) which converge to u in Wp1 (U ) and in
C(U ), see the proof of Theorem 5.3.3 in [Eva10]. Hence, T un = un |U tends to
u|U uniformly on U and to T u in Lp (U, d), so that T u = u|U .
1 (U ) N (T ) is a consequence of the
2a) We next observe that the inclusion W
p
p1 (U ) by
continuity of T since T vanishes on Cc (U ) and this space is dense in W

56

4.1. CLOSED OPERATORS AND THEIR SPECTRA, REVISITED

definition. To prove the converse, we start with the model case that v Wp1 (V+ )
has a compact support in V+ and T v = 0. Our density results yield vn C 1 (V+ )
converging to v in Wp1 (V+ ), and hence T vn = vn |V 0 in Lp (V ), as n .
Observe that
Z yd
|d vn (y 0 , s)| ds,
|vn (y 0 , yd )| |vn (y 0 , 0)| +
0
p
 Z yd
0
0
p
0
p
|d vn (y , s)| ds
|vn (y , yd )| 2 |vn (y , 0)| + 2
0

for y 0 V and yd 0. Integrating over y 0 und employing Holders inequality, we


obtain
Z
Z Z yd
Z
p1
0
p
0
0
p
0
|d vn (y 0 , s)|p ds dy 0 .
|vn (y , 0)| dy + 2xd
|vn (y , yd )| dy 2
V

We can now let n and arrive at


Z Z
Z
p1
0
p
0
|v(y , yd )| dy 2xd
(4.1)
V

yd

|d v(y 0 , s)|p ds dy 0 .

We next use a cutoff argument to obtain a support in the interior of V+ . Choose a


function C (R+ ) such that = 0 on [0, 1] and = 1 on [2, ). Set n (s) =
(ns) for s 0 and n N, and define wn = n v on V+ . Observe that wn v
in Lp (V+ ) as n . It holds j wn = n j v for j = 1, . . . , d 1 and d wn =
n d v + n0 (n)v. Estimate (4.1) further implies
Z
|wn
V+

v|pp dy

Z
c
0

Z
c
0

Z
c
0

2/nZ

|1 n |

|v|pp dy 0 ds

+ cn

V
2/nZ

|v|pp dy 0 ds

V
2/nZ

+ cn

|v|pp dy 0 ds + c

Z
0

2/nZ

2/n

p1

0
sZ

s
0
2/n Z

|v(y 0 , s)|p dy 0 ds

|d v(y 0 , )|p dy 0 d ds

|d v(y 0 , )|p dy 0 d 0

for some constants c > 0. Because of v Wp1 (V+ ), the above integrals tend to 0 as
n , and so wn v in Wp1 (V+ ) as n . Since wn = 0 for yd (0, 1/n], we
can mollify wn to obtain a function w
bn Cc (V+ ) such that kw
bn wn k1,p 1/n.
1
This means that w
bn v in Wp (V+ ) as n .
ej and Vej and the functions
2b) We come back to u W 1 (U ) and consider the sets U
p

j and j from step 1). Let vj = (j u) (1


j ). First, observe that the trace of vj
1
to the set Vj is given by (T u) j if u C(U ), in addition. By continuity one can
extend this identity to all u Wp1 (U ). Let T u = 0. Then we can apply part 2a) to
vj and obtain w
bnj Cc1 (Vj+ ) converging to vj in Wp1 (Vj+ ). The function
u
bn =

m
X

ej )
w
bnj (j |U U

j=1

57

CHAPTER 4. SELFADJOINT OPERATORS

thus belongs to Cc1 (U ) and converges to u in Wp1 (U ) as n . Since u


bn has compact
support, we can mollify u
bn to a function un Cc (U ) with kb
un un k1,p 1/n. This
p1 (U ).
means that un u in Wp1 (U ) as n , and hence u W
1 (U ) is 0 at the boundary in the sense of trace.
We can thus say that u W
p
Observe that for a Borel set U we can define the trace of u Wp1 (U ) on
by T u = R T u, where R : Lp (U, d) Lp (, d) is the contractive restriction
map. Finally, for p > d we have Wp1 (U ) , C(U ) by Theorem 3.24 and Corollary
3.22 so that Theorem 4.8 is trivial in this case. Theorem 3.10 further implies that
W11 (a, b) , C([a, b]) for d = 1.
Example 4.9. Let X = Lp (0, 1), 1 p < , and A = with
D(A) = {u Wp1 (0, 1) : u(0) = 0}.
If u Wp1 (0, 1) and u = u for some C, then u = e (cf. Example 4.2). Since
e 6 D(A), we obtain p (A) = . Further, the Volterra operator V given by
Z t
V f (t) =
f ( )d
0

for f X and t [0, 1] maps into D(A) and AV f = f , thanks to Theorem 3.10.
Hence, A has the bounded inverse V and is thus closed due to Lemma 1.5. Theorem
4.6 and Example 4.7 now yield that (A) = .
Example 4.10. Let X = Lp (0, 1), 1 p < , and A = 2 with
p1 (0, 1).
D(A) = Wp2 (0, 1) W
(A is the one-dimensional Dirichlet-Laplacian.) Let un D(A) and u, v X be
given such that un u and 2 un v in X as n . Example 4.1 shows that
1 (0, 1) since W
1 (0, 1) W 2 (0, 1)
un u in Wp2 (0, 1) and 2 u = v. Moreover, u W
p
p
p
is closed in Wp2 (0, 1). As a result, u D(A) and Au = v so that A is closed.
Suppose that = 2 is an eigenvalue of A with eigenfunction 0 6= u D(A).
Then u Wp1 (0, 1) , C([0, 1]) by Theorem 3.10, so that u C 1 ([0, 1]) (cf. Example
4.2). Moreover, 2 u = u C([0, 1]) and so u C 2 ([0, 1]). It follows that u =
ae + be 6= 0, for some a, b C. The boundary conditions u(0) = u(1) = 0 imply
that a + b = 0 and ae + be = 0, which yields a = b 6= 0, e = e and 6= 0.
Hence, Re = Re , so that iR\{0}. We thus obtain e = e and |e | = 1,
which gives e {1} and so i(Z\{0}). Hence, = 2 = 2 k 2 for some k N.
Conversely, = 2 k 2 is an eigenvalue with eigenfunction
u(t) = exp(ikt) exp(ikt) = 2 sin(kt).
Given f X we set
Z
Z t
Z
Z
Z
1
1 t
t 1
1 1
1 1
|ts|f (s)ds = t f (s)ds
sf (s)ds
f (s)ds+
sf (s)ds
Rf (t) =
2 0
2 0
2 0
2 t
2 t

58

4.1. CLOSED OPERATORS AND THEIR SPECTRA, REVISITED

for t [0, 1]. As in Example 4.2 we see that Rf W 1 (0, 1) Lp (0, 1) and
1
Rf (t) =
2

1
f (s)ds
2

f (s)ds
t

for t [0, 1]. Consequently, Rf W 2 (0, 1) and 2 Rf = f Lp (0, 1). Thus Rf


Wp2 (0, 1). Since affine functions h satisfy h00 = 0, the function
R0 f (t) = Rf (t) (1 t)Rf (0) tRf (1),
1 (0, 1) and AR0 f = f . Hence, R0
for t (0, 1), belongs to D(A) = Wp2 (0, 1) W
p
is the inverse of A. Theorem 4.6 and Example 4.7 now yield that (A) = p (A) =
{ 2 k 2 : k N}.
Example 4.11. Let X = L2 (Rd ) and A = 4 = 11 +. . .+dd with D(A) = W22 (Rd ).
Then A is closed.
Proof. Let u Cc3 (Rd ) and k, l {1, . . . , d}. Integration by parts yields
Z
Z
Z
2
kkl uk2 =
kl ukl udx =
kkl ul udx =
kk ull udx,
Rd

d
X

kkl uk22 =

k,l=1

Rd

d
X

Z
Rd

k=1

kk u

d
X

Rd

Z
ll udx =
Rd

l=1

4u4udx = k4uk22 .

By Exercise 10, we have

kuk22,2 c0 kuk22 +

d
X

kkl uk22

k,l=1

for some constant c0 > 0, so that the above equality yields



()
kuk22,2 c0 kuk22 + k4uk22 2c0 kuk2A
for all u Cc3 (Rd ). For v W22 (Rd ) we further have
kvk2A

kvk22

d
X

!2
kkk vk2

kvk22 + d

k=1

d
X

kkk vk22 dkvk22,2 ,

k=1

by Cauchy-Schwarz inequality in Rd . If u W22 (Rd ), then there are un Cc (Rd )


converging to u in W22 (Rd ) and thus also for kkA by the above estimate. Using also
(), we obtain

kuk22,2 = lim kun k22,2 2c0 lim kun k2A = 2c0 kuk2A .
n

As a result, the graph norm of A is equivalent to the complete norm kk2,2 on D(A) =
W22 (Rd ). So (D(A), kkA ) is a Banach space and thus A is closed by Lemma 1.5.

59

CHAPTER 4. SELFADJOINT OPERATORS

Example 4.12. Let 1 < p < , X = Lp (Rd ), and A = 4 with D(A) = Wp2 (Rd ).
The Calder
on-Zygmund estimate says that the graph norm of A is equivalent to

kk2,p on Cc (Rd ), see e.g. Corollary 9.10 in [Gil98]. As in Example 4.11 one then
concludes that A is closed.
There are related results for X = L1 (Rd ) and X = C0 (Rd ), see Theorem 5.8 in
[Tan97] and Theorem 31.7 in [Lun95]. In theses case p {1, }, the descriptions
of the domains are more complicated and not just Sobolev (or C k -) spaces. To
2 (B(0, 1)) such that
indicate the difficulties, we note that there is a function u 6 W
4u L (B(0, 1)), namely
(
(x2 y 2 ) ln(x2 + y 2 ) for (x, y) 6= (0, 0),
u(x, y) =
0
for (x, y) = (0, 0).
Then the second derivative
xx u(x, y) = 2 ln(x2 + y 2 ) +

4x2
(6x2 2y 2 )(x2 + y 2 ) 4x2 (x2 y 2 )
+
2
+y
(x2 + y 2 )2

x2

y
is unbounded around (0, 0), but 4u(x, y) = 8 xx2 +y
2 is bounded.

Example 4.13. Let 1 < p < , U Rd be bounded and open with U C 2 ,


X = Lp (U ), and A = 4 with
p1 (U )
D(A) = Wp2 (U ) W
(Dirichlet-Laplacian). Then A is closed and has a compact resolvent. In fact,
the closedness can be proved as in Example 4.10, using Example 4.12. In view of
Example 4.7 it suffices to show that A is bijective. This is done e.g. in Theorem 9.15
of [Gil98].

4.2

Selfadjoint operators and their spectra

Let X and Y be Hilbert spaces with scalar product (|) . Let T L(X, Y ) and A be
a linear densely defined operator from X to Y . We define the Hilbert space adjoints
T 0 and A0 as in the Banach space case by


for all x X, y Y :
x T 0 y = (T x|y) ,
D(A0 ) = {y Y : z X x D(A) : (Ax|y) = (x|z) },

A0 y := z

As in Remark 1.23 one sees that A0 is well-defined, linear and closed. We say that
T is unitary if T is invertible with T 1 = T 0 . If X = Y , we call T selfadjoint if
T = T 0 and normal if T T 0 = T 0 T . Note that if T is unitary or selfadjoint, then
T and I T are normal for all C. Also, a densely defined linear operator
is selfadjoint if A = A0 (in particular, D(A) = D(A0 ) and A must be closed) and
normal if AA0 = A0 A.

60

4.2. SELFADJOINT OPERATORS AND THEIR SPECTRA

Let : X X be the Riesz isomorphism given by ((x))(y) = (y|x) for all


x, y X. For C and X = Y , we then obtain
IX T = (IX T 0 )1 ,
IX A = (IX A0 )1 ,
with D(A ) = D(A0 ), see 4.4 in [FA]. So Theorem 1.24 implies that
(4.2)

(A) = (A ) = (A0 )

and r (A) = p (A ) = p (A0 ),

where the bars mean complex conjugation. Moreover,


R(, A0 ) = 1 R(, A ) = 1 R(, A) = R(, A)0 ,
for all (A).
Recall that kT 0 k = kT k for every T L(X, Y ), T 00 = T and (ST )0 = T 0 S 0 for
S L(Y, Z).
Proposition 4.14. Let X and Y be Hilbert spaces. If T L(X, Y ), then kT 0 T k =
kT T 0 k = kT k2 . Let T L(X) be normal. We then have kT k = r(T ), and thus T = 0
if (T ) = {0}.


0 1
2
has the spectrum (T ) =
Remark. In X = C , the Jordan matrix T =
0 0
{0}, but kT k = 1. The equality kT k = r(T ) will be the key to the deeper properties
of selfadjoint (or normal) operators.
Proof of Proposition 4.14. For x X we have
kT xk2 = (T x|T x) = T 0 T x|x) kT 0 T kkxk2 ,
kT k2 = sup kT xk2 kT 0 T k kT 0 kkT k = kT k2 .
kxk1

Hence, kT k2 = kT 0 T k. Similarly, one obtains kT k2 = kT 0 k2 = kT T 0 k.


Next, let T be normal. From the first part we then deduce that
kT 2 k2 = kT 2 (T 2 )0 k = kT T T 0 T 0 k = kT T 0 (T T 0 )0 k
= kT T 0 k2 = kT k4 ,
n

so that kT 2 k = kT k2 . Iteratively it follows that kT 2 k = kT k2 for all n N. Using


Theorem 1.16, we conclude that
n

r(T ) = lim kT m k /m = lim kT 2 k2


1

= kT k.

Definition 4.15. Let A and B be linear operators from a Banach space X to a


Banach space Y . We say that B extends A (and write A B) if D(A) D(B) and
Ax = Bx for all x D(A).
Next, let X be a Hilbert space. A linear operator A on X is called symmetric if
we have (Ax|y) = (x|Ay) for all x, y D(A).

61

CHAPTER 4. SELFADJOINT OPERATORS

Remark 4.16. a) We have A B for linear operators if and only if gr(A) gr(B).
It then holds A = B if and only if D(B) D(A).
b) Let A B for linear operators on the Banach spaces X and Y . If A is
surjective and B is injective, then A = B. As a consequence, if X = Y and there is
a C such that I A is surjective and I B is injective, then A = B.
Proof. Let x D(B) and set y = Bx. Then there is a z D(A) such that Bx =
y = Az = Bz. Since B is injective, we obtain x = z D(A) so that A = B.
c) If A is densely defined and symmetric on a Hilbert space X, then A A0 and
A is selfadjoint if D(A0 ) D(A).
Lemma 4.17. Let A be symmetric and , R. Set = + i. For x D(A) it
holds (Ax|x) R and
kx Axk2 = kx Axk2 + ||2 kxk2 ||2 kxk2 .
In particular, if A is closed, then ap (A) R. If (A)\R, then kR(, A)k
1
|Im | .
Proof. For x D(A) we have (Ax|x) = (x|Ax) = (Ax|x) so that (Ax|x) = (x|Ax)
is real. From this fact we deduce that
kx Axk2 = (x Ax + ix|x Ax + ix)
= kx Axk2 + 2 Re (ix|x Ax) + kixk2
= kx Axk2 + 2 Re(ikxk2 i (x|Ax) ) + ||2 kxk2
= kx Axk2 + ||2 kxk2 ||2 kxk2 .
In particular, 6 ap (A) if Im = 6= 0.
If (A)\R and y X, write x = R(, A)y D(A). We then calculate
kyk2 = kx Axk2 |Im |2 kxk2 = |Im |2 kR(, A)yk2
which yields the final inequality in the lemma.
Theorem 4.18. Let X be a Hilbert space and A be densely defined, closed and
symmetric.
a) Then (A) is either a subset of R or (A) = C or (A) = { C : Im 0}
or (A) = { C : Im 0}.
b) The following assertions are equivalent.
(1) A = A0 ,
(2) (A) R,
(3) iI A0 and iI + A0 are injective,

62

4.2. SELFADJOINT OPERATORS AND THEIR SPECTRA

(4) (iI A)D(A) and (iI + A)D(A) are dense.


c) Let A be selfadjoint. Then we have
kR(, A)k

(4.3)

1
|Im |

for 6 R. Further, (A) = ap (A) and A has no selfadjoint extension B 6= A.


Proof. a) Suppose that there would exists (A) and (A) with Im > 0
and Im > 0. The line segment from to must contain a point (A). Then,
Im > 0 and ap (A) by Proposition 1.19, which contradicts Lemma 4.17 since
A is symmetric. A similar fact holds if Im < 0 and Im < 0. Assertion a) thus
follows.
b) Let A be selfadjoint. Lemma 4.17 yields ap (A) R. Due to (4.2) we also
have r (A) = p (A0 ) = p (A). Hence, r (A) = p (A) R. From Proposition 1.19
we thus deduce (A) R; i.e., (1) implies (2). The implication (2) (3) is obvious.
Equation (4.2) also shows that i p (A0 ) if and only if i r (A) so that (3) and
(4) are equivalent. Let (4) (and thus (3)) hold. Due to Lemma 4.17 and Proposition
1.19, the range of iI A is closed. In view of (4), iI A is then surjective. Due to
(3), iI A0 is injective, and hence A = A0 thanks to Remark 4.16b). We also have
verified the implication (4) (1).
c) Let A = A0 . Then (A) = (A0 ) R so that assertion c) follows from Lemma
4.17 and Remark 4.16.
1 (0, ). For
Example 4.19. a) Let X = L2 (0, ) and A = i on D(A) = W
2

u D(A) and v Cc (0, ), we have


Z
Z
Z
0
0
(Au|v) = i
(u)vdx = i
uv dx =
uiv dx = (u|Av) .
0

For v D(A) there are vn Cc (0, ) converging to v in W21 (0, ), so that vn v


and Avn Av in X as n . Hence, (Au|v) = (u|Av) for all u, v D(A), i.e.,
A is symmetric. By the exercises, (iA) = { C : Re 0} so that (A) =
i(iA) = { C : Im 0} (use Exercise 4). Consequently, A is not selfadjoint.
b) Let X = L2 (R) and A = i with D(A) = W21 (R). As in part a) one shows
that A is symmetric. Example 4.2 implies that (A) = i(iA) = i2 R = R. Hence,
A is selfadjoint.
c) Let X = L2 (R) and A = 2 with D(A) = W22 (R). For u D(A) and
v Cc (R) we have
Z
Z
2
2
(Au|v) = ( u)v = (1)
uv 00 = (u|Av) .
R

If v D(A) there exist some vn Cc (R) with vn v and Avn = vn00 2 v = Av


in X as n . Hence (Au|v) = (u|Av) and A is symmetric. Moreover, (A) = R
by Example 4.3 so that A is selfadjoint.

63

CHAPTER 4. SELFADJOINT OPERATORS

Lemma 4.20 (Greens formula). Let U Rd be bounded and open with U C 2 ,


1 < p < , u Wp2 (U ) and v Wp20 (U ). We then have
Z
Z
Z
Z
u vd.
( u)vd
u4vdx +
(4u)vdx =
U

(In the boundary integrals we write


P u and v instead of T u and T v, is the outer
unit normal of U , and u = dj=1 j T j u.)
Proof. Greens formula holds for u, v C 2 (U ) by Analysis 3. For u Wp2 (U ) and
v Wp20 (U ) there are un , vn C 2 (U ) such that un u in Wp2 (U ) and vn v in
Wp20 (U ) as n , due to Corollary 3.23 and Theorem 3.24. Theorem 4.8 further
0
yields that un |U T u in Lp (U, d) and vn |U T v in Lp (U, d). Since
j : Wq2 (U ) Wq1 (U ) is continuous for q = p, p0 and j {1, . . . , d}, we further obtain
0
j un |U T j u in Lp (U, d) and j vn |U T j v in Lp (U, d) as n . The
assertion now follows by approximation.
We note that the result is also true for p {1, }, with a modified proof.
Example 4.21. a) Let U Rd be open and bounded with U C 2 and let A1 = 4
1 (U ). For u, v D(A1 ) we have T u = T v = 0 so that
with D(A1 ) = W22 (U ) W
2
Lemma 4.20 yields
Z
Z
(A1 u|v) = (4u)v =
u4v = (u|A1 v) ;
U

i.e., A1 is symmetric. In Example 4.13 we have seen that 0 (A1 ) (see also Example
4.10 for d = 1). So A1 is selfadjoint due to Theorem 4.18.
2 (0, 1) and A1 from part
b) On X = L2 (0, 1) consider A0 = 2 with D(A0 ) = W
2
a) with U = (0, 1). As in a) we see that A0 is symmetric. Since A0 $ A1 and
(A1 ) R, Remark 4.16 implies that C\R (A0 ). Hence, A0 is not selfadjoint
and (A0 ) = C.
Example 4.22. Let X = L2 (Rd ) and A = 4 with D(A) = W22 (Rd ). For u, v
Cc (Rd ) we have (Au|v) = (u|Av) . Since Cc (Rd ) is dense in W22 (Rd ), we obtain
as before the symmetry of A. To compute the spectrum of A we use the Fourier
transform F given by
Z
d
u
() = Fu() = (2) /2
eix f (x)dx,
Rd

for Rd and f L1 (Rd ). It is known that F can be extended from L1 (Rd )L2 (Rd )
to a unitary operator F on L2 (Rd ). Let u Cc (Rd ). It then holds
Fi j u = i j u
,

(4.4)

see e.g. 5 in [FA]. These facts imply


kuk22,2 ' kuk22 +

d
X

ki j uk22 = k
uk22 +

i,j=1

d
X

kFi j uk22

i,j=1

64

4.2. SELFADJOINT OPERATORS AND THEIR SPECTRA

(4.5)

Z 

|
u()|2 +

Rd

d
X


|i j |2 |
u()|2 d ' k(1 + ||22 )
uk22 ,

i,j=1

where a ' b means that c1 a b c2 a for some constants c1 , c2 > 0 and all a, b.
Let u W22 (Rd ). There are un Cc (Rd ) converging to u in W22 (Rd ). Then
u
n u
in L2 (Rd ), and by (4.5), the sequence ((1 + ||22 )
un )n is Cauchy in L2 (Rd ).
2
2
Hence, the functions (1 + ||2 )
un converge to (1 + ||2 )
u in L2 (Rd ) as n .
Consequently, (4.5) and (4.4) hold on W22 (Rd ) by approximation. If u L2 (Rd )
satisfies ||22 u
L2 (Rd ), we have for Cc (Rd ) that
Z
ui j dx = (u|i j ) = (Fu|Fi j ) = (
u|i j F)
Rd
Z
1
F 1 (i j u
)dx.
)|) =
= (i j u
|F) = F (i j u
Rd

In other words, there exists the weak derivatives i j u = F 1 (i j u


) L2 (Rd )
2
d
for all i, j {1, . . . , d}, and thus u W2 (R ). We have shown the important
characterization
W22 (Rd ) = {u L2 (Rd ) : ||22 u
L2 (Rd )}
for the Sobolev space with p = 2.
Let f L2 (Rd ) and C. By the above results, u L2 (Rd ) belongs to W22 (Rd )
and satisfies u 4u = f on Rd if and only if ||2 u
L2 (Rd ) and ( + ||22 )
u = f
1
d

on R . The latter implies that u


= +||2 f =: mf . If 6 R , then m L (Rd )
 2
||22
1
2
d

f
belongs to L2 (Rd ), ||22 u
= +||
so that the function u := F 1 +||
2
2 f L (R )
2
2
and ( + ||2 )
u = f. As a result, (A) and (A) R .
2

If < 0, then m is unbounded. Suppose that mg L2 (Rd ) for all g L2 (Rd ).


Then the (closed) multiplication operator M : g 7 mg in L2 (Rd ) were defined on
L2 (Rd ) and thus bounded by the closed graph theorem. But then (M ) = m(Rd )
would be bounded, which is wrong. So there is an f L2 (Rd ) such that mf 6 L2 (Rd )
and so (A). Summing up, (A) = R .
Theorem 4.23. Let A be densely defined and selfadjoint on the Hilbert space X
and let B be symmetric with D(A) D(B). Assume there are constants c > 0 and
[0, 1/2) such that kBxk ckxk + kAxk for all x D(A). Then A + B with
D(A + B) = D(A) is selfadjoint.
Proof. By Theorem 4.18, we have it (A) for all t R\{0}. Take (0, 1 2)
(0, 1) and x X. Using (4.3) we estimate
kB(itI A)1 xk kA(itI A)1 xk + ck(itI A)1 xk
= kit(itI A)1 x xk + ck(itI A)1 xk
 |t|

c
+ 1 kxk + kxk (1 )kxk,

|t|
|t|

65

CHAPTER 4. SELFADJOINT OPERATORS

c
whenever |t| 12
. Theorem 1.26 now implies that it (A + B) for such t.
Finally, A + B is symmetric since

((A + B)x|y) = (Ax|y) + (Bx|y) = (x|Ay) + (x|By) = (x|(A + B)y)


for all x, y D(A). So, A + B is selfadjoint due to Theorem 4.18.
Actually, in the above theorem it suffices to assume that < 1, see Theorem X.13
in [Ree75].
Example 4.24. Let X = L2 (R3 ), A = 4 with D(A) = W22 (R3 ) and V : D(A) X
be given by V u(x) = |x|b 2 u(x) for some b R. Since kp = 4 > d = 3 we have
D(A) , C0 (R3 ) due to Sobolevs embedding.
Let 0 < 1. Using polar coordinates and Example 4.11, we compute
Z
Z
Z
|u(x)|2
|u(x)|2
2
2
2
|V u| dx = b
dx
+
b
2
2 dx
B(0,) |x|2
R3 \B(0,) |x|2
R3
Z
Z 2
r
b2
2
|u|2 dx
drkuk + 2
c
2
r

3
R \B(0,)
0
ckuk22,2 +

b2
b2
2
2
2
kuk

ckAuk
+
ckuk
+
kuk22 ,
2
2
2
2
2

for constants c > 0 independent of u and . Moreover, V is symmetric on D(A) since


Z
b
(V u|v) =
u(x)v(x)dx = (u|V v)
R3 |x|2
for all u, v D(A). So Theorem 4.23 implies that A + V is selfadjoint if we take a
small > 0.
The spectra (A+V ) and p (A+V ) and the eigenfunctions of A+V are computed
in 7.3.4 of [Tri97], where b > 0. The above operator A = 4 + V is used in physics
to describe the hydrogen atom, see also Example 4.33 below.

4.3

The spectral theorems

The following results and proofs hold for normal operators with minor modifications.
Also, the separability assumption made below can be removed. See e.g. Corollaries X.2.8 and X.5.4 in [DS88b] or Theorems 13.24, 13.30 and 13.33 in [Rud91].
Let T L(Z) for a Banach space Z and p(z) = a0 +a1 z +. . .+an z n be a complex
polynomial. We then define the operator polynomial
(4.6)

p(T ) = a0 I + a1 T + . . . + an T n L(Z).

This gives a map p 7 p(T ) from the space of polynomials to L(Z). For selfadjoint T
on a Hilbert space one can extend this map to all f C((T )), as seen in the next
theorem. We set p1 (z) = z.

66

4.3. THE SPECTRAL THEOREMS

Theorem 4.25 (Continuous functional calculus). Let T L(X) be selfadjoint on a


Hilbert space X. There is exactly one map T : C((T )) L(X), f 7 f (T ), such
that
(C1) (f + g)(T ) = f (T ) + g(T ),
(C2) kf (T )k = kf k (hence, T is injective),
(C3)

1(T ) = I and p1 (T ) = T ,

(C4) (f g)(T ) = f (T )g(T ) = g(T )f (T ),


(C5) f (T )0 = f (T ),
hold for all f, g C((T )) and , C. In particular, we have T (p) = p(T ) for
each polynomial p, where p(T ) is given by (4.6).
Proof. We first show the properties (C1)-(C5) for polynomials p(t) = a0 + a1 t + . . . +
an tn and q(t) = b0 + b1 t + . . . + bn tn with t R and the map p 7 p(T ) defined by
(4.6). Here, any aP
j , bj C may be equal to 0. Clearly, (C1) and (C3) hold in this
case, and p(T )0 = nj=0 aj (T j )0 = p(T ) since T = T 0 . Moreover,
(pq)(T ) =

2n  X
X

n
n

X
X
aj bk T m =
aj T j
bj T k = p(T )q(T ).

m=0 0j,kn
j+k=m

j=0

k=0

It also holds that (pq)(T ) = (qp)(T ) = q(T )p(T ), so that (C4) is shown for polynomials. Properties (C4) and (C5) imply that p(T ) is normal. Hence, Proposition 4.14
and Lemma 4.26 below yield
kp(T )k = max{|| : (p(T ))} = max{|| : p((T ))} = kpk .
Since (T ) is compact, the polynomials are dense in C((T )) by Weierstra approximation theorem. We can thus extend the map p 7 p(T ) to a linear isometry
T : f 7 f (T ) from C((T )) to L(X). By continuity, also (C4) and (C5) hold for
T on C((T )).
If there is another map : C((T )) L(X) satisfying (C1)-(C5), then (p) =
p(T ) = T (p) for all polynomials by (C1), (C3) and (C4), so that = T .
Lemma 4.26. Let T L(Z) for a Banach space Z and let p be a polynomial. Then
(p(T )) = p((T )).
Proof. See Theorem 5.3 below.
Corollary 4.27. Let T L(X) be selfadjoint and f C((T )). Then the following
asssertions hold.
(C6) If T x = x for some x X and C, then f (T )x = f ()x,

67

CHAPTER 4. SELFADJOINT OPERATORS

(C7) f (T ) is normal,
(C8) (f (T )) = f ((T )) (spectral mapping theorem),
(C9) f (T ) is selfadjoint if and only if f is real valued.
Proof. Take a sequence of polynomials pn converging P
to f uniformly. P
Let T x = x.
n
j
Property (C6) holds for a polynomial since p(T )x = j=0 aj T x = nj=0 aj j x =
p()x. Hence,
f (T )x = lim pn (T )x = lim pn ()x = f ()x.
n

From (C5) and (C4) we deduce that f (T )f (T )0 = f (T )f (T ) = f (T )f (T ) =


f (T )0 f (T ) so that f (T ) is normal.
We next show (C8). If 6 f ((T )), then g =
(C4) yield

1
f

C((T )). Thus (C3) and

(I f (T ))g(T ) = g(T )(I f (T )) = g(T )(1 f )(T ) = ((1 f )g)(T ) = 1(T ) = I.


Hence, (f (T )). Let = f () for some (T ). Then n := pn () belongs
to (pn (T )) for all n N by Lemma 4.26. Moreover, n I pn (T ) I f (T ).
Suppose that I f (T ) were invertible. Then also n I pn (T ) would be invertible
for large n by Theorem 1.26. This is impossible, and so (f (T )).
For the last assertion, observe that f (T ) = f (T )0 if and only if (f f )(T ) = 0 if
and only if f f = 0, because T is injective.
Corollary 4.28. Let n N and T = T 0 L(X) with (T ) R+ (in this case
one writes T = T 0 0). Then there is a uniquely determined selfadjoint operator
W L(X) with (W ) R+ such that W n = T .
Proof. Consider w(t) = t1/n for t (T ) R+ and define W := w(T ). Then
W n = w(T )n = w(T ) . . . w(T ) = (w . . . w)(T ) = p1 (T ) = T . Properties (C8)
and (C9) imply that W = W 0 0. For the proof of the uniqueness of W we refer to
Korollar VII.1.16 in [Wer05].
Theorem 4.29 (The compact case). Let X be a Hilbert space with dim X = ,
T L(X) be compact and selfadjoint and A be densely defined, closed and selfadjoint
on X having a compact resolvent. Then the following assertions hold.
a) There is an index set J {, N, {1, . . . , N } : N N} and eigenvalues n 6= 0,
n J, such that
n : n J} R
(T ) = {0} {
where n 0 as n if J = N. Further, there is an orthonormal basis of
N (T ) = R(T ) consisting of eigenvectors of T for n , n J. The eigenspaces
En (T ) = N (n I T ) are finite-dimensional and the orthogonal projection Pn onto
En (T ) commutes with T , for all n J. Finally, the sum
X
T =
n Pn ,
nJ

68

4.3. THE SPECTRAL THEOREMS

converges in L(X).
b) It holds
(A) = p (A) = {n : n N} R
with |n | as n , and there is an orthonormal basis of X consisting of
eigenvectors of A. The eigenspaces En (A) = N (n I A) are finite-dimensional, and
we have Qn D(A) D(A) and AQn x = Qn Ax for all x D(A) and n N, where
Qn is the orthogonal projection onto En (A). Finally, the sum
Ax =

n Qn x

n=1

converges in X for all x D(A).


Proof. 1) Theorem 2.11, 4.6 and 4.18 show the assertions on the spectra (except for
the claim that (A) is infinite) and that dim En (T ) and dim En (A) are finite for all
n. Let x, y D(A) be eigenvectors for j 6= k of A. Then
j (x|y) = (Ax|y) = (x|Ay) = k (x|y) ,
so that (x|y) = 0. Similarly, one sees that Ej (T ) Ek (T ) if j 6= k.
2) Concerning T , we first consider the case J = N. Using the Gram-Schmidt
algorithm in each En (T ), we can construct an orthonormal system S consisting of
eigenvectors of T for the eigenvalues n 6= 0, n N. Note that S R(T ). Let 1n =
1{n } C((T )) and n = 11 + . . . + 1n for each n N. Then 1n (T )2 = 12n (T ) =
1n (T ). Moreover, (C4) and (C9) yield that T 1n (T ) = 1n (T )T , 1n (T )0 = 1n (T ) and
(n I T )1n (T ) = ((n 1 p1 )1n )(T ) = ((n n )1n )(T ) = 0.
If n v = T v for some v X, we further deduce 1n (T )v = 1n (n )v = v from (C6).
As a result, 1n (T ) is a selfadjoint projection onto En (T ). For x X and y
N (1n (T )), we thus have (1n (T )x|y) = (x|1n (T )y) = 0 so that 1n (T ) is orthogonal,
i.e., 1n (T ) = Pn and n (T ) = P1 + . . . + Pn for all n N. Employing also (C2), we
further conclude
kT n (T )T k = kT T n (T )k = k((1 n )p1 )(T )k
= k(1 n )p1 k = sup |j | 0,
jn+1

P
as n . Since T 1n (T ) = n Pn , we obtain T =
n=1 n Pn with convergence in
L(X). Moreover, it follows that n (T )y lin S converges to y as n for all
y R(T ). Therefore, S is an orthonormal basis of R(T ) due to e.g. Theorem 2.16
in [FA]. Finally, it holds N (T ) = R(T 0 ) = R(T ) (see e.g. Proposition 4.41 in [FA])
since T = T 0 and X is reflexive.
If J = , then T = 0 because of Proposition 4.14 and the assertions hold trivially.
If J = {1, , N } is finite, then (1 N )p1 = 0 on (T ) and thus the above
arguments work also in this case with a finite sum. So the assertions concerning T
are shown.

69

CHAPTER 4. SELFADJOINT OPERATORS

3) Fix t (A) R and note that 0


/ p (R(t, A)). Let x X be orthogonal to
all eigenvectors v of A. If v is an eigenvector of A with eigenvalue , then 1 (tI
1
A)v =: w is an eigenvector of R(t, A) for the eigenvalue = t
. Vice versa any
eigenvector w of R(t, A) with eigenvalue defines the eigenvector v = R(t, A)w of
A with eigenvalue = t 1/. (See Proposition 1.20.) Observe that R(t, A)0 =
R(t, A0 ) = R(t, A) is selfadjoint, compact and injective. By Step 2, the operator
R(t, A) possesses an orthonormal base of eigenvectors w. In particular, R(t, A) and
thus A have infinitely many eigenvalues since dim X = and the eigenspaces of
R(t, A) are finite dimensional. Moreover, 0 = (x|v) = (R(t, A)x|w) for all these
eigenvectors w. Since these vectors span X, we obtain R(t, A)x = 0 and hence x = 0.
Consequently, the eigenvectors v of A span X. As for self-adjoint matrices one shows
the orthogonality of eigenvectors of A to different eigenvalues. One can then construct
an orthonormal basis of X consisting of eigenvectors v of A, see Lemma 2.14 and
Theorem 2.16 of [FA].
Let vn,j (j = 1, . . P
. , mn ) be eigenvectors of A forming an orthonormal basis of
n
En (A) so that Qn x = m
j=1 (x|vn,j ) vn,j is the orthogonal projection onto En (A), for
each n P
N. As a result, {vn,j : j = 1, . . . , mn ; n N} is an orthonormal basis of X
and x =
n=1 Qn x for all x X. Next, for x D(A) we have
Qn Ax =
=

mn
X
j=1
mn
X

(Ax|vn,j ) vn,j =

mn
X

(x|Avn,j ) vn,j =

j=1
mn
X

(x|vn,j ) n vn,j =

j=1

mn
X

(x|n vn,j ) vn,j

j=1

(x|vn,j ) Avn,j = AQn X.

j=1

We thus obtain
A

n
X
k=1

Qk x =

n
X

Qk Ax

k=1

Qk Ax =

k=1

AQk x =

k=1

k Qk x,

k=1

as n , so that the closedness of A yields the last assertion.


Remark 4.30. a) In the above proof we also obtain that
Ax =

n=1

mn
X

(x|vn,j ) vn,j

j=1

for all x D(A). An analogous result holds for T .


b) Let T be selfadjoint and compact such that N (T ) is separable. Let zk , k
J0 , be an orthonormal basis for N (T ), where J0 Z could be empty. Denote
by l the non-zero eigenvalues of T (repeated according to their multiplicity) with
corresponding orthonormal basis of eigenvectors {wl : l J1 }. The union {bj : j
J 0 } of {zk : k J0 } and {wl : l J1 } is an orthonormal basis of X, where J 0 = J0 J1 .
By e.g. Theorem 2.18 in [FA], the map
: X `2 (J 0 ),

x = ((x|bj ) jJ 0 ),

70

4.3. THE SPECTRAL THEOREMS

P
is unitary with 1 ((j )jJ 0 ) =
jJ 0 j bj . Moreover, the transformed operator
1
2
0
T acts on ` (J ) as the multiplication operator
X
X
T 1 (j ) = T
j bj =
j j bj = (j j )jJ 0 ,
jJ 0

jJ 0

where j = 0 if j J0 . Hence, T 1 is represented as infinite diagonal matrix


with diagonal elements j . Analogous results hold for A from Theorem 4.29, see e.g.
Theorems 4.5.1-4.5.3 in [Tri97].
c) We point out that many of the important orthonormal bases in analysis are
constructed by means of part b) of the above theorem applied to specific differential
operators A, see [Tri97].
There are (at least) two ways to generalize the above diagonalization result to
the non-compact case. Below we present one of them.
Theorem 4.31 (The multiplication representation). Let T L(X) be selfadjoint on
a separable Hilbert space X. Then there is a measure space (, A, ), a measurable
function h : (T ) and a unitary operator U : X L2 () such that
T x = U 1 hU x,
for all x X.
Proof (partly sketched). 1) Let v1 X. We define the linear subspaces
Y1 := {f (T )v1 : f C((T ))} and X1 = Y 1
of X. Since T f (T )v1 = (p1 f )(T )v1 Y1 for every f C((T )), we obtain T Y1 Y1
and so T X1 X1 . We introduce the map
1 : C((T )) C,

1 (f ) = (f (T )v1 |v1 ) ,

which is linear and bounded because |1 (f )| kf (T )kkv1 k2 = kf k due to (C2). If


f 0, then (f (T )) R+ by (C8). So we can deduce from Corollary 4.28 that




1
1
1
(f (T )v1 |v1 ) = f (T ) /2 v1 f (T ) /2 v1 = kf (T ) /2 v1 k2 0.
The Riesz representation theorem of C((T )) now gives a positive measure 1 on
B((T )) such that
Z
1 (f ) =
f d1 ,
(T )

for all f C((T )) L2 (1 ), see e.g. Theorem 2.14 of [Rud87]. For any x =
f (T )v1 Y1 , we define V1 x := f L2 (1 ). We compute
Z
kV1 xk2L2 (1 ) =
|f |2 d1 = 1 (f f ) = (f f )(T )v1 |v1 )
(T )

71

CHAPTER 4. SELFADJOINT OPERATORS

= f (T )0 f (T )v1 |v1 ) = (f (T )v1 |f (T )v1 ) = kxk2X .


In particular, if x = f (T )v1 = g(T )v1 for some g C((T )), then kf gk22 =
k(f (T ) g(T ))v1 k2X = 0, and so f = g in L2 (1 ). As a result, V1 : Y1 L2 (1 ) is
a linear isometric map and can be extended to a linear isometry U1 : X1 L2 (1 ).
Observe that C((T )) R(U1 ). Riesz theorem also yields that 1 is r egular1 so
that C((T )) is dense in L2 (1 ) due to e.g. Theorem 3.14 in [Rud87]. Hence, the
isometry U1 is bijective and thus unitary by e.g. Proposition 2.24 in [FA]. Finally,
we have
U1 T f (T )v1 = V1 (p1 f )(T )v1 = p1 f = p1 U1 f (T )v1 ,
for all f C((T )). By density, we obtain that T x = U11 p1 U1 x for all x X1 .
2) We are done if there is an v1 X such that X1 = X. In general this is
not true. Using Zorns Lemma (see Theorem I.2.7 of [DS88a]), we instead find
an orthogonal system of spaces Xi as in step 1) which span X. To that aim, we
introduce the collection E of all sets E having as elements at most countably many
closed subspaces Xj X of the type constructed in step 1) such that Xi Xj for all
Xi 6= Xj in E. The system E is ordered via inclusion of sets. Let C be a chain
S in E,
i.e., a subset of E such that E F or F E for all E, F C. We put C = EC E.
Cleary, E C for all E C. Let Y, Z C. Then Y and Z are closed subspaces
of X as constructed in 1) and there are E, F C such that Y E and Z F .
We may assume that E F and thus Y, Z F are orthogonal (if Y 6= Z). As a
result, C contains pairwise orthogonal subspaces of X. If x y have norm 1, then
kx yk2 = kxk2 + kyk2 = 2. The separability of X thus implies that C contains at
most countably many subspaces, i.e., C E and so C is an upper bound of C. Zorns
Lemma now gives a maximal element M = {Xj : j J} in E, where J N and Xj
are pairwise orthogonal subspaces as constructed in step 1).
Assume that there is a v X \ {0} being orthogonal to all these Xj . Let Z be
the closure of the subspace of all vectors f (T )v with f C((T )). Let g C((T ))
and x = g(T )vi Xi for some i J, where vi generates Xi as in step 1). We then
obtain
(x|f (T )v) = (f (T )0 g(T )vi |v) = ((f g)(T )vi |v) = 0
since (f g)(T )vi Xi . By density, it follows that Z is orthogonal to all Xi and thus
M {Z} E. The maximality on M now implies that Z M which is only possibly
if v = 0, which was excluded. As a result, X is the L
closed linear span of the elements
in the orthogonal subspaces Xj . (One writes X = jJ Xj .) We now define
[
X
=
(T ) {j} R2 , A = B(), (A) =
j (Aj )
jJ

jJ

P
with AjP
{j} = A ((T ) {j}), h() = if = {j} and U x = jJ Uj xj
if x =
jJ xj and xj Xj . It can be seen that , A, , h and U satisfy the
assumptions.
1

A positive measure is called regular, if


(B) = inf{(O) : B O, O is open} = sup{(K) : K B, K is compact}

holds for all B in the corresponding Borel sigma algebra.

72

4.3. THE SPECTRAL THEOREMS

S We add an observation to the above proof. Let (T ) \ p (T ). Then =


jJ { {j}} is a null set. In fact, otherwise the charcteristic function
f of would be nonzero and so x = U 1 f 6= 0 would be an eigenvector of T for
the eigenvalue , since T x = U 1 f S
= x. We further note that in the proof of
Theorem 4.31 one could
S also take = jJ (Tj ) {j}, where Tj = T |Xj . It can be
shown that (T ) = jJ (Tj ).
The above representation of bounded selfadjoint operators as multiplication operators now leads to a multiplication representation and to a Bb functional calculus
for (possibly) unbounded selfadjoint operators A. Here Bb ((A)) is the Banach space
of bounded Borel functions on (A) endowed with the supremum norm. We use this
space instead of L ((A)) to avoid certain technical problems.
Theorem 4.32 (Unbounded A). Let A be a closed, densely defined and selfadjoint
operator on a separable Hilbert space X. Then the following assertions hold.
a) There is a measure space (, A, ), a measurable function h : (A) and a
unitary operator U : X L2 () such that
D(A) = {x X : hU x L2 ()}

and

Ax = U 1 hU x.

b) There is a contractive map : Bb ((A)) L(X), (f ) = f (A), satisfying


(C1) and (C3)-(C5) where p1 (T ) = T in (C3) is replaced by r (A) = R(, A)
for r (z) = ( z)1 and every (A). Moreover, if fn Bb ((A)) are
uniformly bounded and converge to f Bb ((A)) pointwise, then fn (A)x f (A)x
as n for all x X. Finally, for x D(A) and f Bb ((A)) we have
f (A)x D(A) and Af (A)x = f (A)Ax.
Proof. a) Let t (A) R. (If (A) = R, take t = i and use the version of
Theorem 4.31 for the normal operator R(i, A), see e.g. Satz VII.1.25 in [Wer05].)
Then R(t, A) L(X) is selfadjoint and can be represented as R(t, A) = U 1 mU on
a space L2 (, ) as in Theorem 4.31. Recall that Proposition 1.20 yields (A) =
t [(R(t, A))\{0}]1 . Set
h( {j}) = t

1
1
= t (A),
m( {j})

for j J and (R(t, A))\{0}. The sets {0{j}} have measure 0 in view of the
obervation above the theorem, due to the injectivity of R(t, A). We can thus extend
h by 0 to a measurable function on . Let x D(A). We put y = tx Ax X.
Since x = R(t, A)y = U 1 mU y, we obtain
hU x = hmU y = (tm 1)U y L2 ()
U

hU x = tU

(with equalities a.e.),

mU y y = tx y = Ax.

If x X satisfies hU x L2 (), then we put y = U 1 (t1 h)U x X and obtain


mU y = (tm mh)U x = U x. Therefore, x = U 1 mU y = R(t, A)y D(A) and part
a) is proved.

73

CHAPTER 4. SELFADJOINT OPERATORS

b) We define a functional calculus : f 7 f (A) for A by setting


f (A)x = U 1 (f h)U x
for f Bb ((A)) and x X. We further put Mf = (f h) for L2 (). It is
straightforward to check that f (A) L(X), is linear, 1(A) = I and (C5) is true.
Let (A). We have hU r (A)x = h(r h)U x = h(1 h)1 U x L2 () for all
x X. So part a) yields that r (A)X D(A) and (I A)r (A) = I. Similarly,
one sees that r (A)(x Ax) = x for all x D(A), and so (C3) is shown. The
contractivity and property(C4) hold since kf (A)k = kMf k kf k (cf. Example
1.54 in [FA]) and
(f g)(A)x = U 1 (f h)(g h)U x = U 1 (f h)U U 1 (g h)U x = f (A)g(A)x,
where g Bb ((A)) and x X.
Let f, fn Bb ((A)) be uniformly bounded by c such that fn f pointwise as
n . For every x X, we have fn (A)x f (A)x = U 1 ((fn f ) h)U x. Since
(fn f ) h 0 pointwise and |((fn f ) h)U x| 2c|U x|, Lebesgues convergence
theorem shows that ((fn f ) h)U x tends to 0 in L2 () and so fn (A)x f (A)x in
X as n .
Let x D(A). The above results yield that
g := h(f h)U x = (f h)U U 1 hU x = (f h)U Ax L2 ().
On the other hand, g = hU U 1 (f h)U x = hU f (A)x. Part a) thus implies that
f (A)x D(A) and
Af (A)x = U 1 hU f (A)x = U 1 g = U 1 (f h)U Ax = f (A)Ax.
Example 4.33 (Schr
odingers equation). Let X be a Hilbert space and H be a
closed, densely defined and selfadjoint operator on X. For a given x D(H) we look
for functions u C 1 (R, X) C(R, [D(H)]) satisfying
(4.7)

d
u(t) = iHu(t),
dt

t R,

u(0) = x.

(H is called Hamiltonian.) An example for this setting is X = L2 (R3 ) and H =


(4 + |x|b 2 ) with D(H) = W22 (R3 ), see Example 4.24. For t R, we consider the
bounded function ft () = eit , R. Theorem 4.32 allows us to define U (t) =
ft (H) L(X). From (C4) and (C5) it follows that
U (t)U (s) = (ft fs )(H) = ft+s (H) = U (t + s) = U (s)U (t),
U (0) = f0 (H) = I,
U (t)0 = ft (H) = ft (H) = U (t).
Taking s = t, we see that there exists U (t)1 = U (t) = U (t)0 , and so U (t) is
unitary. Since kft k1 = 1 and t 7 ft () is continuous for all t, R, the function
R 3 t 7 U (t)z X is continuous with U (0)z = z for each z X.

74

4.3. THE SPECTRAL THEOREMS

Let x D(H). We set y = x Hx for some (H), so that x = R(, H)y =


r (H)y. We then obtain
1
1
(U (t)x U (s)x) =
(ft (H) fs (H))r (H)y
ts
ts


1
(ft fs )r (H)y =: gt,s (H)y
=
ts
i
fs ()

|<
sup(H) |

for all t 6= s. Observe that gt,s ()

=: m()fs () as t s for all (H)

and kgt,s k kmk =


that there exists

for all t 6= s. So Theorem 4.32 shows

d
U (t)x = m(H)U (t)( x Hx) = U (t)m(H)( x Hx).
dt
Moreover, m(T )y = U (m h)U 1 y = U ((ip1 r ) h)U 1 y = U ihU 1 U (r
h)U 1 y = iHR(, H)y = iHx due to Theorem 4.32. Hence, we arrive at
d
U (t)x = iU (t)Hx = HU (t)x,
dt
using Theorem 4.32 once more. Due to these equations, u = U ()x belongs to
C 1 (R+ , X) C(R+ , [D(H)]) and solves (4.7).
Let v C 1 (R+ , X) C(R+ , [D(H)]) be another solution of (4.7). For 0 s t
and h 6= 0 we compute
1
(U (t s h)v(s + h) U (t s)v(s)) U (t s)(v 0 (s) + iHv(s))
h


1
0
= U (t s h)
(v(s + h) v(s)) v (s) + (U (t s h) U (t s))v 0 (s)
h
1

(U (t s h) U (t s))v(s) U (t s)iHv(s) 0
h
as h 0. For any y X we thus obtain
Consequently

d
ds

(U (t s)v(s)|y) = 0, since v 0 = iHv.

(U (t)x|y) = (U (t)v(0)|y) = (U (0)v(t)|y) = (v(t)|y) ,


which gives u(t) = v(t) for all t 0. Thus the strongly continuous unitary group
(U (t))tR solves (4.7) uniquely.

75

Chapter 5

Holomorphic functional calculus


We come back to the case of Banach spaces X and Y .

5.1

The bounded case

Let U C be open, f : U Y be holomorphic (i.e., complex differentiable) and


U be a piecewise C 1 curve with parametrization : [a, b] U . We define the
curve integral
Z
Z b
f dz =
f ((t)) 0 (t)dt

as a Banach space-valued Riemann integral (having the


results and
R same definition,
R
proofs as for Y = R in Analysis 1). It also holds T f dz = T f dz for all T
L(Y, Z). The index of a closed curve (i.e., (a) = (b)) is given by
Z
1
dw
n(, z) =
2i w z
for all z C\.
If is closed, Y = C and n(, z) = 0 for all z 6 U , then Cauchys integral
theorem and formula
Z
(5.1)
f dz = 0,

Z
f (w)
1
(5.2)
dw = n(, z)f (z),
2i w z
hold for all z U \. (This is the socalled homologuous version of these results, see
e.g. Theorem IV.5.4 and IV.5.7 in [Con78].) For a general Banach space (5.1) and
(5.2)
hold for the functions z 7 hf (z), y i for every y Y . Hence, we have e.g.
R
h f dz, y i = 0 for all y Y so that (5.1) holds in Y due to a corollary of the
Hahn-Banach theorem. Similarly one deduces (5.2) for a general Banach space Y .
For a compact K C we introduce the space
H(K) ={f : there exists an open set D(f ) C such that K D(f )

76

5.1. THE BOUNDED CASE

and f : D(f ) C is holomorphic}.


Let K U C, K be compact, and U be open. By Proposition VIII.1.1 in [Con78]
and its proof there exists an admissible curve for K and U (or, in U \ K) which
means that U \K is piecewise C 1 , n(, z) = 1 for all z K and n(, z) = 0 for
all z C\U . Here it is allowed that = 1 n is the finite union of closed
curves j , where we set
Z
Z
Z
f dz.
f dz + . . . +
f dz =

Let T L(X), f H((T )), and be admissible for (T ) and D(f ). We then
define
Z
1
(5.3)
f (T ) :=
f ()R(, T )d L(X).
2i
This integral exists in the Banach space L(X) since 7 f ()R(, T ) is holomorphic
1
one sees the similarity of (5.3)
on (T ) D(f ) . Writing R(, T ) as T
and (5.2), but here R(, T ) does not exist on (T ), whereas in (5.2) the function
1
w 7 wz
is defined on C\{z}.
If 0 is another admissible curve for (T ) and D(f ), then we set 00 = (0 ),
where denotes the inversion of the orientation. We then have
(
1 1 = 0, z (T ),
n(00 , z) = n(, z) n(0 , z) =
0 0 = 0, z C\D(f ).
So we can apply (5.1) on U = D(f )\(T ) obtaining
Z
Z
Z
0=
f ()R(, T )d =
f ()R(, T )d
f ()R(, T )d.
00

Consequently, (5.3) does not depend on the choice of the admissible curve. We recall
1
that r (z) = z
and p1 (z) = z for , z C with 6= z.
Theorem 5.1. Let T L(X) for a Banach space X. Then the map
: H((T )) L(X),

f 7 f (T ),

defined by (5.3) is linear and satisfies


(P1) kf (T )k c() sup |f ()|
(P2)

for a constant c() > 0,

1(T ) = I, p1 (T ) = T , r (T ) = R(, T ),

(P3) f (T )g(T ) = g(T )f (T ) = (f g)(T ),


(P4) f (T ) = f (T ),
(P5) if fn f uniformly on compact subsets of D(f ), then fn (T ) f (T ) in L(X)
as n ,

77

CHAPTER 5. HOLOMORPHIC FUNCTIONAL CALCULUS

for all (T ) and f, g, fn H((T )) with D(fn ) = D(f ) for every n N.


is the only linear map from H((T )) to L(X) satisfying (P1)-(P3). For a
polynomial p, the operators p(T ) in (5.3) and in (4.6) coincide.
Proof. It is clear that f 7 f (T ) is linear. Property (P1) follows from
kf (T )k

1
`() supkR(, T )k sup|f ()| =: c() sup|f ()|.
2

Replacing here f (T ) by f (T ) fn (T ) = (f fn )(T ) we also deduce (P5). To check


(P4), we recall that (T ) = (T ) and R(, T ) = R(, T ) from Theorem 1.24.
Hence,
Z
1

f ()R(, T ) d = f (T ).
f (T ) =
2i
We next show (P3). We choose a bounded open set U C such that (T ) U
U D(f ) D(g) and admissible curves f in U \(T ) and g in (D(f ) D(g))\U .
We then have n(f , ) = 0 for all g C\U and n(g , ) = 1 for all f U .
Using the resolvent equation, Fubinis theorem in L(X) (see e.g. Theorem X.6.16 in
[AmE08]) and (5.2) in C, we compute
Z
Z
1
1
f (T )g(T ) =
f ()R(, T )
g()R(, T )dd
2i f
2i g

 Z Z
1
1 2
f ()g()
=
(R(, T ) R(, T ))dd
2i

f g
Z
Z
1
1
g()
=
f ()R(, T )
dd
2i f
2i g
Z
Z
1
1
f ()
+
g()R(, T )
dd
2i g
2i f
Z
1
=
f ()g()R(, T )d = (f g)(T ).
2i f
This identity also yields (f g)(T ) = (gf )(T ) = g(T )f (T ).
To check (P2), we take f = 1 with D(f ) = C. We choose 0 = B(0, 2kT k).
Theorem 1.16 then leads to
Z
Z X

1
1
R(, T )d =
T n n1 d
1(T ) =
2i 0
2i 0
n=0
Z

X
1
=
Tn
n1 d = I,
2i 0
n=0

R
since the series converges in L(X) uniformly on 0 and 0 m d is equal to 2i if
m = 1 and equal to 0 for m Z\{1}. The property p1 (T ) = T is shown similarly.
For (T ), consider f (z) = z. The previous results imply f (T ) = I T
and f (T )r (T ) = r (T )f (T ) = (r f )(T ) = 1(T ) = I so that r (T ) = R(, T ).

78

5.1. THE BOUNDED CASE

Let : H((T )) L(X) be any linear map satisfying the assertions (P1)-(P3).
The linearity, (P2) and (P3) imply that (p) = p(T ) = a0 I + a1 T + . . . + am T m
for every polynomial p with p(z) = a0 + a1 z + . . . + am z m . If f H((T )) and
D(f )\(T ) is admissible, then there are polynomials pn converging uniformly
to f on the compact set . Hence, pn (T ) (f ) by (P1) and thus = .
Example 5.2. Let E = C(K) for a compact set K Rd and let m C(K). We
define M = m for E. Proposition 1.14 shows that M L(E), (M ) = m(K),
1
and R(, M ) = m
for all (M ).
Let f H(m(K)), be an admissible curve in D(f )\m(K), E, and x K.
Using that the map 7 (x) is continuous and linear from E to C and Cauchys
formula (5.2) for = m(x) m(K), we compute
Z
Z
1
1
f ()R(, M )d (x) =
f ()(R(, M ))(x)d
(f (M ))(x) =
2i
2i
Z
1
1
=
f ()
d(x) = f (m(x))(x).
2i
m(x)
As a result, f (M ) = (f m) is also a multiplication operator.
Theorem 5.3. Let T L(X) and f H((T )). Then the spectral mapping theorem
holds:
(f (T )) = f ((T )).
Proof. Let 6 f ((T )). Then g = 11f H((T )) and g(T )(I f (T )) = (I
f (T ))g(T ) = (g(1 f ))(T ) = I so that (f (T )).
()
Conversely, let = f () for some (T ). We set g(z) = f (z)f
for z
z
0
2
D(f )\{} and g() = f (). Since f C by its holomorphy, g is holomorphic on
D(f ) and g(z)(z ) = f (z) for all z D(f ). Suppose that (f (T )). Since

g(T )(I T ) = (I T )g(T ) = (g(1 p1 ))(T ) = (1 f )(T ) = I f (T ),


(I T ) would then have the inverse g(T )(I f (T ))1 which is impossible. Hence,
(f (T )).
Example 5.4. Let A L(X). For t R we set ft (z) = etz and define etA =
ft (A) L(X). As in Example 4.33, one sees that e(t+s)A = etA esA = esA etA ,
e0A = I = etA etA for all t, s R. Moreover, t 7 etA belongs to C 1 (R, L(X)) with
d tA
= AetA = etA A and the map u(t) = etA x is the unique solution in C 1 (R, X) of
dt e
d
u(t) = Au(t)
dt

(t R),

u(0) = x,

where x X is given. (See also the exercises.) Theorem 5.3 further yields





r etA = max || : (etA ) = et(A) = max et Re : (A) = ets(A) ,

79

CHAPTER 5. HOLOMORPHIC FUNCTIONAL CALCULUS

where s(A) := max{Re : (A)} is the spectral bound of A. Now, if s(A) < 0
(i.e. (A) { : Re < 0}), then we deduce from Theorem 1.16 that
1 > r(eA ) = lim k(eA )n k /n = lim kenA k /n .
1

So we can fix some N N such that keN A k =: q < 1. Writing any given t 0 as
t = hN + for some h N0 and 0 < N we estimate

h
ln q 
ketA k = k eN A e A k q h ke A k max ke A k exp N h
M ewt 0
0 N
N
as t , where w := lnNq > 0 and M := max0 N ke A ke|ln q| . So spectral information on the given operator A implies the exponential decay ku(t)k M ewt kxk
of the solutions u.
Remark. Let S L(X) and P = P 2 L(X) be a projection with SP = P S. Set
X1 = R(P ) and X2 = N (P ). Then, X = X1 X2 by e.g. Lemma 1.63 of [FA].
Moreover, if y = P x R(P ), then Sy = SP x = P Sx R(P ). If x N (P ), then
P Sx = SP x = 0 and so Sx N (P ). As a result, SXj Xj and the restrictions
S|Xj L(Xj ) are well defined for j = 1, 2.
2 for two
Theorem 5.5 (Spectral projection). Let T L(X) and (T ) = 1
disjoint closed sets j 6= in C. Then there is a projection P L(X) with f (T )P =
P f (T ) for all f H((T )) such that (Tj ) = j for j = 1, 2, where Tj = T|Xj
L(Xj ), X1 = R(P ) and X2 = N (P ). Moreover, X = X1 X2 and R(, Tj ) =
R(, T )|Xj for (T ) = (T1 ) (T2 ). We further have
(5.4)

P =

1
2i

Z
R(, T )d,
1

where 1 is an admissible curve for 1 and any open set U1 1 such that U1 2 = .
Proof. There are open sets Uj with U 1 U 2 = and j Uj for j = 1, 2. Define
h H((T )) by h = 1 on U1 and h = 0 on U2 . We set P = h(T ) L(X). We then
deduce P 2 = h2 (T ) = h(T ) = P and f (T )P = P f (T ) for all f H((T )) from (P3)
and (P2). The above remark shows that X = X1 X2 holds and that the operators
Tj = T|Xj L(Xj ) are well defined.
The formula (5.4) follows by choosing = 1 2 , where j are admissible curves
for j and Uj for j = 1, 2. Let 6 1 . We may shrink U1 so that 6 U1 since P
does not depend on the choice of and thus not on the choice of U1 . We define
1
g(z) = z
for z U1 and g(z) = 0 for z U2 . Then g H((T )) and
g(T )(I T ) = (I T )g(T ) = ((1 p1 )g)(T ) = h(T ) = P.
Setting R = g(T )|X1 L(X1 ), we thus obtain
R(IX1 T1 ) = (IX1 T1 )R = IX1 .

80

5.2. SECTORIAL OPERATORS

This means that (T1 ), and so (T1 ) 1 . Similarly, one sees that (T2 ) 2 .
In particular, (T1 ) and (T2 ) are disjoint. Let (T1 ) (T2 ). Given x X,
we have unique x1 X1 and x2 X2 such that x = x1 + x2 . If x T x = 0,
then 0 = x1 T1 x1 + x2 T2 x2 X1 X2 so that xj N (I Tj ) for j = 1, 2
and so x = 0. Given y X, we define xj = R(, Tj )yj Xj for j = 1, 2. Setting
x = x1 + x2 , we derive
x T x = x1 T1 x1 + x2 T2 x2 = y1 + y2 = y.
We have shown that (T ), R(, T )|Xj = R(, Tj ) and
2 (T1 )(T

2,
(T ) = 1
2 ) 1
which implies that (Tj ) = j for j = 1, 2.
Example 5.6 (Exponential dichotomy). In the setting of Example 5.4 assume that
2 where 1 { C : Re < 0} and
(A) iR = . Hence, (A) = 1
2 { C : Re > 0}. Let P be the spectral projection of A for 1 and define A1
and A2 as the restrictions of A to X1 = R(P ) and X2 = N (P ), respectively, as in
Theorem 5.5. Let 1 be given as in Theorem 5.5. Observe that for x X1 we have
Z
1
tA
tA
e x = e P x = (ft h)(A)x =
et R(, A)xd
2i 1
Z
1
et R(, A1 )xd = etA1 x,
=
2i 1
where ft () = et for t R and h is given by the proof of Theorem 5.5. In the same
way one derives etA x = etA2 x for all x X2 and t 0. Since (A1 ) = 1 , we have
s(A1 ) < 0 and so Example 5.4 shows that ketA x1 k M et kx1 k for all t 0 and
x1 X1 and some constants M, > 0.
= { :
We further have (A2 ) = 2 and so s(A2 ) < 0. Note that the curve
} is admissible for (A) = (A). Substituting = , we conclude that
Z
Z
1
1
tA
t
1
e
=
e (I A) d =
et (I (A))1 d = et(A)
2i
2i
for all t R. For x2 X2 we thus obtain
etA2 x2 = etA x2 = et(A) x2 = et(A2 ) x2
0

so that ketA2 x2 k M 0 e t kx2 k for all t 0 and some constants M 0 , 0 > 0.


Summing up, we have decomposed X into etA -invariant subspaces where etA decays
exponentially in forward and in backward time, respectively.

5.2

Sectorial operators

For (0, ] we define the open sector


= { C\{0} : |arg | < }.
We also set /2 =: C . Note that = C\R .

81

CHAPTER 5. HOLOMORPHIC FUNCTIONAL CALCULUS

Definition 5.7. A closed operator A is called sectorial of angle (0, ] if there is


a constant K > 0 such that (A) and
kR(, A)k

K
||

holds for all . If there is an R such that A I is sectorial of angle ,


then we say that A is sectorial with constants (, K, ).
Note that a sectorial operator of angle is also sectorial of angle 0 (0, ).
Example 5.8. Let X be a Hilbert space and A be closed, densely defined and
selfadjoint on X. We further suppose that (A) (, ] for some R. Then
A I =: A is selfadjoint with (A ) R . Take ( 2 , ) and . Since
R(, A )0 = R(, A ), the operator R(, A ) is normal. Propositions 4.14 and 1.20
then yield
kR(, A )k = r(R(, A )) = max{|| : ( (A ))1 {0}}
(
1
,
Re 0,
= d(, (A ))1 ||1
|Im | , Re < 0.
If Re < 0, we can write = ||ei for some ( 2 , ). We then have
sin( ) sin( ) > 0, and thus
kR(, A )k

1
sin()

||

=:

|Im |
||

K
||

for all . As a result, A is sectorial with constants (, K , ) for all < .


Example 5.9. Let X = C([0, 1]) and Ak u = u00 for k = 0, 1 with
D(A0 ) = {u C 2 ([0, 1]) : u(0) = u(1) = 0},
D(A1 ) = {u C 2 ([0, 1]) : u0 (0) = u0 (1) = 0}.
Let and = 2 for some C with Re > 0. Let f X. As in Example
4.10, it holds
u D(A0 ), u A0 u = f
if and only if
if and only if

u C 2 ([0, 1]), u00 = 2 u f on [0, 1], u(0) = u(1) = 0,


Z 1
1
t
t
u(t) = ae + be
+
e|ts| f (s)ds and
2 0
Z 1
1
u(0) = a + b +
es f (s)ds = 0 and
2 0
Z
e 1 s

u(1) = ae + be +
e f (s)ds = 0.
2 0

82

5.2. SECTORIAL OPERATORS

The numbers a = a(f, ) and b = b(f, ) satisfy the two boundary conditions above
if and only if
!
R 1 s




1
2
e
f
(s)ds
1
a(f, )
e
1
0
=
R 1
b(f, )
e 1
e e
e2 0 es f (s)ds
!
R1
1
e 0 (es es )f (s)ds
R 1 s
=
.
2(e e )
e es )f (s)ds
0 (e e
We thus obtain (A0 ) and
2

R( , A0 )f (t) = a(f, )e

+ b(f, )e

1
+
2

e|ts| f (s)ds

for all 2 = C\R , Re > 0, f X and t [0, 1].


Fix ( 2 , ). Take and thus ; i.e., = ||ei with 0 || < 2 .
2

We then obtain Re = || cos || cos 2 . So we can estimate


Z t
kf k
Re
e Re |r| dr
sup
kR(, A0 )f k |a(f, )e
| + |b(f, )| +
2|| t[0,1] t1
Z 1

kf k
Re s
Re s

ds
e
+
e
2|||e e | 0
Z 1

kf k
+
(eRe e Re s + e Re eRe s )ds +
|| Re
0
kf k
=
(eRe 1 + 1 e Re )
2 Re |||e e |

kf k
+ eRe (1 e Re ) + e Re (eRe 1) +
|| Re
1
 (eRe e Re ) + (eRe e Re )

cos(/2)

kf
k
+
1

||2
2(eRe e Re )
=

2
cos(/2)

||

kf k .

Hence, A0 is sectorial for all angles < , where K as .


As in Example 4.10 one sees that (A0 ) = p (A0 ) = { 2 k 2 : k N} with
eigenfunctions uk (t) = sin(kt). Note that D(A0 ) = {u X : u(0) = u(1) = 0} =
6 X.
In a similar way, one shows that A1 is sectorial for every angle < with
(A1 ) = p (A1 ) = { 2 k 2 : k N0 } with eigenfunctions uk (t) = cos(kt). Here
D(A1 ) is dense in X. (See the exercises.)
Example 5.10. Let X = Lp (R), 1 p < , and Au = u0 for D(A) = Wp1 (R). In
Example 4.2 we have seen that (A) = iR and kR(, A)k Re1 for Re > 0. If
(0, 2 ) and we have |Re | = || cos so that A is sectorial for each angle
< 2 . If = s + i for s > 0, then kR(s + i, A)k 1/s due to Theorem 1.13 and so
A is not sectorial of angle = 2 .

83

CHAPTER 5. HOLOMORPHIC FUNCTIONAL CALCULUS

Remark. Let 1 < p < , U Rd be open with U C 2 , and X = Lp (Rd ). The


operators
A0 = 4,

1 (U ),
D(A0 ) = Wp2 (U ) W
p

A1 = 4,

D(A1 ) = {u Wp2 (U ) : u :=

d
X

j T j u = 0

on U },

j=1

are sectorial on X with angle > /2. Here T : Wp1 (U ) Lp (U ) is the trace
operator and is the outer unit normal. There are variants for the spaces X = L1 (Rd )
and X = C(U ) as well as for more general differential operators and boundary
conditions. See e.g. Chapter 3 in [Lun95] and also [Tan97].
Let A be sectorial of angle ( 2 , ) with constant K. Take any r > 0 and
( 2 , ). We define
1 = { = 1 (s) = (s)ei : < s r},
2 = { = 2 () = rei : },
3 = { = 3 (s) = sei : r s < },
= (r, ) = 1 2 3 .
For t > 0, we introduce the operator
Z
Z
1
1
t
tA
e R(, A)d = lim
et R(, A)d,
(5.5)
e =
R 2i R
2i
where R = B(0, R) for R > r. We first have to show that the limit in (5.5)
exists in L(X).
Lemma 5.11. Under the above assumptions, the integral in (5.5) converges absolutely in L(X) and gives an operator etA L(X) which does not depend on the
choice of r > 0 and ( 2 , ). Moreover, ketA k M for all t > 0 and a constant
M = M (K, ) > 0.
Proof. Since kR(, A)k
Z
|

K
||

ke R(, A)kd| K
R

on , we can estimate

exp(ts Re ei ) i
|e |ds
|sei |

Z R
exp(tr Re ei )
exp(ts Re ei ) i
i
+K
|ire
|d
+
K
|e |ds
|rei |
|sei |

r
 Z ts cos

Z
e
tr cos
K 2
ds +
e
d
s
r

!
Z
d
e
tr
K 2
(t cos )
+ 2e
t cos
rt|cos |
Z

=: Kc(r, t, ),

84

5.2. SECTORIAL OPERATORS

for all R, t > 0, where we substituted = st cos . Thus the limit in (5.5) exists
absolutely in L(X) by the majorant criterium, and ketA k Kc(r, t, ). If we take
r = 1/t, then c(1/t, t, ) = c() does not depend on t > 0.
So it remains to check that the integral in (5.5) is independent of r > 0 and
( 2 , ). To this aim, we define 0 = (r0 , 0 ) for some r0 > 0 and 0 ( 2 , ). We
+

further set 0R = 0 B(0, R) and choose R > r, r0 . Let CR


and CR
be the circle arcs
0
from the endpoint of R to that of R in {Im > 0} and {Im < 0}, respectively.

(If = 0 , then CR
contain just one point.) Then SR = R CR
(0R ) (CR
)
is a closed curve in the starshaped domain . So (5.1) shows that
Z
et R(, A)d = 0.
SR

Let = min{, 0 } ( 2 , ). It holds


Z
k

CR

Z

et R(, A)dk

etR Re e


K
i
K|0 |etR cos 0,
|iRe
|d
|Rei |

as R 0. So we conclude that
Z
Z
t
e R(, A)d = lim
et R(, A)d
R R

Z
Z
= lim
et R(, A)d =
et R(, A)d,
R 0
R

as asserted.
Theorem 5.12. Let A be sectorial of angle > 2 . Define etA as in (5.5) for t > 0,
and set e0A = I. Then the following assertions hold.
a) etA esA = esA etA = e(t+s)A for all t, s 0.
b) The map t 7 etA belongs to C 1 ((0, ), L(X)). Moreover, etA X D(A),
d tA
= AetA and kAetA k Ct for a constant C > 0 and all t > 0. We also have
dt e
tA
Ae x = etA Ax for all x D(A) and t 0.
c) etA x converges as t 0 if and only if x D(A). In this case, it holds etA x x
as t 0.
Proof. a) Let t, s > 0. Take 0 < r < r0 and 2 < 0 < < . Set = (r, ) and
0 = (r0 , 0 ). Using the resolvent equation and Fubinis theorem, we compute
Z
Z
1
tA sA
t
e e =
e
es R(, A)R(, A)dd
(2i)2
0
Z
Z
1
1
es
t
=
e R(, A)
dd
2i
2i 0
Z
Z
1
1
et
s
+
e R(, A)
dd.
2i 0
2i

85

CHAPTER 5. HOLOMORPHIC FUNCTIONAL CALCULUS

0 = {z = Rei : 0 2 0 }
Fix and take R > max{r, r0 , ||}. We define CR
0
0
0
0
and SR = R CR . Since n(SR , ) = 1, Cauchys formula (5.2) yields
Z
es
1
d = es .
2i SR0

We further have
Z
0R

Z
|
0
CR

es
d

Z
0

es
d

and

es Re
2R
es
0
d| 2R sup
esR cos
0
0 | |

R ||
CR

as R . Consequently,
e

1
=
2i

Z
0

es
d.

Closing R with the circle arc CR = {z = Rei : 2 }, one verifies in the


same way that
Z
et
0=
d.

We thus conclude that
tA sA

e e

1
=
2i

et es R(, A)d = e(t+s)A = esA etA .

b) LetR x X, t > 0, > 0, and R > r. Observe that the Riemann


sums for R et R(, A)d converge in [D(A)] since 7 R(, A) is continuous in
L(X, [D(A)]). We thus obtain
Z
Z
Z
Z
t
t
t
(5.6) A
e R(, A)d =
e AR(, A)d =
e R(, A)d
et dI.
R

Let CR = { = Rei : 2 }. Using (5.1), one shows as in part a) that


Z
Z
t
|
e d| = |
et d| 2R sup etR cos 2ReR cos 0
R

CR

as R , uniformly for t . Moreover, as in the proof of Lemma 5.11 (with


r = 1/t) we estimate
!
Z
Z
Z
|s| ts cos
1 cos
t
|
ke R(, A)kd| K 2
e
ds +
e
d
1
|s|
t
R
t

2K
2eK
C

+
=: .
t|cos |
t
t

86

5.2. SECTORIAL OPERATORS

Therefore, the right hand side of (5.6) converges to


Z
et R(, A)d

as R . Since A is closed, it follows that etA X D(A) and


Z
1
C
AetA =
et R(, A)d,
kAetA k
2i
2t
for all t > 0. In a similar way one sees that
Z
Z
Z
t
t
e R(, A)d| 2K
| e R(, A)d
R

ets cos ds

2K R cos
e
0
|cos |

as R , uniformly for t . As a result,


Z
Z
d
t
e R(, A)d =
et R(, A)d
d R
R
converges in L(X) uniformly for t , and so t 7 etA L(X) is continuously
d tA
e = AetA . For x D(A), it further holds
differentiable for t > 0 with dt
Z
1
tA
Ae x = lim
et R(, A)Axd = etA Ax.
R 2i R
c) Let x D(A), R > r, and t > 0. As in part a), Cauchys formula (5.2) yields
Z t
Z
1
e
1
et
d = lim
d = 1
R 2i R 0
2i
Observing that R(, A)x x = R(, A)Ax, we conclude that
Z
Z t
1
1
1
e
tA
t
e xx=
e (R(, A) )xd =
R(, A)Axd.
2i

2i
Since the integrand is bounded by ||c 2 on for all t (0, 1], Lebesgues convergence
theorem implies that there exists the limit
Z
1
1
lim(etA x x) =
R(, A)Axd =: z.
t0
2i
Let KR = {Rei : }. Cauchys theorem (5.1) shows that
Z
1
R(, A)Axd = 0.

R (KR )
Since also

Z
k
KR

1
2RK
R(, A)Axdk
kAxk 0

R2

as R , we arrive at z = 0. Because of the uniform boundedness of etA , it follows


that etA x x as t 0 for all x D(A).
Conversely, if etA x y as t 0, then y D(A) by part b). Moreover,
R(1, A)etA x = etA R(1, A)x tends to R(1, A)x as t 0, since R(1, A)x D(A).
We thus obtain R(1, A)y = R(1, A)x, and so x = y belongs to D(A).

87

CHAPTER 5. HOLOMORPHIC FUNCTIONAL CALCULUS

Remark 5.13. If A I = A is sectorial of angle > 2 for some R, then


Z
Z
1
1
t tA
t(+)
e e
=
e
R( + , A)d =
et R(, A)d =: etA
2i
2i +
holds for all t > 0. It is easy to see that etA has the analogous properties as in the
case = 0.
Corollary 5.14. Let A be sectorial with constant (, K, ), where > 2 and let
x D(A). Then u(t) = etA x, t 0, is the unique solution in C 1 ((0, ), X)
C((0, ), [D(A)]) C(R+ , X) of the initial value problem
(5.7)

u0 (t) = Au(t),

t > 0,

u(0) = x.

Proof. Existence follows from Theorem 5.12 and Remark 5.13. Let v be another
solution of (5.7). Let 0 < s t t. Theorem 5.12 then implies that
d (ts)A
e
v(s) = e(ts)A Av(s) + e(ts)A v 0 (s) = 0.
ds
As in Example 4.33, this fact yields e(t)A v() = eA v(t ). Letting 0, one
obtains that etA x = v(t).
Example 5.15. Consider X = C([0, 1]), A = 00 , D(A) = { C 2 ([0, 1]), 0 (0) =
0 (1) = 0}. Let u0 X. Then u(t) = etA u0 belongs to

C(R+ , X) C (0, ), C 2 ([0, 1])) C 1 ((0, ), X)
and solves the following partial differential equation

t u(t, x) = xx u(t, x), t > 0, x [0, 1],


x u(t, 0) = x u(t, 1), t > 0,

u(0, x) = u0 (x),
x [0, 1].

88

Bibliography
[AdF07] R.A. Adams, J.F. Fournier: Sobolev Spaces. 2nd edition. Academic Press,
Amsterdam, 2007.
[Alt06]

H.W. Alt: Lineare Funktionalanalysis. 5te Auflage. SpringerVerlag, Berlin, 2006.

[AmE08] H. Amann, J. Escher: Analysis III. Birkhauser, Basel, 2008.


[Con78] J.B. Conway: Functions of One Complex Variable. 2nd edition. Springer
Verlag, Heidelberg, 1978.
[Con90] J.B. Conway: A Course in Functional Analysis. 2nd edition. Springer
Verlag, New York, 1990.
[DS88a] N. Dunford, J.T. Schwartz: Linear Operators, Part 1: General Theory.
Wiley-Interscience, New York, 1988.
[DS88b] N. Dunford, J.T. Schwartz: Linear Operators, Part 2: Spectral Theory, Self
Adjoint Operators in Hilbert Space. Wiley-Interscience, New York, 1988.
[Eva10] L.C. Evans: Partial Differential Equations. 2nd edition. Amercian Mathematical Society, Providence (RI), 2010.
[Gil98]

D. Gilbarg, N.S. Trudinger: Elliptic Partial Differential Equations of Second Order. 2nd edition. SpringerVerlag, Berlin, 1998.

[Jos08]

J. Jost: Partielle Differentialgleichungen. SpringerVerlag, Berlin, 2008.

[Kel67]

O.D. Kellogg: Foundations of Potential Theory. SpringerVerlag, Berlin,


1967.

[Lun95] A. Lunardi: Analytic Semigroups and Optimal Regularity in Parabolic Problems. Birkh
auser, Basel, 1995.
[Ree75] M. Reed, B. Simon: Fourier Analysis, Self-Adjointness. Academic Press,
New York, 1975.
[Ren04] M. Renardy, R. C. Rogers: An Introduction to Partial Differential Equations. SpringerVerlag, New York, 2004.

89

BIBLIOGRAPHY

[Rud87] W. Rudin: Real and Complex Analysis. McGrawHill, Singapore, 1987.


[Rud91] W. Rudin: Functional Analysis. McGrawHill, Boston, 1991.
[FA]

R. Schnaubelt: Functional Analysis. Class Notes, Karlsruhe, 2010.

[Tan97] H. Tanabe: Functional Analytic Methods for Partial Differential Equations.


Dekker, New York, 1997.
[Tri97]

H. Triebel: Higher Analysis. Huthig Pub Ltd, 1997.

[Wer05] D. Werner: Funktionalanalysis. 5te erweiterte Auflage. SpringerVerlag,


Berlin, 2005.

90

You might also like