Part II

Functional Analysis II
April 3, 2013

1
Functional Analysis II

12 Spectral theory for bounded operators on complex Hilbert

spaces
12.1 Notation
We recall some basic notation and definitions.

Definition 12.1.1. Let E be a Banach space over K and let T ∈ L (E) be a continuous
linear mapping on E.

i) The spectrum of T is the set σ(T ) ⊂ K defined as

σ(T ) := {λ ∈ K : λI − T does not have a bounded inverse}.

ii) λ ∈ K is called an eigenvalue, if ker(λI − T ) 6= {0}, i.e. if there is an x 6= 0 with

T x = λx.

iii) If λ is an eigenvalue, the elements of ker(λI − T ) are called eigenvectors of T asso-

ciated with λ.

We recall the basic properties of σ(T ):

i) λ ∈ σ(T ) ⇒ |λ| ≤ kT k;

ii) σ(T ) is closed, i.e. the resolvent set ̺(T ) := K \ σ(T ) is open;

iii) If K = C, then σ(T ) 6= ∅.

We shall study linear operators on complex Hilbert spaces. Following definition recalls
the basic types of such operators.

Definition 12.1.2. Let H be a Hilbert space over C and let T ∈ L (H).

i) The operator T is called positive, if hT x, xi ≥ 0 for all x ∈ H. We sometimes write

this as T ≥ 0;

ii) There is a unique operator T ∗ ∈ L (H), such that hT x, yi = hx, T ∗ yi for all x, y ∈ H.
This operator is called the adjoint operator to T ;

iii) If T = T ∗ , T is called self-adjoint.

iv) If at least T T ∗ = T ∗ T , then T is called a normal operator.

Few basic properties of these classes of operators may be found in Exercises. We give
one example here.

Example 12.1.3. We consider the shift operator S on ℓ2 , which is defined by

S((x1 , x2 , . . . )) = (0, x1 , x2 , . . . ).

We show that

σp (S) = ∅, σap (S) = {λ ∈ C : |λ| = 1}, σ(S) = {λ ∈ C : |λ| ≤ 1},

where σp (S) is the set of eigenvalues of S (the so-called point spectrum) and σap (S) =
{λ ∈ C : inf{kSx − λxk : kxk = 1} = 0} is the approximative spectrum.

2
 12 Spectral theory for bounded operators on complex Hilbert spaces

Let Sx = λx. Then 0 = λx1 and xj = λxj+1 for j ∈ N. If λ 6= 0, this implies x1 = 0,

which implies x2 = 0 and so on, i.e. x = 0. If λ = 0, x = 0 follows as well. Hence, there
are no eigenvalues of S.
We know, that σ(S) ⊂ {λ ∈ C : |λ| ≤ 1} (as kSk = 1). Let on the other hand λ be a
complex number with |λ| ≤ 1. Then the equation

(S − λI)z = (−λz1 , z1 − λz2 , z2 − λz3 , . . . ) = (1, 0, 0, . . . )

shows that λ 6= 0 and z = (−1/λ, −1/λ2 , −1/λ3 , . . . ), which does not lie in ℓ2 . We
conclude, that there is no z ∈ ℓ2 with (S − λI)z = (1, 0, 0, . . . ), i.e. (S − λI) does not map
ℓ2 onto ℓ2 , and is therefore not invertible. Therefore, σ(S) = {λ ∈ C : |λ| ≤ 1}.
If |λ| < 1 and x ∈ ℓ2 , we have kSx−λxk ≥ kSxk−kλxk = (1−|λ|)kxk, i.e. λ 6∈ σap (S).
Finally, if |λ| = 1, we put
1
xn = √ · (1, λ−1 , λ−2 , . . . , λ−(n−1) , 0, . . . ).
n
It follows, that kxn k = 1 for all n ∈ N and

kSxn − λxn k = (2/n)1/2 → 0

as n → ∞, i.e. σap (S) = {λ ∈ C : |λ| = 1}.

The main aim of this section is to generalize the Spectral Theorem for compact self-
adjoint operators (Theorem 11.8) to bounded self-adjoint (or even normal) operators.

12.2 Functional calculus

Through the rest of this section, we shall work only with a Hilbert space H over C.
Definition 12.2.1. Let T ∈ L (H). Then W (T ) = {hT x, xi : kxk = 1} is called a
numerical range of T .
Lemma 12.2.2. Let T ∈ L (H). Then σ(T ) ⊂ W (T ).
Proof. Let λ 6∈ W (T ) and put d := dist(λ, W (T )) = inf{|λ − µ| : µ ∈ W (T )} > 0. Then
for kxk = 1 we have

d ≤ |λ − hT x, xi| = |hλx − T x, xi| ≤ k(λI − T )xk · kxk = k(λI − T )xk.

Therefore, λI − T is injective and (λI − T )−1 : ran(λI − T ) → H with norm at most 1/d.
Hence, ran(λI − T ) and H are isomorphic and ran(λI − T ) is closed.
Let us assume, that there is x0 ∈ ran(λI − T )⊥ with kx0 k = 1. Then we obtain

0 = h(λI − T )x0 , x0 i = λ − hT x0 , x0 i

and λ ∈ W (T ), which is a contradiction. Hence ran(λI − T ) = H and λ ∈ ̺(T ).

If T is self-adjoint, then hT x, xi ∈ R for every x ∈ H and the previous lemma gives

another proof of σ(T ) ⊂ R. A bit more is even true.
Corollary 12.2.3. If T ∈ L (H) is self-adjoint, then

σ(T ) ⊂ [m(T ), M (T )],

where m(T ) = inf{hT x, xi : kxk = 1} and M (T ) = sup{hT x, xi : kxk = 1}. If T ≥ 0 (i.e.

T is positive), then σ(T ) ⊂ [0, ∞).

3
Functional Analysis II

Functional calculus is a way, how to define f (T ) for (at least some) functions f and
Xn
operators T . Surely, if T ∈ L (H) and p(z) = ak z k is a polynomial, we would like to
k=0
n
X
have p(T ) = ak T k .
k=0
Moreover, for H = Cn and T self-adjoint (i.e. a hermitian matrix), we may consider
the diagonalization
T = U −1 DU,
∞
X
where U is unitary4 and D is diagonal. Let f (z) = ak z k be holomorph on the whole
k=0
C, then (at least formally)
∞ ∞ ∞ ∞
!
X X X X
k −1 k −1 k −1 k
f (T ) = ak T = ak (U DU ) = ak U D U =U ak D U
k=0 k=0 k=0 k=0
= U −1 f (D)U,

where  
f (d1 ) 0 0 ...
 0 f (d2 ) 0 ... 
 
f (D) =  . . . .. 
 . . .
. . . . 
0 0 . . . f (dn )
is again a diagonal matrix. We see, that to define f (T ), f needs to be defined only on
σ(T ) and “f (σ(T )) = σ(f (T ))”, the so-called Spectral Mapping Theorem.

Theorem 12.2.4. (Continuous Functional Calculus) Let T ∈ L (H) be self-adjoint. Then

there is a unique mapping
Φ : C(σ(T )) → L (H),
such that

i) Φ(z)5 = T, Φ(1) = I;

ii) Φ is an involutive homomorphism of algebras, i.e.

a) Φ is linear;
b) Φ is multiplicative: Φ(f g) = Φ(f ) ◦ Φ(g);
c) Φ is involutive: Φ(f )∗ = Φ(f );

iii) Φ is continuous.

Φ is called continuous functional calculus of T and we denote f (T ) := Φ(f ) for f ∈

C(σ(T )).

Proof. Uniqueness: Φ(z n ) = T n and Φ is unique on all polynomials. Furthermore, σ(T ) ⊂

[m(T ), M (T )] is compact and polynomials are dense in C(σ(T )). 6 Due to continuity, Φ
is unique on all C(σ(T )).
4
i.e. U ∗ U = U U ∗ = I
5
By z we denote of course the identity mapping z → z on C and by 1 the mapping z → 1.
6
We assume, that you know, that polynomials are dense in C(I) for every bounded interval I.

4
 12 Spectral theory for bounded operators on complex Hilbert spaces

n
X n
X
Existence: We set Φ(f ) = ak T k for f (z) = ak z k a polynomial. If we show continuity
k=0 k=0
of Φ on polynomials, then there would be a unique extension, which we denote by Φ again.
We shall use the Spectral Mapping Theorem for polynomials (cf. Exercise 1.4)

σ(Φ(f )) = f (σ(T )) = {f (λ) : λ ∈ σ(T )}

to obtain

kΦ(f )k2 = kΦ(f )∗ Φ(f )k = kΦ(f f )k

= sup{|λ| : λ ∈ σ(Φ(f f ))}
= sup{|f f (λ)| : λ ∈ σ(T )}
= sup{|f (λ)|2 : λ ∈ σ(T )} = kf k2∞ .

We refer to Exercise 1.1.2 for the first equality, the second equality follows from a simple
calculation for polynomials, the third one is based on Exercise 2.3 and the fact that Φ(f f )
is self-adjoint.
All the required properties are then proven by approximation. Let us assume (for
example) that pn ⇉ f and qn ⇉ g for some polynomials pn , qn and f, g ∈ C(σ(T )). Then
we get
Φ(f g) ← Φ(pn qn ) = Φ(pn ) ◦ Φ(qn ) → Φ(f ) ◦ Φ(g).
The proof of the other properties is similar.

Theorem 12.2.5. (Properties of the continuous functional calculus)

Let T ∈ L (H) be self-adjoint and let f → f (T ) be the continuous functional calculus for
f ∈ C(σ(T )). Then

i) kf (T )k = kf k∞ := supλ∈σ(T ) |f (λ)|.

ii) If f ≥ 0 on σ(T ), then f (T ) ≥ 0, i.e. f (T ) is positive.

iii) If T x = λx for some x ∈ H, then also f (T )x = f (λ)x.

iv) σ(f (T )) = f (σ(T )), i.e. the Spectral Mapping Theorem holds for all f ∈ C(σ(T )).

v) {f (T ) : f ∈ C(σ(T ))} is a commutative Banach algebra of operators.

vi) All f (T ) are normal; if f is real, then f (T ) is self-adjoint.

Proof. (i) was proven for polynomials already before. For general pn ⇉ f we have
kΦ(f )k ← kΦ(pn )k = kpn k∞ → kf k∞ .
Let f ≥ 0 and take g ∈ C(σ(T )) with g 2 = f and g ≥ 0. Then

hf (T )x, xi = hg(T )x, g(T )∗ xi = hg(T )x, g(T )xi = hg(T )x, g(T )xi = kg(T )xk2 ≥ 0.

(iii) is again clear for polynomials, through approximation it follows for all f ∈ C(σ(T )).
Also (iv) was already discussed for polynomials. Let µ 6∈ f (σ(T )). Then g := (f − µ)−1 ∈
C(σ(T )) and g(f − µ) = (f − µ)g = 1. Hence, we get

g(T )(f (T ) − µI) = (f (T ) − µI)g(T ) = I.

5
Functional Analysis II

Hence µ ∈ ̺(f (T )). This shows the “⊂” part in (iv).

Let on the other hand µ = f (λ) for some λ ∈ σ(T ) and choose polynomials pn with
kpn − f k∞ ≤ 1/n. Then
|f (λ) − pn (λ)| ≤ 1/n and kf (T ) − pn (T )k ≤ 1/n.
We know that pn (λ) ∈ σ(pn (T )), i.e. there exists xn ∈ H with kxn k = 1 and k(pn (T ) −
pn (λ)I)xn k ≤ 1/n. Finally, from
k(f (T ) − µI)xn k ≤ k(f (T ) − pn (T ))xn k + k(pn (T ) − pn (λ)I)xn k + k(pn (λ) − µ)Ixn k
≤ 1/n + 1/n + 1/n = 3/n
we see, that f (T ) − µI is not cont. invertible, i.e. µ ∈ σ(f (T )). Finally, (v) is clear now
and (vi) follows from f (T )∗ = f (T ) and f (T )∗ f (T ) = f f (T ) = f f (T ) = f (T )f (T )∗ .

Let us now have a look on the spectral decomposition theorem for compact self-adjoint
operators and how it could be generalized to bounded operators. One obvious obstacle
would be that the spectrum does not have to be countable any more, i.e. we will have to
replace the sums by certain (appropriately interpreted) integrals. Furthermore, we have
now spent some time to define f (T ) for (at least some) functions f on σ(T ). We therefore
rewrite
X∞ ∞
X
Tx = λk hx, xk ixk = λk Ek x,
k=0 k=0
where Ek x = hx, xk ixk is the projection of x onto the linear span of xk . And if we put
fk (λj ) = δj,k , we get even
∞
X
fk (T )x = fk (λn )hx, xn ixn = hx, xk ixk = Ek x,
n=0

where (for simplicity) we assumed that all the λk ’s are different.

We will therefore write the spectral decomposition of a bounded self-adjoint operator as
an integral with respect to a certain system of projections. Unfortunately, fk do not have
to be continuous on σ(T ). We therefore have to extend the definition of f (T ) to bounded
measurable functions f . This shall be done through duality. For this sake, we shall need
(another) Riesz Representation Theorem, and for this we shall need some notation.
Let (K, τ ) be a compact topological space. We denote by B the smallest σ-algebra con-
taining τ (the so-called Borel σ-algebra). As usual, f : K → C is called Borel measurable,
if f −1 (U ) ∈ B for every U ⊂ C open.
The interaction between measures on B and the topology τ is described by the following
regularity conditions.
A finite measure µ on (K, B) is said to be a Radon measure, if
µ(B) = inf{µ(G) : G ⊃ B, G open} for every B ∈ B,
µ(G) = sup{µ(T ) : T ⊂ G, T compact} for every open G ∈ τ.
Signed Radon measure is a σ-additive mapping on B with µ = µ+ − µ− , where both µ+
and µ− are Radon measures. Unfortunately, this representation is not unique. Indeed, if
ν is another bounded positive Radon measure, we get µ = µ+ − µ− = (µ+ + ν) − (µ− + ν).
Nevertheless, one of these representations is “minimal” (i.e. µ+ and µ− are singular to
each other).7
7
The decomposition of the measure space goes under the name Hahn decomposition and the decompo-
sition of the measure is known as Jordan or Hahn-Jordan decomposition.

6
 12 Spectral theory for bounded operators on complex Hilbert spaces

Using this minimal decomposition, we denote by |µ| = µ+ + µ− the total variation of

µ. If µ is complex-valued, we furthermore decompose it into the real and complex part.
Finally, we denote by M(K) the space of all signed Radon measures on K with
kµkM(K) = |µ|(K).
Theorem 12.2.6. (Yet Another Riesz Representation Theorem)

C(K)∗ ≈ M(K),
R
where µ(f ) = K f dµ. Furthermore, positive functionals (i.e. those with f ≥ 0 ⇒ ϕ(f ) ≥
0) correspond to non-negative measures.
Lemma 12.2.7. Let H be a complex Hilbert space and let Q : H → C be a function. Then
the following are equivalent.
i) There exists exactly one operator A ∈ L (H) such that Q(x) = hAx, xi for all x ∈ H.

ii) There is a constant C > 0 such that |Q(x)| ≤ Ckxk2 , Q(x + y) + Q(x − y) =
2Q(x) + 2Q(y) and Q(λx) = |λ|2 Q(x) for all x, y ∈ H and all λ ∈ C.
Proof. It is easy to show that (i) ⇒ (ii). If (ii) is satisfied, we put (motivated by polar-
ization identity)

Ψ(x, y) = 1/4(Q(x + y) − Q(x − y) + iQ(x + iy) − iQ(x − iy))

and X
Ax = Ψ(x, eγ )eγ ,
γ∈Γ

where (eγ )γ∈Γ is some orthonormal basis of H. The rest follows by simple calculations.
Finally, let us observe that Ψ(x, y) = hAx, yi. Uniqueness of A follows from Exercises.

Remark 12.2.8. (The main idea of Borel-measurable functional calculus)

Let T ∈ L (H) be self-adjoint and let x ∈ H be fixed. We consider the mapping

f → hf (T )x, xi, C(σ(T )) → C.

This mapping is linear, non-negative (i.e. f ≥ 0 implies hf (T )x, xi ≥ 0 for all x ∈ H).
Therefore, there exists a non-negative Radon measure E x on σ(T ), such that
Z
hf (T )x, xi = f dE x for all f ∈ C(σ(T )).
σ(T )
Z
If g is a bounded Borel-measurable function on σ(T ), the integral gdE x converges
σ(T )
and we shall show (using the Lemma above) that it is equal to some hGx, xi for all x ∈ H
and some G ∈ L (H). This G will then be defined to be g(T ).
Let us now elaborate on the program just stated.
Theorem 12.2.9. Let T ∈ L (H) be self-adjoint.
i) Let x ∈ H. Then there is a non-negative Radon measure E x , such that
Z
hf (T )x, xi = f dE x
σ(T )

for every f ∈ C(σ(T )).

7
Functional Analysis II

ii) Let g be a bounded Borel-measurable function on σ(T ). Then there is a unique

G ∈ L (H) with Z
hGx, xi = gdE x
σ(T )

for all x ∈ H. If g is real, G is self-adjoint, if g ≥ 0, G is positive.

Proof. The proof is based on Lemma 12.2.7. Obviously, the operator is unique (hGx, xi =
hG̃x, xi for all x ∈ H implies G = G̃). Furthermore, g ≥ 0 implies hGx, xi ≥ 0, i.e.
positivity, and g real implies hGx, xi is real, i.e. G is self-adjoint.
So, we have to prove the existence. We put
Z
Q(x) = gdE x .
σ(T )

Then
Z Z
x
|Q(x)| ≤ kgk∞ dE = kgk∞ 1dE x = kgk∞ h1(T )x, xi = kgk∞ kxk2 . (12.1)
σ(T ) σ(T )

For every f ∈ C(σ(T )), we obtain

Z Z Z Z
x+y x−y x
f dE + f dE =2 f dE + 2 f dE y
σ(T ) σ(T ) σ(T ) σ(T )
| {z } | {z } | {z } | {z }
hf (T )(x+y),x+yi hf (T )(x−y),x−yi 2hf (T )x,xi 2hf (T )y,yi

and Z Z
f dE λx = |λ|2 f dE x
σ(T ) σ(T )
| {z } | {z }
hf (T )(λx),λxi |λ|2 ·hf (T )x,xi

Due to the uniqueness of the Riesz Representation Theorem, we get E x+y + E x−y =
2E x + 2E y and E λx = |λ|2 E x . This verifies the assumptions of Lemma 12.2.7 and finishes
the proof.

Yet another theorem from the measure theory we shall use is that the set of bounded
Borel-measurable functions is the smallest set, which includes bounded continuous func-
tions and is closed under pointwise limits of such functions.

Theorem 12.2.10. Let K ⊂ C be compact. Let B d (K) be the Banach space of bounded
Borel-measurable functions on K equipped with the supremum norm. Let C(K) ⊂ U ⊂
B d (K) be a set of functions with the following property

(fn )n∈N ⊂ U with f (t) := lim fn (t) existing everywhere and sup kfn k∞ < ∞ ⇒ f ∈ U.
n→∞ n∈N

Then U = B d (K).

Proof. The proof is based on measure theory. We refer to Werner, Lemma VII.1.5 for
details.

Theorem 12.2.11. (Borel measurable functional calculus) Let T ∈ L (H) be self-adjoint.

Then there exists a unique mapping Φ : B d (σ(T )) → L (H) with

i) Φ(z) = T, Φ(1) = I;

8
 12 Spectral theory for bounded operators on complex Hilbert spaces

ii) Φ is an involutive homomorphism of algebras;

iii) Φ is continuous;

iv) fn ∈ B d (σ(T )) with supn kfn k∞ < ∞ and fn (t) → f (t) for every t ∈ σ(T ) implies
hΦ(fn )x, yi → hΦ(f )x, yi for every x, y ∈ H.

Proof. Uniqueness: (i), (ii) and (iii) imply uniqueness on C(σ(T )) and Theorem 12.2.10
on all B d (σ(T )).
Existence: We define Φ(g) = G, where G is the operator from Theorem 12.2.9. We have
to verify (i) − (iv). We know that
Z
hf (T )x, xi = f dE x
σ(T )

for all f ∈ C(σ(T )), i.e. Φ(f ) = f (T ) for all f ∈ C(σ(T )), where f (T ) is the continuous
functional calculus. This implies (i).
Let g be real valued. Then we obtain

kΦ(g)k = kGk = sup{|hGx, xi| : kxk = 1} ≤ sup{kgk∞ · kxk2 : kxk = 1} = kgk∞ ,

where we have used (12.1), the fact, that G is self-adjoint and Corollary 11.6. If g is
complex-valued, we split it first into its real and imaginary part.
The proof of (iv) follows from
Z Z
x
hΦ(fn )x, xi = fn dE → f dE x = hΦ(f )x, xi,
σ(T ) σ(T )

which holds by Lebesgue Dominated Convergence Theorem, and by polarization identity.

(ii) follows from a bit tricky limit procedure, which is quite often used in measure theory.
For example, to show that Φ(f g) = Φ(f ) ◦ Φ(g), we proceed as follows. If f, g ∈ C(σ(T )),
then the result is known. If g ∈ C(σ(T )), we set U := {f ∈ B d (σ(T )) : Φ(f g) =
Φ(f ) ◦ Φ(g)}. Then we know that C(σ(T )) ⊂ U and U is closed under pointwise limits of
uniformly bounded sequences. According to Theorem 12.2.10, U = B d (σ(T )).
Finally, if f ∈ B d (σ(T )), we put V := {g ∈ B d (σ(T )) : Φ(f g) = Φ(f )◦Φ(g)}. We have just
shown that C(σ(T )) ⊂ V , furthermore, V is again closed on pointwise limits of uniformly
bounded functions, i.e. V = B d (σ(T )) and the statement is true for all f, g ∈ B d (σ(T )).
Other properties follow in the same way.

Remark 12.2.12. (iv) in Theorem 12.2.11 can be improved to fn (T )x → f (T )x for every

x ∈ H. This follows from

kfn (T )xk2 = hfn (T )x, fn (T )xi = hfn (T )∗ fn (T )x, xi

= h(f n fn )(T )x, xi → h(f f )(T )x, xi = kf (T )xk2

w
and fn (T )x −
→ f (T )x.

Remark 12.2.13. (Analytic functional calculus) We were able to define f (T ) for

(rather) ugly functions (i.e. all Borel-measurable functions) and very nice operators (i.e.
self-adjoint operators on a Hilbert space over C). It comes probably as no surprise that
for nicer function we can define f (T ) also for more general classes of operators.

9
Functional Analysis II

∞
X
Let X be a Banach space and let T ∈ L(X). Let f (z) = an z n be a potential series
n=0
with convergence radius bigger then r(T ). Then f is analytic on σ(T ) and we can define
∞
X
f (T ) := an T n .
n=0

One can show that

The series for f (T ) converges in L(X) and f (T ) is therefore well-defined.

∞
X
If g(z) = bn z n is another potential series with radius of convergence bigger then
n=0
r(T ), then (f g)(T ) = f (T )g(T ).

The Spectral mapping theorem σ(f (T )) = f (σ(T )) holds.

If T is a self-adjoint operator on a Hilbert space over C, then this definition coincides

with the continuous (and also with the Borel-measurable) functional calculus.

Remark 12.2.14. (Dunford’s analytic functional calculus) The analytic functional

calculus described above allows to study spectral theory also in the frame of Banach spaces.
Nevertheless, the basis for such a theory is different. If f is holomorph on a neighborhood
Ω of σ(T ) (i.e. on an open set containing σ(T )), then we may define
I
1
f (T ) := f (λ)Rλ (T )dλ,
2πi Γ

where Γ ⊂ Ω is a suitable curve, which goes around σ(T ) exactly one times, and Rλ (T ) is
the resolvent mapping of T . Of course, this idea is heavily inspired by the Cauchy integral
known from complex analysis.

12.3 Spectral theorem for bounded self-adjoint operators

For spectral theorem, f (T ) is especially important for f = χA , where A ⊂ R is a Borel
measurable set.

Lemma 12.3.1. Let T ∈ L (H) be self-adjoint and let A ⊂ σ(T ) be a Borel set. Then
EA := χA (T ) is an orthogonal projection.

Proof. Due χ2A = χA , we have EA = EA

2 and E ∗ = E follows from χ = χ and Exercise
A A A A
3.1.

Lemma 12.3.2. Let T ∈ L (H) be self-adjoint. Then

i) E∅ = χ∅ (T ) = 0, Eσ(T ) := χσ(T ) (T ) = I;

ii) For pairwise disjoint sets A1 , A2 , · · · ⊂ σ(T ), and for x ∈ H, we have

∞
X
χAi (T )x = χ∪∞ A (T )x.
i=1 i
i=1

iii) χA (T )χB (T ) = χA∩B (T ) for Borel sets A, B ⊂ σ(T ).

10
 12 Spectral theory for bounded operators on complex Hilbert spaces

Proof. The proof follows from the properties of the Borel-measurable functional calculus.

Remark 12.3.3. Let us note, that

∞
X
χAi (T ) 6= χ∪∞ A (T ),
i=1 i
i=1

i.e. the identity must be interpreted in the “weak sense”.

Definition 12.3.4. Let Σ be the σ-algebra of the Borel sets of R. A mapping

E : Σ → L (H), E(A) = EA

is called a spectral measure if all EA are orthogonal projections and

i) E∅ = 0, ER = I;

ii) For pairwise disjoint sets A1 , A2 , · · · ∈ Σ, and for x ∈ H, we have

∞
X
EAi (x) = E∪∞ A (x).
i=1 i
i=1

A spectral measure E has compact support, if there is a compact set K with EK = I.

Let us state some simple properties of spectral measures.
i) EA + EB = EA∩B + EA\B + EB = EA∩B + EA∪B holds for every A, B ⊂ R.

ii) If A ⊂ B ⊂ R, then EB\A = EB − EA is itself a projection, i.e. EB ≥ EA and (with

the help of Exercises) EB EA = EA EB = EA .
2 +E E = E E
iii) If A and B are disjoint, we get EA + EB = EA∪B , i.e. EA A B A A∪B = EA
due to the previous point; hence EA EB = 0.
iv) EA EB = EA (EA∩B + EB\A ) = EA EA∩B + EA EB\A = EA∩B holds (due to the
previous points) for all A, B ⊂ R.
∗ E x, yi = hE E x, yi =
v) Finally, if A, B ⊂ R are disjoint, then hEA x, EB yi = hEB A B A
h0, yi = 0, i.e. EA and EB project onto mutually orthogonal subspaces.
“In some sense, E is a measure with values in L (H) instead of [0, ∞).”
For the
R sake of spectral decomposition theorem, we need to elaborate a way, how to
calculate f dE ∈ L (H) for E a spectral measure and a complex-valued f . Of course, for
Xn
R
f = χA , we define f dE = E(A) = EA ∈ L (H). For f = αi χAi a simple function,
i=1
n
X
R
we put f dE = αi EAi ∈ L (H). Finally, for a general bounded Borel-measurable
i=1
function f we consider uniformly convergent sequence fn ⇉ f of simple functions and
define Z Z
f dE = lim fn dE.
n→∞

We have to check, that this definition is consistent. It is quite easy to see that the
integral of a simple function f does not depend on the form in which f was written. The
limit procedure is then justified by the following Lemma:

11
Functional Analysis II

Lemma 12.3.5. For a simple function f , we have

Z

f dE ≤ kf k∞ .

n
X
Proof. Let kxk = 1 and let f = αi χAi with mutually disjoint A1 , . . . , An . Then
i=1

Z  2 Xn
2
n
X

f dE (x) = αi EAi (x) = |αi |2 · kEAi (x)k2

i=1 i=1
n
X
≤ max |αi |2 · kEAi (x)k2 ≤ kf k2∞ · kxk2 = kf k2∞ .
i=1,...,n
i=1

R This lemma shows that if (fn )n∈N is a Cauchy-sequence in k · k∞ (i.e. fn ⇉ f ), then

( fn dE)n is a Cauchy-sequence in L (H), Rits limit exists and it does not depend on the
choice of the sequence (fn )n∈N . Therefore, f dE is well defined.
We summarize our findings in the following theorem:

Theorem 12.3.6. (Properties of integration

R with respect to the spectral measure)
Let E be a spectral measure. Then f dE ∈ L (H)R exists for every bounded Borel mea-
surable
Z function
f ∈ B d (R). The mapping f → f dE is linear and continuous, i.e.
R
f dE ≤ kf k∞ . If f is real, then f dE is self-adjoint. If K ⊂ R is compact with

EK = I, then it is enough if f is defined and bounded on K.

Remark 12.3.7. On one hand, we have for each T ∈ L (H) self-adjoint the spectral
measure EA := χA (T ). On the otherR hand, we can define for every spectral measure with
compact support the operator T := λdEλ , i.e. we integrate the identity function λ → λ
with respect to E. We shall show that these two operations are actually inverse to each
other.

Theorem 12.3.8. Let E be a spectral measure R on R with compact support. Let T ∈ L (H)
be the self-adjoint operator defined by T = λdEλ . Then the mapping
Z
Ψ:f → f dE, Ψ : B d (σ(T )) → L (H)
σ(T )

is the Borel-measurable functional calculus from Theorem 12.2.11. Especially, we have

Eσ(T ) = I.

Proof. We sketch first the proof that Eσ(T ) = I. Let f (λ) := λ and let fδ ⇉ f be simple
n(δ)
X
functions fδ = αi,δ χAi,δ with |fδ (λ) − λ| ≤ δ on the support of E and A1,δ , . . . , An(δ),δ
i=1 Z Z


mutually disjoint. Let µ ∈ ̺(T ). As ̺(T ) is open and fδ dE − T = (fδ (λ) − λ)dE

≤
R R
R Pn(δ)
δ, we conclude, that R fδ dE − µI = i=1 αi,δ EAi,δ − µI is invertible for 0 < δ ≤ δ0 (for
δ0 > 0 small enough). On the other hand, the inverse of such an operator has a norm
max{|αi,δ − µ|−1 : 1 ≤ i ≤ n(δ), EAi,δ 6= 0} → k(T − µI)−1 k. Hence, EAi,δ = 0 for

12
 12 Spectral theory for bounded operators on complex Hilbert spaces

|αi,δ − µ| small and we conclude that EU = 0 for some small neighborhood of µ ∈ U.

Hence E̺(T ) = 0 and Eσ(T ) = I.
To conclude the proof, we have to verify that the mapping Ψ has all the properties de-
scribed in Theorem 12.2.11. We know already, that it is linear, continuous, multiplicative
(start with characteristic functions, then simple functions and then limits) and involutive.
The condition (iv) from Theorem 12.2.11 reads in our case as
Z   Z  
fn dE x, y → f dE x, y

R
and follows from the fact, that the mapping g →R h( gdE)x, yiRis in C(σ(T ))∗ and may be
therefore represented by a measure νx,y , i.e. h( gdE)x, yi = gdνx,y . The proof is then
finished by Lebesgue dominated convergence theorem.

Hence, integrationR with respect to a spectral measure E is a functional calculus of the

self-adjoint operator λdEλ . The spectral theorem says, that actually every self-adjoint
operator may be written in this way - i.e. there is a one-to-one correspondence between
bounded self-adjoint operators and spectral measures with compact support.

Theorem 12.3.9. (Spectral theorem for bounded self-adjoint operators)

Let T ∈ L (H) be self-adjoint. Then there exists a unique spectral measure E on R with
compact support with Z
T = λdEλ .
σ(T )
R
The mapping f → f (T ) = f (λ)dEλ coincides with the Borel-measurable functional
calculus and Z
hf (T )x, yi = f (λ)dhEλ x, yi.
σ(T )

By hEλ x, yi we denote the (complexed-valued) measure A → hEA x, yi.

Proof. Let E be the spectral measure R associated to T as described above, i.e. E : A →

χA (T ). We have to show, that S := σ(T ) λdEλ is equal to T . Let f be a simple function
with kf (λ) − λk∞ = supλ∈σ(T ) |f (λ) − λ| ≤ δ. Then

kT − Sk ≤ kT − f (T )k + kf (T ) − f (S)k + kf (S) − Sk.

We estimate all the three summands as follows. Let Ψ be the Borel-measurable functional
calculus of T . Then Ψ(λ) = T and Ψ(f ) = f (T ). Hence

kT − f (T )k = kΨ(λ) − Ψ(f )k = kΨ(λ − f (λ))k ≤ kf (λ) − λk∞ ≤ δ,

where we have used the boundedness of Borel-measurable P functional calculus.

Next we estimate kf (T ) − f (S)k. We write f (t) = i αi χAi (t) with Ai ⊂ R disjoint.
Then, by definition of EAi , we get.
X X
f (T ) = αi χAi (T ) = αi EAi .
i i

Furthermore, according to Theorem 12.3.8 applied to S, we know that

Z Z X X
f (S) = f (λ)dEλ = αi χAi (λ)dEλ = αi EAi . (12.2)
R R i i

13
Functional Analysis II

Hence, kf (T ) − f (S)k = 0.
Finally, we estimate kf (S)−Sk. We use again (12.2) combined with the norm estimate
from Theorem 12.3.6 and obtain
Z

kf (S) − Sk =
(f (λ) − λ)dEλ ≤ kf (λ) − λk∞ ≤ δ

Hence, kS − T k ≤ δ for δ > 0 arbitrary small, i.e.R S = T.

The properties of the mapping f → f (T ) := σ(T ) λdEλ follow from Theorem 12.3.8,
once the identity S = T was proven. Finally, the identity for hf (T )x, yi is a matter of
definition for f = χA , and also for simple functions. Finally, taking a limit, it follows for
every bounded Borel measurable f .

Let us summarize the structure of this section into the following diagram.

T ∈L (H)
Weierstraß Theorem Riesz
self-adjoint −−−−−−−−−−−−−−→ Cont. F.C.(Thm. 12.2.4) −−−−→ Borel FC (Thm. 12.2.11)
density of polynomials


EA :=χA (T )y
R
T ∈L (H) T := λdEλ R Thm. 12.3.6
self-adjoint ←−−−−−−− f dE ←−−−−−−− Spectral measure
Thm. 12.3.8

12.4 Few applications of spectral theorem

Theorem 12.4.1. Let S ∈ L (H) be self-adjoint, g : σ(S) → R and f : R → R be
Borel-measurable and bounded. Then

(f ◦ g)(S) = f (g(S)).

Proof. First, note that f ◦ g is again Borel-measurable. Furthermore, g(S) is self-adjoint

and, therefore, f (g(S)) can be defined.
It is enough to consider f = χA , the rest follows as usually. Then

χA ◦ g = χg−1 (A)

and we have to show that χg−1 (A) (S) = χA (g(S)). Let F be the spectral measure of S
and E the spectral measure of g(S); then we have to show that Fg−1 (A) = EA for all Borel
sets A, or, equivalently,
hFg−1 (A) x, yi = hEA x, yi
for all x, y ∈ H and all Borel sets A.
Let x, y ∈ H be fixed and let us consider the measures
1
νx,y : A → hFg−1 (A) x, yi,
2
νx,y : A → hEA x, yi,
µx,y : A → hFA x, yi.
1 (A) = µ
Hence, νx,y −1
x,y (g (A)). We use the transformation law for the (complex-valued)
1
measures νx,y and µx,y
Z Z
1
hdνx,y = (h ◦ g)dµx,y

14
 12 Spectral theory for bounded operators on complex Hilbert spaces

in the form
Z Z Z
1
h(λ)dνx,y (λ) = (h ◦ g)dµx,y = h(g(λ))dhFλ x, yi

and obtain
Z Z Z Z
λ dνx,y (λ) = g(λ) dhFλ x, yi = hg(S) x, yi = λ dhEλ x, yi = λn dνx,y
n 1 n n n 2
(λ).

1 and ν 2 coincide on polynomials, and due to Theorems of Riesz and

The measures νx,y x,y
Weierstraß, they coincide on C(σ(g(S))) and also as measures.

Remark 12.4.2. We recall the notion of the push-forward measure. Let µ be a measure
on X1 and f : X1 → X2 a mapping. Then f∗ (µ)(B) = µ(f −1 (B)) for B ⊂ X2 is
called the push-forward measure of µ under f . To develop the complete theory (which
includes the transformation law we have used just now) one has to take care also about
the corresponding σ-algebras.

Corollary 12.4.3. Let T ∈ L (H) be positive and let n ∈ N. Then there exists a unique
positive S ∈ L (H) with S n = T.

Proof. The functions fn : t → t1/n are continuous, bounded and non-negative on σ(T ) ⊂
[0, ∞). We set S := fn (T ). Then S ≥ 0 and from t1/n . . . t1/n = t, it follows S n = T. Let
gn (t) = tn , then (fn ◦ gn )(t) = t for t ∈ [0, ∞) ⊃ σ(S) and

S = (fn ◦ gn )(S) = fn (gn (S)) = fn (S n ) = fn (T )

shows the uniqueness of S.

Corollary 12.4.4. (Polar decomposition)

To an operator T ∈ L (H), we define |T | = (T ∗ T )1/2 . Furthermore, there is a partial
isometry U with T = U |T |.

Proof. (Proof is the same as for compact operators).

We see that k|T |xk2 = hx, T ∗ T xi = hT x, T xi = kT xk2 . Therefore U (|T |x) = T x is an
isometry of ran |T | to ran T . This can be extended to U : ran |T | → ran T . Finally, we put
U = 0 on (ran |T |)⊥ = ker |T | = ker T.

12.5 Spectral theorem for normal operators

We shall proceed similarly to the case of compact normal operators (cf. Exercises).
Let T ∈ L (H) be a normal operator. Using

T + T∗ T − T∗
T = +i ,
2 } | {z
| {z 2i }
=:T1 =:T2

we can decompose T = T1 +iT2 , where T1 and T2 are self-adjoint and commuting operators.
Let E be the spectral measure associated with T1 and F the spectral measure associated
with T2 . We shall show that EA FB = FB EA for all sets A, B ⊂ R. We show first that
EA T2 = T2 EA . The same argument repeated for the pair (EA , T2 ) instead of (T1 , T2 ) then
gives the general statement.

15
Functional Analysis II

We observe that T1n T2 = T2 T1n , i.e. p(T1 )T2 = T2 p(T1 ) for all polynomials and, by
limiting arguments, also for all ϕ ∈ C(σ(T1 )). Finally, we consider ϕn → χA , which
implies ϕn (T1 )x → EA (x) for every x ∈ H. This finally gives

T2 (EA x) = T2 ( lim ϕn (T1 )x) = lim T2 (ϕn (T1 )x) = lim ϕn (T1 )(T2 x) = EA (T2 x).
n→∞ n→∞ n→∞

If Ω ⊂ C is a product of two sets, i.e. Ω = {x + iy : x ∈ A and y ∈ B} for some

A, B ⊂ R, we put G(Ω) := EA FB = FB EA . Using standard arguments from measure
theory, G can be extended to a σ-additive mapping on the Borel sets of C. Let us note,
that it has also a compact support.
From
Z Z Z Z
T1 = λdEλ = Re zdGz and T2 = λdFλ = Im zdGz ,
R C R C

we get Z
T = T1 + iT2 = zdGz .
C
This finishes a sketch of the proof of the following theorem:

Theorem 12.5.1. (Spectral theorem for bounded normal operators)

Let T ∈ L (H) be normal. Then there exists a unique spectral measure G on C with
compact support with Z
T = λdGλ .
σ(T )
R
The mapping f → f (T ) = f (λ)dGλ defines a Borel-measurable functional calculus and
Z
hf (T )x, yi = f (λ)dhGλ x, yi.
σ(T )

16
13 Spectral theory for unbounded operators
Also in this section, H denotes a Hilbert space over C.

13.1 Definitions and Motivation

Many important operators on Hilbert spaces appearing for example in physics are not
bounded. This is especially the case for many differential operators. Therefore, one is
interested in analysis of operators defined on dom(T ), which is only a subspace of H. Of
course, if T ∈ L (H), then dom(T ) = H. On the other hand, if T is closed and defined on
all H, it is bounded by the closed graph theorem. Hence, an unbounded closed operator
is never defined on the whole H. Many (unbounded) operators are closed (or at least
closable) and, therefore, the flexibility of dom(T ) 6= H is essential for the theory.

Definition 13.1.1. i) The linear mapping T : dom(T ) ⊂ H → H is called an operator,

if dom(T ) is a linear subspace of H. We say, that T is densely defined if dom(T ) = H.

ii) Graph of an operator T is the set G(T ) = {(T x, x) : x ∈ dom(T )} ⊂ H × H,

where H × H = {(x, y) : x, y ∈ H} is a Hilbert space with the scalar product
h(u, v), (x, y)iH×H = hu, xi + hv, yi.

iii) We call T closed operator if G(T ) is closed (in H × H).

iv) We say that T is closable if G(T ) is a graph of some linear operator T0 . This operator
is then called closure of T . We shall denote this by T0 = T .

v) An operator S : dom(S) ⊂ H → H is called extension of T if G(T ) ⊂ G(S) or,

equivalently, if dom(T ) ⊂ dom(S) and Sx = T x for all x ∈ dom(T ). We shall denote
this by T ⊂ S.

vi) T = S if T ⊂ S and S ⊂ T , i.e. if dom(T ) = dom(S) and Sx = T x for all

x ∈ dom(T ).

The domain of T , i.e. dom(T ), is an essential part of the definition of every operator
T . For example, if S and T are two (unbounded) operators on H, we set

dom(S + T ) = dom(S) ∩ dom(T ), (S + T )(x) = Sx + T x,

dom(ST ) = {x ∈ dom(T ) : T x ∈ dom(S)}, (ST )(x) = S(T (x)).

Definition 13.1.2. An operator T : dom(T ) → H is called symmetric if

hT x, yi = hx, T yi for all x, y ∈ dom(T ).

Theorem 13.1.3. (Hellinger-Toeplitz) Let T be a symmetric operator defined on whole

dom(T ) = H. Then T is bounded, i.e. T ∈ L (H).

Proof. Let xn → 0 and T xn → z. We show that z = 0 and apply the Closed graph
theorem. But this follows easily from
D E
hz, zi = lim T xn , z = lim hT xn , zi = lim hxn , T zi = 0.
n→∞ n→∞ n→∞

Let us mention that in this case T = T ∗ .8

8
Where have we used that dom(T ) = H?

17
Functional Analysis II

Next we define the adjoint of a densely defined operator T .

Definition 13.1.4. Let T be a densely defined operator. We define

dom(T ∗ ) = {y ∈ H : x → hT x, yi is continuous on dom(T )}.

For y ∈ dom(T ∗ ), the mapping x → hT x, yi may be extended to a continuous mapping on

all H and (according to the Riesz theorem) represented as x → hx, zi, z ∈ H. We denote
this z by T ∗ y. Due to the density of dom(T ), z is unique. T ∗ is called the adjoint operator
to T . If T = T ∗ , then T is called self-adjoint.

We have hT x, yi = hx, T ∗ yi for all x ∈ dom(T ) and all y ∈ dom(T ∗ ).

Example 13.1.5. Let H = L2 (0, 1) and define

dom(T ) = C0∞ (0, 1) = {f ∈ C ∞ (0, 1) : f has compact support},

T f = if ′ .

Since C0∞ (0, 1) is dense in L2 (0, 1), T is densely defined. It is easy to check, that T is
symmetric:
Z 1 Z 1 Z 1
′ ′
hT f, gi = if (t)g(t)dt = −i f (t)g (t)dt = f (t)ig′ (t)dt = hf, T gi.
0 0 0

However, T is not self-adjoint - the same calculation goes through also for f, g ∈ C01 (0, 1),
which means, that f → hT f, gi is continuous also for g ∈ C01 (0, 1) ⊂ dom(T ∗ ).
Therefore, there comes a question: what is the closure of T and the adjoint of T ?

Remark 13.1.6. In the sequel, we shall encounter the space AC([0, 1]) of absolutely
continuous functions on [0, 1]. Let us recall their definition and basic facts about them.
A function f : [0, 1] → R (or C) is called absolutely continuous, if for every ε > 0 there is
a δ > 0, such that
Xm
|f (bj ) − f (aj )| < ε
j=1
P
for every 0 ≤ a1 < b1 ≤ a2 < b2 ≤ · · · ≤ am < bm ≤ 1 with j (bj − aj ) < δ. Obviously, ev-
ery Lipschitz function on [0, 1] is absolutely continuous. Using the notion of total variation,
one shows that every absolutely continuous function is a difference of two nondecreasing
absolutely continuous functions. This, combined with the fact that nondecreasing func-
tions are differentiable almost everywhere, then gives that absolutely continuous functions
are also differentiable almost everywhere, i.e. it makes sense to ask if their derivative is
an element of L1 ([0, 1]).
We shall also need the following properties:
R1 R1 ′
(a) (partial integration): If f, g ∈ AC([0, 1]), then 0 f g′ = [f (x)g(x)]x=1x=0 − 0 f g.

(b) f ∈ AC([0, 1]) if, and

R x only if, f is differentiable almost everywhere, f ′ ∈ L1 ([0, 1])
and f (x) = f (0) + 0 f ′ (t)dt.

Example 13.1.7. Let H = L2 ([0, 1]). We define

dom(T ) = {f ∈ AC([0, 1]) : f ′ ∈ L2 ([0, 1]) and f (0) = f (1) = 0}.

18
 13 Spectral theory for unbounded operators

The set dom(T ) is dense in H and T is therefore densely defined.

We shall show that the operator T f = f ′ for f ∈ dom(T ) is closed and unbounded on
H. Unboundedness follows by considering the sequence fn (x) = n1 sin(πnx) and calculating
the norms kfn k and kfn′ k.
Let us now suppose that fn ∈ dom(T
Rx ), fn → f in H and
R xT fn → g in H. We know
that (absolute continuity) fn (x) = 0 fn′ and we put h(x) = 0 g. Then
Z 1 Z 1 1/2
|fn (x) − h(x)| ≤ |fn′ − g| ≤ |fn′ 2
− g| →0
0 0

as n → ∞. Therefore, fn ⇉ h on [0, 1] and hence also fn → h in H, and finally f = h

a.e. We conclude that fn ⇉ f , f ∈ AC([0, 1]) and f (0) = f (1) = 0. We therefore have
f ∈ dom(T ) and T f = g, i.e. T is closed.
Next, we look for T ∗ . We shall show that

dom(T ∗ ) = {h ∈ AC([0, 1]) : h′ ∈ L2 ([0, 1])} and T ∗ h = −h′ .

Let us suppose first, that h ∈ AC([0, 1]) and h′ ∈ L2 ([0, 1]). Then
Z 1 Z 1 Z 1
′ 1 ′
hT f, hi = f h̄ = [f h̄]0 − f h̄ = − f h̄′ = hf, −h′ i,
0 0 0

hence f → hT f, hi is continuous and h ∈ dom(T ∗ ).

Let on the other hand g ∈ dom(T ∗ ) and h = T ∗ g. For f ∈ dom(T ), we get
Z 1 Z 1  Z x x=1 Z 1 Z x 
′ ∗ ′
f ḡ = hT f, gi = hf, T gi = hf, hi = f h̄ = f (x) h̄ − f (x) h̄ dx
0 0 0 x=0 0 0
Z 1 Z x 
′
=− f (x) h̄ dx.
0 0

We see, that  Z 
x
′
f (·), g(x) + h =0
0
Rx
for all f ∈ dom(T ). In other words, the function x → g(x) + R0 h is orthogonal to f ′ for all
x
f ∈ dom(T ). Therefore, this function is constant and (since 0 h is absolutely continuous)
g is also in AC([0, 1]) and g ′ = −h ∈ H. Furthermore, this shows that T ∗ g = h = −g ′ .

Example 13.1.8. If H = L2 ([0, 1]),

dom(T ) = {f ∈ AC([0, 1]) : f ′ ∈ L2 ([0, 1]), f (0) = f (1) = 0} and T f = if ′

we obtain in a similar way

dom(T ∗ ) = {h ∈ AC([0, 1]) : h′ ∈ L2 ([0, 1])} and T ∗ h = ih′ .

Therefore, T ⊂ T ∗ and T is symmetric but not self-adjoint.

Example 13.1.9. Finally, if H = L2 ([0, 1]),

dom(T ) = {f ∈ AC([0, 1]) : f ′ ∈ L2 ([0, 1]), f (0) = f (1)} and T f = if ′ ,

we get
dom(T ∗ ) = dom(T ) and T = T ∗ .

19
Functional Analysis II

If T is densely defined and symmetric, then T ⊂ T ∗ and dom(T ) ⊂ dom(T ∗ ) ⊂ H, and

T ∗∗ may be defined.

Theorem 13.1.10. Let T : dom(T ) → H be densely defined. Then

i) T ∗ is closed.

ii) If T ∗ is densely defined, then T ⊂ T ∗∗ .

iii) If T ∗ is densely defined, then T ∗∗ is the closure of T .

Proof. (i): We have to show that yn ∈ dom(T ∗ ), yn → y ∈ H and T ∗ yn → z implies

y ∈ dom(T ∗ ) and z = T ∗ y. We have

hT x, yi = lim hT x, yn i = lim hx, T ∗ yn i = hx, zi for every x ∈ dom(T ).

n→∞ n→∞

Hence x → hT x, yi is continuous, y ∈ dom(T ∗ ) and T ∗ y = z.

(ii): If x ∈ dom(T ) and y ∈ dom(T ∗ ), then hT x, yi = hx, T ∗ yi. Hence y → hx, T ∗ yi is
continuous and x ∈ dom(T ∗∗ ). Furthermore,

hx, T ∗ yi = hT ∗∗ x, yi for all y ∈ dom(T ∗ ).

We conclude that T x = T ∗∗ x for all x ∈ dom(T ), i.e. T ⊂ T ∗∗ .

(iii): We show that
G(T ) = G(T ∗∗ ).
The inclusion “⊂” follows from (i) and (ii). Let (v, u) ∈ G(T )⊥ . Then hx, ui + hT x, vi = 0
for all x ∈ dom(T ). Then v ∈ dom(T ∗ ) and T ∗ v = −u.
For (T ∗∗ z, z) ∈ G(T ∗∗ ), we have

h(T ∗∗ z, z), (v, u)i = hz, ui + hT ∗∗ z, vi = hz, ui + hz, T ∗ vi = hz, u + T ∗ vi = 0,

hence (v, u) ∈ G(T ∗∗ )⊥ and the second inclusion follows.

Corollary 13.1.11. Let T : dom(T ) → H be densely defined.

i) T is symmetric if, and only if, T ⊂ T ∗ . Then we have T ⊂ T ∗∗ ⊂ T ∗ = T ∗∗∗ and

also T ∗∗ is symmetric.

ii) T is closed and symmetric if, and only if, T = T ∗∗ ⊂ T ∗ .

iii) T is self-adjoint if, and only if, T = T ∗ = T ∗∗ .

Example 13.1.12. We shall present an example of an operator with a non-densely defined

adjoint. Let H = L2 (−1, 1) and let

dom(T ) = {f ∈ C ∞ (−1, 1) ∩ L2 (−1, 1) : |f (j) (0)| ≤ Cf 2−j j! for all j ≥ 0},

∞
X f (j)(0) j
(T f )(x) = x , −1 < x < 1.
j!
j=0

Roughly speaking, T sends f to its Taylor series, and the domain contains only functions
for which this series converges uniformly and absolutely. We shall show that T is a densely
defined operator with dom(T ∗ ) = {0}.
We proceed in the following way

20
 13 Spectral theory for unbounded operators

i) If T is an arbitrary linear operator and ker(T ) is dense, then dom(T ∗ ) = ker(T ∗ ) :

Let y ∈ dom(T ∗ ). Then x → hT x, yi is a continuous mapping on x ∈ dom(T ). As
ker(T ) ⊂ dom(T ) ⊂ H is a dense subset of H, this mapping is continuous and equal
to 0 on a dense subset of H. Hence, T ∗ y = 0.

ii) If ran(T ) is also dense, then dom(T ∗ ) = {0}:

Let again y ∈ dom(T ∗ ). Then (by previous point) T ∗ y = 0 and the mapping
x → hT x, yi is equal to zero for all x ∈ dom(T ). This means, that hx̃, yi = 0 for x̃
from a dense subset of H, hence y = 0.

iii) Finally, ker(T ) and ran(T ) are dense for the operator above:
Obviously, ker T ⊃ C0∞ ((−1, 1) \ {0}), i.e. ker T contains all functions vanishing on
some neighborhood of zero and this set is dense in H. Furthermore, the range of
T includes all polynomials, as the constant Cf might be arbitrary large, hence also
ran(T ) is dense in H.

13.2 Spectral properties of unbounded operators

Definition 13.2.1. Let T : dom(T ) ⊂ H → H be densely defined.

i) The set

̺(T ) := {λ ∈ C : λI − T : dom(T ) → H has a bounded inverse in H}9

is called the resolvent set of T .

ii) The mapping R : ̺(T ) → L (H), Rλ = R(λ) := (λI − T )−1 is called the resolvent
mapping.

iii) σ(T ) := C \ ̺(T ) is called the spectrum of T .

Example 13.2.2. Let Ω ⊂ Rn be measurable and let f : Ω → R a measurable function.

We denote by

dom(Mf ) = {g ∈ L2 (Ω) : f g ∈ L2 (Ω)},

Mf (g)(t) = f (t)g(t), t∈Ω

the pointwise multiplication operator on the Hilbert space H = L2 (Ω).

Then it holds

i) Mf is densely defined:
Let An := {x ∈ Ω : |f (x)| ≤ n}. Then gχAn ∈ dom(Mf ) for all n ∈ N and all g ∈ H.
And gχAn → g in H finishes the argument.

ii) Mf is self-adjoint (and, therefore, Mf is also closed):

R
Let h ∈ dom(Mf ). Then g → hMf g, hi = Ω f (t)g(t)h(t)dt = hg, Mf hi is continuous,
h ∈ dom(Mf∗ ) and Mf∗ h = Mf h. If, on the other hand, h ∈ dom(Mf∗ ), then the
mapping above is continuous, and f h ∈ H follows.
9
This means, that there is an operator S ∈ L (H), such that (T − λI)S = I and S(T − λI) is identity
on dom(T ).

21
Functional Analysis II

iii) σp (Mf ) = {z ∈ R : f −1 (z) has positive measure} are the eigenvalues (the so-called
point spectrum) of Mf and every eigenvalue of Mf has infinite multiplicity:
If the set Az := f −1 (z) has positive measure, then {g ∈ H : supp g ⊂ Az } is
an infinite-dimensional subspace of eigenvectors. If, on the other hand, Mf g(t) =
f (t)g(t) = zg(t) for almost every t ∈ (−1, 1), we get f (t) = z almost everywhere on
the support of g and the claim follows.

iv) z ∈ C is in σ(Mf ) if, and only if, for every ε > 0, the set {x ∈ Ω : |f (x) − z| < ε} =
f −1 ({w ∈ C : |w − z| < ε}) has positive measure:
Let z ∈ C be such that An := f −1 ({w ∈ C : |w − z| < 1/n}) has positive measure
for every n ∈ N and consider gn 6= 0 with supp gn ⊂ An . Then gn ∈ dom(Mf ) and
Z 1 Z
2 2 kgn k2
k(Mf − zI)gn k = |(f (t) − z)gn (t)| dt = |(f (t) − z)gn (t)|2 dt ≤
−1 An n2

and Mf − zI cannot be boundedly invertible, i.e. z ∈ σ(Mf ).

Let on the other hand z ∈ C and ε > 0 be such that f −1 ({w ∈ C : |w − z| < ε}) has
measure zero. Then Mf − zI can be inverted by M 1 .
f (t)−z

If T is not closed, then (T − λI) is not closed for every λ ∈ C, hence G(T − λI) is
not closed. On the other hand, if (T − λI)−1 would be bounded (i.e. in L (H)), then
G((T − λI)−1 ) would be closed and, therefore, also G(T − λI) would be closed.
This can be summarized by saying that σ(T ) = C whenever T is not closed.

Theorem 13.2.3. Let T : dom(T ) → H be densely defined. Then

i) ̺(T ) is open.

ii) The resolvent mapping is analytic and the resolvent identity

Rλ − Rµ = (µ − λ)Rλ Rµ

holds.

iii) σ(T ) is closed.

Warning: σ(T ) can be unbounded, or empty.

Proof. (Same as for bounded operators!)

i) Let λ0 ∈ ̺(T ) and |λ − λ0 | < k(λ0 I − T )−1 k−1 . Then

λI − T = (λ0 I − T ) + (λ − λ0 )I = (λ0 I − T ) [I − (λ − λ0 )(λ0 I − T )−1 ] .

| P
{z }
∞
[... ]−1 = n=0 ((λ−λ0 )(λ0 I−T )
−1 )n

Hence, λI − T is invertible.

ii) Formal calculation gives immediately

(λI − T )(µI − T )[(λI − T )−1 − (µI − T )−1 ] = (µ − λ)I

and the similar identity for multiplication from the right side. So, only the (easy)
inspection of domains is missing.

iii) follows from (i).

22
 13 Spectral theory for unbounded operators

Theorem 13.2.4. Let T be a densely defined operator. Then ker(T ∗ ) = ran(T )⊥ , where

ker(T ∗ ) = {x ∈ dom(T ∗ ) : T ∗ x = 0},

ran(T ) = {T x : x ∈ dom(T )}.

Proof. y ∈ ker(T ∗ ) if, and only if, hT x, yi = 0 for all x ∈ dom(T ) and if, and only if
y ⊥ ran(T ).

Theorem 13.2.5. Let T be a densely defined symmetric operator, z ∈ C \ R. Then the

following statements are equivalent:
i) T is self-adjoint.

ii) T is closed and ker(T ∗ − zI) = ker(T ∗ − z̄I) = {0}.

iii) ran(T − zI) = ran(T − z̄I) = H.

Proof. We shall use the fact, that if T is densely defined and z ∈ C, then (T − zI)∗ =
T ∗ − z̄I, cf. Exercises.
(i) implies (ii): T = T ∗ and T ∗ is closed, therefore also T is closed. Let x ∈ ker(T ∗ −
zI) = ker(T − zI) (as T = T ∗ ). Then

zhx, xi = hzx, xi = hT x, xi = hx, T ∗ xi = hx, zxi = z̄hx, xi,

i.e. x = 0. The same holds for ker(T ∗ − z̄I).

(ii) implies (iii): By Theorem 13.2.4, we have ran(T − zI)⊥ = ker(T ∗ − z̄I) = {0}. So, it
is enough to show, that ran(T − zI) is closed.
Let yn ∈ ran(T − zI), i.e. yn = (T − zI)xn for some xn ∈ dom(T ) and we suppose that
yn → y. Let z = a + ib with b 6= 0 and let u ∈ dom(T ). Then we have

k(T − zI)uk2 = h(T − a − ib)u, (T − a − ib)ui = k(T − aI)uk2 + b2 kuk2 ≥ b2 kuk2 ,

as h−ibu, (T − aI)ui + h(T − aI)u, −ibui = 0 as u ∈ dom(T ) ⊂ dom(T ∗ ). Hence kuk ≤

1/|b| · k(T − zI)uk.
We apply this inequality for u = xm − xn and observe, that (xn )n∈N is a Cauchy-sequence
and, therefore, x = limn→∞ xn exists. Then T xn = yn + zxn converges also, to y + zx. As
T is closed, we get x ∈ dom(T ), T x = y + zx, or y = (T − zI)x ∈ ran(T − zI).
(iii) implies (i): Let x ∈ dom(T ∗ ). By assumption (iii), we can find y ∈ dom(T ), such
that (T − zI)y = (T ∗ − zI)x = (T ∗ − zI)y (as T ⊂ T ∗ ). Hence, x − y ∈ ker(T ∗ − zI),
which (see Theorem 13.2.4) is equal to ran(T − z̄I)⊥ = {0}. Therefore, x = y ∈ dom(T ),
dom(T ) ⊂ dom(T ∗ ) and T = T ∗ .

Corollary 13.2.6. Let T be self-adjoint. Then σ(T ) ⊂ R.

Proof. Let z ∈ C\R. By Theorem 13.2.5 we have ker(T −zI) = 0 and ran(T −zI) = H.

Example 13.2.7. We present an example, where the change of the domain of the operator
has a big impact on its spectrum. This demonstrates once more, how important is the
right choice of the domain.
We consider again the operator f → if ′ on L2 (0, 1), on the following domains:

dom(S) = {f ∈ AC([0, 1]) : f ′ ∈ L2 (0, 1)},

dom(T ) = {f ∈ dom(S) : f (0) = 0}.

23
Functional Analysis II

It is quite easy to prove (similar to previous arguments) that both S and T are closed.
We show that
σ(S) = C, σ(T ) = ∅.
The claim on σ(S) is very easy to confirm. Just notice that ez (x) = e−izx ∈ dom(S)
for all z ∈ C and (S − z)ez = 0.
To find σ(T ), fix z ∈ C and let
Z x
(Rz f )(x) = −ie−izx eizt f (t)dt.
0

This is well defined for all f ∈ L2 (0, 1) and (Rz f )(x) is absolutely continuous function of
x ∈ [0, 1]. Hence Rz f ∈ L2 (0, 1) as well. An easy calculation shows that

(Rz f )′ (x) = −iz(Rz f )(x) − if (x).

Hence (Rz f )′ ∈ L2 (0, 1) and (Rz f )(0) = 0, i.e. Rz f ∈ dom(T ).

Moreover (T − zI)Rz = I, so ran(T − zI) = H. Similar arguments (integration by
parts) show that Rz (T − zI)f = f for all f ∈ dom(T ). Together, z ∈ ̺(T ) and σ(T ) = ∅.

13.3 Spectral theorem for unbounded operators

The main aim of this section is to prove the following theorem.
Theorem 13.3.1. (Spectral theorem for unbounded operators) Let T : dom(T ) → H be
a self-adjoint operator. Then there exists a unique spectral measure E on Borel sets of
σ(T ), such that Z
T = λdEλ ,
σ(T )
i.e. Z
hT x, yi = λdhEλ x, yi for all x ∈ dom(T ) and all y ∈ H.
σ(T )

Remark 13.3.2. i) Formally, it looks very similar to the case of bounded operators.
R
ii) σ(T ) might be an unbounded subset of R, therefore σ(T ) λdEλ integrates an un-
bounded function over an unbounded subset of R.
iii) Let E be a spectral measure on R and f be an unbounded (but everywhere finite)
Borel-measurable function. We let
Z
Df := {x ∈ H : |f (t)|2 dhEt x, xi < ∞}. (13.1)
R

Recall, that A → hEA x, xi is a non-negative measure on R.

iv) Furthermore, one has to show that
Z
hTf x, yi = f (t)dhEt x, yi
R

makes sense and that dom(Tf ) = Df .

Lemma 13.3.3. Df is a dense subset of H. For all x ∈ Df and y ∈ H we have
Z Z 1/2
2
|f (t)|d|hEt x, yi| ≤ kyk · |f (t)| dhEt x, xi . (13.2)
R R

24
 13 Spectral theory for unbounded operators

Here, the integration with respect to d|hEt x, yi| means the integration with respect to
the variation of the measure µx,y : A → hEA x, yi, i.e. |µx,y |.

Proof. (Sketch)
First, we show that Df is a linear set. From

hEA (x + y), x + yi = kEA (x + y)k2 ≤ (kEA xk + kEA yk)2 ≤ 2(kEA xk2 + kEA yk2 )

we deduce that µx+y,x+y ≤ 2µx,x + 2µy,y , where again µx,x (A) = hEA x, xi. It follows, that
x, y ∈ Df implies x + y ∈ Df . Similar calculation shows also that cx ∈ Df for c ∈ C.
Second, we show that Df is dense. Let ωn = {t ∈ R : |f (t)| < n} ⊂ R and let
y ∈ ran(Eωn ). Then
Z Z
2
hEωnc y, yi = 0 and |f (t)| dhEt y, yi = |f (t)|2 dhEt y, yi ≤ n2 kyk2 < ∞.
R ωn

Therefore, y ∈ Df . Furthermore, Eωn x → x due to kx − Eωn xk2 = kEωnc xk2 = µx,x (ωnc ) →
0.
Finally, let us prove the inequality (13.2). First, we observe that (for x 6= 0)

|µx,y (A)| = |hEA x, yi| = |hEA x, EA yi| ≤ kEA xk · kEA yk.

As ( n )
X n
[
|µx,y |(A) = sup |µx,y (Ak )| : Ak = A, A1 , A2 , . . . , Ak disjoint ,
k=1 k=1
we get
( n n
)
X [
|µx,y |(A) ≤ sup kEAk xk · kEAk yk : Ak = A, A1 , A2 , . . . , Ak disjoint
k=1 k=1
 !1/2 !1/2 
 X
n n
X n
[ 
≤ sup kEAk xk2 · kEAk yk2 : Ak = A, A1 , A2 , . . . , Ak disjoint
 
k=1 k=1 k=1

≤ kEA xk · kEA yk.

n
X
For f a simple function f = αi χAi with Ai ’s disjoint, the inequality (13.2) now follows
i=1
by the Cauchy-Schwartz inequality:
Z n
X n
X
|f (t)|d|hEt x, yi| = |αi | · |µx,y |(Ai ) ≤ |αi | · kEAi xk · kEAi yk
R i=1 i=1
n
!1/2 n
!1/2
X X
≤ |αi |2 kEAi xk2 · kEAi yk2
i=1 i=1
Z 1/2
2
≤ kyk · |f (t)| dhEt x, xi .
R

For a general function f , we consider a sequence of simple functions |fn | ր |f | and apply
Levi’s Theorem (sometimes also called Lebesgue monotone convergence theorem).

Theorem 13.3.4. Let E be a spectral measure on R and let f be a Borel-measurable

function on R. Let Df be defined by (13.1).

25
Functional Analysis II

i) There exists a unique linear operator Tf on H with dom(Tf ) = Df and

Z
hTf x, yi = f (t)dhEt x, yi for all x ∈ Df and all y ∈ H.
R

ii) For all x ∈ Df Z

2
kTf xk = |f (t)|2 dhEt x, xi.
R

iii) Tf ◦ Tg ⊂ Tf g and dom(Tf ◦ Tg ) = Dg ∩ Df g .

iv) Tf∗ = Tf¯ and Tf is closed.
Proof. (Sketch) Uniqueness is clear, as Tf is determined by (i). To show the existence,
one considers the map Z
y→ f (t)dhEt x, yi, x ∈ Df .
R
This is a bounded (semi-)linear mapping on H and may be therefore represented by a
z ∈ H (which we call Tf x := z) such that
Z Z Z 1/2

hz, yi =
f (t)dhEt x, yi and kzk = y →
f (t)dhEt x, yi ≤ 2
|f (t)| dhEt x, xi .
R R H′ R

The linearity of Tf follows by linearity of x → dhEt x, yi.

One part (“≤”) of the identity for kTf xk2 follows by Lemma 13.3.3. The other inequality
follows for fn bounded and detailed study of the inequality in Lemma 13.3.3. The rest (as
well as (iii) and (iv)) follows by truncation and limits.

Theorem 13.3.5. Let f : R → C be measurable. Then dom(Tf ) = H if, and only if, f is
essentially bounded w.r.t. to E.
Proof. Let x ∈ H, then hER x, xi ≤ kxk2 . If f is bounded, we integrate in (13.1) a bounded
function over a finite measure space, hence x ∈ Df .
If, on the other hand, dom(Tf ) = Df = H, then Tf ∈ L (H) by Closed Graph Theorem.
Furthermore, if ωn := {t ∈ R : |f (t)| ≥ n}, then Theorem 13.3.4 (ii) implies that kTf xk ≥
nkxk for x ∈ ran(Eωn ). Therefore Eωn = 0 for n large.
Proof. (Sketch of the proof of Theorem 13.3.1).
Let R := (T − iI)−1 ∈ L (H). Then RR∗ = R∗ R, i.e. R is normal (and R∗ = (T + iI)−1 ∈
L (H)). According to the spectral theorem for bounded normal operators,
Z
R = (T − iI)−1 = zdFz ,
σ(R)

for some spectral measure F on C (R is normal, not self-adjoint).

Let ϕ(t) = 1/(t − i). We want to “change the variables” from z to t. We put E(M ) :=
F (ϕ(M )). The fact that ϕ(i) is not defined is not really disturbing, because we aim for
M ⊂ R anyway. R R 1
This is a new spectral measure and zdFz = t−i dEt . Then one has to show that
Z
T − iI = R−1 = (t − i)dEt ,
R
which implies T = tdEt . One has to show also that E is supported on σ(T ). Finally,
uniqueness follows from the fact, that it is also possible to recover F from E and F is
unique, therefore also E is unique.

26
14 Distributions and Fourier transform
14.1 The space S (Rn ) and the Fourier transform on S (Rn )
We shall use the usual notation from vector analysis, i.e. let n ∈ N, α ∈ Nn0 be a multiindex
and x = (x1 , . . . , xn ) ∈ Rn , then we write

xα := xα1 1 . . . xαnn ,
|α| := α1 + · · · + αn ,
∂ |α|
D α := ∂1α1 . . . ∂nαn = α1 ,
∂x1 . . . ∂xαnn
q
|x| := x21 + · · · + x2n ,
α! := α1 ! . . . αn !.

Definition 14.1.1. Let n ∈ N.

i) We put
S (Rn ) = {ϕ ∈ C ∞ (Rn ) : kϕk(k,l) < ∞ for all k, l ∈ N0 },
where X
kϕk(k,l) = sup (1 + |x|2 )k/2 |D α ϕ(x)|.
x∈Rn
|α|≤l

ii) A sequence (ϕj )j∈N ⊂ S (Rn ) is said to converge in S (Rn ) to ϕ ∈ S (Rn ) if, and
only if,
lim kϕj − ϕk(k,l) = 0 for all k, l ∈ N0 .
j→∞

S
This will be written as ϕj → ϕ.
Remark 14.1.2. The space S (Rn ) is a vector space (in the algebraic sense), but it is
not a Banach space (or normed space). It can be shown, that it is actually impossible to
introduce a norm S (Rn ), such that the convergence in this norm would be equivalent to
convergence, which we have just described. Nevertheless, it is possible to equip S (Rn )
with a topology, such that the convergence in this topology and S -convergence are equiv-
alent. Together with this topology, S (Rn ) becomes a topological vector space. We refer to
the Book of Rudin [4], which contains a lot of details on such spaces. On the other hand,
the space is metrizable, cf. Exercises.
S
Theorem 14.1.3. i) Let (ϕk )k∈N ⊂ S (Rn ) with ϕk → ϕ ∈ S (Rn ). Then ϕk → ϕ
also in Lp (Rn ) for every 0 < p ≤ ∞.
S S
ii) Let (ϕk )k∈N ⊂ S (Rn ) with ϕk → ϕ ∈ S (Rn ). Then D α ϕk → D α ϕ for every
α ∈ Nn0 .
S S
iii) Let (ϕk )k∈N ⊂ S (Rn ) with ϕk → ϕ ∈ S (Rn ). Then xα ϕk → xα ϕ for every α ∈ Nn0 .
S
iv) Let ϕ ∈ S (Rn ) and let τh ϕ(x) = ϕ(x − h). Then τh ϕ → ϕ for h → 0.
R
v) Let ϕ, ψ ∈ S (Rn ). Then ϕψ and (ϕ ∗ ψ)(x) = Rn ϕ(x − y)ψ(y)dy belong both to
S (Rn ) and it holds

D α (ϕ ∗ ψ) = (D α ϕ) ∗ ψ = ϕ ∗ (D α ψ).

27
Functional Analysis II

Proof. We have
|ϕj (x) − ϕ(x)| ≤ (1 + |x|2 )−k/2 kϕj − ϕk(k,0) .
Choosing k large enough, integrating over x ∈ Rn and using kϕj − ϕk(k,0) → 0 gives (i).
Similarly, (ii) and (iii) follows by using only the definition of S (Rn ).
To prove (iv), we have to estimate kϕ(x + h) − ϕ(x)k(k,l) . We estimate using Taylor
theorem
|D α ϕ(x + h) − D α ϕ(x)| ≤ cα |h| · sup sup |D α+β ϕ(y)|
|β|=1 y∈[x,x+h]

and obtain
kτh ϕ − ϕk(k,l) ≤ ck,l |h| · kϕk(k,l+1) , |h| ≤ 1.
Finally, we give the proof of (v). Let ϕ, ψ ∈ S (Rn ). Then (ϕ ∗ ψ) is well defined, i.e.
the integral converges for every x ∈ Rn . The identity for derivatives of (ϕ ∗ ψ) follows by
Lebesgue’s dominated convergence theorem. Indeed, let ej = (0, . . . , 0, 1, 0, . . . , 0) with 1
in the jth entry and zeros everywhere else.
(ϕ ∗ ψ)(x + hej ) − (ϕ ∗ ψ)(x)
D ej (ϕ ∗ ψ)(x) = lim
h→0
Z h
1
= lim [ϕ(x + hej − y) − ϕ(x − y)]ψ(y)dy
h→0 h Rn
Z
ϕ(z + hej ) − ϕ(z)
= lim · ψ(x − z)dz
h→0 Rn h
Then
ϕ(z + hej ) − ϕ(z)
− D ej ϕ(z) → 0
h
for every z ∈ Rn as h → 0 and since this expression is uniformly bounded by a constant
depending only on ϕ, its integral with respect to the measure ψ(x − z)dz converges to 0
as h → 0. The general case follows by repeating this argument by induction.
Next we show, that ϕ ∗ ψ belongs also to S (Rn ). We have (for N > n arbitrary)
Z
|(ϕ ∗ ψ)(x)| ≤ CN (1 + |x − y|2 )−N/2 (1 + |y|2 )−N/2 dy, (14.1)
Rn

where CN depends only on ϕ, ψ and N > n. The part of the integral in (14.1) over the set
{y ∈ Rn : |x − y| > |x|/2} is bounded by
Z
(1 + (|x|/2)2 )−N/2 (1 + |y|2 )−N/2 dy ≤ BN (1 + |x|2 )−N/2 ,
y:|x−y|>|x|/2

where BN depends only on N and the dimension. On the other hand, if |x|/2 > |y − x|,
then |y| > |x|/2 and the part of the integral in (14.1) over the set {z ∈ Rn : |y −x| < |x|/2}
may be estimated from above by
Z
(1 + |x − y|2 )−N/2 (1 + (|x|/2)2 )−N/2 dy ≤ BN (1 + |x|2 )−N/2 .
y:|x−y|<|x|/2

Therefore, (ϕ ∗ ψ)(x) decays at infinity as (1 + |x|2 )−N/2 - and this holds for every N > n.
Using D α (ϕ ∗ ψ) = (D α ϕ ∗ ψ) and replacing ϕ by D α ϕ, we obtain by the same argument,
that also D α (ϕ∗ψ) decays faster then the reciprocal of any polynomial, and ϕ∗ψ ∈ S (Rn )
follows.
Finally, that ϕψ belongs to S (Rn ) follows directly by the Leibniz rule.

28
 14 Distributions and Fourier transform

Definition 14.1.4. Let ϕ ∈ S (Rn ). Then

Z
1
(Fϕ)(ξ) = ϕ(x)e−ihx,ξi dx, ξ ∈ Rn
(2π)n/2 Rn

is called the Fourier transform of ϕ. Furthermore,

Z
−1 1
(F ϕ)(ξ) = ϕ(x)eihx,ξi dx, ξ ∈ Rn
(2π)n/2 Rn

is called the inverse Fourier transform of ϕ.

We will prove later on, that the terminology used is correct, especially, that F −1 really
is an inverse of F.
Furthermore, let us mention that many authors use another normalization of Fourier
transform, i.e. Z
f (x)e−2πihx,ξi dx.
Rn

Although this difference is harmless in general, it makes comparison between different

books and textbooks rather inconvenient, and a lot of care is necessary.

Theorem 14.1.5. i) Let ϕ ∈ S (Rn ). Then Fϕ ∈ S (Rn ) and F −1 ϕ ∈ S (Rn ).

ii) Let ϕ ∈ S (Rn ). Then

D α (Fϕ)(ξ) = (−i)|α| F(xα ϕ)(ξ), α ∈ Nn0 , ξ ∈ Rn , (14.2)

ξ α (Fϕ)(ξ) = (−i)|α| F(D α ϕ)(ξ), α ∈ Nn0 , ξ ∈ Rn . (14.3)

S
iii) Let (ϕj )∞ n
j=1 ⊂ S (R ) with ϕj → ϕ. Then

S S
Fϕj → Fϕ and F −1 ϕj → F −1 ϕ.

Proof. If ϕ ∈ S (Rn ) and α ∈ Nn0 , then one gets from Theorem 14.1.3 that xα ϕ ∈ S (Rn )
and D α ϕ ∈ S (Rn ). Furthermore, we get by Lebesgue’s dominated convergence theorem
Z
∂ 1
(Fϕ)(ξ) = (−i)xl e−ihx,ξi ϕ(x)dx = (−i)F(xl ϕ(x))(ξ).
∂ξl (2π)n/2 Rn

By iteration, we get (14.2). To prove (14.3), we use partial integration in xl -direction with
intervals tending to R to obtain
Z Z
i ∂ −ihx,ξi −i ∂ϕ
ξl (Fϕ)(ξ) = n/2
(e )ϕ(x)dx = n/2
(x)e−ihx,ξi dx
(2π) Rn ∂xl (2π) Rn ∂xl
 
∂ϕ
= (−i)F (ξ).
∂xl

These formulas can be easily iterated, cf.

∂2 ∂
(Fϕ)(ξ) = (−i) (F(xl ϕ(x))(ξ)) = (−i)2 F(xm xl ϕ(x))(ξ).
∂ξm ∂ξl ∂ξm

29
Functional Analysis II

Finally, by (14.2) and (14.3), we obtain10

kFϕk(k,l) ≤ c sup max max |ξ β | · |D α (Fϕ)(ξ)|
ξ∈Rn |β|≤k |α|≤l

≤ c max max kF(D β (xα ϕ(x)))k∞

|β|≤k |α|≤l

≤ c max max kD β (xα ϕ(x))k1

|β|≤k |α|≤l

≤ c max max kxα (D β ϕ)(x))k1

|β|≤k |α|≤l

≤ c kϕk(l+n+1,k) , for all ϕ ∈ S (Rn ).

Theorem 14.1.6. i) Let ε > 0 and let ϕ ∈ S (Rn ). Then

F(ϕ(ε·))(ξ) = ε−n F(ϕ)(ξ/ε), ξ ∈ Rn .

ii) Let h ∈ Rn and ϕ ∈ S (Rn ). Then we denote by (τh ϕ)(x) = ϕ(x − h) the translation
operator. And the formula F(τh ϕ) = e−ihh,ξi Fϕ holds for all ϕ ∈ S (Rn ).
iii) Let h ∈ Rn and ϕ ∈ S (Rn ). Then we denote by (Mh ϕ)(x) = eihh,xi ϕ(x) the modu-
lation of ϕ. And the formula F(Mh ϕ) = τh (Fϕ) holds for every ϕ ∈ S (Rn ).
iv) Let ϕ, ψ ∈ S (Rn ). Then the convolution of ϕ and ψ satisfies the formula
F(ϕ ∗ ψ)(ξ) = (2π)n/2 (Fϕ)(ξ) · (Fψ)(ξ), ξ ∈ Rn .

v) If ϕ ∈ S (Rn ) and ψ ∈ S (Rm ), then (ϕ ⊗ ψ)(x, y) = ϕ(x)ψ(y) ∈ S (Rn+m ) is the

tensor product of ϕ and ψ. And the formula F(ϕ⊗ ψ)(ξ, η) = (Fϕ)(ξ)(Fψ)(η) holds
for every ξ ∈ Rn and every η ∈ Rm .
vi) Let ϕ, ψ ∈ S (Rn ). Then F(ϕψ) = (2π)−n/2 F(ϕ) ∗ F(ψ).
vii) It holds
2 /2 2 /2
F(e−|x| )(ξ) = e−|ξ| , ξ ∈ Rn .
Proof. (i) - (v) are easy consequences of the Definitions. Due to the product structure, it
is enough to show (vii) in n = 1.
Let Z ∞
1 2
h(s) = 1/2
e−t /2 e−its dt, s ∈ R.
(2π) −∞
Then
Z Z ∞
2 1 −(t2 +u2 )/2 1 2
h(0) = e d(t, u) = (2πr) e−r /2 dr = 1,
2π R2 2π 0
i.e. h(0) = 1, is the famous Gaussian integral evaluated using polar coordinates in R2 . On
the other hand, by integration by parts we get
Z ∞ Z ∞
1 2 i d  −t2 /2  −its
h′ (s) = √ e−t /2 e−its (−it)dt = √ e e dt
2π −∞ 2π −∞ dt
Z ∞
−s 2
=√ e−t /2 e−its dt = −sh(s).
2π −∞
10
Let us note, that the meaning of c may change from one line to the other, and that it may depend on
k and l, but not on ϕ ∈ S (Rn ).

30
 14 Distributions and Fourier transform

2 2
Solving this equation, we observe that h(s) = h(0)e−s /2 = e−s /2 , which finishes the
proof.
We leave the proof of (vi) open in the moment and come back to that once we establish
the properties of the inverse Fourier transform.

Obviously, similar statements can be proven for F −1 in the same way. So far we know
that FS (Rn ) ⊂ S (Rn ) and F −1 S (Rn ) ⊂ S (Rn ).

Theorem 14.1.7. Let ϕ ∈ S (Rn ). Then

ϕ = F −1 (Fϕ) = F(F −1 ϕ).

Furthermore, F and F −1 both map S (Rn ) one-to-one onto itself.

Proof. When trying to evaluate F −1 (Fϕ) using Fubini’s theorem, one runs into non-
convergent integrals. Therefore, we use the following trick. We evaluate
Z Z
1 2 |ξ|2 /2 1
(Fϕ)(ξ)eihx,ξi e−ε dξ → (Fϕ)(ξ)eihx,ξi dξ = F −1 (Fϕ)(x)
(2π)n/2 Rn (2π)n/2 Rn

as ε → 0+ .
2 2
Let ψ(x) = e−ε |x| /2 . Then

 2
 |y|2
Fψ(y) = ε−n F e−|x| /2 (y/ε) = ε−n e− 2ε2 .

Furthermore, we get by change of variables (even for arbitrary ϕ, ψ ∈ S (Rn ))

Z Z
1
(Fϕ)(ξ)eihx,ξi ψ(ξ)dξ = ϕ(y)e−ihy−x,ξi ψ(ξ)d(ξ, y) (14.4)
Rn (2π)n/2 Rn ×Rn
Z Z
= ϕ(y)Fψ(y − x)dy = ϕ(x + y)Fψ(y)dy.
Rn Rn

2 |x|2 /2
This, applied to ψ(x) = e−ε then gives
Z Z
1 ihx,ξi −ε2 |ξ|2 /2 ε−n |y|2
− 2
(Fϕ)(ξ)e e dξ = ϕ(x + y)e 2ε dy
(2π)n/2 Rn (2π)n/2 Rn
Z
1 2
= n/2
ϕ(x + εz)e−|z| /2 dz.
(2π) Rn

By Lebesgue dominated convergence theorem, this tends to ϕ(x) as ε → 0+ .

Finally, we come to the proof of (vi) in Theorem 14.1.6. Let us observe that in (14.4),
we have shown actually that

(2π)n/2 F −1 (Fϕ · ψ) = ϕ ∗ F −1 ψ.

We plug in ψ = Fg and ϕ = f and the statement follows.

31
Functional Analysis II

14.2 The space S ′ (Rn ) and the Fourier transform on S ′ (Rn )

Definition 14.2.1. The space S ′ (Rn ) is the set of all continuous complex linear func-
tionals on S (Rn ). That means, T : S (Rn ) → C belongs to S ′ (Rn ) if, and only if,
i) T (λϕ + µψ) = λT (ϕ) + µT (ψ) for all λ, µ ∈ C and ϕ, ψ ∈ S (Rn );
S
ii) T ϕj → T ϕ whenever ϕj → ϕ.
The elements of the space S ′ (Rn ) are called tempered distributions.
Remark 14.2.2. S ′ (Rn ) is again a linear space, as we put (λT +µS)(ϕ) = λT (ϕ)+µS(ϕ)
for all T, S ∈ S ′ (Rn ) and all λ, µ ∈ C. Although one could again equip S ′ (Rn ) with a
topology and turn it into a topological vector space, for our purposes it is enough to equip
it with weak convergence. We say, that
S′
Tj ⇀ T ⇔ Tj (ϕ) → T (ϕ) for all ϕ ∈ S (Rn ).
Theorem 14.2.3. Let T be a linear functional on S (Rn ). Then T ∈ S ′ (Rn ) if, and only
if
|T (ϕ)| ≤ ckϕk(k,l) (14.5)
for some c > 0, k, l ∈ N0 and all ϕ ∈ S (Rn ).
Proof. If T satisfies (14.5), then T ∈ S ′ (Rn ). To prove the converse, we proceed by
contradiction, i.e. for all c > 0, k, l ∈ N0 , there is ϕc,k,l ∈ S (Rn ) with 1 = |T (ϕc,k,l )| >
S
ckϕc,k,l k(k,l) . We consider the sequence ϕk = ϕk,k,k . We obtain, that ϕk → 0 and (by
T ∈ S ′ (Rn )) also T (ϕk ) → 0, which is a contradiction.

Let f be an integrable function, i.e. f ∈ L1 (Rn ). Then the mapping

Z
Tf : ϕ → Tf (ϕ) = f (x)ϕ(x)dx, ϕ ∈ S (Rn )
Rn

is well-defined, linear and due to

Z 
 
 f (x)ϕ(x)dx ≤ kf k1 · kϕk∞ = kf k1 · kϕk(0,0)

Rn

is also continuous, hence T ∈ S ′ (Rn ). With a slight modification of the arguments, the
same holds for also for f ∈ Lp (Rn ) for all 1 ≤ p ≤ ∞. However, it does not hold for p < 1,
and not for Lloc n
1 (R ).
Every distribution, which is equal to Tf for some f ∈ L1 (Rn ), is called regular dis-
tribution. Every distribution, which is not regular is called singular distribution. In this
sense, we may identify L1 (Rn ) with the set of regular distributions, which is a subset of
S ′ (Rn ). The following lemma shows, that this identification is also one-to-one.
Lemma 14.2.4. Let f, g ∈ L1 (Rn ) and let Tf (ϕ) = Tg (ϕ) for all ϕ ∈ S (Rn ). Then f = g
a.e.
Proof. There is a lot of different proofs in the literature. We shall present one which uses
the technique of mollification, which we shall encounter also later on. Obviously, it is
enough to show that if Th (ϕ) = 0 for some h ∈ L1 (Rn ) and all ϕ ∈ S (Rn ), then h = 0
a.e.
Let ω ∈ S (Rn ) Rbe a smooth non-negative symmetric function with support in B :=
{y −n
R : |y| ≤ 1} and Rn ω(x)dx = 1. We define ωε (x) = ε ω(x/ε). Observe that also
Rn ωε (x)dx = 1. We observe the following facts

32
 14 Distributions and Fourier transform

i) h ∗ ωε is a continuous function on the whole Rn for each ε > 0.

ii) Using Fubini’s theorem (and the symmetry of ω) we get quickly (as ωε ∗ ϕ ∈ S (Rn ))
Z Z
0= h(x)(ωε ∗ ϕ)(x)dx = ϕ(y)(h ∗ ωε )(y)dy.
Rn Rn

iii) As ωε ∗ h is a continuous function, it must be positive on some neighborhood of

any point in which it is positive. Considering ϕ ∈ S (Rn ) with support in this
neighborhood and the equation above, we obtain (ωε ∗ h)(x) = 0 for all x ∈ Rn .

iv) Finally, we use that, for every t > 0, h ∈ L1 (Rn ) may be written as h = h1 + h2 ,
where h1 is continuous with compact support and kh2 k1 ≤ t.

v) The proof is finished by

khk1 = kh − ωε ∗ hk1 = kh1 + h2 − ωε ∗ h1 − ωε ∗ h2 k1

≤ kh1 − ωε ∗ h1 k1 + kh2 k1 + kωε ∗ h2 k1 ,
| {z } | {z }
(∗) ≤2t

where
Z
(∗) = |h1 (x) − ωε ∗ h1 (x)|dx
ZR  Z
n

 
=  [h1 (x) − h1 (x − y)]ωε (y)dy  dx

ZR Rn
n
Z
≤ ωε (y) |h1 (x) − h1 (x − y)|dxdy
Rn Rn
≤ sup kh1 (·) − h1 (· − y)k1 .
y:|y|≤ε

Due to the bounded support of h1 and its uniform continuity, the last expression
goes to zero as ε → 0. Hence, choosing ε, t > 0 small enough, khk1 is arbitrary small
and, therefore, equal to zero, i.e. h = 0 a.e.

Any finite Borel measure µ on Rn is a tempered distribution via

Z
µ(ϕ) = ϕ(x)dµ(x).
Rn

We show how one can extend the convolution also to measures. Let first ϕ, ψ ∈ S (Rn )
and f ∈ S (Rn ) as well. We denote by µ the measure on Rn , which has density ϕ with
respect to the Lebesgue measure. Furthermore, λ corresponds to ψ in the same way. Then
Z Z Z
(ϕ ∗ ψ)(f ) = (ϕ ∗ ψ)(x)f (x)dx = f (x) ϕ(x − y)ψ(y)dydx
ZR n R n
ZR n

= f (x)ϕ(x − y)ψ(y)dydx = f (z + y)ϕ(z)ψ(y)dzdy

Rn ×Rn Rn ×Rn
Z
= f (z + y)dµ(z)dλ(y).
Rn ×Rn

33
Functional Analysis II

We observe that the last expression makes sense for all Borel measures µ and λ and
f ∈ C0 (Rn ), i.e. continuous functions which tend to zero at infinity.
Hence, for two Borel measures µ and λ on Rn , we define µ ∗ λ to be the unique measure
on Rn , such that (using the Riesz representation theorem)
Z Z
f d(µ ∗ λ) = f (x + y)dµ(x)dλ(y) (14.6)
Rn Rn ×Rn

for all f ∈ C0 (Rn ). Using standard arguments, this holds also for all bounded Borel-
measurable functions, i.e. fξ (x) = eihx,ξi . Therefore, we obtain immediately the formula

F(µ ∗ λ)(ξ) = (2π)n/2 (Fµ)(ξ)(Fλ)(ξ), ξ ∈ Rn .

Furthermore, kµ ∗ λk ≤ kµk · kλk, where the norms denote total variation of measures.
Finally, if µ is absolutely continuous with respect to the Lebesgue measure, then the same
is true for µ ∗ λ. To see this, consider f = χE , where the Lebesgue measure of E is zero.
Then Z
0= f (x + y)dµ(x)
Rn
for all y ∈Rn , and (µ ∗ λ)(E) = 0 follows.
Finally, (14.6) implies that

(µ ∗ λ)(E) = (µ ⊗ λ)({(x, y) ∈ Rn × Rn : x + y ∈ E})

for every Borel-measurable set E ⊂ Rn .

Definition 14.2.5. Let T ∈ S ′ (Rn ), ϕ, ψ ∈ S (Rn ) and α ∈ Nn0 . Then we put

(D α T )(ϕ) = (−1)|α| T (D α ϕ),

(FT )(ψ) = T (Fψ), (F −1 T )(ψ) = T (F −1 ψ)

and
(ϕT )(ψ) = T (ϕψ). (14.7)
Actually, we use (14.7) to define ϕT for every (infinitely-differentiable) function ϕ, such
that ϕψ ∈ S (Rn ) for every ψ ∈ S (Rn ). This applies for example to the polynomials on
Rn or to (linear combinations of) functions of the type eihθ,xi .

Using Theorem 14.2.3, one shows that D α T, (ϕT ), FT, F −1 T ∈ S ′ (Rn ).

Remark 14.2.6. Let us sketch the proof (given by Schwartz) that it is impossible to
introduce a continuous multiplication of distributions.
R Let ψ ∈ S (Rn ) be compactly
supported non-negative function with ψ(0) = 1, ψ = 1. Then ψk (x) := kn ψ(kx) converge
to δ in S ′ (Rn ) as k → ∞. By the definition of multiplication of distribution δ and test
function ψk , we obtain

(δ · ψk )(ϕ) = δ(ψk ϕ) = ψk (0)ϕ(0) = kn ϕ(0) → ∞

for all ϕ ∈ S (R) with ϕ(0) > 0. On the other hand, assuming that it is possible to define
continuous multiplication in S ′ (Rn ) implies that (δψk )(ϕ) → (δ · δ)(ϕ) = δ2 (ϕ), i.e. a
contradiction.

Theorem 14.2.7. Let T ∈ S ′ (Rn ).

34
 14 Distributions and Fourier transform

i) Then F(F −1 T ) = F −1 (FT ) = T . Furthermore, both F and F −1 map S ′ (Rn )

one-to-one continuously onto itself.

ii) Let α ∈ Nn0 . Then xα T ∈ S ′ (Rn ) and D α T ∈ S ′ (Rn ). Furthermore,

F(D α T ) = i|α| xα (FT ) and F(xα T ) = i|α| D α (FT ).

iii) Let ε > 0 and let T ∈ S ′ (Rn ). Then

T (ε·)(ϕ) := T (ε−n ϕ(·/ε)), ϕ ∈ S (Rn )

is the dilation of T and

F(T (ε·)) = ε−n F(T )(·/ε).

iv) Let h ∈ Rn and T ∈ S ′ (Rn ). Then we denote by (τh T )(ϕ) = T (ϕ(· + h)), ϕ ∈
S (Rn ), the translation operator. The formula F(τh T ) = e−ihh,ξi FT holds for all
T ∈ S ′ (Rn ).

v) Let h ∈ Rn and T ∈ S ′ (Rn ). Then we denote by (Mh T )(ϕ) = T (eihh,xi ϕ), ϕ ∈

S (Rn ), the modulation of T. And the formula F(Mh T ) = τh (FT ) holds for every
T ∈ S ′ (Rn ).

Proof. We obtain

F(F −1 T )(ϕ) = (F −1 T )(Fϕ) = T (F −1 (Fϕ)) = T (ϕ)

for every ϕ ∈ S (Rn ), i.e. F(F −1 T ) = T. The other identities follow in the same manner.

14.3 Fourier transform on L1 (Rn ) and L2 (Rn )

There is another natural definition of Fourier transform, which turns out to be essentially
equivalent to definition given above.

Definition 14.3.1. Let f ∈ L1 (Rn ). Then we define

Z
1
Ff (ξ) = f (x)e−ihx,ξi dx, ξ ∈ Rn .
(2π)n/2 Rn

This integral converges absolutely for every ξ ∈ Rn .

Theorem 14.3.2. Let f ∈ L1 (Rn ) and ξ ∈ Rn .

i) If g(x) = f (x)eihα,xi , then Fg(ξ) = Ff (ξ − α).

ii) If g(x) = f (x − α), then Fg(ξ) = Ff (ξ)e−ihα,ξi .

iii) If g ∈ L1 (Rn ) and h = f ∗ g, then Fh(ξ) = (2π)n/2 Ff (ξ) · Fg(ξ).

iv) If g(x) = f (−x), then Fg(ξ) = Ff (ξ).

v) If g(x) = f (x/λ), then Fg(ξ) = λn Ff (λξ).

vi) If g(x) = −ixj f (x), and g ∈ L1 (Rn ), then Ff is differentiable at ξ and ∂j (Ff )(ξ) =
Fg(ξ).

35
Functional Analysis II

Proof. The proof is done by direct substitutions to the Definition, the last statement
follows by Lebesgue dominated convergence theorem.

Let us mention, that the properties of F and F are very similar. Nevertheless, the
identities above are understood in pointwise sense, the properties of F were proven in the
distributional sense.
We show that F = F on L1 (Rn ) ֒→ S ′ (Rn ) (and we shall after that use only the letter
F for any of the Fourier transforms).
Theorem 14.3.3. Let f ∈ L1 (Rn ). Then Ff is a regular distribution and Ff = Ff in
the distributional sense.
Proof. If f ∈ L1 (Rn ), then Ff is obviously bounded, cf.
Z 
 
|Ff (ξ)| =
1  f (x)e−ihx,ξi
dx  ≤ kf k1 , ξ ∈ Rn .
(2π) n/2   (2π)n/2
Rn

Therefore, TFf may be interpreted as an element of S ′ (Rn ).

On the other hand, let f ∈ L1 (Rn ) and ϕ ∈ S (Rn ). Then
Z
(Ff )(ϕ) = f (Fϕ) = f (x)(Fϕ)(x)dx
Z R
n
Z
1
= f (x) ϕ(y)e−ihx,yi dydx
(2π)n/2 Rn Rn
Z Z
1
= ϕ(y) n/2
f (x)e−ihx,yi dxdy
(2π)
ZR R
n n

= ϕ(y)Ff (y)dy = TFf (ϕ),

Rn

i.e. (Ff )(ϕ) = (Ff )(ϕ) for all ϕ ∈ S (Rn ) are equal in the distributional sense.

Theorem 14.3.4. Let f ∈ L1 (Rn ). Then Ff ∈ C0 (Rn ), i.e. Ff ∈ C(Rn ) and lim Ff (x) =
|x|→∞
0.
Proof. The boundedness of Ff was already discussed above. The continuity follows from
Lebesgue convergence theorem (with |f (x)| as an integrable majorant):
Z Z
1 −ihξ+h,xi 1
f (x)e dx → f (x)e−ihξ,xi dx as h → 0.
(2π)n/2 Rn (2π)n/2 Rn
Finally, if n = 1 and f = χ[a,b] with −∞ < a < b < ∞, we get
Z b
1 −iξt 1 e−iξa − e−iξb
Fχ[a,b] (ξ) = e dt = →0
(2π)1/2 a (2π)1/2 iξ
Q
if ξ → ±∞. In the same way, if n > 1 and g(x) = nj=1 χ[aj ,bj ] (xj ) on Rn , then
Yn
1 e−iξj aj − e−iξj bj
Fg(ξ) = →0
(2π)n/2 j=1 iξj

if |ξ| → ∞. The same conclusion therefore holds also for finite sums of step functions of
intervals. Finally, we use that every function f ∈ L1 (Rn ) might be approximated by such
a finite sum function h to arbitrary precision in the L1 (Rn )-norm. We obtain
kf − hk1
|Ff (ξ)| ≤ |F(f − h)(ξ)| + |Fh(ξ)| ≤ + |Fh(ξ)|.
(2π)n/2

36
 14 Distributions and Fourier transform

As the last summand goes to zero and the first can be made arbitrary small, we obtain
the conclusion.

The non-distributional approach to Fourier transform now proceeds as follows

i) The inversion theorem is proven for f ∈ L1 (Rn ) and Ff ∈ L1 (Rn ), and holds
pointwise a.e.

ii) Ff is now defined on L1 (Rn ).

iii) One shows that for f ∈ L1 (Rn ) ∩ L2 (Rn ), the following Plancherel identity holds:
kFf k2 = kf k2 .

iv) As L1 (Rn ) ∩ L2 (Rn ) is dense in L2 (Rn ), there is a unique extension of F from

L1 (Rn ) ∩ L2 (Rn ) to L2 (Rn ), which will be denoted by F again.

v) It follows that F is a Hilbert space isomorphism of L2 (Rn ) onto L2 (Rn ).

Theorem 14.3.5. (Fourier inversion theorem on L1 (Rn ))

Let f ∈ L1 (Rn ) and Ff ∈ L1 (Rn ). Then
Z
1
f (x) = (Ff )(ξ)eihξ,xi dξ
(2π)n/2 Rn

for almost every x ∈ Rn .

Proof. We know from Lemma 14.2.4 that if Th (ϕ) = 0 for h ∈ L1 (Rn ) and all ϕ ∈ S (Rn ),
then h = 0 a.e. This can be easily generalised. Let u ∈ L1 (R n n
−|x| 2 /2 n
R ) and v ∈ L∞ (R ).
Furthermore, let ψ(x) =Re . Then (u + v)ψ ∈ L1 (R ) and Rn (u + v)ϕ = 0 for all
ϕ ∈ S (R ) implies also Rn [(u + v)ψ]ϕ = 0 for all ϕ ∈ S (Rn ). Hence (u + v)ψ = 0 a.e.
n

and u + v = 0 a.e. follows.

Using Fubini’s theorem we get immediately
Z Z Z
1 −ihx,ξi
f (x)(Fϕ)(x)dx = f (x)ϕ(ξ)e d(x, ξ) = (Ff )(ξ)ϕ(ξ)dξ
Rn (2π)n/2 Rn ×Rn Rn

for f ∈ L1 (Rn ) and ϕ ∈ S (Rn ) (actually, even ϕ ∈ L1 (Rn ) can be allowed). The same
formula holds also for F −1 . Finally, we obtain
Z Z Z Z
−1 −1 −1
[F (Ff )]ϕ = (Ff )(F ϕ) = f (F (Fϕ)) = f ϕ.

As F −1 (Ff ) ∈ L∞ (Rn ) and f ∈ L1 (Rn ), we obtain F −1 (Ff ) = f a.e.

As the zero function is in L1 (Rn ), we obtain the following Corollary.

Corollary 14.3.6. Let f ∈ L1 (Rn ) and let Ff = 0 a.e. Then f = 0 a.e.

Theorem 14.3.7. Let f ∈ L1 (Rn ) ∩ L2 (Rn ), then

kf k2 = kFf k2 .

Proof. For f ∈ L1 (Rn ) ∩ L2 (Rn ), let h = f ∗ f¯˜. Then we have the following

i) h ∈ L1 (Rn );

37
Functional Analysis II

ii) h is continuous at 0; indeed

Z Z
h(x) − h(0) = f (y)f˜(x − y)dy − f (y)f˜(−y)dy
ZR Rn
n

= f (y)[f (y − x) − f (y)]dy.
Rn

Writing f = f1 + f2 , where f1 is continuous with compact support and kf2 k2 ≤ t,

we get
Z Z
|h(x) − h(0)| ≤ |f (y)| · |f1 (y − x) − f1 (y)|dy + |f (y)| · |f2 (y − x) − f2 (y)|dy.
Rn Rn

The first integral tends to zero (using again the compact support of f1 and its uniform
continuity), the second is bounded by 2tkf k2 < ∞.
iii) Fh = (2π)n/2 |Ff |2 ≥ 0;
iv) Finally, we obtain
Z Z
2 |ξ|2 /2
kFf k22 = (2π) −n/2
(Fh)(ξ)dξ = lim (2π) −n/2
(Fh)(ξ)e−ε dξ
Rn ε→0 Rn
Z
2 2
= lim ε−n (2π)−n/2 h(x)e−|x| /(2ε ) dx
ε→0 R n
Z
= h(0) = f (x)f^(−x)dx = kf k22 ,
Rn
where we used the continuity of h at zero and
 Z 
 −n −|x|2 /(2ε2 )


lim ε (2π) −n/2
h(x)e dx − h(0)
ε→0 Rn
Z
−n −n/2 2 2
= lim ε (2π) |h(x) − h(0)|e−|x| /(2ε ) dx
ε→0
ZR
n

2 2
≤ lim ε−n (2π)−n/2 |h(x) − h(0)|e−|x| /(2ε ) dx
ε→0 x:|x|≥t
Z
−n −n/2 2 2
+ lim ε (2π) |h(x) − h(0)|e−|x| /(2ε ) dx.
ε→0 x:|x|≤t

The first integral goes (for every t > 0) to zero11 , the second can be made arbitrary
small by choosing t > 0 small.

14.4 Few applications of Fourier transform

14.4.1 Bochner’s Theorem
We say, that a complex-valued function Φ : Rn → C is positive-definite if the matrix
(Φ(xj − xk ))N j,k=1 is positive semi-definite for every N ∈ N and every choice of points
x1 , . . . , xN ∈ Rn , i.e. if
XN
cj ck Φ(xj − xk ) ≥ 0
j,k=1
2
/2ε2
11
The integrated functions go pointwise to zero, and as sup ε−n e−|x| is finite, the integrable
|x|≥t,0<ε<1
majorant is |h(x) − h(0)| ∈ L1 (Rn ).

38
 14 Distributions and Fourier transform

for every N ∈ N, every x1 , . . . , xN ∈ Rn and every c = (c1 , . . . , cN ) ∈ CN .

Let Φ be a Fourier transform of a finite positive Borel measure ν, i.e.
Z
1
Φ(ξ) = e−ihx,ξi dν(x), ξ ∈ Rn .
(2π)n/2 Rn
We observe, that each such function is positive-definite by
N
X N
X Z
1
cj ck Φ(xj − xk ) = cj ck · n/2
e−ihxj −xk ,ξi dν(ξ)
j,k=1 j,k=1
(2π) R n

 
Z X N
1 
= cj ck e−ihxj −xk ,ξi  dν(ξ)
(2π)n/2 Rn j,k=1
 
Z X N
1
=  cj e−ihxj ,ξi ck e−ihxk ,ξi  dν(ξ)
(2π)n/2 Rn j,k=1
 2
Z X N 
1  
 −ihxj ,ξi 
=
(2π)n/2 Rn  j=1
cj e  dν(ξ) ≥ 0.


This gives immediately the easy part of the following theorem.

Theorem 14.4.1. The Fourier transform of every positive Borel measure on Rn is a
positive-definite function. On the other hand, if Φ : Rn → C is continuous and positive-
definite, then Φ is a Fourier transform of a finite positive Borel measure.
Open problem: If f : Rn → C is a positive and positive-definite continuous function
(i.e. f (x) is real, f (x) ≥ 0 for all x ∈ Rn and f = Fµ for some positive finite Borel
measure µ) with f (0) = 1, then
N
X |c1 + · · · + cN |2
cj ck f (xj − xk ) ≥
N
j,k=1

for all x1 , . . . , xN ∈ Rn and all c = (c1 , . . . , cN ) ∈ CN .

14.4.2 Translation invariant spaces

A subspace M of L2 (Rn ) is called translation invariant if f ∈ M implies fα (x) := f (x −
α) ∈ M for all α ∈ Rn .
The task we are going to investigate in this section sounds:

Describe the closed translation invariant subspaces of L2 (Rn ).

We shall proceed as follows. We denote by M̂ the image of M under F. Then M̂ is a

closed subspace of L2 (Rn ) (recall that F is an isometry on L2 (Rn )). From the properties
of Fourier transform, we obtain that fˆ ∈ M̂ implies also e−ihα,xi fˆ ∈ M̂ . Hence, M̂ is
invariant under multiplication with e−ihα,xi .
Let E ⊂ Rn be any measurable subset of Rn . Surely, if M̂ = {ϕ ∈ L2 (Rn ) : ϕ =
0 a.e. on E}, then M̂ is closed under multiplications with e−ihα,xi . Moreover, M̂ is closed
(and M̂ ⊥ = {ψ ∈ L2 (Rn ) : ψ = 0 a.e. on Rn \ E}). If M is the inverse image of M̂ , under
Fourier transform, then M has exactly the desired properties.

39
Functional Analysis II

One may now conjecture, that every closed translation invariant subspace is obtained
exactly in this manner. So, to every closed translation invariant subspace M we have to
construct a set E ⊂ Rn , such that f ∈ M if, and only if, Ff (x) = 0 a.e. on E. The
obvious construction
\ \
E= Ef = {x ∈ Rn : Ff (x) = 0}
f ∈M f ∈M

runs into serious difficulties, when we realize that each Ef is defined only up to set of
measure zero and that there are uncountably many f ∈ M . Hence, we lose every control
about E.
So, let M be a closed translation invariant subspace of L2 (Rn ) and let M̂ be its image
under Fourier transform. Let P denote the orthogonal projection onto M̂ . Hence, to each
f ∈ L2 (Rn ) there is P f ∈ M̂ , such that f − P f ⊥ M̂ , i.e.

(f − P f ) ⊥ P g, f, g ∈ L2 (Rn )

and also
(f − P f ) ⊥ P g(x)e−ihα,xi , f, g ∈ L2 (Rn ), α ∈ Rn .

This is equivalent to
Z
(f − P f )(x)P g(x)e−ihα,xi dx = 0, f, g ∈ L2 (Rn ), α ∈ Rn ,
Rn

i.e. that F((f − P f )P g) = 0. As both f − P f and P g belong to L2 (Rn ), their product is

in L1 (Rn ) and the uniqueness theorem on Fourier transform gives that (f − P f )P g = 0
almost everywhere. This remains true also if we replace P g by P g, hence

f · P g = (P f ) · (P g), f, g ∈ L2 (Rn ).

Interchanging f and g leads finally to

f · P g = P f · g, f, g ∈ L2 (Rn ).

This may be (roughly speaking) interpreted as that (P f )/f is constant for all f ∈ L2 (Rn ).
To avoid the devision by zero here, we consider some strictly positive function in L2 (Rn ),
2
for example g(x) = e−|x| will do, and put

(P g)(x)
ϕ(x) := , x ∈ Rn .
g(x)

Then we get P f = ϕ · f a.e. on Rn . Now we observe that

ϕ2 · g = ϕ · P g = P 2 g = P g = ϕ · g,

i.e. ϕ2 = ϕ. This means, that ϕ = 0 or ϕ = 1 a.e. and we put E := {x ∈ Rn : ϕ(x) = 0}.

Now f ∈ M̂ if, and only if, f = P f = ϕ · f gives that M̂ consists of exactly those functions
which vanish a.e. on E. This finishes the proof of the claim.
Let us mention that characterization of shift translation invariant spaces (i.e. space
invariant under integer translations) was an open problem until recently.

40
 14 Distributions and Fourier transform

14.4.3 Fast Fourier Transform

Let ZN be the set of the N th roots of unity, i.e. the set
ZN = {1, e2πi/N , e2·2πi/N . . . , e(N −1)·2πi/N }.
Then ZN with the usual complex multiplication is an Abelian (i.e. commutative) group.
Furthermore, ZN is isomorph to {0, 1, . . . , N } equipped with summation modulo N . And
it is also isomorph to Z/N Z, the set of equivalence classes based on the reminder when
dividing by N .
Let us put
el (k) = e2πilk/N for l = 0, 1, . . . , N − 1 and k = 0, 1, . . . , N − 1.
Furthermore, we denote by V the vector space of complexed-valued functions on ZN with
N
X −1
hF, GiV = F (k)G(k),
k=0
N
X −1
kF k2V = |F (k)|2 .
k=0

Through a simple calculation, one obtains immediately

Lemma 14.4.2. hel , em iV = N · δm,l .
√
Therefore, e∗l = el / N , l = 0, 1, . . . , N − 1, is an orthonormal basis of V .
Hence, for every F ∈ V , we get
N
X −1
F = hF, e∗n iV e∗n ,
n=0
N
X −1
kF k2 = |hF, e∗n iV |2 .
n=0

The nth Fourier coefficient of F is defined as

N −1
1 X 1
an = F̂ (n) = F (k)e−2πikn/N = √ hF, e∗n i.
N N
k=0

Theorem 14.4.3. For F ∈ V we have

N
X −1 N
X −1 √ N
X −1 N
X −1
F (k) = hF, e∗n iV e∗n (k) = N an e∗n (k) = an en (k) = an e2πink/N ,
n=0 n=0 n=0 n=0
N
X −1 NX−1 N
X −1
1 1 1
|an |2 = |hF, e∗n iV |2 = kF k2 = |F (k)|2 .
N N N
n=0 n=0 k=0

Naive way how to compute F̂ (0), . . . , F̂ (N − 1) from F (0), . . . , F (N − 1) given and

ωN = e−2πi/N given is
N −1
N 1 X kr
ak (F ) := F (r)ωN .
N
r=0
2 , . . . , ω N −1
It involves N − 2 multiplications to get ωN and each aN
N k needs N + 1 multipli-
cations and N − 1 additions. Therefore, we need 2N 2 + N − 2 ≤ 2N 2 + N operations.

41
Functional Analysis II

Theorem 14.4.4. (Fast Fourier Transform) Given ωN = e−2πi/N for N = 2n we need at

most
4 · 2n · n = 4N log2 (N ) = O(N log N )
operations to calculate all Fourier coefficients of F .
Proof. Let #(M ) be the minimum number of operations needed to calculate all Fourier
coefficients on ZM . We claim that

#(2M ) ≤ 2#(M ) + 8M

provided ω2M = e−2πi/(2M ) is given.

Using the claim, the rest follows by induction. For N = 21 = 2, we need less than 8
operations to calculate

aN
0 (F ) = 1/2(F (1) + F (−1)), aN
1 (F ) = 1/2(F (1) − F (−1)).

If the statement is true up to N = 2n−1 , we get

#(2N ) ≤ 2 · 4 · 2n−1 (n − 1) + 8 · 2n−1 = 8n2n−1 = 4n2n .

Proof of the claim:

We need at most 2M operations to get ω2M 2 , . . . , ω 2M −1 . Furthermore, for F defined
2M
on Z2M , we consider F0 and F1 defined on ZM , which are given by F0 (r) = F (2r) and
F1 (r) = F (2r+1). We also assume, that we were able to calculate their Fourier coefficients
(in ZM ) in #(M ) operations.
The claim is then proven by the following calculation (0 ≤ k ≤ 2M − 1)12
2M −1 M −1 M −1
!
2M 1 X kr 1 1 X k(2l) 1 X k(2m+1)
ak (F ) = F (r)ω2M = F (2l)ω2M + F (2m + 1)ω2M
2M 2 M M
r=0 l=0 m=0
M −1 M −1
!
1 1 X 1 X
kl km k
= F0 (l)ωM + F1 (m)ωM ω2M
2 M M
l=0 m=0
1 M 
= ak (F0 ) + aMk (F k
1 2M .
)ω
2

14.4.4 Uncertainty principle

The uncertainty principle in its most simple form is the following inequality (sometimes
called Heisenberg-Pauli-Weyl inequality).
Theorem 14.4.5. If f ∈ L2 (R) and a, b ∈ R are arbitrary, then
Z ∞ 1/2 Z ∞ 1/2
2 2 2 2 1
(x − a) |f (x)| dx · (ξ − b) |Ff (ξ)| dξ ≥ kf k22 . (14.8)
−∞ −∞ 2
Let us give first a simple proof for f ∈ S (R). By substitution (i.e. by considering
e−ihx,bi f (x − a) instead of f ), we may restrict ourselves to a = b = 0. Furthermore, we
suppose that kf k2 = 1. Then we get
Z ∞ Z ∞ Z ∞ 
2 d 2 ′ ′
1= |f (x)| dx = − x |f (x)| dx = − xf (x)f (x) + xf (x)f (x) dx,
−∞ −∞ dx −∞
12
Let us observe, that aM M
k (F0 ) = ak−M (F0 ) if k ≥ M.

42
 14 Distributions and Fourier transform

where we have used partial integration and the identity |f |2 = f f. Hence

Z ∞ Z ∞ 1/2 Z ∞ 1/2
′ 2 2 ′ 2
1≤2 |x| · |f (x)| · |f (x)|dx ≤ 2 x |f (x)| dx |f (x)| dx .
−∞ −∞ −∞

The proof is then finished by Parseval’s identity:

kf ′ k2 = kF(f ′ )k2 = kξ · Ff (ξ)k2 .
Theorem 14.4.6. Let A and B be (possibly unbounded) self-adjoint operators on a Hilbert
space H. Then
1
k(A − aI)f k · k(B − bI)f k ≥ |h[A, B]f, f i|
2
for all a, b ∈ R and all f in the domain of AB and BA. Furthermore, [A, B] = AB − BA
is the commutator of A and B.
Proof.
h[A, B]f, f i = h{(A − aI)(B − bI) − (B − bI)(A − aI)}f, f i
= h(B − bI)f, (A − aI)f i − h(A − aI)f, (B − bI)f i
= 2i Imh(B − bI)f, (A − aI)f i
and the statement follows by applying the Cauchy-Schwartz inequality.

The proof of Theorem 14.4.5 then follows by choosing

Xf (x) = xf (x), P f (x) = if ′ (x).
If f is the domain of P X and XP (and let us observe, that the left-hand side of (14.8)
is infinity if this is not the case), we obtain
[X, P ]f (x) = ixf ′ (x) − i(xf (x))′ = −if (x)
and
kf k22 1
= |h−if (x), f (x)i| ≤ k(X − aI)f k2 · k(P − bI)f k2 ,
2 2
where
k(P − bI)f k2 = kF(P − bI)f k2 = k(ξ − b)Ff (ξ)k2

14.5 More on convolutions

When defining the convolution of a distribution and a test function, we have two pos-
sibilities. The first one is based on the calculation (for Tf a regular distribution and
ϕ, ψ ∈ S (Rn ))
Z Z Z
(Tf ∗ ϕ)(ψ) = (f ∗ ϕ)(x)ψ(x)dx = ψ(x) f (y)ϕ(x − y)dydx
Rn Rn Rn
Z Z Z 
= f (y) ψ(x)ϕ(x − y)dxdy = Tf ψ(x)ϕ(x − y)dx ,
Rn Rn Rn

the other on the observation that

Z
(f ∗ ϕ)(x) = f (y)ϕ(x − y)dy = Tf (ϕ(x − ·)) = Tf (τx ϕ̃), x ∈ Rn .
Rn

We shall use the first formula as the definition and afterwards, we show that it coincides
with the second one.

43
Functional Analysis II

Definition 14.5.1. Let T ∈ S ′ (Rn ) and ϕ ∈ S (Rn ). Then we define

(ϕ ∗ T )(ψ) = (T ∗ ϕ)(ψ) = T (ϕ̃ ∗ ψ), ψ ∈ S (Rn ),

where ϕ̃(x) = ϕ(−x) is the reflexion of ϕ.

Example 14.5.2. Let T = δx0 . Then ϕ ∗ T is the function x → ϕ(x − x0 ). This follows
quickly from
Z
(ϕ ∗ T )(ψ) = δx0 (ϕ̃ ∗ ψ) = (ϕ̃ ∗ ψ)(x0 ) = ψ(x)ϕ(x − x0 )dx.
Rn

It follows, that convolution with δ0 is the identity operator.

Due to the impossibility of multiplication inside the space of distributions and due to
the properties of the Fourier transform (see below), impossibility of convolutions of general
distributions follows.
Proposition 14.5.3. Let T ∈ S ′ (Rn ) and ϕ ∈ S (Rn ). Then T ∗ ϕ is a regular distribu-
tion and
(T ∗ ϕ)(x) = T (τx ϕ̃), x ∈ Rn .
Proof. We have to show that
Z  Z 
(T ∗ ϕ)(ψ) = T (ϕ̃ ∗ ψ) = T ϕ̃(· − y)ψ(y)dy =T (τy ϕ̃)(·)ψ(y)dy
Rn Rn

is equal to Z
T (τy ϕ̃(·))ψ(y)dy,
Rn
i.e. that we may interchange the integration and the application of T . To do this, it is
enough to prove the convergence of the Riemann sums of the integral under discussion in
S (Rn ). We sketch the arguments but leave out the details.
For each N ∈ N, we partition [−N, N ]n into (2N 2 )n cubes Qm , m = 1, . . . , (2N 2 )n
with centers ym . One has to show that the functions
1 X
x→ (τym ϕ̃)(x)ψ(ym )
Nn m

converge in S (Rn ) to Z
x→ (τy ϕ̃)(x)ψ(y)dy
Rn
as N → ∞. Although not difficult, we leave out the technical details.

Definition 14.5.4. Let T ∈ S ′ (Rn ). Then the support of T is the intersection of all
closed sets K ⊂ Rn , such that T (ϕ) = 0 for all ϕ ∈ S (Rn ) with supp ϕ ⊂ Rn \ K.
The support of the distribution δx0 is exactly the set {x0 }.
Lemma 14.5.5. Let T ∈ S ′ (Rn ). Then T has compact support if, and only if, there
exists a constant C > 0 and l, N ∈ N such that
X
|T (ϕ)| ≤ C sup |D α ϕ(x)|, ϕ ∈ S (Rn ).
|x|≤N |α|≤l

44
 14 Distributions and Fourier transform

Proof. Let the estimate be satisfied and let ψ ∈ S (Rn ) with supp ψ ⊂ Rn \ {x ∈ Rn :
|x| ≤ N } = {x ∈ Rn : |x| > N }. Then obviously |T (ψ)| = 0 and we conclude that T has
compact support.
Let on the other hand T have compact support, i.e. supp T ⊂ {x :∈ Rn : |x| ≤ M }.
We consider an infinitely differentiable function ψ with ψ(x) = 1 for |x| ≤ M and ψ(x) = 0
for |x| ≥ M + 1. Using Theorem 14.2.3, we get

|T (ϕ)| = |T (ϕψ) + T (ϕ(1 − ψ))| = |T (ϕψ)|

X
≤ c kϕψk(k,l) = c sup (1 + |x|2 )k/2 |D α (ϕψ)(x)|
x∈Rn
|α|≤l
X
=c sup (1 + |x|2 )k/2 |D α (ϕψ)(x)|
|x|≤M +1 |α|≤l
X
≤ c′ sup |D α ϕ(x)|
|x|≤M +1 |α|≤l

and we obtain the statement with N = M + 1.

Theorem 14.5.6. Let T ∈ S ′ (Rn ) and ϕ ∈ S (Rn ). Then

i) T ∗ ϕ is a C ∞ -function,

ii) D α (T ∗ ϕ) = (D α T ) ∗ ϕ = T ∗ (D α ϕ) for all α ∈ Nn0 .

iii) For all multiindices α ∈ Nn0 there exist constants Cα , kα > 0, such that

|D α (T ∗ ϕ)(x)| ≤ Cα (1 + |x|2 )kα /2 , x ∈ Rn . (14.9)

iv) If T has compact support, then T ∗ ϕ is in S (Rn ).

Proof. We already know that T ∗ϕ is a regular distribution given pointwise by (T ∗ϕ)(x) :=

T (τx ϕ̃), x ∈ Rn . Next, we show that this is a C ∞ -function. Let ej = (0, . . . , 0, 1, 0, . . . , 0)
be the canonical unit vector in Rn with jth coordinate equal to 1 and zeros otherwise. We
shall use (cf. Exercises) that the test functions

τhej ψ − ψ
−h

converge to D ej ψ in S (Rn ) for all ψ ∈ S (Rn ) as h → 0. This gives for ψ = τx ϕ̃

τx+hej ϕ̃ − τx ϕ̃ S
→ (−1) · D ej (τx ϕ̃) = (−1) · τx (D ej ϕ̃)
h
and, after applying T ,
 
(T ∗ ϕ)(x + hej ) − (T ∗ ϕ)(x) τx+hej ϕ̃ − τx ϕ̃
=T → T (−τx (D ej ϕ̃)).
h h

We combine this with T ∗ (D ej ϕ)(x) = T (τx (D ^ ej ϕ)) = T (−τ (D ej ϕ̃)) = T (−D ej (τ ϕ̃)) =
x x
e
(D T ) ∗ ϕ(x) and finish the proof of (ii) for α = ej . Iterating this identity, we obtain that
j

T ∗ ϕ is infinitely differentiable and that (ii) holds for every α ∈ Nn0 .

45
Functional Analysis II

The proof of (14.9) is based on Theorem 14.2.3. We obtain

]
|D α (T ∗ ϕ)(x)| = |T ∗ (D α ϕ)(x)| = |T (τx (D α ϕ))|
X
]
≤ c kτx (D α ϕ)k 2 k/2
(k,l) ≤ c sup (1 + |y| )
]
|D β τx (D α ϕ)(y)|
y∈Rn
|β|≤l
X
2 k/2 α+β
≤ c sup (1 + |y| ) |τx (D ϕ̃)(y)|
y∈Rn
|β|≤l
X
= c sup (1 + |y|2 )k/2 |(D α+β ϕ̃)(x + y)|
y∈Rn
|β|≤l
X
≤ c sup (1 + |z − x|2 )k/2 |(D α+β ϕ̃)(z)|
z∈Rn
|β|≤l
X
≤ c (1 + |x|2 )k/2 · sup (1 + |z|2 )k/2 |(D α+β ϕ̃)(z)| ≤ c (1 + |x|2 )k/2 kϕ̃k(k,|α|+l) .
z∈Rn
|β|≤l

Finally, to show that T ∗ ϕ ∈ S (Rn ) for T with compact support, we use Lemma
14.5.5. We get for arbitrary k ∈ N
X
|(T ∗ ϕ)(x)| = |T (ϕ(x − ·))| ≤ C sup |Dyα ϕ(x − y)|
|y|≤N |α|≤l

≤ Cϕ′ sup (1 + |x − y|2 )−k/2 ≤ Cϕ′′ (1 + |x|2 )−k/2

|y|≤N

for |x| ≥ 2N. Hence (T ∗ ϕ)(x) decays faster then any polynomial. The same argument
applied to D α (T ∗ ϕ) = T ∗ (D α ϕ) yields the same for all the derivatives of T ∗ ϕ. Hence
T ∗ ϕ ∈ S (Rn ).

Finally, we show that products and convolutions of distributions behave under Fourier
transform exactly as we expect them to do.
Theorem 14.5.7. Let T ∈ S ′ (Rn ) and ϕ ∈ S (Rn ). Then
F(ϕT ) = (2π)−n/2 F(ϕ) ∗ F(T ),
F(ϕ ∗ T ) = (2π)n/2 F(ϕ)F(T ).
Proof. We obtain for every ψ ∈ S (Rn )
F(ϕT )(ψ) = (ϕT )(Fψ) = T (ϕF(ψ))
and
g ∗ ψ) = T (F(Fϕ
F(ϕ) ∗ F(T )(ψ) = FT (Fϕ g ∗ ψ)) = (2π)n/2 T (ϕF(ψ)).

The other identity follows in the same manner.

14.6 Paley-Wiener Theorem

Paley-Wiener’s Theorem is in general any statement connecting the decay of a function
or a distribution with the smoothness of its Fourier transform (and vice versa). Although
proven by Paley and Wiener for square-integrable functions first, it was adapted to test
functions and distributions by Schwartz (and Hörmander). We give one important example
of such a theorem below.
Another remarkable fact about Paley-Wiener theorems is that they connect the theory
of distributions with the theory of functions of several complex variables. Therefore, we
first (re-)introduce some notation from complex analysis.

46
 14 Distributions and Fourier transform

Definition 14.6.1. Let Ω ⊂ Cn be an open set and let f be a complex-valued function on

Ω. Then f is said to be holomorphic in Ω, if it is holomorphic in each variable separately.
A function f of n complex variables is said to be entire if it is holomorphic on whole Cn .
For any z = (z1 , . . . , zn ) ∈ Cn , we shall write x = Re z = (Re z1 , . . . , Re zn ) ∈ Rn and
y = Im z ∈ Rn for real and imaginary part of z.
Let T be a tempered distribution on Rn , i.e. T ∈ S ′ (Rn ), with compact support.
Then the FourierR transform of T was defined in Definition 14.2.5. However, the definition
Ff (ξ) = (2π)1n/2 Rn f (x)e−ihx,ξi dx, f ∈ S (Rn ), suggests that we could define also
!
e−ih·,ξi
(FT )(ξ) := T , (14.10)
(2π)n/2

whenever the right-hand side makes sense. As e−ihx,·i is not in S (Rn ), some care is
of course necessary. If T has compact support, then T (e−ihx,·i ) might be defined as
T (e−ihx,·i ϕ(·)) for any ϕ ∈ S (Rn ) with ϕ = 1 on supp T. Obviously, this “definition” does
not depend on the choice of ϕ, i.e. T (e−ihx,·i ϕ(·)) = T (e−ihx,·i ψ(·)) if both ϕ, ψ ∈ S (Rn )
with ϕ = ψ = 1 on supp T.
First, using ψ ∈ S (Rn ) with ψ(x) = 1 for x ∈ supp T , we get immediately that
FT = F(ψT ) = (2π)−n/2 Fψ ∗ FT
is a C ∞ -function (cf. Theorem 14.5.6). Furthermore, with Φ ∈ S (Rn ) with FΦ = ψ, we
get
(2π)n/2 FT (ξ) = (FT ∗ Fψ)(ξ) = (FT ∗ Φ̃)(ξ) = FT (τξ Φ)
= T (F(τξ Φ)) = T (e−ih·,ξi ψ),
which we defined as T (e−ih·,ξi ). Therefore, also (14.10) is justified.
Let us observe that (14.10) makes sense also for z ∈ Cn instead of ξ ∈ Rn . This
mapping (which is an extension of Fourier transform to complex arguments) is usually
called Fourier-Laplace transform and is defined by
!
e−ih·,zi
(FT )(z) := T , z ∈ Cn . (14.11)
(2π)n/2
Note, that we denote it by F again.
Its properties are given in the following version of (Schwartz-)Paley-Wiener’s theorem.
Theorem 14.6.2. Let T ∈ S ′ (Rn ) have compact support. Then its Fourier-Laplace
transform is an entire function with polynomial growth on Rn .
Proof. First, we observe that
!
ej FT (z + hej ) − FT (z) 1 e−ihx,z+hej i − e−ihx,zi
D (FT )(z) = lim = lim T .
h→0 h (2π)n/2 h→0 h
The fact, that FT is an entire function, now follows from the compact support of T and
the convergence of the sequence
e−ihx,z+hej i − e−ihx,zi
x→ , x ∈ Rn ,
h
to −ixj e−ihx,zi in C ∞ (Rn ). This gives D ej (FT )(z) = T (−ixj e−ihx,zi ). This argument
might be iterated and we obtain D α (FT )(z) = T ((−ixj )α e−ihx,zi ) for all α ∈ Nn0 . Moreover,
the polynomial growth on Rn follows from (14.9).

47
Functional Analysis II

14.7 Fundamental solution

Theorem 14.7.1. Let T ∈ S ′ (Rn ) be supported at the singleton {x0 }. Then there exists
a number k ∈ N0 and complex numbers aα for |α| ≤ k, such that
X
T = aα D α δx0 .
|α|≤k

Proof. We may assume, that x0 = 0. We shall use Lemma 14.5.5. We assume that
X
|T (ϕ)| ≤ C sup |D α ϕ(x)|
|x|≤N |α|≤l

for all ϕ ∈ S (Rn ). First we show, that if ϕ ∈ S (Rn ) satisfies

(D α ϕ)(0) = 0, |α| ≤ l, (14.12)

then T ϕ = 0. We fix one such a ϕ ∈ S (Rn ) and we chose ψ ∈ C ∞ (Rn ) with ψ = 0 on

{x ∈ Rn : |x| ≤ 1} and ψ = 1 on {x ∈ Rn : |x| ≥ 2}. We put ψε (x) := ψ(x/ε) for ε > 0.

|T (ϕ)| = |T (ϕ(ψε + (1 − ψε )))| ≤ |T (ϕψε )| + |T (ϕ(1 − ψε ))|

X
≤ 0 + C sup |D α (ϕ(1 − ψε ))(x)|
|x|≤N |α|≤l
X
≤ C sup |D α (ϕ(1 − ψε ))(x)|
|x|≤2ε |α|≤l

and it is rather easy to see (using (14.12)), that the last expression tends to zero as ε → 0.
Now, let ϕ ∈ S (Rn ) be arbitrary and let η ∈ S (Rn ) be a function that is equal to 1
in a neighborhood of origin. Then (using Taylor’s polynomial)
 
X D α ϕ(0)
ϕ(x) = η(x)  xα + ψ(x) + (1 − η(x))ϕ(x),
α!
|α|≤l

where ψ ∈ S (Rn ) satisfies ψ(x) = O(|x|l+1 ) as |x| → 0.

Now T (ηψ) = 0 due to the first part of the proof and T ((1 − η)ϕ) = 0 due to the
support property of T . Hence
X D α ϕ(0) X
T (ϕ) = T (η(x)xα ) = aα (D α δ0 )(ϕ)
α!
|α|≤l |α|≤l

for aα = (−1)|α| T (η(x)xα )/α!.

Corollary 14.7.2. Let T ∈ S ′ (Rn ) with FT supported at the singleton {ξ0 }. Then T is a
linear combination of functions ξ α eihξ,ξ0 i . In particular, if FT is supported at the origin,
then T is a polynomial.

The Laplacian is the partial differential operator

n
X ∂2T
∆(T ) = , T ∈ S ′ (Rn ).
j=1
∂x2j

Solutions of the Laplace equation ∆(T ) = 0 are called harmonic distributions.

48
 14 Distributions and Fourier transform

Corollary 14.7.3. Let T ∈ S ′ (Rn ) be a harmonic distribution. Then T is a polynomial.

Proof. Taking the Fourier transform, we observe that |ξ|2 ·FT = 0. Hence, FT is supported
at the origin and, therefore, must be a polynomial.

Liouville’s classical theorem says that every bounded harmonic function must be con-
stant. We observe that it follows directly from Corollary 14.7.3.
A distribution T ∈ S ′ (Rn ) is called a fundamental solution of a partial differential
operator with constant coefficients L if L(T ) = δ0 .

Theorem 14.7.4. For n ≥ 3, we have

2π n/2
∆(|x|2−n ) = −(n − 2)ωn−1 δ0 = −(n − 2) δ0
Γ(n/2)

and for n = 2,
∆(log |x|) = 2πδ0 .
Here, ωn−1 stands for the (n − 1)-dimensional measure of the unit sphere in Rn .

Proof. We use the Green’s identity

Z Z  
∂u ∂v
v∆u − u∆vdx = v −u ds,
Ω ∂Ω ∂ν ∂ν

where Ω is a domain in Rn with smooth boundary and ∂/∂ν denotes the derivative with
respect to the outer normal vector to ∂Ω.
Let Ωε = {x ∈ Rn : ε < |x| < 1/ε}. Furthermore, we choose (for n ≥ 3) v(x) = |x|2−n
and u = ϕ ∈ S (Rn ). This leads to
Z Z
2−n
∆(|x| )(ϕ) = v(∆ϕ) = v(x)∆ϕ(x)dx = lim |x|2−n ∆ϕ(x)dx
Rn ε→0 Ωε

and
Z Z Z  
2−n 2−n ∂|x|2−n ∂ϕ(x)
|x| ∆ϕ(x)dx = ∆(|x| )ϕ(x)dx − ϕ(x) − |x|2−n ds.
Ωε Ωε ∂Ωε ∂ν ∂ν

The first integral is equal to zero as ∆(|x|2−n ) = 0 for x 6= 0.

The second integral splits into two parts, integral over {x ∈ Rn : |x| = ε} and {x ∈
n
R : |x| = 1/ε}. We observe quickly that
Z  
∂|x|2−n ∂ϕ(x)
ϕ(x) − |x|2−n ds → 0
{x∈Rn :|x|=1/ε} ∂ν ∂ν

as ε → 0. This is due to
Z  
∂|x|2−n 2−n ∂ϕ(x)
ϕ(x) − |x| ds
{x∈Rn :|x|=1/ε} ∂ν ∂ν
Z  
n−1 n−2 ∂ϕ(x)
= ϕ(x)(2 − n)ε −ε ds
{x∈Rn :|x|=1/ε} ∂ν

and the rapid decay of ϕ ∈ S (Rn ) and all its derivatives at infinity.

49
Functional Analysis II

As for the integral over {x ∈ Rn : |x| = ε}, it is equal to

Z Z
∂ϕ(x)
(2 − n)ε1−n ϕ(x)ds + ε2−n ds.
{x∈Rn :|x|=ε} {x∈Rn :|x|=ε} ∂ν

Finally, the first integral tends to (2 − n)ϕ(0)ωn−1 , where ωn−1 is the (n − 1)-dimensional
measure of the unit sphere in Rn , and the second integral tends to zero.
Let us note, that the formula ωn−1 = 2π n/2 /Γ(n/2) follows easily by the following
trick. Using polar coordinates in Rn and the substitution t := r 2 , we get
Z Z ∞ Z
√ 2 2 ωn−1 ∞ −t n/2−1 ωn−1
( π)n = e−|x| dx = ωn−1 e−r r n−1 dr = e t dt = Γ(n/2).
Rn 0 2 0 2
The proof for n = 2 is very similar.

Remark 14.7.5. Due to the previous Corollary, the fundamental solution of Laplacian is
not unique. We may add to any of the fundamental solutions above a harmonic polynomial
and the Laplace operator of such a distribution would not change.
Although one could develop a full theory of fundamental solutions in the spirit of
the book of Rudin (with Theorem of Malgrange-Ehrenpreis ensuring the existence of a
fundamental solution for any partial differential operator with constant coefficients as the
main highlight), we shall give a couple of examples of this technique instead.
Let us assume that L(T ) = δ0 , i.e. that T is a fundamental solution to L. Then the
solution of the equation L(u) = f can be obtained by convolution.

L(T ∗ f ) = (LT ) ∗ f = δ0 ∗ f = f.

Of course, the interpretation of T ∗ f can cause problems if T ∈ S ′ (Rn ) and f is not

smooth enough, i.e. f 6∈ S (Rn ).
Remark 14.7.6. Using previous theorem and the above presented arguments, one may
immediately write down the solution (or, better said, all the solutions) to the equation
∆T = ϕ, at least for ϕ ∈ S (Rn ). The obvious disadvantage is of course, that the usual
physical problems are defined only on some domain of Rn . We shall deal with this problem
later on when considering distributions on domains.
We shall now give another application of the technique of fundamental solution. Let
us consider the heat equation:

ut (x, t) − kuxx (x, t) = 0, x ∈ R, 0 < t < ∞,

u(x, 0) = g(x).

The fundamental solution of these equation will be obtained by choosing g = δ0 .

Let us sketch the way to the fundamental solution. We apply the Fourier transform to
this equations in x variable (i.e. for every t fixed). This gives

Fx [ut (x, t) − kuxx (x, t)] = 0, Fx [u(x, 0)] = Fx [δ0 ],

i.e.

vt (ξ, t) + kξ 2 v(ξ, t) = 0, v(ξ, 0) = (2π)−1/2 ,

where now v(ξ, t) = Fx (u(x, t))(ξ). We solve these equations for each ξ ∈ R separately. We
fix ξ ∈ R and put z(t) = v(ξ, t) and obtain z ′ (t) + kξ 2 z(t) = 0 and z(0) = (2π)−1/2 . This

50
 14 Distributions and Fourier transform

2 2 2
leads to z(t) = (2π)−1/2 e−kξ t and v(ξ,√t) = (2π)−1/2 e−kξ t . If we denote h(s) = e−s /2 , i.e.
Fh = h, we get v(ξ, t) = (2π)−1/2 h( 2ktξ) and finally the fundamental solution √ to the
one-dimensional heat equation Φ(x, t) = (Fξ−1 v(ξ, t))(x) = (2π)−1/2 (2kt)−1/2 h(x/ 2kt) =
 
1 x2
√ exp − .
4πkt 4kt
We now claim that the solution to the heat equation with a general right hand side g
is then given as the convolution of Φ with g in the x variable, i.e.
Z
u(x, t) = Φ(x − y, t)g(y)dy, −∞ < x < ∞, 0 < t < ∞.
R

Indeed, at least formally, we obtain

Z
ut (x, t) − kuxx (x, t) = [Φt (x − y, t) − kΦxx (x − y, t)] g(y)dy = 0
R

and Z Z
u(x, 0) = Φ(x − y, 0)g(y)dy = δ0 (x − y)g(y)dy = g(x).
R R
Finally, let us mention that the fundamental solution to the n-dimensional heat equa-
tion is obtained simply as a tensor product of n one-dimensional fundamental solutions,
i.e.  
1 |x|2
Φ(x, t) = exp − , x ∈ Rn , 0 < t < ∞.
(4πkt)n/2 4kt

14.8 Spaces D(Ω) and D ′ (Ω)

Unfortunately, the spaces S (Rn ) and S ′ (Rn ) together with Fourier transform F are not
well suited to deal with partial differential equations on domains. That is the reason, why
there is a way (and historically, it was actually introduced before the Schwartz space and
its dual), which allows to deal with distributions on domains.
Definition 14.8.1. Let Ω ⊂ Rn be an open set.
i) We define the space of test functions

D(Ω) = {ϕ ∈ C ∞ (Ω) : supp ϕ is a compact subset of Ω}.

ii) A sequence (ϕj )j∈N ⊂ D(Ω) is said to converge to ϕ ∈ D(Ω) in D(Ω), if there is a
compact set K ⊂ Ω, such that

supp ϕj ⊂ K, for all j ∈ N

and
D α ϕj ⇉ D α ϕ for all α ∈ Nn0 .
D
We write ϕj → ϕ to denote the convergence in D(Ω).

iii) D ′ (Ω) is the collection of all complex-valued linear continuous functionals over D(Ω),
i.e. T : D(Ω) → C belongs to D ′ (Ω) if, and only if,

T (λ1 ϕ1 + λ2 ϕ2 ) = λ1 T (ϕ1 ) + λ2 T (ϕ2 ) for all λ1 , λ2 ∈ C and ϕ1 , ϕ2 ∈ D(Ω),

D
T (ϕj ) → T (ϕ) whenever ϕj → ϕ.

The elements of D ′ (Ω) are called distributions.

51
Functional Analysis II

iv) The space D ′ (Ω) is turned into a vector space by setting

(λ1 T1 + λ2 T2 )(ϕ) = λ1 T1 (ϕ) + λ2 T2 (ϕ) for all λ1 , λ2 ∈ C, T1 , T2 ∈ D ′ (Ω) and ϕ ∈ D(Ω).

v) Finally, we also equip the space D ′ (Ω) with the notion of convergence. Namely, we
D′
say that Tj converges to T in D ′ (Ω), Tj ⇀ T , if

Tj (ϕ) → T (ϕ) for all ϕ ∈ D(Ω).

The spaces D(Ω) are much more flexible due to the inclusion of the domain into their
definition. Unfortunately, a lot of the algebraic structure is lost, for example Fourier
transforms and convolutions do not have an easy counterpart on D(Ω). Also modulations
and translations are only of a limited use on Ω.
Nevertheless, one can define the derivative of a distribution T ∈ D ′ (Ω) by (D α T )(ϕ) =
(−1)|α| T (D α ϕ) and its product with a smooth function ψ ∈ C ∞ (Ω) as (ψT )(ϕ) = T (ψϕ).
If Ω = Rn , we obtain two spaces of test function, and two spaces of distributions. Obvi-
ously, D(Rn ) ⊂ S (Rn ) (including a continuous embedding) and, consequently, S ′ (Rn ) ⊂
D ′ (Rn ), again also in the sense of continuous embedding.
The space D(Ω) allows very well to explain what is meant by weak solutions to PDE’s.
Let us for example consider the equation

∂u(t, x) ∂u(t, x)
+ = 0.
∂t ∂x
In the classical setting this means that we are looking for a function u ∈ C 1 (R2 ), such
that the equation holds pointwise. In the distributional sense, we require that
 
∂u(t, x) ∂u(t, x)
+ (ϕ) = 0
∂t ∂x

for all ϕ ∈ D(R2 ). But this is equivalent to

  Z
∂ϕ(t, x) ∂ϕ(t, x)
u + = u(t, x) (ϕt (t, x) + ϕx (t, x)) d(t, x) = 0.
∂t ∂x R2

Whenever u is smooth enough, these two notions coincide. But there are functions (i.e.
u(t, x) = |t − x|), which do satisfy the weak formulation, but can not satisfy the strong
formulation due to their lack of differentiability.

52
15 Introduction to harmonic analysis
15.1 Approximation of identity
Theorem 15.1.1. Let Ω ⊂ Rn be a domain. The set of continuous functions with compact
support contained in Ω is dense in Lp (Ω), 1 ≤ p < ∞.

Proof. We shall need two facts from measure theory.

i) Lebesgue measure λ in Rn is regular, i.e. λ(A) = inf{λ(G) : G ⊃ A, G open}.

ii) The space of step functions, i.e. span{χA : A ⊂ Ω, A measurable}, is dense in Lp (Ω)
for every 1 ≤ p < ∞.
S
We first consider open sets Ωj ⊂ Ω, j ∈ N, such that Ωj ⊂ Ωj ⊂ Ωj+1 ⊂ Ω and ∞ j=1 Ωj =
13
Ω. Let us take f ∈ Lp (Ω). Then f χΩj → f in Lp (Ω) and we may restrict ourselves
to f ∈ Lp (Ω) with compact support in Ω. Due to the second property of the Lebesgue
P
measure, this function may be approximated by a step function K k=1 ̺k χAk with Ak ⊂
supp f. So, it is enough to approximate characteristic functions χB with B compact in Ω.
Using the first property of the Lebesgue measure, we may restrict ourselves to bounded
open sets G ⊂ G ⊂ Ω. Then the sequence of functions x → max(0, 1 − k dist(x, G)) gives
the desired approximation.

Lemma 15.1.2. Let f ∈ Lp (Rn ), 1 ≤ p < ∞. Then f (· + h) → f (·) in Lp (Rn ) if h → 0.

Proof. If f is continuous with compact support, then the result follows by uniform con-
tinuity of f and the Lebesgue dominated convergence theorem. If f ∈ Lp (Rn ), we may
find for every t > 0 a continuous function g with compact support such that kf − gkp < t.
Then

kf (· + h) − f (·)kp ≤ kf (· + h) − g(· + h)kp + kg(· + h) − g(·)kp + kg(·) − f (·)kp

≤ 2t + kg(· + h) − g(·)kp

and the conclusion follows.

Theorem 15.1.3. The family of functions (Kε )ε>0 ⊂ L1 (Rn ) is called the approximation
of identity, if
R
(K1) Rn |Kε (x)|dx ≤ C < ∞ for all ε > 0,
R
(K2) Rn Kε (x)dx = 1 for all ε > 0,
R
(K3) limε→0+ |x|>δ |Kε (x)|dx = 0 for all δ > 0.

Then
R
i) If K ∈ L1 (Rn ) with Rn K(x)dx = 1, then Kε (x) = ε−n K(x/ε) is an approximation
of identity.

ii) If (Kε )ε>0 is an approximation of the identity, then

lim kKε ∗ f − f kp = 0
ε→0+

for every 1 ≤ p < ∞ and f ∈ Lp (Rn ).

13
For example the sets Ωj := {x ∈ Ω : |x| < j and dist(x, ∂Ω) > 1/j} will do.

53
Functional Analysis II

R
Proof. (i) Let K ∈ L1 (Rn ) with Rn K(x)dx = 1. Then we get immediately
Z Z Z Z
−n −n
ε K(x/ε)dx = K(x)dx = 1 and ε |K(x/ε)|dx = |K(x)|dx = kKk1 < ∞.
Rn Rn Rn Rn

As for the third point, we have

Z Z
|Kε (x)|dx = |K(y)|dy → 0
|x|>δ |y|>δ/ε

as ε → 0+ , due to the Lebesgue dominated convergence theorem.

(ii) We calculate for p > 1 and its conjugated index p′ with 1/p + 1/p′ = 1 using
Hölder’s inequality (if p = 1, the calculation becomes slightly simpler)
Z Z Z p
 
p
kKε ∗ f − f kp = p
|(Kε ∗ f )(x) − f (x)| dx =  Kε (y)f (x − y)dy − f (x) dx

Rn Rn
p R
n
Z Z
 
 
=  n Kε (y)[f (x − y) − f (x)]dy  dx
Rn R
Z Z p
′
≤ |Kε (y)|1/p+1/p · |f (x − y) − f (x)|dy dx
Rn Rn
Z Z Z p/p′
p
≤ |Kε (y)| · |f (x − y) − f (x)| dy · |Kε (y)|dy dx
Rn Rn Rn
Z Z
′
≤ C p/p |Kε (y)| · |f (x − y) − f (x)|p dydx
ZR Rn
n

p/p′
=C |Kε (y)| · kf (· − y) − f (·)kpp dy
Rn
(Z Z )
p/p′
≤C |Kε (y)| · kf (· − y) − f (·)kpp dy p
+2 kf kpp |Kε (y)|dy
|y|≤δ |y|>δ

for every δ > 0. Using (K3) and previous Lemma, we obtain the conclusion of the
theorem.

Definition 15.1.4. Let Ω ⊂ Rn be a domain. Then Cc∞ (Ω) denotes the set of infinitely-
differentiable functions compactly supported in Ω.

It was shown in exercises that this class is actually non-empty.

Theorem 15.1.5. Cc∞ (Ω) is dense in Lp (Ω) for every 1 ≤ p < ∞ and every domain
Ω ⊂ Rn .

Proof. Let f ∈ Lp (Ω). First, we approximate f by a continuous and compactly supported

g (i.e. kf − gkp ≤ h) and then consider the functions
R ωε ∗ g, where ωε (x) = ε−n ω(x/ε)
and ω ∈ Cc (R ) has compact support and ω = 1. It follows that ωε ∗ g ∈ Cc∞ (Rn )
∞ n

(the support property is clear, the differentiability was proven in a much more general
setting in Theorem 14.5.6). Together with the formula kωε ∗ g − gkp → 0, the conclusion
follows.

15.2 Maximal operator

For x ∈ Rn and r > 0, we denote by B(x, r) the ball in Rn with center at x and radius r,
i.e. B(x, r) = {y ∈ Rn : |x − y| < r}.

54
 15 Introduction to harmonic analysis

Let f be a locally integrable function on Rn . Then we define the Hardy-Littlewood

maximal operator of f by
Z
1
M f (x) = sup |f (y)|dy.
r>0 |B(x, r)| B(x,r)

Here stands |B(x, r)| for the Lebesgue measure of B(x, r).
Let us note, that maximal operator is not linear, but is sub-linear, i.e.
Z
1
M (f + g)(x) = sup |f (y) + g(y)|dy
r>0 |B(x, r)| B(x,r)
Z Z !
1
≤ sup |f (y)|dy + |g(y)|dy
r>0 |B(x, r)| B(x,r) B(x,r)

≤ (M f )(x) + (M g)(x).

We shall study the mapping properties of the operator M in the frame of Lebesgue
spaces Lp (Rn ), 1 ≤ p ≤ ∞. If f ∈ L∞ (Rn ), then
Z Z
1 kf k∞
|f (y)|dy ≤ 1dy = kf k∞
|B(x, r)| B(x,r) |B(x, r)| B(x,r)

holds for every x ∈ Rn and every r > 0 and kM f k∞ ≤ kf k∞ follows. To deal with other
p’s, we need some more notation first.
Let f ∈ L1 (Rn ) and let α > 0. Then
Z
n
α · |{x ∈ R : |f (x)| > α}| = αdy
{x∈Rn :|f (x)|>α}
Z Z
≤ |f (y)|dy ≤ |f (y)|dy = kf k1 .
{x∈Rn :|f (x)|>α} Rn

The set of measurable functions f on Rn with

kf k1,w := sup α · |{x ∈ Rn : |f (x)| > α}| < ∞

α>0

is denoted by L1,w (Rn ) and called weak L1 . We have just shown that L1 (Rn ) ֒→ L1,w (Rn ).
Let us mention that k · k1,w is not a norm, but it still satisfies

kf + gk1,w = sup α · |{x ∈ Rn : |f (x) + g(x)| > α}|

α>0
 
n n
≤ sup α |{x ∈ R : |f (x)| > α/2}| + |{x ∈ R : |g(x)| > α/2}|
α>0
 
≤ 2 sup α/2 · |{x ∈ Rn : |f (x)| > α/2}| + |{x ∈ Rn : |g(x)| > α/2}|
α>0
≤ 2(kf k1,w + kgk1,w ).

1
Finally, we observe, that the function x → ∈ L1,w (Rn ) \ L1 (Rn ).
kxkn
The main aim of this section is to prove the following theorem.

Theorem 15.2.1. Let f be a measurable function on Rn . Then

i) If f ∈ Lp (Rn ), 1 ≤ p ≤ ∞, then the function M f is finite almost everywhere.

55
Functional Analysis II

ii) If f ∈ L1 (Rn ), then M f ∈ L1,w (Rn ) and

kM f k1,w ≤ Akf k1 ,

where A is a constant which depends only on the dimension (i.e. A = 5n will do).

iii) If f ∈ Lp (Rn ) with 1 < p ≤ ∞, then M f ∈ Lp (Rn ) and

kM f kp ≤ Ap kf kp ,

where Ap depends only on p and dimension n.

The proof is based on the following covering lemma.
Lemma 15.2.2. Let E be a measurable subset of Rn , which is covered by the union of a
family of balls (B j ) with uniformly bounded diameter. Then from this family we can select
a disjoint subsequence, B1 , B2 , B3 , . . . , such that
X
|Bk | ≥ C|E|.
k

Here C is a positive constant that depends only on the dimension n; C = 5−n will do.
Proof. We describe first the choice of B1 , B2 , . . . . We choose B1 so that it is essentially
as large as possible, i.e.
1
diam(B1 ) ≥ sup diam(B j ).
2 j
The choice of B1 is not unique, but that shall not hurt us.
If B1 , B2 , . . . , Bk were already chosen, we take again Bk+1 disjoint with B1 , . . . , Bk and
again nearly as large as possible, i.e.
1
diam(Bk+1 ) > sup{diam(B j ) : B j disjoint with B1 , . . . , Bk }.
2
In this way, we get a sequence B1 , B2 , . . . , Bk , . . . of balls. It can be also finite, if there
were no j
P balls B disjoint with B1 , B2 , . . . , Bk . P
If k |Bk | = ∞, then the conclusion of lemma is satisfied and we are done. If k |Bk | <
∞, we argue as follows.
We denote by Bk∗ the ball with the same center as Bk and diameter five times as large.
We claim that [
Bk∗ ⊃ E,
k
P P
which then immediatelySgives that |E| ≤ k |Bk∗ | = 5n k |Bk |.
We shall show that k Bk∗ ⊃ B j for every j. This is clear if B j is one P of the balls in
the preselected sequence. If it is not the case, we obtain diam(Bk ) → 0 (as k |Bk | < ∞)
and we choose the first k with diam(Bk+1 ) < 12 diam(B j ). That means, that B j must
intersect one of the balls B1 , . . . , Bk , say Bk0 . Obvious geometric arguments (based on
the inequality diam(Bk0 ) ≥ 1/2 · diam(B j )) then give that B j ⊂ Bk∗0 . This finishes the
proof.

Proof. (of Theorem 15.2.1). Let α > 0 and let us consider Eα := {x ∈ Rn : M f (x) > α}.
From the definition of M we obtain, that for every x ∈ Eα , there is a ball Bx centered at
x such that Z
|f (y)|dy > α|Bx |.
Bx

56
 15 Introduction to harmonic analysis

Hence, we get |Bx | < kf k1 /α and ∪x∈Eα Bx ⊃ Eα and the balls (Bx )x∈Eα satisfy the
assumptions of Lemma 15.2.2. Using this covering lemma, we get a sequence of disjoint
balls (Bk )k , such that
X
|Bk | ≥ C|Eα |.
k

We therefore obtain
Z XZ X
kf k1 ≥ |f (y)|dy = |f | > α |Bk | ≥ αC|Eα |,
∪k Bk k Bk k

which may be rewritten as supα>0 α · |Eα | = kM f k1,w ≤ C1 kf k1 .

This proves the first assertion of the theorem for p = 1 and the second assertion.
We now consider 1 < p < ∞. The proof follows from the information on the endpoints,
i.e. from
kM f k1,w ≤ Ckf k1 and kM f k∞ ≤ kf k∞ .

Actually, the Theorem of Marcinkiewicz says, that every operator with these two proper-
ties is then automaticaly bounded on all Lp (Rn ), 1 < p < ∞. Furthermore, this is a corner
stone of the so-called interpolation theory.
We shall restrict ourself to the necessary minimum needed to prove the theorem. This
is based on the technique of splitting the function into good and bad part, idea elaborated
in detail by Calderón and Zygmund.
Let α > 0 and put f1 (x) := f (x) if |f (x)| > α/2 and f1 (x) := 0 otherwise. Due to
|f (x)| ≤ |f1 (x)| + α/2, we have also M f (x) ≤ M f1 (x) + α/2 and also

{x ∈ Rn : M f (x) > α} ⊂ {x ∈ Rn : M f1 (x) > α/2}.

Due to the second part of the theorem

2A
|Eα | = |{x ∈ Rn : M f (x) > α}| ≤ |{x ∈ Rn : M f1 (x) > α/2}| ≤ kf1 k1
Z α
2A
≤ |f (y)|dy.
α {x∈Rn :|f (x)|>α/2}

We use the information of the size of the level sets of M f to estimate the Lp -norm of M f .
Z Z ∞
kM f kpp = M f (x)p dx = |{x ∈ Rn : M f (x)p > α}|dα
Rn 0
Z ∞
= |{x ∈ R : M f (x) > α1/p }|dα
n
0
Z ∞
=p β p−1 |{x ∈ Rn : M f (x) > β}|dβ
0
Z Z !
∞
2A
≤p β p−1 |f (y)|dy dβ
0 β {x∈Rn :|f (x)|>β/2}
Z Z 2|f (y)| Z
p−2 2Ap
= 2Ap |f (y)| β dβdy = |f (y)| · |2f (y)|p−1 dy
R n 0 p − 1 Rn
Z
2p Ap
= |f (y)|p dy,
p − 1 Rn

57
Functional Analysis II

where we used Fubini’s theorem and the substitution β := α1/p with dα = pβ p−1 dβ. This
gives the first and the third statement of the theorem with
 1/p
5n p
Ap = 2 , 1 < p < ∞.
p−1

Corollary 15.2.3. (Lebesgue’s differentiation theorem)

Let f be locally integrable on Rn . Then
Z
1
lim f (y)dy = f (x)
r→0 |B(x, r)| B(x,r)

holds for almost every x ∈ Rn .

Proof. We may cover the set of “bad” points

( Z Z )
n 1 1
x ∈ R : lim f (y)dy does not exist or lim f (y)dy 6= f (x)
r→0 |B(x, r)| B(x,r) r→0 |B(x, r)| B(x,r)

by
∞
( Z )
[ 1 1
n
x ∈ R : lim sup f (y)dy − f (x) >
r→0 |B(x, r)| B(x,r) k
k=1

united with
∞
( Z )
[ 1 1
n
x ∈ R : lim inf f (y)dy − f (x) < − .
r→0 |B(x, r)| B(x,r) k
k=1

It is therefore enough to show, that each of these sets has measure zero. Let us fix k ∈ N
and decompose f = g + h, where g ∈ C(Rn ) and khk1 ≤ t, t > 0.
Obviously, Z
1
lim g(y)dy = g(x), x ∈ Rn ,
r→0 |B(x, r)| B(x,r)

which implies
( Z )
n 1 1
x ∈ R : lim sup f (y)dy − f (x) >
r→0 |B(x, r)|B(x,r) k
( Z )
1 1
= x ∈ Rn : lim sup h(y)dy − h(x) >
r→0 |B(x, r)| B(x,r) k
   
1 1
⊂ x ∈ Rn : M h(x) > ∪ x ∈ Rn : |h(x)| > .
2k 2k

The measure of the first set is smaller than 2kAkhk1 and the measure of the second is
smaller than 2kkhk1 . As khk1 might be chosen arbitrary small, the measure of the original
set is zero. The same argument works for lim inf instead of lim sup as well.

58
 15 Introduction to harmonic analysis

Theorem 15.2.4. Let ϕ be a function which is non-negative, radial, decreasing (as func-
tion on (0, ∞)) and integrable. Then

sup |ϕt ∗ f (x)| ≤ kϕk1 M f (x),

t>0

where again ϕt (x) = t−n ϕ(x/t), x ∈ Rn .

Proof. Let in addition ϕ be a simple function. Then it can be written as
k
X
ϕ(x) = aj χBrj (x),
j=1

with aj > 0 and rj > 0. Then

k
X 1
ϕ ∗ f (x) = aj |Brj | χB ∗ f (x) ≤ kϕk1 M f (x),
|Brj | rj
j=1
P
since kϕk1 = j aj |Brj |. As any normalized dilation of ϕ satisfies the same assumptions
and has the same integral, it satisfies also the same inequality. Finally, any function
satisfying the hypotheses of the Theorem can be approximated monotonically from below
by a sequence of simple radial functions. This finishes the proof.

15.3 Calderón-Zygmund decomposition

The aim of this is to present the decomposition method of Calderón and Zygmund. In its
simple form, it was already used in the proof of boudnedness of the maximal operator M .
Theorem 15.3.1. Given a function f , which is integrable and non-negative, and given a
positive number λ, there exists a sequence14 (Qj ) of disjoint dyadic cubes such that
[
i) f (x) ≤ λ for almost every x 6∈ Qj ;
j
[  1
 
ii)  Qj  ≤ kf k1 ;
λ
j
Z
1
iii) λ < f ≤ 2n λ.
|Qj | Qj

Proof. We denote by Qk the collection of dyadic cubes with side length 2−k , k ∈ Z.
Furthermore, we define
X  1 Z 
Ek f (x) = f χQ (x).
|Q| Q
Q∈Qk

We define also

Ωk := {x ∈ Rn : Ek f (x) > λ and Ej f (x) ≤ λ if j < k}

That is, x ∈ Ωk if k is the first index with Ek f (x) > λ. As the integrability of f implies
Ek f (x) → 0 for k → −∞, such k always exists. The sets Ωk are clearly disjoint and each
can be written as the union of cubes in Qk . Together, these cubes form the system (Qj ).
14
possibly finite, or even empty

59
Functional Analysis II

This gives the third statement of the theorem. The first follows by Lebesgue differen-
tiation theorem: indeed, Ek f (x) ≤ λ for all k ∈ Z implies f (x) ≤ λ at almost every such
point. The second follows just by

[  X Z
  1X 1
 Qj  = |Qj | ≤ f ≤ kf k1 .
λ Qj λ
j j j

15.4 Interpolation theorems

We shall present two basic interpolation theorems, the Riesz-Thorin interpolation theorem
and Marcinkiewicz interpolation theorem.
The operator T mapping measurable functions to measurable functions is called sub-
linear, if

|T (f0 + f1 )(x)| ≤ |T f0 (x)| + |T f1 (x)|,

|T (λf )(x)| = |λ| · |T f (x)|, λ ∈ C.

Let (X, µ) and (Y, ν) be measure spaces and let T be a sub-linear operator mapping
Lp (X, µ) into a space of measurable functions on (Y, ν). We say that T is strong type
(p, q) if it is bounded from Lp (X, µ) into Lq (Y, ν). We say, that it is of weak type (p, q),
q < ∞, if

kT f kq,w := sup λ · ν 1/q ({y ∈ Y : |T f (y)| > λ}) ≤ Ckf kp , f ∈ Lp (X, µ).
λ>0

Theorem 15.4.1. (Marcinkiewicz interpolation theorem)

Let (X, µ) and (Y, ν) be measure spaces, 1 ≤ p0 < p1 ≤ ∞, and let T be a sublinear
operator from Lp0 (X, µ) + Lp1 (X, µ) to the measurable functions on Y that is weak (p0 , p0 )
type and weak (p1 , p1 ) type. Then T is strong (p, p) for p0 < p < p1 .

Proof. Let λ > 0 be given and let f ∈ Lp (X, µ). Then we decompose f into f = f0 + f1
with

f0 = f χ{x:|f (x)|>λ} ,
f1 = f χ{x:|f (x)|≤λ} .

The case p1 = ∞ appeared implicitly already in the proof of the boundedness of Hardy-
Littlewood maximal operator, so we suppose that p1 < ∞. Then we have

ν({y ∈ Y : |T f (y)| > λ}) ≤ ν({y ∈ Y : |T f0 (y)| > λ/2}) + ν({y ∈ Y : |T f1 (y)| > λ/2})

and
 p i
2Ai
ν({y ∈ Y : |T fi (y)| > λ/2}) ≤ kfi kpi , i = 0, 1.
λ

60
 15 Introduction to harmonic analysis

We combine them to get

Z ∞
kT f kpp = p λp−1 ν({x : |T f (x)| > λ})dλ
Z0 ∞ Z
p−1−p0 p0
≤p λ (2A0 ) |f (x)|p0 dµdλ
0 x:|f (x)|>λ
Z ∞ Z
+p λp−1−p1 (2A1 )p1 |f (x)|p1 dµdλ
0 x:|f (x)|≤λ
Z Z |f (x)|
= p(2A0 )p0 |f (x)|p0 λp−1−p0 dλdµ
X 0
Z Z ∞
p1 p1
+ p(2A1 ) |f (x)| λp−1−p1 dλdµ
X |f (x)|
= Ckf kpp .

Following theorem belongs also to the classical heart of interpolation theory. As we

shall not need it in the sequel, we state it without proof.

Theorem (Riesz-Thorin Interpolation)

Let 1 ≤ p0 , p1 , q0 , q1 < ∞, and for 0 < θ < 1 define p and q by
1 1−θ θ 1 1−θ θ
= + and = + .
p p0 p1 q q0 q1
If T is a linear operator from Lp0 + Lp1 to Lq0 + Lq1 , such that

kT f kq0 ≤ M0 kf kp0 for f ∈ Lp0

and
kT f kq1 ≤ M1 kf kp1 for f ∈ Lp1 ,
then
kT f kq ≤ M01−θ M1θ kf kp for f ∈ Lp .

15.5 Hilbert Transform

The Hilbert transform of a measurable function f on R is defined as
Z
1 f (x − y)
Hf (x) = dy.
π R y
As the integral does not converge absolutely, it has to be interpreted in an appropriate
limiting sense, which uses its cancelation property, i.e.
Z
1 f (x − y)
Hf (x) = lim dy.
ε→0+ π y:|y|>ε y

With this definition, Hf (x) makes sense for all smooth functions, especially for f ∈ S (R).
15

15
Observe, that the restriction to n = 1 is both natural and essential for the cancelation property.
Furthermore, from now on, we shall denote the Fourier transform of a function f also by the more usual
fb.

61
Functional Analysis II

If we define the distribution

Z
1 ϕ(x)
p.v. (ϕ) := lim dx, ϕ ∈ S (R),
x ε→0+ x:|x|>ε x

1 1
then Hf := p.v. ∗ f. This formula suggests that we look for the Fourier transform of
π x
1
Hf. Therefore, we regularize 1/x. This can be done using complex distributions π(x±iε)
and letting ε → 0 or (and that is what we shall do) by defining
1 x
Qt (x) := · .
π t 2 + x2
1
Obviously, limt→0 Qt (x) = πx holds pointwise and, as we shall show below, also in S ′ (R).
As
Z ∞
1
F −1 (sgn(x)e−a|x| )(ξ) = √ sgn(x)e−a|x| · eixξ dx
2π −∞
 Z 0 Z ∞ 
1 x(a+iξ) x(−a+iξ)
=√ − e dx + e dx
2π −∞ 0
  r
1 −1 1 2 iξ
=√ + = · 2 ,
2π a + iξ a − iξ π a + ξ2
we obtain
b t (ξ) = √−i sgn(ξ)e−t|ξ| ,
Q
2π
√
we have also limt→0 Q bt (ξ) = −i sgn(ξ)/ 2π. As this convergence is uniform on compact
sets, it holds also in S ′ (R). Finally, due to the continuity of Fourier transform, we obtain
 
1 1 ∧ bt = −i√sgn(·) .
p.v. = [lim Qt ]∧ = lim Q
π x t→0 t→0 2π
Theorem 15.5.1. In S ′ (R),
1 1
lim Qt (x) = p.v. .
t→0 π x
Proof. For each ε > 0, the functions ψε (x) = x−1 χ{y:|y|>ε} (x) are bounded and define
tempered distributions with limε→0+ ψε = p.v. x1 . Therefore, it is enough to show that

lim (πQt − ψt ) = 0
t→0

in S ′ (R). This follows by

Z Z
xϕ(x) ϕ(x)
(πQt − ψt )(ϕ) = dx − dx
R t 2 + x2 x:|x|>t x
Z Z  
xϕ(x) x 1
= 2 2
dx + 2 2
− ϕ(x)dx
x:|x|<t t + x x:|x|>t t + x x
Z Z
xϕ(xt) ϕ(tx)
= 2
dx − dx
x:|x|<1 1 + x x:|x|>1 x(1 + x2 )

for ϕ ∈ S (R). As t → 0, we apply Lebesgue dominated convergence theorem and use the
symmetry of the integrands to conclude, that the limit is zero.

62
 15 Introduction to harmonic analysis

S ′ (R) 1 1
Qt −−−−→ π p.v. x
t→0
 
 
Fy Fy

S ′ (R) −i√
sgn(·)
√−i sgn(ξ)e−t|ξ| −−−−→ ,
2π t→0 2π

Summarizing, one defines the Hilbert transform Hf for f ∈ S (Rn ) by any of these
formulas:
1 1
Hf = p.v. ∗ f,
π x
Hf = lim Qt ∗ f,
t→0
(Hf ) (ξ) = −i sgn(ξ)fˆ(ξ).
∧

Using the third expression, we can extend the definition of H to L2 (R) and it holds

kHf k2 = k(Hf )∧ k2 = k − i sgn(·)fˆk2 = kfˆk2 = kf k2 ,

H(Hf ) = (−i sgn(·)(Hf )∧ )∨ = ((−i sgn(·))2 fˆ)∨ = (−fˆ)∨ = −f,
hHf, Hgi = hf, gi, by polarization,
Z Z
Hf · g = − f · Hg,

where the last identity follows from

hHf, gi = h(Hf )∧ , ĝi = h−i sgn(·)fˆ, ĝi = hfˆ, sgn(·)ĝi = hfˆ, −(Hg)∧ i = −hf, Hgi

and the simple fact that Hg = Hg.

Theorem 15.5.2. For f ∈ S (R), the following is true.

(i) (Kolmogorov) H is weak type (1, 1),

C
kHf k1,w ≤ Ckf k1 , i.e. |{x ∈ R : |Hf (x)| > λ}| ≤ kf k1 , λ > 0, f ∈ L1 (Rn ).
λ

(ii) (M. Riesz) H is of strong type (p, p) for every 1 < p < ∞, i.e.

kHf kp ≤ Cp kf kp .

Proof. Step 1.: We show the weak type (1, 1) by exploiting Theorem 15.3.1. Let f be
non-negative and let λ > 0, then Theorem 15.3.1 gives a sequence of disjoint intervals
(Ij ), such that
[
f (x) ≤ λ for a.e. x 6∈ Ω = Ij ,
j
1
|Ω| ≤ kf k1 ,
Zλ
1
λ< f ≤ 2λ.
|Ij | Ij

63
Functional Analysis II

Given this decomposition of R, we decompose f into “good” and “bad” part defined by
 !
f (x),Z x 6∈ Ω, X X Z
1
g(x) = 1 b(x) = bj (x) = f (x) − f χIj (x).
 f, x ∈ Ij , |Ij | Ij
|Ij | Ij j j

Then g(x) ≤ 2λ almost everywhere, and bj is supported on Ij and has zero integral. Since
Hf = Hg + Hb, we have

|{x ∈ R : |Hf (x)| > λ}| ≤ |{x ∈ R : |Hg(x)| > λ/2}| + |{x ∈ R : |Hb(x)| > λ/2}|.

We estimate the first term using the L2 -boundedness of H by

Z Z Z Z
4 2 4 2 8 8
|{x ∈ R : |Hg(x)| > λ/2}| ≤ 2 |Hg(x)| dx = 2 g(x) dx ≤ g(x)dx = f (x)dx.
λ R λ R λ R λ R
S
Let 2Ij be the interval with the same center as Ij and twice the length. Let Ω∗ = j 2Ij .
Then |Ω∗ | ≤ 2|Ω| and

|{x ∈ R : |Hb(x)| > λ/2}| ≤ |Ω∗ | + |{x 6∈ Ω∗ : |Hb(x)| > λ/2}|

Z
2 2
≤ kf k1 + |Hb(x)|dx.
λ λ R\Ω∗
P
As |Hb(x)| ≤ j |Hbj (x)|, it is enough to show that

XZ
|Hbj (x)|dx ≤ Ckf k1 .
j R\2Ij

Denote the center of Ij by cj and use that bj has zero integral to get

Z Z Z 
 b (y) 
 j 
|Hbj (x)|dx =  dy  dx
R\2Ij R\2Ij  Ij x − y 
Z    
Z 1 1
 
=  bj (y) − dy  dx
R\2Ij  Ij x − y x − cj 
Z Z !
|y − cj |
≤ |bj (y)| dx dy
Ij R\2Ij |x − y| · |x − cj |
Z Z !
|Ij |
≤ |bj (y)| 2
dx dy
Ij R\2Ij |x − cj |

The last inequality follows from |y −cj | < |Ij |/2 and |x−y| > |x−cj |/2. The inner integral
equals 2, so
XZ XZ
|Hbj (x)|dx ≤ 2 |bj (y)|dy ≤ 4kf k1 .
j R\2Ij j Ij

Our proof of the weak (1, 1) inequality is for non-negative f , but this is sufficient since
an arbitrary real function can be decomposed into its positive and negative parts, and a
complex function into its real and imaginary parts.

64
 15 Introduction to harmonic analysis

Step 2.: Since H is weak type (1, 1) and strong type (2, 2), it is also strong type (p, p)
for 1 < p < 2. If p > 2, we apply duality, i.e.
Z  
 
 
kHf kp = sup  Hf · g : kgkp′ ≤ 1
ZR  
 
= sup  f · Hg : kgkp′ ≤ 1
R

≤ kf kp · sup kHgkp′ : kgkp′ ≤ 1 ≤ Cp′ kf kp .

The strong (p, p) inequality is false if p = 1 or p = ∞; this can easily be seen if we let
f = χ[0,1] . Then  
1  x 
Hf (x) = log  ,
π x − 1
and Hf is neither integrable nor bounded.

15.6 BMO
Hilbert transform acts rather badly on L∞ (R). Not only is H unbounded on L∞ (R), it
can not be easily defined on a dense subset of L∞ (R). The definition
Z
1 f (x)
Hf (y) = dx
π R y−x
runs for f ∈ L∞ (R) into troubles for x near y and x near infinity. If we look onto
differences, the situation changes to
Z  
′ 1 1 1
Hf (y) − Hf (y ) = f (x) − dx.
π R y − x y′ − x
This improves the situation for x near infinity, as 1/(x − y) − 1/(x − y ′ ) = O(1/x2 ) in this
case.
We say that f and g are equivalent modulo a constant if f (x) = g(x) + C for some
(complex) constant C and almost every x ∈ R. Given f ∈ L∞ (R) and y ∈ R. Then we
take an open interval B ⊂ R with center at zero containing y. Then f χB ∈ L2 (R) and we
define Hf (y) to be
Z  
1 1 1
Hf (y) := H(f χB )(y) + f (x) + dx.
π R\B y−x x
The first term is defined by the L2 definition of H, the integral in the second term converges
absolutely. The definition depends on B, but choosing another interval B ′ ⊃ B with the
center at the origin leads to a difference
Z   Z  
1 1 1 1 1 1
H(f χB )(y) + f (x) + dx − H(f χB ′ )(y) − f (x) + dx
π R\B y−x x π R\B ′ y−x x
Z  
1 1 1
= H(f χB − f χB ′ )(y) + f (x) + dx
π B ′ \B y−x x
Z   Z
1 1 1 1 f (x)
= −H(f χB ′ \B )(y) + f (x) + dx = dx,
π B ′ \B y−x x π B ′ \B x
which does not depend on y. This defines Hf modulo constant for f ∈ L∞ (R). Of course,
such a definition does not allow to measure Hf in the usual norms, as Lp . Instead of that,
we need a space of functions defined modulo constants.

65
Functional Analysis II

Definition 15.6.1. (Bounded mean oscillation). Let f : Rn → C be a function defined

modulo a constant. The BM O (or Bounded Mean Oscillation) norm of f is defined
Z  Z 
1  1 
n
kf |BM O(R )k := sup f − f 
 |B| B 
B |B| B

where B ranges over all balls. Note that if one shifts f by a constant, the BM O norm
is unchanged, so this norm is well-defined for functions defined modulo constants. We
denote by BM O(Rn ) the space of all functions with finite BM O norm.

Example 15.6.2. Let f (x) = sgn(x). Let |y| < a/2. We take B = (−a, a) and apply the
definition of Hf as presented above. This gives (for y > 0)
Z a Z  
sgn(x) 1 1
πHf (y) = p.v. dx + sgn(x) + dx
−a y − x (−∞,−a)∪(a,∞) y−x x
! Z
Z −a   Z ∞ 
sgn(x) 1 1 1 1
= lim dx − − dx + − dx
ε→0+ (−a,a)\(y−ε,y+ε) y − x −∞ x x−y a x x−y
| {z } | {z
x=∞ }
a
− ln a+y x  a
=ln x−y =− ln a−y
x=a
Z 0 Z y−ε Z a 
1 1 1 a a
= lim dx − dx − − ln − ln
−a x − y x−y y+ε x − y a+y a−y
ε→0+
0
 
y y a−y a a
= lim ln + ln − ln − ln − ln
ε→0+ a+y ε ε a+y a−y
= 2 ln y − 2 ln a.

If y < 0, similar calculation applies as well. Hence, πHf (y) = 2 ln |y| − 2 ln a. Hence,
ignoring the constant,
2
H(sgn x) = ln |x|.
π
Let us observe, that
Z  Z  Z
1  
f − 1 
f ≈ inf
1
|f − c|
|B|  |B| B  c∈R |B| B
B

holds for every ball B with universal constants. Indeed, the left-hand side is obviously
larger than the right hand side. On the other hand, we get as well
Z  Z  Z Z  Z 
1  1  1 1  1 
f − f  ≤ |f − c| + c − f 
|B| B  |B| B |B| B |B| B  |B| B
Z Z  Z 
1 1  1 
= |f − c| +  (c − f )
|B| B 
|B| B |B| B
Z Z  Z 
1 1 1
≤ |f − c| + |c − f |
|B| B |B| B |B| B
Z
2
= |f − c| .
|B| B

Theorem 15.6.3. (H maps L∞ (R) into BM O(R))

Let f ∈ L∞ (R). Then
kHf |BM O(R)k . kf k∞ .

66
 15 Introduction to harmonic analysis

Proof. Due to the observation above, it is enough to show that for every ball B, there is
a constant cB such that Z
1
|Hf − cB | . kf k∞ .
|B| B
We split Hf = H(f χ2B ) + H(f χR\2B ), where 2B is a ball with the same center as B, but
twice the radius. First we get
Z Z 1/2 Z 1/2 Z 1/2
1 1 1 kf k∞ p
|H(f χ2B )| ≤ |H(f χ2B )|2 · 1 ≤p |f |2 ≤p |2B| . kf k∞ .
|B| B |B| B B |B| 2B |B|
This deals with the “local” part of Hf . For the “global” part, observe that for x ∈ B we
have (modulo constant)
Z  
1 1
H(f χR\2B )(x) = f (y) − dy,
y:y6∈2B x−y γ −y

where γ is the center of B.

Z Z   
1  1 1 
 f (y) − dy  dx
|B| B  R\2B x−y γ−y 
Z Z   
1  1 1 
≤ f (y) −  dydx
|B| B R\2B  x−y γ−y 
Z Z  
kf k∞  1 1 
≤  − dydx
|B| B R\2B  x − y γ − y 

By shifting, we may assume, that γ = 0, B = (−a, a) and 2B = (−2a, 2a). We estimate

Z aZ ∞ Z a Z ∞ Z
1 x 1 1 a ∞ 1
dydx ≤ xdx dy ≤ dy ≤ c
2a 0 2a |y(x − y)| 2a 0 2a |y(y − a)| 4 2a (y − a)2
(and similarly for the remaining parts). Altogether, this gives
Z
1
|H(f χR\2B )| . kf k∞ .
|B| B
Adding the two facts, we obtain the claim.

15.7 Singular integrals

Hilbert transform is the most important example of the so-called singular integrals. These
are convolutions with kernel K, which might have a singularity at origin.
Theorem 15.7.1. (Calderón-Zygmund) Let K ∈ S ′ (Rn ) be a tempered distribution,
which is associated to a locally integrable function on Rn \ {0} and satisfies
(CZ1) |K̂(ξ)| ≤ A, ξ ∈ Rn ,
Z
(CZ2) |K(x − y) − K(x)|dx ≤ B, y ∈ Rn .
kxk2 ≥2kyk2

Then, for 1 < p < ∞,

kK ∗ f kp ≤ Cp kf kp and kK ∗ f k1,w ≤ Ckf k1 .

67
Functional Analysis II

The proof copies very much the proof of the same statement for the Hilbert transform,
and we leave out the details.
The condition (CZ2) is sometimes called Hörmander condition. By the help of mean
value theorem, it is satisfied for example if
C
k∇K(x)k2 ≤ , x 6= 0.
kxkn+1
2
An important and non-trivial generalisation of the theory of singular integrals is given
by considering the vector-valued analogues. By this, we mean the following.
• H, H̃ are (complex) Hilbert spaces.
• For 0R< p < ∞, Lp (Rn → H) is the set of measurable functions f : Rn → H, such
that Rn kf (x)kpH dx < ∞.
Z
n
• Let K : R → L (H, H̃). Then T f (x) = K(y)f (x − y)dy takes values in H̃.
Rn

• Under same (just appropriately interpreted) conditions as above, T is bounded from

Lp (Rn → H) into Lp (Rn → H̃). Especially, the gradient condition above is still valid
in this case.

15.8 Khintchine inequality

We denote by
rn (t) := sign sin(2n πt), t ∈ [0, 1], n ∈ N0 .
the Rademacher functions.
The system (rn )∞ n=0 forms an orthonormal system in L2 (0, 1), but it is not a basis
(consider i.e. the function f (t) = 1 − 2t).
Theorem 15.8.1. Let p ∈ [1, ∞). Then there are positive constants Ap and Bp such that
Xm
!1/2 Z 1 Xm
p !1/p
 Xm
!1/2
2   2
Ap |an | ≤  an rn (t) dt ≤ Bp |an |
n=1 0  
n=1 n=1
holds for every m ∈ N and every sequence of real numbers a1 , . . . , am .
Proof. By Ap and Bp we denote the best possible constants (which are actually known,
but we shall derive slightly weaker estimates). Furthermore, orthogonality of Rademacher
functions gives immediately A2 = B2 = 1. Finally, due to monotonicity of the Lp -norms,
we have Ar ≤ Ap and Br ≤ Bp for r ≤ p.
So, it is enough to show that A1 > 0 and B2k < ∞ for all k ∈ N.
We start with B2k . Let us observe that
Z 1 Xm
2k
 Z 1 X m
!2k
 
E :=  an rn (t) dt = an rn (t) dt
0 n=1  0 n=1
X Z 1
(2k)! α1
= αm
a . . . am r1α1 (t) . . . rm
αm
(t)dt
α1 ! . . . αm ! 1 0
|α|=2k
X Z 1
(2k)! 2α1
= a . . . am2αm
r12α1 (t) . . . rm
2αm
(t)dt
(2α1 )! . . . (2αm )! 1 0
|α|=k
X (2k)!
= a2α1 . . . a2α
m ,
m
(2α1 )! . . . (2αm )! 1
|α|=k

68
 15 Introduction to harmonic analysis

where we have used the multinomial theorem (a generalisation of the binomial theorem to
a bigger number of summands) and the fact that
Z 1
r1α1 (t) . . . rm
αm
(t)dt
0

is equal to zero if some of the αi ’s is odd and equal to 1 if all of them are even.
Let α = (α1 , . . . , αm ) ∈ Nm
0 be integers with |α| = k, then

2k α1 ! . . . αm ! = (2α1 α1 !) . . . (2αm αm !) ≤ (2α1 )! . . . (2αm )!.

This implies
(2k)! X k!
E≤ a2α1 . . . a2α
m
m
k
2 k! α1 ! . . . αm ! 1
|α|=k
m
!k
(2k)! X 2 (2k)!
= k |an | = k kak2k 2
2 k! 2 k!
n=1

and  1/(2k)
1/(2k) (2k)!
E ≤ kak2 .
2k k!
Hence, the statement holds with 16

 1/(2k)
(2k)!
B2k := .
2k k!
Finally, we have to show the existence of A1 > 0. We proceed by a nice duality trick
using the (alreadyPmproven) first part of this theorem.
Let f (t) := n=1 an rn (t). By Hölder’s inequality for p = 3/2 and p′ = 3, we have
Z 1 Z 1 Z 1 2/3 Z 1 1/3
2 2/3 4/3 4
|f (t)| dt = |f (t)| · |f (t)| dt ≤ |f (t)|dt · |f (t)|
0 0 0 0
Z 1 2/3 Z 1 2/3
4/3 4/3 4/3 4/3
≤ |f (t)|dt B4 · kak2 = |f (t)|dt B4 · kf k2 .
0 0

Therefore,
Z 1 2/3 Z 1 1/3
−4/3 2
|f (t)|dt ≥ B4 |f (t)| dt ,
0 0
that is Z 1/2
Z 1 1
|f (t)|dt ≥ B4−2 2
|f (t)| dt = B4−2 kak2 .
0 0

Hence, A1 ≥ B4−2 .

Remark 15.8.2. Stochastic reformulation of Khintchine’s inequalities sounds as follows.

Let εi , i = 1, . . . , m be independent variables with P(εi = 1) = 1/2 and P(εi = −1) = 1/2.
Let 1 ≤ p < ∞. Then there are constants Ap , Bp such that for every a1 , . . . , am ∈ R 17
 p !1/p
Xm 
 
Ap kak2 ≤ E  ai εi  ≤ Bp kak2 .
 
i=1
16
√
By Stirling’s formula, one can show quite easily that B2k grows as 2k for k → ∞.
17
Also a1 , . . . , am ∈ C can be considered with slightly modified proof.

69
Functional Analysis II

Choosing p large enough, this estimate gives very quickly the so-called tail bound estimates
on sum of independent Rademacher variables, i.e. the assymptotic estimates of
  !
Xm 
 
P  ai εi  > t
 
i=1
for t → ∞.
We use this reformulation of Khintchine’s inequalities to give another proof of Theorem
15.8.1.
Proof. (of the upper estimate in Theorem 15.8.1).
We normalize to kak2 = 1. Then
m
! m m m
X Y Y Y
E exp ai εi = E exp(ai εi ) = E exp(ai εi ) = cosh(ai ).
i=1 i=1 i=1 i=1

Using Taylor’s expansion, one obtains cosh(aj ) ≤ Hence, exp(a2j /2).

m
! m
X Y
E exp ai εi ≤ exp(a2i /2) . 1,
i=1 i=1
and by Markov’s inequality
m
! m
! ! m
! !
X X X
P ai εi > λ = P exp ai εi > exp(λ) = P exp ai εi − λ > 1
i=1 i=1 i=1
m
!
X
≤ E exp ai εi − λ . e−λ .
i=1
m  !
X 
 
By symmetry of εi ’s, we get also P  ai εi  > λ . e−λ . The rest then follows by
 
i=1
distributional representation of the Lp -norm.
Khintchine’s inequalities have an interesting application in operator theory. Let 1 ≤
p < ∞ and let T : Lp (Rn ) → Lp (Rn ) be a bounded linear operator. Then

X N 1/2  X
N 1/2
2 2
|T fj | ≤ cp |fj | , (15.1)

j=0 p j=0 p
where the constant cp depends only on p and kT k.
The proof follows by considering Rademacher functions r1 , . . . , rN and

XN 1/2 p Z XN p/2 Z Z 1 XN p/p

|T fj |2 = |(T fj )(x)|2 dx ≤ c | T fj (x)rj (t)|p dt dx
Rn Rn 0
j=0 p j=1 j=1
Z 1Z N
X Z 1Z X
N 
=c | T fj (x)rj (t)|p dxdt = c |T fj rj (t) (x)|p dxdt
0 Rn j=1 0 Rn j=1
Z 1Z N
X Z X
N p/2
≤ ckT kp | fj (x)rj (t)|p dxdt ≤ c′ |fj (x)|2 dx
0 Rn j=1 Rn j=1

 X
N 1/2 p

= cpp |fj |2 .

j=0 p
By letting N → ∞, the same result holds also for infinite sums.

70
 15 Introduction to harmonic analysis

15.9 Littlewood-Paley Theory

Let {Ij } be a sequence of intervals on the real line, finite or infinite, and let {Sj } be the
sequence of operators defined by (Sj f )∧ (ξ) = χIj (ξ)fˆ(ξ). Later on, we shall concentrate
on the dyadic decomposition of R (strictly speaking of R \ {0}) given by

Ij = (−2j+1 , −2j ] ∪ [2j , 2j+1 ), j ∈ Z. (15.2)

Furthermore, we denote Sj∗ = Sj−1 + Sj + Sj+1 . Let us observe that this implies Sj∗ Sj =
Sj Sj∗ = Sj .
Finally, we addopt this concept also to smooth dyadic decompositions. Let ψ ∈ S (R)
be non-negative, have support in 1/2 ≤ kξk2 ≤ 4 and be equal to 1 on 1 ≤ kξk2 ≤ 2. Then
we define
ψj (ξ) = ψ(2−j ξ) and (S̃j f )∧ (ξ) = ψj (ξ)fˆ(ξ), ξ ∈ R.
Theorem 15.9.1. (Littlewood-Paley Theory) Let 1 < p < ∞.
i) Then there exist two constants Cp > cp > 0, such that
!1/2
X

cp kf kp ≤ |Sj f |2 ≤ Cp kf kp .

j p

The same holds for Sj∗ .

ii) There exists a constant Cp > 0 such that

!1/2
X

|S̃j f |2 ≤ Cp kf kp .

j p

X
iii) Finally, if |ψ(2−j ξ)|2 = 1, then there is also a constant cp > 0, such that
j

!1/2
X
2
cp kf kp ≤ |S̃j f | .

j p

Proof. Step 1.
We know, that (Sj f )∧ = χIj fˆ, where Ij was defined by (15.2). We define

Ij− := (−2j+1 , −2j ], Ij+ := [2j , 2j+1 ), j∈Z

and split Sj = Sj− + Sj+ , where (Sj− f )∧ = χI − fˆ and (Sj+ f )∧ = χI + fˆ.

j j
We observe, that
1

χI + (x) = sgn(x − 2j ) − sgn(x − 2j+1 ) for (almost) all x ∈ R,
j 2
and
1 
Sj+ f = (χI + fˆ)∨ = (sgn(· − 2j )fˆ)∨ − (sgn(· − 2j+1 )fˆ)∨ .
j 2
Finally, we write

sgn(ξ − 2j )fˆ(ξ) = τ2j [sgn(ξ)fˆ(ξ + 2j )] = τ2j [sgn(ξ) · τ−2j fˆ(ξ)],

71
Functional Analysis II

leading to

(sgn(ξ − 2j )fˆ(ξ))∨ = M2j (sgn ·τ−2j fˆ)∨ = (2π)−1/2 · M2j (sgn(·)∨ ∗ (τ−2j fˆ)∨ )
√   !
2π 1 1
= (2π)−1/2 M2j p.v. ∗ M−2j f
−i π x
  
1 1
= iM2j p.v. ∗ M−2j f
π x
= iM2j HM−2j f.

Using the boundedness of H on Lp (R) for 1 < p < ∞, we immediately obtain that
kSj f kp ≤ ckf kp , and the same is true also for Sj+ and Sj− .
Step 2.
We combine Step 1. with (15.1) to obtain
 
X 1/2 
1  X  1/2  X  1/2 

|Sj+ fj |2 ≤ |(sgn(ξ − 2j )fˆj (ξ))∨ |2 + |(sgn(ξ − 2j+1 )fˆj (ξ))∨ |2
2  
j∈Z p j∈Z p j∈Z p
 
1  
X 2
1/2


≤ |M2j HM−2j fj | + ...
2  
j∈Z p
 
1 
X 1/2


= |HM−2j fj |2 + ...
2  
j∈Z p
 
 X 1/2  X 1/2

≤ cp |M−2j fj |2 + . . . = c′p |fj |2 .
 
j∈Z p j∈Z p

The same holds of course for Sj− and, therefore, also for Sj .
Step 3.
This, together with the identity Sj = Sj S̃j implies

X 1/2 X 1/2 X 1/2

|Sj f |2 = |Sj S̃j f |2 ≤ |S̃j f |2 .

j∈Z p j∈Z p j∈Z p

Step 4.
This shows, that the second inequality in part (i) of the theorem follows from (ii). There-
fore, we prove (ii) now.
Let Ψ̂ = ψ and Ψj (x) = 2j Ψ(2j x). Then Ψ̂j = ψj and S̃j f = Ψj ∗ f. It is enough to
prove that the vector-valued mapping

f → (S̃j f )j

is bounded from Lp to Lp (ℓ2 ). If p = 2, this follows by Plancherel theorem:

!1/2 2 Z
X X

|S̃j f |2
= |ψj (ξ)|2 · |fˆ(ξ)|2 dξ ≤ 3kf k22 .
R
j 2 j

The proof for p 6= 2 is based on the Hörmander’s condition for vector-valued singular
integrals, i.e. that
kΨ′j (x)kℓ2 ≤ Ckxk−2
2 .

72
 15 Introduction to harmonic analysis

Using that Ψ ∈ S (R), we obtain

 1/2
X X X X C
 |Ψ′j (x)|2  ≤ |Ψ′j (x)| = 22j |Ψ′ (2j x)| ≤ C 22j max(1, 2j |x|)−3 ≤ .
|x|2
j j j j

Step 5.
Finally, we prove the first inequality in part (i) and part (iii) of the theorem. Surprisingly
enough, they follow quite quickly from previous steps and duality.
The identity
X 1/2

|Sj f |2 = kf k2 (15.3)

j∈Z 2

follows by Plancherel’s theorem and, by polarization, also

Z X Z
Sj f · Sj g = fg
R j R

follows. Using this, and the first part of the theorem for p′ with 1/p + 1/p′ = 1 allows the
following estimate.
Z  
 
 
kf kp = sup  f g : kgkp′ ≤ 1
 R  
Z X 
 
= sup  Sj f · Sj g  : kgkp′ ≤ 1
 R  
j
 
 X 1/2 X 1/2 

≤ sup |Sj f |2 · |Sj g|2 : kgkp′ ≤ 1
 
j∈Z p j∈Z p

X 1/2

≤ cp |Sj f |2 .

j∈Z p

Part (iii) of the theorem follows in exactly the same way - the assumption of the theorem
gives exactly the identity (15.3).

73
16 Selected topics
16.1 Riesz representation theorem
We prove the Riesz Representation Theorem, which was used to construct the Borel mea-
surable functional calculus. We follow essentially [14], where one can find a lot more to
measure and integration theory.

16.1.1 Outer measures

Definition 16.1.1. By an outer measure on a set X we understand a set function γ,
which assigns to every subset A ⊂ X a non-negative number γ(A) ∈ [0, ∞], such that the
following conditions are satisfied:
i) γ(∅) = 0,

ii) if A ⊂ B, then γ(A) ≤ γ(B),

P
iii) γ(∪An ) ≤ γ(An ) for all An ⊂ X.
We describe a construction (due to Carathéodory) which constructs a measure from
an outer measure.
Definition 16.1.2. A set M ⊂ X is said to be γ-measurable (in the sense of Carathéodory)
if
γ(T ) = γ(T ∩ M ) + γ(T \ M )
for each set T ⊂ X (in other words, M splits additively each set T ⊂ X). The collection
of all γ-measurable sets will be denoted by M(γ).
Let us note the inequality γ(T ) ≤ γ(T ∩ M ) + γ(T \ M ) always holds. To prove
measurability, it is therefore sufficient to show that γ(T ) ≥ γ(T ∩ M ) + γ(T \ M ) for all
T ⊂ X with γ(T ) < ∞.
Theorem 16.1.3. M(γ) is a σ-algebra and γ is a complete measure on M(γ).
Proof. The proof will be divided into several steps.
Step 1.: It is easy to check that X ∈ M(γ), that X \ M ∈ M(γ) provided M ∈ M(γ) and
that A ∈ M(γ) whenever γ(A) = 0.
Step 2.: Let M, N ∈ M(γ). We would like to show that M ∩ N ∈ M(γ). Let T ⊂ X be
arbitrary. We have
γ(T ) = γ(T ∩ M ) + γ(T \ M )
and
γ(T ∩ M ) = γ(T ∩ M ∩ N ) + γ((T ∩ M ) \ N )
together with

γ(T \ (M ∩ N )) = γ((T \ (M ∩ N )) ∩ M ) + γ((T \ (M ∩ N )) \ M )

= γ((T ∩ M ) \ N ) + γ(T \ M ).

We put these three inequalities together and obtain

γ(T ) = γ(T ∩ M ∩ N ) + γ((T ∩ M ) \ N ) + γ(T \ M )

= γ(T ∩ M ∩ N ) + γ(T \ (M ∩ N ))
= γ(T ∩ (M ∩ N )) + γ(T \ (M ∩ N )).

74
 16 Selected topics

Since M(γ) is closed under complements and under finite intersections, it is also closed
under finite unions.
Step 3.: To show that γ is σ-additive on M(γ), choose Mn ∈ M(γ) pairwise disjoint.
Setting T = M1 ∪ M2 and using γ-measurability of M1 , we obtain
γ(M1 ∪ M2 ) = γ(M1 ) + γ(M2 ).
Thus γ is finitely additive. Further
∞ k k
! ∞
!
X X [ [
γ(Mn ) = lim γ(Mn ) = lim γ Mn ≤γ Mn ,
k→∞ k→∞
n=1 n=1 n=1 n=1

and since the reverse inequality always holds, we reach the conclusion. S
Step 4.: Let now Mn ∈ M(γ) be pairwise disjoint. Our aim is to show that ∞ n=1 Mn ∈
M(γ). Choosing a test set T ⊂ X, we have
k
! k
! ∞
! k
[ [ [ X
γ(T ) = γ T \ Mn + γ T ∩ Mn ≥ γ T \ Mn + γ(T ∩ Mn )
n=1 n=1 n=1 n=1

for each k ∈ N. Since γ is σ-additive (Step 3), it follows

∞
! ∞
[ [
γ(T ) ≥ γ T \ Mn + γ(T ∩ Mn ),
n=1 n=1

which is what we wanted as the second inequality is trivial.

16.1.2 The Theorem

In this section, (K, τ ) is a compact topological vector space. Although Riesz Representa-
tion Theorem holds even in a more general setting, we shall restrict to this and discuss
the extensions separately later on.
Theorem 16.1.4. Urysohn’s lemma (for locally compact spaces).
If K is a compact set and U an open subset of a locally compact space X, K ⊂ U ⊂ X,
then there exists a continuous function f and a compact set L with
K ⊂ L ⊂ U, 0 ≤ f ≤ 1, f = 1 on K, f = 0 on X \ L.
Theorem 16.1.5. Dini’s property.
If (fn ) is a monotone sequence of continuous functions on a compact space X, which
converges pointwise to a continuous function, then the convergence of (fn ) is uniform on
X.
Theorem 16.1.6. (Daniell’s property) Let A be a positive linear functional on C(K). Let
fn ∈ C(K) with fn ց 0. Then Afn → 0.
Proof. From positivity of A (i.e. f ≥ 0 implies Af ≥ 0) and linearity of A, the monotonic-
ity of A follows (i.e. f ≤ g implies Af ≤ Ag). Therefore, the limit b := limn→∞ Afn ≥ 0
always exists. By Dini’s theorem, fn ⇉ 0 and, therefore,Pthere exists a sequence (nk ),
such that kfnkPk∞ ≤ k−2 for each k ∈ N. Thus the series k fnk converges uniformly on
K and if f := k fnk , then f ∈ C(K). Finally, we obtain for each k ∈ N
k k
!
X X
kb ≤ Afni = A fni ≤ Af
i=1 i=1

and b = 0 follows.

75
Functional Analysis II

Let us stress, that we did not suppose continuity of A in this theorem.

Theorem 16.1.7. (Riesz Representation Theorem)

Let A be a positive linear functional on C(K). Then there
R exists a unique complete Radon
measure µ on K, such that C(K) ⊂ L1 (µ) and Af = K f dµ for all f ∈ C(K).

Proof. The proof is based on the following idea. We first define

µ∗A (G) := sup{Af : f ∈ C(K) : 0 ≤ f ≤ 1, f = 0 on K \ G}, G is open,

µ∗A (E) := inf{µ∗A (G) : G open, G ⊃ E}, E ⊂ K arbitrary.

As µ∗A is monotone on open sets, the second line of this definition extends the first line
and the notation is correct. Such a set function is called outer Radon measure (associated
to A). We shall show, that this mapping is really an outer measure (in the sense described
in Definition 16.1.2), and we shall prove its regularity properties. The construction of
Carathéodory will then give a σ-algebra (and we shall show, that it contains all open sets)
on which µ∗A is a measure. Finally, we shall prove, that this measure has the property
from the statement of the theorem. Let us work out this program.
Step 1.: We shall show the following

i) µ∗A (C) = inf{Ag : g ∈ C(K), 0 ≤ g ≤ 1, g = 1 on C} for every compact set C ⊂ K.

ii) µ∗A (G) = sup{µ∗A (C) : C compact, C ⊂ G} for every open set G ⊂ K.

iii) µ∗A is an outer measure.

Step 1.(i): Let C ⊂ K be a compact set and let g ∈ C(K), 0 ≤ g ≤ 1, g = 1 on C. First we

show that µ∗A (G) ≤ Ag. Fix ε ∈ (0, 1) and denote G := {x ∈ K : g(x) > 1 − ε}. Obviously,
g Ag
C ⊂ G. If f ∈ C(K), 0 ≤ f ≤ 1 and f = 0 on K \ G, then f ≤ 1−ε and Af ≤ 1−ε holds
∗ Ag
for every f with such properties. Hence µA (G) ≤ 1−ε and

Ag ≥ (1 − ε)µ∗A (G) ≥ (1 − ε)µ∗A (C).

We see that Ag ≥ µ∗A (C), and thus

µ∗A (C) ≤ inf{Ag : g ∈ C(K), 0 ≤ g ≤ 1, g = 1 on C}.

To prove the reverse inequality, select an open set G ⊃ C. Urysohn’s lemma provides a
function g ∈ C(K), 0 ≤ g ≤ 1, g = 0 on K \ G and g = 1 on C. The definition of µ∗A (G)
for G open gives Ag ≤ µ∗A (G). Finally, we observe that

µ∗A (C) = inf{µ∗A (G) : G open, G ⊃ C}

≥ inf{Ag : g ∈ C(K), 0 ≤ g ≤ 1, g = 1 on C}.

Step 1.(ii): Suppose we are given an open set G ⊂ K and ε > 0. Let f ∈ C(K) be
a function such that 0 ≤ f ≤ 1, f = 0 on K \ G. Since fn := min(f, 1/n) ց 0, an
appeal to Daniell’s property gives the existence of an n ∈ N, such that Afn < ε. If
C := {x ∈ K : f (x) ≥ 1/n}, then C is a compact subset of K. Due to (i), there exists
g ∈ C(K) such that 0 ≤ g ≤ 1, g = 1 on C and Ag ≤ µ∗A (C) + ε. Since f − fn ≤ g, we get

Af ≤ Afn + Ag ≤ µ∗A (C) + 2ε.

76
 16 Selected topics

Hence,

µ∗A (G) = sup{Af : f ∈ C(K) : 0 ≤ f ≤ 1, f = 0 on K \ G}

≤ sup{µ∗A (C) : C compact, C ⊂ G} + 2ε,

and we finish the proof of one part of (ii). The second one follows directly by monotonicity
of µ∗A .
Step 1.(iii): Finally, we prove that µ∗A is an outer measure. Clearly, we have µ∗A (∅) = 0
and µ∗A (S) ≤ µ∗A (T ) whenever S ⊂ T. It remains to show that µ∗A is σ-subadditive.
To this end, let ε > 0 and C1 , C2 ⊂ K be two given compact sets. By (i), we find
fj ∈ C(K) such that fj = 1 on Cj and Afj ≤ µ∗A (Cj ) + ε; j = 1, 2. Then (by (i) again)

µ∗A (C1 ∪ C2 ) ≤ A(min(f1 + f2 , 1)) ≤ A(f1 + f2 ) = Af1 + Af2 ≤ µ∗A (C1 ) + µ∗A (C2 ) + 2ε

and we see that µ∗A is subadditive on compact sets.

Next, consider open sets G1 , G2 ⊂ K and pick a compact set C ⊂ G1 ∪ G2 . For any
point x ∈ C, there is a neighborhood Vx whose closure is either in G1 or in G2 . Thanks
to the compactness of C, we obtain finite collections of open sets (Vi1 )i , (Vi2 )i such that
S S S
Vij ⊂ Gj and i Vi1 ∪ i Vi2 ⊃ C. Set Cj := C ∩ i Vij ; j = 1, 2. Then Cj are compact,
Cj ⊂ Gj and C = C1 ∪ C2 . Hence, µ∗A (C) ≤ µ∗A (C1 ) + µ∗A (C2 ) ≤ µ∗A (G1 ) + µ∗A (G2 ) and it
readily follows that µ∗ is subadditive on open sets. S∞
Sn let (Gn ) be a sequence of open sets. Chose a compact set C ⊂ i=1 Gi . Then
Now,
C ⊂ i=1 Gi for some n ∈ N, and
n
! n ∞
[ X X
∗ ∗
µA (C) ≤ µA Gi ≤ µ∗A (Gi ) ≤ µ∗A (Gi )
i=1 i=1 i=1
S P∞ ∗
and from (ii), we conclude that µ∗A ( ∞ i=1 Gi ) ≤ i=1 µA (Gi ).
Finally,
S let (E
P ∗n ) be a system of arbitrary subsets of K. We wish to show that
∗
µA ( En ) ≤ µA (En ). It is clearly sufficient to assume that µ∗A (En ) < ∞ for all n.
Given ε > 0, we find open sets Gn such that En ⊂ Gn and µ∗A (Gn ) < µ∗A (En ) + 2−n ε.
Then [  [  X
µ∗A En ≤ µ∗A Gn ≤ µ∗A (En ) + ε.

As ε > 0 was arbitrary, µ∗A is subadditive as needed.

Step 2.: We show, that every Borel subset of K is µ∗A measurable (in the sense of
Carathéodory). It is sufficient to prove measurability of open subsets. Notice, that we
have µ∗A (G1 ∪ G2 ) = µ∗A (G1 )+ µ∗A (G2 ) for open disjoint sets G1 and G2 . This follows easily
from definition of µ∗A (G) for G open. Indeed, there are fj ∈ C(K), j = 1, 2 with 0 ≤ fj ≤ 1,
fj = 0 on K\Gj and µ∗A (Gj ) ≤ Afj +ε. Then f1 +f2 ∈ C(K) and 0 ≤ f1 +f2 ≤ 1, f1 +f2 = 0
on K \ (G1 ∪ G2 ) gives µ∗A (G1 ∪ G2 ) ≥ A(f1 + f2 ) = A(f1 ) + A(f2 ) ≥ µ∗A (G1 ) + µ∗A (G2 ) − 2ε.
The reverse inequality appeared in Step 1. (iii).
Now given an open set G ⊂ K, and a test set T ⊂ K and ε > 0, we approximate T ∩ G
and T \ G be disjoint open sets.
Hence, there exists open sets V1 and V2 , such that

V1 ⊃ T, µ∗A (V1 ) < µ∗A (T ) + ε,

V2 ⊃ V1 \ G, µ∗A (V2 ) < µ∗A (V1 \ G) + ε.

We can also find a compact set C ⊂ V1 ∩ G such that µ∗A (C) + ε > µ∗A (V1 ∩ G) and
an open set W with a compact closure W such that C ⊂ W ⊂ W ⊂ V1 ∩ G. Set

77
Functional Analysis II

W0 := (V1 ∩ V2 ) \ W . Then W0 is an open set, W ∩ W0 = ∅, W ∪ W0 ⊂ V1 , V1 \ G ⊂ W0

and µ∗A (W ) + ε > µ∗A (V1 ∩ G). Thus

µ∗A (T ∩ G) + µ∗A (T \ G) ≤ µ∗A (V1 ∩ G) + µ∗A (V1 \ G) < µ∗A (W ) + ε + µ∗A (W0 )
= µ∗A (W ∪ W0 ) + ε ≤ µ∗A (V1 ) + ε < µ∗A (T ) + 2ε.

Step 3.: We denote by MA := M(µ∗A ) the σ-algebra of measurable sets of µ∗A . The
restriction of µ∗A to MA will be denoted by µA . We shall show, that this measure satisfies
the statement of the theorem.
First, µA is a complete Radon measure. As we restricted ourselves to K compact, it
is easy to see that C(K) ⊂ L1 (µA ). Hence, we only have to show that
Z
Af = f dµA , f ∈ C(K).
K

Let f ∈ C(K) be given. Without restriction, we may assume that 0 ≤ f ≤ 1. For n ∈ N

and k = 0, 1, . . . , n we denote

fk := min(f, k/n), Gk := {x ∈ K : f (x) > k/n}.

Using the definition of µ∗A (Gk−1 ) and n(fk − fk−1 ) = 0 on K \ Gk−1 , we get
1
≤ A(fk − fk−1 ) ≤ µA (Gk−1 )
n
for each k = 1, . . . , n. On the other hand, as n(fk − fk−1 ) = 1 on Gk , we first obtain
µA (C) ≤ nA(fk − fk−1 ) for every compact C ⊂ Gk , which implies also µA (Gk ) ≤ nA(fk −
fk−1 ) for all k = 1, . . . , n.
As χGk ≤ n(fk − fk−1 ) ≤ χGk−1 , we get also
Z
1 1
µA (Gk ) ≤ (fk − fk−1 )dµA ≤ µA (Gk−1 )
n K n
for each k = 1, . . . , n.
Thus (using f0 = 0 and fn = f )
 Z  X  Z 
   n 
Af − f dµA  =  A(fk − fk−1 ) − (fk − fk−1 )dµA 
  
K k=1 K
n
X 1 1 1
≤ (µA (Gk−1 ) − µA (Gk )) = µA (G0 ) = µA ({x ∈ K : f (x) > 0}).
n n n
k=1

Since RµA ({x ∈ K : f (x) > 0}) < ∞ and n may be chosen arbitrarily large, we get
Af = K f dµA .
Step 4.: Finally, we prove the uniqueness.
If ν is a complete Radon measure on K satisfying the statement of the theorem, we have
Z
Af = f dν ≤ χG dν = ν(G)
K

for any open set G ⊂ K and f ∈ C(K) with 0 ≤ f ≤ 1, f = 0 on K \ G. Hence,

µA (G) = µ∗A (G) ≤ ν(G). On the other hand, by a similar argument and by the point
(i) above, we have µA (C) ≥ ν(C) for each compact set C ⊂ K and from regularity of
µA and ν it immediately follows that they coincide on Borel sets, which concludes the
assertion.

78
 16 Selected topics

16.2 Lemma of Johnson and Lindenstrauss

In dieser Vorlesung werden einige Eigenschaften Zufälliger Matrizen untersucht. Das
Hauptergebnis ist dann der Beweis des Lemmas von Johnson und Lindenstrauss.

Wir benutzen nur normalverteilte Zufallsvariablen N (0, 1), d.h. Zufallsvariablen mit
Wahrscheinlichkeitsdichte
1 2
p(x) = √ · e−x /2 .
2π
2 √
Lemma 16.2.1. (i) Sei ω ∼ N (0, 1). Dann gilt E (eλω ) = 1/ 1 − 2λ für −∞ < λ <
1/2.

(ii) (2-Stabilität der Normalverteilung) Sei k eine natürliche Zahl, sei λ = (λ1 , . . . , λk ) ∈
Rk und seien ω1 , . . . , ωk unabhängige identisch
P verteilte (u.i.v.) N (0, 1) Zufallsvari-
ablen. Dann gilt λ1 ω1 + · · · + λk ωk ∼ ( ki=1 λ2i )1/2 · N (0, 1).
√
Proof. Der Beweis von (i) folgt aus der Rechnung (y := 1 − 2λ · x)
Z ∞ Z ∞ Z ∞
1 λx2 −x2 /2 1 (λ−1/2)x2 1 2 dy 1
√ e ·e dx = √ e dx = √ e−y /2 · √ =√ .
2π −∞ 2π −∞ 2π −∞ 1 − 2λ 1 − 2λ

Um (ii) zu beweisen, reicht es nur den Fall k = 2 zu betrachten. Der Rest folgt dann
durch Induktion. Sei also λ = (λ1 , λ2 ) ∈ R2 , λ 6= 0 fest und seien ω1 und ω2 u.i.v. N (0, 1)
Zufallsvariablen. Wir setzen S := λ1 ω1 + λ2 ω2 . Sei t ∈ R beliebig. Wir berechnen
Z
1 2 2
P(S ≤ t) = e−(x +y )/2 dxdy
2π (x,y):λ1 x+λ2 y≤t
Z
1 2 2
= e−(x +y )/2 dxdy
2π x≤c;y∈R
Z
1 2
=√ e−x /2 dx.
2π x≤c
2 2
In dem letzten Schritt haben wir die Rotationsinvarianz der Funktion (x, y) → e−(x +y )/2
ausgenutzt. Den Wert c berechnen wir aus der Skizze (!Tafel!) als die Länge der Normalen
von Null auf die Gerade {(x, y) : λ1 x + λ2 y = t}. Es ergibt sich
 
λ1 t λ2 t t
c=
λ2 + λ2 λ2 + λ2 = pλ2 + λ2 .
,
1 2 1 2 1 2

Wir erhalten also

Z q 
1 −x2 /2 2 2
P(S ≤ t) = √ √ 2 2 e dx = P λ1 + λ2 · ω ≤ t .
2π λ1 +λ2 ·x≤t

Falls ω1 , . . . , ωk (auch abhängige) normalverteilte Zufallsvariablen sind, dann gilt E(ω12 +

· · · + ωk2 ) = k. Falls ω1 , . . . , ωk auch unabhängig sind, konzentriert sich der Wert von
ω12 +· · ·+ωk2 auch sehr stark um k. Dieser Effekt (auch Konzentration des Maßes genannt)
wird genau im nächsten Lemma beschrieben.

79
Functional Analysis II

Lemma 16.2.2. Sei k eine natürliche Zahl und seien ω1 , . . . , ωk unabhängige normalverteilte
Zufallsvariablen. Sei 0 < ε < 1. Dann gilt
k 2 /2−ε3 /3]
P(ω12 + · · · + ωk2 > (1 + ε)k) ≤ e− 2 [ε

und
k 2 /2−ε3 /3]
P(ω12 + · · · + ωk2 < (1 − ε)k) ≤ e− 2 [ε .

Proof. Wir beweisen nur die erste Ungleichung. Der Beweis der zweiten Ungleichung folgt
sehr ähnlich. Wir setzen β := 1 + ε > 1 und rechnen

P(ω12 + · · · + ωk2 > βk) = P(ω12 + · · · + ωk2 − βk > 0) = P(λ(ω12 + · · · + ωk2 − βk) > 0)
= P(exp(λ(ω12 + · · · + ωk2 − βk)) > 1)
≤ E exp(λ(ω12 + · · · + ωk2 − βk)),

wobei λ > 0 eine beliebige positive reelle Zahl ist und erst später bestimmt wird. Wir
haben im letzten Schritt die Markov-Ungleichung benutzt. Weiter benutzen wir die Eigen-
schaften der Exponentialfunktion und die Unabhängigkeit von ω1 , . . . , ωk . Es ergibt sich
2 2 2
E exp(λ(ω12 + · · · + ωk2 − βk)) = e−λβk · E eλω1 · · · eλωk = e−λβk · (E eλω1 )k

und mit Hilfe von Lemma 16.2.1 erhalten wir schließlich (für 0 < λ < 1/2)

E exp(λ(ω12 + · · · + ωk2 − βk)) = e−λβk · (1 − 2λ)−k/2 .

Wir suchen jetzt einen Wert von 0 < λ < 1/2, für den der letzte Ausdruck möglichst klein
wird. Wir berechnen also die Ableitung von e−λβk · (1 − 2λ)−k/2 und setzen sie gleich Null.
So erhalten wir

e−λβk · (−βk) · (1 − 2λ)−k/2 + e−λβk · (−k/2) · (1 − 2λ)−k/2−1 · (−2) = 0,

(−β) · (1 − 2λ) + 1 = 0,
1 − 2λ = 1/β,
1 − 1/β
λ= .
2
Wir sehen auch sofort, dass für diesen Wert von λ wirklich die Abschätzungen 0 < λ < 1/2
erfüllt sind. Wir benutzen diesen Wert von λ und erhalten
1−1/β β−1
P(ω12 + · · · + ωk2 > βk) ≤ e− 2
·βk
· (1 − (1 − 1/β))−k/2 = e− 2
k
· β k/2
εk k
= e− 2 · e 2 ln(1+ε) .

Das Ergebnis folgt dann aus der Ungleichung

t2 t3
ln(1 + t) ≤ t − + , −1 < t < 1.
2 3

Remark 16.2.3. Überlegen Sie sich, dass die Unabhängigkeit von ω1 , . . . , ωk in Lemma
16.2.2 nicht weggelassen werden darf.

80
 16 Selected topics

Lemma 16.2.4 (Johnson und Lindenstrauss, 1984). Sei 0 < ε < 1 und seien k, n und d
natürliche Zahlen mit
k ≥ 4(ε2 /2 − ε3 /3)−1 ln n.
Dann existiert für jede Menge P = {x1 , . . . , xn } eine Abbildung f : Rd → Rk , so dass

(1 − ε)kxi − xj k22 ≤ kf (xi ) − f (xj )k22 ≤ (1 + ε)kxi − xj k22 , i, j ∈ {1, . . . , n}. (16.1)

Proof. Wir setzen

 
ω1,1 . . . ω1,d
1 1  . .. ..  x,
f (x) = √ · M x = √  .. . . 
k k
ωk,1 . . . ωk,d

wobei (ωu,v ) u.i.v. N (0, 1) Zufallsvariablen sind. Wir zeigen, dass diese Wahl von f
mit positiver Wahrscheinlichkeit (16.1) erfüllt. Damit wäre die Existenz einer solchen
Abbildung bewiesen.
x −x
Sei i, j ∈ {1, . . . , n} beliebig mit xi 6= xj . Dann setzen wir y = kxii−xjjk2 und berech-
nen die Wahrscheinlichkeit, dass die rechte Ungleichung in (16.1) nicht erfüllt ist. Wir
benutzen die 2-Stabilität der Normalverteilung und erhalten

P(kf (xi ) − f (xj )k22 > (1 + ε)kxi − xj k22 ) = P(kf (y)k2 > (1 + ε)) = P(kM yk22 > (1 + ε)k)
= P((ω1,1 y1 + · · · + ω1,d yd )2 + · · · + (ωk,1 y1 + · · · + ωk,d yd )2 > (1 + ε)k)
= P(ω12 + · · · + ωk2 > (1 + ε)k)

für unabhängige normalverteilte ω1 , . . . , ωk . Mit Hilfe von Lemma 16.2.2 läßt sich der
letzte Ausdruck von oben durch
k 2 3
e− 2 [ε /2−ε /3]
abschätzen. Dieselbe Abschätzung gilt auch für die linke Ungleichung in (16.1) und für
n

alle 2 Paare {i, j} ⊂ {1, . . . , n} mit i 6= j. Die Wahrscheinlichkeit, dass eine der Ungle-
ichungen in (16.1) nicht gilt, ist also höchstens
 
n k 2 3 k 2 3 k
2· · e− 2 [ε /2−ε /3] < n2 · e− 2 [ε /2−ε /3] = exp(2 ln n − [ε2 /2 − ε3 /3]) ≤ e0 = 1
2 2

für k ≥ 4(ε2 /2 − ε3 /3)−1 ln n.

Remark 16.2.5. Wie ändert sich die Wahrscheinlichkeit, dass f = √1 · M die Formel
k
(16.1) erfüllt, wenn k größer wird, d.h. wenn

k ≥ c(ε2 /2 − ε3 /3)−1 ln n

mit c > 4?

Remark 16.1. Der Beweis des Lemmas 16.2.4 benutzt die sogenannte Probabilis-
tische Methode: Die Existenz von einem Objekt (in unserem Fall der Abbildung
f : Rd → Rk mit (16.1)) wird dadurch bewiesen, dass man zeigt, dass zufällig kon-
struierte Objekte mit positiver Wahrscheinlichkeit die gewünschten Eigenschaften
erfüllen. Als Urvater dieser Methode wird meistens Paul Erdös bezeichnet, es wurde
aber auch schon vorher benutzt. Bis jetzt gibt es keine konstruktive Beweise von
Lemma 16.2.4.

81
Functional Analysis II

Der ursprüngliche Beweis von Johnson und Lindenstrauss ist viel mehr geometrisch.
Man beweist, dass eine (richtig normierte) Projektion auf einen zufälligen k-dim.
Unterraum von Rd mit positiver Wahrscheinlichkeit (16.1) erfüllt.

Seit ca. 10-15 Jahren gibt es viele Varianten von diesem Beweis (meistens durch
spezifische Anwendungen motiviert).

i) Die Auswertung von f (x) soll möglichst schnell sein (in unserem Fall nur k × d,
sonst d ln(k) durch FFT, Fast Fourier Transform)
ii) Die Anzahl der Zufallsvariablen soll minimiert werden (in unserem Fall k × d)
iii) Die Implementierung soll möglichst einfach sein
iv) u.s.w.

82
References
[1] H. W. Alt, Lineare Funktionalanalysis. Eine anwendungsorientierte Einführung, 6. Aufl., Springer, 2012

[2] J. B. Conway, A course in functional analysis, 2. Aufl., Springer, 1990.

[3] H. Heuser, Funktionalanalysis. Theorie und Anwendung, 4. Aufl., Teubner, 2006

[4] W. Rudin, Functional Analysis, McGraw-Hill, 1991.

[5] D. Werner, Funktionalanalysis, 7. Aufl., Springer, 2011.

[6] O. Christensen, Frames and bases: an introduction course, Birkhäuser, 2008

[7] E. M. Stein, Singular integrals and differentiability properties of functions, Princeton

Mathematical Series, No. 30 Princeton University Press, Princeton, N.J., 1970.

[8] E. M. Stein and G. Weiss, Introduction to Fourier analysis on Euclidean spaces, Prince-
ton Mathematical Series, No. 32. Princeton University Press, Princeton, N.J., 1971.

[9] H. Triebel, Higher analysis, University Books for Mathematics, Johann Ambrosius
Barth Verlag GmbH, Leipzig, 1992.

[10] D. D. Haroske and H. Triebel, Distributions, Sobolev spaces, elliptic equations, EMS
Textbooks in Mathematics. European Mathematical Society (EMS), Zürich, 2008.

[11] K. Gröchenig, Foundations of time-frequency analysis, Applied and Numerical Har-

monic Analysis. Birkhäuser Boston, Inc., Boston, MA, 2001.

[12] L. Grafakos, Classical Fourier analysis, Second edition, Graduate Texts in Mathe-
matics, 249. Springer, New York, 2008.

[13] L. Grafakos, Modern Fourier analysis, Second edition, Graduate Texts in Mathemat-
ics, 250. Springer, New York, 2009.

[14] J. Lukeš and J. Malý, Measure and Integral, Matfyzpress, 2005; available at
”http://www.mff.cuni.cz/fakulta/mfp/download/books/lukes-maly - measure and integral.pdf”.

83