CMMS 11 FOURIER ANALYSIS

Yu. Safarov

Notation
If ck , k = 1, 2, . . . , are real numbers then

lim sup ck = lim sup {ck } , lim inf ck = lim inf {ck } ,
m→∞ k>m m→∞ k>m

∅ is the empty set;

{x1 , x2 , . . . } denotes the set with the elements x1 , x2 , . . . ;
N = 1, 2, 3, . . . is the set of positive interges;
R is the set of real numbers;
R+ is the set of nonnegative real numbers, that is, A is a subset of B;
∞ is the symbol of infinity;
R̂+ is the set R+ ∪ {+∞};
R̂ is the set R ∪ {+∞} ∪ {−∞}.

If A and B are some sets then

A ∪ B = {x : x ∈ A or x ∈ B} is the union of A and B;
A ∩ B = {x : x ∈ A and x ∈ B} is the intersection of A and B;
A ⊂ B means that x ∈ A implies x ∈ B, that is, A is a subset of B;
B ⊃ A is the same as A ⊂ B;
B \ A = {x : x ∈ B and x ∈ / A};
A
2 is the set of all subsets of A.

We shall always fix a set E (usually E = Rn ) and deal with subsets B of E and
complex functions f defined on E. Then
B c = E \ B is the complement of B in E.
If E is a metric space then
B = {x ∈ Rn : dist (x, B) = 0} is the closure (completion) of the set B.
The support of a function f on E is defined as the closure of the set {x ∈ E :
f (x) 6= 0}.

For a real function f we define

f+ (x) = max {f (x), 0} and f− (x) = max {−f (x), 0}.

Then f+ and f− are non-negative functions such that f = f+ − f− and f+ f− ≡ 0.

Typeset by AMS-TEX
1
2 CMMS 11 FOURIER ANALYSIS

1. Measures and measure spaces

σ-algebras.
Let E be an arbitrary set, and A be a set of subsets of E.
Definition 1.1. The set A is said to be a σ-algebra over E if
(1) ∅ ∈ A and E ∈ A;
(2) if B ∈ A then B c ∈ A;
(3) if Bk ∈ A, k = 1, . . . , then ∪∞
k=1 Bk ∈ A.
c
Since ∩∞ ∞ c c
k=1 Bk = (∪k=1 Bk ) and B1 \ B2 = B1 ∩ B2 , (1)–(3) imply
(4) if Bk ∈ A, k = 0, 1, . . . , then ∩∞
k=1 Bk ∈ A;
(5) if B1 ∈ A and B2 ∈ A then B1 \ B2 ∈ A.
The conditions (3) and (4) mean that σ-algebra is closed under countable unions
and intersections. Since we can take Bk = ∅ for k > N , a σ-algebra contains also
all the finite unions and intersections of its elements.
The set A which contains only ∅ and E is a σ-algebra over E. The set A
containing all the subsets of E is also a σ-algebra over E. Of course, most of
interesting σ-algebras lie somewhere in between these two extreme examples.
Examples 1.2.
(1) Let E = {a, b, c, d}. Then the set A = {∅, {a, b}, {c, d}, E} is a σ-algebra.
(2) Let E = {a, b, c} and A = {∅, {a}, E}. Then A is not a σ-algebra. Indeed,
the set ({a})c = {b, c} is not an element of A.
(3) Let E = N and A = Nf ∪ Nc , where Nf is the set of all finite subsets of N
and the set Nc ⊂ 2N is defined as follows:

X ∈ Nc if and only if X c = N \ X ∈ Nf .

Then A is not a σ-algebra over E. The main peculiarity of this example is

that the set A satisfies the conditions (1), (2) of our definition and is closed
under finite unions and intersections. However, A does not satisfy the last
condition (3). Indeed, the set B = {2, 4, 6, 8, . . . } is the union of elements
{2k} ∈ Nf but B ∈ / Nf and B ∈ / Nc .

Exercise 1.3. Let Ai , i ∈ I, be some σ-algebras over E where I is some set (not
necessarily countable). Prove that ∩i∈I Ai is a σ-algebra over E.
Let A0 be a set of subsets of E. Since 2E is a σ-algebra, it is clear that there
always exists at least one σ-algebra containing all the sets from A0 .
Definition 1.4. Denote by E(A0 ) the intersection of all σ-algebras over E which
contain all the sets from A0 . Then E(A0 ) is a σ-algebra (see Exercise 1.3) which is
called the σ-algebra generated by A0 .
In other words, E(A0 ) is the minimal σ-algebra over E which contains all the
sets from A0 .
Example 1.5. Let E = {a, b, c}, A0 = {{a}}. Then E(A0 ) = {∅, E, {a}, {b, c}}.
Indeed, since the set A1 = {∅, E, {a}, {b, c}} is a σ-algebra which contains {a}
inclusion E(A0 ) ⊂ A1 holds. Since E(A0 ) is a σ-algebra and {a} ∈ E(A0 ) we have
{∅, E} ⊂ A0 , {b, c} = ({a})c = E \ {a} and hence A1 ⊂ A0 .
 CMMS 11 FOURIER ANALYSIS 3

Definition 1.6. Let E be a metric space. The Borel σ-algebra over E is the σ-
algebra generated by all open sets. We will denote this σ-algebra by B(E). The
sets B ∈ B(E) are said to be the Borel sets.
Example 1.7.
(1) Obviously, every open or closed set is a Borel set.
(2) Since any countable set {x1 , x2 , . . . } coincides with the union of closed sets
{xk }, the countable sets are also Borel sets.

Definition of a measure space. Complete measure spaces.

The reason for introducing σ-algebras is that they are natural objects on which
we can define measures.
Definition 1.8. The map µ : A → R̂+ is a measure if
(1) µ(∅) = 0; P∞
(2) µ is countably additive, that is µ (∪∞
k=1 Bk ) = k=1 µ(Bk ) for all Bk ∈ A
such that Bk ∩ Bm = ∅ , m 6= k.
If µ(E) < ∞ then µ is said to be a finite measure. If µ(E) = 1 then µ is called
a probability measure. The triple (E, A, µ), where E is an arbitrary set, A is a
σ-algebra over E and µ is a measure on A, is called a measure space.
Example 1.9. Let A be the set of all subsets of E.
(1) Let us fix a point x ∈ E and define the map µ as follows:

1, if x ∈ B,

µ(B) =
0, if x ∈
/ B.

Then µ is a measure. This measure is called the δ-measure at the point x

and is denoted δx .
(2) Let µ(B) coincides with the number of elements of the set B if B is finite
set and µ(B) = +∞ if B is infinite set. Then µ is a measure. This measure
is called the counting measure.

Definition 1.10. The measure space (E, A, µ) is said to be complete if every subset
B1 of every set B ∈ A with µ(B) = 0 also belongs to A (and then, by Lemma 1.12,
µ(B1 ) = 0).
Examples 1.11.
(1) If A = 2E (as in Example 1.9) then the measure space (E, A, µ) is always
complete.
(2) Let E = {a, b}, A = {∅, E}, and µ(∅) = µ(E) = 0. Then the measure
space (E, A, µ) is not complete.

Lemma 1.12. Let (E, A, µ) be a measure space and let A, B, B1 , B2 , . . . be ele-

ments of A. Then the following statements hold true.
(1) If A ⊂ B thenP µ(A) 6 µ(B).
∞
(2) µ(∪∞ B
k=1 k ) 6 k=1 µ(Bk ).
(3) If B1 ⊃ B2 ⊃ B3 ⊃ . . . , ∩∞
k=1 Bk = ∅ , and µ(B1 ) < ∞, then µ(BN ) → 0
as N → ∞.
4 CMMS 11 FOURIER ANALYSIS

Proof.
(1) Since B = A ∪ (B \ A), A ∩ (B \ A) = ∅ and the function µ is nonnegative,
we obtain µ(B) = µ(A) + µ(B \ A) > µ(A).
(2) Since

∪∞ m=k
k=1 Bk = (B1 ) ∪ (B2 \ B1 ) ∪ (B3 \ (B1 ∪ B2 )) ∪ . . . (Bk+1 \ ∪m=1 Bm ) . . .

and the sets

B1 , B 2 \ B1 , B3 \ (B1 ∪ B2 ), ... Bk+1 \ ∪m=k

m=1 Bm , ...

are mutually disjoint, from countable additivity of the measure µ it follows that

µ(∪∞
k=1 Bk ) = µ(B1 ) + µ(B2 \ B1 ) + µ(B3 \ (B1 ∪ B2 ) + . . .

+ µ(Bk+1 \ ∪m=k
m=1 Bm ) + . . .

Now the required inequality follows from the first statement.

(3) Since ∩∞
k=1 Bk = ∅, we have

Bm = (Bm \ Bm+1 ) ∪ (Bm+1 \ Bm+2 ) ∪ (Bm+2 \ Bm+3 ) . . .

for all m ∈ N. The sets Bk \ Bk+1 are mutually disjoint, and therefore
∞
X
µ(Bm ) = µ (Bk \ Bk+1 ) (1.1)
k=m

for all m. In particular, from

P∞ this equality and the first statement of the theorem
it follows that the series k=1 µ (Bk \ Bk+1 ) converges, which implies that (1.1)
vanishes as N → ∞. 
 CMMS 11 FOURIER ANALYSIS 5

2. Lebesgue’s measure
Outer measure.
Definition 2.1. Let (E, A, µ) be a measure space. The map ν : A → R̂+ is a
outer measure if
(1) ν(∅) = 0;
∞
(2) the inequality ν (∪∞
P
k=1 B k ) 6 ν(Bk ) holds for all Bk ∈ A.
k=1

Lemma 1.12 implies that every measure is a outer measure.

Example 2.2. Let A be the set of all subsets of E. Define the map µ as follows:
1, if B 6= ∅,

µ(B) =
0, if B = ∅.
Then µ is an outer measure. One can show that µ is not a measure if E contains
more than one element.
Definition 2.3. If a set P ⊂ Rn coincides with the direct product of some open
intervals (a1 , b1 ), (a2 , b2 ), . . . (an , bn ) then we say that P ⊂ Rn is an open rectangle
and denote
Voln (P ) = (b1 − a1 )(b2 − a2 ) . . . (bn − an ).

Definition 2.4. Let B ⊂ Rn . Denote

∞
X
µout (B) = inf Voln (Pk ) ,
k=1

where the infimum is taken over all countable families of rectangles {Pk }k∈N such
that B ⊂ ∪k∈N Pk .

Definition of Lebesgue’s measure.

Theorem 2.5.
n
(1) The map µout defined on the σ-algebra 2R (that is, on the σ-algebra of all
subsets of Rn ) is an outer measure.
(2) For any rectangle P the equality µout (P ) = Voln (P ) holds.
(3) Let B(Rn ) be the Borel σ-algebra, N the family of sets B ⊂ Rn such that
µout (B) = 0, and Ln be the σ-algebra generated by N ∪ B(Rn ). Let µn be
the restriction of µout to Ln . Then µn is a measure and the measure space
(Rn , Ln , µn ) is complete.

Defintion 2.6. The measure µn defined in Theorem 2.5 is called the Lebesgue
measure. The sets B ⊂ B(Rn ) are said to be Lebesgue measurable.
Of course, Lebesgue’s measure is the most natural measure in Rn . It is invariant
with respect to Euclidean transformations (translations and rotations). Under the
additional assumption that the measure of the unit cube is equal to 1, it is the only
measure (defined on the σ-algebra Ln ) which has this property.
6 CMMS 11 FOURIER ANALYSIS

The Cantor set.

It is clear that any countable set has Lebesgue measure zero. However, it is
not immediately clear that there are uncountable subsets of R1 whose Lebesgue
measure is zero. In this subsection we construct such a set. Namely, we start with
the set B0 = [0, 1] and let C1 be the set obtained by removing the open middle
third of C0 ;

[− − − − − − − − − − − − −(· · · · · · · · · · · · · · · · · · · · · · · · · · ··) − − − − − − − − − − − −−]

C1

(i.e., C1 = C0 \ (1/3, 2/3)). Next, let C2 be the set obtained from C1 after removing
the open middle third of each of the (two) intervals of which C1 is disjoint union;

[− − − − (· · · · · · ··) − − − −(· · · · · · · · · · · · · · · · · · · · · · · · · · ··) − − − −(· · · · · · ··) − − − −]

C2

More generally, given Ck which is a union of 2k disjoint, closed intervals, let Ck+1
be the set which one gets by removing the open middle third of each of these closed
intervals. Finally, set C = ∩∞
k=0 Ck .
The set C is called the Cantor set. It is closed, and does not contain any
isolated points (i.e., any neighbourhood U of a point x ∈ C contains also other
points from C). There is a one-to-one correspondence between the points from C
and the sequences of 0 and 1 (the first k numbers in the sequence indicate the closed
interval in Ck which contains x). Therefore C is uncountable.
The measure of C0 \ C is equal to
 2  3
1 1 2 1 2 1 2 1 1
+ · + · + · + ... = · = 1,
3 3 3 3 3 3 3 3 1 − 2/3

so the measure C is equal to zero.

 CMMS 11 FOURIER ANALYSIS 7

3. Measurable functions
In this section we fix a σ-algebra A over the set E and say that a set is measurable
if it belongs to our σ-algebra A.
Definition 3.1. We say that the real function f : E → R̂ is measurable (or, more
precisely, that f is a measurable function with respect to the σ-algebra A) if for
any real a the inverse image
f −1 ([−∞, a)) = {x ∈ Rn : f (x) < a}
is an element of A.
In the case where E = Rn and A = Ln we say that f is Lebesgue measurable,
and in the case where E is a metric space and A = B(E) we say that f is Borel
measurable (or, simply, that f is a Borel function). From inclusion B(Rn ) ⊂ Ln it
follows that any Borel function f : Rn → R̂ is Lebesgue measurable.
Remark 3.2. A measurable function can be infinite at some points, and such points
also form a measurable set. Indeed,
f −1 (−∞) = ∩∞
k=1 f
−1
([−∞, −k)) , f −1 (+∞) = ∪∞
k=1 (f
−1
([−∞, k)) )c .
Example 3.3. Let E be a metric space and let f be a continuous functon. Then f
is a Borel function. Indeed for any a ∈ R1 the set f −1 ((−∞, a)) is open and hence
is a Borel set.
Theorem 3.4. For any function f : E → R̂ the following two conditions are
equivalent:
(1) f is measurable
(2) the inverse images of all intervals (including degenerate and infinite inter-
vals) are elements of A.

Simple functions.
From the measure-theoretic standpoint, the most elementary functions are those
taking only a finite number of values.
Definition 3.5. The function ϕ : E → R1 is said to be simple if it takes a finite
number of values ak and for each ak the inverse image ϕ−1 (ak ) is a measurable set.
One can easily prove the following lemma.
Lemma 3.6.
(1) A simple function is measurable.
(2) Let ψ1 , ψ2 , . . . ψm be simple functions and let F : Rm → R1 be a continuous
function. Then the function F (ψ1 , ψ2 , . . . ψm ) is a simple function.
Lemma 3.6 immediately implies
Corollary 3.7. P
Let ψ1 , ψ2 , . . . ψm be simple functions and let a1 , a2 , . . . am be real
m
numbers. Then k=1 ak ψk is a simple function.
In other words, Corollary 3.7 says that a linear combination of simple functions
is a simple function.
The following lemma implies that every measurable function can be approxi-
mated by simple functions.
8 CMMS 11 FOURIER ANALYSIS

Lemma 3.8. If f is measurable then there exists a sequence of simple functions

ψk such that ψk (x) → f (x) for each x as k → ∞. Moreover, one can choose ψk in
a such way that for all x and k the estimates

(ψk )+ (x) 6 (ψk+1 )+ (x) 6 f+ (x), (ψk )− (x) 6 (ψk+1 )− (x) 6 f− (x)

hold. Furthermore, if the function f is bounded then we can choose a sequence ψk ,

k = 1, 2, 3, . . . which converges to f uniformly on E.
Proof. Since f = f+ − f− , it is sufficient to prove the lemma assuming that f and
ψk are non-negative (then the general result is obtained by considering the linear
combinations of the simple functions corresponding to f+ and f− ).
Let f be non-negative. We define

ϕk (x) = max {m/2k : m/2k 6 f (x)} (4.1)

m∈Z+

where Z+ = {0, 1, 2, . . . }; in other words

ϕk (x) = m/2k if m/2k 6 f (x) < (m + 1)/2k , m = 0, 1, 2, . . .

Obviously, 0 6 ϕk (x) 6 ϕk+1 (x) 6 f (x).
Since f is measurable, the inverse images
ϕ−1
k (m/2 k
) = f −1
[m/2k
, (m + 1)/2k
) are measurable sets.
If f is bounded then every function ϕk can take only a finite number of values
(R)
0 , 1/2k , . . . , m0 /2k (where m0 /2k < sup f ). Thus, ϕk are simple functions, and

0 6 ϕk (x) 6 ϕk+1 (x) 6 f (x)

for all x and k. We have |f (x) − ϕk (x)| 6 2−k , so ϕk → f uniformly as k → ∞.

Let f be unbounded. We define

ϕk (x) , ϕk (x) 6 k ,
ψk (x) =
k, ϕk (x) > k .

Then ψk are simple functions and

0 6 ψk (x) 6 ψk+1 (x) 6 f (x) , ∀ x, k .

Moreover, ψk (x) → f (x) for each fixed x. Indeed,

if f (x) is finite then ψk (x) = ϕk (x) for sufficiently large k, so ψk (x) → f (x);
if f (x) is infinite then, obviously, ψk (x) → +∞.
The proof is complete. 

A convergence theorem.
We shall deduce all the basic properties of the measurable functions from the
following theorem.
 CMMS 11 FOURIER ANALYSIS 9

Theorem 3.9. Let ϕ1 , ϕ2 , . . . be a sequence of measurable functions. Then

(1) g1 (x) = sup {ϕk (x)} and g2 (x) = inf {ϕk (x)} are measurable functions;
k k
(2) f1 (x) = lim sup ϕk (x) and f2 (x) = lim inf ϕk (x) are measurable functions;
(3) if ϕk (x) → f (x) for every x as k → ∞ then f is measurable.
Proof.
(1) If a ∈ R1 then from definitions of supremum and infimum it follows that
g1−1 (−∞, a] = ∩k ϕ−1
k (−∞, a]
g2−1 [a, +∞) = ∪k ϕ−1
k ([a, +∞).

Now the required statement follows from Theorem 3.4 and the definition of mesura-
bility.
(2) is an easy consequence of (1) and equalities
lim sup ϕk (x) = inf sup {ϕm (x)}, lim inf ϕk (x) = sup inf {ϕm (x)}.
k m>k k m>k

(3) immediately follows from (2). 

Properties of measurable functions.

Lemma 3.8 and Theorem 3.9 imply that a function f is measurable if and only if
f is the pointwise limit of a sequence of simple functions. Therefore we can obtain
various results concerning the measurable functions by proving them first for the
simple functions and then taking the pointwise limit. In particular, we obtain
Theorem 3.10. Let F (x1 , . . . , xp ) be a continuous function of p real variables.
Then for any measurable functions f1 , . . . , fp the function F (f1 , . . . , fp ) is measur-
able.
Proof. From lemma 3.8 it follows that for every m there exists a sequence of simple
functions ψm,k such that ψm,k (x) → fm (x) for each x as k → ∞. Lemma 3.6
implies that for any k ∈ N the function
Fk (x) = F (ϕ1,k (x), . . . , ϕp,k (x))
is simple. Since F is continuous, we have
F (f1 (x), . . . , fp (x)) = lim F (ϕ1,k (x), . . . , ϕp,k (x)) = lim Fk (x)
k→∞ k→∞

for all x. Thus, F (f1 , . . . , fp ) is the pointwise limit of a sequence of simple functions
and therefore F (f1 , . . . , fp ) is measurable. 
Theorem 3.10 immediately implies
Corollary 3.11. If the functions f and g are measurable then |f |, f + g and f g
are also measurable.

Complex-valued functions.
A complex-valued function is said to be measurable (simple) if its real and com-
plex parts are measurable (simple) functions. Clearly, all results of this section can
be easily extended to the complex-valued functions.
10 CMMS 11 FOURIER ANALYSIS

4. Construction of integrals
In this section we fix a measure space (E, A, µ).
Definition 4.1. Denote by Σ0 (E, A, µ) (or simply by Σ0 ) the set of all simple
functions ϕ such that µ({x ∈ E : ϕ(x) 6= 0} < ∞.
Definition 4.2. If ϕ ∈ Σ0 then we define
Z m
X
ϕ dµ = ak µ(Ωk )
E k=1

where a1 , a2 , . . . , am are the nonzero values of the function f and Ωk = ϕ−1 (ak ) =
{x ∈ E : ϕ(x) = ak }.
Example 4.3. Let (E, A, µ) = (R1 , L1 , µ1 ) and

0, if x ∈ (−∞, −1),



1, if x ∈ [−1, 1],


ϕ(x) =


 3, if x ∈ (1, 2),
0, if x ∈ [2, +∞).


Then the function ϕ belongs the set Σ0 and

Z
ϕ dµ1 = µ1 [−1, 1] + 3µ1 (1, 2) = 2 + 3 = 5 .
R1

Lemma 4.4. Let f ∈ Σ0 and g ∈ Σ0 . Then for any (real) constants α and β
Z Z Z
(αf + βg) dµ = α f dµ + β g dµ .

Proof. Obviously, αf dµ = α f dµ for all simple functions f and α ∈ R1 . There-

R R

fore it is sufficient to prove that

Z Z Z
(f + g) dµ = f dµ + g dµ . (5.1)

Assume that f and g take the values a1 , . . . , al and b1 , . . . , bp respectively and

denote f −1 (ak ) = Ωk , g −1 (bm ) = Ω
e m . Then f + g is a simple function taking the
values ak + bm on the sets Ωk ∩ Ω e m , and by definition
Z X
(f + g) dµ = (ak + bm ) µ(Ωk ∩ Ω e m)
k,m
X X
= ak µ(Ωk ∩ Ω
e m) + bm µ(Ωk ∩ Ω
e m ) . (5.2)
k,m k,m

e m form an exact non-overlapping cover of Rn ,

Since the sets Ω
X
µ(Ω ∩ Ω
e m ) = µ(Ω)
m
 CMMS 11 FOURIER ANALYSIS 11

for all Borel sets Ω. Therefore the first term in the right hand side of (5.2) coincides
with Z
X
ak µ(Ωk ) = f dµ .
k

By analogy, the second term in the right hand side of (5.2) is equal to

X Z
am µ(Ω
e m) = g dµ .
m

This implies (5.1). 

Definition of integral.
Definition 4.5. Let f be a non-negative measurable function. We define
Z Z
f dµ = sup ϕ dµ ,
E

where the supremum is taken over all functions

R ϕ ∈ Σ0Rsuch that 0 6 ϕ 6 f. If E0
is a measurable subset of E then we define E0 f dµ = E f0 dµ, where

x ∈ E0 ,

f (x),
f0 (x) =
0, x∈
/ E0 .

Further on in this
R section we shallR always assume that the integrals are taken
over E and write f dµ instead of E f dµ.
One can easily prove that Definition 4.5 does not contradict to Definition 4.2,
that is, both definitions give the same result in the case where f is a non-negative
function from Σ0 .
Definition 4.6. Let f be a measurable function. We define
R R R
(1) if R f+ dµ = +∞ andR f− dµ < ∞ then R f dµ = +∞ ;
(2) if R f+ dµ < ∞ and R f− dµ = +∞ thenR f dµ = R −∞ ; R
(3) if f+ dµ < ∞ and f− dµ < ∞ then f dµ = f+ dµ − f− dµ.
In the last case the function f is said to be integrable. The class of integrable
functions is denoted by L1 (E, A, µ) (or simply L1 ).
Remark
R 4.7. Note that a measurable function f belongs to L1 if and only if
|f | dµ < ∞.
Theorem 4.8 (elementary properties of the integral). Let f and g be inte-
grable functions. Then
R R
(1) Rf dµ 6 Rg dµ whenever f 6 g;
(2) R| f dµ| 6 R|f | dµ;
(3) R af dµ =R a f dµ Rfor every real constant a;
(4) f dµ + g dµ = f + g dµ.
12 CMMS 11 FOURIER ANALYSIS

Proof. The first three results immediately follow from the definition. The last
result is less obvious. It is deduced from Lemma 4.4 with the use of Lemma 3.8
and Theorem 5.2 
The above definition of integral is due to Lebesgue. Note that Lebesgue’s integral
of a non-negative measurable functionR always exists (and
R may be infinite). But that
R f+ dµ = ∞ and f− dµ = ∞ then we cannot
is not true in the general case. If
say anything about the value of f dµ.
Example R4.9. Let f (x) = xR−1 sin x , x ∈ R1 , and µ be Lebesgue’sR measure on
∞ ∞ ∞
R1 . Then 1 f+ dµ = ∞ and 1 f− dµ = ∞, so Lebesgue’s integral 1 f dµ does
not exist. But the Riemann integral is well-defined and finite (this can be easily
proved by integrating by parts).
In a sense, Lebesgue’s definition of integrals is more restrictive as far as the
behaviour at infinity is concerned. On the other hand, it allows us to integrate very
non-smooth functions, for which the Riemann integral does not exist. Moreover, it
works for an arbitrary measure µ whereas the Riemann integral is defined only for
Lebesgue’s measure.
Example
R 4.10. Let µ be the δ-measure at the point x (Example 1.9). Then
f dµ = f (x).
 CMMS 11 FOURIER ANALYSIS 13

5. Convergence of integrals

Monotone convergence theorem.

In this section we fix a measure space (E, A, µ) and assume that it is complete.
We shall need the following technical lemma.
Lemma 5.1. Let ϕ ∈ Σ0 be a non-negative simple function and fk , k ∈ 1, 2, 3 . . .,
be non-negative measurable functions such that
fk (x) 6 fk+1 (x) , lim fk (x) > ϕ(x)
k→∞
R R
for all x ∈ E. Then lim fk dµ > ϕ dµ.
k→∞

Proof. Given δ > 0, denote Ek = {x ∈ E : fk (x) > ϕ(x) − δ}. Then, under
conditions of the lemma, Ek ⊂ Ek+1 , and ∪k Ek = E . Therefore (E \ Ek ) ⊃
(E \ Ek+1 ) and ∩k (E \ Ek ) = ∅. Now Lemma 1.12(3) implies that µ(E \ Ek ) → 0
as k → ∞.
In view of Theorem 4.8(1) and Lemma 4.4 we have
Z Z Z
fk dµ > fk dµ > (ϕ − δ) dµ
Ek Ek
Z Z Z Z Z Z
= ϕ dµ − ϕ dµ − δ dµ > ϕ dµ − C dµ − δ dµ
E E\Ek Ek E E\Ek Ek
Z
= ϕ dµ − C µ(E \ Ek ) − δ µ(Ek )

where C = max f (x). Since the last two terms in the right hand side can be
x∈E
made arbitrarily small by choosing large k and small δ respectively, the required
inequality follows. 
Theorem 5.2 (the monotone convergence theorem). Let fm be non-negative
measurable functions such that
fm (x) 6 fm+1 (x) , lim fm (x) = f (x)
m→∞
R R
for all x ∈ E. Then lim fm dµ = f dµ .
m→∞
Proof. The condition fm 6 fm+1 implies that fk 6 f and hence, by Theorem
4.8(1), Z Z
fk dµ 6 f dµ , ∀k = 1, 2, . . .

On the other hand, in view of Lemma 5.1, we have

Z Z
ϕ dµ 6 lim fm dµ.
m→∞

for all ϕ ∈ Simp0 such that 0 6 ϕ 6 f . By Definition 4.5 this implies that
Z Z
lim fm dµ > f dµ.
m→∞


14 CMMS 11 FOURIER ANALYSIS

Lebesgue’s dominated convergence theorem.

Lemma 5.3 (Fatou’s lemma). Let fk be non-negative measurable functions.
Then Z Z
lim inf fk dµ 6 lim inf fk dµ . (5.1)

Proof. Denote hm (x) = inf {fk (x)} Then

k>m

fm (x) > hm (x), hm (x) 6 hm+1 (x) (5.2)

for all x ∈ E and m ∈ N, Now (5.2)) and the monotone convergence theorem imply
Z Z Z
lim inf fk dµ = lim hk dµ = lim hk dµ
k→∞ k→∞
Z Z
= lim inf hk dµ 6 lim inf fk dµ .


Corollary 5.4. Let fk be real measurable functions.
R Assume that there exists a
measurable function g such that fk (x) 6 g(x) and g dµ < ∞. Then
Z Z
lim sup fk dµ > lim sup fk dµ.

Proof. If we apply Fatou’s lemma to the functions (g − fk ) then (5.1) turns into
the required inequality. 
Theorem 5.5 (Lebesgue’s dominated convergence theorem). Let fk be mea-
surable functions, and fk (x) → f (x) for all x. Assume thatR there exists a non-
negative measurable function g such that |fk (x)| 6 g(x) and g dµ < ∞. Then
Z
lim |f − fk | dµ = 0 , (5.5)
k→∞
Z Z
lim fk dµ = f dµ . (5.6)
k→∞

Proof. Since |fk (x)| 6 g(x), we have |f (x)| 6 g(x) and |f (x) − fk (x)| 6 2g(x). By
Corollary 5.4
Z Z Z
lim |f − fk | dµ 6 lim |f − fk | dµ = lim |f − fk | dµ = 0 ,
k→∞ k→∞ k→∞

which implies (5.5). In view of Theorem 4.8(2) and Theorem 4.8(4)

Z Z Z
| f dµ − fk dµ | 6 |f − fk | dµ

for all k ∈ N. Therefore (5.5) implies (5.6).

 CMMS 11 FOURIER ANALYSIS 15

Further properties of the integrals.

Using Theorems 5.2 and 5.5 one can prove that many results, which are valid
for the integrals of simple functions, remain valid for the integrals of arbitrary
integrable functions f . Indeed, in view of Lemma 3.8 any measurable function can
be approximated by simple functions, and then one can often take the limit. In
particular, f, g ∈ L1 then
R R
(1) | f g dµ| 6 (sup |f |) |g| dµ,

1/2 R 2
1/2
|f |2 dµ
R R
(2) | f g dµ| 6 |g| dµ ,
(3) for all p > 1
Z 1/p Z 1/p Z 1/p
p p p
|f + g| dx 6 |f | dx + |g| dx (Minkowski’s inequality)

etc.

Complex-valued functions. R R R
If f is a complex-valued function then one defines f dµ = Re f dµ+ Im f dµ.
Using the corresponding results on real-valued functions, one can easily prove sim-
ilar statements for complex-valued functions.
16 CMMS 11 FOURIER ANALYSIS

6. Comparison of measures.
The notion “almost everywhere”.
Definition 6.1. We say that an x-dependent statement about quantities on the
measure space (E, A, µ) holds almost everywhere (a.e.) if the set B of x for which
the statement fails is an element of A and µ(B) = 0. If E0 ∈ A then we say that
an x-dependent statement holds almost everywhere on E0 if the subset B ⊂ E0 of
x for which the statement fails is an element of A and µ(B) = 0.
Of course, the notion “almost everywhere” depends on the measure µ.
Example 6.2. Let µ be the δ-measure at the point x0 , and f and g be some
measurable functions. Then f > g a.e. if and only if f (x0 ) > g(x0 ).
Example 6.3. Let µ be Lebesgue’s measure on R1 , and f and g be measurable
functions on R1 such that f (x) = g(x) for all irrational x. Then f = g a.e.
Theorem 6.4. Let (E, A, µ) be a measure space, E0 ∈ A and f, g be non-negative
measurable functions on E0 . Then
R
(1) E0 f dµ = 0 if and only if f (x) = 0 a.e. on the set E0 ;
R R
(2) E0 f dµ = E0 g dµ whenever f = g a.e. on the set E0

Proof. This Theoremt is an easy consequence of definitions.

Since the integral over a set of measure zero is equal to zero, all the results
from Section 5 remain valid if the corresponding conditions are fulfilled
a.e. In particular, Theorem 5.5 is valid if fk → f a.e. and |fk | 6 g a.e.
The examples above show that two functions f and g which are equal a.e. may
take different values on a rather rich set. In particular, if µ is Lebesgue’s measure
on R1 then f (x) may not concide with g(x) on the set of rational numbers, or even
on an uncountable set (for example, on the Cantor set). The typical problem is to
prove that for a given f there exists (or does not exist) a smooth function g such
that f = g a.e.
Proposition 6.5. Let µ be Lebesgue’s measure and f be a measurable function on
R1 . Assume that f has a jump at x0 ∈ R1 , that is, the limits f (x0 +0) and f (x0 −0)
exist and f (x0 + 0) 6= f (x0 − 0). Then f cannot coincide a.e. with a function which
is continuous at x0 .
Proof. Let f = g a.e. Then any interval of the form (x0 − ε, x0 ) or (x0 , x0 + ε)
contains at least one point x at which f (x) = g(x) (otherwise f (x) 6= g(x) on a
set of positive Lebesgue’s measure). Therefore there exist sequences x− k ↑ x0 and
+ − − + +
xk ↓ x0 such that g(xk ) = f (xk ) and g(xk ) = f (xk ) for all k = 1, 2, . . . Then

lim g(x− +
k ) = f (x0 − 0) 6= f (x0 + 0) = lim g(xk ) ,
k→∞ k→∞

so the function g is not continuous at x0 . 

Singular and absolutely continuous measures.

 CMMS 11 FOURIER ANALYSIS 17

Definition 6.6. Let A be a σ-algebra over E and let µ1 and µ2 be measures defined
on A. We say that
(1) the measures µ1 and µ2 are (mutually) singular if there exists a set Ω ∈ A
such that µ1 (Ω) = µ2 (Ωc ) = 0.
(2) the measure µ1 is absolutely continuous with respect to the measure µ2 if
µ1 (Ω) = 0 whenever µ2 (Ω) = 0.

Example 6.7. Let ν1 coincides with the Lebesgue measure µn and ν2 coincides
with the restriction of the δ-measure at the point x0 to the σ-algebra Ln of Lebesgue
mesurable sets. Then the measures ν1 and ν2 are mutually singular. Indeed, we
can take Ω = {x0 }.
If two measures are singular then they live on non-overlapping subsets of E and,
in fact, have nothing to do with one another. In particular, it can happen that
f > g a.e. with respect to µ1 and f < g a.e. with respect to µ2 .
If µ1 is absolutely continuous with respect to µ2 then any statement which holds
µ2 -a.e. is also valid µ1 -a.e. Moreover, under some additional restriction the measure
µ1 can be represented as an integral with respect to the measure µ2 (see Theorem
6.10).
Definition 6.8. We say that the measure space (E, A, µ) (or the measure µ) is
σ-finite if E can be represented as the union of a countable collection of the sets
Ek , k = 1, 2, . . . such that Ek ∈ A and µ(Ek ) < ∞ for all k ∈ N.
Example 6.9. The Lebegue measure µn is σ-finite. Indeed, we can take Ek =
(−k, k)n , k ∈ N.
Theorem 6.10 (Radon–Nikodym theorem). Let A be a σ-algebra over the set
E and let µ1 and µ2 be measures defined on A. Assume that the measure µ2 is
σ-finite. Then µ1 is absolutely continuous with respect to µ2 if and only if there
exists a measurable non-negative function ρ such that
Z
µ1 (E0 ) = ρ dµ2 (6.1)
E0

for all sets E0 ∈ A.

The function ρ from Theorem 6.10 is said to be the density or the Radon–
Nikodym derivative of the measure µ1 with respect to µ2 . The density ρ is “almost
uniquely”
R defined.
R Indeed, if ρe is another measurable function satisfying (6.1) then
Ω
ρ dµ2 = Ω ρe dµ2 for all Borel sets Ω, which implies that ρe = ρ a.e. with respect
to µ2 .
Remark 6.11. One can prove that the function µ1 defined by (6.1) is always (even
if µ2 is not σ-finite) a measure which is absolutely continuous with respect to µ2 .

Decomposition of a Borel measure on R1 .

Let µ be a Borel measure on R1 , and

µ ([0, λ)) , λ > 0,
N (λ) =
−µ ([λ, 0)) , λ 6 0.
18 CMMS 11 FOURIER ANALYSIS

Then N is a non-decreasing measurable function on R1 such that

µ ([λ1 , λ2 )) = N (λ2 ) − N (λ1 ) , ∀ λ 1 , λ 2 ∈ R1 .
Since the intervals of the form [λ1 , λ2 ) generate the Borel σ-algebra, the measure µ
is uniquely determined by N .
Definition 6.12. The function N is said to be the distribution function of the
measure µ.
Definition 6.13. We say that the Borel measure µ is continuous if µ ({x}) = 0 for
all points x.
By Lemma 1.12 lim µ ([λ1 , λ2 )) = 0 and lim µ ([λ1 , λ2 )) = µ ({λ1 }). Therefore
λ1 ↑λ2 λ2 ↓λ1
the measure µ is continuous if and only if the corresponding distribution function
is continuous.
Definition 6.14. We say that the Borel measure µ is discrete if there exists a
countable set of points x1 , x2 , . . . such that Ω ∩ {x1 , x2 , . . . } = ∅ implies µ(Ω) = 0.
A discrete Borel measure can be represented as a sum of the form
X
µ = ck δxk , (6.2)
k

where δxk are the delta-measures at the points xk and ck = µ ({xk }). The corre-
sponding distribution function is constant on the intervals which do not contain the
points xk and has the jumps ck at the points xk .
Clearly, if µ1 is continuous and µ2 is discrete then µ1 and µ2 are mutually singular
(we can take Ω = {x1 , x2 , . . . }).
Proposition 6.15. Any Borel measure µ on R1 can be uniquely represented as the
sum µd + µc , where µd is a discrete measure and µc is a continuous measure.
Proof. Let Ω be the set of points x for which µ ({x}) > ε, and
Ω = ∪ε>0 Ωε = {x : µ ({x}) > 0} .
Since the measure of a bounded set is finite, the intersection of Ωε with any bounded
interval contains only a finite number of points. This implies that the sets Ωε are
countable. Therefore Ω is also countable. Let Ω = {x1 , x2 , . . . }, and µd be the
discrete measure defined by (6.2). Then µc = µ − µd is a continuous measure, and
µ = µd + µc . The points xk and the constants ck are determined uniquely by the
measure µ, so µd and µc are uniquely defined. 
Proposition 6.15 is not very helpful, since we do not have any explicit represen-
tation for an arbitrary continuous measure µc . It is easy to see that a measure
is continuous if it is absolutely continuous with respect to Lebesgue’s measure.
However, the converse is not true.
Example. Let N (λ) be the continuous function on [0, 1] such that
N (λ) = 1/2 as λ ∈ (1/3, 2/3),
N (λ) = 1/4 as λ ∈ (1/9, 2/9) and N (λ) = 3/4 as λ ∈ (7/9, 8/9),
etc. Then N (0) = 0, N (1) = 1, and N is constant on any open interval from
[0, 1] \ C, where C is the Cantor set (see Section 2). The corresponding measure µ
lives on the Cantor set, and for any fixed point x we have µ ({x}) = 0. Therefore µ
is neither discrete nor absolutely continuous with respect to Lebesgue’s measure.
 CMMS 11 FOURIER ANALYSIS 19

Definition 6.16. A continuous measure which is not absolutely continuous with

respect to Lebesgue’s measure is said to be singular continuous.
Theorem 6.17 (Lebesgue decomposition theorem). Any Borel measure µ on
R1 can be uniquely represented as the sum µd + µa + µs , where µd , µa and µs are
mutually singular measures such that µd is discrete, µs is singular continuous and
µa is absolutely continuous with respect to Lebesgue’s measure.
Of course, with the preceding notation µc = µa + µs . By Theorem 6.10 the abso-
lutely continuous component µa of the measure µ can be represented as an integral
with respect to Lebesgue’s measure. In general situation there is no good represen-
tation for the singular component µs . However, the Borel measures which appear
in “real” problems usually do not have the singular component. For such mea-
sures Theorem 6.17 allows one to obtain explicit formulae in terms of the densities
(corresponding to µa ) and the δ-measures (corresponding to µd ).
20 CMMS 11 FOURIER ANALYSIS

7. The Lebesgue spaces

Given a measure space (E, A, µ), a real number p ∈ [1, ∞) and a complex-valued
measurable function f on E, we define

Z 1/p
p
kf kLp (E,µ) = |f | dµ .

Obviously, as p varies kf kLp (E,µ) provides different estimates of the size f as it is

“seen” by the measure µ. In particular, kf kLp (E,µ) cannot detect any properties of
f which occur on sets having µ-measure zero.

Remark. In applications the most important Lp -norms are those corresponding to

p = 1 and p = 2. Another important object is the L∞ -norm of a measurable
function f , which is defined by

kf kL∞ (E,µ) = inf {M > 0 : |f | 6 M a.e. } .

However, we shall consider only finite p.

Definition 7.1. We say that f ∈ Lp (E, µ) if f is measurable and kf kLp (E,µ) < ∞.
When E = Rn and µ is the Lebesgue measure, we omit µ and simply write Lp (Rn ).

We shall identify two functions from Lp (E, µ) which coincide µ-a.e. (which
means, in fact, that we assume the elements of Lp (E, µ) to be the classes of func-
tions). Since kf kLp (E,µ) = 0 implies f = 0 a.e., this means that f = 0 as an element
of Lp (E, µ) whenever kf kLp (E,µ) = 0. By Minkowski’s inequality

kf + gkLp (E,µ) 6 kf kLp (E,µ) + kgkLp (E,µ) .

From here it follows that the metric defined on kf kLp (E,µ) by

dist (f, g) = kf − gkLp (E,µ) (7.1)

satisfies the triangle inequality. We shall consider Lp (E, µ) as a metric space with
the metric (7.1). In particular, fk → f in Lp (E, µ) if kf − fk kLp (E,µ) → 0.

Remark 7.2. Strictly speaking, the elements of Lp are equivalence classes of func-
tions. One can say “the function f belongs to Lp ” (Definition 7.1) but bearing
in mind that there are many other functions which coincide with f as elements of
Lp . In particular, if we modify f on a set of measure zero then the new function
coincides with the initial one as an element of the Lp space. Therefore one should
be carefull making statements like “the function f ∈ Lp (Rn ) is equal to zero at the
origin” (changing the function f at the origin we do not change it as an element of
Lp (Rn )).
By Lebesgue’s dominated convergence theorem, if fk → f µ-a.e. and there is
a function g ∈ Lp (E, µ) such that |fk (x)| 6 g(x) then fk → f in Lp (E, µ). The
converse is not true; convergence in Lp (E, µ) does not imply convergence µ-a.e.
However, we have the following
 CMMS 11 FOURIER ANALYSIS 21

Theorem 7.3. If kf − fk kLp (E,µ) → 0 then there is a subsequence fkj which con-
verges to f µ-a.e. Moreover, if {fk } is a Cauchy sequence in Lp (E, µ) then there
exist a subsequence fki and f ∈ Lp (E, µ) such that kf − fkj kLp (E,µ) → 0 and
fkj → f µ-a.e.
Proof. Assume that {fk } is a Cauchy sequence in Lp (E, µ), that is,
Z
p
kfm − fk kLp (E,µ) = |fm − fk |p dµ → 0 , k, m → ∞ .

If fkj → fe µ-a.e. then by Theorem 3.9 f is measurable and by Fatou’s lemma

kf − fkj kpLp (E,µ) 6 lim inf kfkl − fkj kpLp (E,µ) 6 sup kfkl − fkj kpLp (E,µ) → 0
l→∞ l>j

as j → ∞. This obviously implies f ∈ Lp (E, µ). Therefore it is sufficient to prove

only that any Cauchy sequence in Lp (E, µ) contains a subsequence which converges
µ-a.e. to some function f .
If {fk } is a Cauchy sequence in Lp (E, µ) then

lim sup µ ({x : |fm − fk | > ε}) = 0

k→∞ m>k

for each fixed ε. Let us choose 1 6 k1 < k2 < . . . in such a way that

sup µ {x : |fm (x) − fkj (x)| > 2−j−1 } 6 2−j−1 .

m>kj

If |fkj+1 (x) − fkj (x)| < 2−j−1 for all j > i then

l−1
X l−1
X
|fkl (x) − fki (x)| 6 |fkj+1 (x) − fkj (x)| < 2−j−1 < 2−i , ∀l > i.
j=i j=i

Therefore sup |fkl (x) − fki (x)| > 2−i implies sup |fkj+1 (x) − fkj (x)| > 2−j−1 ,
l>i j>i
and we obtain
 
−i
µ {x : sup |fkl (x) − fki (x)| > 2 }
l>i
 
−j−1
6 µ {x : sup |fkj+1 (x) − fkj (x)| > 2 }
j>i
 
−j−1
6 µ ∪ {x : |fkj+1 (x) − fkj (x)| > 2 }
j>i
∞
X ∞
X
µ {x : |fkj+1 (x) − fkj (x)| > 2−j−1 } 2−j−1 6 2−i . (7.2)


6 6
j=i j=i

Assume that the sequence fkj (x) is not convergent. Then there exists ε > 0 for
which the inequality
sup |fkl (x) − fki (x)| > ε
l>i
22 CMMS 11 FOURIER ANALYSIS

holds with an arbitrarily large i. Therefore the set of points at which {fkj } does
not converge is a subset of

Ωj = ∩ {x : sup |fkl (x) − fki (x)| > 2−i }

i>j l>i

for each fixed j. In view of (7.2)

∞
X   ∞
X
−i
µ(Ωj ) = µ {x : sup |fkl (x) − fki (x)| > 2 } 6 2−i .
i=j l>i i=j

Therefore µ(Ωj ) → 0 as j → ∞, which implies that {fkj } converges almost every-

where to some function f . 
Corollary 7.4. The space Lp (E, µ) is complete.
Proof. Given a Cauchy sequence {fk }, we choose a subsequence fkj which converges
in Lp (E, µ) to some function f ∈ Lp (E, µ) (Theorem 7.3). Then

kf − fk kLp (E,µ) 6 kf − fkj kLp (E,µ) + kfkj − fk kLp (E,µ)

for all k and kj , so

lim kf − fk kLp (E,µ) 6 lim kf − fkj kLp (E,µ) + lim kfkj − fk kLp (E,µ) = 0 .
k→∞ kj →∞ k,kj →∞


Definition 7.5. We say that a function ϕ on R1 is a step function if ϕ takes only
a finite number of values a1 , . . . , am and for each ak 6= 0 the inverse image ϕ−1 (ak )
is a (possibly degenerate) interval.
Theorem 7.6. The set of step functions is dense in Lp (R1 ) for all 1 6 p < ∞.
The proof of this theorem is based on the following
Lemma 7.7. Any open set U ∈ R1 is the union of a countable collection of mutu-
ally disjoint open intervals.
Proof. If U ⊂ R1 is open and x ∈ U , let Ix be the maximal connected component
of U containing x. Since U is open, Ix is an open interval. Obviously, for any
x, y ∈ U , either Ix = Iy or Ix ∩ Iy = ∅.
We have U = ∪x Ix where the union is taken over all the rational x. Since the
set of rational numbers is countable (and the intervals Ix either coincide or do not
intersect), we have obtained the required representation of U . 
Proof of Theorem 7.6. From the definition of Lebesgue’s integral it follows that any
function f ∈ Lp (R1 ) is approximated in Lp (R1 ) by compactly supported simple
functions. Any compactly supported simple function is a linear combination of the
characteristic functions of some measurable sets. Therefore it is sufficient to prove
that the characteristic function χ of an arbitrary measurable set Ω ∈ R1 can be
approximated by step functions.
By Theorem 2.5 for all ε > 0 there exist open sets Ωε ⊃ Ω such that µ1 (Ωε \Ω) 6
ε. Since the characteristic functions χε of the sets Ωε converge in Lp (R1 ) to χ as
ε → 0, we can assume without loss of generality that Ω is an open set.
 CMMS 11 FOURIER ANALYSIS 23

By Lemma 7.7 an open set Ω coincides with the union of a countable collection
of open intervals Ik , k = 1, 2, . . . If χk is the characteristic function of Ik then the
Pj
ej = k=1 χk converge to the characteristic function of Ω in Lp (R1 ) as
functions χ
j → ∞. 
Let C0∞ (R1 ) be the space of smooth functions with compact supports. Obviously,
the characteristic function of a bounded interval I can be approximated in Lp (R1 )
by C0∞ -functions fk (for example, we can choose fk in such a way that fk (x) = 1
for all k and x ∈ I, 0 6 fk (x) 6 1 for all x, and fk (x) = 0 if dist(x, I) > k −1 ).
Therefore Theorem 7.6 implies
Corollary 7.8. The space C0∞ (R1 ) is dense in Lp (R1 ) for all 1 6 p < ∞.
24 CMMS 11 FOURIER ANALYSIS

8. Distributions
Convergence in S(R1 ).
We say that the sequence of functions ϕk ∈ S(R1 ) converges to ϕ ∈ S(R1 ) in
S(R1 ) if
dm (ϕ − ϕk )
sup | xl | → 0, k → ∞,
x dxm
for all l, m = 0, 1, . . . . If ϕk → ϕ in S(R1 ) then for all l and m we have

dm (ϕ − ϕk )
| m
| 6 ck,l,m (1 + x2 )−l ,
dx
where ck,l,m → 0 as k → ∞. This implies, in particular, that ϕk → ϕ in Lp (R1 ) for
all p.

The dual space S 0 (R1 ).

A map f : S(R1 ) → C1 is said to be a functional on S(R1 ). The value of
functional f on the function ϕ ∈ S(R1 ) is denoted by hf, ϕi. We say that the
functional f is linear if

hf, αϕ + βψi = α hf, ϕi + β hf, ψi , ∀ϕ, ψ ∈ S(R1 ) , ∀ α, β ∈ C1 ,

and f is continuous if ϕk → ϕ implies hf, ϕk i → hf, ϕi.

Definition 8.1. The linear continuous functionals on S(R1 ) are said to be the
(temperate) distributions.
The distributions form a linear space. If ϕ, ψ ∈ S(R1 ) and α, β ∈ C1 , we define
the distribution αf + βg by

hαf + βg, ϕi = α hf, ϕi + β hg, ϕi , ∀ ϕ ∈ S(R1 ) .

The linear space of distributions is denoted by S 0 (R1 ).

1 1
Example R 8.2. If f ∈ Lp (R ) with p > 1 then the functional on S(R ) defined by
hf, ϕi = f ϕ dx is a distribution. Obviously, if two functions f1 and f2 define the
same distribution then f1 = f2 almost everywhere (with respect to the Lebesgue
measure). This allows us to identify the functions f ∈ Lp (R1 ) with distributions.
Further on we shall use the same notation f for the function f ∈ Lp (R1 ) and for
the corresponding distribution.
Example 8.3. If µ is a Borel measure on R1 such that

µ ({x : |x| 6 R}) 6 C(RN + 1)

R
with some positive constants C and N then hµ, ϕi = ϕ dµ is a distribution. One
can prove that Z Z
ϕ dµ1 = ϕ dµ2 , ∀ ϕ ∈ S(R1 )

if and only if µ1 = µ2 . Therefore the Borel measures can be also identified with the
corresponding distributions.
 CMMS 11 FOURIER ANALYSIS 25

Example 8.4. Let x ∈ R1 be a fixed point. The distribution δx defined by

hδx , ϕi = ϕ(x) , ∀ ϕ ∈ S(R1 )

is said to be the δ-function at x. The δ-function is one of the simplest distributions;

its value on ϕ depends only on the value of ϕ at one fixed point. We have
Z
hδx , ϕi = ϕ(x) dδx , ∀ ϕ ∈ S(R1 ) ,

we δx in the right hand side stands for the δ-measure at x (Example 1.9).

Operations with distributions.

One can do with the distributions almost all the same things as with the functions
from S(R1 ). The basic idea is as follows. Assume that we are going to extend a
linear operator T in the space S(R1 ) to the space S 0 (R1 ). First, we take ψ ∈ S(R1 )
and write Z
hT ψ, ϕi = T ψ ϕ dx , ∀ ϕ ∈ S(R1 )

(here we consider the function T ψ as a distribution). Then we try to find a linear

operator T 0 in S(R1 ) such that
Z
hT ψ, ϕi = ψ T 0 ϕ dx , ∀ ϕ ∈ S(R1 ) .

Finally, for f ∈ S(R1 ) we define T f by

hT f, ϕi = hf, T 0 ϕi , ∀ ϕ ∈ S(R1 ) . (8.1)

Obviously, (8.1) defines a linear functional T f on S(R1 ). If T 0 is continuous in

S(R1 ) then ϕk → ϕ implies hT f, ϕk i → hT f, ϕi. In this case the functional T f is
continuous, so T f ∈ S 0 (R1 ).
Definition 8.5. Let h be an infinitely smooth function on R1 such that

|dk h/dxk | 6 ck (1 + x2 )mk , ∀ k = 0, 1, . . . , (8.2)

with some constants ck > 0 and mk > 0. Then ϕ → h ϕ is a continuous operator

in S(R1 ), and for f ∈ S 0 (R1 ) we define hf by

hhf, ϕi = hf, h ϕi , ∀ ϕ ∈ S(R1 ) .

Definition 8.6. The differentiation is a continuous operator in S(R1 ), so for f ∈

S 0 (R1 ) we define f 0 by

hf 0 , ϕi = − hf, ϕ0 i , ∀ ϕ ∈ S(R1 ) .

In the same manner we can define many other operators in S 0 (R1 ), in particular,
the operator f (x) → g(x) = f (y(x)).
26 CMMS 11 FOURIER ANALYSIS

Example 8.7. If h is a continuously differentiable function satisfying (8.2) with

k = 0, 1 then the derivative of h ∈ S 0 (R1 ) coincides with the usual derivative f 0 .
Example 8.8. Let 
0, x < 0,
f (x) =
1, x > 0.
Then for all ϕ ∈ S(R1 ) we have
Z Z ∞
0 0
− hf, ϕ i = − f ϕ dx = − ϕ0 dx = ϕ(0) .
0

Therefore f 0 coincides with the δ-function at x = 0.

Example 8.9. Let
−∞ < a1 < a2 < · · · < am < ∞ .
Assume that f is a piecewise continuous function, which is equal to zero as x < a1
or x > am and is continuously differentiable on the intervals (ak , ak+1 ). Let

ck = f (ak + 0) − f (ak − 0)

be the jumps of f at the points ak , and let

f 0 (x), x ∈ (ak , ak+1 ) , k = 1, . . . , m − 1,


g(x) =
0, x < a1 or x > am .

Integrating by parts we obtain

Z Z m
X
0 0
− hf, ϕ i = − f ϕ dx = g ϕ dx + ck ϕ(ak ) .
k=1

Pm
Therefore f 0 = g + k=1 ck δak where δak are the δ-functions at the points ak .
Example 8.10. Let f = δx . Then hf 0 , ϕi = −ϕ0 (x). The distribution f 0 cannot
be described in any simpler way. This distribution is called the derivative of the
δ-function at x.

Supports of distributions.
Generally speaking, one cannot say what is the value of a distribution at one
fixed point. However, we can single out a class of distributions which coincide on
an open set.
Definition 8.11. We say that the distribution f vanishes on an open set Ω and
write f |Ω = 0 if hf, ϕi = 0 for all ϕ ∈ S(R1 ) with supp ϕ ⊂ Ω. We say that f
coincides with a distribution g on Ω if (f − g)|Ω = 0 .
In particular, we say that the distribution f coincides with a function g on Ω if
(f − g)|Ω = 0 in the sense of distributions.
 CMMS 11 FOURIER ANALYSIS 27

Definition 8.12. For f ∈ S 0 (R1 ) we define supp f = R1 \ Ωf , where Ωf is the

maximal open set on which f is equal to zero (i.e., Ωf is the union of all open sets
Ω such that f |Ω = 0).
Example 8.13. The supports of all the derivatives of the δ-function at x coincide
with the point x.
The support of a function coincides with the support of the corresponding distri-
bution modulo a set of Lebesgue’s measure zero. If h is a function satisfying (8.2)
then
supp(hf ) ⊂ (supp h) ∩ (supp f ) , ∀ f ∈ S 0 (R1 ) .
In particular, if h = 0 in a neighbourhood of supp f then hf = 0. That is not
necessarily true if h = 0 only on supp f .
Example 8.14. Let f = δ00 be the derivative of the δ-function at zero, and
h(x) = x. Then h = 0 on supp f . However,

hhf, ϕi = − (xϕ)0 |x=0 = − ϕ(0) ,

so xδ00 = −δ0 . In the same manner we obtain that x2 δ00 = −xδ0 = 0.

The set of distributions with compact supports is denoted by E 0 (R1 ). We can
choose C0∞ -functions ψj such that j ψj (x) ≡ 1 (it is called a partition of unity).
P

Then an arbitrary distribution f ∈ S 0 (R1 ) is represented as a sum of distributions

fj = ψj f with compact supports.
Theorem 8.15. If f ∈ E 0 (R1 ) then
m
X dj gj
f = , (8.3)
j=0
dxj

where gj are some continuous functions.

Theorem 8.15 implies that any distribution f ∈ S 0 (R1 ) can be represented as a
sum of distributions fj of the form (8.3). However, the number of derivatives m in
the representation of fj may be depending on j and going to infinity as j → ∞.
28 CMMS 11 FOURIER ANALYSIS

9. Fourier transform in S 0 (R1 )

We shall need the following
Lemma 9.1. The Fourier transform is continuous in S(R1 ).
Proof. Let ψ̂k (ξ) be the Fourier transforms of ψk ∈ S(R1 ). We have to prove that
ψk → 0 in S(R1 ) implies

dm ψ̂k
sup | ξ l | → 0, k→∞ (9.1)
ξ dξ m

(with the notation of Section 8 ψk = ϕ − ϕk ). Since

dm fˆ
 l\
d (xm f )

l
ξ m
= (−i)l+m , ∀f ∈ S(R1 ) ,
dξ dxl
l
and f → xm ddxfl is a continuous operator in S(R1 ), it is sufficient to obtain (9.1)
only for l = m = 0. This follows from the estimates
Z
−1/2
sup |ψ̂k (ξ)| = (2π) |e−ixξ ψk (x) dx |
ξ
Z Z
−1/2 −1/2
6 (2π) |ψk (x)| dx 6 ck (2π) (1 + x2 )−1 dx ,

where ck = sup|(1 + x2 )ψk (x)| → 0 as ψk → 0 in S(R1 ). 

x

If f ∈ S(R1 ) then

hfˆ, ϕi = hf, ϕ̂i , ∀ ϕ ∈ S(R1 ) . (9.2)

Let now f be a distribution. Then, in view of Lemma 9.1, (9.2) defines a linear
continuous functional fˆ on S(R1 ).
Definition 9.2. The Fourier transform of f ∈ S 0 (R1 ) is the distribution fˆ defined
by (9.2).
Obviously, (9.2) holds if f belongs to Lp (R1 ), p = 1, 2, and if the Fourier trans-
form is understood in the Lp sense. Therefore the Fourier transform in S 0 (R1 )
coincides with the Fourier transform in Lp (R1 ) if f ∈ Lp (R1 ).
By analogy, the inverse Fourier transform F −1 f of f ∈ S 0 (R1 ) is defined by

hF −1 f, ϕi = hf, F −1 ϕi , ∀ ϕ ∈ S(R1 ) .

We have
hF −1 fˆ, ϕi = hfˆ, F −1 ϕi = hf, F
\ −1 ϕi = hf, ϕi ,

for all ϕ ∈ S(R1 ), which means that F −1 fˆ = f . Since F −1 f (ξ) = fˆ(−ξ), all the
results concerning the Fourier transform in S 0 (R1 ) can be easily reformulated for
the inverse Fourier transform.
 CMMS 11 FOURIER ANALYSIS 29

Example 9.3. If δ̂0 is the Fourier transform of the δ-function at x = 0 then

Z
−1/2
hδ̂0 , ϕi = ϕ̂(0) = (2π) ϕ(x) dx .

Therefore δ̂0 is the function identically equal to (2π)−1/2 . In the same way we obtain
that F −1 δ0 = δ̂0 . This implies that the direct and inverse Fourier transforms of the
function which is identically equal to a, coincide with a (2π)1/2 δ0 .
Proposition 9.4. For all f ∈ S 0 (R1 )

fb0 = (iξ) fˆ , c = i (fˆ)0 .

xf (9.3)

Proof. The equalities (9.3) are valid for all the functions from S(R1 ). Therefore for
all ϕ ∈ S(R1 ) we have

c = i hξ fˆ, ϕi ,
hfb0 , ϕi = hf 0 , ϕ̂i = − hf, (ϕ̂)0 i = i hf, xϕi

c , ϕi = hxf, ϕ̂i = hf, xϕ̂i = −i hf, ϕb0 i = i h(fˆ)0 , ϕi .

hxf
This implies (9.3) for all f ∈ S 0 (R1 ). 
From Example 9.3 and Proposition 9.4 it follows that the Fourier transforms of
polynomials are linear combinations of the δ-function at x = 0 and its derivatives.
Definition 9.5. We say that a distribution f belongs to the Sobolev class H s (R1 )
with s ∈ R1 if the Fourier transform fˆ is a function such that (1 + ξ 2 )s/2 fˆ(ξ) ∈
L2 (R1 ).
The inner product and the norm in H s (R1 ) is defined for the negative s in the
same way as for s > 0. If s < 0 then, generally speaking, the elements of H s (R1 )
are not functions but distributions. The value of s indicates how non-smooth the
distributions are. By Proposition 9.4 f ∈ H s (R1 ) implies dk f /dxk ∈ H s−k (R1 ),
so the differentiation makes the distribution more non-smooth. For example, the
derivative of a δ-function is a more non-smooth distribution than the δ-function
itself.

Applications to differential equations.

Let us consider the differential equation
m
X
ck Dk f = 0 , (9.4)
k=1

where Dk = (−i)k dk /dxk and ck are some constants. We assume that f is a

distribution and understand the derivatives in the sense of distributions. Then,
taking the Fourier transform, we obtain

P (ξ) fˆ(ξ) = 0 , (9.5)

Pm k
where P (ξ) = k=1 ck ξ is the symbol of the corresponding differential operators.
30 CMMS 11 FOURIER ANALYSIS

Clearly, if P is not identically equal to zero and fˆ is a function then (9.5) implies
fˆ = 0 almost everywhere. Thus, we cannot obtain from (9.5) any non-trivial
solution of (9.4) if we assume fˆ to be a function. But the equation (9.5) can have
non-trivial solutions in the class of distributions. For example, if P (ξ) has a zero
of the second order at the point η then (9.5) is fulfilled for fˆ = αδη + βδη0 with all
α, β ∈ C1 . Taking the inverse Fourier transform we obtain non-trivial solutions of
(9.4), which appear to be infinitely smooth functions.

Divergent integrals.
We have defined the Fourier transform for any distribution f ∈ S 0 (R1 ). In
particular, the distribution f can be a bounded function. This means, in fact, that
we have defined the integral Z
e−ixξ f (x) dx

for an arbitrary bounded measurable function f . Of course, the integral does not
converge in the classical sense, but it can be understood as a distribution in ξ.
This idea can be generalized in different directions. It often happens that a
divergent integral depending on a parameter can be interpreted as a distribution in
this parameter. Moreover, the corresponding distribution may well be a function
even if the integral does not converge.
Example 9.6. Let f be a bounded measurable function. Let us consider the
integral Z
2
eitx f (x) dx . (9.6)

We formally write
!
Z Z 1 Z
2 2 d 2
e−itx f (x) dx = e−itx f (x) dx + ix−2 e−itx f (x) dx (9.7)
−1 dt |x|>1

and accept the right hand side of (9.7) as the definition of the integral (9.6). The
first term in the right hand side of (9.7) is a function of t, the second term is a
derivative of a function which we understand as a distribution. One can prove
that if f is infinitely differentiable with uniformly bounded derivatives, then the
distribution on the right hand side of (9.7) coincides with an infinitely smooth
function for t 6= 0.