Measure Theory
Measure Theory
an introduction
Bibliography 145
CHAPTER 1
(2) From An ↓ A it follows that (A1 \ An ) ↑ (A1 \ A), and hence by (1)
µ(A1 \ An ) → µ(A1 \ A).
Since An ⊂ A1 , A ⊂ A1 , and µ(A1 ) < ∞ this means
µ(A1 ) − µ(An ) → µ(A1 ) − µ(A),
(see Lemma 1.3), from which the claim follows.
In particular, if (Ω, A, µ) is a finite measure space, then it follows from Theorem
1.13 that
(1.6) An ↓ ∅ ⇒ µ(An ) ↓ 0.
There is a partial converse.
Theorem 1.14. Let A be a σ-algebra and let µ be a finitely additive measure
on Ω such that
(1) µ(Ω) < ∞;
(2) µ satisfies (1.6).
Then µ is a measure.
µ(Ω0 ) < ∞.
Lemma 1.15. Let (Ω, A, µ) be a measure space. The following statements are
equivalent:
(1) there is a sequence Ωn ∈ A such that
∞
[
(1.7) µ(Ωn ) < ∞, Ω= Ωn ;
n=1
Proposition 1.18. Let (Ω, A, µ) be a σ-finite measure space and let the se-
quence Ωn ∈ A be such that (1.8) holds. Define µn : A → [0, ∞] by
µn (A) = µ(A ∩ Ωn ), A ∈ A.
Then µn defines a finite measure on Ω and µn (A) ↑ µ(A) for all A ∈ A.
Proof. It is clear by the text preceding Lemme 1.15 that µn is a measure
since Ωn ∈ A, and that this measure µn is finite since µ(Ωn ) < ∞. Now observe
(A ∩ Ωn ) ↑ A, which implies that
µn (A) = µ(A ∩ Ωn ) ↑ µ(A), A ∈ A,
due to Theorem 1.13.
3. Generators of σ-algebras
There are various ways to construct σ-algebras on a set Ω. A typical situation
is: given a collection E of subsets of Ω find the smallest σ-algebra which contains
the collection E.
Lemma 1.19. The intersection of a nonempty family of σ-algebras on a set Ω
is a σ-algebra.
Proof. Let Aα with α ∈ I, some index set,Tbe a collection of σ-algebras.
T Since
Ω in all Aα , it also belongs to the intersection α∈I Aα . If A belongs to α∈I Aα ,
Ac belong to all Aα and hence Ac belongs to α∈I Aα . Now let
T
then A and also T
A , n ∈ N, be in α∈I Aα . Then for each α ∈ I all sets An belong to Aα and hence
Sn∞
n=1 An belongs to Aα and thus to α∈I Aα .
T
Proposition 1.20. Let E be a collection of subsets of a set Ω. Then there is
precisely one σ-algebra A such that
(1) E ⊂ A;
(2) if B is a σ-algebra with E ⊂ B, then A ⊂ B.
Proof. The family of σ-algebras which contain E is nonempty as the power
set P (Ω) is such a σ-algebra. By Lemma 1.19 the intersection of this family is a
σ-algebra and it contains E, hence (1) has been shown. Item (2) is trivial.
Let A0 be another σ-algebra with the properties (1) and (2). Then clearly
A ⊂ A0 and, likewise, A0 ⊂ A. Hence a σ-algebra with the properties (1) and (2)
is uniquely determined.
Definition 1.21. Let E be a collection of subsets of a set Ω. The unique σ-
algebra in Proposition 1.20 is said to be generated by E, denoted by σ(E), and E is
said to be the generator of this σ-algebra.
Let E and F be collections of subsets of a set Ω. If E ⊂ F then it is clear that
σ(E) ⊂ σ(F). A σ-algebra A may be generated by different collections of subsets
of Ω.
By adding more structure to the set Ω one can obtain “natural” generators.
Definition 1.22. Let Ω be a topological space. The σ-algebra generated by all
open sets of Ω is the Borel σ-algebra B(Ω); its elements are called Borel measurable
subsets.
Note that countable intersections of open sets and countable unions of closed
sets are Borel sets. Such sets are sometimes called Gδ and Fσ sets, respectively.
In the following the only topological spaces that will be considered are the spaces
Ω = Rd with d ≥ 1. For d = 1 the σ-algebra B(R) will be denoted by B when no
confusion arises; likewise for d > 1 the σ-algebra B(Rd ) will be denoted by Bd .
Proposition 1.23. The Borel σ-algebra B on R is generated by
(1) the collection of closed subsets in R;
(2) the collection of intervals (−∞, b], b ∈ R;
(3) the collection of intervals (a, b], a, b ∈ R, a < b.
Moreover, the Borel σ-algebra Bd on Rd is generated by
(4) the collection of closed subsets;
(5) the collection of half-spaces {(x1 , . . . , xd ) : xi ≤ b} for some index i and
some b ∈ R;
(6) the collection of rectangles (a1 , b1 ] × · · · × (ad , bd ] where ai , bi ∈ R, ai < bi ,
and 1 ≤ i ≤ d.
Proof. Let B1 , B2 , and B3 be the σ-algebras, generated by the sets in (1),
(2), and (3), respectively. The Borel σ-algebra B includes the family of open sets
of R and, since it is closed under complementation, it also includes the family of
closed subsets of R. But then it also contains the σ-algebra B1 generated by the
closed subsets of R: B ⊃ B1 . The sets (−∞, b], b ∈ R, are closed and belong to B1 ;
hence B1 ⊃ B2 . Observe that for all a < b ∈ R:
(a, b] = (−∞, b] ∩ (−∞, a]c .
Since B2 is closed under complementation, each interval (a, b] belongs to B2 , and
therefore B2 ⊃ B3 . Each open subset of R belongs to B3 , due to its dyadic decom-
position; hence B3 ⊃ B. The first part of the proposition now follows from
B ⊃ B1 ⊃ B2 ⊃ B3 ⊃ B.
The second part is proved in the same way.
In a similar way one can show the following useful observation.
Proposition 1.24. The Borel σ-algebra Bd is generated by
(1) the collection of compact subsets in Rd ;
(2) the collection of rectangles [a1 , b1 ] × · · · × [ad , bd ] where ai , bi ∈ R, ai < bi ,
and 1 ≤ i ≤ d.
4. Dynkin systems
The study of generators of a σ-algebra is facilitated by the notion of a Dynkin
system. A Dynkin system is weaker than a σ-algebra: it does not require closedness
under countable unions, but only closedness under countable unions of pairwise
disjoint sets.
Definition 1.25. A collection D of subsets of Ω is a Dynkin system if
8 1. ALGEBRAS AND MEASURES
(1) Ω ∈ D;
(2) A ∈ D ⇒ Ac ∈ D; S∞
(3) An ∈ D, n ∈ N, pairwise disjoint ⇒ n=1 An ∈ D.
It is obvious that a σ-algebra is a Dynkin system. However, not every Dynkin
system is a σ-algebra.
The following lemma and proposition can be proved in the same way as for the
corresponding results for σ-algebras.
Lemma 1.26. The intersection of a nonempty family of Dynkin systems is a
Dynkin system.
Proposition 1.27. Let E be a collection of subsets of Ω. Then there is precisely
one Dynkin system D such that
(1) E ⊂ D;
(2) if F is a Dynkin system with E ⊂ F, then D ⊂ F.
Definition 1.28. Let E be a collection of subsets of a set Ω. The unique Dynkin
system in Proposition 1.27 is said to be generated by E, denoted by d(E), and E is
said to be the generator of this Dynkin system.
Let E be a collection of subsets of Ω. The σ-algebra σ(E) contains E. Since
σ(E) is a Dynkin system, which contains E, it follows that
(1.9) d(E) ⊂ σ(E).
It is of great practical interest to know when equality holds in (1.9). The following
lemmas will lead to the main theorem concerning this question.
Lemma 1.29. The following statements are equivalent:
(1) D is a σ-algebra;
(2) D is a Dynkin system which is closed under intersections.
Proof. (1) ⇒ (2) This implication is clear.
(2) ⇒ (1) Let D be a Dynkin system which is closed under intersections. First
note that for A, B ∈ D one has
A ∪ B = (Ac ∩ B c )c ∈ D,
as Ac , B c ∈ D and Ac ∩ B c ∈ D by the extra assumption. Hence D is an algebra
and thus D is closed under finite unions. It will now be shown that D is closed
under countable unions. Let An ∈ D and introduce the pairwise disjoint sets A0n
as in (1.2). Then A0n = An \ (A1 ∪ · · · ∪ An−1 ) ∈ D, and
∞
[ ∞
[
An = A0n ∈ D,
n=1 n=1
since D is a Dynkin system. Therefore D is a σ-algebra.
Lemma 1.30. Let E be a collection of subsets of Ω. Then the collection
(1.10) DD = {A ∈ P (Ω) : A ∩ D ∈ d(E)} when D ∈ d(E),
is a Dynkin system. Moreover, if E is closed under intersections, then
(1.11) d(E) ⊂ DD when D ∈ d(E),
and, consequently, d(E) is closed under intersections.
4. DYNKIN SYSTEMS 9
The following corollaries are typical applications of the main result, Theorem
1.31. It suffices to prove a statement concerning the σ-algebra generated by a set
10 1. ALGEBRAS AND MEASURES
5. Completion of σ-algebras
Let (Ω, A, µ) be a measure space and let A ∈ A have measure 0. Since the
measure µ is monotone it follows that µ(B) = 0 when B ⊂ A and B belongs to A.
However, it is not necessarily true that subsets of a set of measure 0 are measurable.
Definition 1.34. A measure space (Ω, A, µ) is complete if every subset of a
set of measure 0 is measurable (and, hence, is itself a set of measure 0).
Definition 1.35. Let (Ω, A, µ) be a measure space. Let Z be the class of all sets
N such that to each N there corresponds some F ∈ A with N ⊂ F and µ(F ) = 0.
5. COMPLETION OF σ-ALGEBRAS 11
Note that the class Z in Definition 1.35 is closed under countable unions. To
see this, let Nn ∈ Z. Then there exist Fn ∈ A with Nn ⊂ Fn and µ(Fn ) = 0. Hence
∞ ∞ ∞ ∞
!
[ [ [ [
Nn ⊂ Fn , Fn ∈ A, µ Fn = 0,
n=1 n=1 n=1 n=1
S∞
which shows that n=1 Nn ∈ Z.
Definition 1.36. Let (Ω, A, µ) be a measure space and let Z be the class in-
troduced in Definition 1.35. Let Ā be the class of all sets of the form E ∪ N with
E ∈ A and N ∈ Z. The completion µ̄ of µ is defined on Ā by
µ̄(E ∪ N ) = µ(E), E ∈ A, N ∈ Z.
E ∪ N = E ∪ [F ∩ (F \ N )c ] = (E ∪ F ) ∩ (E ∪ (F \ N )c ),
so that
(E ∪ N )c = (E ∪ F )c ∪ ((F \ N ) ∩ E c ).
Now E ∪ F ∈ A implies that (E ∪ F )c ∈ A, while (F \ N ) ∩ E c ⊂ F . Hence
(E ∪ N )c ∈ Ā. Thus Ā is closed under complements. To see that Ā is closed under
countable unions, let En ∈ A and Nn ∈ Z. Then
∞ ∞ ∞
! !
[ [ [
(En ∪ Nn ) = En ∪ Nn ∈ Ā,
n=1 n=1 n=1
An = En ∪ Nn , En ∈ A, Nn ∈ Z, µ̄(An ) = µ(En ).
12 1. ALGEBRAS AND MEASURES
S∞
Now let An ∈ A, then n=1 An ∈ A, so that
∞ ∞
!
[ [
0
(An ∩ Ω ) = An ∩ Ω0 ∈ A 0 .
n=1 n=1
Hence A is a σ-algebra on Ω .
0 0
Construction of measures
How to provide an arbitrary set Ω with measurable sets and a measure? One
way to do this is by means of the notion of outer measure. Outer measures are
defined on all subsets of Ω and they are required to be monotone and countably
subadditive. In this context a subset A ⊂ Ω is defined to be measurable if for
all subsets Z ⊂ Ω the outer measure is additive on the disjoint union of Z ∩ A
and Z ∩ Ac . The restriction of the outer measure to the measurable sets gives a
measure space which is complete (subsets of sets of measure 0 are measurable).
In this chapter measurable sets and measures are constructed via outer measures.
In particular the Lebesgue measure and the Lebesgue-Stieltjes measure will be
introduced.
measurable.
15
16 2. CONSTRUCTION OF MEASURES
2. Lebesgue measure on Rd
The notion of an outer measure for Rd is introduced via the volume of a (closed)
d-dimensional rectangle R of the form
R = [a1 , b1 ] × · · · × [ad , bd ], ai , bi ∈ R, ai ≤ bi , i = 1, . . . , d,
whose volume is defined by `(R) = (b1 − a1 ) · · · (bd − ad ).
Definition 2.6. Let A ⊂ Rd . Then m∗ (A) ∈ [0, ∞] is defined by
( ∞ ∞
)
X [
∗ d
(2.13) m (A) = inf `(Rn ) : Rn ⊂ R closed rectangle, A ⊂ Rn .
n=1 n=1
∗
P∞ Rn , n ∈ N, form a cover of A and m (A) is the infimum of the
The rectangles
total volumes n=1 `(Rn ) of all possible covers Rn .
Theorem 2.7. The extended real-valued function m∗ : P (Rd ) → [0, ∞] in
(2.13) is an outer measure on Rd , called the Lebesgue outer measure.
Proof. It will be shown that Definition 2.2 applies. It is clear that m∗ (∅) = 0.
Moreover, if A ⊂ B ⊂ R, then a covering of B is also a covering of A. Hence it
follows that
A ⊂ B ⇒ m∗ (A) ≤ m∗ (B).
18 2. CONSTRUCTION OF MEASURES
1Here the so-called ”grosse Umordnung” is used: give an enumeration of the cover and take
the sum via this enumeration.
2. LEBESGUE MEASURE ON Rd 19
Definition 2.9. The Lebesgue measurable sets of Rd are the measurable sets
defined by the Lebesgue outer measure m∗ in (2.13), and Lebesgue measure is the
restriction of m∗ to the Lebesgue measurable sets. The corresponding complete
measure space is denoted by (Rd , Md , m).
Let a ∈ Rd , then m∗ ({a}) = 0. Hence the subset {a} ⊂ Rd is Lebesgue
measurable with Lebegue measure 0. As a consequence any subset of the form
{an ∈ Rd : n ∈ N} is Lebesgue measurable with Lebesgue measure 0.
Let A = {x ∈ Rd : xi = 0} for a fixed 1 ≤ i ≤ d; i.e., A ⊂ Rd is a d − 1
dimensional hyperspace. Then A is Lebesgue measurable and m(A) = 0. To show
this, it suffices to treat the subspace A = Rd−1 × {0}. For any ε > 0 one has
∞ h ε
[ ε i
A⊂ [−n, n]d−1 × − · (2n)−(d−1) · 2−n , · (2n)−(d−1) · 2−n ,
n=1
2 2
Decompose each Rn with respect to the rectangle R into an almost disjoint finite
union of rectangles (i.e., these rectangles have disjoint interiors):
N
[n
Rn = Ren ∪ Tn,j , Ren = Rn ∩ R ⊂ R, Tn,j ⊂ Rc .
j=1
Therefore
∞
X ∞
X XNn
m∗ (Z ∩ R) + m∗ (Z ∩ Rc ) ≤ `(R
en ) + `(Tn,j ) < m∗ (Z) + ε.
n=1 n=1 j=1
where #Cn denotes the number of elements of Cn . Since Q is orthogonal one has
that Q(B) = B. In addition, Q(C1 ) ⊃ Q(C2 ) ⊃ · · · ⊃ Q(B) = B, cf. Theorem 1.13,
so that
X
(2.20) m(B) = lim m(Q(Cn )) = lim m(Q(R)).
n→∞ n→∞
R∈Cn
dn
Note that Q(R0 ) is the union of 2 oblique rectangles Q(R) with R ∈ Λn . All
these oblique rectangles Q(R), R ∈ Λn , are translates of each other. According to
Proposition 2.14 Lebesgue measure is translation invariant and thus
m(Q(R)) = 2−dn m(Q(R0 )).
Hence it follows from the last term in (2.20) that
(2.21) m(B) = lim 2−dn · m(Q(R0 )) · #Cn .
n→∞
Comparing (2.19) and (2.21), one concludes that m(Q(R0 )) = `(R0 ) as desired.
Now assume R0 is a rectangle of the form R0 = [a1 , b1 ] × · · · × [ad , bd ], with
ai < bi for each i. Let x = (a1 , . . . , an ) and, for each n ∈ N, let
1 1
Sn = a1 , 1 + b1 × · · · × ad , 1 + bd ,
n n
2. LEBESGUE MEASURE ON Rd 23
where the first and sixth equalities follow from Theorem 1.13, the second and fifth
equalities follow from Proposition 2.14, the third equality follows from the linearity
of Q and the fourth equality follows from the fact that Sn − x is a rectangle of the
form [0, c1 ) × · · · [0, cd ) (with ci = −ai + (1 + n1 )bi ), which we have already treated.
Now fix A ⊂ Rd . For any ε > 0, there exist rectangles R1 , R2 , . . . such that
A ⊂ ∪∞n=1 Rn and
X∞
`(Rn ) ≤ m∗ (A) + ε.
n=1
Then, since Q(A) ⊂ ∪∞
n=1 Q(Rn ), we have
∞
X ∞
X
m∗ (Q(A)) ≤ m(Q(Rn )) = `(Rn ) ≤ m∗ (A) + ε.
n=1 n=1
∗ ∗
This shows that m (Q(A)) ≤ m (A), and the reverse inequality follows by symme-
try and the fact that Q−1 is also orthogonal.
Now assume that A is Lebesgue measurable, then for any set Z ⊂ Ω
m∗ (Z ∩ Q(A)) + m∗ (Z ∩ (QA)c )
= m∗ (Z ∩ Q(A)) + m∗ (Z ∩ Q(Ac ))
(2.22)
= m∗ (Q−1 Z ∩ A) + m∗ (Q−1 Z ∩ Ac )
= m∗ (Q−1 (Z)) = m∗ (Z),
which implies that Q(A) is Lebesgue measurable. The converse follows by symme-
try.
where λ1 , . . . , λd are the necessarily positive eigenvalues of T > T . Define the linear
transformation S by
p p
S = V > DV where D = diag ( λ1 , . . . , λd ).
Then S is symmetric, invertible, and S 2 = V > D2 V = V > ∆V = T > T . Define
W = T S −1 . Since S −1 is symmetric, it follows that
W > W = S −1 T > T S −1 = S −1 S 2 S −1 = I,
and therefore W is orthogonal. Hence T can be written as
T = W S = W V > DV = U DV with U = W V >.
Note that U is orthogonal, while D is diagonal with positive entries.
Step 2. The dilation of a cover of A is a cover of the dilation of A. A dilated
rectangle has an area that is obtained by multiplying the area of the rectangle by
det D. Hence one obtains
m∗ (D(A)) = (det D) m∗ (A).
Step 3. The decomposition of T as in (2.25) and a successive argument involving
Lemma 2.15 and Step 2 shows
since C is a σ-algebra.
Thus in both cases it follows that A ∈ C. Therefore M ⊂ C.
Hence in each of the above cases there exists an open set O ⊂ Rd such that
A ⊂ O and m∗ (O \ A) < ε, so that (2) is satisfied.
(2) ⇒ (1) For every n ∈ N there
T∞exists an open set On ⊂ Rd such that A ⊂ On
∗
and m (On \ A) < 1/n. With O = n=1 On one has O Lebesgue measurable, while
A ⊂ O. It follows for every n ∈ N that
∞
!
\ 1
m∗ (O \ A) = m∗ (On \ A) ≤ m∗ (On \ A) ≤ .
n=1
n
Proof. Only the implication (⇒) needs a proof. If A measurable then there
exists an open set O ⊂ Rd and a closed set F ⊂ Rd such that F ⊂ A ⊂ O and
m(O \ F ) < ε. Since A is bounded also F is bounded and hence compact; moreover
there exists an open ball B around 0 such that A ⊂ B. Then O1 = O ∩ B is a
bounded open set and A ⊂ O1 , while O1 \ F ⊂ O \ F , so that m(O1 \ F ) < ε.
4. Lebesgue-Stieltjes measure on R
The construction of the Lebesgue measure on Rd has an analog on R when the
outer measure is defined in terms of a non-decreasing function.
Proposition 2.23. Let F : (a, b) → R be a non-decreasing real-valued function.
Then for x ∈ (a, b) the limits F (x−) and F (x+) exist and
sup F (t) = F (x−) ≤ F (x) ≤ F (x+) = inf F (t).
a<t<x x<t<b
Moreover, the function F has at most countably many discontinuities x ∈ (a, b),
which are jump discontinuities: F (x−) < F (x+). Furthermore
a<x<y<b ⇒ F (x+) ≤ F (y−).
Proof. Let x ∈ (a, b), then the set {F (t) : a < t < x} is bounded above by
F (x). Hence
A = sup F (t) ≤ F (x).
a<t<x
Now choose ε > 0. Then by the supremum property there exists δ > 0 such that
a<x−δ <x and A − ε < F (x − δ) ≤ A.
The monotonicity property of F shows that for all t with x − δ < t < x
A − ε ≤ F (x − δ) ≤ F (t) ≤ A.
Hence it has been shown that A = F (x−). A similar reasoning holds for F (x+).
Now let a < x < y < b, then the monotonicity of F in fact gives
F (x+) = inf F (t) = inf F (t), F (y−) = sup F (t) = sup F (t).
x<t<b x<t<y a<t<y x<t<y
(1) ∅ ∈ S;
(2) A, B ∈ S ⇒ A ∩ B ∈SS;
n
(3) A, B ∈ S ⇒ A \ B = k=1 Ck , where C1 , . . . , Cn ∈ S are pairwise disjoint.
Theorem 2.39 (Carathéodory extension theorem). Let S be a semi-ring on Ω
and let µ : S → [0, ∞] be a pre-measure. Then µ has an extension to a measure
on σ(S), the σ-algebra generated by S. If the pre-measure on S is σ-finite, then its
extension to σ(S) is unique.
Let us now show how both the Lebesgue measure and the Lebesgue-Stieltjes
measure can be obtained from this theorem.
Example 2.40 (Lebesgue measure). The collection of rectangles of Rd of the
form (a1 , b1 ] × · · · × (ad , bd ] forms a semi-ring S as is easily verified (we need to take
semi-open intervals here, in contrast to what was done in the course of the chapter,
in order to guarantee that the difference of two members of S produces a disjoint
union of members of S). Denote the function which assigns the volume `(R) to each
block R ∈ S by µ. Then the function µ is a pre-measure.
Example 2.41 (Lebesgue-Stieltjes measure). Let the collection of subsets S of
R be defined by
S = {(a, b] : −∞ < a ≤ b < ∞}.
Then S is a semi-ring which generates the Borel σ-algebra on R. Let F : R → R be
a non-decreasing and right-continuous function and define µ : S → [0, ∞] by
µ((a, b]) = F (b) − F (a).
Then µ is a pre-measure on S etc.
We will not give the full proof of Theorem 2.39, as much of it would be a rep-
etition of the work we have done in constructing the Lebesgue measure on Rd . We
will simply indicate the steps involved, stating the corresponding results without
proof.
Step 1b: From a pre-measure on a ring to an outer measure on the power set
The results in this step are sometimes stated only for the case that R is an
algebra.
Definition 2.45. Let R be a ring on Ω and let µ : R → [0, ∞] be a pre-measure
on R. Define for any set E ⊂ Ω
( ∞ ∞
)
X [
(2.33) µ∗ (E) = inf µ(An ) : E ⊂ An , A n ∈ R .
n=1 n=1
∗
Proposition 2.46. The set function µ : P (Ω) → [0, ∞] in (2.33) is an outer
measure on Ω with the property
µ∗ (A) = µ(A), A ∈ R.
Measurability of functions
1. Measurable mappings
Definition 3.1. Let (Ω, A) and (Ω0 , A0 ) be measurable spaces. The mapping
f : Ω → Ω0 is measurable, or, more precisely, (A, A0 )-measurable, if f −1 (A0 ) ∈ A
for any A0 ∈ A0 .
Let the mappings f : (Ω, A) → (Ω0 , A0 ) and g : (Ω0 , A0 ) → (Ω00 , A00 ) be measur-
able. Then it is clear from the definition that also the composition
g ◦ f : (Ω, A) → (Ω00 , A00 )
is measurable.
Theorem 3.2. Let (Ω, A) and (Ω0 , A0 ) be measurable spaces and let E0 be a
generator for A0 . Then the following statements are equivalent:
(1) the mapping f : Ω → Ω0 is measurable;
(2) f −1 (E 0 ) ∈ A for any E 0 ∈ E0 .
Proof. (1) ⇒ (2) This is clear.
(2) ⇒ (1) Define the collection F0 ⊂ A0 by
F0 = { A0 ∈ A0 : f −1 (A0 ) ∈ A }.
Clearly Ω0 ∈ F0 , since f −1 (Ω0 ) = Ω. If A0 ∈ F0 , then
f −1 (Ω0 \ A0 ) = Ω \ f −1 (A0 ) ∈ A.
If A0n ∈ F0 , then
∞ ∞
!
[ [
f −1
A0n = f −1 (A0n ) ∈ A.
n=1 n=1
Hence F0 is a σ-algebra in Ω0 . Due to the assumption E0 ⊂ F0 , then also
A0 = σ(E0 ) ⊂ F0 ;
so that A0 = F0 . Hence f is measurable.
35
36 3. MEASURABILITY OF FUNCTIONS
2. Measurable functions
A real-valued function f : Ω → R has been called measurable in the context
of a σ-algebra A on Ω and the Borel σ-algebra B on R; cf. Definition 3.4. It
will be necessary to widen this definition to include extended real-valued functions
f : Ω → R. Here R denotes the extended real line which is provided with the
usual ordering and topology. The Borel σ-algebra B(R) is then defined as the
smallest σ-algebra containing all open sets in R. The following notation B = B(R)
will be used throughout. If (Ω, A) is a measurable space, the extended real-valued
function f on Ω is measurable if it is measurable as an extended real-valued function
f : (Ω, A) → (R, B).
Proposition 3.8. Let f be an extended real-valued function from the measur-
able space (Ω, A) to (R, B). Then the following statements are equivalent:
(1) f is measurable;
(2) the set f −1 (O) belongs to A for each open set O ⊂ R.
The Borel σ-algebra B is generated by half-open and half-closed intervals in R.
In the present context the following result is equivalent to Proposition 3.8.
Corollary 3.9. The extended real-valued function f : (Ω, A) → (R, B) is mea-
surable if and only if for each c ∈ R one, and hence all, of the following assertions
are satisfied:
(1) {ω ∈ Ω : f (ω) < c} ∈ A;
(2) {ω ∈ Ω : f (ω) ≤ c} ∈ A;
(3) {ω ∈ Ω : f (ω) > c} ∈ A;
(4) {ω ∈ Ω : f (ω) ≥ c} ∈ A.
Proposition 3.10. Let f, g : (Ω, A) → (R, B) be measurable extended real-
valued functions and let A ∈ A. Then the sets
{ω ∈ Ω : f (ω) < g(ω)}, {ω ∈ Ω : f (ω) ≤ g(ω)}, {ω ∈ Ω : f (ω) = g(ω)},
are measurable.
Proof. Note that f (ω) < g(ω) if and only if there is a number r ∈ Q with
f (ω) < r < g(ω). This observation shows the identity
[
{ω ∈ Ω : f (ω) < g(ω)} = {ω ∈ Ω : f (ω) < r} ∩ {ω ∈ Ω : r < g(ω)},
r∈Q
By the previous reasoning concerning the supremum the functions gk are mea-
surable, and then the same reasoning for the infimum show that lim supn∈N fn is
measurable. The result for lim inf n∈N fn follows in an analogous way.
Hence f + g is measurable.
If h is measurable, then also h2 is measurable. To see this observe that for
c≤0
{ω ∈ Ω : h(ω)2 < c} = ∅,
while for c > 0
√ √
{ω ∈ Ω : h(ω)2 < c} = {ω ∈ Ω : − c < h(ω) < c}.
{ω ∈ Ω : f /g(ω) < c}
is the union of
and
{ω ∈ Ω : g(ω) < 0} ∩ {ω ∈ Ω : f (ω) > cg(ω)}.
Hence f /g is measurable on its domain of definition.
are measurable, since constant functions are measurable. Hence with f also |f | =
f + + f − is measurable. Conversely, if f + and f − are measurable, then also f =
f + − f − is measurable.
Remark 3.17. Many of the notions introduced so far play a role in probability
theory. Here is a very brief outline. Let (Ω, A, µ) be a probability space, i.e., (Ω, A, µ)
is a σ-algebra and µ(Ω) = 1. The measurable sets A ∈ A are called events in the
context of probability. The probability of an event A ∈ A is defined by µ(A) and
the conditional probability of A ∈ A given B ∈ A is defined by µ(A ∩ B)/µ(B). The
events A, B ∈ A are said to be independent if µ(A ∩ B) = µ(A)µ(B). A random
variable is just a function f : (Ω, A) → (R, B) which is measurable.
40 3. MEASURABILITY OF FUNCTIONS
Example 3.20. Let A ⊂ Ω be a measurable set and define the extended real-
valued function ∞A from Ω to R by
(
∞, ω ∈ A,
(3.1) ∞A (ω) =
0, ω 6∈ A.
Then clearly ∞A is measurable and the sequence fn = n1A is a monotone sequence
of simple functions such that fn → ∞A pointwise.
Example 3.22. Let (Ω, A, µ) be a measure space and let f and g be extended
real-valued functions from Ω to R. Then f = g a.e. means that there is a set A ∈ A
such that
(3.2) f (ω) = g(ω) for all ω ∈ A,
while µ(Ac ) = 0.
Almost everywhere equality of functions gives rise to an equivalence relation.
Lemma 3.23. Let (Ω, A, µ) be a measure space. Let f , g, and h be extended
real-valued functions from Ω to R, such that f = g a.e. and g = h a.e. Then f = h
a.e.
Proof. Since f = g a.e. and g = h a.e., there are sets A, B ∈ A such that
f (ω) = g(ω) for ω ∈ A, and g(ω) = h(ω) for ω ∈ B,
while µ(Ac ) = µ(B c ) = 0. Define C = A ∩ B, so that it is clear that C ∈ A and
µ(C c ) = µ((A ∩ B)c ) = 0. Observe that
f (ω) = g(ω) = h(ω) for ω ∈ C,
so that f = h a.e.
Note that the set A in (3.2) need not be the maximal subset of Ω on which
the functions f and g have equal values. However, if the functions f and g are
measurable one has the following result.
Lemma 3.24. Let (Ω, A, µ) be a measure space and let f and g be measurable
extended real-valued functions from Ω to R. Then
f = g a.e. ⇔ µ({ω ∈ Ω : f (ω) 6= g(ω)}) = 0.
Proof. Define the set N ⊂ Ω by
N = {ω ∈ Ω : f (ω) 6= g(ω)}.
It is clear that the sets N and N c belong to A.
(⇒) Assume that f = g a.e. Hence there exists A ∈ A such that f (ω) = g(ω)
for all ω ∈ A, while µ(Ac ) = 0. Note that N ⊂ Ac and hence µ(N ) = 0.
(⇐) For all ω ∈ N c one has f (ω) = g(ω). Since (N c )c = N and µ(N ) = 0 it
therefore follows that f = g a.e.
Let (Ω, A, µ) be a measure space. A sequence of extended real-valued functions
fn from Ω to R converges pointwise almost everywhere to an extended real-valued
function f from Ω to R, i.e.,
lim fn = f a.e.,
n→∞
if there exists a set A ∈ A such that
fn (ω) → f (ω) for all ω ∈ A,
c
while µ(A ) = 0.
Lemma 3.25. Let (Ω, A, µ) be a measure space and let fn and f be measurable
extended real-valued functions from Ω to R, Then
lim fn = f a.e. ⇔ µ({ω ∈ Ω : lim fn (ω) 6= f (ω)}) = 0.
n→∞ n→∞
4. PROPERTIES WHICH ARE VALID ALMOST EVERYWHERE 43
Proof. Let A ⊂ Ω be a set such that fn (ω) → f (ω) for all ω ∈ A, while
µ(Ac ) = 0. Then fn 1A converges to f 1A . Note that fn 1A is measurable so that
the pointwise limit f 1A is measurable. Note that f 1A (ω) = f (ω) for ω ∈ A while
µ(Ac ) = 0. Hence f is measurable by Theorem 3.26.
Let (Ω, A, µ) be a measure space and let (Ω, Ā, µ̄) be its completion. Extended
real-valued functions on Ω which are measurable with respect to A are also mea-
surable with respect to Ā. If an extended real-valued function on Ω is measurable
with respect to Ā, then there is an extended real-valued function measurable with
respect to A which comes very close to the original function. The closeness is mea-
sured in terms of the complete measure. This is where the notion “equality almost
everywhere” comes up naturally.
Theorem 3.28. Let (Ω, A, µ) be a measure space and let (Ω, Ā, µ̄) be its com-
pletion.
(1) Let f : (Ω, A) → (R, B) be a measurable extended real-valued function.
Then f is also measurable with respect to (Ω, Ā).
44 3. MEASURABILITY OF FUNCTIONS
Integrability of functions
Example 4.2. Let f : R → R be the characteristic function Rof the set Q and
let m be the Lebesgue measure. Then f is a simple function and R f dm = 0.
Proposition 4.3. Let (Ω, A, µ) be a measure space. Let f, g : (Ω, A) → (R, B)
be nonnegative simple functions and α ≥ 0. Then αf and f +g are also nonnegative
simple functions and
R R
(1) RΩ αf dµ = α ΩRf dµ, α ≥R 0;
(2) Ω (f + g)Rdµ = Ω f Rdµ + Ω g dµ;
(3) f ≤ g ⇒ Ω f dµ ≤ Ω g dµ.
Proof. Assume in general that
r
X
f= αk 1Ak with Ak = f −1 (αk ), k = 1, . . . , r,
k=1
45
46 4. INTEGRABILITY OF FUNCTIONS
and
s
X
g= βl 1 B l with Bl = g −1 (βl ), l = 1, . . . , s.
l=1
(1) If α = 0 the statement is trivial. If α > 0 note that αf is simple with values
αα1 , . . . , ααr ∈ R, and
Ak = (αf )−1 (ααk ), k = 1, . . . , r,
hence Z r
X r
X Z
αf dµ = ααk µ(Ak ) = α αk µ(Ak ) = α f dµ.
Ω k=1 k=1 Ω
Let γ be a value of f + g and let Jγ be the set of all pairs (k, l) with γ = αk + βl .
Then
[ X
Dγ := (f + g)−1 (γ) = Ck,l with µ(Dγ ) = µ(Ck,l ).
(k,l)∈Jγ (k,l)∈Jγ
Definition 4.5. Let (Ω, A, µ) be a measure space and let the extended real-
valued function
R f : (Ω, A) → (R, B) be measurable and nonnegative. Then the
integral Ω f dµ is defined by
Z Z
f dµ = sup g dµ : 0 ≤ g ≤ f, g simple ∈ [0, ∞].
Ω Ω
Note that if f in this definition is a simple function, then the definition agrees
with Definition 4.1. The following result is preliminary version of what will be
called the monotone convergence theorem.
Lemma 4.6. Let (Ω, A, µ) be a measure space and let the extended real-valued
function f : (Ω, A) → (R, B) be measurable and nonnegative. Let fn : (Ω, A) →
(R, B) be a sequence of nonnegative simple functions such that fn ↑ f . Then
Z Z
fn dµ ↑ f dµ.
Ω Ω
Therefore, let the nonnegative simple function g with g ≤ f have the representation
r
X
g= γk 1Ck , Ck = g −1 (γk ),
k=1
where the sets Ck are disjoint and measurable. Choose 0 < ε < 1 and define
Ωn = {ω ∈ Ω : fn (ω) ≥ (1 − ε)g(ω)}.
48 4. INTEGRABILITY OF FUNCTIONS
Then clearly the S sets Ωn are measurable, and fn ≤ fn+1 implies Ωn ⊂ Ωn+1 .
∞
Furthermore Ω = n=1 Ωn . To see this let ω ∈ Ω and assume that for all n one has
ω 6∈ Ωn , so that for all n ∈ N one has
fn (ω) < (1 − ε)g(ω).
Then it follows that fn (ω) < (1 − ε)f (ω), which contradicts fn (ω) ↑ f (ω). Thus
one may conclude Ωn ↑ Ω. Now observe that
Z Z Z
fn dµ ≥ fn 1Ωn dµ ≥ (1 − ε) g1Ωn dµ
Ω Ω Ω
Z r
!
X
= (1 − ε) γk 1Ck ∩Ωn dµ
Ω k=1
r
X
= (1 − ε) γk µ(Ck ∩ Ωn ),
k=1
However, f and fn are also measurable with respect to Ā; cf. Theorem 3.28.
Therefore an application of Lemma 4.6 in the context of (Ω, Ā, µ̄) shows that
Z Z
fn dµ̄ ↑ f dµ̄.
Ω Ω
1. INTEGRALS OF NONNEGATIVE MEASURABLE FUNCTIONS 49
2. Integrable functions
Let the extended real-valued function f : (Ω, A) → (R, B) be measurable. Then
the functions f + and f − are nonnegative measurable functions and
(4.3) f = f + − f −, |f | = f + + f − .
Definition 4.13. Let (Ω, A, µ) be a measure space and let the extended real-
valued function f : (Ω, A) → (R, B) be measurable. Then the integral Ω f dµ is
R
defined by Z Z Z
f dµ = f + dµ − f − dµ,
Ω Ω Ω
if at least one of the integrals in the right-hand side is finite. The function f is said
to be integrable if both integrals in the right-hand side are finite, or, equivalently, if
Z
|f | dµ < ∞.
Ω
2. INTEGRABLE FUNCTIONS 51
Proposition 4.14. Let (Ω, A, µ) be a measure space, let the extended real-
valued functions f, g : (Ω, A) → (R, B) be integrable, and let α ∈ R. Then αf and
f + g (when defined) are also integrable and
R R
(1) RΩ αf dµ = α ΩRf dµ, α ∈RR;
(2) Ω (f + g)Rdµ = Ω f Rdµ + Ω f dµ;
(3) fR≤ g ⇒ ΩRf dµ ≤ Ω g dµ;
(4) | Ω f dµ| ≤ Ω |f | dµ.
Proof. Items (1), (2), and (3) follow from Definition 4.13 and Proposition 4.8
as applied to f + , g + , f − , and g − . For (4) observe that
Z Z Z Z Z Z
+ − + −
f dµ = f dµ − f dµ ≤ f dµ + f dµ = |f | dµ,
Ω Ω Ω Ω Ω Ω
due to (4.3).
Proposition 4.15. Let (Ω, A, µ) be a measure space and let the extended real-
valued functions f, g : (Ω, A) → (R, B) beRmeasurable.
R Assume that f integrable
and that g = f a.e. Then g integrable and Ω g dµ = Ω f dµ.
Proof. Assume that f is integrable. It follows from g = f a.e. that
g + = f + a.e. and g − = f − a.e.
By Corollary 4.11 it follows that
Z Z Z Z
g + dµ = f + dµ and g − dµ = f − dµ,
Ω Ω Ω Ω
and each of the integrals is finite. Hence g is integrable and the result follows.
Corollary 4.16. Let (Ω, A, µ) be a measure space and let the extended real-
valued function f : (Ω, A) → (R, B) be integrable. Then
(4.4) µ({ω ∈ Ω : |f (ω)| = ∞}) = 0,
and there exists an integrable real-valued function g : (Ω, A) → (R, B) such that
Z Z
(4.5) g = f a.e. and g dµ = f dµ.
Ω Ω
Proof. Let N = {ω ∈ Ω : |f (ω)| = ∞}, then it follows from Corollary 4.12
that µ(N ) = 0, so that (4.4) holds. Now define the function g by
g = f 1Ω\N .
Then g is measurable
R and gR= f a.e. Hence it follows from Proposition 4.15 that g
is integrable and Ω g dµ = Ω g dµ.
In Theorem 3.28 the measurability of a function was studied in terms of the
completion of a measure space. A similar question is now treated for the integra-
bility of a function. Corollary 4.7 will play an important role.
Theorem 4.17. Let (Ω, A, µ) be a measure space and let (Ω, Ā, µ̄) be its com-
pletion.
(1) Let the extended real-valued function f : (Ω, A, µ) → (R, B) be integrable.
Then f is also integrable with respect to (Ω, Ā, µ̄), and
Z Z
(4.6) f dµ = f dµ̄.
Ω Ω
52 4. INTEGRABILITY OF FUNCTIONS
(2) Let the extended real-valued function f : (Ω, Ā) → (R, B) be integrable
with respect to µ̄. Then the the extended real-valued function g : (Ω, A) →
(R, B) in Theorem 3.28 is integrable with respect to µ, and
Z Z
(4.7) g dµ = f dµ̄.
Ω Ω
The integrability of g with respect to µ̄ implies that g + and g + are integrable with
respect to µ̄. Then the above identies show that g + and g + are integrable with
respect to µ. Therefore it follows that g is integrable with respect to µ and that
Z Z
g dµ = g dµ̄.
Ω Ω
Definition 4.18. Let (Ω, A, µ) be a measure space and let the extended real-
valued function f : (Ω, A) → (C, B) be measurable. Then the function f is said to
be integrableR if the real part f1 and and the imaginary part f2 are integrable and
the integral Ω f dµ is defined as
Z Z Z
f dµ = f1 dµ + i f2 dµ, f = f1 + if2 .
Ω Ω Ω
Note that the properties (1), (2), and (4) in Proposition 4.14 remain valid with the
understanding that now in (1) α ∈ C. To see that the inequality in (4) remains
valid, let
Z Z
iθ
f dµ = e f dµ,
Ω Ω
3. INTEGRATION WITH RESPECT TO LEBESGUE MEASURE 53
see Theorem 4.17 and Theorem 3.28. In the following a function f : Rd → R will
be called Lebesgue integrable over Rd if it is integrable with respect to (Rd , Md , m)
or with respect to (Rd , Bd , mB ). Except for sets of Lebesgue measure 0 there is no
difference in the two approaches and no distinction will be made between m and
mB when there is no need. Note that for functions f : Rd → R which are defined
only on a subset of Rd similar considerations are valid when the notion of trace
σ-algebra is introduced; cf. Proposition 1.39.
Any continuous function f : Rd → R is Borel measurable (and hence Lebesgue
measurable). If the continuous function f : Rd → R has compact support, then f
is clearly Borel measurable. Furthermore the function f is integrable. To see this
let the support K ⊂ Rd of f be compact, then
Z Z Z
|f | dm = |f | dm ≤ M dm = M m(K) < ∞,
Rd K K
Proof. Since the function f is continuous, it is bounded on [a, b]. Hence the
Lebesgue integrability is clear.
4. CONVERGENCE THEOREMS 55
(1) Let x ∈ (a, b). Then for every ε > 0 there exists δ > 0 so that |t − x| < δ
implies that |f (t) − f (x)| < ε. For 0 < h < δ consider
Z x+h Z x !
F (x + h) − F (x) 1
− f (x) = f (t) dt − f (t) dt − f (x)
h h a a
1 x+h
Z
= [f (t) − f (x)] dt,
h x
since [a, x] and (x, x + h] are disjoint measurable sets. If −δ < h < 0, the last
relation holds analogously. Note that
Z x+h Z x+h
[f (t) − f (x)] dt ≤ |f (t) − f (x)| dt ≤ |h|ε.
x x
Consequently,
F (x + h) − F (x)
− f (x) ≤ ε.
h
Hence F 0 (x) = f (x). A similar argument holds for the endpoints.
R x (2) Let F be continuously differentiable on [a, b] with F 0 = f . Define G(x) =
0
a
f (t) dt. Then by (1) G = f , so that G − F is constant on [a, b]. In particular,
Rb
F (b) − F (a) = G(b) − G(a) = a f (t) dt.
The lemma shows that as far as the Lebesgue integral of a continuous function
is concerned one obtains the value via the difference of a primitive at the end points.
Some more rules of calculation are as before.
Corollary 4.20. Let the real-valued functions h, k : [a, b] → R be continuously
differentiable, then
Z b Z b
0
h(t)k (t) dt = h(b)k(b) − h(a)k(a) − h0 (t)k(t) dt.
a a
4. Convergence theorems
Let (Ω, A, µ) be a measure space. Let fn be a sequence of integrable functions
which
R converges
R pointwise to a measurable function f . In order to conclude that
Ω
fn dµ → Ω
f dµ, i.e., to interchange limit and integration, one needs conditions
to ensure the convergence of the integrals and the limiting identity.
56 4. INTEGRABILITY OF FUNCTIONS
By means of the monotone convergence theorem, the result in Lemma 4.19 can
be extended to the situation where I ⊂ R is a half-open and half-closed interval
and f : I → R is nonnegative and continuous.
Lemma 4.27 (Lebesgue integration on the real line). Let f : (a, b] → R or
f : [a, b) → R be nonnegative and continuous with primitive F . Then
Z Z
f dm = lim [F (b) − F (α)] or f dm = lim [F (β) − F (a)],
(a,b] α&a [a,b) β%b
respectively.
Proof. It has been shown that for a ≤ β < b one has
Z
f dm = F (β) − F (a).
[a,β]
P∞
Then the series n=1 fn converges almost everyhere to an integrable extended real-
valued function and
∞ ∞ Z
Z !
X X
(4.12) fn dµ = fn dµ.
Ω n=1 n=1 Ω
Proof. In order to apply the dominated convergence theorem, define the fol-
lowing functions
N
X ∞
X
(4.13) gN = fn , h= |fn |.
n=1 n=1
Then clearly |gN | ≤ h. By Proposition 4.32 and the assumption (4.11) one has
Z ∞ Z
X
h dµ = |fn | dµ < ∞,
Ω n=1 Ω
cf. Proposition 4.32. However one can see, for instance via the calculus of residues,
that Z R
sin t π
lim dt = .
R→∞ 0 t 2
Note that the mere existence of the limit is easily deduced by integration by parts
on the interval [1, R] and the dominated convergence theorem.
Proof. From (i) and (ii) one sees that the function F is continuous at x0 ;
cf. Theorem 4.37. In order to show that F is differentiable at x0 , let (hn ) be any
sequence in R \ {0} with hn → 0, and define
f (x0 + hn , ω) − f (x0 , ω)
fn (ω) = , f0 (ω) = (∂/∂x)|x0 f (x, ω), ω ∈ Ω.
hn
Then (ii) implies with the mean value theorem that for almost all ω ∈ Ω there
exists a number θn (ω) with |θn (ω)| ≤ |hn | such that
fn (ω) = (∂/∂x)|x0 +θn (ω) f (x, ω).
By (i) the functions fn are integrable, and by (iii) the functions fn are dominated by
the integrable function g for almost all ω ∈ Ω. Hence by the dominated convergence
theorem Z Z
0
F (x0 ) = lim fn dµ = f0 (ω) dµ.
n→∞ Ω Ω R
Hence F is differentiable at x0 and its derivative is given by Ω f0 dµ.
(2) The function f is integrable with respect to ν if and only if the function
f ◦ ϕ is integrable with respect to µ, in which case (4.17) holds.
Proof. Since ϕ is measurable the expression in (4.16) is well-defined. Clearly
ν(B) ≥ 0 for all B ∈ A0 and ν(∅) = 0. So it remains to check the countable
64 4. INTEGRABILITY OF FUNCTIONS
Product measures
The area of a rectangle in the plane is the product of the length of the sides.
This is the basis for the introduction of two-dimensional Lebesgue measure. Like-
wise the double integral of a continuous function over a rectangle can be computed
as two iterated one-dimensional integrals. The question is now if two-dimensional
Lebesgue measure can be seen as the product of one-dimensional Lebesgue mea-
sures. In a more general context, it can be shown that the product of two measure
spaces provides a product measure, so that an integral of a function on the product
space can be computed as iterated integrals over the individual components of the
product space. However it turns out that the product measure is in general not
complete. This chapter gives a concise introduction to product measures and their
completions.
65
66 5. PRODUCT MEASURES
∞
[
Aω1 = (An )ω1 ∈ A2 .
n=1
Next assume that µ2 is σ-finite. Then there exists a sequence Ω2,n ∈ A2 such
that
µ2 (Ω2,n ) < ∞, Ω2,n ↑ Ω2 .
Recall from Proposition 1.18 that
µ2,n (A2 ) = µ2 (A2 ∩ Ω2,n ), A2 ∈ A2 ,
defines a finite measure µ2,n : A2 → [0, ∞] on Ω2 with the property
µ2,n (A2 ) ↑ µ2 (A2 ), A2 ∈ A2 .
By what has just been shown it follows that for every A ∈ A1 ⊗ A2 the function
ω1 7→ µ2,n (Aω1 ),
is A1 -measurable. Now observe that
µ2,n (Aω1 ) ↑ µ2 (Aω1 ),
so that the function in the right-hand side, as the pointwise limit of A1 -measurable
functions, is A1 -measurable itself.
The following main result gives the definition of the product measure and shows
in what sense it is unique. In fact the theorem itself is the first form of the Fubini-
Tonelli theorems.
Theorem 5.7 (Product measure). Let (Ω1 , A1 , µ1 ) and (Ω2 , A2 , µ2 ) be σ-finite
measure spaces and let (Ω1 × Ω2 , A1 ⊗ A2 ) be the corresponding measurable product
space. Then
Z Z
(5.6) µ(A) = µ2 (Aω1 )µ1 (dω1 ) = µ1 (Aω2 )µ2 (dω2 ), A ∈ A1 ⊗ A1 ,
Ω1 Ω2
2. Fubini-Tonelli theorems
The unique product measure as given in Theorem 5.7 is the basis for the Fubini-
Tonelli theorems. The key ingredient is the formula in (5.6) for a measurable set
A ∈ A1 ⊗ A1 , which can be rewritten as an identity with iterated integrals:
Z Z Z
1A d(µ1 ⊗ µ2 ) = 1Aω1 d µ2 µ1 (dω1 )
Ω1 ×Ω2 Ω Ω
Z 1 Z 2
= 1A 2 d µ1 µ2 (dω2 ),
ω
Ω2 Ω1
involving sections of sets. From now on these identities will be used in the following
equivalent form involving sections of functions:
Z Z Z
1A d(µ1 ⊗ µ2 ) = (1A )ω1 dµ2 dµ1 (ω1 )
Ω1 ×Ω2 Ω Ω
(5.9) Z 1 Z 2
ω2
= (1A ) dµ1 dµ2 (ω2 ),
Ω2 Ω1
cf. (5.3). Recall that any measurable function f : (Ω1 × Ω2 , A1 ⊗ A2 ) → (R, B) has
measurable sections; cf. Proposition 5.5.
Theorem 5.10 (Fubini-Tonelli). Let (Ω1 , A1 , µ1 ) and (Ω2 , A2 , µ2 ) be σ-finite
measure spaces. Assume that f : (Ω1 × Ω2 , A1 ⊗ A2 ) → (R, B) is a measurable
nonnegative extended real-valued function. Then
(1) the function ω1 7→ Ω2 fω1 dµ2 is nonnegative and A1 -measurable;
R
Moreover,
Z Z Z
f d(µ1 ⊗ µ2 ) = fω1 dµ2 dµ1 (ω1 )
Ω1 ×Ω2 Ω Ω
(5.10) Z 1 Z 2
= f ω2 dµ1 dµ2 (ω2 ),
Ω2 Ω1
which proves the first equality in (5.10); the other equality is clear by symmetry.
It is clear from the pointwise limiting relation fn ↑ f that also (fn )ω1 ↑ fω1 for
any ω1 ∈ Ω1 . By Proposition 5.5 sections of measurable functions are measurable,
2. FUBINI-TONELLI THEOREMS 71
again by Theorem 4.26. The final step is integration of the last result (5.12) over
Ω1 , which is allowed when the functions in the left-hand side and in the right-hand
side are measurable in ω1 . To see that these functions are measurable observe that
the identity
Z Z
(1A )ω1 dµ2 = 1Aω1 dµ2 = µ2 (Aω1 )
Ω2 Ω2
and
Z Z
ω2
|f | dµ1 dµ2 (ω2 ).
Ω2 Ω1
and
Z Z Z
(5.15) f − d(µ1 ⊗ µ2 ) = (f − )ω1 dµ2 dµ1 (ω1 ).
Ω1 ×Ω2 Ω1 Ω2
72 5. PRODUCT MEASURES
where in the penultimate step use has been made of (5.4). Likewise one finds
Z Z Z
|f | d(µ1 ⊗ µ2 ) = |f ω2 | dµ1 dµ2 (ω1 ).
Ω1 ×Ω2 Ω2 Ω1
This completes the proof.
Theorem 5.12 (Fubini-Tonelli). Let (Ω1 , A1 , µ1 ) and (Ω2 , A2 , µ2 ) be σ-finite
measure spaces. Assume that f : (Ω1 × Ω2 , A1 ⊗ A2 ) → (R, B) is an integrable
extended real-valued function. Then the set
(5.16) F1 = { ω1 ∈ Ω1 : fω1 not integrable with respect to µ2 }
has µ1 -measure 0, and the set
(5.17) F2 = { ω2 ∈ Ω2 : f ω2 not integrable with respect to µ1 }
has µ2 -measure 0. Furthermore
Z Z Z
f d(µ1 ⊗ µ2 ) = fω1 dµ2 dµ1 (ω1 )
Ω1 ×Ω2 Ω1 \F1 Ω2
Z Z
= f ω2 dµ1 dµ2 (ω2 ).
Ω2 \F2 Ω1
has µ1 -measure 0. Due to (5.4) it is clear that the set in (5.18) is equal to the set
F1 in (5.16).
By definition the integral of the integrable function f is given by
Z Z Z
f d(µ1 ⊗ µ2 ) = f + d(µ1 ⊗ µ2 ) − f − d(µ1 ⊗ µ2 ),
Ω1 ×Ω2 Ω1 ×Ω2 Ω1 ×Ω2
or, equivalently, to
Z Z Z
(f + )ω1 dµ2 − (f − )ω1 dµ2 dµ1 (ω1 ),
Ω1 \F1 Ω2 Ω2
Since each of the integrals in the right-hand side is finite for ω1 ∈ Ω1 \ F1 , the
right-hand side is equal by definition to
Z
fω1 dµ2 .
Ω2
This proves the first identity in the theorem. The proof of the second identity is
similar.
3. Product measures on Rp × Rq
The notions of Borel measurable sets and Lebesgue measurable sets with their
corresponding measures have been developed for the space Rd . Again for conve-
nience the notations Md and Bd will be used for the corresponding σ-algebras; the
Lebesgue measure in these cases is denoted by md . Now the products of these
notions will be considered.
Proposition 5.13 (Product of Borel algebras). The product Bp ⊗ Bq of the
σ-algebras Bp and Bq satisfies
(5.19) Bp ⊗ Bq = Bp+q .
Moreover, the product measure mp ⊗ mq on Bp ⊗ Bq is well-defined and satisfies
(5.20) mp ⊗ mq = mp+q ,
where mp+q is Lebesgue measure on Bp+q .
Proof. First it will be shown that Bp ⊗ Bq ⊂ Bp+q . Let A1 ∈ Bp and
A2 ∈ Bq , and let π1 and π2 be the canonical projections from Rp+q to Rp and
to Rq , respectively. These projections are continuous and thus Borel measurable.
Therefore it follows from
A1 × A2 = (A1 × Rq ) ∩ (Rp × A2 ) = π1−1 (A1 ) ∩ π2−1 (A2 ),
that A1 × A2 is a Borel set in Rp+q . Therefore
Bp ⊗ Bq ⊂ Bp+q .
For the reverse inclusion, recall that Bp+q is generated by intervals of the form
A1 × A2 with A1 ⊂ Rp and A2 ∈ Rq , and note that A1 × A2 ∈ Bp ⊗ Bq . Since
Bp ⊗ Bq is a σ-algebra, one obtains Bp+q ⊂ Bp ⊗ Bq . This establishes (5.19).
To see that the product measure mp ⊗ mq is well-defined on Bp ⊗ Bq it suf-
fices to note that mp and mq are σ-finite measures on the σ-algebras Ap and
74 5. PRODUCT MEASURES
Aq , respectively; cf. Theorem 5.7. Note that on sets of the form A × B where
A = [a1 , b1 ] × · · · × [ap , bp ] and B = [c1 , d1 ] × · · · × [cq , dq ] the measures coincide:
(mp ⊗ mq )(A × B) = mp (A)mq (B) = mp+q (A × B).
Now apply Proposition 1.24 with Corollary 1.33.
Note that with abuse of language these identities are frequently encountered in the
following form:
Z Z Z ! Z Z !
f dxdy = f (x, y) dy dx = f (x, y) dx dy,
Kp ×Kq Kp Kq Kq Kp
which is alright as long as one knows what one is doing. The following result is a
useful variation as one often integrates over a subset A ⊂ Kp × Kq which is Borel
measurable. Recall that
Z Z
f dmp+q = (f 1A ) dmp+q .
A Kp ×Kq
The version of the Fubini-Tonelli theorem in Corollary 5.14 is the one that is
often presented in calculus texts.
In general, even if the repeated integrals of a function f with respect to the mea-
sures mp and mq exist, the integrability of f with respect to the product measure
mp+q is not guaranteed. The following simple example shows this.
Example 5.15. On the square (0, 1) × (0, 1) consider the function
x2 − y 2
f (x, y) = .
(x2 + y 2 )2
Clearly the function f is continuous and hence it is Borel measurable. Provide the
horizontal and vertical axis with the Borel σ-algebra and consider the square with
3. PRODUCT MEASURES ON Rp × Rq 75
the product measure. Observe that the sections of f are also Borel measurable and
that Z 1 Z 1
1 1
f (x, y) dm(x) = − 2 , f (x, y) dm(y) = 2 .
0 y + 1 0 x +1
Hence repeated integration gives
Z 1 Z 1 Z 1 Z 1
π π
f (x, y) dm(y) dm(x) = , f (x, y) dm(x) dm(y) = − ,
0 0 4 0 0 4
so that the order of integration matters. Clearly the above function f is not inte-
grable on the square, i.e.,
Z
|f (x, y)| dm2 (x, y) = ∞,
(0,1)×(0,1)
Since all these sets are open, hence Lebesgue measurable in Rp+q , one obtains
mp+q ((O1 × O2 ) \ (K1 × K2 ))
≤ mp+q ((O1 \ K1 ) × O2 )) + mp+q ((O1 × (O2 \ K2 )))
= mp (O1 \ K1 )mq (O2 ) + mp (O1 )mq (O2 \ K2 )
≤ ε/2 + ε/2 = ε.
where each A1,n and A2,m is bounded and measurable, so that by the earlier argu-
ment
∞
[
A1 × A2 = A1,n × A2,m
n,m=1
p+q
is Lebesgue measurable in R .
Finally, observe that
Bp+q = Bp ⊗ Bq ⊂ Mp ⊗ Mq ⊂ Mp+q .
According to Theorem 2.19, Mp+q is the completion of Bp+q .
To see that the product measure mp ⊗ mq is well-defined on Mp ⊗ Mq it suffices
to note that mp and mq are σ-finite measures on the σ-algebras Mp and Mq ,
respectively; cf. Theorem 5.7. The restriction of mp ⊗ mq to Bp+q = Bp ⊗ Bq
coincides with mp+q . Hence the completion of mp ⊗ mq is mp+q .
In particular, they are simultaneously finite. Now each of these integrals will be
written in terms of the function f . It is clear from (5.23) that
Z Z
|g| d(µ1 ⊗ µ2 ) = |f | dλ.
Ω1 ×Ω2 Ω1 ×Ω2
To treat the remaining integrals observe that the function g satisfies (5.22). Asso-
ciated with the set N ∈ L in (5.22) is the set M ∈ A1 ⊗ A2 in (5.25). By Lemma
5.18 there exist an A1 -measurable set H1 ⊂ Ω1 with µ1 (H1 ) = 0 such that (5.26)
holds, and an A2 -measurable set H2 ⊂ Ω2 with µ2 (H2 ) = 0 such that (5.27) holds.
Hence for each ω1 ∈ Ω1 \ H1 one has the identity
Z Z Z Z
|gω1 | dµ2 = |gω1 | dµ2 + |gω1 | dµ2 = |gω1 | dµ2 ,
Ω2 Mω1 Ω2 \Mω1 Ω2 \Mω1
where the last step follows from (5.26). Note that in the last integral (ω1 , ω2 ) 6∈ M
(otherwise it would follow that ω2 ∈ Mω1 ), so that certainly (ω1 , ω2 ) 6∈ N , see
(5.25). Therefore in the last integral one has g(ω) = f (ω) and gω1 = fω1 . Hence
for each ω1 ∈ Ω1 \ H1
Z Z Z
|gω1 | dµ2 = |fω1 | dµ2 = |fω1 | dµ2 .
Ω2 Ω2 \Mω1 Ω2
Observe that
Z Z Z Z
|gω1 | dµ2 dµ1 (ω1 ) = |gω1 | dµ2 dµ1 (ω1 )
Ω1 Ω2 Ω1 \H1 Ω2
Z Z
= |fω1 | dµ2 dµ1 (ω1 ).
Ω1 \H1 Ω2
As in the proof of Theorem 5.19 for each ω1 ∈ Ω1 \ H1 one has the identity
Z Z Z Z
gω1 dµ2 = gω1 dµ2 + gω1 dµ2 = gω1 dµ2 .
Ω2 Mω1 Ω2 \Mω1 Ω2 \Mω1
Observe that in the last integral (ω1 , ω2 ) 6∈ M (otherwise it would follow that
ω2 ∈ Mω1 ), so that certainly (ω1 , ω2 ) 6∈ N , i.e., f (ω) = g(ω) and gω1 = fω1 .
Now note that the union G1 ∪H1 is A1 -measurable and µ1 (G1 ∪H1 ) = 0. Hence
one obtains from (5.28) and (5.30)
Z Z
f dλ = g d(µ1 ⊗ µ2 )
Ω1 ×Ω2 Ω1 ×Ω2
Z Z
= gω1 dµ2 dµ1 (ω1 )
Ω1 \G1 Ω2
Z Z
= gω1 dµ2 dµ1 (ω1 )
Ω1 \(G1 ∪H1 ) Ω2
Z Z !
= gω1 dµ2 dµ1 (ω1 )
Ω1 \(G1 ∪H1 ) Ω2 \Mω1
Z Z !
= fω1 dµ2 dµ1 (ω1 )
Ω1 \(G1 ∪H1 ) Ω2 \Mω1
Z Z
= fω1 dµ2 dµ1 (ω1 ),
Ω1 \(G1 ∪H1 ) Ω2
where in the last step Proposition 4.15 has been used, so that for ω1 ∈ Ω1 \(G1 ∪H1 )
it follows that fω1 is integrable with respect to µ2 . By taking L1 = G1 ∪ H1 the
proof is complete for one case. The other case follows similarly.
5. Convolutions
The Fubini-Tonelli theorems will be used to derive a few facts concerning con-
volutions. The following general observation is useful.
Lemma 5.21. Let the real-valued functions f, g : Rd → R be Lebesgue measur-
able. Then the real-valued function
(x, y) → f (x)g(y)
from Rd × Rd to R is Lebesgue measurable.
80 5. PRODUCT MEASURES
Proof. Note that the function (x, y) 7→ f (x) is Lebesgue measurable. To see
this observe that for every c ∈ R the set
{(x, y) ∈ Rd × Rd : f (x) ≤ c} = {x ∈ Rd : f (x) ≤ c} × R
is measurable in the product space. Likewise the function (x, y) 7→ g(y) is Lebesgue
measurable. In addition, products of measurable functions are measurable.
Corollary 5.22. Let the real-valued functions f, g : Rd → R be Lebesgue
measurable. Then the real-valued function
(x, y) → f (x − y)g(y)
is Lebesgue measurable on R × Rd .
d
for the case that f is differentiable almost everywhere and f 0 is Lebesgue integrable,
is not valid; but what comes in its place is very rewarding not only from a theoretical
point of view but also from an applied point of view. The results in this chapter are
based on the fact that every monotone function is differentiable almost everywhere
with respect to Lebesgue measure and the derivative is Lebesgue integrable on any
compact interval. The main new notions here are functions of bounded variation
and absolutely continuous functions; later they will find an natural place in the
abstract decomposition theory of ordinary and signed or complex measures. There
are also higher-dimensional versions of the topics in this chapter, but they lead too
far away from the present purposes.
1. Vitali covers
Definition 6.1. Let E ⊂ R be a set. A Vitali cover of E is a collection V of
closed intervals with positive length with the property that for every x ∈ E and for
every η > 0 there is an interval I ∈ V with x ∈ I and m(I) < η.
Lemma 6.2. Let E ⊂ R be a set with m∗ (E) < ∞ and let V be a Vitali cover
of E. For each ε > 0 there are disjoint intervals I1 , . . . , IN ∈ V such that
m∗ (E \ (I1 ∪ · · · ∪ IN )) < ε.
Proof. Choose an open set O ⊂ R such that E ⊂ O and m(O) < ∞; cf. ??.
Observe that the collection W defined by
W = {I ∈ V : I ⊂ O}
is a Vitali cover of E. Choose an interval I1 ∈ W and continue by induction in the
following way. Assume that the disjoint intervals I1 , . . . , In have been chosen, then
either E ⊂ I1 ∪ · · · ∪ In and the selection process stops, or E \ (I1 ∪ . . . In ) 6= ∅.
In the last case let kn be the supremum of the lengths of the intervals in W that
have an empty intersection with I1 ∪ · · · ∪ In . Then choose such an interval In+1
with length greater than kn /2. If this process continues one finds mutually disjoint
closed intervals In , n ∈ N, from W such that
∞ ∞
!
X [
(6.1) m(Ik ) = m Ik ≤ m(O) < ∞,
k=1 k=1
83
84 6. INTEGRATION AND DIFFERENTIATION ON R
so that the sum converges. Now let ε > 0, then there exists N ∈ N such that
X∞
m(Ik ) < ε,
k=N +1
and let Jk be a closed interval with the same center as Ik but with five times its
length. It is now claimed that
∞
[
(6.2) E \ (I1 ∪ · · · ∪ IN ) ⊂ Jk .
k=N +1
that s(m∗ (E) − ε) > 2C, the above estimate leads to a contradiction. Therefore it
follows that m∗ (E) = 0, which implies that the set E is Lebesgue integrable with
measure 0.
(2) Due to (6.3) it suffices to show that the set F defined by
F = {x ∈ (a, b) : D+ f (x) < D+ f (x)}
is Lebesgue measurable with measure 0. For each x ∈ F there exist rx , sx ∈ Q with
D+ f (x) < rx < sx < D+ f (x).
S
Thus it is clear that F ⊂ r,s∈Q Ers , where
Ers = {x ∈ (a, b) : D+ f (x) < r < s < D+ f (x)}.
In order to show that F has measure 0 it is therefore sufficient to show that for
fixed r, s ∈ Q the set E = Ers has measure 0. Choose ε > 0 and let O ⊂ R be an
open set with
E ⊂ O, m(O) < m∗ (E) + ε;
86 6. INTEGRATION AND DIFFERENTIATION ON R
which leads to
M
X N
X
(6.4) s (vj − uj ) < r (yk − xk ).
j=1 k=1
To majorize the sum in the right-hand side of (6.4) observe that by construction
N
X
(6.5) (yk − xk ) ≤ m(O) < m∗ (E) + ε.
k=1
1A function h : (a, b) → R is said to have a strict local maximum at c ∈ (a, b) if there exist
pc , qc ∈ R such that for all x with pc < x < qc and x 6= c one has h(x) < h(c). The mapping
ι : c → (pc , qc ) from the points where h has a strict local maximum to their corresponding
intervals is injective. To see this, let h have local maxima at c and d, c 6= d, and assume that
(pc , qc ) = (pd , qd ); this implies that g(d) < g(c) and g(c) < g(d), a contradiction. Without loss
of generality one may take pc , qc ∈ Q, so that ι is an injective mapping into the countable set
Q × Q. Hence it follows that the number of points where h has a strict local maximum is at most
countable.
88 6. INTEGRATION AND DIFFERENTIATION ON R
≤ f (b) − f (a),
where the last inequality follows from the monotonicity of f . Hence the Lebesgue
integrability of f 0 follows.
Example 6.5. Recall the construction of the Cantor set C in Example 2.13.
The Cantor function will be defined as follows. The set C1 is [0, 1] with (1/3, 2/3)
removed. For x ∈ (1/3, 2/3) define f (x) = 1/2. The set C2 is C1 with (1/9, 2/9)
and (7/9, 8, 9) removed. For x ∈ (1/9, 2/9) define f (x) = 1/4 and for x ∈ (7/9, 8/9)
define f (x) = 3/4. Continue by induction. The set Cn−1 is made up of 2n−1
intervals, each of length 1/3n−1 . The set Cn is obtained from Cn−1 by removing
the middle third of each of these intervals. On these deleted intervals define f by
1 3 5 2n − 1
, , , . . . , ,
2n 2n 2n 2n
respectively. Finally the function f is defined on all of [0, 1] by setting f (0) =
0, f (1) = 1, and for x ∈ C \ {0, 1} by setting f (x) = limt↑x f (t). Then f is
nondecreasing and continuous. It follows from Theorem 6.3 that f is differentiable
almost everywhere.
Actually, the construction of f gives more. Let x ∈ [0, 1] \ C, then x belongs to
an open interval where f is constant. Hence f is differentiable at x and f 0 (x) = 0.
Thus f 0 = 0 almost everywhere, so that
Z 1
f 0 (t) dt = 0 < 1 = f (1) − f (0),
0
3. TERMWISE DIFFERENTIATION OF SERIES OF MONOTONE FUNCTIONS 89
i.e., strict inequality may happen in Theorem 6.4. Finally it is claimed that f is
not differentiable on C. Let x ∈ C and assume that f is differentiable at x. Then
for any sequences αn , βn with
αn ↑ x and βn ↓ x,
one must have that the limit
f (βn ) − f (αn )
lim
n→∞ βn − αn
exists in R, as one can see from
f (βn ) − f (αn ) 0 f (βn ) − f (x) 0
− f (x) = λn − f (x)
βn − α n βn − x
f (αn ) − f (x) 0
+ (1 − λn ) − f (x) ,
αn − x
where λn = (βn − x)/(βn − αn ) is a bounded sequence. However one sees that
for x ∈ C it follows that x ∈ Cn for each n ∈ N. In other words x ∈ [an , bn ]
for some interval [an , bn ] in Cn (which is made up of 2n such intervals), so that
bn − an = 1/3n . Note that by definition
1
f (bn +) − f (an −) = ,
2n
which shows that
f (bn ) − f (an ) 3n
∼ n.
bn − an 2
E ⊂ [a, b] the set of points where f and all fn are differentiable. Due to Theorem
6.3 the set [a, b] \ E has measure 0. For x ∈ E also Rn is differentiable so that
n
X
f 0 (x) = fk0 (x) + Rn0 (x) x ∈ E,
k=1
Definition 6.8. Let (rk ), k ∈ N, be a countable set in an interval (a, b), and
let hk > 0 satisfy
X∞
hk < ∞.
k=1
Define δ by
δ = min{ |c − rk | : 1 ≤ k ≤ N },
so that δ > 0. Let c < x < c + δ. If for some k one has c < rk ≤ x, then clearly
k ≥ N + 1, which leads to
X X
f (x) − f (c) = hk ≤ hk < ε.
k: c<rk ≤x k=N +1
Now let c − δ < x < c. If for some k one has x ≤ rk ≤ c, then clearly k ≥ N + 1,
which leads to
X X
f (c) − f (x) = hn ≤ hk < ε.
n: x<rn ≤c k=N +1
92 6. INTEGRATION AND DIFFERENTIATION ON R
(3) It will be shown that at r` the function f is right continuous and that f
has a jump h` . Choose ε > 0. Let N ∈ N be so large that N > r` and
∞
X
hk < ε.
k=N +1
Define δ by
δ = min{ |r` − rk | : 1 ≤ k ≤ N, k 6= ` },
so that δ > 0. Let r` < x < r` + δ, then as in the proof of (2)
f (x) − f (r` ) < ε.
Now let r` − δ < x < r` then
X
f (r` ) − f (x) = hk
k: x<rk ≤r`
so that X
f (r` ) − f (x) − h` = hk < ε,
k: x<rk <r`
Now Proposition 6.13 and Theorem 6.4 have the following direct consequence.
Corollary 6.14. Let f : [a, b] → R be a real-valued function of bounded
variation. Then f is differentiable almost everywhere and f 0 is Lebesgue integrable.
Definition 6.15. A real-valued function f : [a, b] → R is absolutely continuous
if for every ε > 0 there exists δ > 0 such that for each finite sequence of pairwise
disjoint subintervals [ak , bk ] ⊂ [a, b], 1 ≤ k ≤ n,
Xn Xn
(6.12) (bk − ak ) < δ ⇒ |f (bk ) − f (ak )| < ε.
k=1 k=1
Now observe that this estimate holds for any choice of finite partition Pk of each
interval [ak , bk ]. This leads to
n n s(k)
X X X
Vabkk (f ) = sup |(f (xkj ) − f (xkj−1 )| ≤ ε.
Pk j=1
k=1 k=1
5. DIFFERENTIATION AND INTEGRATION 95
< ε + nδ = 2ε.
Therefore f is absolutely continuous. Hence f is of bounded variation, which implies
that f is differentiable a.e. and that f 0 is integrable.
It remains to show that f 0 = g a.e. This will be done in two steps: first the
case that g is bounded will be treated and then the general case will be reduced to
the bounded case.
First assume that g is integrable and bounded on [a, b]: |g(u)| ≤ K. Then
f (t + h) − f (t) 1 t+h
Z
= g(u) du ≤ K
h h t
and
f (t + h) − f (t)
→ f 0 (t) a.e.
h
5. DIFFERENTIATION AND INTEGRATION 97
For every x ∈ [a, b] it follows from the dominated convergence theorem that
Z x
1 x
Z
0
f (t) dt = lim [f (t + h) − f (t)] dt
a h→0 h a
" Z #
1 x+h 1 a+h
Z
= lim f (t) dt − f (t) dt
h→0 h x h a
= f (x) − f (a),
where in the last step the continuity of f has been used. Hence for every x ∈ [a, b]
it follows that Z x
[f 0 (t) − g(t)] dt = 0,
a
so that g = f 0 a.e.
Next consider the general case where g is integrable. From the definition of the
integral it follows that one may assume that g is nonnegative and integrable. In
this case define the functions gn by
gn (x) = min (n, g(x)).
Then g − gn ≥ 0 and observe that
Z x
fn (x) = [g(t) − gn (t)] dt
a
is a nondecreasing function on [a, b]. Therefore fn is differentiable almost every-
where with nonnegative integrable derivative fn0 . Furthermore observe that the
function gn is bounded so that it follows from the previous step in the proof that
Z x
d
gn (t) dt = gn (x) a.e.
dx a
By rewriting f as follows
Z x
f (x) = γ + fn (x) + gn (t) dt,
a
one sees that for all n ∈ N
f 0 (x) ≥ gn (x) a.e.
which gives
f 0 (x) ≥ g(x) a.e.
Integration of this inequality gives
Z b Z b
f 0 (t) dt ≥ g(t) dt = f (b) − f (a).
a a
Rb
Since by Theorem 6.4 f (b) − f (a) ≥ a f 0 (t) dt one concludes
Z b
[f 0 (t) − g(t)] dt = 0.
a
However f 0 ≥ g a.e. so that one sees that f 0 = g a.e.
The next result is the main theorem of the calculus in the context of Lebesgue
integration: it shows precisely the connection between differentiation and integra-
tion; cf. Lemma 4.19.
98 6. INTEGRATION AND DIFFERENTIATION ON R
Proof. The functions F and G are absolutely continuous on [a, b] and hence
they are bounded: |F (x)| ≤ A and |G(x)| ≤ B. Thus it follows from
|F (x)G(x) − F (y)G(y)| = |F (x)G(x) − F (x)G(y) + F (x)G(y) − F (y)G(y)|
≤ |F (x)||G(x) − G(y)| + |G(y)||F (x) − F (y)|
≤ A|G(x) − G(y)| + B|F (x) − F (y)|,
that the product F G is absolutely continuous. Hence F G is differentiable a.e. and
(F G)0 = F 0 G + F G0 a.e.
0 0
Since F = f and G = g a.e. the statement now follows from Theorem 6.21.
Note that by Theorem 6.21 the previous result could also be stated in the
following more familiar equivalent form.
Corollary 6.25. Let the real-valued functions F, G : [a, b] → R be absolutely
continuous, then
Z b Z b
G(t)F 0 (t) dt + G0 (t)F (t) dt = F (b)G(b) − G(a)F (a).
a a
CHAPTER 7
Integration in Rd
Proof. Let A ∈ A and let g be its indicator function. Then it follows that
g ◦ ϕ−1 = 1ϕ(A) .
101
102 7. INTEGRATION IN Rd
2. Non-linear transformations in Rd
Let X ⊂ Rd and Y ⊂ Rd be open subsets. A bijection ϕ from X onto Y is
called a diffeomorphism if ϕ and its inverse ϕ−1 are both C r , r ≥ 1. It follows from
(ϕ ◦ ϕ−1 )(y) = y, y ∈ Y , and the chain rule that
(7.11) ϕ0 (ϕ−1 (y))(ϕ−1 )0 (y) = I, y ∈ Y,
0 −1 −1 0
where ϕ (ϕ (y)) and (ϕ ) (y) are linear transformations in Rd . A direct conse-
quence of (7.11) is
(7.12) det(ϕ0 ◦ ϕ−1 )(y) det(ϕ−1 )0 (y) = 1, y ∈ Y.
In particular the linear transformations ϕ (ϕ (y)) and (ϕ−1 )0 (y) are invertible for
0 −1
Assume that a is the center of Q and that the length of its sides is r, so that
m(Q) = rd . Since Q is a closed cube in Rd it is compact and since ϕ is continuous,
the image ϕ(Q) is compact and thus Lebesgue measurable. It can be seen from the
generalized mean-value theorem that ϕ(Q) ⊂ Y lies in a cube with center ϕ(a) and
length of its sides at most r maxu∈Q kj(u)k. Hence it follows that
d d
(7.15) m(ϕ(Q)) ≤ r max kj(u)k = max kj(u)k m(Q).
u∈Q u∈Q
For any invertible linear transformation T in Rd apply the result in (7.15) to the
diffeomorphism T −1 ◦ ϕ. By the chain rule the derivative of T −1 ◦ ϕ is given by the
matrix T −1 ϕ0 . Hence one obtains
d
−1 −1
(7.16) m(T (ϕ(Q))) ≤ max kT j(u)k m(Q).
u∈Q
which lies almost everywhere below the function (1 + ε)| det j | by relation (7.21).
Hence it follows from (7.22) that
Z
d+1
(7.23) m(ϕ(Q)) ≤ (1 + ε) | det j | dm.
Q
Since the inequality in (7.23) is true for every ε > 0 the claim (7.14) has been
shown.
Proof. The result has been shown for closed cubes in X in Lemma 7.6. The
proof of the general case will be given in a number of steps. In Step 1 the result
for closed is extended (via dyadic cubes) to open sets in X. Steps 2 and 3 then
succesively treat the case of measurable sets which are conveniently bounded by
compact sets and the case of general measurable sets in X. The notation j = ϕ0
will still be used as in the proof of Lemma 7.6.
Step 1. Let O be an open set in X ⊂ Rd . It will be shown that ϕ(O) (being
Lebesgue measurable) satisfies
Z
(7.25) m(ϕ(O)) ≤ | det j | dm.
O
Recall that any open subsetSO ⊂ X can be written asS∞a countable union of pairwise
∞
disjoint dyadic cubes O = n=1 Qn . Then ϕ(O) = n=1 ϕ(Qn ), a pairwise disjoint
union of measurable sets. Since the complement of each dyadic cube Qn in its
closure Qn is of lower dimension than d, it follows that m(Qn \ Qn ) = 0 and thus
the relation (7.14) holds also for dyadic cubes. Therefore
X∞ X∞ Z Z
m(ϕ(O)) = m(ϕ(Qn )) ≤ | det j | dm = | det j | dm,
n=1 n=1 Qn O
Let M = maxu∈H | det j(u) |. Since A is measurable there exists compact sets Kn
and open sets On such that
1
Kn ⊂ A ⊂ On ⊂ H, m(On \ Kn ) <
Mn
(if necessary intersect the open set from Corollary 2.22 with H 0 to satisfy the
inclusion On ⊂ H). Since ϕ is a diffeomorphism, clearly ϕ(Kn ) is compact and
ϕ(On ) is open. Moreover by Step 1
m(ϕ(On )) − m(ϕ(Kn )) = m(ϕ(On \ Kn ))
Z
1
≤ | det j | dm ≤ M m(ϕ(On \ Kn )) ≤ ,
On \Kn n
which confirms that ϕ(A) is measurable. Furthermore
1
m(ϕ(A)) ≤ m(ϕ(On )) ≤ m(ϕ(Kn )) +
Z Z n
1 1
≤ | det j | dm + ≤ | det j | dm + ,
Kn n A n
which proves the claim (7.26).
Step 3. Let A be a Lebesgue measurable set in X ⊂ Rd . It will be shown that
ϕ(A) is Lebesgue measurable and that
Z
m(ϕ(A)) ≤ | det j | dm.
A
2. NON-LINEAR TRANSFORMATIONS IN Rd 107
Note that here and in the following examples the choice of the angles is as follows:
for a vector (x, y, z) ∈ R3 first the angle θ is chosen with respect to the z-axis;
for the corresponding projection in the xy-plane the angle ϕ is then chosen with
respect to the y-axis. Note that in the literature different choices may occur.
Clearly the following identities hold:
p
x2 + y 2 + z 2 = r2 , (cos ϕ) x = (sin ϕ) y, (sin θ) z = (cos θ) x2 + y 2 .
Hence (7.38) produce spheres, planes, and cones when the point (r, ϕ, θ) varies in
R3 with the parameter r, ϕ, or θ fixed, respectively. When restricted to the open
set (0, ∞) × (−π, π) × (0, π) the corresponding mapping in (7.38) is injective. The
corresponding transformation formula in (7.32) is expressed as
dxdydz = r2 sin θ drdϕdθ.
Prolate spheroidal coordinates for R3 . The prolate spheroidal coordinates are
based on the formulas:
x = sinh ρ sin ϕ sin θ,
(7.39) y = sinh ρ cos ϕ sin θ, (ρ, ϕ, θ) ∈ R3 .
z = cosh ρ cos θ,
due to (7.51). This shows that the definition of the measure µ in (7.49) is indepen-
dent of the representation.
Recall that a manifold M in Rd is a subset of Rd which can be viewed as a col-
lection of patches, which “overlap differentiably”. The totality of the corresponding
coordinate charts forms a countable atlas for the manifold. A set A ⊂ M is called
measurable if each point of M is contained in a patch whose intersection with A
is measurable in that patch. In fact, the measurable sets thus defined form a σ-
algebra on the manifold M , independent of the particular choice of atlas. In order
to obtain a measure on M let the atlas be given by patches Mi , parametrized by
αi : Xi ⊂ Rk → Rd with Mi = αi (Xi ), and denote the corresponding measure by
µi , i ∈ N. Let A ⊂ M be measurable, then the inductively defined sets
n
!
[
A1 = A ∩ M1 , An+1 = (A ∩ Mn+1 ) \ Ai
i=1
S∞
are measurable (in each patch) and disjoint, such that A = i=1 Ai . With this
decomposition one defines the function µ : M → [0, ∞] by
X∞
µ(A) = µi (Ai ).
i=1
The definition of the function µ is independent of the particular choice of atlas.
Moreover, it turns out that µ is a measure on the measurable sets of M .
CHAPTER 8
Let (hn ) be a Cauchy sequence in the semi-normed space X. Hence for εk = 2−k
there exists Nk so that
(8.1) khn − hm k ≤ 2−k , n, m ≥ Nk .
Thus there is a subsequence (hnk ), such that
(8.2) khnk+1 − hnk k ≤ 2−k , k ∈ N.
Now define the elements fk = hnk+1 − hnk , so that by construction
∞
X
(8.3) kfk k < ∞,
k=1
P∞
i.e., the series k=1 fk converges absolutely. The connection between the Cauchy
sequence and the elements fn can be expressed by the identity
N
X
(8.4) hnN +1 − hn1 = fk .
k=1
P∞from (8.4) that the subsequence hnk converges in X if and only if the
It follows
series k=1 fk converges in X.
Proposition 8.3. Let X be a semi-normed linear space. Then the following
statements are equivalent:
(1) X is complete;
(2) every absolutely convergent series in X converges in X.
Proof. (1) ⇒ (2) The partial sums of an absolutely convergent series form a
Cauchy sequence.
(2)P⇒ (1) Let (hn ) be a Cauchy sequence in X and construct the associated
∞
series k=1 fk as above. Since the series converges absolutely, cf. (8.3), it follows
from the assumption that the series converges. Hence (8.4) implies that the subse-
quence hnk converges in X, which by Lemma 8.1 implies that the Cauchy sequence
converges in X.
The semi-normed linear space X can be turned into a normed linear space
by forming appropriate equivalence classes relative to the closed subspace N. For
h, k ∈ X define h ∼ k if h − k ∈ N. It is easily seen that this gives an equivalence
2. MINKOWSKI AND HÖLDER INEQUALITIES 119
For p = ∞ the space L (Ω) is the set of all measurable complex-valued functions
∞
3. Completeness
The completeness of the spaces Lp (Ω), 1 ≤ p ≤ ∞, will now be considered. For
the case 1 ≤ p < ∞ the equivalence in Proposition 8.3 will be used. As a bonus
there is a statement about pointwise convergence.
Theorem 8.12. Let (Ω, A, µ) be a measure P∞space and let 1 ≤ p < ∞. Let
fn ∈ Lp (Ω),
P∞ n ∈ N, and assume that the series k=1 fk converges absolutely. Then
the series k=1 fk converges in Lp (Ω), i.e., there exists a function f ∈ Lp (Ω), such
that
X∞
(8.16) fk = f,
k=1
where the series converges in L (Ω). In addition, the series in (8.16) converges
p
and hn ∈ L (Ω). It follows from (8.19) and (8.20) that h ∈ Lp (Ω) with
p
(8.21) khkp ≤ M,
and, in particular, h is finite almost everywhere. Hence, according to (8.18), the
series
X∞
|fk (ω)|
k=1
3. COMPLETENESS 123
Proof. Choose ε > 0 then there exists a function ϕ ∈ Cc (Rd ) such that
kf − ϕkp < ε. Then
Z Z
|f (x + t) − f (t)|p dm(t) ≤ |f (x + t) − ϕ(x + t)|p dm(t)
R d R d
Z Z
+ |ϕ(x + t) − ϕ(t)|p dm(t) + |ϕ(t) − f (t)|p dm(t).
Rd Rd
The last and the first integral (translation invariance) in the righthand side are at
most εp . Let K be the support of ϕ and define K1 by
K1 = {x ∈ Rd : d(x, K) ≤ 1}
Then K1 is a compact set and for |x| ≤ 1 the support of the integrand in the middle
integral is contained in K1 . On this compact set ϕ uniformly continuous so there
exists 0 < δ < 1 such that |ϕ(x + t) − ϕ(t)| ≤ ε for |x| < δ.
CHAPTER 9
Decompositions of measures
2. Majorization
Definition 9.3. Let (Ω, A) be a measurable space and let ν and µ be measures
on (Ω, A). Then ν is majorized by µ if
ν(A) ≤ µ(A), A ∈ A.
Let (Ω, A, µ) be a measure space and let h be a measurable nonnegative function
on Ω with 0 ≤ h ≤ 1 almost everywhere with respect to µ. Then
Z
ν(A) = h dµ, A ∈ A,
A
defines a measure ν which is majorized by µ; cf. Theorem 4.39 and the text following
the theorem.
3. ABSOLUTELY CONTINUOUS MEASURES 129
is a bounded linear functional on the Hilbert space L2 (Ω, A, µ). Hence by the Riesz
representation theorem there exists a function h, square integrable with respect to
µ, such that Z Z
f dν = f h dµ, f ∈ L2 (Ω, A, µ).
Ω Ω
Note that h is uniquely defined up to a set of µ-measure 0. With f = 1A and A ∈ A
this representation reduces to
Z
(9.4) ν(A) = h dµ, A ∈ A.
A
and it is clear that these sets are measurable, disjoint, and that Ω = F ∪ G.
Furthermore, define the function h by
hλ
(9.10) h= 1F ,
hµ
then clearly h is measurable and nonnegative.
The behavior of the measure λ relative to the sets F and G will be considered
separately. By (9.8) and (9.10) one obtains
Z Z Z
(9.11) λ(A) = hλ dα = hµ h dα = h dµ, A ∈ A, A ⊂ F,
A A A
where in the last equality the identity (9.5) has been used. Next observe that
Z
(9.12) µ(G) = hµ dα = 0, λ(G) = 0,
G
where the first identity is due to (9.8) and (9.9), and the second identity follows
from the first identity and the fact that λ µ.
Now let A ∈ A, and decompose A according to Ω = F ∪ G, so that
A = (A ∩ F ) ∪ (A ∩ G).
Hence, by A ∩ G ⊂ G and by the second identity of (9.12),
λ(A) = λ(A ∩ F ) + λ(A ∩ G) = λ(A ∩ F ).
By A ∩ F ⊂ F and (9.11) it follows then that
Z
λ(A) = λ(A ∩ F ) = h dµ.
A∩F
Since A ∩ G ⊂ G it then follows from the first identity in (9.12) that
Z Z Z
λ(A) = h dµ + h dµ = h dµ,
A∩F A∩G A
which gives (9.6).
Now assume that the measures λ and µ are σ-finite. Then there exist measur-
able sequences Ln , Mn of subsets of Ω with λ(Ln ) and µ(Mn ) finite for all n ∈ N,
and
[∞ [∞
Ω= Ln , Ω = Mn .
n=1 n=1
Without loss of generality assume that the sets Ln are pairwise disjoint and that the
sets Mn are pairwise disjoint; cf Lemma
S∞ 1.4. Then clearly both λ and µ are finite
on Ln ∩ Mm , n, m ∈ N, and Ω = n,m=1 (Ln ∩ Mm ). Rearrange these intersections
into a sequence of pairwise disjoint measurable sets Ωn on which λ and µ are finite,
and note that
∞
[
(9.13) Ω= Ωn .
n=1
Relative to this decomposition of the set, define for each component the collection
An = { A ∩ Ωn : A ∈ A },
which is a σ-algebra on Ωn ; cf. Theorem 1.39. As the restrictions of λ and µ to Ωn
are finite, the Radon-Nikodym theorem for finite measures may be applied. Hence
132 9. DECOMPOSITIONS OF MEASURES
4. Lebesgue decomposition
In order to state the Lebesgue decomposition theorem one needs the following
rather geometric notions of measures being concentrated on a set and of mutually
singular measures.
Definition 9.9. Let (Ω, A) be a measurable space and let λ be a measure. The
measure λ is said to be concentrated on a set S ∈ A if λ(S c ) = 0.
Lemma 9.10. Let (Ω, A) be a measurable space, let λ be a measure, and let S
be a set in A. The following statements are equivalent:
(1) The measure λ is concentrated on S.
(2) For all A ∈ A the inclusion A ⊂ S c implies λ(A) = 0.
(3) For all A ∈ A the identity λ(A) = λ(A ∩ S) holds.
Proof. Let S be a set in A. It is clear for all A ∈ A that
A = (A ∩ S) ∪ (A \ S) where (A ∩ S) ∩ (A \ S) = ∅.
Hence λ(A) = λ(A ∩ S) + λ(A \ S).
(1) ⇒ (2) This is trivial.
(2) ⇒ (3) Since A \ S ⊂ S c the assumption implies that λ(A \ S) = 0, which
leads to λ(A) = λ(A ∩ S).
(3) ⇒ (1) The assumption λ(A) = λ(A ∩ S) for all A ∈ A implies that λ(S c ) =
λ(S ∩ S c ) = 0. Hence λ is concentrated on S.
Definition 9.11. Let (Ω, A) be a measurable space and let λ1 and λ2 be mea-
sures. Then λ1 and λ2 are mutually singular, denoted by λ1 ⊥ λ2 , if λ1 is concen-
trated on S1 ∈ A, λ2 is concentrated on S2 ∈ A, and S1 ∩ S2 = ∅.
Lemma 9.12. Let (Ω, A) be a measurable space and let λ1 and λ2 be measures.
Then the following statements are equivalent:
(1) λ1 and λ2 are mutually singular.
4. LEBESGUE DECOMPOSITION 133
It is clear that λa and λs are measures, cf. Proposition 1.39, and that λ = λa + λs .
First observe that the definition implies that
(9.19) λs (F ) = λ(F ∩ G) = 0.
Hence from (9.18) and (9.19) it is clear that λs is concentrated on F c = G and
that µ is concentrated on Gc = F . Therefore λs and µ are mutually singular, i.e.,
λs ⊥ µ. Next let A ∈ A and assume that µ(A) = 0, which implies that
Z
hµ dα = µ(A) = 0.
A
Lemma 9.15. Let λ be a measure on the measurable space (R, B). Then the
real-valued function f : R → R defined by
(9.23) f (x) = λ((−∞, x]), x ∈ R.
is non-decreasing and λ = mf .
Definition 9.16. Let λ be a measure on R and let x ∈ R. Then λ has a
derivative L ∈ R with respect to the Lebegue measure m if for every ε > 0 there
exists δ > 0 such that
λ(I)
− L < ε,
m(I)
for every open interval I with x ∈ I and m(I) < δ. If the derivative exists at x ∈ R,
it will be denoted by (Dm λ)(x).
Lemma 9.17. Let f : R → R be a non-decreasing real-valued function, let λ
be the corresponding measure, and let x ∈ R. Then the following statements are
equivalent:
(1) f is differentiable at x and f 0 (x) = L;
(2) λ is differentiable with respect to m at x and (Dm λ)(x) = L.
Proof. (1) ⇒ (2) Assume that f is differentiable at x and f 0 (x) = L. Let
s < x < t and let I = (s, t). Then
λ(s, t] f (t) − f (s)
−L= − L,
m(s, t] t−s
which shows (2).
(2) ⇒ (1) Assume that λ is differentiable with respect to m at x and that
(Dm λ)(x) = L. Then for each ε > 0 there exists δ > 0 such that for any s < x < t
with t − s < δ one has
f (t) − f (s) λ(s, t]
−L = − L < ε.
t−s m(s, t]
Furthermore the assumption implies that λ({x}) = 0, so that f is continuous at x.
Hence f is differentiable at x and f 0 (x) = L.
Let f : R → R be a non-decreasing real-valued function with corresponding
measure λ on the measurable space (R, B) as in (9.22). Then according to Theorem
9.14 the measure λ has the Lebesgue decomposition
λ = λa + λs , λa m, λs ⊥ m,
and, in addition, λa ⊥ λs . Here λa has the representation
Z
λa (A) = h dm, A ∈ B.
A
where h is the Radon-Nikodym derivative of λ with respect to m. There is an
intimate connection between the Radon-Nikodym derivative h and the derivative
Dm λ.
Theorem 9.18. Let f : R → R be a non-decreasing real-valued function with
corresponding measure λ on the measurable space (R, B) as in (9.22). Then for
m-almost all x ∈ R the derivative (Dm λ)(x) exists. In fact, for m-almost all x ∈ R
one has (Dm λ)(x) = h(x).
136 9. DECOMPOSITIONS OF MEASURES
and [ [
A∩ {E : E ∈ E} = {A ∩ E : E ∈ E} (distributive law),
\ \
A∪ {E : E ∈ E} = {A ∪ E : E ∈ E} (distributive law).
Furthermore, one has
[ c \
{E : E ∈ E} = {E c : E ∈ E} (de Morgan’s law),
\ c [
{E : E ∈ E} = {E c : E ∈ E} (de Morgan’s law).
It is customary to denote unions and intersections of a collection of subsets of a
universe via an index set as follows
[ \
Eα , Eα ,
α∈I α∈I
and, in particular, when the collection is finite, one uses
[n \n
Ei , Ei .
i=1 i=1
Let En , n ∈ N, and E be subsets of a universe Ω. The notations En ↑ E or En ↓ E
will denote that En is monotonically increasing or En is monotonically decreasing
in the sense that En ⊂ En+1 or En+1 ⊂ En for n ∈ N, and that
∞
[ ∞
\
En = E or En = E,
n=1 n=1
respectively.
Let Ω and Ω0 be sets and define the product Ω × Ω0 by
Ω × Ω0 = { (ω, ω 0 ) : ω ∈ Ω, ω 0 ∈ Ω0 }.
For A, B ∈ Ω and C, D ∈ Ω0 one has
(A ∪ B) × (C ∪ D) = (A × C) ∪ (A × D) ∪ (B × C) × (B × D),
(A ∩ B) × (C ∩ D) = (A × C) ∩ (B × D),
(A × C)c = (Ac × Ω0 ) ∪ (Ω × C c ),
(A \ B) × C = (A × C) \ (B × C),
(A × C) \ (B × D) = [(A \ B) × C] ∪ [A × (C \ D)].
Now let f be a function from Ω to Ω0 . For E ∈ Ω and G ∈ Ω0 one uses the
following notations:
f (E) = { ω 0 ∈ Ω0 : ω 0 = f (ω) for some ω ∈ E },
and
f −1 (G) = { ω ∈ Ω : f (ω) ∈ G }.
If E is a collection of subsets of Ω, then
[ [
f {E : E ∈ E} = {f (E) : E ∈ E},
\ \
f {E : E ∈ E} ⊂ {f (E) : E ∈ E},
and if G is a collection of subsets of Ω0 , then
[ [
f −1 {G : G ∈ G} = {f −1 (G) : G ∈ G},
2. THE EXTENDED REAL LINE 139
\ \
f −1 {G : G ∈ G} = {f −1 (G) : G ∈ G}.
there is a subsequence xni of xn such that xni → b and b is the smallest number
with this property. The numbers a and b are called the upper limits of (xn ):
a = lim sup xn , b = lim inf xn .
Then lim inf xn ≤ lim sup xn , and lim inf xn = − lim sup(−xn ). If lim xn exists in
R, then
lim xn = lim sup xn = lim inf xn .
Furthermore
xn ≤ yn ⇒ lim inf xn ≤ lim inf yn and lim sup xn ≤ lim sup yn .
3. Open sets in Rd
The real line R is provided in the usual way with a topology via open sets or via
open intervals. Recall that a subset of R is an interval if and only if it is non-empty
and contains each point between
T any two ofSits points. Hence if Iγ is a non-empty
class of intervals of R and γ Iγ 6= ∅, then γ Iγ is an interval. The following fact
plays a fundamental role.
Theorem 10.1. Every non-empty open set of R is the union of countably many
disjoint open intervals.
Proof. Let O be a non-empty open subset of R and let x ∈ O. Since O is
open there is an open interval with midpoint x which is completely contained in O.
Let Ix be the union of all open intervals which contain x and which are contained
in O. Then
(1) Ix is an open interval which contains x and which is contained in O;
(2) Ix contains every open interval which contains x and which is contained
in O;
(3) if y ∈ Ix , then Iy = Ix .
Let x, y ∈ O be distinct points. Then either Ix and Iy are disjoint or they are
identical; for if they are not disjoint then there exists z ∈ Ix ∩ Iy , so that Iz = Ix
and Iz = Iy , which gives Ix = Iy .
Now let I be the class of all disjoint open intervals of the form Ix for points in
x ∈ O. For each such interval Ix choose a rational point r ∈ Ix such that Ir = Ix .
Hence I is countable.
A dyadic interval of R of length 1/2m is an interval of the form
1
(10.1) (k, k + 1], k ∈ Z, m ∈ N ∪ {0}.
2m
Dyadic intervals of a given length are pairwise disjoint and cover R: every point of
R belongs to a unique dyadic interval of given length. Dyadic intervals of different
length intersect precisely when one is contained in the other. Parallel to Theorem
10.1 one has a result for dyadic intervals.
Theorem 10.2. Every non-empty open set O of R is the union of countably
many pairwise disjoint dyadic intervals with closure in O.
This statement allows a generalization to Rd . For each m ∈ N ∪ {0} let Fm be
the family of d-dimensional rectangles
k1 k1 + 1 kd kd + 1
(10.2) , × ··· × , , k1 , . . . , kd ∈ Z.
2m 2m 2m 2m
4. DER GROSSE UMORDNUNGSSATZ 141
integer with this property, then Q ∈ Gm , so that x belongs to the right-hand side
of (10.3). This completes of the identity (10.3).
Proof. It suffices to show (1) ⇔ (3) and the associated statements, as the
rest follows by symmetry.
(1) ⇒ (3) Let M ∈ N, and choose N and K so large that a1 , . . . , aM occur
among
{ank : n ≤ N, k ≤ K}.
Then it follows that
∞ ∞
M N K
! !
X X X X X
|am | ≤ |ank | ≤ |ank | ,
m=1 n=1 k=1 n=1 k=1
which implies
∞ ∞ ∞
!
X X X
|am | ≤ |ank | .
m=1 n=1 k=1
(3) ⇒ (1) Choose K and fix n. Choose M so large that an1 , . . . , anK occur
among a1 , . . . , aM . Then it follows
K
X M
X ∞
X
|ank | ≤ |am | ≤ |am |,
k=1 m=1 m=1
which implies
∞
X ∞
X
|ank | ≤ |am |.
k=1 m=1
Let N ≥ 1 and choose K so that for all n = 1, . . . , N
∞ K ∞ K
X X X X 1
|ank | − 1/N < |ank | or |ank | − |ank | < .
N
k=1 k=1 k=1 k=1
This implies
∞ ∞ ∞
!
X X X
|ank | ≤ |am | + 1.
n=1 k=1 m=1
Now assume the equivalent conditions (1) and (3), so that by what has been
shown above:
∞ ∞ ∞ ∞
!
X X X X
|am | ≤ |ank | ≤ |am | + 1.
m=1 n=1 k=1 m=1
Replace in these inequalities ank and am by Cank and Cam with C > 0. Hence it
follows that
∞ ∞ ∞ ∞
!
X X X X 1
|am | ≤ |ank | ≤ |am | + .
m=1 n=1 m=1
C
k=1
4. DER GROSSE UMORDNUNGSSATZ 143
145