Professional Documents
Culture Documents
Heinrich v. Weizsäcker
Revised Translation
2004/2008
Fachbereich Mathematik
Technische Universität Kaiserslautern
2 Measure Theory (October 20, 2008)
Preface
The aim of this text is to give a short introduction into the basic concepts
of the Lebesgue theory of measure and integration on general spaces in order to
enable the reader to follow Master courses in Probability and in Functional anal-
ysis. The main advantage of Lebesgue’s theory over more elementary concepts
like the Riemann integral is its power in dealing with countable limit operations,
both for the measure (viz. generalized volume) of sets, and for the integral of
functions. These become essential features in higher real analysis and proba-
bility. The price is some additional work in constructing suitable domains of
definition for measure and integral.
There are two main approaches to this theory. The first approach starts with
measures and then constructs the integrals, in analogy to the classical passage
from volume to integration. Conversely, the second approach constructs in
several steps the space of integrable functions and then a set has finite measure
simply if the indicator function of this set is integrable. In this text we follow
the first line, even though it takes a little longer. The main reason for this choice
is that in Probability the measurable sets (’events’) have an important role by
themselves and this path of arguments gives additional information about them.
The common goal of both approaches are the limit theorems of chapter 4 (in
particular the dominated convergence theorem). A reader who knows these limit
theorems can start directly with chapter 5 and check the definitions and results
of the first four chapters only when needed.
Up to minor modifications this text is a translation of the original German
version [?]. Only chapter 2 has been revised more thoroughly. For most math-
ematical words and phrases we introduce both the English and German term.
Contents
1 Measures 5
1.1 σ-Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Null-Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 An example of a non-measurable set . . . . . . . . . . . . . . . . 9
2 Construction of measures 11
2.1 Set algebras and contents . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Outer measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Proof of the Measure Extension Theorem . . . . . . . . . . . . . 16
2.4 Monotone classes and more on uniqueness . . . . . . . . . . . . . 19
3 Measurable functions 21
4 The Integral 25
4.1 The elementary integral . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 The Integral for non-negative measurable and for integrable func-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 The Limit Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 30
5 Fubini’s Theorem 33
6 Convergence in measure 37
6.1 Comparison with other types of convergence . . . . . . . . . . . . 37
6.2 A Metric for Local Convergence in Measure . . . . . . . . . . . . 39
7 The Lp -spaces 43
3
4 Measure Theory (October 20, 2008)
Chapter 1
Measures
1.1 σ-Algebras
A measure measures the size of sets. Not all sets can be measured in this sense
as we shall see in section 1.4. On the other hand the class of sets which are mea-
surable should be sufficiently rich. In particular we want to keep measurability
of sets if we perform simple operations like taking the complement or taking
(countable) unions and intersections. This leads to the following definition.
Remark: 1. Properties a) and b) imply that the empty set ∅ belongs also
to F, since ∅ = S c .
2. Properties b) and c) imply that for every sequence A1 , A2 , · · · of elements of
T∞
F the Sintersection n=1 An is in F, because first Acn ∈ F for each n due to b),
hence n Acn ∈ F due to c) and finally b) yields
∞
à ∞
!c
\ [
An = Acn ∈ F.
n=1 n=1
The remainder of this section shows the typical procedure by which most
σ-algebras are found. If one has several different σ-algebras over the same set
1 Ac denotes the complement of A in S.
5
6 Measure Theory (October 20, 2008)
S one can consider their intersection, i.e. the system of subsets of S which are
members of each of the given σ-algebras.
Proof. We should verify that the intersection has the three properties a)-c)
in Definition 1.1. We restrict
T ourselves to condition c). Let A1 , A2 , · · · be a
sequence of elements of S i∈I F i . Thus An ∈ Fi for all n ∈ N and all i ∈ I. For
fixed i ∈ I it follows
S that T n ∈ Fi since Fi is a σ-algebra. This applies to
n∈N A
each i, and hence n∈N An ∈ i∈I Fi .
Remarks 1. Thus a set A belongs to σ(E) if and only every σ-algebra which
contains all elements of E must also contain A. This concept is similar to the
concept of the linear hull of a set E of vectors in a vector space V , which is the
smallest vector space containing all vectors in E or, the intersection of all linear
subspaces of V which contain E as a subset. But while the elements of the linear
hull can be described explicitly as the linear combinations of the elements of E,
an analogous explicit representation of the elements of σ(E) by the elements of
E is not possible in general.
Definition 1.3 Let (S, d) be a metric space. Let O(S) be the system of all open
subsets of S. The σ-algebra generated by O(S) is called the Borel σ-algebra
of S and is denoted by B(S). Its elements are called the Borel sets of S.
So starting from open or closed sets one can form at liberty countable unions,
intersections and complements without ever leaving the realm of Borel sets.
1.2 Measures
Definition 1.4 Let S be a set and let F be a σ-algebra on S. A map µ : F →
[0, ∞] is called a measure (Maß), if µ(∅) = 0 and if µ is σ-additive, i.e. for
every sequence of pairwise disjoint sets Fn ∈ F one has
Ã∞ ! ∞
[ X
µ Fn = µ(Fn ). (1.1)
n=1 n=1
and in particular
µ(B \ A) = µ(B) − µ(A).
3. A measure is sub-additive: For any two not necessarily disjoint sets
A, B in F one has
µ(A ∪ B) ≤ µ(A) + µ(B),
because µ(A ∪ B) = µ(A) + µ(B\A) ≤ µ(A) + µ(B).
Notation. We write An ↑ A if the sequence (An ) of sets is monotonically
increasing with union A; The notation An ↓ A is used similarly.
1.3 Null-Sets
Null-sets are very useful because for most purposes one can ignore everything
that happens on them.
A M B = (A ∪ B) \ (A ∩ B). (1.3)
The symmetric difference consists of precisely those points at which the two
sets differ, i.e. which are in one of the two sets but not in the other. If the
symmetric difference A M B of two sets is a µ-null-set then A and B are called
µ-equivalent. If moreover both sets are measurable then µ(A∪B) = µ(A∩B) =
µ(A) = µ(B).
Definition 1.7 A measure µ, and the associated measure space (S, F, µ) are
complete (vollständig), if every µ-null set belongs to the σ-algebra F.
Proof. The discussion preceding definition 1.7 shows that the extension of
µ to Fµ is unambiguous. Two sets A0 and A ∈ F differ at exactly the same
points at which the complements A0c and Ac differ. This implies that Fµ is
stable under complements. If (A0n ) is a sequence of sets in Fµ then there are
sets An ∈ F such that A0n MSAn ⊂ Nn ∈ F where µ(N S n ) = 0 for each n. All
points at which the setSA0 = n A0n differs
S from A = n An ∈ F are contained
in the µ-null-set N = Nn and thus n A0n ∈ Fµ . So Fµ is a σ-algebra. If the
A0n are disjoint then the sets An ∩ N c = A0n ∩ N c are also disjoint and hence
∞
X ∞
X ∞
X
µ(A0n ) = µ(An ) = µ(An ∩ N c ) = µ(A ∩ N c ) = µ(A) = µ(A0 ).
n=1 n=1 n=1
Theorem 1.4 (Vitali set) There is a set E ⊂ [0, 1] with the following prop-
erty: Let λ be a measure on a σ-algebra L over [0, 1] such that λ([0, 1]) = 1 and
for all F ∈ L we have F + x mod 1 ∈ L for x ∈ [0, 1] and
The right-hand side is positive and finite. But on the left we have added in-
finitely many times the same value λ(E). Both for λ(E) = 0 and for λ(E) > 0
we get a contradiction.
Construction of measures
Even though it is clear how to measure the size of simple sets by volume or
similar finitely additive functionals, it is not obvious that these definitions can
be extended to a measure on a whole σ-algebra of sets. The main result in this
chapter is Carathéodory’s extension theorem (theorem 2.2). As a special case
this contains the construction of Lebesgue measure in Euclidean space.
There are several approaches to this theorem. Our proof is based on an
approximation argument. It can easily be reframed to work also in the functional
analytic context.1
This shows that an algebra is a σ-algebra if and only if it is closed under forming
countable disjoint sequences.
theory of fractals, cf. e.g. the script [?] ’Fractal sets and Preparation to Geometric Measure
Theory’ (http://www.mathematik.uni-kl.de/ wwwstoch/2002w/skriptrev.pdf). A proof via
approximation from inside is also possible, see the script [?]
11
12 Measure Theory (October 20, 2008)
Example 2.1 Let A be the set algebra on R which is generated by all intervals
which are closed on the right and open on the left. Then A consists precisely of
all sets of the form
[n
A= Ji (2.1)
i=1
where each Ji is an interval of one of the three types (−∞, b], (a, b] or (a, ∞).
It is called the interval algebra over R.
A ring R is also stable under the operation M (cf. 1.2). Moreover (1.3) shows
that A ∩ B = (A ∪ B) \ (A M B) and hence a ring is stable under intersections. If
one takes M as addition and ∩ as multiplication then a ring in our sense becomes
a ’commutative ring’ in the algebraic sense.2
Example 2.2 An example of a ring is the interval ring R of all bounded sets
in the interval algebra A of Example 2.1.
Obviously a ring is a set algebra if and only if it has the set S as an element.
Example 2.3 Let A (resp. R) be the above algebra (resp. ring) of unions of
intervals. One can assume that in the representation (2.1) the intervals Ji are
disjoint. If F : R → R is a non-decreasing function then we get a content mF
on A resp. R by the formula
n
[ n
X
mF ( Ji ) = F (bi ) − F (ai ), (2.4)
i=1 i=1
where for unbounded intervals we use the limit values F (−∞) = limx↓−∞ F (x)
and F (∞) = limx↑∞ F (x) in the obvious way. It is called Stieltjes content.
For F (x) = x we get the usual geometric size of a set based on the length of
intervals.
Clearly if a content can be extended to a measure it must be σ-additive for
all sets for which it is defined.
Definition 2.4 A content m on a ring R is called σ-additive if for every
disjoint sequence (An ) in R whose union is in R, one has
∞
[ ∞
X
m( An ) = m(An ). (2.5)
n=1 n=1
N
[ N
[
RN = RN ∩ Bnc ⊂ Rn \ Bn
n=1 n=1
14 Measure Theory (October 20, 2008)
R = (a1 , b1 ] × · · · (an , bn ].
This is a very mild condition. For example the volumeSfunction on the ring
generated by rectangular sets in Rd is σ-finite since Rd = k (−k, k]d . We need
this condition in order to ensure uniqueness in the following theorem which is
the central result in this chapter.
Theorem 2.2 Let m be a σ-additive and σ-finite content on the ring R. Then
there is a unique extension of m to a measure on σ(R).
The remainder of this chapter is devoted to the proof of this theorem and
some additional comments. In a first round the reader may pass directly to the
next chapter.
Definition 2.6 Let S be a set. A map µ∗ from the power set 2S to [0, ∞]
is called an outer measure (äußeres Maß) on S, if it has the following
properties:
1. µ∗ (∅) = 0,
2. A ⊂ B implies µ∗ (A) ≤ µ∗ (B) (monotonicity),
Measure Theory (October 20, 2008) 15
S∞ P∞
3. µ∗ ( n=1 En ) ≤ n=1 µ∗ (En ) (σ-sub-additivity).
Thus all properties of an outer measure are verified. The final inequality in the
assertion is clear since every R ∈ R forms an admissible covering of itself on the
right hand side of (2.6).
µ∗|R = m.
S
Proof.
S Let A ∈ R be given. Suppose A ⊂ i Ri and Ri ∈ R. Then
A = i A ∩ Ri . As m is σ-sub-additive and monotone we get
∞
X ∞
X
m(A) ≤ m(A ∩ Ri ) ≤ m(Ri ).
i=1 i=1
Taking the infimum over these coverings gives m(A) ≤ µ∗ (A). The converse
inequality was part of the proposition.
holds. Moreover the function µ∗ and the set operations ∪, ∩, \ are continuous
for d: If both d(An , A) → 0 and d(Bn , B) → 0 then µ∗ (An ) → µ∗ (A), d(An ∪
Bn , A ∪ B) → 0, and similarly for ∩ and \.
which in turn implies the continuity of µ∗ . Let now ◦ denote one of the above
set operations. Let A, B and A0 , B 0 be two pairs of sets. If a point x belongs to
A ◦ B but not to A0 ◦ B 0 or conversely, then x is at least in one of the symmetric
difference sets A M A0 or B M B 0 . More formally
(A ◦ B) M (A0 ◦ B 0 ) ⊂ (A M A0 ) ∪ (B M B 0 ).
µ∗ (A ∪ B) = lim m(An ∪ Bn )
= lim m(An ) + lim m(Bn ) − lim m(An ∩ Bn )
= lim µ∗ (An ) + lim µ∗ (Bn ) − lim µ∗ (An ∩ Bn )
= µ∗ (A) + µ∗ (B) + µ∗ (∅) = µ∗ (A) + µ∗ (B).
4. We claim that µ∗|Mf is σ-additive, and moreover for every disjoint se-
S
quence
P∞ (An ) in Mf the union A = An also belongs to Mf if and only if
∗ ∗
n=1 µ (An ) < ∞. Clearly if A ∈ Mf (or more generally if µ (A) < ∞) then
by step 2 the partial sums of the series are all dominated by µ∗ (A) and the
series converges. Conversely, assume that the series converges. Then
N
[ N
[ ∞
[ ∞
X
d(A, An ) = µ∗ (A M An ) = µ∗ ( An ) ≤ µ∗ (An ) → 0
n=1 n=1 n=N +1 n=N +1
6. We note in passing that in the case where R is an algebra and m(S) < ∞
3
the proof is now complete: The class Mf then is an algebra according to step
2. By step 3, the series in the statement of step 4 always are bounded and thus
Mf is closed under countable disjoint unions. Hence Mf is a σ-algebra which
contains both R = Rf and N , i.e. σ(R ∪ N ) ⊂ Mf . Further µ∗ is σ-additive
on Mf , in particular it induces a measure extension of m on σ(R) which is
complete because N is closed under taking subsets.
II. 7. The remainder of the proof is devoted to the extension to the un-
bounded case. Consider the system of sets
M̄ = {A ⊂ S : A ∩ B ∈ Mf f or all B ∈ Mf }. (2.10)
We can now give the uniqueness argument for theorem 2.2: Let m be σ-finite
and let (Ek ) be an increasing sequence in R of finite content whose union is the
whole set S. Let A ∈ σ(R). Then
ν(A) = lim ν(A ∩ Ek ) = lim µ(A ∩ Ek ) = µ(A).
k k
Measure Theory (October 20, 2008) 19
Example 2.5 (Counting measure) Let R be the family of all finite subsets of
a set S. Let m assign to each finite set its cardinality. The corresponding outer
measure assigns to each infinite set the value +∞. The empty set ∅ is the only
null-set and σ(R) = σ(R ∪ N ) consists of all countable subsets of S and their
complements. Let us look into the proof of theorem 2.3 in this case. The system
Mf defined there agrees with the system of finite sets whereas the σ-algebra
M̄ defined in (2.10) coincides with the full power set. We see that the proof
actually sometimes produces a measure extension on a larger σ-algebra than
just on σ(R ∪ N ).
Example 2.6 Here is a counter-example even if the total mass of the measures
is fixed. Let S = {1, 2, 3, 4} and let the system E consist of the two sets {1, 2}
and {2, 3}. Then all four singletons {1}, {2}, {3}, {4} can easily be produced
via algebra operations from these two sets, and hence σ(E) is the power set of
S. The two weight assignments {1/4, 1/4, 1/4, 1/4} and {1/2, 0, 1/2, 0} clearly
induce two different measures µ and ν on the power set S with µ(S) = ν(S) = 1
and µ(A) = ν(A) = 1/2 for both sets A ∈ E.
a) S ∈ D.
b) If A, B ∈ D and A ⊂ B then B \ A ∈ D.
c) If
S∞(An ) is an increasing sequence of elements of D, then their union
n=1 An belongs D.
Theorem 2.5 (Uniqueness criterion for measures) Let the system E be stable
under intersections and generate the σ-algebra F. Let µ and ν be two measures
on F which agree on E. If the set S is the union of an increasing sequence of
sets Ek ∈ E with µ(Ek ) = ν(Ek ) < ∞ then µ = ν.
Proof. First of all it is easy to verify that for each k the system of sets
Dk = {A ∈ F : µ(A ∩ Ek ) = ν(A ∩ Ek ) is a monotone class which contains E.
Thus by the monotone class theorem µ(A ∩ Ek ) = ν(A ∩ Ek ) for all k and all
A ∈ F. This implies the assertion by continuity from below of both measures.
A typical application is
Corollary 2.2 A finite measure on the Borel subsets of a metric space is uniquely
determined by its values on open sets.
Chapter 3
Measurable functions
f −1 (G) ∈ F f or all G ∈ G.
For functions with one-dimensional range often the values ±∞ occur in limit
procedures. Therefore it is convenient to consider also extended real valued
functions. The following notation is convenient.
Notation: Let f, g : S −→ [−∞, ∞] be two functions and let a ∈ R. Then
denote by {f ≤ a} the set {x ∈ S|f (x) ≤ a} = f −1 ((−∞; a]), by {f = g} the
set {x ∈ S|f (x) = g(x)} and introduce similarly the notations
Moreover we recall that the extended real line [−∞, ∞] is a metric space
when endowed with the metric
|x − y|
d : [−∞, ∞] × [−∞, ∞] → [0, 1], d(x, y) = .
1 + |x − y|
It is not hard to see that the mapping f : [−∞, ∞] → [−1, 1] given by f (x) =
x
1+|x| is homeomorphic, and hence the above metric defines the usual topology
on [−∞, ∞].
For extended real valued functions measurability has a simpler description.
21
22 Measure Theory (October 20, 2008)
A = {A ⊆ Y |f −1 (A) ∈ F}.
where E denotes the class of intervals of the form [−1, a), [−1, a], (a, 1], and
[a, 1] for a ∈ (−1, 1). The inclusion ⊇ is clear, and the reverse inclusion follows
from the fact that
Since the singleton sets {−∞} and {+∞} are Borel measurable, the previous
criterion admits an obvious analogue for extended real-valued functions.
Theorem 3.1 Let f, g be extended real valued functions. Then the sets {f ≤
g}, {f < g}, {f = g} are in F.
S
Proof. We have {f < g} = r∈Q {f ≤ r} ∩ {g > r} ∈ F, and thus also
{f ≤ g} = {g < f }c ∈ F and {f = g} = {g ≤ f } ∩ {f ≤ g} ∈ F.
Measure Theory (October 20, 2008) 23
Lemma 3.3 Let (S, F ), (Y, G), (Z, H) be measurable spaces. Let f : Y → Z, g :
S → Y be measurable maps. Then the composition f ◦ g : S → Z is measurable.
This holds for all boxes and lemma 3.1 together with theorem 1.2 implies the
measurability of f .
Theorem 3.2 Let (S, F) be a measurable space and let {fn }n∈N be a sequence of
measurable extended real valued functions on S. a) Then the (pointwise defined)
functions supn fn , inf n fn , lim supn→∞ fn , lim inf n→∞ fn are measurable.
b) The set F = {x : lim fn (x) exists} is in F and limn 1F fn is measurable.
Definition
Pn 3.3 Let (S, F) be a measurable space. A function of the form
f (x) = i=1 αi 1Ai (x) with αi ∈ R, Ai ∈ F is called an elementary func-
tion or measurable step function (meßbare Treppenfunktion).
Remark 3.2 Clearly the measurable step functions form a vector space. A
measurable step function takes only finitely many different values Sβ1 , . . . , βm .
The sets Bj = {f = βj } are disjoint P and measurable and S = Bj . So f
also admits the representation f (x) = βj 1Bj (x), where the sets Bj form a
(disjoint) partition of the set S.
Proof. Let
n
n2
X −1
fn (x) = k2−n 1{k2−n ≤f <(k+1)2−n } + n1{f ≥n} , (3.1)
k=0
i.e. if f (x) < n, then fn (x) is the largest member of the grid 2−n N0 which is
situated on the left of f (x), and fn (x) = n if f (x) ≥ n. Because these grids
are successively contained in each other we get fn (x) ≤ fn+1 (x). Moreover
fn ≤ f . Finally f (x) − fn (x) ≤ 2−n , as soon as f (x) ≤ 2n · n · 2−n = n and
fn (x) = n ↑ ∞ = f (x) if f (x) = ∞. Therefore (fn ) converges point-wise to
f.
The next result is the only place in this section where a measure appears.
Sometimes functions are not defined at all points of the set S. Then the
following convention is helpful.
The Integral
In this chapter (S, F, µ) is a measure space. We define the integral with respect
to µ and study some of its properties.
The representation 4.1 is not unique, because the sets of the form {f = α}
can be split in many different ways into disjoint measurable parts. But
Lemma 4.1 The integral in Definition 4.1 does not depend on the choice of the
representation in (4.1).
Pm
Proof. Let f = j=1 βj 1Bj be a second representation of S f with disjoint
S
sets Bj ∈ F . Without loss of generality we may assume S = i Ai = j Bj
since the coefficient 0 is admitted. Moreover αi = βj for all pairs (i, j) of indices
with Ai ∩ Bj 6= ∅, because both coefficients agree with the common value which
the function f takes on this intersection. The additivity of µ shows
n
X n X
X m m X
X n m
X
αi µ(Ai ) = αi µ(Ai ∩ Bj ) = βj µ(Ai ∩ Bj ) = βj µ(Bj ).
i=1 i=1 j=1 j=1 i=1 j=1
25
26 Measure Theory (October 20, 2008)
For the sake of simplicity of this lemma we have assumed in Definition 4.1
that the sets Ai in the representation (4.1) are disjoint. Part a) of the following
theorem shows in particular that the formula (4.2) is also valid if the Ai are not
disjoint.
Proof. The proofs for a) and b) follow from the fact that for any two
elementary functions there is a common partition of S into disjoint measurable
subsets on which
Pm both functions are constant.
c) Let f = j=1 αj 1Aj , and assume without loss of generality that αj > 0 for
all j = 1, . . . , m. We consider the sets Bjn,k = {x ∈ Aj : fn (x) > αj − k1 }. For
fixed k we get Bjn,k ↑ Aj as n → ∞ because each x ∈ Aj is in Bjn,k for eventually
1
all n. Moreover, P if k is 1large enough to make αj − k > 0 for all j = 1, . . . , m,
we have fn ≥ j (αj − k ) 1B n,k by the positivity assumption, and thus
j
Z X Xm
1 1
lim fn dµ ≥ lim (αj − ) µ(Bjn,k ) = (αj − ) µ(Aj ).
n n k j=1
k
R P R
As k → ∞ we get limn fn dµ ≥ αj µ(Aj ) = f dx. The converse inequality
follows from the monotonicity of the integral.
From Definition 4.2 and lemma 4.2 it is straightforward to see that the
integral for non-negative measurable functions is again additive, monotone and
positively homogeneous.
We now admit functions which also take negative values. Every function
f : S → [−∞, ∞] has the decomposition f = f + − f − , where f + = max(f, 0)
and f − = max(−f, 0). Then |f | = f + + f − . All these functions are measurable
if f is measurable (cf. corollary 3.1).
R
Definition 4.3 a) A measurable function f : S → [−∞, ∞] with |f | dµ < ∞
is called µ-integrable (µ-integrierbar). For a µ-integrable function the inte-
gral is defined by Z Z Z
f dµ := f dµ − f − dµ.
+
b) First suppose that µ(N ) = 0 for N = {f 6= 0}. Then for every elementary
function
R g with g ≤ |f | we have also µ{g 6= 0}R = 0 and hence by Definition 4.1
gdµ = 0. Then according
R to Definition 4.2 |f |dµ = 0.
Conversely assume |f |dµ = 0. One has {|f | > 1/n} ↑ {|f | > 0} and therefore
by Markov’s inequality
Z
µ({|f | > 0}) = lim µ({|f | > 1/n}) ≤ lim n |f |dµ = 0.
n n
28 Measure Theory (October 20, 2008)
c) The first assertion follows by letting a tend to +∞ in (4.3). For the second
assertion note that
Z Z Z Z Z Z
−( f + dµ + f − dµ) ≤ f + dµ − f − dµ ≤ f + dµ + f − dµ.
R R
d) Clearly if |f | ≤ h and h is integrable then |f | dµ ≤ hdµ < ∞.
Example 4.3 (Riemann integrals). Let [a, b] be a real interval and let f :
Pk
[a, b] → R be Riemann integrable. In the special case where f = i=1 αi 1(ai ,bi ]
Pk
is a step function clearly the Riemann sum i=1 αi (bi − ai ) coincides with the
elementary Lebesgue integral. In the general case there are two sequences (f n )
and (f n ) of step functions such that for f = inf n f n and f = supn f n we have
Z b Z b
f ≤ f ≤ f and sup f n (x)dx = inf f n (x)dx.
n a n a
The two functions f and f are -according to theorem 3.2- Borel measurable.
Since they are bounded and on a bounded interval the constants are integrable
they areRLebesgue-integrable
R by proposition
R 4.1 d). Monotonicity of the integral
implies f dλ1 = f dλ1 . Thus f − f dλ1 = 0 and by proposition 4.1 d)
f − f = 0 a.e.. Therefore f = f a.e. and proposition 4.2 shows that f is
Lebesgue-integrable and
Z Z Z b Z b
1 1
f dλ = f dλ = f (x) dx = f (x) dx.
a a
R∞
However a so-called improper Riemann integral like in the formula 0 sinx x dx =
R∞
π/2 typically is not a Lebesgue-integral, e.g. in this example 0 | sinx x | dx = ∞,
i.e. the integrand is not λ1 -integrable in the sense of Definition 4.3.
Part a) of the following lemma follows from proposition 4.1 c). Part b) is
obvious.
30 Measure Theory (October 20, 2008)
Proof. Theorem 3.3 shows that for every n ∈ N there exists a sequence
(fnk )k∈N of elementary functions such that fnk ↑k→∞ fn . Let us consider the
elementary function gk = max(f1k , . . . , fkk ). For each k we have gk+1 ≥ gk and
f ≥ gk . Moreover supk gk ≥ supk≥n fnk = fn for each n. So f ≥ supk gk ≥
supn fn = f , i.e. gk ↑ f . Since the gn are elementary lemma 4.2 gives
Z Z Z
f dµ = lim gn dµ ≤ lim fn dµ.
n→∞ n→∞
R R
The converse inequality f dµ ≥ limn→∞ fn dµ follows from the monotonicity
of the integral.
Remark 4.1 The additivity of the integral and the theorem of monotone con-
vergence together show that every non-negative measurable function f gives rise
to a new measure f µ : F → [0, ∞], defined by
Z
f µ(A) = 1A f dµ.
In fact for every sequence (An ) of pairwise disjoint sets with union A
Z X ∞ ∞ Z
X ∞
X
f µ(A) = ( 1An )f dµ = 1An f dµ = f µ(An ).
n=1 n=1 n=1
Example 4.4 In general the inequality in Fatou’s lemma is strict even if the
functions converge point-wise and are in L1 (µ). We consider the standard
Lebesgue measure space ([0, 1], B([0, 1]), λ|B([0, 1])). Let f = 0 and fn = n1(0, n1 ] .
Then fn → f point-wise. On the other hand
Z 1 Z 1
1
lim fn (x) dx = lim n = 1 6= 0 = f (x)dx.
n 0 n n 0
The following theorem is perhaps the most useful single result in measure
theory.
and Z
lim |f − fn | dµ = 0.
n→∞
Proof. The union N of the countably many null-sets {fn 9 f }, {|fn | > h}
and {|h| = ∞} is a null-set. Without changing the integrals or the integrability
of the functions involved we can set all function to vanish on N . Let us consider
the auxiliary functions gn = 2h − (f − fn ). Then limn gn = 2h and gn ≥ 0, so
by Fatou’s lemma
Z Z Z
2h dµ = lim gn dµ ≤ lim inf gn dµ
n n
Z Z Z
¡ ¢
= 2h dµ − lim sup f dµ − fn dµ .
n
32 Measure Theory (October 20, 2008)
¡R R ¢
Therefore lim supn f dµ − fn dµ ≤ 0. Similarly we can apply Fatou’s
lemma to the non-negative
¡R Rfunctions
¢ 2h + f − fn , and get in the same way the
estimate lim inf n f dµ − fn dµ ≥ 0. Together this gives the first of the two
limit assertions. The second one follows if we apply the first one to the sequence
(|f − fn |) which is dominated by 2h and converges everywhere to 0.
As a corollary one gets the following rule for differentiation under the integral
sign.
Theorem 4.6 Let (a, b) an open interval and let f : (a, b)×S → R be a function
such that
i) for fixed t ∈ (a, b) is the function f (t, ·) is µ-integrable on S,
ii) for fixed x ∈ S the function f (·, x) is continuously differentiable on (a, b),
∂
iii) there is a µ-integrable h : S → [0, ∞] such that | ∂t f (t, x)| ≤ h(x) for all
x ∈ S and all t ∈ (a, b).
R
Then the integral f (t, x)dµ(x) is continuously differentiable in t and
Z Z
∂ ∂
f (t, x) dµ(x) = f (t, x) dµ(x). (4.6)
∂t S S ∂t
Proof. Let (tn ) be sequence of points which converges to some t ∈ (a, b).
Consider the measurable functions gn defined by gn (x) = f (tn ,x)−f
tn −t
(t,x)
. By the
∗
mean-value theorem for each x there is a point tn (x) ∈ (a, b) such that gn (x) =
∂ ∗
∂t f (tn (x), x). By hypothesis we can conclude |gn (x)| ≤ h(x). Furthermore
∂
gn (x) → ∂t f (t, x) for all x ∈ S. The theorem of dominated convergence gives
R R Z Z
S
f (tn , x) dµ(x) − S f (t, x) dµ(x) ∂
= gn (x) dµ → f (t, x)dµ.
tn − t ∂t
i.e. the relation (4.6). Similarly the continuity of the right-hand side in (4.6)
follows because iii) implies
Z Z Z
∂ ∂ ∂
lim f (tn , x) dµ(dx) = lim f (tn , x) dµ(dx) = f (t, x) dµ(dx).
n ∂t n ∂t ∂t
Chapter 5
Fubini’s Theorem
Theorem 5.1 (Fubini) Let (X, F, µ) and (Y, G, ν) two σ-finite measure spaces.
Let F ⊗ G be the σ-algebra on the cartesian product X × Y which is generated
by the ’measurable rectangles’ F × G with F ∈ F, G ∈ G. Then
a) There is a unique measure µ ⊗ ν on the σ-algebra F ⊗ G, such that
for all F ∈ F, G ∈ G.
b) Let f : X × Y → [0, ∞] be a F ⊗ G-measurable function. For every x ∈ X the
sectional
R function f (x, ·) : y 7→ f (x, y) is G-measurable and the function x 7→
Y
f (x, y) dν(y) is F-measurable. Similarly the R sectional function f (·, y) is F-
measurable for each y ∈ Y and the map y 7→ X f (x, y) dµ(x) is G-measurable.
Moreover
Z Z Z
f (x, y) dµ ⊗ ν(x, y) = f (x, y) dν(y) dµ(x) (5.1)
X×Y X Y
Z Z
= f (x, y) dµ(x) dν(y).
Y X
In this case f (x, ·) ∈ L1 (ν) for µ-almost all x and f (·, y) ∈ L1 (µ) for ν-almost
all y and the formula in b) is valid in the sense that the inner integrals denote
an arbitrary measurable function on the respective null-sets where they are not
defined.
Before the proof of the theorem let us write down two important corollaries:
33
34 Measure Theory (October 20, 2008)
Proof. 1. Suppose first that µ(X) < ∞ and ν(Y ) < ∞. Consider the
system D of all subsets A of X × Y for which i) the section sets Ax are in G for
all x ∈ X and ii) the map x 7→ ν(Ax ) is F-measurable. Clearly X × Y ∈ D.
The set operation A 7→ Ax commutes with countable unions and the difference
operator \, moreover ν is a finite measure. These two facts imply that D is a
monotone class in the sense of section 2.4. Obviously D contains the system E
of all ’measurable rectangles’ F × G with F ∈ F , G ∈ G. The intersection of
two such sets is again Rin E. So theorem 2.4 applies and gives F ⊗ G ⊂ D. So
the map µ ⊗ ν : A 7→ X ν(Ax ) dµ(x) is well-defined and the σ-additivity of ν
and the theorem of monotone convergence show that it is a measure on F ⊗ G.
Moreover
Z Z
µ ⊗ ν(F × G) = ν((F × G)x ) dµ(x) = ν(G) dµ(x) = µ(F )ν(G).
X F
holds µ-a.e. and therefore the identity (5.1) carries over from f + and f − to f .
The second identity in b) follows by exchanging the roles of the variables x, y.
5. Finally we get rid of the finiteness condition on the measures. Let (Qn ) and
(Rn ) be two sequences in F resp. G, such that µ(Qn ) < ∞, ν(Rn ) < ∞ and
Measure Theory (October 20, 2008) 35
Remark 5.1 We have used monotone classes. If one does not want to rely on
section 2.4 one can get almost the same result as follows. Let R be the ring
of finite disjoint unions of measurable rectangles. It is easy to extend the set
function F × G 7→ µ(F )ν(G) by additivity to a content m on R R and to see
that the function x 7→ ν(Rx ) is F-measurable and m(R) = X ν(Rx ) dµ for
all R ∈ R. This implies that m is σ-additive and σ-finite. Let µ ⊗ ν denote
the unique extension to F ⊗ G. Let now f ≥ 0 be F ⊗ G-measurable. The
only difficulty is that without
R monotone classes it is not so easy to prove the
F-measurability of x 7→ f (x, y) dν(y). However the Fµ -measurability can be
verified. For this let first A ∈ F ⊗ G. Then for each n there is a sequence
S (Rin )i
of (without loss ofPgenerality disjoint) sets in R such that A ⊂ i Rin =: An
n 1 1
and µ ⊗ ν(A) ≥ i m(Ri ) − n = µ ⊗ ν(An ) − n . The class D in part 1.
of the above proof contains the Rin and their disjoint union An . Similarly,
applying the same argument to Ac we find a set An ∈ D such that An ⊂ A
and µ ⊗ ν(An ) ≥ µ ⊗ ν(A) − n1 . Then using the same argument as for the
Lebesgue-integrability of Riemann integrable functions we see that x 7→ ν(Ax )
is Fµ -measurable and the first identity in corollary 5.1 holds. The rest of the
proof works essentially like before.
4. We give an example that the uniqueness of the product measure may get
lost if one of the measures is no longer σ-finite. In view of the arguments of
remark 5.1 it also illustrates the importance of σ-finiteness in the uniqueness
part of Carathéodory’s theorem 2.2. Let X = Y = [0, 1] and F = G = B([0, 1]).
Let λ be Lebesgue measure on F and let ζ be the counting measure on G. As
in the beginning of the proof of theorem 5 we can define two measures ρ1 and
ρ2 on F ⊗ G by Z 1
ρ1 (B) = ζ(B x ) dλ(x),
0
Z 1 X
ρ2 (B) = λ(By ) dζ(y) = λ(By ).
0 y
whereas X X
ρ2 (D) = λ(Dy ) = 0 = 0.
y y
Chapter 6
Convergence in measure
Besides point-wise convergence or a.e. convergence there are several other ways
to express that a sequence of measurable functions converges. The Lp -spaces
of the next chapter will give a full scale of such concepts. Here we discuss
convergence in measure which is particularly useful in probability theory. The
reader who wants to pass directly to the next chapter will need the completeness
result of theorem 6.4 b) which can be understood independently of the other
parts of this section.
37
38 Measure Theory (October 20, 2008)
Theorem 6.1 Between the three concepts in Definition 6.1 the following im-
plications hold:
a) If fn −→ f in the mean then fn −→ f in measure.
b) If fn −→ f µ a.e. then fn 1A −→ f 1A in measure for every A ∈ F with
µ(A) < ∞.
The converse implications are wrong even for finite measure spaces.
S∞ P∞
N ∩ Am0 ⊂ N ∩ Am ⊂ l=m Bl and µ(N ∩ Am0 ) ≤ l=m 2−l = 2 · 2−m for all
m ≥ m0 . Hence µ(N ∩ Am0 ) = 0 for all m0 and finally µ(N ) = 0 as desired.
2. Conversely suppose the subsequence condition holds. Let A ∈ F be a set of
finite measure and ε > 0. Choose a sub-sequence such that
lim sup µ(A ∩ {|fn − f | > ε}) = lim µ(A ∩ {|fnk − f | > ε}).
n k→∞
Corollary 6.1 Let h ∈ L1 (µ) and let Mh be the set all measurable functions g
with |g| ≤ h. Then for sequences in Mh convergence in mean and local conver-
gence in measure agree.
Theorem 6.3 (Egorov). Let µ(S) < ∞ and let fn → f µ-a.e.. Then for each
ε > 0 there is a set A ∈ F with µ(Ac ) < ε such that fn converges to f even
uniformly on A.
1
Proof. Let gn = supk≥n |fk − f |. Then gn ↓ 0 µ-a.e., i.e. µ({g
P∞n > m }) → 0
for all m ∈ N. Then we find a sequence of indices such that m=1 µ({gnm >
1 ∞ 1 c
m }) < ε. Let A = ∩m=1 {gnm ≤ m }. Then µ(A ) < ε. For the verification of
1
uniform convergence let η > 0. Choose m ≥ η . Then for all x ∈ A and n ≥ nm
1
one gets |fn (x) − f (x)| ≤ gnm (x) ≤ m ≤ η.
the sequence (fnk )(x) converges for k → ∞ to a finite real number f (x) for all
x ∈ A. Outside of A let f (x) = 0. The measurability
P∞ of the fn and of A imply
the measurability of f . Moreover |f − fnk | ≤ l=k |fnl+1 − fnl | everywhere and
by (6.2) and (6.3)
∞
X
ρ(|f − fnk |) ≤ ρ(|fnl+1 − fnl |) < ∞
l=k
For the proof of (6.4) P∞assume that in this calculation the last series converges.
Then the function P∞ i=1 (h ∧ gi ) is integrable and hence a.e. finite. Let x be a
point such that i=1 (h(x) ∧ gi (x)) converges. ThenP gi (x) < h(x) for eventually
∞
all
P∞ indices because h(x) > 0 and therefore the series i=1 gi (x) converges. Thus
g
i=1 i is finite a.e.. Therefore d is a complete pseudo-metric.
P P
If n d(fn , f ) < ∞ then the series n fn (x) − f (x) converges absolutely µ-a.e.
because of (6.4) and in particular fn → f µ-a.e..
Assume that limn d(fn , f ) = 0. In order to prove local convergence in measure
we verify the sub-sequence criterion P of theorem 6.2. Let (fnk ) be a sub-sequence.
Then every sub-sub-sequence with l d(fnk l , f ) < ∞ converges to f µ-a.e..
Conversely assume local convergence in measure of fn to f or |fn −f | to 0. Then
the same is true for the functions |fn − f | ∧ h. By corollary
R 6.1 these functions
converge also in mean, i.e. by definition d(fn , f ) = fn − f | dµ → 0.
42 Measure Theory (October 20, 2008)
Chapter 7
The Lp-spaces
Definition 7.1 Let (S, F, µ) be a measure space and let f be an extended real-
valued F-measurable function. We define for p ∈ [1, ∞] the number kf kp by
Z
¡ ¢1/p
kf kp := |f |p dµ (1 ≤ p < ∞)
Further let Lp (µ) = { f : f is F-measurable and kf kp < ∞}. Lp (µ) denotes the
space of all µ-equivalence classes of elements of Lp (µ).
43
44 Measure Theory (October 20, 2008)
which implies the assertion. In the last step we have used the general relation
µZ ¶ q1 µZ ¶ p−1
p
p−1 (p−1)q p
kh kq = h dµ = h dµ = khkp−1
p . (7.2)
1 1
Corollary 7.1 Let µ(S) < ∞ and p < q. Then kf kp ≤ kf kq µ(S) p − q for all
measurable functions f and in particular Lq (µ) ⊂ Lq (µ).
1 1
Proof. Let s be the number for which s = p − 1q . The constant function
1
s
1 is in L (µ) with k1ks = µ(S) . Hölder’s inequality yields kf kp = kf · 1kp ≤
s
Theorem 7.2 For each p ∈ [1, ∞] the space Lp (µ) with the norm k · kp is a
Banach space.
Measure Theory (October 20, 2008) 45
and hence µ(B) < ∞ is equivalent to 1B ∈ Lp (µ). Moreover the theorem 4.5 of
dominated convergence holds also in Lp : If a sequence of measurable functions
satisfies |fn | ≤ h ∈ Lp (µ) and fn → f point-wise, then kfn − f kp → 0. In fact
|fn − f |p ≤ p
R h and we can apply the original theorem 4.5 to these functions and
conclude |fn − f |p dµ → 0.
Proof. Let L be the smallest closed linear subspace of Lp (µ) which contains
the indicator functions 1E , E ∈ E. We want to show L = Lp (µ). Fix a number
k ∈ N. We consider the class Dk of all sets F ∈ F such that 1F ∩Ek ∈ L. For
every E ∈ E we have E ∩ Ek ∈ E and hence 1E∩Ek ∈ L by assumption. So
E ⊂ Dk .
We claim that Dk is a monotone class. Since S ∩ Ek = Ek ∈ E we have S ∈
Dk . Let F, G ∈ Dk such that G ⊂ F . Then 1(F \G)∩Ek = 1F ∩Ek − 1G∩Ek ∈ L
since L is a linear space. Thus F \ G ∈ Dk . Let Fn be an increasing sequence
in Dk and Fn ↑ F . Then Fn ¡∩ Ek ↑ F ∩ Ek¢ as n → ∞ and µ(F ∩ Ek ) < ∞.
Thus k1F ∩Ek − 1Fn ∩Ek kpp = µ (F \ Fn ) ∩ Ek → 0 as n → ∞ and hence 1F ∈ L
since L is closed. Therefore F ∈ Dk . The monotone class theorem 2.4 shows
that F ⊂ Dk for all k. Let now F ∈ F be any measurable set with µ(F ) < ∞.
Then F ∩ Ek ↑ F and as above we see that 1F = limk 1F ∩Ek ∈ L.
Finally let f ∈ Lp (µ). By theorem 3.3 there is a sequence of elementary
functions (fn )n which increases to f + . Each fn is in L as a linear combination
of indicators of sets of finite measure. Then fn → f + in Lp (µ) by the above
extended dominated convergence and hence f + ∈ L and similarly f − ∈ L and
f ∈ L.
Theorem 7.3 Let 1 ≤ p < ∞. Let µ be a measure on B(Rd ) for which bounded
sets have finite measure. Then the continuous functions with compact support
are dense in Lp (µ). Similarly for a finite measure µ on the Borel sets of a
metric space the bounded continuous functions are dense in Lp (µ).
Proof. In the first situation let L be the closure of Cc (Rd ) in Lp (µ). Let
U be an open set of finite µ-measure. There is a sequence (fn ) in Cc (Rd ) such
46 Measure Theory (October 20, 2008)
Corollary 7.2 (Lusin’s Theorem) Let µ be as in the second part of the pre-
vious theorem. Let f be measurable and real-valued. Let A be a Borel set of
finite measure. Then for each ε > 0 there is a set K ⊂ A with µ(K) > µ(A) − ε
such that f is continuous on K.
S
Proof. We may assume that f vanishes outside A. Since A = n A ∩ {|f | <
n} we may assume that f is bounded and in particular f ∈ L1 (µ). There is a
sequence (fn ) of continuous functions which converges in L1 (µ) to f . According
to theorems 6.1 and 6.2 we can assume that fn → f µ-a.e.. By Egorov’s theorem
6.3 for each ε > 0 there is a measurable set K ⊂ A with µ(K) > µ(A) − ε such
that the fn converge uniformly to f on K. Since each fn is continuous the same
is true for the restriction of f to K.
Chapter 8
The Radon-Nikodym
Theorem
dν
The µ-equivalence class of dµ is uniquely determined by ν:
Proposition 8.1 Let ν, ν 0 be two measures with the the Radon-Nikodym den-
0
dν
sities f = dµ and f 0 = dν 0 0
dµ . Then ν ≤ ν if and only if f ≤ f µ-a.e.. Two
measurable functions are Radon-Nikodym densities of the same measure ν if and
only if they are µ-equivalent.
Proof. Suppose that ν ≤ ν 0 , yet µ({f > f 0 }) > 0. Then proposition 4.1
part b) applied to the function 1{f >f 0 } (f − f 0 ) would give the contradiction
Z Z
ν({f 0 > f }) = f dµ > f 0 dµ = ν 0 ({f 0 > f }).
{f >f 0 } {f >f 0 }
47
48 Measure Theory (October 20, 2008)
Proof. b) → a) is trivial.
a) → b). If b) does not hold then there are some ε > 0, a sequence (δn ) of positive
reals converging to 0 and a sequence (An ) of measurable sets withP µ(An ) < δn
and ν(An )T≥ εSfor all n. Without loss of generality we may
P assume n δn < ∞.
Let A = n m≥n S Am . Then µ(A) = 0 because of µ(An ) < ∞. On the
other
S hand ∞ > ν( m≥n Am ) ≥ ν(An ) ≥ ε for all n and thus ν(A) ≥ ε since
m≥n Am ↓ A as n → ∞. Thus a) fails as well.
R
class of h. Thus by `(h) = h dν a linear functional on L2 (µ) is defined which
fulfils the assumption
R of the theorem
R of Fischer-Riesz. So there is a function
f ∈ L2 (µ) with h dν = `(h) = hf dµ for all h ∈ L2 (µ). Let A ∈ F . Since µ
is finite we have 1A ∈ L2 (µ) and so
Z
ν(A) = `(1A ) = f dµ,
A
dν
i.e. the RN-density dµ is the quotient fg of the RN-densities with respect to ρ.
4. Finally let µ and ν be σ-finite with ν ¿ µ. According to lemma 6.1 there are
two strictly positive measurable functions g ∈ L1 (µ) and h ∈ L1 (ν). Then the
two measures gµ and hν (cf. (8.1)) are finite and it is easy to check that they
have the same null-sets as the measures µ and ν, respectively. Thus hν ¿ gµ
and there is a RN-density f = dhνdgµ . Then using (8.3) again we get for all A ∈ F
Z Z Z Z
1 1 1 g
ν(A) = 1A h dν = 1A dhν = 1A f dgµ = f dµ,
h h h A h
g
i.e. hf is the desired Radon-Nikodym derivative of ν w.r.t. µ.
ν = νac + νs (8.4)
R
νs (A) = ν(A ∩ Dc ). Then νs ⊥ µ since µ(Dc ) = {f =0} f dρ = 0 and νs (D) = 0.
Finally
Z Z Z
g g
νac (A) = ν(A ∩ D) = g dρ = 1D f dρ = 1D dµ.
A∩D A f A f
= f dρ + f dρ = ν(A ∩ D) + ν(A ∩ Dc )
A∩{f ≥0} A∩{f <0}
where clearly in the last two sums the first term is non-negative and the second
term is non-positive for all A. Thus ν + = ν(D ∩ ·) and ν − = −ν(Dc ∩ ·) provide
the indicated decomposition.
Every representation ν = ν + − ν − as difference of two singular measures is
minimal in the above sense: Suppose ν = κ1 − κ2 and let D be a measurable
set such that ν + (Dc ) = 0 = ν − (D). Then
Definition 8.5 The space of signed measures on the measurable space (S, F) is
denoted by M(S, F). For ν ∈ M(S, F) the measure |ν| = ν + + ν − is called the
Measure Theory (October 20, 2008) 51
dν + dν + dν dν − dν d|ν|
( ) = , ( )− = , | |= . (8.6)
dµ dµ dµ dµ dµ dµ
dν
Proof. For the uniqueness of dµ note that the proof of proposition 8.1 carries
over to our situation. In order to show ν + ¿ µ and ν − ¿ µ let µ(N ) = 0.
Then µ(N ∩ D) = µ(N ∩ Dc ) = 0 and hence ν + (N ) = ν(N ∩ D) = 0 and
+
dν dν dν −
ν − (N ) = −ν(N ∩ Dc ) = 0. The function f = dµ defined by dµ = dνdµ − dµ
has the desired properties since according to the proof of the Hahn-Jordan
decomposition f ≥ 0 on D and f < 0 on Dc .