Chapter 1
Chapter 1
MOTIVATION
Rb
In calculus, one introduces the notion of Riemann integral a f (x)dx of a real valued function
f : [a, b] → R defined on a closed bounded interval [a, b]. Continuous functions are Riemann
integrable (or more generally, functions with an at most countable set of discontinuities) as
are monotone functions (or more generally, functions of bounded variation). Despite its many
merits, unfortunately, the Riemann integral has some serious drawbacks.
For example, let (fn )n∈N be a sequence of Riemann integrable functions on [0, 1] which
converges pointwise to a function f . Then, it may happen that f is not Riemann integrable,
and even if f is Riemann integrable, it may happen that
Z 1 Z 1 Z 1
lim fn (x)dx 6= f (x)dx = lim fn (x)dx ,
n→∞ 0 0 0 n→∞
i.e. the integral of the limit is not equal to the limit of the integrals.
For a concrete example of the first phenomenon, let (rk )k∈N be an enumeration of the rational
numbers in [0, 1], i.e. {rk ; k ∈ N} = Q ∩ [0, 1] and rk 6= rl whenever k, l ∈ N, k 6= l. For n ∈ N,
the function (
1, x ∈ {r1 , . . . , rn },
fn : [0, 1] → R, x 7→
0, else
R1
is continuous except at x ∈ {r1 , . . . , rn } and thus Riemann integrable, 0 fn (x)dx = 0. Addi-
tionally, the sequence (fn )n∈N converges pointwise to the function
(
1, x ∈ Q ∩ [0, 1],
f : [0, 1] → R, x 7→
0, x ∈ (R\Q) ∩ [0, 1],
which is not Riemann integrable (upper Riemann sums are all equal to 1, lower Riemann sums
are all equal to 0)!
For a concrete example of the second phenomenon mentioned above, for n ∈ N let
fn : [0, 1] → R, x 7→ n2 x(1 − x2 )n .
Then, fn is Riemann integrable since it is continuous and by an elementary calculation we have
Z 1 Z 1
2 n2 n2
fn (x)dx = n x(1 − x2 )n dx = − (1 − x2 )n+1 |1x=0 = →n→∞ ∞.
0 0 2(n + 1) 2n + 2
On the other hand, due to fn (0) = fn (1) = 0 and
fn (x) = xn2 exp n ln(1 − x2 ) , x ∈ (0, 1),
5
6 CHAPTER 1. MOTIVATION
We would like to overcome these deficiencies of the Riemann integral by introducing a more
flexible notion of integral. To get some inspiration, we briefly recall the definition of the
Riemann integral.
For [a, b], a partition Z(a, b) is a finite subset Z(a, b) = {a0 , a1 , . . . , an } such that a = a0 <
a1 < . . . < an = b. Its width is defined as max1≤j≤n (aj − aj−1 ) and denoted by |Z(a, b)|. Given
a partition Z(a, b) = {a0 , a1 , . . . , an } of [a, b], ξ = (ξ1 , . . . , ξn ) ∈ Rn is called a subordinate
choice of supporting points if ξj ∈ [aj−1 , aj ], 1 ≤ j ≤ n. Finally, for a function f : [a, b] → R,
a partition Z(a, b) of [a, b], and a subordinate choice of supporting points ξ = (ξ1 , . . . , ξn ), the
corresponding Riemann sum is defined as
n
X
R(f, Z(a, b), ξ) := f (ξj )(aj − aj−1 ).
j=1
Then, a function f : [a, b] → R is called Riemann integrable if there is s ∈ R such that for every
ε > 0 there is δ > 0 such that for each partition Z(a, b) of [a, b] of width less than δ and every
subordinate choice of supporting points ξ it holds
i.e. in decomposing the domain of f into subintervals, evaluating f at an arbitrary point from
each subinterval, multiplying this value with the length of the corresponding subinterval and
summing up.
Alternatively, one could decompose the range of f : [a, b] → R instead of its domain and
perform a similar procedure. In order to make this a little more precise, let us assume that
our function f : [a, b] → R takes only values in [0, M ] for some M > 0. For a partition
Z(0, M ) = {m0 , m1 , . . . , mn } of [0, M ], 0 = m0 < m1 . . . < mn = M we define the pre-image of
[mj−1 , mj ), j = 1 . . . , n, under f as
so, by definition, for every x ∈ f −1 ([mj−1 , mj )) we have mj−1 ≤ f (x) < mj - we emphasize
that f −1 ([mj−1 , mj )) may be empty. However, the subset f −1 ([mj−1 , mj )) is not an interval
in general. Thus, it is not clear how to define the “length” of the set f −1 ([mj−1 , mj )). Let us
assume for the moment that we can define the length of f −1 ([mj−1 , mj )) in a reasonable way.
We denote it by ` (f −1 ([mj−1 , mj ))). Then, analogous to the definition of Riemann sums, we
define the “range decomposition sum”
n
X
mj−1 ` f −1 ([mj−1 , mj )) 1 .
L(f, Z(0, M )) := (1.1)
j=1
1
Instead of the mj−1 ’s we could additionally have chosen an arbitrary subordinate choice of supporting
7
Analogously to the Riemann integral, a function f : [a, b] → [0, M ] will be called “range
decomposition integrable”, if the limit lim|Z(0,M )|→0 L(f, Z(0, M )) exists.
In order to make the above idea rigorous, we thus need to generalize the notion of “length
of an interval” in a reasonable way to all subsets of R. More general than for a function
f : [a, b] → [0, M ], the above idea can be easily transferred to non-negative bounded functions
f defined on a (compact) subset K of Rd . Then, f −1 ([mj−1 , mj )) = {x ∈ K; f (x) ∈ [mj−1 , mj )}
and instead of the problem to extend the notion of length to an arbitrary subset of R (in order to
define ` (f −1 ([mj−1 , mj )))), we have to define the d-dimensional volume vold (f −1 ([mj−1 , mj )))
of f −1 ([mj−1 , mj )), i.e. we need to extend the notion of d-dimensional volume to arbitrary
subsets of Rd , where for a rectangle
d
Y
R = [a1 , b1 ] × · · · × [ad , bd ] = [ak , bk ]
k=1
its (d-dimensional) volume is vold (R) := dk=1 (bk − ak ). For d = 1, this is obviously the length
Q
of the “rectangle” [a, b].
Thus, we would like to extend the notion of volume from rectangles to all subsets A of Rd ,
and thereby allow infinite volume, such that the following reasonable properties hold.
(vol-iii) For every sequence (Ak )k∈N of subsets of Rd with Ak ∩ Aj = ∅, k 6= j (i.e. (Ak )k∈N is
pairwise disjoint), we have
∞
! ∞
[ X
vold Ak = vold (Ak ).
k=1 k=1
(vol-iv) Whenever C ∈ Rd×d is a orthogonal matrix and b ∈ Rd , for all A ⊆ Rd it holds vold (A) =
vold (C(A) + b), where C(A) + b = {Cx + b; x ∈ A}. (Since mappings of the form
T : Rd → Rd , x 7→ Cx + b are called motions, we request that the volume we wish to
exist should be motion invariant.)
Unfortunately, as follows from the next theorem, our wish cannot be fulfilled. In the next
theorem, we denote the power set of Rd , i.e. the set of all subsets of Rd , by P(Rd ).
Theorem. Let µ : P(Rd ) → [0, ∞] be a mapping which satisfies the following properties.
(µ.1) µ(∅) = 0.
For each sequence (An )n∈N in P(Rd ) of pairwise disjoint subsets of Rd it holds µ(∪n∈N An ) =
(µ.3) P
∞
n=1 µ(An ).
to get a complete analogue for the Riemann sum, however, we want to keep it as simple as possible in order to
elaborate the main idea.
8 CHAPTER 1. MOTIVATION
Proof. In order to prove the claim, by property (µ.1) it suffices to show that µ(Rd ) = 0. Since
the proof of this statement is quite technical, let us devide it into several steps.
Step 1. In this step we construct a rather complicated sequence (An )n∈N subset of Rd which
eventually will be used to derive µ(Rd ) = 0. In order to do so, we define an equivalence relation
on (0, 1]d as follows.
∀ x, y ∈ (0, 1]d : x ∼ y :⇔ x − y ∈ Qd ,
i.e. two vectors in Rd with entries in (0, 1] are considered to be equivalent if every component
of their difference is a rational number. It is easy to see that ∼ defines an equivalence relation
on (0, 1]d . As usual, for x ∈ (0, 1]d we denote its coset by [x], i.e. [x] = {y ∈ (0, 1]d ; x ∼ y}. As
with every equivalence relation, the set of all cosets defines a partition of (0, 1]d . We use this
equivalence relation to define a subset A of (0, 1]d . We pick precisely one element from each of
the different cosets and put them into the set A. More formally, A ⊆ (0, 1]d is defined by the
following properties
(A.2) ∀ a, b ∈ A, a 6= b : (a ∼ b ⇒ a = b).
rk + a = x = rl + a0 .
Since by step 2 the (An )n∈N are pairwise disjoint, it follows from property (µ.3) that
! ∞
[ X
µ An = µ(An ). (1.3)
n∈N n=1
Additionally, by definition An = rn +A, so that property (µ.4) implies µ(An ) = µ(rn +A) = µ(A)
for every n ∈ N. Combining this with equations (1.2), (1.3) we conclude
∞
X
d
µ(A) ≤ µ (−1, 2]d .
µ (0, 1] ≤
n=1
d
Once we show µ (−1, 2] < ∞, the second of the previous chain of inequalities
implies µ(A) =
d
0 which by the first inequality gives thedesired conclusion µ (0, 1] = 0.
Thus, it remains to show µ (−1, 2]d < ∞. In order to do so, we first observe that (−1, 2]d
equals the pairwise disjoint union of the 3d sets x + (0, 1]d , x ∈ {−1, 0, 1}d , i.e. the components
of the vector x ∈ Rd are either −1, 0, or 1. (For d = 1 we obviously have
(−1, 2] = (−1, 0] ∪ (0, 1] ∪ (1, 2] = (−1 + (0, 1]) ∪ (0 + (0, 1]) ∪ (1 + (0, 1])
and the sets on the right hand side are pairwise disjoint; for d > 1, we leave it to the reader to
do verify the claim.) Let x1 , . . . , x3d be an enumeration of {−1, 0, 1}d . We define
(
xn + (0, 1]d , if n ≤ 3d
Bn :=
∅, if n > 3d ,
so that (Bn )n∈N is a pairwise disjoint sequence of subsets of Rd with ∪n∈N Bn = (−1, 2]d . Thus,
by property (µ.3) it follows
∞ 3d ∞ 3d
X X X X
d d
µ (0, 1]d < ∞,
µ (−1, 2] = µ(Bn ) = µ xn + (0, 1] + µ(∅) =
n=1 n=1 n=3d +1 n=1
where we have used properties (µ.1) and (µ.4) in the third step of the above chain of equalities,
d
and property (µ.5) in the last. Thus, the proof of µ (0, 1] = 0 is complete.
Step 5. We now show that µ(Rd ) = 0. As explained at the beginning of the proof, this will
finish the proof of the theorem. Similar as in step 4, it is not hard to see that
[
Rd = z + (0, 1]d
z∈Zd
and that for z, z 0 ∈ Zd , z 6= z 0 , we have z + (0, 1]d ∩ z 0 + (0, 1]d = ∅. Clearly, Zd is countable,
While it is not possible to extend the notion of d-dimensional volume to all subsets of Rd ,
we shall try to extend it to a system of subsets of Rd which is as large as possible and which
contains every open subset of Rd as well as every closed, in particular, every compact, subset
of Rd . This task will lead to the notion of a σ-algebra, more precisely, the Borel σ-algebra of
Rd . σ-algebras will be introduced in Chapter 2. It will turn out that the generalization of the
d-dimensional volume we are looking for is a co-called measure (on the Borel σ-algebra on Rd ).
Measures will be studied systematically in Chapter 3. There, it will be proved that there is
at most one way to extend the notion of d-dimensional volume to the Borel σ-algebra of Rd .
The problem of existence of such an extension will be addressed in a very abstract setting in
Chapter 4. The results of Chapter 4 will be used in Chapter 5 to prove the (unique) existence
of a measure, the so-called one dimensional Lebesgue measure, on the Borel σ-algebra of R
which extends the notion of “length” of intervals to arbitrary subsets of R belonging to its
Borel σ-algebra.
In Chapter 6, we will come back to our initial intention, to introduce a more flexible no-
tion of integral than the Riemann integral for functions f : [a, b] → [0, M ] by decomposing
their range instead of their domain and by considering “range decomposition sums”, see equa-
tion (1.1), instead of Riemannian sums. However, since we saw that the notion of length
cannot be extended to all subsets of R, i.e. the one dimensional Lebesgue measure is not
defined on the power set P(R), we have to introduce a new class of functions f for which
f −1 ([mj−1 , mj )) belongs to the Borel σ-algebra of R, for arbitrary [mj−1 , mj ), so that its
(generalized)
Pn length ` (f −1 ([mj−1 , mj )))), which appears in the “range decomposition sum”
−1
j=1 mj−1 ` (f ([mj−1 , mj ))), is correctly defined. This new class of functions consists of the
so-called measurable functions. We first study systematically properties of measurable func-
tions before we define integrals with respect to an arbitrary measure of measurable functions
and thoroughly investigate their properties in Chapter 6.
With this new notion of integrals with respect to an arbitrary measure, we will be able to
extend (in a unique way) the notion of d-dimensional volume to all subsets of Rd which belong
to its Borel σ-algebra in Chapter 7, i.e. we will then prove the existence of the d-dimensional
Lebesgue measure. Having at hand the notion of an integral with respect to an arbitrary
measure since Chapter 6, this will finally allow us to introduce the notion of an integral for
(measurable) functions defined on subsets of Rd .
Certain vector spaces of integrable functions are studied in Chapter 8. These so-called Lp -
spaces turn out to be Banach spaces, i.e. normed vector spaces in which every Cauchy sequence
converges, and they appear naturally in real analysis, stochastics, functional analysis, and the
theory of partial differential equations.
An analogue of the substitution rule for Riemann integrals from calculus is proved in Chap-
ter 9 for integrals with respect to d-dimensional Lebesgue measure. This so-called change of
variables formula motivates the definition of surface measures of smooth submanifolds of Rd ,
as one encounters in vector analysis and differential geometry.
These lecture notes are intended as a self-study course on the fundamentals of real analy-
sis. Throughout each chapter you find links to several short online-exercises from the OPAL
“leveling-up” course. The intention of these online-exercises is to provide you with the oppor-
tunity to check your self-study success after you have read/studied a section from the lecture
notes by giving you an immediate feedback. In addition, at the end of each chapter you find
several problems. These are generally of a more advanced nature than the questions from the
online-exercises. While no solutions will be provided, you are strongly encouraged to try to
solve some of these problems since they will help you to get an even better grasp on the content
of the lecture notes than by only taking the online-exercises.