21-720: Measure and Integration

Fall 2023
Notes
Daniel Hathcock

Contents
1 Introduction 3

2 Riemann Integral (is insufficient) 4

2.1 Riemann Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Shortcomings of the Riemann Integral . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Measures 8
3.1 Outer Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Measurable Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Measurable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5 Simple Functions, Egorov’s, and Lusin’s Theorems (Missed for Banff) . . . . . . . . 24

4 Integration 25
4.1 Non-negative functions (Missed for Covid) . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Lebesgue Integration of Arbitrary Sign . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3 Modes of Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5 Product Spaces 37
5.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Integrals on Product Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3 Exchanging Iterated Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

6 Lp spaces 48
6.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.2 Properties of Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

1
 6.3 Topological Properties of Lp -spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

7 Duals and Decompositions of Lp spaces 57

7.1 Radon-Nikodym . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
7.2 Riesz Representation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.3 Lebesgue Decomposition Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.4 Besicovitch Derivation and Lebesgue Differentiation Theorem . . . . . . . . . . . . 68

2
Introduction
Note. Lectures will not always follow the textbook - in particular, in the beginning, lecture will
deal with function from Rn to Rd , rather than just R to R.

Note. Why do we care?

• 17th century: Newton and Leibnitz were beginning to create calculus. They had a notion of
integration by calculating area, volume, center of mass.
• 1850’s: Riemann created a notion of infinitesimal calculus. This created a more formal notion
of integration and differentiation. This gave us
ˆ x
F (x) := f (t) dt =⇒ F ′ (x) = f (x)
a

and
ˆ b
′
F (x) = f (x) =⇒ f (t) dt = F (b) − F (a).
a

• 19th century: Cauchy gave us rigorous analysis. Notions of limits/derivatives/integration/-

continuity. In this analysis, integration is valid only for continuous functions.
Question: how can we extend the notion of integration to non-continuous functions?
This brings us to the Lebesgue integral / Lebesgue measure.

Note. Today: Lebesgue integration is essential for the study of:

• Functional analysis
• Calculus of variations
• Differential equations
• Probability theory
• ...

3
Riemann Integral (is insufficient)

2.1 Riemann Integral

´b ´b
Note. Consider functions f : [a, b] → R, with a < b. We will define a
f= a
f (t) dt.
We will partition the interval [a, b] by choosing a collection P of points a = x0 < x1 < . . . < xn =
b.

Definition 2.1.1 (Lower/Upper Riemann Sum). Suppose that f is bounded. Define the lower
Riemann sum L(f, P, [a, b]) as
n
X
L(f, P, [a, b]) := (xj − xj−1 ) · inf f.
[xj−1 ,xj ]
j=1
Pn
Similarly, the upper Riemann sum is U (f, P, [a, b]) := j=1 (xj − xj−1 ) · sup[xj−1 ,xj ] f

Observation. We always have L(f, P, [a, b]) ⩽ U (f, P, [a, b]) for any partition P of [a, b]. Moreover,
the lower and upper sums are always real numbers (since f is bounded).

Proposition 2.1.1. Consider P, P ′ two partitions of [a, b], with P ′ finer than P. Then

L(f, P, [a, b]) ⩽ L(f, P ′ , [a, b]) ⩽(∗) U (f, P ′ , [a, b]) ⩽ U (f, P, [a, b]).

Proof. (*) follows from the observation. The other two inequalities are symmetric, so we prove
just the first one.
If P = {a = x0 , x1 , . . . , xn = b}, and P ′ = {a = x′0 , . . . , x′N = b}, then for any xj , we have some
xj−1 = x′k < . . . < x′k+m = xj . Therefore
m
X
(xj − xj−1 ) inf f= (x′k+i − x′k+i−1 ) inf f
[xj−1 ,xj ] [xj−1 ,xj ]
i=1
Xm
⩽ (x′k+i − x′k+i−1 ) inf f.
[xk+i−1 ,xk+i ]
i=1

Summing over j yields L(f, P, [a, b]) ⩽ L(f, P ′ , [a, b]).

Corollary 2.1.2. If P, P ′ are arbitrary partitions of [a, b], then

L(f, P, [a, b]) ⩽ U (f, P ′ , [a, b])

Proof. Simply take a refinement of both partitions P ′′ := P ∪ P ′ , and apply Proposition 2.1.1.

Definition 2.1.2 (Lower/Upper Riemann Integral). f : [a, b] → R bounded. The lower Rie-
mann integral is

L(f, [a, b]) := sup L(f, P, [a, b]),

P

4
and similarly for the upper Riemann integral U (f, [a, b]) := inf P U (f, P, [a, b]).
Moreover, f is said to be Riemann integrable if L(f, [a, b]) = U (f, [a, b]). Then we call the
intgral
ˆ b
f := L(f, [a, b]) = U (f, [a, b]).
a

Observation. We always have L(f, [a, b]) ⩽ U (f, [a, b]) (by taking inf and sup of Corollary 2.1.2).

Theorem 2.1.3. f : [a, b] → R is continuous, then f is Riemann integrable.

Remark. Note that we did not say that f is bounded, since this is implied by continuity on a
compact interval. In fact, it also implies that f is uniformly continuous.

Definition 2.1.3 (Uniform Continuity). A function f is uniformly continuous on [a, b] if for

every ε > 0, there exists δ > 0 such that for for any s, t ∈ [a, b] we have

|s − t| < δ =⇒ |f (s) − f (t)| < ε

Proof of Theorem 2.1.3. We claim that U (f, [a, b]) ⩽ L(f, [a, b]). Fix ε > 0, and choose δ from
the uniform continuity of f . There is some n ∈ N such that b−an
< δ. Define partition P = {a =
b−a
x0 , . . . , xn = b} with intervals of length n . Then

U (f, [a, b]) − L(f, [a, b]) ⩽ U (f, P, [a, b]) − L(f, P, [a, b])
n
" #
X
⩽ (xj − xj−1 ) sup f − inf
[xj−1 ,xj ] [xj−1 ,xj ]
j=1

⩽ (b − a) · ε → 0

2.2 Shortcomings of the Riemann Integral

Note. Some issues we will see:
1. Riemann integral does not handle discontinuous functions well.
2. Riemann integral does not handle unbounded functions.
3. Riemann integral does not handle limits of functions well, i.e.
ˆ b ˆ b
?
fk → f =⇒ fk → f
a a

Example 2.2.1 (1). The rational indicator function f : [0, 1] → R defined by

(
1 if x ∈ Q
f (x) =
0 ifx ̸∈ Q

5
Then for any subinterval [xj−1 , xj ] ⊆ [0, 1], we have inf [xj−1 ,xj ] f = 0 and sup[xj−1 ,xj ] f = 1. And so
for any partitions P, P ′ , we get

L(f, P, [0, 1]) = 0 < U (f, P ′ , [0, 1]) = 1

Despite the fact that intuitively, f should have 0 area under its graph, since there are very few
rational numbers.

Example 2.2.2 (2). Consider

(
√1 x>0
x
f (x) :=
0 x=0

Then for any x > 0, if we consider the subinterval [0, x], we have sup[0,x] f = +∞, so U (f, [a, b]) =
+∞. So f is not integrable.
On the other hand, if we consider the same function with f : [a, 1] → R, then the integral is
well-defined, and we have
ˆ 1
√
lim f = lim(2 − 2 a) = 2
a↓0 a a↓0

Definition 2.2.1 (Convergence of functions). X a set, and fk : X → R, f : X → R. We say {fk }

converges pointwise to f if for all x ∈ X,

fk (x) →k→∞ f (x)

We say that {fk } converges uniformly to f if for all ε > 0, there is some k0 such that ∀x ∈ X,
∀k ⩾ k0 ,

|fk (x) − f (x)| < ε

Proposition 2.2.1. If fk : [a, b] → R are continuous and converge uniformly to f on [a, b], then f
is continuous.

Proof. Exercise.
´b ´b
Example 2.2.3 (3). If fk → f pointwise, then it is not true that a
fk → a
f.
For example, enumerate the rationals on [0, 1] as {r1 , r2 , . . .}. For k ∈ N, define
(
1 x ∈ {r1 , . . . , rk }
fk (x) :=
0 o.w.
´1
Then 0 fk = 0 (check!). On the other hand, fk → f pointwise where f = 1Q is the rational
indicator. But f is not Riemann integrable!
We could even make such an example where all of fk and f are continuous (homework!).

6
Theorem 2.2.2 (“Dominated Convergence Theorem” (for Riemann integrals)). fk : [a, b] → R
are Riemann integrable, and there is some M such that |fk (x)| ⩽ M for all x ∈ [a, b], k ∈ N. Say
fk → f pointwise for some f : [a, b] → R. If f is Riemann integrable, then
ˆ b ˆ b
f = lim fk .
a k→∞ a

Remark. The proof of this is not so easy. It will instead follow as a corollary of the dominated
convergence theorem for Lebesgue integrals (which does not need the assumption that f is inte-
grable).

7
Measures

3.1 Outer Measure

Note. To improve our notion of integral, we want to have a more general way of measuring subsets
of the domain of a function f : X → R. That is, some function on P(X) := {F : F ⊆ X}.

Note. Some notation:

• We will be dealing with RN .
• For x, y ∈ RN ,
N
X
x y :=
• xi y i
i=1

√
• ∥x∥ := x•x
• B(x, r) := {y ∈ RN : ∥x − y∥ < r} (open ball of radius r > 0).
• Q(x, r) := {y ∈ RN : ∥x − y∥∞ ⩽ r} (closed cube of side-length r > 0).
• I ⊆ R interval (like [a, b], (a, b], (a, b)). If x, y ∈ I then tx + (1 − t)y ∈ I for t ∈ [0, 1].
• R ⊆ RN is a rectangle I1 × . . . × IN .

Note. We start by saying that we want the outer measure (length) of an interval I to be sup I −inf I.
More generally, for a set X and subset E ⊆ X, we assign a measure to E by “approximating” E by
“elementary sets” for which we know their measure. Let E ⊆ P(X) be the family of elementary
sets.
We need to cover every E ⊆ X with (hopefully countably S many) sets from E . For this purpose,
assume that E ⊃ {∅, X1 , . . . , Xn , . . .} such that X = ∞
n=1 Xn . Assume also that we have some
function ρ : E → [0, +∞] with ρ(∅) = 0.
Then we will define (informally) an outer measure as µ∗ : P(X) → [0, +∞] as, for E ⊆ X,
(∞ )
X
µ∗ (E) := inf ρ(En )
ES
1 ,E2 ,...
E⊆ ∞ n=1 En
n=1

Example 3.1.1. Some examples of outer measures:

1. Lebesgue Outer Measure - X = RN , E is the family of all rectangles R. Then define
ρ(R) := meas(R) = length(I1 ) × . . . × length(IN ) if R = I1 × . . . × IN . Then the outer
measure is:
(∞ ∞
)
X [
L0N := inf ρ(Rn ) : Rn rectangles, such that E ⊆ Rn
n=1 n=1

8
 2. Lebesgue-Stieltjes Outer Measure - X = I ⊆ R is an open interval. f : I → R an
increasing function, and E := {(a, b] : a, b ∈ I, a ⩽ b}. Then ρ is defined as (a, b] 7→
f (b) − f (a).
Then the Lebesgue-Stieltjes outer measure of E ⊆ I generated by f is
(∞ ∞
)
X [
µ∗f (E) := inf (f (bn ) − f (an )) : an ⩽ bn ∈ I, E ⊆ (an , bn ]
n=1 n=1

3. Hausdorff Outer Measure - Let 0 ⩽ s < +∞, and 0 < δ ⩽ ∞, X = RN . Define

Eδ := {E ⊆ RN : diam(E) < δ}. Then ρs : Eδ → [0, +∞) maps
 s
diam E
E 7→ αs
2
for some constant αs . Then define for E ⊆ RN :
(∞ ∞
)
X  diam En s [
s
Hδ (E) := inf αs : diam En < δ, E ⊆ En .
n=1
2 n=1

diam En 0


If s = 0, sum over all En ̸= ∅ and define 2
= 1 even if diam En = 0.
Observe that the function δ 7→ Hδs (E) is decreasing in δ (since Eδ′ ⊂ Eδ for δ ′ < δ). So we
may define the s-dimensional Hausdorff outer measure as
H0s (E) := lim Hδs (E).
δ↓0

To be more specific
π s/2
αs :=
Γ( 2s + 1)
´∞
where Γ is the Euler-Gamma function: Γ(t) := 0 e−x xt−1 dx (and in particular Γ(n) =
(n − 1)! for n ∈ N). Then |B(x, r)| = αN rN for an N -dimensional ball.

Definition 3.1.1 (Outer Measure). X ̸= ∅, µ∗ : P(X) → [0, +∞] is an outer measure if

(i) µ∗ (∅) = 0.
(ii) (Monotonicity) E ⊆ F implies that µ∗ (E) ⩽ µ∗ (F ).
(iii) (Countable Subadditivity) µ∗ ( ∞
S P∞ ∗
n=1 En ) ⩽ n=1 µ (En ).

Proposition 3.1.1. X ̸= ∅, E ⊆ P(X), ∅ ∈ E , and X = ∞ n=1 Xn for Xn ∈ E . Say ρ : E →

S
[0, +∞] such that ρ(∅) = 0. Then µ : P(X) → [0, +∞] (defined as before with respect to ρ) is
∗

an outer measure and for E ∈ E , we have µ∗ (E) ⩽ ρ(E).

nP o
∞ ES
1 ,E2 ,...
Proof. Recall that µ∗ (E) := inf n=1 ρ(En ) : E⊆ n=1 En . Then the final property in the state-
∞

ment of the proposition is obvious (E may be covered by itself). Now we check the properties of
outer measure:

9
 (i) µ∗ (∅) ⩽ ρ(∅) = 0.
(ii) If F ⊆ Fn and Fn ∈ E , then E ⊆ Fn , so then
S S

∞
X
∗
µ (E) ⩽ ρ(Fn )
n=1

Taking the inf over the collection {Fn }, we get µ∗ (E) ⩽ µ∗ (F ).

assume not. µ∗ (En ) < +∞ for every n. Then for every
(iii) If the RHS is +∞, we are done, so S
n, we can find Fk,n such that En ⊆ ∞k=1 Fk,n such that
∞
X ε
ρ(Fk,n ) ⩽ µ∗ (En ) +
k=1
2n
S∞ S∞ S∞
(for any ε > 0). Therefore n=1 En ⊆ n=1 Fk,n (which is a countable number of sets).
k=1

⩾ 0, ∞
P P∞ P∞ P∞
Exercise 3.1.1 (Fubini’s). For ak,n k=1 n=1 ak,n = n=1 k=1 ak,n

Therefore
∞
! ∞ ∞ X
∞ ∞  ∞
[ X X X ε X
µ∗ En ⩽ ρ(Fk,n ) = ρ(Fk,n ) = µ∗ (En ) + ⩽ ε + µ∗ (En )
n=1 n,k=1 n=1 k=1 n=1
2n n=1

(where the first equality uses Fubini’s). Now let ε ↓ 0.

Proposition
S∞ P∞ E ∈ E iff ρ is countably subadditive. That is, if
3.1.2. µ∗ (E) = ρ(E) for every
E ⊆ n=1 En for E, En ∈ E , then ρ(E) ⩽ n=1 ρ(En ).

Proof. (⇐): We already know that µ∗ (E) ⩽Sρ(E). Assume countable subadditivity. We claim
that ρ(E) ⩽ µ∗ (E) for all E ∈ E . Let E ⊆ ∞
n=1 En for En ∈ E . Then countable subadditivity
implies that
∞
X
ρ(E) ⩽ ρ(En ).
n=1

Take the inf of the RHS, and we get the claim.

(⇒) Let E ⊆ ∞ n=1 En for En ∈ E . We claim that subadditivity of ρ holds. We have
S

∞
X
ρ(E) = µ∗ (E) ⩽ ρ(En ),
n=1

the first inequality is the assumption, and the second is the definition of inf.

Note. In the textbook, the Lebesgue outer measure on R is defined as:

(∞ ∞
)
X [
|A| := inf length(In ) : In open intervals, A ⊆ In
n=1 n=1

(it uses only open intervals!). This is equivalent to our definition of L01 (A).

10
Proposition 3.1.3. |A| = L01 (A) for all A ⊆ R.

Proof. The ⩾ direction is immediate, since L01 may use any intervals, including open ones.
So weSshow that |A| ⩽ L01 (A). If the RHS is ∞, we are done, so assume not. Fix ε > 0, and let
A⊆ ∞ n=1 In , with In not necessarily open, such that
∞
X
length(In ) ⩽ L01 (A) + ε
n=1

Let an = inf In and bn = sup In , and let I˜n := (an − ε

,b
2n+1 n
+ ε
2n+1
). Then clearly
∞
[
A⊆ I˜n
n=1

and also
∞
X ∞
X
|A| ⩽ length(I˜n ) = 2ε + length(In ) = L01 (A) + 2ε.
n=1 n=1

Now let ε ↓ 0.
S∞ P∞
Proposition 3.1.4. If R, Rn are rectangles with R ⊆ n=1 Rn , then meas(R) ⩽ n=1 meas(R).
Therefore, L0N (R) = meas(R).

Proof. Homework (show that the meas of rectangles is countably subadditive).

Proposition 3.1.5. If A ⊆ R is countable, then L01 (A) = 0.

P∞ Say A = {a1 , a2 , . . .}. We can simply take In = [an , an ], having length 0. Then L0 (A) ⩽
1
Proof.
n=1 0 = 0.

Remark. Conversely, if L01 (A) = 0, this does not mean that A is countable! For example, the
Cantor set is uncountable with measure 0.

Proposition 3.1.6. In RN :
(i) Translation invariance: LnN (x + E) = L0! (E) for all x ∈ RN and E ⊆ RN .
(ii) Homogeneity: L0N (λE) = |λ|N L0N (E) for λ ∈ R, E ⊆ RN .

Note. We would like to have when A ∩ B = ∅, then µ∗ (A ∪ B) = µ∗ (A) + µ∗ (B) (of course we
have ⩽ by subadditivity). But unfortunately this is not always true.
We will have to assume the Axiom of Choice: if E ⊆ P(X) whose elements are disjoint
nonempty sets of X, then there exists a V ⊂ X such that V contains exactly one element of each
set in E .

Proposition 3.1.7 (No disjoint additivity). There exists A, B ⊆ R such that A ∩ B = ∅, but
|A ∪ B| < |A| + |B|.

11
Proof. Consider the outer Lebesgue measure |·| on the interval X = [−1, 1]. We define an equiva-
lence relation a ∼ b if a − b ∈ Q. Denote an equivalence class as [a] := {b ∈ [−1, 1] : a − b ∈ Q}.
We will invoke the axiom of choice with E = {[a] : a ∈ [−1, 1]}. This gives a set V ⊆ [−1, 1] such
that V contains exactly one element from each equivalence class. Now produce an enumeration of
the rationals in [−2, 2]. That is [−2, 2] ∩ Q = {r1 , r2 , . . .}. We claim that
∞
[
[−1, 1] ⊆ (rk + V )
i=1

Indeed, if a ∈ [−1, 1], then let v ∈ [a] ∩ V , and a − v ∈ Q ∩ [−2, 2]. So there is some rk = a − v,
and thus a ∈ rk + V .
Now, we have
∞
X
2 = |[−1, 1]| ⩽ |rk + V | (subadditivity of O.M.s)
n=1
X∞
= |V | (translation invariance)
n=1

which implies that |V | > 0. But now observe that (ri + V ) ∩ (rj + V ) ̸= ∅ if i ̸= j. And for all
n ∈ N, nk=1 (rk + V ) ⊆ [−3, 3], and so by the monotonicity of O.M.s,
S

n
[
(rk + V ) ⩽ 6
k=1

Now assume for contradiction that the outer Lebesgue measure is additive for disjoint sets. Then
Sk Pk
by induction, we have for A1 , . . . , Ak with Ai ∩ Aj = ∅, then i=1 Ai = i=0 |Ai |. But this
gives a contradiction, since |V | is a positive number for which n · |V | ⩽ 6 for any n, which is
impossible.

Note. So the outer Lebesgue measure is “too good to be true”. To resolve this issue, we must
restrict the family of sets to which we apply µ∗ . For the outer Lebesgue measure, these will be
Lebesgue measurable sets. And, by the above proof, the set V will not be Lebesgue measurable!

Remark (Impossibility of a measure on all of P(R)). We might hope for some other µ which
satisfies all of the properties we want. This is impossible!

Proposition 3.1.8. There does not exist µ : P(R) → [0, +∞] such that
(a) µ(I) = length(I) for open intervals I ⊆ R.
(b) µ( ∞
S P∞
k=1 Ak ) = k=1 µ(Ak ) if Ai ∩ Aj = ∅ for i ̸= j.

(c) µ(t + A) = µ(A) for t ∈ R, A ⊆ R.

Proof. To do this, we will show that (a), (b), and (c) imply translation invariance, monotonicity,
subadditivity, and the property that if I is a closed interval, then µ(I) = sup I − inf I. If this is
true, then the exact same proof used above shows that µ must fail to have property (b).

12
Translation invariance is precisely property (c).
For monotonicity, by (a), we have µ(∅) = 0. If A ⊆ B, then B = A ∪ (B \ A), which is a disjoint
union. Therefore by (b),

µ(B) = µ(A) + µ(B \ A) ⩾ µ(A)

For subadditivity, consider A1 , A2 , . . . (not necessarily disjoint). Define A′i by A′1 = A1 , A′2 =
A2 \ A1 ,
k−1
!
[
A′k := Ak \ Aj
j=1
S∞ S∞
which are all pairwise disjoint, and j=1 Aj = j=1 A′j . Therefore,
∞
! ∞
! ∞ ∞
[ [ X X
µ Aj =µ A′j = µ(A′j ) ⩽ µ(Aj )
j=1 j=1 j=1 j=1

where the final inequality is by monotonicity (proved above).

Finally, for a < b, we have µ([a, b]) ⩽ b − a + 2ε for every ε, by monotonicity ([a, b] ⊆ (a − ε, b + ε))
and property (a).

3.2 Measurable Spaces

Definition 3.2.1 (Measurable Set). Let X ̸= ∅, and µ∗ : P(X) → [0, +∞] and outer measure.
A set E ⊆ X is said to be µ∗ -measurable if:

∀F ⊆ X, µ∗ (F ) = µ∗ (F ∪ E) + µ∗ (F \ E) (M)

Remark. Some properties we can immediately see:

1) For all F ⊆ X, we always have µ∗ (F ) ⩽ µ∗ (F ∪ E) + µ∗ (F \ E).
2) If E is µ∗ -measurable, then if A ∩ E = ∅, then µ∗ (A ∪ E) = µ∗ (A) + µ∗ (E). This is because
(M) implies

µ∗ (A ∪ E) = µ∗ ((A ∪ E) ∩ E ) + µ∗ ((A ∪ E) \ E )
| {z } | {z }
=E =A

So we denote by M ∗ the set of µ∗ -measurable sets:

M ∗ := {E ⊆ X : E is µ∗ -measurable}

It turns out that M ∗ is a σ-algebra!

Definition 3.2.2 (σ-algebra). X ̸= ∅. We say M ⊆ P(X) is a σ-algebra if

13
 (i) ∅ ∈ M .
(ii) E ∈ M =⇒ X \ E ∈ M .
S∞
(iii) En ∈ M for n ∈ N, implies n=1 En ∈ M .

Definition 3.2.3 (Regular). X ̸= ∅, an outer measure µ∗ : P(X) → [0, +∞] is said to be

regular if for all F ⊆ X, there exists E ∈ M ∗ such that F ⊆ E and µ∗ (F ) = µ∗ (E).

Definition 3.2.4 (Measurable Space). (X, M ) with M a σ-algebra on X is called a measurable

space. E ∈ M are called M -measurable.

Proposition 3.2.1. (X, M ) a measurable space. Then

(a) X ∈ M .
(b) D, E ∈ M implies that D ∪ E, D ∩ E, D \ E ∈ M .
(c) En ∈ M implies ∞ n=1 En ∈ M .
T

Proof. (a) is a consequence of (i) and (ii).

For (b), we can use (iii) with E1 = E and E2 = D, so D ∪ E ∈ M . Moreover, X \ (D ∩ E) =
(X \D)∪(X \E). Each of these are in M by (ii), and then again (ii) and (iii) give that D∩E ∈ M .
(c) is an exercise.

Note. We will later prove Carathéodory’s Theorem: X ̸= ∅, µ∗ an outer measure. Then M ∗

is a complete σ-algebra (i.e., for any F ⊆ E with µ∗ (E) = 0, then F ∈ M ∗ ).

Example 3.2.1. Some examples of σ-algebras:

1) {∅, X} is a σ-algebra for any X.
2) P(X) is a σ-algebra.
3) M := {E ⊆ X : E countable or X \ E countable}.
4) If RN , let LnN be the Lebesgue outer measure. The corresponding M ∗ is called the σ-algebra
(by Carathéodory’s theorem) of Lebesgue measurable sets. For these sets, we simply write
L N (E) := L0N (E).
Question: In RN , are rectangles, balls, etc. Lebesgue measurable?
5) Let X be a topological space. Then β(X) is the smallest σ-algebra that contains all open
subsets of X, called the Borel σ-algebra.
Observe that this implies that F ⊆ X which are closed are also in β(X).

Proposition 3.2.2. X a set, A ⊆ P(X). Then the intersection of all σ-algebras that contain all
elements of A is the smallest σ-algebra that contains the elements of A (that is, it is a σ-algebra).

Proof. The intersection is not empty, since P(X) is a σ-algebra.

14
Let MA be the intersection, so A ⊆ MA . We just prove the three σ-algebra properties:
(i) ∅ is in all σ-algebras, so is in MA .
(ii) E ⊆ MA , and if M ′ is any σ-algebra that contains A, then E ∈ M ′ =⇒ X \ E ∈ M ′ . This
holds for all M ′ , so X \ E ∈ MA
(iii) Similar.

Note. A natural question is whether β(X) ⊆ M ∗ (for some M ∗ on a topological space)?

Definition 3.2.5 (Metric Outer Measure). For (X, d) a metric space, µ∗ an outer measure, then
µ∗ is said to be a metric outer measure if

µ∗ (E ∪ F ) = µ∗ (E) + µ∗ (F )

for all E, F ⊆ X such that dist(E, F ) := inf{d(x, y) : x ∈ E, y ∈ Y } > 0.

Proposition 3.2.3. (X, d) a metric space, µ∗ a metric outer measure. Then every Borel subset
of X is µ∗ -measurable.

Proof. β(X) is generated by closed sets, so it is sufficient to show that all closed sets are in M ∗ .
If C ⊆ X closed, then given F ⊆ X, we want to show

µ∗ (F ) = µ∗ (F ∩ C) + µ∗ (F \ C)

We know we have (⩽) by subadditivity.

Now we show (⩾). If µ∗ (F ) = ∞, then it is true. If µ∗ (F ) < ∞, write

E0 = {x ∈ F \ C : dist(x, C) ⩾ 1}
 
1 1
En = x ∈ F \ C : ⩽ dist(x, C) ⩽
n+1 n
S∞ S∞
Since C is closed, if x ∈ F \C, then dist(x, C) > 0, so x ∈ n=0 En , so F \C = n=0 En . Moreover,
if x ∈ E2k and y ∈ E2k+2 , then
1 1
⩽ d(x, C) ⩽ d(x, y) + d(y, C) < d(x, y) +
2k + 1 2k + 2
1 1
so d(x, y) ⩾ 2k+1 + 2k+2 =: η > 0. So we have proved that dist(E2k , E2k+2 ) > 0. But now using
the definition of metric outer measure, we get additivity! Therefore
∞
k
!
X [
∗ ∗
µ (E2j ) = µ E2j ⩽ µ∗ (F ) < ∞
j=0 j=0

for all k. And similarly for kj=1 µ∗ (E2j−1 ) ⩽ µ∗ (F ) < ∞. So both series are convergent, and so
P
P ∞ ∗
j=0 µ (Ej ) is also convergent.

15
Now we claim that dist(F ∩ C, nj=0 Ej ) > 0 for all n. Indeed, for x ∈ F ∩ C and y ∈ nj=1 Ej ,
S S
1
then n+1 ⩽ d(y, C) ⩽ d(y, x). We now have
∞
!
[
∗ ∗ ∗ ∗
µ (F ∩ C) + µ (F \ C) = µ (F ∩ C) + µ Ej
j=0
∞
n
!
[ [
= µ∗ (F ∩ C) + µ∗ Ej ∪ Ej
j=0 j=n+1
n
! ∞
!
[ [
⩽ µ∗ (F ∩ C) + µ∗ Ej + µ∗ Ej
j=0 j=n+1
n
! ∞
[ X
⩽ µ∗ (F ∩ C) + µ∗ Ej + µ∗ (Ej )
j=0 j=n+1
 n
[  ∞
X
= µ∗ (F ∩ C}) ∪
| {z Ej + ∗
µ (Ej )
⊆F j=0 j=n+1
| {z }
⊆F
∞
X
⩽ µ∗ (F ) + µ∗ (Ej )
j=n+1
P∞
for all n. Now let n → ∞ and use the fact that j=0 µ∗ (Ej ) converges to see that µ∗ (F ∩ C) +
µ∗ (F \ C) ⩽ µ∗ (F ).

Corollary 3.2.4. Borel sets in RN are L0N -measurable and H0s -measurable (for 0 ⩽ s < ∞).

Proof. We just show that they are metric outer measures in RN . Hausdorff measure is an exercise.
For Lebesgue measure: E, F ⊆ RN with d := dist(E, F ) > 0. We want to show that L0N (E ∪ F ) =
L0N (E) + L0N (F ). We have (⩽) always, so we show (⩾).
Let Rn ⊆ RN be rectangles such that E ∪ F ⊆ ∞
S
n=1 Rn . If necessary, partition each Rn into
d
smaller rectangles so that all have diameter < 2 .
Therefore if Rm ∩ E =
̸ ∅, then Rm ∩ F = ∅, and likewise Rm ∩ F ̸= ∅ =⇒ Rm ∩ E = ∅.
Therefore
X∞ X X
meas Rn ⩾ meas Rn + meas Rn ⩾ L0N (E) + L0N (F )
m=1 Rm ∩E̸=∅ Rm ∩F ̸=∅

Take the inf on LHS over all coverings of E ∪ F , so we get L0N (E ∪ F ) ⩾ L0N (E) + L0N (F ).

Note. So we have now proved that

β(RN ) ⊆ {Lebesgue measurable sets} ⊆ P(RN )
Might any of these be equal? Well we know the second subset is strict, since there are non-
measurable sets. It turns out the first is also strict.
The example V of a non-measurable set in R can be extended to RN by just taking V ×V ×. . .×V .

16
Note. Consider the Cantor set D (recall, we get this from removing the middle third of [0, 1], then
the middle third of each remaining closed piece, and so on). It is closed, so is measurable. It is
also uncountable, but has measure 0.
The Cantor function f : D → [0, 1] is defined to be increasing, it has the same value at the
endpoints of each interval that is removed. f extends continuously onto f : [0, 1] → [0, 1].
In HW#2, we will prove that there are Lebesgue measurable sets that are not Borel.
Also in HW#2, we will show that a function between two topological spaces f : (X, T ) → (Y, T ′ )
is continuous, then the preimage of a Borel set is Borel.
Also in HW#2, we will show that f : RN → R continuous does not necessarily map Lebesgue
measurable sets to Lebesgue measurable sets.
However, if E is Borel, then f (E) is Lebesgue measurable (though it may not be Borel). More
generally:

Theorem 3.2.5. If X is a complete separable metric space, f : X → Rd continuous (it suffices

to be Borel rather than continuous: i.e., the preimage of a Borel set is Borel). If E is Borel, then
f (E) is Lebesgue measurable.

Remark. As a corollary to all of this, one can show that there exist sets E ⊆ R which are Lebesgue
measurable, but E + E := {x + y : x ∈ E, y ∈ E} is not Lebesgue measurable.

3.3 Measurable Functions

Definition 3.3.1 (Measurable function). Let X and Y be nonempty, and M , N σ-algebras on
X and Y , respectively. f : X → Y is said to be measurable if for all F ∈ N , f −1 (F ) ∈ M .

Remark. Some conventional notation:

• When we say f : (X, M ) → RN is measurable (without specifying anything else about the
σ-algebra on RN ), we mean ∀B ⊆ RN Borel, f −1 (B) ∈ M .
• Likewise, if we say f : Rd → RN is Lebesgue measurable, we mean ∀B ⊆ RN Borel, f −1 (B)
is Lebesgue meas. in Rd .
• And if f : Rd → RN is said to be Borel meas., then ∀B ⊆ RN Borel, f −1 (B) is Borel.

Proposition 3.3.1. (X, M ), (Y, N ) with N the smallest σ-algebra that contains a family F ⊆
P(Y ). Then f : X → Y is measurable iff

∀F ∈ F , f −1 (F ) ∈ M

Proof. Let Z := {F ∈ N : f −1 (F ) ∈ M }. We know that F ⊆ Z. It suffices to prove that Z is a

σ-algebra (exercise).

Remark. So, to test whether a function f : (X, M ) → RN is measurable, we just need to check
whether the preimages of open intervals of the form (a, +∞) for every a are measurable (since

17
these open intervals generate the Borel σ-algebra).

Example 3.3.1. Some examples:

• If M = {∅, X}, then f : X → R is measurable iff f is constant.
• If M = P(X), then any f : X → R is measurable.

Proposition 3.3.2 (Continuous functions are measurable). (X, T ), (Y, T ′ ) top. spaces. Then
f : X → Y continuous implies measurable.

Proof. β(T ′ ) is generated by open sets. Let F ⊆ Y be open, then if f is continuous, f −1 (F ) is

open in X, so is in β(T ).

Proposition 3.3.3 (Increasing functions are measurable). f : R → R increasing means that f is

measurable.

Proof. We want to show that f −1 ((a, +∞)) ∈ β(R) for all a ∈ R. Well let b := inf f −1 ((a, +∞)).
Either this is empty (b = +∞), or we have

f −1 ((a, +∞)) = (b, +∞) ∈ β(R) or f −1 ((a, +∞)) = [b, +∞) ∈ β(R)

Example 3.3.2. Let (X, M ) be a meas. space. E ⊆ X, and

(
1 x∈E
χE (x) :=
0 x ̸∈ E

the characteristic function of E. This is map (X, M ) → R. Is this measurable? Well for any
F ⊆ R, we have


 E 0 ̸∈ F, 1 ∈ F

X \ E 0 ∈ F, 1 ̸∈ F
χ−1
E (F ) =


 X 0, 1 ∈ F
0, 1 ̸∈ F

∅

E ({1}) = E, we have that χE is measurable iff E ∈ M . That is, iff E is measurable!

Since χ−1

Proposition 3.3.4 (Composing measurable functions is measurable). (X, M ), (Y, N ), (Z, D) meas.
spaces. If f : X → Y , g : Y → Z both measurable, then

g◦f :X →Z

is measurable.

Proof. Easy.

18
Remark. If f : (X, β(X)) → (R, β(R)) is a Borel function, and g : R → R is Lebesgue measurable,
this does not imply that g ◦ f : X → R is measurable!!
This is because Borel measurable is a stronger requirement than Lebesgue measurable.
On the other hand, if f : (X, M ) → R is measurable (in any form), and g : R → R is Borel, then
g ◦ f is measurable.

Definition 3.3.2 (Borel subset of the extended line). Say B ⊆ [−∞, +∞] is Borel if B ∩ R is
Borel. That is, B ∩ R ∈ β(R).

Remark. f : (X, M ) → [−∞, +∞] measurable if f −1 ((a, +∞]) ∈ M for all a ∈ R.

Corollary 3.3.5. If (X, M ) a meas. space and f : E → R measurable, then f 2 , |f |, f + , f − , λf

are all measurable.

Proof. Just let g(t) = t2 , or g(t) = |t|, etc. These are all continuous, so measurable, and we can
compose with f .

Definition 3.3.3. (X, M ), (Y, N ). Define M ⊗ N ⊆ P(X × Y ) as the smallest σ-algebra that
contains E × F for all E ∈ M , F ∈ N . And of course this definition can be extended to a product
of finitely many.

Proposition 3.3.6. If we have (X, M ), (Y1 , N1 ), . . . , (Yn , Nn ) then f : X → Y1 × . . . × Yn is

measurable iff

πi ◦ f : X → Y i

is measurable for all i = 1, . . . , n (where πi is the projection (y1 , . . . , yn ) 7→ yi ).

Proof. (⇒) Observe πi is measurable. So if f is measurable, then all of its projections are mea-
surable.
(⇐) Exercise.

Proposition 3.3.7 (Operations on measurable functions). (X, M ) meas space, and f, g : X → R

measurable. Then f + g, f · g, min(f, g), max(f, g) are measurable.

Proof. Consider the function X → R2 defined by x 7→ (f (x), g(x)). This is measurable by the
previous proposition. Then we simply compose with the function g : R2 → R mapping (s, t) 7→ s+t,
for example, which is continuous, thus measurable.

Proposition 3.3.8. (X, M ) meas space, fn : X → [−∞, +∞] measurable for all n ∈ N, then
supn fn , inf n fn , lim sup fn and lim inf fn are all measurable. Also, if

E := {x ∈ X : {fn (x)} has a limit}

then E is a measurable set in M .

19
 S∞
Proof. Well (supn fn )−1 ((a, +∞]) = n=1 fn−1 ((a, +∞]), and similarly for inf n fn .
Similarly, lim supn fn = inf n (supk>n fk )), which is measurable by the above. Similarly for lim inf.
Finally, E = ∞
T S∞ T −1
n=1 j=1 k⩾j (fj − fk ) ((??)) which is measurable.

Proposition 3.3.9. (X, M ), (Y, N ) measurable. f : X → Y measurable, and g : X → Y .

µ∗ : M → [0, +∞] is a complete outer measure (i.e. if E ∈ M with µ∗ (E) = 0, then for F ⊆ E,
F ∈ M ).
Suppose that f = g almost everywhere (that is, there exists E ∈ M with µ∗ (E) = 0 such that
f (x) = g(x) for all x ̸∈ E). Then g is also measurable.

Proof. Let E be the set of measure 0 where f and g differ. Let F ∈ M . Then

g −1 (F ) = {x ∈ X \ E : g(x) ∈ F } ∪ {x ∈ E : g(x) ∈ F }
= (f −1 (F ) \ E) ∪ (g −1 (F ) ∩ E)
| {z } | {z }
∈M ⊆E =⇒ ∈M

so g is measurable.

3.4 Measures
Definition 3.4.1 (Measure). X ̸= ∅, M a σ-algebra on X. Then µ : (X, M ) → [0, +∞] is a
measure if
• µ(∅) = 0
• µ( ∞
P∞
n=1 En for En ∈ M with En ∩ Em = ∅.
S
n=1 En ) =

Then (X, M , µ) is called a measure space.

Example 3.4.1. Some examples:

1. Counting Measure: in R,
(
#E E finite
µ(E) :=
+∞ E infinite

2. Dirac measure for x0 ⊆ X, δx0 : (X, M ) → R defined by

(
1 x0 ∈ E
E 7→
0 otherwise

Proposition 3.4.1. (X, M , µ) a measure space, and D, E ∈ M with D ⊆ E, then we have

(i) Monotonicity: µ(E) ⩽ µ(E).
(ii) µ(E \ D) = µ(E) − µ(D) if µ(D) < +∞.

20
Proof. Exercise.

Proposition 3.4.2 (Countable Subadditivity). (X, M , µ) a measure space, and En ∈ M . Then

∞
[ ∞
X
µ( En ) ⩽ µ(En )
n=1 n=1

Proof. Haven’t we done this already?

Proposition 3.4.3. (X, M , µ) meas space, and E1 ⊆ E2 ⊆ . . . measurable sets. Then

∞
[
µ( En ) = lim µ(En )
n→∞
n=1

Proof. Exercise.

Proposition 3.4.4. (X, M , µ) meas space, and E1 ⊇ E2 ⊇ . . . measurable sets and µ(E1 ) < +∞.
Then
∞
\
µ( En ) = lim µ(En )
n→∞
n=1

Proof. Exercise.

Remark. Why do we need µ(E1 ) < +∞ above? Well let En = [n, +∞). ObserveT that µ(E1 ) = +∞,
but otherwise the assumptions of the above proposition hold. But we have ∞
n=1 En = ∅, whereas
lim µ(En ) = ∞.

Proposition 3.4.5. If D, E ∈ M , µ(D ∩ E) < ∞, then µ(D ∪ E) = µ(D) + µ(E) − µ(D ∩ E).

Proof. Exercise.

Theorem 3.4.6 (Carathéodory’s Theorem). X ̸= ∅, µ∗ : P(X) → [0, +∞] an outer measure.

M ∗ the set of µ∗ -measurable sets. Then
1. M ∗ is a σ-algebra.
2. µ∗ : M ∗ → [0, +∞] is a complete measure, i.e. if E ∈ M ∗ with µ∗ (E) = 0, then any
F ⊆ E is also in M ∗ .

Proof. (Step 1): we claim that ∅ ∈ M ∗ . We know that µ∗ (∅) = 0, so let F ⊆ X. Then
µ∗ (F ∩ ∅) + µ∗ (F \ ∅) = µ∗ (F ) clearly.
(Step 2): we claim that E ∈ M ∗ implies X \ E ∈ M ∗ . Fix F ⊆ X. Then

µ∗ (F ∩ (X \ E)) + µ∗ (F \ (X \ E)) = µ∗ (F \ E) + µ∗ (F ∩ E) = µ∗ (F )

21
(Step 3): first, we claim that if E1 , E2 ∈ M ∗ , then E1 ∪ E2 ∈ M ∗ . Well fix F ⊆ X, and we want
to show µ∗ (F ) ⩾ µ∗ (F ∩ (E1 ∪ E2 )) + µ∗ (F \ (E1 ∪ E2 )). If µ∗ (F ) = +∞, we are done, so assume
not. Now we have

µ∗ (F ) ⩾ µ∗ (F ∩ E1 ) + µ∗ (F \ E1 )
⩾ µ∗ (F ∩ E1 ) + µ∗ ((F \ E1 ) ∩ E2 ) + µ∗ ((F \ E1 ) \ E2 )
 
⩾ µ (F ∩ E1 ) ∪ (F \ E1 ) ∩ E2 + µ∗ ((F \ E1 ) \ E2 )
∗

 
∗
= µ F ∩ (E1 ∪ E2 ) + µ(F \ (E1 ∪ E2 ))

Now by induction we have finite unions are also in M ∗ .

(Step 4): we claim finite additivity holds. If E1 , E2 ∈ M ∗ , with E1 ∩ E2 = ∅, then let F ⊆ X.
Since
     
∗ ∗ ∗
µ F ∩ (E1 ∪ E2 ) = µ F ∩ (E1 ∪ E2 ) ∩ E1 + µ F ∩ (E1 ∪ E2 ) \ E1

we have
 
∗
µ F ∩ (E1 ∪ E2 ) = µ∗ (F ∩ E1 ) + µ∗ (F ∩ E2 )

In particular, if F = X, then we get additivity between E1 and E2 . Again, by induction we now

have finite additivity.
(Step 5): let En ∈ M ∗ mutually disjoint. We know that for every m, m n=1 En ∈ M . Fix F ⊆ X.
∗
S
Then
 [m   m
[ 
∗ ∗ ∗
µ (F ) = µ F ∩ ( En ) + µ F \ ( En )
n=1 n=1
m
X  m
[ 
∗ ∗
= µ (F ∩ En ) + µ F \ ( En )
n=1 n=1
m
X  ∞
[ 
∗ ∗
⩾ µ (F ∩ En ) + µ F \ ( En ) (*)
n=1 n=1

Now letting m → ∞, we get

∞
X  ∞
[ 
∗ ∗ ∗
µ (F ) ⩾ µ (F ∩ En ) + µ F \ ( En )
n=1 n=1
 ∞
[   ∞
[ 
⩾ µ∗ F ∩ ( ∗
En ) + µ F \ ( En )
n=1 n=1
S∞
M ∗ . Moreover, we also get immediately from (*) and letting F = ∞
S
hence n=1 En ∈ n=1 En that
∗
S ∞ P ∞ ∗
µ ( n=1 En ) = n=1 µ (En ) (countable additivity).

22
(Step 6): if En ∈ M ∗ not necessarily mutually disjoint, then we still get that their countable
union is in M ∗ by “disjointifying”.
(Step 7): we finally show completeness. Let E ∈ M with µ∗ (E) = 0, and E ′ ⊆ E. Fix F ⊆ X.
We want to show µ∗ (F ) = µ∗ (F ∩ E ′ ) + µ∗ (F \ E ′ ). Well

µ∗ (F
| ∩
′
{zE}) + µ
∗
(F \ E ′ ) ⩽ µ∗ (F )
| {z }
⊆E ⩽µ∗ (F )

and we already know (⩾) by subadditivity. So we are done.

Note. Recall: there exists f continuous and E measurable such that f (E) is not measurable.
Here is one way to show that: take a nonmeasurable set E ⊆ [0, 1], and the continuous function
f : [0, 1] → [0, 1] which maps onto the Cantor set D (this is the Cantor function). Recall
L N (D) = 0.
So let A := D ∩ f −1 (E). Then L0N (A) = 0, so by completeness of L0N , we have A is measurable.
But also by onto, f (A) = E which is nonmeasurable!

Theorem 3.4.7 (Borel Regularity). A ⊆ RN , then

1. (Outer regularity) L0N (A) = inf{L0N (G) : A ⊆ G, G open}
2. (Inner regularity) If A Lebesgue measurable, then L0N (A) = sup{L0N (K) : K ⊆ A, K compact}.

Proof. We look at each part:

1. If L0N (A) = ∞, we are done, so assume
S not. We want
P∞ to show (⩾). Fix ε > 0, and consider
rectangles Rn ⊆ RN such that A ⊆ ∞ R
n=1 n and n=1 meas(R n ) ⩽ L 0
N
(A) + 2ε .
ε
Let Sn be an open rectangle such that Rn ⊆ Sn and meas(Sn ) ⩽ meas(Rn ) + 2n
. Then we
have
∞
[
A⊆ Sn := G
n=1

and G is open. But then

∞
X ε
inf{L0N (G) : A ⊆ G, G open} ⩽ L (G) ⩽
N
meas(Rn ) + ⩽ L0N (A) + ε
n=1
2

and let ε → 0.
2. We proceed in steps: (Step 1): if A is bounded and measurable, then A ⊆ B(0, R) (some
open ball of radius R), so L N (A) < +∞. Moreover, there is some compact F with B(0, R) ⊇
F ⊇ A.
Fix ε > 0. By outer regularity, there exists open G such that F \ A ⊆ G, and

L N (G) ⩽ L N (F \ A) + ε. (**)

23
 Since A is measurable, L N (F ) = L N (F \ A) + L N (F ∩ A), so

L N (A) = L N (F ) − L N (F \ A) ⩽ L N (F ) − L N (G) + ε

by (**). Let K := F \ G, which is compact and K ⊆ A. Moreover, as F ⊆ K ∪ G, we have

L N (A) ⩽ L N (K) + L N (G) − L N (G) + ε = L N (K) + ε

Now let ε → 0, and we get inner regularity for bounded measurable A.

(Step 2): now suppose A is an unbounded measurable set. Define for each k: Ak :=
A ∩ B(0, k), which is measurable and bounded. Moreover Ak ↗ A. Therefore, by Proposi-
tion 3.4.3, L N (Ak ) → L N (A).
If L N (A) = ∞, then L N (Ak ) → ∞, and by Step 1, there are Kk ⊆ Ak ⊆ A compact such
that L N (Kk ) + 1 ⩾ L N (Ak ), so L N (Kk ) → ∞. Therefore the sup is infinite.
Otherwise, if L N (A) < ∞, then fix ε > 0. There exists a k such that L N (A) ⩽ L N (Ak )+ 2ε .
Then let K ⊆ Ak ⊆ A compact with L N (Ak ) ⩽ L N (K) + 2ε (by Step 1). Then L N (A) ⩽
L N (K) + ε, and we let ε → 0.

3.5 Simple Functions, Egorov’s, and Lusin’s Theorems (Missed for

Banff )
Whoops (See Olha’s very nice notes)

24
Integration

4.1 Non-negative functions (Missed for Covid)

Whoops again. It is the definition and properties of integral. Can’t be that important, right? (See
Olha’s notes again)

4.2 Lebesgue Integration of Arbitrary Sign

Note. We have f : X → [−∞, ∞] now. Define f + := max(f, 0) and f − := max(−f, 0). Then
clearly f = f + − f − and |f | = f + + f − .
Observe that f is measurable if and only if f ± are measurable.

Definition 4.2.1 (Lebesgue Integrable).

´ (X, M, µ)
´ measure space, f : X → [−∞, ∞] measurable.
If at least one of the two integrals X f + dµ and X f − dµ is finite, then we define the Lebesgue
integral of f to be
ˆ ˆ ˆ
f dµ := +
f dµ − f − dµ
X X X

If both are finite, then f is said to be integrable. If X is a topological space then f is said to be
locally integrable if f |K is integrable for all K ⊆ X compact.
´
Remark. f : X → [−∞, ∞] is Lebesgue integrable iff X |f |dµ < ∞.
´ + ´ −
Proof. (⇒) if f is integrable,
´ then X
f dµ and X
f dµ are both finite, so clearly their sum is
finite. This is exactly X |f |dµ.
´ ´
(⇐) If X |f |dµ < ∞, then note that f ± ⩽ |f |, so each integral X f ± dµ is finite by monotonicity.
Thus f is integrable.
1 1
´ 4.2.2 (L ). We write L (X) := {f : X → [−∞, ∞] | f is integrable}. We write
Definition
∥f ∥L1 := X |f |dµ.

Note. Recall that if f : [a, b] → R is continuous, then f is Riemann integrable. Here, we will show
that this is “almost” an “if and only if” statement. We won’t prove it now.

Theorem 4.2.1 (Lebesgue-Vitali Theorem). f : [a, b] → R bounded. Then f is Riemann inte-

grable iff
L 1 ({x ∈ [a, b] : f not continuous at x}) = 0.
Furthermore, f is Riemann integrable implies that f is Lebesgue integrable, and
ˆ b ˆ
f (x)dx = f dL 1
a [a,b]

where the left is a Riemann integral, and the right is a Lebesgue integral.

25
 n 1 N N
´ f : R → R, f ∈ L (R ), let ε > 0. There exist g : R → R continuous such
Proposition 4.2.2.
that ∥f − g∥L1 = Rn |f − g|dx < ε.
´
Proof. f = f + − f − . We know RN
f ± dµ < +∞. Fix ε > 0, then there exist si : RN → R simple
for i = 1, 2 such that
ˆ ˆ
+ ε
f dx ⩽ s1 dx +
RN RN 4
ˆ ˆ
− ε
f dx ⩽ s2 dx +
RN RN 4
Pn1 Pn2
We can write s1 , s2 in standard form: s1 = k=1 ak χAk and s2 = k=1 bk χBk . And we know that
n1
X ˆ
ak L (Ak ) =
N
s1 dx < ∞
k=0 Rn

(assume that all ak > 0). Therefore L N (Ak ) < ∞ for each k. By inner and outer regularity,
(1) (1) (1) (1) (1) (1)
we have some Gk open and Fk compact such that Fk ⊆ Ak ⊆ Gk with L N (Gk \ Fk ) <
ε 1
4n1 2 max{ak }
.
(1) (1)
Now use Urysohn’s lemma: there exists some φk : RN → [0, 1] continuous such that φk |F (1) ≡ 1
k
(1) (2) (2) (2)
and φk |RN \G(1) ≡ 0. We can do the same for s2 (get Gk , Fk , φk ).
k

Then
n1
X n2
X
(1) (2)
g := ak φk − bj φ j
k=1 j=1

26
 ´
is continuous. We claim that RN |f − g|dx < ε. Indeed:
ˆ ˆ Xn1 n2
X
!
(1) (2)
|f − g|dx = f+ − f− − ak φ k − bj φj
RN RN k=1 j=1
ˆ " n1
X n2
X
#
(1) (2)
⩽ f+ − ak φk + f− − bj φj
RN k=1 j=1
ˆ ˆ n1
X n1
X
+ (1)
⩽ f − s1 dx + ak χ A k − ak φk dx
RN RN k=1 k=1
ˆ ˆ n2
X n2
X
− (2)
+ f − s2 dx + bk χBk − bk φk dx
RN RN k=1 k=1
n1 ˆ n2 ˆ
ε X (1)
X (2)
⩽ + ak χAk − φk dx + bk χBk − φk dx
2 k=1 RN k=1 RN
1 n ˆ n2 ˆ
ε X (1)
X (2)
⩽ + ak χ Ak − φk dx + bk χBk − φk dx
2 k=1 (1) (1)
Gk \Fk | {z } k=1
(2) (2)
Gk \Fk
⩽2
n1 n2
ε X (1) (1)
X (2) (2)
⩽ + ak · 2L N (Gk \ Fk ) + bk · 2L N (Gk \ Fk )
2 k=1 k=1
⩽ε
Example 4.2.1. Let f : [π, +∞) → R defined by x 7→ sinx x which is bounded and continuous. We
claim this is not Lebesgue integrable! Indeed:
ˆ ∞ ∞ ˆ 2kπ ∞ ˆ 2kπ ∞
sin x X sin x X 1 X 1
dx = dx ⩾ |sin x|dx = 2 =∞
π x k=1 (2k−1)π
x k=1
(2k − 1)π (2k−1)π k=1
(2k − 1)π
´ℓ
But on the other hand, we can show that limℓ→∞ π sinx x dx is finite! Indeed,
ˆ ℓ  ℓ ˆ ℓ ˆ ℓ
sin x − cos x cos x 1 cos x
lim dx = lim − 2
dx = − lim dx
ℓ→∞ π x ℓ x π π x π ℓ→∞ π x2
´ℓ x
We claim that ℓ 7→ π cos x2
dx is a Cauchy sequence. For t < s, we have
ˆ t ˆ s ˆ s ˆ s
cos x cos x cos x 1 1 1
2
dx − 2
dx = 2
dx ⩽ 2
dx = − + → 0
π x π x t x t x s t
as s, t → ∞.

Proposition 4.2.3 (More properties of integrals). (X, M, µ) meas space, f, g : X → [−∞, ∞]

measurable, α, β ∈ R. Then
(i) If f, g integrable, then αf + βg is also. And
ˆ ˆ ˆ
(αf + βg)dµ = α f dµ + β gdµ
X X X

27
 ´ ´
(ii) f, g are integrable, f ⩾ g µ-a.e. then X f dµ ⩽ X gdµ.
´ ´ ´ ´
(iii) X f dµ ⩽ X |f |dµ. (this follows because f ⩽ |f |, so X f dµ ⩽ X |f |dµ).
(iv) If f is integrable, then

µ({x ∈ X : |f (x)| = +∞}) = 0

| {z }
=:E

(again, this follows because |f | ⩾ χE |f |).

´ ´
(v) f = g µ-a.e. then f is integrable iff g is, and X
f dµ = X
gdµ.

Proof. Exercise.

Corollary 4.2.4. (X, M, µ) meas space. Then (L1 (X), ∥·∥L1 ) (also denoted just ∥·∥1 ) is a normed
space:
• ∥f ∥1 = 0 ⇐⇒ f = 0 µ-a.e.
• ∥λf ∥1 = |λ|∥f ∥1 for λ ∈ R.
• ∥f + g∥1 ⩽ ∥f ∥1 + ∥g∥1 .

Theorem 4.2.5 (Lebesgue Dominated Convergence Theorem). (X, M, µ) meas space, with fn :
X → [−∞, ∞] meas, and the limit limn→∞ fn (x) =: f (x) exists µ-a.e. Assume that there exists an
´ → [0, ∞] such that |fn (x)| ⩽ g(x) µ-a.e for each n. Then f is Lebesgue
integrable function g : X
integrable and limn→∞ X |fn − f |dµ = 0.
´ ´
In particular, X fn dµ → X f dµ.

Remark. Recall from lecture 1: fn : [a, b] → R Riemann integrable, and |fn (x)| ⩽ M for all x ∈
´b ´b
[a, b]. Then there exists limn→∞ fn (x) =: f (x). If f is Riemann integrable, then a f = lim a fn .
We can now prove this:

Proof. fn bounded and Riemann integrable implies that the fn are Lebesgue integrable (by Lebesgue-
´b ´b
Vitali). Let g ≡ M . Then by the LDCT, we have limn→∞ a fn = a f .

Proof of LDCT. Modify, if needed, fn and f on a set of measure 0 so that limn→∞ fn (x) = f (x)
and |fn (x)| ⩽ g(x) for all x ∈ X. By Fatou’s Lemma:
ˆ ˆ
|f |dµ = lim inf |fn |dµ
X X n→∞ ˆ
⩽ lim inf |fn |dµ
n→∞
ˆ X

⩽ gdµ < ∞
X

28
Thus f is integrable. Thus
ˆ ˆ ˆ
gdµ ± f dµ = (g ± f )dµ
X X ˆ X

= lim inf(g ± fn )dµ

X ˆ
⩽ lim inf (g ± fn )dµ (Fatou’s again)
ˆ X ˆ
= gdµ + lim inf (±fn )dµ
X X
´ ´ ´
The two X gdµ cancel, and we have showed that ± X f dµ ⩽ lim inf X (±f )dµ. Because of the
±, we have showed
ˆ ˆ ˆ ˆ
f dµ ⩽ lim inf fn dµ ⩽ lim sup fn dµ ⩽ f dµ
X X X X
´ ´
so they are all equal, and lim X
fn dµ = X
f dµ.
Finally, let’s consider |fn − f | ⩽ 2g. We still have 2g is integrable. We know that |fn − f | → 0
pointwise. So we may apply what we just proved to these new functions. We have proved that
ˆ ˆ
lim |fn − f |dµ = 0dµ = 0
X X

Example 4.2.2. We observe that the existence of g is fundamental! Consider the interval
´ [0, 1],
and take the functions fn = nχ[0,1/n] . Then fn → 0 pointwise a.e. on [0, 1]. Moreover [0,1] fn dx =
´
1 ̸= 0 = [0,1] lim fn dx

Proposition 4.2.6. (X, M, µ) meas space, fn : X → [−∞, ∞] meas functions such that
∞ ˆ
X
|fn |dµ < ∞ (*)
n=1 X

Then the series ∞

P P∞
n=1 fn (x) converges for µ-a.e. x ∈ X. The function f (x) := n=1 fn (x) µ-a.e.
is integrable, and
∞ ˆ
X ˆ
fn dµ = f dµ
n=1 X X

Proof. Recall
∞ ˆ
X ˆ X
∞
|fn |dµ = |fn |dµ (**)
n=1 X X n=1

P the monotone convergence theorem, since |fn | are non-negative) . By (*), the RHS is < ∞.
(using
So ∞ n=1 |fn |(x) < ∞ µ-a.e. Therefore, the series defining f converges (µ-a.e.), as desired.

29
 fn (x) implies that f is measurable. Let g(x) := ∞
Pk P
And f (x) = limk→∞ n=1 n=1 |fn |(x) ⩾ 0. By
Pk
(**), g is integrable. Now n=1 fn (x) ⩽ g(x) µ-a.e. for all k. By the LDCT, we have

ˆ X
k ˆ X
∞ ˆ
lim fn dµ = fn dµ = f dµ
k→∞ X n=1 X n=1 X

Pk ´ P∞ ´
and the LHS is equal to limk→∞ n=1 X
fn dµ = n=1 X
fn dµ.

4.3 Modes of Convergence

Definition 4.3.1. (X, M, µ) meas space. Let fn , f : X → R meas functions.
(i) fn → f pointwise µ-a.e. if there is some E ∈ M with µ(E) = 0 such that limn→∞ fn (x) =
f (x) for all x ̸∈ E.
(ii) fn → f almost uniformly if ∀ε > 0 there is a set E ∈ M with µ(E) < ε such that fn → f
uniformly on X \ E. That is

lim sup |fn (x) − f (x)| = 0

n→∞ x∈X\E

(iii) fn → f in measure if for all ε > 0,

lim µ({x ∈ X : |fn (x) − f (x)| > ε}) = 0

n→∞

´
(iv) fn → f in L1 (X) if ∥fn − f ∥L1 = X
|fn − f |dµ → 0.
(v) fn converges weakly in L1 (X) (written fn ⇀ f ) if for all g : X → R measurable and
bounded,
ˆ ˆ
fn gdµ → f gdµ
X X

Example 4.3.1. X = [0, 1], and let fn (x) = xn . We claim that fn → f for
(
0 x ∈ [0, 1)
f (x) =
1 x=1

almost uniformly, but not uniformly. Well clearly the convergence is not uniform, since the fn are
uniformly continuous, but f is not continuous.
On the other hand, fix ε > 0, and let E = [1 − ε, 1]. Then |E| = ε, and moreover

sup |fn (x) − f (x)| = (1 − ε)n → 0

x∈X\E

Note. We have the following implications:

30
 • fn → f in L1 (X) implies fn → f in measure. Which implies that there exists a subsequence
{fnk } such that fnk → f almost uniformly. This implies fnk → f pointwise a.e.
• fn → f almost uniformly implies fn → f in measure.
• fn → f in L1 (X) implies fn → f weakly in L1 (X).

Lemma 4.3.1 (Chebyshev Inequality). f ∈ L1 (X), c > 0. Then

ˆ
1
µ({x ∈ X : |f (x)| ⩾ c}) ⩽ |f |dµ
c X

Proof. Let E := {x ∈ X : |f (x)| ⩾ c}. Then

ˆ ˆ ˆ
|f |dµ ⩾ |f (x)|dµ ⩾ cdµ = cµ(E)
X E E

Theorem 4.3.2. (X, M, µ) a meas space, fn , f : X → R meas functions.

(i) If fn → f almost uniformly, then fn → f pointwise a.e.
(ii) If fn → f in measure, then there exists a subsequence {fnk }k∈N such that fnk → f almost
uniformly.
(iii) If fn → f in L1 (X), then fn → f weakly in L1 (X).
(iv) If fn → f in L1 (X), then fn → f in measure.
(v) If fn → f almost uniformly, then fn → f in measure.

Proof. Let’s prove each one:

(iv) Let ε > 0. Then by Chebyshev:
ˆ
1
µ({x ∈ X : |fn (x) − f (x)| > ε}) ⩽ |fn − f |dµ
ε X

But by assumption, we have L1 -convergence, so the RHS → 0 as n → ∞.

(v) Fix ε > 0. Then for any δ > 0, by almost uniform convergence we have some set E with
µ(E) < δ such that fn → f uniformly on X \ E. In particular, there is some n0 such that
for n ⩾ n0 , {x ∈ X : |f (x) − fn (x)| > ε} ⊆ E, so
µ({x ∈ X : |f (x) − fn (x)| > ε}) ⩽ µ(E) < δ
Since this holds for any δ, we have µ({x ∈ X : |f (x) − fn (x)| > ε}) → 0 as desired.
(i) fn → f almost uniformly, we WTS fn → f pointwise. For all k ∈ N, find Ek ∈ M such
T∞ that
1
µ(Ek ) < k such that fn → f uniformly on X \ Ek . WLOG, Ek+1 ⊆ Ek . Let E0 := k=1 Ek .
Then
µ(E0 ) = lim µ(Ek ) = 0
k

If x ̸∈ E0 , then there is some k so that x ∈ X \ Ek , where we have uniformly convergence.

So clearly we also have pointwise convergence in X \ Ek , and in particular at x.

31
 (ii) Suppose fn → f in measure. We want to find a subsequence converging almost uniformly.
For all k, limn→∞ µ({x ∈ X : |fn (x) − f (x)| > 21k }) = 0. There exists nk such that for all
n ⩾ nk ,
1 1
µ({x ∈ X : |fn (x) − f (x)| > k
}) ⩽ k+1
2 2
(we can assume the nk are increasing
1
P {nk }. LetP
S∞in k). Then take the sequence Ek := {x ∈ X :
1
|fnk (x) − f (x)| > 2k }, and Fk := j=k Ek . Observe µ(Fk ) ⩽ j⩾k µ(Ej ) ⩽ j⩾k 2j+1 = 21k .
Fix ε > 0, then we claim there exists E with µ(E) < ε and
lim sup |fnk (x) − f (x)| = 0
k→∞ x∈X\E

1
Let k be such that 2k
< ε. Then let E = Fk (observe µ(Fk ) ⩽ ε). Indeed,
1
sup |fnk (x) − f (x)| ⩽ →0
x∈X\Fk 2k
as k → ∞. This shows uniform convergence on X \ Fk .
(iii) For any bounded measurable g, say |g(x)| ⩽ M for every x. Then
ˆ ˆ ˆ ˆ
fn g dµ − f g dµ ⩽ |g| · |fn − f | dµ ⩽ M · |fn − f | dµ → 0
X X X X
´ ´
and hence X
fn g dµ → X
f g dµ as desired.

Remark. We can also show that if fn → f in L1 (X), and also f ∈ L1 (X), then in addition to
a subsequence that converges pointwise, we can also get some g : X → [0, ∞] such that g is
integrable, and |fnk (x)| ⩽ g(x) µ-a.e.
´
Proof. If we define w(x) := ∞ 1
P
k=1 |f n k
(x) − f (x)|, where n k is such that |f − f | dµ <
X nk 2k
. This
gives that w(x) < ∞ a.e. Then let g(x) := w(x) + |f (x)|. Then we have
|fnk (x)| ⩽ |fnk (x) − f (x)| + |f (x)| ⩽ w(x) + |f (x)| = g(x)
Remark. Egoroff’s theorem says that pointwise convergence in a space with µ(X) < ∞ does imply
almost uniform convergence.

Example 4.3.2. Pointwise convergence does not imply convergence in measure, nor convergence
almost uniformly (so clearly does not imply L1 (X) convergence).
Let X = R, M = β(R) and µ = L 1 . Then fn (x) := χ[n,∞) (x). So for all x, fn (x) → 0. But
fn ̸→ 0 in measure, since for 0 < ε < 1, we have
L 1 ({x ∈ X : |fn (x) − 0| > ε}) = ∞
for all n.
Also fn ̸→ 0 almost uniformly (i.e. for all ε > 0, there exists G ∈ β(R) with L 1 (G) < ε such that
fn → 0 uniformly on R \ G). But for any G ∈ β(R) with L 1 (G) < ε, we have L 1 ([n, ∞) \ G) >
L 1 ([n, ∞)) − L 1 (G) = ∞. Let x ∈ [n, ∞) \ G, then supx∈R\G |fn (x) − 0| ⩾ |fn (xn )| = 1 ̸→ 0.

32
Example 4.3.3. Also, convergence in measure (or L1 (X)) does not imply convergence pointwise
a.e. Let X = [0, 1) now, and M = β(R).
She writes / erases too damn fast...Couldn’t catch this :(

Definition 4.3.2. (X, M, µ) meas space, F a family of measurable function f : X → [−∞, ∞].
F is said to be equi-integrable if ∀ε > 0, ∃δ > 0 such that wherever E ∈ M, µ(E) < δ,
ˆ
sup |f | dµ < ε
f ∈F E

Remark. Equi-integrability does not imply integrability!

ε
Suppose that exists M ∈ [0, ∞) with |f (x)| ⩽ M µ-a.e. for all f ∈ F. Fix ε > 0, and δ := M
. If
E ∈ M with µ(E) < δ, then
ˆ
sup |f | dµ ⩽ M µ(E) ⩽ ε
f ∈F E

But if, say, f (x) = 1, it is bounded by a constant, but is not integrable over all of R (its integral
is infinite).

Remark. But the converse is true: if f ∈ L1 (X), then the singleton

´ family {f } is equi-integrable.
Let En := {x ∈ X : |f (x)| > n}, and fn := f χEn . Note that X |f | dµ < ∞, so |f (x)| < ∞ a.e. So
WLOG we can say f : X → R.
Therefore fn (x) → 0 pointwise. And since |fn (x)| ⩽ |f | ∈ L1 , then by LDCT,
ˆ ˆ ˆ
lim |f | dµ = lim |fn | dµ = 0=0
n→∞ En n→∞ X X
´
So we have that there exists nε such that ∀n ⩾ nε , En
|f | dµ < 2ε . Let δ = ε
2nε
, and consider
E ∈ M with µ(E) < δ. Then
ˆ ˆ ˆ
|f | dµ ⩽ |f | dµ + |f | dµ
E Enε E\Enε
ε
⩽ + nε µ(E \ Enε )
2
⩽ε

Theorem 4.3.3 (Vitali’s Convergence Theorem). (X, M, µ) meas space, fn , f : X → R integrable.

Then fn → f in L1 if and only if all of the following hold
(i) fn → f in measure
(ii) {fn } equi-integrable.
´
(iii) For all ε > 0, ∃G ∈ M such that µ(G) < ∞ and supn X\G
|fn | dµ ⩽ ε.

Remark. (iii) above is trivial if µ(X) < ∞ (just take G = X).

33
Also, we already know that f in L1 implies that f satisfies (ii). To see f also satisfies (iii), take
Gn := {x ∈ X : n1 ⩽ |f (x)| ⩽ n}. Then χGn ∪{f =0} |f | ↗ |f |. So by LMCT, we have
ˆ ˆ ˆ
|f | = |f | → |f |
Gn ∪{f =0} Gn X
´ ´
So X\Gn
|f | dµ → 0. By Chebyshev, µ(G) ⩽ n X
|f | < ∞.
´
Proof. Assume that fn → f in L1 (i.e. X
|fn − f | dµ → 0).
(i) We already proved convergence in L1 implies convergence in measure.
(ii) Suppose that {fn } not equi-integrable. Then ´there exists some ε > 0 so that for all δ > 0,
∃G with µ(G) < δ and n = n(ε, δ) such that G |fn | dµ ⩾ ε. Consider this ε > 0, let δ = k1 .
Exists Gk with µ(Gk ) < k1 and nk so that
ˆ
|fnk | dµ ⩾ ε
Gk

Then fnk → f in L1 . Extract further a subsequence (relabel) and some g : X → R integrable

so that |fnk (x)| ⩽ g(x) a.e. Since {g} is equi-integrable, we get that there exists some δ > 0
so that ∀G with µ(G) < δ,
ˆ
g dµ < ε
G

For k >> 1 such that k1 < δ, µ(Gk ) < 1

k
< δ, so
ˆ ˆ
ε⩽ |fnk | dµ ⩽ g dµ < ε
Gk Gk

which is a contradiction.
´
(iii) Fix ε > 0. Let n0 >> 1 such that X |fn0 − f | dµ < 2ε . Since f1 , . . . , fn0 ∈´ L1 (X), for each
one we can find some measurable Gi ∈ M with µ(Gi ) < ∞ such that X\Gi |fi | dµ < ε.
´
Similarly, we can find some Gf with X\Gf |f | dµ < 2ε .
Set G := Gf ∪ G1 ∪ . . . ∪ Gn0 −1 . Then still µ(G) < ∞. And for n ⩽ n0 − 1, we have
ˆ
|fn | dµ < ε.
X\G

Whereas for n ⩾ n0 , we have

ˆ ˆ ˆ
ε ε
|fn | dµ ⩽ |fn − f | dµ + |f | dµ ⩽ + =ε
X\G X\G X\G 2 2

Now we must prove the other direction. Suppose

´ that (i), (ii), and (iii) hold. Now fix ε > 0. By
(ii), ∃δ > ´0 so that if µ(E) < δ, then supn E |fn | dµ < 2ε . By (iii), ∃Eε so that µ(Eε ) < ∞ such
that supn X\Eε |fn | dµ ⩽ 2ε .

34
We know that f ∈ L1 , ´so satisfies (ii) and (iii). Then we can make δ potentially smaller so that
µ(E) < δ implies also ´ E |f | dµ < 2ε . Likewise, we can take Eε larger if necessary (but still finite
measure) so that also X\Eε |f | dµ ⩽ 2ε .
Case 1: if µ(Eε ) = 0, then
ˆ ˆ ˆ ˆ
|fn − f | dµ = |fn − f | dµ ⩽ |fn | dµ + |f | dµ < ε
X X\Eε X\Eε X\Eε

Case 2: if µ(Eε ) > 0, then we use the fact that fn → f in measure. So

ε
µ({x ∈ X : |fn (x) − f (x)| > }) → 0
µ(Eε )
Let n0 such that n ⩾ n0 satisfies
ε
µ({x ∈ X : |fn (x) − f (x)| > }) < δ
µ(Eε )
Now take
ˆ ˆ ˆ
|fn − f | dµ = |fn − f | dµ + |fn − f | dµ
X X\Eε Eε
| {z }
⩽ε
ˆ ˆ
⩽ε+ |fn − f | dµ + |fn − f | dµ
ε ε
Eε ∩{|fn −f |⩽ µ(E )
} Eε ∩{|fn −f |> µ(E )
}
| {zε } ε

ε
⩽µ(Eε )· µ(E =ε
ˆ ε)

=ε+ε+ ε |fn − f | dµ
{|fn − f | > }
µ(Eε )
| {z }
⩽δ

⩽ 3ε
Proposition 4.3.4 (Exam 1 Question 4(i) :(). Consider {fn }, and suppose (ii) and (iii) hold above
from the Vitali Convergence Theorem. Then fn ∈ L1 (X) for all n.

Proof. Let Fm := {x ∈ X : |fn (x)| > m}. Then Fm ↘ ∅. Fix A ∈ M with µ(A) < ∞. Then also
A ∩ Fm ↘, and µ(A ∩ F1 ) ⩽ µ(A) < ∞, so µ(A ∩ Fm ) → 0.
´
Fix ε > 0, find δ > 0 such that µ(E) <´ δ implies supn E |fn | dµ < ε (by equi-integrability). Let
G ∈ M such that µ(G) < ∞ and supn X\G |fn | dµ < ε. Choose m >> 1, so µ(G ∩ Fm ) < δ.
Then
ˆ ˆ ˆ
|fn | dµ = |fn | dµ + |fn | dµ
X X\G G
ˆ ˆ
<ε+ |fn | dµ + |fn | dµ
G\Fm G∩Fm

⩽ ε + mµ(G \ Fm ) + ε
⩽ 2ε + mµ(G) < ∞

35
Theorem 4.3.5 (Dunford-Pettis Theorem).
´ (X, M, µ) meas space, F a family of integrable func-
tions f : X → R such that supf ∈F X |f | dµ < ∞
Then every sequence of functions in F admits a subsequence converging weakly in L1 (X) to
some function in L1 (X) if and only if F is equi-integrable and satisfies property (iii) of Vitali’s
Convergence Theorem.

Remark. We won’t prove the above theorem, but it is a fundamental property of equi-integrability.
Also note that by the proposition above, we don’t actually need to assume that the f ’s in F are
integrable in the backward direction! This is implied.

Theorem 4.3.6. (X, M, µ) meas space, F a family of integrable functions. Consider the following
conditions:
(i) F is equi-integrable.
´
(ii) limt→∞ supf ∈F {x∈X:|f |>t} |f | dµ = 0.

(iii) (De La Vallee Poussin Criterium) there exists an increasing function g : [0, ∞) → [0, ∞]
such that
g(t)
lim =∞
t→∞ t
´
and supf ∈F X g(|f |) dµ < ∞.
´
Then (ii) ⇐⇒ (iii) and both imply (i). Moreover, if supf ∈F X
|f | dµ < ∞, then all three
conditions are equivalent.

Proof. We prove a bunch of implications:

´
• (ii) =⇒ (i): Fix ε > 0. By (ii), choose tε >> 1 such that supf ∈F {x:|f |>tε }
|f | dµ < 2ε . Let
δ = 2tεε , and let G such µ(G) < δ. Then
ˆ ˆ ˆ
ε
|f | dµ = |f | dµ + |f | dµ ⩽ + tε µ(G) ⩽ ε
G G∩{|f |>tε } G∩{|f |⩽tε } 2

• (ii) =⇒ (iii): We construct an increasing sequence ki ∈ Z such that

ˆ
1
sup |f | dµ < i+1
f ∈F {|f |>ki } 2

For ℓ ∈ N, let bℓ be the number of nonnegative integers i such that ki < ℓ. Then bℓ ↗ ∞ as
ℓ → ∞. Define

g(t) := tbℓ

36
 g(t)
for t ∈ [ℓ, ℓ + 1). Observe that t
⩾ b⌊t⌋ → ∞. Now we check:
ˆ ∞ ˆ
X
g(|f |) dµ = g(|f |) dµ
X ℓ=0 {ℓ⩽|f |<ℓ}

X∞ ˆ
= bℓ |f | dµ
ℓ=0 {ℓ⩽|f |<ℓ+1}
∞ X ˆ
X
= |f | dµ
ℓ=0 i:ki <ℓ {ℓ⩽|f |<ℓ+1}
∞ Xˆ
X
= |f | dµ
i=0 ℓ>ki {ℓ⩽|f |<ℓ+1}
∞ ˆ
X
= |f | dµ
i=0 {|f |>ki }
∞
X 1
⩽ i+1
<∞
i=0
2
´
• (iii) =⇒ (ii) We have such a function g. Let M < ∞ with M := supf ∈F X
g(|f |) dµ. Fix
ε > 0, find tε >> 1 such that g(t) t
⩾ Mε . Let t ⩾ tε , then
ˆ ˆ ˆ
ε ε
|f | dµ ⩽ g(|f |) dµ ⩽ g(|f |) dµ ⩽ ε
{|f |>t} {|f |>t} M X M

g(|f |) M
where the first inequality follows because |f |
⩾ ε
(since |f | > t ⩾ tε ).
´
• (i) =⇒ (ii) with extra assumption: Now we assume that C := supf ∈F X
|f | dµ < ∞.
We also assume equi-integrability, which is (i).
C
Fix ε > 0, let tε := (δ from equi-integrability). Let t ⩾ tε . Then
δ
ˆ
1 C
µ({x ∈ X : |f | > t}) ⩽ |f | dµ ⩽ <δ
t X t
Now by equi-integrability, we have
ˆ
sup |f | dµ < ε
f ∈F {x:|f (x)|>t}

Product Spaces

5.1 Definitions
Definition 5.1.1 (Tensor Product of Measurable Spaces). (X, M), (Y, N ) measurable spaces.
Then M ⊗ N ⊆ P(X × Y ) is the smallest σ-algebra that contains all sets of the form E × F with
E ∈ M and F ∈ N .

37
Remark. If X, Y are topological spaces, then we can consider their product topology X × Y . Then
it happens that

β(X) ⊗ β(Y ) ⊆ β(X × Y )

If X, Y are separable metric spaces, then we have equality! β(X) ⊗ β(Y ) = β(X × Y ).

Note. (X, M, µ) and (Y, N , ν) meas spaces. We will define elementary sets in the product space:

E := {F × G : F ∈ M, G ∈ N }

and ρ : E → [0, ∞] maps F × G 7→ µ(F )ν(G). Then we define

(∞ ∞
)
X [
∗
(µ × ν) (E) := inf µ(Fn )ν(Gn ) : {Fn } ⊆ M, {Gn } ⊆ N , E ⊆ Fn × Gn
n=1 n=1

for any E ⊆ X × Y . Recall from the beginning of the class that we know (µ × ν)∗ is an outer
measure. Then Caratheodory’s Theorem gives that

M × N := {all (µ × ν)∗ -measurable sets}

is a σ-algebra, and (µ × ν)∗ |M×N is a measure (denoted (µ × ν))!

Theorem 5.1.1. (X, M, µ), (Y, N , ν) meas spaces. F ∈ M, G ∈ N . Then F × G is (µ × ν)∗ -

measurable, and (µ × ν)(F × G) = µ(F )µ(G). Therefore, M ⊗ N ⊆ M × N .

Proof. First observe: F1 , F2 ∈ M and G1 , G2 ∈ N , then

(F1 × G1 ) ∩ (F2 × G2 ) = (F1 ∩ F2 ) × (G1 ∩ G2 )

and



(F1 × G1 ) \ (F2 × G2 ) = (F1 \ F2 ) × G1 ∪ (F1 ∩ F2 ) × (G1 \ G2 )

Observe that ρ(F × G) = µ(F )ν(G). Therefore, after showing F × G is measurable, we must
simply show that ρ is subadditive.
Claim: F × G is (µ × ν)∗ -measurable. Fix E ⊆ X × Y , we want to show

(µ × ν)∗ (E) = (µ × ν)∗ (E ∩ (F × G)) + (µ × ν)∗ (E \ (F × G))

(we only need show ⩾). Assume that E ⊆ ∞

S
n=1 Fn × Gn for some {Fn } ⊆ M and {Gn } ⊆ N .
Then using the identities above,
∞
[
E \ (F × G) ⊆ (Fn × Gn ) \ (F × G)
n=1
[∞
= ((Fn \ F ) × Gn ) ∪ ((Fn ∩ F ) × (Gn \ G))
n=1

38
and
∞
[ ∞
[
E ∩ (F × G) ⊆ (Fn × Gn ) ∩ (F × G) = (Fn ∩ F ) × (Gn ∩ G).
n=1 n=1

So

(µ × ν)∗ (E \ (F × G)) + (µ × ν)∗ (E ∩ (F × G))

X∞ X∞ ∞
X
⩽ µ(Fn \ F )ν(Gn ) + µ(Fn ∩ F )ν(Gn \ G) + µ(Fn ∩ F )ν(Gn ∩ G)
n=1 n=1 n=1
∞
X ∞
X
= µ(Fn \ F )ν(Gn ) + µ(Fn ∩ F )ν(Gn )
n=1 n=1
X∞
= µ(Fn )ν(Gn )
n=1

Now taking the inf over all such coverings of E by elementary sets gives the desired result. This
shows F × G is measurable.

S∞Well consider F, Fn ∈ M and G, Gn ∈ N (so their products are all in

Claim: ρ is subadditive.
E ) such that F × G ⊆ n=1 Fn × Gn . Well we have
∞
[
F ×G= (Fn ∩ F ) × (Gn ∩ G)
n=1

Now consider

χF (x)χG (y) = χF ×G (x, y)

X∞
⩽ χ(Fn ∩F )×(Gn ∩G) (x, y)
n=1
∞
X
= χFn ∩F (x)χGn ∩G (y)
n=1

Now fix ŷ ∈ Y . Then the function mapping X → [0, ∞] defined by x 7→ χFn ∩F (x)χGn ∩G (ŷ) is
measurable. Hence, since countable sums of measurable functions are measurable, we then have
that
∞
X
x 7→ χFn ∩F (x)χGn ∩G (ŷ)
n=1

39
is also measurable. Hence
ˆ
µ(F )χG (ŷ) = χG (ŷ)χF (x) dµ(x)
X
ˆ X∞
⩽ ( χFn ∩F (x)χGn ∩G (ŷ)) dµ(x)
X n=1

Xˆ
∞
= χFn ∩F (x)χGn ∩G (ŷ) dµ(x)
n=1 X
∞
X
= µ(Fn ∩ F )χGn ∩G (ŷ)
n=1

(we exchange sum andPintegral by monotone convergence theorem!) Now we repeat the argument
∞
on the function y 7→ n=1 µ(Fn ∩ F )χGn ∩G (y) (which is measurable). We have now that
ˆ
µ(F )ν(G) = µ(F ) χG (y) dν(y)
Y
ˆ X
∞
⩽ µ(Fn ∩ F )χGn ∩G y dν(y)
Y n=1
∞
X ˆ
= µ(Fn ∩ F ) χGn ∩G (y) dν(y)
n=1 Y
∞
X
⩽ µ(Fn )ν(Gn )
n=1

Corollary 5.1.2 (Regularity). (X, M, µ), (Y, N , ν) meas spaces, E ⊆ X × Y . Then there exists
C ∈ M ⊗ N such that E ⊆ C and
(µ × ν)∗ (E) = (µ × ν)(C)

Proof. ⩽ is easy.
S∞ (k) (k)
And, for any k ∈ N, let E ⊆ n=1 Fn × Gn such that
∞
X
∗ 1
µ(Fn(k) )ν(G(k)
n ) ⩽ (µ × ν) (E) +
n=1
k
T∞ S∞ (k) (k)
Then let C := k=1 n=1 Fn × Gn ⊇ E. Then for all k, we have
∞
X
∗ 1
(µ × ν)(C) ⩽ µ(Fn(k) )ν(G(k)
n ) ⩽ (µ × ν) (E) +
n=1
k

5.2 Integrals on Product Spaces

Definition
S∞ 5.2.1 (σ-finite). (X, M, µ) meas space, µ is finite if µ(X) < ∞. µ is σ-finite if
X = n=1 Xn for Xn ∈ M with µ(Xn ) < ∞ for all n.

40
Definition 5.2.2 (Cross-section). For a set E ⊆ X × Y , for x ∈ X and y ∈ Y the cross-sections
are defined as
Ex := {y ∈ Y : (x, y) ∈ E} ⊆ Y and Ey := {x ∈ X : (x, y) ∈ E} ⊆ X
Proposition 5.2.1. If f : X × Y → R which is M ⊗ N measurable, we can consider for x ∈ X,
the function [f ]x : Y → R mapping y 7→ f (x, y) and for y ∈ Y , the function [f ]y : X → R mapping
x 7→ f (x, y). Then [f ]x , [f ]y are measurable.

Proof. B ⊆ R Borel. Then

([f ]x )−1 (B) = {y ∈ Y : [f ]x (y) ∈ B} = {y ∈ Y : f (x, y) ∈ B} = [f −1 (B)]x .
We know that f −1 (B) is measurable. Hence, to show that [f ]x is measurable, it suffices to show
that if E ∈ M ⊗ N , then Ex ∈ N (and likewise that Ey ∈ M to show that [f ]y is measurable).
Let F := {E ∈ M ⊗ N : ∀x ∈ X, Ex ∈ M and ∀y ∈ Y, Ey ∈ N }. We just need show
1. A ∈ M, B ∈ N implies A × B ∈ F (easy).
2. F is a σ-algebra, hence F = M ⊗ N .
These are left as an exercise.

Definition 5.2.3 (Dynkin Class). X ̸= ∅, then the collection D ⊆ P(X) is called Dynkin class
on X if
(i) X ∈ D
(ii) E, F ∈ D with E ⊆ F , then F \ E ∈ D
(iii) En ∈ D with En ↗. Then ∞
S
n=1 En ∈ D.

Theorem 5.2.2. Let F ⊆ P(X) be a family closed under finite intersections. Then the σ-algebra
generated by F coincides with the Dynkin class generated by F (i.e. the smallest Dynkin class
that contains F).

Remark. We do not prove this.

Theorem 5.2.3. (X, M, µ), (Y, N , ν) meas spaces with µ and ν complete. Let E ∈ M × N have
σ-finite µ × ν measure. Then
(1) For µ-a.e. x ∈ X, the cross sections Ex ∈ N are measurable, and for ν-a.e. y ∈ Y the cross
sections Ey ∈ M are measurable.
(2) Furthermore, the functions x 7→ ν(Ex ) and y 7→ µ(Ey ) are measurable.
(3) And
ˆ ˆ
(µ × ν)(E) = µ(Ey ) dν(y) = ν(Ex ) dµ(x)
ˆY ˆ ˆX ˆ
= χE (x, y) dµ(x) dν(y) = χE (x, y) dν(y) dµ(x)
Y X X Y

41
Proof. We will rewrite the three parts of the theorem as:
(i) ∀y ∈ Y , x 7→ χE (x, y) is measurable (i.e., χEy is measurable for all y).
´ ´
(ii) y 7→ X χE (x, y) dµ(x) = X χEy (x) dµ(x) = µ(Ey ) is measurable.
´
(iii) (µ × ν)(E) = Y µ(Ey ) dν(y)
(observe that (i) implies (1) by taking (χEy )−1 ({1}) = Ey . So if χEy is measurable, then Ey ∈ M).
Everything is symmetric in x and y, so we may repeat the proof with x and y flipped to prove the
entire theorem. We proceed in a sequence of steps:
Step 1: Assume that µ × ν is finite (in this case, no need for completeness). Let D be the family
of all sets D ∈ M × N satisfying (i), (ii), and (iii). We claim that D is a Dynkin class. First,
observe F ∈ M, G ∈ N , then F × G ∈ D. In particular, if we can show that D is a Dynkin class,
then by the previous theorem, D ⊇ M ⊗ N .
To show that D is a Dynkin class, we have that X × Y ∈ D since X ∈ M and Y ∈ N . Now, let
E1 , E2 ∈ D with E1 ⊆ E2 . We want to show E2 \ E1 ∈ D. First note that
ˆ
(µ × ν)(Ei ) = µ((Ei )y ) dν(y) < ∞
Y

(by finiteness). Then

(µ × ν)(E2 \ E1 ) = (µ × ν)(E2 ) − (µ × ν)(E1 )

ˆ ˆ
= µ((E2 )y )dν(y) − µ((E1 )y )dν(y)
ˆY Y

= [µ((E2 )y ) − µ((E1 )y )] dν(y)

ˆy
= µ((E1 \ E1 )y )dν(y)
Y

This proves property (iii). (i) and (ii) are easy. This shows that D satisfies the second property of
Dynkin class.
Now we show that third property of Dynkin class. If En ↗ with En ∈ D, we want to show
S ∞
n=1 En ∈ D. For (i), fix y ∈ Y , and observe

x 7→ χEn (x, y)

is measurable. Therefore x 7→ χS En (x, y) is measurable. Property (ii) is similar. For property

42
(iii), we have
∞
[
(µ × ν)( En ) = lim(µ × ν)(En )
n
n=1
ˆ
= lim µ((En )y )dν(y)
n
ˆ Y

= lim µ((En )y )dν(y)

Y n
ˆ ∞
[
= µ( (En )y )dν(y)
Y n=1
ˆ ∞
[
= µ(( En )y )dν(y)
Y n=1

(used the LMCT twice!) as desired. So D is Dynkin class. This concludes step 1!
Step 2: Assume that µ and ν are complete, and µ × ν finite. We want to show that if E ∈
M × N , then E satisfies properties (i)-(iii). There exists some R ∈ M ⊗ N such that E ⊆ R and
(µ × ν)(E) = (µ × ν)(R) (by regularity). Since (µ × ν)(E) < ∞ (why? oh, finiteness!), then also
(µ × ν)(R \ E) = 0.

´ by regularity, find some S ∈ M ⊗ N with R \ E ⊆ S with 0 = (µ × ν)(R \ E) = (µ × ν)(S) =

´Again
Y X S
χ (x, y)dµ(x)dν(y) ´(by Step 1!). Since this is an integral which is non-negative and equal to
0, we must have µ(Sy ) = X χS (x, y)dµ(x) = 0 for ν-a.e. y ∈ Y .
Now, R \ E ⊆ S implies (R \ E)y = Ry ´\ Ey ⊆ Sy . Therefore,
´ by completeness of µ, Ry \ Ey is
measurable with
´ measure 0 (ν-a.e.)! So χ
X R
(x, y)dµ(x) = χ (x, y)dµ(x) (for ν-a.e. y ∈ Y ).
X E
Therefore y 7→ X χE (x, y)dµ(x) is measurable (i)-(ii).
Finally, for (iii) we see

(µ × ν)(E) = (µ × ν)(R)
ˆ ˆ
= χR (x, y)dµ(x)dν(y)
ˆY ˆX
= χE (x, y)dµ(x)dν(y)
Y X

This concludes Step 2.

Step 3: Now we assume only that µ and ν areScomplete, and E ∈ M×N is σ-finite. Can we show
E satisfies (i)-(iii)? Well we can write E = ∞
n=1 En with En ∈ M × N and (µ × ν)(En ) < ∞.
WLOG we can assume the En are mutually disjoint.
We will prove (i)-(iii) for each E ∩ En . Then (i)-(iii) hold for E. This is easy to prove for (i) and

43
(ii). For (iii), note that
∞
X
(µ × ν)(E) = (µ × ν)(E ∩ En )
n=1
∞ ˆ ˆ
X
= χE∩En (x, y)dµ(x)dν(y)
n=1 X Y
ˆ ˆ X
∞
= χE∩En (x, y)dµ(x)dν(x)
Y X n=1
ˆ ˆ
= χE (x, y)dµ(x)dν(y)
Y X

(using LMCT) as desired. Therefore, we may WLOG assume that (µ × ν)(E) < ∞. Find
Fn ∈ M, Gn ∈ N such that E ⊆ ∞
S
n=1 (F n × Gn ) (WLOG assume pairwise disjoint) such that
X
µ(Fn )ν(Gn ) ⩽ (µ × ν)(E) + 1

So for all n, µ(Fn )ν(Gn ) = (µ × ν)(Fn × Gn ) < ∞. We may consider

(M × N )n := M × N |Fn ×Gn := {D ∩ (Fn × Gn ) : D ∈ M × N }

Then (µ × ν)|(M×N )n is finite! So by Step 2, we have (for example for (iii)) that
ˆ ˆ
(µ × ν)(E ∩ (Fn × Gn )) = χE∩(Fn ×Gn ) (x, y)dµ(x)dν(y)
Y X

and
X
(µ × ν)(E) = (µ × ν)(E ∩ (Fn × Gn ))
Xˆ ˆ
n

= χE∩(Fn ×Gn ) (x, y)dµ(x)dν(y)

Y X
ˆn ˆ X
= χE∩(Fn ×Gn ) (x, y)dµ(x)dν(y)
Y X
ˆ ˆ n

= χE (x, y)dµ(x)dν(y)
Y X

Makes perfect sense, right? This class is psychotic. This shit is so mind-numbingly boring.

5.3 Exchanging Iterated Integrals

Remark. Is it the case that for σ-finite µ, ν,

µ, ν complete ⇐⇒ µ × ν complete?

44
Not exactly! It depends on which measurable space we are working over. Recall that Caratheodory’s
theorem implies automatically that µ × ν is complete over the space of (µ × ν)∗ -measurable sets,
which is how we defined M × N .
On the other hand, even if µ and ν are complete, µ × ν may not be complete over M ⊗ N . In
particular, if µ = ν = L 1 , and M = N are both the set of Lebesgue-measurable sets, then let V
be the (non-measurable) Vitali set. We have that

(µ × ν)∗ ({0} × V ) = 0

But on the other hand, {0} × V ∈ ̸ M ⊗ N , since Proposition 5.2.1 would then imply that all
cross-sections are measurable. In particular it would imply that ({0} × V )0 = V ∈ N . Clearly
this is a contradiction.

Remark. Conversely, if µ, ν are σ-finite, then µ × ν being complete over M ⊗ N does imply that
µ and ν are complete.
To see this, assume first that µ and ν are finite (the entire spaces X and Y are finite). Take
F ∈ M with µ(F ) = 0, and consider any F ′ ⊆ F . Then

(µ × ν)∗ (F ′ × Y ) ⩽ (µ × ν)∗ (F × Y ) = µ(F )ν(Y ) = 0

since F ∈ M and Y ∈ N . Hence F ′ × Y ∈ M ⊗ N by the assumed completeness. Now again by

Proposition 5.2.1, (F ′ × Y )y = F ′ ∈ M, so µ is complete. We can do the same to show that ν is
also complete.
To extend this to σ-finite µ and ν, just write X = ∞
S S∞
n=1 Xn and Y = n=1 Yn for finite Xn , Yn ,
and take intersections.

Theorem 5.3.1 (Tonelli’s Theorem). (X, M, µ), (Y, N , ν) meas spaces, µ, ν complete and E ⊆
X × Y is σ-finite in µ × ν-measure. If f : X × Y → [0, ∞] (µ × ν)-measurable with support E,
then
(1) For µ-a.e. x ∈ X, the function [f ]x : Y → [0, ∞] mapping y 7→ f (x, y) is measurable.
´
(2) g : X → [0, ∞] mapping x 7→ Y f (x, y) dν(y) is measurable.
(3) Same for Y .
Moreover,
ˆ ˆ ˆ  ˆ ˆ 
f (x, y) d(µ × ν)(x, y) = f (x, y) dν(y) dµ(x) = f (x, y) dµ(x) dν(y)
X×Y X Y Y X

Proof. Step 1: If f is simple: f = m

P
n=1 cn χEn for cn ⩾ 0, then Theorem 5.2.3 implies all of the
parts of this theorem (by linearity of integrals).
Step 2: So let f be an arbitrary measurable function. Then there exists sn : X ×Y → [0, ∞] simple
functions sn ↗ f pointwise. For (1)-(3), we just note that [sn ]x ↗ [f´]x pointwise, and they are
measurable by Step 1, so [f ]x is measurable. Likewise, let gn map x 7→ Y sn (x, y) dν(y). These are

45
 ´ ´
measurable by Step 1, and for any x, we have lim gn (x) = lim Y
sn (x, y) dν(y) = Y
f (x, y) dν(y) =
g(x) by LMCT. So, g = lim gn is also measurable.
Finally by LMCT (thrice):
ˆ ˆ
f (x, y) d(µ × ν)(x, y) = lim sn (x, y) d(µ × ν)(x, y)
X×Y
ˆ ˆ
X×Y

= lim sn (x, y) dν(y) dµ(x)

ˆ ˆX Y
= lim sn (x, y) dν(y) dµ(x)
ˆ ˆ
X Y

= f (x, y) dν(y) dµ(x)

X Y

Remark. In fact, Tonelli’s theorem doesn’t necessarily need that µ and ν are σ-finite; it only needed
that the support of f has σ-finite µ × ν measure.

Theorem 5.3.2 (Fubini’s Theorem). (X, M, µ), (Y, N , ν) meas spaces, µ, ν complete. f : X ×
Y → [−∞, ∞] is (µ × ν)-integrable. Then
(1) For µ-a.e. x ∈ X, [f ]x (y) := f (x, y) is ν-integrable.
(2) For ν-a.e. y ∈ Y , [f ]y (x) := f (x, y) is µ-integrable.
´
(3) F (x) := Y f (x, y) dν(y) is integrable.
(4) Same for Y .
Moreover, again,
ˆ ˆ ˆ  ˆ ˆ 
f (x, y) d(µ × ν)(x, y) = f (x, y) dν(y) dµ(x) = f (x, y) dµ(x) dν(y) (*)
X×Y X Y Y X

´
Proof. Write f = f + − f − , and suppose X×Y |f (x, y)| d(µ × ν) < ∞. S Observe that this implies
that E := {x : |f (x, y)| > 0} is σ-finite wrt µ × ν. This is because E = ∞
n=1 En where
 
1
En = < |f (x, y)| < n
n
all have finite measure by Chebyshev’s inequality.
Now apply Tonelli’s Theorem to f + and f − (which have σ-finite support E) and add them to get
(*).

Remark. Tonelli’s (and thus also Fubini’s) doesn’t need that µ and ν are complete if f is M ⊗ N
measurable (Fubini’s still needs integrable).

Note. Here are some sample applications of Tonelli’s and Fubini’s theorems:

Theorem
´ ´ ∞M,p−1
5.3.3. (X,
p
ν) meas space, µ complete. 1 < p < ∞, and f : X → R measurable.
Then X |f | dµ = p 0 s µ({x ∈ X : |f (x)| > s}) ds.

46
Proof. If there is some s0 > 0 such that ´µ({x ∈ X : |f (x)| > s0 }) = ∞, then it is true for all
s < s0 , so the RHS = ∞, and the LHS ⩾ {|f |>s0 } |f |p dµ = ∞.
Thus, we may assume µ({x ∈ X : |f (x)| > s}) < ∞ for all s ∈ R⩾0 . Therefore, X0 := {|f | > 0} is
σ-finite, so restrict f to X0 . We now apply Tonelli’s theorem with the spaces (X0 , M|X0 , µ|M|X0 )
and ((0, ∞), Lebesgue σ-algebra, L 1 ):
ˆ ∞ ˆ ∞ ˆ 
p−1 p−1
p s µ({x ∈ x0 : |f (x)| > s}) ds = p s χ{|f |>s} (x) dµ(x) ds
0 0 X0
ˆ ˆ ∞ 
p−1
=p s χ{|f |>s} (x) ds dµ
X0 0
ˆ ˆ |f (x)|
!
= psp−1 ds dµ
X0 0
ˆ
= |f (x)|p dµ
ˆX0
= |f (x)|p dµ
X
´∞
Definition 5.3.1. Γ-function is defined as Γ(t) := 0
xt−1 e−x dx for t > 0.
π N/2
Theorem 5.3.4. Let N ⩾ 1, then L N (B(0, R)) = αN RN where αN = Γ(1+N/2)
.

π N/2
Proof. Suffices to prove L N (B(0, 1)) =: αN = Γ(1+N/2)
. Write RN +1 = {(x, y) : x ∈ RN , y ∈ R},
and set D := {(x, y) ∈ RN +1 : ∥x∥2 < y}. Consider
ˆ ˆ ∞ˆ ˆ ∞ ˆ ∞
√ N N
−y
e dxdy = √
−y
e dxdy = e L (B(0, y))dy = αN
−y N
e−y y 2 dy = αN Γ(1+ )
D 0 B(0, y) 0 0 2

On the other hand, by Fubini’s

ˆ ˆ ˆ ∞
−y
e dxdy = e−y dydx
N 2
D
ˆR ∥x∥
= (−e−y |∞ ∥x∥2
)dx
ˆRN

2
= e−∥x∥ dx
ˆR
N

2 2 2
= e−x1 −x2 −···−xN dx
RN
ˆ N
−x2
= e dx =: I N
R

I N
These two calculations show that αN = Γ(1+N/2) . To calculate I, we know that for N = 2, α2 = π,
2
√
while Γ(1 + 2/2) = 1. Hence π = I , so I = π.

47
Lp spaces

6.1 Definitions
Definition 6.1.1 (Normed Space). (X, ∥·∥) is a normed space if X is a vector space, and the
norm ∥·∥ maps X → [0, ∞] such that
(i) ∥x∥ = 0 iff x = 0
(ii) ∥tx∥ = |t| · ∥x∥ for all t ∈ R, x ∈ X.
(iii) ∥x + y∥ ⩽ ∥x∥ + ∥y∥, for all x, y ∈ X.

Remark. Normed spaces induce metric spaces with distance defined by d(x, y) = ∥x − y∥.

Definition 6.1.2 (Banach Space). A normed space is Banach if Cauchy sequences converge.

Definition 6.1.3 (M p ). (X, M, µ) meas space, and p ⩾ 1. Then for E ∈ M,

 ˆ 
p p
M (E) := f : E → (−∞, ∞) f meas and |f | dµ < ∞
E
´
1/p
We write ∥f ∥M p (E) := E
|f |p dµ .
For p = ∞,
 
∞
M (E) := f : E → (−∞, ∞) f meas and bounded

and ∥f ∥M ∞ (E) := supx∈E |f (x)|.

Remark. We often write ∥f ∥p to mean ∥f ∥M p .

Note. We will show that ∥f ∥M p is (almost) a norm. Property (ii) is easy. To see property (iii), we
will prove Holder’s inequality, and then the Minkowski inequality.

Definition 6.1.4 (Holder Conjugate Exponent). For 1 ⩽ p ⩽ ∞, the conjugate exponent is:
 p
 p−1
 1<p<∞
′
p := ∞ p=1

1 p=∞


1 1
So that p
+ p′
= 1.

Theorem 6.1.1 (Holder’s Inequality). Let f, g : X → (−∞, ∞) be measurable. Then

(1) For 1 < p < ∞:
ˆ ˆ 1/p ˆ 1/p′
p p′
|f g| dµ ⩽ |f | dµ |g| dµ = ∥f ∥p ∥g∥p′
X X X

48
 (2) For p = 1:
ˆ ˆ  
|f g| dµ ⩽ |f | dµ sup|g(x)|
X X x∈X

(3) For p = ∞:
ˆ   ˆ 
|f g| dµ ⩽ sup|f (x)| |g| dµ
X x∈X X

Proof. We prove (1). If ∥f ∥p = 0 or ∥g∥p′ = 0, then RHS = 0. But in this case, f ≡ 0 a.e. (or g),
´
and so f g ≡ 0 a.e. Therefore X |f g| dµ = 0, and Holder’s inequality holds.
1
So, WLOG, ∥f ∥p , ∥g∥p′ > 0, and are finite. Now, using the fact that log is concave and p
+ p1′ = 1,
we have for any a, b ⩾ 0
 
1 p 1 p′ 1 1
log a + ′b ⩾ · p log(a) + ′ · p′ log(b) = log(ab)
p p p p
And since log is increasing, this also implies that
1 p 1 p′
a + ′ b ⩾ ab
p p
This is called Young’s inequality. In particular, let a = |f (x)| and b = |g(x)|. Then integrate
to get
ˆ ˆ ˆ
1 p 1 ′
|f g| dµ ⩽ |f | dµ + ′ |g|p dµ < ∞
X p X p X
So, in particular, f g ∈ M 1 (X). Now take the map f 7→ tf for t > 0. We have
ˆ ˆ
1 1 ′
t · |f g| dµ = |tf g| dµ ⩽ ∥tf ∥pp + ′ ∥g∥pp′
X X p p
so dividing by t, we get
ˆ
tp−1 1 ′
|f g| dµ ⩽ ∥f ∥pp + ′ ∥g∥pp′ =: h(t)
X p tp
p′ /p
∥g∥p′
Taking a derivative in t to see that h(t) is minimized at t = ∥f ∥p
. In particular, plugging in,

ˆ (p−1)·p′ /p
1 ∥g∥p′ 1 ∥f ∥p p′
|f g| dµ ⩽ · ∥f ∥pp + ′ ∥g∥ p′
X p ∥f ∥p−1
p
p′ ∥g∥pp′ /p
1 ∥g∥p′ p 1
= · p−1 ∥f ∥p + ′ ∥f ∥p ∥g∥p′
p ∥f ∥p p
= ∥f ∥p ∥g∥p′

49
 ′
(since p′ − pp = 1). Part (2) (and (3)) are easily proved by observing that |f g| ⩽ |f | · ∥g∥∞ .
Therefore by monotonicity of integrals
ˆ ˆ ˆ ˆ  
|f g| dµ ⩽ |f | · ∥g∥∞ dµ = |f | dµ · ∥g∥∞ = |f | dµ sup|g(x)|
X X X X x∈X

Theorem 6.1.2 (Minkowski Inequality (Triangle inequality)). ∥f + g∥p ⩽ ∥f ∥p + ∥g∥p

Proof. For p = 1 and p = ∞, the inequality is obvious. Let p ∈ (1, ∞), and assume that
∥f ∥p , ∥g∥p < ∞ (else the inequality is trivial).
Step 1: Since t 7→ tp is convex, then for a, b > 0 we have
 p
p p 1 1
(a + b) = 2 a + b ⩽ 2p−1 ap + 2p−1 bp
2 2

Step 2: Hence, we get

ˆ ˆ ˆ ˆ 
p p p−1 p p
|f + g| dµ ⩽ (|f | + |g|) dµ ⩽ 2 |f | dµ + |g| dµ < ∞
X X X X
p
So, in particular, f + g ∈ M (X).
Step 3:
ˆ
∥f + g∥pp = |f + g|p dµ
ˆX
= |f + g| · |f + g|p−1 dµ
ˆX ˆ
⩽ |f | · |f + g| dµ + |g| · |f + g|p−1 dµ
p−1

X X
ˆ 1/p ˆ 1/p′ ˆ 1/p ˆ 1/p′
p (p−1)p′ p (p−1)p′
⩽ |f | |f + g| + |g| |f + g|
X X X X
  ˆ 1/p′
p
= ∥f ∥p + ∥g∥p · |f + g|
X
  ′
= ∥f ∥p + ∥g∥p · ∥f + g∥p/p
p

where the second inequality is Holder’s, and the equality after that is because (p − 1)p′ = p.
′
Dividing by ∥f + g∥pp/p , we get
p− pp′
∥f + g∥p = ∥f + g∥p ⩽ ∥f ∥p + ∥g∥p
Note. Finally, we must check that property (i) of normed spaces holds. But it doesn’t!
ˆ
|f |p dµ = 0 ̸ =⇒ f ≡ 0
X

It only implies that f = 0 a.e. Thus, in M p (X), we will introduce an equivalence relation: f ∼ g
iff f = g a.e. Therefore f = 0 a.e. means that f ∼ 0.

50
Definition 6.1.5 (Lp space). Let 1 ⩽ p < ∞. We may now define Lp (X) := M p (X)/ ∼. This is
´
1/p
a normed space with ∥[f ]∥Lp := ∥f ∥M p = ∥f ∥p = X |f |p dµ .

Remark. To define L∞ , we must be slightly more careful.

Definition 6.1.6 (Essential Supremum). (X, M, µ) meas space. For measurable f ,

ess sup f := inf{t ∈ R : f (x) ⩽ t µ-a.e.}

Remark. Hence, if ess sup f = M < ∞, then f ⩽ M a.e. (Exercise to prove this).

Definition 6.1.7 (L∞ space). L∞ (X) := {[f ] | f : X → [−∞, ∞], f measurable, ess sup|f | < ∞}.
This defines a normed space with norm ∥[f ]∥L∞ := ∥[f ]∥∞ = ess sup f .

Remark. Holder’s inequality still holds for p = 1 and p = ∞ replacing sup with ess sup.

Theorem 6.1.3. If µ(X) < ∞, and 1 ⩽ p < q ⩽ ∞, then Lq (X) ⊆ Lp (X).

Proof. Let f ∈ Lq (X). Then

ˆ ˆ
p
|f | dµ = 1 · |f |p dµ
X X
⩽ ∥|f |p ∥q/p · ∥1∥(q/p)′
ˆ p/q ˆ 1/(q/p)′
q
= |f | dµ · 1 dµ
X X
1/(q/p)′
= ∥f ∥pq · µ(X) <∞

by Holder’s inequality.

6.2 Properties of Lp spaces

Note. Last time we proved that if µ(X) < ∞, then for 1 ⩽ p < q ⩽ ∞, we have Lq (X) ⊆ Lp (X).

Lemma 6.2.1. (X, M, µ) meas space, and f ∈ L∞ (X)∩Lp (X) for all (large) p, then ∥f ∥p → ∥f ∥∞
a.e.

Proof. First assume that µ(X) < ∞. Then |f | ⩽ ∥f ∥∞ . Hence

ˆ 1/p
p
∥f ∥p = |f | dµ ⩽ ∥f ∥∞ µ(X)1/p → ∥f ∥∞
X

So lim supp ∥f ∥p ⩽ ∥f ∥∞ . Now let r < ∥f ∥∞ , and E := {x ∈ X : |f (x)| > r}. Then 0 < µ(E) < ∞,
´ ´
and X |f |p dµ ⩾ E |f |p dµ ⩾ rp µ(E), so ∥f ∥p ⩾ rµ(E)1/p . Hence

lim inf∥f ∥p ⩾ r

for all such r. Now take r → ∥f ∥∞ , and we get lim inf∥f ∥p ⩾ ∥f ∥∞ .

51
Now what if µ(X) = ∞? Then if ∥f ∥∞ = 0 (f = 0 a.e.), we are good. Otherwise, 0 < ∥f ∥∞ < ∞.
For 0 < ε < ∥f ∥∞ , define

D := {x ∈ X : |f (x)| ⩾ ∥f ∥∞ − ε}

Then µ(D) > 0, so (by Chebyshev):

ˆ ˆ
∞> |f | dµ ⩾ |f |p dµ ⩾ (∥f ∥∞ − ε)p µ(D)
p

X D

So µ(D) < ∞ and ∥f ∥p ⩾ (∥f ∥∞ − ε)µ(D)1/p , hence we get lim inf∥f ∥p ⩾ ∥f ∥∞ .

∥f ∥∞
Now fix ε > 0, and define f˜(x) := f (x)
∥f ∥∞ +ε
. Then 0 ⩽ f˜ ⩽ ∥f ∥∞ +ε
. We have
ˆ ˆ p
p 1/p
( |f | dµ) = (∥f ∥∞ + ε)( f˜ dµ)1/p
X X

´ p
We claim there exists some p̄ such that ∀p ⩾ p̄, ( X
f˜ dµ)1/p ⩽ 1. If so, then lim sup∥f ∥p ⩽
∥f ∥∞ + ε (so let ε → 0).
To prove the claim, we have
ˆ ˆ ˆ
˜
p0 +p
˜
p0
˜
p p0 ∥f ∥∞ p
f dµ = f · f dµ ⩽ f˜ dµ · ( )
X X ∥f ∥∞ + ε
∥f ∥∞ p 1
If p is sufficiently large, then ( ∥f ∥ ) ⩽ p , hence the RHS is ⩽ 1.
∞ +ε ∥f˜∥p00

Theorem 6.2.2. (X, M, µ) meas space, with 1 ⩽ p < q < ∞. Then

(i) Lp (X) ̸⊆ Lq (X) iff X contains measurable sets of arbitrarily small (positive) measure.
(ii) Lq (X) ̸⊆ Lp (X) iff X contains measurable sets of arbitrarily large (finite) measure.

Proof. Fix such p < q.

(i) (⇒) Assume that Lp (X) ̸⊆ Lq (X). There exists f ∈ Lp (X) with f ̸∈ Lq (X). For every n,
define

En := {x : |f (x)| ⩾ n}

(so En ↘). Then

ˆ ˆ
p
∞> |f | dµ ⩾ |f |p dµ ⩾ np µ(En )
X En

by Chebyshev’s. So divide by np , and conclude µ(En ) → 0. So they are arbitrarily small.

We must show they have nonzero measure.

52
 Suppose that µ(En ) = 0. Then
ˆ ˆ ˆ ˆ
q q p
∞= |f | dµ = |f | dµ ⩽ nq−p
|f | dµ ⩽ |f |p dµ < ∞
X X\En X\En X

since f ∈ Lp (X) with f ̸∈ Lq (X), a contradiction.

(⇐) Conversely, assume that we have arbitrarily small sets. WePwant to construct some
(pairwise P np ) ↘ 0. Then we can define f :=
P q disjoint) En with µ(E cn χEn with cn ↗ ∞ so
that cn µ(En ) = ∞ while cn µ(En ) < ∞.
Let F0 ∈ M with 0 < µ(F0 ) ⩽ 1. Then find Fn satisfying 0 < µ(Fn ) < 31 µ(Fn−1 ). Define
[
En := Fn \ Fk
k⩾n+1

1
Then the En are pairwise disjoint, and µ(En ) ⩽ µ(Fn ) ⩽ 3n
µ(F0 ). Likewise, µ(En ) > 0, or
else
[ X 1
µ(Fn ) ⩽ µ( Fk ) ⩽ µ(Fk ) ⩽ µ(Fn )
k⩾n+1 k⩾n+1
2

1
a contradiction. Now define cqn := cqn µ(En ) =
P
nµ(En )
. This immediately gives us that
P1
n
= ∞ as desired. Also,
X X 1 X 1 p
cpn µ(En ) = p/q p/q
µ(En ) = p/q
(µ(En ))1− q
n µ(En ) n
p
We can observe that µ(En ) ∼ 1
3n
, so for large enough n, we certainly have 1
np/q
µ(En )1− q ⩽ 1
n2
,
so the sum converges!
q p q p 1 1
S L (X) ̸⊆ L (X). Let f ∈ L and f ̸∈ L . Fn := {x : n+1 ⩽ |f (x)| ⩽ n }, and
(ii) (⇒) Assume
let F∞ = Fn . If µ(F∞ ) < ∞, then
ˆ ˆ ˆ ˆ
p p p
|f | dµ = |f | dµ + |f | ⩽ µ(F∞ ) + |f |q dµ < ∞
X {|f |⩽1} {|f |>1} {|f |>1}

since f ∈ Lq . But f ̸∈ Lp , so this is a contradiction! So µ(F∞ ) = ∞. Now apply Chebyshev’s

to get
ˆ ˆ
q 1
∞> |f | dµ ⩾ |f |q dµ ⩾ µ(Fn )
X Fn (n + 1)q

Thus µ(Fn ) < ∞. So we just define Gn := nk=1 Fk . These all have finite measure, and
S
µ(Gn ) → µ(F∞ ) = ∞.
(⇐) Assume there are sets with arbitrarily large measure. We wantPto find some mutu-
ally disjoint sets En with µ(En ) → P
∞, and some cn ↘ 0 such that cpn µ(En ) = ∞ and
q
P
cn µ(En ) < ∞. Then just let f = cn χEn .
Exercise to finish this one.

53
Theorem 6.2.3. Lp (X) is a Banach space (a normed space in which Cauchy sequences converge).

Lemma 6.2.4. Let V be a normed space. V is complete iff for any sequence {vn } ⊆ V such that
P Pℓ
∥vn ∥ < ∞, there exists v ∈ V such that limℓ→∞ n=1 vn − v = 0.

Pℓ
Proof. (⇒) Let zℓ := n=1 vn . We claim that {zn } is a Cauchy sequence. Indeed
ℓ+p ℓ+p
X X X
∥zℓ+p − zℓ ∥ = vn ⩽ ∥vn ∥ ⩽ ∥vn ∥ → 0
n=ℓ+1 n=ℓ+1 n⩾ℓ+1

Hence, it converges, and so the partial sums of vn converge as desired.

(⇐) Let {vn } be a Cauchy sequence. ∀ε > 0 there exists n0 such that for all n ⩾ n0 and all p, we
have ∥vn+p − vn ∥ < ε.
1
We claim that {vn } converges. If ε = 2k+1
, then there exists nk+1 > nk so that ∀n, m ⩾ nk+1 , we
have
1
∥vn − vm ∥ <
2k+1
(base case is ε = 21 ). So then we have vnk+1 − vnk < 1
2k
, so we can write
∞
X X 1
vnk+1 − vnk ⩽ <∞
k=1
2k

Therefore, by our assumption, there exists some v ∈ V such that

m
X
(vnk+1 − vnk ) − v → 0
k=1

This telescopes: −vn1 + vnm+1 − v → 0. That is, vnm+1 → v + vn1 . But {vn } is Cauchy, so a
subsequence converging implies that {vn } converges (exercise)!. Hence, vn → v + vn1 as well.

Proof of Theorem 6.2.3. Let 1 ⩽ p < ∞. We already

Pknow that Lp (X) is a normed space, so
p ∞
we will use the lemma. Let fn ∈ L (X) such that n=1 ∥fn ∥p < ∞. We want to show that
Pℓ
n=1 fn − f → 0 as ℓ → ∞.
p

Let g(x) := ( ∞ p
P
n=1 |fn (x)|) . Then
"ˆ ℓ
#1/p ℓ ℓ
X X X
( |fn |)p dµ = |fn | ⩽ ∥fn ∥p
X n=1 n=1 n=1
p

by triangle inequality (Minkowski). Now the LMCT tells us (since ( ℓn=1 |fn |)p ↗ g)
P

ˆ ˆ ℓ
X ℓ
X ∞
X
g dµ = lim ( |fn |)p dµ ⩽ lim( ∥fn ∥p )p = ( ∥fn ∥p )p < ∞
X n→∞ X n=1 n=1 n=1

54
So g(x) < ∞ a.e. on X. So ∞
P
n=1 fn (x) is absolutely convergent a.e. Define
(P
∞
n=1 fn (x) g(x) < ∞
f (x) :=
0 otherwise

f is measurable, and |f (x)|p ⩽ g(x) a.e, hence f ∈ Lp (X). Now limℓ→∞

Pℓ
n=1 fn (x) = f (x) a.e.,
so for a.e. x ∈ X, we have
ℓ p ℓ
!
X X
fn (x) − f (x) ⩽ Cp ( |fn (x)|)p + |f (x)|p ⩽ 2Cp g(x)
n=1 n=1

for some constant Cp (since (a + b)p ⩽ Cp (ap + bp ). See the beginning of proof for Minkowski’s).
Now we may apply LDCT (since g is integrable):
ˆ ℓ
X
p

fn (x) − f (x) dµ → 0
X n=1

6.3 Topological Properties of Lp -spaces

Theorem 6.3.1. (X, M, µ) meas space. Then family of all simple functions in Lp (X) (with
1 ⩽ p ⩽ ∞) is dense in Lp (X). That is, for all f ∈ Lp (X), there exists simple functions sn ∈ Lp (X)
with sn → f in Lp (X).

Proof. We proved this for L1 (X) a while ago. Should be the same! Exercise.

Definition 6.3.1. (X, T ) a topological space. X is normal if C1 , C2 ⊆ X closed with C1 ∩C2 = ∅,

then there exists open sets U1 , U2 ⊆ X such that C1 ⊆ U1 and C2 ⊆ U2 with U1 ∩ U2 = ∅.

Remark. Metric spaces (and thus normed spaces) are normal! Just let

U1 := {x ∈ X : dist(x, C1 ) < dist(x, C2 )}

and U2 := {x ∈ X : dist(x, C1 ) > dist(x, C2 )}. These are open!

Theorem 6.3.2 (Urysohn’s). (X, T ) is normal iff C1 , C2 closed with C1 ∩ C2 = ∅, then there
exists a continuous function φ : X → [0, 1] such that φ ≡ 1 on C1 and φ ≡ 0 on C2 .

Theorem 6.3.3. (X, M, µ) meas space with X a normal topological space and B(X) ⊆ M.
Assume that for all E ∈ M with µ(E) < ∞,

µ(E) = sup{µ(C) : C closed with C ⊆ E} = inf{µ(A) : A open with E ⊆ A}

Then Lp (X) ∩ Cb (X) (Cb (X) is the set of continuous and bounded functions) is dense in Lp (X).
If in addition, X is a metric space, then Lp (X) ∩ Cc (X) is dense in Lp (X) (Cc (X) are continuous
functions with compact support).

55
Proof. Apparently using Urysohn’s theorem.

Definition 6.3.2 (Dual). Let Y be a normed space. The dual of Y , written Y ′ is the (vector)
space of all linear and continuous maps L : Y → R.

Remark. Some basic properties of the dual (that we won’t prove):

|L(y)|
Lemma 6.3.4. Y ′ is a normed space with ∥L∥Y ′ := sup∥y∥Y ⩽1 |L(y)| = supy̸=0 ∥y∥
.
Also, Y ′ is always a Banach space.

Definition 6.3.3 (Bidual). The Bidual of Y is (Y ′ )′ , which we write as Y ′′ .

Remark. In some loose sense, Y ⊆ Y ′′ . In particular, we can embed Y in Y ′′ using the embedding
J : Y → Y ′′ mapping y 7→ J(y) : Y ′ → R where J(y) maps L 7→ L(y). Y is called reflexive if
J is a bijection. This immediately implies that Y is Banach, since Y ′′ is Banach. Moreover, this
embedding preserves norms!

Note. In a normed space Y , B(y, r) is compact iff dim Y < ∞.

Y = Lp (R) is not finite-dimensional! In particular, if we have some fn with supn ∥fn ∥p < ∞, it
does not imply there exists a subsequence with ∥fnk − f ∥p → 0.
To get this type of guarantee, we need a weaker form of convergence.

Definition 6.3.4. Let (Y, ∥·∥) be a normed space. The weak topology in Y is the smallest
topology on Y that renders continuous every L ∈ Y ′ (recall that these L are linear and already
continuous wrt Y !). i.e., it is the smallest topology that contains all sets of the form L−1 (U ) for
L ∈ Y ′ and U open in R.
If yn converges to y in the weak topology, we write yn ⇀ y.

Remark. yn → y in ∥·∥ implies that yn ⇀ y, but not the other way around!

Theorem 6.3.5. In a Banach space, closed balls are sequentially compact wrt weak topology iff
the space if reflexive.

Proposition 6.3.6. (Y, ∥·∥) normed space. yn ⇀ iff ∀L ∈ Y ′ , L(yn ) → L(y).

´ ´
Remark. Recall that we defined weak convergence in L1 as fn ⇀ f whenever X fn g dµ → X f g dµ
for every g ∈ L∞ (X).
This is special case of the above proposition! In particular, one can show that (L1 (X))′´≡ L∞ (X).
That is, L ∈ (L1 (X))′ if and only if ∃g ∈ L∞ (X) such that for all f ∈ L1 (X), L(f ) = X f g dµ.

Theorem 6.3.7 (Riesz Representation Theorem). (X, M, µ) meas space, and 1 ⩽ p < ∞. Let p′
p
be the conjugate exponent of p (p′ = p−1 ). Then every continuous linear functional L : Lp (X) → R
′
is represented by a unique g ∈ Lp (X) in the sense that for all f ∈ Lp (X),
ˆ
L(f ) = f g dµ (*)
X

56
 ′
Moreover, ∥L∥(Lp )′ = ∥g∥Lp′ . Conversely, if g ∈ Lp (X), then (*) determines an element L ∈
(Lp (X))′

Remark. Typically, this does not hold for p = ∞. In particular, we usually have L1 (X) ⊊
(L∞ (X))′ . So L1 (X) is not reflexive.
Thus, in L1 (X), even if we have a sequence of uniformly bounded functions, it does not guarantee
a weakly convergence subsequence! Recall Dunford-Pettis: it says that to get a weakly convergent
subsequence, we need more than just bounded. We also need equi-integrability (and some other
stuff).
But for Lp (X) for 1 < p < ∞, we do get weak convergent subsequences!

Corollary 6.3.8. 1 < p < ∞, if supn ∥fn ∥p < ∞, then there exists a subsequence {fnk } and
′
f ∈ Lp (X) such that fnk ⇀ f in Lp . i.e., for all g ∈ Lp ,
ˆ ˆ
fnk g dµ → f g dµ
X X

Definition 6.3.5. (Y, ∥·∥), Ln ∈ Y ′ . {Ln } has weak-* convergence to L (written Ln ⇀ L in

weak-*) if ∀y ∈ Y , Ln (y) → L(y).

Theorem 6.3.9. (Y, ∥·∥) separable normed space. If supn ∥Ln ∥Y ′ < ∞, then there exists a subse-
quence with Lnk ⇀ L in weak-*.

Corollary 6.3.10. (X, M, µ). Suppose that L1 (X) is separable, and we have fn ∈ L∞ (X) with
supn ∥fn ∥∞ < ∞, then there exists a subsequence and some f ∈ L∞ with fnk ⇀ f in weak-*.

Duals and Decompositions of Lp spaces

7.1 Radon-Nikodym
Definition 7.1.1 (Absolutely Continuous). (X, M) measurable space, µ, ν : M → [0, ∞] mea-
sures. The measure ν is said to be absolutely continuous wrt µ (written ν << µ) if ν(E) = 0
whenever µ(E) = 0.

Example 7.1.1. µ a measure. If f ∈ L1 (X), f ⩾ 0, then

ˆ
ν(E) := f dµ
E

is a measure with ν << µ.

Theorem 7.1.1 (Radon-Nikodym Theorem). (X, M) measurable space, µ, ν : M → [0, ∞] mea-

sures on the space, and µ is σ-finite. If ν is absolutely continuous wrt µ, then there exists a unique
(up to sets of measure µ = 0) f : X → [0, ∞] measurable such that
ˆ
ν(E) = f dµ
E

57
In this case, f is called the Radon-Nikodym derivative of ν wrt µ. We write
dν
f=
dµ

Note. The ultimate goal (which we will use the Radon-Nikodym theorem to prove): for any ν, µ
with µ being σ-finite, we want to write ν as a sum of two measures: νac + νs . νac will be absolutely
continuous wrt µ, and νs is “singular” wrt µ. Singular is some notion of “orthogonality” to µ.

Lemma 7.1.2. (X, M) measurable space, µ, ν measures. For E ∈ M:

ˆ ˆ 
′ ′ ′
νac (E) := sup f dµ f : X → [0, ∞] measurable and f dµ ⩽ ν(E ) for all E ⊆ E, E ∈ M
E E′

Then
• νac is a measure
• νac << µ
• For every E ∈ M, the supremum in the definition of νac (E) is attained by a function fE .
• If νac is σ-finite, then fE is independent of E. That is, there is some f such that for all E,
ˆ
νac (E) = f dµ
E

Proof. Observe that if µ(E) = 0, then νac (E) = 0, so νac << µ is immediate.
Step 1: νac is a measure. We clearly have νac (∅) = 0. To show monotonicity (we will use this to
prove
´ countable additivity), take E1 , E2 with E1 ⊆ E2 . Let f : X → [0, ∞] measurable such that
E′
f dµ ⩽ ν(E ′ ) for all E ′ ⊆ E1 . Define
(
f (x) x ∈ E1
g(x) :=
0 x ∈ X \ E1

g is measurable. If E ′ ∈ M, then
ˆ ˆ ˆ ˆ
g dµ = g dµ + g dµ = f dµ ⩽ ν(E ′ ∩ E1 ) ⩽ ν(E ′ )
E′ E ′ ∩E1 E ′ \E1 E ′ ∩E1

So g is admissible for νac (E2 ). Hence

ˆ ˆ ˆ ˆ
f dµ = f dµ + 0 dµ = g dµ ⩽ νac (E2 )
E1 E1 E2 \E1 E2

Take the sup over f on the LHS and we get νac (E1 ) ⩽ νac (E2 ).
S
Now we prove countable
P additivity. Let {En } pairwise disjoint (and write E := En ). We want
to show νac (E) = νac (En ). To see the (⩽) direction, let f : X → [0, ∞] measurable such that

58
´ ´
E′
f dµ ⩽ ν(E ′ ) for all E ′ ⊆ E. Then for all n, E′
f dµ ⩽ ν(E ′ ) for all E ′ ⊆ En , so f is also
admissible for En . Hence
ˆ X∞ ˆ X∞
f dµ = f dµ ⩽ νac (En )
E n=1 En n=1
P∞
by LMCT. Now take the sup on LHS to get νac (E) ⩽ n=1 νac (En ).
Now to show the (⩾) direction, WLOG say νac (E) < ∞. By the monotonicity we proved above,
νac (E´n ) ⩽ νac (E) < ∞. Given ε > 0, for every n we can find measurable fn : X → [0, ∞] such
that E ′ fn dµ ⩽ ν(E ′ ) whenever E ′ ⊆ En , and
ˆ
ε
νac (En ) ⩽ fn dµ + n
En 2

(by the definition of the sup used to define νac ). Define f := ∞ ′

P
n=1 χEn fn . For E ⊆ E, write
ˆ ∞ ˆ
X
f dµ = fn dµ
E′ n=1 E ′ ∩En

X∞
⩽ ν(E ′ ∩ En )
n=1
[
= ν( (E ′ ∩ En ))
= ν(E ′ )

So f is admissible for νac (E). Therefore we have

∞ ∞ ˆ ˆ
X X ε
νac (En ) ⩽ ( fn dµ + n ) = f dµ + ε ⩽ νac (E) + ε
n=1 n=1 En 2 E

Let ε → 0 and we get the desired inequality. This completes the proof of countable additivity, so
νac is a measure!
Step 2: Fix E ∈ M. We want to find fE . First observe that if f, g are admissible for E, then
max(f, g) is still admissible:
ˆ ˆ ˆ
max(f, g) dµ = f dµ + g dµ ⩽ ν(E ′ ∩ {f ⩾ g}) + ν(E ′ ∩ {f < g}) = ν(E)
E′ E ′ ∩{f ⩾g} E ′ ∩{f <g}

Therefore,
´ there exists an increasing sequence of admissible functions fn : X → [0, ∞] such that
limn E fn dµ = νac (E). So, we just take f := limn fn . f is admissible because by LMCT, for every
E ′ we have
ˆ ˆ
f dµ = lim fn dµ ⩽ ν(E ′ )
E′ n E′
´ ´
Identically, we have that E
f dµ = limn E
fn dµ = νac (E) as desired for the third point of the
lemma.

59
Step 3: Finally, we want to show that if νac is σ-finite, then there is one f that works for every
E. Assume first that νac
´ is finite. Now apply Step 2 with ´ E = X to get some f admissible (for
X) such that νac (X) = X f dµ. We claim that νac (E) = E f dµ for every E ∈ M.
Observe that since f is admissible for X, then f is admissible for any set E. Hence
ˆ
νac (E) ⩾ f dµ (*)
E

So it remains to show (⩽).

´
Suppose for contradiction that there exists
´ some E such that νac (E) > E f dµ. But since (*)
works for any set, we get νac (X \ E) ⩾ X\E f dµ. But also νac is a measure, so
ˆ ˆ ˆ
νac (X) = νac (E) + νac (X \ E) > f dµ + f dµ = f dµ = νac (X)
E X\E X

a contradiction!
Now assume more generally that νac is σ-finite. Write X = ∞
S
n=1 Xn with νac (Xn ) < ∞, pairwise

P n we have some fn such that νac (E ∩

disjoint.´ Then by the above proof for finite X, for every
Xn ) = E´fn dµ for every E ∈ M. Now just let f := ∞ n=1 χXn fn (exercise to show this satisfies
νac (E) = E f dµ for every E).

Definition 7.1.2 (Signed measure). (X, M) measurable space, a signed measure is a function

λ : M → [−∞, ∞]

such that
(i) λ(∅) = 0
(ii) λ takes at most one of the two values −∞ and +∞. i.e., either λ : M → [−∞, ∞) or
λ : M → (−∞, ∞] (hence we can add λ(E1 ) + λ(E2 ) and it always makes sense).
(iii) If En ∈ M, are pairwise disjoint, then
∞
! ∞
[ X
λ En = λ(En )
n=1 n=1

Definition 7.1.3. (X, M) measurable space and λ : M → [−∞, ∞] a signed measure. A set
E ∈ M is said to be positive (resp. negative) if ∀F ⊆ E with F ∈ M, λ(F ) ⩾ 0 (resp.
λ(F ) ⩽ 0).

Lemma 7.1.3. (X, M) measurable space, λ : M → [−∞, ∞] a signed measure. Write

λ+ (E) := sup{λ(F ) : F ⊆ E, F ∈ M} ⩾ 0

λ− (E) := − inf{λ(F ) : F ⊆ E, F ∈ M} = sup{−λ(F ) : F ⊆ E, F ∈ M} ⩾ 0

(Observe also that λ− = (−λ)+ , since −λ is also a signed measure). Then λ+ and λ− are non-
negative measures. Moreover, if λ : M → [−∞, ∞) and λ− (E) < ∞, then:

60
 • for every E ∈ M,

λ+ = sup{λ(F ) : F ⊆ E, F ∈ M, λ− (F ) = 0} (*)

• λ+ is finite.
• λ = λ+ − λ− .
(and similarly if λ : M → (−∞, ∞]).

Proof. Step 1: we first prove that λ+ is a measure. Clearly λ+ (∅) = λ(∅) = 0. Next, observe that
if E1 ⊆ E2 , then clearly λ+ (E1 ) ⩽ λ+ (E2 ) (we will use this monotonicity to help prove countable
additivity).
Next assume {En } are pairwise disjoint. Let E := n En . We want to show λ+ (E) = n λ+ (En ).
S P
First show (⩾). If there is some n such that λ+ (En ) = ∞, then we are done by monotonicity, so
assume all λ+ (En ) < ∞. Fix ε > 0. For all n, there exists Fn ⊆ En such that λ(Fn ) ⩾ λ+ (En )− 2εn .
Moreover, since λ is a measure,
[ X X
λ+ (E) ⩾ λ( Fn ) = λ(Fn ) ⩾ ( λ+ (En )) − ε
n n n

and let ε → 0.
Now to show (⩽): let F ∈ M such that F ⊆ E (so it is admissible for the sup defining λ+ (E)).
And define Fn = F ∩ En . Then
[ X
λ(F ) = λ( Fn ) = λ(Fn ) ⩽ λ+ (En )
n n

Now take the sup on the LHS over F .

Step 2: now assume that λ : M → [−∞, ∞) and λ− (E) < ∞. Observe that (*) holds if: for any
F ∈ M with F ⊆ E and λ− (F ) > 0, then there exists some G ⊆ F with G ∈ M and λ− (G) = 0
and λ(G) ⩾ λ(F ). So this is what we will prove.
Consider some such F ⊆ E (i.e., λ− (F ) > 0). If λ(F ) ⩽ 0, then take G = ∅. So assume λ(F ) > 0.
Since F ⊆ E and we assumed λ− (E) < ∞, then also λ− (F ) < ∞. Proceed by induction. There
exists G1 ∈ M with G1 ⊆ F such that −λ(G1 ) ⩾ λ− (F ) − 1 and −λ(G1 ) ⩾ 0 (G1 exists by
the definition of λ− (F )). Set F1 := F \ G1 . Then λ(F1 ) = λ(F ) − λ(G1 ) ⩾ λ(F ). Observe that
−λ− ⩽ λ, and hence λ− (F1 ) = λ− (F ) − λ− (G1 ) ⩽ λ− (F ) + λ(G1 ) ⩽ 1.
Now for induction, assume we have Fn−1 ⊆ F with Fn−1 ∈ M, λ(Fn−1 ) ⩾ λ(F ) and λ− (Fn−1 ) ⩽
1
n−1
. To construct Fn : if λ− (Fn−1 ) = 0, take G := Fn−1 , and we are done. Else, λ− (Fn−1 ) > 0,
so there exists a Fn ⊆ Fn−1 such that −λ(Gn ) ⩾ 0 and −λ(Gn ) ⩾ λ− (Fn−1 ) − n1 . Then let
Fn := Fn−1 \ Gn . Then we (identically to G1 ) have
1
λ(Fn ) ⩾ λ(F ) and λ− (Fn ) ⩽
n
T
Finally, take G := n Fn . Since λ(Fn ) ∈ [0, ∞) and Fn ⊆ Fn−1 , then λ(G) = limn λ(Fn ) ⩾ λ(F ).
Moreover, λ− (G) = limn λ− (Fn ) = 0. This proves (*)

61
Step 3: Now we show λ(E) = λ+ (E) − λ− (E). Observe that λ(E) > −∞, otherwise −λ(E) = ∞,
hence λ− (E) = ∞, a contradiction to our Step 2 assumption! So
λ+ (E) − λ(E) = sup{λ(F ) : F ⊆ E, F ∈ M} − λ(E)
= sup{λ(F ) − λ(E) : F ⊆ E, F ∈ M}
= sup{−λ(E \ F ) : F ⊆ E, F ∈ M}
= sup{−λ(G) : G ⊆ F, G ∈ M}
= λ− (E)

Restatement of Radon-Nikodym:

Theorem 7.1.4 (Radon-Nikodym Theorem). (X, M) measurable space, µ, ν : M → [0, ∞] mea-

sures on the space, and µ is σ-finite. If ν is absolutely continuous wrt µ, then there exists a unique
(up to sets of measure µ = 0) f : X → [0, ∞] measurable such that
ˆ
ν(E) = f dµ
E

Proof. Step 1: If µ, ν are finite, then ν finite implies that νac is finite, which implies that νac is
σ-finite. So apply Lemma 7.1.2: there exists f : X → [0, ∞] measurable such that
ˆ
νac (E) = f dµ
E
´ ´
for all E. We claim ν(E) = E f dµ also. For E ∈ M, define ν ′ (E) := ν(E) − E f dµ. By
definition of νac , we have ν ′ (E) ⩾ 0. Also ν ′ is a measure, and ν ′ <<
´ µ (since ν << µ). We want
′ ′
to show ν ≡ 0. If not, let E0 ∈ M be such that ν (E0 ) = ν(E0 ) − E0 f dµ > 0. This then implies
µ(E0 ) > 0. We can find ε > 0 with ε << 1 such that ν ′ (E0 ) > εµ(E0 ). Now
(ν ′ − εµ)− (E0 ) = (εµ − ν ′ )+ (E0 )
= sup{εµ(G) − ν ′ (G) : G ⊆ E0 , G ∈ M}
⩽ εµ(E0 ) < ∞
And also (ν ′ − εµ)+ (E0 ) ⩾ ν ′ (E0 ) − εµ(E0 ) > 0. By Lemma 7.1.3, there exists F0 ⊆ E0 such that
F0 ∈ M with (ν ′ − εµ)− (F0 ) = 0 and (ν ′ − εµ)(F0 ) > 0. That is, ν ′ (F0 ) > εµ(F0 ). ν ′ << µ, and
ν ′ (F0 ) > εµ(F0 ) imply that µ(F0 ) > 0 also.
If G ⊆ F0 , then (ν ′ − εµ)
´
−
(F0 ) = 0 implies (ν ′ − εµ)− (G) = 0 implies (ν ′ − εµ)(G) ⩾ 0. So
εµ(G) ⩽ ν ′ (G) = ν(G) − G f dµ. Equivalently,
ˆ
f + εχF0 dµ ⩽ ν(G)
G

For any E ∈ M,
ˆ ˆ ˆ
(f + εχF0 ) dµ = f dµ + (f + εχF0 ) dµ
E E\F0 E∩F0

⩽ ν(E \ F0 ) + ν(E ∩ F0 ) = ν(E)

62
 ´
In particular, we showed for all E ∈ M, E (f + ´ εχF0 ) dµ ⩽ ν(E).´ But then this function is
admissible for νac (E). So in particular, νac (X) ⩾ X (f + εχF0 ) dµ = X f dµ + εµ(F0 ) = νac (X) +
εµ(F0 ) > νac (X), a contradiction.
Now we will prove
´ ´ there exists another g : X → [0, ∞] measurable such
uniqueness. Suppose that
that ν(E) = E g dµ for all E ∈ M. So E (g − f ) dµ = 0 for every E (observe this uses the fact
that µ, ν are finite). Take E := {x ∈ X : g(x) ⩾ f (x)}. Then
ˆ
g − f dµ = 0
E | {z }
⩾0

which implies that g − f ⩽ 0 µ-a.e. Symmetrically, we can take E := {x ∈ X : g(x) ⩽ f (x)} to

see that g − f ⩾ 0 µ-a.e. In particular, g = f µ-a.e.
Step 2: Now suppose µ is finite and ν σ-finite. Let X = ∞
S
n=1 Xn with Xn ∈ M and pairwise
disjoint, and ν(Xn ) < ∞. Apply Step 1 to µ : M|Xn → [0, ∞] and ν : M|Xn → [0, ∞] which are
both finite. Then for each n, there exists a unique f : Xn → [0, ∞] such that
ˆ
ν(E ∩ Xn ) = fn dµ
E∩Xn
P
for E ∈ M. Let f (x) := n χXn (x) · fn (x). Then
∞
X Xˆ Xˆ ˆ
ν(E) = ν(E ∩ En ) = fn dµ = f dµ = f dµ
n=1 n E∩En n E∩En E

by LMCT. Then uniqueness follows from Step 1 by restricting to each Xn .

Step 3: Now assume µ is finite and ν is arbitrary (with ν << µ). WLOG say ν(X) = ∞, or
else apply Step 1. Define T := sup{µ(E) : E ∈ M, ν(E) < ∞}. T < ∞ because µ(X) < ∞. We
can findSan increasing sequence of En ∈ M with ν(En ) < ∞ such that lim µ(En ) = T . Define
E∞ := n En , so µ(E∞ ) = lim µ(En ) = T (since the sequence is increasing).
Now ν : M|E∞ → [0, ∞] is σ-finite, so apply Step 2. There exists a unique f∞ : E∞ → [0, ∞] such
that
ˆ
ν(E) = f∞ dµ
E

for every E ∈ M|E∞ .

Claim 1: ν : M|X\E∞ takes values only in {0, ∞}. Suppose not: we have some F ⊆ X \ E∞ such
that 0 < ν(F ) < ∞. In particular, by absolute continuity, µ(F ) > 0 as well. Observe that En ∪ F
is admissible for T , hence

µ(E∞ ) = T ⩾ µ(En ∪ F ) = µ(En ) + µ(F )

for every n. Let n → ∞, then T ⩾ T + µ(F ) > T , a contradiction. This proves Claim 1.
Claim 2: if F ⊆ X \ E∞ with µ(F ) > 0, then ν(F ) = ∞. If ν(F ) < ∞, then again En ∪ F is
admissible for T , and we use an identical argument to get a contradiction.

63
Finally, define
(
f∞ (x) x ∈ E∞
f (x) :=
∞ x ∈ X \ E∞

Now let E ∈ M. If E ⊆ E∞ , we are done. Otherwise, if µ(E ∩ (X \ E∞ )) > 0, then by Claim 2,

ν(E ∩ (X \ E∞ )) = ∞, so
ν(E) = ν(E ∩ (X \ E∞ )) + ν(E ∩ E∞ ) = ∞
´ ´
while E f dµ ⩾ E∩(X\E∞ f dµ = ∞, as desired. On the other hand, if µ(E ∩ (X \ E∞ )) = 0, then
ν(E ∩ (X \ E∞ )) = 0. So
ˆ ˆ
ν(E) = ν(E ∩ E∞ ) = f dµ = f dµ
E∩E∞ E

as desired.
´
Finally, we prove uniqueness: suppose there is another g : X → [0, ∞] with ν(E) = E g dµ for all
E ∈ M. By uniqueness in Step 2, f = g µ-a.e. on E∞ . We claim that g = ∞ µ-a.e. on X \ E∞ .
Assume ∃F ⊆ X \ X∞ such that µ(F ) > 0 S and g < ∞ on F . But this implies ν(F ) = ∞, so
write Fn := {x ∈ F : g(x) ⩽ n}. Then F = n Fn and Fn ↗ F . So limn µ(Fn ) = µ(F ) > 0. In
particular, there must be some n0 >> 1 such that µ(Fn0 ) > 0. Again, this implies ν(Fn ) = ∞.
But then
ˆ
∞ = ν(Fn ) = g dµ ⩽ n0 µ(Fn0 ) < ∞
Fn0

a contradiction.
S
Step 4: Assume only that µ is σ-finite and ν << µ. Just write X = n Xn pairwise disjoint with
µ(Xn ) < ∞, and apply Step 3. Proceed as in Step 2.

7.2 Riesz Representation Theorem

Remark. Observe we can write the norm for [Lp (X)]′ in two ways:
L(f )
∥L∥[Lp (X)]′ := sup |L(f )| = sup
f ∈Lp (X) f ̸=0 ∥f ∥Lp
∥f ∥p ⩽1

Moreover, it turns out the norm of some g ∈ Lp (X) can be written as

ˆ
∥g∥Lp = sup f g dµ
′
f ∈Lp (X) X
normfp′ ⩽1

“If I were you, I would try to prove this. Like, tonight” – Irene Fonseca the class before the
midterm.
Note: proof uses Holder’s?

64
 ′
Theorem 7.2.1 (Riesz). If 1 ⩽ p < ∞, then [Lp (X)]′ = Lp (X) in the sense that L ∈ [Lp (X)]′
′
(i.e. L : Lp → R linear and continuous) if and only if there exists (a unique) g ∈ Lp (X) such that
for all f ∈ Lp (X),
ˆ
L(f ) = f g dµ
X

and furthermore,
∥L∥[Lp (X)]′ := sup |L(f )| = ∥g∥Lp′ (X)
f ∈Lp (X)
∥f ∥p ⩽1

(this final piece follows from the remark above).

Proof. Step 1: Assume for a moment that µ is finite. Assume also that L ⩾ 0, i.e. L(f ) ⩾ 0
whenever f ⩾ 0 a.e. For E ∈ M, define ν : M → R by E 7→ L(χE ).
We claim that ν is a non-negative measure. L(χ∅ ) = L(0)
S = 0 (by linearity). ν is finitely additive:
p
if En pairwise disjoint, then χ En → χE in L if E = En (exercise. Remark: this is where the
S

proof fails for p = ∞). Thus

m
X m
[ ∞
[
ν(En ) = ν( En ) = L(χ Sm
n=1 En ) → L(χE ) = ν( En )
n=1 n=1 n=1
P∞
since χSmn=1 En
→ χE in Lp , while L is continuous with respects to Lp . Thus n=1 ν(En ) = ν(E),
so ν is a measure!
Now, if µ(E) = 0, then χE = 0 µ-a.e. The 0-function is in Lp (X), so L(χE ) = L(0) = 0, and thus
ν(E) = 0. This shows that ν << µ.
´
Now apply Radon-Nikodym: there exists a unique g ⩾ 0 such that L(χE ) = ν(E) = X χE g dµ.
This gives us what we want for any f = χE (and´it extends trivially to any simple function f = s).
Now we claim that for any f ∈ Lp (X), L(f ) = X f g dµ.
If f ∈ Lp (X) with f ⩾ 0, we can find sn ⩾ 0 simple such that sn ↗ f in Lp In particular, spn ⩽ f p ,
and f p is integrable. Now we have,
ˆ ˆ
L(f ) ← L(sn ) = sn g dµ → f g dµ
X X

The left arrow is since sn → f in L and L is continuous wrt Lp . The right arrow is by LMCT. So
p

the LHS and RHS are equal, as desired.

If f is signed, just write f = f + − f − , and do the above twice (use the fact that L is a linear map).
′
Next, we prove that g ∈ Lp (X). Note that L is continuous at 0. In particular, for every ε > 0,
there exists δ > 0 such that whenever ∥f − 0∥Lp ⩽ δ we have |L(f ) − L(0)| = |L(f )| ⩽ ε. In
particular, ∥f ∥p ⩽ δ =⇒ |L(f )| ⩽ ε. Let ε = 1, and f ∈ Lp (X) with f ̸≡ 0. We then have
!
f ∥f ∥p ∥f ∥p 1
|L(f )| = L δ · ⩽1· = ∥f ∥p
∥f ∥p δ δ δ

65
Hence, sup{ |L(f
∥f ∥
)|
: f ∈ Lp (X) \ {0}} ⩽ 1
δ
< ∞. And this is equivalent to writing
p

∥L∥[Lp (X)]′ := sup{|L(f )| : f ∈ Lp (X), ∥f ∥p ⩽ 1} < ∞

But then by the remark/exercise above, we have that ∥g∥p′ = ∥L∥[Lp (X)]′ < ∞ (since we just
showed L can be written in terms of g).
′
Finally, we will prove uniqueness of g. Suppose we have g1 , g2 ∈ Lp (X) both representing L.
Then, in particular,
ˆ
f (g1 − g2 ) dµ = 0
X

for all f ∈ Lp (X). In particular (since µ is finite), we may take f = χE for any E ∈ M. Let
E + := {x : g1 (x) ⩾ g2 (x)}. Then χE + (g1 − g2 ) ⩾ 0 a.e., and thus we must have χE + (g1 − g2 ) = 0
a.e. Identically, we can show that E − := {x : g1 (x) ⩽ g2 (x)} also has measure 0. So g1 and g2 can
differ only on a set of measure 0 (something wrong here, but you get the idea).
Step 2: Now let L be general (no longer assume L ⩾ 0). We can write L = L+ − L− so that
L+ , L− are both linear and continuous functions Lp (X) → R, and both non-negative operators.
Suppose first f ∈ Lp (X) and f ⩾ 0. Define

L+ (f ) := sup{L(g) : 0 ⩽ g ⩽ f }

L− (f ) := − inf{L(g) : 0 ⩽ g ⩽ f }
For general f ∈ Lp , define L+ and L− linearly by writing f = f + − f − . We can prove that L+ and
L− are linear, continuous, and non-negative, and that L = L+ − L− (all of these are an exercise!).
Now apply Step 1 to L+ and L− separately to get g1 and g2 , and choose g = g1 − g2 .
Step 3: do the usual tricks for µ being σ-finite.

7.3 Lebesgue Decomposition Theorem

Definition 7.3.1 (Mutually Singular). (X, M) measurable space, µ, ν non-negative measures. We
say µ and ν are mutually singular, written µ ⊥ ν, if there exist disjoint measurable sets Xµ , Xν
such that X = Xµ ∪ Xν and for every E ∈ M,

µ(E) = µ(E ∩ Xµ ) and ν(E) = ν(E ∩ Xν )

(intuition: it means their supports are disjoint: the sets that one sees are not seen by the other).

Lemma 7.3.1. (X, M) measurable space, µ, ν non-negative measures. Define for E ∈ M

νs (E) := sup{ν(F ) : F ⊆ E, F ∈ M, µ(F ) = 0}

Then
• νs is a measure

66
 • For each E ∈ M, the supremum is attained by a measurable set.
• Moreover, if νs is σ-finite (in particular, if ν is σ-finite), then νs ⊥ µ.

Proof. Step 1: νs is a measure: exercise!

Step 2: Next we show sup = max. Fix E ∈ M, if νs (E) = 0, then F = ∅, so suppose that
νs (E) > 0. Then we can find tn ↗ such that 0 < tn <Sνs (E) and tn ↗ νs (E). Find Fn ⊆ E such
that µ(Fn ) = 0 and tn < ν(Fn ) ⩽ νs (E). Let En := nk=1 Fk . Then En ⊆ E, and µ(En ) = 0, so
En is admissible for νs (E). Hence

νs (E) ⩾ ν(En ) ⩾ ν(Fn ) > tn → νs (E)

S
so Sνs (E) ⩾ lim ν(En ) ⩾ lim inf ν(Fn ) ⩾ νs (E), while lim ν(En ) = ν( En ), so we must have
ν( En ) = νs (E), as desired.
Step 3: Now assume that νs is finite (then generalize to σ-finite). There exist Xs ∈ M such that
νs (X) = ν(Xs ) and µ(Xs ) = 0. We claim for all E ∈ M,

νs (E) = νs (E ∩ Xs ) and µ(E) = µ(E ∩ (X \ Xs ))

The second part is obvious since X \ Xs is measurable:

µ(E) = µ(E ∩ (X \ Xs )) + µ(E ∩ Xs ) = µ(E ∩ (X \ Xs ))

since µ(Xs ) = 0. For the first part, fix an E ∈ M. Notice that step 2 implies there is some
Es ∈ M such that Es ⊆ E with ν(E) = νs (Es ) and µ(Es ) = 0. First, observe that ν(Es \ Xs ) = 0,
or else Es ∪ Xs is admissible for νs (X). Then

∞ > νs (X) ⩾ ν(Es ∪ Xs ) = ν(Xs ) + ν(Es \ Xs ) > ν(Xs ) = νs (X)

| {z }
>0

a contradiction. So ν(Es \ Xs ) = 0, and

νs (E) = ν(Es ) = ν(Es ∩ Xs ) + ν(Es \ Xs ) = ν(Es ∩ Xs ) ⩽ νs (Ex ∩ Xs ) ⩽ νs (E)

Theorem 7.3.2 (Lebesgue Decomposition Theorem). (X, M) measurable space, µ, ν : M →

[0, ∞] measures, and µ is σ-finite. Then we can write ν = νac + νs with νac << µ.
Moreover, if ν is σ-finite, then νs ⊥ µ, and the decomposition is unique, i.e. ν = ν̄ac + ν̄s for some
measures ν̄ac , ν̄s with ν̄ac << µ and ν̄s ⊥ µ, then ν̄ac = νac and ν̄s = νs .

Proof. It suffices to show that ν = νac + νs (where νac and νs are from the previous lemmas), and
to show uniqueness.
Step 1: First, we show

ν = νac + νs (*)

67
Fix E ∈ M, find Es ∈ M such that Es ⊆ E and µ(Es ) = 0 and νs (E) = ν(Es ). If ν(Es ) = ∞,
then ν(E) = ∞ so (*) holds.
So assume ν(Es ) < ∞. We claim ν|M|E\Es << µ|M|E\Es . Let F ⊆ E \ Es with µ(F ) = 0. We want
to show ν(F ) = 0. If not, ν(F ) > 0, then Es ∪ F is admissible for νs (E). So
∞ > νs (E) =⩾ ν(Es ∪ F ) = ν(Es ) + ν(F ) > ν(Es ) = νs (E)
a contradiction.
Therefore by Radon-Nikodym, there exists a unique f˜ such that for all Ẽ ∈ M|E\Es , ν(Ẽ) =
´
Ẽ
f˜ dµ. So
ˆ
νac (E \ Es ) ⩾ f˜ dµ = ν(E \ Es ) ⩾ νac (E \ Es )
E\Es

which implies that ν(E \ Es ) = νac (E \ Es ) = νac (E). Hence, finally,

ν(E) = ν(E \ Es ) + ν(Es ) = νac (E) + νs (E)

Step 2: now we assume that ν is σ-finite (first, assume ν is finite), and we must prove uniqueness.
Let Xν̄s ∈ M such that ν(Xν̄s ) = 0 and ν̄s (E) = ν̄s (E ∩ X¯nus ) for all E ∈ M. If E ⊆ X \ X̄ν̄s ,
since ν(E) = ν̄ac (E) + ν̄s (E) = ν̄ac (E), then ν(E) = ν̄ac (E). Hence, ν|X\Xν̄s = ν̄ac |X\Xν̄s . For all
E ∈ M,
ν̄ac (E) = ν̄ac (E \ Xν̄s ) = ν(E \ Xν̄s ) ⩾ νac (E \ Xν̄s ) = νac (E)
and hence ν̄ac ⩾ νac . Symmetrically, νac ⩾ ν̄ac . So we are done if ν < ∞. Use the usual ways to
extend to σ-finite ν.

Definition 7.3.2 (Abs Continuous and Singular for Signed Measures). (X, M) measurable space,
µ : M → [0, ∞] a measure and λ : M → [−∞, ∞] a signed measure.
(i) Write λ is absolutely continuous wrt µ, λ << µ, if µ(E) = 0 implies λ(E) = 0.
(ii) Write λ and µ are mutually singular, λ ⊥ µ, if there exist disjoint sets Xλ , Xµ ∈ M such
that X = Xλ ∪ Xµ and for all E ∈ M,
λ(E) = λ(E ∩ Xλ ) and µ(E) = µ(E ∩ Xµ )
Remark. Also, observe that λ << µ implies that λ+ , λ− << µ.

Note. We we can easily extend the Lebesgue Decomp Theorem to signed measures (she didn’t
actually state it clearly...): λac := (λ+ )ac ´− (λ− )ac and λs := (λ+ )s − (λ− )s . Then λ = λac + λs and
there is an f := f + − f − with λac (E) = E f dµ. Write f =: dλdµac .

7.4 Besicovitch Derivation and Lebesgue Differentiation Theorem

Note. ν : β(RN ) → [0, ∞] finite on compact sets (this is called a Radon measure). By Lebesgue
Decomp Theorem, ν = νac + νs with νac << L N , νs ⊥ L N .

68
 ´
By Radon-Nikodym: νac (E) = E f (x)dx for all E ∈ β(RN ). Since νac (E) is finite on compact
sets, this integral is finite over compact sets. We say that f ∈ L1loc (RN ).
Then by the Lebesgue Differentiation Theorem (below), there is an E0 so that if x ∈ RN \ E0 , then
ˆ
νac (B(x, r)) 1 dν
lim N = lim f (y)dy =: f (y)dy = f (x) = (x)
r→0 L (B(x, r)) r→0 αN r N B(x,r) B(x,r) dµ

Theorem 7.4.1 (Lebesgue Differentiation Theorem). f : RN → R, f ∈ L1loc (RN ). Then there

exists Borel set E0 ⊆ RN , L N (E0 ) = 0 such that for all x ∈ RN \ E0 ,
ˆ
1
lim |f (y) − f (x)|dy = 0
r→∞ αN r N B(x,r)

In particular,
ˆ
1
lim f (y)dy = f (x)
r→∞ αN r N B(x,r)

(why?)

Remark. If g = f L N -a.e., then exists F0 measurable, L N (F0 ) = 0, such that f (y) = g(y) for
y ∈ RN \ F0 (in fact, we can assume F0 Borel). Then

|f (y) − f (x)|dy = |g(y) − f (x)|dy = |g(y) − f (x)|dy

B(x,r) B(x,r)\F0 B(x,r)

so if the LHS → 0, then RHS → 0.

We want something stronger than the above (in what way is this stronger?):

Theorem 7.4.2 (Besicovitch Derivation Theorem). ν : β(RN ) → [0, ∞] meas, finite on compact
sets. Then exists a Borel set F0 ⊆ RN , L N (F0 ) = 0 such that for all x ∈ RN \ F0

dνac ν(B(x, r))

(x) = lim N (*)
dL N r→0 L (B(x, r)

and
νs (B(x, r))
lim =0 (**)
r→0 L N (B(x, r))

Remark. Notice that (*) follows from (**), since

 
dνac νac (B(x, r)) νac (B(x, r)) νs (B(x, r)) ν(B(x, r))
(x) = lim N = lim + N = lim N
N r→0 L (B(x, r)) r→0 L (B(x, r))
N L (B(x, r)) r→0 L (B(x, r))
|dL{z }
f (x)

Please god release me from this misery. I have no idea what the fuck is going on anymore.

69
Proof. It suffices to prove (**). It can be shown that νs is regular. That is, if E ∈ β(RN ) then
νs (E) = inf{νs (U ) : E ⊆ U, U open} = sup{νs (K) : K ⊆ E, K compact}. We know νs ⊥ L N , so
there exists E0 ∈ β(RN ) with

L N (RN \ E0 ) = νs (E0 ) = 0

For t > 0, define Et := {x ∈ E0 : lim supr→0 Lν(B(x,r))

N (B(x,r)) > t} (i.e., the set of points in E0 where

(**) fails by t). Et ⊆ E0 implies that νs (Et ) = 0. By outer regularity, for any ε > 0 we can find
an open set U such that Et ⊆ U and νs (U ) < ε. Let K ⊆ Et compact.
If x ∈ K, then there exists 0 < rx < 1 such that B(x, rx ) ⊆ U and LνsN(B(x,rx ))
(B(x,rx ))
> t. Now we use
Vitali’s Covering Theorem (see below to recall). K being compact (giving a finite subcover)
plus Vitali’s covering theorem gives that
ℓ
[
K⊆ B(xk , 3rxk )
k=1

Hence
ℓ
X
L N (K) ⩽ L N (B(xk , rxk ))
k=1
ℓ
X
=3 N
L N (B(xk , rxk ))
k=1
ℓ
X 1
⩽ 3N νs (B(xk , rxk ))
k=1
t
ℓ
!
3N [
= νs B(xk , rxk )
t k=1
3N 3N ε
⩽ νs (U ) <
t t
In particular, for any K ⊆ Et compact, we have L N (K) ⩽ O(ε) (for every ε), so L N (K) = 0.
But by inner regularity of the Lebesgue measure, we then must have that L N (Et ) = 0.
Finally, just let F0 := ∞n=1 E1/n . We have L (E) = 0, and for all x ̸∈ E,
N
S

νs (B(x, r)) 1
lim sup ⩽ →0
L (B(x, r))
N n

Recall:

Theorem 7.4.3 (Vitali’s Covering Theorem). G ⊆ RN , and {B(x, rx )}x∈G with 0 < rx < R.
Then there exists a countable family of pairwise disjoint balls {B(xn , rxn )}n∈N such that
∞
[
G⊆ B(xn , 3rxn )
n=1

70
Definition 7.4.1 (Hardy-Littlewood Maximal Function). Let f : RN → R, f ∈ L1loc . The maxi-
mal function of f is M (f ) : RN → [0, ∞] defined as

M (f )(x) := sup |f (y)|dy

r>0 B(x,r)

Theorem 7.4.4. f ∈ Lp (RN ), 1 ⩽ p < ∞. Then

(i) Weak L1 inequality: If p = 1, then for all t > 0,
ˆ
3N 3N
L ({x ∈ R : M (f )(x) > t}) ⩽
N N
|f |dy = ∥f ∥1
t RN t
And in particular, M (f ) < ∞ almost everywhere.
(ii) If 1 < p ⩽ ∞, then M (f ) ∈ Lp , and

∥M (f )∥p ⩽ C(N, p)∥f ∥p

for some constant C(N, p). (Note, this is false for p = 1).

Proof. For part (i): let t > 0, and define Et := {x ∈ RN : M (f )(x) > t}. Take K ⊆ Et compact.
If x ∈ K, then (by defn of maximal function) there exists a ball B(x, rx ) with rx > 0 such that

|f |dy > t
B(x,rx )

WLOG we may assume that 0 ⩽ rx < R (this requires a nontrivial proof, but we won’t prove it).
Now we apply Vitali’s covering theorem and compactness of K: we have
ℓ
[
K⊆ B(xk , 3rk )
k=1

where the B(xk , rk ) are pairwise disjoint. So

ℓ
X
L N (K) ⩽ L N (B(xk , 3rk ))
k=1
ℓ
X
= 3N L N (B(xk , rk ))
k=1
ℓ ˆ
N
X 1
⩽3 |f (y)|dy
k=1
t B(xk ,rk )
ˆ
3N
= |f (y)|dy
t Sℓ
B(xk ,rk )
ˆ k=1

3N
⩽ |f |dy.
t RN

71
 3N
´
Finally, by inner regularity of L N , we get L N (Et ) ⩽ |f |dy, as desired.
t RN

Moreover, if we look at the set {x ∈ RN : M (f )(x) = ∞} = {x ∈ RN : M (f )(x) > n}. So

T

ˆ
3N
L ({x ∈ R : M (f )(x) = ∞}) = lim L (En ) ⩽
N N N
|f |dy = 0
n→∞ n RN
So, in particular, we have showed that the maximal function of f is in fact finite almost everywhere!
For part (ii): p = ∞ is trivial (in fact, M (f )(x) ⩽ ∥f ∥∞ for all x ∈ RN ), so let 1 < p < ∞. For
t > 0, define
(
f (x) if |f (x)| > 2t
ft (x) :=
0 otherwise

We claim that ft ∈ L1 (RN ). Indeed:

ˆ ˆ
|ft (y)|dy = |f (y)|dy
RN {x:|f (x)|>t/2}
ˆ
2 p−1 t
=( ) |f (y)|( )p−1 dy
t 2
ˆ{x:|f (x)|>t/2}
2
⩽ ( )p−1 |f (y)|p dy
t {x:|f (x)|>t/2}

⩽∞

Also, |f | ⩽ |ft | + 2t , so

t
M (f ) ⩽ M (ft ) +
2
Therefore,
t
{x ∈ RN : M (f )(x) > t} ⊆ {x ∈ RN : M (ft ) > }
2
and we can apply part (i):
ˆ ˆ
3N 2 3N 2
L ({x ∈ R : M (f )(x) > t}) ⩽
N N
|ft |dy = |f (y)|dy
t RN t {x:|f (x)|>t/2}

72
Finally, recall that we proved (as a consequence of Tonelli’s theorem):
ˆ ˆ ∞
p
(M (f )(y)) dy = L N ({x ∈ RN : (M (f )(x))p > s})ds
RN
ˆ ∞
0

=p tp−1 L N ({x : M (f )(x) > t})dt (change variables t := s1/p )

0
ˆ ∞ ˆ 
N p−2
⩽ 3 2p t |f (y)|dy dt
0 {x:|f (x)|>t/2}
ˆ ˆ !
2|f (y)|
= 3N 2p |f (y)| tp−2 dt dy
RN 0
p+1 ˆ
p2
= 3N |f (y)|p dy
p − 1 RN
= C(N, p)∥f ∥pp

We restate:

Theorem 7.4.5 (Lebesgue Differentiation Theorem). f : RN → R, f ∈ L1loc (RN ). Then there

exists Borel set E0 ⊆ RN , L N (E0 ) = 0 such that for all x ∈ RN \ E0 ,

lim |f (y) − f (x)|dy = 0

r→∞ B(x,r)

In particular,

f (y)dy → f (x)
B(x,r)

(why? Exercise)

Proof. We will assume only that f ∈ L1 . If we can do this, then this will imply the theorem for
f ∈ L1loc by multiplying f by a characteristic function on a (compact) ball (this is now in L1loc ).
We can do this for balls covering all of RN , then get an E0 for each of these and take their union.
´
Fix ε > 0. Let gε ∈ Cc (RN ) such that RN |f − gε | dµ < ε. gε is uniformly continuous, so for all
η > 0, there exists δ = δ(η) such that if x, y ∈ RN with ∥x − y∥ < δ, then |gε (x) − gε (y)| < η. So
if 0 < r < δ,
ˆ
1
|gε (y) − gε (x)| dµ ⩽ η dµ = η
B(x,r) αN rN B(x,r)

In particular, for every x ∈ RN ,

|gε (y) − gε (x)| dµ → 0

B(x,r)

73
We have
ˆ
1
lim sup |f (y) − f (x)|dy
αN rN B(x,r)
ˆ
1
= lim sup |f (y) − gε (y) + gε (y) − f (x) + gε (x) − gε (x)|dy
αN rN B(x,r)
" #
⩽ lim sup |f (y) − gε (y)|dy + |gε (y) − gε (x)|dy + |gε (x) − f (x)|
B(x,r) B(x,r)
| {z }
→0
= M (f − gε )(x) + |gε (x) − f (x)|

For t > 0, define

Gt := {x ∈ RN : lim sup |f (y) − f (x)|dy > t}

B(x,r)

Et,ε := {x ∈ RN : M (f − gε )(x) > t}

Ft,ε := {x ∈ RN : |gε (x) − f (x)| > t}
So we have proved that G2t ⊆ Et,ε ∪ Ft,ε . Now, using part (i) from the theorem about Maximal
functions, we know
ˆ
C(N ) C(N )
L (Et,ε ) ⩽
N
|f − gε |dy ⩽ ε
t RN t
And by Chebychev’s we get
ˆ
1 ε
L (Ft,ε ) ⩽
N
|f − gε |dy ⩽
t RN t
So in total, we have
C(N ) + 1
L N (G2t ) ⩽ L N (Et,ε ) + L N (Ft,ε ) ⩽ ε
t
Now letting ε → 0, we have that L N (G2t ) = 0. And recall that t was arbitrary, so this holds for
all t.
Finally, we define our desired set:
∞
[
E0 := G1/n
n=1

First, L N (E0 ) = 0. And if x ̸∈ E0 , then

1
lim sup |f (y) − f (x)|dy ⩽
B(x,r) n
for all n, so the limit is 0.

74