Patrick Lei

All notes

Analysis I

Lectures by Robin Young, UMass, Fall 2018.

Real Analysis 1
Math 623
Notes by Patrick Lei, May 2020
Lectures by Robin Young, Fall 2018

University of Massachusetts, Amherst

Disclaimer

These notes are a May 2020 transcription of handwritten notes taken during lecture in Fall 2018. Any errors are mine and not the instructor’s. In addition, my notes are picture-free (but will include commutative diagrams) and are a mix of my mathematical style (omit lengthy computations, use category theory) and that of the instructor. If you find any errors, please contact me at plei@umass.edu.

1 Lebesgue Measure

1.1 Motivation for the Course

We will consider some examples that show us some fundamental problems in analysis.

Example 1.1 (Fourier Series).

Let f be periodic on [0,2⁢π]. Then if f is sufficiently nice, we can write

where an=12⁢π⁢∫02⁢πf⁡(x)⁢ei⁢n⁢x⁢dx. If this is possible, then we have Parseval’s identity:

However, the function space we are working in on the RHS is not complete because there exist sequences an such that the associated function |f|2 is not Riemann-integrable. This tells us that we need a better notion of integration.

Example 1.2 (Dirichlet Function).

Enumerate the rationals in the interval [0,1] by rn. Then define

Then the limit of fN is simply the indicator function f=χℚ. Here, the fN are Riemann integrable, but f is not.

Next, we can consider the problem of characterizing which functions are Riemann-integrable.

Theorem 1.3.

A function f is Riemann-integrable if and only if the set of points where f is not continuous has measure 0.

Fourth, we may consider lengths of curves. Suppose we have a fractal (this will remain undefined) such as the Cantor set, a coastline, or broccoli. What is is the correct dimension for such a set? When is the “length” of such a curve finite?

Finally, recall the fundamental theorem of calculus: For a differentiable function F,

and for an integrable function f,

When do these two results apply?

Question 1.4.

All of these examples tell us that the fundamental problem in analysis is the exchange of limits. We will see that many of the results in the course are of this form.

1.2 Basic Notions

First recall the very basic notions of set theory: unions, intersections, complements, relations, and functions. We use the usual definitions of image, preimage, and partial and total orders.1 We will also assume the axiom of choice, so in particular, we have the following results:

Hausdorff Maximal Principle:

Every partially ordered set has a maximal totally ordered subset.

Zorn’s Lemma:

If X is a partially ordered set where every chain has an upper bound, then X has a maximal element.

Well Ordering Principle:

Ever set admits a well-ordering.

Next, we will recall some basic properties of Euclidean space.

Theorem 1.5 (Bolzano-Weierstrass).

Every bounded sequence in ℝn has a limit point. Note that the lim sup is an accumulation point.

Next recall that ℝn with the Euclidean distance ρ is a metric space. In fact, ℝn is a complete normed metric space, and the triangle inequality is implied by Cauchy-Schwarz. Recall that open balls are given by Br⁢(x)={y∣ρ⁡(x,y)<r} and closed balls are given by B¯r⁢(X)⁢{y∣ρ⁡(x,y)≤r}. Then a set G is open if around any x∈G we can find an open ball Br⁢(x)⊂G and a set F is closed if Fc is open. We can then define the interior, closure, and boundary of a set in the usual way.

Example 1.6.

Consider ℚ⊂ℝ. Then ℚ∘=∅ and ℚ¯=ℝ, so ℚ is a countable dense subset of ℝ. Therefore, ℝ is separable.

Convergence is defined in the usual way.

Lemma 1.7.

Let X be a metric space and E⊂X. Then the following are equivalent:

  1. 1.

    x∈E¯;

  2. 2.

    For all ε>0, Be⁢(x)∩E≠∅;

  3. 3.

    There exists a sequence {xn}⊂E such that xn→x.

1.3 Rectangles

We now build towards defining the Lebesgue measure. First, we call any set of the form

a closed rectangle. We define the volume |R| of any rectangle by

It should be obvious what a cube and an open rectangle are.

Definition 1.8.

A collection of rectangles is almost disjoint if their interiors are disjoint.

Lemma 1.9.

If a rectangle R is an almost disjoint union R=⋃Rj, then |R|=∑|Rj|.

Now let ε>0 small enough. Then Rε=∏[ai+ε,bi−ε]⊂R∘⊂R, As ε→0, we obtain |R∘|=|R|.

Corollary 1.10.

Let R⊂⋃i=1nRi, where the union is not necessarily almost disjoint. Then |R|≤∑|Ri|.

Our goal is to extend this notion of volume to open sets in terms of rectangles.

Lemma 1.11.

In ℝ, any open set G is a disjoint union of countable many open intervals.

Proof.

Let x∈G. Then there is a maximal interval Ix⊆G. Then if y∈Ix, note that Ix=Iy. Thus G=⋃x∈GIx. However, each interval must contain a rational number, so there are only countably many of them. □

Therefore, if G⊆ℝ is open, we can declare its measure to be μ⁡(G)=∑|Ij|, where G is a disjoint union of the intervals Ij. Note that this is not convenient in higher dimension, so we use closed rectangles instead. Note that any open set in ℝ is a countable union of closed rectangles G=⋃[aj+1/n,bj−1/n], but this union is not even almost disjoint.

Theorem 1.12.

In ℝd, any open set G can be written as an almost disjoint union G=⋃j=1∞Fj of closed rectangles. Given this, we expect m⁡(G)=∑j=1∞|Fj|.

Proof.

Grid ℝd by the points in ℤd. Then each cube Q of length 1 satisfies either Q⊆G,Q⊆G¯c, or neither. At the first step, include all Q⊆G and exclude all Q⊆G¯c. Then we reduce ℤd to 12⁢ℤd and proceed by induction. □

Example 1.13 (Cantor Set).

This is not really an example of rectangles, but is an important example for this course. It is a closed, compact, totally disconnected set where every point is a limit point. Define the sets

  • •

    C0=[0,1];

  • •

    C1=[0,1/3]∪[2/3,1];

  • •

    C2=[0,1/9]∪[2/9,1/3]∪[2/3,7/9]∪[8/9,1],

and then continue the process. Then we define the Cantor set to be C=⋂k=1∞Ck.

First, note that C is an intersection of closed sets, so it is closed. Then given x,y∈C, note that there exists k such that x−y>23k, so x,y are in different intervals in Ck. Clearly there are no isolated points. Finally, note that C is the set of all real numbers with no 1 in their ternary expansions, so it is uncountable. However, note that Cc is a disjoint union of open intervals with measure

Therefore, we expect m⁡(C)=0.

1.4 Outer Measure

It is reasonable to expect that if G is an almost disjoint union G=⋃jQj then m⁡(G)=∑j|Qj|. Instead of imposing this inequality for general E, we will make it an inequality, so we drop the condition of being almost disjoint. Define for any E⊂ℝd the outer measure

We can show that m∗⁢(C)=0 directly. Note that Ck is a union of 2k closed subintervals of length 3−k, so we see that m∗⁢(C)≤(23)k for all k. Therefore, m∗⁢(C)=0. For another example, it is easy to show that if P is countable, then m∗⁢(P)=0. To see this, let ε>0 and then take the rectangle of width ε2i centered at p. Then m∗⁢(P)≤2⁢ε.

Properties of outer measure. Here are some properties of the outer measure:

  1. 1.

    (Monotonicity) If E1⊆E2, then m∗⁢(E2)≤m∗⁢(E2). As a corollary, bounded sets have finite outer measure.

  2. 2.

    (Countable subadditivity) If E⊆⋃jEj, then m∗⁢(E)≤∑jm∗⁢(Ej).

    Proof.

    Assume that each m∗⁢(Ej)<∞. Then for all j, there exist countably many cubes Qj,k such that Ej⊆⋃kQj,k, so we have

    Then any E⊆⋃jEj⊆⋃j,kQj,k, so

  3. 3.

    For any E⊆ℝd, we can write

    Proof.

    It is easy to see that m∗⁢(E)≤infm∗⁢(G). In the other direction, assume m∗⁢(E) is finite. Then E⊆⋃Qj, so ∑|Qj|−ε≤m∗⁢(Ej)≤∑|Qj|. Choose Gj open such that Qj¯⊆Gj and |Gj|≤|Qi|+ε2k.

    Now E⊆⋃Gj and G=⋃Gj, so

  4. 4.

    (Wannabe additivity) If E=E1+E2 and d⁡(E1,E2)>0, then m∗⁢(E)=m∗⁢(E1)+m∗⁢(E2). Here d⁡(A,B)=inf{d(x,y)∣x∈A,y∈B}.

    Proof.

    Choose 0<δ<d⁡(E1,E2) and choose Qj such that E⊆⋃Qj, m∗⁢(E)>∑|Qj|−ε, and diam⁡(Qj)<δ. If E1∩Qj≠∅, then E2∩Qj=∅. Therefore, there exist J1,J2 such that E1⊆⋃j∈J1Qj and E2⊆⋃i∈J2Qi, so

  5. 5.

    If E is a countable union of almost disjoint cubes, then m∗⁢(E)=∑|Qj|.

    Proof.

    For each j, pick Qj~⊆Qj∘ such that |Q~j|>|Qj|−ε2j. Then ⋃Q~j⊆E and are disjoint and a finite distance apart, so for each N, we have ∑1Nm∗⁢(Q~j)≤m∗⁢(E), and therefore ∑j=1n|Qj|−ε≤m∗⁢(E), and thus ∑j=1∞|Qj|≤m∗⁢(E). □

    Note we cannot conclude finite additivity for general disjoint sets E1,E2.

1.5 Measurable Sets

Definition 1.14.

A subset E⊆ℝn is measurable if for all ε>0, there exists G⊇E such that m∗⁢(G∖E)<ε.

If E is measurable, define its Lebesgue measure as m⁡(E)=m∗⁢(E). Here are some properties of measurable sets:

  1. 1.

    Open sets are measurable;

  2. 2.

    If m∗⁢(E)=0, then E is measurable;

  3. 3.

    Countable unions of measurable sets are measurable;

  4. 4.

    Closed sets are measurable;

  5. 5.

    If E is measurable, then so is Ec;

  6. 6.

    Countable intersections of measurable sets are measurable.

This says that measurable sets form a σ-algebra. In addition, we should note that if F is closed and K is compact such that F∩K=∅, then d⁡(F,K)>0.2

Theorem 1.15.

If Ej are measurable and disjoint, then m⁡(⋃Ej)=∑m⁡(Ej).

Proof.

Assume that the Ej are bounded. Then there exist Fj closed with Fj⊆Ej such that m⁡(Ej∖Fj)<ε2j for all j. Then for finite N, ⋃j=1NFj is compact, so m⁢(⋃Fj)=∑m⁡(Fj). Therefore m⁡(E)≥∑j=1Nm⁡(Fj)≥∑j=1Nm⁡(Ej)−ε uniformly in N. Therefore, m⁡(E)≥∑m⁡(Ej). The other inequality is from subadditivity.

Next, if some of the Ej are unbounded, set Qj to be the cube with sides [−k,k]. Then set S1=Q1 and Sj+1=Qj+1∖Qj. Finally, set Ej,k=Ej∩Sk. Then each Ej,k is bounded, so we have

Corollary 1.16.
  1. 1.

    If Ek↗E=⋃Ek with Ek measurable, then m⁡(E)=limm⁡(Ek);

  2. 2.

    If Ek↘F=⋂Ek and some Ek has finite measure, then limm⁡(Ek)=m⁡(F).

Proof.
  1. 1.

    Set Dk=Ek+1∖Ek. Then the Dk are disjoint and measurable, so E=⋃Dk. Therefore,

  2. 2.

    Set Bk=Ek∖Ek+1. Then Ek0=(⋂Ek)∪(⋃k≥k0Bk). Then m⁡(Ek0)<∞ by assumption, so

    Therefore m⁢(⋂Ek)=limm⁡(EN).

□

Theorem 1.17.

If E is measurable, then for all ε>0:

  1. 1.

    There exists an open G⊇E such that m⁡(G∖E)<ε;

  2. 2.

    There exists closed F⊆E such that m⁡(E∖F)<ε;

  3. 3.

    If m⁡(E)<∞, then there exists compact K⊆E such that m⁡(E∖K)<ε;

  4. 4.

    If m⁡(E)<∞, then there exists F=⋃j=1NQj such that m⁡(E⁢△⁢F)<ε.

Proof.

We only need to prove the last two, the first two are by definition.

  1. 3.

    There exist a closed F with m⁡(E∖F)<ε/2. Then for all n, set Kn=B¯n∩F. Then the E∖Kn are measurable, and converge down to E∖F, so there exists n such that m⁡(E∖Kn)<ε/2+m⁡(E∖F)<ε.

  2. 4.

    Note there exists G=⋃j=1∞Qj containing E such that m⁡(G)<m⁡(E)+ε/2. Then there exists N such that ∑j<N|Qj|<ε/2, so we can set F=⋃i=1NQj. Then

Corollary 1.18.

E is measurable if and only if

  1. 1.

    E differs from a Gδ set by a null set;

  2. 2.

    E differs from an Fσ set by a null set.

Proof.
  1. 1.

    For all n there exists Gn open such that m⁡(Gn∖E)<1/n and Gn⊇E. Therefore, set S=⋂Gn. Clearly E⊆S and m⁡(S∖E)=0 by Corollary 1.16.

  2. 2.

    The argument is the same.

□

Other Properties of Lebesgue Measure

  1. 1.

    m is translation invariant;

  2. 2.

    If E is measurable, then m⁡(λ⁢E)=λd⁢m⁢(E);

  3. 3.

    m is reflection invariant.

Definition 1.19.

A set Σ⊂𝒫⁡(X) is a σ-algebra if it is closed under complements and countable unions.

The Borel Sets are the smallest σ-algebra containing the open balls.

Lemma 1.20.

If E⊆ℝ is measurable and m⁡(E)>0, then the set of differences E⊖E={x−y∣x,y∈E} contains an open interval around 0.

Proof.

There exists an open set G⊇E such that m⁡(G)≤(1+ε)⁢m⁢(E). Then G is a disjoint union G=⋃k=1∞ of open intervals. Set Ek=Ik∩E. Then we know that ∑|Ik|≤(1+ε)⁢∑m∗⁢(Ek), so there exists k0 such that |Ik0|≤m∗⁢(Ek0). Then for small δ if Ek0 and Ek0+δ are disjoint, then m⁡(Ek0∪Ek0+δ)≥21+ε⁢|Ik0|. However, if δ<|Ik0|, then Ik0∪Ik0+δ is an interval. Thus m⁡(Ek0∪Ek0+δ)≤|Ik0|+δ. Finally, we see that δ≥1−ε1+ε⁢|Ik0|, so Ek0∩Ek0+δ≠∅ for small enough δ. □

Theorem 1.21 (Vitali).

There is a non-measurable set S⊂[0,1].

Proof.

Enumerate the rationals in [0,1] by rk. Now consider the set ℝ/ℚ and choose one representative of each equivalence class using the axiom of choice.3 Call this set P. Next, note that P⊖P contains no interval because it contains no rationals. If m∗⁢(P)=0, then [0,1]⊆⋃r∈ℚP+r, which would imply that m⁡([0,1])=0. Thus P must be non-measurable. □

Corollary 1.22.

Any set with positive outer measure contains a non-measurable set.

Proof.

Let P be as in Theorem 1.21. Then let Ar=A∩(P+r). At least one of these sets must be non-measurable. □

1.6 Measurable Functions

Our goal will be integration. We expect that if E is a measurable set, then ∫χE=m⁡(E). More generally, we want

Functions of this form are called simple. We also want to have some results where the integral of a limit is equal to the limit of the integral. Similar to continuity, we can try to define f to be measurable if f−1⁢(E) is measurable for all measurable E. However, this is too hard to check, we find a minimal condition.

Definition 1.23.

A function f:ℝd→ℝ is measurable if for all a, f−1⁢(−∞,a) is measurable.

Here are some properties of measurable functions:

  1. 1.

    f is measurable if and only if f−1⁢(G) is measurable for all open G. Alternatively, f−1⁢(F) is measurable for all closed F.

  2. 2.

    Continuous functions are measurable. In addition, if f is measurable and finite-valued and Φ is continuous, then Φ∘f is measurable.

  3. 3.

    If {fn} is a sequence of measurable functions, then supfn,inffn,lim inffn,lim supfn, and limfn (if it exists) are all measurable.

  4. 4.

    If f,g are measurable, then f+g,f⁢g are both measurable.

  5. 5.

    Define f=g almost everywhere if the set on which they differ has measure 0. If f=g almost everywhere, then f is measurable if and only if g is measurable.

We will now prove some approximation theorems for measurable functions. First, we attempt to approximate them by simple functions.

Theorem 1.24.

If f≥0 is measurable on ℝd, then there exists an increasing sequence φk of simple functions converging pointwise to f.

Proof.

Define the sequence of functions

Clearly FN⁢(x)↗f⁡(x) for all x and FN⁢(ℝ)⊆[0,N]. Now define the set

and the simple function

Clearly fN,M is increasing in M. We also see that 0≤FN⁢(x)−FN,M⁢(x)≤1M. Now set N=M=2k. Then we see that

for all x. □

Corollary 1.25.

Suppose f is measurable. Then there exists a sequence φk simple such that |φk⁢(x)|↗|f⁡(x)| and limφk⁢(x)=f⁡(x).

Proof.

Write f⁡(x)=f+⁢(x)−f−⁢(x), where f+⁢(x)=max⁡{0,f⁡(x)} and f−⁢(x)=max⁡{−f⁡(x),0}. The rest is left to the reader. □

Theorem 1.26.

If f is measurable, there exists a sequence ψk of step functions such that ψk→f almost everywhere.

Proof.

First, there exist φk simple converging to f everywhere. Write φk=∑j=1Nkaj,k⁢χEj,k. Then for all Ej,k, there exists a finite collection of cubes ℚℓ such that m⁢(E⁢△⁢⋃Qℓ)<ε. Therefore, there exist almost disjoint rectangles R~p such that ⋃Qℓ=⋃R~p. Shrink the R~p to obtain disjoint closed rectangles Rp such that m⁢(E⁢△⁢⋃Rp)<2⁢ε. Then there exist step functions ψk and measurable sets Fk such that

  1. 1.

    m⁡(Fk)<12k;

  2. 2.

    {φk≠ψk}⊆Fk.

Then set F=lim supFk, so F has measure 0. Then for all x∉F, we have

Theorem 1.27 (Egorov).

Let fk be a sequence of measurable functions on E with m⁡(E)<∞ and fk→f pointwise. Then for all ε there exists a closed set Aε⊆E with m⁡(E∖Aε)<ε and such that fk→f uniformly on Aε.

Proof.

For given n,k, set

If we fix n, then Ekn⊆Ek+1n and Ekn↗E. Therefore there exists kn such that m⁡(E∖Eknn)<12n. Now choose N such that 21−N<ε2 and set A~ε=⋂n≥NEknn. Then

Choose Aε⊆A~ε such that m⁡(A~ε∖Aε)<ε2. Then m⁡(E∖Aε)<ε. To prove uniform convergence on Aε, choose M such that M≥N and 1M≤δ. Then if x∈Aε, x∈EknM, so

Theorem 1.28 (Lusin).

Let f be measurable and finite-valued on E and m⁡(E)<∞. Then for all ε, there exists Fε⊆E closed such that m⁡(E∖Fε)<ε and f|Fε is continuous.

Proof.

First, there exists a sequence of step functions fn such that fn→f pointwise. Then remove thin boundaries of rectangles to find En such that m⁡(En)<12n and such that fn is continuous off of En. By Egorov, there exists Aε/3 closed such that fn→f uniformly on Aε/3 and m⁡(E∖Aε/3)<ε/3. Next, let F′=Aε/n∖⋃n≥NEn, where 21−N<ε3. Now if m≥N, fn is continuous on F′ and fn→f uniformly, so f is continuous on F′. Finally, approximate F′ by a closed set Fε⊆F′ such that m⁡(F′∖Fε)<ε3. □

Finally, we will construct a non-measurable function. First, we will define the Cantor function fc:[0,1]→[0,1] which is given by

Note that f is monotone and continuous on [0,1] and constant on Cc.

Then define f⁡(x)=fc⁢(x)+x, so f:[0,1]→[0,2] is a continuous bijection. On any excluded interval I, f sends I to I+fc⁢(I). Therefore, m⁢(f⁢(⋃I))=1, so m⁡(f⁡(C))=1. Then there exists a non-measurable set S⊆f⁡(C). Then χS is not measurable.

2 Integration

2.1 Defining the Integral

Our goal is to integrate measurable functions. If f=∑ak⁢χEk is simple, then we define

To check that this is well-defined, if we can write f=∑bℓ⁢χFℓ, then we can consider E1∩E2,E1∖E2,E2∖E1. This gives a canonical representation of a simple function (where all of the Ej are disjoint). It is easy to see that we have linearity, monotonicity, and the triangle inequality. We want to extend this to a larger class of functions.

Next, we will consider bounded measurable functions on sets of finite measure. If f is supported on a set of finite measure and f is bounded, we define

for some sequence φn of simple functions converging to f. We need to show this is independent of the sequence.

Lemma 2.1.

If φn→f pointwise on E with m⁡(E)<∞, and f is bounded, then

  1. 1.

    The limit lim∫φn exists;

  2. 2.

    If f=0 almost everywhere, then lim∫φn=0.

Proof.

For the first part, we use Egorov. There exists a closed set Aε⊆E such that φn→f uniformly on Aε and m⁡(E∖Aε)<ε. Let In=∫Eφn. Then

On Aε, we have uniform convergence, so

which clearly converges to 0. Then the second terms are ∫E∖Aε|φn−φm|⁢dx≤2⁢M⁢m⁢(E∖Aε), which also converges to 0, where M=sup|f|. Thus In is Cauchy, so it converges.

If f=0 almost everywhere, then

so limIn=0. □

By Lemma 2.1, integration is well-defined. Again, we have the desired properties. Before we extend the definition of the integral to all functions, we will prove a convergence theorem. This is the first exchange of limits result.

Theorem 2.2 (Bounded Convergence Theorem).

If fn is a sequence of functions, uniformly bounded by M and uniformly supported on E of finite measure, and it fn→f pointwise almost everywhere, then f is measurable and ∫f=lim∫fn. In fact, ∫|f−fn|→0 as n→∞.

Proof.

The proof of this is the same as proof of the consistency of the definition of the integral. □

Corollary 2.3.

If f≥0 is bounded and supported on E with m⁡(E)<∞, and if ∫f=0, then f=0 almost everywhere.

Proof.

Set fk⁢(x)=f⁡(x)⁢χ{f⁡(x)≥1/k}. Then fk is bounded, supported on E, and fk≤f. Then

so m⁢({f⁡(x)>0})=0. □

Theorem 2.4.

If f is Riemann-integrable on [a,b], then it is Lebesgue integrable, and the integrals coincide.

Proof.

Translate the condition about partitions into step functions φi≤f≤ψi, where φi↗f and ψi↘f. Clearly the ψi,φi are uniformly bounded. Therefore, by bounded convergence, φ~,ψ~ are both measurable and bounded, and ∫φ~=∫ψ~, so ψ~=φ~ almost everywhere. Then clearly φ~=f almost everywhere, so f is measurable, and by monotonicity, ∫Rf=∫Lf. □

Finally, we can define integration for all nonnegative functions. We define

For any measurable E, define ∫Ef=∫f⋅χE. Note that it is easy to check that the definition makes sense. If ∫f<∞, we say that f is integrable.

Properties of the Integral.

  1. 1.

    Linearity;

  2. 2.

    Monotonicity;

  3. 3.

    Additivity on disjoint sets;

  4. 4.

    If g is integrable and 0≤f≤g, then f is integrable;

  5. 5.

    If f is integrable, then f⁡(x)<∞ almost everywhere;

  6. 6.

    If ∫f=0, then f=0 almost everywhere.

Note that if fn→f almost everywhere, then ∫fn does not necessarily converge to ∫f. For example, if we take fn=n⁢χ[0,1/n], then ∫fn=1, but f=0.

2.2 Some Convergence Results

We will state some convergence results where we can swap limits and integrals. We will also give the general definition of the integral here.

Lemma 2.5 (Fatou).

If fn is a sequence of functions with fn≥0 and if f⁡(x)=lim inffn⁢(x), then

Proof.

Suppose g bounded satisfies m⁡(supp⁡g)<∞ and 0≤g<f. Then set gn=min⁡{g,fn}. Then ∫g is bounded, so by bounded convergence, ∫gn→∫g. Also, ∫gn≤∫fn, so ∫g≤liminffn. Because ∫f=sup∫g, we get the desired result. □

Corollary 2.6.

If fn≤f with fn⁢(x)→f⁡(x) almost everywhere, then lim∫fn=∫f.

Proof.

By monotonicity, lim sup∫fn≤∫f≤lim inf∫fn. □

Corollary 2.7 (Monotone Convergence).

If fn⁢(x)↗f⁡(x) almost everywhere, then lim∫fn=∫f.

Corollary 2.8.

Given ak⁢(x)≥0 measurable, then ∫∑ak⁢(x)≤∑∫ak⁢(x). Moreover, if ∑∫ak⁢(x) converges, ∑ak⁢(x) converges almost everywhere.

Proof.

Let fn be the partial sums. Then by monotone convergence, if ∑∫ak converges, ∫∑ak also converges, so it is integrable. Thus ∑ak converges almost everywhere. □

Example 2.9 (Borel-Cantelli).

Given Ek measurable with ∑m⁡(Ek)<∞, then m⁡(lim¯⁢Ek)=0. To see this, set ak=χEk and use the previous result.

Example 2.10.

Let

Then ∫|x|=εf⁡(x)⁢dx≤cε for all ε>0.

To see this, we will bound ∫|x|≥εf. Consider the annulus Ak={x∣2k⁢ε≤|x|<2k+1⁢ε}. Then set

Then set g⁡(x)=∑ak⁢(x). Clearly f≤g, so ∫f is finite if ∫g is. However, we see that

We now give the general definition of the integral. Let f be a measurable function and write f⁡(x)=f1⁢(x)−f2⁢(x), where both f1,f2 are nonnegative. Then we want ∫f=∫f1−∫f2. Clearly this is defined if at most one of the two integrals on the RHS is infinite.

Definition 2.11.

f is integrable if both ∫f1,∫f2 are finite.

Theorem 2.12.

The integrable is linear, additive, monotone, and satisfies the triangle inequality.

Lemma 2.13.

Let f be integrable. Then

  1. 1.

    Given ε>0, there exists a ball B of finite measure such that ∫Bc|f|<ε;

  2. 2.

    For all ε>0, there exists δ such that whenever m⁡(E)<δ, then ∫E|f|<ε.

Proof.
  1. 1.

    We assume f≥0. The for each N consider BN={|x|≤N} and fn=χBn. Then we see that fn↗f, so by monotone convergence, ∫fn→∫f<∞. Therefore, there exists N such that

  2. 2.

    Again assume f≥0. Then set En={f⁡(x)≤n} and fn=f⋅χEn. Again, fn↗f and f is integrable, so there exists N such that f−fn<ε2 for all n≥N. Then choose δ such that N⁢δ<ε2. Now if m⁡(E)<δ, we see that

Theorem 2.14 (Dominated Convergence).

If fn→f almost everywhere and |fn|≤|g| almost everywhere for some integrable g, then f is integrable and ∫|fn−f|→0 as n→∞.

Proof.

For N≥0, set EN={x∣|x|⁢g⁢(x)≤N}. Then define gn=g⋅χEn. As in the lemma above, there exists N such that ∫ENc⁢g<ε.

Note that fk⁢χEn≤n for all k and is supported on En, which has finite measure. Also, fk⁢χEn→f⁢χEn. By bounded convergence,

for k large enough, so

We now briefly consider functions f:ℝd→ℂ. Then note that f=u+i⁢v, so |f|=u2+v2. Also, because a+b≤a+b, |f|≤|u|+|v|. We conclude that f is integrable if and only if u,v are both integrable.

2.3 Vector Space of Integrable Functions

Let L1 denote the set of integrable functions. Define the L1 norm ‖f‖1=∫|f|. Note that this is only a seminorm, so we set f∼g if f=g almost everywhere.

Theorem 2.15.

L1 is a normed vector space.

Theorem 2.16 (Riesz-Fischer).

L1 is a Banach space.1

Proof.

Suppose {fn} is Cauchy in L1. Then for all k choose nk+1 such that ‖fnk+1−fnk‖<12k. Then we define

Note that g is nonnegative, so by monotone convergence,

so by dominated convergence, f is integrable. In addition, ‖f−fnk+1‖→0, so we have a convergent subsequence. But then because the sequence is Cauchy and by the triangle inequality, it converges to f. □

Corollary 2.17.

If fn→f in L1, there exists a subsequence fnk such that fnk→f almost everywhere.

Definition 2.18.

We say fn converges to f in measure if m⁢{x∣|fn⁢(x)−f⁡(x)|>ε}→0.

Example 2.19.

Write natural numbers n as n=k+2v, where 0≤k<2v. Set fn=χ[k2v,k+12v]. Then this converges to 0 in measure but not pointwise anywhere.

Theorem 2.20.

If fn→f in measure, there exists a subsequence fnk→f almost everywhere.

Proof.

Given v, there exists nv such that m⁢{x∣|fn⁢(x)−f⁡(x)|≥2−v}≤2−v for n≥nv. Then choose

Now if x∉Ev, then fnv⁢(x)→f⁡(x) for x∉⋂k=1∞⋃v=k∞Ev. However, ∑m⁡(Ev)<∞, so this is a set of measure 0. □

Remark 2.21.

In Fatou’s lemma, monotone convergence, and dominated convergence, we can replace pointwise convergence by convergence in measure.

Recall that in a topological space X, a set A is dense if A∩𝒪≠∅ for all open 𝒪. In a normed space, this is equivalent to saying that for all x∈X and ε>0, there exists a∈A such that ‖a−x‖<ε.

Theorem 2.22.

The following classes of functions are dense in L1:

  1. 1.

    Simple functions;

  2. 2.

    Step functions;

  3. 3.

    Continuous functions of compact support.

Proof.

Take f≥0.

  1. 1.

    We know that there exists a sequence of increasing simple functions φn↗f pointwise, but f is integrable, so ‖f−fn‖=∫f−fn→0.

  2. 2.

    We approximate simple functions by step functions. Let φ=∑ai⁢χEi be simple. Then there exist disjoint closed rectangles Rj such that m⁢(E⁢△⁢⋃Rj)<ε. This bounds ∫χE−χRj.

  3. 3.

    It is easy to approximate χR by continuous functions. Just build side ramps.

□

2.4 Symmetries

First, note that the integral is translation invariant. To see this, we can translate step functions. In addition, it is easy to see that nd⁢∫f⁡(n⁢x)⁢dx=∫f⁡(x)⁢dx.

Examples 2.23.

First, note that

Similarly,

Example 2.24 (Convolution).

Given integrable f,g, define their convolution f∗g by

It is easy to see that f∗g=g∗f.

Now write d=d1+d2, so ℝd=ℝd1×ℝd2. Then for fixed y write fy=f⁡(−,y). Similarly, for E⊆ℝd, write Ey={x∣(x,y)∈E}⊆ℝd1. Define fx and Ex similarly. The main issue is that if E is measurable, then Ey,Ex might not be.

Theorem 2.25 (Fubini).

Suppose f⁡(x,y) is integrable on ℝd=ℝd1×ℝd2. Then

  1. 1.

    For almost every y∈ℝd2, fy is integrable on ℝd1;

  2. 2.

    The function ∫ℝd1fy⁢(x) is integrable on ℝd2;

  3. 3.

    ∫ℝd2∫ℝd1fy⁢(x)⁢dx⁢dy=∫ℝdf⁡(x,y).

Proof.

Let ℱ be the set of integrable functions where all three conditions hold. We show that ℱ=L1. First, we show that ℱ≠0.

Given any rectangle R⊆ℝd, it is easy to see that χR∈ℱ. If we write R=R1×R2, then χRy⁢(x)=χR1⁢(x)⋅χR2⁢(y). Then the first condition holds everywhere. For the second condition, note that

For the third condition, note that

Therefore, all step functions are in ℱ.

Now we will show that

  1. 1.

    ℱ is closed under linear combinations;

    To see this, let fk∈𝒻 and ak∈ℝ. Then use linearity of the integral.

  2. 2.

    ℱ is closed under limits.

    We do this for monotone limits. Note that for all f∈ℱ, there exists A of measure 0 such that fy is not integrable for y∈A. If fk∈ℱ and fk↗f and f is integrable, we want to show that f∈ℱ. Let Ak correspond to fk. Then their union has measure 0, so fy is integrable almost everywhere.

    Now set gk⁢(y)=∫fky↗∫fy=g. Now each gx is integrable and gk↗g, so ∫gk=∫fk↗∫f, so ∫g<∞ and thus g<∞ almost everywhere. Thus fy is integrable almost everywhere, so f∈ℱ.

    Recall that if f is integrable, there exists a set φk of simple functions converging up to f. Each simple is a linear combination of χE for some measurable E of finite measure, so we prove χE∈ℱ.

This reduces the problem to measurable sets:

  1. 3.

    Gδ sets;

    If E is contained in the boundary of a rectangle, the result is obvious. Next, suppose E is a finite union of almost disjoint rectangles Rj. Then write χE as a linear combination of the χR and the χP, where P⊆∂R. Thus χE∈ℱ.

    Now if E is open and of finite measure, write E=⋃Qj, where the Qj are almost disjoint cubes. Then write χE=∑χQj, which is a monotone limit. Finally, if E=⋂Gk for Gk open, then χE is also a monotone limit.

  2. 4.

    Null sets;

    Recall that there exists a Gδ set B of measure 0 such that E⊆B. Then χB∈ℱ, so ∫∫χBy=0, so ∫χBy=0 almost everywhere. Because Ey⊆By for all y, Ey is measurable and null almost everywhere, so χE∈ℱ.

  3. 5.

    Limiting procedure;

    For any measurable E, E differs from a Gδ set by a null set.

Finally, for general f, write f=f+−f−, which are both integrable and nonnegative. □

Theorem 2.26 (Tonelli).

Fubini extends to measurable functions for which ∫f is defined but infinite.

Corollary 2.27 (Cavaleiri’s Principle).

Let E be a measurable subset of ℝd1×ℝd2. Then for almost every y, the slice Ey is measurable, and m⁡(E)=∫m⁡(Ey).

3 Differentiation

3.1 Integration and Differentiation

Recall the fundamental theorem of calculus:

  1. 1.

    For f continuous, the function F⁡(x)=∫axf⁡(x)⁢dx is differentiable and F′=f.

  2. 2.

    If F′ is Riemann-integrable, then F⁡(b)=F⁡(a)=∫abF′⁢(x)⁢dx.

This roughly tells us that integration is anti-differentiation. We can replace differential equations with formulations in terms of integrals. Then recall that

What happens when we generalize to higher dimensions and consider 1m⁡(B)⁢∫Bf? Define f∗⁢(x)=sup1m⁡(B)⁢∫|f|. Note that f∗ need not be integrable.

Theorem 3.1.
  1. 1.

    f∗ is measurable;

  2. 2.

    f∗<∞ almost everywhere;

  3. 3.

    There exists A depending only on d such that for all α,

Proof.
  1. 1.

    Define Eα={x∈ℝd∣f∗⁢(x)>α}. Then note that x∈Eα if and only if

    for some ball B∋x. In fact, for all y∈B, y∈Eα, so Eα is open.

  2. 2.

    This is a corollary of part 3.

  3. 3.

    The proof of this relies on the Vitali Covering Lemma, which is stated below.

    Let Eα={x∣f∗⁢(x)>α}. Then we know x∈Eα if and only if there exists B∋x such that m⁡(B)<1α⁢∫B|f|. Given a compact K⊆Eα, write K⊆⋃x∈KBx and extract a finite subcover. Then we use Vitali, so

    Because K is arbitrary, we have the desired result.

□

Lemma 3.2 (Vitali’s Covering Lemma).

Let B={B1,…,BN} be a finite set of open balls. Then there exists a set of disjoint open balls {Bi1,…,Bik}⊂B such that ⋃Bj⊆⋃3⋅Bik.

Proof.

Inductively choose Bik to be the ball of largest radius not disjoint to thre previous balls and not contained in ⋃3⋅Bij. We started with a finite collection, so the process terminates. □

Theorem 3.3 (Lebesgue Differentiation Theorem).

If f is integrable, then for almost every x,

In ℝ, this implies ddx⁡F⁡(x)=f⁡(x).

Proof.

For α>0, define the set

We will show that m⁡(Eα)=0.

Let ε>0. Recall that continuous functions of compact support are dense in L1, so choose a continuous function with compact support g such that ‖f−g‖<ε. Then g is uniformly continuous, so 1m⁡(B)⁢∫g⁡(y)⁢dy→g⁡(x) as m⁡(B)→0. Now use the triangle inequality:

Therefore, we have

Let Fα be the set where (f−g)∗⁢(x)>α and Gα where |f−g|>α. By Theorem 3.1, m⁡(Fα)≤Aα⁢‖f−g‖, and by Chebyshev, m⁡(Gα)≤1α⁢‖f−g‖. □

Corollary 3.4.

For almost every x, f∗⁢(x)≥|f⁡(x)|.

Also note that we can replace integrability by local integrability in these results. If E is measurable, define x to be a point of density of E if

Corollary 3.5.

Let E be measurable. Then

  1. 1.

    Almost every x∈E is a point of density;

  2. 2.

    Almost every x∉E is not a point of density.

In a related notion, we can define the Lebesgue set ℒ⁡(f) of f as the set of all x such that f⁡(x)<∞ and

Clearly if x∈ℒ⁡(f), then the Lebesgue differentation theorem holds. Also, if f is continuous at x, then x∈ℒ⁡(f).

Corollary 3.6.

For any f∈Lloc1, almost every x belongs to ℒ⁡(f).

Proof.

For rationals r∈ℚ, consider hr=|f⁡(x)−r|. Then there exists Er with m⁡(Er)=0 such that the Lebesgue differentiation theorem holds for hr for all x∉Er. Thus if x∉E=⋃Er, then there exists r such that |f⁡(x)−r|<ε for all ε. For this ε,r, choose δ such that if m⁡(B)<δ, then

so 1m⁡(B)⁢∫|f⁡(y)−r|<|f⁡(x)−r|+ε. Therefore

Note that any locally integrable function has different representatives up to sets of measure 0. We know that for almost every x, f⁡(x)=lim1m⁡(B)⁢∫Bf and if f=f¯ almost everywhere, ∫Bf=∫Bf¯, but ℒ⁡(f)≠ℒ⁡(f¯) up to a set of measure 0.

On the Lebesgue set, we can generalize average. Instead of balls, we can consider general sets U. Note that 1m⁡(U)⁢∫Uf=∫1m⁡(U)⁢χU⁢f, which is morally equal to ∫K⁢f, where the kernel K satisfies ∫K=1. We will actually consider families of such kernels Kδ.

3.2 Kernels

Consider sets U for which 1m⁡(U)⁢∫Uf converges.

Definition 3.7.

A family of sets Uα shrinks regularly to x if there exists c>0 such that for all B∋x, there exists α such that x∈Uα⊆B, and m⁡(Uα)≥c⋅m⁡(B).

Corollary 3.8.

If f is locally integrable and Uα shrinks regularly to x∈ℒ⁡(f), then

Proof.

Note that

Note that E is measurable if and only if for all A, m∗⁢(A)=m∗⁢(A∩E)+m∗⁢(A∩E)c. Also, note that 1m⁡(B)⁢∫Bf=1m⁡(B)⁢∫RdχB⋅f=∫f⋅1m⁡(B)⁢χB. Instead of multiplication, we will convolce. We will show later that ∗ commutes with differentiation.

Note that f∗χ is a restriction of f. Far from x, f∗χ=0 and f∗χ=f near x. However, χE is not usually differentiable. Instead, we will convolve with different kernels.

Definition 3.9.

A good kernel Kδ satisfies

  1. 1.

    ∫Kδ=1 for all δ>0.

  2. 2.

    ∫|Kδ|≤A uniformly. Also, for all ξ>0,

    as δ→0.

Example 3.10.

We will construct a smooth bump function. Let

Now set f2⁢(x)=f1⁢(x)⋅f1⁢(1−x). Then set φ⁡(x)=f2⁢(x+1/2)∫f2. This is a smooth even function with integral 1 supported on [−1/2,1/2]. Now we will introduce scaling, to define a family ϕε:ℝd→ℝ defined by

where K is chosen such that ∫φε⁢(x)=1. If we define fε=f∗ϕε, then fε is a smooth approximation to f that is C∞. In addition, the derivative of fε is D⁢ϕε∗f.

The bottom line is that for all ε>0, ϕε∗f is C∞. Also, (ϕε∗f)⁢(x)→f⁡(x) as ε→0 for all Lebesgue points of f.

We call a kernel Kδ an approximation to the identity if it satisfies

  1. 1.

    ∫Kδ=1;

  2. 2.

    |Kδ⁢(x)|≤A⁢δ−d for all δ>0;

  3. 3.

    |Kδ⁢(x)|≤A⁢δ|x|d+1 for all δ>0,x∈ℝd,x≠0.

Therefore, Kδ→∞ as δ→0 for x=0. If x≠0, Kδ⁢(x)→0 as δ→0. Note that the Dirac mass1 is a “generalized function” which satisfies ∫δ=1, δ⁡(x)=0 for x≠0, and ∫δ⋅f⁡(x)=f⁡(0).

Example 3.11 (Poisson Kernel).

Consider the upper half plane (or equivalently the disk). Then define

for x∈ℝ,y>0. Then for x≠0, Py→0 as y→0. On the disk, define

Note that Δ⁢P=0 away from the origin. If we convolve u=P∗g, u solves Δ⁢u=0 on y>0 and u⁡(x,0)=g⁡(x).

Example 3.12 (Heat Kernel).

Define the heat kernel

Again, if x≠0, Ht⁢(x)→0 as t→0. If x=0, then Ht⁢(0)→∞. Also, for t>0, ∫ℝdHt⁢(x)⁢dx=1. Note that for t>0, the heat equation ut=Δ⁢u with initial conditions u⁡(0,x)=g⁡(x) is solved by u⁡(t,x)=Ht∗g.

Because the heat equation is linear, since δ=limHt is a convolution identity, ∗ commutes with ∂x, so given f, set u=Ht∗f. Because differentiation commutes with ∗, we see that

but at t=0, then δ∗f=f.

Example 3.13 (Fejer Kernel).

This kernal is defined by

It is related to the Fourier transform. Recall that the Fourier transform f~ is defined by

and its inverse

Note that f⁡(x)⁢e2⁢π⁢i⁢x⁢ξ as a function of ξ is a wave.

Theorem 3.14.

If f is measurable and x∈ℒ⁡(f), then (f∗Kδ)⁢(x)→f⁡(x) as δ→0.

Proof.

We calculate

We will show that ∫|f⁡(x−y)−f⁡(x)|⁢|Kδ⁢(y)|⁢dy→0. We will use properties to Kδ and split the integral into parts |y|≤δ and |y|>δ. Define for x∈ℒ⁡(f),

where Wd=m⁡(B0⁢(1)). We then need Lemma 3.15, which is stated after this proof. This yields boundedness on (0,R]. Then for large e, we use the triangle inequality:

Turning back to our original problem, note that

The first term is given by

On the annulus, we will estimate |Kδ|≤C⁢δ|x|d+1. Then we can write

so

For large k, A⁡(2k+1⁢δ)≤M, so for any N, we see that

Therefore, given ε, choose N such that c⁢Wd⁢2d⁢M⁢2−N<ε2. Then for all sufficiently small δ,

Lemma 3.15.

If f∈L1 and x∈ℒ⁡(f), then A⁡(r) is continuous on (0,∞) and A⁡(r)→0 as r→0.

Proof.

Because the integral is absolutely continuous, A⁡(r) is continuous: absolute continuity of the integral means that for all ε>0, there exists δ>0 such that if m⁡(U)<δ, then ∫U|f|<ε.

Now consider small annuli. For r>0 fixed, the annuli have measure Wd⁢((r+δ)d−(r−δ)d), which is small. Then A⁡(r)→0 is just the statement that x∈ℒ⁡(f). □

3.3 Differentiation

Let f:ℝ→ℝ be monotone increasing. Given only two points xL,xR, for any increasing sequence xL≤⋯<x−1<x0<x1<…<xR, we have

In particular, this series converges absolutely, so there are only countably many discontinuities, and moreover ∑jf⁡(xj,+)−f⁡(xj,−)<∞, where f⁡(xj,+)=limx>xjf⁡(x) and f⁡(xj,−) is defined similarly.

We know that f⁡(x,−)≤f⁡(x)≤f⁡(x,+) for all x and that the three values are equal if and only if f is continuous at x. Then ∑f⁡(xk,+)−f⁡(xk,−)<∞ means that there are countably many jumps and that their sum is finite. Now assume that f is continuous and increasing on [a,b].

Note that the Cantor function is continuous, monotone, and takes f:[0,1]→[0,1]. Recall that fc is constant on each excluded interbal, so fc′=0 almost everywhere. Therefore, ∫01fc′⁢(x)⁢dx=0, but fc⁢(1)−fc⁢(0)=1.

Theorem 3.16.

If f is continuous and monotone on [a,b], then the derivative exists almost everywhere and

Before we prove the theorem, we need to define the Dini derivatives. Define

Then the Dini numbers are defined everywhere and that D−≤D−,D+≤D+. Then f is differentiable at x if and only if D+≤D− and D−≤D+ and all these values are finite.

Lemma 3.17 (Rising Sun Lemma).

Let G be continuous on [a,b]. Then the set {x∣∃h>0⁢(G⁡(x+h)>G⁡(x))} is a relatively open union of disjoint intervals ⋃i(ai,bi), and for all x∈(ai,bi), G⁡(ai)=G⁡(bi)>G⁡(x).

Proof.

It is clear that the set is open. Now if G⁡(bi)>G⁡(ai), then we can extend the interval to the left, and similarly to the right if G⁡(ai)>G⁡(bi). □

Lemma 3.18 (Vitali’s Covering Lemma).

A collection 𝒢 of open/closed interbals is a Vitali cover of E if for all ε>0 and x∈E, there exists G∈𝒢 such that x∈G and |G|<ε. Let E be of finite measure and G a Vitali cover. Then for all ε>0, there exists a finite set of disjoint Gn∈𝒢 such that m⁡(E)≤∑|Gj|+ε.

Proof.

Assume without loss of generality that each G∈𝒢 is closed. Fine 𝒪 of finite measure with E⊆𝒪. Now choose a series of intervals as follows. Suppose that G1,…,Gn have already been chosen. Then let

Either E⊆⋃Gi or there exists Gn+1∈𝒢 contained in 𝒪 and disjoint from the other sets such that |Gn+1|>12⁢kn>0. Continue the process. If this does not stop, then choose N such that ∑N+1∞|Gj|<ε5. Let R=E∖⋃j=1NGj.

To show that m⁡(R)<ε, choose x∈R. Because the Gj are closed, there exists G∈𝒢 such that x∈G and |G|<12⁢d⁢(x,⋃1NGj). Thus G does not intersect any of the Gj. From our construction, there exists n>N such that |G|≤kn≤2⁢|Gn+1|. however, |Gn+1| tends to 0, so there exists a smallest n>N such that G∩Gn is nonempty. For this n, if mn is the midpoint of Gn, we see that |x−mn|≤|G|+12⁢|Gn|. However, |G|≤2⁢|Gn|, so |x−mn|≤52⁢|Gn|. Therefore x∈Jn for some Jn with |Jn|=5⁢|Gn|, so

Proof of Theorem 3.16.

We want to show that D+⁢F≤D−⁢F almost everywhere and D+⁢F<∞ amost everywhere. Fix γ and set Eγ={x∣D+⁢F⁢(x)>γ}. We will show that Eγ is measurable. Note that Δh⁢F is measurable and f is continuous, so replace h with 1/n. Set G⁡(x)=F⁡(x)−γ⁢x.

The idea is that on Eγ, D+⁢F⁢(x)>γ. Then there exists h0 such that F⁡(x+h)−F⁡(x)h>γ for 0<h<h0. Thus F⁡(x+h)−F⁡(x)>γ⁢h, so G⁡(x+h)>G⁡(x). G is not monotone but is continuous. We will now eliminate {D+⁢F=∞} and {D+⁢F>D−⁢F}. Note that D+⁢F⁢(x)>γ if and only if Gγ⁢(x+h)>Gγ⁢(x) for some h. By rising sun, we know that Eγ={D+⁢F>γ} is open and of finite measure, so m⁢({D+⁢F=∞})=0.

Now choose R>r rational and set E={x∣D+⁢F>R>r>D−⁢F}. We will show that m⁡(E)=0. If not, then there exists an open 𝒪⊇E with m⁡(𝒪)<m⁡(E)⁢Rr. On some connected component In of 𝒪, apply rising sun to G⁡(x)=−F⁡(−x)+r⁢x on −In. Then we obtain intervals (ck,dk), so wet (ak,bk)=(−dk,−ck). Then we see that −F⁡(bk)−r⁢bk=−F⁡(ak)−r⁢ak>−F⁡(x)−r⁢x for x∈(ak,bk).

To show that m⁡(E)=0, find an open 𝒪⊇E such that m⁡(𝒪)<m⁡(E)+ε. Then for all x∈E, there exists an interval [x−h,x]⊆𝒪 suchthat F⁡(x)−F⁡(x−h)<r⁢h. This is a Vitali covering.

Now choose I1,…,IN such that m⁡(E)≤∑|In|+ε. Summing over this finite subcover, we see that

Then if we set A to be the interior of ⋃Ij, then for all y∈A there exists δ such that (y,y+δ)⊆Ij and F⁡(y+δ)−F⁡(y)>R⁢δ. By Vitali, there exists J1,…,Jm of the form Ji=[yi,yi+δi] such that m⁡(⋃Ji)≥m⁡(A)−ε≥m⁡(E)−2⁢ε. For each Ji, we have F⁡(yi+δi)−F⁡(yi)>R⁢δi, so adding, we get that

Each Ji is contained in an Iℓ, so we can sum over i to see that

Because ε is arbitrary and R⁢m⁢(E)≤r⁢m⁢(E), we see that m⁡(E)=0. □

We now claim that F′ is measurable. In addition, F′ is integrable and

To see this, note that F′ is an almost everywhere limit of measurable functions. By Fatou, we see that

Because F is continuous,

for n large enough, so n⁢∫bb+1/nF⁡(y)⁢dy→F⁡(b). The same argument holds for the integral near a.

3.4 Bounded Variation

Given [a,b], create a finite partition a=x0<x1<⋯<xn=b. Then set

Then define the increasing, decreasing, and total variations of f by

We say f is of Bounded Variation if T⁢V⁢(f) is finite.

Lemma 3.19.

If f is BV, then T⁢V⁢(f)=I⁢V⁢(f)+D⁢V⁢(f) and f⁡(b)−f⁡(a)=I⁢V⁢(f)−D⁢V⁢(f).

Proof.

Choose a finite partition. It is easy to see that f⁡(b)−f⁡(a)=p−n, so for all partitions, n+f⁡(b)−f⁡(a)=p. Then n+f⁡(b)−f⁡(a)≤I⁢V⁢(f), but this means D⁢V⁢(f)+f⁡(b)−f⁡(a)≤I⁢V⁢(f). Similarly, if we take the supremum over n first, we get the inequality reversed. The first part of the lemma should be easy. □

Theorem 3.20.

A function f is BV if and only if it is the difference of two bounded monotone increasing functions.

Proof.

If f is BV, set g⁡(x)=I⁢V[a,x]⁢(f) and h⁡(x)=D⁢V[a,x]⁢(f). Because f is BV, g,h are monotone increasing and bounded, and by Lemma 3.19, f⁡(x)−f⁡(a)=I⁢V[a,x]⁢(f)−D⁢V[a,x]⁢(f)=g⁡(x)−h⁡(x). On the other hand, if f=g−h with g,h monotone and bounded, then T⁢V⁢(f)≤T⁢V⁢(g)+T⁢V⁢(h) is finite. □

Corollary 3.21.

If f is BV, then f′ exists almost everywhere.

Now we will consider jumps. Assume f is bounded and monotone increasing. At each point, the limits f⁡(x+)=limx→x+f⁡(x) and f⁡(x−) (defined analogously) exist. By monotonicity, f⁡(x+)≥f⁡(x)≥f⁡(x−). Define the jump set to be the set J={x∣f⁡(x−)<f⁡(x+)}. We know that J is countable, so enumerate it by J={x1,x2,…}. For each n, define [f]n=f⁡(xn+)−f⁡(xn−). At each n, we have f⁡(xn)=f⁡(xn−)+θ⁢[f]n for some 0≤θ≤1. Now define

and define the jump function of f by Jf⁢(x)=∑n[f]n⁢jn⁢(x).

Lemma 3.22.
  1. 1.

    Jf is monotone increasing;

  2. 2.

    f−Jf is monotone increasing and continuous;

  3. 3.

    Jf is differentiable almost everywhere and Jf′⁢(x)=0 almost everywhere.

Proof.
  1. 1.

    This is obvious.

  2. 2.

    The series defining Jf converges uniformly and JF⁢(x)≤F⁡(b)−F⁡(a), so if x∉J, each jn⁢(−) is continuous at x, so Jf is continuous at x. If x is a jump, then Jf⁢(x)=∑xk<x[f]k⁢jk⁢(x)+∑xk>x[f]k⁢jk⁢(x)+[f]n⁢jn⁢(x). In particular, Jf⁢(x−)=∑xk<x[f]k,Jf⁢(x+)=∑xk>x[f]k, and Jf⁢(x)=Jf⁢(x−)+θn⁢[f]n.

    Now f⁡(x+)−Jf⁢(x+)−(f⁡(x−)−Jf⁢(x−))=0 at x=xn. Also, f⁡(x)−Jf⁢(x)=0 provided that f−Jf is increasing, because it is continuous. If x<y, note that

    so we see that f−Jf is increasing.

  3. 3.

    Note that Jf is discontinuous on a countable set. Given any ε>0, there exists a closed F such that J⊆F∘ and m⁡(F)<ε. Then Fc is open and Jf is constant on each component of Fc. Therefore, the set where Jf′ does not exist or is nonzero has measure less than ε for all ε. □

Morally, we say that Jf′⁢(x)=∑[f]n⁢Hn′, where H is the Heaviside step function, and H′=δ.

We now consider the question of which continuous monotone functions satisfy the Fundamental Theorem of Calculus. We consider properties of ∫axf⁡(t)⁢dt, where f is integrable on [a,b]. We will refer to the integral as a function F⁡(x).

Lemma 3.23.

F is continuous and of bounded variation.

Proof.

Continuity is easy. Use the fact that for all ε, there exists δ such that if m⁡(E)<δ, then ∫E|f|<ε. To show that F is BV, choose a=x0<x1<⋯<xn=b. Then we see that

which is finite. □

Lemma 3.24.

If f is integrable and ∫axf⁡(t)⁢dt=0 for all x, then f⁡(x)=0 almost everywhere.

Proof.

Set E=E+∪E−, where E+−{f>0} and E−={f<0}. If m⁡(E+)>0, then there exists closed F⊆E+ with m⁡(F)>0. Then G=Fc is relatively open. Either we have ∫abf⁡(t)⁢dt≠0 or 0=∫abf⁡(t)⁢dt=∫Gf⁡(t)⁢dt+∫Ff⁡(t)⁢dt. Then ∫Gf≠0, so write G=⋃(ai,bi). Therefore, at least one of the ∫aibif⁡(t)⁢dt is nonzero, so either ∫aaif⁡(t)⁢dt or ∫abif⁡(t)⁢dt is nonzero. □

Theorem 3.25.

If f is integrable and if F⁡(x)=F⁡(a)+∫axf⁡(t)⁢dt, then F′⁢(x)=f⁡(x) almost everywhere.

Proof.

If f is bounded by K, then we know F is BV, so F′ exists almost everywhere. Let fn⁢(x)=Δ1/n⁢F⁢(x), so

Also, fn→F′ almost everywhere. By bounded convergence, we then have

so F′⁢(x)=f⁡(x) almost everywhere.

For the general case, assume that f>0. Then set fn⁢(x)=min⁡{f⁡(x),n}. Then f−fn≥0, so Gn⁢(x)=∫axf⁡(x)−fn⁢(x)⁢dx is monotone increasing. Thus it has nonnegative derivative almost everywhere. By the above, we know that

almost everywhere. Therefore

almost everywhere. Thus ∫abF′⁢(x)⁢dx≥∫abfn⁢(x)⁢dx, so ∫abf⁡(x)⁢dx≤lim inf∫abfn⁢(x)⁢dx≤∫abF′⁢(x)⁢dx, so then

3.5 Absolute Continuity

Definition 3.26.

A function f:ℝ→ℝ is absolutely continuous if for all ε>0, there exists δ>0 such that if xi,xi′ for a finite set of non overlapping intervals and if ∑i|xi′−xi|<δ, then

Note that absolute continuity implies uniform continuity. From the previous lemma, then integrability of f means that ∫axf is absolutely continuous.

Lemma 3.27.

If f is absolutely continuous, then it is also BV.

Proof.

Let f be absolutely continuous and choose ε=1. Then given any partition x0<x1<⋯<xn, choose K>1+b−aδ. Add in many partition elements so that [a,b] is covered by K sets of subinterbals each of length less than δ. Now

Therefore T⁢V⁢(f)≤K. □

As a consequence, if f is absolutely continuous, then its derivative exists almost everywhere.

Lemma 3.28.

If f is absolutely continuous on [a,b] and if f′=0 almost everywhere, then f is constant.

Proof.

Fix t∈(a,b) and set E=(a,t)∩{f′=0}, so m⁡(E)=t−a. Let two parameters be given: ξ an upper bound for f′ on E and ε a bound for m⁡((a,t)∖E). Use absolute continuity on m⁡((a,t)∖E): Choose δ according to the definition of f depending on ε.

If x∈E, then f′⁢(x)=0. Thus there exists h such that [x,x+h]⊆[a,t] and |f⁡(x+h)−f⁡(x)|<ξ⁢h, so E⊆⋃x∈E[x,x+h]. Now use Vitali’s covering lemma, so there exists a finite non-overlapping set of interals and B such that m⁡(B)<δ and

Order the xk,yk and set y0=a,xn+1=t. Now we have

Theorem 3.29 (Fundamental Theorem of Calculus, Part 2).

A function F has integrable f such that

if and only if F is absolutely continuous.

Proof.

Suppose that F is absolutely continuous. Then F is BV, so F⁡(x)=F1⁢(x)−F2⁢(x) for monotone increasing F1,F2. Therefore F′=F1′+F2′ exists almost everywhere and F1′⋅F2′=0. Therefore

Set G⁡(x)=∫axF′⁢(t)⁢dt and h⁡(x)=F⁡(x)−G⁡(x). Then h is also absolutely continuous and G′⁢(x)=F′⁢(x) almost everywhere from the definition of G, so h′=0 almost everywhere. Thus h is constant and must be equal to F⁡(a). □

We will now characterize functions of bounded variation.

  1. 1.

    If F is BV, then F=I⁢V⁢(F)−D⁢V⁢(F).

  2. 2.

    F has a jump part JF=∑k[F]k⁢jk⁢(x), where [F]k=F⁡(xk+)−F⁡(xk−).

  3. 3.

    Fc=F−JF is continuous and the difference of two monotone continuous functions. Thus Fc′ exists almost everywhere. If we set

    then Fa⁢c is absolutely continuous. Then the function Fs⁢c=Fc−Fa⁢c is monotone increasing, but Fs⁢c′=0 almost everywhere, so it does not satisfy FTC (behaves like the Cantor function).

In conclusion, we can write F=JF+Fa⁢c+Fs⁢c.

4 General Measures

Let X be a set and Σ be a σ-algebra of X. Then M is a measure on (X,Σ) if M is a nonnegatie σ-additive function on Σ.

Examples 4.1.
  1. 1.

    The Lebesgue measure as defined previously;

  2. 2.

    The counting measure;

  3. 3.

    The discrete measure on Σ=𝒫⁡(X) where each element is assigned a weight;

  4. 4.

    Let X be an uncountable set, set Σ to be the countable or cocountable sets, and then let μ⁡(E)=0 when E is countable and μ⁡(E)=1 otherwise.

Lemma 4.2.

The following properties hold for all measures:

  1. 1.

    Monotonicity;

  2. 2.

    Countable subadditivity;

  3. 3.

    If Ei∈Σ and μ⁡(Ei)<∞, then μ⁢(⋂Ei)=limμ⁡(En) if the Ei are nested.

Proof.
  1. 1.

    Observe that B=A∪(B∖A) and then use additivity and nonnegativity.

  2. 2.

    Let Fi=En∖(⋃i<nEi). Then μ⁢(⋃Ek)=∑μ⁡(Fk)≤∑μ⁡(Ek).

  3. 3.

    Note that E1=E∪⋃k=1∞Ek∖Ek+1, so

We say that μ is finite if μ⁡(X)<∞ and that μ is σ-finite if X is a countable union of sets of finite measure.

Example 4.3.
  1. 1.

    The Lebesgue measure is σ-finite (just take [a,a+1] for a∈ℤ).

  2. 2.

    The counting measure is finite if and only if S is finite.

  3. 3.

    Then discrete measure if finite if and only if countably many points have positive weights and the sum of the weights is finite. It is σ-finite if and only if only countably many points have positive weight and each weight is finite.

  4. 4.

    This is a finite measure. In fact, any measure such that μ⁡(X)=1 is a probability measure.

4.1 Outer Measures

An outer measure is a countable subadditive function defined on 𝒫⁡(X) such that m∗⁢(∅)=0. Our main question is to extract a measure from an outer measure when we have no recourse to open sets, unlike in the Lebesgue case.

Definition 4.4.

E is (Caratheodory)-measurable if for all A⊆X, μ∗⁢(A)=μ∗⁢(E∩A)+μ∗⁢(Ec∩A).

Note that

  1. 1.

    This condition is equivalent to μ∗⁢(A)≥μ∗⁢(A∩E)+μ∗⁢(A∩Ec);

  2. 2.

    The condition holds for the emptyset;

  3. 3.

    E is measurable if and only if Ec is.

This tells us that Σc={E∣E⁢ is measurable} is a σ-algebra and that μ∗ is a measure on Σc.

To prove the last claim, we need to show countable additivity. First, we will do it for finite unions:

Also, additivity on disjoint sets is easy. Now we will consider countable disjoint unions. Set Fn=⋃j=1nEj and F=⋃i=1∞Ej. Then

Therefore, m∗⁢(Fn∩A)=∑j=1nm∗⁢(Ej∩A), so m∗⁢(A)≥∑j=1nm∗⁢(Ej∩A)+m∗⁢(Fnc∩A), so if we allow n→∞, then

Now let X be a metric space. As before, the Borel sets are the smallest σ-algebra containing all open balls. An outer measure is a metric outer measure if it has the separation property.

Theorem 4.5.

If m∗ is a metric outer measure, then the Borel sets are measurable and m∗|ℬX is a measure.

Proof.

We show that closed sets are measurable. Assume F is closed and m∗⁢(A)<∞. Then set

Then clearly the An are increasing. We will show that limm∗⁢(An)=m∗⁢(Fc∩A).

Set Bn=An+1∩Anc. Then d⁡(Bn+1,An)≥1n−1n+1. Then we see that

Then we see that m∗⁢(A)≥∑oddm∗⁢(Bn) and m∗⁢(A)≥∑evenm∗⁢(Bn), so in particular, ∑nm∗⁢(Bn)≤2⁢m∗⁢(A). Then we obtain

Because the sum is finite, we have the desired result. □

Theorem 4.6.

Let μ be a Borel measure that is finite on compact sets. Then for every E∈ℬ and ε>0, there exists an open G⊇E and closed F⊆E such that μ⁡(G∖E)<ε and μ⁡(E∖F)<ε.

Proof.

Let 𝒷 be a collection of sets satisfying both properties. We show this is a σ-algebra containing ℬ. Clearly, it is closed under complements. Thus suppose Ek∈𝒷 for each k. First, there exist Gk⊇Ek open such that μ⁡(Gk∖Ek)<ε2k, so their union is open and μ⁡(G∖E)<ε.

Similarly, there exists Fk⊆Ek suchthat μ⁡(Ek∖Fk)<ε2k, but the union F∗=⋃Fk is not closed. We prove that there exists a closed F⊆F∗ such that μ⁡(F∗∖F)<ε. Assume that Fk are increasing, so choose x0 and set Bn. Then let An=B¯n∖Bn−1, so write F∗=⋃n(F∗∩An), where F∗∩An=limkFk∩An=⋃k(Fk∩An).

For each n, there exists k⁡(n) such that μ⁡(F∗∖Fk⁡(n)∩An)<ε2n with μ⁡(B¯n)<∞. Then we see that F=⋃n≥1(Fk⁡(n)∩An), so μ⁡(F∗∖F)<ε and each F∩B¯k is closed, so F is closed.

Finally, we show that open G∈𝒷. Set Fk={x∈B¯k∣d⁡(x,Gc)≥1k}. Then we use the argument above. □

4.2 Results and Applications

Definition 4.7.

A premeasure is a measure-like function on an algebra A (closed under complement and finite union).

We see that a premeasure defines an outer measure. Define f to be measurable if the preimage of any open set is a Borel set and define f to be simple if

for En measurable. Then given a measure μ, definee

where ∫ψ⁢dμ=∑an⁢μ⁢(En).

Theorem 4.8.

μ∗|ΣA is a measure where ΣA is the smallest σ-algebra containing A.

We have the following results, in analogy with the Lebesgue case.

Theorem 4.9 (Egorov).

Let fk be defined on E with μ⁡(E)<∞. Then for all ε>0, there exists Aε⊆E with μ⁡(E∖Aε)<ε and such that fk→f uniformly on Aε.

Lemma 4.10 (Fatou).

If f is nonnegative, then ∫lim inffn⁢dμ≤lim inf∫fn⁢dμ.

As before, if f=f+−f− with both f+,f−≥0, then ∫f⁢dμ is defined provided that not both ∫f+⁢dμ and ∫f−⁢dμ are infinite. Then f is

  1. 1.

    Integrable if ∫X|f|⁢dμ<∞. We say f∈L1⁢(X,μ);

  2. 2.

    Square-integrable if ∫Xf2⁢dμ<∞. We say that f∈L2⁢(X,μ).

Theorem 4.11 (Monotone Convergence).

If fn≥0 and fn↗f almost everywhere, then ∫f⁢dμ=lim∫fn⁢dμ.

Theorem 4.12 (Dominated Convergence).

If fn→f almost everywhere and |fn|≤g for some integrable g, then ∫X|fn−f|⁢dμ→0.

Example 4.13.

Consider the Stieltjes integral. Recall that the Riemann integral is ∫f⁢dx, and the Stieltjes integral is ∫f⁢dg for some monotone g. Finally, the Lebesgue integral is the correct definition of integration.

To construct a measure associated to a monotone function F, first normalize by changing F at the jumps to ensure right continuity.

Theorem 4.14.

Given a monotone, right-continuous F, there exists a unique Borel measure μF such that μF⁢([a,b])=F⁡(b)−F⁡(a) for all a<b. Moreover, if μ is a Borel measure that is finite on compact sets, there exists a unique F (up to constants) such that μ=μF.

Proof.

The first part is straightforward, but painful (Robin’s words), so it is not given here. On the other hand, if μ is a given Borel measure and is finite on compact sets, how do we define F? We want μ⁡((a,b])=F⁡(b)−F⁡(a). If x>0, then we must have μ⁡((0,x])=F⁡(x) and similarly for x<0. It is easy to see that this function is monotone, so we show right continuity:

□

4.3 Signed Measures

Note that we can add measures and multiply them by nonnegative constants. How can we obtain full linearity?

Definition 4.15.

A function ν on a σ-algebra Σ is a signed measure if

  1. 1.

    ν⁡(∅)=0;

  2. 2.

    ν does not take on both ±∞;

  3. 3.

    ν is additive on disjoint sets. In particular, if ν⁡(E) is finite and E=⋃Ej, then ∑ν⁡(Ej) converges absolutely.

We say that A⊆X is positive with respect to ν if A∈Σ and for each E⊆A with E∈Σ, we have ν⁡(E)≥0. Define similar for negativity. Then a set N is null if it is both positive and negative. However, having measure 0 does not imply that a set is null.

Lemma 4.16.

Subsets and countable unions of positive sets are positive.

Proof.

Subsets of positive sets are positive by definition. For the second part, if E⊆⋃An, set En=E∩An∩An−1c∩⋯∩A1c. Then ν⁡(En)≥0 and ν⁡(E)=∑ν⁡(En)≥0. □

Lemma 4.17.

If E satisfies 0<ν⁡(E)<∞, then there exists a positive A⊆E with ν⁡(A)>0.

Proof.

If E is positive, we are done. Otherwise, let n1 be the smallest integer such that there exists E1⊆E with ν⁡(E1)<−1n. Inductively, choose nk to be the smalles integer such that there exists Ek⊆E∖(⋃j<kEj) with ν⁡(Ek)<−1nk. This process may or may not terminate, so set A=E∖⋃Ek. Then clearly A has positive measure because ν⁡(A)=ν⁡(E)−∑ν⁡(Ek)>0 and ∑ν⁡(Ek) converges absolutely. Therefore ∑1nk<−∑ν⁡(Ek) is finite, so either the process stops or nk→∞, so ν⁡(A)<∞.

To show that A is positive, it is clear if the process stops. Otherwise, for all ε>0, there exists k such that 1nk−1<ε. Then A⊆E∖⋃j=1nkEj, so A contains no subset with measure less than −ε. □

Theorem 4.18 (Hahn Decomposition Theorem).

Given a signed measure ν, there exists A positive and B negative such that X=A∪B and A∩B=∅.

Proof.

Assume that ν does not take the value ∞. Define λ=supA⁢ positiveν⁡(A). Note that λ≥0. Then choose Ai positive such that λ=limν⁡(Ai) and set A=⋃Ai is positive and that λ≥ν⁡(A).

Also, A∖Ai⊆A, so each ν⁡(A∖Ai)≥0. Thus ν⁡(A)≥ν⁡(Ai) for all i, so ν⁡(A)=λ. Set B=Ac. We will show that B is negative. To see this, if E⊆B with ν⁡(E)>0, then A∪E is positive with measure strictly greater than λ. □

Note that this A,B are not unique. Then define the positive and negative parts of ν by ν+⁢(E)=ν⁡(E∩A) and ν−⁢(E)=−ν⁡(E∩B). These are both measures. Then we see that ν=ν+−ν−. This decomposition is unique. Also, define the total variation of ν as |ν|=ν++ν−.

Note that |ν| is defined because ν+,ν− cannot both be infinite. If |ν|<∞, we say that ν is finite. If two measures μ1,μ2 have sets A1,A2 such that A1∩A2=∅,A1∪A2=X and μ1⁢(A2)=μ2⁢(A1)=0, then we call them mutually singular and write μ1⟂μ2.

Example 4.19.

The discrete measure is mutually singular with the Lebesgue measure. The Cantor measure is also mutually singular with the Lebesgue measure.

The opposite notion is absolute continuity of measures.

Definition 4.20.

We say that ν is absolutely continuous with respect to μ, written ν≪μ, if for all E with μ⁡(E)=0, ν⁡(E)=0.

Example 4.21.

If E is measurable and f is integrable, then νf⁢(E)=∫Ef⁢dμ is absolutely continuous with respect to μ.

Theorem 4.22 (Radon-Nikodym).

Let (X,ℬ,μ) be a measure space with Borel sets and suppose μ is σ-finite and ν≪μ. Then there exists a nonnegative μ-measurable f such that for all E∈ℬ, ν⁡(E)=∫Ef⁢dμ. Here, f is called the Radon-Nikodym derivative of ν, sometimes written f=[dνdμ].

Before we prove this result (following Royden), we need two lemmas:

Lemma 4.23.

If Bα is a collection of Borel sets ordered by some countable α∈D satisfying Bα⊆Bβ for α<β, then there exists a measure f such that f≤α on Bα and f≥α on Bαc.

Proof.

Given x∈X, set f⁡(x)=inf{α∣x∈Bα}. Then if x∈Bα, f⁡(x)≤α. Then if x∉Bα, f⁡(x)≥α because x∉Bβ for β<α. To show that f is measurable, note that

which is measurable. □

Lemma 4.24.

Let α∈D countable and Bα∈ℬ such that if α<β, then μ⁡(Bα∖Bβ)=0. Then the same conclusion as in Lemma 4.23 holds.

Proof.

Let C=⋃α<βBα∖Bβ. Set Bα′=Bα∪C. Note that C has measure 0. Then apply Lemma 4.23 to the Bα′. □

Proof of Radon-Nikodym.

Assume μ is finite. For all rational α, consider the signed measure ν−α⁢μ. By the Hahn decomposition theorem, there exist Aα,Bα=Aαc with ν−α⁢μ≥0 on Aα and ν−α⁢μ≤0 on Bα. Then set A0=X,B0=∅ and note that Bα∖Bβ=Bα∩Aβ, so ν−α⁢μ≤0 on Bα and ν−β⁢μ≥0 on Aβ. Therefore, if β>α, then μ⁡(Bα)∖Bβ=0, so by Lemma 4.24, there exists f such that f≥α on Aα and f≤α on Bα. Because B0=∅, f≥0.

We now show that ∫Ef⁢dμ=ν⁡(E). Fix n and set

Then E is the disjoint union E=⨆k=0∞Ek∪E∞, so

Now Ek⊆Bk+1n∩Akn, so kn≤f≤k+1n on Ek, so

By definition of B,A, we know that kn⁢μ⁢(Ej)≤ν⁡(Ek)≤k+1n⁢μ⁢(Ek), so

Then add and take n→∞. □