21-720 Measure and Integration, Fall 2014 Notes typeset by Adam Gutter Lecture 1 - 08/25/2014 Book Recommendations • For grad students, use Folland. In-depth, comprehensive, challenging. • Also good but simpler is Cohn. Cohn is a very nice read. • Rudin is good, but very dense. Quotes - ”My favourite book is Folland.” ”These books are all isomorphic.” Here’s some motivation for the development of the Lebesgue measure/Lebegue integral: Challenge - Let pf n q be a sequence of functions defined on r0, 1s,0 ď f n ď 1 for all n. Show that if f n Ñ 0 pointwise, then lim nÑ8 f n ş 1 0 f n “ 0. Using Riemann integration, this is easy if f n Ñ 0 uniformly. Not true if we only have pointwise convergence. However, the problem becomes easy again if we use the theory of Lebesgue integration. Intuition - The Riemann integral is calculated by subdividing the domain of a function, then draw- ing rectangles to approximate the area under a curve. The Lebesgue integral is calculated by subdividing the range of a function. Approximate f by a function of the form n ř i“1 a i χ Ai , where χ S pxq“ # 1 x P S 0 x R S , where the sets A i are disjoint and ”nice” and f « a i on each A i . We then approximate ż b a f « n ÿ i“1 a i ¨ ”lengthpA i q” For example, we might have a 1 㨨¨ă a n , A i “ f ´1 pra i ,a i`1 qq. To get some intuition for this analogy, consider a banker counting money. One way to do this is to add the bills together in sequence. This is like Riemann integration. Lebesgue integration is analogous to sorting the bills into stacks of different denomination, then counting the size of each stack. We will develop a notion of ”size” for a set A, denoted by λpAq. In one dimension, this will correspond to length. In two dimensions, this will correspond to area. In three dimensions, this will correspond to volume, etc. There are some things we must be careful about, lest we get in trouble. Banach-Tarski Paradox - There is a positive integer n, disjoint sets A 1 ,...,A n Ď R 3 , and maps R 1 ,...,R n : R 3 Ñ R 3 that are combinations of rotations, reflections, and translations such that Ť 1ďiďn A i “ Bp0, 1q and ď 1ďiďn R i pA i q“ Bp0, 1qY Bp2, 1q Quote - ”With this problem, I’ve solved world hunger. Give me one loaf of bread, I’ll give you two!” Issue - Although it seems like we’re doubling the volume of the ball, there ends up being no sensible notion of volume for the sets A i , so we can’t identify the volume of the ball with the sum of the volumes of the pieces. We’ll formalize this later. Remark - This can be done with n “ 5 (and this is optimal). Definition - Let X be a set. We say Σ Ď P pXq is a σ-algebra on the set X if
76
Embed
Notes typeset by Adam Gutter - CMUgautam/teaching/2014-15/720-measure/pdfs/a… · Notes typeset by Adam Gutter Lecture 1 - 08/25/2014 Book Recommendations For grad students, use
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 1 - 08/25/2014
Book Recommendations
• For grad students, use Folland. In-depth, comprehensive, challenging.
• Also good but simpler is Cohn. Cohn is a very nice read.
• Rudin is good, but very dense.
Quotes - ”My favourite book is Folland.” ”These books are all isomorphic.”
Here’s some motivation for the development of the Lebesgue measure/Lebegue integral:Challenge - Let pfnq be a sequence of functions defined on r0, 1s, 0 ď fn ď 1 for all n. Show that if
fn Ñ 0 pointwise, then limnÑ8
fnş1
0fn “ 0.
Using Riemann integration, this is easy if fn Ñ 0 uniformly. Not true if we only have pointwiseconvergence. However, the problem becomes easy again if we use the theory of Lebesgue integration.
Intuition - The Riemann integral is calculated by subdividing the domain of a function, then draw-ing rectangles to approximate the area under a curve. The Lebesgue integral is calculated by
subdividing the range of a function. Approximate f by a function of the formnř
i“1
aiχAi , where
χSpxq “
#
1 x P S
0 x R S, where the sets Ai are disjoint and ”nice” and f « ai on each Ai. We then
approximateż b
a
f «nÿ
i“1
ai ¨ ”lengthpAiq”
For example, we might have a1 ă ¨ ¨ ¨ ă an, Ai “ f´1prai, ai`1qq.To get some intuition for this analogy, consider a banker counting money. One way to do this
is to add the bills together in sequence. This is like Riemann integration. Lebesgue integration isanalogous to sorting the bills into stacks of different denomination, then counting the size of eachstack.
We will develop a notion of ”size” for a set A, denoted by λpAq. In one dimension, this willcorrespond to length. In two dimensions, this will correspond to area. In three dimensions, this willcorrespond to volume, etc.
There are some things we must be careful about, lest we get in trouble.
Banach-Tarski Paradox - There is a positive integer n, disjoint sets A1, . . . , An Ď R3, and mapsR1, . . . , Rn : R3 Ñ R3 that are combinations of rotations, reflections, and translations such thatŤ
1ďiďn
Ai “ Bp0, 1q andď
1ďiďn
RipAiq “ Bp0, 1q YBp2, 1q
Quote - ”With this problem, I’ve solved world hunger. Give me one loaf of bread, I’ll give you two!”
Issue - Although it seems like we’re doubling the volume of the ball, there ends up being no sensiblenotion of volume for the sets Ai, so we can’t identify the volume of the ball with the sum of thevolumes of the pieces. We’ll formalize this later.Remark - This can be done with n “ 5 (and this is optimal).
Definition - Let X be a set. We say Σ Ď PpXq is a σ-algebra on the set X if
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
1. H P Σ
2. A P Σ ùñ Ac P Σ
3. pAnq Ď Σ ùñŤ
nAn P Σ
Intuition - Think of Σ as all subsets of X to which we can assign a meaningful notion of volume.These will be our ”measurable” sets.Remark - Conditions 1 and 2 imply that X P Σ. Conditions 2 and 3 imply that Σ is closed undercountable intersections.
Example - Σ “ tH, Xu, Σ “ PpXq.
Proposition - If Σ1 and Σ2 are σ-algebras on X, then so is Σ1XΣ2. In fact, if tΣαuαPA is a collectionof σ-algebras on a set X, then so is
Ş
αPA
Σα.
Proof - Easy exercise.
Definition - Let E Ď PpXq be arbitrary. Denote by σpEq the σ-algebra on X generated by E . Thisis defined to be the intersection of all σ-algebras on X that contain E .
By the previous proposition, this definition is justified. Intuitively, σpEq is the smallest σ-algebraon X that contains E .
Example/Definition - Take X “ Rd, E “ topen setsu. We call σpEq the Borel σ-algebra on X and
denote it by BpRdq.Remark - BpRdq “ σpthypercubes in Rduq (exercise).Remark - BpRdq “ σptclosed setsuq.
Definition - Let X be a set, Σ a σ-algebra on X. We say µ : Σ Ñ r0,8s is a positive measure onpX,Σq if
1. µpHq “ 0.
2. Given disjoint sets pAnq Ď Σ, we have µpŤ
ně1Anq “
ř
ně1µpAnq
Remark - In general, (2) doesn’t imply (1). Why? Given A P Σ, we have
µpAq “ µpAYHq “ µpAq ` µpHq
We’d like to conclude that µpHq “ 0, but this is only true if µpAq ă 8. Thus, 2 implies 1 if thereis A P Σ such that µpAq ă 8.
Example (Counting measure) - Let X be any set, Σ “ PpXq, and define µpAq “ |A|, the cardinalityof A. That this is a measure is left as an exercise.Example (δ measure) - Let a0 P X, Σ “ PpXq. Define
δa0pXq “
#
1 x P A
0 x R A
That this is a measure is left as an exercise. We call this the δ measure at a0.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 2 - 08/27/2014
Goal - Construct the Lebesgue measure, which will be a measure on pRd,BpRdqq. This measure willsatisfy
λppa1, b1q ˆ ¨ ¨ ¨ ˆ pad, bdqq “ź
1ďiďd
pbi ´ aiq
for all ai ă bi. That is, the measure will agree on cells in Rd.
Goal 2 - Define a more robust notion of integration. Abstractly, given a measure space pX,Σ, µq:
1. If s : X Ñ R is of the form s “ř
1ďiďN
aiχAi , where the sets Ai P Σ are disjoint, defineş
Xs dµ “
ř
1ďiďN
aiµpAiq.
2. Given an arbitrary function f : X Ñ R, approximate f by functions from (1) to defineş
Xf dµ.
Construction of Lebesgue Measure
Definition - We say I Ď R is a cell if it is a finite interval, i.e. I is of the form pa, bq, pa, bs, ra, bq, orra, bs, where a, b P R.Definition - We say I Ď Rd is a cell if I “ I1 ˆ ¨ ¨ ¨ ˆ Id, where each Ii is a cell in R. Given anonempty cell I in Rd, define
`pIq “ź
1ďiďd
psuppIiq ´ infpIiqq
and define `pHq “ 0. (Intuitively, this is volume.)
Definition - Given A Ď Rd arbitrary, define
λ˚pAq “ inftÿ
ně1
`pInq | A Ďď
ně1
In, In cellsu
We call λ˚ the Lebesgue outer measure. Although λ˚ is defined on all of PpRdq, it is only countablyadditive on a proper subset of PpRdq (as we shall see).
Properties of λ˚
1. λ˚pHq “ 0
2. A Ď B ùñ λ˚pAq ď λ˚pBq (monotonicity)
3. If Ai Ď Rd, then λ˚pŤ
iě1
Aiq ďř
iě1
λ˚pAiq (countable subadditivity)
Proof - (1) and (2) are immediate. (Quote - ”For the entire semester, there’s only one trick we keepusing, which we’ll use now.”) For (3), let ε ą 0. For each i, let pIi,nq be a cover of Ai by cells suchthat
ÿ
ně1
`pIi,nq ď λ˚pAiq `ε2i
Clearly tIi,nui,nPN is a countable cover ofŤ
iě1
Ai by cells, so by definition, we have
λ˚pď
iě1
Aiq ďÿ
i,n
`pIi,nq ďÿ
i
pλ˚pAiq `ε2i q “
ÿ
iě1
λ˚pAiq ` ε
Since ε ą 0 was arbitrary, let εÑ 0` to get λ˚pŤ
iě1
Aiq ďř
iě1
λ˚pAiq.
Next time, we’ll show two things. First, that λ˚pIq “ `pIq for any cell I. Next, we’ll show thatλ˚ has a property called separated additivity. That is, if the distance between two sets A,B Ď Rd ispositive, then λ˚pAYBq “ λ˚pAq ` λ˚pBq.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 3 - 08/29/2014
Definition - We say A,B Ď Rd are separated if dpA,Bq ą 0, where
dpA,Bq “ inft|a´ b| | a P A, b P Bu
Proposition 1 - If A,B Ď Rd with dpA,Bq ą 0, then λ˚pAYBq “ λ˚pAq ` λ˚pBq.Proof - We may assume λ˚pAq ` λ˚pBq ă 8 (otherwise, the result is clear by monotonicity). Bysubadditivity, we have λ˚pA Y Bq ď λ˚pAq ` λ˚pBq, so it suffices to show that λ˚pA Y Bq ěλ˚pAq ` λ˚pBq.
Denote δ “ dpA,Bq and let tIku be a cover of AYB by cells. By subdividing each Ik into cells ofdiameter at most δ
2 , we obtain a new collection tJku of cells. Note thatř
k
`pIkq ěř
k
`pJkq since the
subdivision process may result in new cells that are duplicated (and we only count each cell once).Since diampJkq ă
δ2 for each k, then Jk X A ‰ H implies Jk X B “ H and vice-versa. Now
partition tJku “ tJ1ku Y tJ
2ku, where J 1k XA ‰ H and J2k XA “ H for each k. Now
ď
k
J 1k Ě A,ď
k
J2k Ě B ùñÿ
k
`pIkq ěÿ
k
`pJkq “ÿ
k
`pJ 1kq `ÿ
k
`pJ2k q ě λ˚pAq ` λ˚pBq
Taking the infimum over all covers tIku, we have λ ˚ pAYBq ě λ˚pAq ` λ˚pBq, as desired.Remark - For sets A and B that aren’t separated, we don’t have disjoint additivity. Intuitively, theproblem is with the boundary of A and B. If A, B are sufficiently complicated, then there is no wayto divide a cover of A Y B into a cover of A and a cover B (like we did above). Covering A maynecessarily entail covering some of B and/or vice-versa, so we get double-counting.
Proposition 2 - If I is a cell, then λ˚pIq “ `pIq.Lemma - λ˚pAq “ inft
ř
n`pInq | A Ď
Ť
nIn, In open cellsu “: µ˚pAq.
Proof - Clearly λ˚pAq ď µ˚pAq. We need to show that µ˚pAq ď λ˚pAq. We may assume λ˚pAq ă 8,otherwise this is clear.
Let ε ą 0 and let pInq be a cover of A by cells such thatř
n`pInq ď λ˚pAq ` ε. For each n, let
Jn Ě In be an open cell with `pJnq ´ `pInq ăε
2n . Now
µ˚pAq ďÿ
n
`pJnq ďÿ
n
`pInq ` ε ď λ˚pAq ` 2ε
Since ε ą 0 was arbitrary, then µ˚pAq ď λ˚pAq, as desired.
Proof (of Proposition 2) - First assume that I is closed. Clearly λ˚pIq ď `pIq by definition. We needto show λ˚pIq ě `pIq. By Lemma 1, it’s enough to consider covers of I by open cells.
Say pInq is a cover of I by open cells in Rd. We need to show thatř
ně1`pInq ě `pIq. Since I
is closed, it is compact, so there is N such that I ĎŤ
1ďnďN
In. Extend the faces of the cells tInu
to hyperplanes and use these to subdivide I into new cells pJnq. Clearlyř
n`pJnq “ `pIq (by the
distributive law for multiplication in R), andř
n`pJnq ď
ř
n`pInq. Thus,
ř
n`pInq ě `pIq. Taking the
infimum over covers of I by open cells, we have λ˚pIq ě `pIq, as desired.If I is not closed, let J Ď I be a closed cell such that `pJq ě `pIq ´ ε. Then
λ˚pIq ě λ˚pJq “ `pJq ě `pIq ´ ε
Taking εÑ 0`, we have the desired result.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Remark (translation invariance) - For any cell I and point x0 in Rd, `pIq “ `pI ` x0q. Therefore,λ˚pAq “ λ˚pA` x0q for any set A and point x0. That is, the Lebesgue outer measure is translationinvariant.Remark - There’s another way to obtain λ˚. Instead of using the volume of cubes, try and definea notion of volume by requiring it to be translation invariant (among other things). Doing this, werecover the Lebesgue measure. More generally, this can be applied to topological groups to definesomething called Haar measure.
We return to the construction of the Lebesgue measure, introducing Caratheodory’s criterionand abstracting our discussion.
Definition - We say µ˚ is an outer measure on X if µ˚ : PpXq Ñ r0,8s and
1. µ˚pHq “ 0
2. A Ď B ùñ µ˚pAq ď µ˚pBq
3. µ˚pŤ
nAnq ď
ř
nµ˚pAnq
Example - λ˚ is an outer measure.
Theorem (Caratheodory) - Let µ˚ be an outer measure on X. Define
Σ “ tE | for all A Ď X, µ˚pAq “ µ˚pAX Eq ` µ˚pAX Ecqu
Then
1. Σ is a σ-algebra.
2. µ æΣ is a measure.
Proof - Next time.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 4 - 09/01/2014Theorem (Caratheodory) - Let µ˚ be an outer measure on X. Define
Σ “ tE | for all A Ď X, µ˚pAq “ µ˚pAX Eq ` µ˚pAX Ecqu
Then
1. Σ is a σ-algebra, the (µ˚-)measurable subsets of X.
2. µ æΣ is a measure.
Proof - We first show that H P Σ and Σ is closed under complementation.
• H P Σ: For any A Ď X, we have µ˚pAq “ 0` µ˚pAq “ µ˚pAXHq ` µ˚pAXXq.
• Given E P Σ, we have µ˚pAq “ µ˚pAX Eq ` µ˚pAX Ecq “ µ˚pAX Ecq ` µ˚pAX pEcqcq.
To show that Σ is closed under countable unions, we first prove some facts about finite unionsand µ˚.
Claim 1 E,F P Σ ùñ E Y F P Σ. (Thus, Σ is closed under finite unions.)
Proof Let A Ď X. Then, appealing to measurability of E,F , we have
µ˚pAX pE X F qq ` µ˚pAX pE X F qcq
“ µ˚pAX pE Y F q X Eq ` µ˚pAX pE Y F q X Ecq ` µ˚pAX pE X F qcq
“ µ˚pAX Eq ` µ˚pAX F X Ecq ` µ˚pAX F c X Ecq
“ µ˚pAX Eq ` µ˚pAX Ecq
“ µ˚pAq
Claim 2 If E1, . . . , En P Σ are pairwise-disjoint and A Ď X, then
µ˚pAX pď
1ďiďN
Eiqq “ÿ
1ďiďN
µ˚pAX Eiq
Proof By induction, it’s enough to consider N “ 2. Let E,F P Σ be disjoint,A Ď X. Now
To show that Σ is closed under countable unions (proving (1)), it’s enough by Claim 1 to showthat Σ is closed under finite disjoint unions. Let pEnq Ď Σ pairwise-disjoint, A Ď X. We wantŤ
nEn P Σ.
By subadditivity, it’s enough to show that µ˚pAq ě µ˚pAXŤ
nEnq`µ
˚pAXpŤ
nEnq
cq. By Claim
2 and monotonicity, we have
µ˚pAq “ µ˚pAXď
1ďnďN
Enq ` µ˚pAX p
ď
1ďnďN
Enqcq
ě µ˚pAXď
1ďnďN
Enq ` µ˚pAX p
ď
ně1
Enqcq
“ÿ
1ďnďN
µ˚pAX Enq ` µ˚pAX p
ď
ně1
Enqcq
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Taking the limit N Ñ8, we have (by subadditivity)
µ˚pAq ěÿ
ně1
µ˚pAX Enq ` µ˚pAX p
ď
ně1
Enqcq ě µ˚pAX p
ď
ně1
Enqq ` µ˚pAX p
ď
ně1
Enqcq
Thus, we have equality above, which proves the result. Moreover, an immediate consequence of thisproof is that µ˚ is countably additive on Σ (take A “
Ť
ně1En).
Let’s return to the construction of the Lebesgue measure on Rd. Let λ˚ be the Lebesgue outermeasure. By Caratheodory,
L :“ LpRdq “ tE Ď Rd | for all A Ď Rd, λ˚pAq “ λ˚pAX Eq ` λ˚pAX Ecqu
is a σ-algebra and λ :“ λ˚ æΣ is a measure, the Lebesgue measure.This seems promising, but there’s one more problem. We don’t know what L is! It may be
trivial!
Proposition - Let I be a cell. Then I P L.
Corollary - BpRdq Ď LpRq.Remark - BpRdq Ĺ LpRdq Ĺ PpRdq. In fact, |BpRdq| “ |R| and |LpRdq| “ |PpRdq|.
Proof (of Proposition) - We need to show that for all A, λ˚pAq ě λ˚pA X Iq ` λ˚pA X Icq. Letε ą 0. Choose a cell I1 Ď I such that λ˚pIzI1q ă ε and dpI1, I
cq ą 0. By separated additivity andmonotonicity of λ˚, we have
λ˚pAX I1q ` λ˚pAX Icq “ λ˚pAX pI1 Y I
cqq ď λ˚pAq p˚q
Moreover,λ˚pAX Iq ď λ˚pAX I1q ` λ
˚pAX pIzI1qq ď λ˚pAX I1q ` ε
Rearranging, we have λ˚pAX I1q ě λ˚pAX Iq ´ ε. By p˚q, we have
We’ve now shown that the Lebesgue measure exists, is nontrivial, and agrees with our intuitivenotion of volume for cells. Next time, we’ll show that the Lebesgue measure is the unique measurethat satisfies our requirements.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 5 - 09/05/2014
Theorem (Uniqueness) - If µ is a measure on pRd,Lq (or pRd,Bq) and µpIq “ λpIq for all cells I,then
µpAq “ λpAq
for all A P L (or B).Proof - =We proceed in steps.
1. Let A P L be arbitrary. Let pInq be a cover of A by cells. Then by countable subadditivityand monotonicty,
µpAq ďÿ
n
µpInq “ÿ
n
λpInq “ÿ
n
`pInq
Taking the infimum over all possible covers, we have µpAq ď λpAq.
2. Now suppose that A P L is bounded, say A Ď I, a bounded cell. By the previous step,
Our approach above used a neat trick, but now we’ll consider a more general, abstract approachto uniqueness.
Uniqueness Abstractly
Let X be some set, Σ a σ-algebra on X, C Ď PpXq. Let µ, ν be measures on pX,Σq such thatµpXq “ νpXq ă 8. Suppose that µ “ ν on C. Does this imply that µ “ ν on σpCq.
As it turns out, the answer is no! For example, consider two sets A and B. We might haveµpAzA X Bq “ µpBzA X Bq “ 0 and µpA X Bq “ 1, while νpAzA X Bq “ νpBzA X Bq “ 1, whileνpAXBq “ 0. Then µ “ ν on tA,Bu, but µ and ν disagree on AXB.
To prove that µ “ ν on σpCq, we need C to be closed under intersections! Now denote D “
tA | µpAq “ νpAqu. Does D have the properties of a σ-algebra that we want?
1. H P D is clear.
2. A P D ùñ Ac P D follows from µpXq “ νpXq ă 8.
3. We can’t generally show that D is closed under finite unions (let alone countable) unless weknow that D is closed under finite intersections! Instead, let’s consider a property that weknow holds of D.
31 If pAnq Ď D is increasing, thenŤ
nAn P D.
21 While we’re at it, let’s generalize (2): If A,B P D with A Ď B, then BzA P D. So, not only isD closed under complements, but it’s closed under relative complements.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
The discussion above motivates a number of definitions.
Definition - We say Π Ď PpXq is a π-system Π ‰ H and
A,B P Π ùñ AXB P Π
Definition - We say Λ Ď PpXq is a λ-system if
1. X P Λ
2. A,B P Λ, A Ď B ùñ BzA P Λ
3. pAnq Ď Λ, An Ď An`1 ùñŤ
nAn P Λ.
Theorem (Dynkin) - If Π is a π-system and Λ is a λ-system, then
Λ Ě Π ùñ Λ Ě σpΠq
Corollary - If µ, ν are finite measures that agree on Π, X P Π, then µ “ ν on σpΠq.Remark 1 - The intersection of an arbitrary family of λ-systems is a λ-system.Remark 2 - If Λ is both a Λ-system and a π-system, then it is a σ-algebra. (Intuition: ”finiteintersection + complements + increasing countable unions = countable unions”)
Proof (of Theorem) - It’s enough to show that λpΠq Ě σpΠq, since Λ Ě λpπq. By Remark 2, it’senough to show that λpπq is a π-system, so then it’s a σ-algebra.
For any C P λpπq, set ΛC “ tA P λpπq | A X C P λpπqu. To show that λpπq is a π-system, weshow that ΛC “ λpπq for each C P λpπq. In particular, it’s enough to show that for any C P λpπq,ΛC is a λ-system that contains π.
Claim ΛC is a λ-system.
Proof 1. Clearly X P ΛC .
2. If A,B P ΛC with A Ď B, then
pBzAq X C “ pB X CqzpAX Cq P λpπq
since AX C,B X C P λpπq.
3. Let pAnq Ď ΛC be increasing. Then
pď
ně1
Anq X C “ď
ně1
pAn X Cq P λpπq
since An X C Ď An`1 X C.
Claim ΛC Ě Π.
Proof We need to show for all B P Π, B X C P λpΠq. If C P Π, this is immediatesince Π is a π-system. If C P λpΠq, fix B P Π. By the previous case, ΛB Ě λpπq, soB X C P λpΠq, so B P ΛC . Thus, ΛC Ě λpπq, so we are done.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 6 - 09/08/2014
Regularity
Definition - Let pX, dq be a metric space, BpXq the Borel σ-algebra of X. We say µ is a regularBorel measure on X if
1. µ is a measure on pX,BpXqq
2. For every A P BpXq, µpAq “ inftµpUq | U Ě A is openu
3. For every open U Ď X, µpUq “ suptµpKq | K Ď U is compactu
4. For every compact K P BpXq, µpKq ă 8.
A measure satisfying condition 2 is called an outer regular measure. A measure satisfying condition3 is called an inner regular measure. A measure satisfying condition 4 is called a Radon measure.
If A P BpXq satisfies condition 2, we say A is inner regular with respect to µ or µ-inner regular.If A satisfies condition 3, we say A is outer regular with respect to µ or µ-outer regular.
Goal: Regularity of λ.
Theorem (Regular of Finite Borel Measures) - Let X be a compact space. Let µ be a finite Borelmeasure on X. Then µ is regular.Proof - Note, µ is Radon since it is finite. Define
Λ “ tA P BpXq | A is µ-inner and µ-outer regularu
It suffices to show that Λ contains all open sets and that Λ is a λ-system, since then Λ Ě BpXq(follows because open sets form a π-system).
Let U Ď X be open. Then U is trivially µ-outer regular. For each n, define Kn “ tx PU |; dpx, U cq ě 1
nu. Then clearly Kn is closed. Since X is compact, then Kn is compact. Note,since U is open, then x P U if and only if dpx, Uq ą 0. Therefore, U “
Ť
ně1Kn. Since Kn Ď Kn`1,
then limnÑ8
µpKnq “ µpUq, thus U is µ-inner regular, so Λ contains all open sets.
We now argue that Λ is a λ-system. Clearly X P Λ since X is both open and compact. Now letA1, A2 P Λ with A1 Ď A2. Fix ε ą 0. Since µ is finite, then for each i “ 1, 2, there are open Ui andcompact Ki such that
Ki Ď Ai Ď Ui, µpAizKiq ă ε, µpUizAiq ă ε
Note that K2zU1 Ď A2zA1 Ď U2zK1, where K2zU1 is compact and U2zK1 is open. Moreover,
Thus, λ is outer regular. To show that λ is inner regular, we reduce to the case that A is bounded.(For the general case, consider the bounded sets AX r´n, nsd.)
Since A is bounded, there is a closed cell I with I Ě A. Fix ε ą 0. By outer regularity, there isopen U Ě IzA such that
We are done, since IzU is compact (closed and bounded).Finally, note that compact sets clearly have finite measure because they are bounded.
Proposition II - For all A P L, there is ε ą 0, U open, and C closed such that
C Ď A Ď U, λpUzCq ă ε
Proof - Case I: Suppose that λpAq ă 8. By Proposition I, given ε ą 0 there are compact K Ď Aand open U Ě A with
λpAq ´ ε2 ă λpKq ď λpAq ď λpUq ă λpAq ` ε
2
Now λpUzKq ă ε and K is closed, so we are done.Case II: A is arbitrary. Set Dn “ Bp0, n` 1qzBp0, nq and write A “
Ť
npAXDnq. By Case I, for
each n and ε ą 0, there are open Un and closed Cn with
Cn Ď AXDn Ď Un, λpUnzCnq ăε
2n
Set U “Ť
nUn, C “
Ť
nCn. Clearly U is open. Moreover, C is closed since Cn Ď Dn for each n.
Finally,λpUzCq ď
ÿ
n
ε2n “ ε
Corollary - Let A Ď Rd. ThenA P L ðñ A “ Fσ YN,
where Fσ is a countable union of closed sets and λpNq “ 0.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Existence of a Nonmeasurable Set
Proposition - LpRq Ĺ PpRqProof - Define an equivalence relation „ on R by x „ y if and only if x ´ y P Q. Partition R intoequivalence classes, say for α P R we write
Cα “ tx P R | x´ α P Qu “ α „
Let A Ď R be a set of representatives for the equivalence classes, i.e. for each α, AXCα is a singleton.
Claim A R LpRq
Proof Suppose not, i.e. A P L. Then R “Ť
qPQpA` qq (disjoint union), so
λpRq “ÿ
qPQλpA` qq “
ÿ
qPQλpAq ùñ λpAq ą 0
We argue that λpAq “ 0, which is a contradiction. By inner regularity, it suffices toshow that λpKq “ 0 for all compact K Ď A. Given such K, set E “
Ť
qPQ,|q|ď1
pK`qq.
Now that E P L and λpEq ă 8 since E is bounded. Now
λpEq “ÿ
qPQ,|q|ď1
λpKq ùñ λpKq “ λpEq “ 0
Remark - There is A Ď R such that
E Ď A, E P L ùñ λpEq “ 0, E Ď Ac, E P L ùñ λpEq “ 0
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 8 - 09/12/2014
Completion
Proposition - Let A Ď Rd. Then
A P LpRdq ðñ DF,G P BpRdq, F Ď A Ď G, λpGzF q “ 0
Proof p ùñ q Suppose A P L. Then for each n P N, there are closed Cn, open Un with
Cn Ď A Ď Un, λpUnzCnq ă1n
Let F “Ť
nCn, G “
Ť
nUn. Now λpGzF q ď λpUnzCnq ă
1n for all n, so λpGzF q “ 0.
p ðù q Write A “ F Y pAzF q. Since AzF Ď GzF and λpGzF q “ 0, then AzF P L, so A P L.
Let pX,Σ, µq be a measure space.Definition - Define N “ tA Ď X | DB P Σ, B Ě A, µpBq “ 0u, the set of all µ-null subsets of X.Definition - We say Σ is complete with respect to µ if Σ Ě N .Remark - The σ-algebra of a measure constructed via an outer measure using the Caratheodoryconstruction is complete.Proof - Let µ be such a measure and let N be any µ-null set, say M Ě N with µpMq “ 0. For anyA P X, we have
µpAq ě µpMq ` µpAXN cq ě µpAXNq ` µpAXN cq,
so N P Σ.Corollary - LpRdq is complete.
Definition - We define the completion Σµ of Σ with respect to µ by
Σµ “ tAYN | A P Σ, N P N u
Definition - Given A P Σµ, define µpAq “ µpBq if A “ B YN , where B P Σ and N P N .Remark - µ is well-defined. (We prove this below.)Lemma - A P Σµ if and only if there are F,G P Σ such that F Ď A Ď G and µpGzF q “ 0.Proof - Let A P Σµ, say A “ B YN where B P Σ, N P N . Since N is µ-null, there is M P Σ withM Ě N and µpMq “ 0. Let F “ B, G “ B YM . Then F Ď A Ď G, µpGzF q ď µpMq “ 0.
Conversely, let F,G P Σ with F Ď A Ď G, µpGzF q “ 0. Write A “ F Y pAzF q. Clearly AzF isµ-null, so A P Σµ.
Proposition - Let pX,Σ, µq be a measure space, pX,Σµ, µq its completion.
1. Σµ is a σ-algebra.
2. µ is a measure on Σµ and µ æΣ“ µ.
3. Σµ is complete with respect to µ.
Proof (1) Clearly H P Σµ. Given A P Σµ, the previous lemma implies there are F,G P Σ withF Ď A Ď G and µpGzF q “ 0. Now F c, Gc P Σ with Gc Ď Ac Ď F c and µpF czGcq “ µpGzF q “ 0.Finally, let pAnq Ď Σµ, say An “ Bn Y Nn, where Bn P Σ and Nn is µ-null. Set B “
Ť
nBn,
N “Ť
nNn. Now B P Σ and by subadditivity, N is µ-null. Thus
Ť
nAn P Σµ.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
(2) To see that µ is well-defined, let A P Σµ, say A “ B1YN1 “ B2YN2, where B1, B2 P Σ andN1, N2 are µ-null. Let Mi Ě Ni for each i, µpMiq “ 0. Now
µpB1q ď µpB2 YM2q ď µpB2q ` µpM2q “ µpB2q
and similarly, µpB2q ď µpB1q, so µpB1q “ µpB2q. Thus, µ is well-defined.Clearly µpHq “ 0. Now let pAnq Ď Σµ be disjoint, say An “ Bn YNn, where Bn P Σ and Nn is
µ-null. WriteŤ
nAn “ B YN , where B “
Ť
nBn P Σ and N “
Ť
nNn is µ-null. Now
µpď
n
Anq “ µpBq “ÿ
n
µpBnq “ÿ
n
µpAnq,
since the sets Bn are disjoint. Thus, µ is a measure, and clearly µ æΣ“ µ.(3) Let N be null with respect to Σµ, say A P Σµ with A Ě N , µpAq “ 0. Let F Ď A Ď G
with µpGzF q “ 0, F,G P Σ. Now µpF q “ µpAq “ µpGq “ 0. In particular, we have H Ď N Ď G,µpGzHq “ µpGq “ 0, so N P Σµ. Thus, Σµ is complete.Remark - Taking the completion does not simply correspond to throwing in all sets of measure zero!Example - µ “ λ, X “ Rd, Σ “ tX,Hu. Then Σµ “ tX,Hu.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 9 - 09/15/2014
Goal: Measurable Functions
Definition - Let pX,Σq be a measure space, pY, τq a topological space (usually R or r´8,8s). Wesay f : X Ñ Y is measurable if f´1pτq Ď Σ, i.e. the preimage of an open set is measurable.
(Remark - The topology on r´8,8s is generated by intervals r´8, aq, pb,8s.)Example - If X is a metric space, with Σ Ě BpXq, then continuous functions are measurable.
Proposition - Let BpY q be the Borel σ-algebra on Y . Then f : X Ñ Y is measurable if and only if
f´1pBpY qq Ď Σ.Proof p ðù q This is clear, since f´1pτq Ď f´1pBpY qq.p ùñ q We prove a useful lemma.
Lemma Let f : X Ñ Y be arbitrary. Then Σ1 “ tA Ď Y | f´1pAq P Σu is aσ-algebra.
Proof Exercise.
By the lemma, if f : X Ñ Y is measurable, then Σ1 is a σ-algebra containing all open sets, thusit contains BpY q.
Proposition - Let f : X Ñ r´8,8s. Then
f is measurable ðñ @a P R, f´1pr´8, aqq P Σ
ðñ @a P R, f´1ppa,8sq P Σ
ðñ @a P R, f´1pra,8sq P Σ
Proof Clearly measurability implies the other three conditions by the previous lemma. Since thehalf-infinite intervals generate the Borel σ-algebra, the reverse implication is clear, since Σ1 “ tA Ďr´8,8s ; f´1pAq P Σu is a σ-algebra.
Proposition - Say f : X Ñ Rd is measurable, g : Rd Ñ Rd is Borel measurable (i.e. g´1pRdq ĎBpRdq). Then g ˝ f is measurable.Proof We have
pg ˝ fq´1pBpRdq Ď f´1pg´1pBpRdqqq Ď f´1pBpRdq Ď Σ
Proposition - Let f, g : X Ñ r´8,8s be arbitrary. Then
f, g are measurable ðñ pf, gq : X Ñ r´8,8s2 is measurable
Proof - Supppose that f, g are measurable. Then for any open U, V Ď r´8,8s, we have
pf, gq´1pU ˆ V q “ f´1pUq X g´1pV q P Σ
In particular, pf, gq´1pIq P Σ for any open cell I Ď r´8,8s. Since the cells generate the Borelσ-algebra, then pf, gqpBpr´8,8s2q Ď Σ, so pf, gq is measurable.
Conversely, if pf, gq is measurable, then so are f “ π1 ˝ pf, gq and g “ π2 ˝ pf, gq (here, πi denotesprojection onto the ith coordinate, which is continuous).Corollary - Let f, g : X Ñ R be measurable. Then f ` g is measurable.Proof - Denote F px, yq “ x ` y, which is continuous, Gpxq “ pfpxq, gpxqq. Then f ` g “ F ˝ G,which is continuous by the previous propositions.
Proposition - Let pfnq be a sequence of measurable functions, fn : X Ñ r´8,8s. Then the followingfunctions are measurable.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
1. sup fn
2. lim sup fn
3. inf fn
4. lim inf fn
Moreover: (5) if fn Ñ f pointwise, then f is measurable.Proof - Denote g “ inf fn. Then tg ě αu “
Ť
ntfn ě αu, so since the fn’s are measurable, so is g.
Now sup fn “ ´ infp´fnq, so sup fn is measurable. Combining (1) and (3), we get (2) and (4). Noweither of (2) or (4) implies (5).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 10 - 09/17/2014
The Devil’s Staircase
Recall - The Cantor set C Ď r0, 1s is compact, has Hausdorff dimension α “ ln 2ln 3 , and Hα denotes
the Hausdorff measure of dimension α.
Definition - The Cantor function (also known as devil’s staircase) is defined by fpxq “ HαpCXr0,xsqHαpCq .
Intuitively, this is the fraction of the Cantor set that lies to the left of x.Remarks
1. f is continuous, since |fpyq ´ fpxq| ď Ci|x´ y|α for some Ci ě 0 and α P p0, 1q (exercise).
2. f is increasing (since Hα is a measure).
Set gpxq “ inftf “ xu “ infty | fpyq “ xu. Note, since f is continuous, then the infimum isattained, so fpgpxqq “ x. However, it’s not hard to see that g isn’t continuous since gp 1
2 q “13 , while
gpyq ě 23 for any y ą 1
2 . Note that
1. For all x P r0, 1s, gpxq P C (exercise).
2. g is strictly increasing, since x ă y ðñ fpgpxqq ă fpgpyqq ùñ gpxq ă gpyq.
3. g is Borel measurable, since
tgpxq ď αu “ tx | gpxq ď αu
“ tx | x “ fpgpxqq ď fpαqu pf is increasing)
“ tx | x ď fpαqu
Proposition - LpRq Ľ BpRq.Proof - Let A Ď r0, 1s, A R L (we know such A exists). Set B “ gpAq. Since B Ď C andλpCq “ 0, then B P L by completeness. On the other hand, B R B, since g´1pBq “ A R B andg is Borel measurable.
Remark - Let h1, h2 : RÑ R, h1 Lebesgue measurable and h2 Borel measurable. Then h1 ˝h2
need not be Lebesgue measurable.
Proof - Let h1 “ χB (B as above), h2 “ g. Then
h1 ˝ h2 “ χB ˝ g “ χg´1pBq “ χA,
which is not measurable since A is not measurable.
Definition - Let pΣ, X, µq be a measure space. We say property P holds almost everywhere(abbreviated a.e.) on some set A if there is a null set N such that P holds on AzN . We sayP holds a.e. if P holds a.e. on X.
Notation - Some say @x instead of a.e. x. We won’t use this notation.
Examples - Almost every real is irrational. The function fpxq “ |x| is differentiable a.e.
Theorem (Rademacher) - Any Lipschitz function f : RÑ R is differentiable almost everywhere.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Example - Continued Fractions - Given x P r0, 1s, one may write x “ 1
a1`1
a2`1...
, where ai P Z
is uniquely determined by x, say ai “ aipxq. Define pnpxqqnpxq
“ 1
a1`1
...1
an`1
in reduced terms. We
know that pnpxqqnpxq
Ñ x as nÑ8. Typically (when x is irrational), qnpxq Ñ 8.
Theorem - limnÑ8
ln qnpxqn exists and is equal to π2
12 lnp2q almost everywhere.
Proof - Comes from ergodic theory.
Proposition - Say pX,Σ, µq is the completion of pX,Σ, µq. Then
f : X Ñ R is Σµ-measurable ðñ Dg : X Ñ R, g is Σ-measurable, f “ g a.e.
Proof - p ðù q Given g, let U Ď R be open. Then
f´1pUq “ ptf “ gu X g´1pUqq Y ptf ‰ gu X f´1pUqq
Since the first set is in Σ and the second is null, then f´1pUq P Σµ.
p ùñ q If f “ χE , where E P Σµ, this is clear, since we may write E “ E1 YN , where E1 P Σand N is null. Now g “ χE1 does the trick. Since the result holds for indicator functions, itholds for simple functions.
The result for measurable functions will follow from a theorem next time about approximatingmeasurable functions by simple functions.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 11 - 09/19/2014
Approximation of Measurable Functions
Definition - We say s : X Ñ R is simple if it is Borel measurable and has finite range.
Definition - Given A P Σ, the characteristic function of A is χA, χApxq “
#
1 x P A
0 x R A.
Remark - If s has a finite range, then s has the form
s “ÿ
1ďiďN
ais´1ptsiuq,
where rangepsq “ ts1, . . . , sNu. Hence, s is simple if and only if s´1ptsiuq is measurable for each i.
Proposition 1 - If f : X Ñ r0,8q is measurable, then there exists an increasing sequence psnq ofsimple functions such that sn Õ f pointwise.Proof - Define Ak,n :“ f´1pr k2n ,
k`12n qq and let sn “
ř
1ďkďn2n
k2nχAk,n ` χtfěnu. Check that this
works.
Proposition 2 - If f : X Ñ R is measurable, then there exists a sequence psnq of simple functionssuch that |sn| ď f for all n and sn Ñ f pointwise.Proof - Denote f` “ maxpf, 0q, f´ “ maxp´f, 0q, so f “ f`´f´. Find psnq and ptnq with sn Õ f`
and tn Õ f´, so now sn ´ tn Ñ f . Moreover,
|sn ´ tn| ď |sn| ` |tn| ď f` ` f´ “ |f |
Theorem (Lusin’s Theorem) - Let X be a metric space, µ a finite regular measure. If f : X Ñ R isµ-measurable, then for every ε ą 0, there is g : X Ñ R continuous such that µptf ‰ guq ă ε.Lemma 1 - For each ε ą 0, there is closed C Ď X such that µpXzCq ă ε and f æC : C Ñ R iscontinuous.Lemma 2 (Tietze Extension Theorem: Easy Version) - Let C Ď X be closed, h : C Ñ R continuous.Then there is continuous H : X Ñ R such that H æC“ h.Proof (Lusin’s Theorem) - Follows from Lemmas 1 and 2.
Proof (Lemma 2) - Define Hpxq “
#
hpxq x P C
infcPCthpcq ` dpx,cq
dpx,Cq ´ 1u x R C. Not hard to check that this is
continuous. Note, this only works if h is bounded below on C. If h is unbounded, then considerarctanphq instead.
Proof (Lemma 1) - Case I: Suppose that f is bounded. Without loss of generality, f : X Ñ r0, 1s.
Set Ak,n “ f´1pr k2n ,k`12n qq, where 0 ď k ď 2n. For all k, n, there exist compact sets Kk,n Ď Ak,n
such thatµpAk,nzKk,nq ă
ε2np2n`1q
Now set Cn “Ť
0ďkď2nKk,n. Note that by construction, we have µpCcnq ă
ε2n by construction, since
µpCcnq ďř
k
µpAk,nzKk,nq. Now set C “Ş
nCn.
Claim This set C works.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Proof Clearly C is closed. Moreover, µpCcq ďř
nµpCcnq “ ε. To see that f æC
is continuous, denote sn “ř
k
k2nχAk,n . Note that sn : Cn Ñ R is continuous (by
the pasting lemma for continuous functions) since Cn “Ť
K
Kk,n and sn æKk,n is
constant. Since C Ď Cn, then sn : C Ñ R is continuous.
On C, |f ´ sn| ď1
2n , so sn Ñ f uniformly. Thus, f is continuous on C.
Remark - This theorem is still true if X “ Rd, µ “ λ because we only need the sets Kk,n to be closed(not necessarily compact) such that λpAk,nzKk,nq ă
ε2np2n`1q and the rest of the proof is identical.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 12 - 09/22/2014
Integration
Definition - Let s : X Ñ R is simple, measurable, and nonnegative, say s “ř
i
aiχAi , where
Ai “ s´1paiq. Defineż
X
s dµ “ÿ
i
aiµpAiq
with the convention 0 ¨ 8 “ 0.Definition - Given f : X Ñ r0,8s measurable, define
ż
X
f dµ :“ sup0ďsďf
s simple meas.
ż
X
s dµ
Definition - Given f : X Ñ r´8,8s measurable, define f` “ maxpf, 0q, f´ “ maxp´f, 0q (note:these are measurable functions since x ÞÑ |x| is continuous and continuous˝measurable is measur-able).
1. Ifş
Xf` dµ,
ş
Xf´ dµ ă 8, we say f is integrable (or L1) and define
ż
X
f dµ “
ż
X
f` dµ´
ż
X
f´ dµ
2. If eitherş
Xf` dµ or
ş
Xf´ dµ is finite, we say f is integrable in the extended sense and defineż
X
f dµ “
ż
X
f` dµ´
ż
X
f´ dµ P r´8,8s
Notation - Depending on whether the measure or measure space are clear from context, we willsometimes write
ż
X
f dµ “
ż
f dµ,
ż
X
f “
ż
f
Example 1 - Let X “ N, µ the counting measure. All sets are measurable, so all functions f : X Ñ Rare measurable. Note a function a : N Ñ R is just a sequence of real numbers. What is
ş
N a dµ?Well, we’d like to say it’s just
ř
nan. However, since the integral is defined as the positive part minus
the negative part, we can only conclude thatş
N a dµ “ř
nan in the case that
ř
|an| ă 8.
Remark - The above example illustrates that we’ve somehow lost cancellation. Even a conditionally
convergent sum (such asř
n
p´1qn
n won’t haveş
N a dµ defined, since we split up a “ a` ´ a´.
Proposition 1 - Let s ě 0 be simple and measurable with spXq “ ta1, . . . , anu, Ai “ s´1paiq. Then
sup0ďtďs
t simple meas.
ż
X
t dµ “ÿ
i
aiµpAiq
Proof - The left-hand side is clearly at least as large as the right-hand side. For the reverse inequality,it’s enough to prove the following lemma.
Lemma Let s, t be simple and measurable with 0 ď s ď t. Thenż
X
s dµ ď
ż
X
t dµ
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Proof Exercise.
Proposition 2 - If f, g are integrable, thenş
Xpf ` gq dµ “
ş
Xf dµ`
ş
Xg dµ.
Proof - We can’t do this yet! Let’s see what goes wrong. Start by assuming that f, g ě 0. Let0 ď s ď f , 0 ď t ď g, so that s` t ď f ` g. Now it’s not hard to show that
ż
X
s dµ`
ż
X
t dµ “
ż
X
ps` tq dµ ď
ż
X
pf ` gq dµ
Taking the supremum over 0 ď s ď f , 0 ď t ď g, we have
ż
X
f dµ`
ż
X
g dµ ď
ż
X
pf ` gq dµ
The problem is that we can’t easily get the reverse! What to do? (To be continued)Remark - To finish this off, we need some sort of limit theorem to take care of the sup in the definitionof
ş
Xf dµ.
Theorem (Lebesgue’s Monotone Convergence Theorem) - Say fn : X Ñ r0,8s are measurable withfn ď fn`1. Denote f “ lim fn (this exists because f is monotone). Then
lim
ż
X
fn dµ “
ż
X
f dµ
Proof - For all n, we haveş
fn ďş
fn`1 ďş
f (this follows from monotonicity of the integral. Inparticular, lim
ş
fn exists and limş
fn ďş
f .We need to show that lim
ş
fn ěş
f . To do this, let 0 ď s ď f simple, measurable. We wantlim
ş
fn ěş
s. First, we take a naıve approach (which we’ll fix in a moment). Write spXq “ta1, . . . , aNu, Ai “ s´1paiq. Set En “ tfn ě su. Then pEnq is increasing and now
ż
X
fn dµ ě
ż
X
χEnf dµ ě
ż
X
χEns dµ “Nÿ
i“1
aiµpAi X Enq
We’d like to take the limit n Ñ 8 and sy µpAi X Enq Ñ µpAiq to get the desired result. Theproblem is that this only works if
Ť
nEn “ X. This may not happen if the sequence pfnq doesn’t
cross s everywhere.The solution is to instead let 0 ă ε ă 1 and set En “ tfn ě p1 ´ εqsu. Now
Ť
nEn “ X. We
modify the above inequality and get
ż
X
fn dµ ě
ż
X
χEnf dµ ě
ż
X
χEnp1´ εqs dµ “ p1´ εqNÿ
i“1
aiµpAi X Enq
Taking the limit n Ñ 8, we have limş
fn ě p1 ´ εqş
s. Take ε Ñ 0` to get limş
fn ěş
s, asdesired.
Proposition (Linearity) - If f, g are integrable, then f ` g is integrable andş
pf ` gq “ş
f `ş
g.Proof - We do this in four steps.
(1) Let s, t ě 0 be simple and integrable. Thenş
ps` tq “ş
s`ş
t (exercise).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
(2) Now let f, g ě 0 be measurable and integrable. Wantş
pf`gq “ş
f`ş
g. Last time, we showedthat there exist increasing sequences psnq, ptnq of nonnegative measurable simple functions withsn Ñ f and tn Ñ g pointwise. By MCT,
lim
ż
sn “ f, lim
ż
tn “ g ùñ
ż
pf ` gq “ lim
ż
psn ` tnq “ limp
ż
sn `
ż
tnq “
ż
f `
ż
g
(3) Lemma - Let f be integrable, f “ g ´ h, where g, h ě 0 are integrable. Then
ż
f “
ż
g ´
ż
h
Proof - Write f “ f` ´ f´ “ g ´ h. Now
f` ` h “ g ` f´ ùñ
ż
f` `
ż
h “
ż
g `
ż
f´ ùñ
ż
f “
ż
pf` ´ f´q “
ż
g ´
ż
h
(4) Now let f, g be arbitrary and integrable, say f “ f` ´ f´, g “ g` ´ g´. Now
pf ` gq “ pf` ` g`q ´ pf´ ` g´q
Applying the lemma from (3), we have
ż
pf ` gq “
ż
pf` ` g`q ´
ż
pf´ ` g´q “ p
ż
f` ´
ż
f´q ` p
ż
g` ´
ż
g´q “
ż
f `
ż
g
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 13 - 09/24/2014
Last time, we defined the integral, proved the Monotone Convergence Theorem, and used it to provelinearity.
Quote - ”The selling point of the Lebesgue integral is the interchangeability ofş
and lim.”
Remark - Note that the MCT fails if we don’t require fn ě 0. Even if we assume fn ě ´1, forexample, consider the case X “ R, fn “ ´
1n . Now
ş
fn “ ´8 for all n, butş
lim fn “ 0.
Goal - If fn ě 0 and lim fn exists, want to say limş
fn “ş
lim fn.
Two Counterexamples
1. Mass escaping to infinity: Set fn “ χrn,n`1q. Then
lim
ż
fn “ 1 ‰ 0 “
ż
lim fn
2. Mass collects at a point: Set fn “ nχp0,
1n s
. Then
lim
ż
fn “ 1 ‰ 0 “
ż
lim fn
To see what goes wrong here, consider the following theorem.
Theorem (Lebesgue Dominated Convergence Theorem) - Suppose that
1. fn : X Ñ r0,8s is measurable.
2. fn Ñ f pointwise.
3. There is integrable F such that |fn| ď F for all n.
Then limş
fn “ş
f .Remark - This addresses the counterexamples above. In the first example, the ’envelope’ of pfnqis the function fpxq “ 1, which is not integrable. In the second example, the envelope of pfnq isfpxq “ 1
x , also not integrable.
To prove LDCT, we first need a lemma.
Fatou’s Lemma - Let fn ě 0 be measurable, fn Ñ f pointwise. Then
lim inf
ż
fn ě
ż
f
Quote - ”Mass can escape, but cannot be created.”Proof - Set gk “ inftfn | k ě nu. Note that
gk Ñ f and pgkq is increasing, nonnnegative, so by MCT,
lim
ż
gk ě
ż
f
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Now fn ě gn for all n, soş
fn ěş
gn by monotonicity. Thus,
lim inf
ż
fn ě lim inf
ż
gn “ lim
ż
gn “
ż
f
Remark - We cannot conclude that limş
fn ěş
f , since limş
fn may not exist.
Proof (LDCT) - Set gn “ f ` fn. Now gn ě 0, so Fatou’s lemma implies
ż
F ` lim inf
ż
fn “ lim inf
ż
gn ě
ż
pF ` fq “
ż
F `
ż
f
Cancellingş
F , we have lim infş
fn ěş
f .Now set hn “ F ´ fn. Since hn ě 0, Fatou’s lemma implies
ż
F ´ lim sup
ż
fn “ lim inf
ż
hn ě
ż
pF ´ fq “
ż
F ´
ż
f
Rearranging, we have lim supş
fn ďş
f . Now
lim sup
ż
fn ď
ż
f ď lim inf
ż
fn ùñ lim
ż
fn “
ż
f
Quote - ”The only work we’ve done to prove this was in the Monotone Convergence Theorem, andeven that was only a few lines. The key is that we’ved coded all these results into the definition ofmeasure.”
Application - The Laplace Transform
Notation - We denoteş
Xfpxq dµpxq “
ş
Xf dµ. For example,
ş1
0x2 dλpxq is the integral of x2 with
respect to λ on r0, 1s. We denote integration with respect to λ by dλpxq “ dx. Why do we do this?IOU - If f : r0, 1s Ñ R is Riemann integrable, then f is L-integrable. Moreover, the Riemannintegral and Lebesgue integral agree with one another.Remark - An improper Riemann integral may disagree with the Lebesgue integral, since the Lebesgueintegral doesn’t allow for cancellation of an infinite positive part and an infinite negative part.
Suppose now that f : r0,8q Ñ R is integrable. Define the Laplace transform of f by
F psq “
ż 8
0
e´stfptq dt
Note: F : p0,8q Ñ R. Here’s an example of where Lebesgue integration is helpful.
Proposition
1. Ifş8
0|f | ă 8, then F is continuous.
2. Ifş8
0t|fptq| ă 8, then F is differentiable (in fact, C1).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Proof - For (1), given s0 P p0,8q, we have limsÑs0
“ limsÑs0
ş8
0e´stfptq dt. We want to apply LDCT. To
get F ps0q on the right-hand side, let psnq be a sequence with sn Ñ s0. Note that |e´sntfptq| ď |fptq|.Since
ş8
0|fptq| dt ă 8, then by LDCT, we have
limnÑ8
F psnq “ F ps0q
To prove (2), write F psq´F ps0qs´s0
“ş8
0e´st´e´s0t
s´s0fptq dt. For any s ‰ s0, the Mean Value Theorem
implies that there is s1 such thate´st´es0t
s´s0“ ´te´s1t
Thus, | e´st´es0t
s´s0| ď t|fptq|, so since
ş8
0t|fptq| dt ă 8, then the LDCT implies that
limsÑs0
F psq´F ps0qs´s0
“
ż 8
0
´te´stfptq dt “ Lptfptqq
By (1), Lptfptqq is continuous, so we are done.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 14 - 09/26/2014
We begin with an application of MCT.
Example (Beppo Levi’s Theorem) - Let pfnq be a sequence of nonnegative measurable functions.Then
ż
X
ÿ
n
fn dµ “ÿ
n
ż
X
fn dµ
Proof - Set sN “ř
1ďnďN
fn. Note, 0 ď sn ď sn`1. Now apply the MCT.
Example - Let φ : RÑ r0,8s withş
R φ ă 8. Define
fpxq “ÿ
mPN,nPZ`φppx´ m
n q2m`nq
Note that
1. f “ 8 at all rationals if φp0q ą 0.
2. f ă 8 a.e. To see this, note that Beppo Levi impliesż
Rf “
ÿ
m,n
ż
Rφppx´ m
n q2m`nq “
ÿ
m,n
2´pm`nqż
Rφ ă 8
In particular, f ă 8 a.e.
We’ll spend most of the next week thoroughly studying convergence. Before that, let’s mentiona few things.
Pushforward Measures
Definition - Let pX,Σ, µq be a measure space, f : X Ñ Y arbitrary. Define Σ1 Ď PpY q by
Σ1 “ tA | f´1pAq P Σu
Define ν on Σ1 by νpAq “ µpf´1pAqq.Proposition - ν is a measure on pY,Σ1q. Moreover, if g : Y Ñ r´8,8s is integrable, then
ż
Y
g dν “
ż
X
g ˝ g dµ
Proof - We’ve already seen that Σ1 is a σ-algebra. Moreover, it’s easy to check that ν is a measure.To prove the statement about integrability, it’s enough to show it for simple functions by the MCTand linearity. In fact, we may consider s “ χA, where A P Σ1. Now
ż
Y
s dν “ νpAq “ µpf´1pAqq “
ż
X
χf´1pAq dµ “
ż
X
s ˝ f dµ
Nowş
Yg dν “
ş
Xg ˝ f dµ for all g positive and simple. Thus, this is true for any integrable g.
Application - Fix x0 P Rd, f : Rd Ñ Rd integrable. Then
ż
Rdfpxq dx “
ż
Rdfpx` x0q dx
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Proof - Let gpxq “ x`x0. Equip the second Rd with the pushforward measure µ induced by pRd, λq.Now µpAq “ λpg´1pAqq “ λpAq, so by translation invariance, µ “ λ. The previous proposition nowyields
ż
Rdfpxq dλpxq “
ż
Rdfpxq dµpxq “
ż
Rdfpx` x0q dλpxq
It’s time for one last remark about integration.
Proposition - Let pX,Σ, µq be complete and let f : X Ñ R be integrable, f “ g a.e. Then g isintegrable and
ş
f “ş
g.Proof - We already know g is measurable. To prove that g is integrable, set h “ f ´ g. Then h “ 0a.e. and h is measurable.
Claimş
h “ 0.
Proof Decompose h “ h` ´ h´. We’ll show thatş
Xh` “
ş
Xh´ “ 0. To do this,
let 0 ď s ď h` be simple, say s “ř
aiχAi , where the sets Ai are disjoint andmeasurable. If ai ‰ 0, then Ai Ď th
` ą 0u, so µpAiq “ 0. Thus,ş
Xs dµ “ 0, so
nowş
Xh` dµ “ 0. Similarly,
ş
Xh´ “ 0. This finishes the proof.
Now h integrable implies f ´ h “ g integrable, andş
g “ş
pf ´ hq “ş
f .Remark - In light of this remark, the MCT and LDCT (for Rd) are still true if we replace fn Ñ fwith fn Ñ f a.e.
Convergence
Theorem (Egorov’s Theorem) - Let pX,Σ, µq be a measure space, µpXq ă 8. Let pfnq be a sequenceof measurable functions with fn Ñ f pointwise a.e. Then for any ε ą 0, there is A Ď X such thatµpXzAq ă ε and fn Ñ f uniformly on A.Remark - In general, we can’t do better, i.e. we can’t get µpAcq “ 0.Proof (of theorem) - For each k, there is n0 such that n ą n0 implies |fn ´ f | ă 1
k a.e. This isbecause
µpď
m
č
něm
t|fn ´ f | ă1k uq “ µpXq
In particular, for each k there is nk such that
µpč
něnk
t|fn ´ f | ă1k uq ě µpXq ´ ε
2k
Now set A “Ş
k
Ak. Then
• µpAq ě µpXq ´ ε.
• Fix k P N. For all a P A, we have a P Ak, so |fnpaq ´ fpaq| ă 1k for all n ě nk. Since k was
arbitrary, then fn Ñ f uniformly on A.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 15 - 09/29/2014
Definition (a.e. pointwise convergence) We say fn Ñ f µ-a.e. if there exists a µ-null set N P Σ suchthat fn Ñ f pointwise on N c, i.e. if tfn Û fu is null.
Definition (convergence in measure) We say fn Ñ f in µ-measure (and write fnµÑ f) if for every
ε ą 0, we havelimnÑ8
µptx P X | |fn ´ f | ą εuq “ 0
Definition (Lp convergence) We say fn Ñ f in L1 ifş
X|fn ´ f | dµÑ 0 as nÑ 8. More generally,
for p ě 1, we say fn Ñ f in Lp (and write fnLpÑ f) if
limnÑ8
ż
|fn ´ f |p “ 0
Later we will define a similar notion for p “ 8.
Intuition - Think of Lp as convergence in a metric space, where the metric is
dpf, gq “ p
ż
X
|f ´ g|p dµq1p
(We need p ě 1, otherwise the triangle inequality fails.)
Example - To see that fn Ñ f a.e. does not imply fnµÑ f , consider fn “ χrn,n`1s. Then fn Ñ 0
a.e. but µpt|fn ´ 0| ą εuq “ 1 for all n and 0 ă ε ă 1.
Proposition 1 - If µpXq ă 8 and fn Ñ f a.e., then fn Ñ f in measure.Proof - Let ε ą 0. Since fn Ñ f a.e., then χt|fn´f |ąεu Ñ 0 a.e. Moreover, χt|fn´f |ąεu ď 1, which isintegrable since µpXq ă 8. By LDCT, we have
limnÑ8
µpt|fn ´ f | ą εuq “ limnÑ8
ż
χt|fn´f |ąεu “ 0
Remark - This can also be proved using Egorov’s theorem.
Question - Does the converse hold? I.e. does fnµÑ f imply fn Ñ f a.e.?
Answer - No! For example, f1 “ χr0,
12 q
, f2 “ χr12 ,1q
, f3 “ χr0,
14 q
, f4 “ χr14 ,
12 q
, etc. Set f “ 0.
Then clearly µpt|fn ´ f | ą εuq Ñ 0 for all ε ą 0, so fn Ñ f in measure. However, for all x P r0, 1q,fnpxq “ 0 infinitely often and fnpxq “ 1 infinitely often, so fn fails to converge a.e.
Proposition 2 - Suppose fn Ñ f in measure. Then there is a subsequence pfnkq Ď pfnq with fnk Ñ fa.e.Proof - Let n1 be such that µpt|fn ´ f | ą 1uq ă 1
2 for n ě n1. Let n2 ą n1 such that µpt|fn ´ f | ą12uq ă
14 . Repeat this process, choosing nk ą nk´1 such that µpt|fnk ´ f | ą
1k uq ă
12k
.
Claim fnk Ñ f a.e.
Proof Let Ak “ t|fnk ´ f | ą 1k u, A “ tx | x P Ak only finitely many timesu. By
definition, if x P A, then fnkpxq Ñ fpxq. Now it suffices to show that µpAcq “ 0.Write
Ac “ tx | x P Ak infinitely oftenu “č
kě1
ď
něk
An
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Now for any k,µpAcq ď µp
ď
něk
Anq ďÿ
něk
12n “
12k´1
Let k Ñ8, then µpAcq “ 0.
Now let’s investigate Lp convergence.
Definition - Let V a vector space. We say ¨ is a norm on V if
1. v ě 0 for all v P V and v “ 0 if and only if v “ 0.
2. For all α P R, αv “ |α|v.
3. u` v ď u ` v for all u, v P V (triangle inequality).
Given a norm ¨ on V , define d : V ˆ V Ñ R by dpu, vq “ u´ v. It’s not hard to check that d isa metric, so now pV, dq is a metric space.
This clearly satsfies properties (2) and (3) of norms, but f1 “ 0 only implies that f “ 0 a.e., notthat f “ 0. This motivates the following definition.Definition - Define an equivalence relation „ on L1 by f „ g if f “ g a.e. Set L1 “ L1 „. Now forF P L1, F represented by some function f , define F 1 “ f1. This is well-defined by definition of„.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 16 - 10/01/2014
Recall - In Lp spaces, convergence fnLpÑ f is defined as lim
nÑ8
ş
X|fn ´ f |
p dµ “ 0.
Exercise - |f | ď f8 a.e.
Henceforth, assume that pX,Σ, µq is complete.Definition - LppX,µq :“ tf : X Ñ r´8,8s | fp ă 8u.
Note that fp “ 0 only implies f “ 0 a.e., not f “ 0. To make ¨p a norm, define an equivalencerelation „ on Lp by f „ g if and only if f “ g a.e. Now define LppX,µq “ LppX,µq „. For F P Lp,define F p “ fp, where f is any representative of F . This is well-defined by definition of „.
Given F,G P LppX,µq, define F `G “ rf ` gs, where F “ f „, G “ g „. This is well-defined.Remark - We can’t meaningfully talk about F pxq since f „ g doesn’t imply fpxq “ gpxq. We canmeaningfully define
ş
AF , tF ą au, etc. In general, we treat elements of Lp as functions, only doing
operations that are constant on equivalence classes.
Goal 1 - Lp is a Banach space.Definition - Let pV, ¨ q be a vector space, dpu, vq “ u´ v. Now pV, dq is a metric space. We saypV, ¨ q is a Banach space if pV, dq is complete.E.g. Rd is a Banach space (with the standard norm).
Theorem For all p P r1,8s, LppX,µq is a Banach space.Proof - We only need to show
1. f ` gp ď fp ` gp
2. Completeness
The rest is easy. IOU this proof.
Lemma (Holder’s Inequality) - Let p, q P r1,8s such that 1p `
1q “ 1. If f P Lp, g P Lq, then
|ş
fg| ď fpgq. In particular, fg P L1.Remark - p “ q “ 2 is Cauchy-Schwarz.Intuition - The relationship between p and q follows by ”counting dimensions”. Say X “ Rd, f, gare dimensionless quantities. Then
ş
Rd fg has dimension Ld, fp has dimension Ldp, and gq has
dimension Ldq. We want Ld “ LdpLdq, which holds if and only if 1p `
1q “ 1.
To formalize this argument (see homework), replace fpxq by fpλxq. Do a change of variablesand then take λ large and λ small to see that the inequality only holds when 1
p `1q “ 1.
Induction Proof (Holder) - The main idea is to show that for real numbers xi, yi, i “ 1, . . . , N , we
have |ř
i
xiyi| ď př
i
|xi|pq1pp
ř
i
|yi|qq1q. Now let s, t be simple functions and write s “
ř
i
xiχEi ,
t “ř
i
yiχEi , where xi, yi P R and the sets tEiu are disjoint. Writing µpEiq “ µpEiq1pµpEiq
1q, we
havest1 “ |
ÿ
i
µpEiqxiyi| ď |ÿ
i
|xi|pµpEiq|
1p|ÿ
i
|yi|qµpEiq|
1q “ sptq
Now the result holds for simple functions, so use approximation and monotone convergence to finishthis off.
Better Proof - First assume that p, q P p1,8q. The proof starts with Young’s Inequality, which is asfollows.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Claim (Young’s Inequality) Let x, y ě 0 and let p, q P p1,8q, 1p `
1q “ 1. Then
xy ď xp
p `yq
q .
Proof This follows from convexity of the function hpzq “ lnpzq, since
lnpxyq “ lnpxpqp `
lnpxqqq ď lnpx
p
p `yq
q q
To prove Holder, first note that if fp “ 0 or gq “ 0, then f “ 0 or g “ 0, so we are done.
Assume now that fp, gq ‰ 0. Set F “ ffp
, G “ ggq
, so that F p “ Gq “ 1. By monotonicity
and Young’s inequality, we have
|
ż
FG| ď
ż
|F ||G| ďş
|F |p
p `
ş
|G|q
q “ 1
Multiplying by fpgq, we have |ş
fg| ď fpgq.The case p “ 1, q “ 8 is left as an exercise.
Lemma (Duality) - Let p P r1,8q, f P Lp, 1p `
1q “ 1. Then
fp “ supgPLqzt0u
1gq
ż
fg p˚q
Remark - If p “ 8, this is still true provided that X is σ-finite.Proof - By Holder, for any g P Lqzt0u, we have 1
gq
ş
fg ď fp. Thus, LHS ě RHS in p˚q. For
the reverse inequality, set F “ ffp
. It’s enough to show that supGq“1
ş
FG “ 1. To get this, choose
G “ |F |p´1signpF q. Clearlyş
FG “ş
|F |p “ 1. Moreover,
ż
|G|q “
ż
|F |pp´1qq “
ż
|F |pq´q “
ż
|F |p “ 1
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 17 - 10/03/2014
IOU - Lp is a Banach space. We still need the triangle inequality and completeness. We’ll prove thefirst today.
Last time, we saw the following:
1. If p, q P r1,8s, 1p `
1q “ 1, then |
ş
fg| ď fpgq.
2. We also have a duality principle: Given p P r1,8q, f P Lp, we have
fp “ supqPLqzt0u
1gq
ż
fg
3. The previous statement holds for p “ 8 if X is σ-finite.
Proposition - Let f, g P Lp. Then f ` g P Lp and
f ` gp ď fp ` gp
Proof - If p “ 8, this is easy. So assume p P r1,8q. The function x ÞÑ |x|p is convex, so
12p |f ` g|
p “ | 12f `12g|
p ď 12 |f |
p ` 12 |g|
p
Since the right-hand side is integrable, so is the left-hand side, thus f ` g P Lp.For the triangle inequality, let q be such that 1
p `1q “ 1. By duality, we have
f ` gp “ suphPLqzt0u
1hq
ż
pf ` gqh
ď suphPLqzt0u
1hq
ż
fh` suphPLqzt0u
1hq
ż
gh
“ fp ` gp
Quote - ”Duality gets you an amazing amount of mileage.”
Before we prove completeness, let’s discuss another useful inequality. On your homework, you’llshow that this can be used to prove Holder’s inequality. First recall that if φ : R Ñ R is convex,then given a finite set of points ci P r0, 1s with
ř
i
ci “ 1, we have φpř
i
cixiq ďř
i
ciφpxiq. Loosely
speaking, Jensen’s inequality extends this result to integrals.Jensen’s Inequality - Let φ be a convex real-valued function, µpXq “ 1. Then for any f P L1pXq,we have
φp
ż
X
f dµq ď
ż
X
φ ˝ f dµ
Remark - If φ is convex, then φ is continuous. In particular, φ is Borel, so φ˝f is measurable. (Thisjustifies well-definedness of the expression on the right.)
Completeness of Lp
Lemma - Let p P r1,8q. Say pfnq Ď Lp such thatř
nfnp ă 8. Then
1.ř
nfn converges pointwise a.e., say to f .
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
2. f P Lp andř
nfn converges in Lp to f .
Proof - Define sN “ř
1ďnďN
fn, tN “ř
1ďnďN
|fn|. For (1), we must show that sN converges and is
finite a.e., for which it’s enough to show that tN converges and is finite a.e.Since the functions ptnq are increasing and nonnegative, we can set F “
ř
n|fn| “ lim
NÑ8tN . By
the triangle inequality, we have
tN p ďÿ
1ďnďN
fnp ďÿ
n
fnp ă 8
Nowş
tpN “ tN pp ď p
ř
nfnpq
p ă 8, so by the MCT, we have
limNÑ8
ż
tpN “
ż
F p ă 8
Thus F P Lp. In particular, F is finite a.e., so now (1) holds.For (2), note that |f | ď F implies fp ď F p, so now f P Lp. Moreover, by the MCT and the
triangle inequalities for Lp and R, we have
f ´ sN p “ ÿ
něN`1
fnp ď ÿ
něN`1
|fn|p “ limmÑ8
ÿ
měněN`1
|fn|p
ď limmÑ8
ÿ
měněN`1
fnp
“ÿ
něN`1
fnp Ñ 0 as N Ñ8
Thus, sN Ñ f in Lp.
Proposition - Lp is complete.Proof - Let pfnq Ď Lp be Cauchy. It’s enough to find a subsequence pfnkq of pfnq such that pfnkqconverges in Lp. To do this, choose n1 ă n2 ă ¨ ¨ ¨ such that fnk`1
´ fnkp ă12k
. By the lemma,since
ř
k
fnk`1´ fnkp ă 1, then
ř
k
pfnk`1´ fnkq converges both a.e. and in Lp. Set
f “ÿ
k
pfnk`1´ fnkq ` fn1
Let the sum telescope, so that fnkLpÑ f .
Exercise - Prove that L8 is complete.
Convergence - We have convergence a.e., convergence in measure, convergence in Lp. What’s therelationship between these notions?
• We know a.e. convergence implies convergence in measure if µpXq ă 8 (but not in general).
• We know convergence in measure implies a.e. convergence along a subsequence.
• In general, a.e. convergence does not imply convergence in Lp, even on finite measure spaces.For example, X “ p0, 1q, fn “ 2nχ
p0,1n q
.
• However, convergence in Lp does imply a.e. convergence along a subsequence (similar to seconditem above).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
• Convergence in Lp implies convergence in measure (next time, using Chebyshev).
• Convergence in measure does not imply convergence in Lp (next time).
Lemma (Chebyshev’s Inequality) - Let f be meaurable. Then for all λ ą 0,
µp|f | ą λq ď 1λf1
Proof - Clearly λχ|f |ąλ ď |f |, so integrate both sides and we’re done.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 18 - 10/06/2014
Proposition - pfnqLpÑ f , p P r1,8q implies fn
µÑ f .
Proof - Use Chebyshev. We have
µpt|fn ´ f | ą εuq “ µpt|fn ´ f |p ą εpuq ď 1
εp fn ´ fpp Ñ 0 as nÑ8
IOU - fnµÑ f + ”mystery assumption” is equivalent to fn
LpÑ f .
Proposition 1 - S “ ts | s simple, µpts ‰ 0uq ă 8u is dense in Lp for all p P r1,8s.Proof - Let f P Lp. If p P r1,8q, let psnq be sequence of simple functions with sn Ñ f a.e. and|sn| ď f a.e. Note, f P Lp implies sn P L
p, so then µptsn ‰ 0uq ă 8 for each n. Now by the LDCT,
sn ´ fpp “
ż
|sn ´ f |p Ñ 0
since |sn ´ f |p ď p|f | ` |f |qp “ 2p|f |p, which is integrable.
If p “ 8, then partition r´f8, f8s into finitely many subintervals tAiu of length at most ε,then set s “
ř
i
xiχf´1pAiq, where xi is the midpoint of Ai. Now |f ´ s| ă ε a.e., so f ´ s8 ď ε.
Proposition 2 - Let X be a σ-compact metric space (i.e. X “Ť
nKn, where the sets pKnq are
increasing and compact). If µ is regular on X, then CcpXq is dense in Lp.Remark/Definition - CcpXq “ tcontinuous functions on X with compact supportu.Proof - First assume that X is compact. Let f P Lp and fix ε ą 0. For each n ě 1, define
fn “
$
’
&
’
%
f |f | ď n
n f ą n
´n f ă ´n
(Think of this as a truncated version of f .) By LDCT, fnLpÑ f since fn ď |f |. Let N sufficiently
large so that fN ´ fp ăε2 . Choose δ ą 0 sufficiently small so that δp2nqp ă p ε2 q
p. By Lusin’stheorem, there is continuous gn such that µptfn ‰ gnuq ă δ. By truncating if necessary, we mayassume |gn| ď n, so now gn has compact support. Moreover,
fn ´ gnpp “
ż
tfn‰gnu
|fn ´ gn|p ď δp2nqp ă p ε2 q
p
Now gn ´ fp ď gn ´ fnp ` fn ´ fp ăε2 `
ε2 “ ε, so we’re done.
The case that X is σ-compact is left as an exercise.
Proposition 3 - If Σ “ σpCq, where C is countable, then LppXq is separable for p P r1,8q.Proof - Homework.
Theorem (Vitali) - Let f P L1pXq, pfnq Ď L1pXq. Then fnL1
Ñ f if and only if
1. fnµÑ f
2. For any ε ą 0, there is δ ą 0 such that µpAq ă δ impliesş
A|fn| ă ε for all n.
3. For any ε ą 0, there is F P Σ such that µpF q ă 8 andş
F c|fn| ă ε for all n.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Quotes ”Condition (2) prevents mass collecting at a point. Condition (3) prevents mass escaping toinfinity.”Remark - Condition (2) is sometimes referred to as equi-integrability. Condition (3) is sometimesreferred to as tightness.Proof (of Vitali) - Suppose that fn Ñ f in L1. We already have fn
µÑ f .
To show equi-integrability, let n0 be such that n ą n0 impliesş
X|f ´ fn| ă
ε2 . Now for any set
A Ď X, we haveż
A
|fn| ď
ż
A
|f ´ fn| `
ż
A
|f | ă
ż
A
|f | ` ε2
If µpAq ă δ, then for any λ ą 0,
ż
A
|f | “
ż
AXt|f |ďλu
|f | `
ż
AXt|f |ąλu
|f | ď λδ `
ż
t|f |ąλu
|f |
By LDCT, we have limλÑ8
ş
t|f |ąλu|f | “ lim
λÑ8
ş
Xχt|f |ąλu|f | “ 0. Now let λ be sufficiently large such
thatş
t|f |ąλu|f | ă ε
4 and let δ ą 0 be sufficiently small such that λδ ă ε4 . Now
µpAq ă δ, n ą n0 ùñ
ż
A
|fn| ă ε
To get the result for n ď n0, repeat the above argument (the one used to minimizeş
A|f |) for the
finitely many functions fn, n ď n0.Next time, we’ll prove that pfnq is tight and prove the converse of the theorem.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 19 - 10/08/2014
Proof (of Vitali, continued) - We need to show that fnL1
Ñ f implies pfnq is tight. Fix ε ą 0. For anyF , we have
ż
F
|fn| ď
ż
F
|f | `
ż
F
|fn ´ f | ă
ż
F
|f | ` ε2
for n sufficiently large, say n ě n0 (where n0 is independent of F ).
Set F “ t|f | ą λu (we’ll choose λ later). By Chebyshev’s inequality, we have µpF q ď f1λ ă 8.
Nowş
F c|f | “
ş
|f |χt|f |ďλu. As λÑ 0`,ş
|f |χt|f |ďλu Ñ 0 by LDCT, so let λ be sufficiently small sothat
ş
|f |χt|f |ďλu ăε2 . Repeat the same trick for the finitely many functions f1, . . . , fn0
, then we’redone.
Conversely, suppose that fnµÑ f , pfnq is equi-integrable, and pfnq is tight. Fix ε ą 0. Then for
any set F and λ ą 0, we have
ż
|fn ´ f | “
ż
F
|fn ´ f | `
ż
F c|fn ´ f |
“
ż
FXt|fn´f |ďλu
|fn ´ f | `
ż
FXt|fn´f |ąλu
|fn ´ f | `
ż
F c|fn ´ f | p˚q
Since pfnq is tight and f P L1, then pfnq Y tfu is tight. Choose F such that µpF q ă 8 andş
F cp|f | ` |fn|q ă
ε3 , which is an upper bound for the third term of p˚q.
The first term is bounded above by λµpF q, so let λ ą 0 sufficiently small such that λµpF q ă ε3 .
This bounds the first term.Since pfnq is equi-integrable and f P L1, then pfnq Y tfu is equi-integrable. Let δ ą 0 such that
µpAq ă δ impliesş
Ap|fn| ` |f |q ă
ε3 . Finally, since fn
µÑ f , let n0 be sufficiently large such that
n ą n0 implies µpt|fn ´ f | ą λuq ă δ. This bounds the second term above by ε3 , finishing the
proof.Remark - If µpXq ă 8, then any family of L1 functions is tight.Remark - Let F P L1, |fn| ď F for all n. Then
1. pfnq is equi-integrable.
2. pfnq is tight.
Proposition
1. Say limλÑ8
supn
ş
t|fn|ąλu|fn| “ 0. Then pfnq is equi-integrable.
2. (de la Vallee-Poussin) Suppose there is increasing φ : r0,8q Ñ r0,8q such that limxÑ8
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
(2) By (1), it’s enough to show that limλÑ0
supn
ş
t|fn|ąλu|fn| “ 0. Let C “ sup
n
ş
Xφp|fn|q ă 8. Let
λ0 be such that λ ą λ0 implies λ ă εφpλq. Now
ż
t|fn|ąλ0u
|fn| ď
ż
t|fn|ąλ0u
εφp|fn|q ď εC
Take the supremum over n and then take the limit λ0 Ñ8, so now
limλ0Ñ8
supn
ż
t|fn|ąλ0u
|fn| “ 0
”Classic” Example - Say µpXq ă 8, fn Ñ f a.e., and supnfnp ă 8 for some p ą 1. Then fn Ñ f
in L1. This follows from (2) above with φpxq “ xp. Why?Since µpXq ă 8, then a.e. convergence implies convergence in measure, and we also have
tightness for free. We get equi-integrability by (2) with φpxq “ xp.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 20 - 10/13/2014
Plan - Signed measures, Hahn decompositionSetting - pX,Σq a measurable space
Definition - Let µ : Σ Ñ p´8,8s or r´8,8q. We say µ is a signed measure if µ is countablyadditive and µpHq “ 0.E.g. Let µ1, µ2 be positive measures, µ2pXq ă 8. Then µ :“ µ1 ´ µ2 is a signed measure.E.g. 2 - Let ν be a positive measure, f : X Ñ R measurable, and suppose that f´ “ maxp0,´fq P
L1. Then define
µpAq “
ż
A
f` dν ´
ż
A
f´ dν, A P Σ
µ is a signed measure.Proof - Countable additivity follows from monotone convergence.
Motivation - We’d like to form a normed vector space of measures. We can’t do this with positivemeasures alone.Remark - This construction has applications to finding solutions of stochastic differential equations.
Definition - Let µ be a signed measure on pX,Σq. We say A P Σ is µ-positive if for every E Ď Awith E P Σ, we have µpAq ě 0, i.e. when µ is restricted to A, we get a positive measure. We defineµ-negative sets similarly.Notation - Denote Ppµq “ tµ-positive setsu, N pµq “ tµ-negative setsu.
Remarks
1. If A P Σ such that µpAq is finite, then µpEq is finite for any E Ď A in Σ. This follows fromwriting
µpAq “ µpEq ` µpAzEq
2. Ppµq and N pµq are closed under countable unions and intersections, as well as (relative)complements.
3. For any P P Ppµq and N P N pµq, P and N are µ-disjoint. That is, for all E Ď P X N ,µpEq “ 0.
Theorem (Hahn Decomposition) - Let µ be a signed measure on pX,Σq. Then there exist P P Ppµqand N P N pµq such that X “ P YN .Proof - Without loss of generality, assume µ ą ´8. Denote
L “ inftµpAq : A P N pµqu
(Note: There is at least one such A, namely A “ H.)
Claim L is finite. Moreover, the infimum is witnessed.
Proof Let pAnq Ď N pµq such that µpAnq Ñ L. Consider the union N “Ť
nNn. By
a previous remark, N P N pµq. Thus, L ď µpNq by definition. On the other hand,µpNq “ µpAnq ` µpNzAnq ď µpAnq for each n, so µpNq ď L. Thus, µpNq “ L. Inparticular, L is finite (since µ ą ´8).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Set P “ XzN . It remains to show that P P Ppµq. We show this by contradiction.Suppose that P R Ppµq. Let A Ď P in Σ such that µpAq ă 0. If A P N pµq, we have a
contradicting the definition of L. If not, the following lemma will yield the same contradiction.
Lemma Let A P Σ, ´8 ă µpAq ă 0. Then there is E Ď A, E P N pµq, such thatµpEq ď µpAq.
Proof The idea is to cut out the parts of A that are positive. We do this as follows.Set
δ1 “ suptµpEq : E Ď A, E P Σu
If δ1 “ 0, then A P N pµq, so take E “ A. If not, δ1 P p0,8s, so choose A1 Ď A inΣ such that µpA1q ě minp δ12 , 1q.
Continue recursively. At step n, we have
(a) δ1 ě δ2 ě ¨ ¨ ¨ ě δn´1 ą 0
(b) disjoint subsets pAkqn´1k“1 of A such that µpAkq ě minp δk2 , 1q.
Define δn “ suptµpEq : E Ď AzpŤ
iAiq, E P Σu. If δn “ 0, stop and set E “
AzpŤ
iAiq. Otherwise, δn P p0,8s, so let An Ď pAzŤ
iAiq with µpAnq ě minp δn2 , 1q.
If this process stops at any point, we are done. Otherwise, set E :“ AzŤ
ně1An.
Why does this work? First write
µpEq “ µpAq ´ÿ
n
µpAnq ă µpAq p˚q
It remains to argue that E P N pµq. This boils down to showing that δn Ñ 0 asn Ñ 8. Let D Ď E in Σ with µpDq ě 0. Then for each n, D Ď Az
Ť
1ďkďn
Ak. In
particular, µpDq ď δn. We are done, provided we show that δn Ñ 0. But this followsfrom p˚q above. Since µpEq is finite, then
ř
nµpAnq is finite, so
ř
nminp δn2 , 1q ă 8,
so δn Ñ 0.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 21 - 10/15/2014
Plan: Jordan decomposition, MpX,ΣqSetting - pX,Σq is a measurable space, µ a signed measure with Hahn decomposition pP,Nq, whereP XN “ H, X “ P YN , P µ-positive and N µ-negative.
Definition - We say signed measures µ1 or µ2 are singular, denoted µ1 K µ2, if there exist disjointX1, X2 P Σ such that for all A P Σ,
µ1pAq “ µ1pAXX1q, µ2pAq “ µ2pAXX2q
In this case, we say µ1 is concentrated on X1 and µ2 is concentrated on X2.Theorem (Jordan Decomposition) - Given µ, there exist unique (positive) measures µ` and µ´ suchthat
(a) µ “ µ` ´ µ´.
(b) At least one of µ`, µ´ is finite.
(c) µ` K µ´.
Remark This uniqueness is in some sense more advantageous than the Hahn decomposition, since wemay able to change the sets P and N in the Hahn decomposition, i.e. P and N are only essentiallyunique. By contrast, µ` and µ´ are unique by (c).Proof - Existence is easy. Let pP,Nq be a Hahn decomposition for µ. Define
µ`pAq “ µpAX P q, µ´pAq “ ´µpAXNq
By definition, µ` and µ´ satisfy (a)-(c).For uniqueness, let µ` and µ´ satisfy (a)-(c). We show that for any A P Σ, we have
µ`pAq “ supEĎA,EPΣ
µpEq p˚q
(ě) For any E Ď A in Σ, we have
µ`pAq ě µ`pEq “ µpEq ` µ´pEq ě µpEq
(ď) Let pX`, X´q be such that µ` is concentrated on X` and µ´ is concentrated on X´. Now
µ`pAq “ µ`pAXX`q “ µpAXX`q ` µ´pAXX`q “ µpAXX`q
Take E “ AXX` and p˚q is proven.From p˚q, we get µ`. To get µ´, simply apply p˚q again, replacing µ by ´µ. This works because
´µ “ µ´ ´ µ`, so p´µq` “ µ´.
Definition - Given the Jordan decomposition µ “ µ` ´ µ´, the total variation of µ is defined by
|µ| :“ µ` ` µ´
If |µ|pXq ă 8, then µ “ |µ|pXq is called the total variation norm of µ.Remark - We’ll se that ¨ is a norm on the space of bounded total variation measures.E.g. - Let ν be a measure, f P L1pνq. Define µpAq “
ş
Af dν “
ş
Af` dν ´
ş
Af´ dν. In this case,
|µ|pAq “
ż
A
|f | dν, µ “
ż
X
|f | dν “ fL1pνq
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Remark/Quote - ”The space of signed measures is an extension of L1-space.”
Theorem - The spaceMpX,Σq of finite signed measures, equipped with the total variation norm, isa Banach space.Proof - Everything but Banach is easy. Let pµnq ĎMpX,Σq be Cauchy, i.e.
µn ´ µm Ñ 0 as n,mÑ8 p˚˚q
In this case, supnµn ă 8 (by the triangle inequality, Cauchy sequences are bounded).
Observe that for any ν PMpX,Σq, we have
ν “ ν` ` ν´ “ ν`pXq ´ ν´pXq
“ supdisjointA1,A2PΣ
tνpAq ´ νpA2qu
“ supdisjointA1,A2PΣ
t|νpA1q| ` |νpA2q|u
ď 2 supAPΣ
|νpAq|
In particular, we have shown that supAPΣ
|νpAq| ď ν ď 2 supAPΣ
|νpAq|. Thus, p˚˚q is equivalent to
supAPΣ
|µn ´ µmpAq| Ñ 0 as n,mÑ8 p˚ ˚ ˚q
Thus for fixed A P Σ, pµnpAqq is a Cauchy sequence of real numbers, so define µpAq :“ limnÑ8
µnpAq.
By p˚ ˚ ˚q, we havesupAPΣ
|µpAq ´ µnpAq| Ñ 0 as n,mÑ8
It remains to show that µ is a signed measure.To get finite additivity, it suffices to consider two disjoint sets A,B P Σ. We have
as nÑ8. Now for σ-additivity, let pAkq Ď Σ be disjoint. Write
|µpď
k
Akq ´mÿ
k“1
µpAkq|
ď pm` 1q supEPΣ
|µpEq ´ µnpEq| ` |µnpď
kěm`1
Akq|
ď pm` 1q supEPΣ
|µpEq ´ µnpEq| ` supEPΣ
|µnpEq ´ µ`pEq| ` |µ`pď
kěm`1
Akq|
Take lim supn
of the right-hand side and the first term disappears. Then take mÑ 8 and the third
term disappears. Finally, take the limit ` Ñ 8 so that the middle term disappears. Thus, µ isσ-additive.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 22 - 10/20/2014
Definition - Let µ, ν be positive measures. We say ν is absolutely continuous with respect to µ,denoted ν ! µ, provided that
µpAq “ 0 ùñ νpAq “ 0 for all A P Σ
Example - Let pX,Σ, µq be a measure space, f : X Ñ r0,8s measurable, and define dν “ f dµ, i.e.νpAq “
ş
Af dµ. (This was on a previous homework.)
Theorem (Radon-Nikodym) - Let µ, ν ě 0 be σ-finite, ν ! µ. Then there exists a unique (up toµ-a.e.) measurable function g : X Ñ r0,8q such that dν “ g dµ.Notation - g is called the Radon-Nikodym derivative of ν with respect to µ. We write g “ dν
dµ .Proof - Case I: Suppose µ, ν ă 8. Define
F “ tf : X Ñ r0,8s | f measurable and
ż
A
f dµ ď νpAq for all A P Σu
Note, F ‰ H since 0 P F .
Claim 1 For any f1, f2 P F , maxpf1, f2q P F .
Proof For all A P Σ,ż
A
maxpf1, f2q “
ż
AXtf1ąf2u
f1 `
ż
AXtf1ďf2u
f2
ď νpAX tf1 ą f2uq ` νpAX tf1 ď f2uq
“ νpAq
Claim 2 Let pfnq Ď F be increasing. Then limnÑ8
fn P F .
Proof Follows by Monotone Convergence Theorem.
Now let α “ supfPF
ş
Xf dµ ď νpXq ă 8. Let pfnq Ď F with
ş
Xfn dµ Ñ α as n Ñ 8. By Claim
1, we may replace each fn with max1ďiďn
tfiu and assume that pfnq is increasing. Set g “ limnÑ8
fn, now
g P F by Claim 2. In particular,ş
Xg dµ ď νpXq ă 8 implies that g ă 8 µ-a.e., so we may assume
g : X Ñ r0,8q.To finish Case 1, it remains to show that dν “ g dµ. To do this, define dν0 “ dν ´ g dµ, i.e.
ν0pAq “ νpAq ´ş
Ag dµ for all A P Σ. Clearly ν0 ě 0 since g P F . We want to show that ν0 ď 0, so
then ν0 “ 0. It’s enough to prove the following claim.
Claim For any ε ą 0, ν0 ď εµ.
Proof Fix ε ą 0. Since ν0 ´ εµ is a signed measure, there is a corresponding Hahndecomposition X “ P Y N . To get ν0 ď εµ, we show that µpP q “ 0. Then byν ! µ, we have νpP q “ 0, so ν0 ´ εµ is negative and we’re done.
To show that µpP q “ 0, it’s enough to show that g ` εχP P F (by definition ofg). Write
ż
A
pg ` εχP q dµ “
ż
A
g ` εµpAX P q
“ νpAq ´ ν0pAq ` εµpAX P q
“ νpAq ´ ν0pAXNq ´ pν0 ´ εµqpAX P q
ď νpAq
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Thus g ` εχP P F , so µpP q “ 0 and we are done.
We still need to show uniqueness. Suppose that dν “ g1 dµ “ g2 dµ. Then for all A P Σ, wehave
νpAq “
ż
A
g1 dµ “
ż
A
g2 dµ ùñ
ż
A
pg1 ´ g2q dµ
Let A “ tg1 ą g2u. We get µptg1 ą g2uq “ 0. By symmetry, µptg1 ă g2uq “ 0. Thus, g1 “ g2 µ-a.e.
Case II: Suppose µ, ν are σ-finite. Write X “Ť
nFn, where Fn Ď Fn`1, µpFnq, νpFnq ă 8. For
each n, define µnpAq “ µpA X Fnq, νnpAq “ νpA X Fnq. Note that µn, νn ă 8 and νn ! µn. ByCase I, for each n there is a unique (µn-a.e.) fn such that dνn “ fn dµn. By this uniqueness, wemust have fn`1 æFn“ fn for each n (µn-a.e.). Thus, g “ lim
nÑ8fn is well-defined and has the desired
property (by the MCT).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 23 - 10/22/2014
Lebesgue Decomposition
Definition - Let µ be a positive σ-finite measure, ν a signed finite measure. We write ν ! µ if
µpAq “ 0 ùñ |ν|pAq “ 0 for all A P Σ
Theorem - Let µ be positive and σ-finite, ν signed and finite, ν ! µ. Then there exists a uniqueg P L1pµq such that dν “ g dµ.Proof - By Jordan, we can write ν “ ν` ´ ν´. Now Radon-Nikodym yields g`, g´ such thatdν˘ “ g˘ dµ. Set g “ g` ´ g´.(Boring) Example - For any finite signed measure µ, we have µ ! |µ|. By Radon-Nikodym, there isg such that dµ “ g d|µ|. But we know this already! If X has Hahn decomposition X “ P YN withrespect to µ, then g “ χP ´ χN .
Later, we’ll see a really surprising and nontrivial application of Radon-Nikodym.
Theorem (Lebesgue Decomposition) - Let µ, ν be positive measures, ν σ-finite. Then there existpositive measures νS and νAC such that
1. ν “ νS ` νAC
2. νAC ! µ
3. νS K µ
Recall - νS K µ if and only if there is N P Σ such that µpNq “ 0, νSpNcq “ 0.
Proof - Case I: Suppose ν is finite. Set
N “ tN P Σ | µpNq “ 0u, α “ suptνpNq | N P N u ă 8
Let pNnq Ď N with νpNkq ą α´ 1k for each k. Set N “
cq “ 0. It remains to show that νAC ! µ.Suppose µpAq “ 0. We need νACpAq “ 0, i.e. νpAzNq “ 0. Since AYN P N , then
α “ νpNq ď νpAYNq ď α
Since ν is finite, then νpAzNq “ 0, so we’re done.Case II: Suppose that ν is σ-finite. Write X “
Ť
nFn, where νpFnq ă 8 and Fn Ď Fn`1. Set
νpnqpAq “ νpA X Fnq, so νpnq ă 8. By Case I, there exists NN Ď Fn and positive measures νpnqS ,
νpnqAC such that νpnq “ ν
pnqS ` ν
pnqAC and µpNnq “ 0, ν
pnqS pN c
nq “ 0.Define N “
Ť
nNn, νSpAq “ νpAXNq, νACpAq “ νpAXN cq, and ”suffer” (i.e. check that it all
works out).
Proposition - The Lebesgue decomposition is unique.
Proof - Suppose that ν “ νp1qS ` ν
p1qAC “ ν
p2qS ` ν
p2qAC . Now
νp1qS ´ ν
p2qS “ ν
p2qAC ´ ν
p1qAC
The left-hand side is singular with respect to µ. The right-hand side is absolutely continuous with
respect to µ. Since these are equal, they must both be equal to zero, so that νp1qS “ ν
p2qS and
νp1qAC “ ν
p2qAC .
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Corollary (Radon-Nikodym-Lebesgue Decomposition) - Let µ, ν be positive, σ-finite measures. Thenthere is unique g : X Ñ r0,8q and unique νS such that
dν “ dνS ` g dµ, νS K µ
Application of Radon-Nikodym: The Dual of Lp
Let X be a Banach space. We denote by X˚ the (continuous) dual of X, defined by
X˚ “ tx˚ | x˚ : X Ñ R linear and continuousu
Theorem - Let p P r1,8q, µ a positive σ-finite measure, 1p `
1q “ 1. Then Lq is isometric to pLpq˚.
Remark - Given g P Lq, define Λg : Lp Ñ R by
fΛgÑ
ż
X
fg dµ
By Holder, this is well-defined. We’ll prove the theorem by showing that g ÞÑ Λg maps Lq to pLpq˚
isometrically.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 24 - 10/24/2014
Theorem - Let p P r1,8q, µ a σ-finite positive measure, 1p `
1q “ 1. Then Lq is isometrically
isomorphic to pLpq˚.Lemma - Let U be a Banach space, T : U Ñ R linear. Then
T is continuous ðñ DC ě 0 s.t. |Tu| ď Cu for all u
Definition - We say T : U Ñ R is bounded if there is C ě 0 such that |Tu| ď Cu for all u P U .Proof (of Lemma) - Say T is bounded. Then |Tu ´ Tv| “ |T pu ´ vq| ď Cu ´ v. In particular, Tis Lipschitz, so T is continuous.
Conversely, suppose T is continuous. Let δ ą 0 such that u ă δ implies |T puq| “ |T puq´T p0q| ă1. Then for all v ‰ 0, we have
|Tv| “ 2vδ |T p
δ2vv q| ă
2δ v
Notation - For u P U , denote by uU the norm of u in U . Similarly, for u˚ P U˚, denote by u˚U˚the norm of u˚ in U˚ (to be defined).Definition - Given u˚ P U˚, define
u˚U˚ :“ supuPU,uU“1
u˚puq
(Note: This is finite since each u˚ is bounded.)Exercise ¨ U˚ is a norm on U˚ and pU˚, ¨ U˚q is a Banach space.
Proof (of Theorem) - Consider the map Λ : Lq Ñ pLpq˚ which takes g P Lq to Λg P pLpq˚ defined
by Λgpfq “ş
fg. The map Λg is clearly linear. Moreover, it is continuous by Holder’s inequalitybecause |Λgpfq| ď gqfp (so it is bounded). Thus, for all g P Lq, Λg P pL
pq˚.It remains to show that Λ is a linear bijective isometry. Linearity is clear. For isometry, we want
ΛgpLpq˚ “ gLq for all g P Lq. Note,
ΛgpLpq˚ “ supfPLp,f‰0
|Λgpfq|fp
“ supfPLp,f‰0
1fp
|
ż
fg| “ gq
(by the duality lemma from long ago). Thus, Λ is an isometry.Since Λ is an isometry, it is automatically injective. We need to show Λ is surjective. Let
φ P pLpq˚. Want g P Lq such that φ “ Λg.Case I: Suppose µ is finite. The main idea is to note that for any A P Σ, χA P L
p, so φpχAq isdefined. Define νpAq “ φpχAq. We show that
1. ν is a signed, finite measure.
2. ν ! µ, so that by Radon-Nikodym, dν “ g dµ
3. φ “ Λg
For (1), note that νpHq “ φp0q “ 0. Given pAnq Ď Σ countable and pairwise disjoint, we havefor each N
νpď
1ďiďN
Aiq “ φpχ Ť
1ďiďN
Aiq “ÿ
1ďiďN
φpχAiq “ÿ
1ďiďN
νpAiq
Thus, we have finite additivity for ν. For countable additivity, it’s enough to show that
ÿ
1ďiďN
χAiNÑ8Ñ χ Ť
iě1Ai in Lp
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Writeχ Ť
iě1Ai ´
ÿ
1ďiďN
χAipp “ χ
Ť
iěN`1
Aipp “
ÿ
iěN`1
µpAiq Ñ 0
Thus, ν is countably additive.For (2), suppose µpAq “ 0. Then χA “ 0 in Lq, so νpAq “ φpχAq “ 0. By Radon-Nikodym,
there is g such that dν “ g dµ.For (3), note that for any simple function s, we have
Λgpsq “
ż
sg dµ “
ż
s dν “ÿ
i
aiνpAiq “ φpsq,
To show that g P Lq, let f P Lp be arbitrary. Let psnq Ď Lp with snLpÑ f . Then by continuity of
φ, φpsnq Ñ φpfq in R. But φpsnq “ş
sn dν Ñş
f dν (definition ofş
). Hence for all f P Lp,
φpfq “
ż
f dν “
ż
fg dµ “ Λgpfq
Now φ bounded implies g P Lq, since
supfp“1
ż
fg dµ “ φpLpq˚ ă 8
Case II: If µ is σ-finite, reduce to Case I.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 25 - 10/27/2014
Last Time - If p P r1,8q, µpXq ă 8, and 1p`
1q “ 1, there is a bijective isometry pLppXqq˚ » LqpXq.
Remark - Why is this useful? Many important normed vector spaces of functions, e.g. pLppXq, ¨pq,are infinite-dimensional. As a result, we don’t get compactness of the closed unit ball (contrast withRn). The Banach-Alaoglu theorem (from functional analysis) says that the closed unit ball of thedual space X˚ is weak-˚ compact (compact with respect to something called the weak-˚ topology).In particular, we have some sort of compactness for LqpXq, since it is the dual space of LppXq.Remark - For p “ 8, pL8q˚ » tν | ν a finitely additive measure, ν ! µu, a ”boring space”. Insearch of a better result, let’s consider a subspace of L8 (which will have a bigger, hopefully moreinteresting space as its dual).
Suppose that X is a compact metric space, so now CpXq Ď L8pXq. (Note: pCpXq, ¨ 8q is aBanach space since a uniform limit of continuous functions is continuous.) Let µ be a finite, signed,regular Borel measure on X. Define
Λµ P CpXq˚, Λµpfq “
ż
X
f dµ for all f P CpXq
Theorem (Riesz) - DenoteM “ tµ | µ a finite, signed, regular Borel measure on Xu. Then the mapΛ :MÑ CpXq˚ defined by
pΛpµqqpfq “ Λµpfq “
ż
X
f dµ
is a bijective linear isometry.Proof - Linearity is immediate. To show that µ is an isometry, first write
µ “ |µ|pXq “ µpP q ´ µpNq, Λµ “ supf8“1
|Λµpfq|,
where pP,Nq is a Hahn decomposition of X with respect to µ, P XN “ H. Clearly µ ě Λµ. Toget the converse, we’d like to choose f “ χP ´ χN above, but this may not be continuous, so useregularity/Lusin’s theorem to approximate f .
It remains to show that µ ÞÑ Λµ is bijective. Since this map is an isometry, it is injective, so onlysurjectivity remains.
Suppose I P pCpXq˚q (”I for integral”). We need to find µ PM such that Ipfq “ş
Xf dµ for all
f P CpXq.Case I: Suppose I is positive. That is, given f P CpXq with f ě 0, we have Ipfq ě 0 (intuitively,
I looks like integration with respect to a positive measure). For any open U Ď X, define
µ˚pUq “ suptIpfq | f P CpXq, 0 ď f ď 1, supppfq Ď Uu
(We’re trying to recover the measure of any Borel set, but our only means of approximation iscontinuous functions.) Now for any A Ď X, define
µ˚pAq “ inftµ˚pUq | U Ě A is openu
It’s not hard to check that µ˚ is indeed an outer measure, so define, as in Caratheodory’s theorem,Σ “ tE | µ˚pAq “ µ˚pAX Eq ` µ˚pAX Ecq for all A Ď Xu. Now µ˚ æΣ is a measure.
Lemma 1 Σ Ě BpXq.
Lemma 2 For all f P CpXq, Ipfq “ş
Xf dµ.
Proof (Lemma 1) It’s enough to show U P Σ for all U Ď X open. Let A Ď X.We need to show that µ˚pAq ě µ˚pAX Uq ` µ˚pAX U cq.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
• If A is open, write A “ V . Fix ε ą 0. Let f, g P CpXq such that 0 ď f, g ď 1,and
supppfq Ď V X U, Ipfq ě µ˚pV X Uq ´ ε2
supppgq Ď V zsupppfq, Ipgq ě µ˚pV zsupppfqq ´ ε2 ě µ˚pV X U cq ´ ε
2
Now f ` g P CpXq, 0 ď f ` g ď 1, and supppf ` gq Ď V , so
µ˚pV q ě Ipf ` gq “ Ipfq ` Ipgq ě µ˚pV X Uq ` µ˚pV X U cq ´ ε
Taking the limit εÑ 0`, we are done.
• Let A be arbitrary. Fix ε ą 0 and let A Ď V Ď X open with µ˚pV q ď µ˚pAq`ε.Now
µ˚pAq ` ε ě µ˚pV q “ µ˚pV X Uq ` µ˚pV X U cq ě µ˚pAX Uq ` µ˚pAX U cq
Taking εÑ 0`, we are done.
Proof (Lemma 2) Need to show f P CpXq implies Ipfq “ş
Xf dµ.
• Step 1: Suppose f ě χA, where A P BpXq. We show that Ipfq ě µpAq. Givenε ą 0, define U “ tf ą 1 ´ εu. Note that U is open and U Ě A. If g P CpXqwith 0 ď g ď 1 and supppgq Ď U , then
g ď f1´ε ùñ Ipgq ď 1
1´εIpfq,
since we are assuming Ip¨q is a positive operator. Taking the supremum over all
such g, we have µpUq ď Ipfq1´ε . Taking the limit εÑ 0`, we have µpAq ď Ipfq.
(Note: Although U “ Upεq, we clearly have A Ď Upεq for any ε ą 0, soµpAq ď µpUpεqq.)
To be continued...
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 26 - 10/29/2014
We continue with the proof of Riesz’s theorem (refer to previous lecture’s notes).
Lemma 2 Ipfq “ş
Xf dµ for all f P CpXq.
Proof • Step 1: We showed that χA ď f implies µpAq ď Ipfq for A P BpXq.• Step 2: Given s “
ř
i
aiχAi ě 0, where tAiu are disjoint, we show that
s ď f ùñ
ż
s
dµ ď Ipfq
To do this, fix ε ą 0. For each i select Ki Ď Ai compact, Ui Ě Ki open suchthat the sets tUiu are disjoint and f
1´ε ě ai on Ui. (Note: Disjointness ispossible since disjoint compact subsets of a metric space can be separated.)Now let φi : X Ñ r0, 1s be continuous such that φi “ 1 on Ki and supppφiq ĎUi. For each i, we have
aiχKi ďφif1´ε ùñ aiµpKiq ď
Ipφifq1´ε
Summing over i, we have
ÿ
i
aiµpKiq ďÿ
i
Ipφifq1´ε “
Ippř
iφiqfq
1´ε ďIpfq1´ε ,
since the supports of the functions tφiu are disjoint. Taking the supremumover compact Ki Ď Ai, we have
ż
s dµ ď Ipfq1´ε
Taking the limit εÑ 0`, we haveş
s dµ ď Ipfq, as desired.
• Step 3: Exercise: If 0 ď f ď t and t is simple, then Ipfq ďş
Xt dµ. (Appeal
to Step 2 and finiteness of µ.)
• Step 4: By the Exercise, given f P CpXq, we have
Ipf`q ě sup0ďsďf`
s simple
ż
X
s dµ “
ż
X
f` dµ “ inf0ďf`ďtt simple
ż
X
t dµ ě Ipf`q
which impliesş
Xf` dµ “ Ipf`q, so
ş
Xf dµ “ Ipfq.
Case II: Let I P CpXq˚ be arbitrary. Given f ě 0, define I`pfq “ suptIpgq | 0 ď g ď f, g PCpXqu. Note I`pfq ě 0 (take g “ 0). Moreover, this is finite since I is bounded. Check that this islinear. Define
I´pfq “ inftIpgq | 0 ď g ď f, g P CpXqu
or alternately define I´ “ I` ´ I (we want I “ I` ´ I´). Now for f P CpXq arbitrary, define
I`pfq “ I`pf`q ´ I`pf´q, I´pfq “ I`pfq ´ Ipfq
Check that everything works out! We get I` P pCpXqq˚ by applying Case I, same for I´.This finishes the proof.
Remark - The decomposition I “ I` ´ I´ corresponds to the Jordan decomposition. Explicitly, ifI “ µ, then I` “ µ` and I´ “ µ´.
Two Extensions of the Theorem
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
1. If X is locally compact and I : CcpXq Ñ R is positive, then there is a unique regular Borelmeasure µ such that Ipfq “
ş
Xf dµ for all f P CcpXq.
2. If X is locally compact, let M be the set of finite, signed, regular Borel measures on X. Wehave CcpXq Ď L8pXq, so we can define C0pXq “ CcpXq in L8pXq. We call C0pXq the set ofcontinuous functions that vanish at infinity. It can be shown that pC0pXqq
˚ “M, where themap
Λ :MÑ pC0pXqq˚, Λµpfq “
ż
X
f dµ
is a bijective linear isometry.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 27 - 10/31/2014
Exercise: Continuity of Regular Measures - Let X be a compact metric space, µ a finite regularpositive Borel measure on X such that µptxuq “ 0 for all x P X. Then for any α P p0, µpXqq, thereis A Ď X with µpAq “ α. (Hint: Use Riesz.)
Product Measures and Fubini’s Theorem
Setting - pX,Σ, µq and pY, τ, νq are measure spaces, µ, ν σ-finite.Definition - We denote Σ ˆ τ “ tA ˆ B | A P Σ, B P τu, the set of all ”rectangles” in A ˆ B. Theproduct σ-algebra on AˆB is defined by
Σb τ :“ σpΣˆ τq
Theorem - There is a unique measure π on pX ˆ Y,Σb τq such that
πpAˆBq “ µpAqνpBq
for all A P Σ, B P τ . We call π the product measure on pX ˆ Y,Σb τq.Remark - This fails if µ or ν is not σ-finite.
We begin by stating a series of theorems which we will then proceed to prove. We take thisapproach because the final result (Fubini’s theorem) plays a decisive role in our method of proof forthe first result (existence of the product measure).
Theorem - Let f : X ˆ Y Ñ r0,8s be Σb τ -measurable.
(a) For all x P X, the map y ÞÑ fpx, yq is τ -measurable. Similarly, for all y P Y , the mapx ÞÑ fpx, yq is Σ-measurable.
(b) The functions x ÞÑş
Yfpx, yq dνpyq and y ÞÑ
ş
Xfpx, yq dµpxq are Σ- and τ -measurable respec-
tively.
(c) (Tonelli)
ż
XˆY
f dπ “
ż
X
´
ż
Y
fpx, yq dνpyq¯
dµpxq “
ż
Y
´
ż
X
fpx, yq dµpxq¯
dνpyq p˚q
Theorem (Fubini) - Suppose f P L1pX ˆ Y, πq.
(a) For a.e. x P X,ş
Yfpx, yq dνpyq is finite and µ-integrable.
(b) For a.e. y P Y ,ş
Xfpx, yq dµpxq is finite and ν-integrable.
(c) The equality p˚q holds and the integrals are finite.
Proof (existence of product measure)Main Idea - Given E Ď X ˆ Y , for each x P X define SxpEq “ ty P Y | px, yq P Eu. For each y P Y ,define TypEq “ tx P X | px, yq P Eu. Now define
πpEq “
ż
X
νpSxpEqq dµpxq
We need to show three things:
1. SxpEq P τ for all x P X.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
2. The function x ÞÑ νpSxpEqq is Σ-measurable.
3. π is a measure.
First, we prove uniqueness. Say π1, π2 are two such measures. If µ, ν are finite, then the settπ1 “ π2u is a λ-system containing Σˆ τ , which is a π-system. By Dynkin’s theorem, tπ1 “ π2u “
Σb τ . If µ, ν are σ-finite, then take limits and use the previous case.For existence, we prove a sequence of lemmas.
Lemma 1 For all E P Σb τ , x P X, y P Y , SxpEq P τ and TypEq P Σ.
Proof Define F “ tE P Σbτ | @x P X, SxpEq P τu. This is a σ-algebra containingthe rectangles Σˆ τ , so F “ Σb τ . The argument is the same for TypEq.
Lemma 2 For all E P Σbτ , the function x ÞÑ νpSxpEqq is Σ-measurable. Similarly,the function y ÞÑ µpTypEqq is τ -measurable.
Proof Denote Λ “ tE P Σ b τ | x ÞÑ νpSxpEqq is Σ-measurableu. This is notobviously a σ-algebra (because of the problem of non-disjoint unions). Instead, weshow that Λ is a λ-system! Since Λ contains the rectangles Σ ˆ τ , we then haveΛ “ Σb τ .
First suppose that µ, ν are finite.
(1) Clearly X ˆ Y P Λ, since νpSxpX ˆ Y qq “ νpY q, a constant function.
(2) Suppose A,B P Λ, A Ď B. We have
νpSxpBzAqq “ νpSxpBqq ´ νpSxpAqq,
a difference of measurable functions, which is measurable. Thus BzA P Λ.
(3) Let pAnq Ď Λ with An Ď An`1. We have
νpSxpď
n
Anqq “ νpď
n
SxpAnqq “ limnÑ8
νpSxpAnqq,
a limit of measurable functions, which is measurable. ThusŤ
nAn P Λ. Hence,
Λ is a λ-system.
If µ, ν are σ-finite, take limits and use the previous case.
To be continued...
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 28 - 11/03/2014
Quotes (A bowl of candy is sitting at the front of the classroom) Student: ”Are we going over theReese’s Representation Theorem with candy?” Professor: ”What’s the connection between Rieszand candy?” (slaps forehead) ”I’m slow.”Proof (of existence of product measure, continued) - We still need to show that πpEq “
ş
XνpSxpEqq dµpxq
is a measure. Let pEnq Ď Σb τ be pairwise-disjoint. Now
πpď
n
Enq “
ż
X
νpSxpď
n
Enqq dµpxq
“
ż
X
ÿ
n
νpSxpEnqq dµpxq
“ÿ
n
ż
X
νpSxpEnqq dµpxq (Beppo-Levi/MCT)
“ÿ
n
πpEnq
Hence, π is a measure. Moreover, for any AˆB P Σˆ τ , we have
πpAˆBq “
ż
X
νpSxpAˆBqq dµpxq “
ż
X
νpBqχApxq dµpxq “ νpBq
ż
A
dµpxq “ µpAqνpBq
Remark - We could have defined the product measure more constructively, starting with some outermeasure that uses coverings of E P Σb τ by countable collections of rectangles pAn ˆBnq Ď Σˆ τ .Then do some sort of Caratheodory argument. There are two drawbacks to this: First of all, thiswould be very tedious. Second of all, this construction makes it difficult to recover the ”nicer”definition of π in terms of integration. We will see that with our definition of π, it doesn’t take toomuch work to prove Tonelli’s and Fubini’s theorems.
Proof (Tonelli) - Let f : X ˆ Y Ñ r0,8s be Σb τ -measurable. We need to show that
(a) For all x P X, y P Y , the maps fpx, ¨q and fp¨, yq are τ -measurable, Σ-measurable respectively.
(b) The mapsş
Yfp¨, yq dνpyq and
ş
Xfpx, ¨q dµpxq are Σ- and τ -measurable respectively.
(c) (Tonelli)
ż
XˆY
f dπ “
ż
X
´
ż
Y
fpx, yq dνpyq¯
dµpxq “
ż
Y
´
ż
X
fpx, yq dµpxq¯
dνpyq p˚q
We proceed in several steps.
1. Say f “ χE , where E P Σb τ . By definition, we haveż
XˆY
f dπ “ πpEq “
ż
X
νpSxpEqq dµpxq “
ż
X
´
ż
Y
χEpx, yq dνpyq¯
dµpxq
and we have already shown that (a) and (b) hold. By symmetry, the function πpEq “ş
YµpTypEqq dνpyq is a measure that satisfies πpA ˆ Bq “ µpAqνpBq. By uniqueness of the
product measure, we have π “ π, so p˚q holds.
2. By linearity and the previous step, p˚q holds if f is a positive simple function. Moreover, (a)and (b) are clearly preserved under positive linear combinations.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
3. Now let f : X ˆ Y Ñ r0,8s be arbitrary Σ b τ -measurable. Let sn Õ f be a sequence ofpositive, simple, Σbπ-measurable functions. By MCT, (a) and (b) clearly hold. For p˚q, write
ż
XˆY
f dπ(MCT)“ lim
nÑ8
ż
XˆY
sn dπ “ limnÑ8
ż
X
ż
Y
snpx, yq dνpyq dµpxq
“
ż
X
´
limnÑ8
ż
Y
snpx, yq dνpyq¯
dµpxq (MCT)
“
ż
X
ż
Y
limnÑ8
snpx, yq dνpyq dµpxq (MCT)
“
ż
X
ż
Y
f dνpyq dµpxq
Proof (Fubini) - Let f P L1pX ˆ Y, πq. Write f “ f` ´ f´. Nowż
XˆY
f dπ “
ż
XˆY
pf` ´ f´q dπ
“
ż
XˆY
f` dπ ´
ż
XˆY
f´ dπ
“
ż
X
ż
Y
f`px, yq dνpyq dµpxq ´
ż
X
ż
Y
f´px, yq dνpyq dµpxq (Tonelli)
“
ż
X
´
ż
Y
f` dνpyq ´
ż
Y
f´ dνpyq¯
dµpxq p˚˚q
“
ż
X
ż
Y
fpx, yq dνpyq dµpxq
Why does step p˚˚q hold? This is because
µptx |
ż
Y
f`px, yq dνpyq “ 8uq “ 0, µptx |
ż
Y
f´px, yq dνpyq “ 8uq “ 0
In particular,ş
Yf`px, yq dνpyq ´
ş
Yf´px, yq dνpyq is well-defined except on µ-null set.
Application of Fubini/Tonelli - Let f : X Ñ r0,8q be measurable, µ σ-finite on X. Then
1.ş
Xf dµ “
ş8
0µptf ą xuq dλpxq
2. If φ : r0,8q Ñ r0,8q is C1, increasing, and satisfies φp0q “ 0, thenż
X
φpfq dµ “
ż 8
0
µptf ą xuqφ1pxq dλpxq
Note that (1) implies (2), since (1) impliesż
X
φpfq “
ż 8
0
µptφpfq ą xuq dλpxq “
ż 8
0
µptf ą xuqφ1pxq dλpxq
by change of variables. To prove (1), define E “ tpx, λq | λ P r0, fpxqsu. Then by Tonelli,ż
XˆRχE dπ “
ż
X
ż 8
0
χE dλ dµpxq “
ż
X
fpxq dµpxq
ż
XˆRχE dπ “
ż 8
0
ż
X
χE dλ dµpxq “
ż 8
0
µptf ě yuq dλpyq
Remark - We also haveş
f “ş8
0µptf ą yuq dλpxq, since we’re only cutting out the graph of the
function (which has measure zero).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 29 - 11/05/2014
Setting - All functions are functions on Rd.Motivation - Consider a continuous but erratic function fptq, e.g. the price of a stock over the courseof a year, which typically moves up and down quite a lot. Instead of studying fptq, we can study a
”moving average” of fptq, defined by F ptq “şt`ε
t´εfptq ¨ 1
2ε dt (note: F ptq is the average value of thefunction fptq over the interval pt´ ε, t` εq). This function generally looks smoother and is easier tointerpret.
Definition - Let f, g be functions on Rd. We define a new function f ˚ g (”f star g” or ”f convolvedwith g”) by
pf ˚ gqpxq “
ż
Rdfpx´ yqgpyq dy
Remark - By making a change of variables, we see that f ˚ g “ g ˚ f .Remark - In the Motivation above, F ptq “ f ˚ p 1
2εχr´ε,εsq.
When is f ˚ g defined? To answer this, we must make assumptions on f and g. For simplicity,let’s suppose first that f, g P L1. Then by Tonelli’s theorem, we have
ż
Rd|f | ˚ |g| dx “
ż
Rd
ż
Rd|fpx´ yq||gpyq| dydx
“
ż
Rd|gpyq|
ż
Rd|fpx´ yq| dxdy
“
ż
Rd|gpyq|f1 dy
“ f1g1
In particular,ş
Rd |f | ˚ |g| dx is finite, so not only is f ˚ g defined a.e., but f ˚ g P L1.The previous result is nice. How does it generalize? That is, given f P Lp and g P Lq, is f ˚ g
defined? Can we find r such that f ˚ g P Lr?We seek an inequality of the form f ˚ gr ď fpgq. As we’ve done previously, let’s try
”counting dimensions”. View f, g as dimensionless functions, ` denotes length, so that Rd hasdimension `d. We note that:
• f ˚ g has dimension `d, so then f ˚ gr has dimension `d`dr.
• fpgq has dimension `dp`dq.
If f ˚ g P Lr, then we should expect d ` dr “
dp `
dq , or 1 ` 1
r “1p `
1q . This agrees with the result
we just derived above; take p “ q “ r “ 1.
Theorem (Young’s Inequality) - Let p, q P r1,8s. If f P Lp, g P Lq, then f ˚ g P Lr, where
1` 1r “
1p `
1q . In fact,
f ˚ gr ď fpgq
Proof - Let r1 be the conjugate of r. By the duality lemma, it’s enough to show that for all h P Lr1
,we have
ż
pf ˚ gqh ď fpgqhr1 p˚q
For simplicity, assume f, g, h ě 0. Otherwise, replace f, g, h by |f |, |g|, |h| and use monotonicity ofthe integral. Let p1, q1 be the conjugates of p, q respectively. We have
1p `
1q “ 1` 1
r ,1p “ 1´ 1
p1 ,1q “ 1´ 1
q1 ùñ1p1 `
1q1 `
1r “ 1 p˚˚q
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
In light of p˚˚q, we would like to apply Holder’s inequality (the general version for n functions,n ě 2), followed by Tonelli’s theorem. To start, write
(Note: Applying Holder’s inequality will separate the three functions f, g, h into three pairs of twofunctions, which will then permit us to use Tonelli’s theorem.) We’d like one of each of the threefactors above to be in Lp
From the first, second, and fourth equation, we require that α “ pr , β “ q
r , γ “ r1
q1 . It is straight-forward to verify that these values satisfy the third, fifth, and sixth equations. By Holder, wehave
ż
Rdpf ˚ gqh ď p
ij
fpx´ yqpgpyqqq1rp
ij
fpx´ yqphpxqr1
q1q1
p
ij
gpyqqhpxqr1
q1p1
“ fprp gqrq fpq1
p gr1q1
r1 gqp1
q hr1p1
r1
“ fα`p1´αqp gβ`p1´βqq ` hγ`p1´γqr1
“ fpgqhr1
This finishes the proof.
Mollification
Definition - Let tφnu be a sequence of functions on Rd. We say tφnu is an approximate identity if
1. φn ě 0 for all n.
2.ş
Rd φn “ 1 for all n¿
3. For each ε ą 0, we have limnÑ8
ş
|x|ąεφnpxq dx “ 0.
Example - Let φ P L1 with φ ě 0,ş
Rd φ “ 1. For any ε ą 0, define φεpxq “1εdφpxε q. Then tφεuεÑ0
is an approximate identity.
Approximate identities are useful because they provide a means to approximate arbitrary func-tions by a sequence of much nicer functions. For example, given an approximate identity consisting ofsmooth functions, we can use these to approximate many non-smooth functions by smooth functions.In particular...
Proposition - Let tφnu be an approimate identity.
1. If f P Lp for some p P r1,8q, then f ˚ φnLpÑ f , i.e. f ˚ φn ´ fp Ñ 0.
2. If f P L8 and f is continuous at some x P Rd, then f ˚ φnpxq Ñ fpxq.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Proof - We prove (2), leaving (1) for next time. Fix ε ą 0. Write
|φn ˚ fpxq ´ fpxq| “ |
ż
φnpyqfpx´ yq dy ´ fpxq|
“ |
ż
φnpyqpfpx´ yq ´ fpxqq dy|
ď
ż
φnpyq|fpx´ yq ´ fpxq| dy
“
ż
|y|ăδ
p¨ ¨ ¨ q dy `
ż
|y|ěδ
p¨ ¨ ¨ q dy
Pick δ ą 0 sufficiently small so that |fpx´ yq ´ fpxq| ă ε2 for all |y| ă δ. This bounds the first term
by ε2 . Now let N be sufficiently large so that n ě N implies
ş
|y|ěδφnpyq dy ď
ε4f8
. This finishes
the proof of (2).
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 30 - 11/07/2014
Remark - Given a measure µ on Rd, define f ˚ µpxq “ş
Rd fpx ´ yq dµpyq. For example, we havef ˚ δ0pxq “ fpxq (i.e. δ0 is the identity for this operation ˚). This justifies the nomenclature”approximate identity”, since f ˚ φn Ñ f “ f ˚ δ0.
Recall - From last time, we still need to prove that f P Lp, p P r1,8q implies f ˚ φnLpÑ f .
Proof - We make use of two facts from previous homeworks. First, that if p P r1,8q and f P Lp, then
τyfLpÑ f as |y| Ñ 0 (here, τyfpxq “ fpx ´ yq). The second is Minkowski’s inequality for integrals.
Fix ε ą 0 and write
f ˚ φnpxq ´ fpxq “
ż
fpx´ yqφnpyq dy ´ fpxq “
ż
pτyfpxq ´ fpxqqφnpyq dy
Now
f ˚ φn ´ fp “ p
ż
|
ż
pτyfpxq ´ fpxqqφnpyq dy|p dxq1p
ď
ż
τyf ´ fpφnpyq dy
“
ż
|y|ăδ
τyf ´ fpφnpyq dy `
ż
|y|ěδ
τyf ´ fpφnpyq dy
Take δ ą 0 small enough so that τy ´ fp ăε2 for all |y| ă δ. Then take N sufficiently large so that
ş
|y|ěδφnpyq dy ă
ε4fp
. This finishes the proof.
Application: Fourier SeriesQuote - ”Whatever I say for the first lecture and a half probably is OK with just the Riemannintegral, but after that you’ll see how much more you can get.”Setup - We consider periodic functions on r0, 1s. Denote enpxq “ e2πinx, n P Z. Define
L2perr0, 1s “ tf : RÑ C | fpx` 1q “ fpxq,
ż 1
0
|f |2 ă 8u
If f, g P L2per, define xf, gy “
ş1
0fg. (Note: To integrate a complex-valued function, simply integrate
the real and imaginary parts separately, then add them back together. ”This is the only sensibleway to integrate such functions.”) This defines an inner product on L2
per that induces the norm
f2 “ş1
0|f |2.
Definition - Given f P L2perr0, 1s, define fpnq “ xf, eny “
ş1
0fpxqe´2πinx dx (Fourier coefficients).
Definition - Given N ě 0, define SNf “ř
|n|ďN
fpnqen (partial sums of Fourier series).
Remark - The set tenunPZ is orthonormal, i.e. xen, emy “ δnm.
Goal 1 - Show that SNfL2
Ñ f for all f P L2per.
One useful observation is that xSNf, f ´ SNfy “ 0 (easy to check). More generally, we have thefollowing:
Lemma - Let pN P spante´N , . . . , eNu. Then xf ´ SNf, pN y “ 0.Proof - By linearity, it’s enough to check this for pN “ em, where |m| ď N . Write
xf, emy “ fpmq, xSNf, emy “ÿ
|n|ďN
fpnqxen, emy “ fpmq
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Corollary 1 (Bessel’s inqeuality) - f`2 ď fL2
Proof - Set EN “ f ´ SNf . Then by the Pythagoras theorem for inner products,
SNf ` EN “ f, ; xSNf,EN y “ 0 ùñ f22 “ SNf22 ` EN
22 ě SNf
22 “
ÿ
|n|ďN
|fpnq|2
Taking the limit N Ñ8, we have the result.
Corollary 2 - Suppose there is a sequence ppN q such that for each N , pN P spante´N , . . . , eNu and
pNL2
Ñ f . Then SNfL2
Ñ f .Proof - It’s enough to show that f´SNf2 ď f´pN . This follows by basic facts about projectionsfor inner product spaces. Explicitly, write
f ´ pN “ pf ´ SNfq ` pSNf ´ pN q ùñ f ´ pN 22 “ f ´ SNf
22 ` SNf ´ pN
22 ě f ´ SNf
22,
where we have used the fact that f ´ SNf and SNf ´ pN are orthogonal (by the lemma).
Let’s relate SNf back to convolutions and approximate identities. Write
SNfpxq “ÿ
|n|ďN
fpnqe2πinx “ÿ
|n|ďN
ż 1
0
fpyqe´2πiny dy e2πinx “
ż 1
0
ÿ
|n|ďN
e2πinpx´yqfpyq dy
This motivates the definition DN pxq “ř
|n|ďN
e2πinx, so that SNf “ DN ˚ f . We call DN the (Nth
order) Dirichlet kernel.By interpreting DN as a geometric series, we may express it as a ratio of sines. What we’d like
is to say that tDNu is an approximate identity. Unfortunately, the functions tDNu are not evenpositive! Next try: Maybe t|DN |u? This also fails, since
ş
|DN | « lnN . What can we do?
Instead of considering the partial sums, let’s try the Cesaro sums! That is, instead of consideringSNf , let’s consider
σNf “1N
Nÿ
n“0
Snf,
the average of the first N partial sums. We may write
σNf “1N
Nÿ
n“0
Dn ˚ f “ p1N
Nÿ
n“0
Dnq ˚ f
Denote KN “1N
Nř
n“0Dn, the Fejer kernel.
Fact - tKNu is an approximate identity!
Corollary - KN ˚ fL2
Ñ f for all f P L2per.
Corollary - SNfL2
Ñ f for all f P L2per.
Quote - ”Even though the functions KN ˚f are a strictly worse approximation to f than the functionsSNf , it’s much easier to show that they converge to f .”
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 31 - 11/10/2014
Motivation/Intuition - The functions tenu are an orthonormal ”basis” of L2per. We expect/hope that
f “ř
nfpnqen.
Remark 1 - Fourier coefficients can be defined much more generally. Certainly the definition extends
to f P L1. In fact, given a finite measure µ on r0, 1s, define µpnq “ş1
0en dµ.
Consequently, if f P Lpr0, 1s for any p P r1,8s, we can define fpnq “ş1
0fpxqenpxq dx. With that
in mind, we ask:
Question - For which p P r1,8s do we have SNfLpÑ f?
Answer - For p “ 8, this fails. E.g. let f P L8 be discontinuous on a set of positive measure. Infact, it is even possible to find continuous f P L8 such that Snf Ñ f in L8. However, on yourhomework you will show that if f P Cαr0, 1s, where α ą 0, then SNf Ñ f in L8.
Proposition If p P p1,8q, f P Lp, then SNfLpÑ f .
Proof - Hard harmonic analysis.Fact - This result is false for p “ 1.
Question 2 - We know f P L2per implies SNf
L2
Ñ f . When do we have pointwise convergence?Answer - Theorem (Carleson-Hunt) - For p ą 1, f P Lp, we have SNf Ñ f a.e.
˚ ˚ ˚
Theorem - The map L2perr0, 1s Ñ `2, f ÞÑ f “ pfpnqq is a bijective linear isometry.
Proof - We showed last time that f P L2 implies f P `2, so the map is well-defined. Moreover, since
SNfL2
Ñ f , thenf`2 “ lim
NÑ8
ÿ
|n|ďN
|fpnq|2 “ limNÑ8
SNfL2 “ fL2
Thus, the map is an isometry, so it is also injective. Linearity is also clear. It remains to provesurjectivity.
Let panq P `2. We’d like to define f “
ř
nanen, but is this in L2
per? Yes, because
ÿ
Mď|n|ďN
anenL2 “ pÿ
Mď|n|ďN
|an|2q12,
so the sequence of partial sums is Cauchy, thus f P L2per.
Next Goal - We seek to establish the following correspondence:
decay of f Ø ”regularity”/”smoothness” of f
Intuition - For n large, en is ”squiggly”. Too much squiggliness limits the smoothness of f .
Prop (Riemann-Lebesgue Lemma) - If f P L1, then f P L8. Moreover, f vanishes at infinity. That
is, lim|n|Ñ8
fpnq “ 0.
Remark - This does not extend to finite measures. That is, µ a finite measure on r0, 1s does notimply µpnq Ñ 0 at infinity. For example, let µ “ δ0, the delta measure centered at 0. Then for any
n, we have µpnq “ş1
0en dδ0 “ enp0q “ 1.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
However, the Riemann-Lebesgue Lemma says that if µ ! λ, then µpnq Ñ 0 at infinity.Proof - The idea is to approximate f by trigonometric polynomials. Let f P L1. Fix ε ą 0. Thenthere is N such that f ´ σNf1 ă ε. For n ą N , we have
fpnq “ pf ´ σNfqpnq ` pσNfqpnq “ pf ´ σNfqpnq
Thus, by Holder’s inequality, we have
|fpnq| “ |
ż 1
0
pf ´ σNfqen dx| ď f ´ σNf1en8 ă ε
Remark - Previously we showed that f P L2 implies f P `2. This can be seen as another regularityresult.
Differentiability
Definition - We say f P L2per is weakly differentiable with weak derivative g if, for all φ P C8per, we
haveż 1
0
fφ1 “ ´
ż 1
0
gφ
(Intuition: We require that integration by parts works.)Example - If f P C1
per, then f 1 is certainly a weak derivative of f (as we would expect).
Proposition - If f P L2per and has weak derivative f 1 P L2
per, then
f 1pnq “ 2πinfpnq
Remark - This is what we would hope for. If f “ř
nfpnqe2πinx, then differentiation term-by-term
gives us the result above.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 32 - 11/12/2014
Last Time - Goal: ”Regularity” of f Ø Decay of f .Proposition - If f P L2
perr0, 1s, and has a weak derivative f 1 P L2perr0, 1s, then f 1pnq “ 2πinfpnq.
Corollary - If f P L2per has a weak derivative in L2
per, then
ÿ
n
p1` |n|2q|fpnq|2 ă 8
Proof - Since f 1 P L2per, then pf 1pnqq P `2. Now use f 1pnq “ 2πinfpnq.
Definition - Let s ě 0. Define the (periodic) Sobolev space of order s by
Hs :“ tf P L2perr0, 1s |
ÿ
n
p1` |n|2qs|fpnq|2 ă 8u
“ tf P L2perr0, 1s | pp1` |n|
2qs2fpnqq P `2u
For f P Hs, denote fHs “ p1` |n|2qs2fpnq`2 .
Remark - For s ě 0, s P N, we have
Hs “ tf | f P L2 and has sth order weak derivatives in L2u
(If s “ 1, note that f P H1 implies that p2πinfpnqq P `2, so this must correspond to some g P L2per.
By definition, this g must be a weak derivative for f . For s ą 1, this follows by induction.)
Theorem (Sobolev Embedding Theorem) - If f P Hs and s ą 12 , then f is continuous. That is, f
agrees a.e. with a continuous function.Corollary - If f P Hs, s ą n` 1
2 , then f P Cn.Proof (of Theorem) - It’s enough to prove the following result:
DC ą 0, f8 ď CfHs @f P Hs X C8per p˚q
That is, the inclusion map Hs X C8per ãÑ L8 is continuous.
Assuming p˚q, let f P Hs and let pφnq Ď Hs X C8per such that φnHsÑ f (e.g. φn “ Snf). Now
pφnq is Cauchy in Hs, so by p˚q, pφnq is Cauchy in L8, thus it converges uniformly. In particular,φn Ñ f uniformly, so f is continuous.
It remains to prove p˚q. For any x P r0, 1s, we have fpxq “ř
nfpnqenpxq. Therefore,
f8 ďÿ
n
|fpnq| “ÿ
n
p1` |n|2qs2|fpnq| ¨ 1p1`|n|2qs2
ď fHspÿ
n
1p1`|n|2qs q
12,
where the final step follows from the Cauchy-Schwarz inequality and the fact that C :“ př
n
1p1`|n|2qs q
12 ă
8 for s ą 12 .
Goal - Lebesgue differentiation
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Suppose µ is a finite measure on Rd and µ ! λ. By Radon-Nikodym, there is f P L1 such thatdµ “ f dλ. We expect that
fpxq “ limrÑ0`
µpBpx,rqqλpBpx,rqq
Lebesgue proved that this is true a.e.Step 1: Considering the maximal function, defined by
Mµpxq :“ suprą0
|µ|pBpx,rqqλpBpx,rqq
If f P L1, then define Mfpxq :“ suprą0
1λpBpx,rqq
ş
Bpx,rq|f | dλ. What can we prove about these objects?
Theorem - If µ is a finite measure, then λptMµ ą αuq ď 3d
α µ. If f P L1, then λptMf ą αuq ď3d
α f1.
Want - To prove this theorem, we’d like f P L1 ùñ Mf P L1. Unfortunately this is false, but wewill get Mf P L1,8, which is enough to prove the theorem. Incidentally, we get something nicer forLp, p ą 1.Fact (Homework) - If f P Lp, p P p1,8s, then Mf P Lp and Mfp ď Cpfp, where Cp dependsonly on d and p, not on f .Lecture 33 - 11/14/2014
Homework Discussion - Given f P C1perr0, 1s, it’s easy to see that f´τhf8 « Ophq, since f´τhf ď
f 18h. On your homework, you’ll show something more general: If f P L2 and has weak derivativeDf P L2, then f ´ τhf2 « Ophq.
Recall - Lebesgue Differentiation Theorem: Given µ a finite signed measure on Rd, µ ! λ, we
have dµdλ pxq “ lim
rÑ0`
µpBpx,rqqλpBpx,rqq holds λ-a.e. To prove this, we first introduce new quantities that will
(eventually) provide upper bounds in our proof.Definition - Let µ be a finite measure on Rd, f P L1pRq. Define
Mµpxq “ suprą0
|µ|pBpx,rqq|λpBpx,rqq
Mfpxq “ suprą0
1λpBpx,rqq
ż
Bpx,rq
|f |
Theorem (Hardy-Littlewood Maximal Inequality) - There is a constant C ą 0 (can choose C “ 3d)
such that if µ is a finite measure on Rd, then for all α ą 0 we have
λptMµ ą αuq ď Cα µ
Proof - IOU.Remark - Given f P L1, we’d like to conclude that Mf P L1, perhaps Mf1 ď Cf1 for someconstant C. This is false! For example, take f “ χr0,1s. For x large, we have Mfpxq ě 1
2x R L1!
Proof (of Lebegue Differentiation) - By Radon-Nikodym, it’s enough to show that for any positive
function f P L1, we have
limrÑ0`
1λpBpx,rqq
ż
Bpx,rq
|f | dλ “ fpxq λ-a.e.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
The main idea is approximate f by a continuous function, since the result clearly holds for continuousfunctions.
Fix f P L1. Define
Ωfpxq “ lim suprą0
1λpBpx,rqq
ż
Bpx,rq
|fpyq ´ fpxq| dy
To prove the result, it’s enough to show that Ωf “ 0 a.e. (sinceş
|f | ďş
|f ´ fpxq| ` fpxq). Notethat if f is continuous, we are done. If not, fix ε ą 0 and write f “ g`h, where g P L1 is continuousand h1 ă ε. Note that
Ωf ď Ωg ` Ωh “ Ωh
To estimate Ωh, it’s enough to consider the following (surprisingly loose) upper bound:
Ωhpxq “ lim suprą0
1λpBpx,rqq
ż
Bpx,rq
|hpyq ´ hpxq| dy
ď
´
suprą0
1λpBpx,rqq
ż
Bpx,rq
|h|¯
` |hpxq|
“ pMh` |h|qpxq
By the Hardy-Littlewood Maximal Inequality and Chebyshev’s Inequality, we have, for α ą 0,
λptΩf ą αuq ď λptΩh ą αuq
ď λptMh ą α2 uq ` λpt|h| ą
α2 uq
ď 3d
α2h1 `1α2h1
ă 2¨3d`2α ¨ ε
The quantity λptΩf ą αuq is independent of ε, so take the limit ε Ñ 0` to get λptΩf ą αuq “ 0.Thus, the set tΩf ą 0u is λ-null, so Ωf “ 0 a.e.
It remains to prove the Hardy-Littlewood Maximal Inequality. To do this, we state a lemma(which we’ll prove next time).
Lemma (Vitali Covering Lemma) - Let E Ď Rd, tB1, . . . , BNu a set of balls such that E ĎŤ
1ďiďN
Bi.
Then there is a subcollection tB1iu Ď tBiu of disjoint balls such that E ĎŤ
i
3B1i.
Notation - If B “ Bpx, rq, then 3B :“ Bpx, 3rq.Proof (Hardy-Littlewood) - It’s enough to prove the result for µ positive (exercise). Let α ą 0and define E “ tMµ ą αu. Fix K Ď E compact. For each x P K, there is rx ą 0 such thatµpBpx, rxqq ą αλpBpx, rqq. These balls cover K, so let tB1, . . . , BNu be a finite subcover of K. ByVitali, there is a disjoint subcollection tB11, . . . , B
1Mu such that K Ď 3B1i. Now
λpKq ďMÿ
i“1
λp3Biq “ 3dMÿ
i“1
λpB1iq ă3d
α
Mÿ
i“1
µpB1iq “3d
α µpMď
i“1
B1iq ď3d
α µ
Taking the supremum over compact K Ď E, we have λpEq ď 3d
α µ, as desired.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 34 - 11/17/2014
Lemma (Vitali) - Let E Ď Rd, tB1, . . . , BNu a collection of balls whose union covers E. Then there
is a disjoint subcollection tB1iu such that the balls t3B1iu cover E. (Here, B “ Bpx, rq ùñ 3B :“Bpx, 3rq.)Proof - By reordering, we may assume that radiuspB1q ě radiuspB2q ě ¨ ¨ ¨ . Choose B11 “ B1.Consider all balls disjoint from B11. Choose B12 to be the one of largest radius. Note that if BkXB
11 ‰
H for some k, then 3B11 Ě Bk.Given tB11, . . . , B
1ku, let B1k`1 be the ball of largest radius that is disjoint from B11, . . . , B
1k (if
there is such a ball). Eventually, this process terminates. Now
• These balls are clearly disjoint.
• These balls cover E. Why? Given Bi, either Bi “ B1j for some j, or Bi meets some B1j oflarger radius, in which case Bi Ď 3B1j .
Corollary (of Lebesgue Differentiation) - Let A P LpRdq. Then
limrÑ0`
λpAXBpx,rqqλpBpx,rqq “ χApxq λ-a.e.
Proof - Apply Lebesgue differentiation with f “ χA.Remark - Given µ a finite measure on Rd, we know
µ ! λ ùñ limrÑ0`
µpBpx,rqqλpBpx,rqq “
dµdλ pxq λ-a.e.
What if µ! λ? Decompose µ “ µS ` µAC , where µAC ! λ and µS K λ. We know
limrÑ0`
µACpBpx,rqqλpBpx,rqq “
dµACdλ pxq λ-a.e.
Claim (Homework) - limrÑ0`
|µS |pBpx,rqqλpBpx,rqq “
#
0 λ-a.e.
8 µS-a.e.
˚ ˚ ˚
The Fundamental Theorem of Calculus
Recall - For Riemann integration, if we define F pxq “şx
0f dx, where f is continuous, then F is
differentiable and F 1 “ f .Theorem - Let f P L1pRq and define F pxq “
şx
0f . Then F is differentiable and F 1 “ f a.e.
Proof - For h ą 0, we have F px`hq´F px´hq2h “ 1
2h
şx`h
x´hf “ µpBpx,hqq
λpBpx,hqq , where µpAq “ş
Af . By the
Lebesgue Differentiation Theorem, we have
fpxq “ dµdλ pxq “ lim
hÑ0`
µpBpx,hqqλpBpx,hqq “ F 1pxq λ-a.e.
Recall - For Riemann integration, if f P C1, thenşb
af 1 “ fpbq´ fpaq. Another way to express this is
f P C1 ùñ @x
ż x
0
f 1 “ fpxq ´ fp0q
We would like to a version of this result for differentiable functions, dropping the C1 assumption.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
• Attempt 1: Is it enough to assume that f is differentiable a.e.? The answer is no, since (forexample), we may have f 1 R L1r0, 1s. For example, let fpxq “ lnx for x P p0, 1s, fp0q “ 0. Thenf is differentiable a.e. with f 1 “ 1
x , butşx
0f 1 “ 8 for all x ą 0. In particular, f 1 R L1r0, 1s.
• Attempt 2: What if we assume f is differentiable a.e. and f 1 P L1? This is still notenough. For example, let fpxq be the Cantor function, which satisfies f 1 “ 0 a.e. In particular,ş1
0f 1 “
ş1
00 “ 0 ‰ 1 “ fp1q ´ fp0q.
To obtain the desired result, we introduce a new notion:
Definition - We say f : ra, bs Ñ R is absolutely continuous if, for any ε ą 0, there is δ ą 0 such thatfor all finite sets tpxi, yiquiďN of disjoint subintervals of ra, bs, we have
Nÿ
i“1
|xi ´ yi| ă δ ùñNÿ
i“1
|fpxiq ´ fpyiq| ă ε
Note - f absolutely continuous implies f continuous.
Proposition - Let g P L1ra, bs such that fpxq ´ fpaq “şx
ag. Then f is absolutely continuous.
Proof - Fix ε ą 0. We know g is equi-integrable since λpra, bsq is finite. Then there is δ ą 0 such
that λpAq ă δ impliesş
A|g| ă ε. Now suppose that tpxi, yiquiďN are disjoint with
Nř
i“1
|xi ´ yi| ă δ.
Then λpNŤ
i“1
pxi, yiqq ă δ, soş
NŤ
i“1pxi,yiq
|g| ă ε. Now
Nÿ
i“1
|fpxiq ´ fpyiq| “Nÿ
i“1
|
ż yi
xi
g| ď
ż
NŤ
i“1pxi,yiq
|g| ă ε
Theorem - A function f : ra, bs Ñ R is absolutely continuous if and only if
1. f is differentiable a.e.
2. f 1 P L1
3. For all x P ra, bs, fpxq “ fpaq `şx
af 1
Proof - We just proved the reverse direction. We will prove the converse next time.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 35 - 11/19/2014
Theorem - A function f : ra, bs Ñ R is absolutely continuous if and only if
1. f is differentiable a.e.
2. f 1 P L1
3. For all x P ra, bs, fpxq “ fpaq `şx
af 1
Proof - We already proved the reverse direction last time. For the converse, suppose that f isabsolutely continuous.
Lemma 1 Suppose f is absolutely continuous and strictly increasing. Then (1),(2), and (3) hold.
Remark We need absolute continuity. Continuity + strictly increasing isn’t enough,e.g. let fpxq “ gpxq ` x, where gpxq is the Cantor function.
Proof Define µpAq “ λpfpAqq. Note, f´1 is increasing, thus measurable, so A P Bimplies fpAq P B.
Claim µ ! λ
Proof Suppose λpAq “ 0. We need to show µpAq “ λpfpAqq “ 0.
Fix ε ą 0 and let K Ď fpAq be compact. We know f´1pKq iscompact and f´1pKq Ď A, so λpf´1pKqq “ 0. Choose δ ą 0 suchthat, given disjoint tpxi, yiquiďN with
ř
1ďiďN
|xi ´ yi| ă δ¡ we haveř
1ďiďN
|fpxiq ´ fpyiq| ă ε. Since λpf´1pKqq, we may choose such
tpxi, yiquiďN which cover f´1pKq. Now
λpKq ďÿ
1ďiďN
λpfpxi, yiqq “ÿ
1ďiďN
|fpxiq ´ fpyiq| ă ε
Since ε ą 0 was arbitrary, then λpKq “ 0. Since K Ď fpAq wasarbitrary compact, then λpfpAqq “ 0.
By Radon-Nikodym, there is g P L1pra, bsq such that dµ “g dλ, so (1) and (2) hold with f 1 “ g. For (3), write
fpxq ´ fpaq “ λppfpaq, fpxqqq “ µppa, xqq “
ż x
a
g dλ “
ż x
a
f 1 dλ
Lemma 2 Suppose f is absolutely continuous and increasing. Then (1), (2), and(3) hold.
Proof Apply Lemma 1 to the function gpxq “ x ` fpxq. Now g is differentiablea.e., g1 P L1, so the same holds for f . For (3), write
fpxq “ gpxq ´ x “ gpaq `
ż x
a
g1 ´ x
“ fpaq ` a`
ż x
a
pf 1 ` 1q ´ x
“ fpaq ` a`
ż x
a
f 1 ` px´ aq ´ x
“ fpaq `
ż x
a
f 1
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Definition - We say f : ra, bs Ñ R has bounded variation if the function
F pxq :“ sup∆
ÿ
xiP∆
|fpxiq ´ fpxi´1q|, x P ra, bs
is finite, where ∆ ranges over all partitions of ra, xs. We call F pxq the variation off .
Remark - If f : ra, bs Ñ R has bounded variation, then f can be written as adifference g´h of increasing functions (in fact, the converse holds). This is because
f “ pF`fq´pF´fq2 , where F˘f
2 are both increasing.
Lemma 3 Suppose f is absolutely continuous. Then (1), (2), and (3) hold.
Proof Claim If f is absolutely continuous, then f has boundedvariation. Moreover, the variation of f is absolutely continuous.
Proof Let F denote the variation of f . Set ε “ 1 and let δ ą 0such that if tpxi, yiquiďN are disjoint and
ř
1ďiďN
|xi ´ yi| ă δ, thenř
1ďiďN
|fpxiq ´ fpyiq| ă 1. Let ∆ be any partition of ra, bs. Refine
it by setting∆1 “ ∆Y ptnδ | n P Zu X ra, bsq
Nowř
xiP∆
|fpxiq ´ fpxi´1q| ă r b´aδ s. In particular, F is finite. The
proof that F is absolutely continuous is left as an exercise.
By the Claim, we may write f “ pF`fq´pF´fq2 , a difference of increasing, ab-
solutely continuous functions. By linearity of the integral and of differentation, wehave the desired result.
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 36 - 11/21/2014
Change of Variables
Theorem - Let U, V Ď Rd be open, φ : U Ñ V a C1 bijection. Then for all f P L1pV q, we haveż
V
f dλ “
ż
U
f ˝ φ|det∇φ| dλ,
where ∇φ “ pBjφiq is the Jacobian of φ.Proof - For A Ď U , A P BpRdq, define µpAq “ λpφpAqq. (Here, µ is defined so that λ on V is thepushforward of µ with respect to φ.) To prove the theorem, we need only prove a series of lemmas.
Lemma 1 For all A P BpUq, φpAq P BpV q. Moreover, µ is a Borel measure on U .
Lemma 2 µ ! λ
Lemma 3 dµdλ “ |det∇φ| λ-a.e.
Once these lemmas are proven, we have the result. This is becauseż
V
f dλ “
ż
U
f ˝ φ dµ (pushforward)
“
ż
U
f ˝ φdµdλ dλ pµ ! λq
“
ż
U
f ˝ φ| det∇φ| dλ
Proof (Lemma 1) Suppose A P BpUq. Define
Λ “ tA P U | φpAq P BpV qu
This is a Λ-system because
• φpUq “ V P BpV q• If A Ď B in Λ, then φpBzAq “ φpBqzφpAq P BpV q (φ is bijective).
• pAnq Ď Λ increasing implies φpŤ
nAnq “
Ť
nφpAnq P BpV q.
Moreover, Λ Ě K :“ tK Ď U | K is compactu, which is a π-system (Λ Ě Kbecause if K Ď U is compact, then so is φpKq Ď V , so φpKq P BpV q). Thus,Λ Ě σpKq “ BpUq. (Compact sets generate the Borel σ-algebra since for anynonempty closed C and x P C, we have C “
Ť
nC X Bpx, nq, a union of compact
sets.)
Proof (Lemma 3) Assuming Lemma 2, we have dµdλ pxq “ Dµpxq :“ lim
rÑ0`
µpBpx,rqqλpBpx,rqq
holds λ-a.e., so it’s enough to prove that Dµpxq “ |det∇φx| for all x P U . Below,we consider simple cases for φ and proceed to generalize.
• Step 0: Suppose φ is linear, i.e. φpxq “ Mx for some invertible M P MdpRq.Then µ “ λ˝φ is translation invariant on Rd and finite on bounded sets, so bya previous result there is C ą 0 such that µpAq “ CλpAq, so now Dµpxq “ C.Let A be the unit cube. Then by a previous homework, we have
C “ µpAq “ |detM | “ |det∇φ| ùñ Dµpxq “ |det∇φx|
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
• Step 1: We may assume without loss that x “ 0 and φpxq “ 0 (we can alwaysreduce to this case by composing φ with translations).
– Case I: Suppose det∇φ0 “ 0, i.e. ∇φ0 is not invertible. We want to showDµp0q “ 0. To do this, denote T “ ∇φp0q, T : Rd Ñ Rd. Since detT “ 0,then dimprangepT qq ă d, so λprangepT qq “ 0.Now let ε ą 0. By definition of differentiation, we have
0 “ lim|x|Ñ0
|φpxq´φp0q´∇φ0px´0q||x| “ lim
|x|Ñ0
|φpxq´Tx||x|
Then there is r0 ą 0 such that |x| ă r0 implies |φpxq´Tx| ă ε|x|. We knowthat for any r, T pBp0, rqq is contained in a pd ´ 1q-dimensional subspaceof Rd and diampT pBp0, rqqq ď T ¨ 2r. Thus,
µpBp0, rqq “ λpφpBp0, rqqq ď pT 2rqd´1 ¨ εr
Now0 ď Dµp0q ď lim
rÑ0`
p2T qd´1εrd
rdλpBp0,1qq“ p2T qd´1ε
Taking the limit εÑ 0`, we have Dµp0q “ 0 “ det∇φ0
To be continued...
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Lecture 37 - 11/24/2014
Theorem - Let U, V Ď Rd be open, φ : U Ñ V a C1 bijection. Then for all f P L1pV q, we have
ż
V
f dλ “
ż
U
f ˝ φ|det∇φ| dλ,
where ∇φ “ pBjφiq is the Jacobian of φ.Proof (continued from last time) - We defined µpAq “ λpφpAqq, a measure on BpUq. We still need
to show that µ ! λ and Dµ :“ limrÑ0`
µpBp¨,rqqλpBp¨,rqq “ |det∇φ|. We consider Dµpxq for x “ 0, φpxq “ 0.
• We’ve already proven that Dµ “ |det∇φ| if φ is linear or det∇φ0 “ 0.
• Case II(a): Suppose ∇φ0 “ I, the identity matrix. Then lim|x|Ñ0
|φpxq´x||x| Ñ 0, so there is r0 ą 0
such that |φpxq ´ x| ă ε|x| for |x| ă r0. Now φpBp0, rqq Ď Bp0, p1` εqrq implies
Remark - It turns out to be difficult to show that Dµ ě 1. What we’d like is to show thatφpBp0, rqq Ě Bp0, p1 ´ εqrq (easy if φ´1 P C1). This involves Brouwer’s fixed-point theoremfrom topology, so we omit the proof here. (For more details, see pp. 150-151 of Rudin’s Realand Complex Analysis, 3rd edition.) If you’re unsatisfied, simply assume that φ´1 P C1 andthen the result follows (see Remark after this proof).
• Case II(b): det∇φ0 ‰ 0. Set T “ ∇φ0 and consider ψ “ T´1φ. Certainly ∇ψ0 “ I, so byCase II(a),
1 “ limrÑ0`
λpψpBp0,rqqqλpBp0,rqq
“ limrÑ0`
λpT´1φpBp0,rqqqλpBp0,rqq
“ |detT |´1 limrÑ0`
λpφpBp0,rqqqλpBp0,rqq
ùñ |detT | “ limrÑ0`
µpBp0,rqqλpBp0,rqq ,
so we are done. This finishes the proof of Lemma 3 from last time.
It remains to show that µ ! λ. By writing U as an increasing union of compact sets, we havethat µ is regular (Homework 3, Question 2). Thus, it suffices to show that if K Ď U is compact andλpKq “ 0, then µpKq “ 0.
Given such K, let ε ą 0 and let W Ě K be open with λpW q ă ε and C :“ supxPW
|∇φx| ă 8 (this
is possible since we may choose W to have compact closure). By the Mean Value Theorem, for anyx, y P Bpx0, r0q ĎW , we have, for i “ 1, . . . , d,
21-720 Measure and Integration, Fall 2014Notes typeset by Adam Gutter
Now cover K with balls tBpx, rxq | x PW,Bpx, 3rxq ĎW u. By compactness, we may pass to a finitesubcover. By the Vitali covering lemma, we may pass to a finite subset of disjoint balls tB1iu suchthat K Ď
Ť
i
3B1i. Now
µpKq “ λpφpKqq ď λpď
i
φp3B1iqq
ďÿ
i
λpφp3B1iqq
ďÿ
i
pdCqd3dλpB1iq
“ pdCqd3dλpď
i
B1iq
ď pdCqd3dλpW q
ă pdCqd3dε
Taking the limit ε Ñ 0`, we have µpKq “ 0, so we are done. This finishes the proof of thetheorem.Remark - Another way to show that the theorme holds for φ´1 P C1 is through an appeal tosymmetry. If we only prove that Dµ ď 1 in Case II(a) above, we end up showing that