Skip to content

Measurable Function

Functions discussed in this chapter are defined on measurable sets, whose value could be generalized real number, i.e. R including ±.

Definition of Measurable Function

Assume function f has defined on measurable set D. If αR, set

{f>α}=D(f>α):={xD:f(x)>α}

is also measurable, then we call f is a measurable function on D.

Example. Continuous function on an interval DR is a measurable function.

  • For an open interval D, this is easy.

  • For an arbitrary interval D, the original image of f>α would differ from open sets with at most 2 points.

Characteristic Function

Assume AX, then define a characteristic function A as

χA(x)={1,xA0,xXA

Measurability of Characteristic Function

For measurable set D, denote χD as its characteristic function, then

{χD>α}={,α1D,0α<1R,α<0

So we could prove χD is measurable function.

Properties

Properties of Measurable Function

Assume f(x) is measurable on E, then the following sets are all measureable.

(i) {fα} (ii) {fα}

(iii) {f<α} (iv) {f=α}

(v) {f<+} (vi) {f=+}

(vii) {f>} (viii) {f=}

  • (i) =E{f>α}

  • (ii) =k=1{f>α1k} (length decrease, so we choose inferior)

  • (iii) using (ii) with same logic of proving (i)

  • (iv) using (i) amd (ii)

  • (v) =k=1{f<k} (length increase, so we choose superior)

  • (vi) use (v) with same logic of proving (i)

  • (vii) same logic of proving (i)

  • (viii) use (vii) with same logic of proving (i).

The above (i), (ii) and (iii) are equivalent definition of measurable function, but (iv) is not.

Example. f is measurable on measurable set D, iff rQ, {f>r} is measurable.

  • "". Let α, α=r.

  • "". Notice that α,{rn}, such that rnR, and r1>r2>, and rnα, so

{f>α}=n=1{f>rn}

which means f is measurable.

When {f=r} is measurable, we could not deduce the same result. Bacause let A to be a non-measurable set, define

f={3,xA2,xA

then rQ, {f=r}= is measurable, but let α=1.5,

{f>1.5}=A

is not measurable, so f is not measurable.

Region Extension

(i) Assume generalized real function f(x) is defined on D1D2. If f(x) is measurable on D1 and D2, then f(x) is also measurable on D1D2.

(ii) Assume f(x) is measurable on D. If AD is also measurable, then f(x) is also measurable on A.

(i) {xD1D2:f>α}={xD1:f>α}{xD2:f>α} and by measure operation.

(ii) {xA:f>α}={xD:f>α}A and by measure operation.

Lemma of functions

Assume f(x), g(x) are measurable on E, then {f>g} is also measurable.

εr=1r,rN+, we have

{f>g}=r=1{f>εr}{εr>g}

so {f>g} is measurable.

Basic Operations

Basic Operations of Measurable Function

Assume f(x), g(x) are measurable function on D, then the following functions are measurable.

(i) cf(x)(cR). (ii) f(x)+g(x). (iii) f(x)g(x).

(i) move c to the side of α.

(ii) same logic with Lemma.

(iii) use 4f(x)g(x)=[f(x)+g(x)]2[f(x)g(x)]2. You only need to consider whether f2(x) are measurable, which is easy to see. Notice that α0,

{f2>α}={f>α}{f<α}

is measurable, and α<0,

{f2>α}=D

is measurable.

Example. Assume f is defined on measurable set D, prove: if f2 is measurable on D and {f>0} is measurable, then f is measurable on D.

  • α0,
{f>α}={f2>α2}{f>0}

is measurable.

  • α<0,
{f>α}={f>0}{α<f0}={f>0}[{f2<α2}{f0}]

is measurable.

Example. Assume f is a measurable function on D, prove: for all open set GR and closed set FR,

f1(G),f1(F)

are measurable.

Known from configuration of set, open set GR, there exists {an} and {bn}, such that an<bn, and G=n=1(an,bn), so

f1(G)=n=1{an<f<bn}

is measurable. Notice that f1(F)=f1(RFc)=Df1(Fc) is also measurable.

Limit superior & limit inferior

Limit superior & limit inferior

Sequence of Function.

Similarly, given a sequence of function {fn(x)}n1 we have limit superior

limnfn=infn1{supkn{fk(x)}}

and limit inferior

limnfn=supn1{infkn{fk(x)}}

Theorem for measurability of Limit

Assume {fn(x)}n1 is a sequence of measurable functions on measurable set D, then the following functions are all measurable.

(1)supn1fn(x),infn1fn(x)(2)limnfn(x),limnfn(x)
  • For expression 1, we can write it as
{supn1fn(x)>α}=n{fn(x)>α}

where the latter one is measurable.

  • For expression 2, we could use conclusion from 1.

Example. Assume fn is a measurable function on D. Prove: Set

{xD:fn converges.}

is measurable.

fn converges, iff its limit superior and inferior are of the same. So we define

F(x)=limnfn,f(x)=limnfn,

both of which are measurable by Theorem for limit. Then consider

{F=f}=D{F>f}

since the latter is measurable according to Lemma of functions, {F=f} is measurable.

Almost Everywhere

Definition of Almost Everywhere of a Proposition

Assume D is measurable, and proposition P(x) is related with points in D. If there exists a zero-measure set ED, such that P(x) holds on DE, then we call P(x) holds almost everywhere on D.

We denote this statement as

P(x) a.e.xD.

Approximation

Approximation with Simple Function

Simple Function

Assume f is a function on measurable set D. If f(D) is composed of finite points {ak}k=1n, and sets Ek={f=ak}(k=1,2,,n) are all measurable, then we call f is a simple function on D.

Apparently, here f(x) could be expressed by finite linear combination of characteristic functions, i.e.

f(x)=k=1nakχEk(x)

Approximate measurable function with sequence of simple functions

If f(x) is measurable on D, then there exists a sequence of simple functions {fk}k=1, such that for every point xD, {fk}k1 converges to f(x). Specially, if

(i) f is non-negative, then {fk}k1 is monotonically increasing.

(ii) f is bounded, then {fk}k1 converges uniformly to f(x).

Use a special construction.

Formulate a sequence of function which converges to f(x). This is a little tricky.

Define

fn(x)={n,f(x)>nk12n,k12nf(x)<k2n,k=n2n+1,n2n,n2nn,f(x)<n

Look the following image for intuitional impression.

Now we show that fn(x) corvergese to f(x). For fixed point xD, if

(i) f(x)=. Then fn(x)=n=f(x). The same for f(x)=.

(ii) <f(x)<+. Then for sufficient large number n, there exists unique kn such that n2n+1knn2n and

kn12nf(x)kn2n

which means fn(x)=kn12n, satisfying

0f(x)fn(x)12n.

Let n, we have fn(x) converges to f(x).

  • Specially, if f(x) is non-negative, for f(x)=+, then fn(x)=n is monotonically increasing.

For f(x)<, if f(x)<n, we have f(x)=k12n for some k such that k12n<f(x)<k2n, then for fn+1(x), we partition this interval into two parts, which are

[2k22n+1,2k12n+1),[2k12n+1,2k2n+1)

the former part fn+1(x)=fn(x)=k12n, while the latter part gives fn+1(x)=2k12n+1>fn(x), so it is monotonically increasing.

for >f(x)n, this is similar to .

  • Specailly, if f(x) is bounded, denote its upper bound as M, so when n>M, we have |f(x)fn(x)|<12n on all xD, which means fn(x)f(x).

Example. Assume f is differentiable on R, prove f(x) is measurable.

f is differentiable, so f and f(x+1n) is continuous, thus measurable. So

f(x)=limnf(x+1n)f(x)1n=Δfn(x)

fn is measurable, so f is also measurable.

Uniform Convergnce a.e.

Lemma: use language of Set to describe convergence

Assume {fk(x)}k1, f(x) are measurable function on a.e.xE, and m(E)<. If fk(x)f(x),a.e.xE, then ε>0, let Ek(ε)={xE:|fk(x)f(x)|ε}, then we have

limjm(k=jEk(ε))=0.

We have the following equivalent statement for point-wise convergence.

fk(x)f(x),a.e.xEε>0,N>0,kN,|fk(x)f(x)|<εε>0,N>0,xk=N{xE:|fk(x)f(x)|<ε}(3)ε>0,xN=1k=N{xE:|fk(x)f(x)|<ε}

Note that we could have an equivalent statement at the last

xε>0N=1k=N{xE:|fk(x)f(x)|<ε}

but ε>0 is not denumerable, so the in the following proof, we have to choose a denumerable set like Q or 1r.

Now we have a equivalent statement of fk(x)f(x),a.e.xE, that is, expression 3. Then its opponent proposition fk(x)f(x) could be expressed by

ε>0,xN=1k=N{xE:|fk(x)f(x)|ε}

That is,

(4){xE:fk(x)f(x)}=ε>0N=1k=NEk(ε)

which has zero-measure according to a.e.xE, so by limit operations in measure, we have

limNm(k=NEk(ε))=m(limNk=NEk(ε))=m(N=1k=NEk(ε))=0.

The first equation holds because Ek(ε) is measurable (fk,f are all measurable functions) and k=NEk(ε) is monotonically decreasing when N. The last equation holds because the left side of equation 4 has zero-measure (which means the item to be unioned should have zero-measure).

Egoroff Theorem

Assume {fk(x)}k1 is a measurable function which is bounded on a.e.xE, and m(E)<. If fk(x)f(x),a.e.xE, then δ>0,  measurable set EδE, such that m(Eδ)δ and

fk(x)f(x),xEEδ.

Choose ε=1r and define

Ek(1r)={xE:|fk(x)f(x)|1r},

then by condition of the theorem and Lemma, δ>0, rR, Nr, NNr,

m[k=NEk(1r)]<δ2r.

Then we have to sum up there measure. That is, let Eδ=r=1k=NEk(1r), which satisfies

m(Eδ)r=1m[k=NEk(1r)]<δ.

Check that

EEδ=r=1k=NEkc(1r)=r=1k=N{xE:|fk(x)f(x)|<1r}

Notice that fkf,a.e.xA, iff supxA|fk(x)f(x)|0(k). If we let A=EEδ, this holds apparently.

Note that the condition m(E)< in the above theorem could not be removed.

Example. fn(x)=xn converges to f(x)={0,0<x<11,x=1 point-wisely, but not uniformly. If we remove a sufficiently small set (1δ,1], then fn(x)f(x) uniformly.

Extension of Measurable Function

The following theorem is the main result of this part.

Lusin Theorem

Assume f(x) is a measurable function a.e. on E, then δ>0, closed set FE, s.t. m(EF)<δ, and f(x) is a continuous function on F.

Using Theorem in Mathematical Analysis for uniform convergent function, that is, the convergent function of uniformly convergent function sequence is continuous(Here notice we have to cut off a closed set from E, which might be infinite). So we have to make use of Egoroff Theorem.

(i) f is a simple function on E.

From conclusion in Property of Simple Function, there exists measurable sets Ei(i=1,2,,N), such that E=i=1NEn, EiEj=, and

f(x)=i=1NαiχEi(x)

From Theorem of Approximation for measurable sets, δ>0, because Ei is measurable, closed set FiEi, such that m(EiFi)<δ/N, and χEi(x)C(Fi), or to be more specific, is constant on Fi. So from F:=i=1NFiE=i=1NEi, we have

m(EF)=m[(i=1NEi)(i=1NFi)]m[i=1N(EiFi)]i=1Nm(EiFi)<δ

(ii) m(E)<.

From conclusion in Approximation with simple function, there exists a sequence of simple function {ψk}k1 such that ψk(x)f(x).

Then since m(E)<, we could use Egoroff Theorem for the above function sequence. That is, δ>0,  measurable function Eδ, such that m(EEδ)<δ/2 and

ψk(x)f(x),xEEδ.

The following method is a little tricky. k, from (i) we could know for simple function ψk(x), there exists closed set FkEEδ, such that m((EEδ)Fk)<δ/2k+1 and ψk(x)C(Fk). Define F=k=1Fk which is still a closed set.

Now we have to prove that the part we remove is small enough. That is, since

EF=E(k=1Ek)[Eδ][(EEδ)(k=1Ek)]

so

m[EF]m[Eδ]+m[(EEδ)(k=1Ek)]δ2+m[k=1((EEδ)Fk)]δ2+k=1m[((EEδ)Fk)]δ2+δ2=δ

Thus, ψk(x) uniformly converges to f(x) on F, by Theorem in Mathematical Analysis for uniform convergent function, meaning f(x)C(F).

(iii) Last situation. m(E)=.

We have to transform this problem into (ii), that is, use finite-measure set to intersect.

Define Ek:=E{xE:k1|x|<k}. Then E=k=1Ek, and EiEj=.

So k, there exists closed sets FkEk, such that m(EkFk)<δ/2k+1, and fC(Fk). Consider define Eδ=k=1Fk (see difference with (ii)) which is just a measurable set, or to be more specific, a Fσ set. But again by using Theorem of Approximation for measurable sets, there exists a closed set FEδ, such that m(EδF)<δ/2.

Now we check the measure of the removal part. Since

EF(EEδ)(EδF),

we have

m(EF)m[EEδ]+m[EδF]=m[E(k=1Fk)]+m[EδF]m[k=1(EkFk)]+δ2k=1m(EkFk)+δ2=δ2+δ2=δ

Thus we are done.

Continuous Extension Theorem

Assume f(x) is a measurable function bounded a.e. on measurable set ER, then δ>0, there closed set FδE and continuous function g(x) on whole space R, such that m(EFδ)<δ and g(x)=F(x),xFδ, and

(5)supxRg(x)=supxFδf(x),infxRg(x)=infxFδf(x).

which naturally satisfies the condition in the theorem.

By Lusin Theorem, δ>0, there exists closed set FE, such that m(EF)<δ, and fC(F). Now the only problem is, how to extend this continuoud function on F to all space F and satisfy the so-called boundary condition 5.

By Conposition of closed set, F could be expressed by RG, where G is a union of denumerable open intervals which are mutually disjoint. That is,

Fc=RF=i=1(ai,bi).

where at most 2 open intervals are infinite intervals. From here we formulate a function on R

g(x)={f(x),xF,f(ai)+f(bi)f(ai)biai(xai),x(ai,bi),f(ai),x(ai,+),f(bi),x(,bi).

Convergence in Measure

Example. nN+, !k,iN+, such that n=2k+i, where 0u2k. Formulate a function sequence

fn(x)=χ[i2k,i+12k](x),n=1,2,,x[0,1]

So for every point x[0,1], fn(x) does not converge.

For the above example, we could consider its convergence in measure, or probability.

Still the above example.

If we define a set

En(ε)={x[0,1]:|fn(x)0|ε},

then its measure

m(En(ε))=12k0

which means for sufficient large n, the frequency of 0 is close to 1. So we call {fn(x)} converges to f(x)=0 in measure.

Definition of convergence in measure

Assume function f and function sequence {fn}n1 are measurable function bounded a.e on D. If δ>0, we have

limnm({|fnf|δ})=0

then we call {fn} converges to f in measure on region D.

Using εN language, we have

ε>0,σ>0,NN+,nN,m({|fnf|>σ})<ε.

Generally speaking, "convergence point-wisely" has no apparent relationship with "convergence in measure", readers could see the following exmaple.

Example. Define fn(x)=χ(n,)(x), f(x)=0, then xR, we have fn(x)f(x). But if we let δ=1/2, then {|fnf|1/2}=(n,), whose measure m({|fnf|1/2})=, which does not converges to f(x) in measure.

Uniqueness of Convergent Function

Uniqueness of Convergent Function

If fn(x) converges to f(x) and g(n) on E in measure, then

f(x)=g(x),a.e.xE.

Because |f(x)g(x)||f(x)fn(x)|+|fn(x)g(x)| (at least one set of the right side should be larger then ε/2 if |f(x)g(x)|>ε), so

{|f(x)g(x)|>ε}{|f(x)fn(x)|>ε2}{|g(x)fn(x)|>ε2}

So by subadditivity, we have

m({|f(x)g(x)|>ε})m({|f(x)fn(x)|>ε2})+m({|g(x)fn(x)|>ε2})0,n

Relationship between two Convergences

Lebesgue's Theorem: Relationship between two Convergences

Assume m(E)<, {fn(x)} is a measurable function sequence a.e. on E. If {fn} converges to function f(x) a.e. on E, where f(x) is a bounded function a.e.xE, then fn(x) converges to f(x) in measure.

Use m(E)< to show that fn(x)f(x), which could gives convergence in measure.

By Egoroff Theorem, because m(E)<, we have point-wise convergent function sequence {fn} also converges uniformly to f(x). That is, we have δ>0, measurable set Eδ, such that m(Eδ)<δ and

fn(x)f(x),xEEδ.

Then by definition of uniform convergence, we have ε>0, N>0, nN,

|fn(x)f(x)|<ε,xEEδ

which means the set

{xE:|fn(x)f(x)|ε}Eδ

so

m({xE:|fn(x)f(x)|ε})m(Eδ)<δ.

and we are done.

Note that m(E)< is not necessary, see that we use the uniform convergence to prove, which is the most important.

Riesz Theorem

Assume function f and function sequence {fn} are measurable function a.e. on E, where f and {fn} are bounded a.e. on E. Then

(i) If {fn} converge to f in measure, then there exists subsequence {fnj} convergent to f a.e.xE.

(ii) If m(E)<, and {fn} converge to f a.e.xE, then {fn} converge to f in measure.

We hope to use δ to get a subsequence of fn.

kN+, let δk=12k, so fnmf implies

limnm({|fnf|δk})=0.

So there exsits nk, such that

m({|fnkf|δk})<12k.

Notice {m({|fnf|δk})}n1 is a number sequence, so n1<n2<.

Define

E:=p=1k=p{|fnkf|12k}

and for every p, measure

m(E)m({|fnkf|12k})<k=p12k=12p1.

let p, we have m(E)=0. Thus for every xDE, i.e.

xp=1k=p{|fnkf|<12k}

which means, p01, such that

|fnkf|<12k,kp0

meaning limkfnk=f.

δ>0, ε>0, since m(D)<, by Egoroff Theorem, there exists measurable set ED, such that m(DE)<ε, and fnf,a.e.xE. That is, there exsits N, such that

|fn(x)f(x)|<δ,xE,nN.

which means

m({|fn(x)f(x)|δ})=M(DE)<ε,n>N

meaning

limnm({|fn(x)f(x)|δ})=0.

Lemma: Sufficient and Necessary Condition for convergence in measure

Assume function f and function sequence {fn} are measurable on E where m(E)<. Then fn(x) converges to f(x) in measure, iff subsequence {fnj}, a subsequence {fnji} such that

fnjif(x),a.e.xE.
  • "".

We prove subsequence {fnj}, fnj converges to f in measure, which holds apparently because fn(x) converges to f(x) in measure.

Then apply Riesz Theorem to fnj and we are done.

  • "". We make use of contradiction. That is, if {fn} does not converges to f in measure, i.e. ε0>0,
limnm({|fnf|ε0})0.

So j,nj, s.t.

(6)m({|fnjf|ε0})>0.

But if there exists a subsequence {fnji} such that

fnjif(x),a.e.xE.

then by Lebesgue's Theorem (m(E)<), we have fnjimf, which contradicts inequation 6.