Real Analysis
Department of Mathematics, Aalto University 2022
Contents
1 Lp spaces 1
1.1 Lpfunctions . . . 1
1.2 Lpnorm . . . 4
1.3 Lpspaces for0<p<1. . . 13
1.4 Completeness ofLp. . . 14
1.5 L∞space . . . 20
1.6 Density of continuous functions . . . 25
1.7 Continuity of translation . . . 29
1.8 LocalLpspace . . . 30
2 The Hardy-Littlewood maximal function 32 2.1 Definition of the maximal function . . . 32
2.2 Hardy-Littlewood-Wiener maximal function theorems . . . 35
2.3 The Lebesgue differentiation theorem . . . 44
2.4 The fundamental theorem of calculus. . . 49
2.5 Points of density . . . 51
2.6 The Sobolev embedding . . . 54
3 Convolutions 58 3.1 Convolution. . . 58
3.2 Approximations of the identity . . . 63
3.3 Pointwise convergence. . . 66
3.4 Convergence inLp . . . 70
3.5 Smoothing properties . . . 72
3.6 Partition of unity . . . 81
3.7 The Poisson kernel . . . 83
4 Differentiation of measures 86 4.1 Covering theorems . . . 86
4.2 The Lebesgue differentiation theorem for Radon measures . . . 96
4.3 The Radon-Nikodym theorem . . . 101
4.4 The Lebesgue decomposition. . . .105
4.5 Lebesgue and density points revisited . . . 108
5 Weak convergence methods 112 5.1 The Riesz representation theorem forLp. . . .112 5.2 The Riesz representation theorem for Radon measures . . . .123 5.3 Weak convergence and compactness of Radon measures . . . . 131 5.4 Weak convergence inLp. . . 135 5.5 Mazur’s lemma. . . 141
spaces in analysis. This section gives basic facts about Lpspaces for general measures. These include Hölder’s inequality, Minkowski’s inequality, the Riesz-Fischer the- orem which shows thatLp is a complete space and the corresponding facts for theL∞space.
L p spaces 1
In this section we study theLpspaces in order to be able to capture quantitative information on the average size of measurable functions and boundedness of operators on such functions. The cases 0<p<1,p=1,p=2, 1<p< ∞andp= ∞ are different in character, but they all play an important role in in Fourier analysis, harmonic analysis, functional analysis and partial differential equations. The spaceL1of integrable functions plays a central role in measure and integration theory. The Hilbert spaceL2of square integrable functions is important in the study of Fourier series. Many operators that arise in applications are bounded in Lpfor 1<p< ∞, but the limit casesL1andL∞require a special attention.
1.1 L p functions
Definition 1.1. Letµbe an outer measure onRn,A⊂Rnaµ-measurable set and f:A→[−∞,∞] aµ-measurable function. Then f∈Lp(A), 1Ép< ∞, if
kfkp= µZ
A|f|pdµ
¶1p
< ∞.
TH E M O R A L: Forp=1, f ∈L1(A) if and only if|f|is integrable in A. For 1Ép< ∞,f∈Lp(A) if and only if|f|pis integrable inA.
Remark 1.2. The measurability assumption on f essential in the definition. For example, letA⊂[0, 1] be a non-measurable set with respect to the one-dimensional Lebesgue measure and considerf: [0, 1]→R,
f(x)=
1, x∈A,
−1, x∈[0, 1] \A.
Thenf2=1 is integrable on [0, 1], butf is not a Lebesgue measurable function.
1
Example 1.3. Letf:Rn→[0,∞], f(x)= |x|−nand assume thatµis the Lebesgue measure. Let A=B(0, 1)={x∈Rn :|x| <1}, Ωn =m(B(0, 1)) and denote Ai= B(0, 2−i+1) \B(0, 2−i),i=1, 2, . . .. Then
Z
B(0,1)|x|−n pdx= X∞ i=1
Z
Ai
|x|−n pdx É
X∞ i=1
Z
Ai
2n pidx (x∈Ai=⇒ |x| Ê2−i=⇒ |x|−n pÉ2n pi)
= X∞ i=1
2n pi|Ai| É X∞ i=1
2n pi|B(0, 2−i+1)|
=Ωn
X∞ i=1
2n pi(2−i+1)n (Ωn= |B(0, 1)|)
=Ωn
X∞ i=1
2n pi−ni+n=2nΩn
X∞ i=1
2in(p−1)< ∞, ifn(p−1)<0⇐⇒p<1. Thusf∈Lp(B(0, 1)) forp<1.
On the other hand, Z
B(0,1)|x|−n pdx= X∞ i=1
Z
Ai
|x|−n pdx Ê
X∞ i=1
Z
Ai
2n p(i−1)dx (x∈Ai=⇒ |x| <2−i+1=⇒ |x|−n p>2n p(i−1))
= X∞ i=1
2n p(i−1)|Ai| =Ωn(2n−1)2−n p X∞ i=1
2n pi2−in (|Ai| = |B(0, 2−i+1)| − |B(0, 2−i)|
=Ωn(2(−i+1)n−2−in)=Ωn(2n−1)2−in)
=C(n,p) X∞ i=1
2in(p−1)= ∞,
ifn(p−1)Ê0⇐⇒pÊ1. Thusf∉Lp(B(0, 1)) forpÊ1. This shows that f∈Lp(B(0, 1))⇐⇒p<1.
If A=Rn\B(0, 1), then we denoteAi=B(0, 2i) \B(0, 2i−1),i=1, 2, . . ., and a similar argument as above shows that
f∈Lp(Rn\B(0, 1))⇐⇒p>1.
Observe that f 6∈L1(B(0, 1)) and f 6∈L1(Rn\B(0, 1)). Thus f(x)= |x|−n is a borderline function inRnas far as integrability is concerned.
TH E M O R A L: The smaller the parameterpis, the worse local singularities anLpfunction may have. On the other hand, the larger the parameterpis, the more anLpfunction may spread out globally.
Example 1.4. Assume thatf:Rn→[0,∞] is radial. Thusf depends only on|x| and it can be expressed asf(|x|), wheref is a function defined on [0,∞). Then
Z
Rnf(|x|)dx=ωn−1 Z ∞
0
f(r)rn−1dr, (1.1)
where
ωn−1= 2πn/2 Γ(n/2)
is the (n−1)-dimensional volume of the unit sphereÇB(0, 1)={x∈Rn:|x| =1}.
Let us show how to use this formula to compute the volume of a ballB(x,r)= {y∈Rn:|y−x| <r},x∈Rn andr>0. By the translation and scaling invariance, we have
rnΩn=rnm(B(0, 1))=m(B(x,r))=m(B(0,r))
= Z
RnχB(0,r)(y)d y= Z
Rnχ(0,r)(|y|)d y
=ωn−1
Z r
0 ρn−1dρ=ωn−1
rn n . In particular, it follows thatωn−1=nΩnand
m(B(x,r))= 2πn/2 Γ(n/2)
rn
n = πn/2 Γ(n2+1)rn. Letr>0. Then
Z
Rn\B(0,r)
1
|x|αdx= Z
Rn
1
|x|αχRn\B(0,r)(x)dx
=rn Z
Rn
1
|rx|αχRn\B(0,r)(rx)dx
=rn−α Z
Rn
1
|x|αχRn\B(0,1)(x)dx
=rn−α Z
Rn\B(0,1)
1
|x|αdx< ∞, α>n, and, in a similar way,
Z
B(0,r)
1
|x|αdx=rn−α Z
B(0,1)
1
|y|αd y=rn−α Z
B(0,1)
1
|x|αdx< ∞, α<n.
Observe, that here we make a change of variablesx=r y.
On the other hand, the integrals can be computer directly by (1.1). This gives Z
Rn\B(0,r)
1
|x|αdx=ωn−1
Z ∞
r ρ−αρn−1dρ
= ωn−1
−α+nρ−α+n
¯
¯
¯
¯
∞ r
=ωn−1
α−nr−α+n< ∞, α>n and
Z
B(0,r)
1
|x|αdx=ωn−1
Z r
0 ρ−αρn−1dρ
= ωn−1
−α+nρ−α+n
¯
¯
¯
¯
r 0
=ωn−1
α−nrn−α< ∞, α<n.
Remarks 1.5:
Formula (1.1) implies following claims:
(1) If|f(x)| Éc|x|−αin a ballB(0,r),r>0, for someα<n, thenf∈L1(B(0,r)).
On the other hand, if |f(x)| Êc|x|−α in B(0,r) for someα>n, then f ∉ L1(B(0,r)).
(2) If|f(x)| Éc|x|−α in Rn\B(0,r) for someα>n, then f ∈L1(Rn\B(0,r)).
On the other hand, if|f(x)| Êc|x|−αin Rn\B(0,r) for someα<n, then f∉L1(Rn\B(0,r)).
Remark 1.6. If f∈Lp(A), then|f(x)| < ∞forµ-almost everyx∈A.
Reason. LetAi={x∈A:|f(x)| Êi},i=1, 2, . . . . Then {x∈A:|f(x)| = ∞}=
\∞ i=1
Ai. By Chebyshev’s inequality
µ({x∈A:|f(x)| = ∞})ɵ(Ai)= Z
Ai
1dµ É
Z
Ai
µ|f| i
¶p
dµ (|f| ÊiinAi) É 1
ip Z
A|f|pdµ
| {z }
<∞
−−−→i→∞ 0. ■
The converse is not true, as the previous example shows.
Remark 1.7. If f∈Lp(A), then {x∈Rn:|f(x)| 6=0} isσ-finite with respect toµ. Reason. LetAi={x∈A:|f(x)| Ê1i},i=1, 2, . . . . Then
{x∈A:|f(x)| 6=0}= [∞ i=1
Ai. By Chebyshev’s inequality
µ(Ai)=µ({x∈A:|f(x)| Ê1i})= Z
Ai
1dµ Éip
Z
Ai|f|pdµ< ∞ (|f| Ê1i inAi)
for everyi=1, 2, . . . . ■
1.2 L p norm
If f∈Lp(A), 1Ép< ∞, the norm off is the number kfkp= kfkLp(A)=
µZ
A|f|pdµ
¶1p . We shall see that this has the usual properties of the norm:
(1) (Nonnegativity) 0É kfkp< ∞,
(2) kfkp=0⇐⇒ f=0µ-almost everywhere, (3) (Homogeneity)ka fkp= |a|kfkp,a∈R, (4) (Triangle inequality)kf+gkpÉ kfkp+ kgkp.
The claims (1) and (3) are clear. For p=1, the claim (4) follows from the pointwise triangle inequality|f(x)+g(x)| É |f(x)| + |g(x)|. Forp>1, the claim (4) is not trivial and we shall prove it later in this section.
Let us recall how to prove (2). Recall that if a property holds except on a set of µmeasure zero, we say that it holdsµ-almost everywhere.
⇐= Assume thatf=0µ-almost everywhere inA. Then Z
A|f|pdµ= Z
A∩{|f|=0}|f|pdµ
| {z }
+ Z
A∩{|f|>0}|f|pdµ
| {z }
=0.
=0, =0,
|f| =0µ-a.e. µ(A∩{|f| >0})=0 Thuskfkp=0.
=⇒ Assume thatkfkp=0. LetAi=©
x∈A:|f(x)| Ê1iª
,i=1, 2, . . .. Then {x∈A:|f(x)| >0}=
[∞ i=1
Ai. By Chebyshev’s inequality
µ(Ai)= Z
Ai
1dµÉ Z
Ai
|i f|pdµÉip Z
A|f|pdµ
| {z }
=0
=0. (i|f| Ê1 in Ai)
Thusµ(Ai)=0 for everyi=1, 2, . . . and µ
̰
[
i=1
Ai
! É
X∞ i=1
µ(Ai)=0.
In other words,f=0µ-almost everywhere inA.
Forµ-measurable functionsfandgon aµ-measurable setA, we are interested in the casef(x)=g(x) forµ-almost everyx∈A, which means that
µ({x∈A:f(x)6=g(x)})=0.
In the case f=gµ-almost everywhere, we do not usually distinguishf fromg.
That is, we shall regard them as equal. We could be formal and consider the equivalence relation
f∼g⇐⇒f=g µ-almost everywhere inA
but this is hardly necessary. In practice, we are thinking f as the equivalence class of all functions which are equal tof µ-almost everywhere inA. ThusLp(A)
actually consists of equivalence classes rather than functions, but we shall not make the distinction. In measure and integration theory we cannot distinguishf fromg, if the functions are equalµ-almost everywhere. In fact, iff=gµ-almost everywhere inA, thenf ∈Lp(A)⇐⇒g∈Lp(A) andkf−gkp=0. In particular, this implies that kfkp= kgkp. On the other hand, ifkf−gkp=0, then f =g µ-almost everywhere inA.
Another situation that frequently arises is that the functionf is defined only almost everywhere. Then we say that f is measurable if and only if its zero extension to the whole space is measurable. Observe, that this does not affect the Lpnorm of f.
Next we show thatLp(A) is a vector space.
Lemma 1.8. (i) If f∈Lp(A), thena f ∈Lp(A),a∈R. (ii) Iff,g∈Lp(A), thenf+g∈Lp(A).
Proof. (1) Z
A|a f|pdµ= |a|p Z
A|f|pdµ< ∞.
(2) p=1 The triangle inequality|f+g| É |f| + |g|implies Z
A|f+g|dµÉ Z
A|f|dµ+
Z
A|g|dµ< ∞. 1<p< ∞ The elementary inequality
(a+b)pÉ(2 max{a,b})p=2pmax{ap,bp}
É2p(ap+bp), a,bÊ0, 0<p< ∞ (1.2) implies
Z
A|f+g|pdµÉ2p µZ
A|f|pdµ+ Z
A|g|pdµ
¶
< ∞.
ä
Remark 1.9. Note that the proof applies for 0<p< ∞. Thus Lp(A) is a vector space for 0<p< ∞. However, it will be a normed space with theLpnorm only for pÊ1 as we shall see later.
Remark 1.10. A more careful analysis gives the useful inequality
(a+b)pÉ2p−1(ap+bp), a,bÊ0, 1Ép< ∞. (1.3) Remarks 1.11:
(1) Iff:A→Cis a complex-valued function, thenf is said to beµ-measurable if and only if Ref and Imf areµ-measurable. We say that f∈L1(A) if Ref∈L1(A) and Imf∈L1(A), and we define
Z
Af dµ= Z
A
Ref dµ+i Z
A
Imf dµ,
whereiis the imaginary unit. This integral satisfies the usual linearity properties. It also satisfies the important inequality
¯
¯
¯
¯ Z
Af dµ
¯
¯
¯
¯É Z
A|f|dµ.
The definition of theLp spaces and the norm extends in a natural way to complex-valued functions. Note that the propertyka fkp= |a|kfkpfor everya∈Cand thusLpis a complex vector space.
(2) The spaceL2(A) is an inner product space with the inner product
〈f,g〉 = Z
A
f g dµ, f,g∈L2(A).
Here gis the complex conjugate which can be neglected if the functions are real-valued. This inner product induces the standardL2norm, since
kfk2= µZ
A|f|2dµ
¶12
= µZ
Af f dµ
¶12
= 〈f,f〉12.
(3) In the special case thatA=Nandµis the counting measure, theLp(N) spaces are denoted bylpand
lp= (
(xi) : X∞ i=1
|xi|p< ∞ )
, 1Ép< ∞.
Here (xi) is a sequence of real (or complex) numbers. In this case, Z
Nx dµ= X∞ i=1
x(i) for every nonnegative functionxonN. Thus
kxkp= ̰
X
i=1
|xi|p
!1p .
Note that the theory ofLpspaces applies to these sequence spaces as well.
Definition 1.12. Let 1<p< ∞. The Hölder conjugate p0 of p is the number which satisfies
1 p+ 1
p0=1.
Forp=1 we definep0= ∞and ifp= ∞, thenp0=1.
Remark 1.13. Note that
p0= p p−1, p=2=⇒p0=2, 1<p<2=⇒p0>2, 2<p< ∞ =⇒1<p0<2, p→1=⇒p0→ ∞, (p0)0=p.
Lemma 1.14 (Young’s inequality). Let 1<p< ∞. Then for everyaÊ0,bÊ0, abÉap
p +bp0 p0, with equality if and only ifap=bp0.
TH E M O R A L: Young’s inequality is a very useful tool in splitting a product to a sum. Morever, it shows where the conjugate exponentp0comes from.
Proof. The claim is obviously true, ifa=0 orb=0. Thus we may assume that a>0 andb>0. Clearly
abÉap p +bp0
p0 ⇐⇒1 p
ap bp0 + 1
p0−ab1−p0Ê0⇐⇒ 1 p
à a b
p0 p
!p
+ 1 p0− a
b
p0 p
Ê0 Lett=a/bp0/pand defineϕ: (0,∞)→R,
ϕ(t)=1 ptp+ 1
p0−t.
Then
ϕ(0)= 1 p0, lim
t→∞ϕ(t)= ∞ and ϕ0(t)=tp−1−1.
Note thatϕ0(t)=0⇐⇒t=1, from which we conclude ϕ(t)Êϕ(1)=1
p+ 1
p0−1=0 for everyt>0.
Moreover,ϕ(t)>0, ift6=1. It follows thatϕ(t)=0 if and only ifa/bp0/p=t=1. ä Remarks 1.15:
(1) Young’s inequality forp=2 follows immediately from (a−b)2Ê0⇐⇒a2−2ab+b2Ê0⇐⇒a2
2 +b2
2 ÊabÊ0.
(2) Young’s inequality can be also proved geometrically. To see this, consider the curves y=xp−1and the inversex=y1/(p−1)=yp0−1. Then
Z a
0
xp−1dx=ap p and
Z b
0
yp0−1d y=bp0 p0 .
By comparing the areas under the curves that these integrals measure, we have
abÉ Z a
0
xp−1dx+ Z b
0
yp0−1d y=ap p +bp0
p0.
Theorem 1.16 (Hölder’s inequality). Let 1<p< ∞and assume thatf∈Lp(A) andg∈Lp0(A). Thenf g∈L1(A) and
Z
A|f g|dµÉ µZ
A|f|pdµ
¶1pµZ
A|g|p0dµ
¶p01 .
Moreover, an equality occurs if and only if there exists a constant csuch that
|f(x)|p=c|g(x)|p0forµ-almost everyx∈A.
TH E M O R A L: Hölder’s inequality is very useful tool in estimating a product of functions.
Remark 1.17. Hölder’s inequality states thatkf gk1É kfkpkgkp0, 1<p< ∞. Ob- serve that for p=2 this is the Cauchy-Schwarz inequality|〈f,g〉| É kfk2kgk2. Proof. Ifkfkp=0, thenf=0µ-almost everywhere inAand thusf g=0µ-almost everywhere inA. Thus the result is clear, ifkfkp=0 orkgkp0=0. We may assume thatkfkp>0 andkgkp0>0. Let
fe= f kfkp
and ge= g kgkp0
. Then
kfekp=
°
°
°
° f kfkp
°
°
°
°p
=kfkp
kfkp =1 and kg˜kp0=1.
By Young’s inequality 1 kfkpkgkp0
Z
A|f g|dµ= Z
A|fe||eg|dµ É
Z
A
µ1
p|fe|p+ 1 p0|ge|p0
¶ dµ É1
p Z
A|fe|pdµ
| {z }
=1
+1 p0 Z
A|ge|p0dµ
| {z }
=1
=1 p+ 1
p0 =1.
An equality holds if and only if Z
A
µ1
p|fe|p+ 1
p0|eg|p0− |feeg|
¶
| {z }
Ê0
dµ=0,
which implies that
1
p|fe|p+ 1
p0|ge|p0− |fege| =0
µ-almost everywhere inA. This shows that an equality occurs in Young’s inequal- ity if and only if|fe|p= |eg|p0 µ-almost everywhere inA. In this case
|f(x)|p= kfkpp
kgkpp00
|g(x)|p0
forµ-almost everyx∈A. ä
WA R N I N G: f∈Lp(A) andg∈Lp(A) does not imply thatf g∈Lp(A).
Reason. Let
f: (0, 1)→R, f(x)= 1
px, g=f,
and assume thatµis the Lebesgue measure. Then f∈L1((0, 1)) andg∈L1((0, 1)), but
(f g)(x)=f(x)g(x)=1
x and f g∉L1((0, 1)).
■
Remarks 1.18:
(1) Forp=2 we have the Cauchy-Schwarz inequality Z
A|f g|dµÉ µZ
A|f|2dµ
¶12µZ
A|g|2dµ
¶12 .
(2) Hölder’s inequality holds for arbitrary measurable functions with the interpretation that the integrals may be infinite. (Exercise)
Lemma 1.19 (Jensen’s inequality). Let 1Ép<q< ∞and assume that 0<
µ(A)< ∞. Then
µ 1 µ(A)
Z
A|f|pdµ
¶1
p
É µ 1
µ(A) Z
A|f|qdµ
¶1
q.
TH E M O R A L: The integral average is an increasing function of the power.
Proof. By Hölder’s inequality Z
A|f|pdµÉ µZ
A|f|
pq p dµ
¶pqµZ
A
1
q q−pdµ
¶q−pq
= µZ
A|f|qdµ
¶pq
µ(A)1−pq. ä
Remark 1.20. If 1Ép<q< ∞andµ(A)< ∞, thenLq(A)⊂Lp(A).
WA R N I N G: Let 1Ép<q< ∞. In general,Lq(A)6⊂Lp(A) orLp(A)6⊂Lq(A).
Reason. Letf: (0,∞)→R,f(x)=xaand assume thatµis the Lebesgue measure.
Then
f∈L1((0, 1))⇐⇒a> −1 and f∈L1((1,∞))⇐⇒a< −1.
Assume that 1Ép<q< ∞. Choosebsuch that 1/qÉb<1/p. Then the func- tionx−bχ(0,1)(x) belongs toLp((0,∞)), but does not belong toLq((0,∞)). On the other hand, the functionx−bχ(1,∞)(x) belongs toLq((0,∞)), but does not belong to
Lp((0,∞)). ■
Examples 1.21:
(1) LetA=(0, 1),µbe the Lebesgue measure and 1Ép< ∞. Define f: (0, 1)→ R,
f(x)= 1 x1p¡
log2x¢2p .
Thenf∈Lp((0, 1)), butf∉Lq((0, 1)) for anyq>p. Thus for everypwith 1Ép< ∞, there exists a function f which belongs toLp((0, 1)), but does not belong to anyLq((0, 1)) withq>p. (Exercise)
(2) Let 1Ép<q< ∞. Assume thatAcontainsµ-measurable sets of arbitrarily small positive measure. Then there are pairwise disjointµ-measurable setsAi⊂A,i=1, 2, . . . , such thatµ(Ai)>0 andµ(Ai)→0 asi→ ∞. Let
f= X∞ i=1
aiχAi,
whereaiÊ0 withai→ ∞asi→ ∞are chosen so that X∞
i=1
aqiµ(Ai)= ∞ and X∞ i=1
apiµ(Ai)< ∞.
Thenf∈Lp(A) \Lq(A). It can be shown, thatLp(A) is not contained in Lq(A) if and only ifAcontains measurable sets of arbitrarily small positive measure. (Exercise)
(3) Let 1Ép<q< ∞. Assume thatAcontainsµ-measurable sets of arbitrarily large measure. Then there are pairwise disjointµ-measurable setsAi⊂A, i=1, 2, . . . , such thatµ(Ai)>0 andµ(Ai)→ ∞asi→ ∞. Let
f= X∞ i=1
aiχAi,
whereaiÊ0 withai→ ∞asi→ ∞are chosen so that X∞
i=1
aqiµ(Ai)< ∞ and X∞ i=1
apiµ(Ai)= ∞.
Then f ∈Lq(A) \Lp(A). It can be shown, thatLq(A) is not contained in Lp(A) if and only if A contains measurable sets of arbitrarily large measure. (Exercise)
Remark 1.22. There is a more general version of Jensen’s inequality. Assume that 0<µ(A)< ∞. Let f∈L1(A) such thata<f(x)<bfor everyx∈A. Ifϕis a convex function on (a,b), then
ϕ µ 1
µ(A) Z
A
f dµ
¶
É 1
µ(A) Z
Aϕ◦f dµ.
The casesa= −∞andb= ∞are not excluded. Observe, that in this case may happen thatϕ◦f is not integrable. We leave the proof as an exercise.
Theorem 1.23 (Minkowski’s inequality). Assume 1Ép< ∞andf,g∈Lp(A).
Thenf+g∈Lp(A) and
kf+gkpÉ kfkp+ kgkp.
Moreover, an equality occurs if and only if there exists a positive constantcsuch that f(x)=c g(x) forµ-almost everyx∈A.
TH E M O R A L: Minkowski’s inequality is the triangle inequality for theLp norm. It implies that theLpnorm, with 1Ép< ∞, is a norm in the usual sense and thatLp(A) is a normed space if the functions that coincide almost everywhere are identified.
Remark 1.24. Elementary inequalities (1.3) and (1.4) imply that kf+gkp=
µZ
A|f+g|pdµ
¶1p
É2p−1p µZ
A
(|f|p+ |g|p)dµ
¶1p
É2
p−1 p
õZ
A|f|pdµ
¶1p +
µZ
A|g|pdµ
¶1p!
=2p−1p (kfkp+ kgkp), 1Ép< ∞.
Observe that the factor 2(p−1)/pis strictly greater than one forp>1 and Minkowski’s inequality does not follow from this.
Proof. p=1 : The triangle inequality, as in the proof of Lemma1.8, shows that kf+gk1É kfk1+ kgk1.
1<p< ∞ : Ifkf+gkp=0, there is nothing to prove. Thus we may assume thatkf+gkp>0. By Hölder’s inequality, we have
Z
A|f+g|pdµÉ Z
A|f+g|p−1|f|dµ+
Z
A|f+g|p−1|g|dµ
(|f+g|p= |f+g|p−1|f+g| É |f+g|p−1(|f| + |g|)) É
µZ
A|f+g|(p−1)p0dµ
¶p01 µZ
A|f|pdµ
¶1p
+ µZ
A|f+g|(p−1)p0dµ
¶1
p0µZ
A|g|pdµ
¶1p . Since (p−1)p0=pand 0< kf+gkp< ∞, we have
µZ
A|f+g|pdµ
¶1−p−1p
É µZ
A|f|pdµ
¶1p +
µZ
A|g|pdµ
¶1p .
It remains to consider when the equality can occur. This happens if there is an equality in the pointwise inequality
|f(x)+g(x)|p= |f(x)+g(x)|p−1|f(x)+g(x)| É |f(x)+g(x)|p−1(|f(x)| + |g(x)|) for µ-almost every x∈A as well as an equality in the application of Hölder’s inequality. An equality occurs in Hölder’s inequality if
c1|f(x)|p= |f(x)+g(x)|p=c2|g(x)|p
forµ-almost everyx∈A. Equality occurs in in the pointwise inequality above if f(x) andg(x) have the same sign. This completes the proof. ä Remark 1.25. It is possible to prove Minkowski’s inequality directly by Young’s inequality instead of Hölder’s inequality (exercise).
Note that the normed spaceLp(A), 1Ép< ∞, is a metric space with the metric d(f,g)= kf−gkp.
1.3 L p spaces for 0 < p < 1
It is sometimes useful to considerLpspaces for 0<p< ∞. Observe that Definition 1.1makes sense also when 0<p<1 and the space is a vector space by the same argument as in the proof of Lemma1.8. However,kfkpis not a norm for 0<p<1.
Reason. Let f,g:R→R, f =χ[0,1
2) and g=χ[1
2,1]. Then f+g=χ[0,1] so that kf+gkp=1. On the other hand,
kfkp=2−1p and kgkp=2−1p. Thus
kfkp+ kgkp=2·2−1p=21−1p<1,
when 0<p<1. This shows thatkfkp+ kgkp< kf+gkp. ■ Thus the triangle inequality does not hold true when 0<p<1, but we have the following result.
Lemma 1.26. Iff,g∈Lp(A) and 0<p<1, thenf+g∈Lp(A) and kf+gkppÉ kfkpp+ kgkpp.
Proof. The elementary inequality
(a+b)pÉap+bp, a,bÊ0, 0<p<1, (1.4) implies
kf+gkpp= Z
A|f+g|pdµÉ Z
A|f|pdµ+
Z
A|g|pdµ= kfkpp+ kgkpp.
ä However,Lp(A) is a metric space with the metric
d(f,g)= kf−gkpp= Z
A|f−g|pdµ
This metric is not induced by a norm, sincekfkppdoes not satisfy the homogeneity required by the norm. On the other hand,kfkpsatisfies the homogeneity, but not satisfy the triangle inequality.
Remarks 1.27:
(1) By (1.2), we have kf+gkpɡ
kfkpp+ kgkpp
¢1p
É21p(kfkp+ kgkp), 0<p<1.
Thus a quasi triangle inequality holds with a multiplicative constant.
(2) Iff,g∈Lp(A), fÊ0,gÊ0, then
kf+gkpÊ kfkp+ kgkp, 0<p<1.
This is the triangle inequality in the wrong direction (exercise).
Remark 1.28. It is possible to define theLpspaces also whenp<0. Aµ-measurable function is inLp(A) for p<0, if
0<
Z
A|f|pdµ< ∞.
If f ∈Lp(A) for p<0, then f 6=0 µ-almost everywhere and|f| < ∞µ-almost everywhere inA. However, this is not a vector space.
1.4 Completeness of L p
Next we prove a famous theorem, which is not only important in the theory ofLp spaces, but has a historical interest as well. The result was found independently by F. Riesz and E. Fisher in 1907, primarily in connection with the Fourier series which culminates in showing completeness ofL2.
Recall that a sequence (fi) of functions fi∈Lp(A), i=1, 2, . . ., converges in Lp(A) to a function f∈Lp(A), if for everyε>0 there existsiεsuch that
kfi−fkp<ε when iÊiε. Equivalently,
ilim→∞kfi−fkp=0.
A sequence (fi) is a Cauchy sequence inLp(A), if for everyε>0 there existsiε such that
kfi−fjkp<ε when i,jÊiε. WA R N I N G: This is not the same condition as
kfi+1−fikp<ε when iÊiε.
Indeed, the Cauchy sequence condition implies this, but the converse is not true (exercise).
CL A I M: If fi→f inLp(A) asi→ ∞, then (fi) is a Cauchy sequence inLp(A).
Reason. Letε>0. Since fi→f in Lp(A) as i→ ∞, there exists iε such that kfi−fkp<ε2 wheniÊiε. Minkowski’s inequality implies
kfi−fjkpÉ kfi−fkp+ kf−fjkp<ε 2+ε
2=ε
wheni,jÊiε. ■
Theorem 1.29 (Riesz-Fischer). For every Cauchy sequence (fi) inLp(A), 1É p< ∞, there existsf∈Lp(A) such thatfi→f inLp(A) asi→ ∞.
TH E M O R A L: Lp(A), 1Ép< ∞, is a Banach space with the normk · kp. In particular,L2(A) is a Hilbert space.
Proof. Assume that (fi) is a Cauchy sequence inLp(A). We construct a subse- quence as follows. Choosei1such that
kfi−fjkp<1
2 when i,jÊi1.
We continue recursively. Suppose thati1,i2, . . . ,ikhave been chosen such that kfi−fjkp< 1
2k when i,jÊik. Then chooseik+1>iksuch that
kfi−fjkp< 1
2k+1 when i,jÊik+1. For the subsequence (fik), we have
kfik−fik+1kp< 1
2k, k=1, 2, . . . . Let
gl=
l
X
k=1
|fik+1−fik| and g= X∞ k=1
|fik+1−fik|. Then
lim
l→∞gl=lim
l→∞
l
X
k=1
|fik+1−fik| = X∞ k=1
|fik+1−fik| =g
and as a limit ofµ-measurable functions gis aµ-measurable function. Fatou’s lemma and Minkowski’s inequality imply
µZ
A
gpdµ
¶1p
= µZ
A
lim inf
l→∞ glpdµ
¶1p
Élim inf
l→∞
µZ
A
glpdµ
¶1p
Élim inf
l→∞
l
X
k=1
µZ
A|fik+1−fik|pdµ
¶1p
É X∞ k=1
1 2k=1.
Thus g∈Lp(A) and consequently g(x)< ∞forµ-almost every x∈A. It follows that the telescoping series
fi1(x)+ X∞ k=1
(fik+1(x)−fik(x))
converges absolutely forµ-almost everyx∈A. Denote the sum of the series by f(x) for thosex∈Aat which it converges and setf(x)=0 in the remaining set of measure zero. Then
f(x)=fi1(x)+ X∞ k=1
(fik+1(x)−fik(x))
=lim
l→∞
à fi1(x)+
l−1
X
k=1
(fik+1(x)−fik(x))
!
=lim
l→∞fil(x)= lim
k→∞fik(x)
for µ-almost every x∈A. Thus there is a subsequence (fik) which converges µ-almost everywhere inA. Next we show that the original sequence converges to
f inLp(A).
CL A I M: fi→f inLp(A) asi→ ∞.
Reason. Letε>0 and let (fik) be a subsequence which converges to f µ-almost everywhere inA. Since (fi) is a Cauchy sequence inLp(A), there existsiεsuch thatkfik−fikp<εwhen i,ikÊiε. For a fixed iÊiε, we have fik−fi→f−fi µ-almost everywhere inAasik→ ∞. By Fatou’s lemma
µZ
A|f−fi|pdµ
¶1p
= µZ
A
lim inf
k→∞ |fik−fi|pdµ
¶1p
Élim inf
k→∞
µZ
A|fik−fi|pdµ
¶1p
Éε. ■
This shows that f−fi∈Lp(A) and thus f =(f−fi)+fi∈Lp(A). Moreover, for everyε>0 there existsiεsuch thatkfi−fkpÉεwheniÊiε. This completes the
proof. ä
WA R N I N G: In general, if a sequence has a converging subsequence, the original sequence need not converge. In the proof above, we used the fact that we have a Cauchy sequence.
We shall often use a part of the proof of the Riesz-Fisher theorem, which we now state.
Corollary 1.30. Iffi→f inLp(A), then there exist a subsequence (fik) such that lim
k→∞fik(x)=f(x) µ-almost every x∈A.
Proof. The proof of the Riesz-Fischer theorem gives a subsequence (fik) and a functiong∈Lp(A) such that
lim
k→∞fik(x)=g(x) µ-almost every x∈A
and fik→g inLp(A). On the other hand, fi→f inLp(A), which implies that fik→f inLp(A). By the uniqueness of the limit, we conclude that f=gµ-almost
everywhere inA. ä
Let us compare the various modes of convergence of a sequence (fi) of functions inLp.
Remarks 1.31:
(1) Iffi→f inLp(A) asi→ ∞, then lim
i→∞kfikp= kfkp. Reason. By Minkowski’s inequality
kfikp= kfi−f+fkpÉ kfi−fkp+ kfkp,
which implieskfikp− kfkpÉ kfi−fkp. By switching the roles off and fi, we havekfkp− kfikpÉ kfi−fkp. Thus
¯
¯kfikp− kfkp
¯
¯É kfi−fkp→0, from which it follows that
lim
i→∞kfikp= kfkp.
■
(2) Iffi→f inLp(A) asi→ ∞, thenfi→f in measure.
Reason. By Chebyshev’s inequality µ({x∈A:|fi(x)−f(x)| Êε})É 1
εp Z
A|fi−f|pdµ
= 1
εpkfi−fkpp i→∞
−−−→0. ■
(3) Iffi→f inLp(A) asi→ ∞, then there exist a subsequence (fik) such that lim
k→∞fik(x)=f(x) µ-almost every x∈A.
Reason. The convergence in measure implies the existence of an almost ev- erywhere converging subsequence. This gives another proof of the previous
corollary. ■
(4) In the casep=1,fi→f inL1(A) asi→ ∞implies not only that
ilim→∞
Z
A|fi|dµ= Z
A|f|dµ but also that
lim
i→∞
Z
A
fidµ= Z
A
f dµ. Reason.
¯
¯
¯
¯ Z
A
(fi−f)dµ
¯
¯
¯
¯É Z
A|fi−f|dµ= kfi−fk1 i→∞
−−−→0.
■
The following example shows that imply pointwise convergence almost eery- where does not implyLpconvergence pointwise convergence andLpconvergence does not imply pointwise convergence almost eerywhere.
Example 1.32. In the following examples we assume thatµis the Lebesgue mea- sure.
(1) fi→f in Lp as i→ ∞does not imply fi→f almost everywhere. Let fi:R→R, fi(x)=χ[i−1,i)(x), i=1, 2, . . ., and f =0. Assume thatµis the Lebesgue measure. Thenfi(x)→0 for everyx∈R. However,kfikp=1 for everyi=1, 2, . . . andkfkp=0. Thus the sequence (fi) does not converge to
f inLp(R), 1Ép< ∞.
(2) fi→f almost everywhere as i→ ∞does not imply fi→f in Lp. Let fi:R→R,
fi(x)=i2χ¡0,1
i
¢(x), i=1, 2, . . . .
Then Z
R|fi(x)|pdx=i2p Z
Rχp
(0,1i)(x)dx=i2p1i=i2p−1< ∞. Thusfi∈Lp(R), 1Ép< ∞,fi(x)→0 for everyx∈R, but
kfikp=i2−1pÊi−−−→ ∞i→∞ . Thus (fi) does not converge inLp(R).
(3) fi→f inLpasi→ ∞does not implyfi→f almost everywhere. Consider the sliding sequence of functionsfi:R→R,
f2k+j(x)=kχh j 2k,j+1
2k
i(x), k=0, 1, 2, . . . , j=0, 1, 2, . . . , 2k−1.
Then
kf2k+jkp=k2−kp−−−−→k→∞ 0,
which implies that fi→0 inLp(R), 1Ép< ∞, as i→ ∞. However, the sequence (fi(x)) fails to converge for everyx∈[0, 1], since
lim sup
i→∞
fi(x)= ∞ and lim inf
i→∞ fi(x)=0
for everyx∈[0, 1]. Note that there are many converging subsequences. For example, f2k+1(x)→0 for everyx∈[0, 1] ask→ ∞.
(4) A sequence can converge inLpwithout converging inLq. Considerfi:R→ R,
fi(x)=1iχ(i,2i)(x), i=1, 2, . . . .
Thenkfikp=i−1+1p,i=1, 2 . . . . Thus f→0 inLp(R), 1<p< ∞, asi→ ∞, butkfik1=1 for everyi=1, 2 . . . , so that the sequence (fi) does not converge inL1(R).
The following discussion clarifies the difference between the pointwise conven- gence andLpconvergence.
Theorem 1.33. Let 1Ép< ∞. Assume that fi∈Lp(A), i=1, 2, . . . , fi→f µ- almost everywhere in Aas i→ ∞. If there exists g∈Lp(A), gÊ0, such that
|fi| Égµ-almost everywhere in Afor everyi=1, 2, . . . , thenf∈Lp(A) and fi→f inLp(A) asi→ ∞.
TH E M O R A L: Pointwise convergence implies the norm convergence inLp if the sequence is uniformly bounded by a function inLp. This is a dominated convergence theorem inLp.
Proof. Since|fi−f| É |fi|+|f| Ég+|f|µ-almost everywhere inAandg+|f| ∈Lp(A), the dominated convergence theorem implies
ilim→∞
Z
A|fi−f|dµ= Z
A
ilim→∞|fi−f|dµ=0.
ä Theorem 1.34. Assume that fi∈Lp(A),i=1, 2, . . . and f∈Lp(A), 1Ép< ∞. If
fi→f µ-almost everywhere in Aand lim
i→∞kfikp= kfkp, thenfi→f inLp(A) as i→ ∞.
Proof. Since|fi| < ∞and|f| < ∞µ-almost everywhere inA, by (1.2), we have 2p(|fi|p+ |f|p)− |fi−f|pÊ0
µ-almost everywhere in A. The assumption fi→f µ-almost everywhere inA implies
ilim→∞(2p(|fi|p+ |f|p)− |fi−f|p)=2p+1|f|p µ-almost everywhere inA. Applying Fatou’s lemma, we obtain
Z
A
2p+1|f|pdµÉlim inf
i→∞
Z
A
¡2p(|fi|p+ |f|p)− |fi−f|p¢ dµ Élim inf
i→∞
µZ
A
2p|fi|pdµ+ Z
A
2p|f|pdµ− Z
A|fi−f|pdµ
¶
=lim
i→∞
Z
A
2p|fi|pdµ+
Z
A
2p|f|pdµ−lim sup
i→∞
Z
A|fi−f|pdµ
= Z
A
2p|f|pdµ+ Z
A
2p|f|pdµ−lim sup
i→∞
Z
A|fi−f|pdµ.
Here we used the facts that if (ai) is a converging sequence of real numbers and (bi) is an arbitrary sequence of real numbers, then
lim inf
i→∞ (ai+bi)=lim
i→∞ai+lim inf
i→∞ bi and lim inf
i→∞ (−bi)= −lim sup
i→∞
bi. SubtractingR
A2p+1|f|pdµfrom both sides, we have lim sup
i→∞
Z
A|fi−f|pdµÉ0.
On the other hand, since the integrands are nonnegative lim inf
i→∞
Z
A|fi−f|pdµÊ0.
Thus
ilim→∞
Z
A|fi−f|pdµ=0.
ä Remark 1.35. Let 1Ép< ∞. Assume that fi∈Lp(A), 0ÉfiÉ fi+1 µ-almost everywhere in A, i=1, 2, . . . . Then the pointwise limit f =limi→∞fi existsµ- almost everywhere inA. The monotone convergence theorem implies
lim
i→∞
Z
A
fipdµ= Z
A
lim
i→∞fipdµ= Z
A
fpdµ. Theorem1.34(or Theorem1.33) implies fi→f inLp(A) asi→ ∞.
TH E M O R A L: An increasing sequence of nonnegative functions inLpcon- verges inLp, if the limit function belongs toLp. This is a monotone convergence theorem inLp.
1.5 L ∞ space
The definition of the L∞ space differs substantially from the definition of the Lp space for 1Ép< ∞. The main difference is that instead of the integration the definition is based on the almost everywhere concept. The classL∞consists of bounded measurable functions with the interpretation that we neglect the behaviour of functions on a set of measure zero.
Definition 1.36. Let A⊂Rn be a µ-measurable set and f :A→[−∞,∞] a µ- measurable function. Thenf∈L∞(A), if there existsM, 0ÉM< ∞, such that
|f(x)| ÉM forµ-almost everyx∈A.
Functions inL∞are sometimes called essentially bounded functions. The essential supremum off is
ess sup
x∈A
f(x)=inf{M:f(x)ÉMforµ-almost everyx∈A}
=inf©
M:µ({x∈A:f(x)>M})=0ª and the essential infimum off is
ess inf
x∈A f(x)=sup{m:f(x)Êmforµ-almost everyx∈A}
=sup©
m:µ({x∈A:f(x)<m})=0ª . TheL∞norm of f is
kfk∞=ess sup
x∈A |f(x)|.