1.4 Lorentz Spaces
1.4.4 The Off-Diagonal Marcinkiewicz Interpolation Theorem
(1.4.16) lies in the spaceLq(0,∞)with respect to the measuretq/p−1dt. This fol- lows from the inequality
f
Lp,q = Z ∞
0
t
q p
Z ∞ t/2
s
q0 p0−1
g∗(s)q0−1ds s
q
dt t
1q
≤C1(p,q) Z ∞
0
(t
1 p0
g∗(t))q0dt t
1q
=C1(p,q) g
q0/q Lp0,q0 <∞,
which is a consequence of Hardy’s second inequality in Exercise 1.2.8 withb=q/p.
Using (1.4.15), we conclude that Z ∞
0
f∗(t)g∗(t)dt≤ T
f
Lp,q ≤C1(p,q) T
g
q0/q
Lp0,q0. (1.4.17) On the other hand, we have
Z ∞
0
f∗(t)g∗(t)dt ≥ Z ∞
0 Z t
t/2
s
q0 p0−1
g∗(s)q0−1ds s g∗(t)dt
≥ Z ∞
0
g∗(t)q0 Z t
t/2
s
q0 p0−1ds
s dt
=C2(p,q) g
q0 Lp0,q0.
(1.4.18)
Combining (1.4.17) and (1.4.18), we obtain g
Lp0,q0≤C(p,q) T
. This estimate is valid only when we have a priori knowledge that
g
Lp0,q0 <∞. Suitably modifying the preceding proof and using that
g Lp0,q0
(KN)<∞, we obtain that g
Lp0,q0
(KN)≤ C(p,q)
T
for allN=1,2, . . .. LettingN→∞, we obtain the required conclusion.
(vii) For a complete characterization of this space, we refer to the article of Cwikel [62].
(viii) The dual ofL∞can be identified with the set of all bounded finitely additive
set functions. See Dunford and Schwartz [77].
Remark 1.4.18.Some parts of Theorem 1.4.17 are false ifXis atomic. For instance, the dual of`p(Z)containsl∞when 0<p<1 and thus it is not{0}.
for someK>0,λ∈C, and all functions f,gin the domain ofT. To avoid triviali- ties, we assume thatK≥1.
Theorem 1.4.19.Let0<r≤∞,0<p06=p1≤∞, and0<q06=q1≤∞and let (X,µ)and (Y,ν)be two measure spaces. Let T be either a quasilinear operator defined on Lp0(X) +Lp1(X)and taking values in the set of measurable functions on Y or a linear operator defined on the set of simple functions on X and taking values as before. Assume that for some M0,M1<∞the following (restricted) weak type estimates hold:
T(χA)
Lq0,∞ ≤M0µ(A)1/p0, (1.4.19) T(χA)
Lq1,∞ ≤M1µ(A)1/p1, (1.4.20) for all measurable subsets A of X withµ(A)<∞. Fix0<θ<1and let
1
p=1−θ p0
+ θ p1
and 1
q=1−θ q0
+θ q1
. (1.4.21)
Then there exists a constant M, which depends on K, p0, p1, q0, q1, M0, M1, r, and θ, such that for all functions f in the domain of T and in Lp,r(X)we have
T(f)
Lq,r ≤M f
Lp,r. (1.4.22)
We note thatLp,∞⊆Lp0+Lp1 (Exercise 1.1.10(c)), and thusT is well defined on Lp,rfor allr≤∞. Ifr<∞andTis linear and defined on the set of simple functions onX, thenT has a unique extension that satisfies (1.4.22) for all f inLp,r(X), since simple functions are dense in this space.
Before we give the proof of Theorem 1.4.19, we state and prove a lemma that is interesting on its own.
Lemma 1.4.20.Let0<p<∞and0<q≤∞. Let T be either a quasilinear oper- ator defined on Lp(X,µ)and taking values in the set of measurable functions of a measure space(Y,ν), or a linear operator initially defined on the space of simple functions on X and taking values as before. Suppose that there exists a constant L>0such that for all A⊆X of finite measure we have
T(χA)
Lq,∞≤Lµ(A)1/p. (1.4.23)
Fixα0<q with0<α0≤ log 2Klog 2. Then for all0<α≤α0 there exists a constant C(p,q,K,α)>0 (depending only on the parameters indicated) such that for all functions f in Lp,α(X)that lie in the domain of T , we have the estimate
T(f)
Lq,∞≤C(p,q,K,α)L f
Lp,α. (1.4.24)
Lemma 1.4.20 is saying that if a quasilinear operator satisfies aLp,1→Lq,∞es- timate uniformly on all characteristic functions, then it must map a Lorentz space Lp,αtoLq,∞for someα<1.
Proof. It suffices to prove Lemma 1.4.20 for f ≥0, since we can express a general function f as
f = (f1−f2) +i(f3−f4), wherefj≥0, and use quasilinearity.
It follows from the Aoki–Rolewicz theorem (Exercise 1.4.6) that for allf1, . . . ,fm we have the pointwise inequality
|T(f1+· · ·+fm)| ≤4 m
∑
j=1|T(fj)|α1 α1
1
≤4 m
∑
j=1|T(fj)|α α1
,
(1.4.25)
where 0<α≤α1andα1satisfies the equation (2K)α1 =2.
The second inequality in (1.4.25) is a simple consequence of the fact thatα ≤α1. Fixα0>0 with
α0≤α1= log 2
log 2K and α0<q.
This ensures that the quasinormed spaceLq/α,∞is normable whenα≤α0. In fact, Exercise 1.1.12 gives that the spaceLs,∞is normable as long ass>1 and for some equivalent norm
f
Ls,∞ we have
f Ls,∞ ≤
f
Ls,∞ ≤ s s−1
f
Ls,∞. Next we claim that for any f ≥0 we have
T(fχA)
Lq,∞≤C(q,α)Lµ(A)1/p f
L∞. (1.4.26)
To prove (1.4.26) first observe that multiplying by a suitable constant, we may as- sume that f≤1. Write
f(x) =
∞
∑
j=1dj(x)2−j in binary expansion, wheredj(x) =0 or 1. Let
Bj={x∈A: dj(x) =1}. ThenBj⊆Aand the function fχAcan be written as the sum
∞
∑
j=12−jχBj.
We use (1.4.25) once and (1.4.3) twice in the following argument. We have T(fχA)
Lq,∞ ≤4
∞
∑
j=1
2−jα|T(χBj)|αα1 Lq,∞
=4
∞
∑
j=12−jα|T(χBj)|α
1 α
Lq/α,∞
≤4
∞
∑
j=1
2−jα|T(χBj)|α
1 α
Lq/α,∞
≤4 ∞
∑
j=1
2−jα
|T(χBj)|α
Lq/α,∞
α1
≤4 q q−α
α1 ∞
∑
j=1
2−jα
|T(χBj)|α Lq/α,∞
α1
=4 q q−α
α1 ∞
∑
j=1
2−jα
T(χBj)
α Lq,∞
α1
≤4 q q−α
α1 L
∞ j=1
∑
2−jαµ(Bj)α/p α1
≤2 q q−α
α1
(1−2−α)−α1Lµ(A)1/p, sinceBj⊆A. This establishes (1.4.26) with
C(q,α) =2 q q−α
α1
(1−2−α)−α1. Now write the function f as
f=
∞ n=−∞
∑
fχAn, whereAnare measurable sets defined by
An={x∈X: f∗(2n+1)<|f(x)| ≤ f∗(2n)}. (1.4.27) Observe that
µ(An) =
{t∈R: f∗(2n+1)<f∗(t)≤f∗(2n)}
=
[2n,2n+1]
=2n,
since f andf∗are equidistributed. Next we have
T(f)
Lq,∞ ≤4
∞
n=−∞
∑
|T(fχAn)|αα1 Lq,∞
=4
∞ n=−∞
∑
|T(fχAn)|α
1 α
Lq/α,∞
≤4
∞ n=−∞
∑
|T(fχAn)|α
1 α
Lq/α,∞
≤4 ∞
n=−∞
∑
|T(fχAn)|α
Lq/α,∞
α1
≤4 q q−α
α1 ∞ n=−∞
∑
|T(fχAn)|α Lq/α,∞
α1
≤4 q q−α
α1 ∞ n=−∞
∑
T(fχAn)
α Lq,∞
α1
≤8 q q−α
α2
(1−2−α)−α1L ∞
n=−∞
∑
f∗(2n)α2nα/p α1
≤8 q q−α
α2
(1−2−α)−α1(log 2)α1L f
Lp,α.
Taking into account the splittingf =f1−f2+i f3−i f4, wherefj≥0, we conclude the proof of the lemma with constant
C(p,q,K,α) =CpK2 q q−α
2/α
(log 2)α1(1−2−α)−α1. (1.4.28) Recall that we have been assuming thatα<min log 2Klog 2 ,q
throughout.
We now continue with the proof of Theorem 1.4.19.
Proof. We assume that p0<p1, since if p0>p1we may simply reverse the roles ofp0andp1. We first consider the casep1<∞. Lemma 1.4.20 implies that
T(f)
Lq0,∞ ≤A0
f
Lp0,m, T(f)
Lq1,∞ ≤A1 f
Lp1,m, (1.4.29)
wherem=1
2min q0,q1, log 2
log 2K
,A0=C(p0,q0,K,m)M0,A1=C(p1,q1,K,m)M1, andC(p,q,K,α)is as in (1.4.28).
Fix a function f inLp,r. Split f =ft+ftas follows:
ft(x) =
(f(x) if|f(x)|> f∗(tγ), 0 if|f(x)| ≤ f∗(tγ),
ft(x) =
(0 if|f(x)|>f∗(tγ), f(x) if|f(x)| ≤ f∗(tγ), whereγis the following nonzero real number:
γ=
1 q0−1q
1 p0−1p =
1 q−q1
1
1 p−p1
1
.
Next, observe that the following inequalities are valid:
(ft)∗(s) ≤
(f∗(s) if 0<s<tγ, 0 ifs≥tγ, (ft)∗(s) ≤
(f∗(tγ) if 0<s<tγ, f∗(s) ifs≥tγ.
It follows from these inequalities that ftlies inLp0,mandftlies inLp1,mfor allt>0.
The sublinearity of the operatorT and (1.4.9) imply T(f)
Lq,r =
t1qT(f)∗(t) Lr(dtt)
≤K
t1q(|T(ft)|+|T(ft)|)∗(t) Lr(dtt)
≤K
t1qT(ft)∗(2t) +t1qT(ft)∗(2t) Lr(dtt )
≤Kar
t1qT(ft)∗(t2)
Lr(dtt)+
t1qT(ft)∗(t2) Lr(dtt)
≤Kar t
1 q−q1
0t
1
q0T(ft)∗(2t) Lr(dtt)
+ t
1 q−1
q1t
1
q1T(ft)∗(t2) Lr(dt
t)
,
(1.4.30)
where
ar=
(1 whenr≥1, 2(1−r)/r whenr≤1.
It follows from (1.4.29) that t
1
q0T(ft)∗(t2) ≤2q10sup
s>0
s
1
q0T(ft)∗(s)≤2q10A0
ft
Lp0,m, (1.4.31) tq11T(ft)∗(t2) ≤2q11sup
s>0
sq11T(ft)∗(s)≤2q11A1 ft
Lp1,m, (1.4.32) for allt>0. Now use (1.4.31) and (1.4.32) to estimate (1.4.30) by
Kar2
1 q0A0
t
1 q−q1
0
ft Lp0,m
Lr(dtt)
+Kar2q11A1
t1q−q11 ft
Lp1,m
Lr(dtt)
,
which is the same as Kar2q10A0
t−γ(p10−1p) ft
Lp0,m
Lr(dtt )
+Kar2
1 q1A1
tγ(
1 p−p1
1) ft
Lp1,m
Lr(dtt )
.
(1.4.33)
Next, we change variablesu=tγin the first term of (1.4.33) to obtain Kar2
1 q0A0
t−γ(
1 p0−1p)
ft Lp0,m
Lr(dt
t)
≤Kar2q10A0
|γ|1/r
u−(
1
p0−1p)Z u 0
f∗(s)ms
m p0 ds
s 1m
Lr(duu)
≤Kar 2
1 q0A0
m|γ|1/r r r(p1
0−1p) Z ∞
0
(sp10 f∗(s))rs−r(
1 p0−1p)ds
s 1r
=Kar 2
1 q0A0
m|γ|1/r 1
1 p0−1p
f
Lp,r,
where the last inequality is a consequence of Hardy’s first inequality in Exercise 1.2.8 withp=r/m≥1 andb= (1/p0−1/p)r.
Similarly, change variablesu=tγin the second term of (1.4.33) to obtain Kar2
1 q1A1
tγ(
1 p−p1
1) ft
Lp1,m
Lr(dtt)
≤Kar2
1 q1A1
|γ|1/r u
1 p−p1
1
Z u 0
f∗(u)mspm1ds s +
Z ∞
u
f∗(s)mspm1 ds s
m1 Lr(duu)
≤Ka2r21−mm 2
1 q1A1
|γ|1/r
(p1 m u
1 p−p1
1 f∗(u)u
1 p1
Lr(duu)
+ u
1 p−p1
1
Z ∞ u
f∗(s)ms
m p1 ds
s m1
Lr(du
u)
)
≤Ka2r21−mm 2
1 q1A1
|γ|1/r
(p1 m f
Lp,r+ r mr(1p−p1
1)
ur(
1 p−1
p1)
f∗(u)ru
r p1
Lr(du
u)
)
=Ka2r21−mm 2
1 q1A1
|γ|1/r
(p1
m + 1
m(1p−p1
1) )
f
Lp,r,
where the last inequality above is Hardy’s second inequality in Exercise 1.2.8 with p=r/m≥1 andb= (1/p−1/p1)r.
We have now shown that
T(f)
Lq,r ≤M f
Lp,r
with constant M=Kar
2
1 q0 +2q11 m|γ|1/r
A0
1
p0−1p +ar21−mm A1
p1+ 1
1 p−p1
1
!
. (1.4.34)
We have been tacitly assuming thatr<∞. The remaining case is a simple con- sequence of the result just proved by letting r→∞, in which casear →1 and
|γ|1/r→1.
We now turn to the casep1=∞. Hypotheses (1.4.19) and (1.4.20) together with Exercise 1.1.16 imply that
T(χA)
Lq,∞≤M01−θM1θµ(A)1/p
for all 0≤θ ≤1. We selectλ ∈(0,1)such that the indices p=pλ andq=qλ defined by (1.4.21) whenθ=λ satisfyp0<p<pλ <∞andqλ is strictly between q0andq1. Then apply the casep1<∞just proved withp0,q0as before andp1=pλ andq1=qλ. The result follows withMas in (1.4.34) except thatp1is replaced by
pλandq1byqλ.
Corollary 1.4.21.Let T be as in the statement of Theorem 1.4.19 and let0<p06=
p1≤∞and0<q06=q1≤∞. If T maps Lp0to Lq0,∞and Lp1to Lq1,∞, and for some 0<θ<1we have
1
p=1−θ p0 + θ
p1, 1
q=1−θ q0 +θ
q1, and p≤q, then T satisfies the strong type estimate
T(f) Lq ≤C
f
Lp for all functions f in the domain of T . Moreover, if T is linear, then it has a bounded extension from Lp(X,µ)to Lq(Y,ν).
Proof. Taker=qin the previous theorem.
Definition 1.4.22.Let 0<p,q≤∞. We call an operatorT ofrestricted weak type (p,q)if it satisfies
T(χA)
Lq,∞≤Cµ(A)1/p
for all subsetsAof a measure space(X,µ)with finite measure. Using this terminol- ogy, Corollary 1.4.21 says that if a quasilinear operatorT is of restricted weak types (p0,q0)and(p1,q1)for some p06=p1andq06=q1, then it is bounded fromLpto Lqwhen p≤q.
We now give examples to indicate why the assumptions p06=p1andq06=q1 cannot be dropped in Theorem 1.4.19.
Example 1.4.23.LetX=Y =Rand T(f)(x) =|x|−1/2
Z 1 0
f(t)dt.
Thenα|{x: |T(χA)(x)|>α}|1/2=21/2|A∩[0,1]|and thusT is of restricted weak types(1,2)and(3,2). But observe that T does not mapL2=L2,2 toLq,2. Thus Theorem 1.4.19 fails if the assumptionq06=q1is dropped. The dual operator
S(f)(x) =χ[0,1](x) Z +∞
−∞ f(t)|t|−1/2dt
satisfiesα|{x: |S(χA)(x)|>α}|1/q≤c|A|1/2whenq=1 or 3, and thus it furnishes an example of an operator of restricted weak types(2,1)and(2,3)that is notL2 bounded. Thus Theorem 1.4.19 fails if the assumptionp06=p1is dropped.
As an application of Theorem 1.4.19, we give the following strengthening of Theorem 1.2.13.
Theorem 1.4.24.(Young’s inequality for weak type spaces) Let G be a locally com- pact group with left Haar measureλ that satisfies (1.2.12) for all measurable sub- sets A of G. Let1<p,q,r<∞satisfy
1
q+1=1 p+1
r. (1.4.35)
Then there exists a constant Bpqr>0such that for all f in Lp(G)and g in Lr,∞(G) we have
f∗g
Lq(G)≤Bpqr g
Lr,∞(G)
f
Lp(G). (1.4.36) Proof. We fix 1<p,q<∞. Sincepandqrange in an open interval, we can find p0<p<p1,q0<q<q1, and 0<θ<1 such that (1.4.21) and (1.4.35) hold. Let T(f) = f∗g, defined for all functions f onG. By Theorem 1.2.13,T extends to a bounded operator fromLp0 toLp1,∞ and fromLq0 toLq1,∞. It follows from the Marcinkiewicz interpolation theorem thatT extends to a bounded operator from
Lp(G)toLq(G).
Exercises
1.4.1.(a) Letgbe a nonnegative simple function on(X,µ)and letAbe a measur- able subset ofX. Prove that
Z
A
g dµ≤ Z µ(A)
0
g∗(t)dt.
(b) (G. H. Hardy and J. E. Littlewood) Forf andgmeasurable on(X,µ), prove that
Z
X
|f(x)g(x)|dµ(x)≤ Z ∞
0
f∗(t)g∗(t)dt.
Compare this result to the classical Hardy–Littlewood result asserting that ifaj,bj>
0, the sum∑jajbjis greatest when bothajandbjare rearranged in decreasing order (for this see Hardy, Littlewood, and P´olya [122, p. 261]).
1.4.2.Prove that if f ∈Lq0,∞∩Lq1,∞for some 0<q0<q1≤∞, then f ∈Lq,s for all 0<s≤∞andq0<q<q1.
1.4.3.(Hunt [134]) Given 0<p,q<∞, fix anr=r(p,q)>0 such thatr≤1,r≤q andr<p. Fort≤µ(X)define
f∗∗(t) = sup
µ(E)≥t
1 µ(E)
Z
E
|f|rdµ 1/r
,
while fort>µ(X)(ifµ(X)<∞) let f∗∗(t) =
1 t Z
X
|f|rdµ 1/r
.
Also define
f
Lp,q= Z ∞
0
t1p f∗∗(t)qdt t
1q . (The function f∗∗and the functionalf →
f
Lp,q depend onr.) (a) Prove that the inequality
(((f+g)∗∗)(t))r≤(f∗∗(t))r+ (g∗∗(t))r is valid for allt≥0. Sincer≤q, conclude that the functional
f →
f
r Lp,q
is subadditive and hence it is a norm whenr=1 (this is possible only ifp>1).
(b) Show that for all f we have
f Lp,q≤
f
Lp,q ≤ p
p−r 1/r
f
Lp,q.
(c) In conjunction with Exercise 1.1.12, conclude thatLp,qis metrizable whenever 0<p,q≤∞and also normable when 1<p<∞and 1≤q≤∞.
1.4.4.(a) Show that on aσ-finite measure space(X,µ)the set of countable linear combinations of simple functions is dense inLp,∞(X).
(b) Prove that simple functions are not dense inLp,∞(R)for any 0<p≤∞.
Hint: Part (b): Show that the function h(x) =x−1/pχx>0 cannot be approxi- mated by a sequence of simple functions Lp,∞. To see this, partition the inter-
val(0,∞)into small subintervals of lengthε>0 and let fε be the step function
∑[1/ε]−[1/ε]f(kε)χ[kε,(k+1)ε](x). Show that for somec>0 we havekfε−fkLp,∞≥c.
1.4.5.Let(X,µ)be a nonatomic measure space. Prove the following facts:
(a) IfA0⊆A1⊆X, 0<µ(A1)<∞, andµ(A0)≤t≤µ(A1), then there exists an Et⊆A1withµ(Et) =t.
(b) Givenϕ(t)continuous and decreasing on[0,∞), there exists a measurable func- tion f onXwith f∗(t) =ϕ(t)for allt>0.
(c) GivenA⊆Xwith 0<µ(A)<∞andgan integrable function onX, there exists a subsetAeofX withµ(A) =e µ(A)such that
Z
Ae
g dµ= Z µ(A)
0
g∗(s)ds.
(d) Given f andgmeasurable functions onX, we have sup
h:dh=df
Z
X
h g dµ
= Z ∞
0
f∗(s)g∗(s)ds, where the supremum is taken over allhequidistributed with f.
Hint: Part (a): Reduce matters to the situation in which A0= /0. Consider first the case that for all A⊆X there exists a subset B of X satisfying 101 µ(A)≤ µ(B)≤ 109 µ(A). Then we can find subsets of A1 of measure in any arbitrar- ily small interval, and by continuity the required conclusion follows. Next con- sider the case in which there is a subset A1 of X such that every B⊆A1 satis- fiesµ(B)<101 µ(A1)orµ(B)>109 µ(A1). Without loss of generality, normalizeµ so that µ(A1) =1. Let µ1=sup{µ(C): C⊆A1,µ(C)<101} and pickB1⊆A1 such that 12µ1≤µ(B1)≤µ1. SetA2=A1\B1and define µ2=sup{µ(C): C⊆ A2,µ(C)<101}. Continue in this way and define setsA1⊇A2⊇A3⊇ · · ·and num- bers101 ≥µ1≥µ2≥µ3≥ · · ·. IfC⊆An+1withµ(C∪An+1)<101, thenC∪Bn⊆An withµ(C∪Bn)<15<109, and hence by assumption we must haveµ(C∪Bn)<101. Conclude that µn+1≤ 12µn and that µ(An)≥ 45 for alln=1,2, . . .. Then the set T∞
n=1An must be an atom. Part (b): First show that whend is a simple right con- tinuous decreasing function on[0,∞)there exists a measurable f on X such that f∗=d. For general continuous functions, use approximation. Part (c): Lett=µ(A) and defineA1={x: |g(x)|>g∗(t)}andA2={x: |g(x)| ≥g∗(t)}. ThenA1⊆A2 andµ(A1)≤t≤µ(A2). Pick Aesuch thatA1⊆Ae⊆A2 andµ(A) =e t=µ(A)by part (a). Then R
Aeg dµ=RXgχ
Aedµ=R0∞(gχ
Ae)∗ds=R0µ(eA)g∗(s)ds. Part (d): Let f =∑Nj=1ajχAj wherea1>a2>· · ·>aN >0 and theAj are pairwise disjoint.
Write f as∑Nj=1bjχBj, wherebj= (aj−aj+1)andBj=A1∪ · · · ∪Aj. PickBejas in part (c). ThenBf1⊆ · · · ⊆BfN and the function f1=∑Nj=1bjχ
Bfj has the same distri- bution function as f. It follows from part (c) thatRX f1g dµ=R0∞f∗(s)g∗(s)ds.The case of a general function f follows from that in which f is simple using Exercise 1.4.1 and approximation.
1.4.6.(Aoki [5]/ Rolewicz [224]) Let ·
be a nonnegative functional on a vector spaceXthat satisfies
x+y ≤K
x +
y
for allxandyinX. (To avoid trivialities, assume thatK≥1.) Then forαdefined by the equation
(2K)α=2 (α≤1), we have
x1+· · ·+xn
α≤4( x1
α+· · ·+ xn
α) for alln=1,2, . . .and allx1,x2,. . .,xninX.
Hint:Quasilinearity implies that kx1+· · ·+xnk ≤max1≤j≤n[(2K)jkxjk]for all
x1, . . . ,xninX(use thatK≥1). DefineH: X→Rby settingH(0) =0 andH(x) =
2j/αif 2j−1<kxkα≤2j. Thenkxk ≤H(x)≤21/αkxkfor allx∈X. Prove by induc- tion thatkx1+· · ·+xnkα≤2(H(x1)α+· · ·+H(xn)α).Suppose that this statement is true whenn=m. To show its validity forn=m+1, without loss of generality assume thatkx1k ≥ kx2k ≥ · · · ≥ kxm+1k. ThenH(x1)≥H(x2)≥ · · · ≥H(xm+1).
Assume that all theH(xj)’s are distinct. Then sinceH(xj)α are distinct powers of 2, they must satisfyH(xj)α≤2−j+1H(x1)α. Then
kx1+· · ·+xm+1kα ≤
1≤j≤m+1max (2K)jkxjkα
≤
1≤j≤m+1max (2K)jH(xj)α
≤
1≤j≤m+1max (2K)j21/α2−j/αH(x1)α
=2H(x1)α
≤2(H(x1)α+· · ·+H(xm+1)α).
We now consider the case thatH(xj) =H(xj+1)for some 1≤j≤m. Then for some integerrwe must have 2r−1<kxj+1kα≤ kxjkα≤2r andH(xj) =2r/α. Next note that
kxj+xj+1kα≤Kα(kxjk+kxj+1k)α≤Kα(2 2r)α≤2r+1. This implies
H(xj+xj+1)α≤2r+1=2r+2r=H(xj)α+H(xj+1)α.
Now apply the inductive hypothesis tox1, . . . ,xj−1,xj+xj+1,xj+1, . . . ,xm and use the previous inequality to obtain the required conclusion.
1.4.7.(Stein and Weiss [264]) Let(X,µ)and(Y,ν)be measure spaces. LetZbe a Banach space of complex-valued measurable functions onY. Assume thatZ is closed under absolute values and satisfies
f Z=
|f|
Z. Suppose thatTis a linear operator defined on the space of measurable functions on(X,µ)and taking values in Z. Suppose that for some constantA>0 we have the restricted weak type estimate
T(χE)
Z≤Aµ(E)1/p
for allE measurable subsets ofX and some 0<p<∞. Then there is a constant C(p)>0 such that
T(f)
Z≤C(p)A f
Lp,1
for all f in the domain ofT.
Hint:Let f =∑Nj=1ajχEj ≥0, wherea1>a2>· · ·>aN >0, µ(Ej)<∞pair- wise disjoint. Let Fj=E1∪ · · · ∪Ej, B0=0, and Bj =µ(Fj) for j≥1. Write
f =∑Nj=1(aj−aj+1)χFj, whereaN+1=0. Then T(f)
Z =
|T(f)|
Z
≤
N
∑
j=1(aj−aj+1) T(χFj)
Z
≤A
N
∑
j=1(aj−aj+1)(µ(Fj))1/p
=A
N−1
∑
j=0aj+1(B1/pj+1−B1/pj )
=p−1A f
Lp,1,
where the penultimate equality follows summing by parts; see Appendix F.
1.4.8.Let 0<p<∞and 0<q1<q2≤∞. Letα,β,q>0.
(a) Show that the functionf(t) =t−α(logt−1)−βχ(0,1)(t)lies inLp,q(R)if and only if eitherp>1/α or bothp=1/αandq>1/β.
(b) Show that the function t−1p(logt−1)−
1
q1χ(0,1)(t) lies in Lp,q2(R) but not in Lp,q1(R).
(c) OnRnconstruct examples to show thatLp,q1 $Lp,q2.
(d) On a general nonatomic measure space(X,µ)prove that theredoes notexist a constantC(p,q1,q2)>0 such that for allf inLp,q2(X)the following is valid:
f
Lp,q1 ≤C(p,q1,q2) f
Lp,q2.
1.4.9.(Stein and Weiss [263]) LetLp(ω)denote the weightedLpspace with mea- sureω(x)dx. LetT be a sublinear operator that maps
T: Lp0(ω0)→Lq0,∞(w), T: Lp1(ω1)→Lq1,∞(w),
for somep06=p1, where 0<p0,p1,q0,q1≤∞andω0,ω1,ωare positive functions.
Suppose that
1
pθ =1−θ p0 + θ
p1, 1
qθ =1−θ q0 + θ
q1.
ThenT maps
Lpθ ω
1−θ p0 pθ
0 ω
pθ1pθ 1
→Lqθ,pθ(ω). Hint:Define
L(f) = (ω1/ω0)p1−p1 0 f and observe that for eachθ∈[0,1],Lmaps
Lpθ ω
1−θ p0 pθ
0 ω
pθ1pθ 1
→Lpθ
(ω0p1ω1−p0)p1−p1 0
isometrically. Then apply the classical Marcinkiewicz interpolation theorem to the sublinear operatorT◦L−1,and the required conclusion easily follows.
1.4.10.(Kalton [147]/Stein and Weiss [266]) Letλnbe a sequence of positive num- bers with∑nλn≤1 and∑nλnlog(1
λn) =K<∞.
(a) Letfnbe a sequence of complex-valued functions inL1,∞(X)such that fn
L1,∞≤ 1 uniformly inn. Prove that∑nλnfnlies in L1,∞(X)with norm at most 2(K+2).
(This property is referred to as thelogconvexityofL1,∞.)
(b) Let Tn be a sequence of sublinear operators that mapL1(X)to L1,∞(Y) with norms
Tn
L1→L1,∞ ≤Buniformly in n. Use part (a) to prove that ∑nλnTn maps L1(X)toL1,∞(Y)with norm at most 2B(K+2).
(c) Givenδ>0 pick 0<ε<δ and use the simple estimate µ {
∞ n=1
∑
2−δnfn>α}
≤
∞ n=1
∑
µ {2−δnfn>(2ε−1)2−εnα}
to obtain a simple proof of the statements in part (a) and (b) whenλn=2−δn,n= 1,2, . . ., and zero otherwise.
Hint:Part (a): For fixedα >0, write fn=un+vn+wn, whereun= fnχ|fn|≤α
2, vn=fnχ|fn|> α
2λn, andwn= fnχα
2<|fn|≤2λnα . Letu=∑nλnun,v=∑nλnvn, andw=
∑nλnwn. Clearly|u| ≤α/2. Also{v6=0} ⊆Sn{|fn|> α
2λn}; henceµ({v6=0})≤α2. Finally,
Z
X
|w|dµ ≤
∑
n
λn Z
X
|fn|χα 2<|fn|≤ α
2λndµ
≤
∑
n
λn
Z α/(2λn)
α/2
dfn(β)dβ+ Z α/2
0
dfn(α/2)dβ
≤K+1.
Usingµ({|u+v+w|>α})≤µ({|u|>α/2}) +µ({|v| 6=0}) +µ({|w|>α/2}), deduce the conclusion.
1.4.11.Construct a sequence of functions fk inL1,∞(Rn)and a function f ∈L1,∞
such that fk−f
L∞→0 but fk
L1,∞→∞ask→∞.
1.4.12.(a) Suppose thatXis a quasi-Banach space and letX∗be its dual (which is always a Banach space). Prove that for allT∈X∗we have
T
X∗= sup
x∈X kxkX≤1
|T(x)|.
(b) Now suppose thatX is a Banach space. Use the Hahn–Banach theorem to prove that for everyx∈Xwe have
x
X= sup
T∈X∗ kTkX∗≤1
|T(x)|.
Observe that this result may fail for quasi-Banach spaces. For example, ifX=L1,∞, every linear functional onX∗vanishes on the set of simple functions.
(c) TakeX =Lp,1 andX∗=Lp0,∞. Then for 1<p<∞both of these spaces are normable. Conclude that
f
Lp,1 ≈ sup
kgkLp0,∞≤1
Z
X
f g dµ ,
f
Lp,∞ ≈ sup
kgkLp0,1≤1
Z
X
f g dµ .
1.4.13.Let 0<p,q<∞. Prove that any function inLp,q(X,µ)can be written as f =
+∞
n=−∞
∑
cnfn,
wherefnis a function bounded by 2−n/p, supported on a set of measure 2n, and the sequence{ck}klies in`qand satisfies
2−1p(log 2)1q {ck}k
`q≤ f
Lp,q≤ {ck}k
`q21p(log 2)1q. Hint:Letcn=2n/pf∗(2n)andfn=c−1n fχAn whereAnis as in (1.4.27).
1.4.14.(T. Tao) Let 0<p<∞, 0<γ<1,A>0, and let fbe a measurable function on a measure space(X,µ).
(a) Suppose that f
Lp,∞≤A. Then for every measurable setE of finite measure there exists a measurable subsetE0ofEwithµ(E0)≥γ µ(E)such that
Z
E0
f dµ
≤CγAµ(E)1−1p, whereCγ= (1−γ)−1/p.
(b) Conversely, if the last condition holds for someCγ,A<∞and all measurable
subsetsEof finite measure, then f
Lp,∞≤cγA, wherecγ=Cγ41/pγ−1
√ 2.
(c) Conclude that f
Lp,∞≈ sup
E⊆X 0<µ(E)<∞
inf
E0⊆E µ(E0)≥12µ(E)
µ(E)−1+1p Z
E0
f dµ .
Hint:Part (a): TakeE0=E\ {|f|>A(1−γ)−1pµ(E)−1p}. Part (b): Givenα>0, note that the set
|f|>α is contained in Re f>√α
2 ∪
Imf>√α
2 ∪
Ref<−√α
2 ∪
Imf<−√α
2 .
For E any of the preceding four sets, let E0 be a subset of it with measure at leastγ µ(E)such as in the hypothesis. Then
R
E0f dµ ≥√α
2γ µ(E), which gives
f
Lp,∞≤41/pγ−1Cγ√ 2A.
1.4.15.Given a linear operator T defined on the set of measurable functions on a measure space(X,µ)and taking values in the set of measurable functions on a measure space(Y,ν), define its “transpose”Ttvia the identity
Z
Y
T(f)g dν= Z
X
Tt(g)f dµ
for all measurable functions f onX andgonY, whenever the integrals converge.
LetT be such a linear operator given in the form T(f)(y) =
Z
X
K(y,x)f(x)dµ(x),
whereKis measurable and bounded by some constantM>0. Suppose thatT maps L1(X)toL1,∞(Y)with norm
T
, andTt mapsL1(Y)toL1,∞(X)with norm Tt
. Show that for all 1<p<∞there exists a constantCpthat depends only onpand is independent ofMsuch thatT mapsLp(X)toLp(Y)with norm
T
Lp(X)→Lp(Y)≤Cp T
1 p
Tt
1−1p
.
Hint:ForR>0, letBRbe the set of all(A,B), whereAis a measurable subset of X withµ(A)≤RandBis a measurable subset ofY withν(B)≤R. LetBR,M be the set of all(A,B)inBRsuch that|K(x,y)| ≤Mfor allx∈Aandy∈B. Also let Mp=Mp(R,M)<∞be the smallest constant such that for all(A,B)∈BR,Mwe have
R
BT(χA)dν
≤Mpµ(A)1pν(B)
1
p0. Letδ >0 and(A,B)∈BR,M. Ifµ(A)≤δ ν(B), use Exercise 1.4.14 to find a B0 with ν(B0)≥12ν(B)such that
R
B0T(χA)dν ≤ c
T
µ(A)≤cδ
1 p0
T
µ(A)1pν(B)
1
p0. Thenν(B\B0)≤12ν(B)and we have
Z
B\B0T(χA)dν
≤Mp2−
1
p0µ(A)1pν(B)p10.
Summing, we obtainMp≤Mp2−
1 p0+cδ
1 p0
T
. Wheneverν(B)≤δ−1µ(A), write
R
BT(χA)dν =
R
ATt(χB)dµ
and use Exercise 1.4.14 to find a set A0 with µ(A0)≥12µ(A)and
R
A0Tt(χB)dµ ≤c
Tt
ν(B). Argue similarly to obtainMp≤ Mp2−1p+cδ−
1 p
Tt
. Pick a suitableδ to optimize both expressions. Obtain that Mp is independent of R and M. Considering B+=B∩ {T(χA)>0} and B−= B∩ {T(χA)<0}, obtain thatRB
T(χA)
dν≤2Mpµ(A)1pν(B)p10 for(A,B)∈BR,M. Use Fatou’s lemma to remove the restriction that(A,B)∈BR,M. Finally, use the characterization of
·
Lp,∞ obtained in Exercise 1.1.12 withr=1 to conclude that T(χA)
Lp,∞≤Cp T
1 p
Tt
1−1p
µ(A)1p.
1.4.16.(Bourgain [29]) Let 0<p0<p1<∞and 0<α,β,A,B<∞. Suppose that a family of sublinear operatorsTkis of restricted weak type(p0,p0)with constant A2−kαand of restricted weak type(p1,p1)with constantB2−kβ for allk∈Z. Show that there is a constantC=C(α,β,p0,p1)such that∑k∈ZTk is of restricted weak type(p,p)with constantC A1−θBθ, whereθ=α/(α+β)and
1
p =1−θ p0
+ θ p1
.
Hint:Estimateµ({|T(χE)|>λ})by the sum∑k≥k0µ({|Tk(χE)|>cλ2α0(k0−k)})+
∑k≤k0µ({|Tk(χE)|>cλ2β0(k−k0)}), wherecis a suitable constant and 0<α0<α, β<β0<∞. Apply the restricted weak type(p0,p0)hypothesis on each term of the first sum, the restricted weak type(p1,p1)hypothesis on each term of the second sum, and choosek0to optimize the resulting expression.
APPENDIX: SOME MULTILINEAR INTERPOLATION
Multilinear maps are defined on products on linear spaces and take values in another linear space. We are interested in the situation that these linear spaces are function spaces. Let(X1,µ1), . . . ,(Xm,µm)be measure spaces, letDjbe spaces of measurable functions onXj, and letT be a map defined onD1× · · · ×Dmand taking values in the set of measurable functions on another measure space(Z,σ). ThenT is called multilinear if for all fj,gjinDjand all scalarsλ we have
|T(f1, . . . ,λfj, . . . ,fm)| = |λ| |T(f1, . . . ,fj, . . . ,fm)|,
T(f1, . . . ,fj+gj, . . . ,fm) = T(f1, . . . ,fj, . . . ,fm) +T(f1, . . . ,gj, . . . ,fm).
If Dj are dense subspaces of Lpj(Xj,µj) andT is a multlinear map defined on
∏mj=1Djand satisfies
T(f1, . . . ,fm)
Lp(Z)≤C f1
Lp1(X1)· · · fm
Lpm(Xm),
for allfj∈Dj, thenT has a bounded extension fromLp1× · · · ×Lpm→Z. The norm of a multilinear mapT : Lp1× · · · ×Lpm→Z is the smallest constantCsuch that the preceding inequality holds and is denoted by
T
Lp1×···×Lpm→Z.
Suppose thatT is defined on∏mj=1Dj, where eachDjcontains the simple func- tions. We say thatTis quasimultilinear if there is aK>0 such that for all 1≤j≤m, all fj,gjinDj, and allλ∈Cwe have
|T(f1, . . . ,λfj, . . . ,fm)| = |λ| |T(f1, . . . ,fj, . . . ,fm)|,
|T(f1, . . . ,fj+gj, . . . ,fm)| ≤ K |T(f1, . . . ,fj, . . . ,fm)|+|T(f1, . . . ,gj, . . . ,fm)|
. In the special case in whichK=1,T is called multisublinear.
1.4.17.LetT be a multilinear map defined on the set of simple functions of the product ofmmeasure spaces(X1,µ1)× · · · ×(Xm,µm)and taking values in the set of measurable functions on another measure space(Z,σ). Let 1≤pjk≤∞for 1≤ k≤mandj∈ {0,1}and also let 1≤pj≤∞for j∈ {0,1}. Suppose thatT satisfies
T(f1, . . . ,fm)
Lp j ≤Mj
f1
Lp j1· · ·
fm
Lp jm, j=0,1,
for all simple functions fkonXk. Let(1/q,1/q1, . . . ,1/qm)lie on the open line seg- ment joining(1/p0,1/p01, . . . ,1/p0m)and(1/p1,1/p11, . . . ,1/p1m)inRm+1. Then for some 0<θ<1 we have
1
q=1−θ p0 + θ
p1, 1
qk =1−θ p0k + θ
p1k, 1≤k≤m.
Prove thatT has a bounded extension fromLq1× · · · ×Lqm toLqthat satisfies T(f1, . . . ,fm)
Lq≤M01−θM1θ f1
Lq1· · ·
fm
Lqm
for all fk∈Lqk(Xk).
Hint:Adapt the proof of Theorem 1.3.4.
1.4.18.Let (X1,µ1), . . . ,(Xm,µm) be measure spaces, let Dj be spaces of mea- surable functions on Xj that contain the simple functions, and let T be a quasi- multilinear map defined onD1× · · · ×Dmthat takes values in the set of measurable functions on another measure space (Z,σ). Let 0<pjk ≤∞ for 1≤ j≤m+1 and 1≤k≤m, and also let 0<pj≤∞for 1≤ j≤m+1. Suppose that for all 1≤j≤m+1,T satisfies
T(χE1, . . . ,χEm)
Lp j,∞ ≤Mµ1(E1)
1
p j1· · ·µm(Em)
1 p jm
for all setsEkof finiteµkmeasure. Assume that the system