Factorization Theory
7.1 Maurey-Nikishin factorization theorems
In this section we shall discuss factorization theory of operators with values in theLp-spaces. Here factorization is related to the notion of change of density.
The first factorization result of this type, essentially discovered by Nikishin [157], establishes a criterion for an operator with values in an Lp(µ)-space to factor through Lq(ν) (q > p), where ν is obtained from µ after a suit- able change of density. Nikishin’s motivation came from harmonic analysis rather than Banach space theory, where versions of this result for translation- invariant operators had been known for some time (e.g., in the work of Stein [210]). However, it was the work of Maurey [144] that combined the ideas of Nikishin with the newly evolving theory of Rademacher type to create a very powerful tool.
The proof given below is based on one presented in [221] but is similar to the proof given by Maurey.
Definition 7.1.1.If (Ω,Σ, µ) is aσ-finite measure space then adensity func- tionhon Ω is a measurable function such thath≥0 a.e. and
h dµ= 1.
Theorem 7.1.2.Let µ be a σ-finite measure on some measurable space (Ω,Σ). Suppose that T is an operator from a Banach space X into Lp(µ) and that 1≤p < q <∞. Suppose0< C <∞. Then the following conditions are equivalent:
(a) There exists a density functionhon Ωsuch that
{h>0}|T x|qh1−q/pdµ 1/q
≤Cx, x∈X, (7.1) and
µ{ω: |T x(ω)|>0, h(ω) = 0}= 0, x∈X. (7.2) (b) For every finite sequence(xk)nk=1 inX,
n
k=1
|T xk|q1/q
p≤C n
k=1
xkq1/q
. (7.3)
Interpretation. Condition (a) is to be interpreted in the sense that each function T x is essentially supported on A ={ω ∈ Ω :h(ω) >0}. Thus the operator Sx := h−1/pT x maps into Lp(Ω, h dµ). However, (a) asserts that S actually maps boundedly into the smaller space Lq(Ω, h dµ). This is the diagram depicting the situation:
X T //
Lp(µ)
Lq(hdµ) //Lp(hdµ)
j
OO
Here j is an isometric embedding ofLp(h dµ) onto the subspace Lp(A, µ) of Lp(µ), defined byj(f) =f h1/p.
7.1 Maurey-Nikishin factorization theorems 167 Of course, at a very small cost we could insist thathis a strictly positive density (i.e., h > 0 a.e.) and drop equation (7.2): simply replace h by (1 + v)−1(h+v) where > 0 and v is any strictly positive density. Then j becomes a genuine isometric isomorphism. In this case, however, the norm of S=h−1/pT is a little greater thanC. Since the precise value ofS is rarely of interest we will often use the theorem is this form. In fact, in a formal sense we could replace (7.1) and (7.2) by
Ω
|T x|qh1−q/pdµ 1/q
≤Cx, x∈X,
with the implicit understanding thatT x= 0 a.e. on the set{ω∈Ω :h(ω) = 0} (i.e., whereh−q/p= 0). We will use this convention later.
Before continuing let us notice that, although we have stated this for gen- eral σ-finite measures, it is enough to prove the theorem under our usual convention that µis a probability measure. Ifµis not a probability measure we choose some strictly positive densityv and set dµ =v dµ; then we define T :X →Lp(µ) by Tx= v−1T x. A quick inspection will show the reader that the statement of the theorem forT implies exactly the same statements for T. Thus we can and do resume our convention that µ is a probability measure.
Proof. (a)⇒(b) Since (Ω, h dµ) is a probability measure space andp < q, the Lp(hdµ)-norm is smaller than the Lq(hdµ)-norm and thus we have
Ω
n
k=1
|T xk|qp/q
dµ 1/p
=
{h>0}
n
k=1
|T xk|qh−q/pp/q
h dµ 1/p
≤ {h>0}
n k=1
|T xk|qh−qph dµ 1/q
= n
k=1
{h>0}|T xk|qh−qphdµ 1/q
≤C n
k=1
xkq1/q
.
(b)⇒(a) Let us assume that C is the best constant so that (7.3) holds.
Then sup
n
k=1
|T xk|q1/q
p
: (xk)nk=1⊂X, n k=1
xkq ≤C−q, n∈N
= 1.
LetW0 be the set of all nonnegative functions inL1 that are bounded above by functions of the form (n
k=1|T xk|q)p/q, where n ∈ N and (xk)nk=1 ⊂ X withn
k=1xkq ≤C−q, i.e.,
0≤f ≤n
k=1
|T xk|qp/q
; letW be the norm closure ofW0.
W0andW have the following property:
(*) Let r = q/p > 1. Given f1, . . . , fn ∈ W0 [respectively, W] and c1, . . . , cn ≥0 withc1+· · ·+cn≤1 then(c1f1r+· · ·+cnfnr)1/r∈W0 [respectively,W].
To prove (*) it suffices to consider the case ofW0.Suppose
0≤fk ≤
⎛
⎝mk
j=1
|T xjk|q
⎞
⎠
p/q
, 1≤k≤n,
wheremk
j=1xjkq ≤C−q for 1≤k≤n.Then we also have 0≤
- n
k=1
ckfkr .1/r
≤
⎛
⎝n
k=1 mk
j=1
|T(ck1qxjk)|q
⎞
⎠
p/q
,
with n
k=1
ck
mk
j=1
xjkq≤C−q, and this establishes (*).
Property (*) immediately yields thatW0(and hence its norm-closureW) is convex. Indeed, iff1, . . . , fn∈W0andc1, . . . , cn≥0 withc1+· · ·+cn= 1 then by (*) we obtain
n j=1
cjfj≤
⎛
⎝n
j=1
cjfjr
⎞
⎠
1/r
∈W0.
Using Mazur’s theorem, W is therefore weakly closed. Note that from the choice of C, we have
sup
f∈W0
f dµ= sup
f∈W
f dµ= 1,
so in particularW is bounded. We next show thatW is weakly compact. This requires to show that it is equi-integrable.
SupposeW is not equi-integrable. Then there is someδ > 0, a sequence (fn)∞n=1in W, and a sequence of disjoint measurable sets (En)∞n=1 such that
En
fndµ > δ >0, n∈N.
7.1 Maurey-Nikishin factorization theorems 169
Thus for anyN we have
δN1−1r ≤N−1r
max(f1, f2, . . . , fN)dµ
≤ ⎛⎝1
N N j=1
fjr
⎞
⎠
1/r
dµ
≤1,
by using (*). This is a contradiction for large enoughN.
HenceW is weakly compact and, since integration is a weakly continuous functional onL1(µ), it follows that there existsh∈W with
h dµ= 1. (7.4)
Now supposef ∈W.On the one hand, for anyτ >0 we have (1 +τ)−1r(hr+τ fr)1r ∈W,
therefore, by property (*),
(hr+τ fr)1rdµ≤(1 +τ)1r. (7.5) On the other hand,
(hr+τ fr)1rdµ≥1 +τ1r
h=0
f dµ. (7.6)
Since 1/r <1, combining (7.5) and (7.6) yields
{h=0}
f dµ= 0. (7.7)
From (7.6) and (7.4) we have
{h>0}
h(1 +τ frh−r)1r −1
τ dµ≤ (1 +τ)1r −1
τ , τ >0.
Lettingτ→0 and using Fatou’s lemma we obtain
{h>0}
frh1−rdµ≤1, f ∈W. (7.8) In particular (7.7) and (7.8) hold forf =C−px−p|T x|pwhen 0=x∈X.
This immediately gives (7.2) and (7.1).
Theorem 7.1.3.Let 1 ≤ p < ∞. Suppose that T is an operator from a Banach space X into Lp(µ). If X has type 2 then there exists a constant C=C(p)such that for every finite sequence(xk)nk=1 inX we have
n
k=1
|T xk|21/2
p≤C n
k=1
xk21/2
.
Proof. By Theorem 6.2.13, for every 1≤p <∞there is a constantc=c(p) such that for any finite set of vectors (xk)nk=1 inX,
n
k=1
|T xk|212
p≤cE n k=1
εkT xk
p≤cTE n k=1
εkxk.
Using Kahane’s inequality and the type 2 ofX, E
n k=1
εkxk≤ E
n k=1
εkxk21/2
≤T2(X)n
k=1
xk212 .
SinceLr(µ) forr≥2 are type-2 spaces, we immediately obtain:
Corollary 7.1.4.
(a) Every operator from a subspace of Lr(µ) (2 ≤r < ∞) into Lp(µ) (1 ≤ p <2) factors through a Hilbert space.
(b) If a Banach spaceX is isomorphic to a closed subspace of bothLp(µ)for some1≤p <2 andLr(µ)for some2< r <∞, thenX is isomorphic to a Hilbert space.
Corollary 7.1.4 follows immediately from Theorems 7.1.2 and 7.1.3. Curi- ously, the isometric version of (b) does not hold. That is, ifX is isometric to a subspace ofLp (1≤p <2) and isometric to a subspace ofLr(2< r <∞), it is not true thatX must be isometric to a Hilbert space. Finite-dimensional counterexamples were given by Koldobsky [112]; however, the following prob- lem is still open (see [114]):
Problem 7.1.5.If an infinite-dimensional Banach space X is isometricto a closed subspace of bothLp for some 1≤p <2 andLr for some 2< r <∞, must X be isometric to a Hilbert space?
To push our results further we need a replacement for Theorem 6.2.13 for exponents other than 2.If 1≤q <2 then it turns out that theq-stable random variables constructed in the previous chapter do very nicely. Indeed, we could have used Gaussians in place of Rademachers in the preceding argument.
Lemma 7.1.6.Let 1 ≤p < q < 2. Suppose that γ = (γj)∞j=1 is a sequence of independent normalized q-stable random variables. Then for any finite se- quence of functions(fj)nj=1 inLp(µ),
7.1 Maurey-Nikishin factorization theorems 171 n
j=1
|fj|q1/q
p
=c E
n j=1
γjfjp
p
1/p
, wherec=c(p, q)>0.
Proof. We recall from Theorem 6.4.18 that there is a constantc =c(p, q) so that
⎛
⎝E n j=1
ajγjp
⎞
⎠
1/p
=c−1
⎛
⎝n
j=1
|aj|q
⎞
⎠
1/q
, (aj)nj=1⊂R. Using Fubini’s theorem,
n j=1
|fj|qpq
dµ=cpE n
j=1
γjfjpdµ, and the lemma follows.
Theorem 7.1.7.Let1≤p <2. Suppose thatT is an operator from a Banach spaceXintoLp(µ). IfXhas typerfor somep < r <2, then for eachq∈(p, r) there exists a constant C such that
n
j=1
|T xj|q1/q
p≤C n
j=1
xjq1/q
, for every finite sequence (xj)nj=1 inX.
Proof. In this proof we will require three mutually independent sequences of independent identically distributed random variables: a sequence (εj)∞j=1 of Rademachers, a sequence (γj)∞j=1 of normalizedq-stable random variables and a sequence (ηj)∞j=1 of normalized r-stable random variables.
Let (xj)nj=1be a finite sequence inX. By the previous lemma, for a certain constantc=c(p, q) we have
n
j=1
|T xj|q1/q
p
=c Eγ
n j=1
γjT xjp
p
1/p
≤cT Eγ
n j=1
γjxjp1/p
. Since the normalizedq-stables are symmetric andX has typer,
Eγ n j=1
γjxjp1/p
=
EγEε n j=1
εjγjxjp1/p
≤ Eγ
⎛
⎝Eε n j=1
εjγjxjr
⎞
⎠
p/r1/p
≤Tr(X) Eγ
n
j=1
|γj|rxjrp/r1/p
. Now notice that
E n j=1
ajηjp=c1 n
j=1
|aj|rp/r
,
for a certain constant 0< c1=E|η1|p which is finite sincep < r.Thus letting c2, c3 be positive constants depending only onp, q, andr,
Eγ
n
j=1
|γj|rxjrp/r
=c−11EγEη n j=1
ηjγjxjp
=c−11EηEγ n j=1
ηjγjxjp
=c2Eη
n
j=1
|ηj|qxjqp/q
≤c2 Eη
n j=1
|ηj|qxjqp/q
=c3 n
j=1
xjqp/q
.
The next result now follows immediately from Theorem 7.1.2:
Theorem 7.1.8.Let X be a Banach space of type r > 1. Suppose that 1 ≤ p < r and that T :X →Lp(µ)is an operator. ThenT factors through Lq(µ) for any p < q < r. More precisely, for each p < q < r there is a strictly positive density function h on Ω so that Sx = h−1/pT x defines a bounded operator fromLp(µ)intoLq(Ω, h dµ).
Note here that there is a fundamental difference between the case of type r <2 and type 2. In the former we only obtain a factorization throughLq(µ) when q < r. Can we do better and takeq=r? The answer is no and to see why we must consider subspaces ofLpfor 1≤p <2. This will be the topic of the next section, but let us mention that an improvement is possible: A later theorem of Nikishin [158] implies that T actually factors through the space
“weakLr.” See [186] and the Problems.
Remark 7.1.9.An examination of the proofs of the theorems of this section shows that the main theorem (Theorem 7.1.8) will also hold if 0 < p < 1, when Lp is no longer a Banach space; in this case we can take r = 1 and