Factorization Theory
7.4 The Kwapie´ n-Maurey theorems for type-2 spaces
This implies that n+l i=1
|x∗(zi)|2= m j=1
|x∗(xj)|2, x∗∈X∗.
Then the vector-valued random variables n+l
i=1γizi andm
j=1γjxj have the same distributions onX. As a consequence,
E n+l i=1
γizi2=E m j=1
γjxj2. (7.18) Now,
E n i=1
γizi21/2
≤ 1 2
E n i=1
γizi+ n+l i=n+1
γizi21/2
+1 2
E n i=1
γizi−
n+l
i=n+1
γizi21/2
= E
n+l
i=1
γizi21/2
= E
m j=1
γjxj21/2
, which completes the proof.
Theorem 7.4.4.Let X andY be Banach spaces and E a closed subspace of X. Suppose T :E →Y is an operator. If X has type 2 and Y has cotype 2 then there is a Hilbert spaceH and operatorsS:X →H,R:H →Y so that RS ≤T2(X)C2(Y)T andRS|E=T.
Proof. We shall prove that for all sequences (zi)ni=1 in E and (xj)mj=1 in X such that
n i=1
|x∗(zi)|2≤ m j=1
|x∗(xj)|2, x∗∈X∗ (7.19) we have
n
i=1
T zi21/2
≤T2(X)C2(Y)Tm
j=1
xj21/2
,
and then we will appeal to the factorization criterion given by Theorem 7.3.4.
The key to the argument is to replace the Rademacher functions in the defi- nition of type and cotype by Gaussian random variables.
7.4 The Kwapie´n-Maurey theorems for type-2 spaces 189 On the one hand, for any (zi)ni=1 ⊂E, using the cotype-2 property of Y we have
n i=1
T zi2≤C2(Y)2 E n i=1
εiT zi2.
Then, if for each N ∈ N we consider (εki)1≤i,k≤N, a sequence of N ×N Rademachers,
n i=1
T zi2≤ C2(Y)2
N E
N k=1
n i=1
εkiT zi2
=C2(Y)2 E n i=1
N k=1
εki
√NT zi2.
Notice that for each 1 ≤ i ≤ n, the random variables εi1, εi2, . . . , εiN are independent and identically distributed, so by the Central Limit theorem, for each i the sequence PNk=1εik
√N
∞
N=1 converges in distribution to a Gaussian, γi. Thus,
lim
N→∞E n
i=1
N k=1
εki
√NT zi2=E n i=1
γiT zi2, and, therefore,
n i=1
T zi2≤C2(Y)2 E n i=1
γiT zi2. (7.20) On the other hand, if we let (εi)∞i=1be a sequence of Rademachers indepen- dent of (γi)∞i=1, for any sequence (xj)mj ⊂X, the symmetry of the Gaussians yields
E m j=1
γixj2=EEε m j=1
εjγjxj2
≤T2(X)2E m j=1
|γj|2xj2
=T2(X)2 m j=1
xj2E|γj|2
=T2(X)2 m j=1
xj2. (7.21)
Suppose that the vectors (zi)ni=1 in E and (xj)mj=1 in X satisfy equa- tion (7.19). Using Lemma 7.4.3 in combination with (7.18), (7.20), and (7.21) we obtain the inequality we need to apply Theorem 7.3.4:
n i=1
T zi2≤C2(Y)2E n i=1
γizi2
≤C2(Y)2T2E n i=1
γizi2
≤C2(Y)2T2E m j=1
γjxj2
≤C2(Y)2T2(X)2T2 m j=1
xj2.
There is a quantitative estimate here that we would like to emphasize:
Definition 7.4.5.IfXandY are two isomorphic Banach spaces, theBanach- Mazur distancebetweenX andY, denotedd(X, Y), is defined by the formula
d(X, Y) = inf
TT−1: T :X →Y is an isomorphism
.
The Banach-Mazur distance is not a distance in the real sense of the term since the triangle law does not hold, butdsatisfies a submultiplicative triangle inequality; that is,
d(X, Z)≤d(X, Y)d(Y, Z)
when X, Y, Z are all isomorphic. If X and Y are isometric then d(X, Y) = 1. The converse holds for finite-dimensional spaces but fails for infinite- dimensional spaces! (see the Problems).
In this language, Kwapie´n’s theorem (Theorem 7.4.1) really states:
Theorem 7.4.6.IfX is a Banach space of type 2 and cotype 2 then d(X, H)≤T2(X)C2(X)
for some Hilbert spaceH.
We have seen (Theorem 6.4.8) that ifp >2 every subspace of Lp which is isomorphic to a Hilbert space is necessarily complemented. Theorem 7.4.4 shows that this phenomenon is simply a consequence of the type-2 property:
Theorem 7.4.7 (Maurey). Let X be a Banach space of type 2. Let E be a closed subspace of X which is isomorphic to a Hilbert space. Then E is complemented inX.
Proof. Since E has cotype 2 the identity map on E can be extended to a projection ofX ontoE.
As we mentioned above, if we specialize the range space in Theorem 7.4.4 to be a Hilbert space then the assertion is a form of the Hahn-Banach theorem
7.4 The Kwapie´n-Maurey theorems for type-2 spaces 191 for Hilbert-space valued operators defined on a type-2 space. An interesting question is whether the extension property in Theorem 7.4.4 actually charac- terizes type-2 spaces:
Problem 7.4.8.SupposeX is a Banach space with the property that for every closed subspace E of X and every operatorT0 :E →H (H a Hilbert space) there is a bounded extensionT :X→H. Must X be a space of type2?
For a partial positive solution of this problem we refer to [28].
Up to now the only spaces that we have considered in the context of type and cotype are theLp-spaces (and their subspaces and quotients). It is worth pointing out that there are many other Banach spaces to which this theory can be applied. Perhaps the simplest examples are the “noncommutative”p- spaces orSchatten ideals. These are ideals of operators on a separable Hilbert space which were originally introduced in 1946 by Schatten and studied in several papers by Schatten and von Neumann; an account is given in [202].
IfH is a separable (complex) Hilbert space we defineSp to be the set of compact operators A : H → H so that the positive operator (A∗A)p/2 has finite trace and we impose the norm
ASp= tr (A∗A)p/2.
It is not entirely obvious, but is true, that this is a norm and that the class of such operators forms a Banach space.
In many ways the structure ofSp resembles that ofp.Thus if 1≤p≤2, Sp has typepand cotype 2, while if 2≤p <∞, Sp has cotypepand type 2 (see [215], [65]). See [5] for the structure of subspaces of Sp.
Recently there has been considerable interest in noncommutative Lp- spaces but even to formulate the definition would take us too far afield.
Problems
7.1.For 1≤r, p <∞, prove that the spacer(p) embeds inLp if and only ifr=p.
7.2.Let pn = 1 + n1. Consider the Banach space X =2(2p
n).Show that 21 does not embed isometrically intoX but thatd(X, X⊕221) = 1.
7.3.Show that any reflexive quotient of aC(K) space has type two.
7.4. The weakLp-spaces, Lp,∞.
Let (Ω, µ) be a probability measure space and 0 < p < ∞. A measurable functionf is said to belong toweakLp, denotedLp,∞,if
fp,∞= sup
t>0
tµ(|f|> t)1/p<∞.
(a) Show thatLp,∞is a linear space and that·p,∞is aquasi-normonLp,∞, i.e., · p,∞satisfies the properties of a norm except the triangle law which is replaced by
f +gp,∞≤C(fp,∞+gp,∞), f, g∈Lp,∞, whereC≥1 is a constant independent off,g.
(b) Show that Lp,∞ is complete for this quasi-norm and hence becomes a quasi-Banach space.
(c) Show that ifp >1, · p,∞ is equivalent to the norm fp,∞,c= sup
t>0
sup
µ(A)=t
t1/p−1
A
|f|dµ.
ThusLp,∞ can be regarded as a Banach space.
(d) Show thatLp,∞⊂Lrwhenever 0< r < p.
7.5 (Nikishin [158]). (Continuation.) SupposeX is a Banach space of type pfor some 1≤p <2. Suppose 1≤r < pand T :X →Lr(µ) is a bounded linear operator.
(a) Show that for some suitable constantC we have the following estimate:
µ !m
j=1
{|T xj| ≥1}1/r
≤C m
j=1
xjp1/p
, x1, . . . , xm∈X.
(b) For any constant K > C consider a maximal family of disjoint sets of positive measure (Ei)i∈I such that we can find xi ∈ X with xi ≤ 1 and
|T xi| ≥K(µ(Ei)−1/p) onEi.Show that this collection is countable and that
i∈I
µ(Ei)≤C K
p−rrp .
(c) Show that given >0 there is a setEwithµ(E)>1−so that the map TEf =χET f is a bounded operator fromX intoLp,∞(µ).
This gives a “factorization” through weakLp; it is possible to obtain a more elegant “change of density” formulation (see [186]). Note that ifX is an arbitrary Banach space andr <1 we get boundedness ofTE into weakL1. 7.6 (Jordan-von Neumann [96]). Show, without appealing to Kwapien’s theorem, that if a Banach space X has type 2 with T2(X) = 1 then X is isometrically a Hilbert space. [Hint:For real scalars, define an inner product by (x, y) =14(x+y2− x−y2).]
7.7.Let µ, ν be σ-finite measures. A linear operator T : Lp(µ) → Lr(ν), 0< r, p <∞, is said to be apositiveoperator iff ≥0 impliesT f ≥0.