Factorization Theory
7.3 Factoring through Hilbert spaces
In the first section of this chapter we saw that ifX has type 2 and 1≤p <2 then any operatorT :X→Lpfactors through a Hilbert space. In this section we are giving a characterization for an operator between Banach spaces to factor through a Hilbert space.
Definition 7.3.1.Suppose thatX andY are Banach spaces. We say that an operatorTfromXtoY factors through a Hilbert spaceif there exist a Hilbert spaceH and operatorsS:X −→H andR:H −→Y verifyingT =RS.
7.3 Factoring through Hilbert spaces 181 We will begin by making some remarks that will lead us to the necessary condition we are seeking. We will only consider real scalars, although at the appropriate moment we will discuss the alterations necessary to handle com- plex scalars. Throughout this section H will denote a generic Hilbert space with a scalar product · .
Suppose we havenarbitrary vectorsx1, . . . , xninH. Given a real orthog- onal matrix A = (aij)1≤i,j≤n, let us consider the new vectors in H defined fromA,
zi= n j=1
aijxj, 1≤i≤n. (7.14) Then,
n i=1
zi2= n i=1
n j=1
aijxj2
= n i=1
n j=1
aijxj, n k=1
aikxk
= n i=1
n j=1
n k=1
aijaikxj, xk
= n j=1
xj, xj
= n j=1
xj2.
Any real n×n matrixA = (aij) defines a linear operator (that will be denoted in the same way)A:n2 −→n2 via
A
⎛
⎜⎜
⎜⎝ s1 s2 ... sn
⎞
⎟⎟
⎟⎠=
⎛
⎜⎜
⎜⎝
a11 a12 . . . a1n a21 a22 . . . a2n ... ... . .. ... an1 an2. . . ann
⎞
⎟⎟
⎟⎠
⎛
⎜⎜
⎜⎝ s1 s2 ... sn
⎞
⎟⎟
⎟⎠.
The matrix (aij)1≤i,j≤n is orthogonal if and only if the operator A is an isometry. If (aij)ni,j=1 is not orthogonal but A ≤ 1, it is an exercise of linear algebra to prove that (aij) can be written as a convex combination of orthogonal matrices. In fact, it is always possible to find orthonormal basis (ej)nj=1 and (fj)nj=1 in n2 so that Aej = λjfj with λj ≥ 0: Just find an orthonormal basis of eigenvectors (ej)nj=1 forAA whereA is the transpose.
Then A =DU where Dfj =λjfj and U ej =fj. U is orthogonal and since 0≤λj ≤1 we can writeDas a convex combination of the orthogonal matrices Vfj=jfj wherej =±1.
Thus, ifx1, . . . , xn, z1, . . . , znare arbitrary vectors inHsatisfying equation (7.14), where(ajk)nj,k=1n2→n2 ≤1, we will have
n i=1
zi2≤ n j=1
xj2.
This can easily be extended to the case of differing numbers ofxj’s andzi’s by adding zeros to one of the two collections of vectors.
Theorem 7.3.2.LetT be an operator from a Banach spaceX into a Banach space Y. Suppose that there exist operators S :X −→ H and R :H −→Y verifying T = RS. If (xj)mj=1 and (zi)ni=1 are vectors in X related by the equation
zi= m j=1
aijxj, 1≤i≤n, (7.15) where(aij)is a realn×m matrix such thatAm2→n2 ≤1, then
n
i=1
T zi21/2
≤ S Rm
j=1
xj21/2
.
Proof. The proof easily follows from the comments we made. Indeed, given x1, . . . , xmandz1, . . . , zninXsatisfying (7.15), since the collections of vectors (Sxj)mj=1 and (Szi)ni=1 lie insideH we have
n i=1
T zi2= n i=1
RSzi2
≤ R2 n i=1
Szi2
≤ R2 m j=1
Sxj2
≤ R2S2 m j=1
xj2.
In light of the previous theorem we want to give an alternative formulation of the property that (xj)mj=1 and (zi)ni=1 are vectors in X related by the equation
zi= m j=1
aijxj, 1≤i≤n,
whereA= (aij) is a realn×mmatrix such thatAm2→n2 ≤1.
7.3 Factoring through Hilbert spaces 183 Proposition 7.3.3.Givenn, m∈Nand any two sets of vectors (xj)mj=1 and (zi)ni=1 in a Banach spaceX, the following are equivalent:
(a)There is a realn×m matrixA= (aij)so that Am2→n2 ≤1 and zi=
m j=1
aijxj, 1≤i≤n;
(b) m j=1
|x∗(zj)|2≤ n i=1
|x∗(xi)|2, for all x∗∈X∗.
Proof. Assume that (a) holds. Then, sinceAm2→n2 ≤1, it follows that n
i=1
|x∗(zi)|2= n i=1
x∗m
j=1
aijxj2= n i=1
m j=1
aijx∗(xj)2≤ m j=1
|x∗(xj)|2.
For the reverse implication, (b)⇒(a), consider the linear operators α:X∗−→m2 , x∗→(x∗(xj))mj=1
and
β :X∗ −→n2, x∗→(x∗(zi))ni=1. The hypothesis says that βx∗m
2 ≤ αx∗n
2 for all x∗ ∈X∗. Thus we can define an operatorA0:α(X∗)→β(X∗) withA0 ≤1 andβ=A0◦α.Then A0 can be extended to an operatorA:m2 →n2 withA ≤1. Let (aij) be the matrix associated withA.
x∗(zi) = m j=1
aijx∗(xj) for all x∗∈X∗, which implies
zi= m j=1
aijxj, i= 1, . . . , n.
The main result of this section is the following criterion:
Theorem 7.3.4.LetX andY be Banach spaces. SupposeEis a closed linear subspace ofX andT :E→Y is an operator. In order that there exist a Hilbert spaceH and operatorsR:X →H, S :H →Y with RS ≤C such that T =RS|Eit is necessary and sufficient that for all sets of vectors(xj)mj=1⊂X and(zi)ni=1⊂E such that
n i=1
|x∗(zi)|2≤ m j=1
|x∗(xj)|2, x∗∈X∗,
we have
n
i=1
T zi21/2
≤C m
j=1
xj21/2
.
In the proof of this result and other ones in the next chapter we will make use of the following lemma. IfAis a subset of real vector space we define
cone (A) = n
j=1
αjaj: a1, . . . , an ∈A, α1, . . . , αn≥0, n= 1,2, . . .
. Lemma 7.3.5.LetV be a real vector space. GivenA,Btwo subsets ofV such that V =cone(B)−cone(A), and two functions φ:A →R,ψ:B →R, the following are equivalent:
(i)There is a linear functionalL onV verifying φ(a)≤ L(a), a∈ A and
ψ(b)≥ L(b), b∈ B.
(ii) If (αi)mi=1, (βj)nj=1 are two finite sequences of nonnegative scalars such
that m
i=1
αiai= n j=1
βjbj for some(ai)mi=1⊂ A,(bj)nj=1⊂ B, then
m i=1
αiφ(ai)≤ n j=1
βjψ(bj).
Proof. The implication (i)⇒(ii) is immediate.
(ii)⇒(i) Let us define the mapp:V →[−∞,∞) as follows:
p(v) = inf n
j=1
βjψ(bj)− m i=1
αiφ(ai)
,
the infimum being taken over all possible representations in the form v = n
j=1βjbj −m
i=1αiai, where α1, . . . , αm, β1, . . . , βn ≥ 0, a1, . . . , am ∈ A, andb1, . . . , bm∈ B.
pis well-defined since V= cone (B)−cone (A). Besides, one easily checks that p is positive-homogeneous and satisfies p(v1+v2) ≤ p(v1) +p(v2) for any v1, v2 in V. In order to prove thatpis a sublinear functional we need to show that p(v)> −∞for every v ∈ V. This will follow ifp(0) = 0. Indeed, p(v) +p(−v)≥p(0), so neither p(v) norp(−v) could be−∞ifp(0) = 0.
7.3 Factoring through Hilbert spaces 185 For each representation of 0 in the form 0 =n
j=1βjbj−m
i=1αiai, by the hypothesis it follows that n
j=1βjψ(bj) ≥ m
i=1αiφ(ai). Therefore, by the definition,p(0)≥0 hencep(0) = 0.
As an consequence of the Hahn-Banach theorem, there is a linear func- tional L onV such that L(v)≤p(v) for everyv ∈ V and so φ(a)≤ L(a) for alla∈ AandL(b)≤ψ(b) for allb∈ B.
Proof of Theorem 7.3.4. We need only show that the condition is sufficient.
Let F(X∗) denote the set of all functions from X∗ to R, and consider the natural mapX→ F(X∗),x→x, whereˆ
ˆ
x(x∗) =x∗(x), x∗∈X∗.
Let V be the linear subspace of F(X∗) of all finite linear combinations of functions of the form ˆxˆz, withx,zin X. That is,
V = N
k=1
λkxˆkzˆk: (λk)Nk=1inR, (xk)Nk=1and (zk)Nk=1 inX,andN∈N . Clearly, the set{xˆ2 : x∈X}spansV since each product ˆxˆzwithxandzin X can be written in the form
ˆ xˆz= 1
4
(ˆx+ ˆz)2−(ˆx−z)ˆ 2 .
We want to construct a linear functionalLonV with the following prop- erties:
0≤ L(ˆx2)≤C2x2, x∈X (7.16) and
T x2≤ L(ˆx2), x∈E. (7.17) To this end, let us apply Lemma 7.3.5 in the caseA=B={xˆ2: x∈X}by putting
φ(ˆx2) =
0 if x∈X\E T x2 if x∈E and
ψ(ˆx2)2=C2x2. Suppose that
n i=1
βi2zˆi2= m j=1
α2jxˆ2j
for some (ˆxj)mj=1, (ˆzi)ni=1vectors inX, and some nonnegative scalars (α2j)mj=1, (βj2)nj=1. Let us supposez1, . . . , zl∈E andzl+1, . . . , zn∈X\E.Then
l i=1
β2izˆi2≤ m j=1
α2jxˆ2j, hence
l i=1
T(βjzi)2≤C2 m j=1
αjxj2.
Thus n
i=1
βi2φ(ˆz2i)≤ m j=1
α2jψ(ˆx2j).
Lemma 7.3.5 yields a linear functionalL onV with φ(ˆx2)≤ L(ˆx2)≤ψ(ˆx2), x∈X.
L, in turn, induces a symmetric bilinear form · onX given by x, z=L(ˆxˆz),
so the mapX −→[0,∞),x→1
x, x=1
L(ˆx2) defines a seminorm onX. Thus,X (modulo the subspace{x; x, x= 0}) endowed with the (now) inner product, is an inner product space, andx0=1
x, xa norm on X. LetH be the completion ofX0 under this norm.H is a Hilbert space.
TakeSto be the induced operatorS:X →Hmappingxto its equivalence class in X0.Then we have
Sx ≤Cx, x∈X.
S has norm one and dense range. By construction, ifx∈E we have T x ≤ Sx,
therefore we can find an operator R0 : S(E)→ Y with R0 ≤ 1 and T = R0S|E.ComposeR0with the orthogonal projection ofH ontoS(E) to create R.
The proof for complex scalars.In the case whenX andY are complex Banach spaces we proceed as first by “forgetting” their complex structure and treating them as real spaces. The argument creates a real symmetric bilinear form · onX which is continuous for the original norm. We can then define a complex inner product by “recalling” the complex structure ofX and setting
(x, z) = 1 2π
2π 0
eiθx, eiθz −iieiθx, eiθzdθ.
We leave it to the reader to check that this induces a complex inner product and that using this to defineH gives the same conclusion.