Introduction
The separable quotient problem asks whether every infinite-dimensional Banach space admits a separable infinite-dimensional quotient. Equivalently, given an infinite-dimensional Banach space X X X , does there exist a closed subspace Y Y Y of X X X such that X / Y X/Y X / Y is separable and infinite dimensional? One of the attractive features of this problem is that its formulation is accessible to anyone who has completed a first course in functional analysis. Every infinite-dimensional Banach space contains a separable infinite-dimensional closed subspace, obtained by taking the closed linear span of a countable linearly independent set. The corresponding question for quotients has proved much more difficult.
Traditionally associated with Banach and Mazur, the problem has occupied a central place in the structural theory of Banach spaces. Among the positive results preceding the work discussed here, Johnson and Rosenthal [14 ] proved in 1972 that every separable infinite-dimensional Banach space has an infinite-dimensional quotient with a Schauder basis. In 2008, Argyros, Dodos and Kanellopoulos [3 ] established that every infinite-dimensional dual Banach space has a separable infinite-dimensional quotient.
The nontrivial case of the separable quotient problem concerns nonseparable spaces, since a separable infinite-dimensional space is already a quotient of itself. In the separable setting, a related question asks whether there exists a separable infinite-dimensional Banach space such that every infinite-dimensional quotient has a nonseparable dual. This question, posed by Johnson and Rosenthal [14 ] , has a positive answer. Such a space is supplied by the construction of Argyros, Arvanitakis and Tolias in [1 Section 6] ; the quotient property is explicitly established by Argyros and Beanland in [2 Section 5, Corollary 22] . Thus nonseparability of the dual may persist in every infinite-dimensional quotient of a separable space.
In the preprint Relative independence of the separable quotient problem , dated 23 September 2026, OpenAI [15 ] establishes two opposite conclusions under additional set-theoretic assumptions. If the continuum c c c is real-valued measurable, every infinite-dimensional Banach space has a separable infinite-dimensional quotient. Under the continuum hypothesis, however, there exists a Banach space X X X of density ℵ 1 \aleph_{1} ℵ 1 which has no separable infinite-dimensional quotient. Together with the relative consistency of a real-valued measurable continuum from a measurable cardinal, these results yield the independence of the separable quotient problem from ZFC, relative to the consistency of ZFC with a measurable cardinal.
The aim of the present paper is to describe the mechanism leading to the space X X X and to offer an alternative access to its definition and the proof of its properties. We separate the functional-analytic argument from the construction of the auxiliary space E E E , and organize the latter in two stages. This makes it possible to see how the different properties of E E E enter the definition of X X X before considering the norming sets which produce them.
The paper is organized into four sections. In Section 1, we state and prove the following criterion, which is Theorem 1.2 below.
Theorem 1.2. Assume that a Banach space E E E has a transfinite Schauder basis ( e α ) α < ω 1 (e_{\alpha})_{\alpha<\omega_{1}} ( e α ) α < ω 1 and satisfies the following conditions.
(i) Setting
X 0 = span ‾ { e α ∗ : α < ω 1 } ⊂ E ∗ , X_{0}=\overline{\operatorname{span}}\{e_{\alpha}^{*}:\alpha<\omega_{1}\}\subset E^{*}, X 0 = span { e α ∗ : α < ω 1 } ⊂ E ∗ , the canonical evaluation map identifies E E E with X 0 ∗ X_{0}^{*} X 0 ∗ .
(ii) The space E E E contains no isomorphic copy of ℓ 1 \ell_{1} ℓ 1 , and every infinite-dimensional closed subspace of E E E has nonseparable dual.
(iii) There exists a closed subspace X ⊂ E ∗ X\subset E^{*} X ⊂ E ∗ such that
X + X 0 ‾ ∥ ⋅ ∥ = E ∗ \overline{X+X_{0}}^{\|\cdot\|}=E^{*} X + X 0 ∥ ⋅ ∥ = E ∗ and the canonical restriction map
R : E ⟶ X ∗ , ( R e ) ( x ) = x ( e ) , R:E\longrightarrow X^{*},\qquad(Re)(x)=x(e), R : E ⟶ X ∗ , ( R e ) ( x ) = x ( e ) , is an onto isomorphism.
Then X X X has no infinite-dimensional separable quotient.
As W. B. Johnson pointed out, the existence of spaces satisfying all the assumptions of this theorem implies CH. Thus CH is necessary for this particular construction of X X X ; the auxiliary space E E E itself is constructed in ZFC. The proof of the criterion is comparatively short. What we find remarkable in the AI contribution is the identification of the assumptions which make this argument possible.
Section 2 explains a second remarkable step. For an auxiliary space E E E with the stated properties, let
Q : E ∗ ⟶ W = E ∗ / X 0 Q:E^{*}\longrightarrow W=E^{*}/X_{0} Q : E ∗ ⟶ W = E ∗ / X 0
be the quotient map. We show that if a bounded operator B : E ∗ → W B:E^{*}\to W B : E ∗ → W satisfies
B ( X 0 ) ‾ = W , ∥ B ∥ < 1 10 , B ∗ ( W ∗ ) ⊂ J E ( E ) , \overline{B(X_{0})}=W,\qquad\|B\|<\frac{1}{10},\qquad B^{*}(W^{*})\subset J_{E}(E), B ( X 0 ) = W , ∥ B ∥ < 10 1 , B ∗ ( W ∗ ) ⊂ J E ( E ) ,
then X = ker ( Q − B ) X=\ker(Q-B) X = ker ( Q − B ) satisfies the assumptions of Theorem 1.2 and hence has no separable infinite-dimensional quotient. Here J E : E ⟶ E ∗ ∗ J_{E}:E\longrightarrow E^{**} J E : E ⟶ E ∗∗ is the canonical embedding. The adjoint range condition follows from a norm-convergence property of B B B on the finite initial sums of the special functionals. Informally, B B B may be viewed as an anti-Q Q Q operator : whereas Q Q Q annihilates X 0 X_0 X 0 , the operator B B B maps X 0 X_0 X 0 onto a norm-dense subspace of W W W . The construction exploits these contrasting behaviours by considering the subspace on which Q Q Q and B B B agree.
In Section 3, we present the properties of E E E which yield the existence of B B B . In contrast with the single construction of the auxiliary space in [15 ] , our description proceeds in two stages. First we consider spaces E p , r E_{p,r} E p , r , where 1 < r < p < ∞ 1<r<p<\infty 1 < r < p < ∞ . For parameters
1 < r 2 < p 2 < r 1 < p 1 < ∞ 1<r_2<p_2<r_1<p_1<\infty 1 < r 2 < p 2 < r 1 < p 1 < ∞
put E i = E p i , r i E_i=E_{p_i,r_i} E i = E p i , r i and W i = E i ∗ / X 0 , i W_i=E_i^*/X_{0,i} W i = E i ∗ / X 0 , i , with quotient maps Q i : E i ∗ → W i Q_i:E_i^*\to W_i Q i : E i ∗ → W i . Under CH, Proposition 3.2 constructs a contraction B 2 , 1 : E 2 ∗ → W 1 B_{2,1}:E_2^*\to W_1 B 2 , 1 : E 2 ∗ → W 1 such that
B 2 , 1 e α ( 2 ) ∗ = Q 1 ϕ σ ( α ) ( 1 ) B_{2,1}e_\alpha^{(2)*}=Q_1\phi_{\sigma(\alpha)}^{(1)} B 2 , 1 e α ( 2 ) ∗ = Q 1 ϕ σ ( α ) ( 1 )
where σ \sigma σ bijectively pairs the biorthogonal functionals of E 2 E_2 E 2 with the indices of the special functionals of E 1 E_1 E 1 . In particular, B 2 , 1 ( X 0 , 2 ) B_{2,1}(X_{0,2}) B 2 , 1 ( X 0 , 2 ) is norm dense in W 1 W_1 W 1 . Moreover, B 2 , 1 B_{2,1} B 2 , 1 maps the finite initial sums of every special functional of E 2 E_2 E 2 to a sequence converging in norm to the image of that functional. Its boundedness follows from the ℓ p 2 \ell_{p_2} ℓ p 2 estimate for the scalar families ( f ( e α ( 2 ) ) ) α (f(e_\alpha^{(2)}))_\alpha ( f ( e α ( 2 ) ) ) α , the upper ℓ r 1 \ell_{r_1} ℓ r 1 estimate in W 1 W_1 W 1 , and the inclusion ℓ p 2 ⊂ ℓ r 1 \ell_{p_2}\subset\ell_{r_1} ℓ p 2 ⊂ ℓ r 1 .
In the second stage we choose
2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1 2>s>p_1>r_1>p_2>r_2>\cdots>1 2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1
and describe the space
E = ( ⨁ n = 1 ∞ E p n , r n ) 2 , s . E=\left(\bigoplus_{n=1}^{\infty}E_{p_n,r_n}\right)_{2,s}. E = ( n = 1 ⨁ ∞ E p n , r n ) 2 , s .
Its dual is generated in norm by the biorthogonal functionals, the vertical special functionals belonging to the component spaces, and an additional family of horizontal special functionals. For n ≥ 1 n\geq1 n ≥ 1 , the preceding construction gives contractions B n + 1 , n : E n + 1 ∗ → W n B_{n+1,n}:E_{n+1}^*\to W_n B n + 1 , n : E n + 1 ∗ → W n because p n + 1 < r n p_{n+1}<r_n p n + 1 < r n . The inequality p 1 < s p_1<s p 1 < s similarly gives a contraction B 1 , 0 : E 1 ∗ → W B_{1,0}:E_1^*\to W B 1 , 0 : E 1 ∗ → W whose images of the biorthogonal functionals generate the horizontal special cosets. Writing ρ n : E ∗ → E n ∗ \rho_n:E^*\to E_n^* ρ n : E ∗ → E n ∗ for restriction and j n : W n → W j_n:W_n\to W j n : W n → W for the natural isometric embedding, the global operator is
B f = τ 0 B 1 , 0 ( ρ 1 f ) + ∑ n = 1 ∞ τ n j n B n + 1 , n ( ρ n + 1 f ) , τ n = 2 − n − 5 . Bf=\tau_0B_{1,0}(\rho_1f)+\sum_{n=1}^{\infty}\tau_nj_nB_{n+1,n}(\rho_{n+1}f),\qquad\tau_n=2^{-n-5}. B f = τ 0 B 1 , 0 ( ρ 1 f ) + n = 1 ∑ ∞ τ n j n B n + 1 , n ( ρ n + 1 f ) , τ n = 2 − n − 5 .
This series converges in operator norm and gives ∥ B ∥ ≤ 1 / 16 \lVert B\rVert\leq1/16 ∥ B ∥ ≤ 1/16 . Every vertical and horizontal special coset occurs, up to a nonzero scalar, as the image of a biorthogonal functional. Consequently B ( X 0 ) = W B(X_0)=W B ( X 0 ) = W . Proposition 3.6 also proves the required convergence on special functionals, and hence completes the definition of X = ker ( Q − B ) X=\ker(Q-B) X = ker ( Q − B ) .
Finally, § 5 defines the norming sets G p , r G_{p,r} G p , r and G G G . An essential ingredient is the σ ρ \sigma_\rho σ ρ coding of [6 Sections 2.3–2.4] , based on the ρ \rho ρ function of [19 ] . In the separable setting an injective coding of finite rational data is available. On ω 1 \omega_1 ω 1 , this is replaced by coding the data after transporting suitable finite ρ \rho ρ -closures to finite subsets of N \mathbb{N} N . Although σ ρ \sigma_\rho σ ρ is not injective, coinciding codes give agreement on the common initial part and ρ \rho ρ -separation outside it. We explain how this yields the eventual disjointness of the weight sets of two distinct infinite special sequences. This property controls the interaction of special functionals and is central to the definitions of the norming sets.
We prove directly several consequences of these definitions, including bounded completeness and the estimates used in constructing B B B . The remaining structural results, notably the assertions about nonseparable duals of block subspaces and norm generation of the duals by the biorthogonal and special functionals, are stated explicitly. Their lengthy proofs use techniques from the theory of hereditarily indecomposable Banach spaces and are outside the scope of this paper. They may be presented separately if necessary.
Throughout the note the scalar field is R \mathbb{R} R . All subspaces described as Banach spaces are norm closed, and, unless another topology is explicitly specified, separability of a subset of a dual space refers to the norm topology. The assertion that a space has no separable quotient means that it has no infinite-dimensional separable Banach quotient.
The structure of X X X and two questions
The space X X X has a peculiar structure. Its definition as ker ( Q − B ) \operatorname{ker}(Q-B) ker ( Q − B ) gives an implicit description of its vectors, and even nonzero elements are obtained indirectly through the lifting argument. Its subspaces and possible decompositions are therefore not transparent. On the other hand, its dual is canonically isomorphic to the WCG space E E E , with its explicitly defined transfinite basis and component projections π J \pi_J π J . This contrast recalls the Argyros–Haydon space [5 ] , which is hereditarily indecomposable and solves the scalar-plus-compact problem, while its dual is isomorphic to ℓ 1 \ell_1 ℓ 1 .
Question 1. Is the space X X X indecomposable?
Here indecomposability means that X X X cannot be written as a topological direct sum of two infinite-dimensional closed subspaces. The space X X X has no infinite-dimensional separable complemented subspace, since a bounded projection onto such a subspace would give a separable infinite-dimensional quotient. Thus any decomposition into two infinite-dimensional summands would necessarily involve two nonseparable spaces. It is unclear whether such a decomposition is possible. To obtain projections on X X X from those on E E E , one needs the latter to be weak-star continuous for the duality induced by X X X .
Question 2. Does there exist, under CH, a nonseparable closed subspace of ℓ ∞ ( N ) \ell_\infty(\mathbb{N}) ℓ ∞ ( N ) with no separable infinite-dimensional quotient?
A sufficient condition for an affirmative answer can be formulated in terms of E E E . In what follows, the weak-star topology on E E E is σ ( E , X ) \sigma(E,X) σ ( E , X ) , induced by the canonical isomorphism R : E → X ∗ R:E\to X^* R : E → X ∗ ; it is the topology coming from the predual X X X . Suppose that there is an infinite-dimensional weak-star closed subspace Y ⊂ E Y\subset E Y ⊂ E whose closed unit ball B Y B_Y B Y , with the norm inherited from E E E , is weak-star separable. Set
M = Y ⊥ = { x ∈ X : x ( y ) = 0 for every y ∈ Y } , Z = X / M . M=Y_\perp=\{x\in X:x(y)=0\text{ for every }y\in Y\},\qquad Z=X/M. M = Y ⊥ = { x ∈ X : x ( y ) = 0 for every y ∈ Y } , Z = X / M .
Weak-star closedness gives M ⊥ = R ( Y ) ⊂ X ∗ M^\perp=R(Y)\subset X^* M ⊥ = R ( Y ) ⊂ X ∗ , so that Z ∗ Z^* Z ∗ is canonically identified with R ( Y ) R(Y) R ( Y ) . In particular, Z Z Z is infinite dimensional. Choose a countable weak-star dense family ( y n ) n ∈ N ⊂ B Y (y_n)_{n\in\mathbb{N}}\subset B_Y ( y n ) n ∈ N ⊂ B Y and define
T : Z ⟶ ℓ ∞ ( N ) , T ( x + M ) = ( x ( y n ) ) n ∈ N . T:Z\longrightarrow\ell_\infty(\mathbb{N}),\qquad T(x+M)=(x(y_n))_{n\in\mathbb{N}}. T : Z ⟶ ℓ ∞ ( N ) , T ( x + M ) = ( x ( y n ) ) n ∈ N .
This map is well defined, and weak-star density gives
∥ T ( x + M ) ∥ ∞ = sup n ∈ N ∣ x ( y n ) ∣ = sup y ∈ B Y ∣ x ( y ) ∣ . \|T(x+M)\|_{\infty}=\sup_{n\in\mathbb{N}}|x(y_n)|=\sup_{y\in B_Y}|x(y)|. ∥ T ( x + M ) ∥ ∞ = n ∈ N sup ∣ x ( y n ) ∣ = y ∈ B Y sup ∣ x ( y ) ∣.
Since R : E → X ∗ R:E\to X^* R : E → X ∗ is an isomorphism, the last expression is equivalent to the quotient norm of x + M x+M x + M . Thus T T T is an isomorphic embedding with closed range. Every quotient of Z Z Z is also a quotient of X X X , so Z Z Z has no separable infinite-dimensional quotient. Consequently Z Z Z is nonseparable and T ( Z ) T(Z) T ( Z ) answers Question 2 affirmatively.
There are infinite-dimensional weak-star closed subspaces Y ⊂ E Y\subset E Y ⊂ E which are weak-star separable: one can take the weak-star closed linear span of a countable linearly independent family in E E E . However, weak-star separability of a dual space does not, in general, imply weak-star separability of its closed unit ball. Indeed, Avilés, Plebanek and Rodríguez [8 ] constructed in ZFC a compact space K K K such that C ( K ) ∗ C(K)^* C ( K ) ∗ is weak-star separable but its closed unit ball is not. The condition on B Y B_Y B Y therefore requires a separate argument.
Acknowledgment
Acknowledgment. This work was prepared in close collaboration with the AI model GPT-6 Astra. Under the author’s systematic supervision, the model developed a substantial part of the statements and proofs presented in this paper. The final manuscript was checked and edited by the author, who takes full responsibility for its content. The author also thanks W. B. Johnson for his comments and remarks.
The goal
For a Banach space Z Z Z and a subspace Y ⊂ Z ∗ Y\subset Z^* Y ⊂ Z ∗ , define the restriction, or evaluation, operator
r Y : Z ⟶ Y ∗ , ( r Y z ) ( y ) = y ( z ) . r_Y:Z\longrightarrow Y^*,\qquad(r_Yz)(y)=y(z). r Y : Z ⟶ Y ∗ , ( r Y z ) ( y ) = y ( z ) .
The assertion that the restriction of Z Z Z to Y Y Y is nonseparable means that the set r Y ( Z ) r_Y(Z) r Y ( Z ) is nonseparable in the norm of Y ∗ Y^* Y ∗ .
Proposition 1.1. Let Z Z Z be an infinite-dimensional Banach space. Suppose that r Y ( Z ) r_Y(Z) r Y ( Z ) is nonseparable for every infinite-dimensional separable closed subspace Y ⊂ Z ∗ Y\subset Z^* Y ⊂ Z ∗ . Then Z Z Z has no infinite-dimensional separable quotient.
Proof. Suppose, to the contrary, that T : Z → S T:Z\to S T : Z → S is a bounded linear surjection onto an infinite-dimensional separable Banach space. Choose an infinite-dimensional separable closed subspace H ⊂ S ∗ H\subset S^* H ⊂ S ∗ . The adjoint T ∗ : S ∗ → Z ∗ T^*:S^*\to Z^* T ∗ : S ∗ → Z ∗ is an isomorphic embedding, so Y = T ∗ ( H ) Y=T^*(H) Y = T ∗ ( H ) is an infinite-dimensional separable closed subspace of Z ∗ Z^* Z ∗ .
Define A : S → Y ∗ A:S\to Y^* A : S → Y ∗ by
( A s ) ( T ∗ h ) = h ( s ) ( s ∈ S , h ∈ H ) . (As)(T^*h)=h(s)\qquad(s\in S,\ h\in H). ( A s ) ( T ∗ h ) = h ( s ) ( s ∈ S , h ∈ H ) .
This is well defined and bounded because T ∗ ∣ H T^*|_H T ∗ ∣ H is bounded below. For z ∈ Z z\in Z z ∈ Z and h ∈ H h\in H h ∈ H ,
( r Y z ) ( T ∗ h ) = ( T ∗ h ) ( z ) = h ( T z ) = ( A T z ) ( T ∗ h ) . (r_Yz)(T^*h)=(T^*h)(z)=h(Tz)=(ATz)(T^*h). ( r Y z ) ( T ∗ h ) = ( T ∗ h ) ( z ) = h ( T z ) = ( A T z ) ( T ∗ h ) .
Thus r Y = A T r_Y=AT r Y = A T and r Y ( Z ) = A ( S ) r_Y(Z)=A(S) r Y ( Z ) = A ( S ) is separable, a contradiction. □ \square □
Theorem 1.2. Assume that a Banach space E E E has a transfinite Schauder basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 and satisfies the following conditions.
X 0 = span ‾ { e α ∗ : α < ω 1 } ⊂ E ∗ , X_0=\overline{\operatorname{span}}\{e_\alpha^*: \alpha<\omega_1\}\subset E^*, X 0 = span { e α ∗ : α < ω 1 } ⊂ E ∗ ,
the canonical evaluation map identifies E E E with X 0 ∗ X_0^* X 0 ∗ .
X + X 0 ‾ ∥ ⋅ ∥ = E ∗ \overline{X+X_0}^{\|\cdot\|}=E^* X + X 0 ∥ ⋅ ∥ = E ∗
and the canonical restriction map
R : E ⟶ X ∗ , ( R e ) ( x ) = x ( e ) , R:E\longrightarrow X^*,\qquad(Re)(x)=x(e), R : E ⟶ X ∗ , ( R e ) ( x ) = x ( e ) ,
is an onto isomorphism.
Then X X X has no infinite-dimensional separable quotient.
Proof. Let Y ⊂ X ∗ Y\subset X^* Y ⊂ X ∗ be an infinite-dimensional separable closed subspace and set F = R − 1 ( Y ) F=R^{-1}(Y) F = R − 1 ( Y ) . Since R R R is an isomorphism, F F F is an infinite-dimensional separable closed subspace of E E E . Put
D F = { x ∣ F : x ∈ X } , C F = { y ∣ F : y ∈ X 0 } ‾ ⊂ F ∗ . D_F=\{x|_F:x\in X\},\qquad C_F=\overline{\{y|_F:y\in X_0\}}\subset F^*. D F = { x ∣ F : x ∈ X } , C F = { y ∣ F : y ∈ X 0 } ⊂ F ∗ .
We first show that C F C_F C F is separable. Every vector in E E E is a norm limit of finitely supported vectors, so its support is countable. Taking the union of the supports of a countable dense subset of F F F gives a countable set S ⊂ ω 1 S\subset\omega_1 S ⊂ ω 1 . By continuity of the biorthogonal functionals, every vector in F F F is supported on S S S . Thus e α ∗ ∣ F = 0 e_\alpha^*|_F=0 e α ∗ ∣ F = 0 for α ∉ S \alpha\notin S α ∈ / S , and
C F = span ‾ { e α ∗ ∣ F : α ∈ S } C_F=\overline{\operatorname{span}}\{e_\alpha^*|_F:\alpha\in S\} C F = span { e α ∗ ∣ F : α ∈ S }
is separable.
Suppose that D F D_F D F were also separable. Then D F + C F ‾ \overline{D_F+C_F} D F + C F would be a separable closed subspace of F ∗ F^* F ∗ . The density of X + X 0 X+X_0 X + X 0 in E ∗ E^* E ∗ implies that every restriction f ∣ F f|_F f ∣ F , with f ∈ E ∗ f\in E^* f ∈ E ∗ , belongs to D F + C F ‾ \overline{D_F+C_F} D F + C F . By Hahn–Banach every element of F ∗ F^* F ∗ has such an extension, whence
F ∗ = D F + C F ‾ . F^*=\overline{D_F+C_F}. F ∗ = D F + C F .
This contradicts condition (ii), so D F D_F D F is nonseparable.
Composition with R ∣ F : F → Y R|_F:F\to Y R ∣ F : F → Y defines an isomorphism
U : Y ∗ ⟶ F ∗ , U ψ = ψ ∘ ( R ∣ F ) . U:Y^*\longrightarrow F^*,\qquad U\psi=\psi\circ(R|_F). U : Y ∗ ⟶ F ∗ , U ψ = ψ ∘ ( R ∣ F ) .
For x ∈ X x\in X x ∈ X and e ∈ F e\in F e ∈ F ,
( U ( r Y x ) ) ( e ) = ( r Y x ) ( R e ) = ( R e ) ( x ) = x ( e ) . (U(r_Yx))(e)=(r_Yx)(Re)=(Re)(x)=x(e). ( U ( r Y x )) ( e ) = ( r Y x ) ( R e ) = ( R e ) ( x ) = x ( e ) .
Consequently U ( r Y ( X ) ) = D F U(r_Y(X))=D_F U ( r Y ( X )) = D F , and r Y ( X ) r_Y(X) r Y ( X ) is nonseparable. Proposition 1.1 gives the conclusion. □ \square □
Remark 1.3. The isomorphism in condition (iii) is specified to be the canonical restriction map: this is the compatibility used in the last display of the proof. An abstract isomorphism X ∗ ≅ E X^*\cong E X ∗ ≅ E alone does not specify that compatibility. For X = ker ( Q − B ) X=\ker(Q-B) X = ker ( Q − B ) the required canonical isomorphism is proved in Proposition 2.5 . The argument of Theorem 1.2 itself uses only the basis, the nonseparability clause of (ii), and (iii); the remaining hypotheses belong to the construction.
Remark 1.4. As W. B. Johnson pointed out, the existence of spaces satisfying all the assumptions of Theorem 1.2 implies CH. Here we use density character, denoted by dens, rather than algebraic dimension. Indeed, choose an infinite-dimensional separable closed subspace Y Y Y of E E E . Its dual is nonseparable, so Stegall’s theorem [18 Theorem 1] gives dens Y ∗ ≥ c \operatorname{dens} Y^* \ge\mathfrak{c} dens Y ∗ ≥ c . The Hahn–Banach theorem makes the restriction map E ∗ → Y ∗ E^* \to Y^* E ∗ → Y ∗ onto, and therefore dens E ∗ ≥ c \operatorname{dens} E^* \ge\mathfrak{c} dens E ∗ ≥ c .
The transfinite basis gives dens E = dens X 0 = ℵ 1 \operatorname{dens} E = \operatorname{dens} X_0 = \aleph_1 dens E = dens X 0 = ℵ 1 . Since X ∗ ≅ E X^* \cong E X ∗ ≅ E , the standard inequality dens X ≤ dens X ∗ \operatorname{dens} X \le\operatorname{dens} X^* dens X ≤ dens X ∗ yields dens X ≤ ℵ 1 \operatorname{dens} X \le\aleph_1 dens X ≤ ℵ 1 . Using X + X 0 = E ∗ X + X_0 = E^* X + X 0 = E ∗ , we obtain
c ≤ dens E ∗ ≤ max { dens X , dens X 0 } ≤ ℵ 1 . \mathfrak{c} \le\operatorname{dens} E^* \le\max\{\operatorname{dens} X,\operatorname{dens} X_0\} \le\aleph_1. c ≤ dens E ∗ ≤ max { dens X , dens X 0 } ≤ ℵ 1 .
Thus c = ℵ 1 \mathfrak{c} = \aleph_1 c = ℵ 1 . The auxiliary space E E E is constructed in ZFC; CH is used to define the subspace X X X with the properties required in (iii).
From some specific properties of E to the space X
A list of properties of E
Let G ⊂ ℓ ∞ ( ω 1 ) G \subset\ell_\infty(\omega_1) G ⊂ ℓ ∞ ( ω 1 ) be a norming set such that
∥ u ∥ G = sup g ∈ G ∣ ∑ α < ω 1 u ( α ) g ( α ) ∣ ( u ∈ c 00 ( ω 1 ) ) \lVert u\rVert_G = \sup_{g\in G}\left|\sum_{\alpha<\omega_1}u(\alpha)g(\alpha)\right| \qquad(u\in c_{00}(\omega_1)) ∥ u ∥ G = g ∈ G sup α < ω 1 ∑ u ( α ) g ( α ) ( u ∈ c 00 ( ω 1 ))
is a finite-valued norm. Let E E E be the completion of ( c 00 ( ω 1 ) , ∥ ⋅ ∥ G ) (c_{00}(\omega_1),\lVert\cdot\rVert_G) ( c 00 ( ω 1 ) , ∥ ⋅ ∥ G ) , and write e α e_\alpha e α for its basis vectors. The space E E E has the following properties.
(1) The family ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 is a transfinite Schauder basis of E E E . Its biorthogonal functionals are denoted by e α ∗ e_\alpha^* e α ∗ . Setting
X 0 = span ‾ { e α ∗ : α < ω 1 } ⊂ E ∗ , X_0=\overline{\operatorname{span}}\{e_\alpha^*:\alpha<\omega_1\}\subset E^*, X 0 = span { e α ∗ : α < ω 1 } ⊂ E ∗ ,
the canonical evaluation map
R 0 : E ⟶ X 0 ∗ , ( R 0 e ) ( y ) = y ( e ) , R_0:E\longrightarrow X_0^*,\qquad(R_0e)(y)=y(e), R 0 : E ⟶ X 0 ∗ , ( R 0 e ) ( y ) = y ( e ) ,
is an onto isometry. Thus E = X ∗ E=X^* E = X ∗ through this identification.
(2) The space E E E is weakly compactly generated, contains no isomorphic copy of ℓ 1 \ell_1 ℓ 1 , and every infinite-dimensional closed subspace F ⊂ E F\subset E F ⊂ E has nonseparable dual F ∗ F^* F ∗ .
(3) There is a family
Φ = { ϕ γ : γ ∈ Γ } ⊂ B E ∗ \Phi=\{\phi_\gamma:\gamma\in\Gamma\}\subset B_{E^*} Φ = { ϕ γ : γ ∈ Γ } ⊂ B E ∗
with the following properties.
(i) For each γ ∈ Γ \gamma\in\Gamma γ ∈ Γ ,
ϕ γ = w ∗ - ∑ n = 1 ∞ ϕ n , γ , \phi_\gamma=w^*\text{-}\sum_{n=1}^{\infty}\phi_{n,\gamma}, ϕ γ = w ∗ - n = 1 ∑ ∞ ϕ n , γ ,
where ( ϕ n , γ ) n (\phi_{n,\gamma})_n ( ϕ n , γ ) n is a successive block sequence of finitely supported functionals. The equality means convergence of the partial sums in σ ( E ∗ , E ) \sigma(E^*,E) σ ( E ∗ , E ) .
(ii) If
Q : E ∗ ⟶ E ∗ / X 0 = W Q:E^*\longrightarrow E^*/X_0=W Q : E ∗ ⟶ E ∗ / X 0 = W
is the quotient operator then
Q ( ϕ γ ) ≠ 0 for all γ ∈ Γ Q(\phi_\gamma)\ne0\qquad\text{for all }\gamma\in\Gamma Q ( ϕ γ ) = 0 for all γ ∈ Γ
(4) The dual space is generated in norm by the biorthogonal functionals and the family Φ \Phi Φ :
E ∗ = span ‾ ( { e α ∗ : α < ω 1 } ∪ { ϕ γ : γ ∈ Γ } ) . E^*=\overline{\operatorname{span}}\left(\{e_\alpha^*:\alpha<\omega_1\}\cup\{\phi_\gamma:\gamma\in\Gamma\}\right). E ∗ = span ( { e α ∗ : α < ω 1 } ∪ { ϕ γ : γ ∈ Γ } ) .
(5) Under the continuum hypothesis there exist a bounded linear operator
B : E ∗ ⟶ W , ∥ B ∥ < 1 10 , B:E^*\longrightarrow W,\qquad\lVert B\rVert<\frac{1}{10}, B : E ∗ ⟶ W , ∥ B ∥ < 10 1 ,
a bijection σ : ω 1 ⟶ Γ \sigma:\omega_1\longrightarrow\Gamma σ : ω 1 ⟶ Γ , and scalars 0 < λ γ < 1 0<\lambda_\gamma<1 0 < λ γ < 1 such that
B ( e α ∗ ) = λ σ ( α ) Q ( ϕ σ ( α ) ) ( α < ω 1 ) . B(e_\alpha^*)=\lambda_{\sigma(\alpha)}Q(\phi_{\sigma(\alpha)})\qquad(\alpha<\omega_1). B ( e α ∗ ) = λ σ ( α ) Q ( ϕ σ ( α ) ) ( α < ω 1 ) .
In addition, for every γ ∈ Γ \gamma\in\Gamma γ ∈ Γ ,
B ( ∑ n = 1 N ϕ n , γ ) = ∑ n = 1 N B ( ϕ n , γ ) → N → ∞ ∥ ⋅ ∥ B ( ϕ γ ) . (1) B\left(\sum_{n=1}^{N}\phi_{n,\gamma}\right)=\sum_{n=1}^{N}B(\phi_{n,\gamma})\xrightarrow[N\to\infty]{\lVert\cdot\rVert}B(\phi_\gamma). \tag*{(1)} B ( n = 1 ∑ N ϕ n , γ ) = n = 1 ∑ N B ( ϕ n , γ ) ∥ ⋅ ∥ N → ∞ B ( ϕ γ ) . (1)
Thus the indices of the biorthogonal functionals and the path cosets are paired bijectively. Only density of the resulting range, not surjectivity of B ∣ X 0 B|_{X_0} B ∣ X 0 , will be used below.
Property 5 is established in Proposition 3.6 from the description of E E E in Section 3.2.
Remark 2.1. Properties 4 and 5 imply
W = span ‾ { Q ϕ γ : γ ∈ Γ } = B ( X 0 ) ‾ . W=\overline{\operatorname{span}}\{Q\phi_\gamma:\gamma\in\Gamma\}=\overline{B(X_0)}. W = span { Q ϕ γ : γ ∈ Γ } = B ( X 0 ) .
Indeed, the biorthogonal functionals vanish under Q Q Q , and every Q ϕ γ Q\phi_\gamma Q ϕ γ is a nonzero scalar multiple of an image of a biorthogonal functional under B B B . All closures in this formula are norm closures.
The duality conditions
We use the following notation for the two duality conditions in the predual argument. Condition (D1) is included in property 1; condition (D2) follows from the stated properties, as proved below.
(D1) The canonical evaluation map
R 0 : E ⟶ X 0 ∗ , ( R 0 e ) ( y ) = y ( e ) , R_0:E\longrightarrow X_0^*,\qquad(R_0e)(y)=y(e), R 0 : E ⟶ X 0 ∗ , ( R 0 e ) ( y ) = y ( e ) ,
is an onto isometry. In particular, E = X 0 ∗ E=X_0^* E = X 0 ∗ through this canonical identification.
(D2) Writing J E : E ⟶ E ∗ ∗ J_E:E\longrightarrow E^{**} J E : E ⟶ E ∗∗ for the canonical embedding, the adjoint of B B B satisfies
B ∗ ( W ∗ ) ⊂ J E ( E ) . B^*(W^*)\subset J_E(E). B ∗ ( W ∗ ) ⊂ J E ( E ) .
Proposition 2.2. Properties 1, 3(i), 4, and 5 imply (D2):
B ∗ ( W ∗ ) ⊂ J E ( E ) . B^*(W^*)\subset J_E(E). B ∗ ( W ∗ ) ⊂ J E ( E ) .
Proof. Fix w ∗ ∈ W ∗ w^*\in W^* w ∗ ∈ W ∗ . The functional
y ⟼ w ∗ ( B y ) ( y ∈ X 0 ) y\longmapsto w^*(By)\qquad(y\in X_0) y ⟼ w ∗ ( B y ) ( y ∈ X 0 )
belongs to X 0 ∗ X_0^* X 0 ∗ . By (D1) there is a unique e ∈ E e\in E e ∈ E such that
w ∗ ( B y ) = y ( e ) ( y ∈ X 0 ) . (2) w^*(By)=y(e)\qquad(y\in X_0). \tag*{(2)} w ∗ ( B y ) = y ( e ) ( y ∈ X 0 ) . (2)
Moreover, ∥ e ∥ ≤ ∥ B ∥ ∥ w ∗ ∥ \lVert e\rVert\leq\lVert B\rVert\lVert w^*\rVert ∥ e ∥ ≤ ∥ B ∥ ∥ w ∗ ∥ . For γ ∈ Γ \gamma\in\Gamma γ ∈ Γ put s N , γ = ∑ n = 1 N ϕ n , γ s_{N,\gamma}=\sum_{n=1}^{N}\phi_{n,\gamma} s N , γ = ∑ n = 1 N ϕ n , γ . Each s N , γ s_{N,\gamma} s N , γ belongs to X 0 X_0 X 0 , since the block functionals are finitely supported. Property 3(i) gives s N , γ → ϕ γ s_{N,\gamma}\to\phi_\gamma s N , γ → ϕ γ in σ ( E ∗ , E ) \sigma(E^*,E) σ ( E ∗ , E ) , whereas (1) gives B s N , γ → B ϕ γ Bs_{N,\gamma}\to B\phi_\gamma B s N , γ → B ϕ γ in the norm of W W W . Therefore
( B ∗ w ∗ ) ( ϕ γ ) = w ∗ ( B ϕ γ ) = lim N → ∞ w ∗ ( B s N , γ ) = lim N → ∞ s N , γ ( e ) = ϕ γ ( e ) = ( J E e ) ( ϕ γ ) . (B^*w^*)(\phi_\gamma)=w^*(B\phi_\gamma)=\lim_{N\to\infty}w^*(Bs_{N,\gamma})
=\lim_{N\to\infty}s_{N,\gamma}(e)=\phi_\gamma(e)=(J_Ee)(\phi_\gamma). ( B ∗ w ∗ ) ( ϕ γ ) = w ∗ ( B ϕ γ ) = N → ∞ lim w ∗ ( B s N , γ ) = N → ∞ lim s N , γ ( e ) = ϕ γ ( e ) = ( J E e ) ( ϕ γ ) .
By (2), the continuous functionals B ∗ w ∗ B^*w^* B ∗ w ∗ and J E e J_Ee J E e also agree on X 0 X_0 X 0 . Property 4 says that X 0 X_0 X 0 and the ϕ γ \phi_\gamma ϕ γ together have norm-dense linear span in E ∗ E^* E ∗ . Hence they agree on all of E ∗ E^* E ∗ , and B ∗ w ∗ = J E e B^*w^*=J_Ee B ∗ w ∗ = J E e . Since w ∗ w^* w ∗ was arbitrary, this proves (D2). □
Remark 2.3. Thus the convergence property (1) is sufficient for (D2). The same proof would work with weak convergence in W W W in place of norm convergence. Neither the small norm of B B B nor property 3(ii) is needed for this particular implication.
We do not assume that B ∗ B^* B ∗ is an isomorphism. The operator to be identified as an isomorphism is the restriction map E → X ∗ E\to X^* E → X ∗ . An arbitrary element of E ∗ E^* E ∗ does not have a restriction to X ∗ X^* X ∗ without first specifying an identification of X ∗ X^* X ∗ with E E E .
The subspace X and its dual
Define
L = Q − B : E ∗ ⟶ W , X = ker L . L=Q-B:E^*\longrightarrow W,\qquad X=\ker L. L = Q − B : E ∗ ⟶ W , X = ker L .
The subspace X X X is closed in E ∗ E^* E ∗ and hence is a Banach space with the inherited norm.
Lemma 2.4. Every w ∈ W w\in W w ∈ W has an L L L -lift u ∈ E ∗ u\in E^* u ∈ E ∗ satisfying
L u = w , ∥ u ∥ ≤ 5 2 ∥ w ∥ . Lu=w,\qquad\lVert u\rVert\leq\frac{5}{2}\lVert w\rVert. Lu = w , ∥ u ∥ ≤ 2 5 ∥ w ∥ . Proof. Choose u 0 ∈ E ∗ u_0\in E^* u 0 ∈ E ∗ with Q u 0 = w Qu_0=w Q u 0 = w and ∥ u 0 ∥ ≤ 2 ∥ w ∥ \lVert u_0\rVert\leq2\lVert w\rVert ∥ u 0 ∥ ≤ 2 ∥ w ∥ . Recursively choose u n + 1 u_{n+1} u n + 1 such that
Q u n + 1 = B u n , ∥ u n + 1 ∥ ≤ 2 ∥ B u n ∥ ≤ 1 5 ∥ u n ∥ . Qu_{n+1}=Bu_n,\qquad\lVert u_{n+1}\rVert\leq2\lVert Bu_n\rVert\leq\frac{1}{5}\lVert u_n\rVert. Q u n + 1 = B u n , ∥ u n + 1 ∥ ≤ 2 ∥ B u n ∥ ≤ 5 1 ∥ u n ∥ . The series u = ∑ n = 0 ∞ u n u=\sum_{n=0}^{\infty}u_n u = ∑ n = 0 ∞ u n converges. Its sum satisfies Q u = w + B u Qu=w+Bu Q u = w + B u , and
∥ u ∥ ≤ 2 1 − 1 / 5 ∥ w ∥ = 5 2 ∥ w ∥ . \lVert u\rVert\leq\frac{2}{1-1/5}\lVert w\rVert=\frac{5}{2}\lVert w\rVert. ∥ u ∥ ≤ 1 − 1/5 2 ∥ w ∥ = 2 5 ∥ w ∥ .
Proposition 2.5. Assume D 1 D1 D 1 . The restriction map
R : E ⟶ X ∗ , ( R e ) ( x ) = x ( e ) , R:E\longrightarrow X^*,\qquad(Re)(x)=x(e), R : E ⟶ X ∗ , ( R e ) ( x ) = x ( e ) , is an isomorphic embedding, with
3 5 ∥ e ∥ ≤ ∥ R e ∥ ≤ ∥ e ∥ ( e ∈ E ) . \frac{3}{5}\lVert e\rVert\leq\lVert Re\rVert\leq\lVert e\rVert\qquad(e\in E). 5 3 ∥ e ∥ ≤ ∥ R e ∥ ≤ ∥ e ∥ ( e ∈ E ) . If D 2 D2 D 2 also holds, then R R R is onto X ∗ X^* X ∗ .
Proof. For y ∈ X 0 y \in X_0 y ∈ X 0 , apply Lemma 2.4 to B y By B y . There is u ∈ E ∗ u \in E^* u ∈ E ∗ such that
L u = B y , ∥ u ∥ ≤ 5 2 ∥ B y ∥ ≤ 1 4 ∥ y ∥ . Lu = By,\qquad\lVert u\rVert\le\frac{5}{2}\lVert By\rVert\le\frac{1}{4}\lVert y\rVert. Lu = B y , ∥ u ∥ ≤ 2 5 ∥ B y ∥ ≤ 4 1 ∥ y ∥ .
Since L y = − B y Ly=-By L y = − B y , the vector x = y + u x=y+u x = y + u belongs to X X X and satisfies
∥ x − y ∥ ≤ 1 4 ∥ y ∥ , ∥ x ∥ ≤ 5 4 ∥ y ∥ . \lVert x-y\rVert\le\frac{1}{4}\lVert y\rVert,\qquad\lVert x\rVert\le\frac{5}{4}\lVert y\rVert. ∥ x − y ∥ ≤ 4 1 ∥ y ∥ , ∥ x ∥ ≤ 4 5 ∥ y ∥ .
Given e ≠ 0 e\ne0 e = 0 and ε > 0 \varepsilon>0 ε > 0 , condition D1 provides y ∈ B X 0 y\in B_{X_0} y ∈ B X 0 with ∣ y ( e ) ∣ > ( 1 − ε ) ∥ e ∥ |y(e)|>(1-\varepsilon)\lVert e\rVert ∣ y ( e ) ∣ > ( 1 − ε ) ∥ e ∥ . For the corresponding x ∈ X x\in X x ∈ X ,
5 4 ∥ R e ∥ ≥ ∣ x ( e ) ∣ > ( 3 4 − ε ) ∥ e ∥ . \frac{5}{4}\lVert Re\rVert\ge|x(e)|>\left(\frac{3}{4}-\varepsilon\right)\lVert e\rVert. 4 5 ∥ R e ∥ ≥ ∣ x ( e ) ∣ > ( 4 3 − ε ) ∥ e ∥ .
Let ε \varepsilon ε decrease to zero. The upper bound ∥ R e ∥ ≤ ∥ e ∥ \lVert Re\rVert\le\lVert e\rVert ∥ R e ∥ ≤ ∥ e ∥ follows from the definition. Thus R R R is injective, has closed range, and has a bounded inverse on its range.
For surjectivity, take an arbitrary f ∈ X ∗ f\in X^* f ∈ X ∗ . By Hahn–Banach it has an extension z ∈ E ∗ ∗ z\in E^{**} z ∈ E ∗∗ . Condition D1 provides e 0 ∈ E e_0\in E e 0 ∈ E with z ∣ X 0 = J E e 0 ∣ X 0 z|_{X_0}=J_Ee_0|_{X_0} z ∣ X 0 = J E e 0 ∣ X 0 . Therefore z − J E e 0 z-J_Ee_0 z − J E e 0 annihilates X 0 X_0 X 0 and factors through Q Q Q : for some w ∗ ∈ W ∗ w^*\in W^* w ∗ ∈ W ∗ ,
z = J E e 0 + Q ∗ w ∗ . z=J_Ee_0+Q^*w^*. z = J E e 0 + Q ∗ w ∗ .
By D2, choose e 1 ∈ E e_1\in E e 1 ∈ E with B ∗ w ∗ = J E e 1 B^*w^*=J_Ee_1 B ∗ w ∗ = J E e 1 . For x ∈ X x\in X x ∈ X we have Q x = B x Qx=Bx Q x = B x , and hence
f ( x ) = z ( x ) = x ( e 0 ) + w ∗ ( Q x ) = x ( e 0 ) + w ∗ ( B x ) = x ( e 0 + e 1 ) . \begin{aligned}
f(x)=z(x)=x(e_0)+w^*(Qx) \\
&=x(e_0)+w^*(Bx)=x(e_0+e_1).
\end{aligned} f ( x ) = z ( x ) = x ( e 0 ) + w ∗ ( Q x ) = x ( e 0 ) + w ∗ ( B x ) = x ( e 0 + e 1 ) .
Thus f = R ( e 0 + e 1 ) f=R(e_0+e_1) f = R ( e 0 + e 1 ) . Since f f f was arbitrary, R ( E ) = X ∗ R(E)=X^* R ( E ) = X ∗ .
Remark 2.6. After R R R has been established as an onto isomorphism, its adjoint R ∗ : X ∗ ∗ → E ∗ R^*:X^{**}\to E^* R ∗ : X ∗∗ → E ∗ is also an onto isomorphism. For x ∈ X x\in X x ∈ X ,
R ∗ ( J X x ) = x ∈ E ∗ , R^*(J_Xx)=x\in E^*, R ∗ ( J X x ) = x ∈ E ∗ ,
because both sides evaluated at e ∈ E e\in E e ∈ E equal ( R e ) ( x ) = x ( e ) (Re)(x)=x(e) ( R e ) ( x ) = x ( e ) . This is the precise compatibility between the two dual identifications.
Density modulo the predual X 0 X_0 X 0
Lemma 2.7. Under properties 4 and 5,
B ( X ) ‾ = W , Q ( X ) ‾ = W , X + X 0 ‾ = E ∗ . \overline{B(X)}=W,\qquad\overline{Q(X)}=W,\qquad\overline{X+X_0}=E^*. B ( X ) = W , Q ( X ) = W , X + X 0 = E ∗ . Proof. Let M = B ( X ) ‾ M=\overline{B(X)} M = B ( X ) , a closed subspace of W W W . Define
A : W ⟶ W / M , A ( L y ) = B y + M ( y ∈ E ∗ ) . A:W\longrightarrow W/M,\qquad A(Ly)=By+M\quad(y\in E^*). A : W ⟶ W / M , A ( L y ) = B y + M ( y ∈ E ∗ ) . This is well defined: two L L L -lifts differ by an element of X X X , and their B B B -images differ by an element of M M M . Lemma 2.4 gives
∥ A ∥ ≤ 5 2 ∥ B ∥ < 1 4 . \lVert A\rVert\le\frac{5}{2}\lVert B\rVert<\frac{1}{4}. ∥ A ∥ ≤ 2 5 ∥ B ∥ < 4 1 . For y ∈ X 0 y\in X_0 y ∈ X 0 , we have L y = − B y Ly=-By L y = − B y , so A ( B y ) = − B y + M A(By)=-By+M A ( B y ) = − B y + M . By Remark 2.1, B ( X 0 ) B(X_0) B ( X 0 ) is dense in W W W . It follows that A A A is the negative of the canonical quotient map W → W / M W\to W/M W → W / M . If M ≠ W M\ne W M = W , that quotient map has norm one, a contradiction. Thus M = W M=W M = W .
:::`
Since Q = B Q = B Q = B on X X X , we also have Q ( X ) ‾ = W \overline{Q(X)} = W Q ( X ) = W . Given f ∈ E ∗ f \in E^* f ∈ E ∗ and η > 0 \eta> 0 η > 0 , choose x ∈ X x \in X x ∈ X with ∥ Q ( f − x ) ∥ < η \|Q(f-x)\| < \eta ∥ Q ( f − x ) ∥ < η . By the quotient norm there is y ∈ X 0 y \in X_0 y ∈ X 0 with ∥ f − x − y ∥ < η \|f-x-y\| < \eta ∥ f − x − y ∥ < η . Hence X + X 0 ‾ = E ∗ \overline{X+X_0} = E^* X + X 0 = E ∗ .
The absence of separable quotients
Theorem 2.8. Assume CH and properties (1)–(5) of § 3 . Then
X = ker ( Q − B ) X = \operatorname{ker}(Q-B) X = ker ( Q − B ) is an infinite-dimensional Banach space with no infinite-dimensional separable quotient.
Proof. Condition (D1) is part of property 1, and Proposition 2.2 establishes (D2). Thus Proposition 2.5 shows that the canonical restriction map R : E → X ∗ R : E \to X^* R : E → X ∗ is an onto isomorphism. In particular, X X X is infinite dimensional. Lemma 2.7 gives
X ‾ + X 0 ‾ = E ∗ . \overline{X}+\overline{X_0}=E^*. X + X 0 = E ∗ .
Together with properties 1 and 2, these facts verify all the hypotheses of Theorem 1.2, which gives the conclusion. □
The space E
The auxiliary space E E E is constructed in ZFC; CH enters when we define the operator B B B and the new predual X X X . The properties needed here include density character ℵ 1 \aleph_1 ℵ 1 , the absence of an isomorphic copy of ℓ 1 \ell_1 ℓ 1 , and the requirement that every infinite-dimensional closed subspace have nonseparable dual. Thus we seek a nonseparable analogue of the Gowers tree space GT [11 ] .
Gowers’s construction combines techniques from the theory of hereditarily indecomposable spaces with phenomena originating in the James tree space [13 ] . We refer to Petsoulas and Raikoftsalis [16 ] for a detailed study of a related Gowers-tree-like space and its operators, and to [1 ] for constructions using the attractors method. The preprint [15 ] presents a nonseparable space with the properties required here. I had expected that a space of this kind could be constructed by combining the separable examples with the methods of my joint work with J. López-Abad and S. Todorčević [6 ] . In particular, some years ago I discussed a possible project in this direction with K. Beanland.
The work of Tsirel’son [20 ] , Figiel and Johnson [10 ] , Schlumprecht [17 ] , and Gowers and Maurey [12 ] provides powerful methods for defining saturated norms. These methods, together with hereditary James-tree constructions, motivate the description below.
We organize the description in two stages. First we consider spaces E p , r E_{p,r} E p , r , with 1 < r < p ≤ ∞ 1 < r < p \le\infty 1 < r < p ≤ ∞ . In the second stage we combine spaces E n = E p n , r n E_n = E_{p_n,r_n} E n = E p n , r n with parameters
2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1 , 2 > s > p_1 > r_1 > p_2 > r_2 > \cdots> 1, 2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1 ,
using the norm denoted by
E = ( ⨁ n = 1 ∞ ⊕ E n ) 2 , s . E = \left(\bigoplus_{n=1}^{\infty} \oplus E_n\right)_{2,s}. E = ( n = 1 ⨁ ∞ ⊕ E n ) 2 , s .
This two-stage presentation differs from the single construction in [15 ] . I have not checked all the proofs concerning E E E in that paper and therefore do not claim to have verified its construction. I do, however, have a complete understanding of how the two spaces are constructed.
The following is a description of the component spaces and a proof of the operator argument that follows from their stated properties. The norming sets defining these norms will be described in a later section.
The spaces E p , r E_{p,r} E p , r for finite p p p
Fix 1 < r < p < ∞ 1<r<p<\infty 1 < r < p < ∞ and put p ′ = p / ( p − 1 ) p'=p/(p-1) p ′ = p / ( p − 1 ) . We consider the completion of c 00 ( ω 1 ) c_{00}(\omega_1) c 00 ( ω 1 ) under a norm ∥ ⋅ ∥ p , r \|\cdot\|_{p,r} ∥ ⋅ ∥ p , r with the properties listed below.
Properties of E p , r E_{p,r} E p , r
(1) The family ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 is a bimonotone, boundedly complete transfinite Schauder basis of E p , r E_{p,r} E p , r .
(2) The set K = { 0 } ∪ { e α : α < ω 1 } K=\{0\}\cup\{e_\alpha:\alpha<\omega_1\} K = { 0 } ∪ { e α : α < ω 1 } is weakly compact, and every e α e_\alpha e α is isolated in its relative weak topology. Consequently E p , r E_{p,r} E p , r is weakly compactly generated (WCG).
(3) Writing ( e α ∗ ) α < ω 1 (e_\alpha^*)_{\alpha<\omega_1} ( e α ∗ ) α < ω 1 for the biorthogonal functionals and setting
X 0 p , r = span { e α ∗ : α < ω 1 } ⊂ E p , r ∗ , X_0^{p,r}=\operatorname{span}\{e_\alpha^*:\alpha<\omega_1\}\subset E_{p,r}^*, X 0 p , r = span { e α ∗ : α < ω 1 } ⊂ E p , r ∗ ,
the canonical map E p , r → ( X 0 p , r ) ∗ E_{p,r}\to(X_0^{p,r})^* E p , r → ( X 0 p , r ) ∗ , given by evaluation, is an onto isometry.
(4) Every infinite-dimensional closed subspace of E p , r E_{p,r} E p , r has nonseparable dual, and E p , r E_{p,r} E p , r contains no isomorphic copy of ℓ 1 \ell_1 ℓ 1 .
(5) The basis vectors satisfy an upper ℓ p ′ \ell_{p'} ℓ p ′ estimate with constant one: for every finitely supported scalar family ( a α ) α < ω 1 (a_\alpha)_{\alpha<\omega_1} ( a α ) α < ω 1 ,
∥ ∑ α < ω 1 a α e α ∥ p , r ≤ ( ∑ α < ω 1 ∣ a α ∣ p ′ ) 1 / p ′ . (3) \left\|\sum_{\alpha<\omega_1}a_\alpha e_\alpha\right\|_{p,r}\leq\left(\sum_{\alpha<\omega_1}|a_\alpha|^{p'}\right)^{1/p'}. \tag*{(3)} α < ω 1 ∑ a α e α p , r ≤ ( α < ω 1 ∑ ∣ a α ∣ p ′ ) 1/ p ′ . (3)
Properties of E p , r ∗ E_{p,r}^* E p , r ∗
(i) For every x ∗ ∈ E p , r ∗ x^*\in E_{p,r}^* x ∗ ∈ E p , r ∗ , the scalar family ( x ∗ ( e α ) ) α < ω 1 (x^*(e_\alpha))_{\alpha<\omega_1} ( x ∗ ( e α ) ) α < ω 1 belongs to ℓ p ( ω 1 ) \ell_p(\omega_1) ℓ p ( ω 1 ) and satisfies
( ∑ α < ω 1 ∣ x ∗ ( e α ) ∣ p ) 1 / p ≤ ∥ x ∗ ∥ . (4) \left(\sum_{\alpha<\omega_1}|x^*(e_\alpha)|^p\right)^{1/p}\leq\|x^*\|. \tag*{(4)} ( α < ω 1 ∑ ∣ x ∗ ( e α ) ∣ p ) 1/ p ≤ ∥ x ∗ ∥. (4)
Indeed, (3) defines a contraction
T : ℓ p ′ ( ω 1 ) ⟶ E p , r , T ( a ) = ∑ α < ω 1 a α e α . T:\ell_{p'}(\omega_1)\longrightarrow E_{p,r},\qquad T(a)=\sum_{\alpha<\omega_1}a_\alpha e_\alpha. T : ℓ p ′ ( ω 1 ) ⟶ E p , r , T ( a ) = α < ω 1 ∑ a α e α .
The series converges in norm, by approximation with finite sums. Its adjoint is given by
T ∗ x ∗ = ( x ∗ ( e α ) ) α < ω 1 ∈ ℓ p ( ω 1 ) , T^*x^*=(x^*(e_\alpha))_{\alpha<\omega_1}\in\ell_p(\omega_1), T ∗ x ∗ = ( x ∗ ( e α ) ) α < ω 1 ∈ ℓ p ( ω 1 ) ,
and ∥ T ∗ ∥ ≤ 1 \|T^*\|\leq1 ∥ T ∗ ∥ ≤ 1 , which proves (4). In particular, the set { α < ω 1 : x ∗ ( e α ) ≠ 0 } \{\alpha<\omega_1:x^*(e_\alpha)\ne0\} { α < ω 1 : x ∗ ( e α ) = 0 } is countable.
This also explains property (2): for every x ∗ ∈ E p , r ∗ x^*\in E_{p,r}^* x ∗ ∈ E p , r ∗ and ε > 0 \varepsilon>0 ε > 0 , only finitely many α \alpha α satisfy ∣ x ∗ ( e α ) ∣ ≥ ε |x^*(e_\alpha)|\geq\varepsilon ∣ x ∗ ( e α ) ∣ ≥ ε . Hence every weak neighborhood of zero contains all but finitely many basis vectors. It follows directly that K K K is weakly compact; the biorthogonal functionals isolate its nonzero points. Since K K K has dense linear span, E p , r E_{p,r} E p , r is WCG.
(ii) There is an uncountable family of distinct functionals
Φ p , r = { φ γ : γ ∈ Γ } ⊂ B E p , r ∗ \Phi_{p,r}=\{\varphi_\gamma:\gamma\in\Gamma\}\subset B_{E_{p,r}^*} Φ p , r = { φ γ : γ ∈ Γ } ⊂ B E p , r ∗
with the following properties. (a) For each γ ∈ Γ \gamma\in\Gamma γ ∈ Γ there is a successive sequence of finitely supported blocks ( ϕ n , γ ) n = 1 ∞ (\phi_{n,\gamma})_{n=1}^{\infty} ( ϕ n , γ ) n = 1 ∞ of the biorthogonal functionals, with ϕ n , γ ∈ B E p , r ∗ \phi_{n,\gamma} \in B_{E_{p,r}^{*}} ϕ n , γ ∈ B E p , r ∗ , such that
ϕ γ = w ∗ ∑ n = 1 ∞ ϕ n , γ . \phi_\gamma= w^* \sum_{n=1}^{\infty} \phi_{n,\gamma}. ϕ γ = w ∗ n = 1 ∑ ∞ ϕ n , γ .
Thus the partial sums converge in σ ( E p , r ∗ , E p , r ) \sigma(E_{p,r}^{*}, E_{p,r}) σ ( E p , r ∗ , E p , r ) .
(b) The family { e α ∗ : α < ω 1 } ∪ Φ p , r \{e_\alpha^{*}:\alpha<\omega_1\}\cup\Phi_{p,r} { e α ∗ : α < ω 1 } ∪ Φ p , r has norm-dense linear span in E p , r ∗ E_{p,r}^{*} E p , r ∗ .
(c) For γ ∈ Γ \gamma\in\Gamma γ ∈ Γ and n ∈ N n\in\mathbb{N} n ∈ N , put
ϕ > n , γ = ϕ γ − ∑ k = 1 n ϕ k , γ = w ∗ ∑ k > n ϕ k , γ . \phi_{>n,\gamma}=\phi_\gamma-\sum_{k=1}^{n}\phi_{k,\gamma}=w^*\sum_{k>n}\phi_{k,\gamma}. ϕ > n , γ = ϕ γ − k = 1 ∑ n ϕ k , γ = w ∗ k > n ∑ ϕ k , γ .
For every finite set F ⊂ Γ F\subset\Gamma F ⊂ Γ there is n 0 ∈ N n_0\in\mathbb{N} n 0 ∈ N such that, for every n > n 0 n>n_0 n > n 0 and every scalar family ( a γ ) γ ∈ F (a_\gamma)_{\gamma\in F} ( a γ ) γ ∈ F ,
∥ ∑ γ ∈ F a γ ϕ > n , γ ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ r ) 1 / r . (5) \left\|\sum_{\gamma\in F}a_\gamma\phi_{>n,\gamma}\right\|\leq\left(\sum_{\gamma\in F}|a_\gamma|^r\right)^{1/r}. \tag*{(5)} γ ∈ F ∑ a γ ϕ > n , γ ≤ γ ∈ F ∑ ∣ a γ ∣ r 1/ r . (5)
(d) Each ϕ γ \phi_\gamma ϕ γ lies outside X 0 p , r X_0^{p,r} X 0 p , r . Equivalently, if
Q p , r : E p , r ∗ ⟶ W p , r : = E p , r ∗ / X 0 p , r Q_{p,r}:E_{p,r}^{*}\longrightarrow W_{p,r}:=E_{p,r}^{*}/X_0^{p,r} Q p , r : E p , r ∗ ⟶ W p , r := E p , r ∗ / X 0 p , r
is the quotient map, then Q p , r ϕ γ ≠ 0 Q_{p,r}\phi_\gamma\ne0 Q p , r ϕ γ = 0 for every γ ∈ Γ \gamma\in\Gamma γ ∈ Γ . This nonvanishing is a separate requirement; it does not follow from (b) and (c) alone.
The tail estimate implies that the quotient family admits an upper ℓ r \ell_r ℓ r estimate with constant one. Indeed, a finite initial sum of each path belongs to X 0 p , r X_0^{p,r} X 0 p , r , so Q p , r ϕ > n , γ = Q p , r ϕ γ Q_{p,r}\phi_{>n,\gamma}=Q_{p,r}\phi_\gamma Q p , r ϕ > n , γ = Q p , r ϕ γ . For a finite F ⊂ Γ F\subset\Gamma F ⊂ Γ , choose n n n as in (c). Then
∥ ∑ γ ∈ F a γ Q p , r ϕ γ ∥ ≤ ∥ ∑ γ ∈ F a γ ϕ > n , γ ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ r ) 1 / r . (6) \left\|\sum_{\gamma\in F}a_\gamma Q_{p,r}\phi_\gamma\right\|\leq\left\|\sum_{\gamma\in F}a_\gamma\phi_{>n,\gamma}\right\|\leq\left(\sum_{\gamma\in F}|a_\gamma|^r\right)^{1/r}. \tag*{(6)} γ ∈ F ∑ a γ Q p , r ϕ γ ≤ γ ∈ F ∑ a γ ϕ > n , γ ≤ γ ∈ F ∑ ∣ a γ ∣ r 1/ r . (6)
Property (b) also gives
W p , r = span ‾ { Q p , r ϕ γ : γ ∈ Γ } . W_{p,r}=\overline{\operatorname{span}}\{Q_{p,r}\phi_\gamma:\gamma\in\Gamma\}. W p , r = span { Q p , r ϕ γ : γ ∈ Γ } .
Consequently the assignment of the unit vector u γ u_\gamma u γ to Q p , r ϕ γ Q_{p,r}\phi_\gamma Q p , r ϕ γ extends to a contraction
S p , r : ℓ r ( Γ ) ⟶ W p , r , S p , r u γ = Q p , r ϕ γ , (7) S_{p,r}:\ell_r(\Gamma)\longrightarrow W_{p,r},\qquad S_{p,r}u_\gamma=Q_{p,r}\phi_\gamma, \tag*{(7)} S p , r : ℓ r ( Γ ) ⟶ W p , r , S p , r u γ = Q p , r ϕ γ , (7)
with norm-dense range.
A step towards the operator B B B
Section 2 uses an operator that sends biorthogonal functionals to scalar multiples of path cosets. On a single component E p , r E_{p,r} E p , r , the estimates above do not by themselves establish boundedness of the corresponding formula: the scalar family ( x ∗ ( e α ) ) α (x^*(e_\alpha))_\alpha ( x ∗ ( e α ) ) α is controlled in ℓ p \ell_p ℓ p , whereas the quotient estimate uses ℓ r \ell_r ℓ r , and r < p r<p r < p . The formal identity from ℓ p \ell_p ℓ p to ℓ r \ell_r ℓ r is not bounded. This is an obstruction to that particular estimate, not a proof that no operator with the prescribed images of e α ∗ e_\alpha^* e α ∗ can exist. The construction resolves the exponent mismatch by mapping between different components.
For the next definition, let
1 < r 2 < p 2 < r 1 < p 1 < ∞ , 1<r_2<p_2<r_1<p_1<\infty, 1 < r 2 < p 2 < r 1 < p 1 < ∞ ,
and abbreviate
E i = E p i , r i , X 0 , i = X 0 p i , r i , W i = E i ∗ / X 0 , i ( i = 1 , 2 ) . E_i = E_{p_i,r_i}, \qquad X_{0,i} = X_0^{p_i,r_i}, \qquad W_i = E_i^*/X_{0,i} \qquad(i = 1,2). E i = E p i , r i , X 0 , i = X 0 p i , r i , W i = E i ∗ / X 0 , i ( i = 1 , 2 ) .
Use superscripts ( i ) (i) ( i ) on the basis vectors, their biorthogonal functionals, and the path families to distinguish the two spaces. Let Q i : E i ∗ → W i Q_i : E_i^* \to W_i Q i : E i ∗ → W i denote the quotient maps.
Under CH the uncountable family Φ p 1 , r 1 \Phi_{p_1,r_1} Φ p 1 , r 1 has cardinality ℵ 1 \aleph_1 ℵ 1 . To see this, note that every x ∗ ∈ E 1 ∗ x^* \in E_1^* x ∗ ∈ E 1 ∗ is determined by the scalar family ( x ∗ ( e α ( 1 ) ) ) α < ω 1 (x^*(e_\alpha^{(1)}))_{\alpha<\omega_1} ( x ∗ ( e α ( 1 ) ) ) α < ω 1 , which has countable support by (4). Under CH there are only ℵ 1 ℵ 0 = ℵ 1 \aleph_1^{\aleph_0} = \aleph_1 ℵ 1 ℵ 0 = ℵ 1 such real-valued families on ω 1 \omega_1 ω 1 . Since the path functionals are distinct and form an uncountable family, ∣ Γ 1 ∣ = ℵ 1 |\Gamma_1| = \aleph_1 ∣ Γ 1 ∣ = ℵ 1 .
Definition 3.1. Assume CH and fix a bijection σ : ω 1 → Γ 1 \sigma: \omega_1 \to\Gamma_1 σ : ω 1 → Γ 1 . For x ∗ ∈ E 2 ∗ x^* \in E_2^* x ∗ ∈ E 2 ∗ set
B 2 , 1 x ∗ = ∑ α < ω 1 x ∗ ( e α ( 2 ) ) Q 1 ϕ σ ( α ) ( 1 ) . (8) B_{2,1}x^* = \sum_{\alpha<\omega_1} x^*(e_\alpha^{(2)})Q_1\phi_{\sigma(\alpha)}^{(1)}. \tag*{(8)} B 2 , 1 x ∗ = α < ω 1 ∑ x ∗ ( e α ( 2 ) ) Q 1 ϕ σ ( α ) ( 1 ) . (8)
In this particular case the sum is understood as the norm limit of the directed net of its finite subsums. The following proposition verifies that it exists and defines a bounded operator B 2 , 1 : E 2 ∗ → W 1 B_{2,1} : E_2^* \to W_1 B 2 , 1 : E 2 ∗ → W 1 .
Proposition 3.2. The operator B 2 , 1 B_{2,1} B 2 , 1 has the following properties.
(1) It is a well-defined bounded linear operator with ∥ B 2 , 1 ∥ ≤ 1 \lVert B_{2,1}\rVert\le1 ∥ B 2 , 1 ∥ ≤ 1 .
(2) For every α < ω 1 \alpha< \omega_1 α < ω 1 ,
B 2 , 1 e α ( 2 ) ∗ = Q 1 ϕ σ ( α ) ( 1 ) , B_{2,1}e_\alpha^{(2)*} = Q_1\phi_{\sigma(\alpha)}^{(1)}, B 2 , 1 e α ( 2 ) ∗ = Q 1 ϕ σ ( α ) ( 1 ) ,
and these images have norm-dense linear span in W 1 W_1 W 1 .
(3) For every γ ∈ Γ 2 \gamma\in\Gamma_2 γ ∈ Γ 2 ,
B 2 , 1 ( ∑ n = 1 N ϕ n , γ ( 2 ) ) = ∑ n = 1 N B 2 , 1 ( ϕ n , γ ( 2 ) ) → N → ∞ ∥ ⋅ ∥ B 2 , 1 ( ϕ γ ( 2 ) ) . (9) B_{2,1}\left(\sum_{n=1}^{N}\phi_{n,\gamma}^{(2)}\right)
=
\sum_{n=1}^{N}B_{2,1}(\phi_{n,\gamma}^{(2)})
\xrightarrow[N\to\infty]{\lVert\cdot\rVert}
B_{2,1}(\phi_\gamma^{(2)}). \tag*{(9)} B 2 , 1 ( n = 1 ∑ N ϕ n , γ ( 2 ) ) = n = 1 ∑ N B 2 , 1 ( ϕ n , γ ( 2 ) ) ∥ ⋅ ∥ N → ∞ B 2 , 1 ( ϕ γ ( 2 ) ) . (9)
Proof. For x ∗ ∈ E 2 ∗ x^* \in E_2^* x ∗ ∈ E 2 ∗ put a α = x ∗ ( e α ( 2 ) ) a_\alpha= x^*(e_\alpha^{(2)}) a α = x ∗ ( e α ( 2 ) ) . By (4), a = ( a α ) α < ω 1 a = (a_\alpha)_{\alpha<\omega_1} a = ( a α ) α < ω 1 belongs to ℓ p 2 ( ω 1 ) \ell_{p_2}(\omega_1) ℓ p 2 ( ω 1 ) and ∥ a ∥ ℓ p 2 ≤ ∥ x ∗ ∥ \lVert a\rVert_{\ell_{p_2}} \le\lVert x^*\rVert ∥ a ∥ ℓ p 2 ≤ ∥ x ∗ ∥ . Since p 2 < r 1 p_2 < r_1 p 2 < r 1 ,
∥ a ∥ ℓ r 1 ≤ ∥ a ∥ ℓ p 2 ≤ ∥ x ∗ ∥ . (10) \lVert a\rVert_{\ell_{r_1}} \le\lVert a\rVert_{\ell_{p_2}} \le\lVert x^*\rVert. \tag*{(10)} ∥ a ∥ ℓ r 1 ≤ ∥ a ∥ ℓ p 2 ≤ ∥ x ∗ ∥ . (10)
Relabel a a a by σ \sigma σ to obtain a ~ = ( a σ − 1 ( η ) ) η ∈ Γ 1 \widetilde{a} = (a_{\sigma^{-1}(\eta)})_{\eta\in\Gamma_1} a = ( a σ − 1 ( η ) ) η ∈ Γ 1 . The contraction S p 1 , r 1 S_{p_1,r_1} S p 1 , r 1 in (7) gives
B 2 , 1 x ∗ = S p 1 , r 1 a ~ . B_{2,1}x^* = S_{p_1,r_1}\widetilde{a}. B 2 , 1 x ∗ = S p 1 , r 1 a .
Finitely supported scalar families are dense in ℓ r 1 ( Γ 1 ) \ell_{r_1}(\Gamma_1) ℓ r 1 ( Γ 1 ) , so this equality proves that the series in (8) converges in norm, independently of the order of summation. It also proves linearity and the useful estimate
∥ B 2 , 1 x ∗ ∥ ≤ ( ∑ α < ω 1 ∣ x ∗ ( e α ( 2 ) ) ∣ p 2 ) 1 / p 2 ≤ ∥ x ∗ ∥ . (11) \lVert B_{2,1}x^*\rVert\le\left(\sum_{\alpha<\omega_1}|x^*(e_\alpha^{(2)})|^{p_2}\right)^{1/p_2} \le\lVert x^*\rVert. \tag*{(11)} ∥ B 2 , 1 x ∗ ∥ ≤ ( α < ω 1 ∑ ∣ x ∗ ( e α ( 2 ) ) ∣ p 2 ) 1/ p 2 ≤ ∥ x ∗ ∥ . (11)
This proves (1). Biorthogonality gives the identity in (2), and density follows because σ \sigma σ is onto and the path cosets span a dense subspace of W 1 W_1 W 1 .
To prove (3), fix γ ∈ Γ 2 \gamma\in\Gamma_2 γ ∈ Γ 2 and set
s N = ∑ n = 1 N ϕ n , γ ( 2 ) , A N = ⋃ n = 1 N supp ϕ n , γ ( 2 ) , A = ⋃ n = 1 ∞ supp ϕ n , γ ( 2 ) . s_N = \sum_{n=1}^{N}\phi_{n,\gamma}^{(2)}, \qquad
A_N = \bigcup_{n=1}^{N}\operatorname{supp}\phi_{n,\gamma}^{(2)}, \qquad
A = \bigcup_{n=1}^{\infty}\operatorname{supp}\phi_{n,\gamma}^{(2)}. s N = n = 1 ∑ N ϕ n , γ ( 2 ) , A N = n = 1 ⋃ N supp ϕ n , γ ( 2 ) , A = n = 1 ⋃ ∞ supp ϕ n , γ ( 2 ) .
The block supports are pairwise disjoint. Evaluating the weak-star expansion at each e α ( 2 ) e_{\alpha}^{(2)} e α ( 2 ) therefore gives
s N ( e α ( 2 ) ) = { ϕ γ ( 2 ) ( e α ( 2 ) ) , α ∈ A N , 0 , α ∉ A N , ϕ γ ( 2 ) ( e α ( 2 ) ) = 0 ( α ∉ A ) . s_N(e_{\alpha}^{(2)}) =
\begin{cases}
\phi_{\gamma}^{(2)}(e_{\alpha}^{(2)}), & \alpha\in A_N,\\
0, & \alpha\notin A_N,
\end{cases}
\qquad\phi_{\gamma}^{(2)}(e_{\alpha}^{(2)})=0 \quad(\alpha\notin A). s N ( e α ( 2 ) ) = { ϕ γ ( 2 ) ( e α ( 2 ) ) , 0 , α ∈ A N , α ∈ / A N , ϕ γ ( 2 ) ( e α ( 2 ) ) = 0 ( α ∈ / A ) .
The scalar family ( ϕ γ ( 2 ) ( e α ( 2 ) ) ) α < ω 1 (\phi_{\gamma}^{(2)}(e_{\alpha}^{(2)}))_{\alpha<\omega_1} ( ϕ γ ( 2 ) ( e α ( 2 ) ) ) α < ω 1 belongs to ℓ p 2 ( ω 1 ) \ell_{p_2}(\omega_1) ℓ p 2 ( ω 1 ) by (4), and A N A_N A N increases to A A A . Applying (11) to ϕ γ ( 2 ) − s N \phi_{\gamma}^{(2)}-s_N ϕ γ ( 2 ) − s N , we obtain
∥ B 2 , 1 ( ϕ γ ( 2 ) ) − B 2 , 1 s N ∥ ≤ ( ∑ α ∉ A N ∣ ϕ γ ( 2 ) ( e α ( 2 ) ) ∣ p 2 ) 1 / p 2 ⟶ 0. \left\|B_{2,1}(\phi_{\gamma}^{(2)})-B_{2,1}s_N\right\|
\leq
\left(\sum_{\alpha\notin A_N}\left|\phi_{\gamma}^{(2)}(e_{\alpha}^{(2)})\right|^{p_2}\right)^{1/p_2}
\longrightarrow0. B 2 , 1 ( ϕ γ ( 2 ) ) − B 2 , 1 s N ≤ α ∈ / A N ∑ ϕ γ ( 2 ) ( e α ( 2 ) ) p 2 1/ p 2 ⟶ 0.
Linearity gives the finite-sum identity in (9) , completing the proof. □
The combined space E
In this section we present properties of the space E = ( ⨁ k = 1 ∞ ⊕ E k ) 2 , s E=(\bigoplus_{k=1}^{\infty}\oplus E_k)_{2,s} E = ( ⨁ k = 1 ∞ ⊕ E k ) 2 , s with E k E_k E k isometric to E p k , r k E_{p_k,r_k} E p k , r k . One of the basic properties of the space is that every horizontal block sequence generates a separable subspace not containing ℓ 1 \ell_1 ℓ 1 and with non separable dual. In the separable setting this type of construction are known since 1998. In particular in [4 ] it is shown that for ( X n ) n ∈ N (X_n)_{n\in\mathbb{N}} ( X n ) n ∈ N a family of separable space there exists a space X = ( ⨁ n ⊕ X n ) h i X=(\bigoplus_n\oplus X_n)_{hi} X = ( ⨁ n ⊕ X n ) hi with the property every horizontal block sequence generates an HI space. This type of spaces are also used in [7 ] . The closest separable analogue to the space E E E are spaces X = ( ⨁ n ⊕ X n ) g t X=(\bigoplus_n\oplus X_n)_{gt} X = ( ⨁ n ⊕ X n ) g t with the property every horizontal block sequence generates an HI space with non separable dual. Such a result do not exist but it is provable.
We now proceed to the description of E E E . Fix parameters
2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1. (12) 2>s>p_1>r_1>p_2>r_2>\cdots>1. \tag*{(12)} 2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1. (12)
In particular, 1 < r k < p k < ∞ 1<r_k<p_k<\infty 1 < r k < p k < ∞ for every k k k , and p k + 1 < r k p_{k+1}<r_k p k + 1 < r k .
Definition 3.3. Let C ⊂ ω 1 C\subset\omega_1 C ⊂ ω 1 be unbounded and let 1 < r < p < ∞ 1<r<p<\infty 1 < r < p < ∞ . The set C C C has order type ω 1 \omega_1 ω 1 . We denote by E p , r ( C ) E_{p,r}(C) E p , r ( C ) the isometric copy of E p , r E_{p,r} E p , r supported on C C C , obtained by transporting the basis along the unique order isomorphism C → ω 1 C\to\omega_1 C → ω 1 .
Fix a partition ( C k ) k ≥ 1 (C_k)_{k\geq1} ( C k ) k ≥ 1 of ω 1 \omega_1 ω 1 into pairwise disjoint unbounded sets. For an interval J ⊂ N J\subset\mathbb{N} J ⊂ N , put
C J = ⋃ k ∈ J C k , π J : c 00 ( ω 1 ) ⟶ c 00 ( C J ) , π J x = x ∣ C J , C_J=\bigcup_{k\in J}C_k,\qquad\pi_J:c_{00}(\omega_1)\longrightarrow c_{00}(C_J),\qquad\pi_Jx=x|_{C_J}, C J = k ∈ J ⋃ C k , π J : c 00 ( ω 1 ) ⟶ c 00 ( C J ) , π J x = x ∣ C J , where the restriction is extended by zero outside C J C_J C J . Write π k = π { k } \pi_k=\pi_{\{k\}} π k = π { k } and, for x ∈ c 00 ( ω 1 ) x\in c_{00}(\omega_1) x ∈ c 00 ( ω 1 ) , put
ind ( x ) = { k ∈ N : π k x ≠ 0 } . \operatorname{ind}(x)=\{k\in\mathbb{N}:\pi_kx\neq0\}. ind ( x ) = { k ∈ N : π k x = 0 } . A successive block sequence ( x n ) n (x_n)_n ( x n ) n is an h-block sequence if the finite sets ind ( x n ) \operatorname{ind}(x_n) ind ( x n ) are successive:
max ind ( x n ) < min ind ( x n + 1 ) ( n ∈ N ) . \max\operatorname{ind}(x_n)<\min\operatorname{ind}(x_{n+1})\qquad(n\in\mathbb{N}). max ind ( x n ) < min ind ( x n + 1 ) ( n ∈ N ) . For a finitely supported functional f = ∑ α a α e α ∗ f=\sum_{\alpha}a_{\alpha}e_{\alpha}^{*} f = ∑ α a α e α ∗ , define
ind ( f ) = { k : supp f ∩ C k ≠ ∅ } . \operatorname{ind}(f)=\{k:\operatorname{supp}f\cap C_k\neq\varnothing\}. ind ( f ) = { k : supp f ∩ C k = ∅ } . An h-block sequence of such functionals is defined in the same way: the supports are successive in ω 1 \omega_1 ω 1 , and their index sets are successive in N \mathbb{N} N .
Theorem 3.4. For the partition ( C k ) k (C_k)_k ( C k ) k and parameters (12), there is a norm ∥ ⋅ ∥ 2 , s \|\cdot\|_{2,s} ∥ ⋅ ∥ 2 , s on c 00 ( ω 1 ) c_{00}(\omega_1) c 00 ( ω 1 ) such that its completion
E = ( c 00 ( ω 1 ) , ∥ ⋅ ∥ 2 , s ) ‾ E=\overline{\left(c_{00}(\omega_1),\|\cdot\|_{2,s}\right)} E = ( c 00 ( ω 1 ) , ∥ ⋅ ∥ 2 , s ) has the following properties. Write E k = E p k , r k ( C k ) E_k=E_{p_k,r_k}(C_k) E k = E p k , r k ( C k ) .
(1) The family ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 is a boundedly complete transfinite Schauder basis of E E E . Setting
X 0 = span { e α ∗ : α < ω 1 } , X_0=\operatorname{span}\{e_\alpha^*:\alpha<\omega_1\}, X 0 = span { e α ∗ : α < ω 1 } , the canonical evaluation map E → X 0 ∗ E\to X_0^* E → X 0 ∗ is an onto isometry.
(2) For every interval J ⊂ N J\subset\mathbb{N} J ⊂ N , the map π J \pi_J π J extends to a contractive projection on E E E , with range
π J ( E ) = span ‾ { e α : α ∈ C J } . \pi_J(E)=\overline{\operatorname{span}}\{e_\alpha:\alpha\in C_J\}. π J ( E ) = span { e α : α ∈ C J } . For every k k k , the range π k ( E ) \pi_k(E) π k ( E ) is naturally isometric to E k E_k E k . We use the notation
E = ( ⨁ k = 1 ∞ E k ) 2 , s E=\left(\bigoplus_{k=1}^{\infty}E_k\right)_{2,s} E = ( k = 1 ⨁ ∞ E k ) 2 , s for this space and its decomposition.
(3) Every normalized h h h -block sequence spans a subspace with nonseparable dual, and its closed linear span contains no isomorphic copy of ℓ 1 \ell_1 ℓ 1 .
(4) The space E E E contains no isomorphic copy of ℓ 1 \ell_1 ℓ 1 , and every infinite-dimensional closed subspace of E E E has nonseparable dual.
To describe the dual, let
ρ k : E ∗ ⟶ E k ∗ , ρ k f = f ∣ E k , ι k : E k ∗ ⟶ E ∗ , ( ι k g ) ( x ) = g ( π k x ) . \rho_k:E^*\longrightarrow E_k^*,\qquad\rho_kf=f|_{E_k},\qquad\iota_k:E_k^*\longrightarrow E^*,\qquad(\iota_kg)(x)=g(\pi_kx). ρ k : E ∗ ⟶ E k ∗ , ρ k f = f ∣ E k , ι k : E k ∗ ⟶ E ∗ , ( ι k g ) ( x ) = g ( π k x ) . The map ι k \iota_k ι k is an isometry, ρ k \rho_k ρ k is a contraction, and π k ∗ = ι k ρ k \pi_k^*=\iota_k\rho_k π k ∗ = ι k ρ k . We identify E k ∗ E_k^* E k ∗ with its image under ι k \iota_k ι k . Then the following properties hold.
(i) For each k ≥ 1 k\geq1 k ≥ 1 , the family
Φ k = { ϕ γ k : γ ∈ Γ k } ⊂ B E k ∗ \Phi^k=\{\phi_\gamma^k:\gamma\in\Gamma_k\}\subset B_{E_k^*} Φ k = { ϕ γ k : γ ∈ Γ k } ⊂ B E k ∗ is the family from the component space E k E_k E k , extended to E E E by ι k \iota_k ι k . In particular, its members have successive finite block expansions
ϕ γ k = ω ∗ k ∑ n = 1 ∞ ϕ n , γ k , supp ϕ n , γ k ⊂ C k , \phi_\gamma^k=\omega_*^k\sum_{n=1}^{\infty}\phi_{n,\gamma}^k,\qquad\operatorname{supp}\phi_{n,\gamma}^k\subset C_k, ϕ γ k = ω ∗ k n = 1 ∑ ∞ ϕ n , γ k , supp ϕ n , γ k ⊂ C k , and
E k ∗ = span ‾ ( { e α ∗ : α ∈ C k } ∪ Φ k ) . E_k^*=\overline{\operatorname{span}}\left(\{e_\alpha^*:\alpha\in C_k\}\cup\Phi^k\right). E k ∗ = span ( { e α ∗ : α ∈ C k } ∪ Φ k ) . Put
X 0 , k = span ‾ { e α ∗ : α ∈ C k } , Q k : E k ∗ ⟶ W k : = E k ∗ / X 0 , k . X_{0,k}=\overline{\operatorname{span}}\{e_\alpha^*:\alpha\in C_k\},\qquad Q_k:E_k^*\longrightarrow W_k:=E_k^*/X_{0,k}. X 0 , k = span { e α ∗ : α ∈ C k } , Q k : E k ∗ ⟶ W k := E k ∗ / X 0 , k . The nonzero cosets Q k ϕ γ k Q_k\phi_\gamma^k Q k ϕ γ k have norm-dense linear span in W k W_k W k and satisfy an upper ℓ r k \ell_{r_k} ℓ r k estimate with constant one.
(ii) There is an uncountable family of distinct functionals
Φ 0 = { ϕ γ 0 : γ ∈ Γ 0 } ⊂ B E ∗ \Phi^0=\{\phi_\gamma^0:\gamma\in\Gamma_0\}\subset B_{E^*} Φ 0 = { ϕ γ 0 : γ ∈ Γ 0 } ⊂ B E ∗ with the following properties.
(a) For every γ ∈ Γ 0 \gamma\in\Gamma_0 γ ∈ Γ 0 , there is an h h h -block sequence ( ϕ n , γ 0 ) n = 1 ∞ (\phi^0_{n,\gamma})_{n=1}^{\infty} ( ϕ n , γ 0 ) n = 1 ∞ of finitely supported functionals in B E ∗ B_{E^*} B E ∗ such that
ϕ γ 0 = w ∗ − ∑ n = 1 ∞ ϕ n , γ 0 . \phi^0_\gamma= w^*-\sum_{n=1}^{\infty}\phi^0_{n,\gamma}. ϕ γ 0 = w ∗ − n = 1 ∑ ∞ ϕ n , γ 0 .
(b) Put
ϕ > n , γ 0 = ϕ γ 0 − ∑ j = 1 n ϕ j , γ 0 . \phi^0_{>n,\gamma}=\phi^0_\gamma-\sum_{j=1}^{n}\phi^0_{j,\gamma}. ϕ > n , γ 0 = ϕ γ 0 − j = 1 ∑ n ϕ j , γ 0 .
For every finite F ⊂ Γ 0 F \subset\Gamma_0 F ⊂ Γ 0 , there is n 0 ∈ N n_0 \in\mathbb{N} n 0 ∈ N such that, for every n > n 0 n>n_0 n > n 0 and every scalar family ( a γ ) γ ∈ F (a_\gamma)_{\gamma\in F} ( a γ ) γ ∈ F ,
∥ ∑ γ ∈ F a γ ϕ > n , γ 0 ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ s ) 1 / s . (13) \left\|\sum_{\gamma\in F}a_\gamma\phi^0_{>n,\gamma}\right\|
\leq\left(\sum_{\gamma\in F}|a_\gamma|^s\right)^{1/s}. \tag*{(13)} γ ∈ F ∑ a γ ϕ > n , γ 0 ≤ γ ∈ F ∑ ∣ a γ ∣ s 1/ s . (13)
(c) Each ϕ γ 0 \phi^0_\gamma ϕ γ 0 lies outside X 0 X_0 X 0 . Thus, for the quotient map
Q : E ∗ ⟶ W : = E ∗ / X 0 , Q:E^*\longrightarrow W:=E^*/X_0, Q : E ∗ ⟶ W := E ∗ / X 0 ,
we have Q ϕ γ 0 ≠ 0 Q\phi^0_\gamma\ne0 Q ϕ γ 0 = 0 for every γ ∈ Γ 0 \gamma\in\Gamma_0 γ ∈ Γ 0 .
(iii) The dual has the norm-span description
E ∗ = span ‾ ( ⋃ k ≥ 1 E k ∗ ∪ Φ 0 ) = span ‾ ( { e α ∗ : α < ω 1 } ∪ ⋃ k ≥ 0 Φ k ) . (14) E^*=\overline{\operatorname{span}}\left(\bigcup_{k\geq1}E_k^*\cup\Phi^0\right)
=\overline{\operatorname{span}}\left(\{e_\alpha^*:\alpha<\omega_1\}\cup\bigcup_{k\geq0}\Phi^k\right). \tag*{(14)} E ∗ = span ( k ≥ 1 ⋃ E k ∗ ∪ Φ 0 ) = span ( { e α ∗ : α < ω 1 } ∪ k ≥ 0 ⋃ Φ k ) . (14)
We record some consequences of these properties that will be used below. First, P m : = π { 1 , … , m } P_m:=\pi_{\{1,\ldots,m\}} P m := π { 1 , … , m } is a contraction and P m x → x P_mx\to x P m x → x in norm for every x ∈ E x\in E x ∈ E . This follows first for finitely supported vectors and then for all x x x by density. Consequently, every f ∈ E ∗ f\in E^* f ∈ E ∗ has the decomposition
f = w ∗ − ∑ k = 1 ∞ ι k ρ k f , ∥ f ∥ = sup m ∥ P m ∗ f ∥ , π k ∗ ( B E ∗ ) = ι k ( B E k ∗ ) . (15) f=w^*-\sum_{k=1}^{\infty}\iota_k\rho_kf,\qquad
\|f\|=\sup_m\|P_m^*f\|,\qquad
\pi_k^*(B_{E^*})=\iota_k(B_{E_k^*}). \tag*{(15)} f = w ∗ − k = 1 ∑ ∞ ι k ρ k f , ∥ f ∥ = m sup ∥ P m ∗ f ∥ , π k ∗ ( B E ∗ ) = ι k ( B E k ∗ ) . (15)
Thus, if the notation E ∗ = ( ∑ k ⊕ E k ∗ ) ( 2 , s ) ∗ E^*=(\sum_k\oplus E_k^*)_{(2,s)^*} E ∗ = ( ∑ k ⊕ E k ∗ ) ( 2 , s ) ∗ is used, the component sums are understood in this weak-star sense. Norm convergence of these sums is not asserted; the norm-span description is (14).
The space E E E is also WCG. Indeed, each K k = { 0 } ∪ { e α : α ∈ C k } K_k=\{0\}\cup\{e_\alpha:\alpha\in C_k\} K k = { 0 } ∪ { e α : α ∈ C k } is weakly compact in E k E_k E k , and hence in E E E , and ∥ e α ∥ ≤ 1 \|e_\alpha\|\leq1 ∥ e α ∥ ≤ 1 . The set
K = { 0 } ∪ ⋃ k ≥ 1 2 − k K k K=\{0\}\cup\bigcup_{k\geq1}2^{-k}K_k K = { 0 } ∪ k ≥ 1 ⋃ 2 − k K k
is weakly compact and has dense linear span in E E E . To see compactness, a weakly open set containing zero contains all 2 − k K k 2^{-k}K_k 2 − k K k for sufficiently large k k k ; the remaining finitely many sets are compact.
For f ∈ E ∗ f\in E^* f ∈ E ∗ , the component estimate (4) gives
( ∑ α ∈ C k ∣ f ( e α ) ∣ p k ) 1 / p k ≤ ∥ ρ k f ∥ ≤ ∥ f ∥ ( k ≥ 1 ) . (16) \left(\sum_{\alpha\in C_k}|f(e_\alpha)|^{p_k}\right)^{1/p_k}
\leq\|\rho_kf\|\leq\|f\|\qquad(k\geq1). \tag*{(16)} ( α ∈ C k ∑ ∣ f ( e α ) ∣ p k ) 1/ p k ≤ ∥ ρ k f ∥ ≤ ∥ f ∥ ( k ≥ 1 ) . (16)
In particular, the set { α : f ( e α ) ≠ 0 } \{\alpha: f(e_{\alpha}) \ne0\} { α : f ( e α ) = 0 } is a countable union of countable sets and is therefore countable.
The structure of the quotient W W W
We first place all the path cosets in the common quotient W W W . The projection π k ∗ \pi_k^* π k ∗ maps X 0 X_0 X 0 onto X 0 , k X_{0,k} X 0 , k , so
E k ∗ ∩ X 0 = X 0 , k . E_k^* \cap X_0 = X_{0,k}. E k ∗ ∩ X 0 = X 0 , k .
Consequently the map
j k : W k ⟶ W , j k ( Q k g ) = Q ( ι k g ) , j_k: W_k \longrightarrow W,\qquad j_k(Q_k g)=Q(\iota_k g), j k : W k ⟶ W , j k ( Q k g ) = Q ( ι k g ) ,
is an isometric embedding. Indeed, extension by ι k \iota_k ι k gives one inequality between the quotient norms, and restriction by ρ k \rho_k ρ k gives the reverse inequality. In particular, every Q ϕ γ k Q\phi_\gamma^k Q ϕ γ k is nonzero, and the upper ℓ r k \ell_{r_k} ℓ r k estimate also holds in W W W .
We identify each W k W_k W k , k ≥ 1 k \ge1 k ≥ 1 , with its image under j k j_k j k , and set
W 0 = span ‾ { Q ϕ γ 0 : γ ∈ Γ 0 } , V = span ‾ ( ⋃ k ≥ 1 W k ) . W_0=\overline{\operatorname{span}}\{Q\phi_\gamma^0:\gamma\in\Gamma_0\},\qquad
V=\overline{\operatorname{span}}\left(\bigcup_{k\ge1}W_k\right). W 0 = span { Q ϕ γ 0 : γ ∈ Γ 0 } , V = span ( k ≥ 1 ⋃ W k ) .
Thus W 0 W_0 W 0 is the closed span of the horizontal special cosets, and V V V is the closed span of the vertical quotient spaces.
Proposition 3.5 (The decomposition of the quotient). The quotient W = E ∗ / X 0 W=E^*/X_0 W = E ∗ / X 0 has the following properties.
(1) For every k ≥ 1 k\ge1 k ≥ 1 , the formula
P k ( Q f ) = Q ( π k ∗ f ) ( f ∈ E ∗ ) P_k(Qf)=Q(\pi_k^*f)\qquad(f\in E^*) P k ( Q f ) = Q ( π k ∗ f ) ( f ∈ E ∗ ) defines a contractive projection from W W W onto W k W_k W k .
(2) We have W = W 0 ⊕ V W=W_0\oplus V W = W 0 ⊕ V , and both associated projections P 0 : W → W 0 P_0:W\to W_0 P 0 : W → W 0 and P V : W → V P_V:W\to V P V : W → V are contractive. More precisely,
∥ v − z ∥ ≥ max { ∥ v ∥ , ∥ z ∥ } ( v ∈ W 0 , z ∈ V ) . (17) \|v-z\|\ge\max\{\|v\|,\|z\|\}\qquad(v\in W_0,\ z\in V). \tag*{(17)} ∥ v − z ∥ ≥ max { ∥ v ∥ , ∥ z ∥ } ( v ∈ W 0 , z ∈ V ) . (17) In particular, any two normalized vectors, one in W 0 W_0 W 0 and the other in V V V , have distance at least one.
(3) The family ( W k ) k ≥ 0 (W_k)_{k\ge0} ( W k ) k ≥ 0 is a Schauder decomposition of W W W : every w ∈ W w\in W w ∈ W has a unique representation
w = w 0 + ∑ k = 1 ∞ w k , w k = P k w ∈ W k ( k ≥ 0 ) , (18) w=w_0+\sum_{k=1}^{\infty}w_k,\qquad w_k=P_k w\in W_k\quad(k\ge0), \tag*{(18)} w = w 0 + k = 1 ∑ ∞ w k , w k = P k w ∈ W k ( k ≥ 0 ) , (18) where the series converges in norm. Thus we may write
W = ( ⨁ k = 0 ∞ W k ) W=\left(\bigoplus_{k=0}^{\infty}W_k\right) W = ( k = 0 ⨁ ∞ W k ) with its inherited quotient norm. The partial-sum projections P 0 + ∑ k = 1 m P k P_0+\sum_{k=1}^{m}P_k P 0 + ∑ k = 1 m P k have norm at most two.
Proof. Recall that P m = π { 1 , … , m } P_m=\pi_{\{1,\ldots,m\}} P m = π { 1 , … , m } . The operators P m P_m P m and T m = I − P m = π { m + 1 , m + 2 , … } T_m=I-P_m=\pi_{\{m+1,m+2,\ldots\}} T m = I − P m = π { m + 1 , m + 2 , … } are contractions on E E E . Their adjoints preserve X 0 X_0 X 0 , since they send each biorthogonal functional either to itself or to zero. They therefore induce contractions on the quotient:
P ^ m ( Q f ) = Q ( P m ∗ f ) , T ^ m ( Q f ) = Q ( T m ∗ f ) . \widehat{P}_m(Qf)=Q(P_m^*f),\qquad\widehat{T}_m(Qf)=Q(T_m^*f). P m ( Q f ) = Q ( P m ∗ f ) , T m ( Q f ) = Q ( T m ∗ f ) .
In particular, T ^ m = I W − P ^ m \widehat{T}_m=I_W-\widehat{P}_m T m = I W − P m . The same argument applied to π k \pi_k π k gives the contractive projection P k P_k P k in (1) . Its range is W k W_k W k , and
P ^ m = ∑ k = 1 m P k . \widehat{P}_m=\sum_{k=1}^{m}P_k. P m = k = 1 ∑ m P k .
For every horizontal special functional ϕ γ 0 \phi_\gamma^0 ϕ γ 0 , we have
P m ∗ ϕ γ 0 ∈ X 0 . (19) P_m^*\phi_\gamma^0\in X_0. \tag*{(19)} P m ∗ ϕ γ 0 ∈ X 0 . (19)
Indeed, the expansion ϕ γ 0 = w ∗ ∑ n ϕ n , γ 0 \phi_\gamma^0=w^*\sum_n\phi_{n,\gamma}^0 ϕ γ 0 = w ∗ ∑ n ϕ n , γ 0 is an h-block expansion, so only finitely many of its blocks meet the first m m m components. By weak-star continuity of P m ∗ P_m^* P m ∗ , for some N N N ,
P m ∗ ϕ γ 0 = ∑ n = 1 N P m ∗ ϕ n , γ 0 . P_m^*\phi_\gamma^0=\sum_{n=1}^{N}P_m^*\phi_{n,\gamma}^0. P m ∗ ϕ γ 0 = n = 1 ∑ N P m ∗ ϕ n , γ 0 .
The right-hand side is finitely supported and hence belongs to X 0 X_0 X 0 . Passing to the quotient and then to the closed linear span gives
P ^ m v = 0 , T ^ m v = v ( v ∈ W 0 ) . (20) \widehat{P}_m v=0,\qquad\widehat{T}_m v=v\quad(v\in W_0). \tag*{(20)} P m v = 0 , T m v = v ( v ∈ W 0 ) . (20)
If z ∈ W 1 + ⋯ + W m z\in W_1+\cdots+W_m z ∈ W 1 + ⋯ + W m , then P ^ m z = z \widehat{P}_m z=z P m z = z and T ^ m z = 0 \widehat{T}_m z=0 T m z = 0 . Consequently, for v ∈ W 0 v\in W_0 v ∈ W 0 ,
∥ v − z ∥ ≥ ∥ T ^ m ( v − z ) ∥ = ∥ v ∥ , ∥ v − z ∥ ≥ ∥ P ^ m ( v − z ) ∥ = ∥ z ∥ . \|v-z\|\geq\|\widehat{T}_m(v-z)\|=\|v\|,\qquad\|v-z\|\geq\|\widehat{P}_m(v-z)\|=\|z\|. ∥ v − z ∥ ≥ ∥ T m ( v − z ) ∥ = ∥ v ∥ , ∥ v − z ∥ ≥ ∥ P m ( v − z ) ∥ = ∥ z ∥.
The union of the finite vertical sums is dense in V V V , so these inequalities extend to every z ∈ V z\in V z ∈ V and give (17) .
Replacing z z z by − z -z − z , we see that W 0 ∩ V = { 0 } W_0\cap V=\{0\} W 0 ∩ V = { 0 } and that the two natural projections on W 0 + V W_0+V W 0 + V are contractive. The sum is closed: if ( v n + z n ) n (v_n+z_n)_n ( v n + z n ) n is Cauchy, with v n ∈ W 0 v_n\in W_0 v n ∈ W 0 and z n ∈ V z_n\in V z n ∈ V , then both ( v n ) n (v_n)_n ( v n ) n and ( z n ) n (z_n)_n ( z n ) n are Cauchy by the same inequality, and their limits belong to the respective closed subspaces. The dual norm-span description (14) implies that W 0 + V W_0+V W 0 + V is dense in W W W . Thus W = W 0 ⊕ V W=W_0\oplus V W = W 0 ⊕ V , proving (2).
For every z ∈ V z\in V z ∈ V we have P ^ m z → z \widehat{P}_m z\to z P m z → z in norm. This holds eventually with equality on each finite vertical sum and extends to V V V by density and the uniform bound ∥ P ^ m ∥ ≤ 1 \|\widehat{P}_m\|\leq1 ∥ P m ∥ ≤ 1 . Write w = v + z w=v+z w = v + z , with v ∈ W 0 v\in W_0 v ∈ W 0 and z ∈ V z\in V z ∈ V . Equation (20) gives
P 0 w + ∑ k = 1 m P k w = v + P ^ m z ⟶ v + z = w . P_0w+\sum_{k=1}^{m}P_kw=v+\widehat{P}_mz\longrightarrow v+z=w. P 0 w + k = 1 ∑ m P k w = v + P m z ⟶ v + z = w .
Each P k P_k P k is the identity on W k W_k W k and vanishes on the other summands: for k ≥ 1 k\geq1 k ≥ 1 this follows from the component projections and (20), and for k = 0 k=0 k = 0 it follows from (2). Applying these bounded projections to any norm-convergent representation proves uniqueness. Finally,
∥ P 0 + ∑ k = 1 m P k ∥ = ∥ P 0 + P ^ m ∥ ≤ 2 , \left\|P_0+\sum_{k=1}^{m}P_k\right\|=\|P_0+\widehat{P}_m\|\leq2, P 0 + k = 1 ∑ m P k = ∥ P 0 + P m ∥ ≤ 2 ,
which completes the proof of (3). The projections onto the horizontal and vertical parts can also be expressed by the norm limits
P 0 w = lim m → ∞ T ^ m w , P V w = lim m → ∞ P ^ m w ( w ∈ W ) . P_0w=\lim_{m\to\infty}\widehat{T}_m w,\qquad P_Vw=\lim_{m\to\infty}\widehat{P}_m w\qquad(w\in W). P 0 w = m → ∞ lim T m w , P V w = m → ∞ lim P m w ( w ∈ W ) .
□ \square □
This argument uses the contractive component projections, the h-block expansions of the horizontal special functionals, and the dual norm-span description. It does not use CH.
The operator B B B
For the horizontal family, finite initial block sums lie in X 0 X_0 X 0 . Hence (13), followed by Q Q Q , gives an upper ℓ s \ell_s ℓ s estimate for ( Q ϕ γ 0 ) γ ∈ Γ 0 (Q\phi_\gamma^0)_{\gamma\in\Gamma_0} ( Q ϕ γ 0 ) γ ∈ Γ 0 . Set
q 0 = s , q k = r k ( k ≥ 1 ) . q_0=s,\qquad q_k=r_k\quad(k\geq1). q 0 = s , q k = r k ( k ≥ 1 ) .
The argument proving (7) now gives contractions
S k : ℓ q k ( Γ k ) ⟶ W , S k u γ = Q ϕ γ k ( k ≥ 0 ) . (21) S_k:\ell_{q_k}(\Gamma_k)\longrightarrow W,\qquad S_k u_\gamma=Q\phi_\gamma^k\qquad(k\geq0). \tag*{(21)} S k : ℓ q k ( Γ k ) ⟶ W , S k u γ = Q ϕ γ k ( k ≥ 0 ) . (21)
where ( u γ ) γ (u_\gamma)_\gamma ( u γ ) γ denotes the unit vector basis. Also, (14) implies
W = span ‾ { Q ϕ γ k : k ≥ 0 , γ ∈ Γ k } . (22) W=\overline{\operatorname{span}}\{Q\phi_\gamma^k:k\geq0,\ \gamma\in\Gamma_k\}. \tag*{(22)} W = span { Q ϕ γ k : k ≥ 0 , γ ∈ Γ k } . (22)
Assume CH. Every functional in E ∗ E^* E ∗ is determined by a countably supported real-valued family on ω 1 \omega_1 ω 1 , by (16). There are ℵ 1 ℵ 0 = ℵ 1 \aleph_1^{\aleph_0}=\aleph_1 ℵ 1 ℵ 0 = ℵ 1 such families under CH. Each Φ k \Phi^k Φ k is uncountable and consists of distinct functionals; therefore ∣ Γ k ∣ = ℵ 1 |\Gamma_k|=\aleph_1 ∣ Γ k ∣ = ℵ 1 for every k ≥ 0 k\geq0 k ≥ 0 . We may fix bijections
σ k : C k + 1 ⟶ Γ k ( k ≥ 0 ) . (23) \sigma_k:C_{k+1}\longrightarrow\Gamma_k\qquad(k\geq0). \tag*{(23)} σ k : C k + 1 ⟶ Γ k ( k ≥ 0 ) . (23)
Thus C 1 C_1 C 1 is assigned to the horizontal family Φ 0 \Phi^0 Φ 0 , and C k + 1 C_{k+1} C k + 1 is assigned to Φ k \Phi^k Φ k for $k\geq 1. The parameter inequalities give exactly the estimates needed for these assignments:
p 1 < s = q 0 , p k + 1 < r k = q k ( k ≥ 1 ) . p_1<s=q_0,\qquad p_{k+1}<r_k=q_k\qquad(k\geq1). p 1 < s = q 0 , p k + 1 < r k = q k ( k ≥ 1 ) .
Figure 1 illustrates the assignment of the families and the component operators in the following proposition.
Figure 1. The assignment and the operators in Proposition 3.6 . The first component supplies the horizontal family, and component k + 1 k+1 k + 1 supplies family k k k . The weights are applied when the component operators are combined to form B B B .
Proposition 3.6. Assume CH and the properties of E E E in Theorem 3.4. For k ≥ 0 k\geq0 k ≥ 0 put τ k = 2 − k − 5 \tau_k=2^{-k-5} τ k = 2 − k − 5 . For f ∈ E ∗ f\in E^* f ∈ E ∗ define
B f = ∑ k = 0 ∞ τ k S k ( ( f ( e σ k − 1 ( γ ) ∗ ) ) γ ∈ Γ k ) . (24) Bf=\sum_{k=0}^{\infty}\tau_kS_k\left((f(e_{\sigma_k^{-1}(\gamma)}^*))_{\gamma\in\Gamma_k}\right). \tag*{(24)} B f = k = 0 ∑ ∞ τ k S k ( ( f ( e σ k − 1 ( γ ) ∗ ) ) γ ∈ Γ k ) . (24) Then B : E ∗ → W B:E^*\to W B : E ∗ → W is a bounded linear operator satisfying:
(1) ∥ B ∥ ≤ 1 / 16 < 1 / 10 \|B\|\leq1/16<1/10 ∥ B ∥ ≤ 1/16 < 1/10 .
(2) For k ≥ 0 k\geq0 k ≥ 0 and α ∈ C k + 1 \alpha\in C_{k+1} α ∈ C k + 1 ,
B ( e α ∗ ) = τ k Q ϕ σ k ( α ) k . (25) B(e_\alpha^*)=\tau_kQ\phi_{\sigma_k(\alpha)}^k. \tag*{(25)} B ( e α ∗ ) = τ k Q ϕ σ k ( α ) k . (25) In particular, B ( X 0 ) ‾ = W \overline{B(X_0)}=W B ( X 0 ) = W .
(3) For every j ≥ 0 j\geq0 j ≥ 0 and γ ∈ Γ j \gamma\in\Gamma_j γ ∈ Γ j ,
B ( ∑ n = 1 N ϕ n , γ j ) = ∑ n = 1 N B ( ϕ n , γ j ) → N → ∞ ∥ ⋅ ∥ B ( ϕ γ j ) . (26) B\left(\sum_{n=1}^{N}\phi_{n,\gamma}^j\right)=\sum_{n=1}^{N}B(\phi_{n,\gamma}^j)\xrightarrow[N\to\infty]{\|\cdot\|}B(\phi_\gamma^j). \tag*{(26)} B ( n = 1 ∑ N ϕ n , γ j ) = n = 1 ∑ N B ( ϕ n , γ j ) ∥ ⋅ ∥ N → ∞ B ( ϕ γ j ) . (26)
B ∗ ( W ∗ ) ⊂ J E ( E ) (27) B^{*}(W^{*}) \subset J_{E}(E) \tag*{(27)} B ∗ ( W ∗ ) ⊂ J E ( E ) (27)
Moreover, with the disjoint index set
Γ = { ( k , γ ) : k ≥ 0 , γ ∈ Γ k } , ϕ ( k , γ ) = ϕ γ k , \Gamma= \{(k,\gamma): k \ge0,\ \gamma\in\Gamma_k\}, \qquad\phi_{(k,\gamma)} = \phi^{k}_{\gamma}, Γ = {( k , γ ) : k ≥ 0 , γ ∈ Γ k } , ϕ ( k , γ ) = ϕ γ k ,
the rule σ ( α ) = ( k , σ k ( α ) ) \sigma(\alpha) = (k,\sigma_k(\alpha)) σ ( α ) = ( k , σ k ( α )) for α ∈ C k + 1 \alpha\in C_{k+1} α ∈ C k + 1 defines a bijection σ : ω 1 → Γ \sigma: \omega_1 \to\Gamma σ : ω 1 → Γ . Taking λ ( k , γ ) = τ k \lambda_{(k,\gamma)} = \tau_k λ ( k , γ ) = τ k , the operator B B B satisfies property 5 of § 3 .
Proof. For k ≥ 0 k \ge0 k ≥ 0 and f ∈ E ∗ f \in E^* f ∈ E ∗ , write
a k ( f ) = ( f ( e σ k − 1 ( γ ) ) ) γ ∈ Γ k , A k f = S k a k ( f ) . a_k(f) = \left(f\left(e_{\sigma_k^{-1}(\gamma)}\right)\right)_{\gamma\in\Gamma_k}, \qquad A_kf = S_ka_k(f). a k ( f ) = ( f ( e σ k − 1 ( γ ) ) ) γ ∈ Γ k , A k f = S k a k ( f ) .
By (16) and p k + 1 < q k p_{k+1} < q_k p k + 1 < q k ,
∥ A k f ∥ ≤ ∥ a k ( f ) ∥ ℓ q k ≤ ∥ a k ( f ) ∥ ℓ p k + 1 = ( ∑ α ∈ C k + 1 ∣ f ( e α ) ∣ p k + 1 ) 1 / p k + 1 ≤ ∥ f ∥ . (28) \lVert A_kf\rVert\le\lVert a_k(f)\rVert_{\ell_{q_k}} \le\lVert a_k(f)\rVert_{\ell_{p_{k+1}}} = \left(\sum_{\alpha\in C_{k+1}} \lvert f(e_\alpha)\rvert^{p_{k+1}}\right)^{1/p_{k+1}} \le\lVert f\rVert. \tag*{(28)} ∥ A k f ∥ ≤ ∥ a k ( f ) ∥ ℓ q k ≤ ∥ a k ( f ) ∥ ℓ p k + 1 = α ∈ C k + 1 ∑ ∣ f ( e α ) ∣ p k + 1 1/ p k + 1 ≤ ∥ f ∥ . (28)
This is precisely the estimate of Proposition 3.2. For k ≥ 1 k \ge1 k ≥ 1 , A k A_k A k is the composition of restriction to E k + 1 E_{k+1} E k + 1 , the operator from that proposition with target W k W_k W k , and the embedding j k j_k j k . For k = 0 k = 0 k = 0 , the same proof uses the upper ℓ s \ell_s ℓ s estimate of the horizontal family in place of the upper ℓ r k \ell_{r_k} ℓ r k estimate.
Each A k A_k A k is therefore a linear contraction. Since
∑ k = 0 ∞ τ k = 1 16 , \sum_{k=0}^{\infty}\tau_k=\frac{1}{16}, k = 0 ∑ ∞ τ k = 16 1 ,
the series B = ∑ k = 0 ∞ τ k A k B=\sum_{k=0}^{\infty}\tau_kA_k B = ∑ k = 0 ∞ τ k A k converges in operator norm and proves (1). Biorthogonality and the disjointness of the C k C_k C k give (25) . Every path coset is thus a nonzero scalar multiple of B ( e α ∗ ) B(e_\alpha^*) B ( e α ∗ ) for a suitable α \alpha α . Equation (22) yields B ( X 0 ) ‾ = W \overline{B(X_0)}=W B ( X 0 ) = W , proving (2). The global bijection and the scalars λ ( k , γ ) \lambda_{(k,\gamma)} λ ( k , γ ) have the asserted properties by their definitions.
We prove (3) simultaneously for the component and horizontal families. Fix j ≥ 0 j \ge0 j ≥ 0 and γ ∈ Γ j \gamma\in\Gamma_j γ ∈ Γ j , and put
ϕ = ϕ γ j , s N = ∑ n = 1 N ϕ n , γ j , F N = ⋃ n = 1 N supp ϕ n , γ j . \phi=\phi_\gamma^j,\qquad s_N=\sum_{n=1}^{N}\phi_{n,\gamma}^j,\qquad F_N=\bigcup_{n=1}^{N}\operatorname{supp}\phi_{n,\gamma}^j. ϕ = ϕ γ j , s N = n = 1 ∑ N ϕ n , γ j , F N = n = 1 ⋃ N supp ϕ n , γ j .
The block supports are pairwise disjoint. The weak-star expansion therefore gives
( ϕ − s N ) ( e α ) = { 0 , α ∈ F N , ϕ ( e α ) , α ∉ F N , ϕ ( e α ) = 0 ( α ∉ ⋃ N F N ) . (\phi-s_N)(e_\alpha)=
\begin{cases}
0, & \alpha\in F_N,\\
\phi(e_\alpha), & \alpha\notin F_N,
\end{cases}
\qquad
\phi(e_\alpha)=0\quad\left(\alpha\notin\bigcup_N F_N\right). ( ϕ − s N ) ( e α ) = { 0 , ϕ ( e α ) , α ∈ F N , α ∈ / F N , ϕ ( e α ) = 0 ( α ∈ / N ⋃ F N ) .
For each fixed k k k , the scalar family ( ϕ ( e α ) ) α ∈ C k + 1 (\phi(e_\alpha))_{\alpha\in C_{k+1}} ( ϕ ( e α ) ) α ∈ C k + 1 belongs to ℓ p k + 1 ( C k + 1 ) \ell_{p_{k+1}}(C_{k+1}) ℓ p k + 1 ( C k + 1 ) . Hence (28) gives
∥ A k ( ϕ − s N ) ∥ ≤ ( ∑ α ∈ C k + 1 ∖ F N ∣ ϕ ( e α ) ∣ p k + 1 ) 1 / p k + 1 ⟶ 0. (29) \lVert A_k(\phi-s_N)\rVert\le\left(\sum_{\alpha\in C_{k+1}\setminus F_N}\lvert\phi(e_\alpha)\rvert^{p_{k+1}}\right)^{1/p_{k+1}}\longrightarrow0. \tag*{(29)} ∥ A k ( ϕ − s N )∥ ≤ α ∈ C k + 1 ∖ F N ∑ ∣ ϕ ( e α ) ∣ p k + 1 1/ p k + 1 ⟶ 0. (29)
This is the same truncation argument as in Proposition 3.2.
Since s N → ϕ s_N\to\phi s N → ϕ weak-star, the uniform boundedness principle gives M : = sup N ∥ ϕ − s N ∥ < ∞ M:=\sup_N\lVert\phi-s_N\rVert<\infty M := sup N ∥ ϕ − s N ∥ < ∞ . For every K ≥ 0 K\ge0 K ≥ 0 ,
∥ B ( ϕ − s N ) ∥ ≤ ∑ k = 0 K τ k ∥ A k ( ϕ − s N ) ∥ + M ∑ k > K τ k . \lVert B(\phi-s_N)\rVert\le\sum_{k=0}^{K}\tau_k\lVert A_k(\phi-s_N)\rVert+M\sum_{k>K}\tau_k. ∥ B ( ϕ − s N )∥ ≤ k = 0 ∑ K τ k ∥ A k ( ϕ − s N )∥ + M k > K ∑ τ k .
For fixed K K K , the finite sum tends to zero by (29) ; the second term tends to zero as K → ∞ K\to\infty K → ∞ . This proves the norm convergence in (26) . The finite-sum identity follows from linearity.
Finally, Theorem 3.4 supplies the canonical identification E = X 0 ∗ E=X_0^* E = X 0 ∗ , the weak-star block expansions, and the norm-span assertion (14) . Together with (26) , these are exactly the hypotheses of Proposition 2.2. It follows that B ∗ ( W ∗ ) ⊂ J E ( E ) B^*(W^*)\subset J_E(E) B ∗ ( W ∗ ) ⊂ J E ( E ) , proving (4). □ \square □
The description of E E E , the WCG argument above, and Proposition 3.6 establish all the properties used in Section 2. Therefore Theorem 2.8 applies: under CH, X = ker ( Q − B ) X = \ker(Q-B) X = ker ( Q − B ) has no infinite-dimensional separable quotient, and the canonical restriction map E → X ∗ E \to X^* E → X ∗ is an onto isomorphism.
The norming sets G p , r G_{p,r} G p , r and G G G of E p , r E_{p,r} E p , r and E E E
In this section we will describe the norming sets for defining the norms of the spaces E p , r E_{p,r} E p , r and E E E . A norming set is a G ⊂ ( c 00 ( ω 1 ) , Q ) G \subset(c_{00}(\omega_1),\mathbb{Q}) G ⊂ ( c 00 ( ω 1 ) , Q ) containing the singletons and for every f ∈ G f \in G f ∈ G we have ∥ f ∥ ∞ ≤ 1 \|f\|_\infty\le1 ∥ f ∥ ∞ ≤ 1 . It induces a norm on c 00 ( ω 1 ) c_{00}(\omega_1) c 00 ( ω 1 ) with the rule
∥ x ∥ G = sup { ∣ f ( x ) ∣ : f ∈ G } \|x\|_G = \sup\{|f(x)| : f \in G\} ∥ x ∥ G = sup { ∣ f ( x ) ∣ : f ∈ G }
If the set G G G is closed in the projection of its elements to the intervals of ω 1 \omega_1 ω 1 then the family { e α } α ∈ ω 1 \{e_\alpha\}_{\alpha\in\omega_1} { e α } α ∈ ω 1 is a transfinite basis for the space X G = span c 00 ( ω 1 , ∥ ⋅ ∥ G ) X_G = \operatorname{span} c_{00}(\omega_1,\|\cdot\|_G) X G = span c 00 ( ω 1 , ∥ ⋅ ∥ G ) .
For the norming sets we fix two strictly increasing sequences of positive integers ( m j ) j (m_j)_j ( m j ) j and ( n j ) j (n_j)_j ( n j ) j , following the rapid-growth parameter choices used in [9 ] . We record explicitly the conditions needed in the arguments below:
∑ j = 1 ∞ 1 m j < 1 , n j 1 / q m j ⟶ ∞ for every 1 < q < ∞ . (30) \sum_{j=1}^{\infty}\frac{1}{m_j}<1,\qquad\frac{n_j^{1/q}}{m_j}\longrightarrow\infty\quad\text{for every }1<q<\infty. \tag*{(30)} j = 1 ∑ ∞ m j 1 < 1 , m j n j 1/ q ⟶ ∞ for every 1 < q < ∞. (30)
The second condition strengthens n j / m j → ∞ n_j/m_j \to\infty n j / m j → ∞ and can be arranged, for example, by choosing n j ≥ m j j n_j \ge m_j^j n j ≥ m j j . For bounded completeness at a fixed p p p , we use it with q = p ′ = p / ( p − 1 ) q=p'=p/(p-1) q = p ′ = p / ( p − 1 ) . The standard parameters in [6 Section 2.1] also satisfy this growth condition.
Definition 4.1. Let G ⊂ c 00 ( ω 1 , Q ) G \subset c_{00}(\omega_1,\mathbb{Q}) G ⊂ c 00 ( ω 1 , Q ) , 1 < p < ∞ 1<p<\infty 1 < p < ∞ , and j ∈ N j\in\mathbb{N} j ∈ N .
(1) We say that an f ∈ G f\in G f ∈ G is the result of the ( p , j ) (p,j) ( p , j ) -operation if
f = 1 m j ∑ k = 1 n j λ k f k with ( λ k ) k ⊂ Q , ∑ k ∣ λ k ∣ p ≤ 1 and f 1 < ⋯ < f n j f=\frac{1}{m_j}\sum_{k=1}^{n_j}\lambda_k f_k\quad\text{with}\quad(\lambda_k)_k\subset\mathbb{Q},\quad\sum_k|\lambda_k|^p\le1\quad\text{and}\quad f_1<\cdots<f_{n_j} f = m j 1 k = 1 ∑ n j λ k f k with ( λ k ) k ⊂ Q , k ∑ ∣ λ k ∣ p ≤ 1 and f 1 < ⋯ < f n j where f 1 , … , f n j f_1,\ldots,f_{n_j} f 1 , … , f n j are nonzero elements of G G G .
(2) If f ∈ G f\in G f ∈ G is the result of ( p , j ) (p,j) ( p , j ) -operation then we say that f f f is a weighted functional with its weight w ( f ) = m j w(f)=m_j w ( f ) = m j . Notice that for a weighted functional f f f its weight is not necessarily unique.
(3) The set G ⊂ ( c 00 ( ω 1 ) , Q ) G\subset(c_{00}(\omega_1),\mathbb{Q}) G ⊂ ( c 00 ( ω 1 ) , Q ) is closed in ( p , j ) (p,j) ( p , j ) -operation if
∀ ( f k ) k = 1 n j successive blocks in G ∀ ( λ k ) ⊂ Q with ∑ k ∣ λ k ∣ p ≤ 1 we have \forall(f_k)_{k=1}^{n_j}\text{ successive blocks in }G\ \ \forall(\lambda_k)\subset\mathbb{Q}\text{ with }\sum_k|\lambda_k|^p\le1\text{ we have} ∀ ( f k ) k = 1 n j successive blocks in G ∀ ( λ k ) ⊂ Q with k ∑ ∣ λ k ∣ p ≤ 1 we have f = 1 m j ∑ k = 1 n j λ k f k ∈ G f=\frac{1}{m_j}\sum_{k=1}^{n_j}\lambda_k f_k\in G f = m j 1 k = 1 ∑ n j λ k f k ∈ G
The σ \sigma σ -ϱ \varrho ϱ coding and special sequences
Special sequences are defined using the σ ϱ \sigma_\varrho σ ϱ coding of [6 Sections 2.3–2.4] , based on a ϱ \varrho ϱ function as in [19 ] . The coding assigns a weight m j m_j m j to a finite sequence of successive rational functionals together with their chosen weights and integer parameters. In the separable setting the family to be coded is countable, so an injective coding is available. On ω 1 \omega_1 ω 1 , the function σ ϱ \sigma_\varrho σ ϱ is not injective; the finite closures associated with ϱ \varrho ϱ provide the agreement on common initial parts that will be used below.
The σ ϱ \sigma_{\varrho} σ ϱ coding
We give the construction in the following five steps. The auxiliary function σ \sigma σ takes integer values, whereas σ ϱ \sigma_{\varrho} σ ϱ will take values in { m j : j ∈ N } \{m_j : j \in\mathbb{N}\} { m j : j ∈ N } , in accordance with our convention for weights. Thus our σ ϱ \sigma_{\varrho} σ ϱ includes the passage from an index to the corresponding weight; in [6 Section 2.4] this passage is written separately.
The ϱ \varrho ϱ function. Fix a function ϱ : [ ω 1 ] 2 → N \varrho: [\omega_1]^2 \to\mathbb{N} ϱ : [ ω 1 ] 2 → N with the following properties:
(i) For α < β < γ < ω 1 \alpha< \beta< \gamma< \omega_1 α < β < γ < ω 1 ,
ϱ ( α , γ ) ≤ max { ϱ ( α , β ) , ϱ ( β , γ ) } . \varrho(\alpha,\gamma) \le\max\{\varrho(\alpha,\beta),\varrho(\beta,\gamma)\}. ϱ ( α , γ ) ≤ max { ϱ ( α , β ) , ϱ ( β , γ )} .
(ii) For α < β < γ < ω 1 \alpha< \beta< \gamma< \omega_1 α < β < γ < ω 1 ,
ϱ ( α , β ) ≤ max { ϱ ( α , γ ) , ϱ ( β , γ ) } . \varrho(\alpha,\beta) \le\max\{\varrho(\alpha,\gamma),\varrho(\beta,\gamma)\}. ϱ ( α , β ) ≤ max { ϱ ( α , γ ) , ϱ ( β , γ )} .
(iii) For every β < ω 1 \beta< \omega_1 β < ω 1 and p ∈ N p \in\mathbb{N} p ∈ N , the set
{ α < β : ϱ ( α , β ) ≤ p } \{\alpha< \beta: \varrho(\alpha,\beta) \le p\} { α < β : ϱ ( α , β ) ≤ p }
is finite.
We also put ϱ ( α , α ) = 0 \varrho(\alpha,\alpha)=0 ϱ ( α , α ) = 0 . The existence of such a function is established in [19 ] ; we fix one throughout.
Finite closures. For a finite set F ⊂ ω 1 F \subset\omega_1 F ⊂ ω 1 , put
p F = max ( { ϱ ( α , β ) : α , β ∈ F , α < β } ∪ { 0 } ) . p_F=\max\bigl(\{\varrho(\alpha,\beta):\alpha,\beta\in F,\ \alpha<\beta\}\cup\{0\}\bigr). p F = max ( { ϱ ( α , β ) : α , β ∈ F , α < β } ∪ { 0 } ) .
For p ∈ N p\in\mathbb{N} p ∈ N , define its p p p -closure by
F ‾ p = F ∪ { α < ω 1 : there is β ∈ F with α < β and ϱ ( α , β ) ≤ p } . \overline{F}^{p}=F\cup\{\alpha<\omega_1:\text{ there is }\beta\in F\text{ with }\alpha<\beta\text{ and }\varrho(\alpha,\beta)\le p\}. F p = F ∪ { α < ω 1 : there is β ∈ F with α < β and ϱ ( α , β ) ≤ p } .
This set is finite by (iii). We call F F F p p p -closed if F ‾ p = F \overline{F}^{p}=F F p = F , and ϱ \varrho ϱ -closed if F ‾ p F = F \overline{F}^{p_F}=F F p F = F , using the same formula when p F = 0 p_F=0 p F = 0 . The closure is monotone and idempotent:
F ⊂ H ⟹ F ‾ p ⊂ H ‾ p , ( F ‾ p ) ‾ p = F ‾ p . F\subset H\ \Longrightarrow\ \overline{F}^{p}\subset\overline{H}^{p},\qquad\overline{\left(\overline{F}^{p}\right)}^{p}=\overline{F}^{p}. F ⊂ H ⟹ F p ⊂ H p , ( F p ) p = F p .
Indeed, if α ≤ β ≤ γ \alpha\le\beta\le\gamma α ≤ β ≤ γ , with ϱ ( α , β ) ≤ p \varrho(\alpha,\beta)\le p ϱ ( α , β ) ≤ p and ϱ ( β , γ ) ≤ p \varrho(\beta,\gamma)\le p ϱ ( β , γ ) ≤ p , then (i) gives ϱ ( α , γ ) ≤ p \varrho(\alpha,\gamma)\le p ϱ ( α , γ ) ≤ p . If F ≠ ∅ F\ne\varnothing F = ∅ and p ≥ p F p\ge p_F p ≥ p F , then
F ‾ p = { max F } p , p F ‾ p ≤ p . (31) \overline{F}^{p}=\{\max F\}^{p},\qquad p_{\overline{F}^{p}}\le p. \tag*{(31)} F p = { max F } p , p F p ≤ p . (31)
For the equality, use F ⊂ { max F } p F\subset\{\max F\}^{p} F ⊂ { max F } p and idempotence. For the inequality, apply (ii) to two members of this closure and max F \max F max F .
Here is the coherence property needed for coding. If H H H and K K K are finite p p p -closed sets with p H , p K ≤ p p_H,p_K\le p p H , p K ≤ p , then for every α ∈ H ∩ K \alpha\in H\cap K α ∈ H ∩ K ,
H ∩ [ 0 , α ] = { α } p = K ∩ [ 0 , α ] . (32) H\cap[0,\alpha]=\{\alpha\}^{p}=K\cap[0,\alpha]. \tag*{(32)} H ∩ [ 0 , α ] = { α } p = K ∩ [ 0 , α ] . (32)
In fact, p H ≤ p p_H\le p p H ≤ p gives H ∩ [ 0 , α ] ⊂ { α } p H\cap[0,\alpha]\subset\{\alpha\}^{p} H ∩ [ 0 , α ] ⊂ { α } p ; the reverse inclusion follows from p p p -closedness. The argument for K K K is the same. Consequently, H ∩ K H\cap K H ∩ K is an initial part of both H H H and K K K (possibly empty).
Increasing enumerations of finite sets
For a nonempty finite set F ⊂ ω 1 F \subset\omega_1 F ⊂ ω 1 , let
π F : { 1 , … , ∣ F ∣ } ⟶ F \pi_F:\{1,\ldots,|F|\}\longrightarrow F π F : { 1 , … , ∣ F ∣ } ⟶ F
be its increasing bijection. If f f f is a scalar-valued function on F F F , write π F ( f ) \pi_F(f) π F ( f ) for the function on N \mathbb{N} N given by
π F ( f ) ( n ) = { f ( π F ( n ) ) , 1 ≤ n ≤ ∣ F ∣ , 0 , n > ∣ F ∣ . \pi_F(f)(n)=
\begin{cases}
f(\pi_F(n)), & 1\le n\le|F|,\\
0, & n>|F|.
\end{cases} π F ( f ) ( n ) = { f ( π F ( n )) , 0 , 1 ≤ n ≤ ∣ F ∣ , n > ∣ F ∣.
Thus π F ( f ) \pi_F(f) π F ( f ) records the values of f f f in the order of F F F , including the zero entries. We use the same notation if f f f is a finitely supported function on ω 1 \omega_1 ω 1 with supp f ⊂ F \operatorname{supp} f\subset F supp f ⊂ F . In particular, rational-valued f f f gives π F ( f ) ∈ c 00 ( N , Q ) \pi_F(f)\in c_{00}(\mathbb{N},\mathbb{Q}) π F ( f ) ∈ c 00 ( N , Q ) .
The countable coding σ \sigma σ
Let Q s \mathcal{Q}_s Q s be the set of all finite sequences
s = ( g 1 , w 1 , ℓ 1 , … , g d , w d , ℓ d ) , d ≥ 1 , s=(g_1,w_1,\ell_1,\ldots,g_d,w_d,\ell_d),\qquad d\ge1, s = ( g 1 , w 1 , ℓ 1 , … , g d , w d , ℓ d ) , d ≥ 1 ,
where g 1 < ⋯ < g d g_1<\cdots<g_d g 1 < ⋯ < g d are nonzero elements of c 00 ( N , Q ) c_{00}(\mathbb{N},\mathbb{Q}) c 00 ( N , Q ) , and ( w i ) i = 1 d (w_i)_{i=1}^d ( w i ) i = 1 d and ( ℓ i ) i = 1 d (\ell_i)_{i=1}^d ( ℓ i ) i = 1 d are strictly increasing sequences of positive integers. Put
ε ( s ) = min { ∣ g i ( n ) ∣ : 1 ≤ i ≤ d , n ∈ supp g i } > 0. \varepsilon(s)=\min\{|g_i(n)|:1\le i\le d,\ n\in\operatorname{supp}g_i\}>0. ε ( s ) = min { ∣ g i ( n ) ∣ : 1 ≤ i ≤ d , n ∈ supp g i } > 0.
Since Q s \mathcal{Q}_s Q s is countable, we may choose an injection σ : Q s → N \sigma:\mathcal{Q}_s\to\mathbb{N} σ : Q s → N satisfying
σ ( s ) > max { w d , ℓ d 2 , ε ( s ) − 2 , max supp g d } . (33) \sigma(s)>\max\{w_d,\ell_d^2,\varepsilon(s)^{-2},\max\operatorname{supp}g_d\}. \tag*{(33)} σ ( s ) > max { w d , ℓ d 2 , ε ( s ) − 2 , max supp g d } . (33)
To do so, enumerate Q s \mathcal{Q}_s Q s and assign to each successive element an unused integer exceeding its displayed bound. The additional bound by w d w_d w d makes the increase of the resulting weights explicit.
The coding σ ϱ \sigma_\varrho σ ϱ
Let Q s ( ω 1 ) \mathcal{Q}_s(\omega_1) Q s ( ω 1 ) consist of the finite sequences
f = ( f 1 , w 1 , ℓ 1 , … , f d , w d , ℓ d ) , \mathbf{f}=(f_1,w_1,\ell_1,\ldots,f_d,w_d,\ell_d), f = ( f 1 , w 1 , ℓ 1 , … , f d , w d , ℓ d ) ,
where f 1 < ⋯ < f d f_1<\cdots<f_d f 1 < ⋯ < f d are nonzero elements of c 00 ( ω 1 , Q ) c_{00}(\omega_1,\mathbb{Q}) c 00 ( ω 1 , Q ) , the sequences ( w i ) (w_i) ( w i ) and ( ℓ i ) (\ell_i) ( ℓ i ) are strictly increasing sequences of positive integers, and
ℓ i ≥ p F i , F i = ⋃ k = 1 i supp f k ( 1 ≤ i ≤ d ) . \ell_i\ge p_{F_i},\qquad F_i=\bigcup_{k=1}^{i}\operatorname{supp}f_k\qquad(1\le i\le d). ℓ i ≥ p F i , F i = k = 1 ⋃ i supp f k ( 1 ≤ i ≤ d ) .
The parameters ℓ i \ell_i ℓ i are chosen subject to these conditions; they are not uniquely determined by the supports. Set
H f = F d ‾ ℓ d , f ^ = ( π H f ( f 1 ) , w 1 , ℓ 1 , … , π H f ( f d ) , w d , ℓ d ) . H_{\mathbf{f}}=\overline{F_d}^{\,\ell_d},\qquad
\widehat{\mathbf{f}}=(\pi_{H_{\mathbf{f}}}(f_1),w_1,\ell_1,\ldots,\pi_{H_{\mathbf{f}}}(f_d),w_d,\ell_d). H f = F d ℓ d , f = ( π H f ( f 1 ) , w 1 , ℓ 1 , … , π H f ( f d ) , w d , ℓ d ) .
The increasing enumeration preserves successiveness, so f ^ ∈ Q s \widehat{\mathbf{f}}\in\mathcal{Q}_s f ∈ Q s . We define
σ ϱ ( f ) = m σ ( f ^ ) . (34) \sigma_\varrho(\mathbf{f})=m_{\sigma(\widehat{\mathbf{f}})}. \tag*{(34)} σ ϱ ( f ) = m σ ( f ) . (34)
Since ( m j ) (m_j) ( m j ) is a strictly increasing sequence of positive integers, m j ≥ j m_j\ge j m j ≥ j , and (32) implies σ ϱ ( f ) > w d \sigma_\varrho(\mathbf{f})>w_d σ ϱ ( f ) > w d .
Proposition 4.2. Suppose f , g ∈ Q s ( ω 1 ) \mathbf{f},\mathbf{g}\in\mathcal{Q}_s(\omega_1) f , g ∈ Q s ( ω 1 ) and σ ϱ ( f ) = σ ϱ ( g ) \sigma_\varrho(\mathbf{f})=\sigma_\varrho(\mathbf{g}) σ ϱ ( f ) = σ ϱ ( g ) . Then the two sequences have the same length d d d , the same weight and integer parameters, and the same transported functionals:
π H f ( f i ) = π H g ( g i ) ( 1 ≤ i ≤ d ) . \pi_{H_{\mathbf{f}}}(f_i)=\pi_{H_{\mathbf{g}}}(g_i)\qquad(1\le i\le d). π H f ( f i ) = π H g ( g i ) ( 1 ≤ i ≤ d ) .
Moreover, H f ∩ H g H_f \cap H_g H f ∩ H g is an initial part of both closures, and
f i ∣ H f ∩ H g = g i ∣ H f ∩ H g ( 1 ≤ i ≤ d ) . \left.f_i\right|_{H_f\cap H_g}=\left.g_i\right|_{H_f\cap H_g}\qquad(1\leq i\leq d). f i ∣ H f ∩ H g = g i ∣ H f ∩ H g ( 1 ≤ i ≤ d ) .
If the support of either f i f_i f i or g i g_i g i is contained in this common initial part, then f i = g i f_i=g_i f i = g i . Write I = H f ∩ H g I=H_f\cap H_g I = H f ∩ H g and p = ℓ d p=\ell_d p = ℓ d . If supp f i ⊄ I \operatorname{supp} f_i\not\subset I supp f i ⊂ I , then also supp g i ⊄ I \operatorname{supp} g_i\not\subset I supp g i ⊂ I . Moreover,
α ∈ H f ∖ I , β ∈ H g ∖ I , α < β ⟹ ϱ ( α , β ) > p . (35) \alpha\in H_f\setminus I,\quad\beta\in H_g\setminus I,\quad\alpha<\beta\quad\Longrightarrow\quad\varrho(\alpha,\beta)>p. \tag*{(35)} α ∈ H f ∖ I , β ∈ H g ∖ I , α < β ⟹ ϱ ( α , β ) > p . (35)
In particular, this inequality holds for α ∈ supp f i ∖ I \alpha\in\operatorname{supp} f_i\setminus I α ∈ supp f i ∖ I and β ∈ supp g i ∖ I \beta\in\operatorname{supp} g_i\setminus I β ∈ supp g i ∖ I with α < β \alpha<\beta α < β . For every j > i j>i j > i , both supp f j \operatorname{supp} f_j supp f j and supp g j \operatorname{supp} g_j supp g j are entirely outside I I I , so the same inequality holds for all α ∈ supp f j \alpha\in\operatorname{supp} f_j α ∈ supp f j and β ∈ supp g j \beta\in\operatorname{supp} g_j β ∈ supp g j with α < β \alpha<\beta α < β .
Proof. The strict increase of ( m j ) (m_j) ( m j ) and the injectivity of σ \sigma σ give f ^ = g ^ \widehat{f}=\widehat{g} f = g . Put p = ℓ d p=\ell_d p = ℓ d , their common last integer parameter. By (30), both closures are p p p -closed and have ϱ \varrho ϱ -diameter at most p p p . Equation (31) shows that their intersection is an initial part of each, with the same increasing enumeration there. Equality of the transported functionals therefore gives equality on the intersection. If supp f i \operatorname{supp} f_i supp f i lies in that intersection, the transported f i f_i f i vanishes at every subsequent position. The same is true of the transported g i g_i g i , so g i g_i g i has no support outside the intersection. Thus f i = g i f_i=g_i f i = g i ; the other case is symmetric.
If supp f i ⊄ I \operatorname{supp} f_i\not\subset I supp f i ⊂ I but supp g i ⊂ I \operatorname{supp} g_i\subset I supp g i ⊂ I , the preceding conclusion would give f i = g i f_i=g_i f i = g i , a contradiction. To prove (34), suppose that α ∈ H f ∖ I \alpha\in H_f\setminus I α ∈ H f ∖ I , β ∈ H g ∖ I \beta\in H_g\setminus I β ∈ H g ∖ I , α < β \alpha<\beta α < β , and ϱ ( α , β ) ≤ p \varrho(\alpha,\beta)\leq p ϱ ( α , β ) ≤ p . The p p p -closedness of H g H_g H g would imply α ∈ H g \alpha\in H_g α ∈ H g , hence α ∈ I \alpha\in I α ∈ I , again a contradiction. Finally, the maxima of supp f i \operatorname{supp} f_i supp f i and supp g i \operatorname{supp} g_i supp g i lie outside I I I . Since I I I is an initial part of both closures and the functionals are successive, the supports of all later functionals lie outside I I I as well. □ \square □
Equality of codes therefore identifies the ordered finite data, but need not identify the ordinal supports. The common initial part is what turns equality of transported values into equality of the original functionals on that part. The restriction to points outside I I I in (34) is essential: a support may meet both I I I and its complement. If α ∈ I \alpha\in I α ∈ I and β ∈ H g ∖ I \beta\in H_g\setminus I β ∈ H g ∖ I , then α < β \alpha<\beta α < β and ϱ ( α , β ) ≤ p \varrho(\alpha,\beta)\leq p ϱ ( α , β ) ≤ p , since both points belong to H g H_g H g .
Special sequences
We pass now to the definition of special sequences and special functionals. Fix once and for all an index j 0 ∈ N j_0\in\mathbb{N} j 0 ∈ N . Every special sequence will start with the same chosen weight m j 0 m_{j_0} m j 0 .
Definition 4.3 (Special sequences). Let G ⊂ c 00 ( ω 1 , Q ) G\subset c_{00}(\omega_1,\mathbb{Q}) G ⊂ c 00 ( ω 1 , Q ) be a norming set and let ( m j , n j ) j (m_j,n_j)_j ( m j , n j ) j be as above. A finite sequence ( ϕ i ) i = 1 d (\phi_i)_{i=1}^d ( ϕ i ) i = 1 d is a special sequence if the following conditions hold:
(1) The functionals ϕ 1 < ⋯ < ϕ d \phi_1<\cdots<\phi_d ϕ 1 < ⋯ < ϕ d are nonzero successive weighted elements of G G G .
(2) There are strictly increasing sequences ( w i ) i = 1 d (w_i)_{i=1}^d ( w i ) i = 1 d and ( ℓ i ) i = 1 d (\ell_i)_{i=1}^d ( ℓ i ) i = 1 d of positive integers such that w i w_i w i is a chosen weight of ϕ i \phi_i ϕ i , w 1 = m j 0 w_1=m_{j_0} w 1 = m j 0 , and
ℓ i ≥ p ⋃ k = 1 i supp ϕ k \ell_i\geq p\bigcup_{k=1}^{i}\operatorname{supp}\phi_k ℓ i ≥ p k = 1 ⋃ i supp ϕ k for 1 ≤ i ≤ d 1\leq i\leq d 1 ≤ i ≤ d .
(3) For 1 ≤ i < d 1\leq i<d 1 ≤ i < d ,
σ ϱ ( ϕ 1 , w 1 , ℓ 1 , … , ϕ i , w i , ℓ i ) = w i + 1 . \sigma_\varrho(\phi_1,w_1,\ell_1,\ldots,\phi_i,w_i,\ell_i)=w_{i+1}. σ ϱ ( ϕ 1 , w 1 , ℓ 1 , … , ϕ i , w i , ℓ i ) = w i + 1 .
The notation w ( ϕ i ) = w i w(\phi_i)=w_i w ( ϕ i ) = w i refers here to the weight chosen as part of the special sequence, since a weighted functional may admit more than one weight. An infinite sequence ( ϕ i ) i ∈ N (\phi_i)_{i\in\mathbb{N}} ( ϕ i ) i ∈ N is special if there are sequences ( w i ) i ∈ N (w_i)_{i\in\mathbb{N}} ( w i ) i ∈ N and ( ℓ i ) i ∈ N (\ell_i)_{i\in\mathbb{N}} ( ℓ i ) i ∈ N such that every finite initial sequence is special with these same parameters.
Definition 4.4 (Special functional). (a) For a norming set G G G as before, which is also closed in the projection of its elements to the intervals of ω 1 \omega_1 ω 1 for every finite special sequence and every interval I I I , we define the special functional ϕ I \phi_I ϕ I as the restriction to the interval I I I of the functional ϕ = ∑ i = 1 d ϕ i \phi=\sum_{i=1}^{d}\phi_i ϕ = ∑ i = 1 d ϕ i .
(b) For an infinite special sequence ( ϕ i ) i ∈ N (\phi_i)_{i\in\mathbb{N}} ( ϕ i ) i ∈ N we denote by Φ = ∑ i = 1 ∞ ϕ i \Phi=\sum_{i=1}^{\infty}\phi_i Φ = ∑ i = 1 ∞ ϕ i the functional acting on c 00 ( ω 1 ) c_{00}(\omega_1) c 00 ( ω 1 ) by the rule
Φ ( x ) = ∑ i ϕ i ( x ) \Phi(x)=\sum_i\phi_i(x) Φ ( x ) = i ∑ ϕ i ( x ) which is well defined.
If for every d ∈ N d\in\mathbb{N} d ∈ N we have that ∑ i ≤ d ϕ i ∈ G \sum_{i\leq d}\phi_i\in G ∑ i ≤ d ϕ i ∈ G then in the dual of ( c 00 ( ω 1 ) , ∥ ⋅ ∥ G ) (c_{00}(\omega_1),\|\cdot\|_G) ( c 00 ( ω 1 ) , ∥ ⋅ ∥ G ) it holds
∥ Φ ∥ ≤ 1 and Φ = w ∗ − ∑ i = 1 ∞ ϕ i \|\Phi\|\leq1 \quad\text{and}\quad\Phi=w^*-\sum_{i=1}^{\infty}\phi_i ∥Φ∥ ≤ 1 and Φ = w ∗ − i = 1 ∑ ∞ ϕ i For a special functional Φ \Phi Φ and I I I an interval we set Φ I \Phi_I Φ I its restriction on the interval I I I .
(c) Let Φ \Phi Φ be a special functional, resulting from a finite or infinite special sequence ( ϕ i ) i ∈ F (\phi_i)_{i\in F} ( ϕ i ) i ∈ F and I I I an interval. The weight of Φ I \Phi_I Φ I
w ( Φ I ) = { w ( ϕ i ) : supp ( ϕ i ) ∩ I ≠ ∅ } w(\Phi_I)=\{w(\phi_i):\operatorname{supp}(\phi_i)\cap I\ne\varnothing\} w ( Φ I ) = { w ( ϕ i ) : supp ( ϕ i ) ∩ I = ∅ }
Proposition 4.5 (Weights of infinite special sequences). Let ( ϕ i ) i ∈ N (\phi_i)_{i\in\mathbb{N}} ( ϕ i ) i ∈ N and ( ψ i ) i ∈ N (\psi_i)_{i\in\mathbb{N}} ( ψ i ) i ∈ N be two distinct infinite special sequences, with chosen weights a i = w ( ϕ i ) a_i=w(\phi_i) a i = w ( ϕ i ) , b i = w ( ψ i ) b_i=w(\psi_i) b i = w ( ψ i ) and integer parameters ( ℓ i ϕ ) (\ell_i^\phi) ( ℓ i ϕ ) , ( ℓ i ψ ) (\ell_i^\psi) ( ℓ i ψ ) . Then there exists an integer n 0 ≥ 1 n_0\geq1 n 0 ≥ 1 such that
w ( ϕ i ) = w ( ψ i ) ( 1 ≤ i ≤ n 0 ) (36) w(\phi_i)=w(\psi_i) \qquad(1\leq i\leq n_0) \tag*{(36)} w ( ϕ i ) = w ( ψ i ) ( 1 ≤ i ≤ n 0 ) (36) and
{ w ( ϕ i ) : i > n 0 } ∩ { w ( ψ i ) : i > n 0 } = ∅ . (37) \{w(\phi_i):i>n_0\}\cap\{w(\psi_i):i>n_0\}=\varnothing. \tag*{(37)} { w ( ϕ i ) : i > n 0 } ∩ { w ( ψ i ) : i > n 0 } = ∅ . (37) In fact, the intersection of the two full weight sets is exactly { a i : 1 ≤ i ≤ n 0 } \{a_i:1\leq i\leq n_0\} { a i : 1 ≤ i ≤ n 0 } and is therefore finite.
Proof. For n ≥ 1 n\geq1 n ≥ 1 , denote the corresponding finite coded sequences by
Φ n = ( ϕ 1 , a 1 , ℓ 1 ϕ , … , ϕ n , a n , ℓ n ϕ ) , Ψ n = ( ψ 1 , b 1 , ℓ 1 ψ , … , ψ n , b n , ℓ n ψ ) . \Phi_n=(\phi_1,a_1,\ell_1^\phi,\ldots,\phi_n,a_n,\ell_n^\phi),\qquad
\Psi_n=(\psi_1,b_1,\ell_1^\psi,\ldots,\psi_n,b_n,\ell_n^\psi). Φ n = ( ϕ 1 , a 1 , ℓ 1 ϕ , … , ϕ n , a n , ℓ n ϕ ) , Ψ n = ( ψ 1 , b 1 , ℓ 1 ψ , … , ψ n , b n , ℓ n ψ ) . By definition, a 1 = b 1 = m j 0 a_1=b_1=m_{j_0} a 1 = b 1 = m j 0 . Since the weights in each sequence are strictly increasing, this common first weight is strictly smaller than every later weight of either sequence. If i , j ≥ 2 i,j\geq2 i , j ≥ 2 and a i = b j a_i=b_j a i = b j , then
σ ϱ ( Φ i − 1 ) = σ ϱ ( Ψ j − 1 ) . \sigma_\varrho(\Phi_{i-1})=\sigma_\varrho(\Psi_{j-1}). σ ϱ ( Φ i − 1 ) = σ ϱ ( Ψ j − 1 ) . Proposition 4.2 gives i = j i=j i = j , equality of all earlier weights, and ℓ k ϕ = ℓ k ψ \ell_k^\phi=\ell_k^\psi ℓ k ϕ = ℓ k ψ for 1 ≤ k < i 1\leq k<i 1 ≤ k < i . Consequently, every equality a i = b j a_i=b_j a i = b j has i = j i=j i = j , and
M = { n ∈ N : a n = b n } M=\{n\in\mathbb{N}:a_n=b_n\} M = { n ∈ N : a n = b n } is an initial segment of N \mathbb{N} N containing $1.
We show that M M M is finite. Suppose, to the contrary, that M M M is infinite. Choose i i i with ϕ i ≠ ψ i \phi_i \ne\psi_i ϕ i = ψ i and set
α = max supp ϕ i , β = max supp ψ i . \alpha=\max\operatorname{supp}\phi_i,\qquad\beta=\max\operatorname{supp}\psi_i. α = max supp ϕ i , β = max supp ψ i .
For every n ∈ M n\in M n ∈ M with n > i n>i n > i , the codes of Φ n − 1 \mathbf{\Phi}_{n-1} Φ n − 1 and Ψ n − 1 \mathbf{\Psi}_{n-1} Ψ n − 1 coincide. Put
p n = ℓ n − 1 ϕ = ℓ n − 1 ψ , I n = H Φ n − 1 ∩ H Ψ n − 1 . p_n=\ell_{n-1}^{\phi}=\ell_{n-1}^{\psi},\qquad I_n=H_{\mathbf{\Phi}_{n-1}}\cap H_{\mathbf{\Psi}_{n-1}}. p n = ℓ n − 1 ϕ = ℓ n − 1 ψ , I n = H Φ n − 1 ∩ H Ψ n − 1 .
Neither supp ϕ i \operatorname{supp}\phi_i supp ϕ i nor supp ψ i \operatorname{supp}\psi_i supp ψ i can be contained in I n I_n I n , since Proposition 4.2 would then give ϕ i = ψ i \phi_i=\psi_i ϕ i = ψ i . As I n I_n I n is an initial part of both closures, α , β ∉ I n \alpha,\beta\notin I_n α , β ∈ / I n . In particular α ≠ β \alpha\ne\beta α = β , since otherwise their common value would belong to both closures. Interchanging the two sequences if necessary, assume α < β \alpha<\beta α < β . The separation assertion (34) gives
ρ ( α , β ) > p n = ℓ n − 1 ϕ ( n ∈ M , n > i ) . \rho(\alpha,\beta)>p_n=\ell_{n-1}^{\phi}\qquad(n\in M,\ n>i). ρ ( α , β ) > p n = ℓ n − 1 ϕ ( n ∈ M , n > i ) .
This is impossible: the left-hand side is fixed, whereas the strictly increasing integer sequence ( ℓ n ϕ ) n (\ell_n^\phi)_n ( ℓ n ϕ ) n tends to infinity. Thus M M M is finite. It follows that M = { 1 , … , n 0 } M=\{1,\ldots,n_0\} M = { 1 , … , n 0 } for some n 0 ≥ 1 n_0\ge1 n 0 ≥ 1 . Since equal weights can occur only at the same index, the intersection of the full weight sets is exactly { a i : 1 ≤ i ≤ n 0 } \{a_i:1\le i\le n_0\} { a i : 1 ≤ i ≤ n 0 } . Both assertions follow. □
The norming set G p , r G_{p,r} G p , r
We now define the norming set of E p , r E_{p,r} E p , r using the operations of Definition 4.1 and the special sequences and functionals of Definitions 4.3 and 4.4 .
Definition 4.6. Let 1 < r < p < ∞ 1<r<p<\infty 1 < r < p < ∞ . The norming set G p , r G_{p,r} G p , r is the smallest subset of c 00 ( ω 1 , Q ) c_{00}(\omega_1,\mathbb{Q}) c 00 ( ω 1 , Q ) with the following properties.
(1) It contains e α ∗ e_\alpha^* e α ∗ for every α < ω 1 \alpha<\omega_1 α < ω 1 , is symmetric, and is closed under restrictions to ordinal intervals:
f ∈ G p , r ⟹ − f ∈ G p , r and f ∣ I ∈ G p , r ( I ⊂ ω 1 an interval ) . f\in G_{p,r}\Longrightarrow-f\in G_{p,r}\quad\text{and}\quad f|_I\in G_{p,r}\quad(I\subset\omega_1\text{ an interval}). f ∈ G p , r ⟹ − f ∈ G p , r and f ∣ I ∈ G p , r ( I ⊂ ω 1 an interval ) . (2) It is closed under finite rational convex combinations.
(3) It is closed under the ( p , j ) (p,j) ( p , j ) -operation for every j ∈ N j\in\mathbb{N} j ∈ N , as defined in Definition 4.1 .
(4) If ( ϕ i ) i = 1 d (\phi_i)_{i=1}^d ( ϕ i ) i = 1 d is a finite special sequence in G p , r G_{p,r} G p , r , in the sense of Definition 4.3 , then
Φ = ∑ i = 1 d ϕ i ∈ G p , r . \Phi=\sum_{i=1}^{d}\phi_i\in G_{p,r}. Φ = i = 1 ∑ d ϕ i ∈ G p , r . (5) Let
Φ i = ∑ k = 1 d i ϕ i , k ∈ G p , r ( 1 ≤ i ≤ d ) , \Phi_i=\sum_{k=1}^{d_i}\phi_{i,k}\in G_{p,r}\qquad(1\le i\le d), Φ i = k = 1 ∑ d i ϕ i , k ∈ G p , r ( 1 ≤ i ≤ d ) , where each ( ϕ i , k ) k = 1 d i (\phi_{i,k})_{k=1}^{d_i} ( ϕ i , k ) k = 1 d i is a specified finite special sequence in G p , r G_{p,r} G p , r with its chosen weights. Let I 1 , … , I d I_1,\ldots,I_d I 1 , … , I d be ordinal intervals. Suppose that the weight sets of their restrictions, as defined in Definition 4.4 (c), are pairwise disjoint:
w ( Φ i ∣ I i ) ∩ w ( Φ j ∣ I j ) = ∅ ( i ≠ j ) . w(\Phi_i|_{I_i})\cap w(\Phi_j|_{I_j})=\varnothing\qquad(i\ne j). w ( Φ i ∣ I i ) ∩ w ( Φ j ∣ I j ) = ∅ ( i = j ) . Then, for every ( λ i ) i = 1 d ∈ Q d (\lambda_i)_{i=1}^d\in\mathbb{Q}^d ( λ i ) i = 1 d ∈ Q d satisfying ∑ i = 1 d ∣ λ i ∣ r ≤ 1 \sum_{i=1}^{d}|\lambda_i|^r\le1 ∑ i = 1 d ∣ λ i ∣ r ≤ 1 ,
∑ i = 1 d λ i Φ i ∣ I i ∈ G p , r . \sum_{i=1}^{d}\lambda_i\Phi_i|_{I_i}\in G_{p,r}. i = 1 ∑ d λ i Φ i ∣ I i ∈ G p , r .
In (5), each weight set is computed from the specified special decomposition and its chosen weights.
The corresponding norm and space are therefore
∥ x ∥ p , r = sup { ∣ f ( x ) ∣ : f ∈ G p , r } , E p , r = ( c 00 ( ω 1 ) , ∥ ⋅ ∥ p , r ) ‾ . \|x\|_{p,r}=\sup\{|f(x)|:f\in G_{p,r}\},\qquad E_{p,r}=\overline{(c_{00}(\omega_1),\|\cdot\|_{p,r})}. ∥ x ∥ p , r = sup { ∣ f ( x ) ∣ : f ∈ G p , r } , E p , r = ( c 00 ( ω 1 ) , ∥ ⋅ ∥ p , r ) .
where the bar denotes completion. Since G p , r G_{p,r} G p , r contains the biorthogonal functionals and is closed under interval restrictions, ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 is a normalized bimonotone transfinite basis of E p , r E_{p,r} E p , r .
uniform bound, the partial sums converge pointwise on E p , r E_{p,r} E p , r and define
Φ = w ∗ ∑ i = 1 ∞ ϕ i ∈ B E p , r ∗ . \Phi= w^*\sum_{i=1}^{\infty}\phi_i \in B_{E_{p,r}^*}. Φ = w ∗ i = 1 ∑ ∞ ϕ i ∈ B E p , r ∗ .
Thus the conclusion supplied by the definition is ∥ Φ ∥ ≤ 1 \|\Phi\| \leq1 ∥Φ∥ ≤ 1 . Equality does not hold for every infinite special sequence. For example, choose successive singletons and construct the coding parameters recursively so that
ϕ i = 2 − i w i e a i ∗ , w 1 = m j 0 , w i + 1 = σ ϱ ( ϕ 1 , w 1 , ℓ 1 , … , ϕ i , w i , ℓ i ) . \phi_i=\frac{2^{-i}}{w_i}e_{a_i}^*,\qquad w_1=m_{j_0},\qquad w_{i+1}=\sigma_\varrho(\phi_1,w_1,\ell_1,\ldots,\phi_i,w_i,\ell_i). ϕ i = w i 2 − i e a i ∗ , w 1 = m j 0 , w i + 1 = σ ϱ ( ϕ 1 , w 1 , ℓ 1 , … , ϕ i , w i , ℓ i ) .
Each ϕ i \phi_i ϕ i is a weighted functional of weight w i w_i w i , obtained by using just one nonzero coefficient in the corresponding ( p , j ) (p,j) ( p , j ) -operation. Then
∑ i ∥ ϕ i ∥ = ∑ i 2 − i w i ≤ 1 m j 0 < 1 , \sum_i\|\phi_i\|=\sum_i\frac{2^{-i}}{w_i}\leq\frac{1}{m_{j_0}}<1, i ∑ ∥ ϕ i ∥ = i ∑ w i 2 − i ≤ m j 0 1 < 1 ,
so this special functional belongs to X 0 p , r X_0^{p,r} X 0 p , r and has norm strictly less than one.
We take Φ p , r = { ϕ γ : γ ∈ Γ } \Phi_{p,r}=\{\phi_\gamma:\gamma\in\Gamma\} Φ p , r = { ϕ γ : γ ∈ Γ } to be the family of all distinct infinite special functionals outside X 0 p , r X_0^{p,r} X 0 p , r , fixing one special expansion ( ϕ i , γ ) i (\phi_{i,\gamma})_i ( ϕ i , γ ) i for each member. The nontriviality, uncountability, and norm-span properties of this family are part of the structural assertions in Section 3; here we verify its tail estimate directly from the norming set.
Fix a finite F ⊂ Γ F\subset\Gamma F ⊂ Γ . By Proposition 4.5, there is n 0 n_0 n 0 such that the weight sets of the tails ( ϕ i , γ ) i > n 0 , γ ∈ F (\phi_{i,\gamma})_{i>n_0,\ \gamma\in F} ( ϕ i , γ ) i > n 0 , γ ∈ F are pairwise disjoint. For N > n > n 0 N>n>n_0 N > n > n 0 , each sum
∑ i = n + 1 N ϕ i , γ \sum_{i=n+1}^{N}\phi_{i,\gamma} i = n + 1 ∑ N ϕ i , γ
is an interval restriction of the finite special functional ∑ i = 1 N ϕ i , γ \sum_{i=1}^{N}\phi_{i,\gamma} ∑ i = 1 N ϕ i , γ . Hence condition (5) of Definition 4.6 gives, for rational coefficients with ∑ γ ∈ F ∣ a γ ∣ r ≤ 1 \sum_{\gamma\in F}|a_\gamma|^r\leq1 ∑ γ ∈ F ∣ a γ ∣ r ≤ 1 ,
∥ ∑ γ ∈ F a γ ∑ i = n + 1 N ϕ i , γ ∥ ≤ 1. \left\|\sum_{\gamma\in F}a_\gamma\sum_{i=n+1}^{N}\phi_{i,\gamma}\right\|\leq1. γ ∈ F ∑ a γ i = n + 1 ∑ N ϕ i , γ ≤ 1.
Letting N → ∞ N\to\infty N → ∞ pointwise on E p , r E_{p,r} E p , r and then using density of rational coefficient families in the real ℓ r \ell_r ℓ r unit ball yields
∥ ∑ γ ∈ F a γ ϕ > n , γ ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ r ) 1 / r ( n > n 0 ) . \left\|\sum_{\gamma\in F}a_\gamma\phi_{>n,\gamma}\right\|\leq\left(\sum_{\gamma\in F}|a_\gamma|^r\right)^{1/r}\qquad(n>n_0). γ ∈ F ∑ a γ ϕ > n , γ ≤ γ ∈ F ∑ ∣ a γ ∣ r 1/ r ( n > n 0 ) .
This proves (5). Since the removed initial sums are finitely supported and belong to X 0 p , r X_0^{p,r} X 0 p , r , the same estimate passes to the quotient and gives (6). The upper ℓ r \ell_r ℓ r estimate thus applies to these tails and to the quotient classes.
Remark 4.9. The construction in [15 Section 5.2] uses averaging operations different from the ( p , j ) (p,j) ( p , j ) -operations used here. The comparison concerns the structural properties used in the operator argument. Our presentation describes the component spaces through the ( p , j ) (p,j) ( p , j ) -operations and then combines them to obtain E E E , as in § 4.3 .
Two properties of E p , r E_{p,r} E p , r
We prove that the basis of E p , r E_{p,r} E p , r is boundedly complete and has an upper ℓ p ′ \ell_{p'} ℓ p ′ estimate with constant one, where p ′ = p / ( p − 1 ) p' = p/(p - 1) p ′ = p / ( p − 1 ) . We first recall the transfinite version of bounded completeness from [6 Definition 1.5] .
Definition 4.10. A transfinite Schauder basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 is boundedly complete if every sequence ( e α n ) n ∈ N (e_{\alpha_n})_{n\in\mathbb{N}} ( e α n ) n ∈ N , with ( α n ) n (\alpha_n)_n ( α n ) n strictly increasing, is boundedly complete in the usual sense. That is, for every scalar sequence ( a n ) n (a_n)_n ( a n ) n ,
sup N ∥ ∑ n = 1 N a n e α n ∥ < ∞ ⟹ ∑ n = 1 ∞ a n e α n converges in norm . \sup_N\left\|\sum_{n=1}^{N}a_ne_{\alpha_n}\right\|<\infty\quad\Longrightarrow\quad\sum_{n=1}^{\infty}a_ne_{\alpha_n}\text{ converges in norm}. N sup n = 1 ∑ N a n e α n < ∞ ⟹ n = 1 ∑ ∞ a n e α n converges in norm .
The next proposition adapts the bounded-completeness argument in the proof of [6 Proposition 1.6] . We state explicitly the norming hypothesis that gives an isometry; it holds for the spaces defined here by finitely supported norming functionals.
Proposition 4.11 (The canonical predual). Let ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 be a bimonotone, boundedly complete transfinite basis of a Banach space Z Z Z , and put
Z 0 = span ‾ { e α ∗ : α < ω 1 } ⊂ Z ∗ . Z_0=\overline{\operatorname{span}}\{e_\alpha^*:\alpha<\omega_1\}\subset Z^*. Z 0 = span { e α ∗ : α < ω 1 } ⊂ Z ∗ . Assume that Z 0 Z_0 Z 0 is 1 1 1 -norming for Z Z Z , namely,
∥ z ∥ = sup { ∣ h ( z ) ∣ : h ∈ Z 0 , ∥ h ∥ ≤ 1 } ( z ∈ Z ) . \|z\|=\sup\{|h(z)|:h\in Z_0,\ \|h\|\leq1\}\qquad(z\in Z). ∥ z ∥ = sup { ∣ h ( z ) ∣ : h ∈ Z 0 , ∥ h ∥ ≤ 1 } ( z ∈ Z ) . Then the restriction map
R : Z ⟶ Z 0 ∗ , ( R z ) ( h ) = h ( z ) , R:Z\longrightarrow Z_0^*,\qquad(Rz)(h)=h(z), R : Z ⟶ Z 0 ∗ , ( R z ) ( h ) = h ( z ) , is an onto isometry.
Proof. The norming hypothesis makes R R R an isometry, so R ( Z ) R(Z) R ( Z ) is norm closed. For 0 ≤ β ≤ ω 1 0\leq\beta\leq\omega_1 0 ≤ β ≤ ω 1 , let P β = P [ 0 , β ) P_\beta=P_{[0,\beta)} P β = P [ 0 , β ) be the initial-interval projection on Z Z Z , and define
T β = P β ∗ ∣ Z 0 , P ^ β = T β ∗ : Z 0 ∗ ⟶ Z 0 ∗ . T_\beta=P_\beta^*|_{Z_0},\qquad\widehat P_\beta=T_\beta^*:Z_0^*\longrightarrow Z_0^*. T β = P β ∗ ∣ Z 0 , P β = T β ∗ : Z 0 ∗ ⟶ Z 0 ∗ .
The maps are well defined and contractive, since
T β e α ∗ = { e α ∗ , α < β , 0 , α ≥ β . T_\beta e_\alpha^*=
\begin{cases}
e_\alpha^*, & \alpha<\beta,\\
0, & \alpha\geq\beta.
\end{cases} T β e α ∗ = { e α ∗ , 0 , α < β , α ≥ β .
They satisfy P ^ β R = R P β \widehat P_\beta R=RP_\beta P β R = R P β . If λ \lambda λ is a limit ordinal, density of the finite linear combinations of the e α ∗ e_\alpha^* e α ∗ in Z 0 Z_0 Z 0 gives
T β h ⟶ T λ h in norm ( h ∈ Z 0 , β ↑ λ ) . (38) T_\beta h\longrightarrow T_\lambda h\text{ in norm}\qquad(h\in Z_0,\ \beta\uparrow\lambda). \tag*{(38)} T β h ⟶ T λ h in norm ( h ∈ Z 0 , β ↑ λ ) . (38)
Fix u ∈ Z 0 ∗ u\in Z_0^* u ∈ Z 0 ∗ and put M = ∥ u ∥ M=\|u\| M = ∥ u ∥ . We prove by transfinite induction that P ^ β u ∈ R ( Z ) \widehat P_\beta u\in R(Z) P β u ∈ R ( Z ) for every β ≤ ω 1 \beta\leq\omega_1 β ≤ ω 1 . The assertion is immediate for β = 0 \beta=0 β = 0 . At a successor step use
P ^ β + 1 u − P ^ β u = u ( e β ∗ ) R e β . \widehat P_{\beta+1}u-\widehat P_\beta u=u(e_\beta^*)Re_\beta. P β + 1 u − P β u = u ( e β ∗ ) R e β .
Suppose now that λ \lambda λ is a limit ordinal and the assertion holds below λ \lambda λ . Write
R z β = P ^ β u ( β < λ ) . Rz_\beta=\widehat P_\beta u\qquad(\beta<\lambda). R z β = P β u ( β < λ ) .
Then ∥ z β ∥ ≤ M \|z_\beta\|\leq M ∥ z β ∥ ≤ M and P β z γ = z β P_\beta z_\gamma=z_\beta P β z γ = z β whenever β < γ < λ \beta<\gamma<\lambda β < γ < λ . We claim that the net ( z β ) β < λ (z_\beta)_{\beta<\lambda} ( z β ) β < λ is norm Cauchy. Otherwise, choose ε > 0 \varepsilon>0 ε > 0 and β 1 < β 2 < ⋯ < λ \beta_1<\beta_2<\cdots<\lambda β 1 < β 2 < ⋯ < λ with
∥ z β n + 1 − z β n ∥ ≥ ε ( n ∈ N ) . \lVert z_{\beta_{n+1}}-z_{\beta_n}\rVert\ge\varepsilon\qquad(n\in\mathbb{N}). ∥ z β n + 1 − z β n ∥ ≥ ε ( n ∈ N ) .
Each difference is supported on [ β n , β n + 1 ) [\beta_n,\beta_{n+1}) [ β n , β n + 1 ) . Choose a finitely supported y n y_n y n in this interval such that
∥ y n − ( z β n + 1 − z β n ) ∥ < ε 2 − n − 2 . \lVert y_n-(z_{\beta_{n+1}}-z_{\beta_n})\rVert<\varepsilon2^{-n-2}. ∥ y n − ( z β n + 1 − z β n )∥ < ε 2 − n − 2 .
The supports of the y n y_n y n are successive, ∥ y n ∥ ≥ ε / 2 \lVert y_n\rVert\ge\varepsilon/2 ∥ y n ∥ ≥ ε /2 , and telescoping gives
sup N ∥ ∑ n = 1 N y n ∥ ≤ 2 M + ε / 4. \sup_N\left\lVert\sum_{n=1}^{N}y_n\right\rVert\le2M+\varepsilon/4. N sup n = 1 ∑ N y n ≤ 2 M + ε /4.
Enumerate ⋃ n supp y n \bigcup_n\operatorname{supp}y_n ⋃ n supp y n increasingly as ( α k ) k (\alpha_k)_k ( α k ) k . Every partial sum of the resulting scalar series in ( e α k ) k (e_{\alpha_k})_k ( e α k ) k is an initial-interval projection of some ∑ n = 1 N y n \sum_{n=1}^{N}y_n ∑ n = 1 N y n and therefore has the same uniform bound. Bounded completeness implies that this scalar series converges. Its successive block sums y n y_n y n must then tend to zero in norm, contradicting ∥ y n ∥ ≥ ε / 2 \lVert y_n\rVert\ge\varepsilon/2 ∥ y n ∥ ≥ ε /2 .
Thus z β → z λ z_\beta\to z_\lambda z β → z λ in norm for some z λ ∈ Z z_\lambda\in Z z λ ∈ Z . For every h ∈ Z 0 h\in Z_0 h ∈ Z 0 , equation (38) gives
( R z λ ) ( h ) = lim β ↑ λ u ( T β h ) = u ( T λ h ) = ( P ^ λ u ) ( h ) . (Rz_\lambda)(h)=\lim_{\beta\uparrow\lambda}u(T_\beta h)=u(T_\lambda h)=(\widehat P_\lambda u)(h). ( R z λ ) ( h ) = β ↑ λ lim u ( T β h ) = u ( T λ h ) = ( P λ u ) ( h ) .
This completes the induction. Since P ^ ω 1 \widehat P_{\omega_1} P ω 1 is the identity on Z 0 ∗ Z_0^* Z 0 ∗ , we obtain u ∈ R ( Z ) u\in R(Z) u ∈ R ( Z ) , as required.
Proposition 4.12. Under the parameter growth condition (30) , the basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 of E p , r E_{p,r} E p , r is boundedly complete.
Proof. Let ( α n ) n (\alpha_n)_n ( α n ) n be strictly increasing and suppose
M : = sup N ∥ ∑ n = 1 N a n e α n ∥ p , r < ∞ . M:=\sup_N\left\lVert\sum_{n=1}^{N}a_ne_{\alpha_n}\right\rVert_{p,r}<\infty. M := N sup n = 1 ∑ N a n e α n p , r < ∞. If the series does not converge in norm, there are ε > 0 \varepsilon>0 ε > 0 and 0 = q 0 < q 1 < q 2 < ⋯ 0=q_0<q_1<q_2<\cdots 0 = q 0 < q 1 < q 2 < ⋯ such that
u k = ∑ n = q k − 1 + 1 q k a n e α n , ∥ u k ∥ p , r ≥ ε ( k ∈ N ) . u_k=\sum_{n=q_{k-1}+1}^{q_k}a_ne_{\alpha_n},\qquad\lVert u_k\rVert_{p,r}\ge\varepsilon\qquad(k\in\mathbb{N}). u k = n = q k − 1 + 1 ∑ q k a n e α n , ∥ u k ∥ p , r ≥ ε ( k ∈ N ) . Choose f k ∈ G p , r f_k\in G_{p,r} f k ∈ G p , r , supported in the ordinal range of u k u_k u k , with f k ( u k ) > ε / 2 f_k(u_k)>\varepsilon/2 f k ( u k ) > ε /2 . This is possible because G p , r G_{p,r} G p , r norms the space, is symmetric, and is closed under interval restrictions. The functionals f k f_k f k are nonzero and successive.
Fix j j j and choose a rational t j t_j t j with
1 2 n j − 1 / p ≤ t j ≤ n j − 1 / p . \frac{1}{2}n_j^{-1/p}\le t_j\le n_j^{-1/p}. 2 1 n j − 1/ p ≤ t j ≤ n j − 1/ p . Since n j t j p ≤ 1 n_jt_j^p\le1 n j t j p ≤ 1 , closure under the ( p , j ) (p,j) ( p , j ) -operation gives
f = t j m j ∑ k = 1 n j f k ∈ G p , r . f=\frac{t_j}{m_j}\sum_{k=1}^{n_j}f_k\in G_{p,r}. f = m j t j k = 1 ∑ n j f k ∈ G p , r . Consequently,
M ≥ ∥ ∑ k = 1 n j u k ∥ p , r ≥ f ( ∑ k = 1 n j u k ) > ε 4 n j 1 − 1 / p m j = ε 4 n j 1 / p ′ m j . M\ge\left\lVert\sum_{k=1}^{n_j}u_k\right\rVert_{p,r}\ge f\left(\sum_{k=1}^{n_j}u_k\right)>\frac{\varepsilon}{4}\frac{n_j^{1-1/p}}{m_j}=\frac{\varepsilon}{4}\frac{n_j^{1/p'}}{m_j}. M ≥ k = 1 ∑ n j u k p , r ≥ f ( k = 1 ∑ n j u k ) > 4 ε m j n j 1 − 1/ p = 4 ε m j n j 1/ p ′ .
The last expression tends to infinity by (30) , a contradiction. Hence the series converges, proving bounded completeness. □
Since G p , r ⊂ B X 0 p , r G_{p,r} \subset B_{X_0^{p,r}} G p , r ⊂ B X 0 p , r and G p , r G_{p,r} G p , r norms E p , r E_{p,r} E p , r , the space X 0 p , r X_0^{p,r} X 0 p , r is 1-norming. Proposition 4.11 therefore gives the canonical onto isometry
E p , r ⟶ ( X 0 p , r ) ∗ , x ⟼ ( h ⟼ h ( x ) ) . E_{p,r} \longrightarrow(X_0^{p,r})^*, \qquad x \longmapsto(h \longmapsto h(x)). E p , r ⟶ ( X 0 p , r ) ∗ , x ⟼ ( h ⟼ h ( x )) .
This proves property (3) from the norming-set construction and bounded completeness. We next establish the estimate on the scalar families of the norming functionals.
Lemma 4.13. For every f ∈ G p , r f \in G_{p,r} f ∈ G p , r ,
∥ f ∥ ℓ p ( ω 1 ) ≤ 1. \|f\|_{\ell_p(\omega_1)} \le1. ∥ f ∥ ℓ p ( ω 1 ) ≤ 1. More precisely, if f f f has chosen weight m j m_j m j , then ∥ f ∥ ℓ p ( ω 1 ) ≤ 1 / m j \|f\|_{\ell_p(\omega_1)} \le1/m_j ∥ f ∥ ℓ p ( ω 1 ) ≤ 1/ m j .
Proof. We use the increasing stages in Remark 4.7. The bound holds for 0 0 0 and the signed singletons. It is preserved by symmetry, interval restrictions, and rational convex combinations. Suppose that all functionals in the data for the next operation have ℓ p \ell_p ℓ p norm at most one.
For a ( p , j ) (p,j) ( p , j ) -operation, the supports of the children f k f_k f k are pairwise disjoint. Hence
∥ 1 m j ∑ k = 1 n j λ k f k ∥ p p = 1 m j p ∑ k = 1 n j ∣ λ k ∣ p ∥ f k ∥ p p ≤ 1 m j p . \left\|\frac{1}{m_j}\sum_{k=1}^{n_j}\lambda_k f_k\right\|_p^p
= \frac{1}{m_j^p}\sum_{k=1}^{n_j}|\lambda_k|^p\|f_k\|_p^p
\le\frac{1}{m_j^p}. m j 1 k = 1 ∑ n j λ k f k p p = m j p 1 k = 1 ∑ n j ∣ λ k ∣ p ∥ f k ∥ p p ≤ m j p 1 .
The same calculation bounds a weighted functional by the reciprocal of its chosen weight, using its specified representation.
For a finite special sum Φ = ∑ i = 1 d φ i \Phi=\sum_{i=1}^{d}\varphi_i Φ = ∑ i = 1 d φ i , the chosen weights w i = w ( φ i ) w_i=w(\varphi_i) w i = w ( φ i ) are distinct. Thus
∥ Φ ∥ p ≤ ∑ i = 1 d ∥ φ i ∥ p ≤ ∑ i = 1 d 1 w i ≤ ∑ j = 1 ∞ 1 m j < 1. \|\Phi\|_p \le\sum_{i=1}^{d}\|\varphi_i\|_p
\le\sum_{i=1}^{d}\frac{1}{w_i}
\le\sum_{j=1}^{\infty}\frac{1}{m_j}<1. ∥Φ ∥ p ≤ i = 1 ∑ d ∥ φ i ∥ p ≤ i = 1 ∑ d w i 1 ≤ j = 1 ∑ ∞ m j 1 < 1.
Finally consider rule (5) in Definition 4.6, and set A i = w ( Φ i ∣ I i ) A_i=w(\Phi_i|_{I_i}) A i = w ( Φ i ∣ I i ) . For the chosen special decomposition,
∥ Φ i ∣ I i ∥ p ≤ ∑ w ∈ A i 1 w . \|\Phi_i|_{I_i}\|_p \le\sum_{w\in A_i}\frac{1}{w}. ∥ Φ i ∣ I i ∥ p ≤ w ∈ A i ∑ w 1 .
Since the A i A_i A i are pairwise disjoint and ∣ λ i ∣ ≤ 1 |\lambda_i|\le1 ∣ λ i ∣ ≤ 1 ,
∥ ∑ i = 1 d λ i Φ i ∣ I i ∥ p ≤ ∑ i = 1 d ∣ λ i ∣ ∑ w ∈ A i 1 w ≤ ∑ j = 1 ∞ 1 m j < 1. \left\|\sum_{i=1}^{d}\lambda_i\Phi_i|_{I_i}\right\|_p
\le\sum_{i=1}^{d}|\lambda_i|\sum_{w\in A_i}\frac{1}{w}
\le\sum_{j=1}^{\infty}\frac{1}{m_j}<1. i = 1 ∑ d λ i Φ i ∣ I i p ≤ i = 1 ∑ d ∣ λ i ∣ w ∈ A i ∑ w 1 ≤ j = 1 ∑ ∞ m j 1 < 1.
This completes the induction and proves the first assertion. Applying the displayed ( p , j ) (p,j) ( p , j ) calculation to any chosen weighted representation, whose children now satisfy the first assertion, proves the final assertion as well. □
Corollary 4.14 (Upper ℓ p ′ \ell_{p'} ℓ p ′ estimate). Let p ′ = p / ( p − 1 ) p'=p/(p-1) p ′ = p / ( p − 1 ) . The basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 of E p , r E_{p,r} E p , r admits an upper ℓ p ′ \ell_{p'} ℓ p ′ estimate with constant one: for every finitely supported scalar family ( a α ) (a_\alpha) ( a α ) ,
∥ ∑ α < ω 1 a α e α ∥ p , r ≤ ( ∑ α < ω 1 ∣ a α ∣ p ′ ) 1 / p ′ . \left\|\sum_{\alpha<\omega_1}a_\alpha e_\alpha\right\|_{p,r}
\le\left(\sum_{\alpha<\omega_1}|a_\alpha|^{p'}\right)^{1/p'}. α < ω 1 ∑ a α e α p , r ≤ ( α < ω 1 ∑ ∣ a α ∣ p ′ ) 1/ p ′ .
Proof. For f ∈ G p , r f \in G_{p,r} f ∈ G p , r , Hölder’s inequality and Lemma 4.13 give
∣ f ( ∑ α a α e α ) ∣ ≤ ∥ f ∥ ℓ p ( ω 1 ) ∥ ( a α ) α ∥ ℓ p ′ ( ω 1 ) ≤ ∥ ( a α ) α ∥ ℓ p ′ ( ω 1 ) . \left|f\left(\sum_{\alpha}a_{\alpha}e_{\alpha}\right)\right|\leq\|f\|_{\ell_p(\omega_1)}\|(a_{\alpha})_{\alpha}\|_{\ell_{p'}(\omega_1)}\leq\|(a_{\alpha})_{\alpha}\|_{\ell_{p'}(\omega_1)}. f ( α ∑ a α e α ) ≤ ∥ f ∥ ℓ p ( ω 1 ) ∥ ( a α ) α ∥ ℓ p ′ ( ω 1 ) ≤ ∥ ( a α ) α ∥ ℓ p ′ ( ω 1 ) .
Taking the supremum over f ∈ G p , r f \in G_{p,r} f ∈ G p , r proves the assertion. □
Theorem 4.15 (Summary of the properties of E p , r E_{p,r} E p , r ). Let 1 < r < p < ∞ 1<r<p<\infty 1 < r < p < ∞ , let p ′ = p / ( p − 1 ) p'=p/(p-1) p ′ = p / ( p − 1 ) , and let E p , r E_{p,r} E p , r be the space defined by G p , r G_{p,r} G p , r . Write Φ p , r = { ϕ γ : γ ∈ Γ } \Phi_{p,r}=\{\phi_{\gamma}:\gamma\in\Gamma\} Φ p , r = { ϕ γ : γ ∈ Γ } for the family of infinite special functionals outside X 0 p , r X_0^{p,r} X 0 p , r , as in Remark 4.8 , and let Q p , r : E p , r ∗ → E p , r ∗ / X 0 p , r Q_{p,r}:E_{p,r}^{*}\to E_{p,r}^{*}/X_0^{p,r} Q p , r : E p , r ∗ → E p , r ∗ / X 0 p , r be the quotient map. Then the following hold.
(a) The basis ( e α ) α < ω 1 (e_{\alpha})_{\alpha<\omega_1} ( e α ) α < ω 1 admits an upper ℓ p ′ \ell_{p'} ℓ p ′ estimate with constant one: for every finitely supported scalar family ( a α ) (a_{\alpha}) ( a α ) ,
∥ ∑ α a α e α ∥ p , r ≤ ( ∑ α ∣ a α ∣ p ′ ) 1 / p ′ . \left\|\sum_{\alpha}a_{\alpha}e_{\alpha}\right\|_{p,r}\leq\left(\sum_{\alpha}|a_{\alpha}|^{p'}\right)^{1/p'}. α ∑ a α e α p , r ≤ ( α ∑ ∣ a α ∣ p ′ ) 1/ p ′ . (b) For every normalized infinite block sequence ( x n ) n (x_n)_n ( x n ) n of the transfinite basis, the space span ‾ { x n : n ∈ N } \overline{\operatorname{span}}\{x_n:n\in\mathbb{N}\} span { x n : n ∈ N } has nonseparable dual.
(c) The biorthogonal functionals and the special functionals generate the dual in norm:
E p , r ∗ = span ‾ ( { e α ∗ : α < ω 1 } ∪ Φ p , r ) . E_{p,r}^{*}=\overline{\operatorname{span}}\left(\{e_{\alpha}^{*}:\alpha<\omega_1\}\cup\Phi_{p,r}\right). E p , r ∗ = span ( { e α ∗ : α < ω 1 } ∪ Φ p , r ) . (d) The family ( Q p , r ϕ γ ) γ ∈ Γ (Q_{p,r}\phi_{\gamma})_{\gamma\in\Gamma} ( Q p , r ϕ γ ) γ ∈ Γ admits an upper ℓ r \ell_r ℓ r estimate with constant one:
∥ ∑ γ ∈ F a γ Q p , r ϕ γ ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ r ) 1 / r \left\|\sum_{\gamma\in F}a_{\gamma}Q_{p,r}\phi_{\gamma}\right\|\leq\left(\sum_{\gamma\in F}|a_{\gamma}|^r\right)^{1/r} γ ∈ F ∑ a γ Q p , r ϕ γ ≤ γ ∈ F ∑ ∣ a γ ∣ r 1/ r for every finite F ⊂ Γ F\subset\Gamma F ⊂ Γ and every scalar family ( a γ ) γ ∈ F (a_{\gamma})_{\gamma\in F} ( a γ ) γ ∈ F . Its linear span is norm dense in E p , r ∗ / X 0 p , r E_{p,r}^{*}/X_0^{p,r} E p , r ∗ / X 0 p , r .
Proof of the estimates. Assertion (a) is Corollary 4.14. The estimate in (d) follows from the tail estimate in Remark 4.8 : removing a finite initial sum does not change the quotient class. The density assertion in (d) follows from (c), since Q p , r Q_{p,r} Q p , r annihilates the biorthogonal functionals.
Assertions (b) and (c) are the structural conclusions whose proofs require the techniques of hereditarily indecomposable and James-tree constructions. For the relevant methods, see the Gowers–Maurey construction [12 ] , the Gowers tree space [11 ] , the work of Petsoulas and Raikoftsalis [16 ] , the attractors method [1 ] , the asymptotic constructions of Deliyanni and Manoussakis [9 ] , and the transfinite methods of [6 ] . Their application to the present norming set requires lengthy technical arguments, whose presentation is outside the purpose of this paper. We state (b) and (c) here without including those proofs. □
Corollary 4.16. Every infinite-dimensional closed subspace of E p , r E_{p,r} E p , r has nonseparable dual.
Proof. Let Y ⊂ E p , r Y\subset E_{p,r} Y ⊂ E p , r be an infinite-dimensional closed subspace. By [6 Proposition 1.3(3)] , there is an infinite-dimensional closed subspace Z ⊂ Y Z\subset Y Z ⊂ Y isomorphic to a block subspace U U U of E p , r E_{p,r} E p , r . Normalizing its generating block sequence does not change U U U . Thus U ∗ U^{*} U ∗ , and hence Z ∗ Z^{*} Z ∗ , is nonseparable by Theorem 4.15 (b). The restriction map Y ∗ → Z ∗ Y^{*}\to Z^{*} Y ∗ → Z ∗ is onto by
the Hahn–Banach theorem. A continuous image of a separable space is separable, so Y ∗ Y^* Y ∗ cannot be separable. □
The norming set G G G of the combined space
We now define the norming set G = G E G=G_E G = G E of the space E = ( ⨁ k E k ) 2 , s E=\left(\bigoplus_k E_k\right)_{2,s} E = ( ⨁ k E k ) 2 , s from Section 3.2. Fix the parameters
2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1 2>s>p_1>r_1>p_2>r_2>\cdots>1 2 > s > p 1 > r 1 > p 2 > r 2 > ⋯ > 1
and a partition ( C k ) k ≥ 1 (C_k)_{k\geq1} ( C k ) k ≥ 1 of ω 1 \omega_1 ω 1 into pairwise disjoint unbounded sets. We use the same sequences ( m j , n j ) j (m_j,n_j)_j ( m j , n j ) j , coding σ ϱ \sigma_\varrho σ ϱ , and first weight m j 0 m_{j_0} m j 0 as above. For every k k k , put
E k = E p k , r k ( C k ) , G k = G p k , r k ( C k ) , E_k=E_{p_k,r_k}(C_k),\qquad G_k=G_{p_k,r_k}(C_k), E k = E p k , r k ( C k ) , G k = G p k , r k ( C k ) ,
where G k G_k G k is obtained by transporting G p k , r k G_{p_k,r_k} G p k , r k along the increasing bijection from ω 1 \omega_1 ω 1 onto C k C_k C k , and extending the resulting functionals by zero outside C k C_k C k .
Recall from Definition 3.3 that C J = ⋃ k ∈ J C k C_J=\bigcup_{k\in J}C_k C J = ⋃ k ∈ J C k for an interval J ⊂ N J\subset\mathbb{N} J ⊂ N , and
ind ( f ) = { k : supp f ∩ C k ≠ ∅ } . \operatorname{ind}(f)=\{k:\operatorname{supp} f\cap C_k\ne\varnothing\}. ind ( f ) = { k : supp f ∩ C k = ∅ } .
Thus ( f i ) i = 1 d (f_i)_{i=1}^d ( f i ) i = 1 d is an h-block sequence if its members are nonzero, f 1 < ⋯ < f d f_1<\cdots<f_d f 1 < ⋯ < f d in the ordinal order, and
max ind ( f i ) < min ind ( f i + 1 ) ( 1 ≤ i < d ) . \max\operatorname{ind}(f_i)<\min\operatorname{ind}(f_{i+1})\qquad(1\leq i<d). max ind ( f i ) < min ind ( f i + 1 ) ( 1 ≤ i < d ) .
The exponent in the horizontal averaging operations below is 2 2 2 ; the exponents p k p_k p k are used inside the component sets G k G_k G k .
Definition 4.17 (Horizontal weights and special sequences). Let A ⊂ c 00 ( ω 1 , Q ) A\subset c_{00}(\omega_1,\mathbb{Q}) A ⊂ c 00 ( ω 1 , Q ) . A nonzero functional f ∈ A f\in A f ∈ A is h-weighted with chosen weight w ( f ) = m j w(f)=m_j w ( f ) = m j if it has a specified representation
f = 1 m j ∑ t = 1 n j μ t f t , ( μ t ) t = 1 n j ∈ Q n j , ∑ t = 1 n j ∣ μ t ∣ 2 ≤ 1 , (39) f=\frac{1}{m_j}\sum_{t=1}^{n_j}\mu_t f_t,\qquad(\mu_t)_{t=1}^{n_j}\in\mathbb{Q}^{n_j},\qquad\sum_{t=1}^{n_j}|\mu_t|^2\leq1, \tag*{(39)} f = m j 1 t = 1 ∑ n j μ t f t , ( μ t ) t = 1 n j ∈ Q n j , t = 1 ∑ n j ∣ μ t ∣ 2 ≤ 1 , (39) where ( f t ) t = 1 n j (f_t)_{t=1}^{n_j} ( f t ) t = 1 n j is an h-block sequence in A A A . We call this a horizontal ( 2 , j ) (2,j) ( 2 , j ) -operation. The representation, and hence the chosen weight, need not be unique.
A finite sequence ( ϕ i ) i = 1 d (\phi_i)_{i=1}^d ( ϕ i ) i = 1 d in A A A is a horizontal special sequence if it is an h-block sequence of h-weighted functionals, with chosen representations and weights w i = w ( ϕ i ) w_i=w(\phi_i) w i = w ( ϕ i ) , and there are strictly increasing positive integers ( ℓ i ) i = 1 d (\ell_i)_{i=1}^d ( ℓ i ) i = 1 d such that
w 1 = m j 0 , ℓ i ≥ p F i , F i = ⋃ t = 1 i supp ϕ t , w_1=m_{j_0},\qquad\ell_i\geq p_{F_i},\qquad F_i=\bigcup_{t=1}^i\operatorname{supp}\phi_t, w 1 = m j 0 , ℓ i ≥ p F i , F i = t = 1 ⋃ i supp ϕ t , and
w i + 1 = σ ϱ ( ϕ 1 , w 1 , ℓ 1 , … , ϕ i , w i , ℓ i ) ( 1 ≤ i < d ) . w_{i+1}=\sigma_\varrho(\phi_1,w_1,\ell_1,\ldots,\phi_i,w_i,\ell_i)\qquad(1\leq i<d). w i + 1 = σ ϱ ( ϕ 1 , w 1 , ℓ 1 , … , ϕ i , w i , ℓ i ) ( 1 ≤ i < d ) . In particular, the weights are strictly increasing. An infinite horizontal special sequence is defined by requiring every finite initial sequence to satisfy these conditions with the same chosen representations, weights, and integer parameters.
These are the special sequences of Definition 4.3 with the additional horizontal requirements just stated. In this subsection, weights of horizontal special functionals always refer to their horizontal ( 2 , j ) (2,j) ( 2 , j ) -representations. A weight arising only from a component ( p k , j ) (p_k,j) ( p k , j ) -representation is not, by itself, such a horizontal weight.
For a specified finite horizontal special sequence, write Φ = ∑ t = 1 d ϕ t \Phi= \sum_{t=1}^{d} \phi_t Φ = ∑ t = 1 d ϕ t . Its restriction Φ ∣ I \Phi|_I Φ ∣ I to an ordinal interval I I I is called a horizontal special functional . As in Definition 4.4, its retained weight set is
w ( Φ ∣ I ) = { w ( ϕ t ) : supp ϕ t ∩ I ≠ ∅ } . (40) w(\Phi|_I) = \{w(\phi_t) : \operatorname{supp}\phi_t \cap I \ne\varnothing\}. \tag*{(40)} w ( Φ ∣ I ) = { w ( ϕ t ) : supp ϕ t ∩ I = ∅ } . (40)
This set is computed from the specified special decomposition and its chosen weights; for the zero restriction it is empty.
Definition 4.18 (The norming set of E E E ). The set G G G is the smallest subset of c 00 ( ω 1 , Q ) c_{00}(\omega_1,\mathbb{Q}) c 00 ( ω 1 , Q ) with the following properties.
(1) It contains G k G_k G k for every k ≥ 1 k \ge1 k ≥ 1 , is symmetric, and is closed under restrictions both to ordinal intervals and to horizontal intervals:
f ∈ G ⟹ − f ∈ G , f ∣ I ∈ G , f ∣ C J ∈ G f \in G \quad\Longrightarrow\quad-f \in G,\quad f|_I \in G,\quad f|_{C_J} \in G f ∈ G ⟹ − f ∈ G , f ∣ I ∈ G , f ∣ C J ∈ G for every interval I ⊂ ω 1 I \subset\omega_1 I ⊂ ω 1 and every interval J ⊂ N J \subset\mathbb{N} J ⊂ N . In particular f ∣ C k ∈ G f|_{C_k} \in G f ∣ C k ∈ G for every k k k . All restrictions are extended by zero outside the indicated set.
(2) It is closed under finite rational convex combinations.
(3) For every j ∈ N j \in\mathbb{N} j ∈ N , it is closed under the horizontal ( 2 , j ) (2,j) ( 2 , j ) -operation (38).
(4) For every finite horizontal special sequence ( ϕ t ) t = 1 d (\phi_t)_{t=1}^{d} ( ϕ t ) t = 1 d in G G G , with its chosen representations,
∑ t = 1 d ϕ t ∈ G . \sum_{t=1}^{d} \phi_t \in G. t = 1 ∑ d ϕ t ∈ G . (5) Let Φ i = ∑ t = 1 d i ϕ i , t \Phi_i = \sum_{t=1}^{d_i} \phi_{i,t} Φ i = ∑ t = 1 d i ϕ i , t , 1 ≤ i ≤ d 1 \le i \le d 1 ≤ i ≤ d , be specified finite horizontal special sums in G G G , and let I 1 , … , I d ⊂ ω 1 I_1,\ldots,I_d \subset\omega_1 I 1 , … , I d ⊂ ω 1 be ordinal intervals. If
w ( Φ i ∣ I i ) ∩ w ( Φ j ∣ I j ) = ∅ ( i ≠ j ) , w(\Phi_i|_{I_i}) \cap w(\Phi_j|_{I_j}) = \varnothing\qquad(i \ne j), w ( Φ i ∣ I i ) ∩ w ( Φ j ∣ I j ) = ∅ ( i = j ) , then
∑ i = 1 d λ i Φ i ∣ I i ∈ G whenever ( λ i ) i = 1 d ∈ Q d , ∑ i = 1 d ∣ λ i ∣ s ≤ 1. \sum_{i=1}^{d} \lambda_i \Phi_i|_{I_i} \in G \qquad\text{whenever} \qquad(\lambda_i)_{i=1}^{d} \in\mathbb{Q}^{d},\quad\sum_{i=1}^{d} |\lambda_i|^{s} \le1. i = 1 ∑ d λ i Φ i ∣ I i ∈ G whenever ( λ i ) i = 1 d ∈ Q d , i = 1 ∑ d ∣ λ i ∣ s ≤ 1.
Proposition 4.20 (The induced norm and the component spaces). For every f ∈ G f \in G f ∈ G , one has ∥ f ∥ ∞ ≤ 1 \|f\|_{\infty} \le1 ∥ f ∥ ∞ ≤ 1 , and
{ f ∣ C k : f ∈ G } = G k ( k ≥ 1 ) . (41) \{f|_{C_k}: f \in G\}=G_k \qquad(k\ge1). \tag*{(41)} { f ∣ C k : f ∈ G } = G k ( k ≥ 1 ) . (41) In particular, f ∣ C k ∈ co ‾ G k f|_{C_k}\in\overline{\operatorname{co}}G_k f ∣ C k ∈ co G k for every f ∈ G f\in G f ∈ G , where the closure is taken in the norm of E k ∗ E_k^* E k ∗ . Consequently,
∥ x ∥ 2 , s = sup { ∣ f ( x ) ∣ : f ∈ G } , E = ( c 00 ( ω 1 ) , ∥ ⋅ ∥ 2 , s ) ‾ (42) \|x\|_{2,s}=\sup\{|f(x)|:f\in G\}, \qquad E=\overline{(c_{00}(\omega_1),\|\cdot\|_{2,s})} \tag*{(42)} ∥ x ∥ 2 , s = sup { ∣ f ( x ) ∣ : f ∈ G } , E = ( c 00 ( ω 1 ) , ∥ ⋅ ∥ 2 , s ) (42) defines a Banach space with a normalized bimonotone transfinite basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 . Every horizontal interval projection π J \pi_J π J is contractive, and π k ( E ) \pi_k(E) π k ( E ) is naturally isometric to E k E_k E k .
Proof. We use the construction stages in Remark 4.19 . The component norming sets satisfy the sup-norm bound. A horizontal weighted functional formed from functionals with this bound has sup norm at most 1 / m j 1/m_j 1/ m j , since its children have disjoint supports and ∣ μ t ∣ ≤ 1 |\mu_t|\le1 ∣ μ t ∣ ≤ 1 . A special sum has distinct weights. In rule (5), all retained weights are distinct across the specified sums and ∣ λ i ∣ ≤ 1 |\lambda_i|\le1 ∣ λ i ∣ ≤ 1 . Thus the sup norm in either special-sum rule is bounded by ∑ j 1 / m j < 1 \sum_j 1/m_j<1 ∑ j 1/ m j < 1 . Symmetry, restrictions, and rational convex combinations preserve the bound, proving the first assertion by induction.
Fix k k k . We prove the stronger inclusion f ∣ C k ∈ G k f|_{C_k}\in G_k f ∣ C k ∈ G k by induction on a finite tree analysis from Remark 4.19 , with terminal functionals in the sets G l G_l G l . At a terminal node the assertion follows from the disjointness of the sets C l C_l C l and the fact that 0 ∈ G k 0\in G_k 0 ∈ G k . It is preserved by symmetry, both types of restriction, and rational convex combinations. Here restrictions of members of G k G_k G k to ordinal intervals correspond to interval restrictions in the transported component norming set.
In a horizontal operation, at most one child has support meeting C k C_k C k . Hence its restriction is either zero or
μ t m j ( f t ∣ C k ) , \frac{\mu_t}{m_j}(f_t|_{C_k}), m j μ t ( f t ∣ C k ) ,
which belongs to G k G_k G k : this set is symmetric, contains zero, and is closed under rational convex combinations. Similarly, at most one term of a horizontal special sequence meets C k C_k C k , so restricting its sum again gives an element of G k G_k G k .
For rule (5), discard the terms whose restrictions to C k C_k C k are zero. For each remaining i i i , precisely one special term can contribute on C k C_k C k , and at most one child in its horizontal representation can contribute. Therefore the restriction of the resulting functional has the form
∑ i λ i μ i w i h i , h i ∈ G k , μ i ∈ Q , ∣ μ i ∣ ≤ 1 , \sum_i\frac{\lambda_i\mu_i}{w_i}h_i,\qquad h_i\in G_k,\quad\mu_i\in\mathbb{Q},\quad|\mu_i|\le1, i ∑ w i λ i μ i h i , h i ∈ G k , μ i ∈ Q , ∣ μ i ∣ ≤ 1 ,
where w i ∈ w ( Φ i ∣ I i ) w_i\in w(\Phi_i|_{I_i}) w i ∈ w ( Φ i ∣ I i ) is the weight of that special term. The functionals h i h_i h i are interval restrictions of component restrictions of children to which the induction hypothesis applies. The weights w i w_i w i are pairwise distinct, and hence
∑ i ∣ λ i μ i w i ∣ ≤ ∑ i 1 w i ≤ ∑ j = 1 ∞ 1 m j < 1. \sum_i\left|\frac{\lambda_i\mu_i}{w_i}\right|\le\sum_i\frac{1}{w_i}\le\sum_{j=1}^{\infty}\frac{1}{m_j}<1. i ∑ w i λ i μ i ≤ i ∑ w i 1 ≤ j = 1 ∑ ∞ m j 1 < 1.
Thus this is a rational absolutely convex combination of elements of G k G_k G k and belongs to G k G_k G k . This completes the induction. The reverse inclusion in (41) follows from G k ⊂ G G_k\subset G G k ⊂ G .
Since G G G contains all the signed singletons and its elements have sup norm at most one, the supremum in (42) is finite and defines a norm, with ∥ e α ∥ 2 , s = 1 \|e_\alpha\|_{2,s}=1 ∥ e α ∥ 2 , s = 1 . The restriction rules make ordinal interval projections and horizontal interval projections contractive. Density of c 00 ( ω 1 ) c_{00}(\omega_1) c 00 ( ω 1 ) then gives the bimonotone transfinite Schauder basis and the bounded extensions of these projections. Finally, for x ∈ c 00 ( C k ) x\in c_{00}(C_k) x ∈ c 00 ( C k ) ,
∥ x ∥ 2 , s = sup f ∈ G ∣ ( f ∣ C k ) ( x ) ∣ = sup g ∈ G k ∣ g ( x ) ∣ = ∥ x ∥ E k . \|x\|_{2,s}=\sup_{f\in G}|(f|_{C_k})(x)|=\sup_{g\in G_k}|g(x)|=\|x\|_{E_k}. ∥ x ∥ 2 , s = f ∈ G sup ∣ ( f ∣ C k ) ( x ) ∣ = g ∈ G k sup ∣ g ( x ) ∣ = ∥ x ∥ E k .
Passing to completions proves the asserted component identification. □
Remark 4.21 (Infinite horizontal special functionals). The analogue of Remark 4.8 holds for horizontal special sequences. Indeed, if ( ϕ i ) i ∈ N (\phi_i)_{i\in\mathbb{N}} ( ϕ i ) i ∈ N is such a sequence, all its finite initial sums belong to G G G . They therefore converge weak-star to
Φ = w ∗ ∑ i = 1 ∞ ϕ i ∈ B E ∗ , \Phi=w^*\sum_{i=1}^{\infty}\phi_i\in B_{E^*}, Φ = w ∗ i = 1 ∑ ∞ ϕ i ∈ B E ∗ , since their values stabilize on finitely supported vectors and their dual norms are at most one.
Let Φ 0 = { ϕ γ 0 : γ ∈ Γ 0 } \Phi^0=\{\phi_\gamma^0:\gamma\in\Gamma_0\} Φ 0 = { ϕ γ 0 : γ ∈ Γ 0 } be the family of distinct infinite horizontal special functionals outside X 0 X_0 X 0 , fixing one special expansion for each member. For a finite F ⊂ Γ 0 F\subset\Gamma_0 F ⊂ Γ 0 , Proposition 4.5 gives an n 0 n_0 n 0 after which these expansions have pairwise disjoint weight sets. Applying rule (5) of Definition 4.18 to finite tails, passing to weak-star limits, and then approximating real coefficient families by rational ones gives
∥ ∑ γ ∈ F a γ ϕ γ > n 0 ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ s ) 1 / s ( n > n 0 ) . \left\|\sum_{\gamma\in F}a_\gamma\phi_{\gamma>n}^0\right\|\leq\left(\sum_{\gamma\in F}|a_\gamma|^s\right)^{1/s}\qquad(n>n_0). γ ∈ F ∑ a γ ϕ γ > n 0 ≤ γ ∈ F ∑ ∣ a γ ∣ s 1/ s ( n > n 0 ) . This proves (13). The same upper ℓ s \ell_s ℓ s estimate holds for the cosets Q ϕ γ 0 Q\phi_\gamma^0 Q ϕ γ 0 , because every finite initial sum belongs to X 0 X_0 X 0 . The uncountability and norm-span properties of the family belong to the structural results whose proofs are omitted here; see Remark 4.25.
Proposition 4.22 (Bounded completeness of the basis of E E E ). Under the parameter growth condition (29), the transfinite basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 of E E E is boundedly complete.
Proof. Suppose, to the contrary, that there are strictly increasing ordinals ( α i ) i (\alpha_i)_i ( α i ) i and scalars ( a i ) i (a_i)_i ( a i ) i for which
S N = ∑ i = 1 N a i e α i , M : = sup N ∥ S N ∥ 2 , s < ∞ , S_N=\sum_{i=1}^{N}a_i e_{\alpha_i},\qquad M:=\sup_N\|S_N\|_{2,s}<\infty, S N = i = 1 ∑ N a i e α i , M := N sup ∥ S N ∥ 2 , s < ∞ , but ( S N ) N (S_N)_N ( S N ) N does not converge in norm. Choose ε > 0 \varepsilon>0 ε > 0 and 0 = q 0 < q 1 < q 2 < ⋯ 0=q_0<q_1<q_2<\cdots 0 = q 0 < q 1 < q 2 < ⋯ such that
x n = ∑ i = q n − 1 + 1 q n a i e α i , ∥ x n ∥ 2 , s ≥ ε ( n ∈ N ) . x_n=\sum_{i=q_{n-1}+1}^{q_n}a_i e_{\alpha_i},\qquad\|x_n\|_{2,s}\geq\varepsilon\quad(n\in\mathbb{N}). x n = i = q n − 1 + 1 ∑ q n a i e α i , ∥ x n ∥ 2 , s ≥ ε ( n ∈ N ) . Put P m = π { 1 , … , m } P_m=\pi_{\{1,\ldots,m\}} P m = π { 1 , … , m } , P 0 = 0 P_0=0 P 0 = 0 , and
D n = [ α q n − 1 + 1 , α q n ] . D_n=[\alpha_{q_{n-1}+1},\alpha_{q_n}]. D n = [ α q n − 1 + 1 , α q n ] . The ordinal intervals D n D_n D n are successive. We distinguish two cases.
Case 1. There are m ∈ N m \in\mathbb{N} m ∈ N , δ > 0 \delta> 0 δ > 0 , and infinitely many n n n for which ∥ P m x n ∥ 2 , s ≥ δ \lVert P_m x_n\rVert_{2,s} \ge\delta ∥ P m x n ∥ 2 , s ≥ δ . Since
∥ P m x n ∥ 2 , s ≤ ∑ k = 1 m ∥ π k x n ∥ 2 , s , \lVert P_m x_n\rVert_{2,s} \le\sum_{k=1}^{m} \lVert\pi_k x_n\rVert_{2,s}, ∥ P m x n ∥ 2 , s ≤ k = 1 ∑ m ∥ π k x n ∥ 2 , s ,
there are a fixed k 0 ≤ m k_0 \le m k 0 ≤ m and a subsequence ( x ν i ) i (x_{\nu_i})_i ( x ν i ) i such that, with η = δ / m \eta= \delta/m η = δ / m ,
z i = π k 0 x ν i , ∥ z i ∥ E k 0 = ∥ z i ∥ 2 , s ≥ η . z_i = \pi_{k_0}x_{\nu_i}, \qquad\lVert z_i\rVert_{E_{k_0}} = \lVert z_i\rVert_{2,s} \ge\eta. z i = π k 0 x ν i , ∥ z i ∥ E k 0 = ∥ z i ∥ 2 , s ≥ η .
Here we used Proposition 4.20. Choose g i ∈ G k 0 g_i \in G_{k_0} g i ∈ G k 0 supported in D ν i ∩ C k 0 D_{\nu_i} \cap C_{k_0} D ν i ∩ C k 0 with g i ( z i ) > η / 2 g_i(z_i) > \eta/2 g i ( z i ) > η /2 . Such a choice is possible by symmetry, the norming property, and closure under ordinal interval restrictions. The functionals g i g_i g i are successive in the component order.
Put p = p k 0 p = p_{k_0} p = p k 0 and p ′ = p / ( p − 1 ) p' = p/(p-1) p ′ = p / ( p − 1 ) . For every j j j , choose a rational t j t_j t j with
1 2 n j − 1 / p ≤ t j ≤ n j − 1 / p . \frac{1}{2}n_j^{-1/p} \le t_j \le n_j^{-1/p}. 2 1 n j − 1/ p ≤ t j ≤ n j − 1/ p .
The component ( p , j ) (p,j) ( p , j ) -operation gives
g = t j m j ∑ i = 1 n j g i ∈ G k 0 ⊂ G . g = \frac{t_j}{m_j}\sum_{i=1}^{n_j}g_i \in G_{k_0} \subset G. g = m j t j i = 1 ∑ n j g i ∈ G k 0 ⊂ G .
Set N j = q ν n j N_j = q_{\nu_{n_j}} N j = q ν n j . The ordinal support restrictions give g i ( S N j ) = g i ( z i ) g_i(S_{N_j}) = g_i(z_i) g i ( S N j ) = g i ( z i ) for i ≤ n j i \le n_j i ≤ n j . Therefore
M ≥ g ( S N j ) = t j m j ∑ i = 1 n j g i ( z i ) > η 4 n j 1 / p ′ m j . M \ge g(S_{N_j}) = \frac{t_j}{m_j}\sum_{i=1}^{n_j}g_i(z_i) > \frac{\eta}{4}\frac{n_j^{1/p'}}{m_j}. M ≥ g ( S N j ) = m j t j i = 1 ∑ n j g i ( z i ) > 4 η m j n j 1/ p ′ .
This contradicts (29).
Case 2. For every m ∈ N m \in\mathbb{N} m ∈ N , ∥ P m x n ∥ 2 , s → 0 \lVert P_m x_n\rVert_{2,s} \to0 ∥ P m x n ∥ 2 , s → 0 as n → ∞ n \to\infty n → ∞ . Inductively choose a subsequence ( x ν i ) i (x_{\nu_i})_i ( x ν i ) i and integers 0 = b 0 < b 1 < b 2 < ⋯ 0 = b_0 < b_1 < b_2 < \cdots 0 = b 0 < b 1 < b 2 < ⋯ so that
∥ P b i − 1 x ν i ∥ 2 , s < ε 2 − i − 2 , ind ( x ν i ) ⊂ { 1 , … , b i } . \lVert P_{b_{i-1}}x_{\nu_i}\rVert_{2,s} < \varepsilon2^{-i-2}, \qquad\operatorname{ind}(x_{\nu_i}) \subset\{1,\ldots,b_i\}. ∥ P b i − 1 x ν i ∥ 2 , s < ε 2 − i − 2 , ind ( x ν i ) ⊂ { 1 , … , b i } .
The second choice is possible because x ν i x_{\nu_i} x ν i has finite support. Set
J i = { b i − 1 + 1 , … , b i } , y i = π J i x ν i . J_i = \{b_{i-1}+1,\ldots,b_i\}, \qquad y_i = \pi_{J_i}x_{\nu_i}. J i = { b i − 1 + 1 , … , b i } , y i = π J i x ν i .
Then ( y i ) i (y_i)_i ( y i ) i is an h-block sequence and
∥ x ν i − y i ∥ 2 , s < ε 2 − i − 2 , ∥ y i ∥ 2 , s > ε / 2. \lVert x_{\nu_i} - y_i\rVert_{2,s} < \varepsilon2^{-i-2}, \qquad\lVert y_i\rVert_{2,s} > \varepsilon/2. ∥ x ν i − y i ∥ 2 , s < ε 2 − i − 2 , ∥ y i ∥ 2 , s > ε /2.
Choose h i ∈ G h_i \in G h i ∈ G with h i ( y i ) > ε / 4 h_i(y_i) > \varepsilon/4 h i ( y i ) > ε /4 , and restrict h i h_i h i to D ν i ∩ C J i D_{\nu_i} \cap C_{J_i} D ν i ∩ C J i . By the two restriction rules, the resulting f i f_i f i belongs to G G G and has the same value on y i y_i y i . The functionals ( f i ) i (f_i)_i ( f i ) i form an h-block sequence.
For every j j j , choose a rational u j u_j u j with
1 2 n j − 1 / 2 ≤ u j ≤ n j − 1 / 2 . \frac{1}{2}n_j^{-1/2} \le u_j \le n_j^{-1/2}. 2 1 n j − 1/2 ≤ u j ≤ n j − 1/2 .
The horizontal ( 2 , j ) (2,j) ( 2 , j ) -operation gives
f = u j m j ∑ i = 1 n j f i ∈ G . f = \frac{u_j}{m_j}\sum_{i=1}^{n_j}f_i \in G. f = m j u j i = 1 ∑ n j f i ∈ G .
Set N j = q ν n j N_j = q_{\nu_{n_j}} N j = q ν n j . The support restrictions imply f i ( S N j ) = f i ( y i ) f_i(S_{N_j}) = f_i(y_i) f i ( S N j ) = f i ( y i ) for i ≤ n j i \le n_j i ≤ n j , so
M ≥ f ( S N j ) = u j m j ∑ i = 1 n j f i ( y i ) > ε 8 n j 1 / 2 m j . M \ge f(S_{N_j}) = \frac{u_j}{m_j}\sum_{i=1}^{n_j}f_i(y_i) > \frac{\varepsilon}{8}\frac{n_j^{1/2}}{m_j}. M ≥ f ( S N j ) = m j u j i = 1 ∑ n j f i ( y i ) > 8 ε m j n j 1/2 .
This contradicts (29) with q = 2 q=2 q = 2 . Both cases are impossible, proving bounded completeness. □
Since G ⊂ B X 0 G \subset B_{X_0} G ⊂ B X 0 and G G G norms E E E , the space X 0 X_0 X 0 is 1-norming for E E E . Proposition 4.11 now gives the canonical onto isometry
E ⟶ X 0 ∗ , x ⟼ ( h ⟼ h ( x ) ) . E \longrightarrow X_0^*, \qquad x \longmapsto(h \longmapsto h(x)). E ⟶ X 0 ∗ , x ⟼ ( h ⟼ h ( x )) .
Thus property 1 follows from the norming-set construction and the preceding proposition.
Theorem 4.23. (Summary of the properties of E E E ). Let E E E be the space defined by G G G , with the parameters and partition fixed in Section 4.3. Let Φ 0 = { ϕ γ 0 : γ ∈ Γ 0 } \Phi^0=\{\phi^0_\gamma:\gamma\in\Gamma_0\} Φ 0 = { ϕ γ 0 : γ ∈ Γ 0 } be the family of infinite horizontal special functionals outside X 0 X_0 X 0 . For k ≥ 1 k\geq1 k ≥ 1 , let Φ k = { ϕ γ k : γ ∈ Γ k } \Phi^k=\{\phi^k_\gamma:\gamma\in\Gamma_k\} Φ k = { ϕ γ k : γ ∈ Γ k } be the family of infinite special functionals of E k E_k E k , extended to E E E by ι k \iota_k ι k . We call the families Φ k \Phi^k Φ k , k ≥ 1 k\geq1 k ≥ 1 , the vertical special families. Put Q : E ∗ → E ∗ / X 0 Q:E^*\to E^*/X_0 Q : E ∗ → E ∗ / X 0 . Then the following hold.
(a) The basis ( e α ) α < ω 1 (e_\alpha)_{\alpha<\omega_1} ( e α ) α < ω 1 admits an upper ℓ 2 \ell_2 ℓ 2 estimate with constant one: for every finitely supported scalar family ( a α ) (a_\alpha) ( a α ) ,
∥ ∑ α a α e α ∥ 2 , s ≤ ( ∑ α ∣ a α ∣ 2 ) 1 / 2 . \left\|\sum_\alpha a_\alpha e_\alpha\right\|_{2,s}\leq\left(\sum_\alpha|a_\alpha|^2\right)^{1/2}. α ∑ a α e α 2 , s ≤ ( α ∑ ∣ a α ∣ 2 ) 1/2 . (b) For every normalized infinite h h h -block sequence ( x n ) n (x_n)_n ( x n ) n , the space span { x n : n ∈ N } \operatorname{span}\{x_n:n\in\mathbb{N}\} span { x n : n ∈ N } has nonseparable dual.
(c) The biorthogonal functionals, together with the horizontal and vertical special families, generate the dual in norm:
E ∗ = span ( { e α ∗ : α < ω 1 } ∪ ⋃ k ≥ 0 Φ k ) = span ( ⋃ k ≥ 1 E k ∗ ∪ Φ 0 ) . E^*=\operatorname{span}\left(\{e_\alpha^*:\alpha<\omega_1\}\cup\bigcup_{k\geq0}\Phi^k\right)=\operatorname{span}\left(\bigcup_{k\geq1}E_k^*\cup\Phi^0\right). E ∗ = span ( { e α ∗ : α < ω 1 } ∪ k ≥ 0 ⋃ Φ k ) = span ( k ≥ 1 ⋃ E k ∗ ∪ Φ 0 ) . (d) The horizontal quotient family ( Q ϕ γ 0 ) γ (Q\phi^0_\gamma)_\gamma ( Q ϕ γ 0 ) γ admits an upper ℓ s \ell_s ℓ s estimate, and for every k ≥ 1 k\geq1 k ≥ 1 the vertical quotient family ( Q ϕ γ k ) γ (Q\phi^k_\gamma)_\gamma ( Q ϕ γ k ) γ admits an upper ℓ r k \ell_{r_k} ℓ r k estimate, all with constant one. Equivalently, setting q 0 = s q_0=s q 0 = s and q k = r k q_k=r_k q k = r k for k ≥ 1 k\geq1 k ≥ 1 ,
∥ ∑ γ ∈ F a γ Q ϕ γ k ∥ ≤ ( ∑ γ ∈ F ∣ a γ ∣ q k ) 1 / q k \left\|\sum_{\gamma\in F}a_\gamma Q\phi^k_\gamma\right\|\leq\left(\sum_{\gamma\in F}|a_\gamma|^{q_k}\right)^{1/q_k} γ ∈ F ∑ a γ Q ϕ γ k ≤ γ ∈ F ∑ ∣ a γ ∣ q k 1/ q k for every k ≥ 0 k\geq0 k ≥ 0 , every finite F ⊂ Γ k F\subset\Gamma_k F ⊂ Γ k , and every scalar family ( a γ ) γ ∈ F (a_\gamma)_{\gamma\in F} ( a γ ) γ ∈ F . Moreover,
E ∗ / X 0 = span { Q ϕ γ k : k ≥ 0 , γ ∈ Γ k } . E^*/X_0=\operatorname{span}\{Q\phi^k_\gamma:k\geq0,\ \gamma\in\Gamma_k\}. E ∗ / X 0 = span { Q ϕ γ k : k ≥ 0 , γ ∈ Γ k } .
Proof of the estimates. For (a), we first check that ∥ f ∥ ℓ 2 ( ω 1 ) ≤ 1 \|f\|_{\ell_2(\omega_1)}\leq1 ∥ f ∥ ℓ 2 ( ω 1 ) ≤ 1 for every f ∈ G f\in G f ∈ G . At the initial stage, f ∈ G k f\in G_k f ∈ G k for some k k k , and Lemma 4.13, transported to C k C_k C k , gives
∥ f ∥ ℓ 2 ≤ ∥ f ∥ ℓ p k ≤ 1 , \|f\|_{\ell_2}\leq\|f\|_{\ell_{p_k}}\leq1, ∥ f ∥ ℓ 2 ≤ ∥ f ∥ ℓ p k ≤ 1 ,
since p k < 2 p_k<2 p k < 2 . For a horizontal ( 2 , j ) (2,j) ( 2 , j ) -operation, disjointness of the children’s supports gives
∥ 1 m j ∑ t μ t f t ∥ 2 2 = 1 m j 2 ∑ t ∣ μ t ∣ 2 ∥ f t ∥ 2 2 ≤ 1 m j 2 , \left\|\frac{1}{m_j}\sum_t\mu_t f_t\right\|_2^2
=\frac{1}{m_j^2}\sum_t|\mu_t|^2\|f_t\|_2^2
\leq\frac{1}{m_j^2}, m j 1 t ∑ μ t f t 2 2 = m j 2 1 t ∑ ∣ μ t ∣ 2 ∥ f t ∥ 2 2 ≤ m j 2 1 ,
provided the bound holds for the children. Thus a horizontally weighted functional has ℓ 2 \ell_2 ℓ 2 norm at most the reciprocal of its chosen weight. The special-sum rules preserve the unit bound by the triangle inequality and ∑ j 1 / m j < 1 \sum_j 1/m_j < 1 ∑ j 1/ m j < 1 ; in rule (5), the retained weight sets are disjoint and each coefficient has absolute value at most one. Symmetry, both types of restriction, and rational convex combinations also preserve the bound. Induction as in Remark 4.19 proves the claim. Applying the Cauchy–Schwarz inequality and taking the supremum over f ∈ G f \in G f ∈ G gives (a).
The horizontal estimate in (d) was proved in Remark 4.21. The vertical estimates follow from Theorem 4.15 (d) and the isometric embeddings j k : W k → W j_k : W_k \to W j k : W k → W described in § 4.5 . The norm-span assertion in the quotient follows from (c).
Assertions (b) and (c) require the corresponding HI and James-tree arguments for the combined construction. The discussion and references following Theorem 4.15 apply here as well; for horizontal constructions involving families of spaces, see also [4 , 7 ] . These lengthy technical proofs are omitted from the present paper. □
Corollary 4.24. Every infinite-dimensional closed subspace of E E E has nonseparable dual.
Proof. Let Y ⊂ E Y \subset E Y ⊂ E be an infinite-dimensional closed subspace. By [6 Proposition 1.3(3)] , Y Y Y contains a closed subspace isomorphic to a block subspace U = span ‾ { u n : n ∈ N } U = \overline{\operatorname{span}}\{u_n : n \in\mathbb{N}\} U = span { u n : n ∈ N } , where ( u n ) n (u_n)_n ( u n ) n is normalized. It suffices to show that U ∗ U^* U ∗ is nonseparable: the same will then hold for the dual of that subspace of Y Y Y , which is a quotient of Y ∗ Y^* Y ∗ by the Hahn–Banach theorem.
Suppose first that, for some k k k , the operator π k ∣ U : U → E k \pi_k|_U : U \to E_k π k ∣ U : U → E k is not strictly singular. There is then an infinite-dimensional closed subspace Z ⊂ U Z \subset U Z ⊂ U on which π k \pi_k π k is an isomorphism onto a closed subspace of E k E_k E k . By Corollary 4.16, applied to the isometric copy E k = E p k , r k ( C k ) E_k = E_{p_k,r_k}(C_k) E k = E p k , r k ( C k ) , this image has nonseparable dual. Hence Z ∗ Z^* Z ∗ is nonseparable, and the surjective restriction map U ∗ → Z ∗ U^* \to Z^* U ∗ → Z ∗ shows that U ∗ U^* U ∗ is nonseparable.
It remains to consider the case where every π k ∣ U \pi_k|_U π k ∣ U is strictly singular. Put P m = ∑ k = 1 m π k P_m = \sum_{k=1}^{m} \pi_k P m = ∑ k = 1 m π k and P 0 = 0 P_0 = 0 P 0 = 0 . Finite sums of strictly singular operators are strictly singular, so each P m ∣ U P_m|_U P m ∣ U is strictly singular. Choose positive numbers ( ε n ) n (\varepsilon_n)_n ( ε n ) n with θ : = ∑ n ε n < 1 / 2 \theta:= \sum_n \varepsilon_n < 1/2 θ := ∑ n ε n < 1/2 . We inductively choose successive normalized finite blocks ( v n ) n (v_n)_n ( v n ) n of ( u n ) n (u_n)_n ( u n ) n and integers 0 = b 0 < b 1 < b 2 < ⋯ 0 = b_0 < b_1 < b_2 < \cdots 0 = b 0 < b 1 < b 2 < ⋯ such that
∥ P b n − 1 v n ∥ 2 , s < ε n , ind ( v n ) ⊂ { 1 , … , b n } . \lVert P_{b_{n-1}}v_n\rVert_{2,s} < \varepsilon_n,\qquad\operatorname{ind}(v_n) \subset\{1,\ldots,b_n\}. ∥ P b n − 1 v n ∥ 2 , s < ε n , ind ( v n ) ⊂ { 1 , … , b n } .
To choose v n v_n v n , restrict P b n − 1 P_{b_{n-1}} P b n − 1 to the infinite-dimensional tail of U U U following the preceding block. Strict singularity gives a unit vector whose image has arbitrarily small norm. Approximation by a finite block, followed by normalization, gives the required v n v_n v n . Then choose b n b_n b n to contain its finite horizontal support.
Set
w n = π { b n − 1 + 1 , … , b n } v n . w_n = \pi_{\{b_{n-1}+1,\ldots,b_n\}}v_n. w n = π { b n − 1 + 1 , … , b n } v n .
The vectors w n w_n w n are successive in the ordinal order, their horizontal supports are successive, and
∥ v n − w n ∥ 2 , s < ε n , ∥ w n ∥ 2 , s > 1 − ε n > 1 / 2. \lVert v_n-w_n\rVert_{2,s}<\varepsilon_n,\qquad\lVert w_n\rVert_{2,s}>1-\varepsilon_n>1/2. ∥ v n − w n ∥ 2 , s < ε n , ∥ w n ∥ 2 , s > 1 − ε n > 1/2.
Thus ( w n ) n (w_n)_n ( w n ) n is a seminormalized h-block sequence. To verify explicitly the equivalence of ( v n ) n (v_n)_n ( v n ) n and ( w n ) n (w_n)_n ( w n ) n , let ( a n ) n (a_n)_n ( a n ) n be finitely supported. Bimonotonicity and ∥ v n ∥ 2 , s = 1 \lVert v_n\rVert_{2,s}=1 ∥ v n ∥ 2 , s = 1 imply
sup n ∣ a n ∣ ≤ ∥ ∑ n a n v n ∥ 2 , s . \sup_n |a_n| \le\left\lVert\sum_n a_n v_n \right\rVert_{2,s}. n sup ∣ a n ∣ ≤ n ∑ a n v n 2 , s .
Consequently,
∥ ∑ n a n ( v n − w n ) ∥ 2 , s ≤ ∑ n ∣ a n ∣ ε n ≤ θ ∥ ∑ n a n v n ∥ 2 , s . \left\|\sum_n a_n(v_n-w_n)\right\|_{2,s}\leq\sum_n|a_n|\varepsilon_n\leq\theta\left\|\sum_n a_nv_n\right\|_{2,s}. n ∑ a n ( v n − w n ) 2 , s ≤ n ∑ ∣ a n ∣ ε n ≤ θ n ∑ a n v n 2 , s .
Since θ < 1 \theta<1 θ < 1 , the map v n ↦ w n v_n\mapsto w_n v n ↦ w n extends to an isomorphism between their closed linear spans. After normalizing ( w n ) n (w_n)_n ( w n ) n , Theorem 4.23(b) shows that its closed span has nonseparable dual. The same is true of span { v n : n ∈ N } ⊂ U \operatorname{span}\{v_n:n\in\mathbb{N}\}\subset U span { v n : n ∈ N } ⊂ U . Hahn–Banach again implies that U ∗ U^* U ∗ is nonseparable, completing the proof. □
References
National Technical University of Athens, Faculty of Applied Sciences, Department of Mathematics, Zografou Campus, 157 80, Athens, Greece.
Email address: sargyros@math.ntua.gr