Introduction
The effective study of continuous mathematics begins with a deceptively simple question: what should it mean to compute a real number? In his landmark paper [29 ] , Turing gave this question a precise answer by replacing a completed infinite expansion with a finite mechanical procedure that produces its digits one by one. Although he used this definition to address the Entscheidungsproblem, the conceptual move reached far beyond its original application. It supplied a lasting paradigm for computable analysis: an infinite object becomes accessible through a uniform process that generates finite approximations. The same shift of viewpoint—from an object in the abstract to the information by which it is presented—also lies at the heart of computable structure theory. The present paper follows this line of thought for L p L^p L p -spaces.
The distinction between an abstract mathematical object and a particular presentation of it is largely suppressed in classical mathematics. Isomorphic groups, isometric metric spaces, and linearly isometric Banach spaces are regarded as the same object, and a theorem is normally stated about the abstract object rather than about a particular choice of coordinates. Computation cannot ignore this distinction. An algorithm never receives an abstract isomorphism type, and an infinite object cannot be handed to it all at once. What the algorithm receives is a concrete presentation : a system of names together with procedures for evaluating the relevant operations, relations, or distances. Computable structure theory studies mathematics through such presentations. It asks for the effective content of classical results: which infinite structures admit algorithmic descriptions, which invariants and constructions can be recovered from those descriptions, whether classically equivalent descriptions can be translated into one another effectively, and how much additional computational information is required when they cannot. In this way the
2020 Mathematics Subject Classification. Primary 03D78; Secondary 03D45, 46B04, 46E30. Key words and phrases. Computable presentations, Lebesgue spaces, approximate ultrahomogeneity. subject connects the computability theory of countable algebraic structures with computable analysis and the study of separable uncountable structures, see [13 ] and [15 ] .
For a countable structure, a computable presentation gives an isomorphic copy on a computable subset of N \mathbb{N} N with computable basic operations and relations. The passage to analysis creates an immediate obstacle. A separable Banach space is normally uncountable, so its vectors cannot be listed by natural numbers. Separability nevertheless supplies a countable interface: if ( v n ) n ∈ N (v_n)_{n\in\mathbb{N}} ( v n ) n ∈ N has dense linear span, then its rational linear combinations form a countable skeleton whose completion is the whole space. Computability requires effective access to the norm on this skeleton; arbitrary vectors and operators are then handled by Cauchy names.
This framework gives rise to three closely related problems. The presentability problem asks which classical structures admit at least one computable presentation. The effective uniqueness problem asks whether all computable presentations of a fixed isomorphism or isometry type are computably equivalent. The recognition problem ranges over presentations and asks how difficult it is to determine whether the presented object belongs to a specified class or has a specified type. Separable Lebesgue spaces provide a particularly revealing test case. Their classical linear-isometric structure is governed by the exponent and the underlying measure algebra, but an arbitrary computable Banach presentation need not display either of these objects. It presents only a dense sequence and the norm geometry generated by that sequence. The exponent must therefore be detected indirectly, if it can be detected at all, through geometric identities, embeddings, and finite-dimensional configurations. Conversely, to build a computable presentation one must make all norm computations effective even when the exponent itself is not computable.
The history leading to our work passes naturally through the three problems above. We begin with effective uniqueness. A computably (presentable) Banach space is computably categorical if any two computable Banach presentations of spaces linearly isometric to it are connected by a surjective linear isometry that is computable with respect to their induced Cauchy representations. Pour-El and Richards [25 ] proved that every separable Hilbert space is computably categorical. The first examples beyond Hilbert spaces arose in finite dimensions. Let ( Ω , Σ , μ ) (\Omega,\Sigma,\mu) ( Ω , Σ , μ ) be a measure space and 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ . The Lebesgue space L r ( Ω , Σ , μ ; F ) L^r(\Omega,\Sigma,\mu;\mathbb{F}) L r ( Ω , Σ , μ ; F ) consists of equivalence classes modulo almost-everywhere equality of measurable f : Ω → F f:\Omega\to\mathbb{F} f : Ω → F for which ∥ f ∥ r = ( ∫ Ω ∣ f ∣ r d μ ) 1 / r < ∞ \|f\|_r = \left(\int_{\Omega} |f|^r\,d\mu\right)^{1/r} < \infty ∥ f ∥ r = ( ∫ Ω ∣ f ∣ r d μ ) 1/ r < ∞ . We abbreviate L r ( [ 0 , 1 ] , λ ; F ) L^r([0,1],\lambda;\mathbb{F}) L r ([ 0 , 1 ] , λ ; F ) to L r [ 0 , 1 ] L^r[0,1] L r [ 0 , 1 ] . Also,
ℓ r N = ( F N , ∥ ( a j ) j < N ∥ r = ( ∑ j < N ∣ a j ∣ r ) 1 / r ) . \ell_r^N = \left(\mathbb{F}^N,\|(a_j)_{j<N}\|_r = \left(\sum_{j<N}|a_j|^r\right)^{1/r}\right). ℓ r N = F N , ∥ ( a j ) j < N ∥ r = j < N ∑ ∣ a j ∣ r 1/ r .
Informally, a real number is computable if a single algorithm can approximate it to any prescribed accuracy. Thus, when p p p is computable, the exponent can be used effectively in norm calculations. McNicholl [23 ] established computable categoricity for ℓ p N \ell_p^N ℓ p N where p ≥ 1 p \ge1 p ≥ 1 is computable.
Clanin, McNicholl, and Stull subsequently obtained an infinite-dimensional counterpart. For computable p p p , the space L p [ 0 , 1 ] L^p[0,1] L p [ 0 , 1 ] has a standard computable presentation given by the step functions with coefficients in Q \mathbb{Q} Q in the real case and in Q + i Q \mathbb{Q}+i\mathbb{Q} Q + i Q in the complex case, and with rational interval endpoints.
Theorem 1.1 (Clanin–McNicholl–Stull [9 ] ). Let F F F be either R \mathbb{R} R or C \mathbb{C} C , and let p ≥ 1 p \ge1 p ≥ 1 be a computable real. If Ω \Omega Ω is an atomless, separable measure space such that L p ( Ω ; F ) ≠ { 0 } L^{p}(\Omega; F) \ne\{0\} L p ( Ω ; F ) = { 0 } , then every computable presentation of L p ( Ω ; F ) L^{p}(\Omega; F) L p ( Ω ; F ) is computably linearly isometric to the standard computable presentation of L p ( [ 0 , 1 ] ; F ) L^{p}([0,1]; F) L p ([ 0 , 1 ] ; F ) . Consequently, L p ( [ 0 , 1 ] ; F ) L^{p}([0,1]; F) L p ([ 0 , 1 ] ; F ) is computably categorical for every computable real p ≥ 1 p \ge1 p ≥ 1 .
We next turn to the recognition problem: how difficult is it to recognize a Lebesgue space from an index? Here a Lebesgue space is a separable Banach space linearly isometric to L p ( Ω ) L^{p}(\Omega) L p ( Ω ) for some measure space Ω \Omega Ω and some exponent p p p . A computable Banach-space presentation may be viewed as a program that, given a rational linear combination of its distinguished dense sequence and a requested precision, computes the norm of that combination. An index is simply a natural-number code for such a program. Neither the presentation nor its index is unique. Given a class of Banach spaces, its index set is the collection of all codes whose programs present members of that class. The position of this set in the arithmetical or analytical hierarchy measures how difficult it is to recognize the class from an effective description. Naively, deciding whether an indexed Banach space is a Lebesgue space appears to require a search through the uncountably many possible surjective linear isometries and hence to have the crude analytic upper bound Σ 1 1 \boldsymbol{\Sigma}^{1}_{1} Σ 1 1 .
Theorem 1.2 (Brown–McNicholl–Melnikov [7 ] ). The following statements hold.
(i) The set of all indices of Lebesgue-space presentations, with the exponent allowed to vary, belongs to Π 3 0 \Pi^{0}_{3} Π 3 0 .
(ii) For every fixed computable real p ≥ 1 p \ge1 p ≥ 1 , the set of all indices of L p L^{p} L p -space presentations is Π 2 0 \Pi^{0}_{2} Π 2 0 -complete.
(iii) For every fixed computable real p ≥ 1 p \ge1 p ≥ 1 , the set of all indices of presentations linearly isometric to L p [ 0 , 1 ] L^{p}[0,1] L p [ 0 , 1 ] is Π 2 0 \Pi^{0}_{2} Π 2 0 -complete.
A one-sided recovery theorem for the uncomputable exponents is also the starting point for our work.
Theorem 1.3 ([7 ] ). Let B B B be a L p L^{p} L p -space (1 ≤ p ≤ 2 1 \le p \le2 1 ≤ p ≤ 2 ) of dimension at least two as a linear space. Uniformly from a presentation of B B B , one can computably enumerate { r ∈ Q : p < r } \{r \in\mathbb{Q}: p < r\} { r ∈ Q : p < r } .
A real number x x x is right-c.e. if { r ∈ Q : x < r } \{r \in\mathbb{Q}: x < r\} { r ∈ Q : x < r } is c.e., and it is left-c.e. if { r ∈ Q : r < x } \{r \in\mathbb{Q}: r < x\} { r ∈ Q : r < x } is c.e. It is computable if it is both left-c.e. and right-c.e. Consequently, if 1 < p < 2 1 < p < 2 1 < p < 2 and an L p L^{p} L p -space of dimension at least two is computably presentable, then p p p must be right-c.e. The restriction on the dimension is necessary, since every one-dimensional L p L^{p} L p -space is linearly isometric to the scalar field and therefore carries no information about p p p .
Theorem 1.4 (McNicholl [24 ] ). The following statements hold.
(i) If 2 ≤ p < ∞ 2 \le p < \infty 2 ≤ p < ∞ and a computably presentable L p L^{p} L p -space has dimension at least two, then p p p is computable.
(ii) If n ≥ 2 n \ge2 n ≥ 2 , 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ , and ℓ n p \ell^{p}_{n} ℓ n p is computably presentable, then p p p is computable.
McNicholl also proved that exponent recovery is at least nonuniform.
Theorem 1.5 ([24 ] ). For every nondegenerate open interval I ⊆ ( 1 , 2 ) I \subseteq(1,2) I ⊆ ( 1 , 2 ) , there is no partial computable function F F F such that, whenever it indexes a computable presentation of an L p L^{p} L p -space of dimension at least two with p ∈ I p \in I p ∈ I , F ( e ) F(e) F ( e ) is defined and indexes a computable Cauchy name for p p p .
The remaining case is that of infinite-dimensional separable L p L^p L p -spaces with 1 < p < 2 1 < p < 2 1 < p < 2 . The one-sided recovery theorem makes right-c.e.-ness of p p p necessary, while the nonuniformity theorem suggests that computability may not be necessary.
Conjecture (McNicholl [24 ] , 2020). If 1 < p < 2 1 < p < 2 1 < p < 2 is right-c.e., then L p [ 0 , 1 ] L^p[0,1] L p [ 0 , 1 ] is computably presentable.
This conjecture was later recorded as Conjecture 6.33 in the textbook by Franklin, Goldbring, and McNicholl [15 ] . It is also connected with the following question in effective duality, posed by Downey and Melnikov in their textbook [13 Question 4.2.116] ; a broader version was also posed by Downey, Greenberg, and Qian [12 Question 1] .
Question 1.6 (Downey–Melnikov). Suppose B B B is a computable Banach space and its dual B ∗ B^* B ∗ is separable. Must B ∗ B^* B ∗ have a computable presentation?
Downey and Melnikov observed that a positive solution of McNicholl’s conjecture would give a negative answer to Question 1.6: for a noncomputable right-c.e. p ∈ ( 1 , 2 ) p \in(1,2) p ∈ ( 1 , 2 ) , take a computable copy B B B of L p ( [ 0 , 1 ] ; R ) L^p([0,1];\mathbb{R}) L p ([ 0 , 1 ] ; R ) . Then B B B is reflexive and, by classical duality, B ∗ ≅ L q ( [ 0 , 1 ] ; R ) B^* \cong L^q([0,1];\mathbb{R}) B ∗ ≅ L q ([ 0 , 1 ] ; R ) , where q = p / ( p − 1 ) > 2 q = p/(p-1) > 2 q = p / ( p − 1 ) > 2 , so B ∗ B^* B ∗ is separable. If B ∗ B^* B ∗ had a computable presentation, Theorem 1.4 (i) would make q q q , and hence p = q / ( q − 1 ) p = q/(q - 1) p = q / ( q − 1 ) , computable, a contradiction. Our main theorem supplies such a counterexample.
Theorem 1.7. Let F \mathbb{F} F be either R \mathbb{R} R or C \mathbb{C} C . If 1 < p < 2 1 < p < 2 1 < p < 2 is right-c.e., then L p ( [ 0 , 1 ] ; F ) L^p([0,1];\mathbb{F}) L p ([ 0 , 1 ] ; F ) has a computable presentation.
The proof combines a locally uniform effective stability theorem with a varying-exponent Fraïssé-style construction. The purely atomic analogue behaves differently. We prove the following complementary result.
Theorem 1.8. Let F \mathbb{F} F be either R \mathbb{R} R or C \mathbb{C} C , and let 1 < p < 2 1 < p < 2 1 < p < 2 . If ℓ p ( F ) \ell^p(\mathbb{F}) ℓ p ( F ) has a computable presentation, then p p p is a computable real.
The proof follows by recovering p p p from the decay of symmetric second differences of the norm. For 1 < p < 2 1 < p < 2 1 < p < 2 , the argument requires only a single atomic coordinate and therefore extends to every separable L p L^p L p -space of dimension at least two with a nonzero atomic part. Together with the known cases, this shows that ℓ p \ell^p ℓ p is computably presentable exactly when p p p is computable, for every 1 ≤ p < ∞ 1 \leq p < \infty 1 ≤ p < ∞ . Combining these results with Theorem 1.7 and the classical classification yields a complete classification of computable presentability for separable L p L^p L p -spaces with 1 ≤ p < ∞ 1 \leq p < \infty 1 ≤ p < ∞ .
As a complementary application of the Fraïssé-style finite-dimensional construction method, we formulate an effective Fraïssé theorem for fixed families of finite-dimensional Banach spaces with effectively enumerable dense families of linear isometric embeddings. Applying it to finite-dimensional ℓ ∞ \ell_\infty ℓ ∞ -spaces gives a computable presentation of the Gurariĭ Banach space.
The paper is organized as follows. Section 2 collects the required background on computable analysis, measure theory and the geometry of Banach spaces. Section 3 establishes the locally uniform effective stability theorem, and Section 4 uses it to prove Theorem 1.7 . Section 5 proves exponent recovery for ℓ p \ell^p ℓ p , extends the argument to spaces with a nonzero atomic part, and completes the classification of computably presentable separable L p L^p L p -spaces. Finally, Section 6 proves the effective Fraïssé theorem for Banach spaces and applies it to the Gurariĭ space.
Preliminaries
A computable Polish presentation is a triple ( X , d , ( a i ) i ∈ N ) (X,d,(a_i)_{i\in\mathbb{N}}) ( X , d , ( a i ) i ∈ N ) , where ( X , d ) (X,d) ( X , d ) is complete, ( a i ) (a_i) ( a i ) is dense, and d ( a i , a j ) d(a_i,a_j) d ( a i , a j ) is a computable real uniformly in i , j i,j i , j . The points a i a_i a i are called special points . A fast Cauchy name for x ∈ X x\in X x ∈ X is a sequence ( i n ) (i_n) ( i n ) such that d ( a i n , a i n + 1 ) < 2 − n − 1 d(a_{i_n},a_{i_{n+1}})<2^{-n-1} d ( a i n , a i n + 1 ) < 2 − n − 1 and x = lim n a i n x=\lim_{n}a_{i_n} x = lim n a i n . A map between computable Polish spaces is computable if a single Turing functional transforms every fast Cauchy name of an input into a fast Cauchy name of its image. For i ∈ N i\in\mathbb{N} i ∈ N and r ∈ Q > 0 r\in\mathbb{Q}_{>0} r ∈ Q > 0 , let B ( i , r ) = { x ∈ X : d ( x , a i ) < r } B(i,r)=\{x\in X:d(x,a_i)<r\} B ( i , r ) = { x ∈ X : d ( x , a i ) < r } be a basic open ball . An open set is c.e. open if it is the union of a c.e. family of basic open balls. The effective compactness principle used throughout the proof is as follows.
Lemma 2.1 (Downey and Melnikov [14 ] ). Let X = ( X , d , ( a i ) i ∈ N ) X=(X,d,(a_i)_{i\in\mathbb{N}}) X = ( X , d , ( a i ) i ∈ N ) be a nonempty computable Polish space. The following are equivalent.
(i) Given n n n , one can compute a finite tuple ( i j , r j ) j < m (i_j,r_j)_{j<m} ( i j , r j ) j < m , where r j ∈ Q > 0 r_j\in\mathbb{Q}_{>0} r j ∈ Q > 0 and r j < 2 − n r_j<2^{-n} r j < 2 − n , such that X = ⋃ j < m B ( i j , r j ) X=\bigcup_{j<m}B(i_j,r_j) X = ⋃ j < m B ( i j , r j ) .
(ii) There is a partial computable functional Γ \Gamma Γ such that, for every index e e e enumerating a c.e. family U e U_e U e of basic open balls, Γ ( e ) \Gamma(e) Γ ( e ) is defined if and only if U e U_e U e covers X X X , and in that case Γ ( e ) \Gamma(e) Γ ( e ) returns a finite subfamily that still covers X X X .
The conversions between witnesses to these conditions are uniform. Any space satisfying these conditions is compact and is called computably compact.
Fact 2.2 ([14 ] ). The following operations are uniform.
(i) Finite products of computably compact spaces are computably compact.
(ii) If X X X is nonempty and computably compact and f : X → R f:X\to\mathbb{R} f : X → R is computable, then the values max X f \max_X f max X f and min X f \min_X f min X f are computable.
We now specialize the preceding framework to Banach spaces. Throughout the paper, fix a scalar field F ∈ { R , C } \mathbb{F}\in\{\mathbb{R},\mathbb{C}\} F ∈ { R , C } . Put Q R = Q \mathbb{Q}_{\mathbb{R}}=\mathbb{Q} Q R = Q and Q C = Q + i Q \mathbb{Q}_{\mathbb{C}}=\mathbb{Q}+i\mathbb{Q} Q C = Q + i Q .
Definition 2.3 (Pour-El and Richards [25 ] ). A computable presentation of a separable F \mathbb{F} F -Banach space $ B B B is a sequence ( b n ) n ∈ N (b_n)_{n\in\mathbb{N}} ( b n ) n ∈ N with dense linear span such that ∥ ∑ j < m q j b i j ∥ \left\lVert\sum_{j<m}q_jb_{i_j}\right\rVert ∑ j < m q j b i j is a computable real, uniformly in m m m , the indices i j i_j i j , and the coefficients q j ∈ Q F q_j\in\mathbb{Q}_{\mathbb{F}} q j ∈ Q F .
A fixed effective enumeration of all finite Q F \mathbb{Q}_{\mathbb{F}} Q F -linear combinations of the b n b_n b n ’s gives the special points of the associated computable Polish presentation. Hence the definition above is equivalent to a computable Polish presentation in which vector addition B × B → B B\times B\to B B × B → B and scalar multiplication F × B → B \mathbb{F}\times B\to B F × B → B , with the standard computable presentation of F \mathbb{F} F , are computable.
Unless explicitly stated, all finite-dimensional spaces below have positive dimension. Write L ( E , F ) \mathcal{L}(E,F) L ( E , F ) for the bounded linear operators and put m ( T ) = inf ∥ x ∥ = 1 ∥ T x ∥ m(T)=\inf_{\lVert x\rVert=1}\lVert Tx\rVert m ( T ) = inf ∥ x ∥ = 1 ∥ T x ∥ . A linear isomorphic embedding is a bounded linear map with m ( T ) > 0 m(T)>0 m ( T ) > 0 ; then m ( T ) = ∥ T − 1 : T ( E ) → E ∥ − 1 m(T)=\lVert T^{-1}:T(E)\to E\rVert^{-1} m ( T ) = ∥ T − 1 : T ( E ) → E ∥ − 1 . A linear map is an isometric embedding if it preserves norms. We write Emb ( E , F ) \operatorname{Emb}(E,F) Emb ( E , F ) for the set of linear isometric embeddings from E E E to F F F , equipped with the operator-norm metric. We write Iso ( E ) \operatorname{Iso}(E) Iso ( E ) for the group of surjective linear isometries from E E E to itself.
Call E = ( F d , ∥ ⋅ ∥ E ) E=(\mathbb{F}^{d},\lVert\cdot\rVert_E) E = ( F d , ∥ ⋅ ∥ E ) effective if its norm is computable uniformly on rational vectors. For nonzero E E E , one can uniformly find positive rationals a E , b E a_E,b_E a E , b E such that
a E ∥ x ∥ 2 ≤ ∥ x ∥ E ≤ b E ∥ x ∥ 2 . (1) a_E\lVert x\rVert_2\leq\lVert x\rVert_E\leq b_E\lVert x\rVert_2. \tag*{(1)} a E ∥ x ∥ 2 ≤ ∥ x ∥ E ≤ b E ∥ x ∥ 2 . (1)
Indeed, effectively choose rationals c j > ∥ e j ∥ E c_j>\|e_j\|_E c j > ∥ e j ∥ E for the standard basis vectors. The triangle and Cauchy–Schwarz inequalities give ∥ x ∥ E ≤ ( ∑ j = 1 d c j 2 ) 1 / 2 ∥ x ∥ 2 \|x\|_E\leq\left(\sum_{j=1}^d c_j^2\right)^{1/2}\|x\|_2 ∥ x ∥ E ≤ ( ∑ j = 1 d c j 2 ) 1/2 ∥ x ∥ 2 . Thus one can effectively choose a rational b E b_E b E satisfying the upper comparison bound; the same estimate is an effective Lipschitz bound for ∥ ⋅ ∥ E \|\mathord\cdot\|_E ∥ ⋅ ∥ E . Hence the minimum of ∥ ⋅ ∥ E \|\mathord\cdot\|_E ∥ ⋅ ∥ E on the computably compact Euclidean unit sphere is a computable real. The minimum is positive, so one can effectively choose a positive rational a E a_E a E below it.
The lemma below is a standard consequence of effective compactness.
Lemma 2.4 (Brattka [5 ] ; Downey and Melnikov [14 ] ). Let E E E and F F F be nonzero effective finite-dimensional spaces. From their norm algorithms and computable names for the matrix entries of T ∈ L ( E , F ) T\in\mathcal{L}(E,F) T ∈ L ( E , F ) , both ∥ T ∥ \|T\| ∥ T ∥ and m ( T ) m(T) m ( T ) are uniformly computable.
Let E E E be k k k -dimensional, where k ≥ 1 k\geq1 k ≥ 1 . A basis ( x i ) i < k (x_i)_{i<k} ( x i ) i < k of E E E is an Auerbach basis if its biorthogonal functionals ( x i ∗ ) i < k ⊆ E ∗ (x_i^*)_{i<k}\subseteq E^* ( x i ∗ ) i < k ⊆ E ∗ satisfy x i ∗ ( x j ) = δ i j x_i^*(x_j)=\delta_{ij} x i ∗ ( x j ) = δ ij and ∥ x i ∥ = ∥ x i ∗ ∥ = 1 \|x_i\|=\|x_i^*\|=1 ∥ x i ∥ = ∥ x i ∗ ∥ = 1 for all i , j < k i,j<k i , j < k .
Lemma 2.5 (Auerbach’s lemma; see Banach [2 ] ). Every nonzero finite-dimensional real or complex normed space E E E has an Auerbach basis. For such a basis ( x i ) i < k (x_i)_{i<k} ( x i ) i < k ,
max i < k ∣ a i ∣ ≤ ∥ ∑ i < k a i x i ∥ ≤ ∑ i < k ∣ a i ∣ ( a 0 , … , a k − 1 ∈ F ) . (2) \max_{i<k}|a_i|\leq\left\|\sum_{i<k}a_i x_i\right\|\leq\sum_{i<k}|a_i|\qquad(a_0,\ldots,a_{k-1}\in\mathbb{F}). \tag*{(2)} i < k max ∣ a i ∣ ≤ i < k ∑ a i x i ≤ i < k ∑ ∣ a i ∣ ( a 0 , … , a k − 1 ∈ F ) . (2) In particular, for every normed space F F F and S , T ∈ L ( E , F ) S,T\in\mathcal{L}(E,F) S , T ∈ L ( E , F ) ,
∥ S − T ∥ ≤ ∑ i < k ∥ ( S − T ) x i ∥ ≤ k max i < k ∥ ( S − T ) x i ∥ . \|S-T\|\leq\sum_{i<k}\|(S-T)x_i\|\leq k\max_{i<k}\|(S-T)x_i\|. ∥ S − T ∥ ≤ i < k ∑ ∥ ( S − T ) x i ∥ ≤ k i < k max ∥ ( S − T ) x i ∥.
Proof of the final assertion. For x = ∑ i a i x i x=\sum_i a_i x_i x = ∑ i a i x i with ∥ x ∥ ≤ 1 \|x\|\leq1 ∥ x ∥ ≤ 1 , (2) gives ∣ a i ∣ ≤ 1 |a_i|\leq1 ∣ a i ∣ ≤ 1 for every i i i ; apply the triangle inequality and take the supremum over x x x . □ \square □
Definition 2.6. Let T : E → F T:E\to F T : E → F be a linear isomorphic embedding. Its logarithmic distortion is
ldis ( T ) = sup x ≠ 0 ∣ log ∥ T x ∥ ∥ x ∥ ∣ . \operatorname{ldis}(T)=\sup_{x\ne0}\left|\log\frac{\|Tx\|}{\|x\|}\right|. ldis ( T ) = x = 0 sup log ∥ x ∥ ∥ T x ∥ . For δ > 0 \delta>0 δ > 0 , write Emb δ log ( E , F ) \operatorname{Emb}^{\log}_{\delta}(E,F) Emb δ l o g ( E , F ) for the set of linear isomorphic embeddings T : E → F T:E\to F T : E → F with ldis ( T ) < δ \operatorname{ldis}(T)<\delta ldis ( T ) < δ .
Fact 2.7. Let S S S and T T T be composable linear isomorphic embeddings. Then ldis ( S T ) ≤ ldis ( S ) + ldis ( T ) \operatorname{ldis}(ST)\leq\operatorname{ldis}(S)+\operatorname{ldis}(T) ldis ( S T ) ≤ ldis ( S ) + ldis ( T ) . Also ldis ( T ) = max { log ∥ T ∥ , log ∥ T − 1 : T ( E ) → E } \operatorname{ldis}(T)=\max\{\log\|T\|,\log\|T^{-1}:T(E)\to E\} ldis ( T ) = max { log ∥ T ∥ , log ∥ T − 1 : T ( E ) → E } , hence ldis ( T − 1 ) = ldis ( T ) \operatorname{ldis}(T^{-1})=\operatorname{ldis}(T) ldis ( T − 1 ) = ldis ( T ) .
For two k k k -dimensional normed linear spaces E E E and F F F , their logarithmic Banach–Mazur distance (defined in [3 ] ) is
d B M ( E , F ) = inf T : E ≅ F log ( ∥ T ∥ ∥ T − 1 ∥ ) , d_{\mathrm{BM}}(E,F)=\inf_{T:E\cong F}\log(\|T\|\|T^{-1}\|), d BM ( E , F ) = T : E ≅ F inf log ( ∥ T ∥∥ T − 1 ∥ ) ,
where the infimum is over all linear isomorphisms from E E E onto F F F . With this logarithmic convention, d B M d_{\mathrm{BM}} d BM is a metric on the linear-isometry classes of k k k -dimensional normed spaces.
Fact 2.8. For nonzero finite-dimensional E E E and F F F of the same dimension,
d B M ( E , F ) = 2 inf T : E ≅ F ldis ( T ) . d_{\mathrm{BM}}(E,F)=2\inf_{T:E\cong F}\operatorname{ldis}(T). d BM ( E , F ) = 2 T : E ≅ F inf ldis ( T ) .
Proof. For every isomorphism T T T , put c = ( ∥ T − 1 ∥ / ∥ T ∥ ) 1 / 2 c = (\lVert T^{-1}\rVert/\lVert T\rVert)^{1/2} c = (∥ T − 1 ∥ / ∥ T ∥ ) 1/2 . Then ldis ( c T ) = 1 2 log ( ∥ T ∥ ∥ T − 1 ∥ ) ≤ ldis ( T ) \operatorname{ldis}(cT) = \frac{1}{2}\log(\lVert T\rVert\lVert T^{-1}\rVert) \le\operatorname{ldis}(T) ldis ( c T ) = 2 1 log (∥ T ∥ ∥ T − 1 ∥) ≤ ldis ( T ) ; take infima. □
The age Age ( E ) \operatorname{Age}(E) Age ( E ) of a Banach space E E E is the collection of linear-isometry classes of its nonzero finite-dimensional subspaces. For k ≥ 1 k \ge1 k ≥ 1 , Age k ( E ) \operatorname{Age}_k(E) Age k ( E ) is its k k k -dimensional part.
Lemma 2.9 [16 Example 2.31] . For every 1 < r < 2 1 < r < 2 1 < r < 2 and k ≥ 1 k \ge1 k ≥ 1 , Age k ( L r [ 0 , 1 ] ) \operatorname{Age}_k(L^r[0,1]) Age k ( L r [ 0 , 1 ]) is closed under Banach–Mazur limits.
Definition 2.10 (see [16 ] ). An infinite-dimensional separable Banach space B B B is approximately ultrahomogeneous if, for every E ∈ Age ( B ) E \in\operatorname{Age}(B) E ∈ Age ( B ) , every γ , η ∈ Emb ( E , B ) \gamma,\eta\in\operatorname{Emb}(E,B) γ , η ∈ Emb ( E , B ) , and every ϵ > 0 \epsilon> 0 ϵ > 0 , there is U ∈ Iso ( B ) U \in\operatorname{Iso}(B) U ∈ Iso ( B ) such that ∥ U γ − η ∥ < ϵ \lVert U\gamma-\eta\rVert< \epsilon ∥ U γ − η ∥ < ϵ .
Lemma 2.11 (Lusky [21 ] ). For every 1 < r < 2 1 < r < 2 1 < r < 2 , the space L r ( [ 0 , 1 ] ; F ) L^r([0,1];\mathbb{F}) L r ([ 0 , 1 ] ; F ) is approximately ultrahomogeneous.
A measurable space is standard Borel if its σ \sigma σ -algebra is generated by a Polish topology. A measurable set A A A with 0 < μ ( A ) < ∞ 0 < \mu(A) < \infty 0 < μ ( A ) < ∞ is an atom if every measurable B ⊆ A B \subseteq A B ⊆ A has measure 0 0 0 or μ ( A ) \mu(A) μ ( A ) ; a measure with no atoms is atomless . The pushforward of μ \mu μ under a measurable map T T T is ( T ∗ μ ) ( E ) = μ ( T − 1 ( E ) ) (T_*\mu)(E) = \mu(T^{-1}(E)) ( T ∗ μ ) ( E ) = μ ( T − 1 ( E )) .
Lemma 2.12 (see [18 Theorem 17.41] ). *Let X X X be a standard Borel space carrying an atomless Borel probability measure μ \mu μ . Then, for every 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ , L r ( X , μ ; F ) L^r(X,\mu;\mathbb{F}) L r ( X , μ ; F ) is linearly isometric to L r ( [ 0 , 1 ] ; F ) L^r([0,1];\mathbb{F}) L r ([ 0 , 1 ] ; F ) .
For N ≥ 1 N \ge1 N ≥ 1 , partition [ 0 , 1 ] [0,1] [ 0 , 1 ] into intervals ( I N , j ) j < N (I_{N,j})_{j<N} ( I N , j ) j < N of length 1 / N 1/N 1/ N . For 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ , the map
( a j ) j < N ↦ N 1 / r ∑ j < N a j 1 I N , j (3) (a_j)_{j<N} \mapsto N^{1/r}\sum_{j<N}a_j1_{I_{N,j}} \tag*{(3)} ( a j ) j < N ↦ N 1/ r j < N ∑ a j 1 I N , j (3)
is a linear isometric embedding of ℓ r N \ell_r^N ℓ r N into L r [ 0 , 1 ] L^r[0,1] L r [ 0 , 1 ] . On [ 0 , 1 ] [0,1] [ 0 , 1 ] , λ \lambda λ denotes Lebesgue measure.
Lemma 2.13 (see [1 Proposition 6.1.3] ). Let 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ , let D n = { I 2 n , j : j < 2 n } \mathcal{D}_n = \{I_{2^n,j}: j < 2^n\} D n = { I 2 n , j : j < 2 n } , and define the dyadic conditional expectation by
P n f = 2 n ∑ A ∈ D n ( ∫ A f d λ ) 1 A . P_nf = 2^n\sum_{A\in\mathcal{D}_n}\left(\int_A f\,\mathrm{d}\lambda\right)1_A. P n f = 2 n A ∈ D n ∑ ( ∫ A f d λ ) 1 A . Then P n P_n P n is a linear projection of norm 1 1 1 on L r [ 0 , 1 ] L^r[0,1] L r [ 0 , 1 ] , and its range is linearly isometric to ℓ r 2 n \ell_r^{2^n} ℓ r 2 n via (3). For every finite-dimensional subspace E ⊆ L r [ 0 , 1 ] E \subseteq L^r[0,1] E ⊆ L r [ 0 , 1 ] , ∥ ( P n − Id ) ↾ E ∥ → 0 \lVert(P_n-\operatorname{Id})\mathbin{\upharpoonright}E\rVert\to0 ∥( P n − Id ) ↾ E ∥ → 0 .
Corollary 2.14. For 1 ≤ r < ∞ 1 \le r < \infty 1 ≤ r < ∞ , every E ∈ Age k ( L r [ 0 , 1 ] ) E \in\operatorname{Age}_k(L^r[0,1]) E ∈ Age k ( L r [ 0 , 1 ]) , and every η > 0 \eta> 0 η > 0 , there are N ≥ k N \ge k N ≥ k and a k k k -dimensional subspace F ⊆ ℓ r N F \subseteq\ell_r^N F ⊆ ℓ r N such that d B M ( E , F ) < η d_{\mathrm{BM}}(E,F) < \eta d BM ( E , F ) < η .
Proof. By Lemma 2.13, the restrictions P n ↾ E P_n\mathbin{\upharpoonright}E P n ↾ E are eventually injective and their norms and inverse norms tend to 1 1 1 . Identifying P n ( E ) P_n(E) P n ( E ) with a subspace of ℓ r 2 n \ell_r^{2^n} ℓ r 2 n therefore gives the assertion. □
The real case of the following theorem is due to Bretagnolle, Dacunha-Castelle, and Krivine [6 ] ; the complex case is due to Herz [17 ] and Rosenthal [27 ] .
Theorem 2.15. If 1 < u ≤ v ≤ 2 1 < u \le v \le2 1 < u ≤ v ≤ 2 , then there is T ∈ Emb ( L v ( [ 0 , 1 ] ; F ) , L u ( [ 0 , 1 ] ; F ) ) T \in\operatorname{Emb}(L^v([0,1];\mathbb{F}),L^u([0,1];\mathbb{F})) T ∈ Emb ( L v ([ 0 , 1 ] ; F ) , L u ([ 0 , 1 ] ; F )) . In particular, Age ( L v [ 0 , 1 ] ) ⊆ Age ( L u [ 0 , 1 ] ) \operatorname{Age}(L^v[0,1]) \subseteq\operatorname{Age}(L^u[0,1]) Age ( L v [ 0 , 1 ]) ⊆ Age ( L u [ 0 , 1 ]) .
We record the measure-theoretic form of the Plotkin–Rudin equimeasurability principle. The real form is due to Plotkin [26 ] , and the complex form to Rudin [28 ] .
Let n ≥ 1 n \ge1 n ≥ 1 , let 1 < r < ∞ 1 < r < \infty 1 < r < ∞ , and let μ \mu μ be a finite positive Borel measure on F n \mathbb{F}^n F n . We say that μ \mu μ has a finite r r r -moment if ∫ F n ∥ w ∥ 2 r d μ ( w ) < ∞ \int_{\mathbb{F}^n} \|w\|_2^r \,d\mu(w) < \infty ∫ F n ∥ w ∥ 2 r d μ ( w ) < ∞ . For such a measure, its r r r -characteristic is the function
μ ^ ( r ) ( s ) = ( ∫ F n ∣ 1 + ∑ i = 1 n s i w i ∣ r d μ ( w ) ) 1 / r , s = ( s 1 , … , s n ) ∈ F n . \widehat{\mu}^{(r)}(s)=\left(\int_{\mathbb{F}^n}\left|1+\sum_{i=1}^n s_iw_i\right|^r\,d\mu(w)\right)^{1/r}, \qquad s=(s_1,\ldots,s_n)\in\mathbb{F}^n. μ ( r ) ( s ) = ( ∫ F n 1 + i = 1 ∑ n s i w i r d μ ( w ) ) 1/ r , s = ( s 1 , … , s n ) ∈ F n .
Theorem 2.16 (Plotkin–Rudin). Let n ≥ 1 n \ge1 n ≥ 1 , and let 1 < r < ∞ 1 < r < \infty 1 < r < ∞ not be a positive even integer. If μ \mu μ and ν \nu ν are finite positive Borel measures on F n \mathbb{F}^n F n with finite r r r -moments and μ ^ ( r ) ( s ) = ν ^ ( r ) ( s ) \widehat{\mu}^{(r)}(s)=\widehat{\nu}^{(r)}(s) μ ( r ) ( s ) = ν ( r ) ( s ) for every s ∈ F n s\in\mathbb{F}^n s ∈ F n , then μ = ν \mu=\nu μ = ν .
An effective stability modulus
We obtain an effective strengthening of the qualitative approximate ultrahomogeneity in Lemma 2.11 : a computable stability modulus for almost isometric embeddings, depending only on the dimension, the desired accuracy, and a compact rational interval of exponents. Throughout this section fix a rational interval I = [ a , b ] ⊂ ( 1 , 2 ) I=[a,b]\subset(1,2) I = [ a , b ] ⊂ ( 1 , 2 ) .
Theorem 3.1. There is a total computable rational-valued function ω I : N ≥ 1 × Q > 0 → Q > 0 \omega_I:\mathbb{N}_{\ge1}\times\mathbb{Q}_{>0}\to\mathbb{Q}_{>0} ω I : N ≥ 1 × Q > 0 → Q > 0 such that, for every r ∈ I r\in I r ∈ I , ε ∈ Q > 0 \varepsilon\in\mathbb{Q}_{>0} ε ∈ Q > 0 , k ≥ 1 k\ge1 k ≥ 1 , E ∈ Age k ( L r [ 0 , 1 ] ) E\in\operatorname{Age}_k(L^r[0,1]) E ∈ Age k ( L r [ 0 , 1 ]) , and γ 0 , γ 1 ∈ Emb ω I ( k , ε ) log ( E , L r [ 0 , 1 ] ) \gamma_0,\gamma_1\in\operatorname{Emb}^{\log}_{\omega_I(k,\varepsilon)}(E,L^r[0,1]) γ 0 , γ 1 ∈ Emb ω I ( k , ε ) l o g ( E , L r [ 0 , 1 ]) , there is U ∈ Iso ( L r [ 0 , 1 ] ) U\in\operatorname{Iso}(L^r[0,1]) U ∈ Iso ( L r [ 0 , 1 ]) satisfying ∥ U γ 0 − γ 1 ∥ < ε \|U\gamma_0-\gamma_1\|<\varepsilon ∥ U γ 0 − γ 1 ∥ < ε .
Note that the algorithm computes the modulus only: it does not compute U U U . To control ∥ U γ 0 − γ 1 ∥ \|U\gamma_0-\gamma_1\| ∥ U γ 0 − γ 1 ∥ , it suffices to make U γ 0 x i U\gamma_0x_i U γ 0 x i close to γ 1 x i \gamma_1x_i γ 1 x i on an Auerbach basis ( x i ) i < k (x_i)_{i<k} ( x i ) i < k of E E E . The small distortions of the embeddings imply that corresponding linear combinations of these two tuples have nearly equal norms. To use this norm information to make the two tuples coordinatewise close after applying a suitable surjective linear isometry to one of them, we associate to each tuple a measure on a compact projective space whose cosine transform records the r r r th powers of these norms. Injectivity of the cosine transform and computable compactness then give an effective bound on the Wasserstein distance between the fixed-mass extensions of the two measures, uniform in r ∈ I r\in I r ∈ I . A coupling produces nearby isometric copies of the tuples on an auxiliary atomless probability space, and Lemma 2.11 transfers this approximation to the original tuples under some U ∈ Iso ( L r [ 0 , 1 ] ) U\in\operatorname{Iso}(L^r[0,1]) U ∈ Iso ( L r [ 0 , 1 ]) .
Let Q k Q_k Q k be the projective space of F k \mathbb{F}^k F k . Its points are one-dimensional subspaces [ u ] = F u [u]=\mathbb{F}u [ u ] = F u , with u u u chosen from the Euclidean unit sphere S ( F k ) S(\mathbb{F}^k) S ( F k ) . Put G F = { θ ∈ F : ∣ θ ∣ = 1 } G_{\mathbb{F}}=\{\theta\in\mathbb{F}:|\theta|=1\} G F = { θ ∈ F : ∣ θ ∣ = 1 } , so that G R = { − 1 , 1 } G_{\mathbb{R}}=\{-1,1\} G R = { − 1 , 1 } and G C G_{\mathbb{C}} G C is the unit circle. Equip Q k Q_k Q k with
d Q ( [ u ] , [ v ] ) = min θ ∈ G F ∥ u − θ v ∥ 2 . d_{Q}([u],[v])=\min_{\theta\in G_{\mathbb{F}}}\|u-\theta v\|_2. d Q ([ u ] , [ v ]) = θ ∈ G F min ∥ u − θ v ∥ 2 .
The identity
d Q ( [ u ] , [ v ] ) 2 = min θ ∈ G F ( ∥ u ∥ 2 2 + ∥ θ v ∥ 2 2 − 2 Re ⟨ u , θ v ⟩ ℓ 2 k ) = 2 − 2 max θ ∈ G F Re ⟨ u , θ v ⟩ ℓ 2 k = 2 − 2 ∣ ⟨ u , v ⟩ ℓ 2 k ∣ \begin{aligned}
d_Q([u],[v])^2
&=\min_{\theta\in G_{\mathbb{F}}}\left(\|u\|_2^2+\|\theta v\|_2^2-2\operatorname{Re}\langle u,\theta v\rangle_{\ell_2^k}\right)\\
&=2-2\max_{\theta\in G_{\mathbb{F}}}\operatorname{Re}\langle u,\theta v\rangle_{\ell_2^k}
=2-2\left|\langle u,v\rangle_{\ell_2^k}\right|
\end{aligned} d Q ([ u ] , [ v ] ) 2 = θ ∈ G F min ( ∥ u ∥ 2 2 + ∥ θ v ∥ 2 2 − 2 Re ⟨ u , θ v ⟩ ℓ 2 k ) = 2 − 2 θ ∈ G F max Re ⟨ u , θ v ⟩ ℓ 2 k = 2 − 2 ⟨ u , v ⟩ ℓ 2 k
gives diam ( Q k ) ≤ 2 \operatorname{diam}(Q_k)\le\sqrt{2} diam ( Q k ) ≤ 2 . Take as special points the classes of unit vectors in Q F k \mathbb{Q}_{\mathbb{F}}^k Q F k . It is clear from Lemma 2.1 (i) that Q k Q_k Q k is computably compact.
For a finite Borel measure σ \sigma σ on Q k Q_k Q k and r > 0 r>0 r > 0 , define the projective cosine transform
( C r σ ) ( z ) = ∫ Q k ∣ z ⋅ u ∣ r d σ ( [ u ] ) , z ∈ F k , (C_r\sigma)(z)=\int_{Q_k}|z\cdot u|^r\,d\sigma([u]),\qquad z\in\mathbb{F}^k, ( C r σ ) ( z ) = ∫ Q k ∣ z ⋅ u ∣ r d σ ([ u ]) , z ∈ F k ,
where z ⋅ u = ∑ j < k z j u j z\cdot u=\sum_{j<k}z_j u_j z ⋅ u = ∑ j < k z j u j . The absolute value makes the integrand independent of the unit representative of [ u ] [u] [ u ] . The following lemma is a consequence of Theorem 2.16 .
Lemma 3.2. For finite positive Borel measures σ \sigma σ and ν \nu ν on Q k Q_k Q k , if 1 < r < 2 1<r<2 1 < r < 2 and C r σ = C r ν C_r\sigma=C_r\nu C r σ = C r ν , then σ = ν \sigma=\nu σ = ν .
Proof. The space Q 1 Q_1 Q 1 is a singleton, so a finite measure on it is determined by its total mass. Hence assume k ≥ 2 k\ge2 k ≥ 2 . Put m = σ + ν m=\sigma+\nu m = σ + ν . We may assume that m ( Q k ) > 0 m(Q_k)>0 m ( Q k ) > 0 .
We first choose a nonzero functional c ∈ ( F k ) ∗ c\in(\mathbb{F}^k)^* c ∈ ( F k ) ∗ such that m ( H c ) = 0 m(H_c)=0 m ( H c ) = 0 where
H c = { [ u ] ∈ Q k : c ( u ) = 0 } . H_c=\{[u]\in Q_k:c(u)=0\}. H c = {[ u ] ∈ Q k : c ( u ) = 0 } .
Indeed, let B B B be the Euclidean unit ball in the underlying real vector space of ( F k ) ∗ (\mathbb{F}^k)^* ( F k ) ∗ , and let L B \mathcal{L}_B L B be Lebesgue measure restricted to B B B . The set { ( c , [ u ] ) ∈ B × Q k : c ( u ) = 0 } \{(c,[u])\in B\times Q_k:c(u)=0\} {( c , [ u ]) ∈ B × Q k : c ( u ) = 0 } is Borel. For each fixed nonzero u u u , the section { c ∈ B : c ( u ) = 0 } \{c\in B:c(u)=0\} { c ∈ B : c ( u ) = 0 } is contained in a proper real linear subspace. It is therefore L B \mathcal{L}_B L B -null. The Fubini-Tonelli theorem now yields
∫ B m ( H c ) d L B ( c ) = ∫ Q k L B { c ∈ B : c ( u ) = 0 } d m ( [ u ] ) = 0. \int_Bm(H_c)\,d\mathcal{L}_B(c)=\int_{Q_k}\mathcal{L}_B\{c\in B:c(u)=0\}\,dm([u])=0. ∫ B m ( H c ) d L B ( c ) = ∫ Q k L B { c ∈ B : c ( u ) = 0 } d m ([ u ]) = 0.
Thus m ( H c ) = 0 m(H_c)=0 m ( H c ) = 0 for almost every c ∈ B c\in B c ∈ B ; choose one such c c c . It is automatically nonzero. In particular, σ ( H c ) = ν ( H c ) = 0 \sigma(H_c)=\nu(H_c)=0 σ ( H c ) = ν ( H c ) = 0 . Complete c c c to a basis
c , d 1 , … , d k − 1 c,d_1,\ldots,d_{k-1} c , d 1 , … , d k − 1
of the F \mathbb{F} F -linear dual ( F k ) ∗ (\mathbb{F}^k)^* ( F k ) ∗ . On Q k ∖ H c Q_k\setminus H_c Q k ∖ H c , define the affine-coordinate map
χ c ( [ u ] ) = ( d i ( u ) c ( u ) ) 1 ≤ i < k ∈ F k − 1 . \chi_c([u])=\left(\frac{d_i(u)}{c(u)}\right)_{1\le i<k}\in\mathbb{F}^{k-1}. χ c ([ u ]) = ( c ( u ) d i ( u ) ) 1 ≤ i < k ∈ F k − 1 .
For v ∈ F k v\in\mathbb{F}^k v ∈ F k , define T ( v ) = ( c ( v ) , d 1 ( v ) , … , d k − 1 ( v ) ) T(v)=(c(v),d_1(v),\ldots,d_{k-1}(v)) T ( v ) = ( c ( v ) , d 1 ( v ) , … , d k − 1 ( v )) , then
w ↦ T − 1 ( 1 , w 1 , … , w k − 1 ) ∥ T − 1 ( 1 , w 1 , … , w k − 1 ) ∥ 2 w\mapsto\frac{T^{-1}(1,w_1,\ldots,w_{k-1})}{\left\|T^{-1}(1,w_1,\ldots,w_{k-1})\right\|_2} w ↦ ∥ T − 1 ( 1 , w 1 , … , w k − 1 ) ∥ 2 T − 1 ( 1 , w 1 , … , w k − 1 )
is the inverse of χ c \chi_c χ c . Hence χ c \chi_c χ c is a homeomorphism from Q k ∖ H c Q_k\setminus H_c Q k ∖ H c onto F k − 1 \mathbb{F}^{k-1} F k − 1 .
Define finite measures on Q k ∖ H c Q_k\setminus H_c Q k ∖ H c by d ρ σ ( [ u ] ) = ∣ c ( u ) ∣ r d σ ( [ u ] ) d\rho_\sigma([u])=|c(u)|^r\,d\sigma([u]) d ρ σ ([ u ]) = ∣ c ( u ) ∣ r d σ ([ u ]) , and d ρ ν ( [ u ] ) = ∣ c ( u ) ∣ r d ν ( [ u ] ) d\rho_\nu([u])=|c(u)|^r\,d\nu([u]) d ρ ν ([ u ]) = ∣ c ( u ) ∣ r d ν ([ u ]) , and let
α = ( χ c ) ∗ ρ σ , β = ( χ c ) ∗ ρ ν . \alpha=(\chi_c)_*\rho_\sigma,\qquad\beta=(\chi_c)_*\rho_\nu. α = ( χ c ) ∗ ρ σ , β = ( χ c ) ∗ ρ ν .
Since 1 < r < 2 1<r<2 1 < r < 2 , we have ∥ w ∥ 2 r ≤ ∑ i = 1 k − 1 ∣ w i ∣ r \|w\|_2^r\le\sum_{i=1}^{k-1}|w_i|^r ∥ w ∥ 2 r ≤ ∑ i = 1 k − 1 ∣ w i ∣ r for w ∈ F k − 1 w\in\mathbb{F}^{k-1} w ∈ F k − 1 . Hence
∫ F k − 1 ∥ w ∥ 2 r d α ( w ) ≤ ∑ i = 1 k − 1 ∫ F k − 1 ∣ w i ∣ r d α ( w ) = ∑ i = 1 k − 1 ∫ Q k ∖ H c ∣ d i ( u ) c ( u ) ∣ r ∣ c ( u ) ∣ r d σ ( [ u ] ) = ∑ i = 1 k − 1 ∫ Q k ∖ H c ∣ d i ( u ) ∣ r d σ ( [ u ] ) < ∞ . \begin{aligned}
\int_{\mathbb{F}^{k-1}}\|w\|_2^r\,d\alpha(w)
&\le\sum_{i=1}^{k-1}\int_{\mathbb{F}^{k-1}}|w_i|^r\,d\alpha(w)\\
&=\sum_{i=1}^{k-1}\int_{Q_k\setminus H_c}\left|\frac{d_i(u)}{c(u)}\right|^r|c(u)|^r\,d\sigma([u])\\
&=\sum_{i=1}^{k-1}\int_{Q_k\setminus H_c}|d_i(u)|^r\,d\sigma([u])<\infty.
\end{aligned} ∫ F k − 1 ∥ w ∥ 2 r d α ( w ) ≤ i = 1 ∑ k − 1 ∫ F k − 1 ∣ w i ∣ r d α ( w ) = i = 1 ∑ k − 1 ∫ Q k ∖ H c c ( u ) d i ( u ) r ∣ c ( u ) ∣ r d σ ([ u ]) = i = 1 ∑ k − 1 ∫ Q k ∖ H c ∣ d i ( u ) ∣ r d σ ([ u ]) < ∞.
The same calculation applies to β \beta β . Thus both measures have finite r r r -moments.
Fix s = ( s i ) 1 ≤ i < k ∈ F k − 1 s=(s_i)_{1\le i<k}\in\mathbb{F}^{k-1} s = ( s i ) 1 ≤ i < k ∈ F k − 1 , and set ℓ s = c + ∑ i = 1 k − 1 s i d i ∈ ( F k ) ∗ \ell_s=c+\sum_{i=1}^{k-1}s_i d_i\in(\mathbb{F}^k)^* ℓ s = c + ∑ i = 1 k − 1 s i d i ∈ ( F k ) ∗ . If ( e j ) j < k (e_j)_{j<k} ( e j ) j < k is the standard basis of F k \mathbb{F}^k F k , put z s = ( ℓ s ( e j ) ) j < k z_s=(\ell_s(e_j))_{j<k} z s = ( ℓ s ( e j ) ) j < k . Then F \mathbb{F} F -linearity gives ℓ s ( u ) = \ell_s(u)= ℓ s ( u ) = ∑ j < k ℓ s ( e j ) u j = z s ⋅ u \sum_{j<k}\ell_s(e_j)u_j=z_s\cdot u ∑ j < k ℓ s ( e j ) u j = z s ⋅ u for every u ∈ F k u\in\mathbb{F}^k u ∈ F k . Then the r r r -characteristic satisfies
( α ^ ( r ) ( s ) ) r = ∫ Q k ∖ H c ∣ 1 + ∑ i = 1 k − 1 s i d i ( u ) c ( u ) ∣ r ∣ c ( u ) ∣ r d σ ( [ u ] ) = ∫ Q k ∖ H c ∣ ℓ s ( u ) ∣ r d σ ( [ u ] ) = ∫ Q k ∣ z s ⋅ u ∣ r d σ ( [ u ] ) = ( C r σ ) ( z s ) . \begin{aligned}
\left(\widehat{\alpha}^{(r)}(s)\right)^r=\int_{Q_k\setminus H_c}\left|1+\sum_{i=1}^{k-1}s_i\frac{d_i(u)}{c(u)}\right|^r|c(u)|^r\,d\sigma([u]) \\
&=\int_{Q_k\setminus H_c}|\ell_s(u)|^r\,d\sigma([u])=\int_{Q_k}|z_s\cdot u|^r\,d\sigma([u])=(C_r\sigma)(z_s).
\end{aligned} ( α ( r ) ( s ) ) r = ∫ Q k ∖ H c 1 + i = 1 ∑ k − 1 s i c ( u ) d i ( u ) r ∣ c ( u ) ∣ r d σ ([ u ]) = ∫ Q k ∖ H c ∣ ℓ s ( u ) ∣ r d σ ([ u ]) = ∫ Q k ∣ z s ⋅ u ∣ r d σ ([ u ]) = ( C r σ ) ( z s ) .
The third equality uses σ ( H c ) = 0 \sigma(H_c)=0 σ ( H c ) = 0 . The identical computation with ν \nu ν gives ( β ^ ( r ) ( s ) ) r = ( C r ν ) ( z s ) \left(\widehat{\beta}^{(r)}(s)\right)^r=(C_r\nu)(z_s) ( β ( r ) ( s ) ) r = ( C r ν ) ( z s ) . The hypothesis C r σ = C r ν C_r\sigma=C_r\nu C r σ = C r ν therefore implies α ^ ( r ) = β ^ ( r ) \widehat{\alpha}^{(r)}=\widehat{\beta}^{(r)} α ( r ) = β ( r ) . Since 1 < r < 2 1<r<2 1 < r < 2 , Theorem 2.16 yields α = β \alpha=\beta α = β . Since χ c \chi_c χ c is a homeomorphism, ρ σ = ( χ c − 1 ) ∗ α \rho_\sigma=(\chi_c^{-1})_*\alpha ρ σ = ( χ c − 1 ) ∗ α and ρ ν = ( χ c − 1 ) ∗ β \rho_\nu=(\chi_c^{-1})_*\beta ρ ν = ( χ c − 1 ) ∗ β , hence ρ σ = ρ ν \rho_\sigma=\rho_\nu ρ σ = ρ ν . To remove the weight, put E n = { [ u ] ∈ Q k : ∣ c ( u ) ∣ ≥ 1 / n } E_n=\{[u]\in Q_k:|c(u)|\ge1/n\} E n = {[ u ] ∈ Q k : ∣ c ( u ) ∣ ≥ 1/ n } . For every Borel A ⊆ Q k A\subseteq Q_k A ⊆ Q k , the function ∣ c ∣ − r |c|^{-r} ∣ c ∣ − r is bounded on E n E_n E n , and hence
σ ( A ∩ E n ) = ∫ A ∩ E n ∣ c ∣ − r d ρ σ = ∫ A ∩ E n ∣ c ∣ − r d ρ ν = ν ( A ∩ E n ) . \sigma(A\cap E_n)=\int_{A\cap E_n}|c|^{-r}\,d\rho_\sigma=\int_{A\cap E_n}|c|^{-r}\,d\rho_\nu=\nu(A\cap E_n). σ ( A ∩ E n ) = ∫ A ∩ E n ∣ c ∣ − r d ρ σ = ∫ A ∩ E n ∣ c ∣ − r d ρ ν = ν ( A ∩ E n ) .
Finally, E n ↑ Q k ∖ H c E_n\uparrow Q_k\setminus H_c E n ↑ Q k ∖ H c , while both measures vanish on H c H_c H c . Continuity from below gives σ ( A ) = ν ( A ) \sigma(A)=\nu(A) σ ( A ) = ν ( A ) for every Borel A A A , and therefore σ = ν \sigma=\nu σ = ν . □
Adjoin to Q k Q_k Q k an isolated cemetery point ∂ \partial ∂ : define K k = Q k ⊔ { ∂ } K_k=Q_k\sqcup\{\partial\} K k = Q k ⊔ { ∂ } , and set d ( ∂ , q ) = 1 d(\partial,q)=1 d ( ∂ , q ) = 1 for q ∈ Q k q\in Q_k q ∈ Q k and d ( p , q ) = d Q ( p , q ) d(p,q)=d_{Q}(p,q) d ( p , q ) = d Q ( p , q ) for p , q ∈ Q k p,q\in Q_k p , q ∈ Q k . Then ( K k , d ) (K_k,d) ( K k , d ) is a computably compact metric space. For M > 0 M>0 M > 0 , let P M ( K k ) \mathcal{P}_M(K_k) P M ( K k ) be the space of finite Borel measures on K k K_k K k having total mass M M M , equipped with the 1 1 1 -Wasserstein metric
W 1 ( μ , ν ) = inf π ∈ Π ( μ , ν ) ∫ d ( x , y ) d π ( x , y ) , W_1(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)}\int d(x,y)\,d\pi(x,y), W 1 ( μ , ν ) = π ∈ Π ( μ , ν ) inf ∫ d ( x , y ) d π ( x , y ) ,
where Π ( μ , ν ) \Pi(\mu,\nu) Π ( μ , ν ) is the set of finite Borel measures on K k × K k K_k\times K_k K k × K k whose first and second marginals are respectively μ \mu μ and ν \nu ν . The point ∂ \partial ∂ is used only to supplement two measures to a common total mass.
Lemma 3.3 (Marcone and Valenti [22 ] ; see Downey and Melnikov [14 Lemma 4.50] ). P M ( K k ) \mathcal{P}_M(K_k) P M ( K k ) is computably compact, uniformly in k k k and rational M > 0 M>0 M > 0 . More precisely, let ( x j ) j ∈ N (x_j)_{j\in\mathbb{N}} ( x j ) j ∈ N consist of ∂ \partial ∂ and the classes of unit vectors in Q F k \mathbb{Q}_{\mathbb{F}}^k Q F k . The rational atomic measures
A k , M = { ∑ j < ℓ a j δ x i j : ℓ ≥ 1 , a j ∈ Q ≥ 0 , ∑ j < ℓ a j = M } \mathcal{A}_{k,M}=\left\{\sum_{j<\ell}a_j\delta_{x_{i_j}}:\ell\ge1,\ a_j\in\mathbb{Q}_{\ge0},\ \sum_{j<\ell}a_j=M\right\} A k , M = ⎩ ⎨ ⎧ j < ℓ ∑ a j δ x i j : ℓ ≥ 1 , a j ∈ Q ≥ 0 , j < ℓ ∑ a j = M ⎭ ⎬ ⎫ form an effective special sequence, with uniformly computable finite basic covers.
For r ∈ I r\in I r ∈ I , z ∈ S ( F k ) z\in S(\mathbb{F}^k) z ∈ S ( F k ) , define
φ r , z ( [ u ] ) = ∣ z ⋅ u ∣ r \varphi_{r,z}([u])=|z\cdot u|^r φ r , z ([ u ]) = ∣ z ⋅ u ∣ r
and φ r , z ( ∂ ) = 0 \varphi_{r,z}(\partial)=0 φ r , z ( ∂ ) = 0 .
The map ( r , z , q ) ↦ φ r , z ( q ) (r,z,q)\mapsto\varphi_{r,z}(q) ( r , z , q ) ↦ φ r , z ( q ) is computable. Since ∣ s r − t r ∣ ≤ b ∣ s − t ∣ |s^r-t^r|\le b|s-t| ∣ s r − t r ∣ ≤ b ∣ s − t ∣ for all s , t ∈ [ 0 , 1 ] s,t\in[0,1] s , t ∈ [ 0 , 1 ] , we have
∣ φ r , z ( q ) − φ r , z ( q ′ ) ∣ ≤ b d ( q , q ′ ) . |\varphi_{r,z}(q)-\varphi_{r,z}(q')|\le b\,d(q,q'). ∣ φ r , z ( q ) − φ r , z ( q ′ ) ∣ ≤ b d ( q , q ′ ) .
A fast W 1 W_1 W 1 -Cauchy name for μ ∈ P M ( K k ) \mu\in\mathcal{P}_M(K_k) μ ∈ P M ( K k ) supplies μ n ∈ A k , M \mu_n\in\mathcal{A}_{k,M} μ n ∈ A k , M with W 1 ( μ n , μ ) < 2 − n W_1(\mu_n,\mu)<2^{-n} W 1 ( μ n , μ ) < 2 − n . Fix any coupling π ∈ Π ( μ n , μ ) \pi\in\Pi(\mu_n,\mu) π ∈ Π ( μ n , μ ) . The preceding Lipschitz estimate gives
∣ ∫ K k φ r , z d μ n − ∫ K k φ r , z d μ ∣ = ∣ ∫ K k × K k ( φ r , z ( q ) − φ r , z ( q ′ ) ) d π ( q , q ′ ) ∣ ≤ b ∫ K k × K k d ( q , q ′ ) d π ( q , q ′ ) . \begin{aligned}
\left|\int_{K_k}\varphi_{r,z}\,d\mu_n-\int_{K_k}\varphi_{r,z}\,d\mu\right|
=\left|\int_{K_k\times K_k}\bigl(\varphi_{r,z}(q)-\varphi_{r,z}(q')\bigr)\,d\pi(q,q')\right| \\
&\le b\int_{K_k\times K_k}d(q,q')\,d\pi(q,q').
\end{aligned} ∫ K k φ r , z d μ n − ∫ K k φ r , z d μ = ∫ K k × K k ( φ r , z ( q ) − φ r , z ( q ′ ) ) d π ( q , q ′ ) ≤ b ∫ K k × K k d ( q , q ′ ) d π ( q , q ′ ) .
Taking the infimum over all π ∈ Π ( μ n , μ ) \pi\in\Pi(\mu_n,\mu) π ∈ Π ( μ n , μ ) yields
∣ ∫ K k ϕ r , z d μ n − ∫ K k ϕ r , z d μ ∣ ≤ b W 1 ( μ n , μ ) < b 2 − n . \left|\int_{K_k}\phi_{r,z}\,\mathrm{d}\mu_n-\int_{K_k}\phi_{r,z}\,\mathrm{d}\mu\right|\le bW_1(\mu_n,\mu)<b2^{-n}. ∫ K k ϕ r , z d μ n − ∫ K k ϕ r , z d μ ≤ b W 1 ( μ n , μ ) < b 2 − n .
Since μ n ∈ A k , M \mu_n\in\mathcal{A}_{k,M} μ n ∈ A k , M , the first integral is a finite sum computable uniformly in r r r , z z z , μ n \mu_n μ n . The displayed error bound is independent of r r r and z z z . Consequently the map ( r , z , μ ) ↦ ∫ K k ϕ r , z d μ (r,z,\mu)\mapsto\int_{K_k}\phi_{r,z}\,\mathrm{d}\mu ( r , z , μ ) ↦ ∫ K k ϕ r , z d μ is computable uniformly in k k k , M M M , r r r , z z z , and μ \mu μ , and hence so is ( r , z , μ , ν ) ↦ ∫ K k ϕ r , z d ( μ − ν ) (r,z,\mu,\nu)\mapsto\int_{K_k}\phi_{r,z}\,\mathrm{d}(\mu-\nu) ( r , z , μ , ν ) ↦ ∫ K k ϕ r , z d ( μ − ν ) . Define
D ( r , μ , ν ) = sup ∥ z ∥ 2 = 1 ∣ ∫ K k ϕ r , z d ( μ − ν ) ∣ = sup ∥ z ∥ 2 = 1 ∣ ( C r μ ∣ Q k ) ( z ) − ( C r ν ∣ Q k ) ( z ) ∣ . D(r,\mu,\nu)=\sup_{\lVert z\rVert_2=1}\left|\int_{K_k}\phi_{r,z}\,\mathrm{d}(\mu-\nu)\right|=\sup_{\lVert z\rVert_2=1}\left|(C_r\mu|_{Q_k})(z)-(C_r\nu|_{Q_k})(z)\right|. D ( r , μ , ν ) = ∥ z ∥ 2 = 1 sup ∫ K k ϕ r , z d ( μ − ν ) = ∥ z ∥ 2 = 1 sup ∣ ( C r μ ∣ Q k ) ( z ) − ( C r ν ∣ Q k ) ( z ) ∣ .
Lemma 3.4. For every k ≥ 1 k\ge1 k ≥ 1 and rational M > 0 M>0 M > 0 , D : I × P M ( K k ) 2 → R D:I\times\mathcal{P}_M(K_k)^2\to\mathbb{R} D : I × P M ( K k ) 2 → R is computable uniformly in k , M k,M k , M .
Proof. By the preceding integral computation, ∣ ∫ ϕ r , z d ( μ − ν ) ∣ \left|\int\phi_{r,z}\,\mathrm{d}(\mu-\nu)\right| ∫ ϕ r , z d ( μ − ν ) is computable uniformly. Since S ( F k ) S(\mathbb{F}^k) S ( F k ) is uniformly computably compact, its maximum D D D is computable by Fact 2.2(ii). □ \square □
Lemma 3.5. Given k ≥ 1 k\ge1 k ≥ 1 and positive rationals M , η M,\eta M , η , one can compute a positive rational τ = τ ( I , k , M , η ) \tau=\tau(I,k,M,\eta) τ = τ ( I , k , M , η ) such that D ( r , μ , ν ) < τ D(r,\mu,\nu)<\tau D ( r , μ , ν ) < τ implies W 1 ( μ , ν ) < η W_1(\mu,\nu)<\eta W 1 ( μ , ν ) < η for all r ∈ I r\in I r ∈ I and μ , ν ∈ P M ( K k ) \mu,\nu\in\mathcal{P}_M(K_k) μ , ν ∈ P M ( K k ) .
Proof. If D ( r , μ , ν ) = 0 D(r,\mu,\nu)=0 D ( r , μ , ν ) = 0 , the cosine transforms of μ ∣ Q k \mu|_{Q_k} μ ∣ Q k and ν ∣ Q k \nu|_{Q_k} ν ∣ Q k agree on S ( F k ) S(\mathbb{F}^k) S ( F k ) . For every finite measure σ \sigma σ on Q k Q_k Q k and z ≠ 0 z\ne0 z = 0 , ( C r σ ) ( z ) = ∥ z ∥ 2 r ( C r σ ) ( z ∥ z ∥ 2 ) (C_r\sigma)(z)=\lVert z\rVert_2^r(C_r\sigma)\left(\frac{z}{\lVert z\rVert_2}\right) ( C r σ ) ( z ) = ∥ z ∥ 2 r ( C r σ ) ( ∥ z ∥ 2 z ) . Thus the transforms agree on all of F k \mathbb{F}^k F k , and Lemma 3.2 gives μ ∣ Q k = ν ∣ Q k \mu|_{Q_k}=\nu|_{Q_k} μ ∣ Q k = ν ∣ Q k . Their common total mass M M M then gives μ = ν \mu=\nu μ = ν . Consequently the computable function
H ( r , μ , ν ) = D ( r , μ , ν ) + max { η − W 1 ( μ , ν ) , 0 } H(r,\mu,\nu)=D(r,\mu,\nu)+\max\{\eta-W_1(\mu,\nu),0\} H ( r , μ , ν ) = D ( r , μ , ν ) + max { η − W 1 ( μ , ν ) , 0 } is strictly positive on the computably compact space I × P M ( K k ) 2 I\times\mathcal{P}_M(K_k)^2 I × P M ( K k ) 2 . Its minimum m m m is therefore positive and computable by Fact 2.2(ii). Compute a rational 0 < τ < m 0<\tau<m 0 < τ < m . Whenever W 1 ( μ , ν ) ≥ η W_1(\mu,\nu)\ge\eta W 1 ( μ , ν ) ≥ η , we have D ( r , μ , ν ) = H ( r , μ , ν ) ≥ m > τ D(r,\mu,\nu)=H(r,\mu,\nu)\ge m>\tau D ( r , μ , ν ) = H ( r , μ , ν ) ≥ m > τ . □ \square □
We now associate with each tuple in L r [ 0 , 1 ] k L^r[0,1]^k L r [ 0 , 1 ] k a finite measure that records its projective directions and their L r L^r L r -mass. Let f = ( f 0 , … , f k − 1 ) ∈ L r [ 0 , 1 ] k f=(f_0,\ldots,f_{k-1})\in L^r[0,1]^k f = ( f 0 , … , f k − 1 ) ∈ L r [ 0 , 1 ] k . Choose Borel representatives, write f ( t ) = ( f 0 ( t ) , … , f k − 1 ( t ) ) f(t)=(f_0(t),\ldots,f_{k-1}(t)) f ( t ) = ( f 0 ( t ) , … , f k − 1 ( t )) , and put A f = { t ∈ [ 0 , 1 ] : f ( t ) ≠ 0 } A_f=\{t\in[0,1]:f(t)\ne0\} A f = { t ∈ [ 0 , 1 ] : f ( t ) = 0 } . The projective spectral measure σ f \sigma_f σ f is the pushforward of ∥ f ( t ) ∥ 2 r d λ ( t ) \lVert f(t)\rVert_2^r\,\mathrm{d}\lambda(t) ∥ f ( t ) ∥ 2 r d λ ( t ) on A f A_f A f under the Borel map t ↦ f ( t ) ∥ f ( t ) ∥ 2 t\mapsto\frac{f(t)}{\lVert f(t)\rVert_2} t ↦ ∥ f ( t ) ∥ 2 f ( t ) . Changing the Borel representatives does not change σ f \sigma_f σ f . Since 1 < r < 2 1<r<2 1 < r < 2 ,
∥ σ f ∥ = σ f ( Q k ) = ∫ A f ∥ f ( t ) ∥ 2 r d λ ( t ) ≤ ∑ i < k ∥ f i ∥ r r < ∞ . \lVert\sigma_f\rVert=\sigma_f(Q_k)=\int_{A_f}\lVert f(t)\rVert_2^r\,\mathrm{d}\lambda(t)\le\sum_{i<k}\lVert f_i\rVert_r^r<\infty. ∥ σ f ∥ = σ f ( Q k ) = ∫ A f ∥ f ( t ) ∥ 2 r d λ ( t ) ≤ i < k ∑ ∥ f i ∥ r r < ∞.
For every z ∈ F k z\in\mathbb{F}^k z ∈ F k , the definition gives
( C r σ f ) ( z ) = ∫ A f ∣ z ⋅ f ( t ) ∣ r d λ ( t ) = ∥ ∑ i < k z i f i ∥ r r . (4) (C_r\sigma_f)(z)=\int_{A_f}|z\cdot f(t)|^r\,\mathrm{d}\lambda(t)=\left\lVert\sum_{i<k}z_if_i\right\rVert_r^r. \tag*{(4)} ( C r σ f ) ( z ) = ∫ A f ∣ z ⋅ f ( t ) ∣ r d λ ( t ) = i < k ∑ z i f i r r . (4)
Given M > 0 M>0 M > 0 with ∥ σ f ∥ ≤ M \lVert\sigma_f\rVert\le M ∥ σ f ∥ ≤ M , define its fixed-mass cemetery extension by
σ ^ f M = σ f + ( M − ∥ σ f ∥ ) δ ∂ ∈ P M ( K k ) , \widehat{\sigma}_f^M=\sigma_f+(M-\lVert\sigma_f\rVert)\delta_{\partial}\in\mathcal{P}_M(K_k), σ f M = σ f + ( M − ∥ σ f ∥) δ ∂ ∈ P M ( K k ) ,
where δ ∂ \delta_{\partial} δ ∂ is the Dirac measure at ∂ \partial ∂ .
Lemma 3.6. Let k ≥ 1 k \ge1 k ≥ 1 , 1 < r < 2 1 < r < 2 1 < r < 2 , M , η > 0 M,\eta> 0 M , η > 0 , f , g ∈ L r [ 0 , 1 ] k f,g \in L^r[0,1]^k f , g ∈ L r [ 0 , 1 ] k , and suppose ∥ σ f ∥ , ∥ σ g ∥ ≤ M \|\sigma_f\|,\|\sigma_g\| \le M ∥ σ f ∥ , ∥ σ g ∥ ≤ M . If W 1 ( σ ^ f M , σ ^ g M ) < η W_1(\widehat{\sigma}_f^M,\widehat{\sigma}_g^M) < \eta W 1 ( σ f M , σ g M ) < η , then there is U ∈ Iso ( L r [ 0 , 1 ] ) U \in\operatorname{Iso}(L^r[0,1]) U ∈ Iso ( L r [ 0 , 1 ]) such that max i < k ∥ U f i − g i ∥ r r < 2 η \max_{i<k}\|Uf_i-g_i\|_r^r < 2\eta max i < k ∥ U f i − g i ∥ r r < 2 η .
Proof. Choose a coupling π \pi π of the two cemetery extensions such that
c = ∫ d ( x , y ) d π ( x , y ) < η . c=\int d(x,y)\,\mathrm{d}\pi(x,y)<\eta. c = ∫ d ( x , y ) d π ( x , y ) < η .
Choose a Borel section s : Q k → S ( F k ) s:Q_k\to S(\mathbb{F}^k) s : Q k → S ( F k ) by selecting the unit representative whose first nonzero coordinate is positive real, and extend it to K k K_k K k by s ( ∂ ) = 0 s(\partial)=0 s ( ∂ ) = 0 . For x , y ∈ K k x,y\in K_k x , y ∈ K k , put
α ( x , y ) = ∑ i < k s i ( x ) s i ( y ) ‾ , ζ ( x , y ) = { α ( x , y ) / ∣ α ( x , y ) ∣ , α ( x , y ) ≠ 0 , 1 , α ( x , y ) = 0. \alpha(x,y)=\sum_{i<k}s_i(x)\overline{s_i(y)},\qquad
\zeta(x,y)=
\begin{cases}
\alpha(x,y)/|\alpha(x,y)|, & \alpha(x,y)\ne0,\\
1, & \alpha(x,y)=0.
\end{cases} α ( x , y ) = i < k ∑ s i ( x ) s i ( y ) , ζ ( x , y ) = { α ( x , y ) /∣ α ( x , y ) ∣ , 1 , α ( x , y ) = 0 , α ( x , y ) = 0.
This Borel phase belongs to G F G_F G F . If x , y ∈ Q k x,y\in Q_k x , y ∈ Q k , then
∥ ζ ( x , y ) s ( x ) − s ( y ) ∥ 2 2 = 2 − 2 ∣ α ( x , y ) ∣ = d Q ( x , y ) 2 . \|\zeta(x,y)s(x)-s(y)\|_2^2=2-2|\alpha(x,y)|=d_{Q}(x,y)^2. ∥ ζ ( x , y ) s ( x ) − s ( y ) ∥ 2 2 = 2 − 2∣ α ( x , y ) ∣ = d Q ( x , y ) 2 .
Since d Q ≤ 2 < 2 d_Q\le\sqrt{2}<2 d Q ≤ 2 < 2 , it follows that
∥ ζ ( x , y ) s ( x ) − s ( y ) ∥ 2 2 ≤ 2 r − 1 d ( x , y ) ( x , y ∈ K k ) . (5) \|\zeta(x,y)s(x)-s(y)\|_2^2\le2^{r-1}d(x,y)\qquad(x,y\in K_k). \tag*{(5)} ∥ ζ ( x , y ) s ( x ) − s ( y ) ∥ 2 2 ≤ 2 r − 1 d ( x , y ) ( x , y ∈ K k ) . (5)
The space
( Ω , P ) = ( K k × K k × [ 0 , 1 ] , M − 1 π ⊗ λ ) (\Omega,\mathbb{P})=(K_k\times K_k\times[0,1],M^{-1}\pi\otimes\lambda) ( Ω , P ) = ( K k × K k × [ 0 , 1 ] , M − 1 π ⊗ λ )
is a standard Borel probability space, and the Lebesgue factor makes its measure atomless. On this space define
F i ( x , y , t ) = M 1 / r ζ ( x , y ) s i ( x ) , G i ( x , y , t ) = M 1 / r s i ( y ) ( i < k ) . F_i(x,y,t)=M^{1/r}\zeta(x,y)s_i(x),\qquad G_i(x,y,t)=M^{1/r}s_i(y)\qquad(i<k). F i ( x , y , t ) = M 1/ r ζ ( x , y ) s i ( x ) , G i ( x , y , t ) = M 1/ r s i ( y ) ( i < k ) .
By Lemma 2.12 , identify L r ( Ω , P ) L^r(\Omega,\mathbb{P}) L r ( Ω , P ) with L r [ 0 , 1 ] L^r[0,1] L r [ 0 , 1 ] , using the same identification for all F i , G i F_i,G_i F i , G i . This preserves the norms of all linear combinations and the distances ∥ F i − G i ∥ r \|F_i-G_i\|_r ∥ F i − G i ∥ r , allowing us to apply the approximate ultrahomogeneity of L r [ 0 , 1 ] L^r[0,1] L r [ 0 , 1 ] below. For every z ∈ F k z\in\mathbb{F}^k z ∈ F k , the first marginal of π \pi π and (4) give
∥ ∑ i < k z i F i ∥ r r = ∫ K k × K k ∣ z ⋅ s ( x ) ∣ r d π ( x , y ) = ∫ K k ∣ z ⋅ s ( x ) ∣ r d σ ^ f M ( x ) = ∥ ∑ i < k z i f i ∥ r r . \left\|\sum_{i<k}z_iF_i\right\|_r^r
=\int_{K_k\times K_k}|z\cdot s(x)|^r\,\mathrm{d}\pi(x,y)
=\int_{K_k}|z\cdot s(x)|^r\,\mathrm{d}\widehat{\sigma}_f^M(x)
=\left\|\sum_{i<k}z_if_i\right\|_r^r. i < k ∑ z i F i r r = ∫ K k × K k ∣ z ⋅ s ( x ) ∣ r d π ( x , y ) = ∫ K k ∣ z ⋅ s ( x ) ∣ r d σ f M ( x ) = i < k ∑ z i f i r r .
The second marginal gives the analogous identity for G G G and g g g . Then (6) yields
∥ F i − G i ∥ r r = ∫ K k × K k ∣ ζ ( x , y ) s i ( x ) − s i ( y ) ∣ r d π ( x , y ) ≤ ∫ K k × K k ∥ ζ ( x , y ) s ( x ) − s ( y ) ∥ 2 r d π ( x , y ) ≤ 2 r − 1 c . \begin{aligned}
\|F_i-G_i\|_r^r
=\int_{K_k\times K_k}|\zeta(x,y)s_i(x)-s_i(y)|^r\,\mathrm{d}\pi(x,y) \\
&\le\int_{K_k\times K_k}\|\zeta(x,y)s(x)-s(y)\|_2^r\,\mathrm{d}\pi(x,y)
\le2^{r-1}c.
\end{aligned} ∥ F i − G i ∥ r r = ∫ K k × K k ∣ ζ ( x , y ) s i ( x ) − s i ( y ) ∣ r d π ( x , y ) ≤ ∫ K k × K k ∥ ζ ( x , y ) s ( x ) − s ( y ) ∥ 2 r d π ( x , y ) ≤ 2 r − 1 c .
The norm identities show that f i ↦ F i f_i\mapsto F_i f i ↦ F i extends to a linear isometric embedding of span { f i : i < k } \operatorname{span}\{f_i:i<k\} span { f i : i < k } into L r [ 0 , 1 ] L^r[0,1] L r [ 0 , 1 ] , and similarly for g i ↦ G i g_i\mapsto G_i g i ↦ G i . Consequently Lemma 2.11 implies that, for every u > 0 u>0 u > 0 , there are V , W ∈ Iso ( L r [ 0 , 1 ] ) V,W\in\operatorname{Iso}(L^r[0,1]) V , W ∈ Iso ( L r [ 0 , 1 ]) with
max i < k ∥ V f i − F i ∥ r < u , max i < k ∥ W g i − G i ∥ r < u . \max_{i<k}\|Vf_i-F_i\|_r<u,\qquad
\max_{i<k}\|Wg_i-G_i\|_r<u. i < k max ∥ V f i − F i ∥ r < u , i < k max ∥ W g i − G i ∥ r < u .
Put ρ = ( 2 r − 1 c ) 1 / r \rho=(2^{r-1}c)^{1/r} ρ = ( 2 r − 1 c ) 1/ r . Since c < η c<\eta c < η and r < 2 r<2 r < 2 , we have ρ < ( 2 η ) 1 / r \rho<(2\eta)^{1/r} ρ < ( 2 η ) 1/ r . Choose u > 0 u>0 u > 0 such that 2 u + ρ < ( 2 η ) 1 / r 2u+\rho<(2\eta)^{1/r} 2 u + ρ < ( 2 η ) 1/ r and set U = W − 1 V U=W^{-1}V U = W − 1 V . Then, for every i < k i<k i < k ,
∥ U f i − g i ∥ r = ∥ V f i − W g i ∥ r ≤ ∥ V f i − F i ∥ r + ∥ F i − G i ∥ r + ∥ G i − W g i ∥ r < 2 u + ρ < ( 2 η ) 1 / r . \begin{aligned}
\|Uf_i-g_i\|_r=\|Vf_i-Wg_i\|_r \\
&\le\|Vf_i-F_i\|_r+\|F_i-G_i\|_r+\|G_i-Wg_i\|_r<2u+\rho<(2\eta)^{1/r}.
\end{aligned} ∥ U f i − g i ∥ r = ∥ V f i − W g i ∥ r ≤ ∥ V f i − F i ∥ r + ∥ F i − G i ∥ r + ∥ G i − W g i ∥ r < 2 u + ρ < ( 2 η ) 1/ r .
Thus U ∈ Iso ( L r [ 0 , 1 ] ) U \in\operatorname{Iso}(L^r[0,1]) U ∈ Iso ( L r [ 0 , 1 ]) has the required property. □
Proof of Theorem 3.1. Given k ≥ 1 k \ge1 k ≥ 1 and ε ∈ Q > 0 \varepsilon\in\mathbb{Q}_{>0} ε ∈ Q > 0 , put
s = min { 1 , ε / k } , η = s 2 / 8. s=\min\{1,\varepsilon/k\}, \qquad\eta=s^2/8. s = min { 1 , ε / k } , η = s 2 /8.
and use Lemma 3.5 with M = 4 k M=4k M = 4 k to compute τ > 0 \tau>0 τ > 0 such that D < τ D<\tau D < τ implies W 1 < η W_1<\eta W 1 < η . Define
δ : = min { 1 , τ / ( 10 k ) } , ω I ( k , ε ) : = δ / 2. \delta:=\min\{1,\tau/(10k)\}, \qquad\omega_I(k,\varepsilon):=\delta/2. δ := min { 1 , τ / ( 10 k )} , ω I ( k , ε ) := δ /2.
Fix r ∈ I r\in I r ∈ I , E ∈ Age k ( L r [ 0 , 1 ] ) E\in\operatorname{Age}_k(L^r[0,1]) E ∈ Age k ( L r [ 0 , 1 ]) , and γ 0 , γ 1 ∈ Emb ω I ( k , ε ) log ( E , L r [ 0 , 1 ] ) \gamma_0,\gamma_1\in\operatorname{Emb}^{\log}_{\omega_I(k,\varepsilon)}(E,L^r[0,1]) γ 0 , γ 1 ∈ Emb ω I ( k , ε ) l o g ( E , L r [ 0 , 1 ]) . Since 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 and δ / 2 < log ( 1 + δ ) \delta/2<\log(1+\delta) δ /2 < log ( 1 + δ ) , for j = 0 , 1 j=0,1 j = 0 , 1 and every x ∈ E x\in E x ∈ E we have
( 1 + δ ) − 1 ∥ x ∥ ≤ ∥ γ j x ∥ ≤ ( 1 + δ ) ∥ x ∥ . (1+\delta)^{-1}\|x\|\le\|\gamma_jx\|\le(1+\delta)\|x\|. ( 1 + δ ) − 1 ∥ x ∥ ≤ ∥ γ j x ∥ ≤ ( 1 + δ ) ∥ x ∥.
Choose an Auerbach basis ( x i ) i < k (x_i)_{i<k} ( x i ) i < k and put f i = γ 0 x i f_i=\gamma_0x_i f i = γ 0 x i , g i = γ 1 x i g_i=\gamma_1x_i g i = γ 1 x i . Since the Auerbach basis is normalized, ∥ x i ∥ E = 1 \|x_i\|_E=1 ∥ x i ∥ E = 1 , so the upper bound on γ 0 \gamma_0 γ 0 gives ∥ f i ∥ r ≤ 1 + δ \|f_i\|_r\le1+\delta ∥ f i ∥ r ≤ 1 + δ for every i < k i<k i < k . Moreover, 0 < r / 2 < 1 0<r/2<1 0 < r /2 < 1 , so subadditivity of u ↦ u r / 2 u\mapsto u^{r/2} u ↦ u r /2 on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) gives, pointwise,
∥ f ( t ) ∥ 2 r = ( ∑ i < k ∣ f i ( t ) ∣ 2 ) r / 2 ≤ ∑ i < k ∣ f i ( t ) ∣ r . \|f(t)\|_2^r=\left(\sum_{i<k}|f_i(t)|^2\right)^{r/2}\le\sum_{i<k}|f_i(t)|^r. ∥ f ( t ) ∥ 2 r = ( i < k ∑ ∣ f i ( t ) ∣ 2 ) r /2 ≤ i < k ∑ ∣ f i ( t ) ∣ r .
Integrating and using the definition of σ f \sigma_f σ f , we obtain
∥ σ f ∥ = ∫ A f ∥ f ( t ) ∥ 2 r d λ ( t ) ≤ ∑ i < k ∫ 0 1 ∣ f i ( t ) ∣ r d λ ( t ) ≤ k ( 1 + δ ) r ≤ k 2 r ≤ 4 k , \|\sigma_f\|=\int_{A_f}\|f(t)\|_2^r\,d\lambda(t)\le\sum_{i<k}\int_0^1|f_i(t)|^r\,d\lambda(t)\le k(1+\delta)^r\le k2^r\le4k, ∥ σ f ∥ = ∫ A f ∥ f ( t ) ∥ 2 r d λ ( t ) ≤ i < k ∑ ∫ 0 1 ∣ f i ( t ) ∣ r d λ ( t ) ≤ k ( 1 + δ ) r ≤ k 2 r ≤ 4 k ,
where the last two inequalities use δ ≤ 1 \delta\le1 δ ≤ 1 and r < 2 r<2 r < 2 . The same argument applies to g g g , using the upper bound on γ 1 \gamma_1 γ 1 . Thus their cemetery extensions of mass 4 k 4k 4 k are defined.
For z ∈ S ( F k ) z\in S(F^k) z ∈ S ( F k ) , put x z = ∑ i z i x i x_z=\sum_i z_ix_i x z = ∑ i z i x i . By Lemma 2.5, ∥ x z ∥ E ≤ ∑ i ∣ z i ∣ ≤ k \|x_z\|_E\le\sum_i|z_i|\le\sqrt{k} ∥ x z ∥ E ≤ ∑ i ∣ z i ∣ ≤ k . By (5) and the bounds on γ 0 , γ 1 \gamma_0,\gamma_1 γ 0 , γ 1 , both ( C r σ f ) ( z ) (C_r\sigma_f)(z) ( C r σ f ) ( z ) and ( C r σ g ) ( z ) (C_r\sigma_g)(z) ( C r σ g ) ( z ) lie between ( 1 + δ ) − r ∥ x z ∥ E r (1+\delta)^{-r}\|x_z\|_E^r ( 1 + δ ) − r ∥ x z ∥ E r and ( 1 + δ ) r ∥ x z ∥ E r (1+\delta)^r\|x_z\|_E^r ( 1 + δ ) r ∥ x z ∥ E r . Since ϕ r , z ( ∂ ) = 0 \phi_{r,z}(\partial)=0 ϕ r , z ( ∂ ) = 0 and 1 < r < 2 1<r<2 1 < r < 2 , taking the supremum over z ∈ S ( F k ) z\in S(F^k) z ∈ S ( F k ) gives
D ( r , σ f 4 k ^ , σ g 4 k ^ ) ≤ k r / 2 ( ( 1 + δ ) r − ( 1 + δ ) − r ) ≤ k ( ( 1 + δ ) 2 − ( 1 + δ ) − 2 ) \begin{aligned}
D(r,\widehat{\sigma_f^{4k}},\widehat{\sigma_g^{4k}})\le k^{r/2}\bigl((1+\delta)^r-(1+\delta)^{-r}\bigr) \\
&\le k\bigl((1+\delta)^2-(1+\delta)^{-2}\bigr)
\end{aligned} D ( r , σ f 4 k , σ g 4 k ) ≤ k r /2 ( ( 1 + δ ) r − ( 1 + δ ) − r ) ≤ k ( ( 1 + δ ) 2 − ( 1 + δ ) − 2 )
≤ 5 k δ ≤ τ / 2 < τ , \le5k\delta\le\tau/2<\tau, ≤ 5 k δ ≤ τ /2 < τ ,
here ( 1 + δ ) 2 − 1 ≤ 3 δ (1+\delta)^2-1\le3\delta ( 1 + δ ) 2 − 1 ≤ 3 δ and 1 − ( 1 + δ ) − 2 ≤ 2 δ 1-(1+\delta)^{-2}\le2\delta 1 − ( 1 + δ ) − 2 ≤ 2 δ for 0 ≤ δ ≤ 1 0\le\delta\le1 0 ≤ δ ≤ 1 . The inverse modulus and Lemma 3.6 therefore give U ∈ Iso ( L r [ 0 , 1 ] ) U\in\operatorname{Iso}(L^r[0,1]) U ∈ Iso ( L r [ 0 , 1 ]) with ∥ U f i − g i ∥ r < 2 η \|Uf_i-g_i\|_r<2\eta ∥ U f i − g i ∥ r < 2 η for all i < k i<k i < k . Since 2 η = s 2 / 4 < 1 2\eta=s^2/4<1 2 η = s 2 /4 < 1 and r < 2 r<2 r < 2 ,
∥ U γ 0 − γ 1 ∥ ≤ k max i < k ∥ U f i − g i ∥ r < k ( 2 η ) 1 / r ≤ k 2 η = k s / 2 < ε . \|U\gamma_0-\gamma_1\|\le k\max_{i<k}\|Uf_i-g_i\|_r<k(2\eta)^{1/r}\le k\sqrt{2\eta}=ks/2<\varepsilon. ∥ U γ 0 − γ 1 ∥ ≤ k i < k max ∥ U f i − g i ∥ r < k ( 2 η ) 1/ r ≤ k 2 η = k s /2 < ε .
This proves the theorem. □
An effective varying-exponent Fraïssé-style construction
Fix a right-c.e. real p ∈ ( 1 , 2 ) p\in(1,2) p ∈ ( 1 , 2 ) , rationals 1 < a < p < b < 2 1<a<p<b<2 1 < a < p < b < 2 , and a computable nonincreasing sequence of rationals q m ↓ p q_m\downarrow p q m ↓ p in I = [ a , b ] I=[a,b] I = [ a , b ] . Using the stability modulus from § 3 , we construct X m = ℓ q m d m X_m=\ell_{q_m}^{d_m} X m = ℓ q m d m with rational bonding maps I m : X m → X m + 1 I_m:X_m\to X_{m+1} I m : X m → X m + 1 . Here an operator between spaces with specified coordinate bases is rational if its matrix has entries in Q F \mathbb{Q}_F Q F . Summable distortions will give a computable limiting norm and a completion B B B .
There are two kinds of requirements. Copy requirements embed each member of a Banach–Mazur dense family ( D e ) e ⊆ Age ( L p [ 0 , 1 ] ) (D_e)_{e}\subseteq\operatorname{Age}(L^p[0,1]) ( D e ) e ⊆ Age ( L p [ 0 , 1 ]) into X m X_m X m with vanishing distortion as m → ∞ m\to\infty m → ∞ . Pair requirements supply, for sufficiently undistorted γ , η : D e → X m \gamma,\eta:D_e\to X_m γ , η : D e → X m , an alternative map J : X m → X m + 1 J:X_m\to X_{m+1} J : X m → X m + 1 with J γ ≈ I m η J\gamma\approx I_m\eta J γ ≈ I m η . We use these alignments to build an approximate intertwining between a cofinal subsequence of the renormed spaces X ˉ m \bar X_m X ˉ m and finite-dimensional subspaces of L p [ 0 , 1 ] L^p[0,1] L p [ 0 , 1 ] . Its limit is a surjective linear isometry B → L p [ 0 , 1 ] B\to L^p[0,1] B → L p [ 0 , 1 ] . We adapt the Fraïssé-sequence construction in [16 ] , proof of Theorem 2.26 to varying exponents, using effective searches and summable distortions to obtain a computable limiting norm. The next lemma aligns finitely many pairs while embedding a common base space into the next exponent.
Lemma 4.1. Let r ′ , r ∈ I ∩ Q r',r\in I\cap\mathbb{Q} r ′ , r ∈ I ∩ Q with r ′ ≤ r r'\le r r ′ ≤ r , and let Y Y Y and a family ( E i ) 1 ≤ i ≤ s (E_i)_{1\le i\le s} ( E i ) 1 ≤ i ≤ s be nonzero effective finite-dimensional spaces in Age ( L p [ 0 , 1 ] ) \operatorname{Age}(L^p[0,1]) Age ( L p [ 0 , 1 ]) . Fix positive rationals θ , ξ i , β i \theta,\xi_i,\beta_i θ , ξ i , β i with β i < ω I ( dim E i , ξ i / 2 ) \beta_i<\omega_I(\dim E_i,\xi_i/2) β i < ω I ( dim E i , ξ i /2 ) . One can effectively find N ≥ 1 N\ge1 N ≥ 1 , a rational operator Φ ∈ Emb β i log ( Y , ℓ r ′ N ) \Phi\in\operatorname{Emb}^{\log}_{\beta_i}(Y,\ell^N_{r'}) Φ ∈ Emb β i l o g ( Y , ℓ r ′ N ) , and finite sets J i \mathcal{J}_i J i of rational operators in Emb β i log ( Y , ℓ r ′ N ) \operatorname{Emb}^{\log}_{\beta_i}(Y,\ell^N_{r'}) Emb β i l o g ( Y , ℓ r ′ N ) such that for all γ , η ∈ Emb β i log ( E i , Y ) \gamma,\eta\in\operatorname{Emb}^{\log}_{\beta_i}(E_i,Y) γ , η ∈ Emb β i l o g ( E i , Y ) , there is J ∈ J i J\in\mathcal{J}_i J ∈ J i such that ∥ J γ − Φ η ∥ < ξ i \lVert J\gamma-\Phi\eta\rVert<\xi_i ∥ J γ − Φ η ∥ < ξ i .
Proof. Put ω i = ω I ( dim E i , ξ i / 2 ) \omega_i=\omega_I(\dim E_i,\xi_i/2) ω i = ω I ( dim E i , ξ i /2 ) and β ^ i = ( β i + ω i ) / 2 \widehat{\beta}_i=(\beta_i+\omega_i)/2 β i = ( β i + ω i ) /2 . Choose a rational v > e θ v>e^\theta v > e θ and set ρ i = ξ i / ( 4 v ) \rho_i=\xi_i/(4v) ρ i = ξ i / ( 4 v ) . We first find finitely many rational centers whose open ρ i \rho_i ρ i -balls cover
K i = { T ∈ L ( E i , Y ) : ∥ T ∥ ≤ e β i , m ( T ) ≥ e − β i } . K_i=\{T\in\mathcal{L}(E_i,Y):\lVert T\rVert\le e^{\beta_i},\ m(T)\ge e^{-\beta_i}\}. K i = { T ∈ L ( E i , Y ) : ∥ T ∥ ≤ e β i , m ( T ) ≥ e − β i } .
For matrices over F \mathbb{F} F , a rational coordinate box is a product of bounded closed intervals with rational endpoints in the matrix coordinates. When F = C \mathbb{F}=\mathbb{C} F = C , the real and imaginary parts of each entry are treated as separate coordinates. It is clear by Lemma 2.1 (i) that rational coordinate boxes are computably compact in the operator-norm metric. Choose a rational coordinate box B i ⊇ K i \mathcal{B}_i\supseteq K_i B i ⊇ K i ; this is effective since every matrix entry of T ∈ K i T\in K_i T ∈ K i has modulus at most e β i b E i / a Y e^{\beta_i}b_{E_i}/a_Y e β i b E i / a Y by (1) . The computability of ∥ T ∥ \lVert T\rVert ∥ T ∥ and m ( T ) m(T) m ( T ) , given by Lemma 2.4 , makes B i ∖ K i \mathcal{B}_i\setminus K_i B i ∖ K i a c.e. open subset of B i \mathcal{B}_i B i . Enumerate finite sets N i \mathcal{N}_i N i of rational centers in Emb β i log ( E i , Y ) \operatorname{Emb}^{\log}_{\beta_i}(E_i,Y) Emb β i l o g ( E i , Y ) such that
B i ⊆ ( B i ∖ K i ) ∪ ⋃ c ∈ N i B ( c , ρ i ) \mathcal{B}_i\subseteq(\mathcal{B}_i\setminus K_i)\cup\bigcup_{c\in\mathcal{N}_i}B(c,\rho_i) B i ⊆ ( B i ∖ K i ) ∪ c ∈ N i ⋃ B ( c , ρ i )
using Lemma 2.1 (ii). This search terminates: K i K_i K i is compact and lies in the open set Emb β i log ( E i , Y ) \operatorname{Emb}^{\log}_{\beta_i}(E_i,Y) Emb β i l o g ( E i , Y ) , whose rational points are dense.
Having found these nets, search over N N N and rational matrices Φ , J i , c , d : Y → ℓ r ′ N \Phi,J_{i,c,d}:Y\to\ell^N_{r'} Φ , J i , c , d : Y → ℓ r ′ N in Emb θ log ( Y , ℓ r ′ N ) \operatorname{Emb}^{\log}_{\theta}(Y,\ell^N_{r'}) Emb θ l o g ( Y , ℓ r ′ N ) such that ∥ J i , c , d c − Φ d ∥ < ξ i / 2 \lVert J_{i,c,d}c-\Phi d\rVert<\xi_i/2 ∥ J i , c , d c − Φ d ∥ < ξ i /2 (c , d ∈ N i c,d\in\mathcal{N}_i c , d ∈ N i ). All these conditions are semidecidable. To prove termination, choose h ∈ Emb ( Y , L r [ 0 , 1 ] ) h\in\operatorname{Emb}(Y,L^r[0,1]) h ∈ Emb ( Y , L r [ 0 , 1 ]) . By Theorem 3.1 , for each c , d ∈ N i c,d\in\mathcal{N}_i c , d ∈ N i there is U i , c , d ∈ Iso ( L r [ 0 , 1 ] ) U_{i,c,d}\in\operatorname{Iso}(L^r[0,1]) U i , c , d ∈ Iso ( L r [ 0 , 1 ]) with ∥ U i , c , d h c − h d ∥ < ξ i / 2 \lVert U_{i,c,d}hc-hd\rVert<\xi_i/2 ∥ U i , c , d h c − h d ∥ < ξ i /2 . Compose h h h and the finitely many U i , c , d h U_{i,c,d}h U i , c , d h with S ∈ Emb ( L r [ 0 , 1 ] , L r ′ [ 0 , 1 ] ) S\in\operatorname{Emb}(L^r[0,1],L^{r'}[0,1]) S ∈ Emb ( L r [ 0 , 1 ] , L r ′ [ 0 , 1 ]) , supplied by Theorem 2.15 . Let W W W be the span of their ranges. A sufficiently fine dyadic conditional expectation P P P satisfies ∥ ( P − Id ) ∣ W ∥ < 1 − e − θ / 2 \lVert(P-\operatorname{Id})|_W\rVert<1-e^{-\theta/2} ∥( P − Id ) ∣ W ∥ < 1 − e − θ /2 by Lemma 2.13 . Put δ = ∥ ( P − Id ) ∣ W ∥ \delta=\lVert(P-\operatorname{Id})|_W\rVert δ = ∥( P − Id ) ∣ W ∥ . Each of the maps V = S h V=Sh V = S h and V = S U i , c , d h V=SU_{i,c,d}h V = S U i , c , d h belongs to Emb ( Y , W ) \operatorname{Emb}(Y,W) Emb ( Y , W ) . Since P P P is contractive, ∥ P V ∥ ≤ 1 \lVert PV\rVert\le1 ∥ P V ∥ ≤ 1 . Moreover, for every x ∈ Y x\in Y x ∈ Y we have
∥ P V x ∥ ≥ ∥ V x ∥ − ∥ ( P − Id ) V x ∥ ≥ ( 1 − δ ) ∥ x ∥ . \lVert PVx\rVert\ge\lVert Vx\rVert-\lVert(P-\operatorname{Id})Vx\rVert\ge(1-\delta)\lVert x\rVert. ∥ P V x ∥ ≥ ∥ V x ∥ − ∥( P − Id ) V x ∥ ≥ ( 1 − δ ) ∥ x ∥ .
Thus P V PV P V is an isomorphic embedding and ldis ( P V ) ≤ − log ( 1 − δ ) < θ / 2 \operatorname{ldis}(PV)\le-\log(1-\delta)<\theta/2 ldis ( P V ) ≤ − log ( 1 − δ ) < θ /2 . Moreover, the contractivity of P P P and the isometry of S S S give
∥ P S U i , c , d h c − P S h d ∥ ≤ ∥ U i , c , d h c − h d ∥ < ξ i / 2. \lVert PSU_{i,c,d}hc-PShd\rVert\le\lVert U_{i,c,d}hc-hd\rVert<\xi_i/2. ∥ P S U i , c , d h c − P S h d ∥ ≤ ∥ U i , c , d h c − h d ∥ < ξ i /2.
Identify ran P P P isometrically with a finite ℓ r ′ \ell_{r'} ℓ r ′ -space by Lemma 2.13 . These maps satisfy the proposed certificate with strict margins, so sufficiently close rational matrices do as well.
Set J i = { J i , c , d : c , d ∈ N i } \mathcal{J}_i=\{J_{i,c,d}:c,d\in\mathbb{N}_i\} J i = { J i , c , d : c , d ∈ N i } . Given γ , η ∈ Emb β i log ( E i , Y ) \gamma,\eta\in\operatorname{Emb}^{\log}_{\beta_i}(E_i,Y) γ , η ∈ Emb β i l o g ( E i , Y ) , choose net centers c , d c,d c , d within ρ i \rho_i ρ i of γ , η \gamma,\eta γ , η . Then ∥ J i , c , d γ − Φ η ∥ < 2 e θ ρ i + ξ i / 2 < ξ i \lVert J_{i,c,d}\gamma-\Phi\eta\rVert<2e^{\theta}\rho_i+\xi_i/2<\xi_i ∥ J i , c , d γ − Φ η ∥ < 2 e θ ρ i + ξ i /2 < ξ i .
We now specify the requirements and construct the limit. Enumerate all tuples ( s , k , N , A ) (s,k,N,A) ( s , k , N , A ) with s ∈ N s\in\mathbb{N} s ∈ N , 1 ≤ k ≤ N 1\leq k\leq N 1 ≤ k ≤ N , and A ∈ M N × k ( Q F ) A\in M_{N\times k}(\mathbb{Q}_{F}) A ∈ M N × k ( Q F ) of rank k k k , arranging that s ( e ) ≤ e s(e)\leq e s ( e ) ≤ e . Set
D e = ( F k ( e ) , x ↦ ∥ A e x ∥ ℓ q s ( e ) N ( e ) ) . (6) D_e=\left(\mathbb{F}^{k(e)},x\mapsto\lVert A_e x\rVert_{\ell^{N(e)}_{q_{s(e)}}}\right). \tag*{(6)} D e = ( F k ( e ) , x ↦ ∥ A e x ∥ ℓ q s ( e ) N ( e ) ) . (6)
This is a uniformly effective family of finite-dimensional spaces.
Lemma 4.2. Every D e D_e D e belongs to Age ( L p [ 0 , 1 ] ) \operatorname{Age}(L^p[0,1]) Age ( L p [ 0 , 1 ]) , and the family ( D e ) e (D_e)_e ( D e ) e is Banach–Mazur dense in every fixed-dimensional part of Age ( L p [ 0 , 1 ] ) \operatorname{Age}(L^p[0,1]) Age ( L p [ 0 , 1 ]) .
Proof. Membership follows from the normalized-indicator embedding (3) and Theorem 2.15 . For density, write E A t = ( F k , x ↦ ∥ A x ∥ ℓ t N ) E_A^t=(\mathbb{F}^k,x\mapsto\lVert Ax\rVert_{\ell^N_t}) E A t = ( F k , x ↦ ∥ A x ∥ ℓ t N ) for a full-rank matrix A ∈ M N × k ( F ) A\in M_{N\times k}(\mathbb{F}) A ∈ M N × k ( F ) . Given E ∈ Age k ( L p ) E\in\operatorname{Age}_k(L^p) E ∈ Age k ( L p ) and η > 0 \eta>0 η > 0 , Corollary 2.14 and a sufficiently small rational perturbation of a basis matrix give a full-rank rational A A A with d B M ( E , E A p ) < η / 2 d_{\mathrm{BM}}(E,E_A^p)<\eta/2 d BM ( E , E A p ) < η /2 . For q s ≥ p q_s\geq p q s ≥ p , the standard comparison of finite-dimensional ℓ t \ell_t ℓ t -norms gives, for every x ∈ F k x\in\mathbb{F}^k x ∈ F k ,
∥ A x ∥ q s ≤ ∥ A x ∥ p ≤ N 1 / p − 1 / q s ∥ A x ∥ q s . \lVert Ax\rVert_{q_s}\leq\lVert Ax\rVert_p\leq N^{1/p-1/q_s}\lVert Ax\rVert_{q_s}. ∥ A x ∥ q s ≤ ∥ A x ∥ p ≤ N 1/ p − 1/ q s ∥ A x ∥ q s . Hence d B M ( E A q s , E A p ) ≤ ( 1 / p − 1 / q s ) log N → 0 d_{\mathrm{BM}}(E_A^{q_s},E_A^p)\leq(1/p-1/q_s)\log N\to0 d BM ( E A q s , E A p ) ≤ ( 1/ p − 1/ q s ) log N → 0 . This proves the density.
Define
β k , h = 1 2 min { 1 , ω I ( k , 2 − h − 4 ) } . (7) \beta_{k,h}=\frac{1}{2}\min\{1,\omega_I(k,2^{-h-4})\}. \tag*{(7)} β k , h = 2 1 min { 1 , ω I ( k , 2 − h − 4 )} . (7)
Starting with d 0 = 1 d_0=1 d 0 = 1 , we recursively construct X m = ℓ q m d m X_m=\ell^{d_m}_{q_m} X m = ℓ q m d m and rational bonding maps I m : X m → X m + 1 I_m:X_m\to X_{m+1} I m : X m → X m + 1 .
Form the finite direct sum
Y m = X m ⊕ q m D 0 ⊕ q m ⋯ ⊕ q m D m , Y_m=X_m\oplus_{q_m}D_0\oplus_{q_m}\cdots\oplus_{q_m}D_m, Y m = X m ⊕ q m D 0 ⊕ q m ⋯ ⊕ q m D m ,
with its concatenated coordinate basis and norm
∥ ( x , u 0 , … , u m ) ∥ Y m = ( ∥ x ∥ X m q m + ∑ e = 0 m ∥ u e ∥ D e q m ) 1 / q m . \lVert(x,u_0,\ldots,u_m)\rVert_{Y_m}
=\left(\lVert x\rVert_{X_m}^{q_m}+\sum_{e=0}^{m}\lVert u_e\rVert_{D_e}^{q_m}\right)^{1/q_m}. ∥( x , u 0 , … , u m ) ∥ Y m = ( ∥ x ∥ X m q m + e = 0 ∑ m ∥ u e ∥ D e q m ) 1/ q m .
This norm is computable uniformly in m m m .
To see that Y m ∈ Age ( L q m [ 0 , 1 ] ) Y_m\in\operatorname{Age}(L^{q_m}[0,1]) Y m ∈ Age ( L q m [ 0 , 1 ]) , first note that, for e ≤ m e\leq m e ≤ m , the map x ↦ A e x x\mapsto A_e x x ↦ A e x followed by (3) embeds D e D_e D e isometrically into L q s ( e ) [ 0 , 1 ] L^{q_{s(e)}}[0,1] L q s ( e ) [ 0 , 1 ] . Since s ( e ) ≤ e ≤ m s(e)\leq e\leq m s ( e ) ≤ e ≤ m , we have q m ≤ q s ( e ) q_m\leq q_{s(e)} q m ≤ q s ( e ) , so Theorem 2.15 gives an isometric embedding of D e D_e D e into L q m [ 0 , 1 ] L^{q_m}[0,1] L q m [ 0 , 1 ] . The space X m = ℓ q m d m X_m=\ell^{d_m}_{q_m} X m = ℓ q m d m also embeds isometrically into L q m [ 0 , 1 ] L^{q_m}[0,1] L q m [ 0 , 1 ] by (3) . Partition [ 0 , 1 ] [0,1] [ 0 , 1 ] into m + 2 m+2 m + 2 intervals of equal length. By affine changes of variable, multiplication by ( m + 2 ) 1 / q m (m+2)^{1/q_m} ( m + 2 ) 1/ q m , and extension by zero, realize these embeddings on distinct intervals. Their sum is an isometric embedding of Y m Y_m Y m into L q m [ 0 , 1 ] L^{q_m}[0,1] L q m [ 0 , 1 ] .
Write a m ∈ Emb ( X m , Y m ) a_m\in\operatorname{Emb}(X_m,Y_m) a m ∈ Emb ( X m , Y m ) and b e , m ∈ Emb ( D e , Y m ) b_{e,m}\in\operatorname{Emb}(D_e,Y_m) b e , m ∈ Emb ( D e , Y m ) for the rational coordinate inclusions. Apply Lemma 4.1 once, with base Y m Y_m Y m , r = q m r=q_m r = q m , r ′ = q m + 1 r'=q_{m+1} r ′ = q m + 1 , θ = 2 − m − 4 \theta=2^{-m-4} θ = 2 − m − 4 , and one pair requirement for each ( e , h ) ∈ { 0 , … , m } 2 (e,h)\in\{0,\ldots,m\}^2 ( e , h ) ∈ { 0 , … , m } 2 , using domain D e D_e D e , β dim D e , h \beta_{\dim D_e,h} β d i m D e , h , and ξ = 2 − h − 3 \xi=2^{-h-3} ξ = 2 − h − 3 . For the resulting Φ : Y m → ℓ q m + 1 N \Phi:Y_m\to\ell^N_{q_{m+1}} Φ : Y m → ℓ q m + 1 N , set
d m + 1 = N , X m + 1 = ℓ q m + 1 N , I m = Φ a m , C e , m = Φ b e , m . d_{m+1}=N,\qquad X_{m+1}=\ell^N_{q_{m+1}},\qquad I_m=\Phi a_m,\qquad C_{e,m}=\Phi b_{e,m}. d m + 1 = N , X m + 1 = ℓ q m + 1 N , I m = Φ a m , C e , m = Φ b e , m .
Then I m ∈ Emb 2 − m − 4 log ( X m , X m + 1 ) I_m \in\operatorname{Emb}^{\log}_{2^{-m-4}}(X_m,X_{m+1}) I m ∈ Emb 2 − m − 4 l o g ( X m , X m + 1 ) . Applying the pair conclusion to a m γ , a m η a_m\gamma,a_m\eta a m γ , a m η and restricting each alternative map to X m X_m X m gives the following two properties:
( P m ) (P_m) ( P m ) : For e , h ≤ m e,h\le m e , h ≤ m and γ , η ∈ Emb β dim D e , h log ( D e , X m ) \gamma,\eta\in\operatorname{Emb}^{\log}_{\beta\dim D_{e,h}}(D_e,X_m) γ , η ∈ Emb β d i m D e , h l o g ( D e , X m ) , there is a rational J ∈ Emb 2 − m − 4 log ( X m , X m + 1 ) J\in\operatorname{Emb}^{\log}_{2^{-m-4}}(X_m,X_{m+1}) J ∈ Emb 2 − m − 4 l o g ( X m , X m + 1 ) such that ∥ J γ − I m η ∥ < 2 − h − 3 \lVert J\gamma-I_m\eta\rVert<2^{-h-3} ∥ J γ − I m η ∥ < 2 − h − 3 .
( C m ) (C_m) ( C m ) : For each e ≤ m e\le m e ≤ m , there is a rational C e , m ∈ Emb 2 − m − 4 log ( D e , X m + 1 ) C_{e,m}\in\operatorname{Emb}^{\log}_{2^{-m-4}}(D_e,X_{m+1}) C e , m ∈ Emb 2 − m − 4 l o g ( D e , X m + 1 ) .
Write I m , t = I t − 1 ⋯ I m I_{m,t}=I_{t-1}\cdots I_m I m , t = I t − 1 ⋯ I m for t > m t>m t > m and I m , m = I d I_{m,m}=I_d I m , m = I d . For x ≠ 0 x\ne0 x = 0 , successive terms of log ∥ I m , t x ∥ X t \log\lVert I_{m,t}x\rVert_{X_t} log ∥ I m , t x ∥ X t differ by less than 2 − t − 4 2^{-t-4} 2 − t − 4 . Thus
∥ x ∥ ∞ = lim t → ∞ ∥ I m , t x ∥ X t (8) \lVert x\rVert_\infty=\lim_{t\to\infty}\lVert I_{m,t}x\rVert_{X_t} \tag*{(8)} ∥ x ∥ ∞ = t → ∞ lim ∥ I m , t x ∥ X t (8)
defines a norm on X m X_m X m , and
e − 2 − m − 3 ∥ x ∥ X m ≤ ∥ x ∥ ∞ ≤ e 2 − m − 3 ∥ x ∥ X m . (9) e^{-2^{-m-3}}\lVert x\rVert_{X_m}\le\lVert x\rVert_\infty\le e^{2^{-m-3}}\lVert x\rVert_{X_m}. \tag*{(9)} e − 2 − m − 3 ∥ x ∥ X m ≤ ∥ x ∥ ∞ ≤ e 2 − m − 3 ∥ x ∥ X m . (9)
Let X ‾ m \overline{X}_m X m denote X m X_m X m with this limit norm. Then I m ∈ Emb ( X ‾ m , X ‾ m + 1 ) I_m\in\operatorname{Emb}(\overline{X}_m,\overline{X}_{m+1}) I m ∈ Emb ( X m , X m + 1 ) . Let B B B be the completion of the union V = ⋃ m X ‾ m V=\bigcup_m\overline{X}_m V = ⋃ m X m . Denote the canonical embeddings by ι m ∈ Emb ( X ‾ m , B ) \iota_m\in\operatorname{Emb}(\overline{X}_m,B) ι m ∈ Emb ( X m , B ) .
Lemma 4.3. B B B has a computable presentation.
Proof. Enumerate the pairs ( m , v ) (m,v) ( m , v ) with v ∈ Q d m v\in\mathbb{Q}^{d_m} v ∈ Q d m . Their images ι m v \iota_m v ι m v are dense in B B B , and any rational linear combination can be represented by a rational vector w w w in a single space X m X_m X m . Choose a rational R > ∥ w ∥ X m R>\lVert w\rVert_{X_m} R > ∥ w ∥ X m and put Δ t = ∥ I m , t w ∥ X t \Delta_t=\lVert I_{m,t}w\rVert_{X_t} Δ t = ∥ I m , t w ∥ X t for t ≥ m t\ge m t ≥ m . These numbers are uniformly computable, and e − 2 − t − 3 Δ t ≤ ∥ ι m w ∥ B ≤ e 2 − t − 3 Δ t e^{-2^{-t-3}}\Delta_t\le\lVert\iota_mw\rVert_B\le e^{2^{-t-3}}\Delta_t e − 2 − t − 3 Δ t ≤ ∥ ι m w ∥ B ≤ e 2 − t − 3 Δ t .
Since 2 − m − 3 ≤ 1 / 4 2^{-m-3}\le1/4 2 − m − 3 ≤ 1/4 , we have Δ t < 2 R \Delta_t<2R Δ t < 2 R ; also e 2 − t − 3 − 1 ≤ 2 − t − 2 e^{2^{-t-3}}-1\le2^{-t-2} e 2 − t − 3 − 1 ≤ 2 − t − 2 . Consequently
∣ ∥ ι m w ∥ B − Δ t ∣ ≤ ( e 2 − t − 3 − 1 ) Δ t ≤ R 2 − t − 1 ≤ R 2 − t . \left|\lVert\iota_mw\rVert_B-\Delta_t\right|\le\left(e^{2^{-t-3}}-1\right)\Delta_t\le R2^{-t-1}\le R2^{-t}. ∣ ∥ ι m w ∥ B − Δ t ∣ ≤ ( e 2 − t − 3 − 1 ) Δ t ≤ R 2 − t − 1 ≤ R 2 − t .
This gives an effective error bound for the norm of every rational linear combination, as required. □
Lemma 4.4. The following properties hold.
(i) For every m m m , X ‾ m ∈ Age ( L p [ 0 , 1 ] ) \overline{X}_m\in\operatorname{Age}(L^p[0,1]) X m ∈ Age ( L p [ 0 , 1 ]) .
(ii) For every E ∈ Age ( L p [ 0 , 1 ] ) E\in\operatorname{Age}(L^p[0,1]) E ∈ Age ( L p [ 0 , 1 ]) and every α > 0 \alpha>0 α > 0 , all sufficiently large t t t admit a map G ∈ Emb α log ( E , X t ) G\in\operatorname{Emb}^{\log}_{\alpha}(E,X_t) G ∈ Emb α l o g ( E , X t ) .
Proof. (i) Fix m m m . For t ≥ m t\ge m t ≥ m , the isometricity of the bonding maps for the limiting norms and (10) give
e − 2 − t − 3 ∥ x ∥ X ‾ m ≤ ∥ I m , t x ∥ X t ≤ e 2 − t − 3 ∥ x ∥ X ‾ m . e^{-2^{-t-3}}\lVert x\rVert_{\overline{X}_m}\le\lVert I_{m,t}x\rVert_{X_t}\le e^{2^{-t-3}}\lVert x\rVert_{\overline{X}_m}. e − 2 − t − 3 ∥ x ∥ X m ≤ ∥ I m , t x ∥ X t ≤ e 2 − t − 3 ∥ x ∥ X m .
Thus the image I m , t ( X m ) I_{m,t}(X_m) I m , t ( X m ) , with the norm inherited from X t X_t X t , has Banach–Mazur distance at most 2 − t − 2 2^{-t-2} 2 − t − 2 from X ‾ m \overline{X}_m X m . Every X t = ℓ q t d t X_t=\ell^{d_t}_{q_t} X t = ℓ q t d t belongs to Age ( L p [ 0 , 1 ] ) \operatorname{Age}(L^p[0,1]) Age ( L p [ 0 , 1 ]) by the normalized-indicator embedding (3) and Theorem 2.15 . Since 2 − t − 3 → 0 2^{-t-3}\to0 2 − t − 3 → 0 , the closedness of the fixed-dimensional age in Lemma 2.9 proves the assertion.
(ii) Fix E ∈ Age ( L p [ 0 , 1 ] ) E\in\operatorname{Age}(L^p[0,1]) E ∈ Age ( L p [ 0 , 1 ]) and α > 0 \alpha>0 α > 0 . Use Lemma 4.2 and Fact 2.8 to choose R ∈ Emb α / 2 log ( E , D d ) R\in\operatorname{Emb}^{\log}_{\alpha/2}(E,D_d) R ∈ Emb α /2 l o g ( E , D d ) . For sufficiently large t t t , we have t − 1 ≥ d t-1\ge d t − 1 ≥ d and 2 − t − 3 < α / 2 2^{-t-3}<\alpha/2 2 − t − 3 < α /2 ; then ( C t − 1 ) (C_{t-1}) ( C t − 1 ) gives the required map G = C d , t − 1 R G=C_{d,t-1}R G = C d , t − 1 R . □
Proof of Theorem 1.7 . By Lemma 4.3, it remains to identify B B B with L p ( [ 0 , 1 ] ; F ) L^p([0,1];\mathbb{F}) L p ([ 0 , 1 ] ; F ) . We do this by an approximate intertwining.
Fix a sequence ( z i ) i ∈ N (z_i)_{i\in\mathbb{N}} ( z i ) i ∈ N dense in the unit ball of L p [ 0 , 1 ] L^p[0,1] L p [ 0 , 1 ] . We construct strictly increasing indices m n m_n m n and f n ∈ Emb ( X ‾ m n , L p [ 0 , 1 ] ) f_n\in\operatorname{Emb}(\overline{X}_{m_n},L^p[0,1]) f n ∈ Emb ( X m n , L p [ 0 , 1 ]) such that
∥ f n + 1 I m n , m n + 1 − f n ∥ X ‾ m n , L p < 3 ⋅ 2 − n − 3 . (10) \left\|f_{n+1}I_{m_n,m_{n+1}}-f_n\right\|_{\overline{X}_{m_n},L^p}<3\cdot2^{-n-3}. \tag*{(10)} ∥ f n + 1 I m n , m n + 1 − f n ∥ X m n , L p < 3 ⋅ 2 − n − 3 . (10)
At the same time, writing E n = span ( f n ( X ‾ m n ) , z 0 , … , z n ) E_n=\operatorname{span}(f_n(\overline{X}_{m_n}),z_0,\ldots,z_n) E n = span ( f n ( X m n ) , z 0 , … , z n ) , we construct maps b n : E n → X ‾ m n + 1 b_n:E_n\to\overline{X}_{m_{n+1}} b n : E n → X m n + 1 with ∥ b n ∥ < 2 \|b_n\|<2 ∥ b n ∥ < 2 and
∥ f n + 1 b n − Id E n ∥ E n , L p < 2 − n − 3 . (11) \left\|f_{n+1}b_n-\operatorname{Id}_{E_n}\right\|_{E_n,L^p}<2^{-n-3}. \tag*{(11)} ∥ f n + 1 b n − Id E n ∥ E n , L p < 2 − n − 3 . (11)
Here E n E_n E n always carries its norm as a subspace of L p [ 0 , 1 ] L^p[0,1] L p [ 0 , 1 ] . Start with m 0 = 0 m_0=0 m 0 = 0 and any f 0 ∈ Emb ( X ‾ 0 , L p [ 0 , 1 ] ) f_0\in\operatorname{Emb}(\overline{X}_0,L^p[0,1]) f 0 ∈ Emb ( X 0 , L p [ 0 , 1 ]) , which exists by Lemma 4.4(i).
Suppose m n , f n m_n,f_n m n , f n have been chosen, and define E n E_n E n as above. Put β = β d m n , n \beta=\beta_{d_{m_n},n} β = β d m n , n . Since X ‾ m n ∈ Age ( L p [ 0 , 1 ] ) \overline{X}_{m_n}\in\operatorname{Age}(L^p[0,1]) X m n ∈ Age ( L p [ 0 , 1 ]) , choose a d m n d_{m_n} d m n -dimensional D e D_e D e and an isomorphism S ∈ Emb β / 4 log ( D e , X ‾ m n ) S\in\operatorname{Emb}^{\log}_{\beta/4}(D_e,\overline{X}_{m_n}) S ∈ Emb β /4 l o g ( D e , X m n ) using Lemma 4.2 and Fact 2.8. Choose a positive rational δ < min { β / 2 , ω I ( dim E n , 2 − n − 3 ) } \delta<\min\{\beta/2,\omega_I(\dim E_n,2^{-n-3})\} δ < min { β /2 , ω I ( dim E n , 2 − n − 3 )} . By Lemma 4.4(ii), we may take t > m n t>m_n t > m n sufficiently large that t ≥ e , n t\geq e,n t ≥ e , n , 2 − t − 3 < δ / 2 2^{-t-3}<\delta/2 2 − t − 3 < δ /2 , and there is g ∈ Emb δ / 2 log ( E n , X t ) g\in\operatorname{Emb}^{\log}_{\delta/2}(E_n,X_t) g ∈ Emb δ /2 l o g ( E n , X t ) . We apply ( P t ) (P_t) ( P t ) with domain D e D_e D e , accuracy index h = n h=n h = n , and maps g f n S g f_n S g f n S and I m n , t S I_{m_n,t}S I m n , t S . Their eligibility follows from
g f n S ∈ Emb δ / 2 + β / 4 log ( D e , X t ) ⊆ Emb β log ( D e , X t ) , g f_n S\in\operatorname{Emb}^{\log}_{\delta/2+\beta/4}(D_e,X_t)\subseteq\operatorname{Emb}^{\log}_{\beta}(D_e,X_t), g f n S ∈ Emb δ /2 + β /4 l o g ( D e , X t ) ⊆ Emb β l o g ( D e , X t ) ,
I m n , t S ∈ Emb 2 − t − 3 + β / 4 log ( D e , X t ) ⊆ Emb β log ( D e , X t ) . I_{m_n,t}S\in\operatorname{Emb}^{\log}_{2^{-t-3}+\beta/4}(D_e,X_t)\subseteq\operatorname{Emb}^{\log}_{\beta}(D_e,X_t). I m n , t S ∈ Emb 2 − t − 3 + β /4 l o g ( D e , X t ) ⊆ Emb β l o g ( D e , X t ) .
Thus ( P t ) (P_t) ( P t ) supplies a rational J ∈ Emb 2 − t − 4 log ( X t , X t + 1 ) J\in\operatorname{Emb}^{\log}_{2^{-t-4}}(X_t,X_{t+1}) J ∈ Emb 2 − t − 4 l o g ( X t , X t + 1 ) satisfying
∥ J g f n S − I t I m n , t S ∥ D e , X t + 1 < 2 − n − 3 . (12) \left\|Jg f_n S-I_tI_{m_n,t}S\right\|_{D_e,X_{t+1}}<2^{-n-3}. \tag*{(12)} ∥ J g f n S − I t I m n , t S ∥ D e , X t + 1 < 2 − n − 3 . (12)
Set m n + 1 = t + 1 m_{n+1}=t+1 m n + 1 = t + 1 and regard b n = J g b_n=Jg b n = J g as a map into X ‾ m n + 1 \overline{X}_{m_{n+1}} X m n + 1 . Since 2 ⋅ 2 − t − 4 = 2 − t − 3 < δ / 2 2\cdot2^{-t-4}=2^{-t-3}<\delta/2 2 ⋅ 2 − t − 4 = 2 − t − 3 < δ /2 , we have
b n ∈ Emb δ / 2 + 2 − t − 3 log ( E n , X ‾ m n + 1 ) ⊆ Emb δ log ( E n , X ‾ m n + 1 ) . b_n\in\operatorname{Emb}^{\log}_{\delta/2+2^{-t-3}}(E_n,\overline{X}_{m_{n+1}})\subseteq\operatorname{Emb}^{\log}_{\delta}(E_n,\overline{X}_{m_{n+1}}). b n ∈ Emb δ /2 + 2 − t − 3 l o g ( E n , X m n + 1 ) ⊆ Emb δ l o g ( E n , X m n + 1 ) .
In particular, ∥ b n ∥ < e δ < 2 \|b_n\|<e^\delta<2 ∥ b n ∥ < e δ < 2 . Composing (13) with S − 1 S^{-1} S − 1 and converting its target norm by (10) gives
∥ b n f n − I m n , m n + 1 ∥ X ‾ m n , X ‾ m n + 1 < e 2 − t − 4 2 − n − 3 ∥ S − 1 ∥ < e 2 − t − 4 + β / 4 2 − n − 3 < 2 − n − 2 , \begin{aligned}
\left\|b_nf_n-I_{m_n,m_{n+1}}\right\|_{\overline{X}_{m_n},\overline{X}_{m_{n+1}}}<e^{2^{-t-4}}2^{-n-3}\|S^{-1}\| \\
&<e^{2^{-t-4}+\beta/4}2^{-n-3}<2^{-n-2},
\end{aligned} ∥ b n f n − I m n , m n + 1 ∥ X m n , X m n + 1 < e 2 − t − 4 2 − n − 3 ∥ S − 1 ∥ < e 2 − t − 4 + β /4 2 − n − 3 < 2 − n − 2 ,
where 2 − t − 4 ≤ 1 / 8 2^{-t-4}\leq1/8 2 − t − 4 ≤ 1/8 and β < 1 \beta<1 β < 1 justify the last bound.
Choose j n ∈ Emb ( X ‾ m n + 1 , L p [ 0 , 1 ] ) j_n\in\operatorname{Emb}(\overline{X}_{m_{n+1}},L^p[0,1]) j n ∈ Emb ( X m n + 1 , L p [ 0 , 1 ]) by Lemma 4.4(i). The maps j n b n j_nb_n j n b n and Id E n \operatorname{Id}_{E_n} Id E n belong to Emb ω I ( dim E n , 2 − n − 3 ) log ( E n , L p [ 0 , 1 ] ) \operatorname{Emb}^{\log}_{\omega_I(\dim E_n,2^{-n-3})}(E_n,L^p[0,1]) Emb ω I ( d i m E n , 2 − n − 3 ) l o g ( E n , L p [ 0 , 1 ]) . Since p ∈ I p\in I p ∈ I and E n ∈ Age ( L p [ 0 , 1 ] ) E_n\in\operatorname{Age}(L^p[0,1]) E n ∈ Age ( L p [ 0 , 1 ]) , Theorem 3.1 supplies W n ∈ Iso ( L p [ 0 , 1 ] ) W_n\in\operatorname{Iso}(L^p[0,1]) W n ∈ Iso ( L p [ 0 , 1 ]) such that
∥ W n j n b n − Id E n ∥ E n , L p < 2 − n − 3 . \left\|W_nj_nb_n-\operatorname{Id}_{E_n}\right\|_{E_n,L^p}<2^{-n-3}. ∥ W n j n b n − Id E n ∥ E n , L p < 2 − n − 3 .
Put f n + 1 = W n j n ∈ Emb ( X ‾ m n + 1 , L p [ 0 , 1 ] ) f_{n+1}=W_nj_n\in\operatorname{Emb}(\overline{X}_{m_{n+1}},L^p[0,1]) f n + 1 = W n j n ∈ Emb ( X m n + 1 , L p [ 0 , 1 ]) . This proves (12). Furthermore,
∥ f n + 1 I m n , m n + 1 − f n ∥ ≤ ∥ f n + 1 ( I m n , m n + 1 − b n f n ) ∥ + ∥ ( f n + 1 b n − Id E n ) f n ∥ < 2 − n − 2 + 2 − n − 3 = 3 ⋅ 2 − n − 3 , \begin{aligned}
\left\|f_{n+1}I_{m_n,m_{n+1}}-f_n\right\|
\leq\left\|f_{n+1}(I_{m_n,m_{n+1}}-b_nf_n)\right\|+\left\|(f_{n+1}b_n-\operatorname{Id}_{E_n})f_n\right\| \\
&<2^{-n-2}+2^{-n-3}=3\cdot2^{-n-3},
\end{aligned} ∥ f n + 1 I m n , m n + 1 − f n ∥ ≤ ∥ f n + 1 ( I m n , m n + 1 − b n f n ) ∥ + ∥ ( f n + 1 b n − Id E n ) f n ∥ < 2 − n − 2 + 2 − n − 3 = 3 ⋅ 2 − n − 3 ,
which establishes (11) and completes the induction.
For x ∈ X ‾ m x \in\overline{X}_m x ∈ X m , define
F ι m x = lim n → ∞ m n ≥ m f n I m , m n x . F\iota_m x=\lim_{\substack{n\to\infty\\m_n\ge m}}f_nI_{m,m_n}x. F ι m x = n → ∞ m n ≥ m lim f n I m , m n x .
The limit exists by (10) and the summability of ( 2 − n − 3 ) (2^{-n-3}) ( 2 − n − 3 ) ; the identifications along the bonding maps make F F F a well-defined linear map on the dense union V V V . Each f n I m , m n f_nI_{m,m_n} f n I m , m n is isometric for the limiting norm, so F F F extends to F ∈ Emb ( B , L p [ 0 , 1 ] ) F\in\operatorname{Emb}(B,L^p[0,1]) F ∈ Emb ( B , L p [ 0 , 1 ]) . Telescoping (10) also gives
∥ F ι m n − f n ∥ X ‾ m n , L p ≤ ∑ j ≥ n 3 ⋅ 2 − j − 3 = 6 ⋅ 2 − n − 3 . (13) \lVert F\iota_{m_n}-f_n\rVert_{\overline{X}_{m_n},L^p}\leq\sum_{j\geq n}3\cdot2^{-j-3}=6\cdot2^{-n-3}. \tag*{(13)} ∥ F ι m n − f n ∥ X m n , L p ≤ j ≥ n ∑ 3 ⋅ 2 − j − 3 = 6 ⋅ 2 − n − 3 . (13)
To prove surjectivity, fix i i i and let n ≥ i n\geq i n ≥ i . Then z i ∈ E n z_i\in E_n z i ∈ E n and ∥ z i ∥ ≤ 1 \lVert z_i\rVert\leq1 ∥ z i ∥ ≤ 1 . By (11) and (13) ,
∥ F ι m n + 1 b n z i − z i ∥ ≤ 6 ⋅ 2 − n − 4 ∥ b n z i ∥ + ∥ f n + 1 b n z i − z i ∥ < 12 ⋅ 2 − n − 4 + 2 − n − 3 = 7 ⋅ 2 − n − 3 → 0. \begin{aligned}
\lVert F\iota_{m_{n+1}}b_nz_i-z_i\rVert\leq6\cdot2^{-n-4}\lVert b_nz_i\rVert+\lVert f_{n+1}b_nz_i-z_i\rVert\\
&<12\cdot2^{-n-4}+2^{-n-3}=7\cdot2^{-n-3}\to0.
\end{aligned} ∥ F ι m n + 1 b n z i − z i ∥ ≤ 6 ⋅ 2 − n − 4 ∥ b n z i ∥ + ∥ f n + 1 b n z i − z i ∥ < 12 ⋅ 2 − n − 4 + 2 − n − 3 = 7 ⋅ 2 − n − 3 → 0.
The range F ( B ) F(B) F ( B ) is closed because F F F is an isometric embedding and B B B is complete. It therefore contains every z i z_i z i , and hence the whole unit ball by density. As F ( B ) F(B) F ( B ) is a linear subspace, this proves F ( B ) = L p [ 0 , 1 ] F(B)=L^p[0,1] F ( B ) = L p [ 0 , 1 ] .
Exponent recovery in the atomic case and the classification
We prove Theorem 1.8 by extracting p p p from symmetric second differences of the norm. The argument uses only one atomic coordinate, so we formulate its geometric estimate for general L p L^p L p -spaces for use in the general classification below.
Lemma 5.1. Let r ∈ R r\in\mathbb{R} r ∈ R . Suppose ( β n ) n ≥ 1 (\beta_n)_{n\geq1} ( β n ) n ≥ 1 is a uniformly computable sequence of positive reals and there are constants c , C > 0 c,C>0 c , C > 0 and N ∈ N N\in\mathbb{N} N ∈ N such that c 2 − n r ≤ β n ≤ C 2 − n r c2^{-nr}\leq\beta_n\leq C2^{-nr} c 2 − n r ≤ β n ≤ C 2 − n r for all n ≥ N n\geq N n ≥ N . Then r r r is computable.
Proof. The reals α n = − n − 1 log 2 β n \alpha_n=-n^{-1}\log_2\beta_n α n = − n − 1 log 2 β n are uniformly computable, and ∣ α n − r ∣ ≤ D / n |\alpha_n-r|\leq D/n ∣ α n − r ∣ ≤ D / n for n ≥ N n\geq N n ≥ N and some fixed D > 0 D>0 D > 0 .
We next record the form of Clarkson’s inequality [8 ] used below, together with an elementary proof.
Lemma 5.2 (Clarkson’s inequality). Let 1 ≤ p ≤ 2 1\leq p\leq2 1 ≤ p ≤ 2 , and let B B B be a real or complex L p L^p L p -space. For all u , v ∈ B u,v\in B u , v ∈ B ,
∥ u + v ∥ B p + ∥ u − v ∥ B p ≤ 2 ( ∥ u ∥ B p + ∥ v ∥ B p ) . (14) \lVert u+v\rVert_B^p+\lVert u-v\rVert_B^p\leq2\left(\lVert u\rVert_B^p+\lVert v\rVert_B^p\right). \tag*{(14)} ∥ u + v ∥ B p + ∥ u − v ∥ B p ≤ 2 ( ∥ u ∥ B p + ∥ v ∥ B p ) . (14) If u ≠ 0 u\ne0 u = 0 , then also
0 ≤ ∥ u + v ∥ B + ∥ u − v ∥ B − 2 ∥ u ∥ B ≤ 2 p ∥ u ∥ B 1 − p ∥ v ∥ B p . (15) 0\leq\lVert u+v\rVert_B+\lVert u-v\rVert_B-2\lVert u\rVert_B\leq\frac{2}{p}\lVert u\rVert_B^{1-p}\lVert v\rVert_B^p. \tag*{(15)} 0 ≤ ∥ u + v ∥ B + ∥ u − v ∥ B − 2 ∥ u ∥ B ≤ p 2 ∥ u ∥ B 1 − p ∥ v ∥ B p . (15) Proof. Put q = p / 2 ∈ ( 0 , 1 ] q=p/2\in(0,1] q = p /2 ∈ ( 0 , 1 ] . For a , b ≥ 0 a,b\geq0 a , b ≥ 0 , ( a + b ) q ≤ a q + b q (a+b)^q\leq a^q+b^q ( a + b ) q ≤ a q + b q . For real or complex scalars α , β \alpha,\beta α , β , concavity of t ↦ t q t\mapsto t^q t ↦ t q and this inequality give
∣ α + β ∣ p + ∣ α − β ∣ p ≤ 2 ( ∣ α + β ∣ 2 + ∣ α − β ∣ 2 2 ) q |\alpha+\beta|^p+|\alpha-\beta|^p\leq2\left(\frac{|\alpha+\beta|^2+|\alpha-\beta|^2}{2}\right)^q ∣ α + β ∣ p + ∣ α − β ∣ p ≤ 2 ( 2 ∣ α + β ∣ 2 + ∣ α − β ∣ 2 ) q = 2 ( ∣ α ∣ 2 + ∣ β ∣ 2 ) q ≤ 2 ( ∣ α ∣ p + ∣ β ∣ p ) . =2\left(|\alpha|^2+|\beta|^2\right)^q\leq2\left(|\alpha|^p+|\beta|^p\right). = 2 ( ∣ α ∣ 2 + ∣ β ∣ 2 ) q ≤ 2 ( ∣ α ∣ p + ∣ β ∣ p ) . Applying this scalar inequality pointwise to representatives of u u u and v v v and integrating proves (15).
Now suppose u ≠ 0 u \ne0 u = 0 , and put S = ∥ u ∥ B p > 0 S=\lVert u\rVert_B^p>0 S = ∥ u ∥ B p > 0 and G ( s ) = s 1 / p G(s)=s^{1/p} G ( s ) = s 1/ p . Concavity of G G G gives the tangent inequality G ( t ) − G ( S ) ≤ G ′ ( S ) ( t − S ) G(t)-G(S)\le G'(S)(t-S) G ( t ) − G ( S ) ≤ G ′ ( S ) ( t − S ) for t ≥ 0 t\ge0 t ≥ 0 . Apply it with t = ∥ u + v ∥ B p t=\lVert u+v\rVert_B^p t = ∥ u + v ∥ B p and t = ∥ u − v ∥ B p t=\lVert u-v\rVert_B^p t = ∥ u − v ∥ B p , and add. By (15),
∥ u + v ∥ B p + ∥ u − v ∥ B p − 2 ∥ u ∥ B p ≤ G ′ ( S ) ( ∥ u + v ∥ B p + ∥ u − v ∥ B p − 2 S ) \lVert u+v\rVert_B^p+\lVert u-v\rVert_B^p-2\lVert u\rVert_B^p
\le G'(S)\left(\lVert u+v\rVert_B^p+\lVert u-v\rVert_B^p-2S\right) ∥ u + v ∥ B p + ∥ u − v ∥ B p − 2 ∥ u ∥ B p ≤ G ′ ( S ) ( ∥ u + v ∥ B p + ∥ u − v ∥ B p − 2 S )
≤ 2 G ′ ( S ) ∥ v ∥ B p = 2 p ∥ u ∥ B 1 − p ∥ v ∥ B p . \le2G'(S)\lVert v\rVert_B^p=\frac{2}{p}\lVert u\rVert_B^{1-p}\lVert v\rVert_B^p. ≤ 2 G ′ ( S ) ∥ v ∥ B p = p 2 ∥ u ∥ B 1 − p ∥ v ∥ B p .
The lower bound in (16) follows from the triangle inequality. □
Lemma 5.3. Let 1 < p < 2 1<p<2 1 < p < 2 , let B B B be an L p L^p L p -space over F \mathbb{F} F , and let x , y ∈ B x,y\in B x , y ∈ B be F \mathbb{F} F -linearly independent. Put φ ( λ ) = ∥ x + λ y ∥ B \varphi(\lambda)=\lVert x+\lambda y\rVert_B φ ( λ ) = ∥ x + λ y ∥ B for λ ∈ F \lambda\in\mathbb{F} λ ∈ F , and for real h > 0 h>0 h > 0 , let
Δ h φ ( λ ) = φ ( λ + h ) + φ ( λ − h ) − 2 φ ( λ ) . \Delta_h\varphi(\lambda)=\varphi(\lambda+h)+\varphi(\lambda-h)-2\varphi(\lambda). Δ h φ ( λ ) = φ ( λ + h ) + φ ( λ − h ) − 2 φ ( λ ) . For n ≥ 1 n\ge1 n ≥ 1 , define
K n = { [ − n , n ] , F = R , { λ ∈ C : ∣ Re λ ∣ , ∣ Im λ ∣ ≤ n } , F = C , K_n=
\begin{cases}
[-n,n], & \mathbb{F}=\mathbb{R},\\
\{\lambda\in\mathbb{C}:|\operatorname{Re}\lambda|,|\operatorname{Im}\lambda|\le n\}, & \mathbb{F}=\mathbb{C},
\end{cases} K n = { [ − n , n ] , { λ ∈ C : ∣ Re λ ∣ , ∣ Im λ ∣ ≤ n } , F = R , F = C , and M n = max λ ∈ K n Δ 2 − n φ ( λ ) M_n=\max_{\lambda\in K_n}\Delta_{2^{-n}}\varphi(\lambda) M n = max λ ∈ K n Δ 2 − n φ ( λ ) . Then 0 < M n ≤ C 2 − n p 0<M_n\le C2^{-np} 0 < M n ≤ C 2 − n p for some C > 0 C>0 C > 0 . If there is a linear isometric identification B ≅ F ⊕ p Y B\cong\mathbb{F}\oplus_p Y B ≅ F ⊕ p Y , with ∥ ( a , w ) ∥ p = ∣ a ∣ p + ∥ w ∥ Y p \lVert(a,w)\rVert^p=|a|^p+\lVert w\rVert_Y^p ∥( a , w ) ∥ p = ∣ a ∣ p + ∥ w ∥ Y p , under which y = ( y 0 , y ′ ) y=(y_0,y') y = ( y 0 , y ′ ) with y 0 ≠ 0 y_0\ne0 y 0 = 0 , then also M n ≥ c 2 − n p M_n\ge c2^{-np} M n ≥ c 2 − n p for some c > 0 c>0 c > 0 and all sufficiently large n n n . If x , y x,y x , y are finite Q F \mathbb{Q}_{\mathbb{F}} Q F -linear combinations of the elements of a computable presentation ( b j ) j ∈ N (b_j)_{j\in\mathbb{N}} ( b j ) j ∈ N of B B B , then ( M n ) (M_n) ( M n ) is uniformly computable.
Proof. Fix h > 0 h>0 h > 0 and put u = x + h y u=x+hy u = x + h y and v = x − h y v=x-hy v = x − h y . These vectors are linearly independent, since x = ( u + v ) / 2 x=(u+v)/2 x = ( u + v ) /2 and y = ( u − v ) / ( 2 h ) y=(u-v)/(2h) y = ( u − v ) / ( 2 h ) are linearly independent. Thus a = ∥ u ∥ B a=\lVert u\rVert_B a = ∥ u ∥ B and b = ∥ v ∥ B b=\lVert v\rVert_B b = ∥ v ∥ B are positive, and u / a u/a u / a and v / b v/b v / b are distinct unit vectors. Since B B B is strictly convex and a / ( a + b ) ∈ ( 0 , 1 ) a/(a+b)\in(0,1) a / ( a + b ) ∈ ( 0 , 1 ) ,
2 ∥ x ∥ B a + b = ∥ a a + b u a + b a + b v b ∥ B < 1. \frac{2\lVert x\rVert_B}{a+b}
=\left\lVert\frac{a}{a+b}\frac{u}{a}+\frac{b}{a+b}\frac{v}{b}\right\rVert_B<1. a + b 2 ∥ x ∥ B = a + b a a u + a + b b b v B < 1. Consequently,
Δ h φ ( 0 ) = ∥ x + h y ∥ B + ∥ x − h y ∥ B − 2 ∥ x ∥ B = a + b − 2 ∥ x ∥ B > 0. \Delta_h\varphi(0)=\lVert x+hy\rVert_B+\lVert x-hy\rVert_B-2\lVert x\rVert_B=a+b-2\lVert x\rVert_B>0. Δ h φ ( 0 ) = ∥ x + h y ∥ B + ∥ x − h y ∥ B − 2 ∥ x ∥ B = a + b − 2 ∥ x ∥ B > 0. Since 0 ∈ K n 0\in K_n 0 ∈ K n , this also gives M n > 0 M_n>0 M n > 0 for every n ≥ 1 n\ge1 n ≥ 1 . Moreover, φ \varphi φ is ∥ y ∥ B \lVert y\rVert_B ∥ y ∥ B -Lipschitz, so Δ h φ \Delta_h\varphi Δ h φ is 4 ∥ y ∥ B 4\lVert y\rVert_B 4 ∥ y ∥ B -Lipschitz, uniformly in h h h . Under the additional assumption on x , y x,y x , y , the norm algorithm computes φ ( λ ) \varphi(\lambda) φ ( λ ) uniformly for λ ∈ Q F \lambda\in\mathbb{Q}_{\mathbb{F}} λ ∈ Q F and computes ∥ y ∥ B \lVert y\rVert_B ∥ y ∥ B . The Lipschitz bound therefore makes φ \varphi φ a computable function on F \mathbb{F} F . Since the sets K n K_n K n are uniformly computably compact, M n M_n M n is computable uniformly in n n n , by Fact 2.2(ii).
For the upper estimate, put δ = dist ( x , F y ) > 0 \delta=\operatorname{dist}(x,\mathbb{F}y)>0 δ = dist ( x , F y ) > 0 . Apply Lemma 5.2 with ξ = x + λ y \xi=x+\lambda y ξ = x + λ y and η = h y \eta=hy η = h y . Since ∥ ξ ∥ B ≥ δ \lVert\xi\rVert_B\ge\delta ∥ ξ ∥ B ≥ δ , we obtain
Δ h φ ( λ ) ≤ 2 p ∥ ξ ∥ B 1 − p ∥ η ∥ B p ≤ 2 p δ 1 − p ∥ y ∥ B p h p . \Delta_h\varphi(\lambda)\le\frac{2}{p}\lVert\xi\rVert_B^{1-p}\lVert\eta\rVert_B^p\le\frac{2}{p}\delta^{1-p}\lVert y\rVert_B^p h^p. Δ h φ ( λ ) ≤ p 2 ∥ ξ ∥ B 1 − p ∥ η ∥ B p ≤ p 2 δ 1 − p ∥ y ∥ B p h p . For the lower estimate, write x = ( x 0 , x ′ ) x=(x_0,x') x = ( x 0 , x ′ ) , choose λ 0 = − x 0 / y 0 \lambda_0=-x_0/y_0 λ 0 = − x 0 / y 0 , and put z = x + λ 0 y = ( 0 , z ′ ) z=x+\lambda_0y=(0,z') z = x + λ 0 y = ( 0 , z ′ ) . Linear independence implies z ≠ 0 z\ne0 z = 0 . Put S 0 = ∥ z ∥ B p = ∥ z ′ ∥ Y p > 0 S_0=\lVert z\rVert_B^p=\lVert z'\rVert_Y^p>0 S 0 = ∥ z ∥ B p = ∥ z ′ ∥ Y p > 0 and d ± ( h ) = ∥ z ± h y ∥ B p − S 0 d_\pm(h)=\lVert z\pm hy\rVert_B^p-S_0 d ± ( h ) = ∥ z ± h y ∥ B p − S 0 . Since z ± h y = ( ± h y 0 , z ′ ± h y ′ ) z\pm hy=(\pm hy_0,z'\pm hy') z ± h y = ( ± h y 0 , z ′ ± h y ′ ) , the direct-sum norm gives
d ± ( h ) = ∣ y 0 ∣ p h p + ∥ z ′ ± h y ′ ∥ Y p − ∥ z ′ ∥ Y p . d_\pm(h)=|y_0|^p h^p+\lVert z'\pm hy'\rVert_Y^p-\lVert z'\rVert_Y^p. d ± ( h ) = ∣ y 0 ∣ p h p + ∥ z ′ ± h y ′ ∥ Y p − ∥ z ′ ∥ Y p .
The function t ↦ ∥ z ′ + t y ′ ∥ Y p t\mapsto\|z'+ty'\|_Y^p t ↦ ∥ z ′ + t y ′ ∥ Y p is convex on R \mathbb{R} R , so its values at h h h and − h -h − h have sum at least twice its value at 0 0 0 . Consequently,
d + ( h ) + d − ( h ) = 2 ∣ y 0 ∣ p h p + ∥ z ′ + h y ′ ∥ Y p + ∥ z ′ − h y ′ ∥ Y p − 2 ∥ z ′ ∥ Y p ≥ 2 ∣ y 0 ∣ p h p . d_+(h)+d_-(h)=2|y_0|^p h^p+\|z'+hy'\|_Y^p+\|z'-hy'\|_Y^p-2\|z'\|_Y^p\ge2|y_0|^p h^p. d + ( h ) + d − ( h ) = 2∣ y 0 ∣ p h p + ∥ z ′ + h y ′ ∥ Y p + ∥ z ′ − h y ′ ∥ Y p − 2∥ z ′ ∥ Y p ≥ 2∣ y 0 ∣ p h p .
We next bound each d ± ( h ) d_\pm(h) d ± ( h ) separately. Write R = ∥ z ∥ B > 0 R=\|z\|_B>0 R = ∥ z ∥ B > 0 and restrict to 0 < h ≤ R / ( 2 ∥ y ∥ B ) 0<h\le R/(2\|y\|_B) 0 < h ≤ R / ( 2∥ y ∥ B ) . The triangle inequality gives ∣ ∥ z ± h y ∥ B − R ∣ ≤ h ∥ y ∥ B \bigl|\|z\pm hy\|_B-R\bigr|\le h\|y\|_B ∥ z ± h y ∥ B − R ≤ h ∥ y ∥ B , hence
R 2 ≤ ∥ z ± h y ∥ B ≤ 3 R 2 . \frac{R}{2}\le\|z\pm hy\|_B\le\frac{3R}{2}. 2 R ≤ ∥ z ± h y ∥ B ≤ 2 3 R .
The derivative of t ↦ t p t\mapsto t^p t ↦ t p is bounded by p ( 3 R / 2 ) p − 1 p(3R/2)^{p-1} p ( 3 R /2 ) p − 1 on [ R / 2 , 3 R / 2 ] [R/2,3R/2] [ R /2 , 3 R /2 ] . Thus the mean value theorem yields
∣ d ± ( h ) ∣ ≤ p ( 3 R / 2 ) p − 1 ∣ ∥ z ± h y ∥ B − R ∣ ≤ p ( 3 R / 2 ) p − 1 ∥ y ∥ B h = : C 1 h . |d_\pm(h)|\le p(3R/2)^{p-1}\bigl|\|z\pm hy\|_B-R\bigr|\le p(3R/2)^{p-1}\|y\|_B h=:C_1h. ∣ d ± ( h ) ∣ ≤ p ( 3 R /2 ) p − 1 ∥ z ± h y ∥ B − R ≤ p ( 3 R /2 ) p − 1 ∥ y ∥ B h =: C 1 h .
In particular, d ± ( h ) = O ( h ) d_\pm(h)=O(h) d ± ( h ) = O ( h ) , with a constant independent of h h h and of the choice of sign.
Put G ( s ) = s 1 / p G(s)=s^{1/p} G ( s ) = s 1/ p . Both S 0 S_0 S 0 and S 0 + d ± ( h ) S_0+d_\pm(h) S 0 + d ± ( h ) belong to the compact interval I = [ ( R / 2 ) p , ( 3 R / 2 ) p ] ⊂ ( 0 , ∞ ) I=[(R/2)^p,(3R/2)^p]\subset(0,\infty) I = [( R /2 ) p , ( 3 R /2 ) p ] ⊂ ( 0 , ∞ ) . Since G ′ ′ ( s ) = − p − 1 p 2 s 1 / p − 2 G''(s)=-\frac{p-1}{p^2}s^{1/p-2} G ′′ ( s ) = − p 2 p − 1 s 1/ p − 2 , the constant C 2 = max s ∈ I ∣ G ′ ′ ( s ) ∣ C_2=\max_{s\in I}|G''(s)| C 2 = max s ∈ I ∣ G ′′ ( s ) ∣ is finite. Taylor’s theorem, applied separately to the two increments, gives
G ( S 0 + d ± ( h ) ) = G ( S 0 ) + G ′ ( S 0 ) d ± ( h ) + r ± ( h ) , G(S_0+d_\pm(h))=G(S_0)+G'(S_0)d_\pm(h)+r_\pm(h), G ( S 0 + d ± ( h )) = G ( S 0 ) + G ′ ( S 0 ) d ± ( h ) + r ± ( h ) ,
∣ r ± ( h ) ∣ ≤ C 2 2 ∣ d ± ( h ) ∣ 2 ≤ C 2 C 1 2 2 h 2 . |r_\pm(h)|\le\frac{C_2}{2}|d_\pm(h)|^2\le\frac{C_2C_1^2}{2}h^2. ∣ r ± ( h ) ∣ ≤ 2 C 2 ∣ d ± ( h ) ∣ 2 ≤ 2 C 2 C 1 2 h 2 .
Using G ( S 0 + d ± ( h ) ) = φ ( λ 0 ± h ) G(S_0+d_\pm(h))=\varphi(\lambda_0\pm h) G ( S 0 + d ± ( h )) = φ ( λ 0 ± h ) and G ( S 0 ) = φ ( λ 0 ) G(S_0)=\varphi(\lambda_0) G ( S 0 ) = φ ( λ 0 ) , we obtain
Δ h φ ( λ 0 ) = G ′ ( S 0 ) ( d + ( h ) + d − ( h ) ) + r + ( h ) + r − ( h ) ≥ 2 G ′ ( S 0 ) ∣ y 0 ∣ p h p − C 2 C 1 2 h 2 = h p ( 2 G ′ ( S 0 ) ∣ y 0 ∣ p − C 2 C 1 2 h 2 − p ) . \begin{aligned}
\Delta_h\varphi(\lambda_0)
&=G'(S_0)(d_+(h)+d_-(h))+r_+(h)+r_-(h)\\
&\ge2G'(S_0)|y_0|^p h^p-C_2C_1^2h^2\\
&=h^p\bigl(2G'(S_0)|y_0|^p-C_2C_1^2h^{2-p}\bigr).
\end{aligned} Δ h φ ( λ 0 ) = G ′ ( S 0 ) ( d + ( h ) + d − ( h )) + r + ( h ) + r − ( h ) ≥ 2 G ′ ( S 0 ) ∣ y 0 ∣ p h p − C 2 C 1 2 h 2 = h p ( 2 G ′ ( S 0 ) ∣ y 0 ∣ p − C 2 C 1 2 h 2 − p ) .
Here G ′ ( S 0 ) ∣ y 0 ∣ p > 0 G'(S_0)|y_0|^p>0 G ′ ( S 0 ) ∣ y 0 ∣ p > 0 , and h 2 − p → 0 h^{2-p}\to0 h 2 − p → 0 as h → 0 + h\to0^+ h → 0 + because p < 2 p<2 p < 2 . Hence, for all sufficiently small h > 0 h>0 h > 0 , C 2 C 1 2 h 2 − p ≤ G ′ ( S 0 ) ∣ y 0 ∣ p C_2C_1^2h^{2-p}\le G'(S_0)|y_0|^p C 2 C 1 2 h 2 − p ≤ G ′ ( S 0 ) ∣ y 0 ∣ p , and therefore Δ h φ ( λ 0 ) ≥ c h p \Delta_h\varphi(\lambda_0)\ge ch^p Δ h φ ( λ 0 ) ≥ c h p where c = G ′ ( S 0 ) ∣ y 0 ∣ p > 0 c=G'(S_0)|y_0|^p>0 c = G ′ ( S 0 ) ∣ y 0 ∣ p > 0 . Finally, λ 0 ∈ K n \lambda_0\in K_n λ 0 ∈ K n and 2 − n 2^{-n} 2 − n is sufficiently small for all sufficiently large n n n , so M n ≥ Δ 2 − n φ ( λ 0 ) ≥ c 2 − n p M_n\ge\Delta_{2^{-n}}\varphi(\lambda_0)\ge c2^{-np} M n ≥ Δ 2 − n φ ( λ 0 ) ≥ c 2 − n p .
A single atom suffices for the lower estimate in Lemma 5.3 , giving the following recovery result.
Theorem 5.4. Let F \mathbb{F} F be either R \mathbb{R} R or C \mathbb{C} C , let 1 < p < 2 1<p<2 1 < p < 2 , and let B B B be a separable L p L^p L p -space over F \mathbb{F} F of dimension at least two. If B B B has a nonzero atomic part and is computably presentable, then p p p is computable.
Proof. Isolate one normalized atomic coordinate to write B ≅ F ⊕ p Y B\cong\mathbb{F}\oplus_p Y B ≅ F ⊕ p Y , where Y ≠ { 0 } Y\ne\{0\} Y = { 0 } . Fix a computable presentation ( b j ) j ∈ N (b_j)_{j\in\mathbb{N}} ( b j ) j ∈ N of B B B and let D D D be its Q F \mathbb{Q}_\mathbb{F} Q F -linear span, which is dense in B B B . Since { 0 } ⊕ p Y \{0\}\oplus_pY { 0 } ⊕ p Y is a proper closed subspace, choose y ∈ D ∖ ( { 0 } ⊕ p Y ) y\in D\setminus(\{0\}\oplus_pY) y ∈ D ∖ ({ 0 } ⊕ p Y ) . Since dim B ≥ 2 \dim B\ge2 dim B ≥ 2 , the subspace F y \mathbb{F}y F y is also proper and closed, so choose x ∈ D ∖ F y x\in D\setminus\mathbb{F}y x ∈ D ∖ F y . Thus y 0 ≠ 0 y_0\ne0 y 0 = 0 in the chosen decomposition and x , y x,y x , y are linearly independent. By Lemma 5.3 , these vectors yield a uniformly computable positive sequence ( M n ) (M_n) ( M n ) with c 2 − n p ≤ M n ≤ C 2 − n p c2^{-np}\le M_n\le C2^{-np} c 2 − n p ≤ M n ≤ C 2 − n p for some c , C > 0 c,C>0 c , C > 0 and sufficiently large n n n . Lemma 5.1 now makes p p p computable. □ \square □
In particular, taking B = ℓ p ( F ) B=\ell^p(\mathbb{F}) B = ℓ p ( F ) proves Theorem 1.8 . We now combine the exponent-recovery results with the classical classification of separable L p L^p L p -spaces. Write E ⊕ p F E\oplus_p F E ⊕ p F for the sum with norm ∥ ( x , y ) ∥ E ⊕ p F = ( ∥ x ∥ E p + ∥ y ∥ F p ) 1 / p \|(x,y)\|_{E\oplus_p F}=(\|x\|_E^p+\|y\|_F^p)^{1/p} ∥ ( x , y ) ∥ E ⊕ p F = ( ∥ x ∥ E p + ∥ y ∥ F p ) 1/ p .
:::
Theorem 5.5 (see [11 ] ). Let 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ , p ≠ 2 p \ne2 p = 2 , and let F \mathbb{F} F be either R \mathbb{R} R or C \mathbb{C} C . Every nonzero separable L p L^p L p -space over F \mathbb{F} F is linearly isometric to exactly one of the following:
(1) ℓ n p ( F ) \ell_n^p(\mathbb{F}) ℓ n p ( F ) for some n ≥ 1 n \ge1 n ≥ 1 ;
(2) ℓ p ( F ) \ell^p(\mathbb{F}) ℓ p ( F ) ;
(3) L p ( [ 0 , 1 ] ; F ) L^p([0,1];\mathbb{F}) L p ([ 0 , 1 ] ; F ) ;
(4) ℓ n p ( F ) ⊕ p L p ( [ 0 , 1 ] ; F ) \ell_n^p(\mathbb{F}) \oplus_p L^p([0,1];\mathbb{F}) ℓ n p ( F ) ⊕ p L p ([ 0 , 1 ] ; F ) for some n ≥ 1 n \ge1 n ≥ 1 ;
(5) ℓ p ( F ) ⊕ p L p ( [ 0 , 1 ] ; F ) \ell^p(\mathbb{F}) \oplus_p L^p([0,1];\mathbb{F}) ℓ p ( F ) ⊕ p L p ([ 0 , 1 ] ; F ) .
At p = 2 p=2 p = 2 , these types collapse to the usual Hilbert-space classification.
Fact 5.6. Every nonzero separable L 2 L^2 L 2 -space over F \mathbb{F} F is linearly isometric to exactly one of ℓ n 2 ( F ) \ell_n^2(\mathbb{F}) ℓ n 2 ( F ) (n ≥ 1 n \ge1 n ≥ 1 ) or ℓ 2 ( F ) \ell^2(\mathbb{F}) ℓ 2 ( F ) .
Theorem 5.7. Let F \mathbb{F} F be either R \mathbb{R} R or C \mathbb{C} C , let 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ , and let B B B be a nonzero separable L p L^p L p -space over F \mathbb{F} F . Then B B B is computably presentable if and only if exactly one of the following alternatives holds:
(i) dim B = 1 \dim B=1 dim B = 1 ;
(ii) dim B ≥ 2 \dim B \ge2 dim B ≥ 2 , 1 < p < 2 1 < p < 2 1 < p < 2 , B B B is nonatomic, and p p p is right-c.e.;
(iii) dim B ≥ 2 \dim B \ge2 dim B ≥ 2 , 1 < p < 2 1 < p < 2 1 < p < 2 , B B B has a nonzero atomic part, and p p p is computable;
(iv) dim B ≥ 2 \dim B \ge2 dim B ≥ 2 and p ∈ { 1 , 2 } p \in\{1,2\} p ∈ { 1 , 2 } ;
(v) dim B ≥ 2 \dim B \ge2 dim B ≥ 2 , 2 < p < ∞ 2 < p < \infty 2 < p < ∞ , and p p p is computable.
Proof. In dimension one, B ≅ F B \cong\mathbb{F} B ≅ F has a computable presentation independently of p p p . Assume dim B ≥ 2 \dim B \ge2 dim B ≥ 2 . Whenever p p p is computable, the standard presentations of all types in Theorem 5.5 are computable, including the indicated L p L^p L p -sums; Fact 5.6 supplies the case p = 2 p=2 p = 2 . If 1 < p < 2 1 < p < 2 1 < p < 2 and B B B is nonatomic, then B ≅ L p ( [ 0 , 1 ] ; F ) B \cong L^p([0,1];\mathbb{F}) B ≅ L p ([ 0 , 1 ] ; F ) . In this case, Theorems 1.3 and 1.7 show that computable presentability is equivalent to p p p being right-c.e. In every remaining case, computable presentability forces p p p to be computable, by Theorem 5.4 for 1 < p < 2 1 < p < 2 1 < p < 2 and Theorem 1.4 (i) for p > 2 p > 2 p > 2 . □ \square □
An effective Fraïssé theorem for Banach spaces
We extend the finite-dimensional construction method of § 4 to a fixed family of Banach spaces under abstract hypotheses. The construction here uses only linear isometric embeddings, so there is no accumulated distortion to control and no need to estimate convergence to a limiting norm. This makes the construction technically simpler than the one in § 4 . It is a Banach-space analogue of the effective Fraïssé constructions for countable structures in Csima, Harizanov, Miller and Montalbán [10 ] . Its application to the Gurarij Banach space gives a complementary computable presentability result.
Throughout this section, all Banach spaces and linear maps are over R \mathbb{R} R . Let A = ( A n ) n ∈ N \mathcal{A}=(A_n)_{n\in\mathbb{N}} A = ( A n ) n ∈ N be a countable family of nonzero finite-dimensional real Banach spaces. A chain of linear isometric embeddings is a sequence U 0 → u 0 U 1 → u 1 ⋯ U_0 \xrightarrow{u_0} U_1 \xrightarrow{u_1} \cdots U 0 u 0 U 1 u 1 ⋯ such that U s ∈ A U_s \in\mathcal{A} U s ∈ A and u s ∈ Emb ( U s , U s + 1 ) u_s \in\operatorname{Emb}(U_s,U_{s+1}) u s ∈ Emb ( U s , U s + 1 ) . Write u r s = u s − 1 ⋯ u r u_r^s=u_{s-1}\cdots u_r u r s = u s − 1 ⋯ u r and u r r = Id U r u_r^r=\operatorname{Id}_{U_r} u r r = Id U r . When the spaces in A \mathcal{A} A are given in fixed coordinates, the chain is called computable if U s = A n s U_s=A_{n_s} U s = A n s for a computable sequence ( n s ) (n_s) ( n s ) and the matrix entries of u s u_s u s are computable uniformly in s s s . The chain is an A \mathcal{A} A -Fraïssé sequence if:
(F1) Every A p A_p A p embeds linearly isometrically into some U s U_s U s .
(F2) For every f ∈ Emb ( U r , A p ) f \in\operatorname{Emb}(U_r,A_p) f ∈ Emb ( U r , A p ) and ϵ > 0 \epsilon>0 ϵ > 0 , there are s > r s>r s > r and i ∈ Emb ( A p , U s ) i \in\operatorname{Emb}(A_p,U_s) i ∈ Emb ( A p , U s ) with ∥ i f − u r s ∥ < ϵ \lVert i f-u_r^s\rVert<\epsilon ∥ i f − u r s ∥ < ϵ .
Theorem 6.1. Suppose that A = ( A n ) n ∈ N \mathcal{A}=(A_n)_{n\in\mathbb{N}} A = ( A n ) n ∈ N satisfies:
(E1) Each A n A_n A n is given in coordinates as ( R d n , ∥ ⋅ ∥ A n ) (\mathbb{R}^{d_n},\|\cdot\|_{A_n}) ( R d n , ∥ ⋅ ∥ A n ) , where d n d_n d n is computable from n n n , and ∥ x ∥ A n \|x\|_{A_n} ∥ x ∥ A n is computable uniformly in n n n and x ∈ Q d n x\in\mathbb{Q}^{d_n} x ∈ Q d n .
(E2) There are computable sequences ( p e ) e ∈ N (p_e)_{e\in\mathbb{N}} ( p e ) e ∈ N and ( q e ) e ∈ N (q_e)_{e\in\mathbb{N}} ( q e ) e ∈ N and T e ∈ Emb ( A p e , A q e ) T_e\in\operatorname{Emb}(A_{p_e},A_{q_e}) T e ∈ Emb ( A p e , A q e ) whose matrix entries are computable uniformly in e e e , such that
Emb c ( A p , A q ) : = { T e : p e = p , q e = q } \operatorname{Emb}_{c}(A_p,A_q):=\{T_e:p_e=p,\ q_e=q\} Emb c ( A p , A q ) := { T e : p e = p , q e = q } is dense in Emb ( A p , A q ) \operatorname{Emb}(A_p,A_q) Emb ( A p , A q ) in the operator norm for every p , q p,q p , q .
(E3) Every two spaces in A \mathcal{A} A embed linearly isometrically into a common member of A \mathcal{A} A .
(E4) For f ∈ Emb ( A r , A p ) f\in\operatorname{Emb}(A_r,A_p) f ∈ Emb ( A r , A p ) , g ∈ Emb ( A r , A q ) g\in\operatorname{Emb}(A_r,A_q) g ∈ Emb ( A r , A q ) , and η > 0 \eta>0 η > 0 , there are A t A_t A t and i ∈ Emb ( A p , A t ) i\in\operatorname{Emb}(A_p,A_t) i ∈ Emb ( A p , A t ) , j ∈ Emb ( A q , A t ) j\in\operatorname{Emb}(A_q,A_t) j ∈ Emb ( A q , A t ) such that ∥ i f − j g ∥ < η \|if-jg\|<\eta ∥ i f − j g ∥ < η .
Then A \mathcal{A} A has a computable A \mathcal{A} A -Fraïssé sequence U 0 → u 0 U 1 → u 1 ⋯ U_0\xrightarrow{u_0}U_1\xrightarrow{u_1}\cdots U 0 u 0 U 1 u 1 ⋯ whose completed direct limit is computably presentable. Any two A \mathcal{A} A -Fraïssé sequences have linearly isometric completed direct limits, hence their common limit Flim ( A ) \operatorname{Flim}(\mathcal{A}) Flim ( A ) has a computable Banach presentation.
Proof. We shall construct indices n r n_r n r and u r ∈ Emb ( A n r , A n r + 1 ) u_r\in\operatorname{Emb}(A_{n_r},A_{n_{r+1}}) u r ∈ Emb ( A n r , A n r + 1 ) , and write U r = A n r U_r=A_{n_r} U r = A n r . For (F1), impose
C p : Emb ( A p , U t ) ≠ ∅ for some t . C_p:\quad\operatorname{Emb}(A_p,U_t)\ne\varnothing\quad\text{for some }t. C p : Emb ( A p , U t ) = ∅ for some t .
By the density in (E2), (F2) reduces to meeting, whenever p e = n r p_e=n_r p e = n r and k ≥ 1 k\ge1 k ≥ 1 , the requirements
E r , e , k : ∥ i T e − u r t ∥ < 2 − k for some t > r , i ∈ Emb ( A q e , U t ) . (16) E_{r,e,k}:\quad\|iT_e-u_r^t\|<2^{-k}\quad\text{for some }t>r,\ i\in\operatorname{Emb}(A_{q_e},U_t). \tag*{(16)} E r , e , k : ∥ i T e − u r t ∥ < 2 − k for some t > r , i ∈ Emb ( A q e , U t ) . (16)
Fix a computable listing without repetitions of all these formal requirements, giving priority to earlier entries. A requirement is satisfied when the construction supplies a witness.
Construction. Set U 0 = A 0 U_0=A_0 U 0 = A 0 and n 0 = 0 n_0=0 n 0 = 0 . At stage s + 1 s+1 s + 1 , act on the highest-priority unsatisfied requirement that is either a C p C_p C p , or an E r , e , k E_{r,e,k} E r , e , k with r ≤ s r\le s r ≤ s and p e = n r p_e=n_r p e = n r .
If it is C p C_p C p , search for a , b a,b a , b with p a = n s p_a=n_s p a = n s , p b = p p_b=p p b = p , q a = q b q_a=q_b q a = q b and set n s + 1 = q a n_{s+1}=q_a n s + 1 = q a , U s + 1 = A q a U_{s+1}=A_{q_a} U s + 1 = A q a , and u s = T a u_s=T_a u s = T a . Then T b ∈ Emb ( A p , U s + 1 ) T_b\in\operatorname{Emb}(A_p,U_{s+1}) T b ∈ Emb ( A p , U s + 1 ) is the required witness.
If it is E r , e , k E_{r,e,k} E r , e , k , search for a , b a,b a , b with p a = q e p_a=q_e p a = q e , p b = n s p_b=n_s p b = n s , q a = q b q_a=q_b q a = q b and ∥ T a T e − T b u r s ∥ < 2 − k \|T_aT_e-T_bu_r^s\|<2^{-k} ∥ T a T e − T b u r s ∥ < 2 − k , and set n s + 1 = q a n_{s+1}=q_a n s + 1 = q a and u s = T b u_s=T_b u s = T b . Since u r s + 1 = T b u r s u_r^{s+1}=T_bu_r^s u r s + 1 = T b u r s , the map T a T_a T a witnesses E r , e , k E_{r,e,k} E r , e , k .
Verification. The index conditions are decidable, and the strict norm test is semidecidable by Lemma 2.4. The first search terminates by (E3) and (E2). For the second, (E4) gives an amalgam with error less than 2 − k − 1 2^{-k-1} 2 − k − 1 ; sufficiently close approximations to its two embeddings from the dense families in (E2) satisfy the displayed test. Thus all stages terminate and the bonding matrices are uniformly computable. Hence every requirement is eventually satisfied. The requirements C p C_p C p give (F1). For (F2), given f ∈ Emb ( U r , A p ) f\in\operatorname{Emb}(U_r,A_p) f ∈ Emb ( U r , A p ) and ϵ > 0 \epsilon>0 ϵ > 0 , choose e , k e,k e , k with p e = n r p_e=n_r p e = n r , q e = p q_e=p q e = p , ∥ f − T e ∥ < ϵ / 2 \|f-T_e\|<\epsilon/2 ∥ f − T e ∥ < ϵ /2 , and 2 − k < ϵ / 2 2^{-k}<\epsilon/2 2 − k < ϵ /2 . A witness i ∈ Emb ( A p , U t ) i\in\operatorname{Emb}(A_p,U_t) i ∈ Emb ( A p , U t ) to E r , e , k E_{r,e,k} E r , e , k then satisfies
∥ i f − u r t ∥ ≤ ∥ f − T e ∥ + ∥ i T e − u r t ∥ < ϵ . \|if-u_r^t\|\le\|f-T_e\|+\|iT_e-u_r^t\|<\epsilon. ∥ i f − u r t ∥ ≤ ∥ f − T e ∥ + ∥ i T e − u r t ∥ < ϵ .
The norm computation in the proof of Lemma 4.3 gives a computable presentation of the completed limit B B B . Indeed, enumerate the images in B B B of all rational coordinate vectors in the spaces U s U_s U s . They are dense, and, writing ι s ∈ Emb ( U s , B ) \iota_s \in\operatorname{Emb}(U_s,B) ι s ∈ Emb ( U s , B ) for the canonical embeddings, each finite rational linear combination satisfies
∥ ∑ j < m α j u s j a j ∥ B = ∥ ∑ j < m α j u s j t a j ∥ U t , t ≥ max j < m s j . (17) \left\lVert\sum_{j<m} \alpha_j u_{s_j}a_j \right\rVert_B = \left\lVert\sum_{j<m} \alpha_j u^t_{s_j}a_j \right\rVert_{U_t}, \qquad t \ge\max_{j<m}s_j. \tag*{(17)} j < m ∑ α j u s j a j B = j < m ∑ α j u s j t a j U t , t ≥ j < m max s j . (17)
The right-hand vector has uniformly computable coordinates, so its norm is uniformly computable by (E1) and (1).
Finally, let ( U s , u s ) (U_s,u_s) ( U s , u s ) and ( V s , v s ) (V_s,v_s) ( V s , v s ) be any two A \mathcal{A} A -Fraïssé sequences. We construct strictly increasing indices r n , s n r_n,s_n r n , s n and
f n ∈ Emb ( U r n , V s n ) , g n ∈ Emb ( V s n , U r n + 1 ) f_n \in\operatorname{Emb}(U_{r_n},V_{s_n}), \qquad g_n \in\operatorname{Emb}(V_{s_n},U_{r_{n+1}}) f n ∈ Emb ( U r n , V s n ) , g n ∈ Emb ( V s n , U r n + 1 )
such that
∥ g n f n − u r n r n + 1 ∥ < 2 − n − 2 , ∥ f n + 1 g n − v s n s n + 1 ∥ < 2 − n − 2 . (18) \lVert g_nf_n-u^{r_{n+1}}_{r_n}\rVert< 2^{-n-2}, \qquad\lVert f_{n+1}g_n-v^{s_{n+1}}_{s_n}\rVert< 2^{-n-2}. \tag*{(18)} ∥ g n f n − u r n r n + 1 ∥ < 2 − n − 2 , ∥ f n + 1 g n − v s n s n + 1 ∥ < 2 − n − 2 . (18)
Set r 0 = 0 r_0=0 r 0 = 0 . Since U 0 ∈ A U_0\in\mathcal{A} U 0 ∈ A , property (F1) for the sequence ( V s , v s ) (V_s,v_s) ( V s , v s ) gives an index s 0 s_0 s 0 and f 0 ∈ Emb ( U 0 , V s 0 ) f_0\in\operatorname{Emb}(U_0,V_{s_0}) f 0 ∈ Emb ( U 0 , V s 0 ) .
Suppose that r n , s n r_n,s_n r n , s n and f n ∈ Emb ( U r n , V s n ) f_n\in\operatorname{Emb}(U_{r_n},V_{s_n}) f n ∈ Emb ( U r n , V s n ) have been chosen. Apply (F2) to the sequence ( U s , u s ) (U_s,u_s) ( U s , u s ) , taking U r n U_{r_n} U r n as the source space, V s n V_{s_n} V s n as the member of A \mathcal{A} A serving as the target, f n f_n f n as the given embedding, and 2 − n − 2 2^{-n-2} 2 − n − 2 as the tolerance. This gives an index r n + 1 > r n r_{n+1}>r_n r n + 1 > r n and g n ∈ Emb ( V s n , U r n + 1 ) g_n\in\operatorname{Emb}(V_{s_n},U_{r_{n+1}}) g n ∈ Emb ( V s n , U r n + 1 ) satisfying the first inequality in (19).
Now apply (F2) to the sequence ( V s , v s ) (V_s,v_s) ( V s , v s ) , with source V s n V_{s_n} V s n , target U r n + 1 ∈ A U_{r_{n+1}}\in\mathcal{A} U r n + 1 ∈ A , given embedding g n g_n g n , and the same tolerance 2 − n − 2 2^{-n-2} 2 − n − 2 . This gives an index s n + 1 > s n s_{n+1}>s_n s n + 1 > s n and f n + 1 ∈ Emb ( U r n + 1 , V s n + 1 ) f_{n+1}\in\operatorname{Emb}(U_{r_{n+1}},V_{s_{n+1}}) f n + 1 ∈ Emb ( U r n + 1 , V s n + 1 ) satisfying the second inequality in (19). This completes the induction.
Let B U B_U B U and B V B_V B V be the respective completed direct limits. Identify U s U_s U s and V s V_s V s with their images under the canonical embeddings, so that the bonding maps ( u s ) (u_s) ( u s ) and ( v s ) (v_s) ( v s ) become inclusions. Since ( r n ) (r_n) ( r n ) and ( s n ) (s_n) ( s n ) are strictly increasing, the unions ⋃ n U r n \bigcup_n U_{r_n} ⋃ n U r n and ⋃ n V s n \bigcup_n V_{s_n} ⋃ n V s n are dense in B U B_U B U and B V B_V B V , respectively. For x ∈ U r n x\in U_{r_n} x ∈ U r n , insert f n + 1 g n f n x f_{n+1}g_nf_nx f n + 1 g n f n x and use (19) to obtain
∥ f n + 1 x − f n x ∥ B V ≤ ∥ f n + 1 ( x − g n f n x ) ∥ B V + ∥ f n + 1 g n f n x − f n x ∥ B V ≤ 2 − n − 2 ∥ x ∥ B U + 2 − n − 2 ∥ f n x ∥ B V = 2 − n − 1 ∥ x ∥ B U . \begin{aligned}
\lVert f_{n+1}x-f_nx\rVert_{B_V}
&\le\lVert f_{n+1}(x-g_nf_nx)\rVert_{B_V}+\lVert f_{n+1}g_nf_nx-f_nx\rVert_{B_V}\\
&\le2^{-n-2}\lVert x\rVert_{B_U}+2^{-n-2}\lVert f_nx\rVert_{B_V}
=2^{-n-1}\lVert x\rVert_{B_U}.
\end{aligned} ∥ f n + 1 x − f n x ∥ B V ≤ ∥ f n + 1 ( x − g n f n x ) ∥ B V + ∥ f n + 1 g n f n x − f n x ∥ B V ≤ 2 − n − 2 ∥ x ∥ B U + 2 − n − 2 ∥ f n x ∥ B V = 2 − n − 1 ∥ x ∥ B U .
For a fixed x x x in ⋃ n U r n \bigcup_n U_{r_n} ⋃ n U r n , this estimate applies for all sufficiently large n n n . Its summable right-hand side shows that ( f n x ) (f_nx) ( f n x ) is Cauchy in B V B_V B V . Define F x = lim n f n x Fx=\lim_n f_nx F x = lim n f n x , starting with any n n n for which x ∈ U r n x\in U_{r_n} x ∈ U r n . The definition is independent of this starting index. Linearity follows by taking limits on a common space containing the vectors involved, and ∥ F x ∥ B V = lim n ∥ f n x ∥ B V = ∥ x ∥ B U \lVert Fx\rVert_{B_V}=\lim_n\lVert f_nx\rVert_{B_V}=\lVert x\rVert_{B_U} ∥ F x ∥ B V = lim n ∥ f n x ∥ B V = ∥ x ∥ B U . Thus F F F extends uniquely to F ∈ Emb ( B U , B V ) F\in\operatorname{Emb}(B_U,B_V) F ∈ Emb ( B U , B V ) . Summing the difference estimate also gives
∥ F x − f n x ∥ B V ≤ ∑ m ≥ n 2 − m − 1 ∥ x ∥ B U = 2 − n ∥ x ∥ B U ( x ∈ U r n ) . \lVert Fx-f_nx\rVert_{B_V}\le\sum_{m\ge n}2^{-m-1}\lVert x\rVert_{B_U}=2^{-n}\lVert x\rVert_{B_U}\qquad(x\in U_{r_n}). ∥ F x − f n x ∥ B V ≤ m ≥ n ∑ 2 − m − 1 ∥ x ∥ B U = 2 − n ∥ x ∥ B U ( x ∈ U r n ) .
To prove surjectivity, fix y ∈ V s j y\in V_{s_j} y ∈ V s j and let n ≥ j n\ge j n ≥ j . Then y ∈ V s n y\in V_{s_n} y ∈ V s n , g n y ∈ U r n + 1 g_ny\in U_{r_{n+1}} g n y ∈ U r n + 1 , and ∥ g n y ∥ B U = ∥ y ∥ B V \lVert g_ny\rVert_{B_U}=\lVert y\rVert_{B_V} ∥ g n y ∥ B U = ∥ y ∥ B V . Applying the preceding tail estimate with index n + 1 n+1 n + 1 and then the second inequality in (19), we get
∥ F ( g n y ) − y ∥ B V ≤ ∥ F ( g n y ) − f n + 1 g n y ∥ B V + ∥ f n + 1 g n y − y ∥ B V ≤ 2 − n − 1 ∥ g n y ∥ B U + 2 − n − 2 ∥ y ∥ B V = 3 ⋅ 2 − n − 2 ∥ y ∥ B V → 0. \begin{aligned}
\lVert F(g_ny)-y\rVert_{B_V}
&\le\lVert F(g_ny)-f_{n+1}g_ny\rVert_{B_V}+\lVert f_{n+1}g_ny-y\rVert_{B_V}\\
&\le2^{-n-1}\lVert g_ny\rVert_{B_U}+2^{-n-2}\lVert y\rVert_{B_V}
=3\cdot2^{-n-2}\lVert y\rVert_{B_V}\to0.
\end{aligned} ∥ F ( g n y ) − y ∥ B V ≤ ∥ F ( g n y ) − f n + 1 g n y ∥ B V + ∥ f n + 1 g n y − y ∥ B V ≤ 2 − n − 1 ∥ g n y ∥ B U + 2 − n − 2 ∥ y ∥ B V = 3 ⋅ 2 − n − 2 ∥ y ∥ B V → 0.
Consequently every vector in the dense union ⋃ j V s j \bigcup_j V_{s_j} ⋃ j V s j lies in the closure of F ( B U ) F(B_U) F ( B U ) . This range is closed because F F F is isometric and B U B_U B U is complete. Hence F ( B U ) = B V F(B_U)=B_V F ( B U ) = B V , which proves uniqueness of Flim ( A ) \operatorname{Flim}(\mathcal{A}) Flim ( A ) . □ \square □
We now apply Theorem 6.1 to the Gurarij Banach space, following Ben Yaacov and Henson [4 ] for the definition and basic properties.
Definition 6.2. A nonzero real Banach space G G G is a Gurarij space if, for every ε > 0 \varepsilon>0 ε > 0 , finite-dimensional real Banach spaces E ⊆ H E\subseteq H E ⊆ H with H ≠ { 0 } H\ne\{0\} H = { 0 } , and u ∈ Emb ( E , G ) u\in\operatorname{Emb}(E,G) u ∈ Emb ( E , G ) , there is v ∈ Emb ε log ( H , G ) v\in\operatorname{Emb}^{\log}_{\varepsilon}(H,G) v ∈ Emb ε l o g ( H , G ) such that v ↾ E = u v\mathord{\upharpoonright}E=u v ↾ E = u .
A separable Gurarij space exists and is unique up to surjective linear isometry [20 ] . It is universal for separable real Banach spaces: every such space embeds linearly isometrically into it [4 Corollary 2.7] . It is also approximately ultrahomogeneous [4 Corollary 2.8] . From now on, G G G denotes the unique separable real Gurarij space.
Question 6.3 ([15 Question A.19] ). Does G G G have a computable presentation?
For every integer n ≥ 1 n\ge1 n ≥ 1 , let ℓ ∞ n ( R ) : = ( R n , ∥ ⋅ ∥ ∞ ) \ell_\infty^n(\mathbb{R}):=(\mathbb{R}^n,\|\cdot\|_\infty) ℓ ∞ n ( R ) := ( R n , ∥ ⋅ ∥ ∞ ) , where ∥ ( x 1 , … , x n ) ∥ ∞ = max 1 ≤ j ≤ n ∣ x j ∣ \|(x_1,\ldots,x_n)\|_\infty=\max_{1\le j\le n}|x_j| ∥ ( x 1 , … , x n ) ∥ ∞ = max 1 ≤ j ≤ n ∣ x j ∣ . These finite-dimensional spaces with the maximum norm will be the building blocks of our construction. A linear contraction is a linear map T T T satisfying ∥ T x ∥ ≤ ∥ x ∥ \|Tx\|\le\|x\| ∥ T x ∥ ≤ ∥ x ∥ for every x x x .
Lemma 6.4. Let E E E be a finite-dimensional real Banach space, let m , n ≥ 1 m,n\ge1 m , n ≥ 1 , and let 0 ≤ δ < 1 0\le\delta<1 0 ≤ δ < 1 . Suppose that a : E → ℓ ∞ m ( R ) a:E\to\ell_\infty^m(\mathbb{R}) a : E → ℓ ∞ m ( R ) and b : E → ℓ ∞ n ( R ) b:E\to\ell_\infty^n(\mathbb{R}) b : E → ℓ ∞ n ( R ) are linear contractions such that ( 1 − δ ) ∥ x ∥ E ≤ min { ∥ a x ∥ ∞ , ∥ b x ∥ ∞ } (1-\delta)\|x\|_E\le\min\{\|ax\|_\infty,\|bx\|_\infty\} ( 1 − δ ) ∥ x ∥ E ≤ min { ∥ a x ∥ ∞ , ∥ b x ∥ ∞ } for all x ∈ E x\in E x ∈ E . Then there are i ∈ Emb ( ℓ ∞ m ( R ) , ℓ ∞ m + n ( R ) ) i\in\operatorname{Emb}(\ell_\infty^m(\mathbb{R}),\ell_\infty^{m+n}(\mathbb{R})) i ∈ Emb ( ℓ ∞ m ( R ) , ℓ ∞ m + n ( R )) and j ∈ Emb ( ℓ ∞ n ( R ) , ℓ ∞ m + n ( R ) ) j\in\operatorname{Emb}(\ell_\infty^n(\mathbb{R}),\ell_\infty^{m+n}(\mathbb{R})) j ∈ Emb ( ℓ ∞ n ( R ) , ℓ ∞ m + n ( R )) such that ∥ i a − j b ∥ ≤ δ \|ia-jb\|\le\delta ∥ ia − j b ∥ ≤ δ .
Proof. The maps a , b a,b a , b are injective, and
a x ↦ ( 1 − δ ) b x , b x ↦ ( 1 − δ ) a x ax\mapsto(1-\delta)bx,\qquad bx\mapsto(1-\delta)ax a x ↦ ( 1 − δ ) b x , b x ↦ ( 1 − δ ) a x
are contractions on a ( E ) a(E) a ( E ) and b ( E ) b(E) b ( E ) , respectively. By applying the Hahn–Banach theorem coordinatewise, these maps extend to contractions
R : ℓ ∞ m ( R ) → ℓ ∞ n ( R ) , S : ℓ ∞ n ( R ) → ℓ ∞ m ( R ) . R:\ell_\infty^m(\mathbb{R})\to\ell_\infty^n(\mathbb{R}),\qquad S:\ell_\infty^n(\mathbb{R})\to\ell_\infty^m(\mathbb{R}). R : ℓ ∞ m ( R ) → ℓ ∞ n ( R ) , S : ℓ ∞ n ( R ) → ℓ ∞ m ( R ) .
Define i ( x ) = ( x , R x ) i(x)=(x,Rx) i ( x ) = ( x , R x ) and j ( y ) = ( S y , y ) j(y)=(Sy,y) j ( y ) = ( S y , y ) , using the maximum norm on the concatenated m + n m+n m + n coordinates. Both maps are isometric since R R R and S S S are contractions. Since R a = ( 1 − δ ) b Ra=(1-\delta)b R a = ( 1 − δ ) b and S b = ( 1 − δ ) a Sb=(1-\delta)a S b = ( 1 − δ ) a , we have i a x − j b x = ( δ a x , − δ b x ) iax-jbx=(\delta ax,-\delta bx) ia x − j b x = ( δ a x , − δ b x ) and ∥ i a x − j b x ∥ ∞ ≤ δ ∥ x ∥ E \|iax-jbx\|_\infty\le\delta\|x\|_E ∥ ia x − j b x ∥ ∞ ≤ δ ∥ x ∥ E . This proves the lemma. □ \square □
Lemma 6.5 ([19 Proposition 2.12(1),(5) and Section 6.1] )1 . A nonzero separable real Banach space B B B is linearly isometric to G G G if and only if it has the following property: for every finite-dimensional real Banach space E E E , integer n ≥ 1 n\ge1 n ≥ 1 , f ∈ Emb ( E , B ) f\in\operatorname{Emb}(E,B) f ∈ Emb ( E , B ) , φ ∈ Emb ( E , ℓ ∞ n ( R ) ) \varphi\in\operatorname{Emb}(E,\ell_\infty^n(\mathbb{R})) φ ∈ Emb ( E , ℓ ∞ n ( R )) , and ε > 0 \varepsilon>0 ε > 0 , there exists g ∈ Emb ( ℓ ∞ n ( R ) , B ) g\in\operatorname{Emb}(\ell_\infty^n(\mathbb{R}),B) g ∈ Emb ( ℓ ∞ n ( R ) , B ) such that ∥ g φ − f ∥ < ε \|g\varphi-f\|<\varepsilon ∥ g φ − f ∥ < ε .
Theorem 6.6. The real Gurarij Banach space G G G has a computable presentation.
::: Proof. Apply Theorem 6.1 to the explicitly indexed family A = ( A n ) n ∈ N \mathcal{A}=(A_n)_{n\in\mathbb{N}} A = ( A n ) n ∈ N where A n = ℓ ∞ n + 1 ( R ) A_n=\ell_{\infty}^{n+1}(\mathbb{R}) A n = ℓ ∞ n + 1 ( R ) . We use the standard coordinate basis of each A n A_n A n . Its norm is uniformly computable because it is the maximum of finitely many absolute values. This verifies (E1).
We next construct the dense computable families of isometric embeddings required by (E2). Let d , e ≥ 1 d,e\ge1 d , e ≥ 1 , and represent a linear map T : ℓ ∞ d ( R ) → ℓ ∞ e ( R ) T:\ell_{\infty}^{d}(\mathbb{R})\to\ell_{\infty}^{e}(\mathbb{R}) T : ℓ ∞ d ( R ) → ℓ ∞ e ( R ) by an e × d e\times d e × d matrix ( t i j ) (t_{ij}) ( t ij ) . Its operator norm is ∥ T ∥ = max 1 ≤ i ≤ e ∑ j = 1 d ∣ t i j ∣ \lVert T\rVert=\max_{1\le i\le e}\sum_{j=1}^{d}|t_{ij}| ∥ T ∥ = max 1 ≤ i ≤ e ∑ j = 1 d ∣ t ij ∣ . Consequently T ∈ Emb ( ℓ ∞ d ( R ) , ℓ ∞ e ( R ) ) T\in\operatorname{Emb}(\ell_{\infty}^{d}(\mathbb{R}),\ell_{\infty}^{e}(\mathbb{R})) T ∈ Emb ( ℓ ∞ d ( R ) , ℓ ∞ e ( R )) if and only if both of the following conditions hold:
(a) every row satisfies ∑ j = 1 d ∣ t i j ∣ ≤ 1 \sum_{j=1}^{d}|t_{ij}|\le1 ∑ j = 1 d ∣ t ij ∣ ≤ 1 ;
(b) for every column j j j , some row has an entry equal to 1 1 1 or − 1 -1 − 1 in column j j j and has zero entries in all other columns.
Indeed, if T T T is isometric, then (a) follows from ∥ T ∥ = 1 \lVert T\rVert=1 ∥ T ∥ = 1 . The image of the j j j th coordinate unit vector has norm one, so some entry in column j j j equals 1 1 1 or − 1 -1 − 1 ; (a) forces all other entries in that row to vanish. Conversely, (a) gives ∥ T x ∥ ∞ ≤ ∥ x ∥ ∞ \lVert Tx\rVert_{\infty}\le\lVert x\rVert_{\infty} ∥ T x ∥ ∞ ≤ ∥ x ∥ ∞ , and the rows in (b) give the reverse inequality for every x x x .
Conditions (a) and (b) are decidable for matrices with rational entries. Enumerating all rational matrices satisfying these conditions, for all d , e d,e d , e , produces isometric embeddings whose matrix entries are uniformly computable.
These embeddings are dense: keep one signed coordinate row as in (b) for each column fixed, and approximate each remaining row by rational points of the unit ball of ℓ 1 d \ell_{1}^{d} ℓ 1 d . The resulting matrices still satisfy (a) and (b) and converge in operator norm. Thus their uniform enumeration verifies (E2).
For (E3), two spaces ℓ ∞ d ( R ) \ell_{\infty}^{d}(\mathbb{R}) ℓ ∞ d ( R ) and ℓ ∞ e ( R ) \ell_{\infty}^{e}(\mathbb{R}) ℓ ∞ e ( R ) embed into ℓ ∞ d + e ( R ) \ell_{\infty}^{d+e}(\mathbb{R}) ℓ ∞ d + e ( R ) by x ↦ ( x , 0 ) x\mapsto(x,0) x ↦ ( x , 0 ) and y ↦ ( 0 , y ) y\mapsto(0,y) y ↦ ( 0 , y ) . For (E4), apply Lemma 6.4 with δ = 0 \delta=0 δ = 0 . All four hypotheses of Theorem 6.1 hold. It follows that there is a computable A \mathcal{A} A -Fraïssé sequence ( U r , u r ) (U_r,u_r) ( U r , u r ) whose completed limit B B B has a computable presentation.
It remains to verify that B B B is G \mathbb{G} G . Identify each U r U_r U r with its canonical image in B B B ; these are increasing subspaces with dense union. We check the property in Lemma 6.5 . Fix a finite-dimensional E E E , f ∈ Emb ( E , B ) f\in\operatorname{Emb}(E,B) f ∈ Emb ( E , B ) , φ ∈ Emb ( E , ℓ ∞ n ( R ) ) \varphi\in\operatorname{Emb}(E,\ell_{\infty}^{n}(\mathbb{R})) φ ∈ Emb ( E , ℓ ∞ n ( R )) , and ϵ > 0 \epsilon>0 ϵ > 0 . Choose 0 < η < min { 1 / 2 , ϵ / 5 } 0<\eta<\min\{1/2,\epsilon/5\} 0 < η < min { 1/2 , ϵ /5 } . Approximating the images under f f f of a basis of E E E in the dense union gives some r r r and a linear map f 0 : E → U r f_0:E\to U_r f 0 : E → U r such that ∥ f 0 − f ∥ < η \lVert f_0-f\rVert<\eta ∥ f 0 − f ∥ < η . Set a = f 0 / ( 1 + η ) a=f_0/(1+\eta) a = f 0 / ( 1 + η ) . Then a a a is a contraction, ∥ a − f ∥ < 2 η \lVert a-f\rVert<2\eta ∥ a − f ∥ < 2 η , and ∥ a x ∥ ≥ ( 1 − 2 η ) ∥ x ∥ E \lVert ax\rVert\ge(1-2\eta)\lVert x\rVert_E ∥ a x ∥ ≥ ( 1 − 2 η ) ∥ x ∥ E for all x ∈ E x\in E x ∈ E . Write U r = ℓ ∞ m ( R ) U_r=\ell_{\infty}^{m}(\mathbb{R}) U r = ℓ ∞ m ( R ) . By Lemma 6.4 , applied to a , φ a,\varphi a , φ with δ = 2 η \delta=2\eta δ = 2 η , there are i ∈ Emb ( U r , ℓ ∞ m + n ( R ) ) i\in\operatorname{Emb}(U_r,\ell_{\infty}^{m+n}(\mathbb{R})) i ∈ Emb ( U r , ℓ ∞ m + n ( R )) and j ∈ Emb ( ℓ ∞ n ( R ) , ℓ ∞ m + n ( R ) ) j\in\operatorname{Emb}(\ell_{\infty}^{n}(\mathbb{R}),\ell_{\infty}^{m+n}(\mathbb{R})) j ∈ Emb ( ℓ ∞ n ( R ) , ℓ ∞ m + n ( R )) satisfying ∥ i a − j φ ∥ < 2 η \lVert ia-j\varphi\rVert<2\eta ∥ ia − j φ ∥ < 2 η . Property (F2) of the sequence, applied to i ∈ Emb ( U r , ℓ ∞ m + n ( R ) ) i\in\operatorname{Emb}(U_r,\ell_{\infty}^{m+n}(\mathbb{R})) i ∈ Emb ( U r , ℓ ∞ m + n ( R )) , gives s > r s>r s > r and h ∈ Emb ( ℓ ∞ m + n ( R ) , U s ) h\in\operatorname{Emb}(\ell_{\infty}^{m+n}(\mathbb{R}),U_s) h ∈ Emb ( ℓ ∞ m + n ( R ) , U s ) such that ∥ h i − u r s ∥ < η \lVert hi-u_r^s\rVert<\eta ∥ hi − u r s ∥ < η . Regard g = h j g=hj g = hj as an element of Emb ( ℓ ∞ n ( R ) , B ) \operatorname{Emb}(\ell_{\infty}^{n}(\mathbb{R}),B) Emb ( ℓ ∞ n ( R ) , B ) . Since u r s u_r^s u r s is the inclusion under our identifications and ∥ a ∥ ≤ 1 \lVert a\rVert\le1 ∥ a ∥ ≤ 1 , we obtain
∥ g φ − f ∥ ≤ ∥ h ( j φ − i a ) ∥ + ∥ ( h i − u r s ) a ∥ + ∥ a − f ∥ < 2 η + η + 2 η = 5 η < ϵ . \lVert g\varphi-f\rVert\le\lVert h(j\varphi-ia)\rVert+\lVert(h i-u_r^s)a\rVert+\lVert a-f\rVert<2\eta+\eta+2\eta=5\eta<\epsilon. ∥ g φ − f ∥ ≤ ∥ h ( j φ − ia )∥ + ∥( hi − u r s ) a ∥ + ∥ a − f ∥ < 2 η + η + 2 η = 5 η < ϵ .
Thus B B B satisfies Lemma 6.5 and is linearly isometric to G \mathbb{G} G . Its computable presentation gives the desired presentation of G \mathbb{G} G . □ \square □
Acknowledgments
The authors were supported by Focused Research Grants of Natural Science Foundation of Jiangsu Province No. BK20243060. The authors used OpenAI’s ChatGPT 5.6 to assist with developing the proofs. More specifically, it first suggested a possible connection between McNicholl’s conjecture and the work of Ferenczi, López-Abad, Mbombo, and Todorčević [16 ] . When subsequently asked whether the relevant arguments could be effectivized, it provided the initial feasibility analysis that led to the approach developed in this paper. All proofs in this paper have been independently verified by the authors without the use of AI tools. The authors take full responsibility for the final content of the paper.
School of Mathematics, Nanjing University, Nanjing, Jiangsu 210093, People’s Republic of China
Email address: xie.ruofei@nju.edu.cn
Email address: xuanheng21@gmail.com