OctoInformBeta

OpenAI preprint, September 2026 · §8. Path functionals on arbitrary subspaces

Theorem 8.2: Binary branches on closed subspaces of the constructed real Banach space

Separated path functionals on infinite-dimensional subspaces.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

For a closed infinite-dimensional subspace FF of the constructed real Banach space EE, consider the restrictions to FF of its bounded path functionals.

Let Γ=ω1\Gamma=\omega_1.✦ Fix a bijection from Γ\Gamma to ℕ×ω1\mathbb N\times\omega_1, and let c(α)c(\alpha) be its first coordinate. Each color class Ck=c−1({k})C_k=c^{-1}(\{k\}) has cardinality ℵ1\aleph_1 and is unbounded in ω1\omega_1. A rational array on Γ\Gamma is a rational-valued function with finite support.✦ The construction uses pure operation families fkf_k, one for each color, and a mixed family fmf_{\mathrm m}.

A norming set KK on Γ\Gamma is a set of rational arrays containing every positive coordinate array δα\delta_\alpha, with |g(α)|≤1|g(\alpha)|\leq1 for every g∈Kg\in K and α∈Γ\alpha\in\Gamma.✦ Here c00(Γ,ℝ)c_{00}(\Gamma,\mathbb R) is the space of finitely supported real-valued functions on Γ\Gamma, and ℓ∞(K)\ell_\infty(K) is the space of bounded real-valued functions on KK, equipped with the supremum norm. Its coordinate evaluation vectors and finite evaluation map are given by eα(g)=g(α),(ιx)(g)=∑α∈Γx(α)g(α),EK=ι(c00(Γ,ℝ))¯ℓ∞(K).e_\alpha(g)=g(\alpha),\qquad (\iota x)(g)=\sum_{\alpha\in\Gamma}x(\alpha)g(\alpha),\qquad E_K=\overline{\iota(c_{00}(\Gamma,\mathbb R))}^{\,\ell_\infty(K)}.(1) The finite support inclusion in (1) is also regarded as a map into EKE_K, with its inherited norm. For g∈Kg\in K, the norming evaluation functional is hg(x)=x(g)h_g(x)=x(g) on EKE_K; the coordinate functional is eα*=hδαe_\alpha^*=h_{\delta_\alpha}. The coordinate predual and rational array evaluation are X0=span⁡ℝ{eα*:α∈Γ}¯EK*,⟨g,x⟩=(∑α∈Γg(α)eα*)(x).X_0=\overline{\operatorname{span}_{\mathbb R} \{e_\alpha^*: \alpha\in\Gamma\}}^{\,E_K^*},\qquad \langle g,x\rangle= \left(\sum_{\alpha\in\Gamma}g(\alpha)e_\alpha^*\right)(x).(2) The rational combination in (2) takes values in X0X_0. The set KK is the full norming set of the pure and mixed construction, and E=EKE=E_K is its real normed completion.

Norming Families and Paths

The exponent parameter of fkf_k is s=k+2s=k+2, and that of fmf_{\mathrm m} is s=2s=2.✦ The parameter sequences are indexed from zero. With recursive parameter state (tn,bn)(t_n,b_n), their definitions are dn=n+11,(t0,b0)=(d02+2d0+2,0),Bn=bn+2⋅2s(s+1)tn,bn+1=Bn,tn+1=tn+(Bn+1)(dn+1+1)+1+dn+1+dn+12,L(s,n)=2s(s+1)tn,m(s,n)=2(s−1)(s+1)tn+dn.\begin{gathered} d_n=n+11,\qquad (t_0,b_0)=(d_0^2+2d_0+2,0),\\ B_n=b_n+2\cdot2^{s(s+1)t_n},\qquad b_{n+1}=B_n,\\ t_{n+1}=t_n+(B_n+1)(d_{n+1}+1)+1+d_{n+1}+d_{n+1}^2,\\ L(s,n)=2^{s(s+1)t_n},\qquad m(s,n)=2^{(s-1)(s+1)t_n+d_n}. \end{gathered}(3) A positive weight jj uses the underlying parameter index j−1j-1: Lf(j)=L(s,j−1)L_f(j)=L(s,j-1) and mf(j)=m(s,j−1)m_f(j)=m(s,j-1) in (3). The sequence dnd_n in (3) is the depth sequence. The exponent scale tnt_n is the first coordinate of the recursive parameter state in (3).✦ For a parameter ss and a zero-based index nn, L(s,n)L(s,n) in (3) is the block length. The normalization factor in (3) is m(s,n)m(s,n) for a parameter ss and a zero-based index nn. For a family ff and a positive weight jj, its family block length is Lf(j)L_f(j). A family ff has normalization factor mf(j)m_f(j) at each positive weight jj.

Arrays u,vu,v have successive supports in a family ff when every α∈supp⁡u\alpha\in\operatorname{supp}u precedes every β∈supp⁡v\beta\in\operatorname{supp}v; for the mixed family one also requires c(α)<c(β)c(\alpha)<c(\beta).✦ A Type I expression in ff of weight j≥1j\geq1 consists of nonzero successive rational children g0,…,gℓ−1g_0,\ldots,g_{\ell-1}, where 1≤ℓ≤Lf(j)1\leq\ell\leq L_f(j).✦ Its value is ϕ=1mf(j)∑h<ℓgh.\phi=\frac{1}{m_f(j)}\sum_{h<\ell}g_h.(4)✦ The pure base at color kk is {0}∪{δα,−δα:α∈Ck}\{0\}\cup\{\delta_\alpha,-\delta_\alpha:\alpha\in C_k\}.✦ Starting with a base BB, the zero stage is BB; each subsequent stage retains the preceding stage and admits Type I values (4) whose children lie in the preceding stage, together with Type II rational arrays formed from original path pieces whose children all lie in the preceding stage.✦ The pure norming set KkK_k is the union of these stages for fkf_k and its pure base.✦ The mixed base is ⋃kKk\bigcup_kK_k, and the full set KK is the union of its stages for fmf_{\mathrm m}. Write KfK_f for the norming set belonging to either family.

A coherent predecessor closure family (Gn(α))α<β,n≥0(G_n(\alpha))_{\alpha<\beta,\,n\geq0} at an ordinal β\beta is a family of finite sets Gn(α)⊆[0,α]G_n(\alpha)\subseteq[0,\alpha] that contain α\alpha, increase with nn, exhaust [0,α][0,\alpha], and satisfy Gn(ξ)=Gn(α)∩[0,ξ]G_n(\xi)=G_n(\alpha)\cap[0,\xi] whenever ξ∈Gn(α)\xi\in G_n(\alpha).✦ An endpoint extension of a coherent predecessor closure family (Gn(α))α<β,n≥0(G_n(\alpha))_{\alpha<\beta,\,n\geq0} is a sequence of finite sets Gn(β)⊆[0,β]G_n(\beta)\subseteq[0,\beta] that contain β\beta, increase with nn, exhaust [0,β][0,\beta], and satisfy Gn(α)=Gn(β)∩[0,α]G_n(\alpha)=G_n(\beta)\cap[0,\alpha] whenever α∈Gn(β)\alpha\in G_n(\beta). Ordinal recursion using the endpoint closure construction gives the finite closure system Sn(β)S_n(\beta). The positive-indexed closures and closure entry ranks are defined by (5), with closure index zero also selecting the first system entry: Fp(β)=Smax⁡{p−1,0}(β),ρ(α,β)={min⁡{p≥1:α∈Fp(β)},α≤β,0,α>β.F_p(\beta)=S_{\max\{p-1,0\}}(\beta),\qquad \rho(\alpha,\beta)= \begin{cases} \min\{p\geq1:\alpha\in F_p(\beta)\},&\alpha\leq\beta,\\ 0,&\alpha>\beta. \end{cases}(5)

For a finite coordinate set AA, the order position pos⁡A(α)\operatorname{pos}_A(\alpha) of a coordinate α\alpha is the number of members of AA below α\alpha.✦ The positional code of a rational array gg on a finite coordinate set AA is the finite set of pairs (pos⁡A(α),g(α))(\operatorname{pos}_A(\alpha),g(\alpha)) for α∈A\alpha\in A, including zero coefficients.✦ A finite path code aa is a finite list of triples consisting of a weight, metadata, and a positional array code.✦ The metadata bound of a finite path code is the maximum of the metadata entries in that code, or zero when the code is the empty list.✦ The injective coding map σ\sigma maps finite path codes to natural numbers, assigning to each code a number greater than its metadata bound.

A raw infinite path PP is a sequence of triples consisting of nonzero rational pieces uiu_i, declared weights jij_i, and metadata pip_i, indexed from zero.✦ Its cumulative support, endpoint, closure, and prefix code are An=⋃i≤nsupp⁡ui,βn=max⁡An,Hn=Fpn(βn),an=[(ji,pi,{(pos⁡Hn(α),ui(α)):α∈Hn}):0≤i≤n].\begin{gathered} A_n=\bigcup_{i\leq n}\operatorname{supp}u_i,\qquad \beta_n=\max A_n,\qquad H_n=F_{p_n}(\beta_n),\\ a_n=\bigl[(j_i,p_i, \{(\operatorname{pos}_{H_n}(\alpha),u_i(\alpha)): \alpha\in H_n\}):0\leq i\leq n\bigr]. \end{gathered}(6) A raw path is admissible when its piece supports are successive in ordinal order, the metadata are strictly increasing and positive, and 1≤jn≤pn,ρ(α,β)≤pn(α,β∈An,α≤β),jn+1=σ(an).1\leq j_n\leq p_n,\qquad \rho(\alpha,\beta)\leq p_n \quad(\alpha,\beta\in A_n,\ \alpha\leq\beta),\qquad j_{n+1}=\sigma(a_n).(7)✦

A finite path PP in a family ff is a raw path with a positive length ℓ\ell and original uncropped Type I expressions at its first ℓ\ell positions, whose values and weights agree with the raw data, whose first ℓ\ell pieces satisfy (7) and support and metadata succession, and whose pieces are successive in ff.✦ Entries beyond the prefix are unrestricted. Its finite path sum is ∑i<ℓui\sum_{i<\ell}u_i.✦ An infinite path PP in a family ff is a raw path with an original uncropped Type I expression at every position, whose values and weights agree with the raw data, whose pieces satisfy (7) and support and metadata succession and are successive in ff, and whose original Type I pieces have all their children in KfK_f.✦ Its prefix P|nP|_n has length n+1n+1 and retains the raw data and original pieces through index nn.✦ The path partial sums and the associated bounded path functional ΦP∈E*\Phi_P\in E^* are specified by ΦP,n=h∑i≤nui,ΦP(x)=limn→∞ΦP,n(x)(x∈E).\Phi_{P,n}=h_{\sum_{i\leq n}u_i},\qquad \Phi_P(x)=\lim_{n\to\infty}\Phi_{P,n}(x)\quad(x\in E).(8)

Reductions and the Binary Tree

An ordinal cut γ\gamma for Γ\Gamma is a countable ordinal or the top cut above all of Γ\Gamma.✦ For a cut γ\gamma, put Iγ={α:α<γ}I_\gamma=\{\alpha:\alpha<\gamma\}.✦ An ordinal and color rectangle A=(I,D)A=(I,D) consists of order-convex sets I⊆ΓI\subseteq\Gamma and D⊆ℕD\subseteq\mathbb N.✦ The support set of an ordinal and color rectangle A=(I,D)A=(I,D), where I⊆ΓI\subseteq\Gamma and D⊆ℕD\subseteq\mathbb N are order-convex, is II in a pure family and I∩c−1(D)I\cap c^{-1}(D) in the mixed family.✦ Ordinal interval crops have D=ℕD=\mathbb N; color interval crops have I=ΓI=\Gamma. A single color crop has D={k}D=\{k\}.

A coordinate set AA is crop stable for KK when g|A∈Kg|_A\in K for every g∈Kg\in K.✦ For a coordinate set AA that is crop-stable for KK and a normer g∈Kg\in K, the cropped normer is the rational array g|Ag|_A; the ambient crop operator of AA on ℓ∞(K)\ell_\infty(K) is the map sending xx to the function g↦x(g|A)g\mapsto x(g|_A). For a coordinate set AA that is crop-stable for KK, the crop operator PAP_A on EKE_K is the restriction of the ambient crop operator of AA to EKE_K.✦ The initial ordinal projection PγP_\gamma uses the initial ordinal crop (Iγ,ℕ)(I_\gamma,\mathbb N), and the single color projection PkcP_k^c uses (Γ,{k})(\Gamma,\{k\}). The family projection is Pγ,fk=PkcPγ,Pγ,fm=Pγ.P_{\gamma,f_k}=P_k^cP_\gamma,\qquad P_{\gamma,f_{\mathrm m}}=P_\gamma.(9)✦

A bounded linear operator TT between normed spaces is strictly singular when, for every closed infinite-dimensional subspace of its domain and every ε>0\varepsilon>0, some vector xx in that subspace satisfies ∥Tx∥<ε∥x∥\|Tx\|<\varepsilon\|x\|.✦ A branch reduction R=(γ,f,H,U)R=(\gamma,f,H,U) of a real subspace F⊆EF\subseteq E consists of a cut with no last coordinate, a family ff, a closed infinite-dimensional subspace H⊆EH\subseteq E, and a bounded linear map U:H→FU:H\to F. Every Pβ|HP_\beta|_H with β<γ\beta<\gamma is strictly singular; in the mixed case every Pkc|HP_k^c|_H is strictly singular as well. The lifting condition, using (9), is Pγ,fUx=x(x∈H).P_{\gamma,f}Ux=x\qquad(x\in H).(10) Write ℛ(F)\mathcal R(F) for the family of reductions satisfying (10) and these conditions.

An admissible block sequence z=(zn)z=(z_n) in ff consists of nonzero finitely supported real vectors with successive ordinal supports. In the mixed family their colors are also successive; in the pure family fkf_k every support coordinate has color kk. For an admissible block sequence z=(zn)z=(z_n) and natural numbers a,ba,b, the initial support of length bb is ⋃n<bsupp⁡zn\bigcup_{n<b}\operatorname{supp}z_n, and the algebraic tail span at aa is span⁡ℝ{zn:n≥a}\operatorname{span}_{\mathbb R}\{z_n:n\geq a\}.

For a branch reduction R=(γ,f,H,U)R=(\gamma,f,H,U), a threshold state (α,l)(\alpha,l) is a pair consisting of an ordinal α<γ\alpha<\gamma and a natural color threshold ll.✦ The block window of a threshold state (α,l)(\alpha,l) for a branch reduction R=(γ,f,H,U)R=(\gamma,f,H,U) is the set of coordinates in [α,γ)[\alpha,\gamma) of color kk when f=fkf=f_k, and of colors at least ll when ff is the mixed family.✦ A block approximation step at stage nn consists of a unit vector x∈Hx\in H, a nonzero finitely supported real vector vv in this window, and a next ordinal threshold η<γ\eta<\gamma strictly beyond its support, subject to ∥x−ιv∥<ϵn,ϵn=1640002−n.\|x-\iota v\|<\epsilon_n,\qquad \epsilon_n=\frac{1}{64000}\,2^{-n}.(11)✦ Starting with a chosen ordinal below γ\gamma and color threshold zero, successive chosen approximation steps and their threshold updates define the approximating block sequence zRz_R. Its nnth vector is the finite vector selected at the nnth threshold state in (11). The geometric approximation error ϵn\epsilon_n at stage nn is the quantity in (11), with indices starting at zero.

The vector support crop AyA_y of a nonzero finitely supported real vector yy on Γ\Gamma is the rectangle whose ordinal interval joins the minimum and maximum of the support of yy and whose color interval joins its minimum and maximum support colors.✦ An exact pair (y,e)(y,e) for an admissible block sequence zz in a family ff, tail index aa, and weight JJ consists of a nonzero vector yy in the algebraic tail span of zz at aa and an original Type I expression ee in ff of weight JJ. Every child of ee belongs to KfK_f; in a pure family yy has that family’s color. Writing ϕ\phi for the value of ee, its remaining conditions are ∥ιy∥≤500,⟨ϕ,ιy⟩≥12,supp⁡ϕ⊆the support set of the vector support crop of y in f.\|\iota y\|\leq500,\qquad \langle\phi,\iota y\rangle\geq\frac12,\qquad \operatorname{supp}\phi\subseteq \text{the support set of the vector support crop of }y\text{ in }f.(12) The pairing in (12) is (2).

An exact pair tree node dd for an admissible block sequence is a tuple consisting of a starting index aa, a positive weight JJ, an exact pair as in (12), a finishing index b≥ab\geq a, and positive metadata p≥Jp\geq J. Its vector is supported on the first bb blocks. For a finite list of such nodes, the accumulated functional support is the union of the supports of their Type I values, and the next starting index is the maximum of their finishing indices, with value zero for the empty list. The node recursion starts with the empty list and appends a chosen next exact pair node at every stage.✦ Denote its indexed nodes by DND_N.

Binary nodes are indexed from zero, with root 00 and parent ⌊(N−1)/2⌋\lfloor(N-1)/2\rfloor for N>0N>0. The ancestor list of a binary node NN is the list of its ancestors in order from the root to NN, including both the root and NN.✦ For a node assignment D=(DN)N∈ℕD=(D_N)_{N\in\mathbb N} and a selector v:ℕ→ℕv:\mathbb N\to\mathbb N, the raw path along the selected nodes Dv(n)D_{v(n)} has pieces equal to their Type I values and weights and metadata equal to the corresponding node entries.✦ The ancestral path code at NN is the prefix code along its ancestor list.✦ The requested weight for the next node is 11 for the empty list; for a list of length N>0N>0 it is σ\sigma of the ancestral code at ⌊(N−1)/2⌋\lfloor(N-1)/2\rfloor.✦

For b∈{0,1}ℕb\in\{0,1\}^{\mathbb N}, let vbv_b be the infinite binary branch starting at the root and following the child selected by bnb_n at depth nn.✦ The admissible branch path P(z,b)P(z,b) has the original uncropped Type I expressions from these nodes; its raw entries are (un,jn,pn)=(Type I value at Dvb(n),weight at Dvb(n),metadata at Dvb(n)).(u_n,j_n,p_n)= \bigl(\text{Type I value at }D_{v_b(n)},\, \text{weight at }D_{v_b(n)},\, \text{metadata at }D_{v_b(n)}\bigr).(13)✦ The node assignment in (13) is the one constructed on zz.

The Claim

Theorem 1. Let FF be a closed infinite-dimensional real linear subspace of EE. There is a branch reduction R=(γ,f,H,U)∈ℛ(F).R=(\gamma,f,H,U)\in\mathcal R(F).(14)✦ For every infinite path PP in either a pure or the mixed family, the path functional of (8) satisfies ∥ΦP∥≤1.\|\Phi_P\|\leq1.(15)✦ For any reduction RR as in (14), let zRz_R be its approximating block sequence and use its branch paths (13). For distinct b,c∈{0,1}ℕb,c\in\{0,1\}^{\mathbb N}, their restrictions satisfy ∥ΦP(zR,b)|F−ΦP(zR,c)|F∥F*≥12004(∥U∥+1).\left\|\Phi_{P(z_R,b)}|_F-\Phi_{P(z_R,c)}|_F\right\|_{F^*} \geq\frac{1}{2004(\|U\|+1)}.(16)✦

What Is Checked

The claim is stated over closed infinite-dimensional real subspaces FF of EE, their branch reductions RR and approximating block sequences zRz_R, and infinite paths in the pure and mixed families.

Each coordinate functional belongs to the coordinate predual (2).✦

Rectangle crops preserve KK✦, and their ambient crop operators preserve the closed evaluation span.✦ Every initial ordinal segment IγI_\gamma is order convex.✦

Every raw prefix support AnA_n in (6) is nonempty.✦ Finite prefix sums of infinite paths belong to KK.✦

Block approximation steps exist for every threshold state and stage in (11).✦ Every requested tree weight is positive.✦