OctoInformBeta

OpenAI preprint, September 2026 · §7. Vectors detected at a prescribed weight

Lemma 7.3: Coherent weight windows and scalar nullity along infinite paths

Thinning block indices to exclude four hits.

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

A finite path tests a block vector through the pieces whose weights lie in a chosen window. Different blocks have different windows, while the testing path stays fixed. The issue is simultaneous detection when the block vectors tend to zero under every infinite path functional.

The coordinate set is Γ=ω1\Gamma=\omega_1, where ω1\omega_1 is the first uncountable ordinal.✦ Fix a bijection π:Γ→ℕ×Γ\pi:\Gamma\longrightarrow\mathbb N\times\Gamma, and let c(α)c(\alpha) be the first coordinate of π(α)\pi(\alpha). The color classes Ck=c−1({k})C_k=c^{-1}(\{k\}) are unbounded and have cardinality ℵ1\aleph_1. An operation family ff is either the pure family of a color kk or the mixed family.✦ Its parameter is s=k+2s=k+2 in the former case and s=2s=2 in the latter.✦ Supports are successive for ff when every coordinate of the first precedes every coordinate of the second, with strictly increasing colors also required in the mixed family.✦

Weights and Norming

The zero-based parameter sequences consist of a depth dnd_n, an exponent scale ts,nt_{s,n}, and an auxiliary sequence bs,nb_{s,n}, with recursion (1). The corresponding positive sequence index is n+1n+1. dn=n+11,(ts,0,bs,0)=(d02+2d0+2,0),bs,n+1=bs,n+2⋅2s(s+1)ts,n,ts,n+1=ts,n+(bs,n+1+1)(dn+1+1)+1+dn+1+dn+12.\begin{aligned} d_n&=n+11,& (t_{s,0},b_{s,0})&=(d_0^2+2d_0+2,0),\\ b_{s,n+1}&=b_{s,n}+2\cdot2^{s(s+1)t_{s,n}},\\ t_{s,n+1}&=t_{s,n}+(b_{s,n+1}+1)(d_{n+1}+1) +1+d_{n+1}+d_{n+1}^2. \end{aligned}(1) The block lengths, normalization factors, and norming weights at parameter index nn are defined by (2), where rs=s/(s−1)=s′r_s=s/(s-1)=s'. Λs,n=2s(s+1)ts,n,μs,n=2(s−1)(s+1)ts,n+dn,ϑs,n=Λs,n1/rsμs,n.\Lambda_{s,n}=2^{s(s+1)t_{s,n}},\qquad \mu_{s,n}=2^{(s-1)(s+1)t_{s,n}+d_n},\qquad \vartheta_{s,n}=\frac{\Lambda_{s,n}^{1/r_s}}{\mu_{s,n}}.(2) For the family ff, write r=rsr=r_s and use the weight parameters in (3). Weight j≥1j\geq1 uses the underlying index j−1j-1; weight zero also uses index zero. Lj=Λs,max⁡{j−1,0},mj=μs,max⁡{j−1,0},θj=ϑs,max⁡{j−1,0}.L_j=\Lambda_{s,\max\{j-1,0\}},\qquad m_j=\mu_{s,\max\{j-1,0\}},\qquad \theta_j=\vartheta_{s,\max\{j-1,0\}}.(3)

A rational array uu on Γ\Gamma is a rational-valued function with finite support.✦ A type I expression ee of weight j≥1j\geq1 for ff consists of nonzero rational arrays vhv_h, 0≤h<ℓ0\leq h<\ell, with successive supports for ff and 1≤ℓ≤Lj1\leq\ell\leq L_j.✦ Its value v(e)v(e) is the normalized sum in (4). v(e)=1mj∑h<ℓvh.v(e)=\frac1{m_j}\sum_{h<\ell}v_h.(4)✦

The pure norming set KkK_k is generated from zero and the signed coordinate arrays of color kk. The mixed norming set KK is generated from ⋃kKk\bigcup_k K_k. In either construction, stage zero is the indicated base; stage n+1n+1 retains stage nn and adjoins values of type I expressions whose children lie in stage nn, and values of path expressions formed from original path pieces whose children all lie in stage nn.✦ Each norming set is the union of its zero-based stages.✦ Write Kf=KkK_f=K_k for the pure family of color kk, and Kf=KK_f=K for the mixed family.✦

A norming set KK on Γ\Gamma is a set of rational arrays containing every positive coordinate array and having all coefficients bounded in absolute value by one.✦ The bounded function space ℓ∞(K)\ell_\infty(K) of a norming set KK is the space of bounded real-valued functions on KK, with the supremum norm.✦ The finite support inclusion ι\iota and the real normed completion EE are given by (5); the coordinate evaluation vector at α\alpha is ι𝟏{α}\iota\mathbf1_{\{\alpha\}}. (ιx)(u)=∑α∈Γu(α)x(α)(u∈K),E=ι(c00(Γ,ℝ))¯ℓ∞(K).(\iota x)(u)=\sum_{\alpha\in\Gamma}u(\alpha)x(\alpha) \quad(u\in K),\qquad E=\overline{\iota(c_{00}(\Gamma,\mathbb R))}^{\,\ell_\infty(K)}.(5)

The norming evaluation functional associated with u∈Ku\in K is point evaluation at uu on EE.✦ The coordinate functional eα*e_\alpha^* is the evaluation associated with the positive coordinate array at α\alpha.✦ The coordinate predual X0X_0 and the rational coordinate combination uˆ∈X0\widehat u\in X_0 are defined in (6); the pairing there is rational array evaluation on EE. X0=span⁡ℝ{eα*:α∈Γ}¯E*,uˆ=∑α∈Γu(α)eα*,⟨u,x⟩=uˆ(x).X_0=\overline{\operatorname{span}_{\mathbb R} \{e_\alpha^*:\alpha\in\Gamma\}}^{\,E^*},\qquad \widehat u=\sum_{\alpha\in\Gamma}u(\alpha)e_\alpha^*,\qquad \langle u,x\rangle=\widehat u(x).(6)

An ordinal and color rectangle is a pair (I,D)(I,D) of order-convex subsets of Γ\Gamma and ℕ\mathbb N.✦ Its support set for a pure family is II; for the mixed family it is I∩c−1(D)I\cap c^{-1}(D).✦ A coordinate set AA is crop-stable for KK when u|A∈Ku|_A\in K for every u∈Ku\in K.✦ For a coordinate set AA crop-stable for KK, the ambient crop operator on ℓ∞(K)\ell_\infty(K) sends yy to the function u↦y(u|A)u\mapsto y(u|_A), and the coordinate crop projection PAP_A on EE is the restriction of this operator to EE. Rectangle projections use the mixed support set. For an ordinal interval II, the interval projection PIP_I uses the rectangle (I,ℕ)(I,\mathbb N).✦ For a crop-stable coordinate set AA and u∈Ku\in K, the cropped norming functional is the coordinate restriction u|Au|_A.

Paths and Windows

The finite closure system has sets Hn(β)H_n(\beta), indexed from zero at each β<ω1\beta<\omega_1, constructed recursively from the closure families at smaller endpoints. A predecessor closure family assigns finite sets Gn(α)G_n(\alpha) for α<β\alpha<\beta.✦ A predecessor closure family assigning finite sets Gn(α)G_n(\alpha) for α<β\alpha<\beta is coherent when these sets contain their endpoints, lie below them, increase to their initial segments, and satisfy Gn(γ)=Gn(α)∩[0,γ]G_n(\gamma)=G_n(\alpha)\cap[0,\gamma] whenever γ∈Gn(α)\gamma\in G_n(\alpha).✦ A coherent endpoint extension (An)(A_n) of a predecessor closure family GG at β\beta is an increasing sequence of finite subsets of [0,β][0,\beta], containing β\beta and exhausting that interval, whose restrictions at its members below β\beta agree with GG.✦ The endpoint construction supplies the sequence used in the recursion. Write Fp(β)=Hmax⁡{p−1,0}(β)F_p(\beta)=H_{\max\{p-1,0\}}(\beta), so a positive closure index pp selects underlying index p−1p-1, and index zero also selects the first entry.✦ The closure entry rank ρ(α,β)\rho(\alpha,\beta) is the first positive pp with α∈Fp(β)\alpha\in F_p(\beta) when α≤β\alpha\leq\beta, and is zero otherwise.✦

A raw infinite path is a sequence (ui,ji,pi)i≥0(u_i,j_i,p_i)_{i\geq0} of nonzero rational pieces, declared natural weights, and natural metadata.✦ Its prefix support SnS_n, endpoint βn\beta_n, and prefix closure DnD_n are given by (7). Sn=⋃i=0nsupp⁡ui,βn=max⁡Sn,Dn=Fpn(βn).S_n=\bigcup_{i=0}^n\operatorname{supp}u_i,\qquad \beta_n=\max S_n,\qquad D_n=F_{p_n}(\beta_n).(7) The order position of α\alpha in a finite set DD is pos⁡D(α)=|{β∈D:β<α}|\operatorname{pos}_D(\alpha)=|\{\beta\in D:\beta<\alpha\}|.✦ The positional array code of a rational array uu on a finite set DD is the collection of pairs (pos⁡D(α),u(α))(\operatorname{pos}_D(\alpha),u(\alpha)) for every α∈D\alpha\in D, including zero coefficients.✦ A finite path code is a finite list of triples consisting of a weight, metadata, and a positional array code.✦ The prefix code cnc_n is the list in (8), using one common closure for all its pieces. cn=[(ji,pi,{(pos⁡Dn(α),ui(α)):α∈Dn})]i=0n.c_n=\left[ \left(j_i,p_i, \{(\operatorname{pos}_{D_n}(\alpha),u_i(\alpha)):\alpha\in D_n\}\right) \right]_{i=0}^n.(8)✦ The maximum metadata of a code is the largest of its metadata entries, with value zero for the empty list.✦ The coding map σ\sigma pairs the canonical natural number code of a finite path code with one plus its maximum metadata.✦ It is injective, and its value exceeds every metadata entry of its code.

A raw path is admissible when its pieces have successive ordinal supports, its metadata are positive and strictly increasing, and it satisfies (9).✦ An admissible prefix of length ℓ\ell of raw path data (ui,ji,pi)i≥0(u_i,j_i,p_i)_{i\geq0} is the restriction to positions below ℓ\ell of data with successive ordinal supports and positive, strictly increasing metadata, satisfying (9) at those positions, with the coding rule imposed when n+1<ℓn+1<\ell. 1≤jn≤pn,ρ(α,β)≤pn(α,β∈Sn,α≤β),jn+1=σ(cn).1\leq j_n\leq p_n,\qquad \rho(\alpha,\beta)\leq p_n \quad(\alpha,\beta\in S_n,\ \alpha\leq\beta),\qquad j_{n+1}=\sigma(c_n).(9)✦

A finite path PP for ff is a positive-length admissible prefix of raw infinite data, together with type I expressions eie_i having values uiu_i and weights jij_i at every position i<ℓ(P)i<\ell(P), whose original pieces are successive for ff.✦ A finite path PP for ff, with type I expressions eie_i, is legal when every child of every eie_i belongs to KfK_f.✦ An infinite path QQ for ff consists of admissible raw data (ui,ji,pi)i≥0(u_i,j_i,p_i)_{i\geq0} and type I expressions eie_i for ff of values uiu_i and weights jij_i at every position, whose original pieces are successive for ff and whose children belong to KfK_f.✦ Its finite prefix Q≤nQ_{\leq n} retains the raw data and the first n+1n+1 expressions, so index zero includes one piece.✦

For a finite path PP and a weight window [J,T][J,T], the finite path sum U(P)U(P) and the weight window array UJ,T(P)U_{J,T}(P) are defined in (10); both use original uncropped rational pieces. U(P)=∑i<ℓ(P)ui,UJ,T(P)=∑i<ℓ(P)J≤ji≤Tui.U(P)=\sum_{i<\ell(P)}u_i,\qquad U_{J,T}(P)=\sum_{\substack{i<\ell(P)\\J\leq j_i\leq T}}u_i.(10) The path partial sum at nn is the norming evaluation functional associated with U(Q≤n)U(Q_{\leq n}).✦ The bounded path functional FQ∈E*F_Q\in E^* is chosen with the pointwise limit property (11). FQ(x)=limn→∞⟨U(Q≤n),x⟩(x∈E).F_Q(x)=\lim_{n\to\infty}\langle U(Q_{\leq n}),x\rangle \qquad(x\in E).(11)✦

For a legal finite path PP, a finitely supported real vector xx, and a weight window [J,T][J,T], a hit at threshold δ\delta is the condition ℋ(P,x,J,T;δ)\mathcal H(P,x,J,T;\delta) in (12). ℋ(P,x,J,T;δ)⇔δ<|⟨UJ,T(P),ιx⟩|.\mathcal H(P,x,J,T;\delta) \quad\Longleftrightarrow\quad \delta<|\langle U_{J,T}(P),\iota x\rangle|.(12)✦ For a finitely supported real vector xx, a real sequence aa, and a natural number NN, the coefficient ℓ1\ell_1 norm and the series tail notation are defined by (13). |x|1=∑α∈supp⁡x|x(α)|,tail⁡(a,N)=∑k=0∞a(k+N),τf(T)=tail⁡((1/mj)j≥0,T+1).|x|_1=\sum_{\alpha\in\operatorname{supp}x}|x(\alpha)|,\qquad \operatorname{tail}(a,N)=\sum_{k=0}^{\infty}a(k+N),\qquad \tau_f(T)=\operatorname{tail}\bigl((1/m_j)_{j\geq0},T+1\bigr).(13)

Coherent Window Blocks

An admissible block sequence z=(zn)n≥0z=(z_n)_{n\geq0} for ff consists of nonzero, finitely supported real vectors with successive supports for ff; in the pure family of color kk, all support coordinates belong to CkC_k.✦ Its embedded vectors are ιzn\iota z_n, and its algebraic tail span at aa is span⁡ℝ{zn+a:n≥0}\operatorname{span}_{\mathbb R}\{z_{n+a}:n\geq0\}.

The interval partition bound for x∈Ex\in E, cutoff NN, constant AA, and exponent rr is the condition in (14), for every 1≤d≤N1\leq d\leq N and every family of ordinal intervals I0,…,Id−1I_0,\ldots,I_{d-1} with every point of IhI_h below every point of IkI_k whenever h<kh<k. ∑h<d∥PIhx∥≤Ad1/r.\sum_{h<d}\|P_{I_h}x\|\leq A d^{1/r}.(14)✦

Fix a block sequence zz, an initial weight J*J_\ast, a real number ε\varepsilon, and a tail index aa. Coherent window blocks (yn,Jn,Tn)(y_n,J_n,T_n) for a block sequence zz, an initial weight J*J_\ast, a real number ε\varepsilon, and a tail index aa consist of an admissible block sequence y=(yn)y=(y_n) in the algebraic tail span of zz at aa, and natural window endpoints Jn,TnJ_n,T_n satisfying (15). Here the closure-rank condition ranges over α∈supp⁡yh\alpha\in\operatorname{supp}y_h and β∈supp⁡yk\beta\in\operatorname{supp}y_k whenever h,k<nh,k<n. 1≤∥ιyn∥≤2,J*<Jn≤Tn,Th<Jk(h<k),ρ(min⁡{α,β},max⁡{α,β})<Jn,|yn|1τf(Tn)≤ε,|yn|1τf(Tn)→0,FQ(ιyn)→0for every infinite path Q.\begin{aligned} 1&\leq\|\iota y_n\|\leq2,& J_\ast&<J_n\leq T_n,\\ T_h&<J_k\quad(h<k),& \rho(\min\{\alpha,\beta\},\max\{\alpha,\beta\})&<J_n,\\ |y_n|_1\tau_f(T_n)&\leq\varepsilon,& |y_n|_1\tau_f(T_n)&\longrightarrow0,\\ F_Q(\iota y_n)&\longrightarrow0&&\text{for every infinite path }Q. \end{aligned}(15) In addition, each ιyn\iota y_n satisfies the interval partition bound (14) with cutoff LJnL_{J_n} and constant 1616.

The Claim

Theorem 1. Fix a pure or mixed family ff.

For coherent window blocks (yn,Jn,Tn)(y_n,J_n,T_n), every δ>0\delta>0 and every infinite N⊆ℕN\subseteq\mathbb N admit the thinning in (16), where PP ranges over legal finite paths for ff. ∃W⊆N,W infinite,#{i∈W:ℋ(P,yi,Ji,Ti;δ)}≤3for every P.\exists\,W\subseteq N,\quad W\text{ infinite},\qquad \#\{i\in W:\mathcal H(P,y_i,J_i,T_i;\delta)\}\leq3 \quad\text{for every }P.(16)✦

More generally, let (xi)(x_i) be finitely supported real vectors with successive ordinal supports. Let Ji≤TiJ_i\leq T_i be natural window endpoints with Ti<JjT_i<J_j when i<ji<j. Suppose there are real constants MM and A≥0A\geq0 such that (17) holds. The first condition ranges over i<ji<j and α,β∈supp⁡xi\alpha,\beta\in\operatorname{supp}x_i; the third ranges over all type I expressions ee for ff whose children belong to KfK_f, with weight w(e)<Jiw(e)<J_i. ρ(min⁡{α,β},max⁡{α,β})≤Jj,∥ιxi∥≤M,|⟨v(e),ιxi⟩|≤Aθw(e),|xi|1τf(Ti)→0,FQ(ιxi)→0for every infinite path Q.\begin{aligned} \rho(\min\{\alpha,\beta\},\max\{\alpha,\beta\})&\leq J_j,\\ \|\iota x_i\|&\leq M,\\ |\langle v(e),\iota x_i\rangle|&\leq A\theta_{w(e)},\\ |x_i|_1\tau_f(T_i)&\longrightarrow0,\\ F_Q(\iota x_i)&\longrightarrow0 \quad\text{for every infinite path }Q. \end{aligned}(17) For every δ>0\delta>0 and infinite N⊆ℕN\subseteq\mathbb N, there is a thinning as in (18), with PP again ranging over legal finite paths for ff. ∃W⊆N,W infinite,#{i∈W:ℋ(P,xi,Ji,Ti;δ)}≤3for every P.\exists\,W\subseteq N,\quad W\text{ infinite},\qquad \#\{i\in W:\mathcal H(P,x_i,J_i,T_i;\delta)\}\leq3 \quad\text{for every }P.(18)✦

In the paper, the result reads: for the block sequence (yi)(y_i) and weight windows [Ji,Ti][J_i,T_i] just constructed, every δ>0\delta>0 and every infinite set of indices admit an infinite subset on which no finite uncropped path has four hits at threshold δ\delta.

What Is Checked

The claims are stated over the full norming completion EE, with legal finite paths for a fixed family ff, the reciprocal-normalization tails in (13), and scalar nullity under every infinite path functional.

The full norming set contains every positive coordinate array✦ and has coefficients of absolute value at most one. Raw prefix supports are nonempty✦, and the sum of every finite prefix of an infinite path belongs to KK.✦ Each infinite path admits the functional in (11) with norm at most one.✦ Rectangle crops preserve the full norming set✦; their ambient crop operators are contractive and preserve the closed evaluation span EE.

Every coordinate functional eα*e_\alpha^* belongs to the coordinate predual X0X_0.✦