OctoInformBeta

OpenAI preprint, September 2026 · §5. Switching leaves and completing the proof

§5 (proof of Theorem 1.1): a common branch vector and pointwise vanishing: Finite witnesses and pointwise vanishing on the closed difference span

Common branch detection excludes normalized minimal counterexamples.

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

The question concerns functionals along an infinite branch of a finite level tree: can they detect a common vector while tending pointwise to zero on the closed span of differences of a dense sequence?

Let XX be a real Banach space. For a subset KK of XX and a selfmap T:K→KT:K\to K, a subset ss of XX is invariant for TT when T(K∩s)⊆sT(K\cap s)\subseteq s.✦ Minimal invariant set data (K,T)(K,T) consist of a nonempty, norm-closed, bounded, convex, norm-separable, weakly compact subset KK of XX and a fixed-point-free nonexpansive selfmap TT of KK, such that every nonempty norm-closed convex invariant subset of KK equals KK.✦ All norms are the original ambient norm.

The closed difference span YY of KK and its whole closed unit ball BYB_Y, viewed as a subset of XX, are Y=span⁡ℝ(K−K)¯,BY={u∈Y:∥u∥≤1}.Y=\overline{\operatorname{span}_{\mathbb R}(K-K)}, \qquad B_Y=\{u\in Y:\|u\|\leq1\}.(1) Weak compactness of a subset of XX means compactness of its image in the weakly topologized ambient space.✦ Write KwK_{\mathrm w} for the copy of KK with the induced weak topology.✦

Normalized minimal invariant data D=(K,T,z)D=(K,T,z) are minimal invariant set data with diam⁡K=1\operatorname{diam}K=1, a norm-dense sequence (zi)i≥0(z_i)_{i\geq0} in KK, metrizable KwK_{\mathrm w}, and weakly compact BYB_Y from (1).✦ Write 𝒟(X)\mathcal D(X) for the class of these data.

Trees and Switching Data

An ordered rooted tree 𝒯\mathcal T has finite levels Nk={0,…,sk−1}N_k=\{0,\ldots,s_k-1\} with sk>0s_k>0 and s0=1s_0=1, and nondecreasing surjective parent maps pk:Nk+1→Nkp_k:N_{k+1}\to N_k.✦ Its root is o=0∈N0o=0\in N_0.✦ For i≤ni\leq n, the ancestor map Ain:Nn→NiA_i^n:N_n\to N_i is the composition of the intervening parent maps, with AnnA_n^n the identity.✦ A coherent branch bb is a choice bk∈Nkb_k\in N_k such that Ain(bn)=biA_i^n(b_n)=b_i whenever i≤ni\leq n.✦ The positive level nodes form the set 𝒩+={(k,v):k≥1,v∈Nk}\mathcal N_+=\{(k,v):k\geq1,\ v\in N_k\}.✦

For functions gk,v:M→ℝg_{k,v}:M\to\mathbb R on a set MM, a prefix detector for v∈Nnv\in N_n is a point d∈Md\in M satisfying 12≤gi,Ain(v)(d)(1≤i≤n).\frac12\leq g_{i,A_i^n(v)}(d)\qquad(1\leq i\leq n).(2)✦ Finite prefix witnesses in a subset BB of MM mean that, at every positive depth, some node has a detector in BB as in (2).

A weighted ordered tree 𝒯\mathcal T is an ordered rooted tree with a positive weight wk,cw_{k,c} for each child c∈Nk+1c\in N_{k+1}, satisfying ∑pk(c)=pwk,c=1\sum_{p_k(c)=p}w_{k,c}=1 for every parent pp.✦ A tree of nonexpansive maps (Fk,v)(F_{k,v}) on a convex set KK, indexed by an ordered rooted tree 𝒯\mathcal T, is a family consisting of a nonexpansive selfmap Fk,v:K→KF_{k,v}:K\to K for each node.✦ For an assignment H:Nk+1→KH:N_{k+1}\to K, the weighted child average CkHC_kH is the map from NkN_k to KK given by CkH(p)=∑pk(c)=pwk,cH(c)C_kH(p)=\sum_{p_k(c)=p}w_{k,c}H(c).✦ For an assignment HH from NnN_n to KK, the cutoff evaluation En(H)E_n(H) is the point of KK obtained at the root by repeated weighted child averaging and application of the node maps strictly above the cutoff.✦ At depth zero, E0(H)=H(o)E_0(H)=H(o).

For an anchor a∈Ka\in K and 0≤t≤10\leq t\leq1, the anchored averaging map StaS_t^a is the selfmap of KK given by Sta(v)=ta+(1−t)T(v)S_t^a(v)=ta+(1-t)T(v).✦ For an anchor a∈Ka\in K and 0<t≤10<t\leq1, the weighted resolvent ℛt(a)\mathcal R_t(a) is the unique fixed point in KK of the anchored averaging map StaS_t^a.✦ For positive integers nn, the discrete resolvent RnR_n is the selfmap of KK defined by Rn=ℛ1/nR_n=\mathcal R_{1/n}.✦ The terminal resolvent sequence (xj)(x_j) at an anchor x∈Kx\in K is the sequence in KK defined by xj=Rj+1(x)x_j=R_{j+1}(x) for j≥0j\geq0.✦ The generated resolvent family is the least family of selfmaps of KK containing the identity and all RnR_n, and closed under composition and convex mixing.✦ For selfmaps F,GF,G of KK, a convex mixture aF+bGaF+bG is the selfmap of KK given by this linear combination, where a,b≥0a,b\geq0 and a+b=1a+b=1.✦

Candidate ordinals start at zero and count discarded candidates. For an ordinal nn, its candidate budget is qnq_n below. An injective labeling ℓ:𝒩+→ℕ\ell:\mathcal N_+\to\mathbb N assigns retained budgets εk,v\varepsilon_{k,v} to positive level nodes. The cumulative budget minimum mkm_k and dyadic tolerance τk\tau_k are defined by qn=2−(n+9),εk,v=qℓ(k,v),m0=1,mk+1=min⁡({mk}∪{εk+1,v:v∈Nk+1}),τk=2−(k+1)mk.\begin{aligned} q_n&=2^{-(n+9)}, &\varepsilon_{k,v}&=q_{\ell(k,v)},\\ m_0&=1, &m_{k+1}&=\min\bigl(\{m_k\}\cup \{\varepsilon_{k+1,v}:v\in N_{k+1}\}\bigr),\\ \tau_k&=2^{-(k+1)}m_k. \end{aligned}(3) For x,b∈Kx,b\in K and v∈Nnv\in N_n, the ideal switched cutoff Hn,vx,bH_{n,v}^{x,b} is an assignment on NnN_n. The ancestor displacement sum Δn(v,u)\Delta_n(v,u) at u∈Ku\in K uses the same point at every proper ancestor: Hn,vx,b(r)={Fn,r(x),r<v,b,r≥v,Δn(v,u)=∑i=0n−1∥Fi,Ain(v)(u)−u∥.H_{n,v}^{x,b}(r)= \begin{cases}F_{n,r}(x),&r<v,\\ b,&r\geq v,\end{cases} \qquad \Delta_n(v,u)=\sum_{i=0}^{n-1} \|F_{i,A_i^n(v)}(u)-u\|.(4)

Switching tree data for D∈𝒟(X)D\in\mathcal D(X) consist of a weighted ordered tree with such an evaluation, an anchor x∈Kx\in K, the labeling ℓ\ell, and functionals fk,v∈X*f_{k,v}\in X^* of norm at most one at the positive level nodes. The maps are generated, and F0,oF_{0,o} is the identity. There are prediction points ak,v∈Ka_{k,v}\in K and an ultrafilter UU on ℕ\mathbb N refining the cofinite filter. With the budgets from (3) and cutoffs from (4), their conditions are ∥Ck(c↦Fk+1,c(x))(p)−x∥≤τk,Δk+1(v,Fk+1,v(x))≤εk+1,v,Ek+1(Hk+1,vx,xj)→Uweakak+1,v,fk+1,v(Fk+1,v(x)−ak+1,v)≥1−εk+1,v,fk+1,v(Fk+1,v(x)−zi)≥1−εk+1,v(0≤i<k+1).\begin{aligned} \|C_k(c\mapsto F_{k+1,c}(x))(p)-x\|&\leq\tau_k,\\ \Delta_{k+1}(v,F_{k+1,v}(x))&\leq\varepsilon_{k+1,v},\\ E_{k+1}(H_{k+1,v}^{x,x_j}) &\xrightarrow[U]{\mathrm{weak}}a_{k+1,v},\\ f_{k+1,v}(F_{k+1,v}(x)-a_{k+1,v})&\geq1-\varepsilon_{k+1,v},\\ f_{k+1,v}(F_{k+1,v}(x)-z_i)&\geq1-\varepsilon_{k+1,v} \qquad(0\leq i<k+1). \end{aligned}(5) These conditions hold at every parent p∈Nkp\in N_k and child v∈Nk+1v\in N_{k+1}; the weak convergence in (5) concerns the evaluated root outputs in the weakly topologized ambient space. Write 𝒮(D)\mathcal S(D) for the class of switching tree data for DD.

The Claim

Theorem 1. Let 𝒯\mathcal T be an ordered rooted tree, let BB be a nonempty subset of a set MM, and let gk,v:M→ℝg_{k,v}:M\to\mathbb R. Suppose there are finite prefix witnesses in BB. There are a coherent branch bb and points dn∈Bd_n\in B such that 12≤gi,bi(dn)(n≥0,1≤i≤n).\frac12\leq g_{i,b_i}(d_n)\qquad(n\geq0,\ 1\leq i\leq n).(6)✦

Let EE be a real normed space, B⊆EB\subseteq E weakly compact, (φi)i≥0⊆E*(\varphi_i)_{i\geq0}\subseteq E^*, and c∈ℝc\in\mathbb R. Suppose every finite initial family of halfspaces {u:φi(u)≥c}\{u:\varphi_i(u)\geq c\} has a common point in BB. There is u∈Bu\in B such that c≤φi(u)(i≥0).c\leq\varphi_i(u)\qquad(i\geq0).(7)✦

Let f∈X*f\in X^* have norm at most one. For y,a,b∈Xy,a,b\in X and e∈ℝe\in\mathbb R, assume ∥y−a∥,∥y−b∥≤1\|y-a\|,\|y-b\|\leq1 and f(y−a),f(y−b)≥1−ef(y-a),f(y-b)\geq1-e. Then |f(a−b)|≤e.|f(a-b)|\leq e.(8)✦

Let (eα)α∈I(e_\alpha)_{\alpha\in I} be a summable real family and β:ℕ→I\beta:\mathbb N\to I injective. Then limn→∞eβ(n)=0.\lim_{n\to\infty}e_{\beta(n)}=0.(9)✦

For any subset C⊆XC\subseteq X with a norm-dense sequence (qi)i≥0(q_i)_{i\geq0} in CC, span⁡ℝ{qi−qj:i,j≥0}¯=span⁡ℝ(C−C)¯.\overline{\operatorname{span}_{\mathbb R}\{q_i-q_j:i,j\geq0\}} =\overline{\operatorname{span}_{\mathbb R}(C-C)}.(10)✦

Let EE be a real normed space, s⊆Es\subseteq E, and (φn)n≥0⊆E*(\varphi_n)_{n\geq0}\subseteq E^* with ∥φn∥≤1\|\varphi_n\|\leq1. Assume φn(u)→0\varphi_n(u)\to0 for every u∈su\in s. For every u∈span⁡ℝs¯u\in\overline{\operatorname{span}_{\mathbb R}s}, limn→∞φn(u)=0.\lim_{n\to\infty}\varphi_n(u)=0.(11)✦

Let D=(K,T,z)∈𝒟(X)D=(K,T,z)\in\mathcal D(X) and 𝒯\mathcal T be an ordered rooted tree. At every node let yk,v∈Ky_{k,v}\in K, fk,v∈X*f_{k,v}\in X^* and ek,v∈ℝe_{k,v}\in\mathbb R. At positive levels assume ∥fk,v∥≤1\|f_{k,v}\|\leq1 and fk,v(yk,v−zi)≥1−ek,vf_{k,v}(y_{k,v}-z_i)\geq1-e_{k,v} for every 0≤i<k0\leq i<k. Suppose the real family (ek,v)(k,v)∈𝒩+(e_{k,v})_{(k,v)\in\mathcal N_+} is summable. There is a positive depth nn such that, for each v∈Nnv\in N_n and d∈BYd\in B_Y, some ii satisfies 1≤i≤n,fi,Ain(v)(d)<12.1\leq i\leq n,\qquad f_{i,A_i^n(v)}(d)<\frac12.(12)✦

For every D∈𝒟(X)D\in\mathcal D(X), 𝒮(D)=⌀.\mathcal S(D)=\varnothing.(13)✦

For every real Banach space XX, 𝒟(X)=⌀.\mathcal D(X)=\varnothing.(14)✦

What Is Checked

The branch assertion (6) is stated over arbitrary real-valued detector functions on finite ordered levels, whereas (7) uses continuous linear functionals on a weakly compact subset of a real normed space. The detector estimate (8) uses the ambient norm, and (9) concerns a summable real family indexed by an arbitrary set. Assertions (10) and (11) use norm-closed real linear spans. Assertions (12), (13) and (14) are stated over normalized minimal invariant data, including weak compactness of the whole ball BYB_Y.

For 0<t≤10<t\leq1, the anchored averaging map StaS_t^a is a strict contraction with contraction bound 1−t1-t in the ambient norm.✦ At every positive integer nn, the resolvent weight satisfies 0<1/n≤10<1/n\leq1.