OctoInformBeta

OpenAI preprint, September 2026 · §4. A tree of maps and separating functionals

§4.2 (setting): earlier candidates and compact prediction families: Weak limits of evaluated root predictions depend on every earlier candidate coefficient

Compact detector families from earlier candidates.

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

Earlier candidate points enter the predictions at a finite frontier of the tree. Their coefficients vary before the next map is chosen, while the predictions are root outputs of the maps already present. Let XX be a real normed space and KK a convex subset of XX. All norms below are the ambient norm of XX. Write XwX_w for the same vector space with its weak topology. Weak compactness of KK means compactness of its image in XwX_w.✦

An ordered rooted tree has nonempty finite levels Nℓ={0,…,mℓ−1}N_\ell=\{0,\ldots,m_\ell-1\}, with m0=1m_0=1, and order-preserving surjections πℓ:Nℓ+1→Nℓ\pi_\ell:N_{\ell+1}\to N_\ell assigning each child its parent.✦ A weighted ordered tree assigns positive real weights wℓ,cw_{\ell,c} to the children, with the normalization in (1) at each parent: ∑c∈Nℓ+1πℓ(c)=pwℓ,c=1(p∈Nℓ).\sum_{\substack{c\in N_{\ell+1}\\\pi_\ell(c)=p}}w_{\ell,c}=1 \qquad(p\in N_\ell).(1)✦

A tree of nonexpansive maps on KK assigns a nonexpansive selfmap Mℓ,vM_{\ell,v} of KK to each node v∈Nℓv\in N_\ell.✦ The cutoff evaluation Ek:KNk→KE_k:K^{N_k}\to K of a tree of nonexpansive maps on KK is the map sending a frontier assignment to the root value obtained by successive weighted collapses of the levels.✦ This evaluation applies node maps at levels ℓ<k\ell<k and does not reapply the maps at level kk; at level zero it takes the value at the root.✦

Earlier Candidate Coefficients

Let II be a finite set indexing every earlier candidate, including discarded candidates, and let PP be a finite frontier. The map α:I→P\alpha:I\to P assigns a parent to each candidate. For θ∈ℝI\theta\in\mathbb R^I, the parent-restricted coefficients gp,i(θ)g_{p,i}(\theta) and parent coefficient masses sp(θ)s_p(\theta) are the real numbers defined in (2): gp,i(θ)={θi,α(i)=p,0,α(i)≠p,sp(θ)=∑i∈Igp,i(θ).g_{p,i}(\theta)= \begin{cases} \theta_i,&\alpha(i)=p,\\ 0,&\alpha(i)\ne p, \end{cases} \qquad s_p(\theta)=\sum_{i\in I}g_{p,i}(\theta).(2) The admissible coefficient set Θα\Theta_\alpha is the subset of ℝI\mathbb R^I whose coefficients are nonnegative and whose parent masses are at most one, as in (3): Θα={θ∈ℝI:θi≥0(i∈I),sp(θ)≤1(p∈P)}.\Theta_\alpha= \{\theta\in\mathbb R^I: \theta_i\ge0\ (i\in I),\quad s_p(\theta)\le1\ (p\in P)\}.(3)✦ Discarded candidates have coefficient zero in the final coefficient choice.

For points yi∈Xy_i\in X and b∈Xb\in X, the predicted parent point rp(b,θ)∈Xr_p(b,\theta)\in X is the affine combination in (4), defined for every θ∈ℝI\theta\in\mathbb R^I: rp(b,θ)=b+∑i∈Igp,i(θ)(yi−b).r_p(b,\theta)=b+\sum_{i\in I}g_{p,i}(\theta)(y_i-b).(4)✦ When yi∈Ky_i\in K, b∈Kb\in K and θ∈Θα\theta\in\Theta_\alpha, the residual input at pp is rp(b,θ)r_p(b,\theta) regarded as a point of KK. This definition applies both to a tree cutoff, where P=NkP=N_k, and to an arbitrary finite frontier.

Evaluated Predictions and Weak Limits

Fix yi∈Ky_i\in K. At a tree cutoff kk, the evaluated prediction Fk(b,θ)∈KF_k(b,\theta)\in K is the root output in (5): Fk(b,θ)=Ek(p↦Mk,p(rp(b,θ))).F_k(b,\theta)= E_k\bigl(p\mapsto M_{k,p}(r_p(b,\theta))\bigr).(5)✦

Assume KK is weakly compact. For an indexed family q:J→Kq:J\to K and an ultrafilter VV on JJ, let LV(q)∈KL_V(q)\in K be a chosen limit of qq along VV, with convergence in XwX_w.✦ Fix sequences (bj)j∈ℕ(b_j)_{j\in\mathbb N} and (dj)j∈ℕ(d_j)_{j\in\mathbb N} in KK, and an ultrafilter UU on ℕ\mathbb N. For θ∈Θα\theta\in\Theta_\alpha, the prediction weak limit zk(θ)∈Kz_k(\theta)\in K is the point defined in (6), with the weak limit taken after evaluating the nonlinear root outputs: zk(θ)=LU(j↦Fk(bj,θ)).z_k(\theta)=L_U\bigl(j\mapsto F_k(b_j,\theta)\bigr).(6)✦ The cutoff detector set Dk⊂XD_k\subset X is the union in (7): Dk={d0,…,dk}∪{zk(θ):θ∈Θα}.D_k=\{d_0,\ldots,d_k\} \cup\{z_k(\theta):\theta\in\Theta_\alpha\}.(7)✦

For an arbitrary finite frontier PP, take selfmaps fp:K→Kf_p:K\to K and an evaluator A:KP→KA:K^P\to K. For b∈Kb\in K, θ∈Θα\theta\in\Theta_\alpha and m∈ℕm\in\mathbb N, the evaluated prediction FA(b,θ)∈KF_A(b,\theta)\in K at the finite frontier PP, its prediction weak limit zA(θ)∈Kz_A(\theta)\in K, and the detector set DA,m⊂XD_{A,m}\subset X are respectively the points and set defined in (8): FA(b,θ)=A(p↦fp(rp(b,θ))),zA(θ)=LU(j↦FA(bj,θ)),DA,m={dj:0≤j<m}∪{zA(θ):θ∈Θα}.\begin{aligned} F_A(b,\theta)&=A\bigl(p\mapsto f_p(r_p(b,\theta))\bigr),\\ z_A(\theta)&=L_U\bigl(j\mapsto F_A(b_j,\theta)\bigr),\\ D_{A,m}&=\{d_j:0\le j<m\} \cup\{z_A(\theta):\theta\in\Theta_\alpha\}. \end{aligned}(8)

The Claim

Theorem 1. Suppose KK is weakly compact and ∥a−b∥≤1\|a-b\|\le1 for all a,b∈Ka,b\in K. Fix the weighted tree of nonexpansive maps, the earlier candidates and their parent assignment at cutoff kk. For b∈Kb\in K and θ,η∈Θα\theta,\eta\in\Theta_\alpha, the evaluated prediction bound is (9): ∥Fk(b,θ)−Fk(b,η)∥≤∑i∈I|θi−ηi|.\|F_k(b,\theta)-F_k(b,\eta)\| \le\sum_{i\in I}|\theta_i-\eta_i|.(9)✦ The prediction weak limits associated with any sequence (bj)(b_j) in KK and any ultrafilter UU on ℕ\mathbb N satisfy (10): ∥zk(θ)−zk(η)∥≤∑i∈I|θi−ηi|.\|z_k(\theta)-z_k(\eta)\| \le\sum_{i\in I}|\theta_i-\eta_i|.(10)✦ For any sequence (dj)(d_j) in KK, the cutoff detector set has the compactness property (11): Dk is compact in the norm topology of X.D_k\text{ is compact in the norm topology of }X.(11)✦

For the finite frontier construction, suppose every fpf_p is nonexpansive and AA preserves every common pointwise norm bound: for H,H′∈KPH,H'\in K^P and ε∈ℝ\varepsilon\in\mathbb R, the inequalities ∥H(p)−H′(p)∥≤ε\|H(p)-H'(p)\|\le\varepsilon for all p∈Pp\in P imply ∥A(H)−A(H′)∥≤ε\|A(H)-A(H')\|\le\varepsilon. For every m∈ℕm\in\mathbb N, the corresponding compactness property is (12): DA,m is compact in the norm topology of X.D_{A,m}\text{ is compact in the norm topology of }X.(12)✦

What Is Checked

The statements concern convex weakly compact subsets of diameter at most one, finite earlier candidate families, and either a weighted ordered tree of nonexpansive maps or a finite frontier evaluator preserving common pointwise norm bounds.