OctoInformBeta

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

§4.3 (setting): inserting final coefficients in a preselected prediction: Discarded candidates receive zero weight in the finite prediction parameter

Final coefficients identify the ideal cutoff.

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 prediction is selected before the retained children of a tree of nonexpansive maps have all been fixed. Its coefficients are indexed by candidate order, which includes candidates that will be discarded.

Let XX be a real normed space and K⊆XK\subseteq X a convex set. An ordered rooted tree TT is a tree with levels of positive finite sizes sℓs_\ell, with s0=1s_0=1, given by (1). Its parent maps πℓ:Tℓ+1→Tℓ\pi_\ell:T_{\ell+1}\to T_\ell are order-preserving and onto. Tℓ={0,…,sℓ−1}.T_\ell=\{0,\ldots,s_\ell-1\}.(1)✦ A weighted ordered tree TT is an ordered rooted tree with a positive weight ωℓ,u\omega_{\ell,u} for each child u∈Tℓ+1u\in T_{\ell+1}, satisfying the normalization in (2) at every parent p∈Tℓp\in T_\ell.✦ ωℓ,u>0,∑u∈Tℓ+1πℓ(u)=pωℓ,u=1.\omega_{\ell,u}>0, \qquad \sum_{\substack{u\in T_{\ell+1}\\\pi_\ell(u)=p}} \omega_{\ell,u}=1.(2)✦

A tree evaluation EE on KK is a family of selfmaps Mℓ,p:K→KM_{\ell,p}:K\to K, one for each node p∈Tℓp\in T_\ell, that are nonexpansive in the original ambient norm: ∥Mℓ,p(x)−Mℓ,p(z)∥≤∥x−z∥\|M_{\ell,p}(x)-M_{\ell,p}(z)\|\leq\|x-z\| for x,z∈Kx,z\in K.✦ The cutoff evaluation ℰℓ(H)\mathcal E_\ell(H) of an assignment H:Tℓ→KH:T_\ell\to K is the point of KK obtained by combining children with their weights and applying the node maps strictly above the cutoff, toward the root.✦ At level zero it returns H(0)H(0).

Candidates and Prediction Parameters

Fix a level kk. A complete candidate ledger LL at level kk consists of a finite candidate ordinal C={0,…,N−1}C=\{0,\ldots,N-1\}, where NN is the number of candidates, an order-preserving parent assignment P:C→TkP:C\to T_k, points yc∈Ky_c\in K for c∈Cc\in C, and an order-preserving injection r:Tk+1→Cr:T_{k+1}\to C satisfying (3).✦ The image of rr consists of the retained candidates. P(r(u))=πk(u)(u∈Tk+1).P(r(u))=\pi_k(u)\qquad(u\in T_{k+1}).(3)

For a retained node vv, its earlier candidate set IvI_v, the inclusion ιv:Iv↪C\iota_v:I_v\hookrightarrow C, and the restricted parent and point families σv\sigma_v and yvy^v are defined in (4). The retained point at u∈Tk+1u\in T_{k+1} is denoted by zuz_u there. Iv={0,…,r(v)−1},ιv(c)=c,σv=P∘ιv,yv=y∘ιv,zu=yr(u).\begin{gathered} I_v=\{0,\ldots,r(v)-1\},\qquad \iota_v(c)=c,\\ \sigma_v=P\circ\iota_v,\qquad y^v=y\circ\iota_v, \qquad z_u=y_{r(u)}. \end{gathered}(4)

The extended final coefficient family λ:C→ℝ\lambda:C\to\mathbb R assigns the retained edge weights to their candidate positions and zero elsewhere.✦ The actual parameter θv\theta^v at a retained node vv is the coefficient family on IvI_v obtained by restricting the extended final coefficient family λ:C→ℝ\lambda:C\to\mathbb R, as in (5).✦ λ(c)={ωk,u,c=r(u) for some u∈Tk+1,0,c∉r(Tk+1),θcv=λ(ιv(c))(c∈Iv).\lambda(c)= \begin{cases} \omega_{k,u},&c=r(u)\text{ for some }u\in T_{k+1},\\ 0,&c\notin r(T_{k+1}), \end{cases} \qquad \theta^v_c=\lambda(\iota_v(c))\quad(c\in I_v).(5)

For a finite candidate set II, a parent assignment σ:I→Tk\sigma:I\to T_k, and θ∈ℝI\theta\in\mathbb R^I, the parent-restricted coefficient gp,i(θ)g_{p,i}(\theta) and the parent coefficient mass mp(θ)m_p(\theta) are given by (6). The admissible coefficient set 𝒞(σ)\mathcal C(\sigma) is the subset of ℝI\mathbb R^I consisting of nonnegative coefficient families with mass at most one at every parent p∈Tkp\in T_k.✦ gp,i(θ)={θi,σ(i)=p,0,σ(i)≠p,mp(θ)=∑i∈Igp,i(θ),𝒞(σ)={θ∈ℝI:θi≥0(i∈I),mp(θ)≤1(p∈Tk)}.\begin{gathered} g_{p,i}(\theta)= \begin{cases} \theta_i,&\sigma(i)=p,\\ 0,&\sigma(i)\ne p, \end{cases} \qquad m_p(\theta)=\sum_{i\in I}g_{p,i}(\theta),\\ \mathcal C(\sigma)= \{\theta\in\mathbb R^I: \theta_i\geq0\ (i\in I),\ m_p(\theta)\leq1\ (p\in T_k)\}. \end{gathered}(6)

For points ai∈Xa_i\in X, a common point b∈Xb\in X, and real coefficients θi\theta_i, the affine residual parent point Rp(a,b;θ)R_p(a,b;\theta) is the point of XX given by (7).✦ With ai,b∈Ka_i,b\in K and θ∈𝒞(σ)\theta\in\mathcal C(\sigma), the cutoff residual input is the affine residual parent point Rp(a,b;θ)R_p(a,b;\theta) in (7).✦ Rp(a,b;θ)=b+∑i∈Igp,i(θ)(ai−b).R_p(a,b;\theta) =b+\sum_{i\in I}g_{p,i}(\theta)(a_i-b).(7)✦ The evaluated prediction 𝒫E,k(σ,a,b;θ)\mathcal P_{E,k}(\sigma,a,b;\theta) for a parent assignment σ:I→Tk\sigma:I\to T_k, points ai,b∈Ka_i,b\in K, and coefficients θ∈𝒞(σ)\theta\in\mathcal C(\sigma) is the root output in (8).✦ 𝒫E,k(σ,a,b;θ)=ℰk(p↦Mk,p(Rp(a,b;θ))).\mathcal P_{E,k}(\sigma,a,b;\theta) =\mathcal E_k\bigl(p\mapsto M_{k,p}(R_p(a,b;\theta))\bigr).(8)✦

The ideal cutoff output Hv,b:Tk+1→KH_{v,b}:T_{k+1}\to K for v∈Tk+1v\in T_{k+1} and b∈Kb\in K is the assignment in (9).✦ Hv,b(u)={zu,u<v,b,v≤u.H_{v,b}(u)= \begin{cases} z_u,&u<v,\\ b,&v\leq u. \end{cases}(9)✦

The Claim

Theorem 1. Let LL be a complete candidate ledger at level kk of a weighted ordered tree, and let EE be a tree evaluation on the convex set KK. For every retained node v∈Tk+1v\in T_{k+1}, the actual parameter satisfies (10): θv∈𝒞(σv).\theta^v\in\mathcal C(\sigma_v).(10)✦ For every b∈Kb\in K, the two root outputs agree as in (11): ℰk+1(Hv,b)=𝒫E,k(σv,yv,b;θv).\mathcal E_{k+1}(H_{v,b}) =\mathcal P_{E,k}(\sigma_v,y^v,b;\theta^v).(11)✦

What Is Checked

The assertions are stated over the complete finite candidate ledger, with its order-preserving retention map, the positive normalized edge weights, and the nonexpansive node maps on a convex subset of the original normed space. Every cutoff residual input with admissible coefficients belongs to KK.