OctoInformBeta

OpenAI preprint, September 2026 · §3. An anchor for successive choices

Lemma 3.5 [Adaptive finite selection]: Finite convex combinations respect changing thresholds, and positive support retains the selection order

Adaptive choices approximate an anchor.

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

Successive choices may change which maps are admissible at the next step. The approximation sought is a convex combination of their values at a prescribed point.

Let KK be a subset of a real normed space XX. The weak carrier KwK_{\mathrm w} is the copy of KK with the topology induced by the weak topology of XX.✦ The norm-to-weak identification ι:K→Kw\iota:K\to K_{\mathrm w} is the canonical bijection sending each point to its weak representative; the underlying norm point π(w)∈X\pi(w)\in X of w∈Kww\in K_{\mathrm w} is its representative in XX. A subset D⊆XD\subseteq X is weakly compact when its image in the weakly topologized ambient space is compact.✦

For a family (Hm)m≥0(H_m)_{m\geq0} of sets of selfmaps of KK and x∈Kx\in K, the weak tail image closure Cn(H,x)C_n(H,x) at n≥0n\geq0 is the closed subset of KwK_{\mathrm w} given by Cn(H,x)={ι(G(x)):G∈Hn+1}¯Kw.C_n(H,x)=\overline{\{\iota(G(x)):G\in H_{n+1}\}}^{\,K_{\mathrm w}}.(1) The anchor tail-limit set is the subset D(H,x)D(H,x) of XX obtained from the closures in (1) by D(H,x)=π(⋂n≥0Cn(H,x)).D(H,x)=\pi\left(\bigcap_{n\geq0}C_n(H,x)\right).(2)✦ An anchor x∈Kx\in K for (Hm)(H_m) is a point whose tail-limit set D=D(H,x)D=D(H,x) from (2) is a nonempty weakly compact subset of KK satisfying x∈co⁡(D)¯∥⋅∥,ι(D)⊆Cm−1(H,x)(m≥1).x\in\overline{\operatorname{co}(D)}^{\,\|\cdot\|}, \qquad \iota(D)\subseteq C_{m-1}(H,x)\quad(m\geq1).(3) The convex hull and its norm closure in (3) are taken in XX.

For a set MM of possible choices, an evaluation map v:M→Kv:M\to K assigns a point to each choice, and Fm⊆MF_m\subseteq M is the family admissible at level mm. Let SS be a set of states with initial state s*s_* and positive integer thresholds mi(s)m_i(s), for i≥0i\geq0 and s∈Ss\in S. An admissible transition rule ℛi⊆S×M×S\mathcal R_i\subseteq S\times M\times S on the state set SS with choices in MM is a family of relations such that every a∈Fmi(s)a\in F_{m_i(s)} has a successor s′s' with ℛi(s,a,s′)\mathcal R_i(s,a,s').

A legal selection of length N>0N>0 consists of states s0,…,sNs_0,\ldots,s_N and choices a0,…,aN−1a_0,\ldots,a_{N-1} satisfying s0=s*,ai∈Fmi(si),ℛi(si,ai,si+1)(0≤i<N).s_0=s_*,\qquad a_i\in F_{m_i(s_i)},\qquad \mathcal R_i(s_i,a_i,s_{i+1})\quad(0\leq i<N).(4)

The Claim

Theorem 1 (Adaptive finite selection). Suppose KwK_{\mathrm w} is metrizable, xx is an anchor for (Hm)(H_m) as in (3), and τ>0\tau>0. Fix a state set and positive thresholds, with an admissible transition rule for M=KKM=K^K and Fm=HmF_m=H_m. There are a positive integer NN, a legal selection with choices G0,…,GN−1G_0,\ldots,G_{N-1} as in (4), and real coefficients β0,…,βN−1\beta_0,\ldots,\beta_{N-1} satisfying (5): βi≥0(0≤i<N),∑i=0N−1βi=1,∥∑i=0N−1βiGi(x)−x∥<τ.\beta_i\geq0\quad(0\leq i<N),\qquad \sum_{i=0}^{N-1}\beta_i=1,\qquad \left\|\sum_{i=0}^{N-1}\beta_iG_i(x)-x\right\|<\tau.(5)✦

More generally, suppose KwK_{\mathrm w} is metrizable, v:M→Kv:M\to K is an evaluation map, Fm⊆MF_m\subseteq M, and z∈Kz\in K and E⊆KE\subseteq K satisfy the density conditions (6): z∈co⁡(E)¯∥⋅∥,ι(E)⊆{ι(v(a)):a∈Fm}¯Kw(m≥1).z\in\overline{\operatorname{co}(E)}^{\,\|\cdot\|},\qquad \iota(E)\subseteq \overline{\{\iota(v(a)):a\in F_m\}}^{\,K_{\mathrm w}}\quad(m\geq1).(6) For every τ>0\tau>0 and every admissible transition rule with positive thresholds and initial state s*s_*, there are a positive integer NN, a legal selection with choices a0,…,aN−1a_0,\ldots,a_{N-1}, and real coefficients satisfying (7): βi≥0(0≤i<N),∑i=0N−1βi=1,∥∑i=0N−1βiv(ai)−z∥<τ.\beta_i\geq0\quad(0\leq i<N),\qquad \sum_{i=0}^{N-1}\beta_i=1,\qquad \left\|\sum_{i=0}^{N-1}\beta_i v(a_i)-z\right\|<\tau.(7)✦

For any finite list u0,…,uN−1∈Xu_0,\ldots,u_{N-1}\in X and nonnegative real coefficients βi\beta_i with ∑i=0N−1βi=1\sum_{i=0}^{N-1}\beta_i=1, there are a positive integer qq and a strictly increasing map e:{0,…,q−1}→{0,…,N−1}e:\{0,\ldots,q-1\}\to\{0,\ldots,N-1\} satisfying (8): βe(j)>0(0≤j<q),∑j=0q−1βe(j)=1,∑j=0q−1βe(j)ue(j)=∑i=0N−1βiui.\begin{gathered} \beta_{e(j)}>0\quad(0\leq j<q),\qquad \sum_{j=0}^{q-1}\beta_{e(j)}=1,\\ \sum_{j=0}^{q-1}\beta_{e(j)}u_{e(j)} =\sum_{i=0}^{N-1}\beta_i u_i. \end{gathered}(8)✦

In the paper, the result reads: for the anchor xx of Proposition 3.4 and τ>0\tau>0, any finite integer mi≥1m_i\geq1 prescribed before the iith choice, depending on the entire preceding history and already fixed auxiliary data, admits a successive selection procedure choosing Gi∈HmiG_i\in H_{m_i} that terminates after finitely many steps NN with coefficients βi≥0\beta_i\geq0 assigned at termination, ∑i=0N−1βi=1\sum_{i=0}^{N-1}\beta_i=1 and ∥∑i=0N−1βiGi(x)−x∥<τ\|\sum_{i=0}^{N-1}\beta_iG_i(x)-x\|<\tau, and deleting the zero coefficients leaves a nonempty positive list in the original selection order.

What Is Checked

The selection statements concern anchors and evaluation families on KK with metrizable weak carrier, subject to admissible state transitions, while the positive-support statement concerns arbitrary finite convex combinations in XX.