OctoInformBeta

OpenAI preprint, September 2026 · §5. Disjoint windows of entropy gain

§5 (setting): separated bands and retention: Finite choices admit a divisible horizon, while depth and target scale determine retained windows

Separated bands from an eventual pair budget.

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

In a self-similar construction on the line, differences between cylinder translations are grouped by their dyadic sizes. The setting concerns finite collections of bands of these sizes and windows selected by contraction depth and a target dyadic scale.

A weighted affine iterated function system SS on a finite alphabet II consists of maps and weights satisfying (1), with signed slopes rir_i and arbitrary real translations tit_i: φi(x)=rix+ti,0<|ri|<1,pi>0,∑i∈Ipi=1.\varphi_i(x)=r_i x+t_i,\qquad 0<|r_i|<1,\qquad p_i>0,\qquad \sum_{i\in I}p_i=1.(1)✦ The affine composition of a word w∈Inw\in I^n is the map φw(x)=rwx+tw\varphi_w(x)=r_wx+t_w in (2); the empty word has slope and translation (1,0)(1,0), and a prefixed letter acts outermost.✦ The word probability pwp_w and contraction depth dwd_w of a word w∈Inw\in I^n are the real numbers in (2); the depth uses base-two logarithms. φ⌀(x)=x,φiw=φi∘φw,(riw,tiw)=(rirw,ritw+ti),pw=∏j=0n−1pwj,dw=−ln⁡|rw|ln⁡2.\begin{gathered} \varphi_\varnothing(x)=x,\qquad \varphi_{iw}=\varphi_i\circ\varphi_w,\qquad (r_{iw},t_{iw})=(r_i r_w,r_i t_w+t_i),\\ p_w=\prod_{j=0}^{n-1}p_{w_j},\qquad d_w=-\frac{\ln|r_w|}{\ln2}. \end{gathered}(2)

A finite probability law ν\nu on the line consists of a finite support Kν⊂ℝK_\nu\subset\mathbb R and strictly positive masses νx\nu_x, x∈Kνx\in K_\nu, whose sum is one.✦ For a finite source set UU, real weights πu\pi_u and an observable f:U→Ef:U\to E, the fiber mass mf(b)m_f(b) at b∈Eb\in E is the sum in (3).✦ When E=ℝE=\mathbb R and the weights are positive and normalized, these fiber masses define the observable distribution of ff.✦ mf(b)=∑u∈Uf(u)=bπu.m_f(b)=\sum_{\substack{u\in U\\f(u)=b}}\pi_u.(3)

The tag fiber Fτ(u)F_\tau(u) of a point u∈Uu\in U under a map τ:U→E\tau:U\to E is the subset in (4)✦; its fiber mass Mτ(u)M_\tau(u) is the total weight of that subset.✦ Fτ(u)={v∈U:τ(v)=τ(u)},Mτ(u)=mτ(τ(u)).F_\tau(u)=\{v\in U:\tau(v)=\tau(u)\},\qquad M_\tau(u)=m_\tau(\tau(u)).(4) For strictly positive weights π\pi and a real observable ff on UU, the conditional observable distribution νf∣τ(u)\nu_{f\mid\tau(u)} is the observable law of ff on Fτ(u)F_\tau(u) with weights πv/Mτ(u)\pi_v/M_\tau(u).✦ For a tag map τ:I→E\tau:I\to E, the word tag sequence τn(w)\tau_n(w) of a word w∈Inw\in I^n is the sequence (τ(w0),…,τ(wn−1))(\tau(w_0),\ldots,\tau(w_{n-1})).✦ The type-conditioned translation law νw\nu_w associated with a word w∈Inw\in I^n is the conditional observable distribution of v↦tvv\mapsto t_v on InI^n, with weights pvp_v, given τn(v)=τn(w)\tau_n(v)=\tau_n(w).✦

The pairs at band ℓ∈ℤ\ell\in\mathbb Z of a finite probability law ν\nu form the set Eℓ(ν)E_\ell(\nu) in (5).✦ The incident pair mass Iw(x)I_w(x) at x∈ℝx\in\mathbb R of weights ww on Eℓ(ν)E_\ell(\nu) is the half-sum in (5).✦ Eℓ(ν)={{x,y}⊂Kν:2−ℓ≤|x−y|<21−ℓ},Iw(x)=12∑e∈Eℓ(ν)x∈ewe.\begin{gathered} E_\ell(\nu)=\bigl\{\{x,y\}\subset K_\nu: 2^{-\ell}\le |x-y|<2^{1-\ell}\bigr\},\\ I_w(x)=\frac12\sum_{\substack{e\in E_\ell(\nu)\\x\in e}}w_e. \end{gathered}(5) A pair weighting ww on Eℓ(ν)E_\ell(\nu) is feasible when we≥0w_e\ge0 for every pair and Iw(x)≤νxI_w(x)\le\nu_x for every x∈Kνx\in K_\nu.✦ The fair-pair capacity cℓ(ν)c_\ell(\nu) is the supremum of total feasible pair weights in (6).✦ cℓ(ν)=supw feasible∑e∈Eℓ(ν)we.c_\ell(\nu)=\sup_{w\text{ feasible}}\sum_{e\in E_\ell(\nu)}w_e.(6)

For a real rate AA and length nn, the first dyadic band ana_n is the integer in (7).✦ For a real depth bound A′A' and integer band endpoints a,Ba,B, the averaged band capacity budget 𝒫A′(a,B;n)\mathcal P_{A'}(a,B;n) is the weighted sum in that display, restricted to words of depth at most A′nA'n.✦ an=⌈An⌉−1,𝒫A′(a,B;n)=∑w∈Indw≤A′npw∑ℓ∈[a,B]∩ℤcℓ(νw).\begin{gathered} a_n=\lceil An\rceil-1,\\ \mathcal P_{A'}(a,B;n)= \sum_{\substack{w\in I^n\\d_w\le A'n}}p_w \sum_{\ell\in[a,B]\cap\mathbb Z}c_\ell(\nu_w). \end{gathered}(7)

Retained Windows

Definition 1. Let σ:ℕ0→ℝ\sigma:\mathbb N_0\to\mathbb R be a depth profile, let A′∈ℝA'\in\mathbb R and n∈ℕ0n\in\mathbb N_0, and let a,B,L∈ℤa,B,L\in\mathbb Z. A retained window index j∈ℕ0j\in\mathbb N_0 for [j,j+n)[j,j+n) at target LL is an index satisfying (8): σ(j+n)−σ(j)≤A′n,a≤L−⌈σ(j)⌉≤B.\sigma(j+n)-\sigma(j)\le A'n,\qquad a\le L-\lceil\sigma(j)\rceil\le B.(8)✦

The Claim

Theorem 1. Let SS be a weighted affine iterated function system with a tag map on its alphabet. Let A>0A>0, A′<AA'<A and c∈ℝc\in\mathbb R. Suppose there is N∈ℕ0N\in\mathbb N_0 such that, for every n≥Nn\ge N, some integer B≥anB\ge a_n satisfies cn≤𝒫A′(an,B;n)cn\le\mathcal P_{A'}(a_n,B;n). For every integer q≥0q\ge0 there are lengths n1,…,nqn_1,\ldots,n_q and integer endpoints B1,…,BqB_1,\ldots,B_q satisfying (9): ni>0,ani≥1,cni≤𝒫A′(ani,Bi;ni)(1≤i≤q),Bh<ani−A′ni−1(1≤h<i≤q).\begin{gathered} n_i>0,\qquad a_{n_i}\ge1,\qquad cn_i\le\mathcal P_{A'}(a_{n_i},B_i;n_i) \quad(1\le i\le q),\\ B_h<a_{n_i}-A'n_i-1\quad(1\le h<i\le q). \end{gathered}(9)✦ For any finite family of positive integer lengths n1,…,nqn_1,\ldots,n_q and integer cutoffs B1,…,BqB_1,\ldots,B_q, there is a common horizon HH satisfying (10): H∈ℕ,H>0,ni∣H,Bi≤H(1≤i≤q).H\in\mathbb N,\qquad H>0,\qquad n_i\mid H,\qquad B_i\le H\quad(1\le i\le q).(10)✦

In the paper, the result reads: for any positive integer qq, choose sufficiently large n1<⋯<nqn_1<\cdots<n_q, with upper band endpoints BniB_{n_i}, for which Proposition 4.2 holds, ani≥1a_{n_i}\ge1, and ani−A′ni−1>Bnha_{n_i}-A'n_i-1>B_{n_h} whenever h<ih<i; choose an even TT with T/2>max⁡iBniT/2>\max_i B_{n_i}; with DD the paper’s block-depth bound, set Ji={kni:0≤k≤⌊T/(2Dni)⌋}J_i=\{kn_i:0\le k\le\lfloor T/(2Dn_i)\rfloor\}; at a target L∈{1,…,T}L\in\{1,\ldots,T\}, retain [j,j+ni)[j,j+n_i), j∈Jij\in J_i, exactly when σ(j+ni)−σ(j)≤A′ni\sigma(j+n_i)-\sigma(j)\le A'n_i and ani≤L−⌈σ(j)⌉≤Bnia_{n_i}\le L-\lceil\sigma(j)\rceil\le B_{n_i}.

What Is Checked

The selection statement (9) is stated over a finite weighted affine system, an arbitrary tag map and an eventual band budget; the horizon statement (10) concerns arbitrary finite families of positive integer lengths and integer cutoffs. Retention in (8) is defined for any real depth profile on the nonnegative integers and integer band and target indices.

Every word probability pwp_w is strictly positive.✦ For any finite source with strictly positive weights, the conditional weights πv/Mτ(u)\pi_v/M_\tau(u) on Fτ(u)F_\tau(u) sum to one.✦