OctoInformBeta

OpenAI preprint, September 2026 · §5. One amplification block: words and overlap

Proposition 5.1: Explicit numerical margins control the score budgets and normalized simulation error

One block advances a trace pair.

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

An amplification block acts on two distinct binary words together with finite laws that approximately simulate their complete deletion traces. A level pair also controls a likelihood score and the overlap between two consecutive repetition counts of its left word. The deletion probability requested for the channel is q∈(0,1)q\in(0,1); each bit survives with probability 1−q1-q. The working parameters are defined in (1). The construction statements call their working rate qq; here it is denoted by δ\delta. δ=min⁡(q,1/2),Λ=δ−3.\delta=\min(q,1/2),\qquad \Lambda=\delta^{-3}.(1)

A binary word xx of length ℓ\ell is a function assigning a bit to each zero-based position 0,…,ℓ−10,\ldots,\ell-1.✦ A deletion mask aa for a binary word xx of length ℓ\ell is a sequence of retention indicators indexed by the positions of xx.✦ The retained subsequence keeps the selected bits in their original order.✦ A complete trace zz of a binary word xx is a finite list of retained bits of xx in their original order, including its surviving length; the bounded trace space with bound MM is the set of all binary lists of lengths zero through MM. The deletion mask law and its pushforward, the complete trace law 𝖣δ(x)\mathsf D_\delta(x), use the weights in (2). For a real-valued function FF on the complete trace space, write 𝔼𝖣δ(x)F\mathbb E_{\mathsf D_\delta(x)}F for its expectation under the complete trace law. pδ(a)=∏j=0ℓ−1(1−δ)ajδ1−aj,aj∈{0,1}.p_\delta(a)=\prod_{j=0}^{\ell-1} (1-\delta)^{a_j}\delta^{1-a_j},\qquad a_j\in\{0,1\}.(2)✦

A finite probability law PP on a finite set Ω\Omega is a nonnegative mass function with total mass one.✦ For a finite probability law PP on Ω\Omega, a real-valued function ff on Ω\Omega, and a nonnegative integer half order kk, the expectation 𝔼Pf\mathbb E_P f and probability-weighted squared moment quantity ∥f∥L2k(P)2\|f\|_{L^{2k}(P)}^2 are real quantities defined by (3), with the squared moment quantity equal to one when k=0k=0. At level ii, let kik_i be the positive integer half-order and let ri=2kir_i=2k_i be the even score exponent; the source-pair schedule below specifies kik_i. 𝔼Pf=∑z∈ΩP(z)f(z),∥f∥L2k(P)2=(∑z∈ΩP(z)|f(z)|2k)1/k.\mathbb E_P f=\sum_{z\in\Omega}P(z)f(z),\qquad \|f\|_{L^{2k}(P)}^2= \left(\sum_{z\in\Omega}P(z)|f(z)|^{2k}\right)^{1/k}.(3)

For probability laws PP and QQ on a common finite outcome space, their overlap ov⁡(P,Q)\operatorname{ov}(P,Q) and total variation distance TV⁡(P,Q)\operatorname{TV}(P,Q) are the real quantities defined by (4). Complete trace overlap uses the common bounded trace space with bound equal to the larger source length. The word xmx^m is the concatenation of mm copies of xx.✦ ov⁡(P,Q)=∑zmin⁡(P(z),Q(z))=1−TV⁡(P,Q),TV⁡(P,Q)=12∑z|P(z)−Q(z)|.\operatorname{ov}(P,Q)=\sum_z\min(P(z),Q(z)) =1-\operatorname{TV}(P,Q),\qquad \operatorname{TV}(P,Q)=\frac12\sum_z|P(z)-Q(z)|.(4)

For a finite probability law PP on Ω\Omega, a decoder f:Ω→{0,1}*f:\Omega\to\{0,1\}^*, and a binary word xx, an approximate trace simulation of xx with error EE is a decoded law satisfying the test bound in (5). |𝔼P(F∘f)−𝔼𝖣δ(x)F|≤Efor every F:{0,1}*→[0,1].\left|\mathbb E_P(F\circ f)-\mathbb E_{\mathsf D_\delta(x)}F\right|\le E \quad\text{for every }F:\{0,1\}^*\to[0,1].(5)✦

A concrete level pair UU with parameters (N,Δ,d,ω,K,i,k)(N,\Delta,d,\omega,K,i,k) is a collection of distinct words x,yx,y of common length ℓ\ell, finite laws P,QP,Q on a common Ω\Omega, a common decoder ff, a likelihood score σ\sigma, a nonnegative simulation error EE, and an integer repetition count mm, subject to the conditions below. Write ℒ(N,Δ,d,ω,K,i,k)\mathcal L(N,\Delta,d,\omega,K,i,k) for the family of these pairs. Both decoded laws satisfy (5) for their respective words, and the remaining conditions are (6). N/2≤ℓ≤N,E/ℓ≤(i+1)Δ,Q(z)=P(z)(1+σ(z)),∥σ∥L2k(P)2≤N−d,ℓ≤m≤Kℓ,ov⁡(𝖣δ(xm),𝖣δ(xm+1))≥ω.\begin{gathered} N/2\le\ell\le N,\qquad E/\ell\le(i+1)\Delta,\\ Q(z)=P(z)(1+\sigma(z)),\qquad \|\sigma\|_{L^{2k}(P)}^2\le N^{-d},\\ \ell\le m\le K\ell,\qquad \operatorname{ov}\bigl(\mathsf D_\delta(x^m), \mathsf D_\delta(x^{m+1})\bigr)\ge\omega. \end{gathered}(6)

The Global Schedule

Let BB be the block parameter and DD the total number of blocks. Block indices are zero-based: index ii refers to the paper’s block number i+1i+1 and dimension bi+1b_{i+1}. The warmup block count WBW_B is an integer, the block gain gg is a real-valued function of the block dimension bb, and the total score exponent djd_j is a real number at level jj, with their values defined by (7). WB=5⋅2B,bi+1={2,i<WB,B,i≥WB,g(b)=b−14−120,dj=115+∑h=0j−1g(bh+1).\begin{gathered} W_B=5\cdot2^B,\qquad b_{i+1}=\begin{cases}2,&i<W_B,\\B,&i\ge W_B,\end{cases} \qquad g(b)=\frac{b-1}{4}-\frac1{20},\\ d_j=\frac{11}{5}+\sum_{h=0}^{j-1}g(b_{h+1}). \end{gathered}(7)

For a size parameter n≥4n\ge4, the scale slack ε\varepsilon, scale exponents Ab(j)A_b(j), scale exponent products πi\pi_i, and backwards target scales NiN_i are real quantities defined by (8). The scale index jj is zero-based, and a subscript ++ on an integer difference denotes truncation at zero. The window exponent κb(j)\kappa_b(j) for block dimension bb, zero-based nesting index jj, and slack ε\varepsilon is given in (8). ε=124B+100(D+1),Ab(j)=(1+ε)2j2b−ε,πi=∏h=iD−1Abh+1(0),Ni=nπi,κb(j)=Ab((j−1)+)−3(2j−1)ε.\begin{gathered} \varepsilon=\frac1{2^{4B+100}(D+1)},\qquad A_b(j)=(1+\varepsilon)\frac{2^j}{2^b}-\varepsilon,\\ \pi_i=\prod_{h=i}^{D-1}A_{b_{h+1}}(0),\qquad N_i=n^{\pi_i},\\ \kappa_b(j)=A_b((j-1)_+)-3(2^j-1)\varepsilon. \end{gathered}(8)

The initial overlap lower bound ω\omega and overlap constant CC in (9) use the rank Schur constant KRK_{\mathrm{R}} and primitive period constant KPK_{\mathrm{P}}. KR=8847362e1−e−1/1024,KP=36KR2,ω=e−(32+3KP)Λ1536,C=240δω.\begin{gathered} K_{\mathrm{R}}=884736\frac{2e}{1-e^{-1/1024}},\qquad K_{\mathrm{P}}=36K_{\mathrm{R}}^2,\\ \omega=\frac{e^{-(32+3K_{\mathrm{P}})\Lambda}}{1536},\qquad C=\frac{2^{40}}{\delta\omega}. \end{gathered}(9)

The source level-pair family ℒi\mathcal L_i consists of concrete level pairs with the parameters in (10). The natural ceiling ⌈u⌉+\lceil u\rceil_+ means the least nonnegative integer at least uu. The half moment order kik_i at level ii is the nonnegative integer in (10). Δ=n−100(dD+B+1),ki=⌈108210B(dD+B+1)log⁡nlog⁡Ni⌉+,ℒi=ℒ(Ni,Δ,di,ω,C2Bi+2,i,ki).\begin{gathered} \Delta=n^{-100(d_D+B+1)},\qquad k_i=\left\lceil10^8 2^{10B}(d_D+B+1) \frac{\log n}{\log N_i}\right\rceil_+,\\ \mathcal L_i=\mathcal L (N_i,\Delta,d_i,\omega,C^{2Bi+2},i,k_i). \end{gathered}(10)

For the block following a pair U∈ℒiU\in\mathcal L_i, set N=Ni+1N=N_{i+1} and b=bi+1b=b_{i+1}. The degree cutoff RR, work size VV, cancellation order ss, fitting threshold aa, leaf-score allowance LL, score weight ww, and gap ζ\zeta are specified in (11). The source gap lower bound is g0(B)=1/(128⋅22B)g_0(B)=1/(128\cdot2^{2B}), so ζ=(di+1)g0(B)\zeta=(d_i+1)g_0(B). R=210B+40,V=dD+R+B+10,s=⌈100V/ε⌉+,a=N−20V,L=N−Ab(0)di/2,k=ki+1,w=6(2k)L,ζ=di+1128⋅22B.\begin{gathered} R=2^{10B+40},\quad V=d_D+R+B+10,\quad s=\lceil100V/\varepsilon\rceil_+,\\ a=N^{-20V},\quad L=N^{-A_b(0)d_i/2},\quad k=k_{i+1},\quad w=6(2k)L,\quad \zeta=\frac{d_i+1}{128\cdot2^{2B}}. \end{gathered}(11)

The integer window sizes are multiples of four. The extended block window sizes TjT_j are nonnegative integer multiples of four defined by (12), where ⌊u⌋+\lfloor u\rfloor_+ is the nonnegative integer floor. Tj={4⌊Nκb(j+1)/4⌋+,j<b,24,j≥b.T_j=\begin{cases} 4\left\lfloor N^{\kappa_b(j+1)}/4\right\rfloor_+,&j<b,\\ 24,&j\ge b. \end{cases}(12)✦

The block arithmetic schedule starts from the input length ℓ\ell and repetition count mm. Its state at nesting index jj records a length ℓj\ell_j, count mjm_j, copy count, and trial count, initialized with copy count one and trial count zero. Write JjJ_j for the root trial count at (N,ℓj,Tj,mj)(N,\ell_j,T_j,m_j) and ν(P,m)\nu(P,m) for the local trial count at (P,m)(P,m). The scheduled block trials tjt_j and successive arithmetic states obey (13). ℓ0=ℓ,m0=m,Pj={Jj(2mj+1)+Tj,j+1=b,⌊NAb(j+1)/ℓj⌋+,j+1≠b,tj={Jj,j+1=b,ν(Pj,mj),j+1≠b,ℓj+1=ℓjPj,mj+1=⌈CmjPj⌉+.\begin{gathered} \ell_0=\ell,\qquad m_0=m,\\ P_j=\begin{cases} J_j(2m_j+1)+T_j,&j+1=b,\\ \lfloor N^{A_b(j+1)}/\ell_j\rfloor_+,&j+1\ne b, \end{cases}\qquad t_j=\begin{cases}J_j,&j+1=b,\\\nu(P_j,m_j),&j+1\ne b,\end{cases}\\ \ell_{j+1}=\ell_jP_j,\qquad m_{j+1}=\lceil Cm_jP_j\rceil_+. \end{gathered}(13)

The Block Budgets

The complete window tree has one leaf at depth zero and ThT_h ordered children above a tree of depth hh. Its leaf count ℓT(h)\ell_T(h) and source sparsity count 𝒩h\mathcal N_h are defined in (14). The latter omits the bottom window factor and uses the nesting parameter ss. ℓT(h)=∏j=0h−1Tj,𝒩h=2s∏j=1hTj.\ell_T(h)=\prod_{j=0}^{h-1}T_j,\qquad \mathcal N_h=2^s\prod_{j=1}^{h}T_j.(14)

A window collection index vv at depth hh is an index for a local or global collection at a nonleaf vertex of a window tree of depth hh. The index set is empty at depth zero; at depth h+1h+1 it contains the two root tags and pairs of a zero-based child index below ThT_h with a depth-hh collection index.✦ Local collections at that depth have tht_h trials, global collections have GhG_h trials, and descendant counts are inherited recursively.✦ The scheduled global counts GhG_h, hidden metadata tail HhH_h, and fitting tail FhF_h are defined by (15); mvm_v denotes the collection trial count. G0=0,Gh+1=⌊Th+1/4⌋th,Hh=∑v a collection at depth hmv>0e−mvω4/16,F0=F1=0,Fh+2=Th+1(2exp[−2(Th/8)2(Th+1+1)th]+Fh+1).\begin{gathered} G_0=0,\qquad G_{h+1}=\lfloor T_{h+1}/4\rfloor\,t_h,\qquad H_h=\sum_{\substack{v\text{ a collection at depth }h\\m_v>0}} e^{-m_v\omega^4/16},\\ F_0=F_1=0,\qquad F_{h+2}=T_{h+1}\left( 2\exp\!\left[-\frac{2(T_h/8)^2}{(T_{h+1}+1)t_h}\right]+F_{h+1}\right). \end{gathered}(15)

The hidden binomial cap hω(u)h_\omega(u) and hidden smoothing bound Sω(u,s)S_\omega(u,s) are the numerical quantities in (16). hω(u)=4uω2+2,Sω(u,s)=2s2(2v)sv,v=4(s+1)uω2+2.h_\omega(u)=\sqrt{\frac4{u\omega^2+2}},\qquad S_\omega(u,s)=2^{s^2}(2v)^sv,\qquad v=\sqrt{\frac{4(s+1)}{u\omega^2+2}}.(16)

An ordered role shape 𝒯\mathcal T of depth hh and branching bound rr is an ordered rooted tree with at most rr children per vertex and a role in {0,1,2}\{0,1,2\} assigned to each child; at depth zero there is one empty shape.✦ Let ρ(r,h)\rho(r,h) count these shapes.✦ For a sequence c=(cj)c=(c_j) of level costs, the binary tree budget ℬh[c]\mathcal B_h[c] and interior label budget Ih(r)I_h(r) are real quantities defined by (17). ℬ0[c]=1,ℬh+1[c]=chℬh[c]2,Ih(r)=(∏j=1hTj)(r−2h+1)+.\mathcal B_0[c]=1,\qquad \mathcal B_{h+1}[c]=c_h\mathcal B_h[c]^2,\qquad I_h(r)=\left(\prod_{j=1}^h T_j\right)^{(r-2^{h+1})_+}.(17)

For d=2sd=2^s and the level costs cjc_j in (18), the forced contribution budget Jh(r)J_h(r) is the real quantity defined by that display. cj=4d6hω(Gj+1)hω(tj),Jh(r)=ρ(r,h+1)Ih(r)d2(r−2h+1)+(2d3hω(th))ℬh[c].\begin{gathered} c_j=4d^6h_\omega(G_{j+1})h_\omega(t_j),\\ J_h(r)=\rho(r,h+1)I_h(r)d^{2(r-2^{h+1})_+} \bigl(2d^3h_\omega(t_h)\bigr)\mathcal B_h[c]. \end{gathered}(18)✦

For smoothing allowance η\eta and fitting threshold aa, the sharp degree budget Dh(r;η,a)D_h(r;\eta,a) and sharp energy budget ℰh(R;η,a,w)\mathcal E_h(R;\eta,a,w) are real quantities defined by (19). Dh(r;η,a)=4⋅3(h+1)r[(ℓT(h+1)r)3(h+1)r(η2+a2)+𝟏{2h+1≤r}Jh(r)],ℰh(R;η,a,w)=∑r=1R−1w2rDh(r;η,a)+4(w2𝒩h)R1−w2𝒩h.\begin{gathered} D_h(r;\eta,a)=4\cdot3^{(h+1)r} \left[\binom{\ell_T(h+1)}r3^{(h+1)r}(\eta^2+a^2) +\mathbf1_{\{2^{h+1}\le r\}}J_h(r)\right],\\ \mathcal E_h(R;\eta,a,w)= \sum_{r=1}^{R-1}w^{2r}D_h(r;\eta,a) +\frac{4(w^2\mathcal N_h)^R}{1-w^2\mathcal N_h}. \end{gathered}(19)

For the probability-weighted score allowance in (11), the centered moment budget UhU_h and clipping cost 𝒞h\mathcal C_h are real quantities defined by (20). Uh=w2𝒩h1−w2𝒩h,𝒞h=Hh+1+Fh+1a+sqrt6(4Uh)k.U_h=\frac{w^2\mathcal N_h}{1-w^2\mathcal N_h},\qquad \mathcal C_h=H_{h+1}+\frac{F_{h+1}}a+sqrt{6(4U_h)^k}.(20)

The numerical margins use the moment envelope M*M_*, weight envelope W*W_*, cap envelope K*K_*, tail denominator τ\tau, and tail exponent pp in (21). M*=(2⋅108210B(B+4)+1)(D+1)2BD,W*=144M*2,K*=200C2B(D+1)+2ω−1,τ=512+9600K*ω−4,p=300(dD+B+1)+40V+5.\begin{gathered} M_*=\bigl(2\cdot10^8 2^{10B}(B+4)+1\bigr)(D+1)2^{BD},\qquad W_*=144M_*^2,\\ K_*=200C^{2B(D+1)+2}\omega^{-1},\qquad \tau=512+9600K_*\omega^{-4},\\ p=300(d_D+B+1)+40V+5. \end{gathered}(21)

The bad role factor AhA_h, forced role factor QhQ_h, smoothing prefactor PsmP_{\mathrm{sm}}, and skeleton prefactor Psk(h)P_{\mathrm{sk}}(h) are real quantities defined by (22). Ah=∑r=0R−14⋅32(h+1)r,Qh=∑r=0R−14⋅3(h+1)rρ(r,h+1),Psm=2s2+s((s+1)K*)s+1,Psk(h)=2d3K*ℬh[j↦4d6K*2].\begin{gathered} A_h=\sum_{r=0}^{R-1}4\cdot3^{2(h+1)r},\qquad Q_h=\sum_{r=0}^{R-1}4\cdot3^{(h+1)r}\rho(r,h+1),\\ P_{\mathrm{sm}}=2^{s^2+s}((s+1)K_*)^{s+1},\qquad P_{\mathrm{sk}}(h)=2d^3K_*\mathcal B_h[j\mapsto4d^6K_*^2]. \end{gathered}(22)

The numerical schedule margins at scale NN mean all the inequalities in (23). Their quantified level index hh is zero-based. 1200K*≤Nε,48d≤Nε,8W*d2≤Nε,Psm≤Nε,384Ah≤Nε(0≤h≤B),192QhW*2h+1Psk(h)≤Nε(0≤h≤B),1536≤Nε,64(K*+1)(D+1)≤Nε,240(K*+1)(W*+1)2dB+2≤Nε.\begin{gathered} 1200K_*\le N^\varepsilon,\quad48d\le N^\varepsilon,\quad 8W_*d^2\le N^\varepsilon,\quad P_{\mathrm{sm}}\le N^\varepsilon,\\ 384A_h\le N^\varepsilon\quad(0\le h\le B),\\ 192Q_hW_*^{2^{h+1}}P_{\mathrm{sk}}(h)\le N^\varepsilon \quad(0\le h\le B),\\ 1536\le N^\varepsilon,\qquad64(K_*+1)(D+1)\le N^\varepsilon,\\ 2^{40}(K_*+1)(W_*+1)^2d^{B+2}\le N^\varepsilon. \end{gathered}(23)✦

The Claim

Theorem 1. Let B≥2B\ge2, i<Di<D, n≥4n\ge4, and 0<δ≤1/20<\delta\le1/2. Let U∈ℒiU\in\mathcal L_i. Assume the numerical schedule margins (23) at N=Ni+1N=N_{i+1} and the tail condition (24). 2Bexp⁡(−Nε/τ)≤n−p.2B\exp(-N^\varepsilon/\tau)\le n^{-p}.(24) Then the next-level source pair family is nonempty, as in (25). ℒi+1≠⌀.\mathcal L_{i+1}\ne\varnothing.(25)✦ For the block parameters and trial schedule of UU, the budget bounds are (26). w2𝒩b−1≤N−ζ/2/8,w2𝒩b−1≤12,64ℰb−1(R;a,a,w)≤N−di+1,Sω(tj,s)≤a(0≤j<b),𝒞b−1≤12,𝒞b−1+Fb≤Δ.\begin{gathered} w^2\mathcal N_{b-1}\le N^{-\zeta/2}/8,\qquad w^2\mathcal N_{b-1}\le\tfrac12,\\ 64\mathcal E_{b-1}(R;a,a,w)\le N^{-d_{i+1}},\qquad S_\omega(t_j,s)\le a\quad(0\le j<b),\\ \mathcal C_{b-1}\le\tfrac12,\qquad \mathcal C_{b-1}+F_b\le\Delta. \end{gathered}(26)✦ Independently, let ℓ,ℓ′,v,j\ell,\ell',v,j be nonnegative integers with ℓ′≥1\ell'\ge1 and vℓ≤ℓ′v\ell\le\ell', and let E,Δ,c,fE,\Delta,c,f be real numbers with Δ≥0\Delta\ge0, E≤(j+1)ΔℓE\le(j+1)\Delta\ell, and c+f≤Δc+f\le\Delta. The normalized error bound is (27). c+vE+fℓ′≤(j+2)Δ.\frac{c+vE+f}{\ell'}\le(j+2)\Delta.(27)✦

In the paper, the result reads: for the global parameters and 1≤j≤D1\le j\le D, a level-(j−1)(j-1) pair with common C,ω,MBC,\omega,M_B admits C′=max⁡(C,Cabs/ω)C'=\max(C,C_{\mathrm{abs}}/\omega) such that every preceding pair remains valid after enlarging CC to C′C', and a new level-jj pair exists with these same enlarged constants, dj=dj−1+(bj−1)/4−0.05d_j=d_{j-1}+(b_j-1)/4-0.05, and Ej≤Δ+(ℓj/ℓj−1)Ej−1E_j\le\Delta+(\ell_j/\ell_{j-1})E_{j-1}, uniformly in the joint regime, with C′C' retained at later levels.

What Is Checked

The statements (25) and (26) concern source level pairs built from actual binary words, finite simulation laws, and the prescribed block arithmetic schedule; (27) concerns real errors and integer length and leaf counts. The deletion mask weights in (2) are nonnegative for rates in [0,1][0,1]✦, and their sum over all masks is one for every real deletion parameter.✦