OctoInformBeta

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

Lemma 5.3: Under the numerical schedule margins, with an error estimate for the second warmup step

Adjacent powers of nested atoms retain overlap.

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 repeatedly assembles a binary word from a pair of binary words called atoms. At nesting level jj, let hjh_j be the repetition count; the comparison is between deletion traces of hjh_j and hj+1h_j+1 copies of the same atom, whose source lengths differ.

Let q∈(0,1]q\in(0,1] be the requested deletion probability, so each bit survives with probability 1−q1-q. Put δ=min⁡(q,1/2)\delta=\min(q,1/2) and Λ=δ−3\Lambda=\delta^{-3}. The construction theorems call their working deletion rate qq; their rate is written δ\delta here. A binary word ww of length ℓ\ell assigns a bit to each position 0,…,ℓ−10,\ldots,\ell-1.✦ A deletion mask aa of a binary word ww is a choice of retained positions; the retained subsequence w|aw|_a lists their bits in order. A complete binary trace is a finite bit string, including its surviving length.✦ The complete trace space 𝒯M=⋃r=0M{0,1}r\mathcal T_M=\bigcup_{r=0}^M\{0,1\}^r consists of bounded traces with bound MM.✦

A finite probability law μ\mu on a finite set Ω\Omega is a nonnegative mass function with total mass one.✦ For a real test function ff on Ω\Omega and a natural half order kk, the expectation and squared moment norm are defined by (1); the latter uses probability weights. 𝔼μf=∑z∈Ωμ(z)f(z),∥f∥L2k(μ)2=(𝔼μ|f|2k)1/k.\mathbb E_\mu f=\sum_{z\in\Omega}\mu(z)f(z),\qquad \|f\|_{L^{2k}(\mu)}^2 =\left(\mathbb E_\mu |f|^{2k}\right)^{1/k}.(1) The point mass at z∈Ωz\in\Omega is the probability law on Ω\Omega that assigns mass one to zz.✦ For probability laws KzK_z on a common finite outcome set, indexed by z∈Ωz\in\Omega, the mixture of μ\mu and KzK_z is the probability law assigning mass ∑zμ(z)Kz(t)\sum_z\mu(z)K_z(t) to each outcome tt.✦ The pushforward f*μf_*\mu of a finite probability law μ\mu under a function ff is the mixture of μ\mu and the point masses at f(z)f(z).✦

The deletion mask law at rate p∈[0,1]p\in[0,1] assigns the weight in (2) to a mask.✦ The complete trace law Tr⁡p(w)\operatorname{Tr}_p(w) of a binary word ww at deletion rate pp is the pushforward of the deletion mask law under a↦w|aa\mapsto w|_a, on any 𝒯M\mathcal T_M with M≥|w|M\ge |w|.✦ The complete trace expectation of a test FF is 𝔼Tr⁡p(w)F\mathbb E_{\operatorname{Tr}_p(w)}F.✦ wt⁡p(a)=∏u=0ℓ−1{1−p,a(u)=1,p,a(u)=0.\operatorname{wt}_p(a) =\prod_{u=0}^{\ell-1} \begin{cases}1-p,&a(u)=1,\\p,&a(u)=0.\end{cases}(2)✦

The overlap of laws μ,ν\mu,\nu on a common finite set is the quantity in (3).✦ Complete trace overlap uses the common cutoff max⁡(|x|,|y|)\max(|x|,|y|) for source words x,yx,y.✦ overlap⁡(μ,ν)=∑tmin⁡{μ(t),ν(t)}=1−TV⁡(μ,ν),TV⁡(μ,ν)=12∑t|μ(t)−ν(t)|.\operatorname{overlap}(\mu,\nu) =\sum_t\min\{\mu(t),\nu(t)\} =1-\operatorname{TV}(\mu,\nu),\qquad \operatorname{TV}(\mu,\nu)=\frac12\sum_t|\mu(t)-\nu(t)|.(3)

An approximate trace simulation (μ,d)(\mu,d) of a binary word ww with error EE consists of a finite law μ\mu on Ω\Omega and a decoder d:Ω→𝒯Md:\Omega\to\mathcal T_M satisfying (4) for every trace test 0≤F≤10\le F\le1.✦ |𝔼μ[F(d(z))]−𝔼Tr⁡δ(w)F|≤E.\left|\mathbb E_\mu[F(d(z))] -\mathbb E_{\operatorname{Tr}_\delta(w)}F\right|\le E.(4)

The Scale Schedule and Input Pair

The rank Schur constant ASchA_{\mathrm{Sch}}, primitive period constant P0P_0, initial overlap lower bound ω\omega, and overlap constant CC are the constants in (5). ASch=8847362e1−e−1/1024,P0=36ASch2,ω=e−(32+3P0)Λ1536,C=240δω.\begin{aligned} A_{\mathrm{Sch}}&=884736\frac{2e}{1-e^{-1/1024}},& P_0&=36A_{\mathrm{Sch}}^2,\\ \omega&=\frac{e^{-(32+3P_0)\Lambda}}{1536},& C&=\frac{2^{40}}{\delta\omega}. \end{aligned}(5) For maximum block dimension BB and total block count DD, the warmup block count wBw_B, block dimensions bi+1b_{i+1}, block gain gg, total score exponents dtd_t, and scale allowance ε\varepsilon are given by (6). Block index ii starts at zero and corresponds to block i+1i+1 in the paper. wB=5⋅2B,bi+1={2,i<wB,B,i≥wB,g(b)=b−14−120,dt=115+∑i=0t−1g(bi+1),ε=124B+100(D+1).\begin{aligned} w_B&=5\cdot2^B,& b_{i+1}&=\begin{cases}2,&i<w_B,\\B,&i\ge w_B,\end{cases}\\ g(b)&=\frac{b-1}{4}-\frac1{20},& d_t&=\frac{11}{5}+\sum_{i=0}^{t-1}g(b_{i+1}),& \varepsilon&=\frac1{2^{4B+100}(D+1)}. \end{aligned}(6) For a scale parameter nn, the scale exponent AA, window exponent κ\kappa, scale exponent product XiX_i, and backwards target scale NiN_i are the quantities in (7). Subtraction in a natural index is truncated at zero. A(b,j)=(1+ε)2j−b−ε,κ(b,j)=A(b,(j−1)+)−3(2j−1)ε,Xi=∏u=iD−1A(bu+1,0),Ni=nXi.\begin{aligned} A(b,j)&=(1+\varepsilon)2^{j-b}-\varepsilon,\\ \kappa(b,j)&=A(b,(j-1)_+)-3(2^j-1)\varepsilon,\\ X_i&=\prod_{u=i}^{D-1}A(b_{u+1},0),&N_i&=n^{X_i}. \end{aligned}(7) The half moment order kik_i and source error allowance Δ\Delta are defined by (8); the even score exponent is ri=2kir_i=2k_i. The notation ⌈⋅⌉+\lceil\cdot\rceil_+ means the natural ceiling truncated at zero. ki=⌈108210B(dD+B+1)log⁡nlog⁡Ni⌉+,Δ=n−100(dD+B+1).k_i=\left\lceil 10^8 2^{10B}(d_D+B+1)\frac{\log n}{\log N_i} \right\rceil_+, \qquad \Delta=n^{-100(d_D+B+1)}.(8)

A concrete level pair UU at scale SS, level ii, error allowance Δ\Delta, score decay exponent dd, overlap threshold ω\omega, count multiplier K0K_0, and half order kk consists of distinct words x,yx,y of a common length ℓ\ell, two finite laws μ,ν\mu,\nu on a common outcome set, a common decoder, a real likelihood score σ\sigma, an error EE, and an integer repetition count mm. The laws simulate x,yx,y, respectively, in the sense of (4), and the remaining conditions are (9). The repeated word xmx^m is the concatenation of mm copies of xx.✦ S/2≤ℓ≤S,0≤E,E/ℓ≤(i+1)Δ,ν(z)=μ(z)(1+σ(z)),∥σ∥L2k(μ)2≤S−d,ℓ≤m≤K0ℓ,overlap⁡(Tr⁡δ(xm),Tr⁡δ(xm+1))≥ω.\begin{aligned} &S/2\le\ell\le S,\qquad 0\le E,\qquad E/\ell\le(i+1)\Delta,\\ &\nu(z)=\mu(z)(1+\sigma(z)),\qquad \|\sigma\|_{L^{2k}(\mu)}^2\le S^{-d},\\ &\ell\le m\le K_0\ell,\qquad \operatorname{overlap}\bigl(\operatorname{Tr}_\delta(x^m), \operatorname{Tr}_\delta(x^{m+1})\bigr)\ge\omega. \end{aligned}(9) A source level pair UU at level ii is a concrete level pair with parameters (10); write EUE_U for its error. S=Ni,d=di,K0=C2Bi+2,k=ki,S=N_i,\qquad d=d_i,\qquad K_0=C^{2Bi+2},\qquad k=k_i,(10) with Δ\Delta from (8) and ω\omega from (5).

Words and Counts Inside a Block

Fix a target scale NN and block dimension bb. The integer window size TjwinT_j^{\mathrm{win}} and extended block windows TjT_j are the sequences in (11), indexed by zero-based nesting level jj. Natural floors are truncated at zero, as indicated by the subscript ++. Tjwin=4⌊Nκ(b,j+1)/4⌋+,Tj={Tjwin,j<b,24,j≥b.T_j^{\mathrm{win}}=4\left\lfloor N^{\kappa(b,j+1)}/4\right\rfloor_+, \qquad T_j=\begin{cases}T_j^{\mathrm{win}},&j<b,\\24,&j\ge b.\end{cases}(11) A block length and overlap state aa is a quadruple of natural numbers recording word length, overlap count, copies, and trials; the last two describe the transition into that state.✦ Starting with the length ℓ=|x|\ell=|x| and overlap count mm of UU, the nested arithmetic states aja_j satisfy (12).✦ Here ℓj,hj\ell_j,h_j are the first two entries of aja_j, JjJ_j is the root trial count at (N,ℓj,Tjwin,hj)(N,\ell_j,T_j^{\mathrm{win}},h_j), and Jloc(Pj,hj)J_{\mathrm{loc}}(P_j,h_j) is the local trial count at copy count PjP_j and overlap count hjh_j. a0=(ℓ,m,1,0),Pj={Jj(2hj+1)+Tjwin,j+1=b,⌊NA(b,j+1)/ℓj⌋+,j+1≠b,tj={Jj,j+1=b,Jloc(Pj,hj),j+1≠b,aj+1=(ℓjPj,⌈ChjPj⌉+,Pj,tj).\begin{aligned} a_0&=(\ell,m,1,0),\\ P_j&=\begin{cases} J_j(2h_j+1)+T_j^{\mathrm{win}},&j+1=b,\\ \lfloor N^{A(b,j+1)}/\ell_j\rfloor_+,&j+1\ne b, \end{cases}\\ t_j&=\begin{cases}J_j,&j+1=b,\\J_{\mathrm{loc}}(P_j,h_j),&j+1\ne b,\end{cases}\\ a_{j+1}&=(\ell_jP_j,\lceil Ch_jP_j\rceil_+,P_j,t_j). \end{aligned}(12) PjP_j and tjt_j are the copy count and block trial count for the transition from level jj to j+1j+1.

For cancellation order ss, put ρ=2s\rho=2^s and let πs\pi_s be the Prouhet binary pattern. The recursive source block words WjW_j are pairs of words defined by (13), where 𝒞\mathcal C denotes complementary-word substitution. Its slot counts PjP_j and nesting levels start at zero. W0(0)=x,W0(1)=y,Wj+1=𝒞(Pj,(⌊Pj/2⌋−1)+,ρ,πs,Wj).\begin{aligned} W_0(0)&=x,&W_0(1)&=y,\\ W_{j+1}&=\mathcal C\left(P_j, (\lfloor P_j/2\rfloor-1)_+,\rho,\pi_s,W_j\right). \end{aligned}(13)✦

Window Budgets and Numerical Margins

For window counts TT and depth ee, the complete window leaf set ℒe\mathcal L_e consists of all paths in the deterministic window tree, including unoccupied paths✦; its cardinality λe\lambda_e is the source leaf count.✦ The window collection set 𝒱e\mathcal V_e indexes the local and global trial collections at every nonleaf vertex.✦ Their recursions are (14), where [Te]={0,…,Te−1}[T_e]=\{0,\ldots,T_e-1\} and the two root tags denote local and global collections. ℒ0={*},ℒe+1=[Te]×ℒe,λe=|ℒe|,𝒱0=⌀,𝒱e+1={local,global}⊔([Te]×𝒱e).\begin{aligned} \mathcal L_0&=\{*\},& \mathcal L_{e+1}&=[T_e]\times\mathcal L_e,& \lambda_e&=|\mathcal L_e|,\\ \mathcal V_0&=\varnothing,& \mathcal V_{e+1}&=\{\mathrm{local},\mathrm{global}\} \sqcup([T_e]\times\mathcal V_e). \end{aligned}(14) For local trial counts τ\tau, the scheduled global trial counts GeG_e obey (15). The collection trial count mvm_v at depth e+1e+1 is τe\tau_e at the local root tag and GeG_e at the global root tag; a collection below a child inherits its depth-ee count.✦ G0=0,Ge+1=⌊Te+1/4⌋τe.G_0=0,\qquad G_{e+1}=\lfloor T_{e+1}/4\rfloor\tau_e.(15) The hidden metadata tail bound HeH_e, fitting tail bound FeF_e, and hidden binomial cap hbinh_{\mathrm{bin}} are the quantities in (16). He(T,τ;ω)=∑v∈𝒱emv>0e−mvω4/16,hbin(m)=4mω2+2,F0(T,τ)=F1(T,τ)=0,Fe+2(T,τ)=Te+1(2exp[−2(Te/8)2(Te+1+1)τe]+Fe+1(T,τ)).\begin{aligned} H_e(T,\tau;\omega) &=\sum_{\substack{v\in\mathcal V_e\\m_v>0}} e^{-m_v\omega^4/16},& h_{\mathrm{bin}}(m)&=\sqrt{\frac4{m\omega^2+2}},\\ F_0(T,\tau)&=F_1(T,\tau)=0,\\ F_{e+2}(T,\tau) &=T_{e+1}\left( 2\exp\left[-\frac{2(T_e/8)^2}{(T_{e+1}+1)\tau_e}\right] +F_{e+1}(T,\tau)\right). \end{aligned}(16) For cancellation order ss, the sparsity count 𝒩e\mathcal N_e, centered moment budget 𝒰e\mathcal U_e, and clipping cost χe\chi_e are defined in (17). Their arguments include a half order kk, fitting threshold aa, and leaf-score norm allowance LL. 𝒩e(T)=ρ∏j=1eTj,z=(12kL)2𝒩e(T),𝒰e(T;k,L)=z1−z,χe(T,τ;k,ω,a,L)=He+1(T,τ;ω)+Fe+1(T,τ)a+6(4𝒰e(T;k,L))k.\begin{aligned} \mathcal N_e(T)&=\rho\prod_{j=1}^eT_j,& z&=(12kL)^2\mathcal N_e(T),& \mathcal U_e(T;k,L)&=\frac{z}{1-z},\\ \chi_e(T,\tau;k,\omega,a,L) &=H_{e+1}(T,\tau;\omega)+\frac{F_{e+1}(T,\tau)}a +\sqrt{6\bigl(4\mathcal U_e(T;k,L)\bigr)^k}. \end{aligned}(17)

A finite ordered role shape of depth ee and branching bound rr is an ordered rooted tree with at most rr children at each vertex and one of three roles 0,1,20,1,2 attached to each child; at depth zero there is one empty shape.✦ The shape and role count S(r,e)S(r,e) is the number of these shapes, without numerical child labels.✦ The binary tree budget ℬe(c)\mathcal B_e(c) for level costs cc and the interior label budget Ie,r(T)I_{e,r}(T) are given by (18). ℬ0(c)=1,ℬe+1(c)=ceℬe(c)2,Ie,r(T)=(∏j=1eTj)(r−2e+1)+.\begin{aligned} \mathcal B_0(c)&=1,& \mathcal B_{e+1}(c)&=c_e\mathcal B_e(c)^2,\\ I_{e,r}(T)&=\left(\prod_{j=1}^eT_j\right)^{(r-2^{e+1})_+}. \end{aligned}(18) The forced contribution budget 𝒥e,r\mathcal J_{e,r} is the quantity in (19); all truncated exponents there are natural integers. cj=4ρ6hbin(Gj+1)hbin(τj),𝒥e,r(T,τ;ω)=S(r,e+1)Ie,r(T)ρ2(r−2e+1)+(2ρ3hbin(τe))ℬe(c).\begin{aligned} c_j&=4\rho^6h_{\mathrm{bin}}(G_{j+1})h_{\mathrm{bin}}(\tau_j),\\ \mathcal J_{e,r}(T,\tau;\omega) &=S(r,e+1)I_{e,r}(T)\rho^{2(r-2^{e+1})_+} \bigl(2\rho^3h_{\mathrm{bin}}(\tau_e)\bigr)\mathcal B_e(c). \end{aligned}(19)✦ For smoothing allowance η\eta, fitting threshold aa, cutoff RR, and score weight ww, the sharp degree budget 𝒟e,r\mathcal D_{e,r} and sharp energy budget ℰe\mathcal E_e are defined by (20). 𝒟e,r(T,τ;ω,η,a)=4⋅3(e+1)r[(λe+1r)3(e+1)r(η2+a2)+𝟏{r≥2e+1}𝒥e,r(T,τ;ω)],ℰe(T,τ;R,ω,η,a,w)=∑r=1R−1w2r𝒟e,r(T,τ;ω,η,a)+4(w2𝒩e(T))R1−w2𝒩e(T).\begin{aligned} \mathcal D_{e,r}(T,\tau;\omega,\eta,a) &=4\cdot3^{(e+1)r}\left[ \binom{\lambda_{e+1}}r3^{(e+1)r}(\eta^2+a^2) +\mathbf1_{\{r\ge2^{e+1}\}}\mathcal J_{e,r}(T,\tau;\omega) \right],\\ \mathcal E_e(T,\tau;R,\omega,\eta,a,w) &=\sum_{r=1}^{R-1}w^{2r}\mathcal D_{e,r}(T,\tau;\omega,\eta,a) +\frac{4\bigl(w^2\mathcal N_e(T)\bigr)^R} {1-w^2\mathcal N_e(T)}. \end{aligned}(20)

The source degree cutoff RR, work size VV, cancellation order ss, moment envelope MM, weight envelope WW, cap envelope KK, tail denominator ZZ, and tail exponent ν\nu are the constants in (21). R=210B+40,V=dD+R+B+10,s=⌈100V/ε⌉+,M=(2⋅108210B(B+4)+1)(D+1)2BD,W=144M2,K=200C2B(D+1)+2ω−1,Z=512+9600Kω−4,ν=300(dD+B+1)+40V+5.\begin{aligned} R&=2^{10B+40},&V&=d_D+R+B+10,&s&=\lceil100V/\varepsilon\rceil_+,\\ M&=\bigl(2\cdot10^8 2^{10B}(B+4)+1\bigr)(D+1)2^{BD},& W&=144M^2,\\ K&=200C^{2B(D+1)+2}\omega^{-1},& Z&=512+9600K\omega^{-4},\\ \nu&=300(d_D+B+1)+40V+5. \end{aligned}(21) The bad role factor FbadF_{\mathrm{bad}}, forced role factor FforceF_{\mathrm{force}}, smoothing prefactor SsmS_{\mathrm{sm}}, and skeleton prefactor PskelP_{\mathrm{skel}} are specified in (22), with ρ=2s\rho=2^s for the source cancellation order in (21). Fbad(e,R)=∑r=0R−14⋅32(e+1)r,Fforce(e,R)=∑r=0R−14⋅3(e+1)rS(r,e+1),Ssm(s,K)=2s2+s((s+1)K)s+1,Pskel(s,e,K)=2ρ3Kℬe(j↦4ρ6K2).\begin{aligned} F_{\mathrm{bad}}(e,R)&=\sum_{r=0}^{R-1}4\cdot3^{2(e+1)r},\\ F_{\mathrm{force}}(e,R)&=\sum_{r=0}^{R-1}4\cdot3^{(e+1)r}S(r,e+1),\\ S_{\mathrm{sm}}(s,K)&=2^{s^2+s}\bigl((s+1)K\bigr)^{s+1},\\ P_{\mathrm{skel}}(s,e,K)&=2\rho^3K\, \mathcal B_e\bigl(j\mapsto4\rho^6K^2\bigr). \end{aligned}(22) The numerical schedule margins at target scale NN are the inequalities (23); the last two apply to every integer 0≤e≤B0\le e\le B. max⁡{1200K,48ρ,8Wρ2,Ssm(s,K),1536,64(K+1)(D+1),240(K+1)(W+1)2ρB+2}≤Nε,384Fbad(e,R)≤Nε,192Fforce(e,R)W2e+1Pskel(s,e,K)≤Nε.\begin{aligned} \max\bigl\{&1200K,\ 48\rho,\ 8W\rho^2,\ S_{\mathrm{sm}}(s,K),\ 1536,\\ &64(K+1)(D+1),\ 2^{40}(K+1)(W+1)^2\rho^{B+2}\bigr\} \le N^\varepsilon,\\ 384F_{\mathrm{bad}}(e,R)&\le N^\varepsilon,\\ 192F_{\mathrm{force}}(e,R)W^{2^{e+1}}P_{\mathrm{skel}}(s,e,K) &\le N^\varepsilon. \end{aligned}(23)✦

The Claim

Theorem 1. Let B≥2B\ge2, 0≤i<D0\le i<D, and n≥4n\ge4, and let UU be a source level pair at level ii and working rate δ∈(0,1/2]\delta\in(0,1/2]. Put N=Ni+1N=N_{i+1} and b=bi+1b=b_{i+1}. Assume the numerical schedule margins (23) at NN and the tail bound 2Bexp⁡(−Nε/Z)≤n−ν2B\exp(-N^\varepsilon/Z)\le n^{-\nu}. Form the words and counts from UU by (12) and (13), using the source cancellation order. The adjacent-power estimate is (24). overlap⁡(Tr⁡δ(Wj(0)hj),Tr⁡δ(Wj(0)hj+1))≥ω(0≤j≤b).\operatorname{overlap}\left( \operatorname{Tr}_\delta\bigl(W_j(0)^{h_j}\bigr), \operatorname{Tr}_\delta\bigl(W_j(0)^{h_j+1}\bigr)\right) \ge\omega\qquad(0\le j\le b).(24)✦ If i<wBi<w_B, use the dimension-two block schedule and put k=ki+1k=k_{i+1}, L=N−A(2,0)di/2L=N^{-A(2,0)d_i/2}, and a=N−20Va=N^{-20V}. Let Tj*=T0T^*_j=T_0 and τj*=t0\tau^*_j=t_0 for every jj, and set χ*=χ0(T*,τ*;k,ω,a,L)\chi^*=\chi_0(T^*,\tau^*;k,\omega,a,L). The second warmup estimate is (25). (12kL)2ρ≤12,χ*≤12,64(h2+1)ρℰ0(T*,τ*;2,ω,a,a,12kL)+2(h2+1)P1(χ*+T0EU)≤14.\begin{gathered} (12kL)^2\rho\le\frac12,\qquad \chi^*\le\frac12,\\ \sqrt{64(h_2+1)\rho\, \mathcal E_0(T^*,\tau^*;2,\omega,a,a,12kL)} +2(h_2+1)P_1\bigl(\chi^*+T_0E_U\bigr)\le\frac14. \end{gathered}(25)✦

In the paper, the result reads: under the input hypotheses of Proposition 5.1 and the specified word and count construction, every 1≤j≤b1\le j\le b satisfies overlap⁡(Tr⁡δ((Uj+)hj),Tr⁡δ((Uj+)hj+1))≥ω\operatorname{overlap}(\operatorname{Tr}_\delta((U_j^+)^{h_j}), \operatorname{Tr}_\delta((U_j^+)^{h_j+1}))\ge\omega, where Uj+U_j^+ is the positive nested atom.

What Is Checked

The claims are stated over finite binary words, finite approximate trace simulations with a probability-weighted likelihood-score moment bound, and the specified integer word and count schedules under the numerical margins and exponential tail bound. The deletion mask weights are nonnegative for rates in [0,1][0,1]✦, and their sum is one for every real deletion parameter.✦