OctoInformBeta

OpenAI preprint, September 2026 · §6. An observation with randomly displaced windows

Lemma 6.1: Leaf errors accumulate over the complete tree, with explicit bounds on fitting and hidden metadata tails

A window oracle simulates root traces.

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 randomly displaced window tree replaces each occurrence of an initial binary word by a leaf observation and records paired observations from the surrounding runs. The root observation is the concatenation of these emissions in their prescribed order. The oracle considered here retains the complete tree, including paths that fail to fit inside their parent windows.

Let q∈(0,1]q\in(0,1] be the requested deletion probability, so that each bit survives with probability 1−q1-q. Set δ=min⁡(q,1/2)\delta=\min(q,1/2) and Λ=δ−3\Lambda=\delta^{-3}; the construction theorems call their working rate qq. The simulation statement also allows an arbitrary rate ρ∈[0,1]\rho\in[0,1].

A binary word ww of length mm is a function assigning a bit to each position in [m]={0,…,m−1}[m]=\{0,\ldots,m-1\}.✦ A deletion mask a∈{0,1}[m]a\in\{0,1\}^{[m]} retains the positions where a(j)=1a(j)=1, in their original order.✦ The mask weight pρ(a)p_{\rho}(a) of a deletion mask a∈{0,1}[m]a\in\{0,1\}^{[m]} is its probability at deletion rate ρ\rho, given by (1).✦ The complete trace law Tr⁡ρ(w)\operatorname{Tr}_{\rho}(w) is the pushforward of this mask law under the retained subsequence map, on the bounded trace space ⋃k=0m{0,1}k\bigcup_{k=0}^{m}\{0,1\}^{k}.✦ A binary trace includes its surviving length. pρ(a)=∏j=0m−1(1−ρ)a(j)ρ1−a(j).p_{\rho}(a)=\prod_{j=0}^{m-1}(1-\rho)^{a(j)}\rho^{1-a(j)}.(1)

A finite probability law μ\mu on a finite set SS is a nonnegative mass function with ∑x∈Sμ(x)=1\sum_{x\in S}\mu(x)=1✦; for a real test function FF on SS, its expectation is 𝔼μF=∑xμ(x)F(x)\mathbb E_{\mu}F=\sum_x\mu(x)F(x).✦ Products μ⊗ν\mu\otimes\nu and coordinate products ⨂jμj\bigotimes_j\mu_j describe independent draws. A mixture of laws KxK_x with mixing law μ\mu has mass ∑xμ(x)Kx(y)\sum_x\mu(x)K_x(y), and their joint law has mass μ(x)Kx(y)\mu(x)K_x(y) at (x,y)(x,y).✦ Write f#μf_{\#}\mu for the pushforward under ff and δx\delta_x for the point mass concentrated at xx.

The complete trace expectation of a test FF is 𝔼Tr⁡ρ(w)F\mathbb E_{\operatorname{Tr}_{\rho}(w)}F.✦ For a finite law μ\mu, a decoder ff into binary traces and a binary word ww, the simulation discrepancy 𝒟ρ(μ,f;w)\mathcal D_{\rho}(\mu,f;w) at deletion rate ρ\rho is the supremum of the expectation differences in (2). Approximate trace simulation with error EE means 𝒟ρ(μ,f;w)≤E\mathcal D_{\rho}(\mu,f;w)\le E✦; exact trace simulation means equality of the decoded and complete trace expectations for every real trace test.✦ 𝒟ρ(μ,f;w)=supF:{0,1}*→[0,1]|𝔼μ[F∘f]−𝔼Tr⁡ρ(w)F|.\mathcal D_{\rho}(\mu,f;w) =\sup_{F:\,\{0,1\}^{*}\to[0,1]} \left|\mathbb E_{\mu}[F\circ f] -\mathbb E_{\operatorname{Tr}_{\rho}(w)}F\right|.(2)

For finite laws on a common outcome space, total variation and overlap use the normalization in (3). For binary words x,yx,y, complete trace overlap uses their two trace laws in the common bounded trace space with cutoff max⁡(|x|,|y|)\max(|x|,|y|).✦ The repeated word wmw^m is the concatenation of mm copies of ww.✦ TV⁡(μ,ν)=12∑z|μ(z)−ν(z)|,ov⁡(μ,ν)=∑zmin⁡{μ(z),ν(z)}=1−TV⁡(μ,ν).\operatorname{TV}(\mu,\nu)=\frac12\sum_z|\mu(z)-\nu(z)|, \qquad \operatorname{ov}(\mu,\nu)=\sum_z\min\{\mu(z),\nu(z)\} =1-\operatorname{TV}(\mu,\nu).(3)

The Window Oracle

Fix a Prouhet order ss, integer sequences of copy counts PjP_j, window counts TjT_j, local trial counts tjt_j and repetition counts hjh_j, and initial words a0,a1a_0,a_1. For a root orientation τ∈{0,1}\tau\in\{0,1\}, the recursive source words WjτW_j^{\tau} are the complementary-word substitutions in (4), where WjW_j denotes the pair of source words at level jj, πs\pi_s is the Prouhet binary pattern and 𝒞\mathcal C denotes complementary-word substitution.✦ Natural subtraction u∸vu\dotminus v means max⁡(u−v,0)\max(u-v,0). W0τ=aτ,Wj+1τ=𝒞(Pj,⌊Pj/2⌋∸1,2s,πs,Wj)τ.W_0^{\tau}=a_{\tau},\qquad W_{j+1}^{\tau} =\mathcal C\left(P_j,\lfloor P_j/2\rfloor\dotminus1, 2^s,\pi_s,W_j\right)^{\tau}.(4)

The scheduled global trial counts and their positions are defined by (5). A global trial at level j+1j+1 has a zero-based pair of indices (v,a)(v,a) with v<⌊Tj+1/4⌋v<\lfloor T_{j+1}/4\rfloor and a<tja<t_j; its single index is vtj+avt_j+a.✦ Its first and second emissions occur at the indicated pre and post positions. g0=0,gj+1=⌊Tj+1/4⌋tj,pre⁡j+1(vtj+a)=v,post⁡j+1(vtj+a)=Tj+1−1−v.\begin{aligned} g_0&=0,&g_{j+1}&=\lfloor T_{j+1}/4\rfloor t_j,\\ \operatorname{pre}_{j+1}(vt_j+a)&=v,& \operatorname{post}_{j+1}(vt_j+a)&=T_{j+1}-1-v. \end{aligned}(5)

The complete window leaf set ℒd\mathcal L_d, trial set 𝒥d\mathcal J_d, filler set ℱd\mathcal F_d and collection set 𝒱d\mathcal V_d are the recursively indexed sets in (6). Levels are numbered from the bottom, and all child and trial indices are zero-based. A collection is one vertex’s local or global trial collection.✦ Write Ld=|ℒd|L_d=|\mathcal L_d|, md=|𝒥d|m_d=|\mathcal J_d| and kd=|ℱd|k_d=|\mathcal F_d| for the source leaf, trial and filler counts. ℒ0={*},𝒥0=ℱ0=𝒱0=⌀,ℒd+1=[Td]×ℒd,𝒥d+1=[td]⊔[gd]⊔([Td]×𝒥d),ℱd+1=([Td]×{0,1})⊔([Td]×ℱd),𝒱d+1={local,global}⊔([Td]×𝒱d).\begin{aligned} &\mathcal L_0=\{*\},\qquad \mathcal J_0=\mathcal F_0=\mathcal V_0=\varnothing,\\ &\mathcal L_{d+1}=[T_d]\times\mathcal L_d,\\ &\mathcal J_{d+1}=[t_d]\sqcup[g_d] \sqcup([T_d]\times\mathcal J_d),\\ &\mathcal F_{d+1}=([T_d]\times\{0,1\}) \sqcup([T_d]\times\mathcal F_d),\\ &\mathcal V_{d+1}=\{\mathrm{local},\mathrm{global}\} \sqcup([T_d]\times\mathcal V_d). \end{aligned}(6)

A window orientation assignment on 𝒥d\mathcal J_d is a function assigning a bit to every trial in 𝒥d\mathcal J_d.✦ At depth zero there is one empty assignment. Choose fixed enumerations of the finite trial, filler and leaf sets. Flattening an assignment gives an array ξ∈{0,1}[md]\xi\in\{0,1\}^{[m_d]}.✦

The nominal prefix and suffix counts pj−,pj+p_j^-,p_j^+, trial bundle reserve ℛj\mathcal R_j and local center cj∈ℤc_j\in\mathbb Z are defined by (7). For a depth bb, the source layout margins ℳb\mathscr M_b require, for every j<bj<b, an even window count Tj≥16T_j\ge16, 2Tj≤Pj2T_j\le P_j, 8⋅2s≤Tj8\cdot2^s\le T_j, an even local trial count tjt_j, and 2ℛj≤min⁡(pj−,pj+)2\mathcal R_j\le\min(p_j^-,p_j^+).✦ pj−=⌊(Pj∸Tj)/2⌋,pj+=Pj∸Tj∸pj−,ℛj=tjhj+⌊tj/2⌋,cj=⌊Tj/2⌋−1+⌊tj/2⌋.\begin{aligned} p_j^-&=\left\lfloor(P_j\dotminus T_j)/2\right\rfloor,& p_j^+&=P_j\dotminus T_j\dotminus p_j^-,\\ \mathcal R_j&=t_jh_j+\lfloor t_j/2\rfloor,& c_j&=\lfloor T_j/2\rfloor-1+\lfloor t_j/2\rfloor. \end{aligned}(7)

The source trial word for a root local trial at depth j+1j+1 is (Wj0)hj+σ(W_j^0)^{h_j+\sigma} for side σ∈{0,1}\sigma\in\{0,1\}; for a root global trial at depth j+2j+2 it is the same word. A trial below a child uses that child’s trial word. The source filler words at depth one are empty. At depth j+2j+2, the filler on side σ\sigma of child uu is the run in (8), where Φ0\Phi_0 and Φ1\Phi_1 denote the source prefix and suffix filler counts and pj0=pj−p_j^0=p_j^-, pj1=pj+p_j^1=p_j^+. Fillers below children use the same recursion. After enumeration, denote the trial words by Vd(ℓ,σ)V_d(\ell,\sigma) and the filler words by Zd(ℓ)Z_d(\ell). Zj+2(u,σ)=(Wj0)Φσ(Tj+1,pjσ,tj,hj,u).Z_{j+2}(u,\sigma) =(W_j^0)^{\Phi_{\sigma}(T_{j+1},p_j^{\sigma},t_j,h_j,u)}.(8)

The recursive leaf orientation starts with a root orientation τ\tau, an incoming integer shift and a sequence of integer centers. At depth zero its value is τ\tau. At a nonleaf, the integer template at the source node anchor determines the orientation passed to a child; that child’s scheduled shift is its incoming shift for the next step. With centers cjc_j from (7), root incoming shift zero and flattened trial assignment ξ\xi, write θd,τ(ξ)∈{0,1}[Ld]\theta_{d,\tau}(\xi)\in\{0,1\}^{[L_d]} for the resulting flattened leaf orientations.

A paired mechanism on a finite side outcome space SS is a collection of laws Fj,Gj,HjF_j,G_j,H_j and weights wj≥0w_j\ge0 indexed by the mdm_d trial slots, satisfying wjHj(x)≤min⁡(Fj(x),Gj(x))w_jH_j(x)\le\min(F_j(x),G_j(x)) for every x∈Sx\in S.✦ A paired report consists of an optional orientation flag e∈{*,0,1}e\in\{*,0,1\} and two chronological outcomes (x,y)∈S2(x,y)\in S^2; ** denotes a hidden flag. The oriented pair laws Qj,σQ_{j,\sigma}, hidden pair masses ηj\eta_j, flagged pair masses Aj,σA_{j,\sigma} and report laws 𝜚j\varrho_j are defined in (9). The joint report law 𝒫\mathcal P on arrays of mdm_d paired reports is the coordinate product of the individual report laws 𝜚j\varrho_j, so reports at distinct trial slots are independent. Qj,0=Fj⊗Gj,Qj,1=Gj⊗Fj,ηj(x,y)=wj2Hj(x)Hj(y),Aj,σ(e,x,y)={ηj(x,y),e=*,Qj,σ(x,y)−ηj(x,y),e=σ,0,e∈{0,1}\{σ},𝜚j(e,x,y)=12∑σ=01Aj,σ(e,x,y),𝒫=⨂j=0md−1𝜚j.\begin{aligned} Q_{j,0}&=F_j\otimes G_j,&Q_{j,1}&=G_j\otimes F_j,\\ \eta_j(x,y)&=w_j^2H_j(x)H_j(y),\\ A_{j,\sigma}(e,x,y)&= \begin{cases} \eta_j(x,y),&e=*,\\ Q_{j,\sigma}(x,y)-\eta_j(x,y),&e=\sigma,\\ 0,&e\in\{0,1\}\setminus\{\sigma\}, \end{cases}\\ \varrho_j(e,x,y)&=\tfrac12\sum_{\sigma=0}^1A_{j,\sigma}(e,x,y),& \mathcal P&=\bigotimes_{j=0}^{m_d-1}\varrho_j. \end{aligned}(9)

For a report array zz, the joint orientation posterior Ψz\Psi_z is the coordinate product of the flag posteriors✦: a hidden flag gives a fair bit and a visible flag σ\sigma gives δσ\delta_{\sigma}.✦ Given finite leaf laws 𝖠0,𝖠1\mathsf A_0,\mathsf A_1 on a finite space EE, the selected leaf law for orientations ϑ\vartheta is ⨂ℓ<Ld𝖠ϑ(ℓ)\bigotimes_{\ell<L_d}\mathsf A_{\vartheta(\ell)}.✦ The canonical source oracle 𝒪d,τ\mathcal O_{d,\tau} is the joint law of reports and leaf observations with mass (10).✦ 𝒪d,τ(z,ℓ)=𝒫(z)∑ξ∈{0,1}[md]Ψz(ξ)∏v=0Ld−1𝖠θd,τ(ξ)(v)(ℓ(v)).\mathcal O_{d,\tau}(z,\ell) =\mathcal P(z)\sum_{\xi\in\{0,1\}^{[m_d]}}\Psi_z(\xi) \prod_{v=0}^{L_d-1}\mathsf A_{\theta_{d,\tau}(\xi)(v)}(\ell(v)).(10)

An emission slot is one side of a trial, a filler, or a leaf.✦ Trial and filler slots have outcomes in SS, and leaf slots have outcomes in EE.✦ The public emission chronology at depth zero contains the sole leaf. At depth j+1j+1 it lists the first local trial emissions, then for each child in increasing order its first scheduled global emissions, prefix filler, child chronology, second scheduled global emissions and suffix filler, followed by the second local trial emissions. Fixed enumerations identify these positions with the flattened emission slots.✦ For decoders f:S→{0,1}*f:S\to\{0,1\}^{*} and g:E→{0,1}*g:E\to\{0,1\}^{*}, the public decoder KdK_d applies ff to trial and filler outcomes and gg to leaf outcomes, then concatenates the resulting words in this chronology.✦

The fitting tail FdF_d is the finite tree bound for paths that fail to fit, defined in (11).✦ For a collection v∈𝒱dv\in\mathcal V_d, let ava_v be its trial count: tjt_j at a local collection at level jj, gjg_j at a global collection, and the corresponding recursive count below a child.✦ For an overlap threshold ω\omega, the hidden metadata tail Hd(ω)H_d(\omega) is the sum in (12); empty collections contribute zero.✦ F0=F1=0,Fj+2=Tj+1(2exp[−2(Tj/8)2(Tj+1+1)tj]+Fj+1).F_0=F_1=0,\qquad F_{j+2}=T_{j+1}\left( 2\exp\left[-\frac{2(T_j/8)^2}{(T_{j+1}+1)t_j}\right]+F_{j+1}\right).(11) Hd(ω)=∑v∈𝒱dav>0exp⁡(−avω4/16).H_d(\omega)=\sum_{\substack{v\in\mathcal V_d\\a_v>0}} \exp(-a_v\omega^4/16).(12)

The Numerical Schedule

Fix maximum block dimension B≥2B\ge2, block count DD and ambient size n≥4n\ge4. The warmup block count is wB=5⋅2Bw_B=5\cdot2^B.✦ At zero-based block index ii, the dimension βi\beta_i corresponds to block number i+1i+1, denoted bi+1b_{i+1} in the paper. The block gain gblkg_{\mathrm{blk}}, total score exponent did_i, scale slack ε\varepsilon, scale exponent a(b,j)a(b,j) and window exponent κ(b,j)\kappa(b,j) are specified in (13). βi={2,i<wB,B,i≥wB,gblk(b)=b−14−120,di=115+∑j=0i−1gblk(βj),ε=124B+100(D+1),a(b,j)=(1+ε)2j−b−ε,κ(b,j)=a(b,j∸1)−3(2j−1)ε.\begin{aligned} \beta_i&=\begin{cases}2,&i<w_B,\\B,&i\ge w_B,\end{cases}& g_{\mathrm{blk}}(b)&=\frac{b-1}{4}-\frac1{20},\\ d_i&=\frac{11}{5}+\sum_{j=0}^{i-1}g_{\mathrm{blk}}(\beta_j),& \varepsilon&=\frac{1}{2^{4B+100}(D+1)},\\ a(b,j)&=(1+\varepsilon)2^{j-b}-\varepsilon,& \kappa(b,j)&=a(b,j\dotminus1)-3(2^j-1)\varepsilon. \end{aligned}(13)

The scale exponent product Γi\Gamma_i and backwards target scale NiN_i are given in (14). The zero-based nesting index jj uses window exponent κ(b,j+1)\kappa(b,j+1). The integer window size and extended block windows at target scale NN and dimension bb are TwinT_{\mathrm{win}} and TT in (15). Floors and ceilings with a subscript ++ are truncated below at zero. Γi=∏j=iD−1a(βj,0),Ni=nΓi.\Gamma_i=\prod_{j=i}^{D-1}a(\beta_j,0),\qquad N_i=n^{\Gamma_i}.(14) Twin(N,b,j)=4⌊Nκ(b,j+1)/4⌋+,Tj={Twin(N,b,j),j<b,24,j≥b.T_{\mathrm{win}}(N,b,j) =4\left\lfloor N^{\kappa(b,j+1)}/4\right\rfloor_+, \qquad T_j=\begin{cases}T_{\mathrm{win}}(N,b,j),&j<b,\\24,&j\ge b. \end{cases}(15)

The rank Schur constant cRc_R, primitive period constant cPc_P, initial overlap threshold ω0\omega_0 and overlap multiplier CC are the constants in (16), evaluated at the working rate δ\delta. cR=88473621−exp⁡(−1/1024)exp⁡(1),cP=36cR2,ω0=exp⁡[−(32+3cP)Λ]1536,C=240δω0.\begin{aligned} c_R&=884736\frac{2}{1-\exp(-1/1024)}\exp(1),&c_P&=36c_R^2,\\ \omega_0&=\frac{\exp[-(32+3c_P)\Lambda]}{1536},& C&=\frac{2^{40}}{\delta\omega_0}. \end{aligned}(16)

A block arithmetic state records its length, overlap count, copy count and trial count.✦ For initial length ℓ\ell and overlap count kk, its initial state is U0=(ℓ,k,1,0)U_0=(\ell,k,1,0). Write ℓj,kj\ell_j,k_j for the first two coordinates of UjU_j. Let JjJ_j be the source root trial count at parameters (N,ℓj,Tj,kj)(N,\ell_j,T_j,k_j), and let Jloc(P,k)J_{\mathrm{loc}}(P,k) denote the source local trial count. The arithmetic transitions and block trial counts tjt_j are those in (17). Pj={Jj(2kj+1)+Tj,j+1=b,⌊Na(b,j+1)/ℓj⌋+,j+1≠b,tj={Jj,j+1=b,Jloc(Pj,kj),j+1≠b,Uj+1=(ℓjPj,⌈CkjPj⌉+,Pj,tj).\begin{aligned} P_j&=\begin{cases} J_j(2k_j+1)+T_j,&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,k_j),&j+1\ne b, \end{cases}\\ U_{j+1}&=(\ell_jP_j,\lceil Ck_jP_j\rceil_+,P_j,t_j). \end{aligned}(17)

A concrete level pair at rate ρ\rho, scale XX, allowance Δ\Delta, decay exponent dd, overlap threshold ω\omega, count multiplier K*K_*, level ii and half moment order kk consists of distinct words x,yx,y of a common length ℓ∈[X/2,X]\ell\in[X/2,X], finite laws μ,ν\mu,\nu with a common decoder ff, a nonnegative simulation error EE, a likelihood score uu and a repetition count mm satisfying (18).✦ Its even score exponent is ri=2kr_i=2k; the squared L2k(μ)L^{2k}(\mu) moment norm is probability-weighted, with the order-zero value defined to be one. 𝒟ρ(μ,f;x)≤E,𝒟ρ(ν,f;y)≤E,E/ℓ≤(i+1)Δ,ν(z)=μ(z)(1+u(z)),ℳk(μ,u)≤X−d,ℳk(μ,u)={(𝔼μ|u|2k)1/k,k>0,1,k=0,ℓ≤m≤K*ℓ,ov⁡(Tr⁡ρ(xm),Tr⁡ρ(xm+1))≥ω.\begin{aligned} &\mathcal D_{\rho}(\mu,f;x)\le E,\qquad \mathcal D_{\rho}(\nu,f;y)\le E,\qquad E/\ell\le(i+1)\Delta,\\ &\nu(z)=\mu(z)(1+u(z)),\qquad \mathcal M_k(\mu,u)\le X^{-d},\\ &\mathcal M_k(\mu,u)=\begin{cases} (\mathbb E_{\mu}|u|^{2k})^{1/k},&k>0,\\ 1,&k=0,\end{cases}\\ &\ell\le m\le K_*\ell,\qquad \operatorname{ov}\bigl(\operatorname{Tr}_{\rho}(x^m), \operatorname{Tr}_{\rho}(x^{m+1})\bigr)\ge\omega. \end{aligned}(18)

A source level pair at index ii is a concrete level pair at rate δ\delta with X=NiX=N_i, d=did=d_i, ω=ω0\omega=\omega_0, and the choices in (19).✦ Δn=n−100(dD+B+1),K*=C2Bi+2,ki=⌈108210B(dD+B+1)log⁡nlog⁡Ni⌉+,ri=2ki.\Delta_n=n^{-100(d_D+B+1)},\qquad K_*=C^{2Bi+2},\qquad k_i=\left\lceil 10^8\,2^{10B}(d_D+B+1)\frac{\log n}{\log N_i} \right\rceil_+,\qquad r_i=2k_i.(19)

The degree cutoff RR, work size VV, cancellation order ss, moment envelope MM, weight envelope WW and cap envelope KK are defined by (20). 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)+2/ω0.\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_0. \end{aligned}(20)

A finite ordered role shape of depth vv and branching bound rr is an ordered rooted tree with at most rr children at each vertex and one role in {0,1,2}\{0,1,2\} on each child; depth zero has one empty shape.✦ Let S(r,v)S(r,v) be the number of these shapes.✦ The bad role factor FbadF_{\mathrm{bad}} and forced role factor FforceF_{\mathrm{force}} are the sums in (21). Fbad(v,R)=∑r=0R−1432(v+1)r,Fforce(v,R)=∑r=0R−143(v+1)rS(r,v+1).F_{\mathrm{bad}}(v,R)=\sum_{r=0}^{R-1}4\,3^{2(v+1)r},\qquad F_{\mathrm{force}}(v,R)=\sum_{r=0}^{R-1}4\,3^{(v+1)r}S(r,v+1).(21)

For level costs cvc_v, the binary tree budget ℬc\mathcal B_c is the sequence ℬc(0)=1\mathcal B_c(0)=1 and ℬc(v+1)=cvℬc(v)2\mathcal B_c(v+1)=c_v\mathcal B_c(v)^2.✦ The smoothing and skeleton prefactors are PsmP_{\mathrm{sm}} and PskP_{\mathrm{sk}} in (22). Psm(s,K)=2s2+s((s+1)K)s+1,Psk(s,v,K)=223sKℬc(v),cj=426sK2.\begin{aligned} P_{\mathrm{sm}}(s,K)&=2^{s^2+s}((s+1)K)^{s+1},\\ P_{\mathrm{sk}}(s,v,K)&=2\,2^{3s}K\,\mathcal B_c(v), \qquad c_j=4\,2^{6s}K^2. \end{aligned}(22)

The numerical schedule margins at scale NN are the inequalities in (23), with the last two required for every 0≤v≤B0\le v\le B.✦ The tail denominator 𝒯\mathcal T and tail exponent ataila_{\mathrm{tail}} are given by (24). max⁡{1200K,482s,8W22s,Psm(s,K),1536,64(K+1)(D+1),240(K+1)(W+1)22s(B+2)}≤Nε,384Fbad(v,R)≤Nε,192Fforce(v,R)W2v+1Psk(s,v,K)≤Nε.\begin{aligned} \max\{&1200K,\ 48\,2^s,\ 8W\,2^{2s},\ P_{\mathrm{sm}}(s,K),\ 1536,\\ &64(K+1)(D+1),\ 2^{40}(K+1)(W+1)^2\,2^{s(B+2)}\} \le N^{\varepsilon},\\ 384F_{\mathrm{bad}}(v,R)&\le N^{\varepsilon},\\ 192F_{\mathrm{force}}(v,R)W^{2^{v+1}}P_{\mathrm{sk}}(s,v,K) &\le N^{\varepsilon}. \end{aligned}(23) 𝒯=512+9600Kω0−4,atail=300(dD+B+1)+40V+5.\mathcal T=512+9600K\omega_0^{-4},\qquad a_{\mathrm{tail}}=300(d_D+B+1)+40V+5.(24)

The Claim

Theorem 1. Fix ρ∈[0,1]\rho\in[0,1], a depth b+1b+1, and the word construction above. Assume the layout margins ℳb\mathscr M_b, the identity Pb=tb(2hb+1)+TbP_b=t_b(2h_b+1)+T_b, an even TbT_b with 4(tb+1)≤Tb4(t_b+1)\le T_b and 8⋅2s≤Tb8\cdot2^s\le T_b, and tj>0t_j>0 for every j<bj<b. Let paired mechanisms have mb+1m_{b+1} slots, and let RjR_j be finite filler laws for j<kb+1j<k_{b+1}. Suppose ff exactly simulates the trial words Vb+1(j,0)V_{b+1}(j,0) from FjF_j, the trial words Vb+1(j,1)V_{b+1}(j,1) from GjG_j, and the filler words Zb+1(j)Z_{b+1}(j) from RjR_j, all at rate ρ\rho. Suppose gg simulates a0,a1a_0,a_1 from 𝖠0,𝖠1\mathsf A_0,\mathsf A_1, respectively, with errors at most E≥0E\ge0. For either root orientation τ\tau, the canonical simulation satisfies (25). 𝒟ρ(𝒪b+1,τ⊗⨂j=0kb+1−1Rj,Kb+1;Wb+1τ)≤Lb+1E+Fb+1.\mathcal D_{\rho}\left( \mathcal O_{b+1,\tau}\otimes\bigotimes_{j=0}^{k_{b+1}-1}R_j, K_{b+1};W_{b+1}^{\tau}\right) \le L_{b+1}E+F_{b+1}.(25)✦

Let B≥2B\ge2, i<Di<D, n≥4n\ge4 and 0<δ≤1/20<\delta\le1/2, and let UU be a source level pair at index ii, with left word length ℓ\ell and overlap count kk. Set N=Ni+1N=N_{i+1}, b=βib=\beta_i, and form Tj,tjT_j,t_j from (15) and (17), starting from (ℓ,k)(\ell,k). Assume the numerical schedule margins (23) at scale NN and 2Bexp⁡(−Nε/𝒯)≤n−atail2B\exp(-N^{\varepsilon}/\mathcal T)\le n^{-a_{\mathrm{tail}}}. The level tail budget is (26). max⁡{Hb(ω0),FbN−20V,Fb}≤n−200(dD+B+1).\max\left\{H_b(\omega_0),\ \frac{F_b}{N^{-20V}},\ F_b\right\} \le n^{-200(d_D+B+1)}.(26)✦

In the paper, the result reads: for the unconditioned surrogate laws μ̃i,ν̃i\widetilde\mu_i,\widetilde\nu_i, their common projection KtreeK_{\mathrm{tree}}, root words Ub+,Ub−U_b^+,U_b^-, incoming leaf error Ei−1E_{i-1}, allowance Δ\Delta and Tr⁡=Tr⁡δ\operatorname{Tr}=\operatorname{Tr}_{\delta}, the distances TV⁡((Ktree)#μ̃i,Tr⁡(Ub+))\operatorname{TV}((K_{\mathrm{tree}})_{\#}\widetilde\mu_i, \operatorname{Tr}(U_b^+)) and TV⁡((Ktree)#ν̃i,Tr⁡(Ub−))\operatorname{TV}((K_{\mathrm{tree}})_{\#}\widetilde\nu_i, \operatorname{Tr}(U_b^-)) are each at most o(Δ)+(∏j=0b−1Tj)Ei−1o(\Delta)+(\prod_{j=0}^{b-1}T_j)E_{i-1}, uniformly in block level ii.

What Is Checked

The simulation is stated over finite side and leaf outcome spaces, paired mechanisms with common-submass domination, and the complete window tree with its layout margins; the tail budget is stated over a source level pair and the numerical schedule margins at its next target scale.

Deletion mask weights sum to one for every real parameter✦ and are nonnegative on [0,1][0,1].✦ For any finite laws F,G,HF,G,H and any real common weight, the paired report masses sum to one. With a nonnegative common weight and common-submass domination, they are also nonnegative.✦ The public chronology contains every emission exactly once, and the rectangular global trial indexing is bijective.✦