OctoInformBeta

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

Lemma 5.2: Supremum and squared counting-measure estimates, with integer translations and a posterior cost for hidden bits

Prouhet convolutions admit binomial smoothing bounds.

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 one amplification block, the smoothing quantity is the convolution of a signed Prouhet motif with a fair binomial mass. Partial observations introduce a posterior smoothing cost determined by the number of bits that remain hidden.

Let q∈(0,1)q\in(0,1) be the requested deletion probability, so each bit survives with probability 1−q1-q. Write δ=min⁡{q,1/2}\delta=\min\{q,1/2\} for the working deletion rate and Λ=δ−3\Lambda=\delta^{-3}; the construction theorems call their working rate qq.

A real Laurent polynomial FF is a finite sum F(z)=∑j∈ℤfjzjF(z)=\sum_{j\in\mathbb Z}f_jz^j with real coefficients, with multiplication determined by zjzk=zj+kz^jz^k=z^{j+k}.✦ Multiplication convolves coefficient sequences. The coefficient norms and the integer translation TcFT_cF, for c∈ℤc\in\mathbb Z, are defined by ∥F∥∞=supj∈ℤ|fj|,∥F∥22=∑j∈ℤfj2,(TcF)(z)=zcF(z).\|F\|_\infty=\sup_{j\in\mathbb Z}|f_j|, \qquad \|F\|_2^2=\sum_{j\in\mathbb Z}f_j^2, \qquad (T_cF)(z)=z^cF(z).(1) The squared ℓ2\ell^2 norm in (1) uses counting measure.

The fair binomial polynomial BnB_n with nn trials and the Prouhet polynomial PsP_s of order ss, for nonnegative integers n,sn,s, are Bn(z)=(1+z2)n,Ps(z)=∏v=0s−1(1−z2v).B_n(z)=\left(\frac{1+z}{2}\right)^n, \qquad P_s(z)=\prod_{v=0}^{s-1}(1-z^{2^v}).(2) The coefficients of BnB_n form the fair binomial mass function, and the signed Prouhet motif is the coefficient sequence σs(j)=[zj]Ps(z)\sigma_s(j)=[z^j]P_s(z); the product index vv starts at zero. The binomial peak bound βn\beta_n is the sequence βn=2n+1,n≥0.\beta_n=\sqrt{\frac{2}{n+1}},\qquad n\geq0.(3)✦

For mm paired trials, a partial observation d=(d0,…,dm−1)d=(d_0,\ldots,d_{m-1}) is a sequence with entries in {0,1,?}\{0,1,?\}, where ?? denotes a hidden bit. For a symbol a∈{0,1,?}a\in\{0,1,?\} and a partial observation dd of mm paired trials, the hidden trial indicator χ(a)\chi(a) and the hidden count N(d)N(d) are the nonnegative integers given by (4). χ(a)={1,a=?,0,a∈{0,1},N(d)=∑i=0m−1χ(di).\chi(a)= \begin{cases} 1,&a=?,\\ 0,&a\in\{0,1\}, \end{cases} \qquad N(d)=\sum_{i=0}^{m-1}\chi(d_i).(4) The observation index ii starts at zero. For a partial observation dd of mm paired trials, a nonnegative integer ss, and a real weight ω\omega, the posterior smoothing cost Cs(d)C_s(d) and the hidden smoothing bound ℋm,s(ω)\mathcal H_{m,s}(\omega) are the nonnegative real quantities given by (5). Cs(d)=2s2(2β⌊N(d)/(s+1)⌋)sβ⌊N(d)/(s+1)⌋,hm,s(ω)=4(s+1)mω2+2,ℋm,s(ω)=2s2(2hm,s(ω))shm,s(ω).\begin{aligned} C_s(d) &=2^{s^2} \bigl(2\beta_{\lfloor N(d)/(s+1)\rfloor}\bigr)^s \beta_{\lfloor N(d)/(s+1)\rfloor},\\ h_{m,s}(\omega) &=\sqrt{\frac{4(s+1)}{m\omega^2+2}},\\ \mathcal H_{m,s}(\omega) &=2^{s^2}\bigl(2h_{m,s}(\omega)\bigr)^s h_{m,s}(\omega). \end{aligned}(5)

The Claim

Theorem 1. Let g,sg,s be nonnegative integers, and put n=⌊g/(s+1)⌋n=\lfloor g/(s+1)\rfloor. For the polynomials in (2) and the peak bound in (3), signed binomial smoothing gives (6): |[zj](PsBg)|≤2s2(2βn)sβnfor every j∈ℤ,∥PsBg∥22≤22s2(2βn)2sβn.\begin{aligned} |[z^j](P_sB_g)| &\leq 2^{s^2}(2\beta_n)^s\beta_n &&\text{for every }j\in\mathbb Z,\\ \|P_sB_g\|_2^2 &\leq 2^{2s^2}(2\beta_n)^{2s}\beta_n. \end{aligned}(6)✦ For every integer cc, the translated squared norm satisfies (7): ∥Ps(TcBg)∥22≤22s2(2βn)2sβn.\|P_s(T_cB_g)\|_2^2 \leq 2^{2s^2}(2\beta_n)^{2s}\beta_n.(7)✦ Let dd be a partial observation of mm paired trials and let ω∈ℝ\omega\in\mathbb R satisfy N(d)≥mω2/2N(d)\geq m\omega^2/2. The hidden smoothing estimate is (8): Cs(d)≤ℋm,s(ω).C_s(d)\leq\mathcal H_{m,s}(\omega).(8)✦

In the paper, the result reads: for the stated Prouhet motif σs*\sigma_{s_*} and an integer translate or reflection π\pi of a fair binomial mass with g≥c0ω2mg\geq c_0\omega^2m trials, where c0>0c_0>0 is absolute, ∥σs**π∥∞≤exp⁡(O((s*+1)2+(s*+1)Λ))m−(s*+1)/2\|\sigma_{s_*}*\pi\|_\infty \leq\exp\!\bigl(O((s_*+1)^2+(s_*+1)\Lambda)\bigr)m^{-(s_*+1)/2}, ∥σs**π∥22≤exp⁡(O((s*+1)2+(s*+1)Λ))m−s*−1/2\|\sigma_{s_*}*\pi\|_2^2 \leq\exp\!\bigl(O((s_*+1)^2+(s_*+1)\Lambda)\bigr)m^{-s_*-1/2}, and ∥π∥∞=O(ω−1m−1/2)\|\pi\|_\infty=O(\omega^{-1}m^{-1/2}) as mm tends to infinity in the construction, with counting measure and constants allowed to depend on c0c_0.

What Is Checked

The claims are stated over finite real Laurent polynomials with the norms and translations of (1), and over partial observations with the hidden indicators of (4) and the costs of (5), subject to the stated lower bound on the hidden count for the posterior estimate.