OctoInformBeta

OpenAI preprint, September 2026 · §4. An initial pair at the working deletion rate

§4: Periodic defects, counted primitive marks, and rank kernels: Exact binomial probabilities and centered differences for uniformly selected slots

Periodic defects through their retained ranks.

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 periodic defect replaces an ordinary binary primitive 0101 by 11. Conditioning on the number of retained primitive slots gives a rank experiment: specified defect slots must occupy specified positions in the retained sequence. Slots, observed ranks, and primitive block positions are numbered from zero.

Let q∈(0,1)q\in(0,1) be the requested deletion probability, so that each bit survives with probability 1−q1-q. The working deletion rate δ\delta and its associated scale Λ\Lambda are the quantities specified in (1); the construction theorems call this working rate qq. δ=min⁡{q,1/2},Λ=δ−3.\delta=\min\{q,1/2\},\qquad \Lambda=\delta^{-3}.(1)

The ordinary run OaO_a, for a∈ℕa\in\mathbb N, is the binary word consisting of aa consecutive ordinary primitives, as in (2). Oa=(01)a.O_a=(01)^a.(2)✦

Write [n]={0,…,n−1}[n]=\{0,\ldots,n-1\}. A finite probability law μ\mu on a finite set AA is a function μ:A→ℝ\mu:A\to\mathbb R with nonnegative values and ∑x∈Aμ(x)=1\sum_{x\in A}\mu(x)=1.✦ For ℓ≤n\ell\le n, a fixed-size subset SS of [n][n] is a subset with exactly ℓ\ell elements.✦ The uniform subset law 𝖴n,ℓ\mathsf U_{n,\ell} assigns the mass in (3) to each such subset. 𝖴n,ℓ(S)=(nℓ)−1,S⊆[n],|S|=ℓ.\mathsf U_{n,\ell}(S)=\binom n\ell^{-1}, \qquad S\subseteq[n],\quad |S|=\ell.(3)✦

For a fixed-size subset SS, the slot at observed rank v∈[ℓ]v\in[\ell], denoted sS(v)s_S(v), is its vvth element in increasing order.✦ An increasing tuple in [n][n] is a tuple p=(p0,…,pr−1)p=(p_0,\ldots,p_{r-1}) with strictly increasing coordinates.✦ For tags p∈[n]rp\in[n]^r and ranks v∈[ℓ]rv\in[\ell]^r, the tagged rank event E(p,v)E(p,v) is the set of subsets specified in (4). E(p,v)={S⊆[n]:|S|=ℓ,sS(vh)=ph for every h∈[r]}.E(p,v)=\{S\subseteq[n]: |S|=\ell,\ s_S(v_h)=p_h \text{ for every }h\in[r]\}.(4)✦

The lower and upper bounds Lj(p)L_j(p) and Uj(p)U_j(p) of the jjth open tuple gap are the slot boundaries in (5); its size is Gj(p)G_j(p). For a rank tuple, the same definitions use ambient size ℓ\ell. Lj(p)={0j=0,pj−1+11≤j≤r,Uj(p)={pj0≤j<r,nj=r,Gj(p)=(Uj(p)−Lj(p))+,0≤j≤r,x+=max⁡{x,0}.\begin{aligned} L_j(p)&=\begin{cases}0&j=0,\\p_{j-1}+1&1\le j\le r,\end{cases} &U_j(p)&=\begin{cases}p_j&0\le j<r,\\n&j=r,\end{cases}\\ G_j(p)&=\bigl(U_j(p)-L_j(p)\bigr)_+, &&0\le j\le r,\qquad x_+=\max\{x,0\}. \end{aligned}(5)

The binomial mass Ba,θ(b)B_{a,\theta}(b), for a,b∈ℕa,b\in\mathbb N and θ∈[0,1]\theta\in[0,1], is the function in (6), with (ab)=0\binom ab=0 when b>ab>a. Ba,θ(b)=(ab)θb(1−θ)(a−b)+.B_{a,\theta}(b)=\binom ab\theta^b(1-\theta)^{(a-b)_+}.(6)✦

Positional Differences

A safe tagged tuple p∈[n]rp\in[n]^r is a tuple of interior tags separated by at least three slots, with the precise conditions in (7). 1≤ph,ph+1<n(h∈[r]),pi+3≤pj(0≤i<j<r).1\le p_h,\quad p_h+1<n\quad(h\in[r]),\qquad p_i+3\le p_j\quad(0\le i<j<r).(7)✦ A positional difference order κ∈{0,1,2}\kappa\in\{0,1,2\} specifies the unchanged, first, or second difference, respectively.✦ The positional difference coefficients cκ(a)c_\kappa(a), for a∈{0,1,2}a\in\{0,1,2\}, are given in (8). cκ(a)a=0a=1a=2κ=0010κ=1−1/201/2κ=21/2−11/2\begin{array}{c|rrr} c_\kappa(a)&a=0&a=1&a=2\\ \hline \kappa=0&0&1&0\\ \kappa=1&-1/2&0&1/2\\ \kappa=2&1/2&-1&1/2 \end{array}(8)✦ For a safe tagged tuple pp and a displacement choice a∈{0,1,2}ra\in\{0,1,2\}^r, the displaced tag tuple pap^a is defined by (9). pha=ph+ah−1(h∈[r]).p^a_h=p_h+a_h-1\qquad(h\in[r]).(9)✦

The Claim

Definition 1 (Periodic defects and rank differences). For m,d∈ℕm,d\in\mathbb N, the single defect period Im,dI_{m,d} is the binary word in (10). Im,d=Od1O(m−1−d)+.I_{m,d}=O_d\,1\,O_{(m-1-d)_+}.(10)✦ For ℓ≤n\ell\le n, tags p∈[n]rp\in[n]^r, and ranks v∈[ℓ]rv\in[\ell]^r, the conditional rank kernel Wn,ℓ(p;v)W_{n,\ell}(p;v) is the probability in (11). Wn,ℓ(p;v)=𝖴n,ℓ(E(p,v)).W_{n,\ell}(p;v)=\mathsf U_{n,\ell}\bigl(E(p,v)\bigr).(11)✦ If pp and vv are strictly increasing and 0<θ<10<\theta<1, this kernel has the binomial product representation (12). Wn,ℓ(p;v)=θrBn,θ(ℓ)∏j=0rBGj(p),θ(Gj(v)).W_{n,\ell}(p;v)= \frac{\theta^r}{B_{n,\theta}(\ell)} \prod_{j=0}^{r} B_{G_j(p),\theta}\bigl(G_j(v)\bigr).(12)✦ For a safe tagged tuple pp, positional orders κ∈{0,1,2}r\kappa\in\{0,1,2\}^r, and an increasing rank tuple vv, the positional rank difference ΔκWn,ℓ(p;v)\Delta^\kappa W_{n,\ell}(p;v) is defined by (13). ΔκWn,ℓ(p;v)=∑a∈{0,1,2}r(∏h=0r−1cκh(ah))Wn,ℓ(pa;v).\Delta^\kappa W_{n,\ell}(p;v)= \sum_{a\in\{0,1,2\}^r} \left(\prod_{h=0}^{r-1}c_{\kappa_h}(a_h)\right) W_{n,\ell}(p^a;v).(13)✦

In the paper, with ambient size parameter NN and uu periods, the result reads: η=.004\eta=.004, m=⌊N1/2−η⌋m=\lfloor N^{1/2-\eta}\rfloor, K=⌊N1/2−2η⌋K=\lfloor N^{1/2-2\eta}\rfloor, and k=⌊δm⌋k=\lfloor\delta m\rfloor, with the remaining side count filling the word to length in [N/2,N][N/2,N]; a period has mm primitive blocks with 11 replacing 0101 at position ⌊m/2⌋+b\lfloor m/2\rfloor+b, where b∈{−1,0,1}b\in\{-1,0,1\}; independently at each primitive, a common empty-output mass δ2/2\delta^2/2 is silently erased, and otherwise a mark in {∅,0,1,01}\{\emptyset,0,1,01\} is returned, with the returned empty mark distinct from silence; conditional on ℓ\ell returned marks among umum slots, the retained subset is uniform, for an increasing tuple jj of defect periods, Wj(v)W_j(v) is the probability that their defect slots occur at specified increasing ranks, and Ar(v)A_r(v) sums Wj(v)W_j(v) over increasing rr-tuples of defect periods; a count is admissible when θ=ℓ/(um)∈[1−δ2,1−δ2/4]\theta=\ell/(um)\in[1-\delta^2,1-\delta^2/4]; coarse gaps γ\gamma have half-integer endpoints and integer internal gaps; with low-degree cutoff R0=3000R_0=3000, the operators Dh=(Th−Th−1)/2D_h=(T_h-T_h^{-1})/2 and Vh=(Th+Th−1−2)/2V_h=(T_h+T_h^{-1}-2)/2 use the shift ThT_h moving the hhth tag one slot, and λ=K/m\lambda=\sqrt{K/m}.

What Is Checked

The definitions are stated over finite binary words, fixed-size subsets with ℓ≤n\ell\le n, and tagged slot and rank tuples, with interior separation imposed for positional differences.

For ℓ≤n\ell\le n, the binomial coefficient (nℓ)\binom n\ell is positive, and the uniform subset law 𝖴n,ℓ\mathsf U_{n,\ell} has nonnegative masses summing to one.