OctoInformBeta

OpenAI preprint, September 2026 · Section 9. Exceptional terms and the power-saving induction

Theorem 9.2: The set functional decays uniformly under prime hypotheses and bounds a fixed power of density

Uniform power decay for the set functional.

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 set BB of integers is square-difference-free if it contains no pair a,ba,b with b=a+t2b=a+t^2 for a positive integer tt. The set functional considered here is a weighted diagonal moment of a residue lift of its indicator. The public interval is [N]={1,…,N}[N]=\{1,\ldots,N\}; a set B⊆[N]B\subseteq[N] is represented in the natural range {0,…,N−1}\{0,\ldots,N-1\} by A={b−1:b∈B}A=\{b-1:b\in B\}. Write αN(A)=|A|/N\alpha_N(A)=|A|/N, with α0(⌀)=0\alpha_0(\varnothing)=0.

The Tuple Laws

The fixed height hh, block cutoff TT, block count mm, list base rr, and interaction parameter δ\delta have the values in (1). h=24,T=80,m=2T+1=161,r=h!2,δ=110h2.h=24,\qquad T=80,\qquad m=2T+1=161,\qquad r=\frac{h!}{2},\qquad \delta=\frac{1}{10h^2}.(1) A permutation word vv of height kk on a block set CC is a function assigning a permutation of {0,…,k−1}\{0,\ldots,k-1\} to each block.✦ For the block set ℬ={0,…,m−1}\mathcal B=\{0,\ldots,m-1\}, the permutation tuple vertices form the set 𝒱\mathcal V in (2); dd is its cardinality. 𝒲(C,k)=Sym⁡({0,…,k−1})C,𝒱=𝒲(ℬ,h),d=(h!)m.\mathcal W(C,k)=\operatorname{Sym}(\{0,\ldots,k-1\})^C, \qquad \mathcal V=\mathcal W(\mathcal B,h),\qquad d=(h!)^m.(2)

An ordered position pair (i,j)(i,j) at height kk is a pair of integers satisfying 0≤i<j<k0\le i<j<k.✦ For a permutation word vv of height kk on CC, a block b∈Cb\in C, and an ordered position pair (i,j)(i,j) at height kk, the position transposition vb;i,jv^{b;i,j} is the word whose permutation at block bb is v(b)(ij)v(b)(i\ j) and whose other permutations agree with vv.✦ When CC is finite, for R⊆CR\subseteq C and a real parameter tt, the transposition interaction HR,tH_{R,t} is the matrix in (3). HR,t(v,w)=∑b∈R∑0≤i<j<ktj−i𝟏{w=vb;i,j}.H_{R,t}(v,w)=\sum_{b\in R}\ \sum_{0\le i<j<k} t^{j-i}\mathbf1_{\{w=v^{b;i,j}\}}.(3)✦

The cyclic shift cc is the permutation of the hh positions that sends each position to the next, with the last returning to zero.✦ For a finite block set CC, functions u,v:C→Gu,v:C\to G into a group GG, a shift g∈Gg\in G, and an integer threshold t≥0t\ge0, threshold shift adjacency is the relation between uu and vv requiring that at least tt blocks satisfy v(b)=u(b)gv(b)=u(b)g.✦

For a real matrix HH on a finite set II and labelings ℓ1,ℓ2:I→K\ell_1,\ell_2:I\to K into a label set KK, the label coincidence pairing PHP_H, index energy EHE_H, and index Boltzmann weight WHW_H are defined by (4). PH(ℓ1,ℓ2)=∑x∈I∑y∈IHx,y𝟏{ℓ1(x)=ℓ2(y)},EH(ℓ)=12PH(ℓ,ℓ),WH(ℓ)=exp⁡(EH(ℓ)).\begin{aligned} P_H(\ell_1,\ell_2)&=\sum_{x\in I}\sum_{y\in I} H_{x,y}\mathbf1_{\{\ell_1(x)=\ell_2(y)\}},\\ E_H(\ell)&=\tfrac12P_H(\ell,\ell),\qquad W_H(\ell)=\exp(E_H(\ell)). \end{aligned}(4)

For an active block set R⊆ℬR\subseteq\mathcal B, an inactive-block address aa is a function assigning a permutation of the hh positions to each block outside RR.✦ The addressed list index set IRI_R, addressed list entry set LRL_R, and inactive-block projection πR\pi_R are given in (5). 𝒜R=Sym⁡({0,…,h−1})ℬ\R,IR={0,…,r|R|−1},LR=𝒜R×IR,πR(v)=v|ℬ\R.\begin{aligned} \mathcal A_R&=\operatorname{Sym}(\{0,\ldots,h-1\})^{\mathcal B\setminus R}, &I_R&=\{0,\ldots,r^{|R|}-1\},\\ L_R&=\mathcal A_R\times I_R, &\pi_R(v)&=v|_{\mathcal B\setminus R}. \end{aligned}(5)

Let ER⊆LR×LRE_R\subseteq L_R\times L_R be the list constraint edges, the prescribed finite set of ordered pairs specifying the square-difference constraints between list values. For a field FF, the strict square indicator sFs_F is one at nonzero squares and zero elsewhere.✦ For a prime pp, list values Y:LR→𝔽pY:L_R\to\mathbb F_p, and a labeling ℓ:𝒱→IR\ell:\mathcal V\to I_R, the square-constraint indicator VRV_R, tuple Boltzmann weight WRW_R, and sampled tuple configuration ZRZ_R occur in the tuple component functional Cp,RC_{p,R} of (6). Both expectations are uniform probability averages over their entire finite spaces, and FF is a real observable on 𝔽p𝒱\mathbb F_p^{\mathcal V}. VR(Y)=∏(a,b)∈ERs𝔽p(Y(b)−Y(a)),WR(ℓ)=WHR,δ(ℓ),ZR(Y,ℓ)(v)=Y(πR(v),ℓ(v)),Cp,R(F)=𝔼Y:LR→𝔽p𝔼ℓ:𝒱→IRVR(Y)WR(ℓ)F(ZR(Y,ℓ)).\begin{aligned} V_R(Y)&=\prod_{(a,b)\in E_R}s_{\mathbb F_p}(Y(b)-Y(a)), &W_R(\ell)&=W_{H_{R,\delta}}(\ell),\\ Z_R(Y,\ell)(v)&=Y(\pi_R(v),\ell(v)),\\ C_{p,R}(F)&=\mathbb E_{Y:L_R\to\mathbb F_p} \mathbb E_{\ell:\mathcal V\to I_R} V_R(Y)W_R(\ell)F(Z_R(Y,\ell)). \end{aligned}(6)

The block subset mixture weight is wp(R)=p−|R|/16w_p(R)=p^{-|R|/16}.✦ The unnormalized tuple mixture ℳp\mathcal M_p and the restricted tuple mixture ℳp,𝒟\mathcal M_{p,\mathcal D} for a finite family 𝒟\mathcal D of block subsets are the counting sums in (7). ℳp(F)=∑R⊆ℬ|R|≤T+1wp(R)Cp,R(F),ℳp,𝒟(F)=∑R∈𝒟wp(R)Cp,R(F).\mathcal M_p(F)=\sum_{\substack{R\subseteq\mathcal B\\|R|\le T+1}} w_p(R)C_{p,R}(F),\qquad \mathcal M_{p,\mathcal D}(F)=\sum_{R\in\mathcal D}w_p(R)C_{p,R}(F).(7)

The terminal family is 𝒟*={R⊆ℬ:|R|=T+1}\mathcal D_* =\{R\subseteq\mathcal B:|R|=T+1\}.✦ The full tuple law μp\mu_p and exceptional tuple sublaw εp\varepsilon_p are normalized by the same total mixture mass; the exceptional mass ηp\eta_p is defined with them in (8). μp(F)=ℳp(F)ℳp(1),εp(F)=ℳp,𝒟*(F)ℳp(1),ηp=εp(1).\mu_p(F)=\frac{\mathcal M_p(F)}{\mathcal M_p(1)},\qquad \varepsilon_p(F)=\frac{\mathcal M_{p,\mathcal D_*}(F)}{\mathcal M_p(1)},\qquad \eta_p=\varepsilon_p(1).(8)

Thresholds and Admissible Primes

For a bijection θ:I→W\theta:I\to W and a finite directed edge set EE on WW, the graph pullback θ*E\theta^*E consists of the pairs (x,y)(x,y) with (θ(x),θ(y))∈E(\theta(x),\theta(y))\in E.✦ The crossing edges of a graph on a disjoint union I⊔KI\sqcup K are the pairs (i,k)∈I×K(i,k)\in I\times K joined in either orientation.✦ The vertex index decomposition at v∈Wv\in W is the bijection σv:{*}⊔(W\{v})→W\sigma_v:\{*\}\sqcup(W\setminus\{v\})\to W sending ** to vv and each other vertex to itself.✦ Write kE(v)k_E(v) for the number of crossing edges of σv*E\sigma_v^*E. The graph conditional energy constant is (9). CE(v)=(2(1/2)kE(v)+4((1/2)kE(v))2)((32kE(v))1/4)2.C_E(v)=\left(\frac{2}{(1/2)^{k_E(v)}}+ \frac{4}{((1/2)^{k_E(v)})^2}\right) \left((32k_E(v))^{1/4}\right)^2.(9)✦

The empty-block list index is zero.✦ The leading list identification ι:𝒱→L⌀\iota:\mathcal V\to L_{\varnothing} sends vv to (π⌀(v),0)(\pi_{\varnothing}(v),0).✦ The leading constraint graph is E0=ι*E⌀E_0=\iota^*E_{\varnothing}✦, and its edge count is e=|E⌀|e=|E_{\varnothing}|.✦ The leading mass lower bound b0b_0 and mass positivity threshold PmassP_{\mathrm{mass}} are defined in (10). b0=2−(e+1),Pmass=max⁡{3,(2e+1e)2}.b_0=2^{-(e+1)},\qquad P_{\mathrm{mass}}=\max\{3,(2^{e+1}e)^2\}.(10) The Boltzmann weight upper bound WmaxW_{\max} determines the conditional error numerator BcondB_{\mathrm{cond}}. The conditional decay constant CcondC_{\mathrm{cond}} and conditional decay threshold PcondP_{\mathrm{cond}} are defined with them in (11). Wmax=exp⁡(d(T+1)(h2)δ2),Bcond=∑v∈𝒱CE0(v)+2mWmax,Ccond=Bcond/b0,Pcond=max⁡{Pmass,3,Ccond32}.\begin{aligned} W_{\max}&=\exp\left(\frac{d(T+1)\binom h2\delta}{2}\right),\\ B_{\mathrm{cond}}&=\sum_{v\in\mathcal V}C_{E_0}(v)+2^mW_{\max}, &C_{\mathrm{cond}}&=B_{\mathrm{cond}}/b_0,\\ P_{\mathrm{cond}}&=\max\{P_{\mathrm{mass}},3,C_{\mathrm{cond}}^{32}\}. \end{aligned}(11) Put K=d/2K=d/2, the half vertex count.✦ The interaction lower bound κ\kappa determines the marked reflection constant CmarkedC_{\mathrm{marked}}. The unmarked reflection constant CunmarkedC_{\mathrm{unmarked}} and reflection positivity threshold PreflP_{\mathrm{refl}} are defined in (12). κ=910δh−1,Cmarked=(exp⁡(κ)−1)K(rT+1)K,Cunmarked=(32d2)1/4Wmax⁡2,Prefl=max⁡{3,(CunmarkedrKCmarked)16}.\begin{aligned} \kappa&=\tfrac9{10}\delta^{h-1}, &C_{\mathrm{marked}}&=\frac{(\exp(\kappa)-1)^K}{(r^{T+1})^K},\\ C_{\mathrm{unmarked}}&=(32d^2)^{1/4}W_{\max}^2, &P_{\mathrm{refl}}&=\max\left\{3, \left(\frac{C_{\mathrm{unmarked}}r^K}{C_{\mathrm{marked}}}\right)^{16}\right\}. \end{aligned}(12)

Fix a finite small prime family (qs)s∈S(q_s)_{s\in S}. Its local quadratic modulus and combined small prime modulus MM are defined by (13). qsquad={8,qs=2,qs,qs≠2,M=∏s∈Sqsquad.q_s^{\mathrm{quad}}=\begin{cases}8,&q_s=2,\\q_s,&q_s\ne2,\end{cases} \qquad M=\prod_{s\in S}q_s^{\mathrm{quad}}.(13) For a finite ambient prime family p=(pj)j∈Jp=(p_j)_{j\in J}, the combined prime family on S⊔JS\sqcup J agrees with qq on SS and pp on JJ.✦ The source prime hypotheses call pp admissible at scale NN when this combined family is injective, contains every prime at most NN, and satisfies (14) for every j∈Jj\in J. pj≥64,pj≥max⁡{Pmass,Pcond,Prefl},gcd⁡(M,pj)=1,ηpj≤pj−4.p_j\ge64,\qquad p_j\ge\max\{P_{\mathrm{mass}},P_{\mathrm{cond}},P_{\mathrm{refl}}\},\qquad \gcd(M,p_j)=1,\qquad \eta_{p_j}\le p_j^{-4}.(14)✦

The Residue Lift and Set Functional

For a finite modulus family pp, its product residue space is Xp=∏j∈Jℤ/pjℤX_p=\prod_{j\in J}\mathbb Z/p_j\mathbb Z.✦ The centered residue delta for a positive integer qq is the function ρq(x)=q𝟏{x=0}−1\rho_q(x)=q\mathbf1_{\{x=0\}}-1 on ℤ/qℤ\mathbb Z/q\mathbb Z.✦ For a support U⊆JU\subseteq J, its conductor pUp_U and centered residue product GUG_U are given in (15). The conductor-bounded support family 𝒰Q\mathcal U_Q contains exactly the complete supports with conductor at most QQ.✦ For a real function ff on the natural numbers, the residue lift component Lp,N,UfL_{p,N,U}f and conductor-truncated residue lift Lp,N,QfL_{p,N,Q}f, also called the truncated lift, are the normalized counting sum and support sum in (15). Here n‾=(nmod⁡pj)j∈J\bar n=(n\bmod p_j)_{j\in J}, and a normalized sum of length zero is zero. pU=∏j∈Upj,GU(z)=∏j∈Uρpj(zj),𝒰Q={U⊆J:pU≤Q},Lp,N,Uf(z)=1N∑n=0N−1f(n)GU(z−n‾),Lp,N,Qf(z)=∑U∈𝒰QLp,N,Uf(z).\begin{aligned} p_U&=\prod_{j\in U}p_j,\qquad G_U(z)=\prod_{j\in U}\rho_{p_j}(z_j),\\ \mathcal U_Q&=\{U\subseteq J:p_U\le Q\},\\ L_{p,N,U}f(z)&=\frac1N\sum_{n=0}^{N-1}f(n)G_U(z-\bar n),\\ L_{p,N,Q}f(z)&=\sum_{U\in\mathcal U_Q}L_{p,N,U}f(z). \end{aligned}(15) For a nonnegative integer NN and a real exponent aa, the power cutoff is the natural-number floor of NaN^a.✦ The cutoff QNQ_N uses the fixed exponent β\beta in (16). β=1100000,QN=⌊Nβ⌋.\beta=\frac1{100000},\qquad Q_N=\lfloor N^\beta\rfloor.(16)

For a real linear functional LL on functions on a set Ω\Omega, its point mass at x∈Ωx\in\Omega is λL(x)=L(𝟏{x})\lambda_L(x)=L(\mathbf1_{\{x\}}).✦ For real linear functionals LjL_j on real observables of XjVX_j^V, where VV is a finite vertex set and each XjX_j is a finite state space, and a real observable FF on (∏j∈JXj)V(\prod_{j\in J}X_j)^V, the tensor product is the counting sum in (17). The product tuple law for a prime family pp is 𝝁p=⨂j∈Jμpj\boldsymbol\mu_p=\bigotimes_{j\in J}\mu_{p_j}.✦ (⨂j∈JLj)(F)=∑(zj)∈∏jXjV(∏j∈JλLj(zj))F(v↦(zj(v))j∈J).\left(\bigotimes_{j\in J}L_j\right)(F) =\sum_{(z_j)\in\prod_jX_j^V} \left(\prod_{j\in J}\lambda_{L_j}(z_j)\right) F\bigl(v\mapsto(z_j(v))_{j\in J}\bigr).(17)✦

VV is a finite vertex set, XX a state space, and LL a real linear functional on real observables of XVX^V. The multilinear integral ℐL\mathcal I_L is the functional of real functions gv:X→ℝg_v:X\to\mathbb R given in (18), and the diagonal integral 𝒟L\mathcal D_L is its value when the same function is used at every vertex.✦ For a fixed ambient prime family pp, the set diagonal functional YpY_p is defined by (18); the indicator 𝟏A(n)\mathbf1_A(n) is one for n∈An\in A and zero otherwise.✦ ℐL((gv)v∈V)=L(z↦∏v∈Vgv(zv)),𝒟L(g)=ℐL((g)v∈V),Yp(N,Q;A)=𝒟𝝁p(Lp,N,Q𝟏A).\begin{aligned} \mathcal I_L((g_v)_{v\in V}) &=L\left(z\mapsto\prod_{v\in V}g_v(z_v)\right),\\ \mathcal D_L(g)&=\mathcal I_L((g)_{v\in V}),\\ Y_p(N,Q;A)&=\mathcal D_{\boldsymbol\mu_p} \bigl(L_{p,N,Q}\mathbf1_A\bigr). \end{aligned}(18)

The Claim

Theorem 1. Fix a finite family (qs)s∈S(q_s)_{s\in S} of distinct primes with combined small prime modulus M≥2M\ge2. There are constants a>0a>0 and D≥1D\ge1, depending only on this family, such that, for every integer N≥1N\ge1 and every finite ambient prime family pp admissible at scale NN, each square-difference-free A⊆{0,…,N−1}A\subseteq\{0,\ldots,N-1\} satisfies (19). Yp(N,QN;A)≤DN−a.Y_p(N,Q_N;A)\le D N^{-a}.(19)✦ For any finite prime family pp with pj≥max⁡{Pmass,Prefl}p_j\ge\max\{P_{\mathrm{mass}},P_{\mathrm{refl}}\} for every jj, every integer N≥0N\ge0 and integer Q≥1Q\ge1, and every A⊆{0,…,N−1}A\subseteq\{0,\ldots,N-1\}, the density comparison is (20). αN(A)d≤Yp(N,Q;A).\alpha_N(A)^d\le Y_p(N,Q;A).(20)✦

In the paper, the result reads: there are absolute constants a0>0a_0>0 and C0<∞C_0<\infty such that the paper’s set functional satisfies Y(N,A)≤C0N−a0Y(N,A)\le C_0N^{-a_0} for every positive integer NN and every square-difference-free A⊆[N]A\subseteq[N].

What Is Checked

The decay assertion ranges over finite prime families satisfying the source prime hypotheses, and the density assertion is stated for real conductor-truncated lifts under product tuple laws with the mass and reflection thresholds.

The component and mixture functionals preserve addition and real scalar multiplication. The same identities hold for restricted mixtures and tensor products.