OctoInformBeta

OpenAI preprint, September 2026 · Section 8. Contraction on ordered intervals

Proposition 8.2: An explicit error estimate and eventual power decay for square-difference-free sets

Ordered interval bounds for good residue lifts.

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

Residue slices of a square-difference-free set give real functions on a product of prime fields. Their multilinear average is taken against normalized good tuple laws, which couple a fixed family of permutation vertices.

For a positive integer NN, write [N]={1,…,N}[N]=\{1,\ldots,N\} for the public integer interval and IN={0,…,N−1}I_N=\{0,\ldots,N-1\} for the natural interval. Translation by one sends A⊆INA\subseteq I_N to A+1⊆[N]A+1\subseteq[N]. A finite set A⊆ℕA\subseteq\mathbb N is square-difference-free when b≠a+m2b\ne a+m^2 for all a,b∈Aa,b\in A and positive integers mm.✦

For a positive integer hh, a permutation word vv on a block set BB is a function assigning a permutation of {0,…,h−1}\{0,\ldots,h-1\} to each block.✦ Write 𝔖h\mathfrak S_h for the symmetric group on {0,…,h−1}\{0,\ldots,h-1\}. The permutation tuple vertices 𝒱\mathcal V and their parameters are specified in (1): h=24,T=80,B={0,…,2T},|B|=161,r=h!/2,δ=110h2,𝒱=(𝔖h)B,d=|𝒱|.\begin{gathered} h=24,\qquad T=80,\qquad B=\{0,\ldots,2T\},\qquad |B|=161,\\ r=h!/2,\qquad \delta=\frac1{10h^2},\qquad \mathcal V=(\mathfrak S_h)^B,\qquad d=|\mathcal V|. \end{gathered}(1) The cyclic shift cc sends each position to the next, returning the last to zero.✦ The cyclic tuple vertex vk∈𝒱v_k\in\mathcal V has value ckc^k in every block.✦

Fix a finite small prime family (qi)i∈I0(q_i)_{i\in I_0}. Its local quadratic moduli m(qi)m(q_i) and combined modulus MM are given in (2); let πi:ℤ/Mℤ→ℤ/m(qi)ℤ\pi_i:\mathbb Z/M\mathbb Z\to\mathbb Z/m(q_i)\mathbb Z be canonical reduction. The strict local square sets 𝒬i\mathcal Q_i in that equation consist of the residue 11 modulo 88 at the prime 22, and nonzero squares at odd primes.✦ m(q)={8,q=2,q,q≠2,M=∏i∈I0m(qi),𝒬i={{1}⊆ℤ/8ℤ,qi=2,{x∈𝔽qi:x≠0,x is a square},qi≠2.\begin{gathered} m(q)=\begin{cases}8,&q=2,\\q,&q\ne2,\end{cases} \qquad M=\prod_{i\in I_0}m(q_i),\\ \mathcal Q_i= \begin{cases} \{1\}\subseteq\mathbb Z/8\mathbb Z,&q_i=2,\\ \{x\in\mathbb F_{q_i}:x\ne0,\ x\text{ is a square}\},&q_i\ne2. \end{cases} \end{gathered}(2) A strict residue pair (u,w)(u,w) modulo MM is a pair of elements of ℤ/Mℤ\mathbb Z/M\mathbb Z satisfying πi(w−u)∈𝒬i\pi_i(w-u)\in\mathcal Q_i for every ii.✦ A residue tuple s:𝒱→ℤ/Mℤs:\mathcal V\to\mathbb Z/M\mathbb Z is good when its consecutive cyclic vertices form strict residue pairs, as in (3): πi(s(vk+1)−s(vk))∈𝒬i(i∈I0,0≤k<24).\pi_i\bigl(s(v_{k+1})-s(v_k)\bigr)\in\mathcal Q_i \qquad(i\in I_0,\ 0\le k<24).(3)

For a finite set A⊆ℕA\subseteq\mathbb N, let 𝟏A\mathbf1_A be its indicator. The normalized residue slice fu:ℕ→ℝf_u:\mathbb N\to\mathbb R at u∈ℤ/Mℤu\in\mathbb Z/M\mathbb Z is the function in (4): fu(n)=M𝟏A(n)𝟏{n≡u(mod⁡M)}.f_u(n)=M\mathbf1_A(n)\mathbf1_{\{n\equiv u\pmod M\}}.(4)✦

Residue Lifts

Let (pj)j∈J(p_j)_{j\in J} be a finite prime family and Xp=∏j∈J𝔽pjX_p=\prod_{j\in J}\mathbb F_{p_j} its product residue space. Write n‾=(nmod⁡pj)j∈J\bar n=(n\bmod p_j)_{j\in J}. For a prime qq, the centered residue delta gq:𝔽q→ℝg_q:\mathbb F_q\to\mathbb R and the centered residue product GUG_U are the real functions in (5): gq(x)=q𝟏{x=0}−1,GU(z)=∏j∈Ugpj(zj)(U⊆J,z∈Xp).g_q(x)=q\mathbf1_{\{x=0\}}-1,\qquad G_U(z)=\prod_{j\in U}g_{p_j}(z_j) \quad(U\subseteq J,\ z\in X_p).(5) For a cutoff Q∈ℕQ\in\mathbb N, a conductor-bounded support U⊆JU\subseteq J is a subset of JJ whose product of primes is at most QQ. The family 𝒰Q\mathcal U_Q and the interval lift component ℓa,L,Uf:Xp→ℝ\ell_{a,L,U}f:X_p\to\mathbb R are defined in (6) for a,L,Q∈ℕa,L,Q\in\mathbb N and f:ℕ→ℝf:\mathbb N\to\mathbb R: 𝒰Q={U⊆J:∏j∈Upj≤Q},ℓa,L,Uf(z)=1L∑0≤t<Lf(a+t)GU(z−a+t¯).\mathcal U_Q=\left\{U\subseteq J:\prod_{j\in U}p_j\le Q\right\}, \qquad \ell_{a,L,U}f(z)=\frac1L\sum_{0\le t<L} f(a+t)G_U(z-\overline{a+t}).(6) The component is zero for L=0L=0. The residue lift component on ILI_L is ℓ0,L,Uf\ell_{0,L,U}f.✦ Partial lifts retain complete exact supports: for a family 𝒯⊆𝒰Q\mathcal T\subseteq\mathcal U_Q, their sum and the conductor-truncated lift are as in (7): ℒa,L𝒯f=∑U∈𝒯ℓa,L,Uf,ℒN,Qf=∑U∈𝒰Qℓ0,N,Uf.\mathcal L^{\mathcal T}_{a,L}f =\sum_{U\in\mathcal T}\ell_{a,L,U}f, \qquad \mathcal L_{N,Q}f =\sum_{U\in\mathcal U_Q}\ell_{0,N,U}f.(7)

For a real parameter RR, the moment hypothesis at (L,Q,R)(L,Q,R) is the uniform probability bound (8), for every a∈ℕa\in\mathbb N, v∈𝒱v\in\mathcal V and 𝒯⊆𝒰Q\mathcal T\subseteq\mathcal U_Q: 𝔼x∈Xp(ℒa,L𝒯fs(v)(x))2m≤R2m(m∈{1,d−1}).\mathbb E_{x\in X_p} \left(\mathcal L^{\mathcal T}_{a,L}f_{s(v)}(x)\right)^{2m} \le R^{2m} \qquad(m\in\{1,d-1\}).(8)

The Good Tuple Laws

For an active block subset S⊆BS\subseteq B, an inactive-block address u∈𝒜S=(𝔖h)B\Su\in\mathcal A_S=(\mathfrak S_h)^{B\setminus S} is a permutation word on B\SB\setminus S.✦ The addressed list indices form 𝒥S={0,…,r|S|−1}\mathcal J_S=\{0,\ldots,r^{|S|}-1\}✦, and addressed list entries form ℰS=𝒜S×𝒥S\mathcal E_S=\mathcal A_S\times\mathcal J_S.✦ The inactive-block projection of a vertex is ρS(v)=v|B\S\rho_S(v)=v|_{B\setminus S}.✦ For a prime pp, given list values Y:ℰS→𝔽pY:\mathcal E_S\to\mathbb F_p and labels j:𝒱→𝒥Sj:\mathcal V\to\mathcal J_S, the sampled tuple configuration is XS,Y,j(v)=Y(ρS(v),j(v))X_{S,Y,j}(v)=Y(\rho_S(v),j(v)).✦

For group-valued addresses u,wu,w on a finite block set, threshold shift adjacency with shift cc and threshold tt is the relation that holds when at least tt blocks satisfy w(b)=u(b)cw(b)=u(b)c.✦ The list constraint edges ES⊆ℰS×ℰSE_S\subseteq\mathcal E_S\times\mathcal E_S specify the square-difference constraints between list values. The nonzero square indicator χp\chi_p and square-constraint indicator VSV_S are given in (9): χp(x)=𝟏{x≠0,x is a square},VS(Y)=∏(u,w)∈ESχp(Y(w)−Y(u)).\chi_p(x)=\mathbf1_{\{x\ne0,\ x\text{ is a square}\}}, \qquad V_S(Y)=\prod_{(u,w)\in E_S}\chi_p(Y(w)-Y(u)).(9) For a vertex v∈𝒱v\in\mathcal V, a block e∈Be\in B and positions 0≤a<b<h0\le a<b<h, the position transposition ve,(ab)v^{e,(a\,b)} composes the permutation of vv at block ee on the right with (ab)(a\,b) and leaves its other blocks fixed.✦ The transposition interaction HSH_S is the matrix in (10): HS(v,w)=∑e∈S∑0≤a<b<hδb−a𝟏{w=ve,(ab)}.H_S(v,w)=\sum_{e\in S}\sum_{0\le a<b<h} \delta^{b-a}\mathbf1_{\{w=v^{e,(a\,b)}\}}.(10)✦ For a real matrix HH on a finite set and labelings j,kj,k of that set, the label coincidence pairing CH(j,k)C_H(j,k) is the counting sum in (11).✦ For a real matrix HH on a finite set and a labeling jj of that set, the index energy EH(j)E_H(j) is one half of the diagonal label coincidence pairing✦, and the index Boltzmann weight WH(j)W_H(j) is its exponential✦; the tuple Boltzmann weight for the transposition interaction HSH_S is WS=WHSW_S=W_{H_S}.✦ CH(j,k)=∑v,wH(v,w)𝟏{j(v)=k(w)},EH(j)=12CH(j,j),WH(j)=exp⁡(EH(j)).C_H(j,k)=\sum_{v,w}H(v,w)\mathbf1_{\{j(v)=k(w)\}},\qquad E_H(j)=\tfrac12 C_H(j,j),\qquad W_H(j)=\exp(E_H(j)).(11) The tuple component average Cp,SC_{p,S} on real observables F:𝔽p𝒱→ℝF:\mathbb F_p^{\mathcal V}\to\mathbb R is the uniform probability average in (12) over list values and labelings: Cp,S(F)=𝔼Y∈𝔽pℰS𝔼j∈𝒥S𝒱VS(Y)WS(j)F(XS,Y,j).C_{p,S}(F)= \mathbb E_{Y\in\mathbb F_p^{\mathcal E_S}} \mathbb E_{j\in\mathcal J_S^{\mathcal V}} V_S(Y)W_S(j)F(X_{S,Y,j}).(12)✦ For a family 𝒟\mathcal D of block subsets, the restricted mixture νp,𝒟\nu_{p,\mathcal D} uses block subset weights p−|S|/16p^{-|S|/16}. Its normalization μp,𝒟\mu_{p,\mathcal D} and the good tuple law μp∘\mu_p^\circ are defined by (13), using the truncated block family 𝒟T\mathcal D_T: νp,𝒟(F)=∑S∈𝒟p−|S|/16Cp,S(F),μp,𝒟(F)=νp,𝒟(F)νp,𝒟(1),𝒟n={S⊆B:|S|≤n},μp∘=μp,𝒟T.\begin{gathered} \nu_{p,\mathcal D}(F)=\sum_{S\in\mathcal D}p^{-|S|/16}C_{p,S}(F), \qquad \mu_{p,\mathcal D}(F)=\frac{\nu_{p,\mathcal D}(F)}{\nu_{p,\mathcal D}(1)},\\ \mathcal D_n=\{S\subseteq B:|S|\le n\},\qquad \mu_p^\circ=\mu_{p,\mathcal D_T}. \end{gathered}(13)

A functional μ\mu on a finite configuration space has point mass μ({y})=μ(𝟏{y})\mu(\{y\})=\mu(\mathbf1_{\{y\}}).✦ The tensor product law Λp\Lambda_p on Xp𝒱X_p^{\mathcal V} and the multilinear integral ℐN,Q\mathcal I_{N,Q} are the counting sums and product observable in (14), where z(j)=(zv(j))v∈𝒱z^{(j)}=(z_v(j))_{v\in\mathcal V}: Λp(F)=∑z∈Xp𝒱F(z)∏j∈Jμpj∘({z(j)}),ℐN,Q(A,s;p)=Λp(z↦∏v∈𝒱ℒN,Qfs(v)(zv)).\begin{aligned} \Lambda_p(F)&=\sum_{z\in X_p^{\mathcal V}} F(z)\prod_{j\in J}\mu_{p_j}^\circ(\{z^{(j)}\}),\\ \mathcal I_{N,Q}(A,s;p)&= \Lambda_p\left(z\mapsto\prod_{v\in\mathcal V} \mathcal L_{N,Q}f_{s(v)}(z_v)\right). \end{aligned}(14)

An ordered position pair at height hh is a pair (a,b)(a,b) of positions in {0,…,h−1}\{0,\ldots,h-1\} with a<ba<b.✦

Thresholds and Errors

The leading list identification sends v∈𝒱v\in\mathcal V to (ρ⌀(v),0)(\rho_\varnothing(v),0), where zero is the unique empty-block list index.✦ The leading constraint graph Γ\Gamma is the pullback of E⌀E_\varnothing under this bijection: a pair (v,w)(v,w) is an edge precisely when its image is an edge of E⌀E_\varnothing.✦ Let e=|E⌀|e=|E_\varnothing| be the leading graph edge count.✦

For a vertex vv of a finite directed graph EE, split the vertices into {v}\{v\} and its complement. Let mE(v)m_E(v) count the crossing edges, counting either orientation once. The graph conditional energy constant gE(v)g_E(v) is defined by (15): gE(v)=(2(1/2)mE(v)+4((1/2)mE(v))2)((32mE(v))1/4)2.g_E(v)=\left(\frac2{(1/2)^{m_E(v)}}+ \frac4{((1/2)^{m_E(v)})^2}\right) \left((32m_E(v))^{1/4}\right)^2.(15)✦ The leading mass lower bound a*a_* and Boltzmann weight upper bound W*W_* are specified in (16), together with the mass positivity threshold PmassP_{\mathrm{mass}}: a*=2−(e+1),W*=exp⁡(d(T+1)(h2)δ2),Pmass=max⁡{3,(2e+1e)2}.a_*=2^{-(e+1)},\qquad W_*=\exp\left(\frac{d(T+1)\binom h2\delta}{2}\right),\qquad P_{\mathrm{mass}}=\max\{3,(2^{e+1}e)^2\}.(16) The conditional error numerator b*b_* determines the conditional decay constant CcondC_{\mathrm{cond}} and conditional decay threshold PcondP_{\mathrm{cond}} through (17): b*=∑v∈𝒱gΓ(v)+2|B|W*,Ccond=b*/a*,Pcond=max⁡{Pmass,3,Ccond32}.b_*=\sum_{v\in\mathcal V}g_\Gamma(v)+2^{|B|}W_*,\qquad C_{\mathrm{cond}}=b_*/a_*,\qquad P_{\mathrm{cond}}=\max\{P_{\mathrm{mass}},3,C_{\mathrm{cond}}^{32}\}.(17)

For positive integers K,HK,H and a real parameter RR, the absolute kernel constant CkerC_{\mathrm{ker}} and ordered interval error ℰord\mathcal E_{\mathrm{ord}} are given in (18): Cker=264+3(4+8π)+156+16,ℰord(d,M,K,H,R)=MdK3dH−1/3(K2MCker+R2).\begin{gathered} C_{\mathrm{ker}}=2^{64}+3(4+8\pi)+156+16,\\ \mathcal E_{\mathrm{ord}}(d,M,K,H,R) =M^dK^{3d}H^{-1/3}\bigl(K^2MC_{\mathrm{ker}}+R^2\bigr). \end{gathered}(18)

For the exponent β\beta specified in (19), the power cutoff is QN=⌊Nβ⌋Q_N=\lfloor N^\beta\rfloor, with natural-number floor.✦ The decay parameters in (19) fix the exponent σ0\sigma_0: β=1100000,γ=164,τ(d)=βγ1000(d+1),σ(d)=γτ(d)4,σ0=σ(d).\beta=\frac1{100000},\qquad \gamma=\frac1{64},\qquad \tau(d)=\frac{\beta\gamma}{1000(d+1)},\qquad \sigma(d)=\frac{\gamma\tau(d)}4,\qquad \sigma_0=\sigma(d).(19)

For a finite vertex set VV and v∈Vv\in V, the vertex index decomposition at vv is the bijection from the disjoint union of a singleton and V\{v}V\setminus\{v\} onto VV that sends the singleton to vv and every other vertex to itself.✦

The Claim

The combined prime family p̃\widetilde p on I0⊔JI_0\sqcup J agrees with qq on I0I_0 and with pp on JJ.✦

Theorem 1. Fix the small prime family qq. For a positive integer NN, suppose p̃\widetilde p is injective, every prime at most NN occurs in p̃\widetilde p, and pj≥64p_j\ge64 and pj≥max⁡{Pmass,Pcond}p_j\ge\max\{P_{\mathrm{mass}},P_{\mathrm{cond}}\} for every j∈Jj\in J. Let A⊆INA\subseteq I_N be square-difference-free, and let ss be a good residue tuple.

Let N,L,H,KN,L,H,K be positive integers, let Q∈ℕQ\in\mathbb N, and let R≥1R\ge1 be real. Assume the moment hypothesis (8) and the scale conditions (20): N≤KL,KLN≤2,NL≤K,HKd≤Q,MKd≤N1/1000,H≤N1/1000,8≤N1/16.\begin{gathered} N\le KL,\qquad \frac{KL}{N}\le2,\qquad \frac NL\le K, \qquad HK^d\le Q,\\ MK^d\le N^{1/1000},\qquad H\le N^{1/1000},\qquad 8\le N^{1/16}. \end{gathered}(20) The ordered global estimate is (21): |ℐN,Q(A,s;p)|≤2d(2K−1R2d+dK−1/64Rd+ℰord(d,M,K,H,R)).|\mathcal I_{N,Q}(A,s;p)| \le2^d\left(2K^{-1}R^{2d}+dK^{-1/64}R^d+ \mathcal E_{\mathrm{ord}}(d,M,K,H,R)\right).(21)✦ There exist C>0C>0 and an integer N0N_0, depending on the fixed family qq, such that for every N≥N0N\ge N_0 and every p,A,sp,A,s satisfying the hypotheses in the first paragraph of this theorem, the ordered contraction estimate (22) holds: |ℐN,QN(A,s;p)|≤CN−σ0.|\mathcal I_{N,Q_N}(A,s;p)|\le C N^{-\sigma_0}.(22)✦

In the paper, the result reads: there are absolute C>0C>0 and N0N_0 such that whenever N≥N0N\ge N_0, A⊆[N]A\subseteq[N] has no positive square difference and ss is good, the bound (23) holds uniformly in AA and ss, for the paper’s prime family 𝒫N\mathcal P_N and cutoff QQ, with good subprobabilities μp−εp\mu_p-\varepsilon_p: |∫∏v∈𝒱ℒ[N],𝒫N,Qfs(v)(zv)d⨂p∈𝒫N(μp−εp)|≤CN−σ0.\left|\int\prod_{v\in\mathcal V} \mathcal L_{[N],\mathcal P_N,Q}f_{s(v)}(z_v)\, d\!\bigotimes_{p\in\mathcal P_N}(\mu_p-\varepsilon_p)\right| \le C N^{-\sigma_0}.(23)

What Is Checked

The assertions are stated over finite injective combined prime families and real residue lifts of square-difference-free subsets of INI_N, with the cyclic strict-square condition on their residue tuples. Each local quadratic modulus m(qi)m(q_i) divides MM✦, and its canonical reduction πi\pi_i is a ring homomorphism. The tuple component averages are real linear functionals; the restricted mixtures and tensor product law preserve sums of observables and multiplication by real scalars.