OctoInformBeta

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

Proposition 9.1: Exceptional subset budgets control the induction coefficient uniformly in the proposed power bound

A uniform step for power-saving induction.

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 square-difference-free set AA of nonnegative integers is a finite set containing no pair a,ba,b with b=a+m2b=a+m^2 for a positive integer mm.✦ At length NN, the induction uses A⊆{0,…,N−1}A\subseteq\{0,\ldots,N-1\}. A set in the public interval {1,…,N}\{1,\ldots,N\} is represented by its translate down by one. The quantity in the induction is a diagonal integral of a conductor-truncated residue lift of 𝟏A\mathbf1_A, with the same ambient prime family at every smaller length.

The Tuple Laws

Put h=24h=24, T=80T=80, m=2T+1=161m=2T+1=161, and r=h!/2r=h!/2. For the position set H={0,…,h−1}H=\{0,\ldots,h-1\} and a block set BB, a permutation word v∈Sym⁡(H)Bv\in\operatorname{Sym}(H)^B is a function assigning to each block a permutation of HH.✦ For the block set I={0,…,m−1}I=\{0,\ldots,m-1\}, the tuple vertex set is 𝒱=Sym⁡(H)I\mathcal V=\operatorname{Sym}(H)^I; write d=|𝒱|d=|\mathcal V|.✦ For an active block subset U⊆IU\subseteq I, an inactive-block address a∈𝒜U=Sym⁡(H)I\Ua\in\mathcal A_U=\operatorname{Sym}(H)^{I\setminus U} is a function assigning a permutation of HH to each block outside UU.✦ The inactive-block projection sends vv to v|I\Uv|_{I\setminus U}.✦ An addressed list index is an integer in LU={0,…,r|U|−1}L_U=\{0,\ldots,r^{|U|}-1\}, and an addressed list entry is a pair in 𝒜U×LU\mathcal A_U\times L_U.

For a finite block set BB, a group GG, an element c0∈Gc_0\in G and a natural threshold tt, threshold shift adjacency is the relation on addresses u,v∈GBu,v\in G^B defined by |{b∈B:v(b)=u(b)c0}|≥t|\{b\in B:v(b)=u(b)c_0\}|\ge t.✦

An ordered position pair is a pair (s,t)∈H2(s,t)\in H^2 with s<ts<t.✦ For a tuple vertex vv, a block bb of II and an ordered position pair (s,t)(s,t), the position transposition vb;s,tv^{b;s,t} is the tuple obtained by replacing v(b)v(b) by v(b)(st)v(b)(s\ t) and leaving the other blocks fixed.✦ The transposition interaction HUH_U on 𝒱\mathcal V, with parameter δ=1/(10h2)\delta=1/(10h^2), is given by (1): HU(v,w)=∑b∈U∑0≤s<t<hδt−s𝟏w=vb;s,t.H_U(v,w)=\sum_{b\in U}\sum_{0\le s<t<h} \delta^{t-s}\mathbf1_{w=v^{b;s,t}}.(1)✦ For a real matrix H0H_0 on a finite set VV, a label set LL and labelings j,k:V→Lj,k:V\to L, the label coincidence pairing CH0C_{H_0}, index energy EH0E_{H_0} and index Boltzmann weight WH0W_{H_0} are defined in (2): CH0(j,k)=∑x∈V∑y∈VH0(x,y)𝟏j(x)=k(y),EH0(j)=12CH0(j,j),WH0(j)=exp⁡(EH0(j)).\begin{aligned} C_{H_0}(j,k)&=\sum_{x\in V}\sum_{y\in V} H_0(x,y)\mathbf1_{j(x)=k(y)},\\ E_{H_0}(j)&=\tfrac12 C_{H_0}(j,j),\qquad W_{H_0}(j)=\exp(E_{H_0}(j)). \end{aligned}(2) The tuple Boltzmann weight on labelings j:𝒱→LUj:\mathcal V\to L_U is WU(j)=WHU(j)W_U(j)=W_{H_U}(j).✦

For a prime pp, the nonzero square indicator sps_p on 𝔽p\mathbb F_p is one on nonzero squares and zero elsewhere.✦ Let EUE_U be the list constraint edge set on 𝒜U×LU\mathcal A_U\times L_U, specifying the square-difference constraints between list values. For list values Y:𝒜U×LU→𝔽pY:\mathcal A_U\times L_U\to\mathbb F_p and a labeling j:𝒱→LUj:\mathcal V\to L_U, the square-constraint indicator VU(Y)V_U(Y) and sampled tuple configuration ΦU(Y,j)\Phi_U(Y,j) are defined by (3): VU(Y)=∏(a,b)∈EUsp(Y(b)−Y(a)),ΦU(Y,j)(v)=Y(v|I\U,j(v)).V_U(Y)=\prod_{(a,b)\in E_U}s_p(Y(b)-Y(a)),\qquad \Phi_U(Y,j)(v)=Y(v|_{I\setminus U},j(v)).(3) The tuple component average Cp,U(F)C_{p,U}(F) of a real observable F:𝔽p𝒱→ℝF:\mathbb F_p^{\mathcal V}\to\mathbb R is the uniform probability average in (4) over all such YY and jj: Cp,U(F)=𝔼Y𝔼jVU(Y)WU(j)F(ΦU(Y,j)).C_{p,U}(F)=\mathbb E_Y\mathbb E_j V_U(Y)W_U(j)F(\Phi_U(Y,j)).(4)✦

The block subset mixture weight is p−|U|/16p^{-|U|/16}.✦ The unnormalized tuple mixture Λp\Lambda_p and its restriction Λp,𝒟\Lambda_{p,\mathcal D} to a family 𝒟⊆𝒫(I)\mathcal D\subseteq\mathcal P(I) are the counting sums in (5): Λp(F)=∑U⊆I|U|≤T+1p−|U|/16Cp,U(F),Λp,𝒟(F)=∑U∈𝒟p−|U|/16Cp,U(F).\Lambda_p(F)=\sum_{\substack{U\subseteq I\\|U|\le T+1}} p^{-|U|/16}C_{p,U}(F),\qquad \Lambda_{p,\mathcal D}(F)=\sum_{U\in\mathcal D} p^{-|U|/16}C_{p,U}(F).(5) The terminal block family is 𝒟term={U⊆I:|U|=T+1}\mathcal D_{\mathrm{term}}=\{U\subseteq I:|U|=T+1\}.✦ The full tuple law μp\mu_p, exceptional tuple sublaw εp\varepsilon_p and exceptional mass ηp\eta_p are defined in (6): μp(F)=Λp(F)Λp(1),εp(F)=Λp,𝒟term(F)Λp(1),ηp=εp(1).\mu_p(F)=\frac{\Lambda_p(F)}{\Lambda_p(1)},\qquad \varepsilon_p(F)=\frac{\Lambda_{p,\mathcal D_{\mathrm{term}}}(F)} {\Lambda_p(1)},\qquad \eta_p=\varepsilon_p(1).(6)

For a real linear functional LL on functions on a finite configuration space, the point mass at zz is L({z})=L(𝟏{z})L(\{z\})=L(\mathbf1_{\{z\}}).✦ The product residue space of a finite family of primes 𝒑=(pj)j∈J\mathbf p=(p_j)_{j\in J} is X𝒑=∏j∈J𝔽pjX_{\mathbf p}=\prod_{j\in J}\mathbb F_{p_j}.✦ The product tuple law μ𝒑=⨂j∈Jμpj\mu_{\mathbf p}=\bigotimes_{j\in J}\mu_{p_j} is the tensor product law specified by the counting sum (7), for real observables FF on X𝒑𝒱X_{\mathbf p}^{\mathcal V}: μ𝒑(F)=∑(zj)∈∏j∈J𝔽pj𝒱(∏j∈Jμpj({zj}))F(v↦(zj(v))j∈J).\mu_{\mathbf p}(F)= \sum_{(z_j)\in\prod_{j\in J}\mathbb F_{p_j}^{\mathcal V}} \left(\prod_{j\in J}\mu_{p_j}(\{z_j\})\right) F\bigl(v\mapsto(z_j(v))_{j\in J}\bigr).(7)✦ For finite sets XX and VV, a real linear functional LL on real observables of XVX^V, and functions (fv)v∈V(f_v)_{v\in V} from XX to ℝ\mathbb R, the multilinear integral is L(z↦∏v∈Vfv(z(v)))L(z\mapsto\prod_{v\in V}f_v(z(v))); the diagonal integral puts the same function at every vertex.

The Lift and Its Cutoff

For a modulus qq, the centered residue delta on ℤ/qℤ\mathbb Z/q\mathbb Z is gq(x)=q𝟏x=0−1g_q(x)=q\mathbf1_{x=0}-1.✦ A subset U⊆JU\subseteq J has conductor qU=∏j∈Upjq_U=\prod_{j\in U}p_j and centered residue product GU(z)=∏j∈Ugpj(zj)G_U(z)=\prod_{j\in U}g_{p_j}(z_j).✦ The conductor-bounded support family at cutoff QQ is 𝒮𝒑(Q)={U⊆J:qU≤Q}\mathcal S_{\mathbf p}(Q)=\{U\subseteq J:q_U\le Q\}.✦ For a positive length NN and a real function ff on the nonnegative integers, the residue lift component ℒ𝒑,N,Uf\mathcal L_{\mathbf p,N,U}f and conductor-truncated residue lift ℒ𝒑,N,Qf\mathcal L_{\mathbf p,N,Q}f are defined by (8): ℒ𝒑,N,Uf(z)=1N∑n=0N−1f(n)GU(z−(nmod⁡pj)j∈J),ℒ𝒑,N,Qf(z)=∑U∈𝒮𝒑(Q)ℒ𝒑,N,Uf(z).\begin{aligned} \mathcal L_{\mathbf p,N,U}f(z) &=\frac1N\sum_{n=0}^{N-1}f(n) G_U\bigl(z-(n\bmod p_j)_{j\in J}\bigr),\\ \mathcal L_{\mathbf p,N,Q}f(z) &=\sum_{U\in\mathcal S_{\mathbf p}(Q)} \mathcal L_{\mathbf p,N,U}f(z). \end{aligned}(8) The set indicator 𝟏A(n)\mathbf1_A(n) is one for n∈An\in A and zero otherwise.✦ The set diagonal functional ℐ𝒑(N,Q,A)\mathcal I_{\mathbf p}(N,Q,A) is defined in (9): ℐ𝒑(N,Q,A)=μ𝒑(z↦∏v∈𝒱ℒ𝒑,N,Q𝟏A(z(v))).\mathcal I_{\mathbf p}(N,Q,A)= \mu_{\mathbf p}\left(z\mapsto\prod_{v\in\mathcal V} \mathcal L_{\mathbf p,N,Q}\mathbf1_A(z(v))\right).(9)✦

The power cutoff at exponent bb is ⌊Nb⌋\lfloor N^b\rfloor.✦ For a natural number kk, define the scale parameters and the abbreviation ℱ𝒑\mathcal F_{\mathbf p} by (10): β=1100000,γ=164,τ(k)=βγ1000(k+1),σ(k)=γτ(k)4,τ=τ(d),σ0=σ(d),ℱ𝒑(N,A)=ℐ𝒑(N,⌊Nβ⌋,A).\begin{gathered} \beta=\frac1{100000},\qquad \gamma=\frac1{64},\qquad \tau(k)=\frac{\beta\gamma}{1000(k+1)},\qquad \sigma(k)=\frac{\gamma\tau(k)}4,\\ \tau=\tau(d),\qquad \sigma_0=\sigma(d),\qquad \mathcal F_{\mathbf p}(N,A)= \mathcal I_{\mathbf p}(N,\lfloor N^\beta\rfloor,A). \end{gathered}(10) For natural numbers k,M,Nk,M,N, the progression transfer factor is 𝒯k,M(N)\mathcal T_{k,M}(N) in (11): 𝒯k,M(N)=(1+M2N2τ(k)−1)k.\mathcal T_{k,M}(N)=\bigl(1+M^2N^{2\tau(k)-1}\bigr)^k.(11)✦ The constant β\beta is 1/1000001/100000. The parameter γ\gamma is 1/641/64. For a natural number kk, the parameter τ(k)\tau(k) is βγ/(1000(k+1))\beta\gamma/(1000(k+1)). For a natural number kk, the parameter σ(k)\sigma(k) is γτ(k)/4\gamma\tau(k)/4.

The Small Prime Conditions

For a finite prime family (ℓs)s∈S(\ell_s)_{s\in S}, the local quadratic modulus msm_s at an index s∈Ss\in S is the integer specified by ms=8m_s=8 when ℓs=2\ell_s=2, and ms=ℓsm_s=\ell_s otherwise; the combined small prime modulus is the integer M=∏s∈SmsM=\prod_{s\in S}m_s. The small prime residue projection πs:ℤ/Mℤ→ℤ/msℤ\pi_s:\mathbb Z/M\mathbb Z\to\mathbb Z/m_s\mathbb Z is canonical reduction.✦ A strict local square modulo msm_s is the residue one when ℓs=2\ell_s=2, and a nonzero square otherwise.✦ Residues x,yx,y modulo MM form a strict quadratic residue pair when πs(y−x)\pi_s(y-x) is a strict local square for every s∈Ss\in S.✦

Let cc be the cyclic shift on HH, and let vk∈𝒱v_k\in\mathcal V have value ckc^k in every block.✦ A good small-prime assignment is a function α:𝒱→ℤ/Mℤ\alpha:\mathcal V\to\mathbb Z/M\mathbb Z for which (α(vk),α(vk+1))(\alpha(v_k),\alpha(v_{k+1})) is a strict quadratic residue pair for each 0≤k<240\le k<24.✦ Write ρ\rho for the associated good small-prime probability.

The combined prime family 𝒑̃\widetilde{\mathbf p} on S⊔JS\sqcup J agrees with ℓs\ell_s on SS and pjp_j on JJ.✦ For B⊆JB\subseteq J, the exceptional mass product is ηB=∏j∈Bηpj\eta_B=\prod_{j\in B}\eta_{p_j}.✦

The Prime Thresholds

The empty-block list index is zero.✦ The leading list identification ι:𝒱→𝒜⌀×L⌀\iota:\mathcal V\to\mathcal A_\varnothing\times L_\varnothing is the bijection v↦(v,0)v\mapsto(v,0).✦ For a bijection e:V→We:V\to W and an edge set E⊆W2E\subseteq W^2, the graph pullback e*Ee^*E consists of pairs (u,v)(u,v) with (e(u),e(v))∈E(e(u),e(v))\in E.✦ The leading constraint graph is E0=ι*E⌀E_0=\iota^*E_\varnothing; its edge count is e0=|E⌀|e_0=|E_\varnothing|.

For a finite vertex set VV and v∈Vv\in V, the vertex index decomposition identifies the singleton {v}\{v\} together with V\{v}V\setminus\{v\} with VV.✦ For a finite vertex set VV, a vertex v∈Vv\in V and an edge set E⊆V2E\subseteq V^2, a crossing edge is a pair in {v}×(V\{v})\{v\}\times(V\setminus\{v\}) whose endpoints form an edge of EE in either orientation.✦ Write m(E,v)m(E,v) for the number of these pairs. The graph conditional energy constant is c(E,v)c(E,v) in (12): c(E,v)=(2(1/2)m(E,v)+4((1/2)m(E,v))2)((32m(E,v))1/4)2.c(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.(12)✦

The Boltzmann weight upper bound WmaxW_{\max}, leading mass lower bound b0b_0, and conditional error numerator CnumC_{\mathrm{num}} are the constants in (13); CcondC_{\mathrm{cond}} is the conditional decay constant: Wmax=exp⁡(d(T+1)(h2)δ2),b0=2−(e0+1),Cnum=∑v∈𝒱c(E0,v)+2mWmax,Ccond=Cnum/b0.\begin{gathered} W_{\max}=\exp\left(\frac{d(T+1)\binom h2\delta}{2}\right),\qquad b_0=2^{-(e_0+1)},\\ C_{\mathrm{num}}=\sum_{v\in\mathcal V}c(E_0,v)+2^mW_{\max},\qquad C_{\mathrm{cond}}=C_{\mathrm{num}}/b_0. \end{gathered}(13) The half vertex count is R=(h!)m/2R=(h!)^m/2, and the interaction lower bound is κ=910δh−1\kappa=\tfrac9{10}\delta^{h-1}. The marked and unmarked reflection constants are specified in (14): Cmarked=(exp⁡(κ)−1)R(rT+1)R,Cunmarked=(32d2)1/4Wmax⁡2.C_{\mathrm{marked}}= \frac{(\exp(\kappa)-1)^R}{(r^{T+1})^R},\qquad C_{\mathrm{unmarked}}=(32d^2)^{1/4}W_{\max}^2.(14) The mass positivity threshold PmassP_{\mathrm{mass}}, conditional decay threshold PcondP_{\mathrm{cond}} and reflection positivity threshold PreflP_{\mathrm{refl}} are defined in (15): Pmass=max⁡{3,(2e0+1e0)2},Pcond=max⁡{Pmass,3,Ccond32},Prefl=max⁡{3,(CunmarkedrR/Cmarked)16}.\begin{aligned} P_{\mathrm{mass}}&=\max\{3,(2^{e_0+1}e_0)^2\},\\ P_{\mathrm{cond}}&=\max\{P_{\mathrm{mass}},3,C_{\mathrm{cond}}^{32}\},\\ P_{\mathrm{refl}}&=\max\{3, (C_{\mathrm{unmarked}}r^R/C_{\mathrm{marked}})^{16}\}. \end{aligned}(15)

The Claim

Theorem 1 (Power-saving induction step). Fix a finite small prime family (ℓs)s∈S(\ell_s)_{s\in S} with M≥2M\ge2. There exist a real E>0E>0 and a positive integer threshold N0N_0, depending only on this family, with the following property. Let N≥N0N\ge N_0 and let 𝒑=(pj)j∈J\mathbf p=(p_j)_{j\in J} be a finite prime family. Suppose the combined family 𝒑̃\widetilde{\mathbf p} is injective and contains every prime at most NN. For every j∈Jj\in J, suppose pjp_j is coprime to MM, pj≥max⁡{64,Pmass,Pcond,Prefl}p_j\ge\max\{64,P_{\mathrm{mass}},P_{\mathrm{cond}},P_{\mathrm{refl}}\}, and ηpj≤pj−4\eta_{p_j}\le p_j^{-4}.

Let 0≤a≤120\le a\le\tfrac12 and D≥0D\ge0. Assume the smaller-scale power bound (16) for every positive integer n<Nn<N and every square-difference-free A′⊆{0,…,n−1}A'\subseteq\{0,\ldots,n-1\}: ℱ𝒑(n,A′)≤Dn−a.\mathcal F_{\mathbf p}(n,A')\le Dn^{-a}.(16) Then every square-difference-free A⊆{0,…,N−1}A\subseteq\{0,\ldots,N-1\} satisfies the induction estimate (17): ℱ𝒑(N,A)≤EN−σ0+𝒯d,M(N)DN−aM2a(1−3ρ4).\mathcal F_{\mathbf p}(N,A) \le EN^{-\sigma_0} +\mathcal T_{d,M}(N)DN^{-a} M^{2a}\left(1-\frac{3\rho}{4}\right).(17)✦

In the paper, the result reads: there are absolute constants K<∞K<\infty and N1≥2N_1\ge2, independent of any proposed induction bound for the induction quantity 𝒴\mathcal Y, such that for every integer N≥N1N\ge N_1 the recurrence (18) holds for the child lengths nBn_B and prime family 𝒫N\mathcal P_N, with qB=∏p∈Bpq_B=\prod_{p\in B}p and uN→0u_N\to0: 𝒴(N)≤KN−σ0+(1+uN)((1−ρ)𝒴(n⌀)+ρ∑⌀≠B⊆𝒫NqB≤NτηB𝒴(nB)).\mathcal Y(N)\le KN^{-\sigma_0}+(1+u_N) \left((1-\rho)\mathcal Y(n_\varnothing)+ \rho\sum_{\substack{\varnothing\ne B\subseteq\mathcal P_N\\q_B\le N^\tau}} \eta_B\mathcal Y(n_B)\right).(18)

What Is Checked

The induction estimate is stated over the real set diagonal functional (9), with the floor cutoff in (10), a fixed small prime family, and the same ambient prime family throughout the smaller-scale hypothesis.

Each local quadratic modulus divides MM.✦

The component and mixture functionals, together with their tensor product, preserve addition and multiplication by real scalars.

Let (pj)j∈J(p_j)_{j\in J} be distinct primes at least five, with ηpj≤pj−4\eta_{p_j}\le p_j^{-4}. For any family ℬ\mathcal B of nonempty subsets of JJ and any real b≤1b\le1, the exceptional power budget is (19): ∑B∈ℬηBqBb≤14.\sum_{B\in\mathcal B}\eta_Bq_B^b\le\frac14.(19)✦ For a positive natural number NN and any real tt, the large exceptional tail bound under these hypotheses is (20): ∑⌀≠B⊆JNt<qBηB≤14N−t.\sum_{\substack{\varnothing\ne B\subseteq J\\N^t<q_B}}\eta_B \le\frac14N^{-t}.(20)✦

For any fixed natural numbers kk and MM, the transfer factor (11) has the natural-number limit (21): limN→∞N∈ℕ𝒯k,M(N)=1.\lim_{\substack{N\to\infty\\N\in\mathbb N}} \mathcal T_{k,M}(N)=1.(21)✦