betaFeedback

WorksMathematics

Complete modulus locking in the exceptional set of the abc conjecture

Abstract

Let Nλ(X)N_\lambda(X) count primitive positive triples a+b=c≤Xa+b=c\le X with rad(abc)<cλ\mathrm{rad}(abc)<c^\lambda. We prove, using inherited anatomic estimates and a new finite exact-rational implication, that Nλ(X)≪λ,εXU+εN_\lambda(X)\ll_{\lambda,\varepsilon}X^{U+\varepsilon} for 0<λ≤10<\lambda\le1, where U=1168000720000012<73125=0.584U=\frac{11680007}{20000012}<\frac{73}{125}=0.584. The new arithmetic-computational step exhausts all third-letter whole-value modulus sets: index 11 is eliminated by a one-Lipschitz argument, leaving 3232 subsets of {2,3,4,5,6}\{2,3,4,5,6\}, and grouped power-difference bounds are used in every fixed degree 2≤e≤62\le e\le6. A branch-and-bound tree with 65,57665{,}576 leaves is verified by an independent standard-library checker which reconstructs every inequality and checks every rational multiplier; floating-point linear programming is used only to discover the tree. We also prove that the enlargement is mathematically essential: an explicit rational point feasible for the old modulus family has D=403/690D=403/690, whereas FullOC gives the collision bound 53/9253/92 at the same point, and a second explicit point gives 7/12≤DFullOC∗7/12\le D^*_{\mathrm{FullOC}}. These are statements about exponent programs, not lower bounds for abc triples. No assertion of abc, Szpiro, finiteness of the exceptional set, or an Arakelov degeneration functor is made.

Introduction and status of the results

For λ>0\lambda> 0, define

Nλ(X)=#{(a,b,c)∈N3:a+b=c≤X, gcd⁡(a,b,c)=1, rad⁡(abc)<cλ}.N_{\lambda}(X)=\#\left\{(a,b,c)\in\mathbb{N}^{3}:a+b=c\leq X,\ \gcd(a,b,c)=1,\ \operatorname{rad}(abc)<c^{\lambda}\right\}.

The abc conjecture predicts that Nλ(X)N_{\lambda}(X) is bounded for each fixed λ<1\lambda< 1. This paper concerns unconditional upper bounds for this counting function; it does not assert the pointwise radical inequality required by abc.

Bernert, Browning, Lichtman and Teräväinen introduced the anatomic exponent program used below and obtained the exponent 3/53/5 for 0<λ<10 < \lambda< 1, together with a budget-dependent bound for 0<λ≤20 < \lambda\leq2 [1]. Li subsequently introduced collision inequalities and radical-budget refinements [2], and then grouped power differences, released coordinates and a whole-value residue bound [3]. The present argument starts from that framework. Its new ingredients are the complete whole-value modulus family, its exact linearization in degrees through 66, and the rational separation and obstruction analysis.

Main arithmetic theorem

Theorem 1.1 (Computer-assisted theorem with exact rational proof data). Put

U=1168000720000012=0.583999999600000240….U=\frac{11680007}{20000012}=0.583999999600000240\ldots.

For every 0<λ≤10 < \lambda\leq1 and every ε>0\varepsilon> 0,

Nλ(X)≪λ,εXU+ε.(1)N_{\lambda}(X)\ll_{\lambda,\varepsilon}X^{U+\varepsilon}. \tag*{(1)}

In particular,

Nλ(X)≪λ,εX73/125+ε=X0.584+ε,73125−U=12500001500.N_{\lambda}(X)\ll_{\lambda,\varepsilon}X^{73/125+\varepsilon}=X^{0.584+\varepsilon},\qquad\frac{73}{125}-U=\frac{1}{2500001500}.

The finite optimization step is not inferred from a floating-point optimum. It is proved by a complete disjunctive tree and positive rational multipliers, independently reconstructed and checked in exact arithmetic. The passage from the exponent program to (1.1), including all logarithmic, divisor and decomposition losses, is given in Section 7.

For orientation, Table 1 lists the stated exponents in the principal sources used here. Decimal values are rounded only for display. No claim of a comprehensive literature optimum is made.

SourceCertified exponentApproximation
BBLT [1]3/53/50.600000000
Li, collision program [2]54/9154/910.593406593
Li, squares and twisted energy [3]589903/1000008589903/10000080.589898281
Theorem 1.111680007/200000120.5839999996

Table 1. Selected general bounds for Nλ(X)N_{\lambda}(X) at the budget endpoint λ=1\lambda= 1; the BBLT 3/53/5 bound is stated for λ<1\lambda< 1.

Relative to the exact exponent of [3], the gain is

5899031000008−U=294916596955000043000024=0.0058982812137….\frac{589903}{1000008}-U=\frac{29491659695}{5000043000024}=0.0058982812137\ldots.

What is new

The argument has five connected outputs.

  1. All fixed degrees through 6. The grouped difference estimate is used with e=2,3,4,5,6e = 2, 3, 4, 5, 6. The underlying uniform count for aUe−bVe=maU^e - bV^e = m in arbitrary fixed degree is already proved in [3 Appendix B]; we reproduce its proof and do not claim a new general-degree counting theorem.

  1. Complete whole-value modulus locking. For fixed remaining collision data, all third-letter sets OC⊆{1,…,6}O_C \subseteq\{1, \ldots, 6\} are considered. A one-Lipschitz deletion argument proves index 1 redundant, leaving the 32 subsets of {2,3,4,5,6}\{2, 3, 4, 5, 6\}. This completeness statement is specific to the whole-value lock and the savings δ0,δ2,…,δ6\delta_0, \delta_2, \ldots, \delta_6; it is not asserted for an arbitrary coordinate-lock program.

  1. A new exact upper certificate. The resulting finite implication has 65,576 leaves, 4,591 geometric nodes and 7,623 collision nodes. Its maximum leaf bound is exactly UU.

  1. Strict separation from the old family. The old modulus family has an exact feasible point with D=403/690D = 403/690. At that point a new FullOC instance gives 53/92<D53/92 < D. Consequently the stronger arithmetic family produces an independent exponent improvement.

  1. A precise remaining program obstruction. A symmetric rational point satisfies all 15 exponent pairs and all 32 FullOC sets with D=7/12D = 7/12. Thus the present program cannot prove an exponent below 7/127/12, and its remaining possible gain below UU is at most U−7/12U - 7/12, without any assertion that this gain is attainable.

The two feasible points are lower obstructions for linear programs. They do not construct integer triples and give no lower bound for Nλ(X)N_{\lambda}(X).

Provenance and logical status

Propositions taken from the literature are labelled as inherited inputs. Section 2 records the BBLT anatomic estimates in the exponent form used by Li. Section 3 reproduces the proofs of the fixed-degree weighted-difference estimate and the arbitrary-unit energy estimate from [3], so that the later collision proof is self-contained. The new work begins with the exhaustive degree-and-modulus implementation of the collision inequality, the index-1 elimination, the finite certificates and the exact program separation.

No floating-point quantity is used as a proof hypothesis. The numerical optimizer discovers branches and candidate multipliers only. The proof consists of the arithmetic inequalities, complete child coverage, and exact rational row and multiplier identities. The main result is therefore a computer-assisted theorem, not a numerical conjecture; it has not been represented here as peer reviewed.

Section 9 explains the boundary with abc and Szpiro. In particular, neither a finiteness assertion for Nλ(X)N_{\lambda}(X) when λ<1\lambda< 1 nor a degeneration functor on a cyclotomic tower is constructed.

Anatomic boxes and the inherited exponent system

This section records the external arithmetic framework. Nothing in it is claimed as new. We use the formulation of [3 Section 2], which in turn cites [1 Propositions 2.1, 2.3, 3.1 and 4.1].

Powerful-coordinate boxes

Write x∼Px \sim P for P<x≤2PP < x \leq2P, with the lowest dyadic block enlarged to include 1. After dyadic decomposition in height and radical, and after fixing a decomposition error ε0>0\varepsilon_0 > 0, there is a fixed d=d(ε0)≥6d = d(\varepsilon_0) \geq6 such that the restricted count is at most Xε0X^{\varepsilon_0} times a constant times the largest relevant box count. In a box, small coefficients are fixed and the three letters have the form

a0∏i=1dxa,ii,b0∏i=1dxb,ii,c0∏i=1dxc,ii,xs,i∼Ps,i.a_0\prod_{i=1}^{d}x_{a,i}^{i},\qquad b_0\prod_{i=1}^{d}x_{b,i}^{i},\qquad c_0\prod_{i=1}^{d}x_{c,i}^{i},\qquad x_{s,i}\sim P_{s,i}.

The equation is

a0∏ixa,ii+b0∏ixb,ii=c0∏ixc,ii.(2)a_0\prod_i x_{a,i}^{i}+b_0\prod_i x_{b,i}^{i}=c_0\prod_i x_{c,i}^{i}. \tag*{(2)}

The small coefficients are pairwise coprime and at most Xε0X^{\varepsilon_0}. We may restrict to solutions whose three displayed letter values are pairwise coprime: in a primitive triple a+b=ca+b=c, a common divisor of any two letters divides the third. Discarding further within-letter restrictions can only enlarge an upper bound.

The dyadic parameters obey the radical and height budgets

2−d−1Xμ≤∏s,iPs,i≤Xμ+3ε0/4,∏iPs,ii≤X,(3)2^{-d-1}X^\mu\le\prod_{s,i}P_{s,i}\le X^{\mu+3\varepsilon_0/4},\qquad\prod_i P_{s,i}^{i}\le X, \tag*{(3)}

and the original third letter satisfies ∏iPc,ii≥X1−ε0/2\prod_i P_{c,i}^{i}\ge X^{1-\varepsilon_0/2}. Here 0<μ≤λ0<\mu\le\lambda is the radical budget of the restricted count. All later estimates are uniform over boxes and frozen coefficients once dd is fixed.

Set

Ps,i=Xαs,i,αs=∑i=1dαs,i,L=αa+αb+αc.(4)P_{s,i}=X^{\alpha_{s,i}},\qquad\alpha_s=\sum_{i=1}^{d}\alpha_{s,i},\qquad L=\alpha_a+\alpha_b+\alpha_c. \tag*{(4)}

and write the nonzero box count as XDX^D. The quantities αs,i\alpha_{s,i}, αs\alpha_s and DD depend on the box, but we suppress this in the notation. Errors below are O(ε0)+o(1)O(\varepsilon_0)+o(1) as X→∞X\to\infty, with dd fixed.

The four inherited estimates

The limiting exponent system uses only coordinates 1≤i≤61\le i\le6 explicitly; the remaining tail is retained through αs−∑i≤6αs,i\alpha_s-\sum_{i\le6}\alpha_{s,i}. All variables are nonnegative.

Proposition 2.1 (Inherited height and radical rows). Every anatomic box satisfies, up to the stated limiting errors,

L≤λ,∑i=16αs,i≤αs,7αs−∑i=16(7−i)αs,i≤1(s∈{a,b,c}).(5)L\le\lambda,\qquad\sum_{i=1}^{6}\alpha_{s,i}\le\alpha_s,\qquad7\alpha_s-\sum_{i=1}^{6}(7-i)\alpha_{s,i}\le1\quad(s\in\{a,b,c\}). \tag*{(5)}

The last row follows from ∑i≤diαs,i≤1\sum_{i\le d}i\alpha_{s,i}\le1 by using i≥7i\ge7 in the tail.

Proposition 2.2 (Inherited pair count). One has

D≤αa+αb,D≤αa+αc,D≤αb+αc.(6)D\le\alpha_a+\alpha_b,\qquad D\le\alpha_a+\alpha_c,\qquad D\le\alpha_b+\alpha_c. \tag*{(6)}

Proposition 2.3 (Inherited Fourier estimate). Define

Fs=max⁡{αs,2+αs,4+αs,6, αs,3+αs,6, αs,5}.(7)F_s=\max\{\alpha_{s,2}+\alpha_{s,4}+\alpha_{s,6},\ \alpha_{s,3}+\alpha_{s,6},\ \alpha_{s,5}\}. \tag*{(7)}

Then

Fa+Fb+Fc≤4L−6D.(8)F_a+F_b+F_c\le4L-6D. \tag*{(8)}

Proposition 2.4 (Inherited geometric estimate). For arbitrary Ia,Ib,Ic⊂{1,…,6}I_a,I_b,I_c\subset\{1,\ldots,6\}, put

Σ0=∑s∈{a,b,c}∑i∈Isαs,i,Σ1=∑s∈{a,b,c}∑i∈Is(i−1)αs,i.(9)\Sigma_0=\sum_{s\in\{a,b,c\}}\sum_{i\in I_s}\alpha_{s,i},\qquad\Sigma_1=\sum_{s\in\{a,b,c\}}\sum_{i\in I_s}(i-1)\alpha_{s,i}. \tag*{(9)}

Then

D≤max⁡{L−Σ0, L−1+Σ1}.D\le\max\{L-\Sigma_0,\ L-1+\Sigma_1\}.

For proofs of (6)–(9), see [1 Propositions 2.3, 3.1 and 4.1]. In the last two estimates the numerator initially contains the actual product ∏s,iPs,i\prod_{s,i} P_{s,i}, before any replacement by the radical budget. This gives LL, rather than merely λ\lambda, in (8) and (9). The factor X−1+O(ε0)X^{-1+O(\varepsilon_0)} in the second geometric alternative comes from the height of the original third letter and is independent of the radical budget. Thus λ\lambda enters these limiting rows only through L≤λL \le\lambda.

For computation, each FsF_s is a nonnegative auxiliary variable bounded below by the three expressions in (7). The complete inequality system is invariant under permutations of the letters. We may therefore impose the symmetry-breaking order

αa,4+2αa,5≥αb,4+2αb,5≥αc,4+2αc,5.(10)\alpha_{a,4}+2\alpha_{a,5} \ge\alpha_{b,4}+2\alpha_{b,5} \ge\alpha_{c,4}+2\alpha_{c,5}. \tag*{(10)}

After relabelling, every feasible point has a representative satisfying (10). The geometric estimate remains valid after such a permutation: its final displayed expression is symmetric, although its proof uses the height of the original third letter.

Uniform power differences and twisted energy

The two arithmetic lemmas in this section are inherited from [3 Sections 3.1–3.2 and Appendix B]. Their proofs are included because the new certificate uses them in degrees and modulus configurations not exhausted there.

Weighted differences of arbitrary fixed degree

Lemma 3.1 (Inherited uniform count; proof reproduced). Fix an integer e≥2e \ge2 and positive real numbers C,ηC,\eta. For T≥2T \ge2, positive integers aa, bb and a nonzero integer mm, the number of positive pairs (U,V)(U,V) satisfying

aUe−bVe=m,a,b,U,V,∣m∣≤TC(11)aU^e-bV^e=m,\qquad a,b,U,V,\lvert m\rvert\le T^C \tag*{(11)}

is Oe,C,η(Tη)O_{e,C,\eta}(T^\eta), uniformly in a,b,ma,b,m. No coprimality hypothesis is required.

Proof. Multiply by ae−1a^{e-1} and put

t=ae−1b,H=ae−1m,θ=t1/e>0,K=Q(θ),n=[K:Q]≤e.t=a^{e-1}b,\qquad H=a^{e-1}m,\qquad\theta=t^{1/e}>0,\qquad K=\mathbb{Q}(\theta),\qquad n=[K:\mathbb{Q}]\le e.

Then θ\theta is an algebraic integer and

(aU)e−tVe=H≠0.(aU)^e-tV^e=H\ne0.

Let ff be the monic minimal polynomial of θ\theta. Since f∈Z[Z]f\in\mathbb{Z}[Z], monic division in Z[Z]\mathbb{Z}[Z] gives

Ze−t=f(Z)g(Z),g∈Z[Z],deg⁡g=e−n.Z^e-t=f(Z)g(Z),\qquad g\in\mathbb{Z}[Z],\qquad\deg g=e-n.

For the algebraic integer α=aU−θV\alpha=aU-\theta V, define

z:=NK/Q(α)=Vnf(aU/V).z:=N_{K/\mathbb{Q}}(\alpha)=V^n f(aU/V).

Evaluating the factorization of Ze−tZ^e-t yields

H=zVe−ng(aU/V).(12)H=zV^{e-n}g(aU/V). \tag*{(12)}

Both factors on the right are integers. Since H≠0H\ne0, the norm zz is a nonzero divisor of HH.

All quantities in (11) have polynomial size in TT. Hence there is C1=C1(e,C)C_1=C_1(e,C) such that every embedding σ:K↪C\sigma:K\hookrightarrow\mathbb{C} satisfies ∣σ(α)∣≤TC1\lvert\sigma(\alpha)\rvert\le T^{C_1}. The product of these absolute values is the nonzero integer ∣z∣\lvert z\rvert, so

T−C1(n−1)≤∣σ(α)∣≤TC1.(13)T^{-C_1(n-1)}\le\lvert\sigma(\alpha)\rvert\le T^{C_1}. \tag*{(13)}

Moreover, the integral ideal (α)(\alpha) has norm ∣z∣\lvert z\rvert dividing ∣H∣\lvert H\rvert.

We count possible ideals uniformly in KK. Above a rational prime pp there are at most ee prime ideals. If ph∥Hp^h\Vert H, an ideal whose norm divides HH has exponent at most hh at each of these primes. Unique factorization of ideals therefore gives at most

∏ph∣H(h+1)e=τ(∣H∣)e(14)\prod_{p^h\mid H}(h+1)^e=\tau(\lvert H\rvert)^e \tag*{(14)}

integral ideals whose norms divide ∣H∣|H|. This bound has no discriminant-dependent constant. Since ∣H∣≤Te|H| \leq T^{e}, the divisor bound makes (14) at most Oe,C,η(Tη/2)O_{e,C,\eta}(T^{\eta/2}).

Fix one generator α0\alpha_0 arising from a counted solution for each principal ideal that occurs. Every other counted generator of that ideal is α=ϵα0\alpha=\epsilon\alpha_0 for a unit ϵ\epsilon. By (13), the vector (log⁡∣σ(ϵ)∣)σ(\log|\sigma(\epsilon)|)_\sigma lies in a cube of side Oe,C(log⁡T)O_{e,C}(\log T).

We need a separation constant uniform over fields of degree at most ee. Algebraic integers of degree at most ee whose conjugates all have modulus at most exp⁡(1)\exp(1) form a finite set: their monic minimal polynomials have integer coefficients bounded in terms of ee. If all conjugates of an algebraic integer have modulus 11, all its positive powers belong to this same finite set, so it is a root of unity. Consequently there is ce>0c_e>0 such that every non-torsion unit of degree at most ee has

max⁡σ∣log⁡∣σ(ϵ)∣∣≥ce.\max_\sigma\left|\log|\sigma(\epsilon)|\right| \geq c_e.

Outside the finite set a conjugate already has modulus greater than exp⁡(1)\exp(1); inside it, take the least positive distance. The number of roots of unity of degree at most ee is bounded by the same finiteness argument.

Apply this separation to the quotient of two units occurring in KK. Distinct unit logarithmic vectors are at distance at least cec_e in the supremum norm. Thus cubes of side ce/2c_e/2 contain at most one vector, and each vector has a bounded number of unit preimages. One ideal therefore contributes Oe,C((1+log⁡T)e)O_{e,C}((1+\log T)^e) generators.

If n≥2n\geq2, then θ\theta is irrational and (U,V)↦aU−θV(U,V)\mapsto aU-\theta V is injective. The ideal and generator counts now prove the lemma after absorbing the logarithm into Tη/2T^{\eta/2}.

Finally, if n=1n=1, then θ\theta is a positive integer. The nonzero integer s=aU−θVs=aU-\theta V divides HH. For each signed divisor ss, substitution gives

(θV+s)e−(θV)e=H.(\theta V+s)^e-(\theta V)^e=H.

This is a nonzero polynomial equation of degree e−1e-1, with leading coefficient eθe−1se\theta^{e-1}s. It has at most e−1e-1 roots, each determining UU. The divisor bound for HH completes the proof. □

Corollary 3.2 (Grouped values). Let ∅≠I⊆{1,…,6}\varnothing\ne I\subseteq\{1,\ldots,6\} and xi,xi′∼Xαix_i,x'_i\sim X^{\alpha_i} for i∈Ii\in I. For fixed nonzero mm of polynomial size in XX, the number of pairs of values

v=∏i∈Ixii,v′=∏i∈I(xi′)i,v−v′=mv=\prod_{i\in I}x_i^i,\qquad v'=\prod_{i\in I}(x'_i)^i,\qquad v-v'=m

is at most Xδ+o(1)X^{\delta+o(1)}, where δ\delta may be the minimum of

δ0=∑i∈Iαi,δe=2∑i∈Ie∣iαi(e≥2 fixed).(15)\delta_0=\sum_{i\in I}\alpha_i,\qquad\delta_e=2\sum_{\substack{i\in I\\ e\mid i}}\alpha_i\quad(e\geq2\text{ fixed}). \tag*{(15)}

Proof. The bound δ0\delta_0 counts vv, after which v′v' is determined. For δe\delta_e, freeze both copies of every coordinate with e∤ie\nmid i. The remaining products have the form UeU^e and VeV^e, so the equation becomes aUe−bVe=maU^e-bV^e=m. Lemma 3.1 applies uniformly, and the frozen choices cost δe\delta_e. A value of UU has at most Xo(1)X^{o(1)} coordinate representations, since every coordinate divides UU; the same holds for VV. If no index in II is divisible by ee, then δe=2δ0\delta_e=2\delta_0 and this alternative is redundant. □

The numerical certificate below uses δe\delta_e precisely for e=2,3,4,5,6e=2,3,4,5,6. The new feature is this systematic use, not Lemma 3.1 itself.

Energy with an arbitrary unit twist

Lemma 3.3 (Inherited energy estimate; proof reproduced). Fix positive integers j,kj,k. Let M≥1M\geq1, let γ\gamma be a unit modulo MM, and let P,Q≥1P,Q\geq1. The number KγK_\gamma of unit-coordinate solutions to

xjyk≡γ(x′)j(y′)k(modM),x,x′∼P,y,y′∼Q,(16)x^j y^k\equiv\gamma(x')^j(y')^k\pmod M,\qquad x,x'\sim P,\quad y,y'\sim Q, \tag*{(16)}

satisfies, for every η>0\eta>0,

Kγ≪j,k,η(PQM)ηPQ(1+P2/M)(1+Q2/M).(17)K_\gamma\ll_{j,k,\eta}(PQM)^\eta PQ\sqrt{(1+P^2/M)(1+Q^2/M)}. \tag*{(17)}

In particular, no coprimality condition on j,kj,k is needed.

Proof. Let G=(Z/MZ)×G=(\mathbb{Z}/M\mathbb{Z})^{\times}, interpreted as the one-element group when M=1M=1. For g∈Gg\in G, let RP(g)R_P(g) count pairs (x,x′)(x,x') with x/x′=gx/x'=g, and define

Sj(t)=∑gj=tRP(g),Sk(t)=∑hk=tRQ(h).S_j(t)=\sum_{g^j=t}R_P(g),\qquad S_k(t)=\sum_{h^k=t}R_Q(h).

Writing g=x/x′g=x/x' and h=y/y′h=y/y', the congruence becomes gjhk=γg^jh^k=\gamma. Therefore

Kγ=∑t∈GSj(t)Sk(γ/t).K_\gamma=\sum_{t\in G}S_j(t)S_k(\gamma/t).

Since t↦γ/tt\mapsto\gamma/t permutes GG, Cauchy–Schwarz gives

Kγ2≤∥Sj∥22∥Sk∥22≤ρj(M)ρk(M)MPMQ.(18)K_\gamma^2\leq\lVert S_j\rVert_2^2\lVert S_k\rVert_2^2\leq\rho_j(M)\rho_k(M)\mathcal{M}_P\mathcal{M}_Q. \tag*{(18)}

where MP=∑gRP(g)2\mathcal{M}_P=\sum_g R_P(g)^2 and ρj(M)\rho_j(M) is the kernel size of g↦gjg\mapsto g^j. The second inequality applies Cauchy–Schwarz on each fibre of the power map.

For an odd prime power the unit group is cyclic, and for a power of 2 it is a product of at most two cyclic groups. Hence

ρj(M)≤j2ω(M)≪j,ηMη,\rho_j(M)\leq j^{2\omega(M)}\ll_{j,\eta}M^\eta,

including M=1M=1. Next, MP\mathcal{M}_P counts xt′≡x′t(modM)xt'\equiv x't\pmod M with all four coordinates in the interval of size PP. Fixing (x,t′)(x,t') gives O(P2)O(P^2) choices. The product x′tx't then lies in one residue class in [1,4P2][1,4P^2], so it takes O(1+P2/M)O(1+P^2/M) integer values. Each value has Po(1)P^{o(1)} factorizations. Thus

MP≪ηP2+η(1+P2/M).\mathcal{M}_P\ll_\eta P^{2+\eta}(1+P^2/M).

The analogous estimate for MQ\mathcal{M}_Q, inserted into (18), proves (17) after reducing the intermediate error exponents. The argument is uniform in γ\gamma. □

If P=XpP=X^p, Q=XqQ=X^q, M≥XwM\geq X^w and M≤XO(1)M\leq X^{O(1)}, then (17) has exponent

κ(p,q;w)=p+q+12((2p−w)++(2q−w)+).(19)\kappa(p,q;w)=p+q+\frac{1}{2}\bigl((2p-w)_++(2q-w)_+\bigr). \tag*{(19)}

Only the lower bound for MM is used in the two main factors; the polynomial upper bound absorbs MηM^\eta.

The collision inequality with a third-letter modulus

We now reproduce the collision formula of [3 Section 4] in the generality needed later. The formula is an inherited arithmetic input; its exhaustive degree-and-modulus deployment and the resulting certificates are new.

Fix distinct letters A,B,CA,B,C. Choose nonempty IA,IB⊆{1,…,6}I_A,I_B\subseteq\{1,\ldots,6\}, a pair j<kj<k in IAI_A, and a possibly empty set OC⊆{1,…,6}O_C\subseteq\{1,\ldots,6\}. Define

sA=∑i∈IAαA,i,sB=∑i∈IBαB,i,p=αA,j,q=αA,k,r=sA−p−q.(20)s_A=\sum_{i\in I_A}\alpha_{A,i},\qquad s_B=\sum_{i\in I_B}\alpha_{B,i},\qquad p=\alpha_{A,j},\qquad q=\alpha_{A,k},\qquad r=s_A-p-q. \tag*{(20)}
fA=αA−sA,fB=αB−sB,fC=∑i∈OCαC,i,wC=∑i∈OCiαC,i.(21)f_A=\alpha_A-s_A,\qquad f_B=\alpha_B-s_B,\qquad f_C=\sum_{i\in O_C}\alpha_{C,i},\qquad w_C=\sum_{i\in O_C}i\alpha_{C,i}. \tag*{(21)}
wB=∑i≤6i∉IBiαB,i+7(αB−∑i≤6αB,i),hB=∑i∈IBiαB,i.(22)w_B=\sum_{\substack{i\leq6\\i\notin I_B}}i\alpha_{B,i}+7\left(\alpha_B-\sum_{i\leq6}\alpha_{B,i}\right),\qquad h_B=\sum_{i\in I_B}i\alpha_{B,i}. \tag*{(22)}

The energy exponent is

K=2r+κ(p,q;wB).(23)K=2r+\kappa(p,q;w_B). \tag*{(23)}

with κ\kappa as in (19). In addition to δ0=sB\delta_0=s_B and the grouped exponents δe\delta_e on letter BB, define

δcoord=min⁡t∈IB{sB−αB,t+(αB,t−wC)+},δvalue=(hB−wC)+.(24)\delta_{\mathrm{coord}}=\min_{t\in I_B}\{s_B-\alpha_{B,t}+(\alpha_{B,t}-w_C)_+\},\qquad\delta_{\mathrm{value}}=(h_B-w_C)_+. \tag*{(24)}

Proposition 4.1 (Inherited collision inequality; proof reproduced). Let δ\delta be the minimum of any nonempty subcollection of δ0\delta_0, δe\delta_e (2≤e≤62 \le e \le6), δcoord\delta_{\mathrm{coord}} and δvalue\delta_{\mathrm{value}}. Then every anatomic box satisfies

D≤fA+fB+fC+max⁡{αC−fC,αC−fC+K+δ2}+o(1).(25)D \le f_A+f_B+f_C+\max\left\{\alpha_C-f_C,\frac{\alpha_C-f_C+K+\delta}{2}\right\}+o(1). \tag*{(25)}

The estimate is uniform over frozen coefficients for fixed dd. Separate savings are combined by taking their minimum, never by adding them.

Proof. Freeze all coordinates of AA outside IAI_A, all coordinates of BB outside IBI_B, and the coordinates of CC in OCO_C. The first two frozen sets include every index above 6. There are at most XfA+fB+fC+o(1)X^{f_A+f_B+f_C+o(1)} choices. In one fixed fibre, write the equation as

ϵAMAu+ϵBMBv=ϵCMCw,ϵA,ϵB,ϵC∈{1,−1},(26)\epsilon_A M_Au+\epsilon_B M_Bv=\epsilon_C M_Cw,\qquad\epsilon_A,\epsilon_B,\epsilon_C\in\{1,-1\}, \tag*{(26)}

where uu, vv are the free monomials and ww is the product of the remaining coordinates of CC. The fixed small coefficient of each letter is included in its MM, and all three MM are positive. By (22) and the weight at least 7 of every tail coordinate,

MB≥XwB,MC≥XwC,M_B\ge X^{w_B},\qquad M_C\ge X^{w_C},

and all quantities are at most XOd(1)X^{O_d(1)}.

Each positive monomial value has Xo(1)X^{o(1)} coordinate representations: each of the bounded number of coordinates divides the value. It therefore suffices to count pairs of values (u,v)(u,v), restoring this factor at the end. Let W\mathcal{W} be the set of possible ww values, so

∣W∣≤XαC−fC+o(1).|\mathcal{W}|\le X^{\alpha_C-f_C+o(1)}.

For w∈Ww\in\mathcal{W}, let nwn_w count permitted pairs (u,v)(u,v), and put N=∑wnwN=\sum_w n_w. If EoffE_{\mathrm{off}} counts ordered collisions between distinct pairs with the same ww, Cauchy–Schwarz on the fibres gives

N2≤∣W∣∑wnw2=∣W∣(N+Eoff),N≪∣W∣+∣W∣Eoff.(27)N^2\le|\mathcal{W}|\sum_w n_w^2=|\mathcal{W}|(N+E_{\mathrm{off}}),\qquad N\ll|\mathcal{W}|+\sqrt{|\mathcal{W}|E_{\mathrm{off}}}. \tag*{(27)}

A nonempty fibre has gcd⁡(MA,MB)=1\gcd(M_A,M_B)=1, and every coordinate of uu is a unit modulo MBM_B by pairwise coprimality of the original letters. Subtracting two equations (26) with the same ww gives

MA(u−u′)=±MB(v−v′).M_A(u-u')=\pm M_B(v-v').

For distinct pairs, both differences are nonzero: if either vanished, the equation would force the other to vanish. In particular u≡u′u\equiv u' (mod MBM_B).

Fix independently the two copies of the coordinates in IA∖{j,k}I_A\setminus\{j,k\}, at a cost X2r+o(1)X^{2r+o(1)}. Since these coordinates are units modulo MBM_B, the remaining congruence is exactly (16) for some unit γ\gamma. Lemma 3.3 is uniform in that unit. Thus the number of possible pairs (u,u′)(u,u') is at most XK+o(1)X^{K+o(1)}, with KK as in (23).

Now fix such a pair (u,u′)(u,u'). The nonzero integer

m=v−v′=±MA(u−u′)/MBm=v-v'=\pm M_A(u-u')/M_B

is fixed. Counting all possible vv gives XsB+o(1)X^{s_B+o(1)}, since v′v' is then determined. Corollary 3.2 gives each grouped bound Xδe+o(1)X^{\delta_e+o(1)}.

There are two residue-class alternatives. In a nonempty fibre, MBM_B is invertible modulo MCM_C. Reducing (26) modulo MCM_C determines the residue class of vv from uu. Also

1≤v≤2∣IB∣XhB.1\le v\le2^{|I_B|}X^{h_B}.

A single residue class in this interval has Od(1+XhB/MC)O_d(1+X^{h_B}/M_C) elements. Restoring divisor-sized coordinate multiplicity gives Xδvalue+o(1)X^{\delta_{\mathrm{value}}+o(1)}. This is a whole-value interval count; it remains valid when an individual coordinate range is larger than the modulus.

Alternatively, fix every coordinate of vv except the one indexed by t∈IBt\in I_B. The congruence becomes xtt≡ξ(modMC)x_t^t\equiv\xi\pmod{M_C} for a unit ξ\xi. Its solutions occupy at most ρt(MC)=Xo(1)\rho_t(M_C)=X^{o(1)} residue classes, by the kernel estimate in Lemma 3.3. Each class contributes O(1+XcB,t/MC)O(1+X^{c_{B,t}}/M_C) integers in the dyadic interval. The other fixed coordinates cost XsB−αB,t+o(1)X^{s_B-\alpha_{B,t}+o(1)}. Taking the minimum over tt proves the coordinate bound. The case MC=1M_C=1 is included.

All these estimates count the same conditional set of pairs (v,v′)(v,v'), so their minimum is valid. Consequently

Eoff≤XK+δ+o(1).E_{\mathrm{off}} \le X^{K+\delta+o(1)}.

Insert this into (27), restore frozen choices and coordinate multiplicities, and take logarithms base XX. The two alternatives in (27) give respectively the diagonal and off-diagonal expressions in (25). No coprimality assumption between different solutions is used. □

Lemma 4.2 (Release of all first-letter coordinates). For fixed jj, kk, IBI_B, OCO_C and a fixed choice of admissible savings, the strongest member of (25) is obtained with IA={1,…,6}I_A=\{1,\ldots,6\}.

Proof. Adding a coordinate of exponent α\alpha to IAI_A subtracts α\alpha from fAf_A and adds α\alpha to rr, hence adds 2α2\alpha to KK. The off-diagonal expression is unchanged, while the diagonal expression decreases by α\alpha. □

The older pairwise difference refinements are repeated instances of Corollary 3.2. For example, IB={2,3}I_B=\{2,3\} gives 2αB,32\alpha_{B,3} and 2αB,22\alpha_{B,2}; IB={3,4}I_B=\{3,4\} gives 2αB,32\alpha_{B,3} and 2αB,42\alpha_{B,4}; and each of {2,5}\{2,5\}, {3,5}\{3,5\}, {4,5}\{4,5\} gives 2αB,52\alpha_{B,5}. Retaining these labels in the data does not require another arithmetic theorem.

Lemma 3.3 permits all 15 pairs 1≤j<k≤61\le j<k\le6. The upper tree in this paper uses only the 11 coprime pairs, which are sufficient for the stated bound. Exact lower-obstruction checks use all 15 pairs. Thus noncoprime pairs are a legal extension, not a hidden hypothesis and not a claimed source of the upper certificate.

Complete modulus locking and elimination of index 1

Throughout the finite programs, Lemma 4.2 is used to set IA={1,…,6}I_A=\{1,\ldots,6\}. The remaining collision data are an ordered pair of distinct letters (A,B)(A,B), an exponent pair (j,k)(j,k), a nonempty IB⊆{1,…,6}I_B\subseteq\{1,\ldots,6\}, and a third-letter modulus set OCO_C.

The old modulus family is

Oold={∅,{1},{2},{3},{4},{5},{6},{2,5},{3,5}}.(28)O_{\mathrm{old}}=\{\varnothing,\{1\},\{2\},\{3\},\{4\},\{5\},\{6\},\{2,5\},\{3,5\}\}. \tag*{(28)}

FullOC starts with all 64 subsets of {1,…,6}\{1,\ldots,6\}. We prove that, for the whole-value lock and the grouped savings used here, it suffices to retain

OFullOC={OC:OC⊆{2,3,4,5,6}},∣OFullOC∣=32.(29)O_{\mathrm{FullOC}}=\{O_C:O_C\subseteq\{2,3,4,5,6\}\},\qquad|O_{\mathrm{FullOC}}|=32. \tag*{(29)}

The deletion inequality

Fix AA, BB, CC, jj, kk, IBI_B and put

δ∗=min⁡{δ0,δ2,δ3,δ4,δ5,δ6}.\delta_*=\min\{\delta_0,\delta_2,\delta_3,\delta_4,\delta_5,\delta_6\}.

This quantity is independent of OCO_C. For any OC⊆{1,…,6}O_C\subseteq\{1,\ldots,6\}, define

H(OC)=fC+min⁡{δ∗,(hB−wC)+}.(30)H(O_C)=f_C+\min\{\delta_*,(h_B-w_C)_+\}. \tag*{(30)}

The diagonal expression in (25) is

fA+fB+fC+αC−fC=fA+fB+αC,f_A+f_B+f_C+\alpha_C-f_C=f_A+f_B+\alpha_C,

so it is independent of OCO_C. With δ=min⁡{δ∗,δvalue}\delta=\min\{\delta_*,\delta_{\mathrm{value}}\}, the off-diagonal expression is

fA+fB+αC+K2+H(OC)2.f_A+f_B+\frac{\alpha_C+K}{2}+\frac{H(O_C)}{2}.

Thus only H(OC)H(O_C) matters when comparing modulus sets.

Proposition 5.1 (Index-1 elimination). If 1∈OC1\in O_C and OC′=OC∖{1}O'_C=O_C\setminus\{1\}, then

H(OC′)≤H(OC).(31)H(O'_C)\le H(O_C). \tag*{(31)}

Consequently, the collision row associated with OC′O'_C is at least as strong as that associated with OCO_C. In the whole-value lock subfamily with savings δ0,δ2,…,δ6\delta_0,\delta_2,\ldots,\delta_6, every set containing 1 may be discarded.

Proof. Write α=αC,1≥0\alpha=\alpha_{C,1}\ge0. Deleting index 1 gives

fC′=fC−α,wC′=wC−α,hB−wC′=hB−wC+α.f'_{C}=f_{C}-\alpha,\qquad w'_{C}=w_{C}-\alpha,\qquad h_{B}-w'_{C}=h_{B}-w_{C}+\alpha.

The function

t↦ϕ(t)=min⁡{δ∗,(t)+}t\mapsto\phi(t)=\min\{\delta_{*},(t)_{+}\}

is nondecreasing and 1-Lipschitz: it is 0 for t≤0t\le0, then tt up to δ∗\delta_{*}, and then constant. Therefore

ϕ(hB−wC′)≤ϕ(hB−wC)+α,\phi(h_{B}-w'_{C})\le\phi(h_{B}-w_{C})+\alpha,

and hence

H(OC′)=fC−α+ϕ(hB−wC′)≤fC+ϕ(hB−wC)=H(OC).H(O'_{C})=f_{C}-\alpha+\phi(h_{B}-w'_{C})\le f_{C}+\phi(h_{B}-w_{C})=H(O_{C}).

The diagonal row is unchanged and the off-diagonal row is monotone in HH, proving the final assertion. □

Applying Proposition 5.1 once to every set containing 1 leaves exactly the 32 sets in (29). This is an exhaustive treatment of the whole-value modulus family, not a heuristic list of promising sets.

Remark 5.2 (Scope of the elimination). The proof uses the independence of δ∗\delta_{*} from OCO_{C} and the one-Lipschitz whole-value expression. It does not automatically eliminate index 1 in a different program whose selected savings include arbitrary coordinate locks or additional modulus-dependent terms. The word “complete” below always refers to the subfamily specified in Proposition 5.1.

The finite collision families

For nonempty OCO_{C}, the tree imposes simultaneously the affine rows coming from

δvalue,δ0,δe(2≤e≤6).\delta_{\mathrm{value}},\qquad\delta_{0},\qquad\delta_{e}\quad(2\le e\le6).

On a fixed sign octant, each is a valid upper bound for the same minimum, so imposing all resulting rows is legitimate. For OC=∅O_{C}=\varnothing, one has wC=0w_{C}=0 and

δcoord=min⁡t∈IB{sB−αB,t+αB,t}=δ0.\delta_{\mathrm{coord}}=\min_{t\in I_{B}}\{s_{B}-\alpha_{B,t}+\alpha_{B,t}\}=\delta_{0}.

The tree uses a uniform coordinate-sign encoding of this fact and additionally retains the older pair bounds and δ2,…,δ6\delta_{2},\ldots,\delta_{6}.

For exact program comparisons, we use all 15 exponent pairs. The upper certificate uses only the 11 coprime pairs. Table 2 counts the collision instances before their diagonal and sign-octant disjunctions; there are six ordered letter pairs and 63 nonempty choices of IBI_{B}.

FamilyExponent pairsModulus setsInstances
Old-family control1596⋅15⋅63⋅9=51,0306\cdot15\cdot63\cdot9=51{,}030
FullOC, all legal pairs15326⋅15⋅63⋅32=181,4406\cdot15\cdot63\cdot32=181{,}440
FullOC, upper tree11 coprime pairs326⋅11⋅63⋅32=133,0566\cdot11\cdot63\cdot32=133{,}056

Table 2. Collision-instance counts in the exact programs.

The old and FullOC control programs below share the base rows, all 64364^{3} geometric choices, released first-letter coordinates, grouped savings and all 15 exponent pairs. They differ only in the modulus family (28) versus (29). Since the upper tree uses a subfamily of FullOC, its upper bound also applies to the all-pair FullOC program.

Piecewise-linear encoding and exact rational certificates

The 25-variable root system

The vector x≥0x\ge0 consists of

(αs,i)s∈{a,b,c}, 1≤i≤6,(αa,αb,αc),(Fa,Fb,Fc),D,(\alpha_{s,i})_{s\in\{a,b,c\},\,1\le i\le6},\qquad(\alpha_{a},\alpha_{b},\alpha_{c}),\qquad(F_{a},F_{b},F_{c}),\qquad D,

for 18+3+3+1=2518+3+3+1=25 variables. The root has the 22 affine rows encoding (5)–(8) and (10). The three expressions in each Fourier maximum are imposed separately as lower bounds on FsF_{s}. The geometric and collision inequalities are disjunctive and are introduced by the tree.

Branch semantics

A geometric node selects (Ia,Ib,Ic)(I_a,I_b,I_c) in (9) and has exactly two children, carrying respectively

D≤L−Σ0,D≤L−1+Σ1.D \le L-\Sigma_0,\qquad D \le L-1+\Sigma_1.

Thus every point satisfying the selected geometric inequality belongs to at least one child.

A collision node selects (A,B,C,j,k,IB,OC)(A,B,C,j,k,I_B,O_C) and a distinguished kB∈IBk_B\in I_B used only for the uniform empty-modulus encoding. It has one diagonal child and eight off-diagonal children. The diagonal child carries

D≤fA+fB+αC.D \le f_A+f_B+\alpha_C.

The eight children cover the closed sign octants of

2p−wB,2q−wB,{hB−wC,OC≠∅,αB,kB,OC=∅.2p-w_B,\qquad2q-w_B,\qquad \begin{cases} h_B-w_C, & O_C\ne\varnothing,\\ \alpha_{B,k_B}, & O_C=\varnothing. \end{cases}

On each octant, κ\kappa, the whole-value lock and the coordinate lock are affine. The corresponding off-diagonal rows for every admissible saving are imposed together. This is valid because Proposition 4.1 permits the minimum of those savings; it is not a choice of a favorable saving after seeing the point. In the empty-modulus case the nonpositive branch of αB,kB\alpha_{B,k_B} is redundant except on its zero boundary, which is harmless.

Every geometric or collision node therefore has complete child coverage. A path from the root collects affine rows Aℓx≤bℓA_\ell x\le b_\ell at its leaf ℓ\ell.

Rational leaf proofs

Lemma 6.1 (Leaf multiplier criterion). Suppose nonnegative rational multipliers yℓy_\ell satisfy

(yℓTAℓ)D=1,(yℓTAℓ)v≥0(v≠D),yℓTbℓ=βℓ.(32)(y_\ell^T A_\ell)_D=1,\qquad(y_\ell^T A_\ell)_v\ge0\quad(v\ne D),\qquad y_\ell^T b_\ell=\beta_\ell. \tag*{(32)}

Then every nonnegative xx satisfying the leaf rows obeys D≤βℓD\le\beta_\ell.

Proof. Multiplying the rows by yℓy_\ell gives

D+∑v≠D(yℓTAℓ)vxv=yℓTAℓx≤yℓTbℓ=βℓ.D+\sum_{v\ne D}(y_\ell^T A_\ell)_v x_v=y_\ell^T A_\ell x\le y_\ell^T b_\ell=\beta_\ell.

All summands on the left are nonnegative by (32) and x≥0x\ge0. □

Floating-point linear programming is used only to select violated branches and discover candidate duals. Candidate multipliers are converted to rational numbers and repaired, if necessary, before storage. Neither an LP status nor a displayed numerical optimum is a proof step. The proof is (32), complete child coverage and Lemma 6.1.

The FullOC upper certificate

Theorem 6.2 (Exact finite implication). Every nonnegative point satisfying the root system, all geometric disjunctions, and the FullOC collision disjunctions with IA={1,…,6}I_A=\{1,\ldots,6\}, the 11 coprime exponent pairs, all 32 modulus sets and δ0,δ2,…,δ6\delta_0,\delta_2,\ldots,\delta_6 satisfies

D≤U=1168000720000012.D\le U=\frac{11680007}{20000012}.

The same conclusion holds for the program enlarged to all 15 exponent pairs.

Computer-assisted proof. The certified tree has the statistics in Table 3. Each leaf stores positive rational multipliers satisfying Lemma 6.1. The exact maximum of its leaf bounds is UU. Since every feasible point belongs to a leaf, this proves the first assertion. Enlarging the exponent-pair family only restricts the feasible set, so the second assertion follows. □

Radical budgetλ=1\lambda= 1
Requested target73/12573/125
Exact maximum leaf bound11680007/2000001211680007/20000012
Leaves65,576
Geometric nodes4,591
Collision nodes7,623
Maximum depth36
LP calls during discovery466,244
Discovery time in the recorded run1965.523944 s

Table 3. Metadata of the main FullOC certificate.

The exact target margin is

73125−U=1250001500>0.\frac{73}{125}-U=\frac{1}{250001500}>0.

The certificate uses δe\delta_e for all e=2,…,6e=2,\ldots,6; its machine-data mode is all_deg.

Independent verification

A separate checker, using only the standard library, reconstructs the root, geometric, collision, sign-octant, whole-value lock and higher-degree rows directly from the formulas. It does not import the search program or its engine. For every node it checks the schema and complete child list; for every leaf it recomputes the dot products in (32) with rational arithmetic; and at the end it recomputes the leaf count and the exact maximum bound.

The recorded run returned

upper=11680007/20000012,leaf=65576,geo=4591,ss=7623,\text{upper} = 11680007/20000012,\qquad\text{leaf} = 65576,\qquad\text{geo} = 4591,\qquad\text{ss} = 7623,

followed by VERIFIED. Three deliberate faults—a changed multiplier, a deleted child and an altered declared maximum—were all rejected. Appendix A gives the command, file hashes and obstruction-checking routines.

Passage from the exponent program to Nλ(X)N_\lambda(X)

We now prove Theorem 1.1. The purpose of this limiting argument is to account explicitly for decomposition errors, divisor losses and the finite but box-dependent nature of the estimates.

Let Nμ∗(Y)N_\mu^*(Y) count primitive triples with

Y/2≤c≤Y,cμ/2≤rad⁡(abc)<cμ.Y/2 \le c \le Y,\qquad c^\mu/2 \le\operatorname{rad}(abc) < c^\mu.

Dyadic decomposition in height and radical gives [1 Section 2]

N1(X)≪(log⁡X)2max⁡1≤Y≤Xmax⁡0<μ≤1Nμ∗(Y).(33)N_1(X) \ll(\log X)^2 \max_{1 \le Y \le X}\max_{0 < \mu\le1} N_\mu^*(Y). \tag*{(33)}

Bounded small-height terms can be added without affecting the asserted bound.

Suppose that the conclusion at λ=1\lambda= 1 fails for some ε>0\varepsilon> 0. Then (33) supplies arbitrarily large YY and some μ≤1\mu\le1 for which

Nμ∗(Y)≥YU+ε/2.N_\mu^*(Y) \ge Y^{U+\varepsilon/2}.

Choose a real D0D_0 with

U<D0<U+ε/2.U < D_0 < U + \varepsilon/2.

For each sufficiently large integer nn, apply the anatomic box decomposition with ε0=1/n\varepsilon_0 = 1/n. The fixed integer d=d(1/n)d = d(1/n) may change with nn, but it is fixed during each individual box estimate. Choosing Y=YnY = Y_n sufficiently large, we obtain a box whose count has exponent Dn≥D0D_n \ge D_0 and for which the errors in all inherited inequalities and in the finite collision family used by the tree are at most O(1/n)O(1/n). This uses uniformity over boxes and frozen coefficients, together with finiteness of the collision family.

Record the 18 first-six coordinates and the three totals. By the height inequalities, they lie in a fixed compact set independently of dd. The pair bounds bound DnD_n. Pass to a convergent subsequence and relabel the letters to satisfy (10). Define the limiting Fourier variables by the maxima in (7).

The limiting point satisfies (5)–(9), all required instances of (25), and D≥D0D \ge D_0. Indeed, these are finitely many closed inequalities involving continuous maxima, minima and positive parts, and their errors tend to zero. The tail enters only through

αs−∑i<6αs,i.\alpha_s-\sum_{i<6}\alpha_{s,i}.

In particular, L≤1L \le1 in the limit even though each finite box has a small budget error, and the actual radical sub-budget μ≤1\mu\le1 is absorbed by the root row L≤1L \le1.

Theorem 6.2 now gives D≤UD \le U, contradicting D0>UD_0 > U. This proves

N1(X)≪εXU+ε.N_1(X) \ll_{\varepsilon} X^{U+\varepsilon}.

The logarithmic and divisor losses have been absorbed into the arbitrary ε\varepsilon. Finally, for 0<λ≤10 < \lambda\le1 and c≥2c \ge2, the condition rad⁡(abc)<cλ\operatorname{rad}(abc) < c^\lambda implies rad⁡(abc)<c\operatorname{rad}(abc) < c. Thus the set counted by Nλ(X)N_\lambda(X) is contained in the set counted by N1(X)N_1(X). This proves Theorem 1.1 in its stated uniform range.

The argument proves no corresponding lower bound. A feasible point of the limiting exponent system is only a possible optimizer of a collection of upper estimates; it need not arise from an anatomic box or from any integer triple.

Strict separation and exact program obstructions

Let Dold∗D^*_{\mathrm{old}} and DFullOC∗D^*_{\mathrm{FullOC}} denote the maxima of DD over the old-family and FullOC limiting programs defined in § 5, both with all 15 exponent pairs. These numbers are optima of finite closed exponent programs. They are not counting exponents attained by actual triples.

An exact feasible point for the old family

Consider the rational point

αa,∙=αb,∙=(146,346,323,346,9230,0),αa=αb=37115,\alpha_{a,\bullet}=\alpha_{b,\bullet}=\left(\frac{1}{46},\frac{3}{46},\frac{3}{23},\frac{3}{46},\frac{9}{230},0\right),\qquad\alpha_a=\alpha_b=\frac{37}{115},
{#eq:8.1}

\alpha_{c,\bullet}=\left(\frac{5}{69},\frac{9}{230},\frac{3}{23},\frac{79}{690},0,0\right),\qquad \alpha_c=\frac{41}{115},

Equation not extracted

D=\frac{403}{690},\qquad F_a=\frac{73}{345},\quad F_b=\frac{3}{23},\quad F_c=\frac{53}{345}.

Equation not extracted

Here αs,∙\alpha_{s,\bullet} denotes (αs,1,…,αs,6)(\alpha_{s,1},\ldots,\alpha_{s,6}). All three tails are zero, L=1L=1, and Fa,Fb,FcF_a,F_b,F_c are auxiliary values satisfying the Fourier lower rows; they need not all be minimal.

Proposition 8.1 (Exact old-family obstruction). The point (8.1) satisfies every base row, all 64364^3 geometric inequalities and all 51,030 old-family collision inequalities. Hence

Dold∗≥403690=0.5840579710144927….(34)D^*_{\mathrm{old}}\ge\frac{403}{690}=0.5840579710144927\ldots. \tag*{(34)}

In particular, the old-family program cannot certify the exponent 0.584.

Exact rational computation. Direct substitution gives zero maximum violation of the base rows. The minimum geometric right-hand side is 403/690403/690, attained for example by

Ia={1},Ib={1,3},Ic={1,2,3}.I_a=\{1\},\qquad I_b=\{1,3\},\qquad I_c=\{1,2,3\}.

Enumeration of the 51,030 old-family collision instances in rational arithmetic gives minimum right-hand side 403/690403/690, with 20 binding instances. One binding instance is

A=a,(j,k)=(2,3),B=c,IB={1,2,3},OC={3,5}.A=a,\qquad(j,k)=(2,3),\qquad B=c,\qquad I_B=\{1,2,3\},\qquad O_C=\{3,5\}.

Thus every required right-hand side is at least DD. The independent script and output listed in Appendix A reproduce these assertions without floating-point arithmetic.

For comparison, the old-family control tree has 49,033 leaves, 5,584 geometric nodes and 5,431 collision nodes. Its independently checked maximum leaf bound is

Dold∗≤1169720012=0.5844993004197482….(35)D^*_{\mathrm{old}}\le\frac{11697}{20012}=0.5844993004197482\ldots. \tag*{(35)}

Thus the old optimum is tightly bracketed, although its exact value is not needed here.

The new family cuts this point

Evaluate the all-pair FullOC family at the same point (8.1). Exact enumeration of all 181,440 instances gives minimum collision right-hand side

5392=0.5760869565217391…,403690−5392=111380.(36)\frac{53}{92}=0.5760869565217391\ldots,\qquad\frac{403}{690}-\frac{53}{92}=\frac{11}{1380}. \tag*{(36)}

There are eight binding instances. One is

A=c,(j,k)=(3,4),B=a,C=b,IB={1,2,3,4},OC={3,4,5}.(37)A=c,\qquad(j,k)=(3,4),\qquad B=a,\qquad C=b,\qquad I_B=\{1,2,3,4\},\qquad O_C=\{3,4,5\}. \tag*{(37)}

We display the substitution to make the separation independently checkable by hand. At (37),

fA=0,fB=9230,fC=27115,αC=37115,f_A=0,\qquad f_B=\frac{9}{230},\qquad f_C=\frac{27}{115},\qquad\alpha_C=\frac{37}{115},
p=323,q=79690,r=77690,wB=946,hB=3746,wC=3946.p=\frac{3}{23},\qquad q=\frac{79}{690},\qquad r=\frac{77}{690},\qquad w_B=\frac{9}{46},\qquad h_B=\frac{37}{46},\qquad w_C=\frac{39}{46}.

Consequently

(2p−wB)+=346,(2q−wB)+=130,(2p-w_B)_+=\frac{3}{46},\qquad(2q-w_B)_+=\frac{1}{30},
κ(p,q;wB)=203690,K=2r+κ=119230.\kappa(p,q;w_B)=\frac{203}{690},\qquad K=2r+\kappa=\frac{119}{230}.

Since hB−wC=−1/23h_B-w_C=-1/23, the whole-value saving is δvalue=0\delta_{\mathrm{value}}=0, and therefore the minimum saving is δ=0\delta=0. The diagonal expression is

fA+fB+αC=83230,f_A+f_B+\alpha_C=\frac{83}{230},

whereas the off-diagonal expression is

fA+fB+fC+αC−fC+K+δ2=63230+12(20230+119230)=5392.f_A+f_B+f_C+\frac{\alpha_C-f_C+K+\delta}{2} =\frac{63}{230}+\frac{1}{2}\left(\frac{20}{230}+\frac{119}{230}\right)=\frac{53}{92}.

Thus (25) gives D≤53/92D\le53/92, which the old feasible point violates.

Theorem 8.2 (Strict FullOC separation). The old and new program optima satisfy

DFullOC∗≤U<73125<403690≤Dold∗≤1169720012.(38)D^*_{\mathrm{FullOC}}\le U<\frac{73}{125}<\frac{403}{690}\le D^*_{\mathrm{old}}\le\frac{11697}{20012}. \tag*{(38)}

In particular, FullOC gives an independent exponent improvement rather than merely improving the numerical solver.

Proof. Theorem 6.2 gives the first inequality. The exact target identity gives U<73/125U<73/125, and

403690−73125=117250>0.\frac{403}{690}-\frac{73}{125}=\frac{1}{17250}>0.

Proposition 8.1 gives 403/690≤Dold∗403/690\le D^*_{\mathrm{old}}, and the old control certificate gives the last inequality. Strictness can also be read directly from (36): a point feasible for the old program is infeasible for FullOC. □

The improvement over the old upper certificate is exactly

1169720012−U=2498003550030030018=0.0004993008197479….\frac{11697}{20012}-U=\frac{24980035}{50030030018}=0.0004993008197479\ldots.

A lower obstruction for the FullOC program

Now consider the symmetric point

αs,∙=(19396,122,322,111,0,0),αs=13(s∈{a,b,c}),\alpha_{s,\bullet}=\left(\frac{19}{396},\frac{1}{22},\frac{3}{22},\frac{1}{11},0,0\right),\qquad\alpha_s=\frac{1}{3}\quad(s\in\{a,b,c\}),
Fa=522,Fb=Fc=322,D=712.F_a=\frac{5}{22},\qquad F_b=F_c=\frac{3}{22},\qquad D=\frac{7}{12}.

Each letter has tail mass

13−(19396+122+322+111)=5396,\frac{1}{3}-\left(\frac{19}{396}+\frac{1}{22}+\frac{3}{22}+\frac{1}{11}\right)=\frac{5}{396},

and L=1L=1.

Proposition 8.3 (Exact FullOC lower obstruction). The point (8.7) satisfies every base row, all 64364^3 geometric inequalities and all 181,440 FullOC collision instances with all 15 exponent pairs. Therefore

712≤DFullOC∗≤U=1168000720000012.(39)\frac{7}{12}\le D^*_{\mathrm{FullOC}}\le U=\frac{11680007}{20000012}. \tag*{(39)}

Exact rational computation. The base rows again have zero maximum violation. The minimum geometric right-hand side is 7/127/12, attained for example by

Ia={1},Ib=Ic={1,3}.I_a=\{1\},\qquad I_b=I_c=\{1,3\}.

The minimum over all FullOC collision right-hand sides is also 7/127/12, with 576 binding instances. A binding choice is

A=a,(j,k)=(3,4),B=b,IB={3},OC=∅.A=a,\qquad(j,k)=(3,4),\qquad B=b,\qquad I_B=\{3\},\qquad O_C=\varnothing.

The independent rational evaluator checks all instances, rather than only these witnesses. □

The remaining gap inside the present program is

U−712=1000015000009=0.0006666662666669…U-\frac{7}{12}=\frac{10000}{15000009}=0.0006666662666669\ldots

Proposition 8.3 rules out every bound below 7/127/12 from this program. It does not assert that DFullOC∗=7/12D^*_{\mathrm{FullOC}}=7/12, nor that any further positive part of the displayed gap is attainable by additional search. Most importantly, neither (8.1) nor (8.7) is a lower-bound construction for Nλ(X)N_\lambda(X).

Scope, limitations and relation to abc–Szpiro

What the theorem does and does not say

Theorem 1.1 is a general upper bound for the exceptional-set counting function. It does not prove the abc conjecture, which requires boundedness of Nλ(X)N_\lambda(X) for every fixed λ<1\lambda<1, and it does not prove any pointwise inequality of the form rad⁡(abc)≫εc1−ε\operatorname{rad}(abc)\gg_\varepsilon c^{1-\varepsilon}. A polynomial upper bound on the number of exceptions is weaker than finiteness.

Likewise, the paper gives no new bound for the discriminant–conductor ratio of an elliptic curve and does not prove Szpiro’s conjecture. No Frey-curve specialization is used to convert the counting theorem into a pointwise Arakelov statement. In particular, the present argument constructs neither a cyclotomic-tower degeneration functor nor an identification between exponent-program variables and arithmetic degrees of such a functor.

The exact feasible points in § 8 have a deliberately restricted meaning. They certify lower bounds for the values of finite exponent programs. They do not provide sequences of anatomic boxes, still less sequences of primitive abc triples. Thus 7/12≤DFullOC∗7/12\le D^*_{\mathrm{FullOC}} is not an arithmetic lower bound for Nλ(X)N_\lambda(X).

Limitations of the present computation

  1. The new tree is run at radical budget λ=1\lambda= 1, with uniformity for 0<λ≤10 < \lambda\le1 obtained by inclusion. The inherited anatomic estimates are available for 0<λ≤20 < \lambda\le2, but no new fixed-budget table beyond 1 is claimed here.

  1. The upper certificate uses 11 coprime exponent pairs. All 15 pairs are arithmetically legal by Lemma 3.3 and are included in the exact obstruction checks, but adding the four noncoprime pairs to the upper search is not claimed to improve UU.

  1. Completeness of the 32 modulus sets is proved for the whole-value lock and the grouped savings specified in Proposition 5.1. It is not a completeness theorem for every possible coordinate-lock extension.

  1. The upper bound UU is not asserted to equal DFullOC∗D^{*}_{\mathrm{FullOC}}. The exact interval is 7/12≤DFullOC∗≤U7/12 \le D^{*}_{\mathrm{FullOC}} \le U. Closing the remaining gap, or beating 7/127/12, requires additional valid inequalities or new arithmetic input.

  1. The BBLT anatomic estimates are external inputs. The collision mechanism and arbitrary-degree weighted-difference theorem are inherited from Li’s rStation manuscripts, whose public preprint status has been stated. The new contribution is the exhaustive certified deployment, elimination, separation and obstruction analysis, not a rebranding of those inputs.

Conclusion

The complete whole-value modulus family lowers the certified exceptional-set exponent to the exact value U<0.584U < 0.584. The improvement is proved twice at different levels: arithmetically, by a complete exact-rational tree for Nλ(X)N_{\lambda}(X), and structurally, by a strict separation between the old and new exponent programs. The old family has a feasible obstruction above 0.584, while FullOC excludes it and closes below 0.584. The symmetric FullOC obstruction at 7/127/12 delineates precisely what remains inside the current method. Reproducing the checker, rather than rerunning the floating-point search, is sufficient to audit every finite implication used in the proof.

Appendix A. Certificate data and reproduction

Data schema

The main files use schema break213-v1. The root records the budget, target, declared maximum, leaf count and tree. A leaf has keys type, y and bound; y maps row identifiers to positive rational multiplier strings. A geometric internal node has I, J, K and children. A collision node has A, pA, B, IB, kB, OC and children. No solver matrix is trusted as input: the checker reconstructs every row from the mathematical formulas.

The raw search label all is normalized to all_deg in the main certificate. This metadata change tells the checker to reconstruct δe\delta_e for e=2,…,6e = 2,\ldots,6; it does not alter the tree, multipliers or leaf bounds. The normalization script is included.

Verification commands

Verification of the supplied certificates and obstructions requires only the Python standard library:

python3 -B check_extended.py cert_fulloc_5840.json.gz --faults
python3 -B check_extended.py cert_old_5845.json.gz --faults
python3 -B check_obstructions.py
sha256sum -c SHA256SUMS

The first two commands must end in VERIFIED, and the third in ALL EXACT OBSTRUCTION CHECKS PASSED. The fault option tests a changed multiplier, a deleted child and an altered declared maximum; all three corrupted certificates must be rejected.

The recorded obstruction check gives

old point / old family:      minimum RHS = 403/690, binding = 20
old point / FullOC:          minimum RHS = 53/92,   binding = 8
symmetric point / FullOC:    minimum RHS = 7/12,    binding = 576

The same script checks the root rows and all 64364^{3} geometric instances for each point.

Optional regeneration requires NumPy and SciPy and is not needed to audit the proof:

BREAK213_SECONDS=2400 python3 -B solve_fulloc_5805.py 1.00 73/125 full5840
python3 -B normalize_fulloc.py cert_full5840_1p00.json cert_fulloc_5840.json
python3 -B check_extended.py cert_fulloc_5840.json --faults

Branch choices and runtime may vary with LP versions. The supplied exact certificate, rather than a fresh optimizer run, is the proof artifact.

Principal SHA-256 hashes

The uncompressed normalized main certificate has hash d29def66fa03ddc512b7ba7938a26295a23a278ce90cfa49ab9f0456a33275c8.

The principal files in the reproduction package have the following hashes.

  • cert_fulloc_5840.json.gz 8b81b3f0e8030f0b98445970dbf95763dd6725feb12d981738bb0ad4c64cdbfd

  • cert_old_5845.json.gz 603cd0be5dd912a3389fcd292ffec36ddd07099d3257504e4d08f149e9106291

  • check_extended.py d3e4256b6ddb9004292519b2a1e1429e39a5203c8998d00ee7bdd9f1a0bfefe7

  • check_obstructions.py 89f616f79f50b20dbb8371db3b8ff6bbfe34ebd8a47ff00c97981060a9af0eb6

  • solve_fulloc_5805.py 6450573601c2d4cab92327ae1652e2692301fff5c68323a76ef81b6ddcd0f36e

  • verify_fulloc_5840.log f731ff73c9255cef2de206c243575304182fe3cdb9bf269ba80408f156dba9d1

The package manifest also covers the old generator, engine, normalization script, raw point diagnostics and all recorded logs.

Audit boundary

A TypeError or candidate-exhaustion event in the search is not treated as a mathematical lower bound. The lower obstructions used in § 8 are proved only after their rational coordinates have been checked against every constraint by check_obstructions.py. Conversely, an LP display such as max_D=0.5840000 is not treated as an upper-bound proof; the upper theorem uses only the completed tree and its independently checked rational identities.

Data availability

All rational certificates, linear-programming witness files and verification scripts for the exponent computations are archived in a separate Zenodo record (DOI: 10.5281/zenodo.24354214). This manuscript contains only the mathematical exposition; the computational reproduction package is hosted externally.

References

  1. [1]Christian Bernert, Tim Browning, Jared Duker Lichtman and Joni Teräväinen, Bounds on the exceptional set in the abc conjecture, arXiv:2410.12234v2, version 2 submitted 9 May 2026 (version 1: 16 October 2024). https://arxiv.org/abs/2410.12234v2.
  2. [2]Hongmin Li, Collection inequalities and the radical budget for the exceptional set in the abc conjecture, rStation, Mathematics paper 17, version 2 (2026). https://rstation.ai/problems/math/papers/17/v2.DOI
  3. [3]Hongmin Li, Squares and twisted energy in the exceptional set of the abc conjecture, rStation, Mathematics paper 28, version 3, 13 November 2026. https://rstation.ai/problems/math/papers/28/v3.

Paper details

vlnc:W_py4v394g
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents