\documentclass[11pt,reqno]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[margin=2.6cm]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools,booktabs,array,enumitem,microtype} \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref} \setlength{\emergencystretch}{3em} \allowdisplaybreaks \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{hypothesis}[theorem]{Hypothesis} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \newcommand{\Z}{\mathbb Z} \newcommand{\N}{\mathbb N} \newcommand{\Q}{\mathbb Q} \newcommand{\C}{\mathbb C} \newcommand{\PP}{\mathbb P} \newcommand{\A}{\mathbb A} \newcommand{\ub}{\bar u} \newcommand{\vb}{\bar v} \newcommand{\Bred}{B_{\mathrm{red}}} \newcommand{\Res}{\operatorname{Res}} \newcommand{\lst}{\ell^{*}} \newcommand{\tst}{t^{*}} \newcommand{\Ncen}{N_k^{\mathrm{cen}}} \title{Counting powerful products of consecutive integers} \author{Danny Plymesser} \date{Version 1.0.1, 24 October 2026} \begin{document} \begin{abstract} We show that the number of $n\le x$ for which $n$ and $n+1$ are both powerful is $O_\varepsilon(x^{1/4+\varepsilon})$, improving the exponent $29/100$ of Reuss. For each fixed $k\ge 3$ we show that the number of $n\le x$ for which $(n+1)(n+2)\cdots(n+k)$ is powerful is $O_{k,\varepsilon}(x^{77/320+\varepsilon})$, and we prove the same exponent uniformly over all widths $k\ge3$. The new ingredient is a per-interval count in Reuss's form of Heath-Brown's approximate determinant method: in lattice-reduced coordinates, the solutions on an auxiliary curve of bidegree $(\alpha,\beta)$ become integral points of height at most $L_1^3T_1^2$ on an affine curve of degree $2\alpha+3\beta$, to which the uniform bound of Castryck, Cluckers, Dittmann and Nguyen applies. The proofs are formalized in Lean~4 with Mathlib and checked by Lean's kernel and an independent kernel; the formal statements take six theorems from the literature as explicit hypotheses. The results and exposition were produced and reviewed using AI systems; no human mathematical review has been performed. Section~\ref{sec:status} describes the verification and its scope. \end{abstract} \maketitle \section{Introduction}\label{sec:intro} A positive integer $m$ is \emph{powerful} (or square-full) if $p^2\mid m$ for every prime $p\mid m$. Every powerful number can be written uniquely as $d^3u^2$ with $d$ squarefree, so there are $\asymp x^{1/2}$ powerful numbers up to $x$. \subsection{Consecutive powerful numbers} Pairs of consecutive powerful numbers exist in abundance: every solution of $X^2-8Y^2=1$ gives the pair $8Y^2,\,X^2$, and this Pell family alone contributes $\asymp\log x$ pairs up to $x$; see \cite{Golomb,Walker} for other families. Erd\H{o}s conjectured that there are no three consecutive powerful numbers \cite{Erdos76,ErdosManitoba}\cite[Problem \#364]{Bloom}, and it has been conjectured that the number of consecutive pairs up to $x$ is $(\log x)^{O(1)}$; see \cite{Erdos76}, \cite[p.~68]{ErdosGraham}, and \cite[(1.15)]{Tao}, \cite[Problem \#365]{Bloom}. Erd\H{o}s also asked whether a product of $k\ge3$ consecutive integers can be powerful \cite[Problem \#137]{Bloom}; by the Erd\H{o}s--Selfridge theorem such a product is never a perfect power \cite{ErdosSelfridge}. Under the $abc$ conjecture the number of consecutive pairs is $x^{o(1)}$ and consecutive triples are finite in number; see \cite[(1.16) and the following discussion]{Tao}. Unconditionally, the trivial bound for \[ N(x):=\#\{n\le x:\ n\text{ and }n+1\text{ are powerful}\} \] is $O(x^{1/2})$. Chan \cite{Chan2012} proved that the number of $n\le x$ with $n$ and $n+l$ both powerful is $\ll d_3(l)x^{2/5}(\log x)^2$ uniformly in $l\ge1$ \cite[Theorem~3]{Chan2012}, and $\ll_l x^{7/19}\log x$ for fixed $l$ \cite[Theorem~4]{Chan2012}. Akta\c{s} and Murty \cite{AktasMurty} gave an $O(x^{2/5})$ bound for $N(x)$. Tao records $x^{2/5}$ as the best known bound \cite[(1.16)]{Tao} and proves an estimate of the same strength, $x^{2/5+o(1)}$, for the non-powerful integers lying in intervals of length at least two with powerful product \cite[Theorem~1.8]{Tao}. Reuss \cite[Theorem~4]{Reuss} proved $N(x)\ll_\varepsilon x^{29/100+\varepsilon}$ with his extension of Heath-Brown's approximate determinant method \cite{HBsqfree,HBAnnals}; the exponent $29/100$ is also recorded in \cite{WongcharoenbhornMeemark}. We are not aware of a published improvement of $29/100$. \subsection{Results} \begin{theorem}\label{thm:k2} For every $\varepsilon>0$, \[ \#\{n\le x:\ n\text{ and }n+1\text{ are powerful}\}\ll_\varepsilon x^{1/4+\varepsilon}. \] \end{theorem} The exponent $1/4$ is the exact limit of our method for $k=2$ (Remark~\ref{rem:quarter}). For $k\ge2$ and $x\ge1$ put \[ N_k(x):=\#\{0\le n\le x:\ (n+1)(n+2)\cdots(n+k)\text{ is powerful}\}. \] For $k\ge3$ the product can be powerful without any single factor being powerful, because primes $p0$, \[ N_k(x)\ll_{k,\varepsilon}x^{77/320+\varepsilon} \qquad\text{and}\qquad \Ncen(x)\ll_{k,\varepsilon}x^{38059/160000+\varepsilon}. \] \end{theorem} Here $77/320=0.240625$ and $38059/160000=0.23786875$. We know of no bound for $N_k(x)$, $k\ge3$, stated in the literature, but two bounds follow at once from the literature: Chan's fixed-shift bound \cite[Theorem~4]{Chan2012} applied to the pair equation of Lemma~\ref{lem:pair} gives $N_k(x)\ll_kx^{7/19}\log x$, and \cite[Theorem~1.8 and (1.16)]{Tao} give $N_k(x)\ll x^{2/5+o(1)}$. Both deductions are ours. Theorem~\ref{thm:k3} improves both exponents. \begin{theorem}[Uniformly over all widths]\label{thm:uniform} For every $\varepsilon>0$, \[ \#\{0\le n\le x:\exists k\ge3,\ P(n,k)\text{ powerful}\} \ll_\varepsilon x^{77/320+\varepsilon}. \] With $k\ge2$ in place of $k\ge3$, the bound is $O_\varepsilon(x^{1/4+\varepsilon})$. The number of non-powerful positive integers at most $x$ which belong to an interval of consecutive positive integers with powerful product is $O_\varepsilon(x^{77/320+\varepsilon})$. The implied constants are not explicit: they depend on the threshold in Theorem~\ref{thm:tao}. \end{theorem} Here $P(n,k)=(n+1)\cdots(n+k)$. The first bound counts the union of starts over all widths, with one constant independent of $k$. The last assertion bounds the quantity in Tao's Theorem~1.8. \subsection{Method} Write $n+i=f_id_i^3u_i^2$ and $n+j=f_jd_j^3u_j^2$. After absorbing the bounded coefficients $f_i,f_j$, each pair of factors gives a solution of \begin{equation}\label{eq:pair} e^3v^2-d^3u^2=h, \end{equation} with $h\ne0$ bounded in terms of $k$ (Lemma~\ref{lem:pair}). Reuss \cite[\S2]{Reuss} splits the range of $\rho=d/e$ into $\asymp M$ intervals $I$, with $\log M=\tfrac98(1+\delta)\log(DE)\log(UV)/\log x$, and shows by the determinant method that on each interval all solutions lie on one auxiliary curve of bounded degree (Lemma~\ref{thm:lemma7}). He then passes to lattice-reduced coordinates $(\lambda,\tau)\in\Z^2\times\Z^2$ with $|\lambda_i|\le L_i$, $|\tau_i|\le T_i$, and counts the points of the auxiliary curve using one factor at a time, which gives $\ll x^{\varepsilon}\min(L_1,T_1)$ solutions per interval. Our new step (Lemma~\ref{lem:P}) also uses the equation \eqref{eq:pair} itself. If $\Gamma\subset\PP^1\times\PP^1$ is a component of bidegree $(\alpha,\beta)$ of the auxiliary curve, the twelve monomials of bidegree $(3,2)$ in $(\lambda,\tau)$ map the solutions on $\Gamma$, at most two-to-one, to integral points of height at most $L_1^3T_1^2$ on the affine curve obtained by slicing the cone over the Segre--Veronese image of $\Gamma$ with the hyperplane $\{\ell=h\}$, where $\ell$ represents the left-hand side of \eqref{eq:pair}. This curve has degree $2\alpha+3\beta$, and the bound of Castryck, Cluckers, Dittmann and Nguyen \cite{CCDN} gives $\ll(L_1^3T_1^2)^{1/(2\alpha+3\beta)}\log(2+L_1T_1)$ solutions, independently of the coefficients. Components with $\min(\alpha,\beta)=1$ are rational curves; they carry $x^{o(1)}$ solutions by Stewart's bound for Thue equations \cite{Stewart} (Proposition~\ref{prop:lowdeg}), once a Riemann--Hurwitz argument shows that the relevant binary form has at least three distinct roots (Lemma~\ref{lem:threeroots}). Summing over intervals with tail bounds for the reduced lattices and the Cauchy--Schwarz inequality gives a per-box exponent $\Bred=m+\tfrac3{10}\lst+\tfrac15\tst$ (Theorem~\ref{thm:Bred}); Reuss's corresponding exponent is $m+\min(\lst,\tst)$ (Remark~\ref{rem:compare}). The final exponents are the suprema of explicit piecewise-quadratic functionals, bounded by exact certificates (Section~\ref{sec:cert}). For $k\ge3$ the proof uses only three pairwise inputs: the Pell count, the Thue count and $\Bred$. For $k=2$ it also uses Reuss's Theorem~1, on windows whose cube-part scales are both at least a fixed power of $x$, and a Heath-Brown-type bound from \cite[\S5]{Reuss}. \subsection{What is not claimed} The exponent $1/4$ of Theorem~\ref{thm:k2} is exactly the supremum of the functional behind it, attained at a degenerate corner where three bounds that were known before this paper meet (Remark~\ref{rem:quarter}). We obtain nothing below $1/4$ for $k=2$. A uniform saving over the exponent $1/10$ for the components of bidegree $(2,2)$ would lower $77/320$, for example to $89/384$ (Remark~\ref{rem:genus1}); we do not prove such a saving. The constants of Theorem~\ref{thm:uniform} are not explicit, because the threshold in Tao's Lemma~3.1 is not; we do not claim explicit constants anywhere. The formal proofs assume the six results from the literature stated in Section~\ref{sec:cited}; they do not prove them. \subsection{Formal status}\label{sec:formalstatus} All three theorems, and the propositions and lemmas of Sections~\ref{sec:setup} to~\ref{sec:proofs}, are proved in Lean~4 (version 4.33.1) with Mathlib. The Lean statements are the theorems above, with the implied constants written as explicit existential quantifiers, and with the results of Section~\ref{sec:cited} as named hypotheses (\texttt{let}-bound propositions assumed in the statement; no axiom is added). Their axiom sets are exactly \texttt{propext}, \texttt{Classical.choice} and \texttt{Quot.sound}. Appendix~\ref{app:lean} lists the Lean declaration for each statement, and Section~\ref{sec:status} describes the checks. \subsection*{Organisation} Section~\ref{sec:setup} sets up the pair equations, and Section~\ref{sec:cited} states the results from the literature with their Lean forms. Sections~\ref{sec:elem} to~\ref{sec:proofs} give the proof. The steps that carry the new argument (Lemma~\ref{lem:threeroots}, Proposition~\ref{prop:lowdeg}, Lemma~\ref{lem:P}, Proposition~\ref{prop:interval}, Theorem~\ref{thm:Bred} and Lemma~\ref{lem:goodtriple}) are proved in full, following the route of the corresponding Lean proof; the remaining steps are sketched. The Lean declarations listed in Appendix~\ref{app:lean} are the reference for every statement. Section~\ref{sec:remarks} collects remarks, and Section~\ref{sec:status} describes verification and the use of AI systems. \section{Set-up}\label{sec:setup} \subsection{Notation} We write $y\sim Y$ for $Y1$ put $U:=x^{1/2}D^{-3/2}$, $V:=x^{1/2}E^{-3/2}$, and call $W=(D,E,x)$ a \emph{window}. Let $\mathcal S_W$ be the set of $(d,e,u,v)\in\N^4$ with $d\sim D$, $e\sim E$, $U1$ and $\log(DE)\asymp\log(UV)\asymp\log x$. Define $M>1$ by $\log M=\frac98\log(DE)\log(UV)/\log x$. Then for $x$ large in terms of $\varepsilon$, the number of $(d,e,u,v)\in\N^4$ with $D0$, with constants depending on $c$ and $\varepsilon$; so it is an instance of Theorem~\ref{thm:reuss1}. The declaration \path{reuss_window_bridge} proves that this open-endpoint form implies the closed-endpoint form $d\sim D$, $e\sim E$ of $\mathcal S_W$: for fixed $(d,e)$ the equation has at most one solution with $u\in(U,2U)$, so the endpoints $d=2D$ and $e=2E$ add $\ll x^{\varepsilon}\min(E,U)+x^\varepsilon\min(D,V)$. Reuss states his Theorem~1 for every fixed $h\ne0$; his proof begins by reducing to $(ev,du)=1$, a reduction that is immediate for $h=\pm1$. We use the theorem only with $h=1$, that is, only for $k=2$. In exponents, Theorem~\ref{thm:reuss1} bounds $\#\mathcal S_W$ by $x^{R(a,b)+\varepsilon}$ with \begin{equation}\label{eq:R} R(a,b):=\min\Big\{\max\Big(\frac{s+m}2,a,b\Big),\ \max\Big(\frac{t+m}2,\ub,\vb\Big)\Big\}. \end{equation} \begin{theorem}[{Castryck--Cluckers--Dittmann--Nguyen \cite[Theorem~3]{CCDN}}]\label{thm:ccdn} For each $N>1$ there is $c(N)$ such that for all $d\ge1$, all integral affine curves $C\subset\A^N_{\overline{\Q}}$ of degree $d$ and all $B\ge1$, \[ N_{\mathrm{aff}}(C,B):=\#\{P\in C(\Q)\cap\Z^N:\ |P_i|\le B\}\le c(N)\,d^3B^{1/d}(\log B+d). \] \end{theorem} The Lean hypothesis is the case $N=12$, for every prime ideal $I\subset\Q[x_1,\dots,x_{12}]$ whose quotient has Krull dimension one; the degree $d$ is the leading coefficient of the Hilbert polynomial of the quotient. It therefore covers curves that are irreducible over $\Q$ but not geometrically irreducible. This is the extension stated in \cite[Corollary~4.1.4 and Remark~4.1.5]{CCDN}: the rational points of such a curve lie on the intersection with a Galois conjugate, at most $d^2$ points. \begin{theorem}[{Stewart \cite[p.~796, consequence of Theorem~1]{Stewart}}]\label{thm:stewart} Let $Q\in\Z[r,s]$ be a binary form of degree $q\ge3$ with content $1$ and nonzero discriminant, and let $m$ be a nonzero integer. Then \[ \#\{(r,s)\in\Z^2:\ \gcd(r,s)=1,\ Q(r,s)=m\}\le 2800\,q^{1+\omega(|m|)}. \] \end{theorem} Stewart's Theorem~1 assumes only content~$1$ and nonzero discriminant, so reducible forms are included, and it counts pairs $(r,s)$, so $(r,s)$ and $(-r,-s)$ are counted separately; the displayed bound is his consequence of Theorem~1 with $g=|m|$, the factor $1+1/(8\varepsilon q)$ disappearing as $\varepsilon\to\infty$ \cite[pp.~795--796]{Stewart}. The Lean hypothesis is this statement, with ``nonzero discriminant'' expressed as squarefreeness of the form over $\Q$. \begin{theorem}[{Heath-Brown \cite[Theorem~15]{HBLNM}, $n=2$; see also \cite[Theorem~2]{ReussThesis}}]\label{thm:hb15} Let $F\in\Z[x_1,x_2]$ be absolutely irreducible of degree $d\ge2$, let $\eta>0$ and $B_1,B_2\ge2$. Let $T=\max B_1^{e_1}B_2^{e_2}$, the maximum over the exponent vectors of the monomials occurring in $F$. Then the integral zeros of $F$ with $|x_i|\le B_i$ all lie on the zero sets of at most $\ll_{d,\eta}T^{\eta}\exp(\log B_1\log B_2/\log T)\log(2\|F\|)$ polynomials of degree $O_{d,\eta}(1)$, each coprime to $F$. \end{theorem} In \cite[Theorem~15]{HBLNM} (Section~2.6 of the author's manuscript of the notes), and in its restatement \cite[Theorem~2]{ReussThesis}, the degree is arbitrary, $B_i\ge1$, the auxiliary polynomials have integer coefficients, are coprime to $F$ and have degree bounded in terms of $d$ and $\eta$, and the last factor is $\log\|F\|$, with $\|F\|$ the largest modulus of a coefficient. Applied to $2F$, which has the same integral zeros, degree, monomials and irreducibility as $F$, the theorem gives the same conclusion with the factor $\log\|2F\|=\log(2\|F\|)$. The Lean hypothesis asks only for complex polynomials not divisible by $F$, implied by integer polynomials coprime to $F$, and assumes $d\ge2$ and $B_i\ge2$, so it is an instance of the theorem applied to $2F$. \begin{theorem}[{Gelfond's inequality, consequence; see \cite[Lemma~1.6.11]{BombieriGubler}}]\label{thm:gelfond} For each $n$ there is $c_n\ge0$ such that if $G\mid F$ in $\Z[x_1,\dots,x_n]$ with $F\ne0$, then $H(G)\le e^{c_n\deg F}H(F)$. \end{theorem} Lemma~1.6.11 of \cite{BombieriGubler} states that if $f=f_1\cdots f_m$ for complex polynomials in $n$ variables, then $2^{-d}\prod_j\ell_\infty(f_j)\le\ell_\infty(f)\le2^d\prod_j\ell_\infty(f_j)$, where $\ell_\infty$ is the largest absolute value of a coefficient and $d$ is the sum of the partial degrees of $f$. If $F=GK$ in $\Z[x_1,\dots,x_n]$ then $\ell_\infty(K)\ge1$ and $d\le n\deg F$, which gives the displayed statement with $c_n=n\log2$; the Lean hypothesis is this statement. \begin{theorem}[{Tao \cite[Lemma~3.1]{Tao}}]\label{thm:tao} If $(N+1)\cdots(N+H)$ is powerful, then $H0$, depending only on $c$ and $\delta$, with the following property. Let $x$ be large in terms of $h$, $c$ and $\delta$, let $W$ be a window with $DE\ge x^c$ and $UV\ge x^c$, and let $M$ be an integer with $x^{c^2}\le M\le x$ and $\log M\ge\frac98(1+\delta)\log(DE)\log(UV)/\log x$. Let $I=(\rho_0,\rho_0(1+M^{-1})]$ with $D/(2E)\le\rho_0\le2D/E$. Then there is a nonzero $C_I\in\Z[\rho,\omega]$ with $\deg_\rho C_I,\deg_\omega C_I\le K_0$ and coefficients at most $x^{\kappa}$, such that $C_I(d/e,v/u)=0$ for all $(d,e,u,v)\in\mathcal S_W$ with $d/e\in I$. \end{lemma} This is \cite[Lemma~7]{Reuss}, whose proof begins with the reduction to $(ev,du)=1$ recalled in Section~\ref{sec:cited}; the reduction is not needed. The approximation $|v/u-(d/e)^{3/2}|\ll|h|(D/E)^{3/2}/x$ and a Taylor expansion make every maximal minor of the matrix $(\rho_j^a\omega_j^b)_{a\le A,b\le B}$ smaller than the reciprocal of its integral denominators, by Cauchy--Binet and a rearrangement bound for $\sum(i\ell_1+k\ell_2)$ over $H$ lattice points; the choice $A\log(DE)\approx B\log(UV)$ makes this work exactly when $\log M$ exceeds $\frac98\log(DE)\log(UV)/\log x$. A kernel vector of the rank-deficient integral matrix gives $C_I$, with the given representations $d_j/e_j$, $v_j/u_j$ as denominators, so no coprimality is used. \subsection{Components of the auxiliary curve}\label{sec:components} Let $G_I(d,e;v,u):=e^{A}u^{B}C_I(d/e,v/u)$, where $A,B\le K_0$ bound the degrees of $C_I$. It is a nonzero bihomogeneous form of bidegree $(A,B)$ with coefficients at most $x^\kappa$, and it vanishes at every solution with $d/e\in I$. Factor $G_I$ into irreducible elements of $\Z[d,e;v,u]$. There are at most $A+B\le2K_0$ non-constant irreducible factors; each is bihomogeneous, and by Theorem~\ref{thm:gelfond} its coefficients are at most $e^{c_4(A+B)}x^\kappa\le x^{\kappa+1}$. Every solution is a zero of one of them. By Gauss's lemma a factor $G$ of bidegree $(\alpha,\beta)$ is irreducible over $\Q$, and it is of one of the following types. \begin{enumerate}[label=(\alph*)] \item $\beta=0$ and $\alpha\ge2$, or $\alpha=0$ and $\beta\ge2$. A rational zero $(w_0,w_1)\ne0$ of a form $G(d,e)$ would give the linear factor $w_1d-w_0e$; so $G$ has no integral zero with $(d,e)\ne0$ (respectively $(v,u)\ne0$), and carries no solution. \item[(b1)] A line $pd+qe=0$ or $pv+qu=0$. \item[(b2)] $\alpha,\beta\ge1$ and $\min(\alpha,\beta)=1$. Then $G$ is irreducible over $\C$. Say $\beta=1$ and write $G=vP(d,e)+uQ(d,e)$. A common factor of $P$ and $Q$ over $\Q$ would split $G$, so $P$ and $Q$ are coprime in $\Q[d,e]$; a Bézout identity for their dehomogenisations (in either chart) shows that they have no common zero in $\PP^1(\C)$. If $G=G_1G_2$ over $\C$, bihomogeneity makes one factor, say $G_1$, of bidegree $(a_1,0)$; if $a_1\ge1$ then $G_1$ vanishes at some $[r:s]$, so $P(r,s)=Q(r,s)=0$, which is impossible; hence $G_1$ is a constant. The case $\alpha=1$ is symmetric. \item[(b3)] $\alpha,\beta\ge2$, so that $2\alpha+3\beta\ge10$. \end{enumerate} \subsection{Rational components}\label{sec:lowdeg} \begin{lemma}\label{lem:threeroots} Let $\alpha,\beta\ge1$. Let $A,B\in\C[r,s]$ be coprime binary forms of degree $\beta$ and $C,E\in\C[r,s]$ coprime binary forms of degree $\alpha$, and put $R:=B^3C^2-A^3E^2$. Suppose $R\ne0$ and that $\beta$ is odd or $\alpha<3\beta/2$. Then $R$ has at least three distinct zeros in $\PP^1(\C)$. In particular this holds whenever $\min(\alpha,\beta)=1$. \end{lemma} \begin{proof} Coprime forms of positive degree have no common zero in $\PP^1(\C)$, and none of $A,B,C,E$ is zero. Suppose that $R$ has at most two distinct zeros. A linear change of the variables $(r,s)$ preserves all hypotheses, so we may assume that every zero of $R$ lies in $\{[0:1],[1:0]\}$. Put $a(t)=A(t,1)$, and $b,c,e$ likewise, so that $r_0(t):=R(t,1)=b^3c^2-a^3e^2$ is a nonzero polynomial whose only root is $0$; thus $r_0=\kappa t^i$ with $\kappa\ne0$. For a nonzero polynomial $p$ write $p=t^{o(p)}p_1$ with $p_1(0)\ne0$. \emph{Claim: $a$ and $b$ are monomials, and $3o(b)+2o(c)=3o(a)+2o(e)$.} Put $f=b^3c^2$ and $g=a^3e^2$, so $f-g=\kappa t^i$, and let $W:=tf'-if$. Since $t(\kappa t^i)'=i\kappa t^i$, also $W=tg'-ig$, and $\deg W\le\min(\deg f,\deg g)$. If $W=0$ then $tf'=if$ forces $f$ to be a multiple of $t^i$, and likewise $g$; then $a,b,c,e$ are monomials and $o(f)=o(g)=i$. If $W\ne0$, the identity $W=b^2c\,\bigl(t(3b'c+2bc')-ibc\bigr)$ shows $u_1:=b_1^2c_1\mid W$, and symmetrically $u_2:=a_1^2e_1\mid W$. Moreover $t^{o(f)}$ and $t^{o(g)}$ divide $W$, hence so does $t^{M_0}$ with $M_0=\max(o(f),o(g))$. A common root $\tau\ne0$ of $u_1$ and $u_2$ would give $f(\tau)=g(\tau)=0$ and so $\kappa\tau^i=0$; thus $u_1$, $u_2$ and $t$ are pairwise coprime and $u_1u_2t^{M_0}\mid W$. Writing $\deg f=o(f)+3\deg b_1+2\deg c_1$ and $\deg g=o(g)+3\deg a_1+2\deg e_1$, the two inequalities $\deg u_1+\deg u_2+M_0\le\deg f$ and $\le\deg g$ add up to $\deg a_1+\deg b_1+(2M_0-o(f)-o(g))\le0$. Hence $a_1,b_1$ are constants and $o(f)=o(g)$, which is the claim. \emph{Conclusion.} Let $p=\deg a=o(a)$ and $q=\deg b=o(b)$. As $A,B$ have no common zero at $[0:1]$, $p=0$ or $q=0$; as they have no common zero at $[1:0]$, $p=\beta$ or $q=\beta$. Similarly $o(c)=0$ or $o(e)=0$. If $p=0$ and $q=\beta$, the claim gives $3\beta+2o(c)=2o(e)$, so $o(c)=0$ and $2o(e)=3\beta$; if $p=\beta$ and $q=0$, it gives $2o(c)=3\beta$. In both cases $\beta$ is even and $3\beta\le2\alpha$, since $o(c),o(e)\le\alpha$, contrary to the hypothesis. Finally, if $\min(\alpha,\beta)=1$ then either $\beta=1$ is odd or $\alpha=1<3\beta/2$. \end{proof} \begin{proposition}\label{prop:lowdeg} Let $\alpha,\beta\ge1$ with $\min(\alpha,\beta)=1$, $K\ge0$ and $\varepsilon>0$. There is $C$, depending only on $h$, $\mathcal G$, $K$, $\alpha$, $\beta$ and $\varepsilon$, such that for every $x\ge1$ and every $G\in\Z[d,e;v,u]$ bihomogeneous of bidegree $(\alpha,\beta)$, irreducible over $\C$, with coefficients at most $x^K$, the number of $(d,e,u,v)\in\N^4$ with $G(d,e,v,u)=0$, $e^3v^2-d^3u^2=h$ and $\gcd(d,e),\gcd(u,v)\in\mathcal G$ is at most $Cx^{\varepsilon}$. The same holds with $C$ independent of $h$, uniformly in $0<|h|\le x$. \end{proposition} By type (b2) above, this applies to every factor of $G_I$ of that type. \begin{proof} Let $\beta=1$. (If $\alpha=1$, write $G=dP(v,u)+eQ(v,u)$, let $(r,s)=(v,u)/\gcd(u,v)$ and put $A=-Q$, $B=P$, $C=r$, $E=s$; then $dB(r,s)=eA(r,s)$ and $vE(r,s)=uC(r,s)$, and the argument below applies verbatim.) Write $G=vP(d,e)+uQ(d,e)$ with forms $P,Q$ of degree $\alpha$, coprime over $\C$ by irreducibility. Let $(d,e,u,v)$ be counted. Then $(d,e)\ne0$ because $h\ne0$; let $\gamma=\gcd(d,e)$ and $(r,s)=(d,e)/\gamma$. Put $A=r$, $B=s$ (as forms of degree $\beta=1$), $C=-Q$ and $E=P$. Then $dB(r,s)=eA(r,s)$, and dividing $G(d,e,v,u)=0$ by $\gamma^\alpha$ gives $vE(r,s)=uC(r,s)$. \emph{Resultants.} For coprime integral forms $P_1,Q_1$ of degree $n\ge1$ with coefficients at most $H$, let $\mathrm{Res}(P_1,Q_1)$ be the product of the resultants of $P_1(t,1),Q_1(t,1)$ and of $P_1(1,t),Q_1(1,t)$. It is a nonzero integer, $|\mathrm{Res}|\le((2n)!)^2H^{4n}$ by the Leibniz expansion of the Sylvester determinants, and for coprime $(r,s)$ the integer $\gcd(P_1(r,s),Q_1(r,s))$ divides it. Let $g_1=\gcd(A(r,s),B(r,s))$ and $g_2=\gcd(C(r,s),E(r,s))$; these divide $\mathrm{Res}_1:=\mathrm{Res}(A,B)$ and $\mathrm{Res}_2:=\mathrm{Res}(C,E)$. Write $(A,B)(r,s)=g_1(a',b')$ and $(C,E)(r,s)=g_2(c',e')$ with coprime pairs. From $dB=eA$ and $vE=uC$ we get $(d,e)=l(a',b')$ and $(v,u)=\mu(c',e')$ with $|l|=\gcd(d,e)$ and $|\mu|=\gcd(u,v)$. Substituting into $e^3v^2-d^3u^2=h$ gives \[ h\,g_1^3g_2^2=R(r,s)\,l^3\mu^2,\qquad R:=B^3C^2-A^3E^2 . \] Hence $R(r,s)$ divides the nonzero integer $M_0:=h\,\mathrm{Res}_1^3\mathrm{Res}_2^2$, and \[ |M_0|\le|h|\,((2\beta)!)^6((2\alpha)!)^4x^{K(12\beta+8\alpha)}. \] In particular $R\ne0$, so by Lemma~\ref{lem:threeroots} $R$ has at least three distinct zeros in $\PP^1(\C)$. \emph{Squarefree kernel.} Let $g\in\Z[t]$ be the primitive polynomial proportional to the product of the distinct monic irreducible factors of $R(t,1)$ over $\Q$, and let $Q_R(r,s):=s^{e_0+\deg g}g(r/s)$, where $e_0=1$ if $R$ vanishes at $[1:0]$ and $e_0=0$ otherwise. Then $Q_R$ is an integral binary form of content $1$, squarefree over $\Q$, of degree $q$ with $3\le q\le3\beta+2\alpha=:Q_0$ (the lower bound counts the zeros of $R$), and $Q_R$ divides $R$ in $\Z[r,s]$ by Gauss's lemma. So $Q_R(r,s)$ divides $M_0$. \emph{Counting.} The solution is determined by $(r,s)$, $\gcd(d,e)$ and $\gcd(u,v)$: the pairs $(a',b')$ and $(c',e')$ are determined by $(r,s)$, and the signs of $l$ and $\mu$ by the positivity of $d,e,u,v$. For each of the $2d(|M_0|)$ values $m'=\pm k$, $k\mid M_0$, Theorem~\ref{thm:stewart} gives at most $2800q^{1+\omega(m')}\le2800\,Q_0\,d(|M_0|)^{Q_0}$ coprime pairs with $Q_R(r,s)=m'$, using $q^{\omega(m')}\le2^{Q_0\omega(M_0)}\le d(|M_0|)^{Q_0}$. Therefore the count is at most $5600\,Q_0\,(\#\mathcal G)^2\,d(|M_0|)^{Q_0+1}$, and the divisor bound $d(n)\ll_\eta n^\eta$ with $|M_0|\ll x^{O(1)}$ gives $Cx^\varepsilon$. The constant depends on $h$ only through $|h|^\eta$; if $|h|\le x$ this factor is absorbed into $x^\varepsilon$. \end{proof} \subsection{Reduced coordinates}\label{sec:reduced} \begin{lemma}\label{lem:reduced} Let $I=(\rho_0,\rho_0(1+M^{-1})]$ with $D/(2E)\le\rho_0\le 2D/E$, and put $\omega_0=\rho_0^{3/2}$. \begin{enumerate}[label=\textup{(\roman*)}] \item If $(d,e,u,v)\in\mathcal S_W$ and $d/e\in I$, then $00$ and $T_1\ge T_2>0$. Then the number of $(\lambda,\tau)\in\Z^2\times\Z^2$ with $\lambda$ and $\tau$ primitive, $|\lambda_i|\le L_i$, $|\tau_i|\le T_i$, $F(\lambda;\tau)=h$ and $([\lambda],[\tau])\in\Gamma$ is at most \[ 2\,c(12)\,d^3B^{1/d}(\log B+d),\qquad d:=2\alpha+3\beta,\quad B:=\max(1,L_1^3T_1^2), \] with $c(12)$ from Theorem~\ref{thm:ccdn}, uniformly in $F$, $\Gamma$ and $h$. \end{lemma} Geometrically, the twelve monomials $\sigma_k$ of bidegree $(3,2)$ embed $\PP^1\times\PP^1$ by $\mathcal O(3,2)$, and the counted points are integral points on the affine slice $\{F=h\}$ of the cone over the image of $\Gamma$, a curve of degree $2\alpha+3\beta$. The proof works with the coordinate ring of this slice directly. \begin{proof} Let $R=\Q[y_0,y_1;y_2,y_3]$, bigraded with $y_0,y_1$ of bidegree $(1,0)$ and $y_2,y_3$ of bidegree $(0,1)$; let $\sigma_k=y_0^iy_1^{3-i}y_2^jy_3^{2-j}$ for $k=3i+j$, $0\le i\le3$, $0\le j\le2$, so that $\sigma(\lambda,\tau)\in\Z^{12}$ denotes the vector of these monomials. Let $G$ be an equation of $\Gamma$, irreducible in $R$, so that $(G)$ is prime. If $G\mid F$, then $F$ vanishes on $\Gamma$ and, as $h\ne0$, nothing is counted. Otherwise the image $\bar F$ of $F$ in the domain $R/(G)$ is nonzero; let $S=(R/(G))[\bar F^{-1}]$ and let $\psi\colon\Q[x_1,\dots,x_{12}]\to S$ send $x_k$ to $h\sigma_k/\bar F$. Its kernel $I$ is a prime ideal. \emph{Hilbert function.} The image under $\psi$ of the polynomials of total degree at most $s$ is $\{\bar N\bar F^{-s}:N\in R_{(3s,2s)}\}$. Indeed, a monomial $x^m$ with $|m|\le s$ maps to $h^{|m|}\sigma^m\bar F^{\,s-|m|}/\bar F^{\,s}$, whose numerator has bidegree $(3s,2s)$; conversely every monomial of bidegree $(3s,2s)$ is a product of $s$ of the $\sigma_k$, and $h\ne0$. Multiplication by $\bar F^{-s}$ is injective, so this image has dimension $\dim(R/(G))_{(3s,2s)}$. The exact sequence $0\to R_{(3s-\alpha,2s-\beta)}\xrightarrow{\ \cdot G\ }R_{(3s,2s)}\to(R/(G))_{(3s,2s)}\to0$ and $\dim R_{(a,b)}=(a+1)(b+1)$ give, for $s\ge\alpha+\beta$, \[ (3s+1)(2s+1)-(3s-\alpha+1)(2s-\beta+1)=(2\alpha+3\beta)\,s+(\alpha+\beta-\alpha\beta). \] So $I$ has degree $d=2\alpha+3\beta$ in the sense of Theorem~\ref{thm:ccdn}. \emph{Dimension one.} Since the Hilbert function grows linearly, any two elements of $\Q[x]/I$ are algebraically dependent: the $(k+1)^2$ products $a^ib^j$ with $i,j\le k$ lie in a filtration piece of dimension $O(k)$. If $\mathfrak p\ni p\ne0$ is a prime and $\Phi(p,b)=0$ with $\Phi\ne0$, cancelling the largest power of the first variable dividing $\Phi$ (as $\Q[x]/I$ is a domain) and reducing modulo $\mathfrak p$ shows that $b$ is algebraic over $\Q$ modulo $\mathfrak p$. Hence the quotient by any nonzero prime is algebraic over $\Q$ and therefore a field, so every nonzero prime is maximal; and $\Q[x]/I$, which is infinite-dimensional and finitely generated, is not a field. So $\dim\Q[x]/I=1$. \emph{The counted points.} Let $(\lambda,\tau)$ be counted. Evaluation at $(\lambda,\tau)$ defines $R/(G)\to\Q$, which sends $\bar F$ to $h\ne0$ and so extends to $S$; composed with $\psi$ it sends $x_k$ to $\sigma_k(\lambda,\tau)$. Hence every element of $I$ vanishes at $w=\sigma(\lambda,\tau)\in\Z^{12}$, and $|w_k|\le L_1^iL_2^{3-i}T_1^jT_2^{2-j}\le L_1^3T_1^2\le B$. \emph{Fibres.} Suppose $\sigma(\lambda,\tau)=\sigma(\lambda',\tau')$ with all four vectors primitive. Some $\tau_1^j\tau_2^{2-j}$ is nonzero, and the coordinates with that $j$ show that the cubic monomial vectors of $\lambda$ and $\lambda'$ are proportional with a nonzero factor; hence $\lambda$ and $\lambda'$ are proportional, and being primitive, $\lambda'=\kappa\lambda$ with $\kappa=\pm1$. Likewise $\tau'=\mu\tau$ with $\mu=\pm1$. A nonzero coordinate $\lambda_1^i\lambda_2^{3-i}\tau_1^j\tau_2^{2-j}$ is multiplied by $\kappa^3\mu^2=\kappa$, so $\kappa=1$. Thus each point $w$ comes from at most the two pairs $(\lambda,\pm\tau)$. Therefore the count is at most twice the number of integral points of height at most $B$ on the curve defined by $I$, and Theorem~\ref{thm:ccdn} with $N=12$ gives the bound. \end{proof} \subsection{The per-interval bound} \begin{proposition}\label{prop:interval} In the setting of this section, for each interval $I$ the number of $(d,e,u,v)\in\mathcal S_W$ with $d/e\in I$ and $\gcd(d,e),\gcd(u,v)\in\mathcal G$ is \[ \ll_{h,c,\delta,\mathcal G,\varepsilon}x^{\varepsilon}\Big(1+\big(L_1(I)^3T_1(I)^2\big)^{1/10}\Big). \] \end{proposition} \begin{proof} By Section~\ref{sec:components} every counted solution is a zero of one of at most $2K_0$ irreducible factors $G$ of $G_I$, each with coefficients at most $x^{\kappa+1}$. We bound the solutions on each factor. (The formal statement gives the bound with $L_1$ and $T_1$ computed from some nonzero vectors of $\Lambda_I$ and $\Lambda'_I$; this implies the form stated, since the reduced vectors $g^{(1)}$, $g'^{(1)}$ are shortest.) Type (a) carries none. For a line $pd+qe=0$, $(d,e)$ is $\gamma$ times a fixed primitive vector with $\gamma\in\mathcal G$, and for each such $(d,e)$ Lemma~\ref{lem:pellthue}(i) gives $\ll x^\varepsilon$ pairs $(u,v)$; for a line $pv+qu=0$, $(v,u)$ takes at most $\#\mathcal G$ values and Lemma~\ref{lem:pellthue}(ii) gives $O(1)$ pairs $(d,e)$ for each. Type (b2) carries $\ll x^\varepsilon$ solutions by Proposition~\ref{prop:lowdeg} with $K=\kappa+1$. Let $G$ be of type (b3), and fix $\gamma=\gcd(d,e)$ and $\gamma'=\gcd(u,v)$ in $\mathcal G$. Put $(d_0,e_0)=(d,e)/\gamma$ and $(v_0,u_0)=(v,u)/\gamma'$; these are primitive, $e_0^3v_0^2-d_0^3u_0^2=h':=h/(\gamma^3\gamma'^2)$ is a nonzero integer, and $G(d_0,e_0,v_0,u_0)=0$ by bihomogeneity. Since $M\le x^{1/2}$ and $x$ is large, $x\ge3|h|M$, so Lemma~\ref{lem:reduced}(i) gives $|d_0-\rho_0e_0|\le4D/M$, $|e_0|\le2E$, $|v_0-\omega_0u_0|\le17V/M$ and $|u_0|\le2U$, and by Lemma~\ref{lem:reduced}(iii) the coordinates $\lambda$, $\tau$ of $(d_0,e_0)$ and $(v_0,u_0)$ are primitive with $|\lambda_i|\le L_i$ and $|\tau_i|\le T_i$. In these coordinates $e^3v^2-d^3u^2$ becomes a bihomogeneous form of bidegree $(3,2)$ and $G$ becomes an irreducible form of the same bidegree, since the changes of variables lie in $\mathrm{GL}_2(\Z)$; the map to $(\lambda,\tau)$ is injective. Lemma~\ref{lem:P} with $h'$ and $d=2\alpha+3\beta\ge10$ bounds the solutions by $2c(12)d^3B^{1/d}(\log B+d)$. Here $B^{1/d}\le1+(L_1^3T_1^2)^{1/10}$, while $L_1\ll x$ and $T_1\ll x$ give $\log B\ll\log x$, and $d\le5K_0$; so this is $\ll x^\varepsilon(1+(L_1^3T_1^2)^{1/10})$. Summing over $(\gamma,\gamma')\in\mathcal G^2$ and over the at most $2K_0$ factors proves the proposition. \end{proof} \section{Summation over intervals}\label{sec:sum} Put $L_*:=\max((DE/M)^{1/2},32(D+E)/M)$ and $T_*:=\max((UV/M)^{1/2},194(U+V)/M)$. \begin{lemma}[Tail counts]\label{lem:tails} For $L\ge32(D+E)/M$, $\#\{I:\ L_1(I)\ge L\}\le720\,DE/L^2$, and for $T\ge194(U+V)/M$, $\#\{I:\ T_1(I)\ge T\}\le7000\,UV/T^2$. \end{lemma} A short lattice vector is a rational approximation to $\rho_0$ or $\omega_0$ with bounded numerator; the points $\rho_0$ are spaced at least $D/(2EM)$ apart, and counting the pairs (denominator, interval) gives the bounds. \begin{lemma}\label{lem:moments} For $0<\theta<2$, $\sum_I L_1(I)^{\theta}\ll_\theta M L_*^{\theta}$ and $\sum_I T_1(I)^{\theta}\ll_\theta M T_*^{\theta}$. \end{lemma} The intervals with $L_1(I)\le L_*$ contribute at most $(3M+1)L_*^\theta$, and by Lemma~\ref{lem:tails} and $DE\le ML_*^2$ the dyadic shell $2^kL_*\le L_1(I)<2^{k+1}L_*$ contributes $\ll ML_*^\theta(2^\theta/4)^k$; the series converges because $\theta<2$. \begin{theorem}[The reduced per-box bound]\label{thm:Bred} Let $h\ne0$, $c\in(0,1)$ and $\mathcal G$ be fixed, and let $W$ be a window with $s,t\ge c$. Define \begin{equation}\label{eq:Bred} \Bred(a,b):=m+\frac3{10}\,\lst+\frac15\,\tst, \end{equation} where $\lst:=\max(\frac{s-m}2,a-m,b-m)$ and $\tst:=\max(\frac{t-m}2,\ub-m,\vb-m)$. Then for every $\varepsilon>0$ the number of $(d,e,u,v)\in\mathcal S_W$ with $\gcd(d,e),\gcd(u,v)\in\mathcal G$ is $\ll_{h,c,\mathcal G,\varepsilon}x^{\max(m,\Bred(a,b))+\varepsilon}$. \end{theorem} \begin{proof} Put $\delta=\min(\varepsilon/2,1/2)$ and $M=\lceil x^{(1+\delta)m}\rceil$. Each solution lies in one of the at most $3M+1$ intervals $I$, so by Proposition~\ref{prop:interval} (with $\varepsilon/2$) the count is at most $\sum_I A\,(1+L_1(I)^{3/10}T_1(I)^{1/5})$ with $A\ll x^{\varepsilon/2}$. For $L,T\ge0$, applying $XY\le X^2+Y^2$ to $X=(L/L_*)^{3/10}$ and $Y=(T/T_*)^{1/5}$ gives \[ L^{3/10}T^{1/5}\le L_*^{3/10}T_*^{1/5}\Bigl(\frac{L^{3/5}}{L_*^{3/5}}+\frac{T^{2/5}}{T_*^{2/5}}\Bigr). \] Summing over $I$ with Lemma~\ref{lem:moments} for $\theta=3/5$ and $\theta=2/5$, the count is $\ll A\,(M+ML_*^{3/10}T_*^{1/5})$. Now $x^{(1+\delta)m}\le M\le2x^{(1+\delta)m}$, $0\le m\le3/16$, $L_*\ll x^{\lst-\delta m/2}$ and $T_*\ll x^{\tst-\delta m/2}$, so $M\ll x^{m+\delta}$ and $ML_*^{3/10}T_*^{1/5}\ll x^{\Bred+\frac34\delta m}\le x^{\Bred+\delta}$. Since $\delta\le\varepsilon/2$, the count is $\ll x^{\max(m,\Bred)+\varepsilon}$. \end{proof} \begin{lemma}\label{lem:Bredm} If $0\le a,b\le\frac13$ and $s>2/27$, then $\Bred(a,b)\ge m$. \end{lemma} \begin{proof} Here $s\le\frac23$, so $t\ge0$. Then $s-m=s(1-\frac98t)\ge0$ because $s\ge2/27$ is equivalent to $t\le\frac89$, and $t-m=t(1-\frac98s)\ge0$ since $s\le\frac23<\frac89$. So $\lst,\tst\ge0$. \end{proof} \begin{lemma}[Balanced form]\label{lem:balanced} If $|a-b|\le\frac23m$, then $\Bred(a,b)=\frac34m+\frac1{10}$. \end{lemma} \begin{proof} $a-m\le\frac{s-m}2$ if and only if $a-b\le m$, and $\ub-m\le\frac{t-m}2$ if and only if $\frac32(b-a)\le m$; with the symmetric conditions this gives $\lst=\frac{s-m}2$ and $\tst=\frac{t-m}2$, and then $\Bred=\frac34m+\frac{3s+2t}{20}=\frac34m+\frac1{10}$ because $3s+2t=2$. \end{proof} At $a=b=1/6$ one has $s=1/3$, $m=3/16$ and $\Bred=77/320$. \section{The optimisation and its certificates}\label{sec:cert} Let the gate $\mathsf G$ be the set of $(a,b)\in[0,1/3]^2$ with $s,t\ge1/10$ and $a,b\ge1/100$, and define \begin{gather*} \Phi_3(a,b):=\min\{s,\ t,\ \Bred(a,b)\ \text{if }(a,b)\in\mathsf G\},\\ \Phi_2^\sharp(a,b):=\min\{s,\ t,\ H_1,\ H_2,\ R\ \text{if }(a,b)\in\mathsf G_0,\ \Bred\ \text{if }(a,b)\in\mathsf G\}, \end{gather*} with $\mathsf G_0=\{(a,b)\in[0,1/3]^2:s,t\ge1/10\}$. \begin{proposition}\label{prop:certs} \begin{enumerate}[label=\textup{(\roman*)}] \item For all $(a_1,a_2,a_3)\in[0,1/3]^3$, $\min\{\Phi_3(a_1,a_2),\Phi_3(a_1,a_3),\Phi_3(a_2,a_3)\}\le 77/320$, with equality at $a_1=a_2=a_3=1/6$. \item For all $(a,b)\in[0,1/3]^2$, $\Phi_2^\sharp(a,b)\le\frac14$, with equality exactly at $(0,\frac14)$ and $(\frac14,0)$. \item If $a,a'\in[19/100,21/100]$, then $(a,a')\in\mathsf G$ and $\Bred(a,a')\le 38059/160000$, with an $O(\eta)$ error on the $\eta$-neighbourhood for $0\le\eta\le1/1000$. \end{enumerate} \end{proposition} These are finite statements about explicit piecewise-quadratic functions, proved in Lean by case analysis on the active terms of each maximum and minimum and by exact polynomial inequalities; part (ii) includes the strip inequality $\min\{s,H_2\}\le\frac14+\frac a4$ for $a\le1/100$. The same bounds are certified independently by exact-rational branch and bound (Python \texttt{fractions}) in the supplementary scripts: each box bound uses the concavity of $m$, $(s+m)/2$ and $(t+m)/2$ in $s$ and the linearity of the remaining terms. Table~\ref{tab:cert} lists the runs; negative controls just below each target fail. The point $(a,b)=(1/200,197/800)$ shows that $R$ is needed outside $\mathsf G$: there $\min\{s,H_2\}=201/800>1/4$. \begin{table}[ht] \centering \footnotesize \begin{tabular}{@{}llllr@{}} \toprule Run & functional & region & target $T$ & result (nodes)\\ \midrule K3\_MIN & $\min_{\mathrm{pairs}}\Phi_3$ & $[0,\frac13]^3$ & $77/320$ & certified (633)\\ K3\_MIN\_CTL & same & same & $0.2406$ & fails near $(\frac16,\frac16,0.171)$\\ K2\_FULL & $\Phi_2^\sharp$ & $[0,\frac13]^2$ & $1/4$ & certified (93)\\ K2\_FULL\_CTL1 & same & same & $\frac14-10^{-6}$ & fails near $(0,\frac14)$\\ CENTRAL & $\min_{\mathrm{pairs}}\Bred$ & $[\frac{19}{100},\frac{21}{100}]^3$ & $38059/160000$ & certified (1)\\ THETA & $\min_{\mathrm{pairs}}\min(s,t,B_{1/12})$ & $[0,\frac13]^3$ & $89/384$ & certified (633)\\ \bottomrule \end{tabular} \caption{Exact-rational certificates from \texttt{certificates.py} and \texttt{endpoint\_cert.py} (K2\_FULL); the outputs, with further runs and controls, are in the supplementary directory. $B_{1/12}$ is defined in Remark~\ref{rem:genus1}.}\label{tab:cert} \end{table} \begin{proposition}[Central small-gap exclusion]\label{prop:centralgap} Let $n\ge1$, $H=n+3$, and suppose $(n+1)(n+2)(n+3)$ is powerful. Suppose, for $r=1,2,3$, that $n+r=f_rd_r^3u_r^2$ with positive integers $u_r$, $f_r\in\{1,2\}$ and odd squarefree $d_r$, and that $H^{19}H^{32/1000}$. \end{proposition} It is proved in Lean without hypotheses. It excludes only the small-gap part of the central box. \section{Proofs of the theorems}\label{sec:proofs} \subsection{Fixed width} \emph{Theorem~\ref{thm:k3}.} Split $n+1\in(X/2,X]$ dyadically and sort the $n$ by $(f_1,f_2,f_3)$ and by the dyadic scales $D_i$ of $d_1,d_2,d_3$. For each of the three pairs, Lemmas~\ref{lem:pair} and~\ref{lem:windows} map the $n$ of a box injectively into $O(1)$ window sets. With $\hat a_i=\max(0,\min(\log D_i/\log X,1/3))$, Proposition~\ref{prop:certs}(i) selects a pair with $s\le77/320$, or $t\le77/320$, or $(\hat a_i,\hat a_j)\in\mathsf G$ and $\Bred\le77/320$; Lemma~\ref{lem:pellthue}, or Theorem~\ref{thm:Bred} with Lemma~\ref{lem:Bredm}, bounds the box, after a Lipschitz transfer of the exponents by $O(1/\log X)$. Summing over $O((\log x)^4)$ boxes gives the bound for $N_k$, and Proposition~\ref{prop:certs}(iii) gives the central subcount. This proof does not use Theorems~\ref{thm:reuss1}, \ref{thm:hb15} or~\ref{thm:tao}. \emph{Theorem~\ref{thm:k2}.} Here $k=2$, $h=1$, and $d=d_1$, $e=d_2$ are squarefree. On boxes with $\hat a,\hat b\ge\eta$ for a small fixed $\eta>0$, Proposition~\ref{prop:certs}(ii) selects one of $s,t,H_1,H_2,R,\Bred$ at most $\frac14$, and Lemmas~\ref{lem:pellthue} and~\ref{lem:hb}, Theorem~\ref{thm:reuss1} (on windows with $D,E\ge x^{c}$) or Theorem~\ref{thm:Bred} give the bound. On thin boxes $\hat a<\eta$, the strip inequality $\min\{s,H_2\}\le\frac14+\frac{\hat a}4$ with Lemmas~\ref{lem:pellthue} and~\ref{lem:hb} gives $x^{1/4+\eta/4+\varepsilon}$, and $\eta$ is arbitrary; the case $\hat b<\eta$ is symmetric. So Theorem~\ref{thm:reuss1} is needed only on windows with both cube-part scales at least a fixed power of $x$. \subsection{Uniformity over all widths}\label{sec:uniform} Put $\theta=77/320$. Summing Theorem~\ref{thm:k3} over $k$ would not suffice; the argument instead selects three consecutive factors with small coefficients and makes the pair bounds uniform in growing shifts. \begin{lemma}[Three consecutive small kernels]\label{lem:goodtriple} For $n\ge0$, $k\ge3$ and $1\le j\le k$, let $f_j=\prod_{p0$. For sufficiently large $x$, uniformly in the integer shift $1\le |h|\le x^{\delta/4}$ and the window $W$, the number $A_W(h)$ of solutions in $\mathcal S_W$ with $\gcd(d,e)=\gcd(u,v)=1$ satisfies $A_W(h)\ll x^{s+\varepsilon}$ and $A_W(h)\ll x^{t+\varepsilon}$, and if also $s,t\ge c$, then $A_W(h)\ll_{c,\delta,\varepsilon}x^{\max(m,\Bred(a,b))+\varepsilon+2\delta}$. \end{lemma} \begin{proof} The first two bounds hold for every $x>1$ uniformly in $0<|h|\le x$. Every solution has $d\le2D$, $e\le2E$ and $u,v\le32x$. For coprime $(d,e)$ the number of coprime $(u,v)$ with $u,v\le32x$ and $e^3v^2-d^3u^2=h$ is at most $2\tau(|h|)\bigl(4\log_2((d^3e^3+1)(32x)^2)+2\bigr)$ (the primitive count behind Lemma~\ref{lem:pellthue}(i)), and for each $(u,v)$ Lemma~\ref{lem:pellthue}(ii) with $G_0=1$ gives at most $2800\cdot3^{1+\omega(|h|)}$ coprime pairs $(d,e)$. Summing over the at most $4DE$ pairs $(d,e)$, respectively $4UV$ pairs $(u,v)$, with $\tau(|h|),3^{\omega(|h|)}\ll_\varepsilon x^{\varepsilon}$, gives the two bounds. For the third bound, $|h|\le x^{\delta/4}$ controls every place where the shift enters Sections~\ref{sec:interval} and~\ref{sec:sum}: it gives $3|h|M\le x$, the hypothesis of Lemma~\ref{lem:reduced}(i); it gives $|h|\le x$, the range of the second form of Proposition~\ref{prop:lowdeg} and of the primitive counts of the previous sentences, which replace Lemma~\ref{lem:pellthue} on the lines of type (b1); and it gives $3+8|h|\le x^{\delta/3}$, so that in Lemma~\ref{thm:lemma7} the approximation of $v/u$ by $(d/e)^{3/2}$ loses at most a factor $x^{\delta/3}$ and the determinant estimate holds with degrees and constants depending only on $c$ and $\delta$. Lemma~\ref{lem:P} does not involve $h$, and since the solutions are primitive no division by $\gcd(d,e)$ or $\gcd(u,v)$ is needed. Hence Proposition~\ref{prop:interval} holds with $\mathcal G=\{1\}$ and an implied constant independent of $h$ in this range, and summing it over the intervals as in the proof of Theorem~\ref{thm:Bred} gives $A_W(h)\ll x^{\max(m,\Bred(a,b))+\varepsilon+\delta}$, which implies the stated bound. \end{proof} \begin{lemma}[Three-term count with small coefficients]\label{lem:uniformtriple} Let $\mathcal Q$ be the positive powerful integers. For every $\varepsilon>0$ there are $\eta>0$ and $C_\varepsilon$ such that, for $y\ge2$ and integers $1\le f_1,f_2,f_3\le y^\eta$, \[ \#\{0\le N\le y:N+r\in f_r\mathcal Q\ (r=1,2,3)\} \le C_\varepsilon y^{\theta+\varepsilon}. \] \end{lemma} \begin{proof} Take $\delta$ and $\varepsilon'$ small in terms of $\varepsilon$, and $\eta$ small in terms of $\delta$ and $\varepsilon$. The $N\le y^\theta$ contribute at most $y^\theta+1$. For $y^\thetay^\theta$ and $F\le y^\eta$, with $\eta$ small in terms of $\delta$ we get $h'\le2F^4\le x'^{\delta/4}$, so Lemma~\ref{lem:uniformwindows} applies to $W'$ uniformly in the shift. \emph{The exponents and the choice of pair.} Let $\hat a_r$ be $\log d_r/\log(N+1)$ clipped to $[0,1/3]$. The exponents $(a,b)$ of $W'$ differ from $(\hat a_i,\hat a_j)$ by $O((1+\log F)/\log N)$, the clipping included, and $s,t,m,\lst,\tst$ and $\Bred$ are Lipschitz in $(a,b)$; this perturbation costs a factor $O(F^{O(1)})$. Proposition~\ref{prop:certs}(i) at $(\hat a_1,\hat a_2,\hat a_3)$ selects a pair with $s\le\theta$, or $t\le\theta$, or $(\hat a_i,\hat a_j)\in\mathsf G$ and $\Bred\le\theta$; in the last case $s,t\ge c=1/20$ for $y$ large, and Lemma~\ref{lem:Bredm} at $(\hat a_i,\hat a_j)$, where $s\ge1/10>2/27$, gives $m\le\Bred$ there. In each case Lemma~\ref{lem:uniformwindows} bounds the primitive solutions in $W'$ with shift $h'$ by $\ll_\varepsilon F^{O(1)}(N+1)^{\theta+\varepsilon'+2\delta}$. \emph{Summation.} The key consisting of the pair, $g$, $q$ and the window $W'$ determines $h'$, and together with the primitive solution it determines $N$. There are $O(F^8(\log y)^3)$ keys. The total is therefore $y^\theta+1+F^{O(1)}(\log y)^3y^{\theta+\varepsilon'+2\delta}$, and with $F\le y^\eta$ the choice of parameters makes the exponent less than $\theta+\varepsilon$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:uniform}] Theorem~\ref{thm:tao} applies to every powerful $P(N,k)$ with $N\ge N_0$ and $k\ge3$: the condition $k{\raggedright}p{5.6cm}>{\raggedright\arraybackslash}p{4.4cm}@{}} \toprule Result & Lean declaration & Named hypotheses\\ \midrule Theorem~\ref{thm:k3} & \path{counting_thm_k3} & Stewart, CCDN, Gelfond\\ Theorem~\ref{thm:k2} & \path{counting_thm_k2_reuss} & Stewart, CCDN, Gelfond, Reuss, Heath-Brown\\ Theorem~\ref{thm:uniform} & \path{counting_thm_uniform_reuss} & the five above and Tao\\ Proposition~\ref{prop:interval} & \path{counting_prop_interval} & Stewart, CCDN, Gelfond\\ Theorem~\ref{thm:Bred} & \path{counting_thm_Bred} & Stewart, CCDN, Gelfond\\ Proposition~\ref{prop:interval}, uniform in $h$ & \path{counting_UPI} & Stewart, CCDN, Gelfond\\ Lemma~\ref{lem:uniformwindows}, third bound & \path{counting_lem_uniformwindows_TB} & Stewart, CCDN, Gelfond\\ Lemma~\ref{lem:uniformtriple} & \path{counting_lem_uniformtriple} & Stewart, CCDN, Gelfond\\ Reuss window convention & \path{reuss_window_bridge} & none\\ \bottomrule \end{tabular} \caption{Release declarations of the counting closure.}\label{tab:finals} \end{table} \begin{table}[ht] \centering\footnotesize \begin{tabular}{@{}>{\raggedright}p{3.4cm}>{\raggedright\arraybackslash}p{11.4cm}@{}} \toprule Statement & Lean declarations\\ \midrule Theorem~\ref{thm:k2} & \texttt{CC.Final.counting\_thm\_k2\_reuss}; \texttt{CC.C12} \texttt{a9994a76}, \texttt{CC} \texttt{6ee9f26a}\\ Theorem~\ref{thm:k3} & \texttt{CC.Final.counting\_thm\_k3}; \texttt{CC.C11} \texttt{d974288e}\\ Theorem~\ref{thm:uniform} & \texttt{CC.Final.counting\_thm\_uniform\_reuss}; \texttt{CC} \texttt{3103e62f}, \texttt{e09c0d72}, \texttt{CC.C23} \texttt{8e090865}, \texttt{AllWidthFromUniformTriple} \texttt{2be59d9f}, \texttt{AllWidthFromPairAndUnion} \texttt{fdb026ef}, \texttt{NonpowerfulMembers} \texttt{ceda7068}\\ Lemmas~\ref{lem:decomp}, \ref{lem:pair} & \texttt{DecompositionAndEnvelope} \texttt{9673eb5d}\\ Lemma~\ref{lem:windows} & \texttt{StrictWindows} \texttt{f0cfdddd} (fixed $h$); growing shifts: \texttt{CC.C21.PcPellWin.growing\_cover}\\ Reuss's window convention & \texttt{CC.Final.reuss\_window\_bridge}\\ Lemma~\ref{lem:pellthue} & \texttt{CC.C08} \texttt{ca19e5d2}\\ Lemma~\ref{lem:hb} & \texttt{CC.C13} \texttt{a92c7c0a}\\ Lemma~\ref{thm:lemma7} & \texttt{Paper3OriginalActualAuxiliaryPolynomial} \texttt{1d348eff}\\ Section~\ref{sec:components} & \texttt{CC.C02} \texttt{7327564c}\\ Lemma~\ref{lem:threeroots} & \texttt{CC.C01.PcCurves.three\_roots\_gen}\\ Proposition~\ref{prop:lowdeg} & \texttt{CC.C01} \texttt{6be17294}; uniform in $h$: \texttt{CC.C16} \texttt{b07245cc}\\ Lemma~\ref{lem:reduced} & \texttt{CC.C09} \texttt{cc873556}, \texttt{CC.C04} \texttt{f5611586}\\ Lemma~\ref{lem:P} & \texttt{CC.C03} \texttt{3ede4992}; type (b3): \texttt{CC.C03} \texttt{f159a569}\\ Proposition~\ref{prop:interval} & \texttt{CC.Final.counting\_prop\_interval}; \texttt{CC.C05} \texttt{d9675218}, \texttt{CC.C06} \texttt{40f41d3d}, \texttt{CC.C07} \texttt{71200785}\\ Lemmas~\ref{lem:tails}, \ref{lem:moments} & \texttt{LatticeTails} \texttt{d31e9b77}, \texttt{LatticeMoments} \texttt{de91d4e3}\\ Theorem~\ref{thm:Bred} & \texttt{CC.Final.counting\_thm\_Bred}; \texttt{CC.C10} \texttt{664a2185}, \texttt{8d943e1c}\\ Lemma~\ref{lem:Bredm} & \texttt{DecompositionAndEnvelope} \texttt{928cf894}\\ Lemma~\ref{lem:balanced} & \texttt{DecompositionAndEnvelope.JocelynExpAlg.balanced\_aux}\\ Proposition~\ref{prop:certs} & (i) \texttt{TripleExponentCertificate} \texttt{d17527ab}; (ii) \texttt{PairExponentCertificate} \texttt{c433f729}; (iii) \texttt{DecompositionAndEnvelope} \texttt{9673eb5d}, \texttt{f1b2bf39}\\ Proposition~\ref{prop:centralgap} & \texttt{Erdos137.WidthThree.CentralSet.SmallGapExclusion.small\_gap\_exclusion}\\ Lemma~\ref{lem:goodtriple} & \texttt{AllWidthFromUniformTriple.JocelynAllWidth.goodtriple} (same statement: \texttt{GoodTriple} \texttt{6ba68896}); $k\ge15$: \texttt{Erdos137.FixedWidth.GoodIndex.three\_consecutive\_small\_kernels}\\ Lemma~\ref{lem:uniformwindows} & \texttt{CC.C08.PcPell.uniform\_s}, \texttt{CC.C08.PcPell.uniform\_t} (first two bounds); \texttt{CC.Final.counting\_lem\_uniformwindows\_TB} (third bound), via \texttt{CC.C19} \texttt{62fb44f0} (\texttt{counting\_UPI}), \texttt{CC.C10} \texttt{8d943e1c}, \texttt{CC.C20} \texttt{88a200ce}\\ Lemma~\ref{lem:uniformtriple} & \texttt{CC.Final.counting\_lem\_uniformtriple}; \texttt{CC.C21} \texttt{b14c9f52}\\ \bottomrule \end{tabular} \caption{Concordance between the statements of the paper and the Lean development.}\label{tab:concordance} \end{table} \clearpage \begin{thebibliography}{99} \bibitem{AktasMurty} K.~Akta\c{s} and M.~R.~Murty, \emph{Fundamental units and consecutive squarefull numbers}, Int.~J. Number Theory \textbf{13} (2017), 243--252. \bibitem{Bloom} T.~F.~Bloom, \emph{Erd\H{o}s problems}, \url{https://www.erdosproblems.com}, Problems \#137, \#364, \#365. \bibitem{BombieriGubler} E.~Bombieri and W.~Gubler, \emph{Heights in Diophantine Geometry}, Cambridge University Press, 2006. \bibitem{CCDN} W.~Castryck, R.~Cluckers, P.~Dittmann and K.~H.~Nguyen, \emph{The dimension growth conjecture, polynomial in the degree and without logarithmic factors}, Algebra Number Theory \textbf{14} (2020), 2261--2294; arXiv:1904.13109. \bibitem{Chan2012} T.~H.~Chan, \emph{Twin squareful numbers}, J.~Aust. Math. Soc. \textbf{93} (2012), 43--51. \bibitem{ErdosManitoba} P.~Erd\H{o}s, \emph{Problems and results on number theoretic properties of consecutive integers and related questions}, in: Proc. Fifth Manitoba Conf. Numer. Math. (1975), 1976, 25--44. \bibitem{Erdos76} P.~Erd\H{o}s, \emph{Problems and results on consecutive integers}, Publ. Math. Debrecen \textbf{23} (1976), 272--282. \bibitem{ErdosGraham} P.~Erd\H{o}s and R.~L.~Graham, \emph{Old and New Problems and Results in Combinatorial Number Theory}, Monographies de L'Enseignement Mathématique \textbf{28}, Geneva, 1980. \bibitem{ErdosSelfridge} P.~Erd\H{o}s and J.~L.~Selfridge, \emph{The product of consecutive integers is never a power}, Illinois J. Math. \textbf{19} (1975), 292--301. \bibitem{Golomb} S.~W.~Golomb, \emph{Powerful numbers}, Amer. Math. Monthly \textbf{77} (1970), 848--852. \bibitem{HBAnnals} D.~R.~Heath-Brown, \emph{The density of rational points on curves and surfaces}, Ann. of Math. (2) \textbf{155} (2002), 553--595. \bibitem{HBLNM} D.~R.~Heath-Brown, \emph{Counting rational points on algebraic varieties}, in: Analytic Number Theory (C.I.M.E. Summer School, Cetraro, 2002), Lecture Notes in Math. \textbf{1891}, Springer, 2006, 51--95; author's manuscript: Oxford Mathematical Institute eprint 676. \bibitem{HBsqfree} D.~R.~Heath-Brown, \emph{Square-free values of $n^2+1$}, Acta Arith. \textbf{155} (2012), 1--13; arXiv:1010.6217. \bibitem{Reuss} Thomas Reuss, \emph{Pairs of $k$-free numbers, consecutive square-full numbers}, arXiv:1212.3150v2 (2014). \bibitem{ReussThesis} Thomas Reuss, \emph{The Determinant Method and Applications}, DPhil thesis, University of Oxford, 2015. \bibitem{Stewart} C.~L.~Stewart, \emph{On the number of solutions of polynomial congruences and Thue equations}, J.~Amer. Math. Soc. \textbf{4} (1991), 793--835. \bibitem{Tao} T.~Tao, \emph{Products of consecutive integers with unusual anatomy}, arXiv:2603.27990v3 (2026). \bibitem{Walker} D.~T.~Walker, \emph{Consecutive integer pairs of powerful numbers and related Diophantine equations}, Fibonacci Quart. \textbf{14} (1976), 111--116. \bibitem{WongcharoenbhornMeemark} W.~Wongcharoenbhorn and Y.~Meemark, \emph{Square-full values of quadratic polynomials}, Bull. Aust. Math. Soc. \textbf{113} (2026), 40--54; arXiv:2405.06968. \end{thebibliography} \end{document}