\documentclass[11pt]{amsart} \usepackage{amsmath} \usepackage{amssymb} \usepackage{amsthm} \usepackage{booktabs} \usepackage{array} \usepackage{hyperref} \newtheorem{theorem}{Theorem} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{proposition}{Proposition}[section] \newtheorem{lemma}[proposition]{Lemma} \newtheorem{htheorem}[proposition]{Theorem} \newtheorem{pcorollary}[proposition]{Corollary} \theoremstyle{definition} \newtheorem{definition}[proposition]{Definition} \newtheorem{conjecture}[proposition]{Conjecture} \theoremstyle{remark} \newtheorem{remark}[proposition]{Remark} \numberwithin{equation}{section} \renewcommand{\theenumi}{\roman{enumi}} \renewcommand{\labelenumi}{(\theenumi)} \setlength{\emergencystretch}{2em} \raggedbottom \newcommand{\NN}{\mathbb{N}} \newcommand{\ZZ}{\mathbb{Z}} \newcommand{\QQ}{\mathbb{Q}} \newcommand{\RR}{\mathbb{R}} \newcommand{\CC}{\mathbb{C}} \newcommand{\DD}{\mathbb{D}} \newcommand{\TT}{\mathbb{T}} \newcommand{\fm}{\mathfrak{m}} \newcommand{\lcm}{\operatorname{lcm}} \newcommand{\den}{\operatorname{den}} \newcommand{\Sol}{\operatorname{Sol}} \newcommand{\Res}{\operatorname{Res}} \newcommand{\file}[1]{{\def\_{\textunderscore\allowbreak}\texttt{#1}}} % The public repository (Lean proofs in its root, certificates in certificates/). \newcommand{\repourl}{\url{https://github.com/telaviv/log-product-conics}} % Macros specific to this paper. \newcommand{\CT}{\operatorname{CT}} \newcommand{\Hs}[1]{\partial^{[#1]}} \newcommand{\Qbar}{\overline{\QQ}} \newcommand{\gn}[1]{\lVert #1\rVert} \newcommand{\Hh}{\mathcal H} \newcommand{\cF}{\mathcal F} \newcommand{\sha}[1]{{\footnotesize\texttt{#1}}} \begin{document} \hypersetup{pdftitle={The product log(1+1/N) log(1+2/N) and rational conics through two logarithms},pdfauthor={Shawn Krisman}} \title[Rational conics through two logarithms]{The product $\log(1+1/N)\log(1+2/N)$ and rational conics through two logarithms} \author{Shawn Krisman} \subjclass[2020]{Primary 11J72; Secondary 11J82, 41A21, 39A06, 65G20} \keywords{irrationality, linear independence, products of logarithms, Hermite--Pad\'e approximation, linear recurrences, creative telescoping, strong four exponentials conjecture, computer-assisted proof, interval arithmetic} \begin{abstract} Put $\lambda_1=\log(1+1/N)$ and $\lambda_2=\log(1+2/N)$. We prove that $\lambda_1\lambda_2$ is irrational for every integer $N\ge84$ and every even $N\ge52$, and more strongly that $1,\lambda_1,\lambda_2,\lambda_1\lambda_2$ are linearly independent over $\QQ$ on the same range. We also prove that $1,\lambda_1,\lambda_2,\lambda_2^2$ are $\QQ$-linearly independent for $N\ge58$, that $1,\lambda_1,\lambda_2,\lambda_1\lambda_2,\lambda_2^2$ are for $N\ge88$, and that $1,\lambda_1,\lambda_2,\lambda_1^2,\lambda_1\lambda_2,\lambda_2^2$ are for $N\ge151$: for $N\ge151$ the point $(\lambda_1,\lambda_2)$ lies on no curve of degree at most two defined over $\QQ$. These are instances, not conditional on the conjecture and with rational coefficients, of what the strong four exponentials conjecture predicts for two logarithms. To our knowledge no published result states any of them for $N$ below about $4.5\cdot10^{11}$. The proofs use explicit two-level Rodrigues forms of type~II Hermite--Pad\'e type, linear recurrences in the index proved by exact creative-telescoping certificates, Pituk's theorem on the growth of solutions of Poincar\'e-type recurrences (read only in restatements, so all results are stated modulo it), exact or ball-arithmetic certificates for the roots of the characteristic polynomials, and a $2$-adic content lemma at odd points that lowers the threshold for even $N$. The irrationality of $\lambda_1\lambda_2$ and the four-term independence for $N\ge103$, and the six-term (conic) independence for $N\ge151$, are also checked in Lean~4 with Mathlib, by arguments that do not use Pituk's theorem. \end{abstract} \maketitle % =========================================================================== \section{Introduction}\label{sec:intro} % =========================================================================== Throughout, $N\ge1$ is an integer and \[ \lambda_1=\log\Bigl(1+\frac1N\Bigr),\qquad \lambda_2=\log\Bigl(1+\frac2N\Bigr),\qquad \theta_N=\lambda_1\lambda_2 . \] For $N=1$ this is $\log2\cdot\log3$, whose irrationality is open. For even $N=2m$ it is $\log(1+1/m)\log(1+1/(2m))$, the product of the logarithms at a pair of points of ratio two. \subsection{Results} \begin{theorem}[modulo Pituk's theorem]\label{thm:ratio2} $\theta_N=\log(1+1/N)\log(1+2/N)$ is irrational for every integer $N\ge84$ and for every even integer $N\ge52$. \end{theorem} For example, $\log(85/84)\log(43/42)$ and $\log(53/52)\log(27/26)$ are irrational. The cases not covered are $N\le51$ and the odd $N$ with $53\le N\le83$. \begin{theorem}[modulo Pituk's theorem]\label{thm:fourterm} For every integer $N\ge84$ and every even integer $N\ge52$, the four numbers $1,\lambda_1,\lambda_2,\lambda_1\lambda_2$ are linearly independent over $\QQ$. \end{theorem} \begin{theorem}[modulo Pituk's theorem]\label{thm:S2} For every integer $N\ge58$, the four numbers $1,\lambda_1,\lambda_2,\lambda_2^2$ are linearly independent over $\QQ$. \end{theorem} \begin{theorem}[modulo Pituk's theorem]\label{thm:C5b} For every integer $N\ge88$, the five numbers $1,\lambda_1,\lambda_2,\lambda_1\lambda_2,\lambda_2^2$ are linearly independent over $\QQ$. \end{theorem} \begin{theorem}[modulo Pituk's theorem]\label{thm:C6} For every integer $N\ge151$, the six numbers $1$, $\lambda_1$, $\lambda_2$, $\lambda_1^2$, $\lambda_1\lambda_2$, $\lambda_2^2$ are linearly independent over $\QQ$. Equivalently, no nonzero polynomial with rational coefficients and of degree at most $2$ vanishes at $(\lambda_1,\lambda_2)$. \end{theorem} Theorem~\ref{thm:fourterm} contains Theorem~\ref{thm:ratio2}. Theorem~\ref{thm:C5b} contains Theorem~\ref{thm:S2} and the four-term statement of Theorem~\ref{thm:fourterm} for $N\ge88$, and Theorem~\ref{thm:C6} contains all the others for $N\ge151$. Each theorem is proved with its own family of forms, and each range is the full range that its family reaches with the proved denominator bounds (Section~\ref{sec:limits}). All results are proved modulo Pituk's theorem, which we have read only in restatements (Section~\ref{sec:status}); Theorems~\ref{thm:ratio2} and~\ref{thm:fourterm} for $N\ge103$ and Theorem~\ref{thm:C6} are also checked in Lean without it. \begin{corollary}\label{cor:examples} \begin{enumerate} \item For $N\ge84$ and for even $N\ge52$: $\lambda_1\lambda_2\notin\QQ+\QQ\lambda_1+\QQ\lambda_2$; in particular \[ \frac{\log(1+1/N)\log(1+2/N)}{\log(1+1/(N+1))}\notin\QQ . \] For $m\ge26$ the numbers $1$, $\log(1+1/m)$, $\log(1+1/(2m))$, $\log(1+1/m)\log(1+1/(2m))$ are $\QQ$-linearly independent. \item For $N\ge58$: $\log^2(1+2/N)/\log(1+1/N)\notin\QQ$. For $m\ge29$: $\log^2(1+1/m)/\log(1+1/(2m))\notin\QQ$. \item For $N\ge88$: $\lambda_1\lambda_2\notin\QQ+\QQ\lambda_1+\QQ\lambda_2+\QQ\lambda_2^2$. \item For $N\ge151$: $\log^2(1+1/N)/\log(1+2/N)\notin\QQ$, and $\lambda_1^2-\lambda_1\lambda_2+\lambda_2^2\notin\QQ+\QQ\lambda_1+\QQ\lambda_2$. \end{enumerate} \end{corollary} \begin{proof} (i) follows from Theorem~\ref{thm:fourterm}, since $\log(1+1/(N+1))=\lambda_2-\lambda_1$, and for $N=2m$ we have $\lambda_1=\log(1+1/(2m))$, $\lambda_2=\log(1+1/m)$. (ii) A relation $\lambda_2^2=c\lambda_1$ with $c\in\QQ$ contradicts Theorem~\ref{thm:S2}; for $N=2m$, $\lambda_2^2/\lambda_1=\log^2(1+1/m)/\log(1+1/(2m))$. (iii) and (iv) follow from Theorems~\ref{thm:C5b} and~\ref{thm:C6}. \end{proof} \subsection{Prior work and what is new}\label{sec:prior} In the coordinates $q=N+1$ of Hata \cite{Hat98} put $f_1=\log(1+1/q)$ and $f_2=\log(1-1/q)$. Then $\lambda_1=-f_2$ and $\lambda_2=f_1-f_2$, so \[ \theta_N=f_2^2-f_1f_2,\qquad\text{while}\qquad f_1f_2=\lambda_1^2-\lambda_1\lambda_2 . \] Hata \cite[Cor.~6.3]{Hat98} proved that $1,f_1,f_2,f_1f_2$ are $\QQ$-linearly independent for every integer $q\ge54$; Lysov \cite{Lys18} proved $f_1f_2\notin\QQ$ for $q\ge33$; and in \cite{Kri26} we extended the four-term statement to $26\le q\le53$, conditionally on the holonomy bound \cite[Thm.~7.0.1]{CDT} of Calegari, Dimitrov and Tang (a preprint). The number $f_1f_2=\lambda_1^2-\lambda_1\lambda_2$ lies on a different hyperbola from $\theta_N=\lambda_1\lambda_2$, so none of these results covers $\theta_N$; Theorem~\ref{thm:C6} contains Hata's four-term statement for $q\ge152$, which is not new. Calegari, Dimitrov and Tang \cite[Thm.~C]{CDT} proved that $1,\log(1+1/m),\log(1+1/n),\log(1+1/m)\log(1+1/n)$ are linearly independent for all integers $m\ne n$ in $\ZZ\setminus\{-1,0\}$ with $|m/n-1|<10^{-6}$; in \cite{Kri26} we widened this to $3\cdot10^{-6}$, conditionally on parts of \cite{CDT} (Thm.~6.0.2, Cor.~9.0.19, Lemma~12.0.1 and the construction of \S14, including an irreducibility asserted there). For even $N=2m$, $\theta_N$ is the product at the pair $(m,2m)$, of ratio $2$. Kawashima \cite{Kaw26} proved a general linear independence criterion for products of polylogarithms (Theorem~2.2 and Corollary~2.3 there; the preprint is not refereed). With $K=\QQ$, $m=r=2$ (so $M=(m+1)^r-1=8$), $\boldsymbol\alpha=(1,-1)$ and $\beta=N+1$, the quantity $V_{v_0}(\boldsymbol\alpha,\beta)$ of \cite[eq.~(V)]{Kaw26} is, in our reading of the LaTeX source of version~1 (we have not checked the proof in \cite{Kaw26}), \begin{align*} V&=(M+1)\log(N+1)-M\log(N+1)\\ &\qquad-\Bigl(M\log2+\tfrac{r(r+1)}2\log(m+1)+r+rM\Bigr)\\ &=\log(N+1)-26.841\ldots, \end{align*} with $26.841\ldots=8\log2+3\log3+2+16$. Since $\mathrm{Li}_1(1/(N+1))=\lambda_1$ and $\mathrm{Li}_1(-1/(N+1))=\lambda_1-\lambda_2$, Corollary~2.3 of \cite{Kaw26} then gives the statement of Theorem~\ref{thm:C6}, and hence of all our theorems, for $N+1>e^{26.841}\approx4.54\cdot10^{11}$. (The choice $\boldsymbol\alpha=(-1,-2)$, $\beta=N$ gives $V=\log N-13\log2-26.841$, hence $N>3.7\cdot10^{15}$.) Only the height terms of \cite[eq.~(V)]{Kaw26} carry the factor $M$; the constant bracket does not. Viola and Zudilin \cite{VZ18} proved that $1,\mathrm{Li}_1(1/z),\mathrm{Li}_2(1/z),\mathrm{Li}_2(1/(1-z))$ are linearly independent over $\QQ$ for every integer $z\ge9$ or $z\le-8$, and for rationals $z=s/r$ or $z=1-s/r$ with $1e^{185}$, and the corrected value is the one given in Section~\ref{sec:prior}. A further AI referee read this paper as a whole (pass with fixes, no fatal or major finding); its fixes are applied. No human expert has read the proofs. See Section~\ref{sec:provenance}. \subsection{Method} All five theorems are proved by one argument, set out in Section~\ref{sec:framework}. For a point $z=N+1$ and a set $\cF$ of target functions (logarithms and products of two logarithms with branch points in $\{-1,0,1\}$) we use explicit polynomials \[ Q_n(t)=\frac1{n!}\frac{d^n}{dt^n}\Bigl[W(t)^n\,\frac1{(2n)!}\frac{d^{2n}}{dt^{2n}}V(t)\Bigr],\qquad W=t^3-t, \] with $V=t^{e_0}(t-1)^{e_1}(t+1)^{e_{-1}}$ chosen so that $Q_nF-P_{F,n}=O(t^{-n-1})$ for every $F\in\cF$, where $P_{F,n}$ is the polynomial part of $Q_nF$ at infinity. The remainders $R_{F,n}(z)$ satisfy a linear recurrence in $n$ with polynomial coefficients. We do not estimate them by a saddle point analysis: an elementary a priori bound excludes the one large characteristic root, Pituk's theorem identifies the growth rate with the modulus of a characteristic root, and a root certificate shows that the remaining roots are smaller than the inverse of the denominators. Backward recursion (the trailing coefficient of the recurrence does not vanish) and one nonzero determinant then give linear independence. The recurrences are proved by exact creative-telescoping certificates. The five theorems differ only in the parameters and certificates (Table~\ref{tab:params}). \subsection*{Outline} Section~\ref{sec:s4} places the results relative to the strong four exponentials conjecture. Section~\ref{sec:framework} is the framework. Sections~\ref{sec:A}--\ref{sec:C6} treat the five families: the ratio-two forms (Theorems~\ref{thm:ratio2} and~\ref{thm:fourterm} for $N\ge103$), the $2$-adic gain (even $N\ge52$), the shape $(2;1,2,\frac54)$ ($84\le N\le102$), and Theorems~\ref{thm:S2}, \ref{thm:C5b}, \ref{thm:C6}. Section~\ref{sec:computer} lists the certificates, Section~\ref{sec:limits} the limits of the method and open problems, and Section~\ref{sec:provenance} provenance and the use of AI tools. % =========================================================================== \section{Relation to the strong four exponentials conjecture}\label{sec:s4} % =========================================================================== Let $\mathcal L$ be the $\QQ$-vector space of logarithms of nonzero algebraic numbers and $\widetilde{\mathcal L}= \Qbar+\Qbar\mathcal L$. \begin{conjecture}[strong four exponentials; {\cite[Conj.~1.5]{W05}}, suggested by D.~Roy]\label{conj:S4} Let $M$ be a $2\times2$ matrix with entries in $\widetilde{\mathcal L}$ whose rows are linearly independent over $\Qbar$ and whose columns are linearly independent over $\Qbar$. Then $M$ has rank $2$. \end{conjecture} (We take the statement and the attribution from the author's copy of \cite{W05}, \S1. The best known result in this direction is Roy's strong six exponentials theorem \cite[\S4, Cor.~2]{Roy92}, as quoted in \cite[Thm.~2.1]{W05}.) \begin{proposition}\label{prop:s4conic} Assume Conjecture~\ref{conj:S4}. Let $\ell_1,\ell_2\in\mathcal L$ be linearly independent over $\QQ$. Then $1,\ell_1, \ell_2,\ell_1^2,\ell_1\ell_2,\ell_2^2$ are linearly independent over $\Qbar$. \end{proposition} \begin{proof} By Baker's theorem on linear forms in logarithms (see \cite{BW07}), $1,\ell_1,\ell_2$ are linearly independent over $\Qbar$. Suppose $q(1,\ell_1,\ell_2)=0$ for a nonzero quadratic form $q\in\Qbar[X_0,X_1,X_2]$. If $q$ has rank $\le2$, then $q=uv$ for linear forms $u,v$ over $\Qbar$, and $u(1,\ell_1,\ell_2)=0$ or $v(1,\ell_1,\ell_2)=0$ contradicts Baker's theorem. If $q$ has rank $3$, then over $\Qbar$ we can write $q=u_1u_4-u_2u_3$ with linear forms $u_i$. The matrix $\bigl(\begin{smallmatrix}u_1&u_2\\u_3&u_4 \end{smallmatrix}\bigr)$ evaluated at $(1,\ell_1,\ell_2)$ has entries in $\widetilde{\mathcal L}$ and determinant $0$. By Conjecture~\ref{conj:S4} its rows or its columns are $\Qbar$-dependent; say $a(u_1,u_2)+b(u_3,u_4)$ vanishes at $(1,\ell_1,\ell_2)$ with $(a,b)\ne0$. By Baker's theorem $au_1+bu_3$ and $au_2+bu_4$ vanish identically, so $q=u_1u_4-u_2u_3\equiv0$ as a polynomial (if, say, $b\ne0$, substitute $u_3=-au_1/b$, $u_4=-au_2/b$), a contradiction. The case of dependent columns is the same. \end{proof} For $N\ge1$ the logarithms $\lambda_1=\log((N+1)/N)$ and $\lambda_2=\log((N+2)/N)$ are $\QQ$-linearly independent: if $a\lambda_1=b\lambda_2$ with integers $a,b$, then $(N+1)^aN^b=(N+2)^bN^a$; a prime dividing $N+1\ge2$ divides neither $N$ nor $N+2$, so $a=0$, and then $b=0$. So Conjecture~\ref{conj:S4} predicts, for every $N\ge1$, the statement of Theorem~\ref{thm:C6} with $\Qbar$ in place of $\QQ$. Following the usual description of plane conics, call a relation $\lambda_1\lambda_2\in\Qbar+\Qbar\lambda_1+\Qbar\lambda_2$ a ``hyperbola'' and a relation $\lambda_2^2\in \Qbar+\Qbar\lambda_1+\Qbar\lambda_2$ a ``parabola'' through the point $(1:\lambda_1:\lambda_2)$. Theorems~\ref{thm:ratio2} and~\ref{thm:fourterm} exclude rational hyperbolas of the first shape, Theorem~\ref{thm:S2} rational parabolas of the second, and Theorem~\ref{thm:C6} every rational conic and line. For $N=1$ the hyperbola case contains $\log2\cdot\log3\notin\QQ$. Our theorems are therefore unconditional explicit instances of consequences of Conjecture~\ref{conj:S4}, with rational (not algebraic) coefficients and for $N$ above explicit thresholds. They are not progress on the conjecture itself. The method is an explicit Hermite--Pad\'e construction at one point; its success depends on the point $z=N+1$ being far from the cut $[-1,1]$ compared with the arithmetic cost of the forms, and it fails by a wide margin at $N=1$ (Section~\ref{sec:limits}). % =========================================================================== \section{The framework}\label{sec:framework} % =========================================================================== \subsection{Coordinates}\label{sec:coords} Put $z=N+1$ and, for $t\in\CC\setminus[-1,1]$ (principal branches), \begin{gather*} f_1(t)=\log\Bigl(1+\frac1t\Bigr),\qquad f_2(t)=\log\Bigl(1-\frac1t\Bigr),\\ L(t)=f_1(t)-f_2(t)=\log\frac{t+1}{t-1},\qquad \theta(t)=-f_2(t)L(t). \end{gather*} These are analytic on $\CC\setminus[-1,1]$ and real on $(1,\infty)$; $f_1,f_2,L$ are $O(1/t)$ and $\theta=f_2^2-f_1f_2$ is $O(t^{-2})$ at $\infty$. The branch points are \[ B_{f_1}=\{0,-1\},\qquad B_{f_2}=\{0,1\},\qquad B_L=\{-1,1\}. \] At $t=z$ we have $f_2(z)=\log(N/(N+1))=-\lambda_1$, $L(z)=\log((N+2)/N)=\lambda_2$ and $f_1(z)=\lambda_2-\lambda_1$. Hence \begin{gather*} \lambda_1=-f_2(z),\quad\lambda_2=L(z),\quad\lambda_1\lambda_2=\theta(z),\\ \lambda_2^2=L(z)^2,\quad\lambda_1^2=f_2(z)^2,\quad -\lambda_1\lambda_2=(f_2L)(z), \end{gather*} and each of the lists in Theorems~\ref{thm:fourterm}--\ref{thm:C6} spans the same $\QQ$-space as $1$ together with the values at $z$ of the target set \begin{equation}\label{eq:targets} \begin{gathered} \cF_{\rm II}=\{f_1,f_2,\theta\},\qquad \cF_{\rm S}=\{f_1,f_2,L^2\},\\ \cF_{\rm B}=\{f_1,f_2,L^2,f_2L\},\qquad\cF_{\rm C}=\{f_1,f_2,f_1^2,f_1f_2,f_2^2\}, \end{gathered} \end{equation} respectively, by an invertible rational change of basis. For instance $(1,\lambda_1,\lambda_2,\lambda_1\lambda_2)= (1,f_1(z),f_2(z),\theta(z))T$ with $T\in\mathrm{GL}_4(\ZZ)$, $\det T=1$; and $\lambda_1^2=f_2^2$, $\lambda_1\lambda_2=f_2^2-f_1f_2$, $\lambda_2^2=f_1^2-2f_1f_2+f_2^2$ at $z$. So it suffices to prove that $1$ and the values $F(z)$, $F\in\cF$, are $\QQ$-linearly independent. Theorem~\ref{thm:ratio2} is the case $\cF=\{\theta\}$. Every target $F$ has a Laurent expansion $F=\sum_{m\ge1}\varphi_mt^{-m}$ at $\infty$. For $F\in\{f_1,f_2,L\}$, $\varphi_m\in\ZZ\cdot\frac1m$ (indeed $\pm\frac1m$ or $\frac2m$); for a product $F=gh$ of two of them, $\varphi_m\in\sum_{a+b=m}\ZZ\cdot\frac1{ab}$. We call these targets of linear and of quadratic type. \subsection{Two-level Rodrigues forms}\label{sec:forms} Write $\Hs k=\frac1{k!}\frac{d^k}{dt^k}$; it maps $\ZZ[t]$ into itself. Let $n\ge1$, let $V\in\ZZ[t]$ have degree $\delta\ge2n$, and put \begin{equation}\label{eq:forms} S=\Hs{2n}V,\qquad Q=\Hs n\bigl[W^nS\bigr]\in\ZZ[t],\qquad W=t(t-1)(t+1). \end{equation} Then $\deg S=\delta-2n$ and $\deg Q=\delta$. For a target $F$ let $P_F=[QF]_+$ be the polynomial part of $QF$ at $\infty$, and $R_F=QF-P_F$, which is analytic on $\CC\setminus[-1,1]$. In all our families $V$ is a product of powers of $t,t-1,t+1$ whose exponents are proportional to $n$, the polynomials form a sequence $Q_n$, and $Q_0=1$, $P_{F,0}=0$. If $V=V_1^n$ for a fixed $V_1$, then \begin{equation}\label{eq:CT} Q_n(z)=\CT_{h_1,h_2}\bigl[\Lambda^n\bigr],\qquad \Lambda=\frac{W(z+h_1)\,V_1(z+h_1+h_2)}{h_1h_2^2}, \end{equation} the constant term of a Laurent polynomial in $h_1,h_2$: the constant term in $h_2$ is $\Hs{2n}V(z+h_1)=S(z+h_1)$, and then the constant term in $h_1$ is $\Hs n(W^nS)(z)$. Equivalently $Q_n(z)=\Res_{u=z}W(u)^n(u-z)^{-n-1}S(u)$ with $S(u)=\Res_{v=u}V(v)(v-u)^{-2n-1}$. \subsection{Contact}\label{sec:contact} \begin{lemma}[contact]\label{lem:contact} Let $V=t^{e_0}(t-1)^{e_1}(t+1)^{e_{-1}}$ with $\deg V\ge2n$, and let $Q$ be as in \eqref{eq:forms}. \begin{enumerate} \item For every $g\in\{f_1,f_2,L\}$, $R_g(t)=O(t^{-n-1})$ at $\infty$. \item Let $g,h\in\{f_1,f_2,L\}$ (possibly $g=h$). Suppose that for every $c\in\{-1,0,1\}$: $e_c\ge2n$ if $c\in B_g\cap B_h$, and $e_c\ge n$ if $c$ lies in exactly one of $B_g,B_h$. Then $R_{gh}(t)=O(t^{-n-1})$. \end{enumerate} \end{lemma} \begin{proof} Let $\Gamma$ be a positively oriented contour around $[-1,1]$. For a target $F$ and $j\ge0$, the coefficient of $t^{-j-1}$ in $R_F$ is $[t^{-1}](t^jQF)=\frac1{2\pi i}\oint_\Gamma t^jQF\,dt$, because $t^jP_F$ is a polynomial. So it suffices to show $\oint_\Gamma t^jQF\,dt=0$ for $0\le j2n$ and $p>\sigma$ is prime, then $p\mid q_i$ for every $i\ge p-n$. If $2n\max(2n,\sigma)$. If $k\ge p-2n$, then $k,2nj}q_i\varphi_{i-j}$ with $i-j\le\deg Q=\delta$. Take a quadratic target. For $p\le2n$, $v_p(\varphi_m)\ge-2\lfloor\log_pm\rfloor\ge-2\lfloor\log_p\delta\rfloor$. For $p>2n$ we have $p^2\ge(2n+1)^2>6n\ge\delta$, so $v_p(a)\le1$ for $a\le\delta$. If $2n\sigma$ (and $p>2n$), then $2p>\sigma+2n=\delta\ge m$, so at most one of $a,b$ is divisible by $p$, $v_p(\varphi_m)\ge-1$, and $v_p(\varphi_m)<0$ forces $i\ge m\ge p$, hence $p\mid q_i$ by (i): $P_F$ is $p$-integral. For $p>\delta$ every $\varphi_m$ with $m\le\delta$ is $p$-integral. The linear case is the same with one factor. (iii) $\log D^{(2)}=2\sum_{p\le2n}\lfloor\log_p\delta\rfloor\log p+\sum_{2n1$ and $M_r\ge\max\{|\Lambda_t(h_1,h_2)|:|t|\le r,\ |h_1|=|h_2|=1\}$. Then for every real $z>r$ and every $k\ge0$, \[ |R_{F,k}(z)|\le C_{F,r}(z)\Bigl(\frac{r^\varsigma M_r}{z^\varsigma}\Bigr)^k,\qquad C_{F,r}(z)=\frac{r\max_{|t|=r}|F(t)|}{z-r} . \] In particular $\limsup_k|R_{F,k}(z)|^{1/k}\le r^\varsigma M_r/z^\varsigma$. \end{lemma} \begin{proof} The function $t^{\varsigma k}R_{F,k}(t)$ is analytic for $|t|\ge r$ and $O(1/t)$ at $\infty$, so by Cauchy's formula for the exterior of the circle, $z^{\varsigma k}R_{F,k}(z)=\frac1{2\pi i}\oint_{|t|=r}t^{\varsigma k}R_{F,k}(t)\frac{dt}{z-t}$. The part $t^{\varsigma k}P_{F,k}(t)/(z-t)$ is analytic in $|t|\le r$ and integrates to $0$, so $R_{F,k}(z)=\frac1{2\pi i}\oint_{|t|=r}Q_k(t)F(t)(t/z)^{\varsigma k}\frac{dt}{z-t}$. Cauchy's estimate for the constant term on $|h_1|=|h_2|=1$ gives $|Q_k(t)|\le M_r^k$ for $|t|\le r$. \end{proof} \subsection{Recurrences}\label{sec:rec} In each family, $Q_k(z)$ satisfies a linear recurrence \begin{equation}\label{eq:rec} \sum_{i=0}^{\varrho}C_i(k,z)\,Q_{k+i}(z)=0\qquad(k\ge0), \end{equation} an identity in $\QQ[z]$ for each integer $k\ge0$, with $C_i\in\ZZ[k,z]$. We write $\mathsf L$ for the operator $(X_k)\mapsto(\sum_iC_i(k,z)X_{k+i})_k$. \begin{lemma}[transfer]\label{lem:transfer} Assume \eqref{eq:rec} holds for all $k\ge0$, that $\deg_zC_i\le\zeta_i$, and that $R_{F,k}=O(z^{-\varsigma k-1})$ for a target $F$. Put $\mu=\max_i(\zeta_i-\varsigma i)$. Then $\mathsf L(P_F)_k=0$ and $\mathsf L(R_F)_k=0$ in $\QQ[z]$ for every $k$ with $\varsigma k+1>\mu$. If moreover $\mathsf L(P_F)_k=0$ is verified for the finitely many remaining $k$, then $\mathsf L(P_F)=\mathsf L(R_F)=0$ for all $k\ge0$, as identities in $z$, hence at every point $z$. \end{lemma} \begin{proof} As Laurent series in $1/z$, $\mathsf L(R_F)_k=\mathsf L(Q)_kF-\mathsf L(P_F)_k=-\mathsf L(P_F)_k$ is a polynomial in $z$, and its $i$-th term is $O(z^{\zeta_i-\varsigma(k+i)-1})$, so it is $O(z^{\mu-\varsigma k-1})$ and vanishes when $\mu-\varsigma k-1<0$. \end{proof} The recurrences \eqref{eq:rec} are proved by one of two kinds of certificate. \emph{(a) A bivariate telescoping certificate} (Section~\ref{sec:A}). With $\Lambda$ as in \eqref{eq:CT}, $\vartheta_1=h_1\partial_{h_1}$ and $\vartheta_2=h_2\partial_{h_2}$, Laurent polynomials $A,B$ in $(h_1,h_2)$ over $\QQ[n,z]$ and a nonzero $d(n)\in\ZZ[n]$ with \begin{equation}\label{eq:cert2} d(n)\sum_{i=0}^{\varrho}c_i(n,z)\Lambda^{i+1}=n\,\vartheta_1(\Lambda)A+\Lambda\,\vartheta_1(A)+n\,\vartheta_2(\Lambda)B+ \Lambda\,\vartheta_2(B). \end{equation} Multiplying by $\Lambda^{n-1}$ ($n\ge1$) gives $d(n)\sum_ic_i\Lambda^{n+i}=\vartheta_1(\Lambda^nA)+\vartheta_2(\Lambda^nB)$, and the constant term of $\vartheta_1$ or $\vartheta_2$ of any Laurent polynomial is $0$. So $d(n)\sum_ic_iQ_{n+i}(z)=0$, which is \eqref{eq:rec} for every $n\ge1$ with $d(n)\ne0$. \emph{(b) A two-step certificate} (Sections~\ref{sec:A125}--\ref{sec:C6}). Write $S_k(u)=\Res_{v=u}\Phi$ with $\Phi=V_k(v)(v-u)^{-2n-1}$ (here $V_k=V$ for index $k$), $w=v-u$ and $\nabla g=\partial_v(g\Phi)/\Phi$, so that $\Res_{v=u}(\nabla g)\Phi=0$ for $g$ meromorphic at $v=u$, and $\partial_u^j\Phi=(2n+1)_jw^{-j}\Phi$. \emph{Step A} consists of two identities of rational functions in $(k,u,v)$, of the form $\sum_ja_j(k,u)(2n+1)_jw^{-j}=\nabla(\cdot)$ and $\Phi_{k+1}/\Phi_k-\sum_jb_j(k,u)(2n+1)_jw^{-j}=\nabla(\cdot)$, where $\Phi_k$ is $\Phi$ at index $k$; taking residues at $v=u$ gives a differential equation $\sum_ja_jS_k^{(j)}=0$ and a shift relation $S_{k+1}=\sum_jb_jS_k^{(j)}$. \emph{Step B} uses the basis $B_{(a,\ell)}=E_k(u-z)^aS_k^{(\ell)}$, $E_k=W(u)^{n}(u-z)^{-n-1}$, of a module over $\QQ(k,z)$ ($0\le a\le2$, $\ell$ below the order of the differential equation), and identities \[ E_{k+1}(u-z)^aS_{k+1}^{(\ell)}=\sum_{(a',\ell')}A_{(a,\ell),(a',\ell')}(k,z)B_{(a',\ell')}+\partial_u\Bigl(E_k\sum_l \gamma_lS_k^{(l)}\Bigr), \] verified as identities of rational functions in $(k,z,u)$ after reduction by the differential equation. Taking residues at $u=z$ gives $Y_{k+1}(z)=A(k,z)Y_k(z)$ for the vector $Y_k=(\Res_{u=z}B_{(a,\ell)})$, whose first entry is $Q_k(z)$. \emph{Step C}: with the row vectors $r_i(k)=e_{(0,0)}A(k+i-1)\cdots A(k)$ we have $Q_{k+i}(z)=r_i(k)Y_k(z)$, and the scalar recurrence is certified by the identity $\sum_iC_i(k,z)r_i(k)=0$ in $\QQ(k,z)^{\dim}$. The finitely many $k$ at which a denominator of $A$ vanishes are covered by direct exact checks of \eqref{eq:rec} on the explicit polynomials. In both cases the search for the certificate (modular linear algebra, interpolation, Chinese remaindering, rational reconstruction) is not part of the proof; only the final exact identity check is. \subsection{Pituk's theorem and the criterion}\label{sec:criterion} \begin{htheorem}[Pituk {\cite[Thm.~2]{Pit02}}; restated in {\cite[Thm.~6.2]{HMW24}}, {\cite[\S4]{IW23}}]\label{thm:pituk} Let $x(k+\varrho)+\sum_{i<\varrho}p_i(k)x(k+i)=0$ for all large $k$, with $p_i(k)\to p_i\in\CC$. Then either $x(k)=0$ for all large $k$, or $\limsup_k|x(k)|^{1/k}=|\lambda|$ for some root $\lambda$ of $\lambda^\varrho+\sum_{i<\varrho}p_i\lambda^i$. \end{htheorem} No condition on $p_0$ or on the moduli of the roots is needed; $\limsup$ cannot be replaced by $\lim$ \cite{HMW24}. \begin{htheorem}[criterion]\label{thm:criterion} Let $z\ge2$ be an integer and $\cF=\{F_1,\dots,F_s\}$ a set of targets. Suppose that for $k\ge0$ there are polynomials $Q_k\in\ZZ[t]$ with $Q_0=1$, and that, writing $P_{j,k}=P_{F_j,k}(z)$ and $R_{j,k}=R_{F_j,k}(z)$: \begin{enumerate} \item[(H1)] \emph{(integrality)} there are integers $D_k\ge1$ and $e_k\ge0$ such that $D_kQ_k(z)$ and all $D_kP_{j,k}$ are integers divisible by $2^{e_k}$, and $\frac1k(\log D_k-e_k\log2)\to\gamma$; \item[(H2)] \emph{(recurrence)} the sequences $Q_k(z)$ and $P_{j,k}$ satisfy \eqref{eq:rec} for all $k\ge0$, where $C_i(k)=C_i(k,z)\in\QQ[k]$ have degree $\le d$ in $k$, $C_\varrho$ has degree exactly $d$, and $C_0(k)\ne0$, $C_\varrho(k)\ne0$ for every integer $k\ge0$; let $\chi(\lambda)=\sum_i[k^d]C_i\cdot\lambda^i$; \item[(H3)] \emph{(a priori bound)} $\limsup_k|R_{j,k}|^{1/k}\le\beta$ for every $j$; \item[(H4)] \emph{(roots)} every root $\lambda$ of $\chi$ satisfies $|\lambda|<\rho$ or $|\lambda|>\beta$, where $\rho\le e^{-\gamma}$; \item[(H5)] \emph{(determinant)} $\det\bigl[Q_k(z),P_{1,k},\dots,P_{s,k}\bigr]_{k=0,\dots,s}\ne0$. \end{enumerate} Then $1,F_1(z),\dots,F_s(z)$ are linearly independent over $\QQ$. If $s=1$ and $F_1(z)\ne0$, (H5) may be omitted. \end{htheorem} \begin{proof} Suppose $a+\sum_jb_jF_j(z)=0$ with integers $a,b_j$. Put $X_k=\sum_jb_jR_{j,k}$. Since $R_{j,k}=Q_k(z)F_j(z)-P_{j,k}$, \[ X_k=-\Bigl(aQ_k(z)+\sum_jb_jP_{j,k}\Bigr). \] By (H1), $J_k=D_kX_k/2^{e_k}$ is an integer. By (H2) and linearity, $\mathsf L(X)=0$. By (H3), $\limsup|X_k|^{1/k}\le\beta$. Divide the recurrence by $C_\varrho(k)$; the coefficients tend to $[k^d]C_i/[k^d]C_\varrho$, so Theorem~\ref{thm:pituk} applies with characteristic polynomial $\chi/[k^d]C_\varrho$. Either $X_k=0$ for all large $k$, or $\limsup|X_k|^{1/k}=|\lambda|$ for a root $\lambda$ of $\chi$ with $|\lambda|\le\beta$; by (H4) then $|\lambda|<\rho\le e^{-\gamma}$, and $\limsup\frac1k\log|J_k|\le\gamma+\log|\lambda|<0$, so the integers $J_k$, and hence $X_k$, vanish for all large $k$. In either case $X_k=0$ for $k\ge k_1$. Since $C_0(k)\ne0$ for every $k\ge0$, the recurrence at $k=k_1-1,\dots,0$ gives $X_k=0$ for all $k\ge0$. Then $aQ_k(z)+\sum_jb_jP_{j,k}=0$ for $k=0,\dots,s$, and (H5) gives $a=b_j=0$. If $s=1$, then $X_0=b_1R_{1,0}= b_1F_1(z)$ since $Q_0=1$ and $P_{1,0}=0$; so $b_1=0$ and then $a=0$. \end{proof} In (H4) we verify, in every family, two counts: $\chi$ has exactly $\varrho-1$ roots in $|\lambda|<\rho$, and exactly $\varrho-1$ roots in $|\lambda|<\beta'$ for some $\beta'>\beta$. These imply (H4): if $\rho<\beta'$ the two discs hold the same $\varrho-1$ roots and the last root has modulus $\ge\beta'>\beta$; if $\beta'\le\rho$, the last root lies outside both discs. \subsection{Root certificates}\label{sec:rootcert} Three methods are used. (a) \emph{Rouch\'e with positivity}: on a circle $|\lambda|=\rho$ one coefficient term of $\chi_z$ dominates the sum of the moduli of the others; for $z$ in a half-line this is checked by positivity of the coefficients of a polynomial in $s=z-z_0$. (b) \emph{Arb boxes}: on a bounded $z$-interval, ball arithmetic shows $\chi_z(\lambda)\ne0$ on $|\lambda|=\rho$ over a finite cover of $[z_1,z_2]\times[0,2\pi)$ by boxes; since the leading coefficient of $\chi_z$ is a nonzero constant, the roots move continuously and the count is constant on the interval, and it is taken at one point by certified root isolation. (c) \emph{Resultants}: \begin{lemma}\label{lem:resultant} Let $\chi_z(\lambda)=\sum_{i=0}^\varrho a_i(z)\lambda^i$ with $a_i\in\RR[z]$ and $a_\varrho$ a nonzero constant, let $\rho>0$, and let $G(z)=\Res_\lambda\bigl(\chi_z(\lambda),\lambda^\varrho\chi_z(\rho^2/\lambda)\bigr)$. If $G$ has no zero on an interval $I\subset\RR$, then for $z\in I$ no root of $\chi_z$ lies on $|\lambda|=\rho$, and the number of roots in $|\lambda|<\rho$ is constant on $I$. \end{lemma} \begin{proof} If $\chi_z(\lambda)=0$ with $|\lambda|=\rho$ and $z$ real, then $\chi_z(\bar\lambda)=0$ and $\bar\lambda=\rho^2/\lambda$, so $\lambda$ is a common root of the two polynomials and $G(z)=0$. The roots depend continuously on $z$ because $a_\varrho$ is a nonzero constant. \end{proof} If all coefficients of $G(z_0+s)$ have the same sign and $G(z_0)\ne0$, then $G$ has no zero on $[z_0,\infty)$. For a radius $\mu/z$ we apply the lemma to $z^\varrho\chi_z(\nu/z)$, a polynomial in $\nu$ with polynomial coefficients and constant leading coefficient, on $|\nu|=\mu$. \subsection{The five families}\label{sec:table} Table~\ref{tab:params} lists the parameters; the remaining sections supply the certificates. In every family the contact order is $\varsigma k$ with $\varsigma=1$ (index $n$), except the shape $(2;1,2,\frac54)$, where the natural index is $m=n/4$ and $\varsigma=4$. The ``margin'' is $\gamma+\log\max|\lambda_i|$ at the least admissible $z$, where the maximum is over the small roots; it enters only through limits. \begin{table}[ht] \caption{The five families. Here $z=N+1$; $\gamma$ is the cost per index (Lemma~\ref{lem:denom}, minus $\log2$ in the $2$-adic case); $\rho\le e^{-\gamma}$ and $\beta'$ are the certified radii of (H4).}\label{tab:params} \small\setlength{\tabcolsep}{4pt} \begin{tabular}{llllll} \toprule family & section & $V$ & $\cF$ & order & $\gamma$\\ \midrule II & \S\ref{sec:A} & $t^n(t-1)^{2n}(t+1)^n$ & $\cF_{\rm II}$ & 5 & 4\\ II$_2$ & \S\ref{sec:twoadic} & same & $\cF_{\rm II}$ & 5 & $4-\log2$\\ II$'$ & \S\ref{sec:A125} & $t^{4m}(t-1)^{8m}(t+1)^{5m}$ & $\cF_{\rm II}$ & 6 & 17\\ S & \S\ref{sec:S2} & $(t^2-1)^{2n}$ & $\cF_{\rm S}$ & 5 & 4\\ B & \S\ref{sec:C5b} & $t^n(t^2-1)^{2n}$ & $\cF_{\rm B}$ & 6 & 5\\ C & \S\ref{sec:C6} & $(t^3-t)^{2n}$ & $\cF_{\rm C}$ & 7 & 6\\ \bottomrule \end{tabular} \medskip \begin{tabular}{lllll} \toprule family & $\rho$ & $\beta'$ & $z$-range & margin\\ \midrule II & $e^{-4}$ & $4000/z$ & $z\ge104$ & $-0.0012$\\ II$_2$ & $2e^{-4}$ & $4000/z$ & odd $z\in[53,103]$ & $-0.0313$\\ II$'$ & $e^{-17}$ & $5.8\cdot10^{14}/z^4$ & $z\in[85,103]$ & $-0.0202$\\ S & $9157/500000$ & $4700/z$ & $z\ge59$ & $-0.00018$\\ B & $673/100000$ & $14000/z$ & $z\ge89$ & $-0.0027$\\ C & $2/807$ & $42000/z$ & $z\ge152$ & $-0.00105$\\ \bottomrule \end{tabular} \end{table} The contact hypotheses of Lemma~\ref{lem:contact} hold in each case: for $\theta=-f_2L$ one needs $(e_0,e_1,e_{-1})\ge (n,2n,n)$; for $L^2$, $e_{\pm1}\ge2n$; for $f_2L$, $(e_0,e_1,e_{-1})\ge(n,2n,n)$; for $f_1^2$, $e_0,e_{-1}\ge2n$; for $f_2^2$, $e_0,e_1\ge2n$; for $f_1f_2$, $e_0\ge2n$ and $e_{\pm1}\ge n$. In the notation of the shape scans of our working notes, a family $(B;b_0,b_1,b_2)$ has $V=t^{b_0n}(t-1)^{b_1n}(t+1)^{b_2n}$ under $\partial^{Bn}$; family II is $(2;1,2,1)$ and II$'$ is $(2;1,2,\frac54)$. % =========================================================================== \section{The ratio-two forms: Theorems~\ref{thm:ratio2} and~\ref{thm:fourterm} for \texorpdfstring{$N\ge103$}{N>=103}}\label{sec:A} % =========================================================================== In this section $V_1=t(t-1)^2(t+1)$, $V=V_1^n=t^n(t-1)^{2n}(t+1)^n$, \[ S_n=\Hs{2n}V_1^n,\qquad Q_n=\Hs n\bigl[W^nS_n\bigr]\in\ZZ[t],\qquad \deg Q_n=4n, \] and $Q_n(z)=\CT_{h_1,h_2}[\Lambda^n]$ with $\Lambda=W(z+h_1)V_1(z+h_1+h_2)/(h_1h_2^2)$ by \eqref{eq:CT}. The targets are $\cF_{\rm II}=\{f_1,f_2,\theta\}$, and we write $P_{F,n}$, $R_{F,n}$; $P_n=P_{\theta,n}$ and $R_n=R_{\theta,n}$. \begin{lemma}\label{lem:A-basic} \begin{enumerate} \item $R_{F,n}(t)=O(t^{-n-1})$ for $F\in\{f_1,f_2,\theta\}$. \item Put $D_n=\prod_{p\le2n}p^{2\lfloor\log_p4n\rfloor}$ and $D_n'=\prod_{p\le2n}p^{\lfloor\log_p4n\rfloor}$, so that $D_n=D_n'^2$. Then $D_nP_{\theta,n}\in\ZZ[t]$, $D_n'P_{f_i,n}\in\ZZ[t]$ for $i=1,2$, and $\frac1n\log D_n\to4$. \item For $r>1$, real $z>r$ and $F\in\cF_{\rm II}$: $|R_{F,n}(z)|\le C_{F,r}(z)(K_r/z)^n$ with $K_r=r(r+1)(r+2)^3(r+3)^3$. In particular $K_r\to3456$ as $r\to1$, and $K_{1.01}=3569.9<4000$. \end{enumerate} \end{lemma} \begin{proof} (i) is Lemma~\ref{lem:contact} with $(e_0,e_1,e_{-1})=(n,2n,n)$. (ii) is Lemma~\ref{lem:denom} with $\delta=4n$, $\sigma=2n$. (iii) is Lemma~\ref{lem:apriori} with $\varsigma=1$: for $|t|\le r$ and $|h_1|=|h_2|=1$ we have $|W(t+h_1)|\le(r+1)(r+2)^2$ and $|V_1(t+h_1+h_2)|\le(r+2)(r+3)^3$. \end{proof} [Exact checks, not used.] The contact is at least $n+1$ for $\theta$ up to $n=512$ by exact series, and exactly $n+1$ for $f_1,f_2$ for $n\le80$; the denominators of (ii) were checked for $n\le150$. \subsection{The recurrence} \begin{proposition}\label{prop:A-rec} There are $c_0,\dots,c_5\in\ZZ[n,z]$, of degree $18$ in $n$ and of degrees $(6,7,8,9,10,6)$ in $z$, with coefficients of at most $83$ bits (file \file{ratio2/cert/rec\_C.pkl}), such that \[ \sum_{i=0}^5c_i(n,z)Q_{n+i}(z)=0\qquad\text{in }\QQ[z],\ \text{for every }n\ge0. \] Moreover: \begin{enumerate} \item $c_0=-256(2n+1)(2n+3)(n+1)^2(n+2)^2\,g_0(n,z)$ and $c_5=4(2n+5)(2n+7)(n+5)^2(2n+9)^2\,g_5(n,z)$, where $g_0,g_5$ have degree $12$ in $n$ and $6$ in $z$, and all $91$ coefficients of $g_0(n,53+s)$ and of $g_5(n,53+s)$, as polynomials in $(n,s)$, are positive. Hence $c_0(n,z)\ne0$ and $c_5(n,z)\ne0$ for all real $n\ge0$ and $z\ge53$. \item $\sum_i[n^{18}]c_i\cdot\lambda^i=\kappa(z)\chi_z(\lambda)$, where \begin{multline*} \kappa(z)=(3425z+3136)\\ \times(5775425z^5-13980709z^4+11418970z^3+905158z^2-3411979z+288463), \end{multline*} $\kappa(53+s)$ has coefficients of one sign, $[n^{18}]c_5=64\kappa(z)$, and \begin{gather*} \chi_z(\lambda)=64\lambda^5-4A_4\lambda^4-4A_3\lambda^3-A_2\lambda^2-64(129z-77)\lambda-1024,\\ A_4=3125z^4-1250z^3-2575z^2+768z+108,\\ A_3=8772z^3-10191z^2+1290z-451,\qquad A_2=25241z^2-27778z+3753. \end{gather*} \end{enumerate} \end{proposition} \begin{proof} \emph{The certificate.} Let $d(n)=\prod_{j\in J}(6n+j)$ with $J=\{0,\dots,14\}\cup\{16,\dots,20\}\cup\{22,\dots,26\}\cup \{28,29\}$. The file \file{ratio2/cert/cert\_exact.pkl} contains Laurent polynomials $A,B$ in $(h_1,h_2)$ with coefficients in $\QQ[n,z]$ ($729$ nonzero columns; numerators of at most $228$ bits) such that \eqref{eq:cert2} holds with these $c_i$ and $d$. The script \file{verify\_cert.py} checks \eqref{eq:cert2} in exact rational arithmetic, comparing all $800$ Laurent monomials. The coefficients in $\QQ[n,z]$ are compared through the Kronecker substitution $n\mapsto X$, $z\mapsto X^{64}$, which is injective here because every $n$-degree that occurs is at most $45<64$. The result is $0$ mismatches. By Section~\ref{sec:rec}(a), and since $d(n)>0$ for $n\ge1$, the recurrence holds for every $n\ge1$. The case $n=0$, and as an independent check all $0\le n\le150$, were verified as exact identities in $\QQ[z]$ on the explicit polynomials (\file{recverify.py}). \emph{The rest.} The factorisations, degrees, positivity and the identity in (ii) are exact polynomial computations (\file{hyp\_check.py}, which also verifies the hash of \file{rec\_C.pkl}). \end{proof} \begin{pcorollary}\label{cor:A-rec} For $F\in\cF_{\rm II}$ and every $n\ge0$: $\sum_ic_i(n,z)P_{F,n+i}(z)=0$ and $\sum_ic_i(n,z)R_{F,n+i}(z)=0$. \end{pcorollary} \begin{proof} Lemma~\ref{lem:transfer} with $\varsigma=1$ and $\mu=\max_i(\deg_zc_i-i)=6$ gives the claim for $n\ge6$. For $n\le40$ the identities $\sum_ic_iP_{F,n+i}=0$ were verified exactly in $\QQ[z]$ for $F=f_1,f_2,\theta$ (\file{check\_forms.py}). \end{proof} \begin{remark}\label{rem:A-numerics} [Numerical; not used.] The roots of $\chi_z$ are exactly the critical values of $\Lambda$ (at $z=5$ they agree to $15$ digits), and the recurrence annihilates $R_n(z)$ to $10^{-1000}$ at $z=5,65,129$. The dominant small root $\lambda_4(z)$ is real; at $z=104$ the moduli of the roots are $e^{-5.92}$, $e^{-5.58}$ (a complex pair), $e^{-4.0012}$ and $e^{23.85}$. The quantity $\log|\lambda_4(z)|+\log z$ increases from $0.6432$ at $z=104$ to $0.65477$ as $z\to\infty$, the latter being $\log|\mu|$ for the relevant root $\mu$ of $12500\mu^4+35088\mu^3+25241\mu^2+8256\mu+1024$. Numerically $R_{f_1,n}(104)$ and $R_{\theta,n}(104)$ follow $\lambda_4$, while $R_{f_2,n}(104)$ decays like the complex pair. \end{remark} \subsection{The roots} \begin{proposition}\label{prop:A-roots} For every real $z\ge104$: (a) $\chi_z$ has exactly $4$ roots in $|\lambda|64\rho^5+4A_3\rho^3+A_2\rho^2+ 64(129z-77)\rho+1024$ holds because, with rational enclosures of $e^{-4}$, the difference of the two sides is a polynomial in $s=z-188$ with positive coefficients. For $z\in[104,188]$, $6146$ Arb boxes in $(z,\arg\lambda)$ (at $200$ bits) show that $\chi_z\ne0$ on $|\lambda|=e^{-4}$; since the leading coefficient $64$ is constant, the number of roots in the disc is constant on $[104,188]$, and it is $4$ at $z=188$. (b) Rouch\'e with the term $4A_4\lambda^4$ on $|\lambda|=4000/z$; after multiplication by a power of $z$ the required inequality is a polynomial in $s=z-104$ with positive coefficients (\file{rootcert.py}; the same check at $z=53+s$ is in \file{hyp\_check.py}). \end{proof} Numerically $\log|\lambda_4(104)|=-4.00118$ and $\log|\lambda_4(103)|=-3.99163$ (Arb): $z=104$ is the least integer for which (a) holds. At $z=104$, $|\lambda_5|\approx2.3\cdot10^{10}$. \subsection{The determinant} \begin{lemma}\label{lem:A-det} $\Delta(t):=\det\bigl[Q_n(t),P_{f_1,n}(t),P_{f_2,n}(t),P_{\theta,n}(t)\bigr]_{n=0,1,2,3}$ equals \begin{multline*} -\tfrac1{675}\bigl(88288200t^6-200412030t^5+142832508t^4\\ +37721844t^3-49389216t^2-2701214t+537872\bigr). \end{multline*} All coefficients of $\Delta(53+s)$ are negative; hence $\Delta(z)<0$ for every real $z\ge53$. The real roots of $\Delta$ are $\approx-0.479$, $-0.132$, $0.083$, $0.750$. \end{lemma} \begin{proof} Exact computation (\file{det4.py}, \file{det\_s125.py}, two independent implementations). For instance $\Delta(104+s)=-(109291171069707975904+6328382704556914018\,s+\dots+88288200\,s^6)/675$. \end{proof} Since row $n=0$ is $(1,0,0,0)$, $\Delta$ is the $3\times3$ determinant of the $P_{F,n}$, $n=1,2,3$. In the $\lambda$-coordinates of Section~\ref{sec:coords} the determinant is the same, since $\det T=1$. \subsection{Proofs for \texorpdfstring{$N\ge103$}{N>=103}} Let $z=N+1\ge104$. We apply Theorem~\ref{thm:criterion} with index $n$, $D_n$ from Lemma~\ref{lem:A-basic}(ii), $e_n=0$, $\gamma=4$; (H2) by Proposition~\ref{prop:A-rec} and Corollary~\ref{cor:A-rec} (all $\deg_nc_i=18$, $[n^{18}]c_5=64\kappa(z) \ne0$, $c_0,c_5\ne0$ for integers $n\ge0$), with $\chi=\kappa(z)\chi_z$; (H3) with $\beta=K_{1.01}/z<4000/z$ (Lemma~\ref{lem:A-basic}(iii)); (H4) with $\rho=e^{-4}$ and $\beta'=4000/z$ (Proposition~\ref{prop:A-roots}). \begin{itemize} \item For Theorem~\ref{thm:ratio2}: $\cF=\{\theta\}$ and $\theta(z)=\theta_N>0$, so (H5) is not needed. \item For Theorem~\ref{thm:fourterm}: $\cF=\cF_{\rm II}$, the common denominator is $D_n$ since $D_n'\mid D_n$, and (H5) is Lemma~\ref{lem:A-det}. \end{itemize} The margin is $4+\log|\lambda_4(104)|=-0.0012$ per index; it enters only through limits. \begin{remark}[the Lean proof]\label{rem:lean} The Lean~4 proof of Theorem~\ref{thm:ratio2} for $N\ge103$ follows this section with the four departures listed in its README, each making the formal proof simpler. (1) Pituk's theorem is replaced by an elementary dichotomy (``cone'') lemma: a solution of the recurrence that is $O(4096^n)$ a priori and not eventually $0$ is $O(s^n)$ with $s=0.01831=52}}\label{sec:twoadic} % =========================================================================== We keep the forms of Section~\ref{sec:A}. For even $N$ the point $z=N+1$ is odd, and the integer forms gain a factor $2^{n-2}$. \emph{Notation.} Write $t=1+2w$, so that the odd integers are the values of $t$ at $w\in\ZZ$. For $g\in\QQ[t]$ let $\gn g=\min_jv_2\bigl([w^j]g(1+2w)\bigr)$, the $2$-adic Gauss valuation on the disc $1+2\ZZ_2$. Put $g[x,t]=(g(t)-g(x))/(t-x)$, let $\int_a^bg(x)\,dx$ denote the formal integral $G(b)-G(a)$ of a polynomial ($G$ an antiderivative), and let $\Hh$ be the harmonic transform $t^i\mapsto H_it^i$, $H_i=\sum_{k\le i}1/k$. Write $\Theta =W^nS_n$, so $Q_n=\Hs n\Theta$. \emph{Elementary facts.} \begin{itemize} \item[(G1)] $\gn{gh}=\gn g+\gn h$ (Gauss' lemma); since $\gn t=0$, $\gn{h/t}=\gn h$ if $t\mid h$. \item[(G2)] $\gn{\Hs kg}\ge\gn g-k$, because $d/dt=\frac12d/dw$ and $\frac1{k!}(d/dw)^k$ preserves $\ZZ_2[w]$. \item[(G3)] $\gn g\ge\min_iv_2([t^i]g)$, because $(1+2w)^i\in\ZZ[w]$. \item[(G4)] $v_2(g(z))\ge\gn g$ for every odd integer $z$. \item[(G5)] $\Hh g(t)=\int_0^tg[x,t]\,dx$. Substituting $x=1+2\xi$ gives $\int_1^tg[x,t]\,dx=\sum_k\hat g_kH_kw^k$, where $\hat g(\xi)=g(1+2\xi)=\sum_k\hat g_k\xi^k$. Hence $\gn{\int_1^tg[x,t]dx}\ge\gn g-\lfloor\log_2\deg g\rfloor$. \end{itemize} \begin{lemma}[Lemma T]\label{lem:T} For every odd integer $z$ and every $n\ge1$: \begin{enumerate} \item $v_2(Q_n(z))\ge2n$; \item $v_2(P_n(z))\ge n-1-\lfloor\log_25n\rfloor-\lfloor\log_2(4n-1)\rfloor$; \item for $n\ge2$, $2^{n-2}$ divides the integers $D_nQ_n(z)$ and $D_nP_n(z)$. \end{enumerate} \end{lemma} \begin{proof} \emph{Step 0 (sizes on the odd disc).} $W(1+2w)=4w(1+w)(1+2w)$ and $V_1(1+2w)=8w^2(1+w)(1+2w)$, so $\gn{W^n}=2n$ and $\gn{V_1^n}=3n$. By (G2), $\gn{S_n}\ge n$, so $\gn \Theta \ge3n$ and $\gn{Q_n}\ge2n$. With (G4) this is (i). \emph{Step 1 (splitting $\Theta$).} $\Theta=UM$ with $U=(t^2-1)^n$ and $M=t^nS_n\in\ZZ[t]$, $\deg M=3n$. Since $(UM)[x,t]=U(t)M[x,t]+M(x)U[x,t]$, (G5) gives $\Hh \Theta =U\cdot\Hh M+\mathrm{Comm}$ with $\mathrm{Comm}(t)=\int_0^tM(x)U[x,t]\,dx$. From $U(t)-U(x)=(t^2-x^2)\sum_{a+b=n-1}(t^2-1)^a(x^2-1)^b$, \begin{gather*} \mathrm{Comm}=\sum_{a+b=n-1}(t^2-1)^a\bigl[t\,\Omega_b(t)+\Omega_b'(t)\bigr],\\ \Omega_b(t)=\int_0^tM(x)(x^2-1)^bdx,\qquad\Omega_b'(t)=\int_0^txM(x)(x^2-1)^bdx . \end{gather*} \emph{Step 2 (the constants vanish).} Let $c_b=\Omega_b(1)$ and $c_b'=\Omega_b'(1)$. Then $c_b=0$ if $2br$; (H4) with $\rho=2e^{-4}=e^{-\gamma}$ and $\beta'=4000/z$; (H5) by Lemma~\ref{lem:A-det}, since $\Delta(z)<0$ for $z\ge53$. Together with Section~\ref{sec:A} this proves Theorems~\ref{thm:ratio2} and~\ref{thm:fourterm} for every even $N\ge52$. % =========================================================================== \section{The shape \texorpdfstring{$(2;1,2,\frac54)$}{(2;1,2,5/4)}: \texorpdfstring{$84\le N\le102$}{84<=N<=102}}\label{sec:A125} % =========================================================================== Let $n=4m$, $V_m=t^{4m}(t-1)^{8m}(t+1)^{5m}$, so that $V_m$ has exponents $(n,2n,\frac54n)$ at $(0,1,-1)$, and \begin{gather*} S_m=\Hs{8m}V_m,\qquad Q_m=\Hs{4m}\bigl[W^{4m}S_m\bigr]\in\ZZ[t],\\ \deg S_m=9m,\qquad\deg Q_m=17m . \end{gather*} We index by $m$. Then $Q_m(z)=\CT_{h_1,h_2}[\Lambda^m]$ with \[ \Lambda=\frac{W(z+h_1)^4(z+h_1+h_2)^4(z+h_1+h_2-1)^8(z+h_1+h_2+1)^5}{h_1^4h_2^8}, \] and $Q_m(z)=\Res_{u=z}\Res_{v=u}W(u)^{4m}V_m(v)(u-z)^{-4m-1}(v-u)^{-8m-1}$. The targets are $\cF_{\rm II}$. \begin{lemma}\label{lem:125-basic} \begin{enumerate} \item $R_{F,m}(t)=O(t^{-4m-1})$ for $F\in\cF_{\rm II}$. \item Put \[ D_m=\prod_{p\le8m}p^{2\lfloor\log_p17m\rfloor}\prod_{8m1$ and real $z>r$, $|R_{F,m}(z)|\le C_{F,r}(z)(r^4M_r/z^4)^m$. We have $M_1=18^4\cdot3^4\cdot4^{13}=5.7063\cdot10^{14}$; at $r=1.001$, $r^4M_r=5.7824\cdot10^{14}<5.8\cdot10^{14}$ (exact rationals), and at $r=1.1$, $r^4M_r=2.075\cdot10^{15}<10^{16}$. \end{enumerate} \end{lemma} \begin{proof} (i) Lemma~\ref{lem:contact} with $(e_0,e_1,e_{-1})=(n,2n,\frac54n)$. (ii) Lemma~\ref{lem:denom} with $\delta=17m=4.25n$ and $\sigma=9m=2.25n$. (iii) Lemma~\ref{lem:apriori} with $\varsigma=4$, since $|W(t+h_1)|\le(r+1)(r+2)^2$ and $|s^4(s-1)^8(s+1)^5|\le(r+2)^4(r+3)^{13}$ for $|s|\le r+2$. \end{proof} [Exact checks.] Contact $\ge4m$ for $m\le40$ (exactly $4m+1$ for $f_1,f_2$); the denominators for $m\le60$. At $m=40$ the exact denominator of $P_{f_i,m}$ has $v_p=1$ at $p=347,353,359\in(320,360]$, so the factor on $(8m,9m]$ is needed. \subsection{The recurrence}\label{sec:125-rec} We use the two-step route of Section~\ref{sec:rec}(b). Let $\Phi=V_m(v)(v-u)^{-8m-1}$, $w=v-u$. \emph{Step A} [\file{stepA.py}, \file{verify\_stepA.py}]. (A1) $\sum_{j=0}^3a_j(m,u)(8m+1)_jw^{-j}=\nabla\bigl((v^3-v)w^{-2}\bigr)$ with $a_0=9m$, $a_1=(10mu+3m-3u)/(8m+1)$, $a_2=-(7mu^2-3mu-4m+6u^2-2)/(2(4m+1)(8m+1))$, $a_3=-W(u)/(2(4m+1)(8m+1))$; so $\sum_ja_jS_m^{(j)}=0$. (A2) $\Phi_{m+1}/\Phi_m-\sum_{j\le2}b_j(m,u)(8m+1)_jw^{-j}=\nabla\bigl((v^3-v)\sum_{s=-7}^7 \eta_s(m,u)w^s\bigr)$, where $\Phi_{m+1}/\Phi_m=v^4(v-1)^8(v+1)^5w^{-8}$; so $S_{m+1}=\sum_{j\le2}b_jS_m^{(j)}$. The $b_j,\eta_s$ have denominator $10368\,u^2(u+1)(m+1)(2m+1)(3m+1)(3m+2)(4m+1)(4m+3)(8m+1)(8m+3)(8m+5)(8m+7)(9m+1)(9m+2)(9m+4)(9m+5)(9m+7)(9m+8)$, nonzero for $m\ge0$. Both identities are verified in $\ZZ[m,u,v]$ after clearing denominators. \emph{Step B} [\file{verify\_sys.py}]. Basis $E_m(u-z)^aS_m^{(\ell)}$, $a,\ell\in\{0,1,2\}$, $E_m=W(u)^{4m}(u-z)^{-4m-1}$; certificates $\gamma_l=P_l(u)/(W(u)^5(u-z)^3)$ with $\deg_uP_l\le45$. The matrix $A$ and the $P_l$ are polynomials in $(m,z)$ divided by $d(m,z)=z^4(z+1)^3L(m)$, where \begin{multline*} L(m)=m(m-3)(m+1)^2(8m+1)^2(4m+1)^2(2m+1)^2(4m+3)^2\\ \times(3m+1)(3m+2)(8m+3)(8m+5)(8m+7)\\ \times\prod_{j\in\{3,\dots,19\}\setminus\{17\}}(17m+j)\prod_{j\in\{1,2,4,5,7,8\}}(9m+j). \end{multline*} The numerators have degree $\le41$ in $m$ and $\le29$ in $z$, rationals of at most $183$ bits, and $645{,}069$ nonzero coefficients over $939$ columns (\file{shape125/cert/sys\_exact.pkl}). All nine identities are verified as identities of rational functions in $(m,z,u)$, coordinate by coordinate as \texttt{fmpz\_mpoly} (a complete equality test). Hence $Y_{m+1}(z)=A(m,z)Y_m(z)$ for integers $m\ge1$, $m\ne3$, and $z\notin\{0,-1\}$ (the zeros of $d(m,z)$); since the scalar recurrence below is a polynomial identity in $z$, the restriction on $z$ is harmless. \emph{Step C} [\file{verify\_scalar.py}, \file{check\_small.py}]. \begin{proposition}\label{prop:125-rec} There are primitive polynomials $C_0,\dots,C_6\in\ZZ[m,z]$, of degree $172$ in $m$ and of degrees $(55,59,63,67,71,75,58)$ in $z$, with coefficients of at most $1042$ bits ($78{,}715$ nonzero coefficients; file \file{shape125/cert/crec\_exact.pkl}), such that $\sum_{i=0}^6C_i(m,z)Q_{m+i}(z)=0$ for every $m\ge0$. \end{proposition} \begin{proof} The identity $\sum_iC_i(m,z)r_i(m)=0$ in $\QQ(m,z)^9$ is verified exactly after clearing denominators. It uses $A(m),\dots,A(m+5)$, so it gives the recurrence for $m\ge4$. For $m=0,\dots,8$ the recurrence was verified exactly on the explicit polynomials. \end{proof} The recurrence was found from the system by exact linear algebra mod primes and reconstructed by CRT over $36$ primes of $61$ bits; it was never guessed from values of $Q_m$. At $z=11$ there is no recurrence of order $\le6$ and degree $\le154$ in $m$ [Numerical]; this is not used. \begin{pcorollary}\label{cor:125-rec} For $F\in\cF_{\rm II}$ and every $m\ge0$: $\sum_iC_iP_{F,m+i}(z)=0=\sum_iC_iR_{F,m+i}(z)$. \end{pcorollary} \begin{proof} Lemma~\ref{lem:transfer} with $\varsigma=4$ and $\mu=\max_i(\deg_zC_i-4i)=55$ gives $m\ge14$. For $m\le30$ the $P$-identities were verified exactly for $f_1,f_2,\theta$ (\file{check\_s125.py}; for $\theta$ also $m\le19$ by \file{check\_small.py}). \end{proof} \emph{Characteristic polynomial, nonvanishing} [\file{final\_checks.py}, \file{crit.py}]. The coefficients of $m^{172}$ in the $C_i(\cdot,z)$ are proportional to the coefficients of a sextic $\chi_6(z,\lambda)$ (file \file{shape125/cert/chi6.pkl}), whose roots are the critical values of $\Lambda$. Its leading coefficient and its constant term are \begin{gather*} -249751218699058108948113661334913024\,(z+1)^3,\\ -25353012004564588029934064107520000000000 . \end{gather*} At every integer $z_0\in[85,103]$ the proportionality is checked exactly, $[m^{172}]C_6(z_0)\ne0$, and $C_0(m,z_0)$ and $C_6(m,z_0)$ have no integer root $m\ge0$ (factorisation over $\ZZ$). \begin{proposition}\label{prop:125-roots} For each integer $z\in[85,103]$, $\chi_6(z,\cdot)$ has six simple roots: five with $|\lambda|10^{16}/z^4$ (in particular $>5.8\cdot10^{14}/z^4$). \end{proposition} \begin{proof} Certified root isolation with \texttt{acb\_poly} at $300$ bits (\file{rootcert125.py}); the bound $10^{16}/z^4$ by \file{rootcert125\_big.py}, which reads \file{chi6.pkl} and checks its hash. Here $\log\inf|\lambda_6|\ge97.4886$ (attained at $z=85$) against $\log(10^{16}/z^4)\le19.07$. \end{proof} The margins $17+\log|\lambda_5(z)|$ are $-0.0202$ ($z=85$), $-0.0665$ ($z=86$), \dots, $-0.7802$ ($z=103$); at $z=84$ the margin is $+0.0266$ and the argument fails. Here $\lambda_4=\bar\lambda_5$ is a complex pair and $\lambda_6\approx e^{97}$--$e^{101}$. \begin{lemma}\label{lem:125-det} $\Delta'(t):=\det[Q_m,P_{f_1,m},P_{f_2,m},P_{\theta,m}]_{m=0,\dots,3}$ has degree $36$; its numerator is $8$ times an irreducible polynomial of degree $36$, and all $37$ coefficients of $\Delta'(85+s)$ are positive. Hence $\Delta'(z)>0$ for every real $z\ge85$. \end{lemma} \begin{proof} Exact computation (\file{det\_s125.py}). The values $\Delta'(z)$, $z=85,\dots,103$, are about $10^{94}$--$10^{97}$. \end{proof} \subsection{Proofs for \texorpdfstring{$84\le N\le102$}{84<=N<=102}} Let $z\in[85,103]$ be an integer. Apply Theorem~\ref{thm:criterion} with index $m$, $D_m$ of Lemma~\ref{lem:125-basic}(ii) (a common denominator, since $D_m'\mid D_m$), $e_m=0$, $\gamma=17$; (H2) by Proposition~\ref{prop:125-rec}, Corollary~\ref{cor:125-rec} and the checks above; (H3) with $\beta=r^4M_r/z^4$, $r=1.001$; (H4) with $\rho=e^{-17}$ and $\beta'=5.8\cdot10^{14}/z^4$ (Proposition~\ref{prop:125-roots}); (H5) by Lemma~\ref{lem:125-det} for Theorem~\ref{thm:fourterm}, and $\theta(z)>0$ for Theorem~\ref{thm:ratio2}. With Section~\ref{sec:A} this proves both theorems for every $N\ge84$. % =========================================================================== \section{The parabola: Theorem~\ref{thm:S2}}\label{sec:S2} % =========================================================================== Here $V_1=(t^2-1)^2$, $V=V_1^n=(t^2-1)^{2n}$, $S_n=\Hs{2n}V_1^n=4^nP_{2n}$ ($P_{2n}$ the Legendre polynomial), $Q_n=\Hs n[W^nS_n]\in\ZZ[t]$, $\deg Q_n=4n$, $Q_n$ even, $Q_n(z)=\CT[\Lambda^n]$ with $\Lambda=W(z+h_1)V_1(z+h_1+h_2)/(h_1h_2^2)$. The targets are $\cF_{\rm S}=\{f_1,f_2,L^2\}$. \begin{lemma}\label{lem:S2-basic} (i) $R_{F,n}=O(t^{-n-1})$ for $F\in\cF_{\rm S}$. (ii) The number $D_n=\prod_{p\le2n}p^{2\lfloor\log_p4n\rfloor}$ is a common denominator of the $P_{F,n}$, and $\frac1n\log D_n\to4$. (iii) $\limsup|R_{F,n}(z)|^{1/n}\le K_r/z$ with $K_r=r(r+1)(r+2)^2(r+3)^4$; $K_{1.001}=4622.61<4700$. \end{lemma} \begin{proof} (i) Lemma~\ref{lem:contact}: for $L^2$ we need $e_{\pm1}\ge2n$. (ii) Lemma~\ref{lem:denom} with $\delta=4n$, $\sigma=2n$. (iii) Lemma~\ref{lem:apriori}, with $|W(s)|\le(r+1)(r+2)^2$ for $|s|\le r+1$ and $|V_1(s)|\le(r+3)^4$ for $|s|\le r+2$. \end{proof} [Exact checks.] Contact and denominators for $n\le30$; at $n=30$ the true profile is $v_p=2$ for $11\le p\le2n$ and $0$ above. \emph{Step A.} (A1) $\sum_{j\le2}a_j(2n+1)_jw^{-j}=\nabla((v^2-1)w^{-1})$ with $a_0=2n$, $a_1=-2u/(2n+1)$, $a_2=-(u^2-1)/(2n+1)$, which gives the Legendre equation $(1-u^2)S_n''-2uS_n'+2n(2n+1)S_n=0$. (A2) $V_1(v)w^{-2}-\sum_{j\le1}b_j(2n+1)_jw^{-j}=\nabla\bigl((v^2-1)\sum_{s=-1}^1\eta_sw^s\bigr)$ with $b_0=2((4n+3)u^2-2n-1)/(n+1)$ and $b_1=2(4n+3)u(u^2-1)/((n+1)(2n+1))$, so $S_{n+1}=b_0S_n+b_1S_n'$ for $n\ge0$. \emph{Step B.} A $6\times6$ system in the basis $E_n(u-z)^aS_n^{(\ell)}$, $a\in\{0,1,2\}$, $\ell\in\{0,1\}$; the certificates $\gamma_l$ are polynomials in $u$ of degree $\le8$; $A$ and the $P_l$ are polynomials in $z$ of degree $\le7$ over $d(n)=(n+1)^2(2n+1)^2(2n+3)(4n+3)(4n+5)$; $108$ nonzero entries, $1491$ coefficients, rationals of at most $20$ bits (\file{s2/cert/sys\_exact.pkl}). All six identities hold exactly, and $Y_{n+1}=A(n,z)Y_n$ for every $n\ge0$. \begin{proposition}\label{prop:S2-rec} There are primitive $C_0,\dots,C_5\in\ZZ[n,z]$ (\file{s2/cert/rec5\_exact.pkl}), even in $z$, of degree $17$ in $n$ and degrees $(4,4,6,6,8,4)$ in $z$ ($396$ terms, at most $70$ bits), with $\sum_iC_i(n,z)Q_{n+i}(z)=0$ for all $n\ge0$. Moreover: \begin{enumerate} \item $C_0=-4096(2n+1)(2n+5)(n+1)^2(n+2)^2g_0$ and $C_5=4(2n+5)(2n+7)(n+5)^2(2n+9)^2g_5$, with $g_i$ of degree $11$ in $n$ and $4$ in $z$; all $60$ coefficients of $g_0(n,59+s)$, and of $g_5(n,59+s)$, have one sign and the constant terms are nonzero, so $C_0,C_5\ne0$ for real $n\ge0$, $z\ge59$; \item $[n^{17}]C_i=\mathrm{COMMON}(z)a_i(z)$ with $\mathrm{COMMON}=180500000z^4+9472000z^2-42467328$ (positive coefficients at $z=59+s$), $a_5=16$, and \begin{multline*} \chi_z(\lambda)=16\lambda^5+(-3125z^4+3000z^2-272)\lambda^4+(1600z^2+1792)\lambda^3\\ +(-5248z^2-5632)\lambda^2+8192\lambda-4096 . \end{multline*} \end{enumerate} \end{proposition} \begin{proof} $\sum_iC_ir_i=0$ in $\QQ(n,z)^6$ is verified exactly (\file{verify\_rec.py}); a perturbed control fails. Since the system holds for all $n\ge0$, so does the recurrence; it was cross-checked on explicit polynomials for $n\le12$. (i) and (ii) are exact computations (\file{check\_lead\_trail.py}). \end{proof} By Lemma~\ref{lem:transfer} ($\varsigma=1$, $\mu=4$) the $P$- and $R$-recurrences hold for $n\ge4$; for $n\le16$ they were checked exactly for all three targets. \begin{proposition}\label{prop:S2-roots} Let $\rho=9157/500000$; then $\log\rho=-4.0000895<-4$. For every real $z\ge59$, $\chi_z$ has exactly $4$ roots in $|\lambda|<\rho$ and exactly $4$ roots in $|\lambda|<4700/z$. \end{proposition} \begin{proof} Lemma~\ref{lem:resultant}: $G_1(z)=\Res_\lambda(\chi_z(\lambda),\lambda^5\chi_z(\rho^2/\lambda))$ has degree $32$, and $G_1(59+s)$ has coefficients of one sign with $G_1(59)\ne0$; at $z=59$ certified root isolation gives $4$ roots inside and one outside. For the radius $4700/z$ the same argument applies to $z^5\chi_z(\nu/z)$ on $|\nu|=4700$ ($G_2$ of degree $50$) (\file{rootcert.py}). \end{proof} The largest real root of $G_1$ is $58.99460$, a double root (exact real-root isolation), so the certificate has $z$-margin $0.0054$ below $59$; floating-point root finders misplace such double roots. At $z=59$ the roots have $\log|\lambda|=-4.019467$ (pair), $-4.000182$ (pair) and $+21.584$; at $z=58$ the pairs are at $-3.982919$ and $-4.002536$ and the certificate fails. \begin{lemma}\label{lem:S2-det} $\det[Q_n,P_{f_1,n},P_{f_2,n},P_{L^2,n}]_{n=0..3}=-\frac{64}{135}z(3506685z^4-1742076z^2+54559)$; every coefficient of the numerator at $z=2+s$ has one sign, so it is nonzero for real $z\ge2$ (real roots $0,\pm0.183,\pm0.681$). \end{lemma} \emph{Proof of Theorem~\ref{thm:S2}.} For $z=N+1\ge59$ apply Theorem~\ref{thm:criterion} with $\gamma=4$, $\beta=K_{1.001}/z$, $\beta'=4700/z$ and $\rho=9157/5000001$. Over about $1.1\cdot10^5$ two-level Rodrigues shapes, refined up to $n=384$, every form tested has $\varpi\le0.43$ [Numerical]: family II has $\varpi=0.150$ (decay $0.5993=-\log|\lambda_4(2)|$, cost $4$), and the best shape found, $(\frac12;\frac32,\frac32,0\,|\,\frac52;3,3,3)$, has $\varpi=0.425$ (decay $3.85$, cost $9.75$, $2$-adic gain $0.695$). Stacking every gain we know (the representation $\log2\cdot\log3=(f_1-f_2)(f_1-2f_2)$ at $z=3$, the $2$-adic content, and the holonomy bound of \cite[Thm.~7.0.1]{CDT} with the best univalent map) gives $\varpi_{\rm eff}\approx0.57$, at most about $0.65$ with speculative covering maps [Numerical and Heuristic]. The residual is a factor of at least $1.75$ in decay over cost, about $1.7$ nats per index at cost~$4$. Closing it needs a new idea, not a tuning of this family. \item \emph{Other ratios and more logarithms.} The same construction applies to $\log(1+a/N)\log(1+b/N)$ with one recurrence per ratio $a/b$. Predicted thresholds in $N$ from the characteristic roots [Numerical]: $N\ge356$ for $(a,b)=(2,3)$, $N\ge1400$ for $(1,3)$, $N\ge1938$ for $(3,4)$, $N\ge7440$ for $(1,4)$; $N\ge411$ for the other parabola $1,\lambda_1,\lambda_2,\lambda_1^2$ with its own forms; and about $N\ge11622$ for $1,\lambda_1,\lambda_2,\lambda_3,\lambda_1\lambda_2$ with $\lambda_3=\log(1+3/N)$. We know no construction of this kind for $\lambda_1\lambda_2\notin\QQ\pi^2$. \item \emph{Algebraic coefficients.} Conjecture~\ref{conj:S4} predicts the statement of Theorem~\ref{thm:C6} with $\Qbar$ in place of $\QQ$, and for every $N\ge1$; our method gives neither. \item \emph{Removing Pituk's theorem.} The Lean proofs of Theorems~\ref{thm:ratio2} and~\ref{thm:fourterm} for $N\ge103$ and of Theorem~\ref{thm:C6} replace Pituk's theorem by elementary growth lemmas (Remark~\ref{rem:lean}). The same should work for the remaining families and ranges, where the large root is equally well separated; it has not been written out. The elementary ODE alternative (continuation of $\sum_nR_n(z)x^n$ across the disc) is also only sketched in our notes. \end{itemize} % =========================================================================== \section{Provenance and use of AI tools}\label{sec:provenance} % =========================================================================== \subsection*{Provenance} The forms of Section~\ref{sec:C6} are those of an earlier unpublished note of ours, whose remainder bound (by a Cauchy estimate) reached only $N\ge5.3\cdot10^6$; the other shapes were found by scanning two-level Rodrigues shapes. The recurrences were found by modular linear algebra and then proved by the exact certificates described above. The external mathematical inputs are Pituk's theorem \cite{Pit02}, used as stated, Baker's theorem (only in Section~\ref{sec:s4}, which is not used in any proof), and the prime number theorem. Nothing in the proofs depends on unpublished work by others; \cite{Kaw26} and \cite{VZ18} are cited only for comparison. \subsection*{Use of AI tools} This work was carried out with substantial assistance from AI systems (Anthropic's Claude models), which were used to search the literature, to develop the proofs, to write the scripts and produce the certificates, to referee the arguments, to formalise part of the result in Lean, and to draft this text. The referee reports were also produced by Claude models, working adversarially on separate copies of the files; they are not a substitute for a human expert's reading. The author directed the project and is responsible for its content. Only the irrationality of $\theta_N$ for $N\ge103$ is verified in a proof assistant. \bibliographystyle{amsplain} \bibliography{refs} \end{document}