\documentclass[11pt]{amsart} \usepackage{amsmath} \usepackage{amssymb} \usepackage{amsthm} \usepackage{booktabs} \usepackage{array} \usepackage{longtable} \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} \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}}} % Visible marker for open items. Every \TODO must be resolved (or the passage deleted) before posting. \newcommand{\TODO}[1]{\textbf{[TODO: #1]}} \begin{document} \hypersetup{pdftitle={Products of two logarithms and arithmetic holonomy bounds},pdfauthor={Shawn Krisman}} \title[Products of two logarithms]{Products of two logarithms and arithmetic holonomy bounds} \author{Shawn Krisman} \subjclass[2020]{Primary 11J72; Secondary 11J82, 34M35, 41A21, 65G20} \keywords{irrationality, linear independence, products of logarithms, Hermite--Pad\'e approximation, arithmetic holonomy bounds, monodromy, computer-assisted proof, interval arithmetic} \begin{abstract} We prove, assuming the arithmetic holonomy bound of Calegari, Dimitrov and Tang (Theorem~7.0.1 of their preprint arXiv:2408.15403), that $1$, $\log(1+1/q)$, $\log(1-1/q)$ and $\log(1+1/q)\log(1-1/q)$ are linearly independent over $\QQ$ for every integer $26\le q\le53$. Together with Hata's theorem for $q\ge54$ this gives the four-term independence for every $q\ge26$. In particular the seven numbers $\log(1+1/N)\log(1+1/(N+1))$, $25\le N\le31$, are irrational; irrationality was previously known for $N\ge32$ (Lysov). The proof feeds Lysov's explicit type~II Hermite--Pad\'e forms into the holonomy bound, and replaces the arithmetic of the forms by an analytic input: the monodromy of a minimal differential operator of order~13, which we show to be irreducible by interval-arithmetic certificates. Conditionally on the published form of the Calegari--Dimitrov--Tang argument for products of two logarithms, we also widen their Theorem~C from $|m/n-1|<10^{-6}$ to $|m/n-1|<3\cdot10^{-6}$, which gives infinitely many new irrational products $\log(1+1/m)\log(1+1/n)$ with $|m-n|\ge2$, and we show by pooling several pairs that at most one exceptional pair class exists with $|m/n-1|<5\cdot10^{-5}$, and at most two with $|m/n-1|<10^{-4}$. All certificates are exact or rigorous ball arithmetic and can be reproduced from the accompanying code at \url{https://github.com/telaviv/log-products-irrationality}. \end{abstract} \maketitle % =========================================================================== \section{Introduction}\label{sec:intro} % =========================================================================== For an integer $q\ge2$ put \[ \theta_q=\log\Bigl(1+\frac1q\Bigr)\log\Bigl(1-\frac1q\Bigr). \] Hata \cite[Cor.~6.3]{Hat98} showed that $1$, $\log(1+1/q)$, $\log(1-1/q)$, $\theta_q$ are linearly independent over $\QQ$ for every integer $q\ge54$, with an effective linear independence measure; in particular $\theta_q\notin\QQ$ for every integer $q\ge54$ \cite[Thm.~0.1]{Hat98}. Lysov \cite{Lys18} found an explicit formula for the common denominator of the type~II Hermite--Pad\'e approximants to $(\log(1+1/z),\log(1-1/z),\log(1+1/z)\log(1-1/z))$ and deduced that $\theta_q$ is irrational for every $q\ge33$ \cite{Lys18,Lys19} (\cite{Lys18} is a preprint; we found no journal version as of 19 November 2026). Calegari, Dimitrov and Tang \cite[Thm.~C, Thm.~14.0.1]{CDT} proved, by a new arithmetic holonomy bound, that $1$, $\log(1+1/m)$, $\log(1+1/n)$ and $\log(1+1/m)\log(1+1/n)$ are linearly independent over $\QQ$ whenever $m\ne n$ are integers in $\ZZ\setminus\{-1,0\}$ with $|m/n-1|<10^{-6}$. In \cite[\S15.1.10]{CDT} they suggest forming ``the generating function of the special linear forms obtained from evaluating a regular sequence of functional Hermite--Pad\'e approximants'' and reading it through their holonomy bounds. This paper carries out that scheme for Lysov's forms, and pushes the constant $10^{-6}$ of \cite[Thm.~C]{CDT}. \subsection*{Dependence on one preprint} Every result in this paper depends on the preprint \cite{CDT} (arXiv:2408.15403v2, 16 September 2024; still the current version on 19 November 2026), which to our knowledge has not been refereed. Theorem~\ref{thm:4T} and Corollary~\ref{cor:F1} use only its general holonomy bound \cite[Thm.~7.0.1]{CDT}, which we apply as stated. Theorems~\ref{thm:Cprime} and \ref{thm:F2} also use the specific construction of \cite[\S14]{CDT} for products of two logarithms, together with \cite[Thm.~6.0.2, Cor.~9.0.19, Lemma~12.0.1]{CDT}. That construction contains a step that is asserted rather than proved in \cite{CDT}, the irreducibility of a certain differential operator $\mathcal L_{a,b}$, and we inherit it; see Section~\ref{sec:Cprime}. Each theorem below names exactly what it assumes. \subsection{Results} \begin{theorem}\label{thm:4T} Assume \cite[Thm.~7.0.1]{CDT}. For every integer $q$ with $26\le q\le53$, the numbers \[ 1,\qquad\log\Bigl(1+\frac1q\Bigr),\qquad\log\Bigl(1-\frac1q\Bigr),\qquad\log\Bigl(1+\frac1q\Bigr)\log\Bigl(1-\frac1q\Bigr) \] are linearly independent over $\QQ$. \end{theorem} With Hata's theorem \cite{Hat98} for $q\ge54$, the same holds for every integer $q\ge26$. Precisely, \cite[Cor.~6.3]{Hat98} states: let $q\ge54$ be an integer; for every $\varepsilon>0$ there is an effectively computable $H_1=H_1(\varepsilon,q)$ such that \[ \Bigl|n_0+n_1\log\Bigl(1+\frac1q\Bigr)+n_2\log\Bigl(1-\frac1q\Bigr)+n_3\theta_q+n_4\Lambda\Bigl(\frac1q\Bigr)\Bigr| \ge H^{-\sigma(q^2)/\tau(q^2)-\varepsilon} \] for all $(n_0,\dots,n_4)\in\ZZ^5$ with $H=\max_{1\le m\le4}|n_m|\ge H_1$. Here $\Lambda(x)=\mathrm{Li}_2((1+x)/2)-\mathrm{Li}_2((1-x)/2)-\log2\cdot\log((1+x)/(1-x))$ \cite[(3.5)]{Hat98}, and $\sigma,\tau$ are explicit \cite[(6.3), (6.4)]{Hat98}; the bound $q\ge54$ comes from Hata's $\tau(k)>0$ for every integer $k\ge2909$ (and $53^2<2909\le54^2$), and $\sigma(k)/\tau(k)\to4$ as $k\to\infty$. Taking $n_4=0$ gives the four-term independence; in fact $1$, $\log(1+1/q)$, $\log(1-1/q)$, $\theta_q$, $\Lambda(1/q)$ are linearly independent over $\QQ$. Hata states this for integers $q$ only: Corollary~6.3 is the case $k=q^2$ of his Theorem~6.2, the analogous statement at $1/\sqrt k$ for every integer $k\ge2909$. Since $\log(1-1/q)=-\log(1+1/(q-1))$, Theorem~\ref{thm:4T} at $q=N+1$ contains the irrationality of the product of two consecutive logarithms $\log(1+1/N)\log(1+1/(N+1))=-\theta_{N+1}$: \begin{corollary}\label{cor:F1} Assume \cite[Thm.~7.0.1]{CDT}. For every integer $N$ with $25\le N\le31$, \[ \log\Bigl(1+\frac1N\Bigr)\log\Bigl(1+\frac1{N+1}\Bigr)\notin\QQ . \] \end{corollary} These are seven new irrational numbers. By Theorem~\ref{thm:4T} for $32\le N\le52$ and Hata's theorem for $N\ge53$ (and, independently, Lysov's theorem \cite{Lys18} for $N\ge32$), the product is irrational for every $N\ge25$. The second group of results concerns CDT's two-parameter family. For $m\in\ZZ\setminus\{-1,0\}$ write $\ell_m=\log(1+1/m)$. \begin{theorem}\label{thm:Cprime} Assume the following parts of \cite{CDT} as published: Theorem~6.0.2, Corollary~9.0.19, Lemma~12.0.1 and \S14, including the irreducibility of the operator $\mathcal L_{a,b}$ and the reduction in the proof of Lemma~14.3.1. Let $m\ne n$ be integers in $\ZZ\setminus\{-1,0\}$ with $|m/n-1|<3\cdot10^{-6}$. Then $1$, $\ell_m$, $\ell_n$ and $\ell_m\ell_n$ are linearly independent over $\QQ$; in particular $\ell_m\ell_n$ is irrational. \end{theorem} \begin{theorem}\label{thm:F2} Make the assumptions of Theorem~\ref{thm:Cprime}. Call a pair $(m,n)$ of distinct integers in $\ZZ\setminus\{-1,0\}$ \emph{exceptional} if $1,\ell_m,\ell_n,\ell_m\ell_n$ are linearly dependent over $\QQ$, and take classes of pairs modulo $(m,n)\sim(n,m)\sim(-m-1,-n-1)$. Then: \begin{enumerate} \item at most one exceptional class has a representative with $|m/n-1|<5\cdot10^{-5}$; \item at most two exceptional classes have a representative with $|m/n-1|<10^{-4}$; \item every exceptional class has $|m/n-1|\ge3\cdot10^{-6}$ in all its representatives (Theorem~\ref{thm:Cprime}); \item in the ranges of (i) and (ii), every exceptional class has $|m-n|\ge2$, by Hata's theorem \cite[Cor.~6.3]{Hat98}. \end{enumerate} \end{theorem} The relation $(m,n)\sim(-m-1,-n-1)$ comes from $\ell_{-k}=-\ell_{k-1}$: the two pairs give the same four numbers up to sign. For $1\le N3\cdot10^5$ and are covered by Hata's theorem. So the new irrational numbers of Theorem~\ref{thm:Cprime} are \begin{equation}\label{eq:newpairs} \ell_N\ell_M,\qquad d=M-N\ge2,\qquad 10^{-6}(M+1)\le d<3\cdot10^{-6}(M+1), \end{equation} that is, $d\cdot10^6/3\max\{\sigma_\fm,\tau(b;e)\}$, such that every $f_i(\varphi(z))$ is the germ of a meromorphic function on $|z|<1$. Then \[ \fm\le\frac{T(\varphi)}{\log|\varphi'(0)|-\tau(b;e)} . \] If all $f_i$ are known to be holonomic, the hypothesis $\log|\varphi'(0)|>\max\{\sigma_\fm,\tau\}$ may be relaxed to $\log|\varphi'(0)|>\tau$. \end{htheorem} (In \cite{CDT} the map is defined on $\DD$ and $T$ is defined by a limit; we only use maps holomorphic on a neighbourhood of $\overline\DD$.) \begin{htheorem}[{\cite[Thm.~6.0.2]{CDT}, partition form}]\label{thm:CDT602} Let $f_1,\dots,f_\fm$ and $(b,e)$ be as in Theorem~\ref{thm:CDT701}, all $f_i$ holonomic. Let $\varphi_0,\dots,\varphi_l$ be holomorphic maps $(\DD,0)\to(\CC,0)$ with $|\varphi_0'(0)|<\dots<|\varphi_l'(0)|$ and $\log|\varphi_l'(0)|>\tau(b;e)$, such that every $f_i(\varphi_k(z))$ is meromorphic on $|z|<1$. Let $0=\vartheta_0<\vartheta_1<\dots<\vartheta_l< \vartheta_{l+1}=\fm$, and let $g:[0,1)\to\RR\cup\{-\infty\}$ be $g(t)=\log|\varphi_k(e^{2\pi i(\fm t-\vartheta_k)/ (\vartheta_{k+1}-\vartheta_k)})|$ for $t\in[\vartheta_k/\fm,\vartheta_{k+1}/\fm)$. Then \[ \fm\le\frac{\displaystyle\int_0^1\!\!\int_0^1\max(g(s),g(t))\,ds\,dt+\frac1\fm\sum_{k=1}^l\vartheta_k^2\log \frac{|\varphi_k'(0)|}{|\varphi_{k-1}'(0)|}}{\log|\varphi_l'(0)|-\tau(b;e)} . \] \end{htheorem} With $l=0$ (the \emph{trivial partition}) the numerator is the rearrangement integral $R(\varphi)=\iint\max(g(s),g(t))\,ds\,dt$ with $g(s)=\log|\varphi(e^{2\pi is})|$. The double integral depends only on the distribution of $g$, which for a partition is the mixture $\sum_k\frac{\vartheta_{k+1}-\vartheta_k}{\fm}\cdot (\text{law of }\log|\varphi_k|\text{ on }\TT)$. We use Theorem~\ref{thm:CDT701} in Section~\ref{sec:proof4T} and Theorem~\ref{thm:CDT602} in Sections~\ref{sec:Cprime} and \ref{sec:F2}. \begin{lemma}\label{lem:Tunivalent} If $\varphi$ is univalent on a neighbourhood of $\overline\DD$ and $\varphi(0)=0$, then $T(\varphi)=\log|\varphi'(0)|$. \end{lemma} \begin{proof} The function $(z,w)\mapsto\log|(\varphi(z)-\varphi(w))/(z-w)|$ is pluriharmonic on a neighbourhood of $\overline\DD^2$ (the quotient is holomorphic and zero-free there, by univalence), so its mean over $\TT^2$ is its value $\log|\varphi'(0)|$ at $(0,0)$; and $\iint_{\TT^2}\log|z-w|=0$. \end{proof} % =========================================================================== \section{Lysov's forms}\label{sec:lysov} % =========================================================================== \subsection{Definitions} For integers $n\ge0$ and $l$ put \[ c_{n,l}=\sum_k(-1)^{n+l}\binom{n+l}{l}\binom{n}{l-k}\binom{2n+k}{2k}\binom{n+k}{k}\binom{4n+2k}{2n+k}, \] the sum over $0\le k\le2n$ with $0\le l-k\le n$ (so $c_{n,l}=0$ unless $0\le l\le3n$), and \[ Q_n(z)=\sum_{l=0}^{3n}c_{n,l}\,z^{2l}\in\ZZ[z], \] an even polynomial of degree $6n$ (Lysov's $Q_{6n}$, \cite[formula~(21) and Lemma~8]{Lys18}; we use the formula as a definition). Let $f_1(z)=\log(1+1/z)$, $f_2(z)=\log(1-1/z)$ and $f_3=f_1f_2$, as Laurent series at $z=\infty$. Let $P_{j,n}$ be the polynomial part of $Q_nf_j$, and $R_{j,n}=Q_nf_j-P_{j,n}=O(z^{-1})$. Since $Q_n$ is even and $f_2(z)=f_1(-z)$, we have $P_{2,n}(z)=P_{1,n}(-z)$. Explicitly, with $A_N=\sum_{i=1}^N(-1)^{i+1}/i$, \begin{equation}\label{eq:P} \begin{aligned} P_{1,n}(z)&=\sum_lc_{n,l}\sum_{m=1}^{2l}\frac{(-1)^{m+1}}{m}z^{2l-m},\\ P_{3,n}(z)&=\sum_lc_{n,l}\sum_{m=1}^{l}\Bigl(-\frac{A_{2m-1}}m\Bigr)z^{2(l-m)}, \end{aligned} \end{equation} because $[z^{-2m}]f_3=-A_{2m-1}/m$ and the odd negative powers of $z$ do not occur in $f_3$. The functions are Cauchy transforms: for $|z|>1$, \[ f_1(z)=\int_{-1}^0\frac{dx}{z-x},\qquad f_2(z)=-\int_0^1\frac{dx}{z-x},\qquad f_3(z)=\int_{-1}^1\frac{\omega(x)\,dx}{z-x}, \] with $\omega$ odd and $\omega(x)=-\log(1+1/x)$ on $(0,1)$. Write $\mu_j$ for the three (signed) measures; their total variations are $1,1,4\log2$. Then $R_{j,n}(z)=\int Q_n(x)\,d\mu_j(x)/(z-x)$. \subsection{Contact orders} \begin{lemma}[Lysov]\label{lem:contact} For all $n\ge0$ and $j=1,2,3$: $\int x^kQ_n(x)\,d\mu_j(x)=0$ for $0\le k<2n$. Equivalently $R_{j,n}(z)=O(z^{-2n-1})$ as $z\to\infty$. \end{lemma} This is \cite[Lemmas~7 and 8]{Lys18}. Since \cite{Lys18} is an unrefereed preprint we give the argument in full; it follows Lysov's. \begin{proof} Put $\mathfrak d=\frac1x\frac d{dx}$ and $\mathfrak D=\frac d{dx}\circ\frac1x$, so that $\mathfrak dx^m=mx^{m-2}$, $\mathfrak Dx^m=(m-1)x^{m-2}$, both operators commute with $x\mapsto-x$, and for smooth $f,g$ on $(0,1]$ \[ \int_\epsilon^1f\,\mathfrak dg=\Bigl[\frac{fg}x\Bigr]_\epsilon^1-\int_\epsilon^1(\mathfrak Df)g,\qquad \int_\epsilon^1f\,\mathfrak Dg=\Bigl[\frac{fg}x\Bigr]_\epsilon^1-\int_\epsilon^1(\mathfrak df)g . \] \emph{Rodrigues form.} We have \[ \begin{gathered} Q_n=\kappa_n\,\mathfrak d^n\bigl[(x^2-1)^nx^{2n}\,\tilde U\bigr],\qquad\kappa_n=\frac1{(2n)!\,(n!)^2},\\ \tilde U=\mathfrak d^n[x^{2n}U],\qquad U=\mathfrak D^{2n}[(x^2-1)^{2n}x^{4n}]. \end{gathered} \] In the variable $w=x^2$ ($\mathfrak d=2\,d/dw$, $\mathfrak D^{2n}=x\,\mathfrak d^{2n}x^{-1}$, $D=d/dw$) this is the pair of identities \[ \begin{aligned} &\text{(i)}&Q_n(w)&=\frac1{n!}D^n\bigl[w^n(w-1)^nV_n(w)\bigr],\\ &\text{(ii)}&V_n(w)&=\frac{16^n}{(2n)!\,n!}D^n\bigl[w^{n+1/2}D^{2n}[w^{2n-1/2}(w-1)^{2n}]\bigr], \end{aligned} \] with $V_n(w)=\sum_k(-1)^kB_kw^k$, $B_k=\binom{2n+k}{2k}\binom{n+k}{k}\binom{4n+2k}{2n+k}$. For (i), $\frac1{n!}D^n[w^{n+k}(w-1)^n]=\sum_j\binom nj(-1)^{n-j}\binom{n+k+j}{n}w^{k+j}$; multiplying by $(-1)^kB_k$ and collecting $w^l$ gives $\sum_{k+j=l}(-1)^{n+l}\binom{n+l}{l}\binom nj B_k=c_{n,l}$. For (ii), the coefficient of $w^k$ on the right is $\frac{16^n}{(2n)!\,n!}\binom{2n}k(-1)^k\frac{\Gamma(2n+k+\frac12)}{\Gamma(k+\frac12)}\frac{(n+k)!}{k!}$, which equals $(-1)^kB_k$ by the Legendre duplication formula for $0\le k\le2n$ (the ratio simplifies to $1$; checked symbolically in \file{i1\_contact.py}), and both sides vanish for $k>2n$. The factors $U$, $x^{2n}U$, $\tilde U$ are even polynomials, and $x^{2n}U$ and $(x^2-1)^nx^{2n}\tilde U$ vanish to order $\ge2n$ at $0$. \emph{Moments.} By the Cauchy-transform representations, $[z^{-s-1}]R_{j,n}=\int x^sQ_n\,d\mu_j$. (For $f_3$ this is the identity $[z^{-2m}]f_3=-A_{2m-1}/m=-2\int_0^1x^{2m-1}\log(1+1/x)\,dx$, which follows from $\int_0^1x^{N-1}\log(1+x)\,dx=A_N/N$ for even $N$ and $\int_0^1x^{N-1}(-\log x)\,dx=1/N^2$; odd negative powers vanish on both sides.) Since $Q_n$ is even, the lemma is equivalent to \[ \begin{aligned} &\text{(O1)}&\int_0^1x^sQ_n\,dx&=0\quad(0\le s\le2n-1),\\ &\text{(O2)}&\int_0^1x^sQ_n\,\ell\,dx&=0\quad(s\text{ odd},\ s\le2n-1), \end{aligned} \] where $\ell(x)=\log(1+1/x)$. Let $h=x^s$ or $h=x^s\ell$ accordingly. \emph{Step 1.} Integrating by parts $n$ times against $\mathfrak d$ gives $\int_0^1Q_nh=\pm\kappa_n\int_0^1(x^2-1)^nx^{2n}\tilde U\cdot r\,dx$ with $r=\mathfrak D^nh$. The boundary terms vanish: $\mathfrak d^{n-1-j}[(x^2-1)^nx^{2n}\tilde U]$ is divisible by $(x^2-1)^{j+1}$ and vanishes to order $2j+2$ at $0$, while $\mathfrak D^jh=O(x^{-2j}|\log x|)$ at $0$. \emph{Step 2.} If $h=x^s$ with $s$ odd, then $r=0$ ($\mathfrak D^n$ kills odd polynomials of degree $\le2n-1$). If $h=x^s$ with $s$ even, $r=c\,x^{s-2n}$. If $h=x^s\ell$, then by induction $\mathfrak D^j[x^s\ell]=\mathfrak D^j[x^s]\,\ell+ (\text{rational})$, since $\ell'=-1/(x(x+1))$; as $\mathfrak D^nx^s=0$, $r$ is rational with poles only at $0$ (order $\le2n$) and $-1$ (order $\le n$), and $r=O(x^{-2})$ at $\infty$. Its odd part is $\mathfrak D^n[x^s\ell_{\mathrm{even}}]$ with $\ell_{\mathrm{even}}=\frac12\log(1-x^{-2})=-\sum_{j\ge1}x^{-2j}/(2j)$; here $\mathfrak D^nx^{s-2j}=0$ for $1\le j\le(s-1)/2$, and the remaining powers are $\le x^{-2n-1}$, so the odd part of $r$ is $O(x^{-2n-1})$. Put $B=x^{2n}(x^2-1)^nr$. In both nontrivial cases $B$ is a polynomial divisible by $(x-1)^n$, of degree $\le4n-2$, whose odd part $x^{2n}(x^2-1)^nr_{\mathrm{odd}}$ has degree $\le2n-1$; and $\int_0^1Q_nh=\pm\kappa_n\int_0^1\tilde U B\,dx$. \emph{Step 3.} Integrating by parts $n$ times against $\mathfrak d$ ($\tilde U=\mathfrak d^n[x^{2n}U]$) gives $\int_0^1\tilde UB=\pm\int_0^1x^{2n}U\,\mathfrak D^nB$; the boundary terms vanish because $\mathfrak D^jB$ is divisible by $(x-1)^{n-j}$ and is $O(x^{-2j})$, while $\mathfrak d^{n-1-j}[x^{2n}U]$ vanishes to order $2j+2$ at $0$. \emph{Step 4.} Integrating by parts $2n$ times against $\mathfrak D$ ($U=\mathfrak D^{2n}E$, $E=(x^2-1)^{2n}x^{4n}$) gives $\int_0^1x^{2n}U\,\mathfrak D^nB=\pm\int_0^1\mathfrak d^{2n}[x^{2n}\mathfrak D^nB]\cdot E$; the boundary terms vanish because $\mathfrak D^{2n-1-j}E$ is divisible by $(x^2-1)^{j+1}$ and vanishes to order $2j+2$ at $0$, while $\mathfrak d^j$ of a polynomial is $O(x^{-2j})$. \emph{Step 5.} $x^{2n}\mathfrak D^nB$ is a polynomial of degree $\le4n-2$ whose odd part $x^{2n}\mathfrak D^n[B_{\mathrm {odd}}]$ vanishes ($\deg B_{\mathrm{odd}}\le2n-1$). An even polynomial of degree $\le4n-2$ is killed by $\mathfrak d^{2n}$, so every integral is $0$. As sanity checks, \file{i1\_contact.py} verifies (i) and (ii) exactly for small $n$, the moments of $f_3$ for $m\le30$, and the contact orders exactly for $n\le24$. \end{proof} \begin{pcorollary}\label{cor:size} $|Q_n(x)|\le(2n+1)2^{20n}$ on $[-1,1]$, and for every integer $q\ge2$ and $j=1,2,3$, \[ |R_{j,n}(q)|\le(2n+1)\Bigl(\frac{2^{20}}{q^2}\Bigr)^n\frac{|\mu_j|}{q-1}. \] \end{pcorollary} \begin{proof} We have \[ \begin{aligned} \sum_l|c_{n,l}|&\le\sum_k\max_l\binom{n+l}l\,2^n\binom{2n+k}{2k}\binom{n+k}k\binom{4n+2k}{2n+k}\\ &\le(2n+1)2^{4n+n+4n+3n+8n}. \end{aligned} \] By Lemma~\ref{lem:contact}, $R_{j,n}(q)=q^{-2n}\int Q_n(x)x^{2n}\,d\mu_j(x)/(q-x)$. \end{proof} \subsection{Denominators} The following lemma is \cite[Lemma~9]{Lys18} at prime-power level; we give the proof in full, so that \cite{Lys18} is used only for the definition of $c_{n,l}$. \begin{lemma}\label{lem:den} For every $n\ge1$ and $j=1,2,3$, the coefficients of $P_{j,n}$ have denominators dividing $[1,\dots,4n]\,[1,\dots,2n]$. \end{lemma} \begin{proof} Write $D_k=[1,\dots,k]$ and fix a prime $p$. We use Kummer's theorem in the form $v_p\binom{A}{B}=\sum_{i\ge1}(\lfloor A/p^i\rfloor-\lfloor B/p^i\rfloor-\lfloor(A-B)/p^i\rfloor)$, each term being $0$ or $1$. In every nonzero summand of $c_{n,l}$ we have $l-n\le k\le2n$. \emph{$P_1$ and $P_2$.} By \eqref{eq:P} the denominators of $P_{1,n}$ divide the numbers $m\le2l\le6n$, with coefficient $c_{n,l}$. If $p\le2n$, then $v_p(m)\le\lfloor\log_p6n\rfloor\le\lfloor\log_p4n\rfloor+1\le v_p(D_{4n}D_{2n})$. If $2n6n$, so $v_p(m)\le1=v_p(D_{4n})$. If $p>4n$ and $p\mid m$, then $4n3n$); and $e_2\le\lfloor\log_p4n\rfloor+\delta_2$, where $\delta_2=1$ if some $p^b\in(4n,2m-1]$ and $\delta_2=0$ otherwise (again $2p^b>6n$). If $\delta_1+\delta_2>0$ then $m>2n$, so $l>2n$ and $k\ge l-n>n$ in every summand. If $\delta_1=1$: $n2n+2$, so the sums over $k$ telescope; the finitely many boundary cases are checked directly (\file{i1\_certs.py}). \emph{Order-4 operator.} In generating-function form the relations say $D_nQ_n=0$, with $D_n$ an operator of order $4$ in $w\,d/dw$ whose leading coefficient $-2(w-1)^2$ does not depend on $n$, and $\alpha(n)Q_{n+1}=E_nQ_n$ with $E_n$ of order $3$. Reducing $E_{n+i}\cdots E_n$ modulo $D_n$ on the right expresses $Q_{n+i}$, $i=0,\dots,4$, through $Q_n,\dots,(w\,d/dw)^3Q_n$; the cofactor vector of the resulting $4\times5$ matrix, made primitive, is $(p_0,\dots,p_4)$, and $\sum_ip_iQ_{n+i}=0$ is re-verified exactly (\file{i1\_opalg.py}). \emph{The numerators.} $\sum_ip_i(n,z^2)P_{j,n+i}=-\sum_ip_i(n,z^2)R_{j,n+i}$. The left side is a polynomial, and by Lemma~\ref{lem:contact} the right side is $O(z^{5-2n})$; so both vanish for $n\ge3$, and $n=0,1,2$ are checked directly. The characteristic polynomial identity is exact (\file{i1\_extras.py}). \end{proof} For $q\ge2$ the prefactor $2^{59}(8w+1)(9w-1)(25w-16)$ does not vanish at $w=q^2$, so the characteristic roots of the recurrence are the roots $\lambda$ of $\chi_{q^2}(-\lambda)$. For $2\le q\le60$ they are real, positive and distinct, and $1/\lambda_1<2^{-20}q^2$ (Arb root isolation, \file{fourterm/check\_y1.py}); we order them $\lambda_1>\lambda_2>\lambda_3>\lambda_4>0$ and put $y_k=1/\lambda_k$, so that $0\max\{6,\tau\}=6$ of Theorem~\ref{thm:CDT701} holds.) But $M_{14}(q)>0$ for every $q$ with $26\le q\le53$; the Arb values are in Table~\ref{tab:4T}, for example $M_{14}(26)=0.0092251638\pm5\cdot10^{-12}$ and $M_{14}(53)=1.3688032\pm3\cdot10^{-8}$ (\file{budget\_4t.py}; $y_2$ is a root of the exact characteristic polynomial of Proposition~\ref{prop:I1}). This contradiction proves Theorem~\ref{thm:4T}, given: \begin{proposition}\label{prop:indep} Let $a\in\CC^3\setminus\{0\}$. The $14$ functions $1,H_a,\vartheta H_a,\dots,\vartheta^{12}H_a$ are linearly independent over $\CC(y)$, hence over $\QQ(y)$. \end{proposition} The rest of this section proves Proposition~\ref{prop:indep}. For $a=(0,0,1)$ it is the case used for Corollary~\ref{cor:F1}, which we also certified separately (Remark~\ref{rem:F1cert}). \subsection{An independence lemma} Fix a base point $y_*\in(y_1,y_2)$ which is a regular point of $L_{13}$, and consider germs at $y_*$; $\gamma$ runs over closed paths at $y_*$ avoiding all singular points of $L_{13}$. Let \[ U_a=\operatorname{span}_\CC\{(\gamma-1)H_a\},\qquad S=\Sol(L_{13})\ (\dim S=13). \] By Proposition~\ref{prop:I1prime}, $U_a\subseteq S$. $U_a$ is invariant under monodromy, since $\delta(\gamma-1)H_a=(\delta\gamma-1)H_a-(\delta-1)H_a$. \begin{lemma}\label{lem:indep} If $\dim U_a=u$, then $1,H_a,\vartheta H_a,\dots,\vartheta^{u-1}H_a$ are linearly independent over $\CC(y)$. \end{lemma} \begin{proof} Suppose $c_0+KH_a=0$ with $K=\sum_{j0$, the margin $M_{14}(q)$ (midpoint of an Arb ball of radius $\le5\cdot10^{-8}$, truncated to six decimals), $y_2$, the base point $y_*$, and the midpoints of the two $13\times13$ determinants of Proposition~\ref{prop:irred} (\file{code/thm1\_cor2/fourterm/summary\_4t.json}). The jump between $q=37$ and $q=38$ is a change of scaled coordinates (base point and scaling), not a failure.}\label{tab:4T} \small \begin{tabular}{rrrrrll} \toprule $q$ & $\fm_{\min}$ & $M_{14}(q)$ & $y_2$ & $y_*$ & det (a) & det (c)\\ \midrule 26 & 14 & 0.009225 & 156.368 & 66 & $1.634\cdot10^{14}\,i$ & $2.307\cdot10^{89}\,i$ \\ 27 & 12 & 0.082012 & 169.118 & 71 & $1.605\cdot10^{13}\,i$ & $1.657\cdot10^{91}\,i$ \\ 28 & 11 & 0.152055 & 182.368 & 77 & $1.985\cdot10^{12}\,i$ & $1.121\cdot10^{93}\,i$ \\ 29 & 10 & 0.219554 & 196.119 & 82 & $2.151\cdot10^{11}\,i$ & $5.668\cdot10^{94}\,i$ \\ 30 & 9 & 0.284686 & 210.369 & 88 & $2.871\cdot10^{10}\,i$ & $2.739\cdot10^{96}\,i$ \\ 31 & 8 & 0.347612 & 225.119 & 95 & $4.584\cdot10^{9}\,i$ & $1.253\cdot10^{98}\,i$ \\ 32 & 7 & 0.408477 & 240.369 & 101 & $6.520\cdot10^{8}\,i$ & $6.072\cdot10^{99}\,i$ \\ 33 & 7 & 0.467411 & 256.119 & 108 & $1.094\cdot10^{8}\,i$ & $5.079\cdot10^{101}\,i$ \\ 34 & 7 & 0.524534 & 272.370 & 114 & $1.669\cdot10^{7}\,i$ & $2.784\cdot10^{103}\,i$ \\ 35 & 6 & 0.579952 & 289.120 & 121 & $2.964\cdot10^{6}\,i$ & $1.637\cdot10^{105}\,i$ \\ 36 & 6 & 0.633764 & 306.370 & 129 & $6.013\cdot10^{5}\,i$ & $1.007\cdot10^{107}\,i$ \\ 37 & 6 & 0.686062 & 324.120 & 136 & $1.120\cdot10^{5}\,i$ & $4.311\cdot10^{108}\,i$ \\ 38 & 5 & 0.736928 & 342.370 & 144 & $7.143\cdot10^{27}\,i$ & $3.168\cdot10^{92}\,i$ \\ 39 & 5 & 0.786438 & 361.120 & 152 & $1.535\cdot10^{27}\,i$ & $6.101\cdot10^{93}\,i$ \\ 40 & 5 & 0.834663 & 380.370 & 160 & $3.367\cdot10^{26}\,i$ & $1.076\cdot10^{95}\,i$ \\ 41 & 5 & 0.881668 & 400.120 & 168 & $7.537\cdot10^{25}\,i$ & $1.747\cdot10^{96}\,i$ \\ 42 & 5 & 0.927512 & 420.371 & 177 & $1.865\cdot10^{25}\,i$ & $2.758\cdot10^{97}\,i$ \\ 43 & 5 & 0.972252 & 441.121 & 185 & $4.342\cdot10^{24}\,i$ & $3.843\cdot10^{98}\,i$ \\ 44 & 4 & 1.015940 & 462.371 & 194 & $1.110\cdot10^{24}\,i$ & $5.232\cdot10^{99}\,i$ \\ 45 & 4 & 1.058624 & 484.121 & 203 & $2.884\cdot10^{23}\,i$ & $9.447\cdot10^{100}\,i$ \\ 46 & 4 & 1.100350 & 506.371 & 213 & $8.133\cdot10^{22}\,i$ & $2.221\cdot10^{102}\,i$ \\ 47 & 4 & 1.141158 & 529.121 & 222 & $2.178\cdot10^{22}\,i$ & $4.203\cdot10^{103}\,i$ \\ 48 & 4 & 1.181090 & 552.371 & 232 & $6.302\cdot10^{21}\,i$ & $8.202\cdot10^{104}\,i$ \\ 49 & 4 & 1.220180 & 576.121 & 242 & $1.847\cdot10^{21}\,i$ & $1.468\cdot10^{106}\,i$ \\ 50 & 4 & 1.258466 & 600.371 & 252 & $5.482\cdot10^{20}\,i$ & $2.422\cdot10^{107}\,i$ \\ 51 & 4 & 1.295978 & 625.121 & 263 & $1.740\cdot10^{20}\,i$ & $4.090\cdot10^{108}\,i$ \\ 52 & 4 & 1.332747 & 650.371 & 273 & $5.295\cdot10^{19}\,i$ & $5.802\cdot10^{109}\,i$ \\ 53 & 4 & 1.368803 & 676.122 & 284 & $1.717\cdot10^{19}\,i$ & $8.436\cdot10^{110}\,i$ \\ \bottomrule \end{tabular} \end{table} \begin{proof}[Proof of Proposition~\ref{prop:indep}, Theorem~\ref{thm:4T} and Corollary~\ref{cor:F1}] By Lemma~\ref{lem:nonzero}, $U_a$ is a nonzero monodromy-invariant subspace of $S$, so $U_a=S$ by Proposition~\ref{prop:irred}; then Lemma~\ref{lem:indep} with $u=13$ gives Proposition~\ref{prop:indep}, and the reduction above gives Theorem~\ref{thm:4T}. Corollary~\ref{cor:F1} is the case $26\le q\le32$ of Theorem~\ref{thm:4T}: if $\theta_q=u/v$ with $u,v\in\ZZ$, $v\ne0$, then $(r_0,a)=(-u,0,0,v)$ is a nonzero relation. \end{proof} \begin{remark}[Independent certificate for Corollary~\ref{cor:F1}]\label{rem:F1cert} Corollary~\ref{cor:F1} was first proved by a different final step: for $26\le q\le32$ the $14$ continuations of $H_3$ along the words $\varnothing,2,3,A2,A3,A^22,A^23,A^32,A^33,A^42,A^43,A^52,A^62,A^72$ have a nonzero $14\times14$ Arb determinant (\file{cert\_det.py}), for example \[ \begin{gathered} \relax[257906186217.156323475542228360\pm3.7\cdot10^{-19}]\,i\quad(q=26),\\ [992625.090850805422458198252898\pm3.9\cdot10^{-25}]\,i\quad(q=32). \end{gathered} \] So $\dim\operatorname{span}\{\gamma H_3\}=14$, and the elementary analogue of Lemma~\ref{lem:indep} gives the independence of $1,H_3,\dots,\vartheta^{12}H_3$. (The passage from coordinate vectors to germs uses $E\ne0$.) Independent reruns at $q=26$ and $q=32$ with initial data that do not use Lemma~\ref{lem:contact} (base $b=2^{-22}$, the crude bound $(2n+1)2^{20n}|\mu_3|/(q-1)$) and a different discretisation reproduce these determinants to all $30$ printed digits, and give the same irreducibility certificates. \end{remark} \begin{remark} Smaller holonomy budgets suffice for larger $q$ ($\fm_{\min}$ in Table~\ref{tab:4T}), but the proof uses $\fm=14$ at every $q$. No case analysis on the relation vector and no input from linear forms in logarithms is needed: only $a\ne0$ is used. \end{remark} % =========================================================================== \section{Widening CDT's range: proof of Theorem~\ref{thm:Cprime}}\label{sec:Cprime} % =========================================================================== \subsection{CDT's construction}\label{sec:CDTconstruction} We recall \cite[\S14]{CDT}. For an odd integer $a$ with $|a|\ge3$ let $A(a,x)=(1-2ax+x^2)^{-1/2}=\sum_ku_k(a)x^k$ and $H(a,x)=A(a,x)\int_0^xA(a,t)\,dt=\sum_kv_k(a)x^k$, so that $v_k(a)/u_k(a)\to\frac12\log\frac{a+1}{a-1}$. For distinct odd $a=2m+1$, $b=2n+1$ put $\lambda_a=\frac12\ell_m$, $\lambda_b=\frac12\ell_n$, and suppose $r_0+r_a\lambda_a+r_b\lambda_b+r_{ab}\lambda_a\lambda_b=0$ with $r\in\ZZ^4$, $(r_a,r_b,r_{ab})\ne0$. CDT form \[ P(x)=\sum_k\bigl(r_0u_k(a)u_k(b)+r_av_k(a)u_k(b)+r_bu_k(a)v_k(b)+r_{ab}v_k(a)v_k(b)\bigr)x^k \] (a combination of Hadamard products) and symmetrise: $G(y)=P(x)+P(x/(x-1))$ with $y=x^2/(x-1)$ \cite[Def.~14.1.9]{CDT}. By \cite[Prop.~14.1.11]{CDT}, $G$ has denominator type $[1,\dots,2k]^2$, satisfies a differential equation $\mathcal L_{a,b}$ whose singular points are $0,4,\infty$ and the four points $Z_4(a,b)$, the $y$-images of the four products $x=(a\pm\sqrt{a^2-1})(b\pm\sqrt{b^2-1})$ (plus apparent singularities at the roots of a polynomial $R_{10}$), and $G$ is overconvergent at the image $\alpha$ of the smallest product. With the seven ``pure'' functions $B_1,\dots,B_7$ of \cite[\S10]{CDT} and ten functions built from $G$ ($G,G',G'',G'''$ and six primitives), \cite[Lemma~14.3.1]{CDT} gives $\fm=17$ functions, linearly independent over $\QQ(y)$, with \[ \begin{gathered} \sigma=(0,2,2,4,\dots,4),\qquad e=(0,0,1;0,0,0,0,0,0;1,1,1,1,1,1,1,1),\\ \tau=\frac{1136}{289}+\frac{78419}{242760}=\frac{1032659}{242760}=4.2538\ldots \end{gathered} \] \cite[(14.5.2)--(14.5.4)]{CDT}. The order of the functions is CDT's: $B_1,B_2,B_3$; $B_4,B_5,G,G',G'',G'''$; $B_6,B_7$ and the six primitives. The map is $\varphi=h\circ\psi$, with \[ h(\mathsf q)=-256\,\mathsf q\prod_{k\ge1}(1+\mathsf q^k)^{24}=-256\Bigl(\frac{\eta(2\tau)}{\eta(\tau)}\Bigr)^{24}, \] the hauptmodul of $\Gamma_0(2)$ in the cusp coordinate $\mathsf q=e^{2\pi i\tau}$, and $\psi:\DD\to\DD$ holomorphic with $\psi^{-1}(0)=\{0\}$; admissibility of $\varphi$ is checked through \cite[Cor.~9.0.19]{CDT}. The constant $10^{-6}$ of \cite[Thm.~C]{CDT} enters only through \cite[Lemma~14.2.4]{CDT} (which needs $|a/b\mp1|<\frac12$), \cite[Lemma~14.2.6]{CDT} (which needs $0<|a/b-1|<10^{-3}$), and the location of the singularities, \cite[\S14.4]{CDT}, which determines the admissible maps. We change only the map. \subsection{Location of the singular values} \begin{lemma}\label{lem:S} Let $0<\varepsilon\le10^{-4}$ and let $m\ne n$ be integers in $\ZZ\setminus\{-1,0\}$ with $|m/n-1|<\varepsilon$; put $a=2m+1$, $b=2n+1$. Then the three genuine singular values (the points of $Z_4(a,b)$ other than $\alpha$) are real with $|y|\ge Y_1:=1/\varepsilon-6$, and $\alpha$ is real with $|\alpha|\le Y_0:=\varepsilon^4/16$. \end{lemma} \begin{proof} Let $A=|a|$, $B=|b|$, $u=a+\operatorname{sgn}(a)\sqrt{a^2-1}$, $v$ likewise; the products are $x\in\{uv,u/v,v/u,1/(uv)\}$, all real since $A,B\ge3$, and $y=x^2/(x-1)=(x-1)+2+1/(x-1)$. By the mean value theorem for $\operatorname{arccosh}$, $\delta:=|\log|u|-\log|v||\le|A-B|/\sqrt{\min(A,B)^2-1}$. For positive pairs $|A-B|/\min(A,B)<\varepsilon/(1-\varepsilon)$, for negative pairs $\le2\varepsilon|n|/(2|n|(1-\varepsilon)-1)\le\varepsilon/(1-1.5\varepsilon)$, using $|n|\ge1/\varepsilon$; and $\min(A,B)\ge2/\varepsilon-3$, so the square-root factor is at most $1+\varepsilon^2/7$. Hence $\delta\le\varepsilon(1+1.6\varepsilon)$, and $x\in\{u/v,v/u\}$ has $|x-1|\le e^\delta-1\le\varepsilon(1+3\varepsilon)$, so $|y|\ge(1-3\varepsilon)/\varepsilon-2-2\varepsilon>1/\varepsilon-6$. For $x=uv$, $|x|\ge(2\min(A,B)-1)^2$ gives $y>1/\varepsilon$; for $x=1/(uv)$, $|\alpha|\le2/|uv|^2<\varepsilon^4/16$. \end{proof} The constant $6$ cannot be dropped entirely: CDT's ``$|y|>1/\varepsilon$'' fails by $O(1)$. At $(m,n)=(-10^6,-10^6-1)$ the smallest genuine singular value has $|y|=10^6-1.50$; at $\varepsilon=3\cdot10^{-6}$ the worst case is $1/\varepsilon-1.83$, at $(m,n)=(-333333,-333334)$. This does not affect \cite[Thm.~C]{CDT}, whose map has room. \subsection{Admissibility of one map for the whole range} Let $\varphi=h\circ\psi$ with $\psi$ entire, $\psi^{-1}(0)=\{0\}$, and suppose \begin{enumerate} \item[(a)] $|\psi|<1$ on $\TT$; \item[(b)] $|\varphi|>2Y_0$ on $\TT$; \item[(c$'$)] $|\varphi|2Y_0$ on $\TT$, the argument principle shows that every point of $U=D(0,2Y_0)$ has exactly one preimage in $\DD$; this is the univalent leaf over $U\ni\alpha$ required by \cite[Cor.~9.0.19]{CDT}, with $Y^0=\{\alpha\}$ and $Y^1$ the three genuine singular values. (Cor.~9.0.19 leaves $U$ free, subject to being a contractible neighbourhood of $0$ in $\CC\setminus\{4\}$ containing $\alpha$ on which $G$ is holomorphic; $G$ is holomorphic on its principal disc, which contains $D(0,2Y_0)$.) So the $17$ pullbacks $f_i\circ\varphi$ are meromorphic on $\DD$, for every pair with $|m/n-1|<\varepsilon$ at once. \subsection{The certificate}\label{sec:certC} We take $\psi(z)=z\exp(\sum_{k=0}^{96}c_kz^k)$ with dyadic coefficients $c_k\in2^{-40}\ZZ$ (the certified object; stored in \file{code/thm3/runs/cert3\_p4\_C\_eps3.0e-06\_eps3.0e-06\_l3.json}), found by an admissibility-aware L-BFGS optimisation of $L-T/17$ (started from lune-shaped maps and from CDT's map). Then $L=\log|\varphi'(0)|=\log256+c_0$ exactly. We apply Theorem~\ref{thm:CDT602} with the four maps $\varphi_k(z)=\varphi(r_kz)$, where \[ (r_0,r_1,r_2,r_3)=\Bigl(\frac1{2048},\frac{99}{256},\frac{215}{256},1\Bigr),\quad (\vartheta_1,\vartheta_2,\vartheta_3)=\Bigl(\frac{61}{200},\frac{231}{200},\frac{1109}{500}\Bigr); \] the $\varphi_k$ are restrictions of $\varphi$, so they are admissible. \emph{Certification} (\file{cert3.py}, \file{cert3d.py}). The circle is cut into dyadic intervals, adaptively. On each interval, an Arb value of $\log\varphi$ at the midpoint (via $\eta$-quotients) and a rigorous bound for $|d\log\varphi/d\vartheta|$, namely $(1+24\sum_kk\rho^k/(1-\rho^k))\cdot\sup|1+zg'(z)|$ with $\rho\ge\sup|\psi|$ on the interval and $g=\sum c_kz^k$, give an upper step function for $\log|\varphi|$, hence an exact rational upper bound for the rearrangement integral of the mixture, and decide (a), (b), (c$'$). The results at $\varepsilon=3\cdot10^{-6}$: \begin{itemize} \item (a), (b), (c$'$) hold; $\max_\TT\log|\varphi|\le13801311154345/2^{40}=12.55222<\log Y_1=12.71688$; \item $L=4.7983738127618231\ldots$, mixture integral $\le9.08327$, penalty $\frac1{17}\sum\vartheta_k^2\log(r_k/r_{k-1})=2.51411\ldots/17$, so that the numerator is $[9.23115527305872\pm1.6\cdot10^{-15}]$; \item hence $17\le\text{numerator}/(L-\tau)\le16.95200$ would be needed: contradiction, with margin $L-\text{numerator}/17-\tau\ge0.0015378$. \end{itemize} The margin is thin ($0.048$ in the function count) but certified, and all slack of the step functions is in the safe direction. Certificates for $\varepsilon=2\cdot10^{-6}$ and $2.5\cdot10^{-6}$ (trivial partition; margins $\ge0.01138$ and $\ge0.00131$) and a certificate for CDT's own map at $10^{-6}$ (margin $\ge0.0308$) serve as cross-checks; only the $3\cdot10^{-6}$ certificate is needed, since Lemma~\ref{lem:S} at $\varepsilon=3\cdot10^{-6}$ covers every pair below it. An independent non-rigorous sampling with a $\mathsf q$-product instead of $\eta$ gives numerator $\approx9.2297$. \subsection{The \texorpdfstring{$\varepsilon$}{epsilon}-dependent steps of CDT's chain}\label{sec:phi6} Lemma~14.2.4 of \cite{CDT} needs $|a/b\mp1|<\frac12$, and Lemma~14.2.6 needs $0<|a/b-1|<10^{-3}$; here $|a/b-1|<\varepsilon(1+2\varepsilon)\le10^{-5}$. The proof of Lemma~14.2.6 uses only $b-a=2(n-m)\in\ZZ\setminus\{0\}$, so $|m-n|\ge2$ changes nothing. \cite[Lemma~14.3.1]{CDT} needs a $10\times10$ determinant $\det(a,b)$ to be nonzero. We recomputed it exactly from the operator of \cite[Prop.~14.1.11]{CDT} and the reduction \cite[(14.3.2)--(14.3.5)]{CDT}, with numerators $S_k$ of the shape \cite[(14.3.6)]{CDT} (one for each basis vector of the unknown $Q$), and proved (Proposition~\ref{prop:detid} below) \[ \det(a,b)(a-b)^{86}=\kappa\,a^6b^6(a+b)^{24}Q_1^4Q_2^4Q_3^4Q_4^2Q_5^2\,\Phi_4\Phi_6\Phi_{14}^3\Phi_{26}\Phi_{79}^3, \] with $\kappa=645700815/37778931862957161709568$, $Q_1=4a^2-4ab+1$, $Q_2=4b^2-4ab+1$, $Q_3=4a^2-4ab+4b^2-3$, $Q_4=16a^2-40ab+16b^2+9$, $Q_5=80a^2-128ab+80b^2-45$, and $\Phi_d$ irreducible of $a$-degree $d$. The factors are handled as follows. \begin{itemize} \item $a$, $b$, $a\pm b$ are nonzero, and the $Q_i$ are odd for odd $a,b$. \item $\Phi_4=80a^4-88a^3b+24a^2b^2-399a^2-88ab^3+786ab+80b^4-399b^2+9$ is odd for odd $a,b$, hence nonzero. (CDT's Lemma~14.2.6 asserts $\Res(R_{10},y(4-y))\ne0$, but its proof treats only $\Phi_{14}$ and $\Phi_{79}$.) \item For $\Phi_{14}$ and $\Phi_{79}$ we use the argument in the proof of CDT's Lemma~14.2.6, which uses only $|b-a|\ge1$ and $|a/b-1|\le10^{-3}$; a larger $|b-a|$ only shrinks the error terms. \item $\Phi_{26}$: Lemma~14.2.4 gives $\Res(R_4,R_{10})\ne0$ by irreducibility and division, for $|a/b-1|<\frac12$. \item $\Phi_6$ divides none of $\Delta_y(R_{10})$, $\Res_y(R_{10},R_4)$, $\Res_y(R_{10},y(y-4))$, contrary to the text of \cite{CDT}, so neither lemma covers it. Explicitly, \[ \begin{aligned} \Phi_6(a,b)={}&59290a^6b^2-185493a^5b^3+249865a^4b^4-185493a^3b^5+59290a^2b^6\\ &+8491a^5b-65541a^4b^2+119686a^3b^3-65541a^2b^4+8491ab^5\\ &-6090a^4+15225a^3b-21000a^2b^2+15225ab^3-6090b^4\\ &+975a^2-4260ab+975b^2+1755, \end{aligned} \] and $\Phi_6(a,b)\equiv16\pmod{32}$ for all odd $a,b$ (check over the $16\times16$ odd residue classes mod $32$). So $\Phi_6(a,b)\ne0$ at every odd pair. (Also $|\Phi_6(a,b)|/|a|^8\ge1864$ whenever $a\ne b$ are integers with $|a/b-1|<1/20$, so $\Phi_6$ has no integer zeros near the diagonal.) \end{itemize} Hence, given CDT's set-up, \cite[Lemma~14.3.1]{CDT} holds for all odd $a\ne b$ with $|a/b-1|<10^{-3}$, with no extra hypothesis. The set of irreducible factors agrees with CDT's printed list; the multiplicities differ, which does not matter. A misprint in \cite{CDT}: the factor ``$1+4b^2-2ab$'', printed in both (14.2.2) and (14.2.3), should read $1+4b^2-4ab$. \begin{proposition}\label{prop:detid} Given the shape \cite[(14.2.1)]{CDT} of the operator of \cite[Prop.~14.1.11]{CDT} and the reduction \cite[(14.3.2)--(14.3.5)]{CDT}, the displayed identity for $\det(a,b)(a-b)^{86}$ holds in $\QQ(a,b)$, hence at every pair of integers $a\ne b$. \end{proposition} \begin{proof}[Proof (computer-assisted, exact)] \emph{The operator.} CDT print only $R_4$, $R_{10}$ and $c_3/c_4$ of their operator $\mathcal L=\sum_{i\le4}c_iD^i$. We interpolate a candidate $\tilde{\mathcal L}\in\QQ[a,b,y][D]$ with $c_4=(y-4)^2y^3R_4R_{10}$ and verify it: the coefficients of $y^k$, $k<63$, of $\tilde{\mathcal L}(G_A)$ vanish identically in $\QQ[a,b]$, where $G_A=\sum_nP_n(a)P_n(b)p_n(y)$ ($P_n$ Legendre polynomials, $p_n(x+w)=x^n+w^n$ for $w=x/(x-1)$), and these $63$ conditions on the $63$ unknowns of CDT's ansatz have rank $63$ (already at $(a,b)=(7,5)$). So the ansatz has a unique solution over $\QQ(a,b)$, and CDT's operator is $\tilde{\mathcal L}$; also $c_3/c_4$ agrees symbolically with CDT's printed formula (\file{f1\_Lsym.py}). \emph{An a priori degree bound.} With $\tilde{\mathcal L}$, the $10$ numerators $S_k=y^4R_{10}^4\sum_ia_iy^i$ obtained from $b_3=(y-4)^2R_4y^k/R_{10}$, $k=0,\dots,9$ (CDT's (14.3.6) is the case of a general $Q=\sum_kq_ky^k$, with one numerator $S_{42}$), are explicit in $\QQ[a,b,y]$. In the variables $t=a-b$ and $b$, the leading coefficient of $R_{10}$ is $-32t^4$, and the remainders of $S_k$ modulo $R_{10}$ give $\det(a,b)=\det M'(t,b)/t^{200}$ with $M'$ a polynomial matrix. Summing the largest entry degrees column by column gives $\deg_b\det M'\le690$ (and $\deg_t\det M'\le685$), using the entries only (\file{f1\_detsym.py}). \emph{The identity.} It is equivalent to $\det M'(t,b)-\kappa\,t^{114}\Pi(t+b,b)=0$, where $\kappa\Pi$ is the right side; this polynomial has $b$-degree at most $\max(690,\deg\Pi)=690$. For each of the $692$ integers $b_0=-346,\dots,345$ we computed $\det M'(t,b_0)\in\QQ[t]$ exactly (fraction-free elimination, every division checked exact) and found it equal to $\kappa t^{114}\Pi(t+b_0,b_0)$ (\file{f1\_identity.py}, about $35$ minutes). A polynomial of $b$-degree $\le690$ that vanishes on $692$ lines $b=b_0$ is zero. \end{proof} An earlier version of this computation interpolated $\det(a,b_0)(a-b_0)^{86}$ on ten lines from $420$ values each, which gives the right answer (it is a consequence of Proposition~\ref{prop:detid}) but, by the bound above, did not by itself prove it (\file{code/thm3/f1/}). \begin{proof}[Proof of Theorem~\ref{thm:Cprime}] A relation for a pair with $|m/n-1|<3\cdot10^{-6}$ gives, by \cite[Prop.~14.1.11, Lemma~14.3.1]{CDT} and Section~\ref{sec:phi6}, $17$ functions as in Section~\ref{sec:CDTconstruction}, linearly independent over $\QQ(y)$ and holonomic. The map of Section~\ref{sec:certC} is admissible for them, and Theorem~\ref{thm:CDT602} gives $17\le16.952$, a contradiction. \end{proof} % =========================================================================== \section{Pooling pairs: proof of Theorem~\ref{thm:F2}}\label{sec:F2} % =========================================================================== Suppose there are $K$ pairwise inequivalent exceptional classes with representatives $(m_k,n_k)$ in the range. Each gives a $G$-function $G_k$ and its ten associated functions (the \emph{$G$-family} of the pair); the pure functions $B_1,\dots,B_7$ do not depend on the pair. The idea is to put all $7+10K$ functions into one holonomy bound. \subsection{Joint independence} \begin{lemma}\label{lem:D} Assume \cite[Lemma~14.2.4]{CDT}. Let $(a,b)$, $(c,d)$ be integer pairs with $|a|,|b|,|c|,|d|\ge2$, $a\ne b$, $c\ne d$, $|a/b-1|<\frac12$, $|c/d-1|<\frac12$. If $Z_4(a,b)\cap Z_4(c,d)\ne\varnothing$, then $\{c,d\}=\{a,b\}$ or $\{c,d\}=\{-a,-b\}$. Moreover $0,4\notin Z_4(a,b)$. \end{lemma} \begin{proof} $Z_4(a,b)$ is the root set of the quartic $R_4(a,b;y)$ of \cite[(14.2.1)]{CDT}, which is irreducible over $\QQ$ by \cite[Lemma~14.2.4]{CDT}, with constant term $1$ and leading coefficient $4(a-b)^2$. Two irreducible quartics with a common root and constant term $1$ are equal. With $\sigma=a^2+b^2$ and $\pi=ab$ the coefficients of $y^4$ and $y^3$ are $c_4=4\sigma-8\pi$ and $c_3=32\pi^2-16\pi\sigma+16\pi-8\sigma$, and $c_3=-2c_4-4\pi c_4$ identically. So $R_4$ determines $\pi$ and $\sigma$, hence $(a+b)^2$ and $(a-b)^2$, hence $\{a,b\}$ up to a common sign. Finally $R_4(0)=1$ and $R_4(4)=(16a^2-40ab+16b^2+9)^2$, which is odd (\file{code/thm4/lemmaD.py}). \end{proof} \begin{lemma}\label{lem:H} Every branch of every function in the $G$-family of $(a,b)$ is holomorphic on $\CC\setminus(\{0,4\}\cup Z_4(a,b))$, and every branch of $B_1,\dots,B_7$ is holomorphic on $\CC\setminus\{0,4\}$. \end{lemma} \begin{proof}[Proof sketch] $P(x)$ is a combination of Hadamard products $f\star g$ with $f\in\{A(a,\cdot),H(a,\cdot)\}$ and $g\in\{A(b,\cdot),H(b,\cdot)\}$, each factor continuable along every path avoiding its two singular points. By Hadamard's multiplication theorem in its global form (deform the contour in $f\star g(x)=\frac1{2\pi i}\oint f(t)g(x/t)\,dt/t$; \cite{Had99} for the star-domain form, and \cite[Thm.~A]{LSS20} (preprint) for the path-wise form: if $f$ continues along every path avoiding a closed discrete set $S$ and $g$ along every path avoiding $T$, then $f\star g$ continues along every path avoiding $\{0\}\cup S\cdot T$), $f\star g$ continues along every path avoiding $0$ and the four products. In $y$, the two roots of $x^2-yx+y=0$ are distinct off $y\in\{0,4\}$ and avoid $\{0\}\cup\{\text{products}\}\cup\{p/(p-1)\}$ exactly off $\{0\}\cup Z_4$, so $G=P(x)+P(x/(x-1))$ continues along every path in $\CC\setminus(\{0,4\}\cup Z_4)$; derivatives and primitives inherit this. The $B_i$ are functions on $\mathbb P^1\setminus\{0,4,\infty\}$ \cite[\S10]{CDT}. \end{proof} \begin{lemma}[cross-pair independence]\label{lem:X} Make the assumptions of Theorem~\ref{thm:Cprime}. Let $(a_k,b_k)$, $k=1,\dots,K$, be odd pairs with $0<|a_k/b_k-1|<10^{-3}$, pairwise inequivalent ($\{a_j,b_j\}\ne\pm\{a_k,b_k\}$ for $j\ne k$), each with a relation vector as above. Then $B_1,\dots,B_7$ and the $K$ $G$-families are linearly independent over $\CC(y)$. \end{lemma} \begin{proof}[Proof sketch] Suppose $\sum_iS_iB_i+\sum_k\sum_lR_{kl}F_{kl}=0$ with $S_i,R_{kl}\in\CC(y)$, $F_{kl}$ the $G$-family of pair $k$. Fix $k$. CDT's proof of Lemma~14.3.1 (following Lemma~12.1.1) uses a loop $\gamma_k$, a conjugate of a small loop around a point $s\in Z_4(a_k,b_k)$, with $G_k\ne\gamma_kG_k$. By Lemma~\ref{lem:D}, $s\notin\{0,4\}\cup Z_4(a_j,b_j)$ for $j\ne k$, so by Lemma~\ref{lem:H} every branch of every $F_{jl}$ ($j\ne k$) and of every $B_i$ is holomorphic at $s$ and fixed by $\gamma_k$. Applying $1-\gamma_k$ leaves $\sum_lR_{kl}(1-\gamma_k)F_{kl}=0$, which is exactly the relation reached in CDT's single-pair proof (the varied primitives $(1-\gamma_k)\int y^e(G_k-T)$ are primitives of $y^e(1-\gamma_k)G_k$, and CDT's argument does not depend on the choice of primitive). CDT's argument, with the determinant of Section~\ref{sec:phi6}, gives $R_{kl}=0$. Doing this for every $k$ leaves $\sum S_iB_i=0$, so $S_i=0$ \cite[Lemma~12.0.1]{CDT}. \end{proof} Pairs sharing an entry, $(a,b)$ and $(a,c)$, are inequivalent, so Lemma~\ref{lem:X} covers them. As a falsification test (not used in the proof), the exact rank modulo two primes $p>2008$ of the union of the series, truncated at $y^{1000}$, with polynomial coefficients of degree $\le D$, is full for $30$ instances ($K=1,2,3$, including shared entries, pairs in the range and degenerate relation vectors), while duplicated pairs give exactly the expected nullity $10(D+1)$ (\file{code/thm4/xrank.py}, \file{runs/xrank.out}). \subsection{The pooled holonomy inequality} By Lemma~\ref{lem:X} we have $\fm=7+10K$ functions, holonomic and linearly independent over $\QQ(y)$, with $\sigma=(0,2,2,4,4,4,4\mid4^{10K})$ and $e=(0,0,1,0,0,1,1\mid(0^4,1^6)^K)$ (now in the order $B_1,\dots,B_7$, then for each pair $G,G',G'',G'''$ and the six primitives; for $K=1$ this is CDT's list reordered within the block $\sigma_i=4$, which changes neither $\tau(b)$ nor $\tau^\sharp(e)$). Evaluating $\tau^\sharp$ of \cite[(6.0.6)]{CDT} at the explicit point $\xi=169/100$ (an upper bound for the minimum, hence a valid $\tau$) gives \[ \tau_2=\frac{1106977681}{262702440}=4.2138081\ldots,\qquad\tau_3=\frac{29601442477}{7082066992}=4.1797744\ldots \] for $K=2$ and $K=3$ (\file{tau\_pool.py}; CDT's $\tau$ for $K=1$ is reproduced by the same code). For the maps of this section the disc condition (c$'$) fails, and we use instead: \begin{lemma}[ray lemma]\label{lem:ray} Let $\varphi$ be holomorphic on a neighbourhood of $\overline\DD$ and $\Gamma=(-\infty,-Y_1]\cup[Y_1,\infty)$. If $\varphi(\TT)\cap\Gamma=\varnothing$, then $\varphi(\overline\DD)\cap\Gamma=\varnothing$. \end{lemma} \begin{proof} For $y_0\in\Gamma$ the number of zeros of $\varphi-y_0$ in $\DD$ is the winding number of $\varphi(\TT)$ about $y_0$. It is constant on each half-line of $\Gamma$ (connected subsets of $\CC\setminus\varphi(\TT)$), and $0$ for $|y_0|>\max_\TT|\varphi|$. \end{proof} With Lemma~\ref{lem:S} (real singular values), (a), (b) and ``$\varphi(\TT)\cap\Gamma=\varnothing$'' make $\varphi$ admissible for all pairs in the range simultaneously. The certificate decides $\varphi(s)\notin\Gamma$ on each interval either by $|\varphi|54$, and Hata's theorem gives the four-term independence. \end{proof} % =========================================================================== \section{What is computer-assisted, and how to reproduce it}\label{sec:computer} % =========================================================================== The folder \file{code/} of the accompanying repository \url{https://github.com/telaviv/log-products-irrationality} contains every script and certificate output used above, in one directory per result; \file{code/README.md} maps each number of this paper to a script and an output file. Software: Python 3.12, python-flint 0.9.0 (FLINT/Arb ball arithmetic), SymPy 1.14, mpmath 1.3, NumPy and SciPy (optimisation only). The computer-assisted steps are: \begin{center} \small \begin{tabular}{p{4.3cm}p{1.4cm}p{3.6cm}p{1.6cm}} \toprule step & kind & script (folder) & time\\ \midrule Relations for $c_{n,l}$; order-4 recurrence; char.\ polynomial (Prop.~\ref{prop:I1}) & exact & \file{i1\_*.py} (\file{thm1\_cor2}) & minutes\\ Contact orders, sanity checks (Lemma~\ref{lem:contact}) & exact & \file{i1\_contact.py} & 10 s\\ $L_{13}$ and I1$'$ per $q$ (Prop.~\ref{prop:I1prime}) & exact & \file{findop.py}, \file{i1\_rec25.py}, \file{fourterm/mkrec.py} & seconds per $q$\\ Budget $M_{14}(q)$ & Arb & \file{fourterm/budget\_4t.py} & seconds\\ Local exponents, Casoratian, $p_0$ roots (Prop.~\ref{prop:exponents}, Lemma~\ref{lem:nonzero}) & exact & \file{exponents.py} (\file{fourterm}) & 15 s\\ Certified continuation, $g_A,g_2,g_3$ & Arb & \file{mono13.py} & $\approx40$ min per $q$\\ Irreducibility (Prop.~\ref{prop:irred}) & Arb & \file{fourterm/irred.py} & 2 s per $q$\\ Rank-14 cross-check (Remark~\ref{rem:F1cert}) & Arb & \file{cert\_det.py} & seconds\\ Map certificate, Theorem~\ref{thm:Cprime} & Arb + exact & \file{cert3.py}, \file{cert3d.py} (\file{thm3}) & $\approx8$ min\\ $\det(a,b)$ identity and $\Phi_6$ (Prop.~\ref{prop:detid}) & exact & \file{f1\_Lsym.py}, \file{f1\_detsym.py}, \file{f1\_identity.py}, \file{f1\_phi6\_nonvanish.py} (\file{thm3/f1}) & 35 min\\ New pairs \eqref{eq:newpairs} & exact & \file{f7\_newpairs.py} & seconds\\ Lemma~\ref{lem:D}; pooled $\tau$; falsifier & exact & \file{lemmaD.py}, \file{tau\_pool.py}, \file{xrank.py} (\file{thm4}) & minutes\\ Pooled map certificates (Theorem~\ref{thm:F2}) & Arb + exact & \file{cert\_pool.py} & $\approx4$ min each\\ \bottomrule \end{tabular} \end{center} The map optimisations (\file{lib\_opt.py}, \file{lib\_range.py}, \file{reopt.py}) and the falsifier are not part of any proof: they produce candidates that the certificates then check. The referees (Section~\ref{sec:provenance}) re-ran the main certificates bit for bit on separate copies (among them the budget and irreducibility certificates of Theorem~\ref{thm:4T}, the rank certificate at $q=29$, the $3\cdot10^{-6}$ map certificate and the $5\cdot10^{-5}$ pooled certificate), and re-ran the continuation step independently, with different initial data and discretisation, at $q=26,32$ (Remark~\ref{rem:F1cert}; outputs in \file{code/thm1\_cor2/runs/referee\_rerun/}) and at $q=40,53$ (the gate referee's reruns of Theorem~\ref{thm:4T}, which reproduce both determinants of Proposition~\ref{prop:irred} to all $30$ printed digits; their outputs are not included in \file{code/}). What the computer establishes is, in each case, a finite statement: an identity of explicit polynomials, a determinant ball that excludes $0$, or an exact rational upper bound for an integral. What it does not establish: the hand arguments (Lemmas~\ref{lem:den}, \ref{lem:indep}, \ref{lem:nonzero}, the reflection criterion, Lemmas~\ref{lem:S}, \ref{lem:D}, \ref{lem:ray}, the reductions), the theorems of \cite{CDT} we use, and the correctness of python-flint's Arb routines (in particular \texttt{acb.modular\_eta} and matrix determinants), which we take on trust. Arb \cite{Arb} is now part of FLINT \cite{FLINT}; python-flint \cite{pyflint} 0.9.0, as used, is linked against FLINT 3.6.0. % =========================================================================== \section{Limits}\label{sec:limits} % =========================================================================== \begin{itemize} \item \emph{$q\le25$ is out of reach of this route.} At $q=25$, $M_{14}(25)\approx-0.0665<0$, and $14$ is the largest budget available for this module (order $13$ plus the constant). Neither the rank nor the irreducibility can help; a new family of forms, a better denominator type or a better map would be needed. \item \emph{$10^{-5}$ is not reached.} For Theorem~\ref{thm:Cprime} the best range margin we found at $4\cdot10^{-6}$ is $-0.013$ (partition form), and even for a single pair the best margin at $10^{-5}$ is $-0.0086$ [numerical, local optima]. We have no certified ceiling: the only rigorous ceiling we know is far from tight. This agrees with \cite[Remark~14.0.4]{CDT}, which says that the constant could be improved but probably not beyond about $10^{-4}$. \item \emph{Pooling stops near $6\cdot10^{-5}$ for $K=2$} on the maps we explored (margins $-0.009$ at $6\cdot10^{-5}$, $-0.031$ at $7\cdot10^{-5}$, $-0.069$ at $10^{-4}$, numerical). \item \emph{The main inherited assumption}, the irreducibility of $\mathcal L_{a,b}$ (asserted in CDT's template Lemma~12.1.1 and implicit in Lemma~14.3.1), is open. Its local system is the push-forward of a rank-$2$ system along the double cover $x\mapsto y$; by Mackey's criterion it would follow from the irreducibility of the rank-2 system and its non-isomorphism with its conjugate under $x\mapsto x/(x-1)$ [heuristic]. \end{itemize} % =========================================================================== \section{Provenance and use of AI tools}\label{sec:provenance} % =========================================================================== \subsection*{Provenance} The mathematical inputs are: Lysov's explicit formula for $Q_n$ \cite{Lys18}, used as a definition (his Lemmas~7--9 are re-proved above); the holonomy bounds and the construction for products of two logarithms of Calegari, Dimitrov and Tang \cite{CDT}, used as stated; Hata's theorem \cite{Hat98}, used only for the extension to $q\ge54$ and for Theorem~\ref{thm:F2}(iv). The scheme of Section~\ref{sec:proof4T} is the one suggested in \cite[\S15.1.10]{CDT}. The recurrence of Proposition~\ref{prop:I1} was found by modular linear algebra and then proved by creative telescoping; the minimal operators were found the same way. Apart from the preprints cited above (\cite{CDT}, \cite{Lys18} and, for Lemma~\ref{lem:H}, \cite{LSS20}), nothing in this paper depends on unpublished work by others. \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, 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. None of the results is verified in Lean or another proof assistant. % =========================================================================== \bibliographystyle{amsplain} \bibliography{refs} \end{document}