\documentclass[11pt]{amsart} \IfFileExists{lmodern.sty}% {\usepackage[T1]{fontenc}\usepackage{lmodern}\usepackage{microtype}}% {\usepackage[expansion=false]{microtype}} \usepackage{amssymb,mathtools} \usepackage{booktabs,array} \usepackage{enumitem} \emergencystretch=2.5em \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref} \hypersetup{pdftitle={Irrationality measures from L\"owner determinants},pdfauthor={Sean Litsey}} % --------------------------------------------------------------------------- % Location of the Lean formalization. \newcommand{\FormalizationURL}{https://github.com/slitsey/catalan-irrationality-measure} \newcommand{\FormalizationLink}{\url{\FormalizationURL}} % --------------------------------------------------------------------------- \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \setcounter{tocdepth}{1} \newcommand{\Z}{\mathbb Z} \newcommand{\Q}{\mathbb Q} \newcommand{\R}{\mathbb R} \newcommand{\C}{\mathbb C} \newcommand{\F}{\mathbb F} \newcommand{\N}{\mathbb N} \newcommand{\den}{\operatorname{den}} \newcommand{\rank}{\operatorname{rank}} \newcommand{\corank}{\operatorname{corank}} \newcommand{\lean}[1]{\nolinkurl{#1}} \renewcommand{\Re}{\operatorname{Re}} \renewcommand{\Im}{\operatorname{Im}} \newcommand{\Lw}{\mathsf L} % Loewner matrix \newcommand{\cO}{\mathcal O} \newcommand{\Id}{\mathrm{Id}} \newcommand{\err}{\operatorname{err}} \newcommand{\cI}{\mathcal I} \newcommand{\Lt}{L_{-3}} % the constant L(2,chi_{-3}) \newcommand{\MuLamU}{31.1973} \newcommand{\MuThirdsU}{9.4913} \newcommand{\MuFifthsU}{23.7756} \newcommand{\MuSeventhsU}{1806.1719} \newcommand{\MuChi}{13.6839} \newcommand{\MuHalvesU}{6.5214} \newcommand{\MuPlaneA}{20.6622} \newcommand{\MuPlaneB}{335.20} \newcommand{\MuPlaneC}{72.726} \newcommand{\MuZetaThree}{6.0382} \newcommand{\MuDipA}{14} \newcommand{\MuDipB}{15.359} \newcommand{\MuDipC}{17.492} \newcommand{\LamFour}{3.5149036} \newcommand{\DFour}{3.7453322} \newcommand{\gFour}{7.1887383} \newcommand{\ExpFour}{31.19724} \newcommand{\LamThree}{3.3104458} \newcommand{\DThree}{4.0295805} \newcommand{\gThree}{6.8255087} \newcommand{\ExpThree}{9.491278} \newcommand{\LamFive}{3.5128488} \newcommand{\DFive}{3.8141613} \newcommand{\gFive}{7.1638755} \newcommand{\ExpFive}{23.77556} \newcommand{\LamSeven}{3.6391158643} \newcommand{\DSeven}{3.6432250312} \newcommand{\gSeven}{7.4218615374} \newcommand{\ExpSeven}{1806.172} \newcommand{\LamHalf}{2.9923824} \newcommand{\DHalf}{3.9790633} \newcommand{\gHalf}{6.4345402} \newcommand{\ExpHalf}{6.521399} \newcommand{\LamChi}{4.0959278} \newcommand{\DChi}{4.6442402} \newcommand{\gChi}{7.5030223} \newcommand{\ExpChi}{13.68385} \newcommand{\LamPl}{3.2227428} \newcommand{\DPl}{3.5503512} \newcommand{\gPl}{6.7690926} \newcommand{\ExpPl}{20.66214} \newcommand{\LamDA}{3.0668400} \newcommand{\ExpDA}{13.99987} \newcommand{\LamDB}{3.1096062} \newcommand{\ExpDB}{15.35830} \newcommand{\LamDC}{3.4140176} \newcommand{\DDC}{3.8109776} \newcommand{\gDC}{6.9434368} \newcommand{\ExpDC}{17.49153} \newcommand{\LamPlB}{3.4768376} \newcommand{\DPlB}{3.4982052} \newcommand{\gPlB}{7.1600631} \newcommand{\ExpPlB}{335.0886} \newcommand{\LamPlC}{3.3496421} \newcommand{\DPlC}{3.4456170} \newcommand{\gPlC}{6.9798542} \newcommand{\ExpPlC}{72.7259} \newcommand{\LamZ}{2.912839} \newcommand{\ExpZ}{6.034884} \newcommand{\MuChiExact}{13.683844} \begin{document} \title[Irrationality measures from L\"owner determinants]{Irrationality measures from L\"owner determinants} \author{Sean Litsey} \email{sean.litsey.math@gmail.com} \date{November 10, 2026} \subjclass[2020]{Primary 11J82; Secondary 11J72, 15A15, 31A15, 68V20} \keywords{Irrationality exponent, $L(2,\chi_{-3})$, $\zeta(3)$, trigamma function, L\"owner matrices, $p$-adic valuations, Hermite interpolation, Andr\'eief's identity, logarithmic energy, formal verification, Lean} \begin{abstract} Let $\mu$ denote the irrationality exponent. One construction, by determinants of rectangular L\"owner matrices of rational combinations of the digamma function and its derivatives at integer nodes, gives irrationality measures on three planes of numbers. For all rationals $a$ and $b$, not both zero, we prove $\mu(aL(2,\chi_{-3})+b\pi^2)\le\MuPlaneA$, $\mu(a\zeta(3)+b\pi^2)\le\MuPlaneB$ and $\mu(a\zeta(3)+b\pi^4)\le\MuPlaneC$. The irrationality of these numbers was proved recently by other methods: on the first plane by Calegari, Dimitrov and Tang, on the second by Q.~Tang and on the third by Lai. On the first plane the best published bound was $24781$, due to Calegari, Dimitrov and Tang. On the second plane, as far as we know, no irrationality measure was known. On the third plane the bound $66.3$ of Leibtag et al.\ is better than ours. At special points of the planes the bounds are lower. We prove $\mu(L(2,\chi_{-3}))\le\MuChi$, and the bounds $\MuThirdsU$, $\MuDipA$, $\MuDipB$ and $\MuDipC$ for $L(2,\chi_{-3})\pm\pi^2/9$, $\pm2\pi^2/27$, $\pm\pi^2/27$ and $\pm\pi^2/18$. For completeness we also prove $\mu(\zeta(3))\le\MuZetaThree$ and $\mu(\pi^2)\le\MuHalvesU$, which are weaker than the known bounds. The construction also reaches numbers off the planes, many of which are differences of two values of the trigamma function. It gives $\mu(G\pm5\pi^2/48)\le\MuLamU$, where $G$ is Catalan's constant, and bounds $\MuFifthsU$ and $\MuSeventhsU$ for the numbers $\psi'(k/5)-\pi^2/6$ and $\psi'(k/7)-\pi^2/6$ with $k$ prime to the denominator. For these numbers we know of no other proof of irrationality. The denominators are bounded prime by prime, and on each plane one table of local bounds and one decay certificate serve every point. Nonvanishing at rational points follows from Hermite interpolation in characteristic zero. The size of the determinants is bounded by Andr\'eief's identity and an energy estimate. The proofs are formalized in Lean~4. \end{abstract} \maketitle \tableofcontents %% ===================================================================== \section{Introduction}\label{sec:intro} \subsection{The results} The \emph{irrationality exponent} $\mu(\xi)$ of a real number $\xi$ is the infimum of the real numbers $\mu$ such that $|\xi-p/q|>q^{-\mu}$ for all integers $p$ and all sufficiently large integers $q$. It does not change when $\xi$ is multiplied by a nonzero rational number or translated by a rational number. A finite bound for $\mu(\xi)$ is an irrationality measure for $\xi$, and it shows that $\xi$ is irrational. Let $\chi_{-3}$ be the nontrivial Dirichlet character modulo $3$. We write \[ \Lt=L(2,\chi_{-3})=\sum_{n\ge0}\Bigl(\frac1{(3n+1)^2}-\frac1{(3n+2)^2}\Bigr)=0.78130\,24128\ldots \] Calegari, Dimitrov and Tang gave the first proof that $\Lt$ is irrational \cite{CDT2024}, and we call $\Lt$ the Calegari--Dimitrov--Tang constant. Our main results are uniform irrationality measures on three planes of numbers. One L\"owner structure, described below, produces all three. The first plane is spanned by $\Lt$ and $\pi^2$ modulo $\Q$, the second by $\zeta(3)$ and $\pi^2$, and the third by $\zeta(3)$ and $\pi^4$. \begin{theorem}[first plane]\label{thm:plane1} Let $a$ and $b$ be rational numbers, not both zero. Then $a\Lt+b\pi^2$ is irrational and \[ \mu(a\Lt+b\pi^2)\le\MuPlaneA . \] \end{theorem} \begin{theorem}[second plane]\label{thm:plane2} Let $a$ and $b$ be rational numbers, not both zero. Then $a\zeta(3)+b\pi^2$ is irrational and \[ \mu(a\zeta(3)+b\pi^2)\le\MuPlaneB . \] \end{theorem} \begin{theorem}[third plane]\label{thm:plane3} Let $a$ and $b$ be rational numbers, not both zero. Then $a\zeta(3)+b\pi^4$ is irrational and \[ \mu(a\zeta(3)+b\pi^4)\le\MuPlaneC . \] \end{theorem} The irrationality of the numbers on these planes, that is, the linear independence of $1$ and the two spanning numbers over $\Q$, was proved in the last two years by methods unrelated to ours. Calegari, Dimitrov and Tang proved it for the first plane with an arithmetic holonomy bound \cite{CDT2024}. Q.~Tang proved it for the second plane in Lean \cite{Tang26}. Lai proved it for the third plane with Fauzan's determinant method and the simultaneous approximations of Zudilin and of Dauguet and Zudilin \cite{Lai2026}. On two of the planes our measures are better than those known. On the first plane the best published bound was $24781$, due to Calegari, Dimitrov and Tang \cite{CDT2025}. As far as we know, Theorem~\ref{thm:plane2} is the first irrationality measure valid on the whole of it. On the third plane, Leibtag et al.\ \cite{LBKWSK26} proved $\mu(a\zeta(3)+b\zeta(4))<66.3$, which is better than Theorem~\ref{thm:plane3}. The third plane is included because its function has a pole of order three, and it shows that the structure handles such a pole. At special points of the planes we prove better bounds. They do not follow from the plane theorems, and each has its own proof. On the first plane we treat one point and four pairs of points. The point $\Lt$ uses its own construction (Section~\ref{sec:chi3-quarter}). The points $\Lt\pm\pi^2/9$ use the construction of the off-plane results below (Section~\ref{sec:proof}). The other three pairs use the functions of the plane with further arithmetic that holds only at these points (\S\ref{ssec:levels}). The pair $\Lt\pm\pi^2/18$ also has its own geometry and decay certificate. \begin{theorem}[special points of the first plane]\label{thm:points1} The following bounds hold. \begin{enumerate}[label=(\roman*),leftmargin=*] \item $\mu(\Lt)\le\MuChi$. \item $\mu(\Lt\pm\frac{\pi^2}9)\le\MuThirdsU$. \item $\mu(\Lt\pm\frac{2\pi^2}{27})\le\MuDipA$. \item $\mu(\Lt\pm\frac{\pi^2}{27})\le\MuDipB$. \item $\mu(\Lt\pm\frac{\pi^2}{18})\le\MuDipC$. \end{enumerate} \end{theorem} At the two ends of the second plane we prove the following bounds (\S\ref{ssec:zeta3} and Section~\ref{sec:proof}). We include them for completeness. They are weaker than the bounds $\mu(\zeta(3))<5.4197$ and $\mu(\pi^2)<4.62675$ of Leibtag et al.\ \cite{LBKWSK26}. \begin{theorem}[the ends of the second plane]\label{thm:ends} $\mu(\zeta(3))\le\MuZetaThree$ and $\mu(\pi^2)\le\MuHalvesU$. \end{theorem} On the line $b=0$ Theorem~\ref{thm:plane3} follows from Theorem~\ref{thm:ends}. \emph{The construction.} All these bounds come from one construction. Let $\psi$ be the digamma function, the logarithmic derivative of Euler's gamma function. The approximations to a number $\xi$ are determinants of L\"owner matrices \cite{Loewner34}, the matrices of divided differences, of a rational combination $F$ of the functions $\psi(z+s)$, $\psi'(z+s)$ and $\psi''(z+s)$ with rational shifts $s$, at integer nodes. The number $\xi$ is $F'(0)$, up to a rational factor. Section~\ref{sec:overview} explains the construction and why it gives measures. \emph{The trigamma form.} The derivative $\psi'(s)=\sum_{n\ge0}(n+s)^{-2}$ is the trigamma function. When $F$ is a difference of two digamma functions, $\xi$ is a difference of two trigamma values, \[ \Psi(s,t)=\psi'(s)-\psi'(t)\qquad(0q^{-\mu_0}$ for all integers $p$ and all sufficiently large integers $q$. Table~\ref{tab:main} collects the results. \begin{table}[ht] \centering \renewcommand{\arraystretch}{1.25} \begin{tabular}{@{}l@{\hspace{2em}}l@{\hspace{2em}}r@{\hspace{2em}}l@{}} \toprule Number & Trigamma form & Bound on $\mu$ & Other bound\\ \midrule $a\Lt+b\pi^2$ & & $\MuPlaneA$ & $24781$ \cite{CDT2025}\\ $a\zeta(3)+b\pi^2$ & & $\MuPlaneB$ & none\\ $a\zeta(3)+b\pi^4$ & & $\MuPlaneC$ & $66.3$ \cite{LBKWSK26}\\ \addlinespace $\Lt$ & $\Psi(\frac13,\frac23)$ & $\MuChi$ & $24781$\\ $\Lt\pm\frac{\pi^2}{9}$ & $\Psi(\frac13,1)$,\; $\Psi(\frac23,1)$ & $\MuThirdsU$ & $24781$\\ $\Lt\pm\frac{2\pi^2}{27}$ & $\Psi(\frac16,\frac13)$,\; $\Psi(\frac23,\frac56)$ & $\MuDipA$ & $24781$\\ $\Lt\pm\frac{\pi^2}{27}$ & $\Psi(\frac13,\frac12)$,\; $\Psi(\frac12,\frac23)$ & $\MuDipB$ & $24781$\\ $\Lt\pm\frac{\pi^2}{18}$ & & $\MuDipC$ & $24781$\\ $\zeta(3)$ & & $\MuZetaThree$ & $5.4197$ \cite{LBKWSK26}\\ $\pi^2$ & $\Psi(\frac12,1)$ & $\MuHalvesU$ & $4.62675$ \cite{LBKWSK26}\\ \addlinespace $G\pm\frac{5\pi^2}{48}$ & $\Psi(\frac14,1)$,\; $\Psi(\frac34,1)$ & $\MuLamU$ & none\\ $\psi'(\frac k5)-\frac{\pi^2}{6}$ & $\Psi(\frac k5,1)$ & $\MuFifthsU$ & none\\ $\psi'(\frac k7)-\frac{\pi^2}{6}$ & $\Psi(\frac k7,1)$ & $\MuSeventhsU$ & none\\ \bottomrule \end{tabular} \caption{The results of this paper. The trigamma form is a rational multiple of the number, where one exists. In the plane rows $a$ and $b$ are rationals, not both zero. The last column gives the best bound from other work that we know. ``None'' means that we know of no irrationality measure. A bound of $1+10^4$ for the first plane follows from an unrefereed supplement to \cite{LBKWSK26} (\S\ref{ssec:background}).}\label{tab:main} \end{table} Other named numbers on the first plane are covered by Theorem~\ref{thm:plane1}, with $\mu\le\MuPlaneA$ in place of $24781$. Examples are $\psi'(\frac13)$, $\psi'(\frac23)$, $\psi'(\frac16)$ and $\psi'(\frac56)$, whose irrationality Calegari, Dimitrov and Tang state explicitly \cite[Corollary~B]{CDT2024}. The Clausen value $\mathrm{Cl}_2(\frac\pi3)=\frac{3\sqrt3}4\Lt$ and the Mahler measure $m(1+x+y)=\frac{3\sqrt3}{4\pi}\Lt$ are not on the plane, and our results say nothing about them. Theorem~\ref{thm:plane2} says nothing about $\zeta(3)/\pi^2$, which is not on the second plane. The nearest numbers that the method misses are discussed in \S\ref{ssec:unreached}. All the theorems above have been verified in the Lean~4 proof assistant, together with every theorem, proposition and lemma used in their proofs. The formal proofs have no unproved steps and use only the standard axioms. Section~\ref{sec:lean} describes them. The remarks and Section~\ref{sec:limits} discuss the method and are not part of the formal proofs. The formalization is available at \FormalizationLink. The use of AI in this work is disclosed at the end of the paper. \subsection{Background}\label{ssec:background} \emph{The constant $\Lt$ and the first plane.} Calegari, Dimitrov and Tang \cite{CDT2024} proved that $1$, $\pi^2$ and $\Lt$ are linearly independent over $\Q$. Their proof applies a new arithmetic holonomy bound to a construction of Zagier. They list the irrationality of $\psi'(\frac13)$, $\psi'(\frac23)$, $\psi'(\frac16)$ and $\psi'(\frac56)$ as a corollary, and remark that $\psi'(\frac n6)$ is irrational for every $n$. In \cite{CDT2025} they made holonomy bounds effective. Theorem~5.1 there gives $\mu(\Lt)\le24781$. The remark that follows it states that the same theorem and proof hold for any $\Q$-linear combination of $\Lt$ and $\pi^2$. So $24781$ is the best published bound at every point of the first plane. Theorem~\ref{thm:plane1} improves it to $\MuPlaneA$ everywhere, and Theorem~\ref{thm:points1} improves it further at these special points. Leibtag et al.\ \cite{LBKWSK26} find numerically that Ap\'ery-like approximations from conservative matrix fields have positive approximation exponents for $\Lt\pm\pi^2/9$. A supplement to their paper uses one such field to give an Ap\'ery-style proof of the theorem of Calegari, Dimitrov and Tang. For every point of the plane it proves an approximation exponent $\delta>10^{-4}$, and it certifies the comparison value $0.000103\ldots$. With the bound $\mu\le1+1/\delta$ of their \S2 the proved exponent gives $\mu\le1+10^4$, a consequence not stated there. The comparison value would give about $9.7\cdot10^3$. The supplement is unrefereed and is not part of the arXiv version. \emph{The values $\zeta(2)$, $\zeta(3)$ and the second plane.} Legendre proved that $\pi^2$ is irrational. Ap\'ery proved that $\zeta(3)$ is irrational and gave a new proof for $\zeta(2)$ \cite{Apery79}, and Beukers gave his integral proofs \cite{Beukers79}. Rhin and Viola obtained $\mu(\zeta(2))<5.441243$ and $\mu(\zeta(3))<5.513891$ with permutation groups of integrals \cite{RhinViola96,RhinViola01}. Zudilin lowered the bound for $\zeta(2)$ to $5.095412$ \cite{Zudilin14}. Giraudin lowered the bounds to $5.0193784$ for $\zeta(2)$ \cite{Giraudin26} and to $5.5138800$ for $\zeta(3)$ \cite{Giraudin26z3}. Leibtag, Barkan, Kalisch, Weinbaum, Shalyt and Kaminer \cite{LBKWSK26} showed that the classical constructions are one-dimensional trajectories in conservative matrix fields. By searching over other trajectories they obtained $\mu(\pi^2)<4.62675$ and $\mu(\zeta(3))<5.4197$. They also proved the uniform bound $\mu(a\zeta(3)+b\zeta(4))<66.3$ on the plane spanned by $\zeta(3)$ and $\pi^4$. This is the closest analogue of Theorem~\ref{thm:plane2}, and it is better than our Theorem~\ref{thm:plane3} on the same plane. The linear forms in $1$, $\zeta(2)$ and $\zeta(3)$ have a longer history. Dauguet and Zudilin \cite{DauguetZudilin14} bounded such forms from below when the coefficients satisfy divisibility and size conditions. They noted that this does not give the linear independence of $1$, $\zeta(2)$, $\zeta(3)$, which remained open. For a fixed number $a\zeta(3)+b\zeta(2)$ their theorem gives no measure. In 2026 Q.~Tang \cite{Tang26} proved the linear independence in Lean. This proof is a formalization with no accompanying written paper. Lai \cite{Lai2026} then proved that $1$, $\zeta(3)$ and $\zeta(4)$ are linearly independent, and Leibtag et al.\ gave a second proof with a uniform measure \cite{LBKWSK26}. As far as we know, no irrationality measure valid on the whole plane of $\zeta(3)$ and $\pi^2$ was known before Theorem~\ref{thm:plane2}. \emph{Catalan's constant.} In October 2026 OpenAI \cite{OpenAI2026} proved that $G$ is irrational and formalized the proof in Lean. Sun's preprint \cite{Sun2026} had claimed a proof of the irrationality of $G$. Wachs \cite{Wachs2026} reports that its main estimate is not justified. Finite irrationality exponents for $G$ were obtained after OpenAI's proof by Jha \cite{Jha2026}, Carreira \cite{Carreira2026}, Bojnal \cite{Bojnal2026}, Xu \cite{Xu2026}, Davis \cite{Davis2026}, the MultiAgency agent swarm \cite{MultiAgency2026} and the author \cite{Litsey2026}. Carreira \cite{Carreira2026b} also records attempts to carry the determinant method to $\beta(4)$. For the arithmetic of $G$ and related values before these results see \cite{RivoalZudilin2003,Zudilin2003,Zudilin02Cheb,KrattRivoal08,KrattZudilin19} and the survey \cite{Zudilin19Survey}. Eskandari, Murty and Nemoto \cite{EMN2025} construct a $2$-dimensional mixed motive over $\Q$ with period $G$. They use it to obtain linear forms in $1$ and $G$. Eskandari \cite{Eskandari2026} improves the exponent of rational approximations to $G$ of Rivoal--Zudilin type from $0.52$ to $0.62$. These are explicit approximations, not a lower bound for all rationals. Our searches found no proof that $G-5\pi^2/48$ or $G+5\pi^2/48$ is irrational. This does not follow from the irrationality of $G$, since it is not known whether $1$, $G$ and $\pi^2$ are linearly independent over $\Q$. \emph{The numbers $\psi'(\frac k5)-\pi^2/6$ and $\psi'(\frac k7)-\pi^2/6$.} Our searches found no irrationality result for them. Prellberg \cite{Prellberg26} proved that $\mathrm{Li}_2(1/5)$ is irrational, but that is a different number. \emph{Recent results.} Since September 2026 a series of new results has appeared in this area. Several of them were obtained with AI systems. OpenAI proved $\mu(\pi)=2$, with a Lean formalization \cite{OpenAIPi2026}. This improves the bound $\mu(\pi)<7.101863$ of Bai \cite{Bai26}, which in turn lowered the bound $7.1033$ of Zeilberger and Zudilin \cite{ZeilbergerZudilin20}. Liu, Jiang and Zhang proved $\mu(\log r)=2$ for every rational $r>0$ with $r\ne1$ \cite{LiuJiangZhang26}, improving the classical bound $\mu(\log2)<3.5746$ of Marcovecchio \cite{Marcovecchio09}. Their result also covers nonzero rational combinations of such logarithms. Prellberg \cite{Prellberg26} proved that $\mathrm{Li}_2(-1/4)$, $\mathrm{Li}_2(1/5)$ and $\mathrm{Li}_2(-1/3)$ are irrational. These are three cases left open in \cite{CDT2024}. The proof uses the integrals of Rhin--Viola and Viola--Zudilin. Fauzan proved that $\zeta(5)$ is irrational, using rationally normalized Hankel determinants \cite{Fauzan2026}. Fasten \cite{Fasten26}, Firsching \cite{Firsching26} and Romik \cite{Romik26} verified the irrationality statement in Lean. Romik's formalization takes the prime number theorem as an axiom. \emph{Determinant methods.} Irrationality proofs built on determinants of large size have a precedent in Zudilin's determinantal criterion with Hankel determinants \cite{Zudilin}. The preprint of Liu, Jiang and Zhang on logarithms \cite{LiuJiangZhang26} uses a separated-weight interpolation determinant. The arguments of Fauzan and Lai \cite{Fauzan2026,Lai2026} use Hankel determinants. Brown \cite{Brown26} proposes a multivariate transfinite-diameter test for Mellin-type integrals and applies it to $\pi^2$. The framework we use is OpenAI's \cite{OpenAI2026}. In it the decay of the determinants is controlled by Cauchy determinants and Andr\'eief's identity, and the denominators are controlled prime by prime. Our determinants are L\"owner matrices, that is, matrices of divided differences. They are a matrix-valued analogue of Pad\'e and Hermite--Pad\'e approximation \cite{NikishinSorokin91}. Formal proofs of irrationality results in Lean also include the irrationality of $\zeta(3)$ by Liu, Zhang and Zhi \cite{LZZ25} and the linear independence of $1$, $\zeta(2)$ and $\zeta(3)$ by Q.~Tang \cite{Tang26}. \subsection{The companion paper} Catalan's constant is also a difference of trigamma values, $G=\frac1{16}\Psi(\frac14,\frac34)$, and the construction applies to this pair. For it, however, our bound for the denominators exceeds our bound for the decay, so we obtain no measure for $G$. The companion paper \cite{Litsey2026} refines OpenAI's determinant instead and proves $\mu(G)\le270$. Our computations for $(\frac14,\frac34)$ are not part of the formal proof. \subsection{What a measure requires} Let $P(x)$ be a polynomial with rational coefficients for which $P(\xi)$ is small. It is a member of a family of polynomials indexed by a size $r$. Suppose that a rational number is very close to $\xi$. Three estimates turn this into a lower bound for its distance to $\xi$. \begin{enumerate}[label=(\roman*)] \item \emph{Denominators.} The coefficients of $P$ have a common denominator of size $e^{(\Lambda+o(1))r^2}$. \item \emph{Smallness on a neighbourhood.} $|P(x)|\le e^{-(D-o(1))r^2}$ for every $x$ with $|x-\xi|\le e^{-\rho r}$, not only at $\xi$. \item \emph{Nonvanishing at the point.} $P$ does not vanish at the rational number for some member of the family whose size can be chosen in terms of its denominator. \end{enumerate} In OpenAI's argument for $G$ the size of the determinant is tied to a prime $p$, and nonvanishing is proved modulo $p$. Here nonvanishing is proved in characteristic zero, so the size $r$ is free. A plane theorem needs more. A number $a\Lt+b\pi^2$ is, up to a rational factor, the value $F'(0)$ for a combination $F$ of three digamma functions whose weights depend on $(a,b)$. The three estimates must then hold with constants that do not depend on the weights. For the second plane the function $F$ has a pole of order two, so the matrices are no longer Cauchy-like at every pole, and both the arithmetic and the analysis need a new argument. For the third plane the pole has order three. \subsection{New ideas} The proofs rest on the following ideas. Section~\ref{sec:overview} explains the first five in more detail. \begin{enumerate}[leftmargin=*] \item \emph{L\"owner determinants of digamma combinations.} The approximations are determinants of L\"owner matrices, the matrices of divided differences, of a combination $F$ of $\psi(z+s)$, $\psi'(z+s)$ and $\psi''(z+s)$ at integer nodes. The rows and columns share a long block of nodes, which carries the variable. A further border of row nodes and column nodes alternates beyond it. The degree in the variable stays small, while the decay is that of a larger node set. This one structure gives all three planes. Only the function $F$ changes. \item \emph{Denominators from local bounds.} For each prime we prove a list of upper bounds for the valuation of the common denominator. Each is an explicit integer expression in a few integers attached to the prime. Integrated against $\log p$, their minimum gives the denominator exponent, and checking it reduces to finitely many comparisons of rational numbers. \item \emph{Local bounds that do not see the weights.} For integer weights the polar data at a prime $p$ depend only on the shifts. All our local bounds for the first plane read only these data. So one table of bounds serves every point of the plane. A prime dividing a weight can only lower the true valuation, and a weight zero removes a pole without changing the bounds. \item \emph{A decay certificate uniform in the weights.} The energy inequalities behind the decay estimate contain no weight. The weights enter only through a constant factor in the external field, which costs $e^{O(r)}$. A weight zero only removes atoms. One trial measure on one contour therefore proves the decay for the whole first plane, and another one for the whole second plane. \item \emph{Nonvanishing in characteristic zero.} A kernel vector of the matrix at a rational point gives a solution of a Hermite interpolation problem. A discrete Wronskian bounds the corank, because one pole of the increment of $F$ is isolated in its class modulo $\Z$. The bound depends only on the number of poles, counted with order. \item \emph{Level returns and a gauge for the special points.} At a prime $p$, the partial sums of $F$ step by the residues at its poles. For special weights this level returns to an earlier value, and whole blocks of entries become $p$-integral. A row operation inside each residue class modulo $p$, by a $p$-integral multiple of a common row, gives some further rows the level of the return. This turns the rank of the reduction modulo $p$ into an assignment bound, and it gives the points of Theorem~\ref{thm:points1}(iii)--(v). \item \emph{Elimination of complete periods for a double pole.} For the second plane the Schur complement of the L\"owner matrix over a union of complete periods modulo $p$ is again of L\"owner type, with generators that are constant on half-periods. This follows from the displacement identity of L\"owner matrices and from $\sum_{v\in\F_p^\times}v^{-1}=0$. A uniform pivot on complete double periods extends it to primes smaller than the border. The required nondegeneracy modulo $p$ is the Wronskian argument of item~5 in characteristic $p$. \item \emph{Jets reduced to simple poles.} At a pole of order $w$ the Gram representation of the L\"owner matrix carries derivatives up to order $w-1$. The atom term is a nondegenerate symmetric bilinear form in the jets, and it splits into $w$ rank-one terms. These splittings, and Cauchy's estimate on small circles, bound every jet determinant by the Cauchy determinant of a fixed function with simple poles. The decay certificate of that function then applies. \item \emph{Assignment bounds for a pole of any order.} For a single shift with a pole of order $h-1$, four explicit systems of potentials in Lemma~\ref{lem:trop} bound the denominators. They use no rank argument. Their minimum gives a denominator bound for the third plane. The elimination of the complete periods below the border, as in item~7, carries over to a pole of any order. It gives a fifth bound. When the border meets no multiple of $p$, this bound lies below the first one by two units for each border pair. We use all five bounds on the third plane. \end{enumerate} The constants of the proofs are listed in Table~\ref{tab:data}. \subsection{Organization} Section~\ref{sec:overview} explains the construction and the main ideas without technicalities, and states the general deduction. Section~\ref{sec:construction} gives the construction. Sections~\ref{sec:content}, \ref{sec:nonvanishing} and~\ref{sec:decay} prove the three estimates for differences of two digamma functions. They also prove nonvanishing and decay for general combinations. Section~\ref{sec:plane1} proves Theorem~\ref{thm:plane1} and the special points of Theorem~\ref{thm:points1}, and Section~\ref{sec:chi3-quarter} treats the point $\Lt$ itself. Section~\ref{sec:plane2} proves Theorem~\ref{thm:plane2}, and Section~\ref{sec:plane3} proves Theorem~\ref{thm:plane3}. Section~\ref{sec:proof} completes the proofs and gives the data. Section~\ref{sec:limits} discusses the limits of the method and the nearest constants that it does not reach. Section~\ref{sec:lean} describes the formalization. The disclosure of AI use follows it. \subsection{Notation and conventions} We write $v_p$ for the $p$-adic valuation. For $P\in\Q[x]$, $\den P$ is the least positive integer whose product with $P$ lies in $\Z[x]$. We write $[0,r)=\{0,1,\dots,r-1\}$. For finite sets $X_0$, $Y_0$ of integers and a matrix $A$ indexed by integers, $A[X_0,Y_0]$ is the submatrix with rows $X_0$ and columns $Y_0$. We write $\|y\|$ for the distance from a real number $y$ to $\Z$, and $\|m\|_p=p\|m/p\|$. %% ===================================================================== \section{Overview of the method}\label{sec:overview} \subsection{L\"owner matrices of digamma differences} Let $F$ be a function analytic near a finite set of distinct points. The \emph{L\"owner matrix} \cite{Loewner34} of $F$ is the matrix of divided differences indexed by the points: \[ \Lw_F[t,t']=\frac{F(t')-F(t)}{t'-t}\ (t\ne t'),\qquad \Lw_F[t,t]=F'(t). \] It is a Cauchy-like structured matrix. Hermite's formula gives $\Lw_F=\frac1{2\pi i}\oint F(z)\,\upsilon(z)\upsilon(z)^{\top}dz$ with $\upsilon_t(z)=1/(z-t)$. A representation of $F$ as a Stieltjes or contour integral therefore exhibits $\Lw_F$ as a Gram matrix of Cauchy kernels. For a rational shift $0q$ be a prime. The indices $k\ge1$ at which the entries of $M$ acquire a denominator divisible by $p$ are the multiples of $p$ and the solutions of $qk\equiv q_s\pmod p$. Let $k_p$ be the least of the latter. Then \[ qk_p=jp+q_s,\qquad 0p_0$ be a prime with $q(r+2m)\max(1+2\alpha,1/\kappa)+o(1)$. All cases are checked in the formal proof. \subsubsection{Primes near $r$: the bounds (LP1) and (LP2)}\label{sssec:lp} For $p$ close to $r$ the reduction of the rescaled matrix has a large block of rank one per residue class. Its corank can be computed. We give the proof of (LP2). The bound (LP1) applies to the primes with $p\le rp$. Even offsets are rows and odd offsets are columns, and together they cover $1\le i\le p-r$. So row and column operations remove that multiple and leave $\begin{psmallmatrix}\tilde x\Id&\mathcal J_Y\\ \mathcal J_X&0\end{psmallmatrix}$. Here it is essential that $x$ is an indeterminate, so that $\tilde x$ is a unit. A Schur complement on $\tilde x\Id$ then shows that the rank is $k+\rank(\mathcal J_X\mathcal J_Y)$. In $\mathcal J_X\mathcal J_Y$ the block between the high rows and the high columns is zero, since their offsets $t-p$ have opposite parities. So $\rank(\mathcal J_X\mathcal J_Y)\le p-r$. In total the rank of the reduction is at most $(r-k)+k+(p-r)=p$, and the corank is at least $r+m-p$. Lemma~\ref{lem:rank} gives $v(\det)\ge(r+m-p)-(2(r-k)+2m+\bar N-p)=-(2(r-k)+3m)$. Deleting a common node changes the potentials by at most one. \end{proof} Under the hypothesis $\bar N\le p+k$ the third pole does not occur, and the proof is shorter. The border nodes beyond $p+k$ only contribute rows and columns that vanish modulo $p$. This observation allows the whole range $\bar N\le2p$, which is automatic on the interval $rm$. For $n=\lfloor m/p\rfloor\ge1$ it applies to the small primes $p\le m$. For these primes only (T1) and (T1$'$) are available among the assignment bounds. Read modulo $p$, the window range is then the arc of residues that receive an odd number of border nodes. The saving over (T1$'$) is the amount, if any, by which the length of the longest cut-free run of this arc exceeds $\|m\|_p$. For other $n$ the window range is empty. The proof uses $F$ only through its poles. For a function with several digamma terms the same statement holds when the cuts are the residues of all the poles and of $r$. \begin{proof} Let $\widetilde\Phi=p^2\Phi$, with independent diagonal variables on the common nodes. Every member is a coefficient of this determinant, so a bound for it bounds every member. Let $ip_0$ be a prime with $q(r+2m)\theta_{\rm end}r$, (C4) gives $\bar N\le p$ and $k_p\ge \bar N$, and (T5) gives $e_p=0$. Fix a unit class modulo $q^*$. By the prime number theorem in arithmetic progressions, \[ \sum_{\substack{\theta_{i-1}r\deg\mathsf w$, this gives $\mathsf w=0$. \end{proof} \subsection{The corank} Put \[ \mathsf D(z)=\prod_{s\in S}(z+s)^{\nu_s},\qquad \mathsf A=\mathsf D\,\delta , \] where $\nu_s$ is the order of $s$. Then $\mathsf D$ has degree $d$, and $\mathsf A$ is a polynomial of degree at most $d-2$, since $\delta(z)=O(|z|^{-2})$ by the balance $\sum_sw_s=0$. Differentiating $\mathsf D\delta=\mathsf A$ gives $\mathsf D'\delta+\mathsf D\delta'=\mathsf A'$. For integers $k\ge0$ we have $\mathsf D(k)\ne0$, $f_{k+1}-f_k=\delta(k)$ and $\vartheta_{k+1}-\vartheta_k=\delta'(k)$. \begin{lemma}[cross product]\label{lem:cross} For two pairs $(\mathsf u_1,\mathsf w_1)$ and $(\mathsf u_2,\mathsf w_2)$ of $\mathcal V$ one has $\mathsf w_1\mathsf u_2=\mathsf w_2\mathsf u_1$. \end{lemma} \begin{proof} The polynomial $\Pi=\mathsf w_1\mathsf u_2-\mathsf w_2\mathsf u_1$ has degree at most $2n-2$. The value conditions give $\Pi(t)=0$ for $t\in Z$. At $t\in C$ the value and jet conditions give \[ \Pi'(t)=(f_t\mathsf u_1'+\vartheta_t\mathsf u_1)\mathsf u_2+f_t\mathsf u_1\mathsf u_2' -(f_t\mathsf u_2'+\vartheta_t\mathsf u_2)\mathsf u_1-f_t\mathsf u_2\mathsf u_1'=0, \] with all polynomials evaluated at $t$. So $\Pi$ has at least $|Z|+|C|=2n$ zeros, counted with multiplicity, and $\Pi=0$. \end{proof} The next lemma is the heart of the argument. It is used twice, with all the conditions and with the value conditions only. \begin{lemma}[no short pairs]\label{lem:noshort} Let $N_0\ge1$ and $0\le c_0\le N_0$ be integers, and let $\mathsf u$, $\mathsf w$ be polynomials of degree at most $\ell$ that satisfy the value conditions on $[0,N_0)$ and the jet conditions on $[0,c_0)$. If \[ 2\ell+d-2<(N_0-1)+\max(c_0-1,0), \] then $\mathsf u=0$. \end{lemma} \begin{proof} Suppose that $\mathsf u\ne0$. Put $\mathsf W(z)=\mathsf w(z+1)\mathsf u(z)-\mathsf w(z)\mathsf u(z+1)$ and \[ \mathsf E(z)=\mathsf D(z)\,\mathsf W(z)-\mathsf A(z)\,\mathsf u(z)\,\mathsf u(z+1). \] By Taylor's formula $\mathsf W=\sum_{j\ge1}(\mathsf w^{(j)}\mathsf u-\mathsf w\mathsf u^{(j)})/j!$. The term $j=1$ has degree at most $2\ell-2$, because the coefficients of $z^{2\ell-1}$ in $\mathsf w'\mathsf u$ and in $\mathsf w\mathsf u'$ are both $\ell$ times the product of the leading coefficients. The terms with $j\ge2$ have degree at most $2\ell-2$ too. Hence $\deg\mathsf E\le2\ell+d-2$. Now we count zeros. Let $k\ge0$ with $k+10$ let $k_\epsilon(z,w)=-\frac12\log(|z-w|^2+\epsilon^2)$. Then $k_\epsilon(z,w)\le-\log|z-w|$. \begin{lemma}[conditional positive definiteness]\label{lem:cpd} Let $z_i\in\C$ be finitely many points and $y_i$ real numbers with $\sum_iy_i=0$. Then $\sum_{i,j}y_iy_jk_\epsilon(z_i,z_j)\ge0$. \end{lemma} \begin{proof} We use Frullani's integral $\log T=\int_0^\infty(e^{-u}-e^{-Tu})\,du/u$ for $T>0$. Together with $\sum_iy_i=0$ it gives $\sum_{i,j}y_iy_j(-\log(|z_i-z_j|^2+\epsilon^2))=\int_0^\infty\frac{e^{-u\epsilon^2}}u\sum_{i,j}y_iy_je^{-u|z_i-z_j|^2}du$. Gaussian kernels are positive definite. Indeed, expanding $e^{2u\langle z,z'\rangle}$ in powers and using the binomial theorem writes the inner sum as a convergent sum of squares. \end{proof} A \emph{trial measure} is a probability measure $\sigma$, symmetric under conjugation, carried by finitely many segments $\mathcal S_i\subset\Gamma\cup[-r_0,0]\cup[0,1]$ with rational endpoints. On each segment it has a density $\sigma_i(y)\,dy$ in the parameter $y\in[0,1]$ of the segment. Here $\sigma_i$ is a polynomial with rational coefficients, and $0\le\sigma_i\le\bar\sigma_i$ for a rational $\bar\sigma_i$. We write $\Theta_\sigma=\sum_i\bar\sigma_i/|\mathcal S_i|$ for the sum of the bounds of its densities with respect to arclength. We compare the free points and the deleted nodes with $|\nu_c|\sigma$ through Lemma~\ref{lem:cpd}. This bounds the free--free energy from below by a sum of one-point terms and the constant $|\nu_c|^2\cI(\sigma)$. Riemann-sum comparisons replace the node sums by the potential $U^{\nu_c}$ and the node self-energy by $\cI(\nu_c)$, up to $O(r\log r)$. With the external field one arrives at the function \[ \Upsilon(z)=|\nu_c|\,U^\sigma(z)-U^{\nu_c}(z)+(1-\varepsilon_0)\pi|\Im z| \] for a small fixed $\varepsilon_0>0$. Here $U^\sigma(z)=\int\log\frac1{|z-w|}\,d\sigma(w)$ is the logarithmic potential of $\sigma$, and $U^{\nu_c}$ is that of $\nu_c$. A \emph{certificate} consists of $\sigma$ and rational numbers $\Upsilon_\Gamma$, $\Upsilon_0$, $J$, $\cI_\sigma$, $\cI_\nu$ with the following properties. \begin{enumerate}[label=(\alph*),leftmargin=*] \item $\Upsilon\ge\Upsilon_\Gamma$ on $\Gamma$. \item $\Upsilon\ge\Upsilon_0$ on $[0,1]$, the scaled position of the deleted common nodes. \item $J\le\int_{-r_0}^0\min(\Upsilon(x)-\Upsilon_\Gamma,0)\,dx$. \item $\cI(\sigma)\le\cI_\sigma$ and $\cI(\nu_c)\ge\cI_\nu$. \end{enumerate} \begin{theorem}[decay on a neighbourhood]\label{thm:decay} Let $00$, $\delta_\epsilon\ge\pi\epsilon\,\Theta_\sigma$, $\delta_{\rm m}>0$ and $\delta_J>0$ be rational, and put \[ \delta=3|\nu_c|^2\delta_\epsilon+4|\nu_c|^2\delta_{\rm m}+2|\mathcal T|\,\delta_J . \] Let $D$ and $\rho$ be rational numbers with \[ D\le2(|\nu_c|\Upsilon_\Gamma+|\mathcal T|J)-|\nu_c|^2\cI_\sigma+\cI_\nu-\delta,\qquad \rho\ge2(\Upsilon_\Gamma-\Upsilon_0). \] Then for every depth $K$, every member of the alternating geometry of parameter $c$ and every real $h$ with $|h|\le e^{-\rho r}$, \[ |P_{r,I}(\xi+h)|\le\exp\bigl(-(D-o(1))r^2\bigr). \] \end{theorem} The loss $\delta$ comes from replacing $\sigma$ by finitely many atoms in the comparison with the free points. The regularisation radius $\epsilon$ of Lemma~\ref{lem:cpd} and the mesh of the atoms lower $\Upsilon_\Gamma$ by $|\nu_c|(\delta_\epsilon+\delta_{\rm m})$. They raise $\cI_\sigma$ by $\delta_\epsilon+2\delta_{\rm m}$. Darboux sums for the integral in (c) lower $J$ by $\delta_J$. Inserted in the expression for $D$, these changes give $\delta$. The tolerances $\delta_{\rm m}$ and $\delta_J$ may be any positive rational numbers, since the atoms can be taken as fine as needed. In the certificates of this paper $\epsilon=10^{-12}$ and $\delta<10^{-9}$. The node energy $\cI(\nu_c)$ is a finite rational combination of the values $\mathcal H(y_k)$, where $\mathcal H(y)=\frac12y^2\log|y|-\frac34y^2$ and the $y_k$ are the rational differences of the endpoints of the two pieces of $\nu_c$. The lower bound $\cI_\nu$ is checked with rational enclosures of the logarithms. These come from the series of $\operatorname{artanh}$ and a reduction to $[1,2)$. The energy $\cI(\sigma)$ is a finite sum of pair energies of polynomial densities on segments. Each is a closed form of the type polynomial times logarithm plus arctangent, and is bounded in the same way. The conditions (a)--(c) are statements about $\Upsilon$ on sets of positive measure. They are proved on a cover by cells. \subsection{The cover by cells}\label{ssec:slope-cert} Each of the conditions (a)--(c) is checked on a finite cover by cells with exact rational arithmetic. On a cell, the potential $U^\sigma$ is split into the contribution of the segments of $\sigma$ near the cell and the rest. A near segment is bounded by a Taylor polynomial in the cell variable with an explicit remainder. For a cell that meets the line of a segment beyond its end, the same bound is used for the segment with its orientation reversed. At a corner of $\Gamma$ the two incident segments are treated together. Far segments are bounded by a multipole expansion with an explicit tail. The potential $U^{\nu_c}$ is an explicit elementary function, and it is bounded above by monotonicity on the cell. Some cells end at a breakpoint of $\sigma$, that is, at an endpoint of one of its segments or at a corner of $\Gamma$. Let $s$ be the distance from the breakpoint along the cell, with $0x^*$. \begin{lemma}[Gram formula]\label{lem:gramjet} For nonnegative integers $t,t'$, with $g=\upsilon_t\upsilon_{t'}$, \[ \Lw_F[t,t']=\frac1{2\pi i}\int_\Gamma g\,E\,dz-\sum_{s,n}\bigl(w'_sg'(x_{s,n})-w_sg(x_{s,n})-w''_sg''(x_{s,n})\bigr). \] \end{lemma} \begin{proof} Put $\tilde F(z)=\sum_s(w_s\psi(1-s-z)-w'_s\psi'(1-s-z)+w''_s\psi''(1-s-z))$. The reflection formula $\psi(1-w)=\psi(w)+\pi\cot\pi w$ and its first two derivatives give $F=\tilde F+E$. The residues of $F\upsilon_t\upsilon_{t'}$ at $t$ and $t'$ give the L\"owner entry, also for $t=t'$. At a head, $\psi$ has residue $-1$, $\psi'$ has principal part $(z-x_{s,n})^{-2}$ and $\psi''$ has principal part $-2(z-x_{s,n})^{-3}$. This gives the atom terms. The function $\tilde F$ is holomorphic to the left of $\Gamma$ and $O(1/|z|)$ there by the balance, so its integral against $\upsilon_t\upsilon_{t'}$ vanishes. The function $E$ decays like $e^{-2\pi|\Im z|}$, again by the balance. \end{proof} The atom term at a double head is a symmetric bilinear form of rank two in the row and column functions, and it splits into two rank-one terms. For functions $g$ and $h$ put $\hat g=g'-\frac{w_s}{2w'_s}g$ and define $\hat h$ likewise. Then \begin{equation}\label{eq:rankone} -w'_s(gh)'+w_sgh=-\frac{w'_s}2\bigl[(g+\hat g)(h+\hat h)-(g-\hat g)(h-\hat h)\bigr]. \end{equation} So the measure $\Omega$ of \eqref{eq:gram} is replaced by a real measure on a parameter space with two labels at each double head, carrying the weights $-w'_s/2$ and $w'_s/2$, and one label at each simple head. Andr\'eief's identity holds on any measure space, so it writes every minor of $\Lw_F$ as an integral of products of two determinants. A repeated label gives two equal rows, so its term vanishes. The same holds at a head of order $w=3$. Write the principal part of $F$ at the head as $\sum_{j\le w}c_j(z-x)^{-j}$. The atom term is then $-\sum_{i,j\ge0}B_{ij}\,g^{(i)}(x)h^{(j)}(x)$ with $B_{ij}=c_{i+j+1}/(i!\,j!)$ and $c_j=0$ for $j>w$. The symmetric matrix $B$ of size $w$ vanishes below its antidiagonal, and its antidiagonal entries are nonzero multiples of $c_w$. So $\det B\ne0$, and over $\Q$ the form $B$ is a sum $\sum_{l=1}^w\lambda_l\,\ell_l\ell_l^{\top}$ of $w$ rank-one terms, with $\lambda_l\ne0$. This gives $w$ labels at the head, with the functions $\sum_j\ell_{l,j}g^{(j)}(x)$ and the weights $-\lambda_l$. For the frames of the third plane the head $-n$ has $c_1=0$, $c_2=\gamma_2$ and $c_3=-2\gamma_3$, so \[ B=\begin{pmatrix}0&\gamma_2&-\gamma_3\\ \gamma_2&-2\gamma_3&0\\ -\gamma_3&0&0\end{pmatrix},\qquad \det B=2\gamma_3^3\ne0 . \] For $w=2$ this is \eqref{eq:rankone}. The determinants now contain derivatives of Cauchy kernels at the heads of order at least $2$, which we call \emph{multiple heads}. We compare them with a frame that has only simple poles. A \emph{reference frame} for $F$ is a frame $\mathcal R$ with simple poles on a set of shifts $S_{\mathcal R}$ with the following properties. It contains every shift of $F$ of positive order. For each shift $s$ of order $w\ge2$ it contains $w-1$ further shifts, the \emph{helpers} of $s$. Their heads lie on the same side of $x^*$ as those of $s$, at distances $\delta$ from them with $\frac14\le\delta\le\frac12$, and at mutual distances at least $\frac14$. It may contain further shifts. The poles of $S_{\mathcal R}$ are separated from $x^*$. For every frame $F^{(2)}_{\gamma_2,\gamma_1}$ we take the shifts $1$, $\frac34$, $\frac12$, with heads $-n$, $-n+\frac14$, $\frac12-n$ for $n\ge1$, and the helper $\frac34$ of the shift $1$. This also serves $\gamma_2=0$, where all orders are at most $1$, and $\gamma_1=0$, where the shift $\frac12$ has order $0$. For the frames $F^{(3)}_{\gamma_2,\gamma_3}$ of the third plane we take the same shifts $1$, $\frac34$, $\frac12$. Now $\frac34$ and $\frac12$ are both helpers of the shift $1$, at distances $\frac14$ and $\frac12$. In the separate proof for $\zeta(3)$ in \S\ref{ssec:zeta3} we take for $F^{(2)}_{1,0}$ the shifts $1$, $\frac12$ of the function $F_{1/2}$, with the helper $\frac12$. Each label of Andr\'eief's integral now has a \emph{reference point}. For a point of $\Gamma$ it is the point itself. For a simple head it is the head, which is a head of $\mathcal R$. For the $w$ labels at a multiple head $x_{s,n}$ they are $x_{s,n}$ and the heads $x_{s^*,n}$ of the helpers $s^*$ of $s$. Distinct labels have distinct reference points. In our frames the only shift of order at least $2$ is $1$, so the multiple heads are the points $-n$ with $n\ge1$. \begin{lemma}[jets]\label{lem:jet} Let $k=O(r)$. Let $t_1,\dots,t_k$ be distinct integers in $[0,\bar N)$, and let $z_1,\dots,z_k$ be an injective list of labels. Each label is a point of $\Gamma$ or a head of $F$ in $[x^*,0]$, and each multiple head may occur with any of its labels. Let $\phi_l$ be the function attached to $z_l$, that is $\upsilon_t(z_l)$ for a point of $\Gamma$ or a simple head, and the combination $\sum_j\ell_{l,j}\upsilon_t^{(j)}$ at the head for a label of a multiple head. Let $x^0_l$ be the reference point of $z_l$. Then \[ \bigl|\det[\phi_l(t_i)]\bigr|\le e^{O(r\log r)}\,\bigl|\det[(x^0_l-t_i)^{-1}]\bigr| , \] where the implied constant depends only on $\Gamma$, on $\mathcal R$ and on the weights of $F$. \end{lemma} \begin{proof} A row at a multiple head is $\sum_j\ell_{l,j}\upsilon_t^{(j)}$ at the head, with fixed rational coefficients $\ell_{l,j}$ and $j\le2$. Give each label of a multiple head its own variable $y_l$, and keep the other labels fixed at their points. By multilinearity the determinant is $\prod_l\bigl(\sum_j\ell_{l,j}\partial_{y_l}^j\bigr)$, over these labels, applied to the Cauchy determinant $\det[(y_l-t_i)^{-1}]$ and evaluated at the heads. The multiple heads are at distance at least $1$ from the nodes. So Cauchy's estimate on circles of radius $\frac18$, one variable at a time, bounds the determinant by $\prod_l\sum_j|\ell_{l,j}|\,j!\,8^j=e^{O(r)}$ times the supremum of the Cauchy determinant over the polydisc. On the polydisc a label of a multiple head is within $e_l$ of its reference point, where $e_l=\frac18$ for the label at $x_{s,n}$ and $e_l=\delta+\frac18$ for the label at a helper head at distance $\delta$. So $e_l\le\frac58$. All other labels have $e_l=0$. Compare each factor of Cauchy's product formula \[ \bigl|\det[(y_l-t_i)^{-1}]\bigr|=\frac{\prod_{l0$. Then $|y_l-y_{l'}|\le d_{ll'}(1+e_l/d_{ll'})(1+e_{l'}/d_{ll'})$. For a node $t_i$ we have $|x^0_l-t_i|\le|y_l-t_i|(1+e_l/|y_l-t_i|)$, and $|y_l-t_i|\ge t_i+\frac78$. Hence the logarithm of the ratio is at most the sum of $e_l/d_{ll'}$ over pairs, plus the sum of $e_l/(t_i+\frac78)$ over the moved labels $l$ and the nodes. Fix a moved label $l$. The reference heads are distinct points of a lattice $\frac1{q_{\mathcal R}}\Z$ in an interval of length $O(r)$. So the sum of $1/d_{ll'}$ over head labels $l'$ is $O(\log r)$, and so is the sum over the nodes. Now fix a label at a point $z$ of $\Gamma$. The heads lie at distance at least a constant $\delta_0>0$ from $\Gamma$, by the separation of the poles from $x^*$. So the sum of $e_l/|x^0_l-z|$ over the moved labels $l$ is again $O(\log r)$. This is the case of mixed configurations, of heads with free points on $\Gamma$. A bound by a constant factor for each such pair would give $e^{O(r^2)}$, so the logarithmic sum matters here. In total the exponent is $O(k\log r)=O(r\log r)$. \end{proof} \begin{theorem}[decay for frames]\label{thm:decayframe} Let $F$ be a frame and $\mathcal R$ a reference frame for it. If all orders of $F$ are at most $1$, $\mathcal R$ may be $F$ itself. Suppose that $\mathcal R$ has a certificate for $\Gamma$ and $\nu_c$ as in Theorem~\ref{thm:decay}, with $|\mathcal T|$ the number of shifts of $\mathcal R$. Then the conclusion of Theorem~\ref{thm:decay} holds for every member of the alternating geometry of parameter $c$ for $F$, with the same $D$ and $\rho$. \end{theorem} \begin{proof} By Lemma~\ref{lem:gramjet}, the rank-one splittings and Andr\'eief's identity, each minor is an integral over the labels of products of two determinants, one for the rows and one for the columns. Lemma~\ref{lem:jet} bounds each of them by the Cauchy determinant at the reference points, at a cost $e^{O(r\log r)}=e^{o(r^2)}$. The label weights are bounded and $|E|\le C_Ee^{-2\pi|\Im z|}$. Distinct labels have distinct reference points, all of them heads of $\mathcal R$ or points of $\Gamma$. So the integral is at most $e^{o(r^2)}$ times the integral of the energy argument above for $\mathcal R$, in which every head of $\mathcal R$ carries an atom of bounded mass. Those are the only inputs of that argument. \end{proof} The certificate of the second plane is that of the reference frame $(1,\frac34,\frac12)$, with $|\mathcal T|=3$. It serves every frame $F^{(2)}_{\gamma_2,\gamma_1}$, including $\gamma_2=0$ and $\gamma_1=0$. The certificate of the third plane is another certificate for the same reference frame, for its own geometry parameter. The certificate of the separate proof for $\zeta(3)$ is that of the function $F_{1/2}$, with $|\mathcal T|=2$. It is the certificate of the row for $\pi^2$. %% ===================================================================== \section{The first plane}\label{sec:plane1} In this section we prove Theorem~\ref{thm:plane1} and parts (iii)--(v) of Theorem~\ref{thm:points1}. We use the frames $F^\pm_{\vec w}$ of \S\ref{ssec:frames} with the alternating geometry of parameter $c=\frac16$, so $\alpha=\frac1{12}$ and $m=\lceil r/12-\frac12\rceil$. The only exception is the pair $\Lt\pm\frac{\pi^2}{18}$, for which we take $c=\frac14$. The point of the section is that the denominator exponent and the decay certificate do not depend on the weights $\vec w$, and that a weight zero is allowed. Nonvanishing is Theorem~\ref{thm:nvframe} with $d\le3$, so one deleted node suffices. \subsection{Polar data} Fix a frame $F^\pm_{\vec w}$ with integer weights. Let $\varrho=2$ on $S^+$ and $\varrho=1$ on $S^-$. For $j\ge1$ the increment is \[ \delta(j-1)=\frac{3w_\pm}{3j-\varrho}+\frac{2w_{1/2}}{2j-1}+\frac{w_1}j , \] and $-\delta'(j-1)=\frac{9w_\pm}{(3j-\varrho)^2}+\frac{4w_{1/2}}{(2j-1)^2}+\frac{w_1}{j^2}$. Let $p\ge5$ be a prime with $3\bar Nr$\\ (D6) & $(\bar N-k_p)_+$ & $p\ge\bar N$, $k_p\ge r$\\ (W) & $2r+2m-(t_2-t_1)$ for each window of Theorem~\ref{thm:window} & as there\\ \bottomrule \end{tabular}} \end{center} In (W) the cuts are the three lattices of singular indices and the residue of $r$. \end{proposition} \begin{proof} (T1) and (T1$'$) are proved as in \S\ref{sssec:trop}. For the others we give potentials, with one function $u$ used on rows and on columns. For (D2), $u=0$ on the common nodes $t0$ and every even $r\ge R(\epsilon)$, every member satisfies $\log\den P_{r,I}\le(\Lambda^*+\epsilon)r^2$. The threshold $R(\epsilon)$ does not depend on the frame. \end{theorem} \begin{proof} Take the product over primes of the bounds of Proposition~\ref{prop:planelocal}. For the primes with $p^2\le3\bar N$, and for $2$ and $3$, every entry has valuation at least $-2\log(3\bar N)/\log p$. These primes contribute $O(r^{3/2})$. Each bound is built from affine functions of $r$, $m$, $p$, $k_A$, $k_B$ and $p\lfloor r/p\rfloor$ by $\max$, $\min$ and truncation, with bounded coefficients. So it differs from $r$ times its homogenisation by a bounded amount. Each integer condition differs from its homogenised form only for $p$ in an interval of bounded length, where (T1) costs $O(r\log r)$. The prime number theorem in progressions modulo $6$ then gives the claim, as in Theorem~\ref{thm:content}. For the primes $p\le t_0r$ we use (T1), which gives $(2+2\alpha)t_0r^2$. On the strips $\theta\le\alpha$ the content is an instance of Theorem~\ref{thm:periodic} for $\alpha=\frac1{12}$, with $K_0=2+2\alpha$. The bounds acting there are closed under translation: (T1) and (T1$'$) grow by $12\nu K_0p$ under the translation of order $\nu$, and it takes the window with $n=\lfloor m/p\rfloor$ to the window with $n+\nu$. So the table is checked on one period of $1/\theta$, and its translates give the pieces of the other strips. \end{proof} On $\theta\le\alpha$ the profile has a closed form. Only (T1), (T1$'$) and the windows act there. Put $s=1/\theta$. Then $\varphi(\theta)=2+2\alpha-\theta\,\Gamma_\kappa(s)$. Here $\Gamma_\kappa(s)$ is the larger of $\|\alpha s\|$ and the longest gap of the lattice $\{0,\kappa_A,\kappa_B\}+\Z$ inside a window. The window is determined by the fractional parts of $s$ and $\alpha s$. This function has period $12$ in $s$, as Theorem~\ref{thm:periodic} predicts. The table has $82$ cells per class on one period. \subsection{Decay} Let $\Gamma$ be the polygon of the formal proof for the alternating geometry of parameter $\frac16$. It crosses the real axis at a point $x^*$ with $x^*\equiv\frac14\pmod1$ for the frames on $S^+$ and $x^*\equiv\frac34\pmod1$ for those on $S^-$. In both cases $x^*$ is at distance at least $\frac14$ from the poles of the shifts. The scaled cover is the same for both. \begin{proposition}\label{prop:planecert} There is a trial measure with $25$ polynomial pieces on $\Gamma$, with a cover by cells satisfying the checks of \S\ref{ssec:slope-cert} with $\varepsilon_0=10^{-6}$ and $|\mathcal T|=3$, for which Theorem~\ref{thm:decay} gives \begin{gather*} D=\frac{99853629795114224867}{28125000000000000000}=3.55035\ldots,\\ \rho=\frac{33845462710966783}{5\cdot10^{15}}=6.76909\ldots \end{gather*} \end{proposition} The certificate contains no weight. By \S\ref{ssec:decaydouble} the decay estimate therefore holds for every frame on $S^+$ and on $S^-$ with these $D$ and $\rho$, also when one weight is zero. The weights change only the threshold in $r$. \subsection{Proof of Theorem~\ref{thm:plane1}} Let $a$ and $b$ be rationals, not both zero. By Lemma~\ref{lem:planeconst} there is a frame on $S^+$ or $S^-$ whose number $\xi$ is a nonzero rational multiple of $a\Lt+b\pi^2$. For $a=0$ it is the frame $F^+_{(0,3,-3)}$, whose shift $\frac13$ has weight zero. Theorem~\ref{thm:planecontent} gives the denominators with $\Lambda^*$, and Proposition~\ref{prop:planecert} gives the decay. Both hold for every frame on $S^\pm$, including those with a weight zero. Theorem~\ref{thm:nvframe} gives nonvanishing with one deleted node, since $d\le3$. For $a=0$ the frame has two poles per period, and the argument is that of two digamma terms. The deduction of Theorem~\ref{thm:deduction} at the even scales gives \[ \mu(\xi)\le\frac\rho{D-\Lambda^*}=20.66213\,7\ldots<\MuPlaneA . \] The strict inequality gives the bound $|a\Lt+b\pi^2-p/q|>q^{-\MuPlaneA}$ for all large $q$, and in particular the irrationality.\qed The thresholds depend on $(a,b)$ through the height of the weights. No uniformity in $(a,b)$ is claimed. \subsection{Level returns}\label{ssec:levels} At special weights the partial sums of the increment return to an earlier level, and then more entries are $p$-integral than Proposition~\ref{prop:planelocal} uses. Fix a frame on $S^\pm$, a class of primes modulo $6$, and a prime $p\ge7$ with $3\bar N3$ be a prime. The increment $g(k)=\frac1{3k-2}-\frac1{2k-1}+\frac1{6k}$ has a pole modulo $p$ exactly when $p$ divides $k$, $3k-2$ or $2k-1$. Let $k_A$ and $k_B$ be the least $k\ge1$ with $p\mid3k-2$ and with $p\mid2k-1$. Then \[ k_B=\frac{p+1}2,\qquad k_A=\begin{cases}\frac{p+2}3,&p\equiv1\pmod6,\\[3pt]\frac{2p+2}3,&p\equiv5\pmod6.\end{cases} \] The residues of $p\,g(k)$ at $k_A$ and $k_B$ are $1$ and $-1$ if $p\equiv1\pmod6$, and $\frac12$ and $-1$ if $p\equiv5\pmod6$. In the first case $k_A6$ with $6\bar N2\bar N+2p$. The increment is \[ \delta(k-1)=\frac{\gamma_1}k-\frac{2\gamma_1}{2k-1}-\frac{\gamma_2}{k^2}\qquad(k\ge1). \] Its poles below $\bar N$ are the double poles $k=jp$ and, if $\gamma_1\ne0$, the simple poles $k=jp+\frac{p+1}2$. Write an integer $t\ge0$ as $t=jp+\omega$ with $0\le\omega\bar N$. The numerator $f_{t'}-f_t$ has valuation at least $-2$, at least $-1$ if $t$ and $t'$ lie in one period, and at least $0$ if they lie in one cell. The diagonal is $-e_t$. The weights $\gamma_1$ and $\gamma_2$ are integers, so a prime dividing them only raises valuations. \end{proof} The primes that divide $\gamma_2$ need one remark. If $p\mid\gamma_2$, the double poles contribute terms $\gamma_2/k^2$ of valuation at least $-1$ to $f_t$, and terms $2\gamma_2/k^3$ of valuation at least $-2$ to $e_t$. So every entry of $M$ has valuation at least $-2$, and Lemma~\ref{lem:trop} with all potentials $1$ gives $e_p(r,I)\le2(r+m)$. The bounds of Proposition~\ref{prop:rows2} below use Lemma~\ref{lem:weights2} only and hold for all $p$. Those of Theorem~\ref{thm:rows2} assume $p\nmid\gamma_2$. Each of them is at least $2(r+m)$, as we check after its proof. So for $p\mid\gamma_2$ the trivial bound $2(r+m)$ takes their place, and every bound used below holds for all frames with $\gamma_2\ne0$. \begin{proposition}[assignment bounds]\label{prop:rows2} Assume $2m\le r$. For every member, \begin{center} {\renewcommand{\arraystretch}{1.2} \begin{tabular}{@{}lll@{}} \toprule & bound & condition\\ \midrule (P1) & $3(r+m)-\|m\|_p$ & none\\ (P4) & $6(r+m-p)+\varepsilon p$ & $p\le r$\\ (P5) & $\varepsilon r+3(\bar N-p)$ & $r\le p\le\bar N$\\ (P6) & $\varepsilon(2r+2m-p+1)$ & $\bar N\le p\le2r$\\ (P7) & $\varepsilon(\bar N-\frac{p-1}2)_+$ & $p\ge2r$\\ \bottomrule \end{tabular}} \end{center} \end{proposition} \begin{proof} All follow from Lemmas~\ref{lem:trop} and~\ref{lem:weights2}. For (P1) take the potentials of (T1$'$), raised by $\frac12$ on every row and every column. Congruent pairs get $3$ and the others at least $2$. For the other bounds take one function $\varphi$ on rows and columns, with $\varphi=0$ on the cell $(0,0)$, $\varphi=\varepsilon$ on $(0,1)$ and $\varphi=3$ on the periods $j\ge1$. Pairs inside the period $0$ have weight at most $\varphi\oplus\varphi$, and any pair with a node in a later period has weight at most $3$. Counting the common nodes twice and the border nodes once gives the four bounds. The hypothesis $2m\le r$ enters twice. For (P5) it gives $p\le\bar N\le2r$, so $(p-1)/22\bar N'$. Let $H=M+p^{-3}T\Id$ on $[0,\bar N')$, so that $H$ restricts to the matrices of the members. Extend the border alternately, so $\hat X=\{t\ge r:t-r\text{ even}\}$ and $\hat Y=\{t\ge r:t-r\text{ odd}\}$. The \emph{pivot} consists of the complete periods below $\kappa p$ and of $2k'$ further complete periods: \begin{align*} \mathcal Q_{\rm r}&=[0,\kappa p)\cup(\hat X\cap[\kappa p,(\kappa+2k')p)),\\ \mathcal Q_{\rm c}&=[0,\kappa p)\cup(\hat Y\cap[\kappa p,(\kappa+2k')p)). \end{align*} We write $H_{\mathcal Q}=H[\mathcal Q_{\rm r},\mathcal Q_{\rm c}]$. \begin{lemma}[displacement and bordering; cf.\ \cite{Fiedler84}]\label{lem:displace} For all nodes $t,t'$ one has $(t-t')H(t,t')=f_t-f_{t'}$, whatever the diagonal. Suppose that $H_{\mathcal Q}$ is invertible. Let $\mathrm{Sw}$ be the Schur complement of $H_{\mathcal Q}$ in the matrix bordered by unit vectors, as in Lemma~\ref{lem:bordered}. Then \begin{enumerate}[label=(\alph*),leftmargin=*] \item every minor of $H$ that contains $\mathcal Q_{\rm r}\setminus A_0$ and $\mathcal Q_{\rm c}\setminus A_0'$ equals $\pm\det H_{\mathcal Q}$ times a minor of $\mathrm{Sw}$ with rows $B_0\cup A_0'$ and columns $B_0'\cup A_0$, where $B_0$ and $B_0'$ are the remaining rows and columns. \item there are vectors $g_z\in\Q(T)^2$, one for each label $z$, such that $(z'-z)\,\mathrm{Sw}(z,z')=g_z\wedge g_{z'}$ for all labels $z\ne z'$. \end{enumerate} \end{lemma} \begin{proof} The first identity is the definition of the L\"owner matrix. With the diagonal matrix $\operatorname{diag}(t)$ of the nodes it says that $\operatorname{diag}(t)H-H\operatorname{diag}(t)=f\mathbf 1^\top-\mathbf 1f^\top$ has rank two. The bordered matrix satisfies the same equation with generators $(1,f_t)$ on actual nodes and $0$ on the border, since $(t-t)\cdot1=0$. A Schur complement of a block preserves this displacement structure, with the generators given below. Laplace expansion along the unit border gives (a). \end{proof} The generators are explicit. On a label $z$ outside the pivot $g_z=(1-H(z,\mathcal Q_{\rm c})H_{\mathcal Q}^{-1}\mathbf 1,\ f_z-H(z,\mathcal Q_{\rm c}) H_{\mathcal Q}^{-1}f)$, and on a pivot label $t$ it is $((H_{\mathcal Q}^{-1}\mathbf 1)_t, (H_{\mathcal Q}^{-1}f)_t)$ up to sign. We rescale them as $\tilde g_z=(p^{-3}g_{z,1},p^{-1}g_{z,2})$ on pivot labels and $\tilde g_z=(g_{z,1},p^2g_{z,2})$ on the others, and the pivot rows and columns of $\mathrm{Sw}$ by $p^{-3}$. The rescaled matrix $\widetilde{\mathrm{Sw}}$ then has entries $p^{-2}(\tilde g_z\wedge\tilde g_{z'})/(z'-z)$. The reduction of the pivot is governed by matrices on period indices. For $\iota\in\{0,1\}$ put $s_\iota(j)=p^2S(j,\iota)$, $s'_\iota(j)=p^3G(j,\iota)$, $\Gamma_\iota(j)=(1,s_\iota(j))$, and \[ B^\iota(l,j)=\frac{s_\iota(j)-s_\iota(l)}{j-l}\ (l\ne j),\qquad B^\iota(j,j)=T-s'_\iota(j). \] Modulo $p$ one has $s_\iota(j)\equiv-\gamma_2h_j$, where $h_j=\sum_{i\le j}i^{-2}$. So the off-diagonal entries of $B^\iota$ vanish modulo $p$ when $p\mid\gamma_2$. This is why $p\nmid\gamma_2$ is assumed from now on. Let $J_X=\{\kappa,\kappa+2,\dots,\kappa+2k'-2\}$, $J_Y=J_X+1$, and $B_X^\iota=B^\iota[[0,\kappa)\cup J_X,[0,\kappa)\cup J_Y]$. Write $\bar B_X$ for its reduction modulo $p$, which does not depend on $\iota$. If $t\equiv t'\pmod p$ then $p^3H(t,t')\equiv B^\iota(j_t,j_{t'})\pmod{p^3}$, where $j_t$ is the period of $t$. If $t\not\equiv t'$ then $p^3H(t,t')\in p\cO$. Hence $p^3H_{\mathcal Q}$ is congruent modulo $p$ to a direct sum of copies of $\bar B_X$ and its transpose, one for each residue class. \begin{theorem}[nondegeneracy]\label{thm:nd} Let $\kappa\ge1$, $k'\ge0$, $p>\kappa+2k'$ and $p\nmid\gamma_2$. Then $\det\bar B_X\ne0$ in $\F_p(T)$. \end{theorem} \begin{proof} For $k'=0$ the determinant is monic in $T$. Let $k'\ge1$. The variable occurs only on the diagonal of the block $[0,\kappa)$, so $\det\bar B_X$ is $\pm\det\bar B[J_X,J_Y]$ times a monic polynomial. Since $\bar B[J_X,J_Y]$ is $-\gamma_2$ times the matrix $[(h_j-h_l)/(j-l)]_{l\in J_X,j\in J_Y}$ and $p\nmid\gamma_2$, it suffices that the determinant of the latter is nonzero in $\F_p$. Suppose that a vector $\mathbf y\ne0$ lies in its kernel. Let $P=\prod_{j\in J_Y}(z-j)$, $\mathsf u=\sum_jy_jP/(z-j)$ and $\mathsf w=\sum_jy_jh_jP/(z-j)$. These polynomials over $\F_p$ have degree at most $k'-1$, $\mathsf u\ne0$, and $\mathsf w(t)=h_t\mathsf u(t)$ on the $2k'$ consecutive nodes $\kappa,\dots,\kappa+2k'-1$. Since $h_{t+1}-h_t=(t+1)^{-2}$, the polynomial \[ \mathsf E(z)=(z+1)^2\bigl(\mathsf w(z+1)\mathsf u(z)-\mathsf w(z)\mathsf u(z+1)\bigr)-\mathsf u(z)\mathsf u(z+1) \] vanishes at the $2k'-1$ nodes other than the last. The top terms cancel, so $\deg\mathsf E\le2k'-2$ and $\mathsf E=0$. Divide $\mathsf u$ and $\mathsf w$ by their greatest common divisor. At $z=-1$ the identity gives $\mathsf u(-1)\mathsf u(0)=0$. If $\mathsf u(x+1)=0$ with $x\ne-1$ then $\mathsf u(x)=0$, and conversely, by coprimality. So every element of $\F_p$ is a root of $\mathsf u$, whose degree is less than $p$. This is a contradiction. \end{proof} The argument is that of Theorem~\ref{thm:nvframe}, in characteristic $p$. The double pole at $-1$ is isolated in its orbit, and the orbit is now finite, which replaces the infinitely many roots by $p$ roots. \begin{proposition}[generators]\label{prop:gen} Let $p>\kappa+2k'$ and $p\nmid\gamma_2$. Then all $\tilde g_z$ lie in $\cO^2$. There are vectors $\mathcal V(\mathbf c)\in\cO^2$ on the cells of the periods outside the pivot, of the period $\kappa-1$ and of the period $\kappa+2k'-1$, such that every label $z$ in these periods satisfies $\tilde g_z\equiv\eta_z\mathcal V(\mathbf c(z))\pmod{p^2}$ with a unit $\eta_z\equiv1\pmod p$, and $\mathcal V(j,0)\equiv\mathcal V(j,1)\pmod p$. Consequently, for two such labels $z\ne z'$, \begin{gather*} v\bigl(\widetilde{\mathrm{Sw}}(z,z')\bigr)\ge-3\ \text{if }z\equiv z'\pmod p,\\ v\bigl(\widetilde{\mathrm{Sw}}(z,z')\bigr)\ge-W_3(\mathbf c(z),\mathbf c(z'))\ \text{otherwise}, \end{gather*} and $v(\widetilde{\mathrm{Sw}}(t,t))\ge-3$ on deleted nodes. \end{proposition} \begin{proof} Put $\widetilde H=p^3H_{\mathcal Q}$. By Theorem~\ref{thm:nd} and the direct sum above, $\widetilde H$ is invertible over $\cO$. We look for the solution $\Theta$ of $\widetilde H\Theta=(\mathbf 1,p^2f)$ modulo $p^2$. Modulo $p^2$ the right side is a function of the cell, and modulo $p$ of the period. Put $\Theta(t)=\eta(t)\mathbf y^\iota(j_t)$ with $\mathbf y^\iota=(B^\iota_X)^{-1}\Gamma_\iota$ on a class oriented like $\hat X$, and the transposed matrix on the others. The congruent part of $\widetilde H(z,\cdot)\Theta$ is then $\eta\,B^\iota(j_z,\cdot)\mathbf y^\iota$, which is $\eta\,\Gamma_\iota(j_z)$ on the pivot. The non-congruent part lies in $p\cO$, and modulo $p^2$ it equals \[ p\sum_{\omega'\ne\omega_z}\frac1{\omega'-\omega_z}\bigl(\Omega_{\omega'}[1,\cdot]-\bar s(j_z)\Omega_{\omega'}[0,\cdot]\bigr), \qquad \Omega_X=\bar\Gamma^\top\bar B_X^{-1}\bar\Gamma , \] with $\Omega_Y=\Omega_X^\top$. Here $\omega_z$ is the residue of $z$, $\Omega_{\omega'}$ is $\Omega_X$ or $\Omega_Y$ according to the orientation of the class of $\omega'$, and bars denote reductions modulo $p$. We use that the pivot consists of complete periods, so every residue carries $\Omega_X$ or $\Omega_Y$ and nothing else. Since $\sum_{\omega'\ne\omega}(\omega'-\omega)^{-1}\equiv0\pmod p$, only the antisymmetric part $\Omega_X-\Omega_X^\top$ survives. It acts on $2$-vectors as a scalar $\delta_{\rm rot}$ times the rotation, so the sum is $p\,\delta_{\rm rot}\sigma(\omega_z)\bar\Gamma(j_z)$, where $\sigma(\omega)$ is the sum of $(\omega'-\omega)^{-1}$ over the residues $\omega'$ of one orientation. Taking $\eta=1-p\,\delta_{\rm rot}\sigma$ gives $\widetilde H\Theta\equiv(\mathbf 1,p^2f)\pmod{p^2}$ on the pivot, and the same computation outside the pivot gives $\tilde g_z\equiv\eta(\Gamma_\iota(j_z)-B^\iota(j_z,\cdot)\mathbf y^\iota)\pmod{p^2}$. These are the vectors $\mathcal V$. They depend on the cell modulo $p^2$, and on the period modulo $p$, since $B^\iota$ and $\Gamma_\iota$ do not depend on $\iota$ modulo $p$. For $k'=0$ we have $\delta_{\rm rot}=0$. By Lemma~\ref{lem:displace}, $\widetilde{\mathrm{Sw}}(z,z')=p^{-2}(\tilde g_z\wedge\tilde g_{z'})/(z'-z)$. The wedge lies in $p^2\cO$ for two labels in one cell and in $p\cO$ for two labels in one period. The difference $z'-z$ has valuation $1$ if $z\equiv z'$ and $0$ otherwise. The diagonal entry on a deleted node is $-p^{-3}\widetilde H^{-1}(t,t)$. \end{proof} \subsection{The local bounds from elimination} We combine Lemma~\ref{lem:displace}(a) with Proposition~\ref{prop:gen}. A potential on the labels bounds the Leibniz terms of the minor of $\widetilde{\mathrm{Sw}}$, and the factor $\det H_{\mathcal Q}$ has valuation exactly $-3|\mathcal Q_{\rm r}|$. \begin{theorem}[local bounds of the second plane]\label{thm:rows2} Let $p\ge5$ with $p^2>2\bar N+2p$ and $p\nmid\gamma_2$, and let $2m\le r$. For every member the following hold. \begin{enumerate}[label=(\roman*),leftmargin=*] \item (E$_1$) If $p>2m$, then $e_p(r,I)\le3r+\sum u+\sum u'$ for all nonnegative potentials $u$ on $X'$ and $u'$ on $Y'$ with $W_3\le u\oplus u'$ on the border. \item (E$_2$) If $p\le r$ and $\bar N\le2p$, then (E$_1$) holds with $3r$ replaced by $3r-(2p-\bar N)$. \item (R) If $p\le2m$, write $2m=kp+s$ with $0\le s\bar N$, since $t\mapsto t+kp$ maps $U_2$ onto $E_v$. Such pairs have $W\le3=1+2$. All other pairs off $E$ have $W\le2$. The sum of the potentials, with the potentials $3$ on deleted nodes and the count of the pivot as in (i), is $3(r+m)-3\|m\|_p+\sum_{t\in E}w(c_b(t))$. \end{proof} Each bound of Theorem~\ref{thm:rows2} is at least $2(r+m)$, which justifies the remark before Proposition~\ref{prop:rows2} on the primes $p\mid\gamma_2$. Indeed (E$_1$) is at least $3r\ge2r+2m$, (E$_2$) is at least $4r+2m-2p\ge2r+2m$ since $p\le r$, and (R) is at least $3r+3m-3\|m\|_p\ge3r\ge2r+2m$. For $k=1$ part (iii) needs only the pivot of (i), with the weight $3$ of congruent pairs kept. For $k\ge2$ it needs the complete double periods. The pivot contracts the window of length $2m$ to one of length $p+s$ or $s$, and the cost of each eliminated pair of periods is exactly $3$ per node. \subsection{The denominator exponent} Homogenise with $\theta=p/r$ as before. The profile $\mathcal P(\theta)$ is the minimum of the homogenised bounds of Proposition~\ref{prop:rows2} and Theorem~\ref{thm:rows2}. For (E$_1$) and (E$_2$) the potentials are constant on the at most three cells cut from the border at the multiples of $\theta/2$. On the band $\frac{2\alpha}{k+1}<\theta\le\frac{2\alpha}k$, the bound (R) is \[ 3+\tfrac32k\theta+\tau(E)\ (k\text{ even}),\qquad 3+6\alpha-\tfrac32(k+1)\theta+\tau(E)\ (k\text{ odd}), \] where $\tau(E)$ is the least over the five vectors $w$ of the sums $\sum w(\mathbf c)\,|\mathbf c\cap E|$, taken over the four cells $\mathbf c$, scaled by $r$, $|\mathbf c\cap E|$ is the length of the intersection, and $E=[1,1+2\alpha-k\theta]$, respectively $E=[1+2\alpha-k\theta,1+\theta]$. Each node of $E$ carries one potential, so there is no factor $\frac12$. Between the points $\theta=2/j$ and $\theta=2(1+2\alpha)/j$ the cell type of $E$ is fixed. So $\mathcal P$ is piecewise affine, with finitely many pieces on every interval bounded away from $0$. On $\theta<2\alpha$ the bounds (P1) and (R) are closed under translation in the sense of Theorem~\ref{thm:periodic}, for $\alpha=\frac3{41}$, with $K_0=3+3\alpha$. The translation of order $\nu$ takes the band $k$ to the band $k+6\nu$. So the table for $\Lambda$ is checked on one period of $1/\theta$, the bands $k=1,\dots,6$, and on the head $\theta\ge2\alpha$. Its translates cover the bands down to $t_1=3\cdot10^6/(1025\cdot10^6+3)$, just below $\alpha/25$. On $(0,t_1]$ we use (P1) without its last term. \begin{theorem}[denominators on the second plane]\label{thm:content2} Let $\alpha=\frac3{41}$ and $\gamma_2\ne0$, with $\gamma_1$ any integer. For every $\epsilon>0$ and every even $r\ge R(\epsilon)$, every member with $|I|\le1$ satisfies $\log\den P_{r,I}\le(\Lambda_2+\epsilon)r^2$, where \[ \Lambda_2=(3+3\alpha)t_1+\int_{t_1}^\infty\mathcal P(\theta)\,d\theta\le\frac{3476837531}{10^9}. \] The threshold does not depend on $(\gamma_2,\gamma_1)$. \end{theorem} \begin{proof} For each prime with $p^2>2\bar N+2p$ the local bounds give $e_p(r,I)\le r\,\mathcal P(p/r)+O(1)$. For $p\mid\gamma_2$ the bounds of Theorem~\ref{thm:rows2} are replaced by the trivial bound $2(r+m)$, which is at most each of them. For the bounds (E$_1$), (E$_2$) and (R) the cells of the integer window and of its homogenisation differ by a bounded number of cells of bounded length, which take the potential $2$. The hypotheses $p>\kappa+2k'$ of Proposition~\ref{prop:gen} hold for $r$ large, since $p>t_1r$ on the bands used. For the other primes every entry has valuation at least $-3\log(2\bar N)/\log p$, which costs $O(r^{3/2})$ in total. The prime number theorem then gives the integral, as in Theorem~\ref{thm:content}, and Theorem~\ref{thm:periodic} gives the pieces of the translated bands from those of the period. All pieces are affine, so the integral is a rational number. It is at most the stated one. \end{proof} For $\gamma_2=0$ we take the frame $F^{(2)}_{0,1}=\psi(z+1)-\psi(z+\frac12)$. Its matrix $M$ is $4$ times that of $F_{1/2}$ in \eqref{eq:frames}, so the coefficient of $x^i$ in a member with $|C\setminus I|$ common nodes is $4^{|C\setminus I|+m-i}$ times the corresponding coefficient for $F_{1/2}$. Its denominators are therefore at most those of $F_{1/2}$. Theorem~\ref{thm:content}, with an admissible table for $q=2$ and $c=\frac6{41}$, bounds them with the exponent $\Lambda_\pi=2.60981\,51766\ldots$, which is less than $\Lambda_2$. \subsection{Decay and proof of Theorem~\ref{thm:plane2}} By Theorem~\ref{thm:nvframe} with $d\le3$, one deleted node suffices for nonvanishing. The decay is Theorem~\ref{thm:decayframe} with the reference frame of shifts $1$, $\frac34$, $\frac12$. \begin{proposition}\label{prop:cert2} Take the reference frame of shifts $1$, $\frac34$, $\frac12$ and $c=\frac6{41}$. There is a trial measure with $49$ polynomial pieces and a cover by $508$ cells satisfying the checks of \S\ref{ssec:slope-cert} with $\varepsilon_0=10^{-6}$ and $|\mathcal T|=3$. For it Theorem~\ref{thm:decay} gives $D\ge\DPlB$ and $\rho\le\gPlB$. \end{proposition} \begin{proof}[Proof of Theorem~\ref{thm:plane2}] Let $a$ and $b$ be rationals, not both zero. Choose a rational $\lambda\ne0$ for which $\gamma_2=\lambda a$ and $\gamma_1=6\lambda b$ are integers. By \S\ref{ssec:frames} the number $\xi$ of $F^{(2)}_{\gamma_2,\gamma_1}$ is $-2\lambda(a\zeta(3)+b\pi^2)$. For $a=0$ we take $\lambda$ with $\gamma_1=1$, so the frame is $\psi(z+1)-\psi(z+\frac12)$. Its trigamma term has weight zero, and it has two simple poles per period. For $b=0$ we take $\lambda$ with $\gamma_2=1$, so the frame is $\psi'(z+1)$. Its shift $\frac12$ has weight zero, and it has one double pole per period. Both are handled exactly as the first plane handles a weight zero, through a frame on the same shifts. The geometry is the same for all $(a,b)$. For $\gamma_2\ne0$, Theorem~\ref{thm:content2} gives the denominators with $\Lambda_2$. For $\gamma_2=0$, the exponent $\Lambda_\pi$ above gives them, and $\Lambda_\pi<\Lambda_2$. Proposition~\ref{prop:cert2} with Theorem~\ref{thm:decayframe} gives the decay for every frame, also when one weight is zero, and Theorem~\ref{thm:nvframe} gives nonvanishing with one deleted node, since $d\le3$. Theorem~\ref{thm:deduction} gives \[ \mu(\xi)\le\frac\rho{D-\Lambda_2}\le\ExpPlB<\MuPlaneB .\qedhere \] \end{proof} Theorem~\ref{thm:ends} gives better bounds at the two ends, by separate proofs. \subsection{The value $\zeta(3)$}\label{ssec:zeta3} For the separate bound of Theorem~\ref{thm:ends} we take $\gamma_1=0$, $\gamma_2=1$ and $c=\frac3{10}$. The frame is $\psi'(z+1)$, with $d=2$ and $\xi=-2\zeta(3)$. Everything above holds with the half $\iota$ set to $0$. The cells are then the periods, $W_3$ takes the values $0$ and $2$, and $\varepsilon=0$ in Proposition~\ref{prop:rows2}. The bounds (P4) and (P5) become $6(r+m-p)$ and $3(\bar N-p)_+$, and (P6), (P7) vanish. In (R) the four cells become two, and the potential vectors are $(0,0,2,2)$ and $(2,2,0,0)$. With $c=\frac3{10}$, and with (P1) without its last term on $(0,\frac3{500}]$, the profile has integral \[ \Lambda_{\zeta(3)}\le\frac{2912839}{10^6}. \] The reference frame is the function $F_{1/2}$, and its certificate is that of the row for $\pi^2$ in Table~\ref{tab:data}. Theorem~\ref{thm:nvframe} with $d=2$ gives nonvanishing with one deleted node. \begin{proof}[Proof of Theorem~\ref{thm:ends} for $\zeta(3)$] The three estimates above and Theorem~\ref{thm:deduction} give $\mu(\zeta(3))\le\rho/(D-\Lambda_{\zeta(3)})\le\ExpZ<\MuZetaThree$. The bound for $\pi^2$ is proved in Section~\ref{sec:proof}, together with Theorem~\ref{thm:offplane}. \end{proof} %% ===================================================================== \section{The third plane}\label{sec:plane3} In this section we prove Theorem~\ref{thm:plane3}. We use the frames $F^{(3)}_{\gamma_2,\gamma_3}$ of \S\ref{ssec:frames}, with rationals $\gamma_2$ and $\gamma_3\ne0$, and the alternating geometry of parameter $c=\frac{13}{100}$. Thus $\alpha=\frac{13}{200}$ and $m=\lceil13r/200-\frac12\rceil$. The pole at the shift $1$ has order three. We combine four assignment bounds with a fifth bound obtained by eliminating complete periods. None of these bounds requires a nonzero coefficient to be a unit modulo $p$. \subsection{Weights} We state the bounds for a single shift with a pole of any order. Let $h\ge2$, and let $c_1,\dots,c_{h-1}$ be rationals, any of which may vanish. Let $M$ be the matrix of \S\ref{ssec:frames} built from the increment \[ \delta(k-1)=\sum_{j=1}^{h-1}c_jk^{-j}\qquad(k\ge1), \] the increment of a rational combination of $\psi(z+1),\psi'(z+1),\dots,\psi^{(h-2)}(z+1)$. The frames of the third plane are the case $h=4$, with $c_1=0$, $c_2=-\gamma_2$ and $c_3=2\gamma_3$. As in Section~\ref{sec:plane2}, write a node as $t=jp+\omega$ with $0\le\omega\bar N$ that divides no denominator of the $c_j$. Let $W(t,t')$ be $0$ for distinct nodes in the same period, $h-1$ for nodes in different periods that are not congruent modulo $p$, and $h$ for distinct congruent nodes. Let $W(t,t)=h\cdot\mathbf 1[t\ge p]$. Then the entry of $\Phi_{r,I}$ at $(t,t')$ has Gauss valuation at least $-W(t,t')$. \end{lemma} \begin{proof} For $t\bar N$, every $k\le\bar N$ has $v_p(k)\le1$. So each term has valuation at least $-(h-1)$, and it is a $p$-integer unless $p\mid k$. If $t$ and $t'$ lie in one period, no multiple of $p$ lies in $(t,t']$. The difference $t'-t$ has valuation $1$ if $t\equiv t'\pmod p$ and $0$ otherwise. The diagonal entry is $x-e_t$, and $e_t$ is a sum of the terms $jc_jk^{-j-1}$ with $k\le t$. These have valuation at least $-h$, and they are $p$-integers for $t2\bar N+2p$ that divides no denominator of the $c_j$. Suppose that $2m2\bar N'$, and let $H=M+p^{-h}T\Id$ on $[0,\bar N')$. The pivot consists of the complete periods $[0,A)$, on the rows and on the columns. Lemma~\ref{lem:displace} uses only the displacement identity, so it holds for this $H$. \emph{The pivot.} Let $S(j)$ be the sum of the terms $c_ik^{-i}$ of $f_t$ with $p\mid k$, for a node $t$ of the period $j$, and $G(j)$ the same for $e_t$. Then $f_t-S(j)$ and $e_t-G(j)$ are $p$-integers. Since $p^2>\bar N'$, the numbers $s(j)=p^{h-1}S(j)$ and $s'(j)=p^hG(j)$ are $p$-integers. Define $B$ on the period indices $[0,\kappa)$ by \[ B(l,j)=\frac{s(j)-s(l)}{j-l}\ (l\ne j),\qquad B(j,j)=T-s'(j). \] If $t\equiv t'\pmod p$, then $p^hH(t,t')\equiv B(j_t,j_{t'})\pmod{p^{h-1}}$. If $t\not\equiv t'$, then $p^hH(t,t')\in p\cO$, because every entry has valuation at least $-(h-1)$ off the diagonal. So $p^hH[[0,A),[0,A)]$ is congruent modulo $p$ to a direct sum of copies of $\bar B$, one for each residue class. The variable $T$ occurs only on the diagonal of $B$, so $\det\bar B$ is monic in $T$, and the pivot is invertible over $\cO$. No coefficient $c_j$ needs to be a $p$-unit. \emph{The weights of the Schur complement.} Since $h\ge3$, the right side $(\mathbf 1,p^{h-1}f)$ of the pivot system is, modulo $p^2$, a function of the period. The pivot has the same complete periods on both sides, so $B$ is symmetric and every residue class carries the same matrix. The proof of Proposition~\ref{prop:gen} then applies with the halves omitted, with $p^3$ and $p^2$ replaced by $p^h$ and $p^{h-1}$, and with $\delta_{\rm rot}=0$. It gives vectors $\mathcal V(j)$ with $\tilde g_z\equiv\eta_z\mathcal V(j_z)\pmod{p^2}$ and $\eta_z\equiv1\pmod p$, for the labels $z$ in the periods $\kappa-1$ and $\kappa$. The rescaled Schur complement has entries $p^{-(h-1)}(\tilde g_z\wedge\tilde g_{z'})/(z'-z)$. So for labels $z\ne z'$ its entry has valuation at least $-h$ if $z\equiv z'\pmod p$, at least $-(h-3)$ if $z\not\equiv z'$ lie in one period, and at least $-(h-1)$ otherwise. On the diagonal it has valuation at least $-h$. \emph{Potentials.} The labels outside $I$ are the border nodes and, when $\bar NA$, let $\chi(t)=1$ for the border nodes on the side of $A$ that has fewer of them, and $\chi(t)=0$ for the others. If $\bar N\le A$, let $\chi=0$. Give a border row the potential $h-3+2\chi$, a border column $2\chi$, a virtual node $h-3$ as a row and $3$ as a column, and a deleted node $h$ on both sides. A virtual node against itself has weight $h=(h-3)+3$. Two labels in one period that are not congruent have weight $h-3$. Two labels on opposite sides of $A$ have weight $h-1$, and exactly one of them has $\chi=1$. A deleted node carries the largest weight $h$. So the potentials dominate the weights. As in the proof of Theorem~\ref{thm:rows2}(i), Lemma~\ref{lem:displace}(a) and the count of the pivot give $e_p(r,I)\le hr+\sum u+\sum u'$, with the sums over the border rows and columns. The border carries $m$ rows and $\beta$ nodes with $\chi=1$, so the sum is $(h-3)m+2\beta$. \end{proof} \subsection{The denominator exponent} Homogenise with $\theta=p/r$ and $m\mapsto\alpha r$ as in \S\ref{ssec:profiles}. For $h\ge3$ and $0<\alpha\le\frac13$ the minimum $\mathcal P_h(\theta)$ of the four homogenised bounds is \begin{center} {\renewcommand{\arraystretch}{1.2} \begin{tabular}{@{}ll@{}} \toprule range of $\theta$ & $\mathcal P_h(\theta)$\\ \midrule $0<\theta\le\frac{1+\alpha}2$ & $h(1+\alpha)-\theta\|\alpha/\theta\|$\\ $\frac{1+\alpha}2\le\theta\le\frac12+\alpha$ & $2h-1+2(h-1)\alpha-2(h-1)\theta$\\ $\frac12+\alpha\le\theta\le1$ & $2h(1+\alpha-\theta)$\\ $1\le\theta\le1+2\alpha$ & $h(1+2\alpha-\theta)$\\ $\theta\ge1+2\alpha$ & $0$\\ \bottomrule \end{tabular}} \end{center} The values agree at the ends of the ranges. The low range is the sawtooth of \S\ref{ssec:checker}. We take the minimum with (Q5). On $\frac1j\le\theta\le\frac1{j-1}$, away from the endpoints, its homogenised form is \[ h+(h-3)\alpha+2\max\bigl(0,\min(j\theta-1,1+2\alpha-j\theta)\bigr). \] At an endpoint we use the next multiple of $p$ in Proposition~\ref{prop:rows3e}. The formal proof integrates an affine majorant of these bounds, rather than the ideal profile. \begin{theorem}[denominators on the third plane]\label{thm:content3} Let $\alpha=\frac{13}{200}$, and let $\gamma_2$ and $\gamma_3\ne0$ be rationals. Let $K\ge0$ and $\epsilon>0$. For every sufficiently large even $r$, every member of $F^{(3)}_{\gamma_2,\gamma_3}$ with $|I|\le K$ satisfies $\log\den P_{r,I}\le(\Lambda_3+\epsilon)r^2$, where \[ \begin{aligned} \Lambda_3&=\frac{2093141963035255548476961845489855904191277848149}{624885257513513299995586450872\cdot10^{18}}\\ &=3.3496420948\ldots \end{aligned} \] \end{theorem} \begin{proof} Apply (Q1)--(Q4) and add (Q5) on the head $\theta\ge2\alpha$. The chambers $\frac1j\le\theta\le\frac1{j-1}$ for $j=2,\dots,8$ cover $[2\alpha,1]$. In each chamber (Q5) has three affine forms: $hr+(h-3)m$ if $\bar N\le jp$, $hr+(h-3)m+2(jp-r)$ if $jp\le r+m$, and $hr+(h-3)m+2(\bar N-jp)$ if $r+m\le jp\le\bar N$. The endpoints are checked with the actual next multiple in Proposition~\ref{prop:rows3e}. Its prime-square condition holds for large $r$ on this head. Oddness of $p$ turns $2m\le p$ into $2m\mu_1>\rho/\eta$. Suppose that $|\xi-p/q|1$. This is a contradiction for large $q$. If $\xi=p/q$ were rational, then for every large $r$ some member would satisfy $q^{-r}\le|\den(P_{r,I})\,P_{r,I}(\xi)|\le e^{-(\eta-o(1))r^2}$, which is impossible. So $\xi$ is irrational. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:offplane}, Theorem~\ref{thm:points1}(ii) and Theorem~\ref{thm:ends} for $\pi^2$] For each of these rows of Table~\ref{tab:main} the function is $F_s$ of \eqref{eq:frames}, and the geometry parameter and the constants are those of Table~\ref{tab:data}. The denominator exponent is given by Theorem~\ref{thm:content} with the table for $\Lambda$ of the formal proof. The decay rates are given by Proposition~\ref{prop:cert}, and nonvanishing by Theorem~\ref{thm:selection}. Theorem~\ref{thm:deduction}, applied at the even scales, gives $\mu(\xi)\le\rho/(D-\Lambda)$, and in every row $\rho/(D-\Lambda)$ is below the stated bound. By \eqref{eq:frames}, $\xi$ is a nonzero rational multiple of $\Psi(s,1)$. Since the inequality is strict, the definition of $\mu$ gives $|\xi-p/q|>q^{-\mu_0}$ for all integers $p$ and all sufficiently large $q$. \end{proof} Theorem~\ref{thm:points1}(i) is proved in Section~\ref{sec:chi3-quarter}, parts (iii)--(v) and Theorem~\ref{thm:plane1} in Section~\ref{sec:plane1}, Theorem~\ref{thm:plane2} and the bound for $\zeta(3)$ in Section~\ref{sec:plane2}, and Theorem~\ref{thm:plane3} in Section~\ref{sec:plane3}. In each case the bound is attained in the same strong form. \begin{table}[ht] \centering {\setlength{\tabcolsep}{1.5pt} \begin{tabular}{@{}lcccccc@{}} \toprule case & $c$ & $\Lambda$ & $D$ & $\rho$ & $\rho/\eta$ & $\mu\le$\\ \midrule plane 1 & $\frac16$ & $\LamPl$ & $\DPl$ & $\gPl$ & $\ExpPl$ & $\MuPlaneA$\\ $(\frac13,\frac23)$ & $\frac12$ & $\LamChi$ & $\DChi$ & $\gChi$ & $\ExpChi$ & $\MuChi$\\ $(\frac k3,1)$ & $\frac3{10}$ & $\LamThree$ & $\DThree$ & $\gThree$ & $\ExpThree$ & $\MuThirdsU$\\ $\Lt\pm\frac{2\pi^2}{27}$ & $\frac16$ & $\LamDA$ & $\DPl$ & $\gPl$ & $\ExpDA$ & $\MuDipA$\\ $\Lt\pm\frac{\pi^2}{27}$ & $\frac16$ & $\LamDB$ & $\DPl$ & $\gPl$ & $\ExpDB$ & $\MuDipB$\\ $\Lt\pm\frac{\pi^2}{18}$ & $\frac14$ & $\LamDC$ & $\DDC$ & $\gDC$ & $\ExpDC$ & $\MuDipC$\\ \addlinespace plane 2 & $\frac6{41}$ & $\LamPlB$ & $\DPlB$ & $\gPlB$ & $\ExpPlB$ & $\MuPlaneB$\\ $\zeta(3)$ & $\frac3{10}$ & $\LamZ$ & $\DHalf$ & $\gHalf$ & $\ExpZ$ & $\MuZetaThree$\\ $(\frac12,1)$ & $\frac3{10}$ & $\LamHalf$ & $\DHalf$ & $\gHalf$ & $\ExpHalf$ & $\MuHalvesU$\\ \addlinespace plane 3 & $\frac{13}{100}$ & $\LamPlC$ & $\DPlC$ & $\gPlC$ & $\ExpPlC$ & $\MuPlaneC$\\ \addlinespace $(\frac k4,1)$ & $\frac15$ & $\LamFour$ & $\DFour$ & $\gFour$ & $\ExpFour$ & $\MuLamU$\\ $(\frac k5,1)$ & $\frac29$ & $\LamFive$ & $\DFive$ & $\gFive$ & $\ExpFive$ & $\MuFifthsU$\\ $(\frac k7,1)$ & $\frac16$ & $\LamSeven$ & $\DSeven$ & $\gSeven$ & $\ExpSeven$ & $\MuSeventhsU$\\ \bottomrule \end{tabular}} \caption{The data of the proofs. A pair $(s,t)$ denotes the function of \eqref{eq:frames} for $\Psi(s,t)$, with $k$ prime to the denominator. For each row the table gives the geometry parameter $c$, the denominator exponent $\Lambda$, the decay rate $D$, the neighbourhood exponent $\rho$ and the bound for $\mu$. The constants are rational numbers, rounded outward to the digits shown: $\Lambda$ and $\rho$ up, $D$ down. The column $\rho/\eta$, with $\eta=D-\Lambda$, is the exact quotient, rounded up. The rows $\Lt\pm\frac{2\pi^2}{27}$ and $\Lt\pm\frac{\pi^2}{27}$ share the decay certificate of plane 1. The row for $\zeta(3)$ shares the certificate of the row $(\frac12,1)$.}\label{tab:data} \end{table} \begin{remark}[relation to linear independence criteria] The deduction is a one-dimensional version of the familiar scheme behind the criteria of Nesterenko \cite{Nesterenko85} and their use by Ball and Rivoal \cite{BallRivoal01}. The nonzero integer $q^{|C\setminus I|}\den(P_{r,I})\,P_{r,I}(p/q)$ plays the role of a small linear form with integer coefficients. There are two differences. The lower bound for the form comes from an arithmetic nonvanishing statement rather than from an analytic asymptotic. The size $r$ of the determinant is chosen in terms of $q$. \end{remark} \subsection{Notes on the individual cases}\label{sec:families} For the rows $(\frac kq,1)$ the function is $F_s$ of \eqref{eq:frames} with $s=k/q$, $\omega=q^{-2}$ and $\xi_s=-q^{-2}\Psi(\frac kq,1)$. The table for $\Lambda$, the trial measure and the cover are data of the formal proof. All shifts with the same denominator share one table for $\Lambda$, indexed by the root classes, and one certificate. Indeed the bounds depend on the shift only through the root class, and the decay depends on it only through the separation of the poles from the contour. For $q=4$ the root classes $j\equiv kp^{-1}\pmod 4$ of $k=1$ and $k=3$ exchange the two classes of primes modulo $4$. For $q=3$ the table uses the primes in the classes modulo $6$, and the root classes $j=1,2$ correspond to first poles near $p/3$ and $2p/3$. For $q=7$ there are six root classes. Here $D-\Lambda$ is only about $4\cdot10^{-3}$, which explains the large exponent. A small further gain in either constant would lower it substantially (Section~\ref{sec:limits}). All certificates take $\varepsilon_0=10^{-6}$ in the function $\Upsilon$. The trial measure for $q=7$ has $22$ polynomial pieces, and its cells on $\Gamma$ at the breakpoints of $\sigma$ use the truncated logarithm of \S\ref{ssec:slope-cert}. For $q=2$ there is one root class, $j=1$, with first pole $k_p=\frac{p+1}2$. On the second plane $D-\Lambda$ is about $0.021$, and on the third plane about $0.096$. This explains the size of their exponents. On the first plane the points $\Lt\pm\frac{2\pi^2}{27}$ and $\Lt\pm\frac{\pi^2}{27}$ gain only in $\Lambda$, since they use the certificate of the plane. The points $\Lt\pm\frac{\pi^2}{18}$ use a longer border, with $c=\frac14$, and their own certificate. %% ===================================================================== \section{Limits of the method}\label{sec:limits} The exponents are determined by three constants through $\rho/(D-\Lambda)$. The derivative of $\rho/(D-\Lambda)$ with respect to $D$ is $-\rho/\eta^2$, so the exponent is most sensitive to the decay where $\eta$ is small. For the shifts $\frac14$ and $\frac34$, $\rho/\eta^2\approx135$. For the shifts $\frac k7$ it is about $4.4\cdot10^5$. On the arithmetic side, the profiles $\mathcal P_j$ are minima of the bounds of Theorem~\ref{thm:local}. Every further local bound lowers $\Lambda$ by the integral of the amount by which it lies below $\mathcal P_j$. The denominator exponent of the method is the integral of the true exponent $e_p$, which the local bounds approximate from above. At third order this approximation is close to exhaustion for the shifts $\frac k5$ and $\frac k7$. We computed the Smith normal forms exactly at scales $r\le2400$ and primes $p\le2m$. There the minimum of the local bounds, with (U$_n$) included, equals the true exponent at all but a few isolated primes. For the shifts $\frac k7$ the invariants of valuation at least $4$ contribute about $5\cdot10^{-6}$ to $\Lambda$. The bounds (U$_n$) gain about $7\cdot10^{-4}$. So at third order the local bounds are essentially exhausted. These computations are not part of the proof. On the analytic side, Remark~\ref{rem:certlimit} explains why a certificate of the present form cannot do better than the constrained equilibrium measure on the chosen contour. A larger decay rate needs a different contour or a different treatment of the energy problem. The geometry parameter $c$ trades the two sides against each other. A longer border increases $D$, through the larger node measure. It also increases $\Lambda$, through the border rows and columns of the assignment bound. \subsection{Constants not reached}\label{ssec:unreached} The construction applies to $\Psi(\frac kq,1)$ for every denominator $q$, but it proves a bound only when $D>\Lambda$. The shifts $\frac k6$ give points of the first plane, through the identities of Section~\ref{sec:intro}. For the next denominators we estimated the gap $D-\Lambda$ at the geometry parameter $c=\frac17$. We took the best local bounds of Section~\ref{sec:content} for $\Lambda$, and a floating-point trial measure for $D$. For $q=11$, $9$ and $8$ the gap is about $-0.41$, $-0.53$ and $-0.56$. For $q=10$ and $q=12$ it is about $-1.26$ and $-1.64$. For the six differences $\Psi(\frac k5,\frac l5)$ with $1\le k