\documentclass[11pt,reqno]{amsart} \usepackage{jaswal} \usepackage{tikz} \newcommand{\Q}{\mathbb{Q}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\SL}{\mathrm{SL}} \newcommand{\GL}{\mathrm{GL}} \newcommand{\tri}[4]{(#1,#2,#3)/#4} \newcommand{\res}[2]{[#1]_{#2}} \newcommand{\Area}{\operatorname{area}} \newcommand{\Jac}{\operatorname{Jac}} \newcommand{\lcm}{\operatorname{lcm}} \newcommand{\famA}{\mathcal{A}} \newcommand{\famB}{\mathcal{B}} \newcommand{\famC}{\mathcal{C}} \newcommand{\listE}{\mathcal{E}} \newcommand{\listF}{\mathcal{F}} \makeatletter \def\LT@makecaption#1#2#3{\LT@mcol\LT@cols c{\hbox to\z@{\hss\parbox[t]\LTcapwidth{\@makecaption{#2}{#3}\endgraf\vskip4pt}\hss}}} \makeatother \setlength{\LTcapwidth}{\textwidth} \begin{document} \title[Obtuse rational lattice triangles]{Obtuse rational lattice triangles} \author{Manvir Jaswal} \address{Toronto, Canada} \email{manvirjaswal@icloud.com} \urladdr{https://manvirjaswal.github.io} \date{November 17, 2026} \subjclass[2020]{Primary 37C83; Secondary 32G15, 11A07, 30F30} \keywords{Rational billiards, lattice triangles, Veech surfaces, Teichm\"uller curves, translation surfaces} \hypersetup{pdftitle={Obtuse rational lattice triangles},pdfauthor={Manvir Jaswal},pdfkeywords={Rational billiards, lattice triangles, Veech surfaces, Teichmüller curves, translation surfaces}} \begin{abstract} A rational triangle is a lattice triangle if the Veech group of its unfolding is a lattice. Answering a question of Hooper, we prove that an obtuse rational triangle is a lattice triangle if and only if its angles are $(\pi/n,\pi/n,(n-2)\pi/n)$ with $n\ge5$, $(\pi/2k,\pi/k,(2k-3)\pi/2k)$ with $k\ge4$, or $(\pi/12,\pi/3,7\pi/12)$. The idea is a first-order test on relative periods. By Möller's theorem, on a Veech surface the difference of two zeros of the holomorphic form is torsion in the factor of the Jacobian with real multiplication, and stays torsion along the Teichmüller curve. So certain relative periods vanish along the Teichmüller disc, and on a triangle unfolding their first derivative is an integral over the plane that can be computed in closed form. For obtuse angle at most $2\pi/3$ we determine for every denominator where the criterion of Mirzakhani and Wright is blind, and exclude all but four of the triangles there by relative periods and by cylinders whose moduli or areas cannot occur on a Veech surface. For obtuse angle above $2\pi/3$ the proof uses a theorem of Larsen, Norton and Zykoski. Parts of the proof are computer-assisted. \end{abstract} \maketitle \contents \section{Introduction}\label{sec:intro} A billiard ball in a polygon whose angles are rational multiples of $\pi$ moves, after unfolding, along straight lines on a compact translation surface: the surface glued from the copies of the polygon under the finite group generated by the linear parts of the reflections in its sides \cite{ZemljakovKatok}. Veech \cite{Veech89} singled out the translation surfaces whose affine automorphisms have derivatives forming a lattice in $\SL_2(\R)$. On such a \emph{Veech surface} the straight-line flow in each direction is either completely periodic or uniquely ergodic, and the number of periodic families of length at most a given bound grows like a constant times the square of the bound. A rational polygon is a \emph{lattice polygon} when its unfolding is a Veech surface. Veech showed that the obtuse isosceles triangles with angles $(\pi/n,\pi/n,(n-2)\pi/n)$, $n\ge 5$, are lattice polygons \cite[Theorem~1.1 and \S8]{Veech89}, and the question of which rational triangles are lattice polygons has been studied since. We write $\tri pqrn$ for the triangle with angles $p\pi/n$, $q\pi/n$, $r\pi/n$, where $p\le q\le r$ are positive integers with $\gcd(p,q,r)=1$ and $n=p+q+r$; the integer $n$ is the \emph{denominator}. The triangle is obtuse exactly when $r>n/2$. It is a \emph{lattice triangle} if the Veech group of its unfolding is a lattice, the unfolding being taken as a translation surface whose only marked points are its cone points of angle above $2\pi$; precise definitions are in \cref{sec:setting}. \begin{theorem}\label{thm:main} An obtuse rational triangle is a lattice triangle if and only if it is $\tri11{n-2}n$ with $n\ge5$, $\tri12{2k-3}{2k}$ with $k\ge4$, or $\tri147{12}$. \end{theorem} The obtuse lattice triangles are those of Veech, of Vorobets and Ward, and Hooper's triangle. This answers a question of Hooper \cite[Open Question~2]{Hooper}. McMullen lists the known lattice triangles and asks whether the list is complete \cite[Theorem~3.9 and Question~3.11]{McMullenBAMS}. The obtuse triangles of his list are all the obtuse lattice triangles, which answers his question for obtuse triangles. \Cref{thm:main} also proves the conjecture, as stated by Angdinata, Chen, Ono, Thorner, Zhang and Zhang, that there are no obtuse lattice triangles other than the known ones \cite[p.~2]{ACOTZZ}. By \cref{thm:main} there are at most two obtuse lattice triangles of each denominator other than $12$, and three of denominator $12$. The proof is assembled in \cref{sec:classification}. The idea of the proof is a first-order test on relative periods. By the criterion of Mirzakhani and Wright in the form of Larsen, Norton and Zykoski (\cref{prop:test}), an obtuse triangle with a witness is not a lattice triangle; a witness is a unit $a$ modulo $n$ with $2a\not\equiv2\pmod n$ satisfying at least two of three inequalities, defined in \cref{ssec:test}. The criterion compares two eigenforms of the cyclic symmetry of the unfolding through the variation of the period matrix. We say that the criterion is \emph{blind} on a triangle with no witness. Where it is blind we use the relative periods instead. On a Veech surface the difference of two zeros of the holomorphic form is a torsion point of the factor of the Jacobian with real multiplication by the trace field, of order dividing a fixed integer at every surface of the Teichmüller curve, by Möller's theorem (\cref{prop:moller}). On the unfolding of a lattice triangle the eigenforms of the cyclic symmetry whose character is a unit modulo $n$ lie in the cohomology of that factor, so the relative period of such an eigenform between two zeros, over a suitable rational relative class, vanishes along the whole Teichmüller disc. Its first derivative is given by integrals over the plane, which can be computed in closed form. For an integer $x$ let $\res xn$ be its residue in $[0,n)$. \begin{theorem}\label{thm:relative} Let $\Delta=\tri pqrn$ be a rational triangle. \begin{enumerate} \item\label{it:TI} If $\Delta$ is obtuse, $n$ is odd, and $p\nmid n$ or $q\nmid n$, then $\Delta$ is not a lattice triangle. \item\label{it:TII} Let $x\in\{p,q,r\}$ with $x\nmid n$ and $g=\gcd(x,n)>1$, and let $b$ be a unit modulo $n$ with $b\equiv 2\pmod{n/g}$, $b\not\equiv 2\pmod n$ and $\res{bp}n+\res{bq}n+\res{br}n=n$. Then $\Delta$ is not a lattice triangle. \end{enumerate} \end{theorem} The condition $\res{bp}n+\res{bq}n+\res{br}n=n$ in \ref{it:TII} says that the eigenspace of character $b$ contains a holomorphic form. Part \ref{it:TI} places no condition on the size of the obtuse angle, and part \ref{it:TII} none on the angles or on the parity of $n$. The proof of \cref{thm:main} has three cases, by the parity of $n$ and the size $r\pi/n$ of the obtuse angle. The criterion rules out every triangle with a witness. When the obtuse angle is at most $2\pi/3$, the \emph{window} of \cref{ssec:blind}, it is blind exactly on three families and a list (\cref{thm:list}): $\famA$ and $\famB$, in which $s=2r-n$ divides $n$ with odd quotient, with $p=1$ in $\famA$, and $2\le p\le s/2$ and a condition on units in $\famB$; $\famC$, in which $s=2$, $4\mid n$ and $p$ is $1$, $2$ or, for $n\equiv4\pmod8$, $4$; and the list $\listE$ of thirty triangles with $n\le48$. Above $2\pi/3$ it is blind exactly on the six one-parameter families (1) to (6) and the seven triples (7) of \cref{thm:lnz41}. For each case the table names the result that says where it is blind, the lattice triangles, and the result that rules out each of the other triangles; \emph{cylinders} means two parallel cylinders whose moduli have an irrational ratio, or one whose share of the area has a Galois conjugate outside $(0,1]$, which cannot occur on a Veech surface (\cref{lem:cylinders}). \begin{center}\small \begin{tabular}{@{}>{\raggedright}p{4.6cm}>{\raggedright}p{4.3cm}>{\raggedright\arraybackslash}p{5.5cm}@{}} \toprule case; lattice triangles & other triangles & ruled out by\\ \midrule $n$ odd (\cref{thm:pq}); $\tri11{n-2}n$ & $p\nmid n$ or $q\nmid n$ & \cref{thm:relative}\ref{it:TI}\\ & $\tri159{15}$ & \cref{thm:relative}\ref{it:TII}\\ & $\tri135{9}$, $\tri37{11}{21}$ & \cref{thm:finite}\\ \addlinespace $n$ even, $r>2n/3$ (\cref{thm:lnz41}); $\tri11{n-2}n$, $\tri12{n-3}n$ & families (3) to (6) of \cref{thm:lnz41} & cylinders: \cite[\S10]{LNZ}, \cref{thm:lnz4,thm:lnz5,thm:lnz6}\\ & the triples of (7) & \cref{lem:TII-single} or \cref{thm:finite}\\ \addlinespace $n$ even, $r\le2n/3$ (\cref{thm:list}); $\tri114{6}$, $\tri125{8}$, $\tri147{12}$ & $\famA$, $\famB$ & \cref{thm:relative} (\cref{thm:AB})\\ & $\famC$ & cylinders: \cref{thm:star,thm:branched,thm:C4}\\ & the even members of $\listE$ & \cref{lem:TII-single} or \cref{thm:finite}\\ \bottomrule \end{tabular} \end{center} \Cref{thm:list}, for every denominator, is the counterpart in the window of the theorem of Larsen, Norton and Zykoski for obtuse angle above $2\pi/3$ (\cref{thm:lnz41}). \Cref{thm:list}\ref{it:list-all}, for finitely many denominators, and \cref{thm:finite} are computer-assisted; the computations are exact and are described in \cref{sec:computations}. \subsection{Earlier work} For right and acute isosceles triangles the answer is known. Kenyon and Smillie proved that a right triangle is a lattice triangle exactly when its smallest angle is $\pi/n$ with $n\ge4$, and an acute isosceles triangle exactly when its apex angle is $\pi/n$ with $n\ge3$ \cite[Theorem~2]{KenyonSmillie} (the `if' parts were known before \cite[p.~67]{KenyonSmillie}). For acute scalene triangles they reduced the question to a number-theoretic statement, which they verified for denominators up to $10{,}000$ \cite[Theorem~1]{KenyonSmillie}. Puchta \cite{Puchta} proved the statement under a coprimality assumption, and the general case is treated in a later preprint \cite{SchlagePuchta}. Apisa determined the orbit closure of the unfolding of every rational isosceles triangle \cite[Theorems~1.4 and 7.7]{Apisa}; in particular the obtuse isosceles lattice triangles are Veech's. What has remained open is the case of obtuse scalene triangles (see \cite[Theorem~3.9 and Corollary~3.10]{McMullenBAMS}). Two kinds of obtuse scalene lattice triangles are known. The triangles with angles $(\pi/2k,\pi/k,(2k-3)\pi/2k)$, $k\ge 4$, were suggested by computations of Boshernitzan reported by Veech \cite[Remark on p.~582]{Veech89}, announced by Vorobets \cite[Theorem~5]{VorobetsNote}, proved to be lattice triangles in \cite[Proposition~4.2 and Theorem~4.4]{Vorobets}, and studied by Ward, who determined their Veech groups \cite[Theorem~A]{Ward}. Bouw and Möller constructed, for every $m0$, and angles $p\pi/n$ at $0$, $q\pi/n$ at $1$. By the law of sines $v=\sin(q\pi/n)e^{ip\pi/n}/\sin(r\pi/n)$; multiplying numerator and denominator of \[ v=\frac{e^{i(p+q)\pi/n}-e^{i(p-q)\pi/n}}{e^{ir\pi/n}-e^{-ir\pi/n}} \] by $e^{ir\pi/n}$ and using $e^{i(p+q+r)\pi/n}=-1$ gives $v=(e^{-iq\vartheta}-1)/(e^{ir\vartheta}-1)\in \Q(\zeta)$. The $2n$ copies of $\Delta$ in $X$ are $e^{ij\vartheta}\Delta$ and $e^{ij\vartheta}\overline\Delta$, $j\in\Z/n$, so their vertices lie in $\Q(\zeta)$, and the holonomy of every saddle connection and of every closed geodesic of $X$ lies in $\Q(\zeta)$, being the holonomy of a path along sides of the copies. This is the observation of Kenyon and Smillie \cite[\S5]{KenyonSmillie}. For a cylinder $Z$ of $X$ let $\rho(Z)=\Area(Z)/\Area(X)$ be its \emph{share} of the area. If $c,w\in \Q(\zeta)$ are the holonomies of the core curve of $Z$ and of a segment crossing $Z$ between singular points on its two boundary components, then \[ \rho(Z)=\frac{\bar cw-c\bar w}{n(v-\bar v)}\quad\text{(up to sign)}, \] since $\Area(Z)=|\operatorname{Im}(\bar cw)|$ and $\Area(X)=2n\cdot\frac12\operatorname{Im}v$. So $\rho(Z)$ is a real element of $\Q(\zeta)$: $\rho(Z)\in \Q(\zeta)^+$. Part \ref{it:cyl-sign} of the next lemma follows in two lines from the proof of Theorem~1 of Fenves with his Lemma~3 \cite{Fenves}; Kenyon and Smillie used a positivity condition of the same kind \cite[Theorem~21 and Proposition~23]{KenyonSmillie}. \needspace{4\baselineskip} \begin{lemma}\label{lem:cylinders} Let $\Delta$ be a rational triangle whose unfolding $X$ has genus at least two, and let $Z$ be a maximal cylinder of $X$, in direction $d$. If $\Delta$ is a lattice triangle, then: \begin{enumerate}[label=\textup{(\roman*)}] \item\label{it:cyl-ratio} any two maximal cylinders of $X$ in direction $d$ have moduli whose ratio is rational; \item\label{it:cyl-sign} $0<\sigma(\rho(Z))\le1$ for every embedding $\sigma$ of $\Q(\zeta)^+$ in $\R$. \end{enumerate} \end{lemma} \begin{proof} A boundary component of $Z$ is a union of saddle connections in direction $d$, so by \cref{prop:veech210} every leaf in direction $d$ is closed or a saddle connection. Then $X$ is the union of the closures of finitely many maximal cylinders $Z_j$ in direction $d$ \cite[Proposition~2.4]{Veech89}, and their moduli $\mu_j$ satisfy $\mu_j=\lambda_j\mu_1$ with $\lambda_j\in\Q_{>0}$ \cite[Remark on p.~582]{Veech89}. A closed leaf of $Z$ meets no saddle connection, so it lies in one of the $Z_j$; both $Z$ and $Z_j$ are connected components of the union of the closed regular leaves in direction $d$ containing that leaf, so $Z=Z_j$. The same holds for every maximal cylinder in direction $d$, so each of them is one of the $Z_j$; this proves \ref{it:cyl-ratio}. Let $c_j\in \Q(\zeta)$ be the holonomy of the core curve of $Z_j$. Then $\Area(Z_j)=\mu_j|c_j|^2$, and the numbers $x_j=c_j/c_1$ lie in $\Q(\zeta)$ and are real, since the $c_j$ are parallel; so $x_j\in \Q(\zeta)^+\setminus\{0\}$, and \[ \rho(Z_i)=\frac{\mu_i|c_i|^2}{\sum_j\mu_j|c_j|^2}=\frac{\lambda_ix_i^2}{\sum_j \lambda_jx_j^2}. \] Both sides are elements of $\Q(\zeta)^+$. For an embedding $\sigma$ of $\Q(\zeta)^+$, each $\sigma(x_j)$ is a non-zero real number, so $\sigma(\rho(Z_i))$ is a quotient of positive numbers whose numerator is one of the terms of the denominator, and $0<\sigma(\rho(Z_i))\le1$. \end{proof} Conversely, a translation surface on which every direction containing a cylinder decomposes into cylinders with pairwise commensurable moduli is a Veech surface \cite[Theorem~1]{Fenves}; so every rational triangle that is not a lattice triangle has a direction containing a cylinder in which $X$ does not decompose into cylinders with pairwise commensurable moduli. \subsection{The criterion of Mirzakhani and Wright}\label{ssec:test} Mirzakhani and Wright proved that the orbit closure of the unfolding has rank at least two (the rank is half the complex dimension of the image of its tangent space in absolute cohomology; a closed orbit has rank one) when the eigenspaces of $T$ of characters $a$ and $2-a$ both contain holomorphic forms, for a unit $a$ with $2a\not\equiv 2\pmod n$ \cite[Theorem~7.5]{MirzakhaniWright}. Larsen, Norton and Zykoski put this in the following form. \begin{proposition}[{Larsen, Norton and Zykoski \cite[Proposition~2.1]{LNZ}}]\label{prop:test} Let $\Delta=\tri pqrn$ be obtuse. If some unit $a$ modulo $n$ with $2a\not\equiv2\pmod n$ satisfies two of the three inequalities \[ \res{ap}n<\res{2p}n,\quad \res{aq}n<\res{2q}n,\quad \res{ar}n<\res{2r}n, \] then $\Delta$ is not a lattice triangle. \end{proposition} Following Larsen, Norton and Zykoski, for an integer $x$ set \[ S_x=\{a\in\Z/n\Z : \res{ax}n<\res{2x}n\}, \] and call a unit $a$ modulo $n$ \emph{usable} if $2a\not\equiv 2\pmod n$. A \emph{witness} for an obtuse triangle is a usable unit $a$ for which at least two of the three inequalities $\res{ap}n<\res{2p}n$, $\res{aq}n<\res{2q}n$, $\res{ar}n<\res{2r}n$ hold, that is, which lies in at least two of $S_p$, $S_q$, $S_r$, where for $p=q$ the first two inequalities are one inequality and count as two, as in \cref{prop:test}. For an obtuse triangle write $s=2r-n$, so that $\res{2p}n=2p$, $\res{2q}n=2q$ and $\res{2r}n=s$. \begin{theorem}[{\cite[Theorem~4.1]{LNZ}}]\label{thm:lnz41} An obtuse triangle $\tri pqrn$ with $r>2n/3$ has a witness if and only if none of the following is the case: \begin{enumerate}[label=(\arabic*)] \item $p=q=1$ \item $p=1$, $q=2$, $n$ is even \item $p=1$, $q=4$, $r\equiv 7\bmod 8$ \item $p=1$, $r=3q+1$ \item $p=2$, $r=3q+2$ \item $p=4$, $r=3q+4$ \item $(p,q,r)$ is one of the following triples: $(1,4,11)$, $(1,3,16)$, $(2,3,17)$, $(1,4,21)$, $(1,8,19)$, $(3,8,29)$, $(2,11,29)$. \end{enumerate} \end{theorem} The members of family (3) with $n>12$ are not lattice triangles: Larsen, Norton and Zykoski find two horizontal cylinders whose moduli have the irrational ratio $4\cos^2(2\pi/n)$ \cite[\S10]{LNZ}. \section{Relative periods}\label{sec:relative} This section proves \cref{thm:relative}, which is applied in \cref{sec:classification} to odd denominators and to the single triangles of \cref{lem:TII-single}, and in \cref{ssec:AB} to the families $\famA$ and $\famB$. On the unfolding of a lattice triangle, the relative period of an eigenform with unit character between two zeros of $\omega$, over a suitable rational relative class, vanishes along the whole Teichmüller disc (\cref{lem:unitW,lem:torsion,lem:flat}). Its first derivative is the difference of the values at these zeros of a solution $G$ of a $\bar\partial$-equation (\cref{lem:first}), and on a triangle unfolding these values are integrals over the plane that can be computed. The obstruction is thus of first order, as is the one behind the criterion of Mirzakhani and Wright, whose proof, as they remark, can be viewed as showing that the $\SL_2(\R)$-orbit of the unfolding is not even tangent to the locus of the surfaces with the endomorphisms that an orbit closure of rank one would force \cite[after the proof of Theorem~7.5]{MirzakhaniWright}. \subsection{Forms on the unfolding}\label{ssec:forms} In this section we allow any placement of the vertices. Let $(x_0,x_1,x_\infty)$ be an ordering of $(p,q,r)$, set $\theta_i=x_i/n$ and $g_i=\gcd(x_i,n)$ for $i\in\{0,1,\infty\}$, and let $X$ be the normalization of the curve $y^n=z^{x_0}(z-1)^{x_1}$, with $\omega=y\,dz/(z(z-1))$ and $T(z,y)=(z,\zeta y)$, $\zeta=e^{2\pi i/n}$. The substitutions $z\mapsto 1-z$ and $z\mapsto 1/z$ (with $y$ multiplied by a suitable constant, respectively by a suitable constant times $1/z$) permute the numerators and multiply $\omega$ by a non-zero constant, so $(X,\omega)$ is the unfolding of $\Delta$ of \cref{ssec:unfolding} up to a constant factor, which changes neither the zeros of $\omega$ nor whether $\Delta$ is a lattice triangle. Above $z=0$, $1$, $\infty$ there are $g_0$, $g_1$, $g_\infty$ points; at a point above $z=0$ there is a local coordinate $u$ with $z=u^{n/g_0}$ and $y/u^{x_0/g_0}$ a unit (as $(y/u^{x_0/g_0})^n=(z-1)^{x_1}$), and similarly above $1$ and above $\infty$ (with $z=u^{-n/g_\infty}$). The form $\omega$ has a zero of order $x_i/g_i-1$ at each point above the vertex $i$, so these points are zeros of $\omega$ exactly when $x_i\nmid n$; the set of zeros is the set $\Sigma$ of \cref{ssec:unfolding}. A meromorphic $1$-form $\eta$ on $X$ has \emph{character} $b\in\Z/n\Z$ if $T^*\eta=\zeta^b\eta$; for instance $\omega$ has character $1$. Let $H^1_b\subset H^1(X,\C)$ be the $\zeta^b$-eigenspace of $T^*$ and $H^{1,0}_b$ the space of holomorphic forms of character $b$. Since $T^*$ is real, complex conjugation maps $H^{1,0}_{-b}$ onto the $\zeta^b$-eigenspace of $T^*$ on antiholomorphic forms, so $H^1_b=H^{1,0}_b\oplus\overline{H^{1,0}_{-b}}$. \begin{lemma}\label{lem:forms} Let $b$ be an integer. \begin{enumerate}[label=\textup{(\alph*)}] \item\label{it:forms-shape} Every non-zero meromorphic form of character $b$ is $y^bR(z)\,dz$ with $R$ rational. It is holomorphic if and only if $R=z^i(z-1)^jh(z)$ with integers $i,j$ and a polynomial $h$ with $h(0)h(1)\ne0$ such that, with $e_0=b\theta_0+i$ and $e_1=b\theta_1+j$, \[ e_0>-1,\qquad e_1>-1,\qquad e_0+e_1+\deg h<-1 . \] For any $R=z^i(z-1)^jh$ its order at a point above $0$ is $(n/g_0)(e_0+1)-1$, above $1$ it is $(n/g_1)(e_1+1)-1$, and above $\infty$ it is $-(n/g_\infty)(e_0+e_1+\deg h+1)-1$. \item\label{it:forms-unit} If $bx_i\not\equiv0\pmod n$ for $i=0,1,\infty$, then $\dim H^{1,0}_b=2-(\{b\theta_0\}+\{b\theta_1\}+\{b\theta_\infty\})$. So $H^{1,0}_b$ has dimension $1$ if $\res{bx_0}n+\res{bx_1}n+\res{bx_\infty}n=n$, spanned by $y^bz^{i_b}(z-1)^{j_b}dz$ with $b\theta_0+i_b=\{b\theta_0\}-1$ and $b\theta_1+j_b=\{b\theta_1\}-1$, and it is zero otherwise. Moreover $\dim H^1_b=1$. \item\label{it:forms-trivial} If $b\not\equiv0$ but $bx_i\equiv 0\pmod n$ for some $i$, then $H^1_b=0$. Also $H^{1,0}_0=0$. \end{enumerate} \end{lemma} \begin{proof} \ref{it:forms-shape} If $\eta$ has character $b$, then $\eta/(y^b\,dz)$ is a $T$-invariant meromorphic function, hence a rational function of $z$. Above a point of $\C\setminus\{0,1\}$ the map $z$ is unramified and $y$ is a unit, so a holomorphic $\eta$ has $R$ without poles off $0$, $1$, $\infty$, that is $R=z^i(z-1)^jh(z)$ as stated. In the local coordinate above $0$ the form is a unit times $u^{bx_0/g_0+in/g_0}\,d(u^{n/g_0})$, of order $(n/g_0)(b\theta_0+i+1)-1$, an integer; it is non-negative if and only if $e_0>-1$. The same computation applies above $1$, and above $\infty$, where $y$ is a unit times $u^{-(x_0+x_1)/g_\infty}$ and $z^i(z-1)^jh(z)\,dz$ is a unit times $u^{-(n/g_\infty)(i+j+\deg h+1)-1}du$. \ref{it:forms-unit}, \ref{it:forms-trivial} For $b\not\equiv0$ these are \cite[Lemma~6.1 and Corollary~6.2]{MirzakhaniWright}, with $k=n$, $q_i=x_i$ and $t(b)=\sum_i\{b\theta_i\}$ in their notation, after the change of variable that puts the three vertices at $0$, $1$, $\infty$: a Möbius transformation sending distinct points $z_0,z_1,z_\infty\in\C$ to $0,1,\infty$, with $y$ replaced by a constant times $y/(z-z_\infty)$, takes their curve $y^n=\prod_i(z-z_i)^{x_i}$ to ours, as $x_0+x_1+x_\infty=n$, and commutes with $T$. The form named in \ref{it:forms-unit} has character $b$ and is holomorphic by \ref{it:forms-shape}: for it $h=1$, $e_0=\{b\theta_0\}-1>-1$, $e_1=\{b\theta_1\}-1>-1$ and $e_0+e_1=-1-\{b\theta_\infty\}<-1$. In \ref{it:forms-trivial} exactly one $bx_i$ is divisible by $n$ (if two were, all three would be, and then $b\equiv0$, as $\gcd(x_0,x_1,x_\infty)=1$), so $\dim H^1_b=t(b)+t(-b)-2=0$. Finally $H^{1,0}_0=0$, because a $T$-invariant holomorphic form descends to $X/\langle T\rangle=\mathbb P^1$. \end{proof} Summing \ref{it:forms-unit} over $b$ gives the genus \eqref{eq:genus}. A holomorphic form of character $b$ is therefore determined up to a constant when it exists, and we write $\eta_b$ for it. \subsection{The Teichmüller disc and Möller's theorem}\label{ssec:moller} \emph{The Teichmüller disc.} Let $(X,\omega)$ be a translation surface with zero set $\Sigma$, and let $\mathbb D$ be the unit disc. For $t\in\mathbb D$ let $X_t$ be the Riemann surface whose underlying topological surface is that of $X$, marked by the identity, and whose holomorphic charts are $w+t\bar w$ off $\Sigma$, where $w$ is a local primitive of $\omega$, and $u_t=u\,(1+t\bar u^{k+1}/u^{k+1})^{1/(k+1)}$ near a zero of order $k$, where $u$ is a coordinate with $u^{k+1}=(k+1)w$. Then $u_t^{k+1}=(k+1)(w+t\bar w)$, so the charts are compatible, and $\omega_t=d(w+t\bar w)$ is a holomorphic form on $X_t$ with the same zeros. The pair $(X_t,\omega_t)$ is the image of $(X,\omega)$ under the real-linear map $v\mapsto v+t\bar v$ of $\C=\R^2$, which has positive determinant $1-|t|^2$; every element of $\GL_2^+(\R)$ is a complex multiple of exactly one such map, so $t\mapsto X_t$ is the Teichmüller disc of $(X,\omega)$, and its image in the moduli space is the Teichmüller curve $C=\mathbb H/\Gamma$ when $(X,\omega)$ is a Veech surface with Veech group $\Gamma$ (we identify $\mathbb D$ with the upper half-plane $\mathbb H$). The identity map $X\to X_t$ is a homeomorphism, not in general smooth at the zeros, and it identifies $H^1(X_t,\Q)$ with $H^1(X,\Q)$ for every $t$. We write $X_0=X$. \emph{The complex surface $\mathbb D\times X$.} The product $\mathbb D\times X$ carries the complex structure with charts $(t,w+t\bar w)$ away from the zeros of $\omega$ and $(t,u_t)$ near a zero. Both charts are holomorphic in $t$ at a fixed point of $X$, so $t\mapsto(t,P)$ is a holomorphic section for every point $P$ of $X$, and the fibre over $t$ is $X_t$. A holomorphic family $\eta_t$ of forms, $\eta_t$ holomorphic on $X_t$, is a holomorphic relative $1$-form on this complex surface; a local primitive $\Phi(t,x)$ along the fibres is a holomorphic function, determined up to adding a holomorphic function of $t$. \emph{The family over the Teichmüller curve.} Let $(X,\omega)$ be a Veech surface of genus at least two whose singular set is its zero set $\Sigma$, with Veech group $\Gamma$, trace field $K=\Q(\operatorname{tr}\Gamma)$. For every subgroup $\Gamma_1\subset\Gamma$ of finite index that is small enough there is a family of curves $f:X_1\to C_1=\mathbb H/\Gamma_1$, the pull-back of the universal family over a finite cover of the moduli space of curves, whose fibre over the image of a point of $\mathbb H=\mathbb D$ is the corresponding surface of the Teichmüller disc \cite[\S1]{Moller}. Let $\mathbb V=R^1f_*\Z$, with its variation of Hodge structure. By \cite[Proposition~2.4]{MollerVHS} it splits as $\mathbb V_\Q=\mathbb W_\Q\oplus\mathbb M_\Q$, a direct sum of rational variations of Hodge structure of weight one, where $\mathbb W_\Q$ has rank $2[K:\Q]$ and the fibre of $\mathbb W_\Q\otimes\R$ at $X$ contains $\operatorname{Re}\omega$ and $\operatorname{Im}\omega$. Möller's factor $A_1$ of $\Jac(X_1/C_1)$ with real multiplication by $K$ is the one given by this splitting: there is a homomorphism $\pi:\Jac(X_1/C_1)\to A_1$ of abelian schemes over $C_1$ whose pull-back maps the rational first cohomology of the fibres of $A_1$ isomorphically onto $\mathbb W_\Q$ (see \cite[(2), and the paragraph after Theorem~3.1]{Moller} and \cite[before Proposition~2.6]{MollerVHS}). By \cite[Theorem~3.1]{Moller} the group of sections of $A_1\to C_1$ is finite. \needspace{5\baselineskip} \begin{proposition}\label{prop:moller} Let $(X,\omega)$ be a Veech surface of genus at least two whose singular set is its zero set $\Sigma$, with trace field $K$. There is a subspace $W\subset H^1(X,\Q)$ of dimension $2[K:\Q]$ such that: \begin{enumerate}[label=\textup{(\alph*)}] \item\label{it:moller-W} $[\omega]\in W\otimes\C$; for every $t\in\mathbb D$, $W\otimes\C=W^{1,0}_t\oplus\overline{W^{1,0}_t}$ with $W^{1,0}_t=(W\otimes\C)\cap H^{1,0}(X_t)$; and the spaces $W^{1,0}_t$ form a holomorphic subbundle of the Hodge bundle over $\mathbb D$; \item\label{it:moller-torsion} for every divisor $D$ of degree $0$ supported on $\Sigma$ there is an integer $N\ge1$ with the following property. Let $\gamma$ be a $1$-chain on $X$ with $\partial\gamma=D$. For every $t\in\mathbb D$ there is $\lambda_t\in H_1(X,\Z)$ with \[ N\int_\gamma\eta=\int_{\lambda_t}\eta\qquad\text{for every holomorphic form $\eta$ on $X_t$ with $[\eta]\in W\otimes\C$.} \] \end{enumerate} \end{proposition} \begin{proof} Take $\Gamma_1\subset\Gamma$ of finite index, acting freely on $\mathbb H$ (Selberg's lemma; the lattice $\Gamma$ is finitely generated), and so small that $f$ exists and that every $P\in\Sigma$ extends to a section $\tilde P$ of $f$ whose value on each fibre is a zero of the form of that fibre: the orders of the zeros are constant along the $\GL_2^+(\R)$-orbit, so the zeros form a curve in $X_1$ that is finite and unramified over $C_1$, and it becomes trivial over a finite unramified covering of $C_1$ (as in the proof of \cite[Lemma~1.2]{Moller}). Then $\mathbb D\cong\mathbb H\to C_1$ is a universal covering, and $C_1\to C$ is unramified as a map of orbifolds, though not in general as a map of the underlying curves: the surface $X$ itself is an orbifold point of $C$ when its Veech group contains a rotation, as it does for a triangle unfolding. Every derivative below is taken in the parameter $t$ of the disc, which is a local coordinate of $C_1$ everywhere. The pull-back of $f$ along $\mathbb D\to C_1$ is a family of curves of genus at least two over the simply connected disc, marked at $X$; a marked family over $\mathbb D$ is the pull-back of the universal family over Teichmüller space by its classifying map, which here is the Teichmüller disc $t\mapsto X_t$. So the pull-back is the family $t\mapsto X_t$ (as a complex surface it is $\mathbb D\times X$, described at the start of this subsection). The sections $\tilde P$ pull back to the constant sections $t\mapsto P$: the pull-back of $\tilde P$ is continuous, its values lie in the finite set $\Sigma$ of zeros of $\omega_t$, and its value at $t=0$ is $P$. Let $W$ be the fibre of $\mathbb W_\Q$ at $X$, a constant subspace of $H^1(X,\Q)$ along the disc, of dimension $2[K:\Q]$. As $\mathbb W_\Q$ is a sub-variation of Hodge structure of weight one and $W\otimes\R$ contains $\operatorname{Re}\omega$ and $\operatorname{Im}\omega$, part~\ref{it:moller-W} holds, and $W^{1,0}_t$ has dimension $[K:\Q]$. Let $D=\sum_PD(P)P$ be a divisor of degree $0$ on $\Sigma$, and $\sigma_D=\sum_PD(P)\,\pi\bigl(\tilde P-\tilde P_0\bigr)$ for a fixed $P_0\in\Sigma$, a section of $A_1\to C_1$; it has finite order $N_1$. Let $t\in\mathbb D$, with image $c\in C_1$, and write $\Jac(X_t)=H^{1,0}(X_t)^\vee/H_1(X,\Z)$ and $A_{1,c}=H^{1,0}(A_{1,c})^\vee/H_1(A_{1,c},\Z)$. The differential $d\pi_c$ is the transpose of $\pi_c^*:H^{1,0}(A_{1,c})\to H^{1,0}(X_t)$; as $\pi_c^*$ is an injective morphism of Hodge structures with image $W$, it maps $H^{1,0}(A_{1,c})$ onto $W^{1,0}_t$ (both have dimension $[K:\Q]$), so the kernel of $d\pi_c$ consists of the functionals that vanish on $W^{1,0}_t$. The cokernels of $(\pi_c)_*:H_1(X,\Z)\to H_1(A_{1,c},\Z)$ are finite and form a local system on $C_1$; let $m$ be the exponent of its fibres. As the sections $\tilde P$ are constant along the disc, $\sigma_D(c)$ is the image under $\pi_c$ of the class of $\eta\mapsto\int_\gamma\eta$, for any chain $\gamma$ with $\partial\gamma=D$. So $N_1\sigma_D(c)=0$ gives $\lambda_t\in H_1(X,\Z)$ with $d\pi_c\bigl(mN_1\int_\gamma\bigr)=d\pi_c\bigl(\int_{\lambda_t}\bigr)$, that is, $mN_1\int_\gamma\eta=\int_{\lambda_t}\eta$ for every $\eta\in W^{1,0}_t$. This is \ref{it:moller-torsion} with $N=mN_1$. \end{proof} For $t\in\mathbb D$ let $A_t=(W^{1,0}_t)^\vee/H_1(X,\Z)$, where a class $\lambda$ acts by $\eta\mapsto\int_\lambda\eta$; by \ref{it:moller-W} it is a complex torus of dimension $[K:\Q]$. Part \ref{it:moller-torsion} says that the image of $D$ in $A_t$, the class of the functional $\eta\mapsto\int_\gamma\eta$ on $W^{1,0}_t$, has order dividing $N$ for every $t$. As the kernel of $d\pi_c$ consists of the functionals that vanish on $W^{1,0}_t$ and the cokernel of $(\pi_c)_*$ is finite, $\pi_c$ factors through the map $\Jac(X_t)\to A_t$ that restricts functionals to $W^{1,0}_t$, and the induced map from $A_t$ to the fibre of $A_1$ at the image of $t$ in $C_1$ is an isogeny. This is the factor $A$ of the Jacobian, of dimension $[K:\Q]$, in which the differences of the zeros stay torsion. When $[K:\Q]$ is the genus, $W=H^1(X,\Q)$ and $A_t=\Jac(X_t)$, so they stay torsion in the whole Jacobian. On the unfolding of a triangle the difference of two zeros of $\omega$ is torsion in the Jacobian, as Mirzakhani and Wright remark \cite[\S1]{MirzakhaniWright} (it also follows from \cref{lem:torsion} below); what a lattice triangle adds is that it stays torsion, in $A$, along the whole Teichmüller disc. In Filip's extension of Möller's theorem to affine invariant manifolds the torsion statement holds for the combinations of zeros annihilated by the relative directions of the manifold, and for Teichmüller curves, which have none, for all of them \cite[Theorem~1.3 and Remark~1.4\,(ii), (iii)]{FilipSplitting}. So the argument below shows that a triangle is not a lattice triangle, not that its orbit closure has rank at least two, and indeed \cref{thm:relative}\ref{it:TII} applies to $\tri136{10}$ and $\tri138{12}$ (\cref{ex:TII}). Delecroix and R\"uth list these two as rank-one orbit closures that are not Teichmüller curves, with a reference to work in preparation \cite[Theorem~4.1]{DelecroixRuth}. A criterion that proves rank at least two cannot apply to a triangle of rank one. Möller's torsion condition has been used to exclude Teichmüller curves on the stable curves over the cusps, by McMullen \cite{McMullenTorsion} and, for a factor of the Jacobian, by Lanneau and Möller \cite[\S3]{LanneauMoller}; here it is read at the surface of the triangle itself. \subsection{Relative periods that vanish along the disc}\label{ssec:three} \begin{lemma}\label{lem:unitW} Let $\Delta$ be a lattice triangle whose unfolding $(X,\omega)$ has genus at least two, and let $W$ be as in \cref{prop:moller}. Then $H^1_b\subset W\otimes\C$ for every unit $b$ modulo $n$. In particular $H^{1,0}_b\subset W^{1,0}_0$ for every unit $b$, and $b=2$ is a unit exactly when $n$ is odd. \end{lemma} \begin{proof} The automorphism $T^*$ of $H^1(X,\Q)$ is defined over $\Q$, so its $\zeta$-eigenspace $H^1_1$ is defined over $\Q(\zeta)$; it contains $[\omega]$ and has dimension $1$ by \cref{lem:forms}\ref{it:forms-unit}, as $1$ is a unit and so is non-trivial at the three vertices. Let $v\in H^1(X,\Q(\zeta))$ span it. Then $v\in\C[\omega]\subset W\otimes\C$ by \cref{prop:moller}\ref{it:moller-W}, so $v\in(W\otimes\C)\cap H^1(X,\Q(\zeta))=W\otimes\Q(\zeta)$, as $W$ is a rational subspace. For a unit $b$ let $\sigma_b$ be the automorphism of $\Q(\zeta)$ with $\sigma_b(\zeta)=\zeta^b$, acting on the coefficients. Then $\sigma_b(v)\in W\otimes\Q(\zeta)$, since $W$ is rational, and $T^*\sigma_b(v)=\sigma_b(T^*v)=\zeta^b\sigma_b(v)$, since $T^*$ is rational. So $\sigma_b(v)$ is a non-zero vector of $H^1_b$, which has dimension $1$ by \cref{lem:forms}\ref{it:forms-unit}; hence $H^1_b\subset W\otimes\C$. \end{proof} \begin{lemma}\label{lem:torsion} Let $X$ be the curve of \cref{ssec:forms} for any rational triangle, and let $\Sigma_0$, $\Sigma_1$, $\Sigma_\infty$ be the sums of the points of $X$ above $z=0$, $1$, $\infty$. \begin{enumerate}[label=\textup{(\alph*)}] \item\label{it:torsion-div} $\operatorname{div}(z)=(n/g_0)\Sigma_0-(n/g_\infty)\Sigma_\infty$ and $\operatorname{div}(z-1)=(n/g_1)\Sigma_1-(n/g_\infty)\Sigma_\infty$. \item\label{it:torsion-quot} Let $P$, $P'$ be two points above the same vertex, say $z=0$, and $g=g_0$. Then $(n/g)(P-P')$ is the divisor of a meromorphic function on $X$. \end{enumerate} Consequently every divisor $D$ of degree $0$ supported above $\{0,1,\infty\}$ is torsion in $\Jac(X)$, and there is a class $\hat\gamma\in H_1(X,\operatorname{supp}D;\Q)$ with $\partial\hat\gamma=D$ such that $\int_{\hat\gamma}\eta=0$ and $\int_{\hat\gamma}\bar\eta=0$ for every holomorphic form $\eta$ on $X$. \end{lemma} \begin{proof} \ref{it:torsion-div} follows from the local coordinates of \cref{ssec:forms}. \ref{it:torsion-quot} The substitutions of \cref{ssec:forms} move any vertex to $z=0$, so we may assume the points lie above $0$. The quotient $X/\langle T^g\rangle$ is the normalization of $y''^g=z^{x_0}(z-1)^{x_1}$ with $y''=y^{n/g}$, which is invariant under $T^g$; with $w=y''/z^{x_0/g}$ it becomes $w^g=(z-1)^{x_1}$. Here $\gcd(x_1,g)=1$, since a prime dividing $g$ and $x_1$ divides $x_0$, $x_1$ and $n$, hence all three numerators. Choosing integers $\alpha$, $\beta$ with $\alpha x_1+\beta g=1$, the function $v=w^\alpha(z-1)^\beta$ satisfies $v^g=z-1$ and $w=v^{x_1}$, so the quotient has genus $0$. The stabilizer of a point above $z=0$ in $\langle T\rangle$ has order $n/g$, since the $g$ points above $0$ form one orbit; it is $\langle T^g\rangle$. So the quotient map is totally ramified, of index $n/g$, at $P$ and at $P'$, and the pull-back of a function on the quotient with divisor (image of $P$) $-$ (image of $P'$) has divisor $(n/g)(P-P')$. The differences of two points above the same vertex, and the divisors $g_\infty\Sigma_0-g_0\Sigma_\infty$, $g_\infty\Sigma_1-g_1\Sigma_\infty$ (of which $n/(g_0g_\infty)$ and $n/(g_1g_\infty)$ times are principal; $g_0$, $g_1$, $g_\infty$ are pairwise coprime because $\gcd(p,q,r)=1$), generate the divisors of degree $0$ supported above $\{0,1,\infty\}$ up to finite index, so every such $D$ has a multiple $ND=\operatorname{div}(F)$. Let $\gamma$ be a $1$-chain with $\partial\gamma=D$. By Abel's theorem $\eta\mapsto N\int_\gamma\eta$ is the period map of some $\lambda\in H_1(X,\Z)$ on holomorphic forms, and $\hat\gamma=\gamma-\lambda/N$ has the stated property; since $\hat\gamma$ is a real class, $\int_{\hat\gamma}\bar\eta=\overline{\int_{\hat\gamma}\eta}=0$. \end{proof} \needspace{4\baselineskip} \begin{lemma}\label{lem:flat} Let $(X,\omega)$ be a Veech surface as in \cref{prop:moller}, $D$ a divisor of degree $0$ supported on $\Sigma$, and $\hat\gamma\in H_1(X,\Sigma;\Q)$ a class with $\partial\hat\gamma=D$ such that $\int_{\hat\gamma}\eta=0$ for every holomorphic form $\eta$ on $X$. Then $\int_{\hat\gamma}\eta=0$ for every $t\in\mathbb D$ and every holomorphic form $\eta$ on $X_t$ with $[\eta]\in W\otimes\C$. \end{lemma} The relative homology $H_1(X,\Sigma;\Z)$ and the homology $H_1(X,\Z)$ do not change along the disc, because the underlying surface and the points of $\Sigma$ do not move; only the complex structure does. \begin{proof} Let $\Lambda\subset\operatorname{Hom}(W,\R)$ be the image of $H_1(X,\Z)$, a lattice, and for $t\in\mathbb D$ let $\iota_t:\operatorname{Hom}(W,\R)\to(W^{1,0}_t)^\vee$ send a functional to the restriction of its complex-linear extension; by \cref{prop:moller}\ref{it:moller-W}, $\iota_t$ is an $\R$-linear isomorphism, and it depends continuously on $t$. Fix a chain $\gamma$ with $\partial\gamma=D$ and let $N$ be as in \cref{prop:moller}\ref{it:moller-torsion}. The functional $\eta\mapsto N\int_\gamma\eta$ on $W^{1,0}_t$ depends continuously on $t$ (see the complex surface $\mathbb D\times X$ in \cref{ssec:moller}), and by \cref{prop:moller}\ref{it:moller-torsion} it equals $\iota_t(\lambda(t))$ with $\lambda(t)\in\Lambda$. So $\lambda(t)=\iota_t^{-1}(N\int_\gamma|_{W^{1,0}_t})$ is a continuous function on the disc with values in the discrete set $\Lambda$, hence constant, $\lambda(t)=\lambda$. Let $\tilde\lambda\in H_1(X,\Q)$ map to $\lambda$ and let $\gamma'=\gamma-\tilde\lambda/N$, a rational relative class with $\partial\gamma'=D$ and $\int_{\gamma'}\eta=0$ for every $\eta\in W^{1,0}_t$ and every $t$. The difference $\gamma'-\hat\gamma$ lies in $H_1(X,\Q)$ and vanishes on $W^{1,0}_0$; being real, it vanishes on $\overline{W^{1,0}_0}$, hence on $W\otimes\C$. As $W$ and $\gamma'-\hat\gamma$ do not depend on $t$, $\gamma'-\hat\gamma$ vanishes on $W^{1,0}_t$ for every $t$, and so does $\hat\gamma$. \end{proof} In this proof the rational relative class $\gamma'=\gamma-\tilde\lambda/N$ vanishes on $W^{1,0}_t$ for every $t$; for affine invariant manifolds the corresponding flatness is shown in the proof of \cite[Theorem~5.2]{FilipSplitting}. \subsection{The first-order formula}\label{ssec:first} \begin{lemma}\label{lem:first} Let $(X,\omega)$ be a translation surface, $X_t$ ($t\in\mathbb D$) as in \cref{ssec:moller}, $S\subset X$ a finite set, and $\hat\gamma\in H_1(X,S;\R)$ a class with $\int_{\hat\gamma}\eta=0$ for every holomorphic form $\eta$ on $X_0$; let $D=\partial\hat\gamma$. Let $\eta_t$ be a holomorphic family of forms, $\eta_t$ holomorphic on $X_t$. On $X_0$ write \[ \frac{\eta_0}{\omega}\,\bar\omega=\bar\partial G+H \] with $H$ antiholomorphic and $G$ a continuous function. Then \[ \frac{d}{dt}\Big|_{t=0}\int_{\hat\gamma}\eta_t=G(D):=\sum_PD(P)\,G(P). \] \end{lemma} Such $G$ and $H$ exist: the left side is a bounded $(0,1)$-form, smooth off the zeros of $\omega$ (at a zero of order $k$, $\eta_0/\omega$ has a pole of order at most $k$ and $\bar\omega$ vanishes to order $k$), $H$ is its antiholomorphic part in the Hodge decomposition, and $G$ is smooth where the left side is and Hölder continuous everywhere, by elliptic regularity; $G$ is unique up to an additive constant, which does not change $G(D)$. The derivative depends on the family only through $\eta_0$, since two families with the same $\eta_0$ differ by $t$ times a holomorphic family plus $O(t^2)$, and $\hat\gamma$ vanishes on holomorphic forms on $X_0$. Kahn and Wright \cite[Theorem~1.1]{KahnWright} give a formula of the same kind for the derivative of the map from a stratum to the moduli space of curves, as an integral against a quadratic differential of the $(0,1)$-part of a closed form, divided by $\omega$, plus terms at the zeros. \begin{proof} On the complex surface $\mathbb D\times X$ of \cref{ssec:moller} the family $\eta_t$ is a holomorphic relative $1$-form; let $\Phi(t,x)$ be a local primitive along the fibres, a holomorphic function determined up to adding a holomorphic function of $t$. Let $\dot\Phi=\partial_t\Phi|_{t=0}$ at fixed $x\in X$: it is defined locally up to a constant, $d\dot\Phi$ is a closed $1$-form on $X$ off the zeros, and for a path $\gamma$ in $X$, $\frac{d}{dt}|_{0}\int_\gamma\eta_t=\int_\gamma d\dot\Phi$. Away from the zeros, $\Phi=F(t,w+t\bar w)$ with $F$ holomorphic in two variables, so $\dot\Phi=F_t(0,w)+F_v(0,w)\bar w$ and $\bar\partial\dot\Phi=F_v(0,w)\,d\bar w=(\eta_0/\omega)\bar\omega$, as $\eta_0=F_v(0,w)\,dw$ and $\omega=dw$. Near a zero of order $k$, $\Phi=\tilde F(t,u_t)$ and $\dot\Phi=\tilde F_t(0,u)+\tilde F_u(0,u)\,\bar u^{k+1}/((k+1)u^k)$, which is continuous at $u=0$. Now $\tilde\eta=d\dot\Phi-dG-H$ is a form of type $(1,0)$ off a finite set, since its $(0,1)$-part is $\bar\partial\dot\Phi-\bar\partial G-H=0$, and it is closed, since $dH=0$; so it is holomorphic there. Near each point of the finite set it is the differential of the continuous function $\dot\Phi-G-(\text{a local primitive of }H)$, which is holomorphic off the point and hence across it; so $\tilde\eta$ is a holomorphic form on $X_0$. Hence $d\dot\Phi=dG+H+\tilde\eta$, and \[ \frac{d}{dt}\Big|_{0}\int_{\hat\gamma}\eta_t=\int_{\hat\gamma}d\dot\Phi=G(D)+\int_{\hat\gamma}H+\int_{\hat\gamma}{\tilde\eta}=G(D), \] the last two integrals vanishing because $\hat\gamma$ is real and vanishes on holomorphic forms on $X_0$. \end{proof} \needspace{4\baselineskip} \begin{corollary}\label{cor:test} Let $\Delta$ be a lattice triangle whose unfolding $(X,\omega)$ has genus at least two, $b$ a unit modulo $n$ with $H^{1,0}_b\ne0$, and write $(\eta_b/\omega)\bar\omega=\bar\partial G_b+H_b$ as in \cref{lem:first}. Then $G_b(D)=0$ for every divisor $D$ of degree $0$ supported on the zeros of $\omega$. \end{corollary} \begin{proof} By \cref{lem:torsion} there is $\hat\gamma\in H_1(X,\operatorname{supp}D;\Q)$ with $\partial\hat\gamma=D$ vanishing on holomorphic forms on $X_0$. By \cref{lem:unitW}, $\eta_b\in W^{1,0}_0$, and as $W^{1,0}_t$ is a holomorphic subbundle (\cref{prop:moller}\ref{it:moller-W}) there is a holomorphic family $\eta_t\in W^{1,0}_t$ with $\eta_0=\eta_b$. By \cref{lem:flat}, $\int_{\hat\gamma}\eta_t=0$ for all $t\in\mathbb D$. Differentiating at $t=0$ (\cref{lem:first}) gives $G_b(D)=0$. \end{proof} \needspace{4\baselineskip} \begin{lemma}\label{lem:residues} Let $\varpi$ be a bounded $(0,1)$-form on a compact Riemann surface $X$, smooth off a finite set, with $\varpi=\bar\partial G$ for a continuous function $G$, and let $\xi$ be a meromorphic $1$-form on $X$ with at most simple poles. Then \[ \int_X\varpi\wedge\xi=-2\pi i\sum_Q G(Q)\operatorname{Res}_Q\xi . \] \end{lemma} \begin{proof} Remove a coordinate disc of radius $\varepsilon$ about each pole of $\xi$ and each point where $\varpi$ is not smooth. On the rest $d(G\xi)=\bar\partial G\wedge\xi=\varpi\wedge\xi$, so by Stokes's theorem $\int\varpi\wedge\xi=-\sum\oint G\xi$ over the circles, oriented counterclockwise. As $\varepsilon\to0$ the left side tends to $\int_X\varpi\wedge\xi$, since $\varpi$ is bounded and $\xi$ has at most simple poles, and $\oint G\xi\to2\pi i\,G(Q)\operatorname{Res}_Q\xi$ by continuity of $G$. \end{proof} Both proofs below rest on one computation. As $T^*\omega=\zeta\omega$ and $T^*\bar\omega=\zeta^{-1}\bar\omega$, a form $\eta_b$ of character $b$ gives \[ T^*\Bigl(\frac{\eta_b}{\omega}\,\bar\omega\Bigr)=\zeta^{b-2}\,\frac{\eta_b}{\omega}\,\bar\omega . \] So the antiholomorphic part $H_b$ is the conjugate of a holomorphic form of the complementary character $b'=2-b$, $G_b$ can be taken with $G_b\circ T=\zeta^{-b'}G_b$, and for a meromorphic form $\xi$ of character $b'$ with at most simple poles the form $(\eta_b/\omega)\bar\omega\wedge\xi$ is $T$-invariant, so that its integral over $X$ is $n$ times an integral over the plane. When $H_b=0$, \cref{lem:residues} expresses the same integral through the values of $G_b$ at the poles of $\xi$. In part \ref{it:TI}, $b=2$, a unit as $n$ is odd; then $b'=0$, $H_2=0$ as $H^{1,0}_0=0$, and $\xi$ is $dz/z$ or $dz/(z-1)$. In part \ref{it:TII}, $b\equiv2\pmod{n/g}$ and $b\not\equiv2\pmod n$ say that $b'x\equiv0$ and $b'\not\equiv0\pmod n$. So $H_b=0$ by \cref{lem:forms}\ref{it:forms-trivial}; $\xi$ can be given a simple pole at each of the $g$ points above $x$ and no other pole; and $\zeta^{-b'}\ne1$, so that, if $\Delta$ is a lattice triangle, \cref{cor:test} applied to $D=TQ-Q$ for a point $Q$ above $x$ forces $G_b$ to vanish at all the points above $x$. \Cref{cor:test} concerns only divisors supported on the zeros of $\omega$, and the points above a vertex $x$ are zeros exactly when $x\nmid n$ (\cref{ssec:forms}): hence the hypothesis $x\nmid n$ in part \ref{it:TII}. \needspace{6\baselineskip} \subsection{Proof of \texorpdfstring{\cref{thm:relative}\ref{it:TI}}{the first part of the theorem on relative periods}}\label{ssec:TI} The two integrals of the next lemma are the cases $(\varepsilon_1,\varepsilon_2)=(0,1)$ and $(1,0)$, with $q_1/k=\theta_0$ and $q_2/k=\theta_1$, of the integral of \cite[Lemma~7.15]{MirzakhaniWright}, which arises there from the pairing of holomorphic forms of complementary characters in the proof of their Proposition~7.3. It is proved non-zero there, and the argument, a comparison of the integrand at symmetric points, also gives the signs. \begin{lemma}\label{lem:signs} Let $\theta_0,\theta_1>0$ with $\theta_0+\theta_1<1/2$, and $\mathcal K(z)=|z|^{2\theta_0-2}|z-1|^{2\theta_1-2}$. The integrals of $\mathcal K\cdot(z-1)$ and $\mathcal K\cdot z$ over $\C$ with respect to the area form $d^2z$ converge absolutely, and \[ \int_\C \mathcal K\cdot(z-1)\,d^2z<0,\qquad\int_\C \mathcal K\cdot z\,d^2z>0 . \] \end{lemma} \begin{proof} Near $0$ and $1$ the exponents $2\theta_i-2$ exceed $-2$, and at infinity the integrands are $O(|z|^{2\theta_0+2\theta_1-3})$ with $2\theta_0+2\theta_1-3<-2$. Write $I_1$ and $I_2$ for the integrals of $\mathcal K\cdot(z-1)$ and $\mathcal K\cdot z$. As $z-(z-1)=1$, $I_2-I_1=\int_\C\mathcal K\,d^2z>0$. Let $f=|z|^{2\theta_0}|z-1|^{2\theta_1}/(\bar z(\bar z-1))$, the function that appears in the proof of \cref{thm:relative}\ref{it:TI} below. Off $\{0,1\}$, \[ \frac{\partial f}{\partial z}=f\cdot\Bigl(\frac{\theta_0}{z}+\frac{\theta_1}{z-1}\Bigr)=\mathcal K\cdot\bigl(\theta_0(z-1)+\theta_1z\bigr), \] and $d(f\,d\bar z)=-2i\,(\partial f/\partial z)\,d^2z$. As $|f|=|z|^{2\theta_0-1}|z-1|^{2\theta_1-1}$, the integrals of $f\,d\bar z$ over the circles $|z|=\varepsilon$, $|z-1|=\varepsilon$ and $|z|=R$ are $O(\varepsilon^{2\theta_0})$, $O(\varepsilon^{2\theta_1})$ and $O(R^{2\theta_0+2\theta_1-1})$. So Stokes's theorem on $\{\varepsilon<|z|\varepsilon\}$ and absolute convergence give, as $\varepsilon\to0$ and $R\to\infty$, $\theta_0I_1+\theta_1I_2=0$. Hence \[ I_1=-\frac{\theta_1}{\theta_0+\theta_1}\,(I_2-I_1)<0,\qquad I_2=\frac{\theta_0}{\theta_0+\theta_1}\,(I_2-I_1)>0.\qedhere \] \end{proof} \begin{proof}[Proof of \cref{thm:relative}\ref{it:TI}] Suppose that $\Delta$ is a lattice triangle. Place the obtuse vertex at $\infty$: $(x_0,x_1,x_\infty)=(p,q,r)$, so that $\theta_0+\theta_1<1/2$. \emph{The form.} Let $\eta_2=y^2\,dz/(z(z-1))$, of character $2$. In \cref{lem:forms}\ref{it:forms-shape} it has $i=j=-1$, $e_0=2\theta_0-1>-1$, $e_1=2\theta_1-1>-1$ and $e_0+e_1=2(\theta_0+\theta_1)-2<-1$, the last exactly because $r>n/2$; so $\eta_2$ is holomorphic. As $n$ is odd, $2$ is a unit and $2x\not\equiv0\pmod n$ for $x=p,q,r$, so $H^{1,0}_2=\C\eta_2$ by \cref{lem:forms}\ref{it:forms-unit}. \emph{The function.} On $X$, \[ \frac{\eta_2}{\omega}\,\bar\omega=y\bar y\,\frac{d\bar z}{\bar z(\bar z-1)}=f\,d\bar z,\qquad f=\frac{|z|^{2\theta_0}|z-1|^{2\theta_1}}{\bar z(\bar z-1)}, \] which is $T$-invariant. In its decomposition $f\,d\bar z=\bar\partial G+H$ the antiholomorphic form $H$ is unique, hence $T$-invariant, and $\bar H\in H^{1,0}_0=0$ (\cref{lem:forms}\ref{it:forms-trivial}); averaging $G$ over $\langle T\rangle$ we may take $G$ to be $T$-invariant, that is, a continuous function of $z\in\mathbb P^1$. Write $G(0)$, $G(1)$, $G(\infty)$ for its values above the three vertices. \emph{Its values.} Apply \cref{lem:residues} with $\varpi=f\,d\bar z$ and $\xi=dz/z$. The form $\xi$ has simple poles above $0$, with residue $n/g_0$ at each of the $g_0$ points, and above $\infty$, with residue $-n/g_\infty$ at each of the $g_\infty$ points, and it is holomorphic above $1$. The integrand $\varpi\wedge\xi=(f/z)\,d\bar z\wedge dz$ is $T$-invariant, $f/z=\mathcal K\cdot(z-1)$ with $\mathcal K$ as in \cref{lem:signs}, and $d\bar z\wedge dz=2i\,d^2z$; so \begin{gather*} 2i\,n\int_\C \mathcal K\cdot(z-1)\,d^2z=-2\pi i\,n\,(G(0)-G(\infty)),\qquad\text{that is,}\\ G(0)-G(\infty)=-\frac1\pi\int_\C \mathcal K\cdot(z-1)\,d^2z>0 . \end{gather*} With $\xi=dz/(z-1)$ and $f/(z-1)=\mathcal K\cdot z$ the same computation gives $G(1)-G(\infty)=-\frac1\pi\int_\C \mathcal K\cdot z\,d^2z<0$ (\cref{lem:signs}). \emph{The contradiction.} The points above $\infty$ are zeros of $\omega$, since $r\nmid n$. By hypothesis the points above $0$, or those above $1$, are zeros too, say those above $i\in\{0,1\}$. The divisor $D=g_\infty\Sigma_i-g_i\Sigma_\infty$ has degree $0$ and is supported on zeros of $\omega$; as $\omega$ has zeros, $X$ has genus at least two, and \cref{cor:test} gives $0=G(D)=g_ig_\infty(G(i)-G(\infty))$, which contradicts the inequalities above. \end{proof} That the triangle is obtuse is needed for the statement and not only for the method: the acute triangles $\tri234{9}$ and $\tri122{5}$ are lattice triangles \cite[Theorems~1 and 2]{KenyonSmillie}, with $n$ odd and a numerator not dividing $n$. \needspace{4\baselineskip} \begin{example}\label{ex:124} The smallest triangle to which \cref{thm:relative}\ref{it:TI} applies and on which the criterion of Mirzakhani and Wright is blind is $\tri124{7}$, the member $s=1$, $n=7$ of the family $\famA$ of \cref{ssec:blind}. Here $\omega$ has a simple zero above $1$ and a zero of order $3$ above $\infty$, $X$ has genus $3$, and with $D=P_1-P_\infty$ the derivative of $\int_{\hat\gamma}\eta_t$ at $t=0$ is $G(1)-G(\infty)=-\frac1\pi\int_\C \mathcal K\cdot z\,d^2z\ne0$ (\cref{lem:signs}). \end{example} \subsection{Proof of \texorpdfstring{\cref{thm:relative}\ref{it:TII}}{the second part of the theorem on relative periods}}\label{ssec:TII} We use the complex beta integral in the form given by Neretin \cite[(1.4)]{Neretin}, who attributes it to Gel'fand, Graev and Vilenkin \cite[Section~II.3.7]{GelfandGraevVilenkin} and Gel'fand, Graev and Retakh \cite[Section~1.6]{GelfandGraevRetakh}. With $z^{u|u'}:=z^u\bar z^{u'}$, which is single-valued when $u-u'\in\Z$, it reads \begin{equation}\label{eq:neretin} B^\C(\alpha|\alpha',\beta|\beta'):=\frac1\pi\int_\C z^{\alpha-1|\alpha'-1}(1-z)^{\beta-1|\beta'-1}\,d^2z=\frac{\Gamma(\alpha)\Gamma(\beta)}{\Gamma(\alpha+\beta)}\cdot\frac{\Gamma(1-\alpha'-\beta')}{\Gamma(1-\alpha')\Gamma(1-\beta')}. \end{equation} Here $z^{u|u'}=|z|^{2u'}z^{u-u'}$, and the integral converges absolutely exactly when $\operatorname{Re}(\alpha+\alpha')>0$, $\operatorname{Re}(\beta+\beta')>0$ and $\operatorname{Re}(\alpha+\alpha'+\beta+\beta')<2$; we use \eqref{eq:neretin} only in this region. \begin{proof}[Proof of \cref{thm:relative}\ref{it:TII}] Suppose that $\Delta$ is a lattice triangle. Place the vertex $x$ at $0$: $x_0=x$, $g_0=g$, and $x_1$, $x_\infty$ are the other two numerators in either order. The $g_0$ points $Q_j=T^jQ_0$ ($j\in\Z/g_0\Z$) above $0$ are zeros of $\omega$, as $x\nmid n$. Fix an integer representing $b$ and let $b'=2-b$. Then $n\nmid b'$ and $(n/g_0)\mid b'$, so $b'\theta_0\in\Z$; and $b'\theta_1\notin\Z$, since otherwise $b'\theta_\infty\in\Z$ too, and then $n$ divides $b'x_0$, $b'x_1$, $b'x_\infty$, hence $b'$, as $\gcd(p,q,r)=1$. \emph{The form and the function.} As $b$ is a unit, $bx_i\not\equiv0\pmod n$ for each $i$, and with $\res{bp}n+\res{bq}n+\res{br}n=n$, \cref{lem:forms}\ref{it:forms-unit} gives $H^{1,0}_b=\C\eta_b$, $\eta_b=y^bz^{i_b}(z-1)^{j_b}dz$, where $a_0=\{b\theta_0\}=b\theta_0+i_b+1$ and $a_1=\{b\theta_1\}=b\theta_1+j_b+1$ lie in $(0,1)$ and $a_0+a_1<1$. The form $\varpi=(\eta_b/\omega)\bar\omega$ satisfies $T^*\varpi=\zeta^{b-2}\varpi=\zeta^{-b'}\varpi$. In its decomposition $\varpi=\bar\partial G+H$ the antiholomorphic form $H$ is unique, so $T^*H=\zeta^{-b'}H$ and $\bar H\in H^{1,0}_{b'}$, which is $0$ by \cref{lem:forms}\ref{it:forms-trivial}, as $b'x_0\equiv0$ and $b'\not\equiv0\pmod n$. So $\varpi=\bar\partial G$, and $G\circ T-\zeta^{-b'}G$ is a constant; adding a constant to $G$ (possible as $\zeta^{-b'}\ne1$) we get $G\circ T=\zeta^{-b'}G$, so $G(Q_j)=\zeta^{-b'j}G(Q_0)$. \emph{What the lattice property gives.} The divisor $D=Q_1-Q_0$ has degree $0$ and is supported on zeros of $\omega$; as $\omega$ has zeros, $X$ has genus at least two, and \cref{cor:test} gives $0=G(D)=(\zeta^{-b'}-1)\,G(Q_0)$, that is, $G(Q_0)=0$. \emph{The value of $G(Q_0)$.} Let $b'_1=\{b'\theta_1\}\in(0,1)$ and \[ \xi=y^{b'}\,z^{-1-b'\theta_0}(z-1)^{b'_1-1-b'\theta_1}\,dz, \] a meromorphic form of character $b'$ (the exponents of $z$ and $z-1$ are integers). By the order formulas of \cref{lem:forms}\ref{it:forms-shape}, with $e_0=-1$, $e_1=b'_1-1$ and $h=1$, it has order $(n/g_1)b'_1-1\ge0$ above $1$ and $(n/g_\infty)(1-b'_1)-1\ge0$ above $\infty$ (both $(n/g_1)b'_1$ and $(n/g_\infty)(1-b'_1)$ are positive integers, as $1-b'_1\equiv b'\theta_\infty\pmod 1$), and a simple pole at each $Q_j$, where $\xi=\mathcal{U}\,dz/z$ with $\mathcal{U}=y^{b'}z^{-b'\theta_0}(z-1)^{b'_1-1-b'\theta_1}$ a function of order $0$; so $\operatorname{Res}_{Q_j}\xi=(n/g_0)\mathcal{U}_j$ with $\mathcal{U}_j=\mathcal{U}(Q_j)\ne0$. Since $T^*\xi=\zeta^{b'}\xi$ and $G\circ T=\zeta^{-b'}G$, the product $G(Q_j)\operatorname{Res}_{Q_j}\xi$ does not depend on $j$, and \cref{lem:residues} gives \[ \int_X\varpi\wedge\xi=-2\pi i\,g_0\,G(Q_0)\,\frac n{g_0}\,\mathcal{U}_0=-2\pi i\,n\,\mathcal{U}_0\,G(Q_0). \] On the other hand $\varpi\wedge\xi$ is $T$-invariant, so $\int_X=n\int_\C$. As $b+b'=2$ and $y\bar y=|z|^{2\theta_0}|z-1|^{2\theta_1}$, \begin{gather*} \varpi\wedge\xi=y\bar y\,\frac{z^{k_0}(z-1)^{k_1}}{\bar z(\bar z-1)}\,d\bar z\wedge dz,\\ k_0=i_b-b'\theta_0=a_0-1-2\theta_0,\quad k_1=j_b+b'_1-b'\theta_1=a_1+b'_1-1-2\theta_1, \end{gather*} with $k_0$, $k_1$ integers; in the notation of \eqref{eq:neretin}, \[ \varpi\wedge\xi=(-1)^{k_1+1}\,z^{a_0-\theta_0-1|\theta_0-1}(1-z)^{a_1+b'_1-\theta_1-1|\theta_1-1}\cdot2i\,d^2z . \] Here $\alpha+\alpha'=a_0>0$, $\beta+\beta'=a_1+b'_1>0$ and $\alpha+\alpha'+\beta+\beta'=a_0+a_1+b'_1<2$, so the integral converges absolutely and \eqref{eq:neretin} gives \[ \int_X\varpi\wedge\xi=2\pi i\,n\,(-1)^{k_1+1}\varkappa,\qquad \varkappa=\frac{\Gamma(a_0-\theta_0)\,\Gamma(a_1+b'_1-\theta_1)\,\Gamma(\theta_\infty)}{\Gamma(a_0+a_1+b'_1-\theta_0-\theta_1)\,\Gamma(1-\theta_0)\,\Gamma(1-\theta_1)} . \] Hence $G(Q_0)=(-1)^{k_1}\varkappa/\mathcal{U}_0$. \emph{It is not zero.} Modulo $1$, $a_0-\theta_0\equiv(b-1)\theta_0\equiv\theta_0$, because $(b-2)\theta_0=-b'\theta_0\in\Z$; $a_1+b'_1-\theta_1\equiv(b+b'-1)\theta_1=\theta_1$; and $a_0+a_1+b'_1-\theta_0-\theta_1\equiv\theta_0+\theta_1=1-\theta_\infty$. None of $\theta_0$, $\theta_1$, $\theta_\infty$ is an integer, so every argument of $\Gamma$ in $\varkappa$ is a real number that is not an integer, and $\varkappa$ is a non-zero real number. This contradicts $G(Q_0)=0$. \end{proof} So under the hypotheses of part \ref{it:TII} the derivative along the Teichmüller disc of the relative period of $\eta_b$ over a rational class bounding $Q_1-Q_0$ is $(\zeta^{-b'}-1)(-1)^{k_1}\varkappa/\mathcal{U}_0$, a non-zero number given in closed form. As $b$ is a unit and $b\equiv2\pmod{n/g}$, the number $n/g$ must be odd: part \ref{it:TII} never speaks at a vertex above which the points are fixed by $T^{n/2}$. \Cref{cor:test}, applied to all units $b$, gives nothing beyond \cref{thm:relative}, on any rational triangle. For a unit $b\not\equiv2\pmod n$ with $H^{1,0}_b\ne0$ and $G_b\circ T=\zeta^{b-2}G_b$, the points above a vertex $x$ are fixed by $T^{\gcd(x,n)}$, so $G_b$ vanishes there unless $(2-b)x\equiv0\pmod n$; hence $G_b=0$ on $\Sigma$ unless $b$ satisfies the hypotheses of part~\ref{it:TII} at some vertex, and then $G_b$ is not constant on $\Sigma$, by the proof above. For $b\equiv2$ ($n$ odd, $\Delta$ obtuse) $G_2$ is constant above each vertex, so it is constant on $\Sigma$ exactly when $p\mid n$ and $q\mid n$, the case part~\ref{it:TI} leaves out; for even $n$, where $2$ is not a unit, only part~\ref{it:TII} remains. \begin{example}\label{ex:TII} For $\tri136{10}$, the vertex $6$ has $g=2$, $6\nmid10$ and $n/g=5$, and $b=7$ is a unit with $7\equiv2\pmod5$, $7\not\equiv2\pmod{10}$ and $\res7{10}+\res{21}{10}+\res{42}{10}=7+1+2=10$. With $6$ at $0$ and $1$ at $1$: $\theta_0=3/5$, $\theta_1=1/10$, $\theta_\infty=3/10$, $a_0=1/5$, $a_1=7/10$, $b'=-5$, $b'_1=1/2$, $k_0=-2$, $k_1=0$, and \[ \varkappa=\frac{\Gamma(-2/5)\,\Gamma(11/10)\,\Gamma(3/10)}{\Gamma(7/10)\,\Gamma(2/5)\,\Gamma(9/10)}. \] For $\tri138{12}$, the vertex $8$ has $g=4$ and $n/g=3$, $b=5$, $\res5{12}+\res{15}{12}+\res{40}{12}=12$, and with $8$ at $0$ and $1$ at $1$: $a_0=1/3$, $a_1=5/12$, $b'=-3$, $b'_1=3/4$, $k_0=-2$, $k_1=0$ and $\varkappa=\Gamma(-1/3)\Gamma(13/12)\Gamma(1/4)/(\Gamma(3/4)\Gamma(1/3)\Gamma(11/12))$. So neither triangle is a lattice triangle. \end{example} \section{Where the criterion is blind}\label{sec:list}\label{ssec:blind} \Cref{thm:list} below says exactly where the criterion is blind when the obtuse angle is at most $2\pi/3$, for every denominator. Only its part~\ref{it:list-all} for even $n$ is used, in the proof of \cref{thm:main}. The sets $S_x$, usable units and witnesses are defined in \cref{ssec:test}. We call the range $n/21$, exactly when $p=3$ and $n=3s$. \end{enumerate} \end{theorem} In particular, for a prime $n\ge5$ the window triangles without a witness are $\tri1{(n-3)/2}{(n+1)/2}n$ and, for $n=11$, $\tri236{11}$; for $n=2\ell$ with $\ell$ an odd prime they are $\tri1{\ell-2}{\ell+1}{2\ell}$ and, for $n=22$, $26$, $34$, also $\tri3{7}{12}{22}$; $\tri19{16}{26}$, $\tri57{14}{26}$; $\tri3{13}{18}{34}$. The window triangles without a witness and with $p=1$ are the members of $\famA$, the members of $\famC$ with $p=1$, and $\tri16{11}{18}$, $\tri18{15}{24}$, $\tri19{16}{26}$, $\tri1{10}{17}{28}$, $\tri1{12}{23}{36}$, $\tri1{16}{23}{40}$. The unit condition $(\mathrm{U})$ is a condition on the units in $S_p$ alone. The member $\tri35{12}{20}$ of $\listE$ satisfies every clause of the explicit condition in part \ref{it:list-B} with $\gcd(p,n)=1$ except $2\le p\le s/2$ ($p=3$, $s=4$, $n/s=5$). The family $\famA$ contains the two lattice triangles $\tri113{5}$ and $\tri114{6}$. Part \ref{it:list-all} is \cref{thm:E-list} for even $n$ and \cref{thm:O-list} for odd $n$. For all but finitely many denominators these rest on \cref{thm:E-large}, proved in \cref{app:even}, and on \cref{thm:O-large}, proved in \cref{ssec:list-odd}; both use the lemmas of \cref{ssec:common}, which hold for every denominator. \needspace{4\baselineskip} \begin{theorem}\label{thm:pq} Let $n$ be odd and let $\Delta=\tri pqrn$ be an obtuse triangle such that $p$ and $q$ divide $n$. Then $pq$ divides $n$; put $I=n/(pq)$ and $s=2r-n$. The triangle $\Delta$ has a witness if and only if $s>pI$, and the triangles with $s\le pI$ are exactly $\tri11{n-2}n$ with $n\ge5$ odd, $\tri135{9}$, $\tri159{15}$ and $\tri3{7}{11}{21}$. \end{theorem} The rest of this section proves \cref{thm:list,thm:pq} and, in \cref{ssec:AB}, applies \cref{thm:relative} to $\famA$ and $\famB$. Throughout, $\Delta=\tri pqrn$ is a window triangle unless said otherwise; then $q\ge(p+q)/2=(n-s)/4$. \subsection{The region where \texorpdfstring{$s$}{s} divides \texorpdfstring{$n$}{n}}\label{ssec:list-B} In this subsection $n$ is arbitrary, $\Delta$ is a window triangle, $s$ divides $n$ and $d=n/s$ is odd. Since $s\le n/3$, $d\ge3$. As $d$ is odd, \[ r=\frac{n+s}2=s\,\frac{d+1}2,\qquad p+q=n-r=s\,\frac{d-1}2 . \] \begin{proposition}\label{prop:region} Let $\Delta$ be a window triangle with $s\mid n$ and $d=n/s$ odd. \begin{enumerate}[label=\textup{(\alph*)}] \item\label{it:region-r} $S_r$ contains no unit modulo $n$. \item\label{it:region-q} If $2p\le s$, a unit $a$ with $\res{ap}n<2p$ lies in $S_q$ if and only if $a\not\equiv2\pmod {d}$. \item\label{it:region-B} If $2\le p\le s/2$, then $\Delta$ has no witness if and only if it satisfies the unit condition $(\mathrm{U})$, that is, if and only if it lies in $\famB$. \item\label{it:region-A} If $p=1$, that is if $\Delta\in\famA$, then $\Delta$ has no witness. \end{enumerate} \end{proposition} \begin{proof} \ref{it:region-r} For every $a$, $ar=s\cdot a(d+1)/2\equiv s\,c_1\pmod{{d}s}$ with $c_1=\res{a(d+1)/2}{d}$, and $0\le sc_1\le s(d-1)1$. \ref{it:region-q} Let $a$ be a unit and $x=\res{ap}n<2p\le s$. Then $x\ge1$, since $n\nmid ap$. Next $a(p+q)=s\cdot a(d-1)/2\equiv s\,c\pmod n$ with $c=\res{a(d-1)/2}{d}$; as $(d-1)/2$ is invertible modulo $d$ ($2\cdot(d-1)/2\equiv-1$) and $d\nmid a$, $1\le c\le d-1$. Hence $aq\equiv sc-x\pmod n$, and since $1\le x2p-s(d-1-c). \] If $c=d-1$ this reads $x>2p$, which is false. If $c\le d-2$, then $2p-s(d-1-c)\le2p-s\le01$, $\Delta$ lies in $\famB$ if and only if $p=3$ and $d=3$; these are the triangles $\tri3{s-3}{2s}{3s}$ with $s\ge7$ and $3\nmid s$, and for them $S_3$ contains the unit $(2,1)$ of $\Z/n=\Z/3\times\Z/s$. \end{enumerate} So on the whole region, $\Delta$ lies in $\famB$ if and only if it satisfies $(\mathrm{B}_1)$ or $(p,d)=(3,3)$. \end{proposition} The proof of part \ref{it:B1-33} uses that $\gcd(p,q,r)=1$; for triples with a common factor $(\mathrm{U})$ can hold with $(p,d)\ne(3,3)$, for instance for $(p,q,r)=(2,2,8)$, $n=12$: $(\mathrm{U})$ holds, vacuously, since $S_2$ contains no usable unit, while $(p,d)=(2,3)$. The triangles that Larsen, Norton and Zykoski construct in the proof of \cite[Proposition~3.7]{LNZ}, to show that the criterion is blind for arbitrarily large $p$, satisfy $(\mathrm{B}_1)$ with $d=2p-1$ and with every number below $2p$ other than $p$ dividing $n$. Condition $(\mathrm{B}_1)$ is the exact form of the pattern they observed experimentally for $p\ne1,2,4$ \cite[p.~7124]{LNZ}. \begin{proof} \ref{it:B1-coprime} Let $p^{-1}$ be the inverse of $p$ modulo $n$. As $a\mapsto ap$ permutes $\Z/n$, $S_p=\{kp^{-1}:0\le k\le2p-1\}$, and $kp^{-1}$ is a unit if and only if $\gcd(k,n)=1$. Next $2kp^{-1}\equiv2\pmod n$ means $2k\equiv2p\pmod n$, that is $k\equiv p$ modulo $n/\gcd(2,n)$; for $1\le k\le2p-1$, $|k-p|1$. A common divisor of $p$ and $s$ divides $q=(p+q)-p$ and $r$, so $\gcd(p,s)=1$. Hence $h=\gcd(p,sd)$ is prime to $s$ and divides $d$; in particular $h$ is odd and $h\ge3$. Write $p=hp'$ and $n=hn'$; then $\gcd(p',n')=1$ and $n'=s\cdot d/h$. \emph{Step 1.} Every $a\equiv1\pmod{n'}$ lies in $S_p$: $\res{ap}n=h\res{ap'}{n'}=hp'=p<2p$, since $p'1$, $(\mathrm{B}_1)$ fails, since it requires $p$ to be a prime not dividing $n$, and a prime with $\gcd(p,n)>1$ divides $n$; so part \ref{it:B1-33} is the statement. If $\gcd(p,n)=1$, $(p,d)=(3,3)$ is impossible, since it gives $3\mid\gcd(p,n)$; so part \ref{it:B1-coprime} is the statement. \end{proof} \subsection{Odd denominators with \texorpdfstring{$p$}{p} and \texorpdfstring{$q$}{q} dividing \texorpdfstring{$n$}{n}}\label{ssec:list-pq} In this subsection $n$ is odd and $\Delta=\tri pqrn$ is any obtuse triangle with $p\mid n$ and $q\mid n$; it need not lie in the window. For odd $n$ a unit $a$ is usable exactly when $a\ne1$, and $\res{2r}n=s$ as before. We prove \cref{thm:pq} at the end of this subsection, with the explicit witness of \cref{prop:pq-wit}, which is \cref{prop:pq-crit}(i) together with \cref{prop:pq-t}. \needspace{4\baselineskip} \begin{proposition}\label{prop:pq-wit} In \cref{thm:pq}, if $s>pI$, a witness is the residue $a$ modulo $n$ with $a\equiv1\pmod{pI}$ and $a\equiv1+I(q+t)/2\pmod{qI}$, where $t$ is any odd integer with $|t|\,pIn$ is $2(p+q)pI$, there is an odd $t$ with $|t|\,pI0$, so \[ N\varphi(D)/D\ge(3D-7)\varphi(D)/(5D)>(3\varphi(D)-7)/5, \] and $\varphi(D)\ge2\cdot4^{k-1}\ge2^{k+1}$, as the least prime of $D$ is at least $3$ and the others at least $5$. Hence $N\varphi(D)/D>(6\cdot2^k-7)/5=2^k-1+(2^k-2)/5\ge2^k-1$, and the number of good $j$ is positive. \end{proof} The smallest triangles where $t=\pm1$ both fail are $\tri1{15}{89}{105}$, with $I=7$, where $t=3$ gives $a=64$, and $\tri1{21}{83}{105}$, with $I=5$, where $t=3$ gives $a=61$. \begin{proposition}\label{prop:pq-list} $s\le pI$ holds exactly for $\tri11{n-2}n$ with $n\ge5$ odd, $\tri135{9}$, $\tri159{15}$ and $\tri37{11}{21}$. \end{proposition} \begin{proof} $s\le pI$ reads $pI(q-1)\le2(p+q)$, and the triangle is obtuse: $2(p+q)4$, and the triangle is $\tri11{I-2}I$ with $I=n\ge5$. If $q\ge3$ and $p=1$: $I(q-1)\le2q+22$, so $I\ge3$; the left half gives $I\le2+4/(q-1)$, so $q\le5$. For $q=3$, $I\le4$, so $I=3$ and the triangle is $\tri135{9}$ ($s=1\le3$). For $q=5$, $I\le3$, so $I=3$ and the triangle is $\tri159{15}$ ($s=3\le3$). If $p\ge3$: $p$ and $q$ are coprime and odd with $p\le q$, so $q\ge p+2$ and $2(p+q)\le4(q-1)$. Then $pI(q-1)\le4(q-1)$, so $p=3$ and $I=1$, and the two inequalities read $3(q-1)\le6+2q<3q$: $6pI$, \cref{prop:pq-t,prop:pq-crit} give a witness. If $s\le pI$, there is no odd $t$ with $|t|\,pIN_0(m)=P_2(31J^2-1)$. If a window triangle $\tri pqrn$ has no witness, then it lies in $\famA$, in $\famB$ or in $\famC$. \end{theorem} \emph{The idea of the proof.} The proof is in \cref{app:even}; the table at the end of this subsection lists its cases and the results that settle them. Suppose that $\Delta$ has no witness. Minkowski's theorem and Jacobsthal's function give $p\le7J^2-1$, and so $7L\le3n$ (Step~1). Step~2 shows that $S_r$ contains no usable unit. The odd elements of $S_r$ fall into rows of $g_{\varsigma}=\gcd(\varsigma,m)$ residues spaced $n/g_{\varsigma}$ apart (\cref{prop:E-Sr}); for $g_{\varsigma}\ge2$, outside a few cases treated apart, the units of a row are spread along it and a usable one lies in $S_q$, a witness, and for $g_{\varsigma}=1$ and $r$ even a witness comes from doubling. For $g_{\varsigma}=1$ and $r$ odd, the products of $r^{-1}$ with the powers of an odd prime $u\nmid m$ with $uLN_0(m)$, and this is where the constant $31=7+24$ comes from. Once $S_r$ has no usable unit, either $\varsigma\mid m$ with $m/\varsigma$ odd, or $\varsigma=1$ with $m$ even (\cref{prop:E-Sr}\ref{it:E-Sr-none}), and Step~3 finds a witness in $S_p\cap S_q$, from the multiplier chain and the rows of $S_p$, for every such triangle outside $\famA$, $\famB$ and $\famC$. The corresponding threshold of Larsen, Norton and Zykoski above $2\pi/3$ is the hypothesis $24j(n)^2<\sqrt n$ of \cite[Lemma~7.9]{LNZ}; here the multiplier chain replaces $\sqrt n$ by $n/P_2(m)$. \needspace{5\baselineskip} \begin{theorem}\label{thm:E-list} Let $n$ be even. A window triangle of denominator $n$ has no witness if and only if it lies in $\famA$, $\famB$ or $\famC$, or is a member of $\listE$ with even denominator (\cref{tab:E}). \end{theorem} \begin{proof} There is no window triangle for $n=2,4$. \emph{Only if.} By \cref{prop:region}\ref{it:region-B}, a triangle without witness with $s\mid n$, $n/s$ odd and $2\le p\le s/2$ lies in $\famB$; so it is enough to show that a window triangle of even denominator without witness lies in $\famA$ or $\famC$, or has $s\mid n$, $n/s$ odd and $2\le p\le s/2$, or is a member of $\listE$ with even denominator. If $m>N_0(m)$, this is \cref{thm:E-large}. The $m\ge3$ with $m\le N_0(m)$ are $2{,}567$ values, the largest $69{,}615$ (\cref{ssec:exceptional}, where the list is proved complete). By \cref{lem:G-box}, which has no hypothesis on $m$, a window triangle of denominator $2m$ without witness has $pq\le4mJ^2$. For the $2{,}567$ values of $m$ an enumeration of the window triangles of denominator $2m$ with $pq\le4mJ^2$ finds that each of them without witness lies in $\famA$ or $\famC$, or has $s\mid n$, $n/s$ odd and $2\le p\le s/2$, or is a member of $\listE$ (\cref{ssec:enumerations}). \emph{If.} For $\famA$ and $\famB$ this is \cref{prop:region}, and for $\famC$ \cref{prop:C-if} below. The members of $\listE$ have no witness by a direct check of their units (\cref{ssec:enumerations}). \end{proof} \begin{proposition}\label{prop:C-if} No member of $\famC$ has a witness. \end{proposition} \begin{proof} Here $4\mid n$, $s=2$, $r=n/2+1$, $q=n/2-1-p$, and $p\le q$ forces $n\ge4p+2$. \emph{$S_r$.} $r$ is odd and $\gcd(r,n)$ divides $2r-n=2$, so $r$ is a unit and $S_r=\{a:ar\equiv0\text{ or }1\pmod n\}=\{0,r^{-1}\}$. As $r^2=n^2/4+n+1\equiv1\pmod n$, $r^{-1}=r=n/2+1$, and $2r\equiv2$: $S_r$ has no usable unit. \emph{$S_p$.} For $p=1$, $S_1=\{0,1\}$. For $p=2$: $\res{2a}n$ is even, so $\res{2a}n<4$ means $2a\equiv0$ or $2$, that is $a\in\{0,n/2,1,n/2+1\}$; $0$ and $n/2$ are not units, and $1$ and $n/2+1$ are not usable. For $p=4$ and $n\equiv4\pmod8$: $\res{4a}n$ is a multiple of $4=\gcd(4,n)$, so $\res{4a}n<8$ means $4a\equiv0$ or $4$, that is $a\equiv0$ or $1\pmod{n/4}$; the eight candidates are $jn/4$ and $jn/4+1$, $j=0,\dots,3$. The $jn/4$ are not units, as $n/4>1$ divides them and $n$; $n/4+1$ and $3n/4+1$ are even, as $n/4$ is odd; and $1$, $n/2+1$ are not usable. So $S_p$ has no usable unit. A witness lies in two of the three sets, hence in $S_p$ or $S_r$: impossible. \end{proof} The statement on $S_p$ is \cite[Proposition~3.5]{LNZ} for $x=1,2,4$. \begin{remark}\label{rem:half-turn} Let $4\mid n$ and $u=1+n/2$. Then $u$ is a unit, $2u\equiv2$ and $u^2\equiv1\pmod n$, so $S_{ux}=uS_x$ and multiplication by $u$ preserves usable units. If $\varsigma$ is odd, exactly one of $p$, $q$ is even, say $e$, the other $o$ is odd, and $\res{up}n$, $\res{uq}n$, $\res{ur}n$ are $e$, $o+n/2$, $\varsigma$ in some order, with sum $n$: a triangle $u\Delta$ which has a witness if and only if $\Delta$ has one, and whose unfolding is the same Riemann surface as that of $\Delta$ with the eigenform of character $u$. The members $\tri2{m-3}{m+1}{2m}$ ($m\ge6$) and $\tri4{m-5}{m+1}{2m}$ ($m\equiv2\bmod4$, $m\ge10$) of $\famC$ are sent to the triangles $\tri12{n-3}n$ and $\tri14{n-5}n$ of families (2) and (3) of \cref{thm:lnz41}; $\tri349{16}$, $\tri58{15}{28}$, $\tri89{23}{40}$ to $\tri14{11}{16}$, $\tri18{19}{28}$, $\tri38{29}{40}$ of item (7). \end{remark} \begin{proof}[Proof of \cref{thm:list}] Part \ref{it:list-all} is \cref{thm:O-list} for odd $n$ ($\famC$ is empty for odd $n$, and the members of $\listE$ with odd denominator are the five of \cref{thm:O-list}) and \cref{thm:E-list} for even $n$. Part \ref{it:list-B} is \cref{prop:region}\ref{it:region-B}, and its explicit form is \cref{prop:B1}. In the sentence after \cref{thm:list}, the prime case is proved at the end of \cref{ssec:list-odd}, and the cases $n=2\ell$ and $p=1$ follow from part \ref{it:list-all}. For $n=2\ell$, $\ell$ an odd prime: $\famC$ needs $4\mid n$; a member of $\famB$ has $s\mid2\ell$ with $n/s$ odd and $s\ge4$, so $s=2\ell>n/3$, which is impossible; $\famA$ needs $s=2$, giving $\tri1{\ell-2}{\ell+1}{2\ell}$; and the members of $\listE$ with denominator twice an odd prime are $\tri37{12}{22}$, $\tri19{16}{26}$, $\tri57{14}{26}$ and $\tri3{13}{18}{34}$. For $p=1$: a member of $\famB$ has $p\ge2$, and the members of $\listE$ with $p=1$ are the six listed there. \end{proof} \begin{center}\small \begin{tabular}{@{}l>{\raggedright}p{4.3cm}>{\raggedright\arraybackslash}p{6.3cm}@{}} \toprule case & condition & settled by\\ \midrule Step 1 & every $\Delta$ without witness & \cref{cor:E-small}: $p\le7J^2-1$; hence $7L\le3n$\\ \multicolumn{3}{@{}l}{Step 2: $S_r$ contains a usable unit, $g_{\varsigma}=\gcd(\varsigma,m)$; a witness is found, or $m\le N_0(m)$}\\ (2a) & $\varsigma\mid m$ & $\varsigma\le8$: \cref{lem:E-smallsigma}; $\varsigma\ge9$: \cref{lem:G-lifts4}\\ (2b) & $\varsigma\nmid m$, $g_{\varsigma}\ge2$ & $g_{\varsigma}\ge4$: \cref{lem:G-lifts4}; $g_{\varsigma}=2$, or $g_{\varsigma}=3$ with $9\mid m$ or $3L\le n$: \cref{lem:G-rows}(iv); $g_{\varsigma}=3$, $9\nmid m$, $3L>n$: \cref{lem:E-three}\\ (2c) & $g_{\varsigma}=1$, $m$ and $\varsigma$ odd & $\varsigma\ge9$: \cref{lem:E-doubling}; $\varsigma\le7$: \cref{lem:E-smallsigma}\\ (2d) & $g_{\varsigma}=1$, $r$ odd & \cref{cor:G-base}\ref{it:G-base-r} and \cref{lem:E-dirichlet}: $m\le N_0(m)$\\ \multicolumn{3}{@{}l}{Step 3: $S_r$ has no usable unit: (i) $\varsigma\mid m$ with $m/\varsigma$ odd, or (ii) $\varsigma=1$ with $m$ even}\\ & (i), $p\le\varsigma$ & $\Delta$ lies in $\famA$ if $p=1$, in $\famB$ if $p\ge2$ (\cref{prop:region}\ref{it:region-B})\\ & (ii), $p\in\{1,2\}$, or $p=4$ and $m\equiv2\pmod4$ & $\Delta$ lies in $\famC$\\ \multicolumn{3}{@{}l}{\quad the other triangles, $p>\varsigma$, $p\ge3$, $g=\gcd(p,m)$: a witness in $S_p\cap S_q$, except in (3d)}\\ (3a) & $g=1$ & \cref{cor:G-base}\ref{it:G-base-p}\\ (3b) & $g\ge2$, $p/g\ge3$ & $g\ge4$: \cref{lem:G-lifts4}; $g\in\{2,3\}$: \cref{lem:G-rows}(iv)\\ (3c) & $g\ge3$, $p/g\le2$ & \cref{lem:G-rowone}, with \cref{lem:jacobsthal}\\ (3d) & $g=2$, $p/g\le2$ & $p=4$, $m\equiv2\pmod4$: in (ii) $\Delta$ lies in $\famC$, and (i) cannot occur\\ \bottomrule \end{tabular} \end{center} \subsection{The families \texorpdfstring{$\famA$}{A} and \texorpdfstring{$\famB$}{B}}\label{ssec:AB} For a member of $\famA$ or $\famB$ write $n=sd$ with $d$ odd, $d\ge3$; then $r=s(d+1)/2$ and $p+q=s(d-1)/2$. Since $\gcd((d+1)/2,d)=1$, $\gcd(r,n)=s$: the $s$ points above the obtuse vertex are zeros of $\omega$ ($r\nmid n$), and $n/\gcd(r,n)=d$. Also $\gcd(p,q,r)=\gcd(p,s)$, since $(d-1)/2$ and $(d+1)/2$ are coprime. \begin{theorem}\label{thm:AB} \begin{enumerate}[label=\textup{(\alph*)}] \item\label{it:AB-A} Part \ref{it:TI} or part \ref{it:TII} of \cref{thm:relative} applies to every member of $\famA$ other than $\tri113{5}$, $\tri114{6}$ and $\tri135{9}$. \item\label{it:AB-B} Part \ref{it:TII} of \cref{thm:relative} applies to every member of $\famB$, at the obtuse vertex. \item\label{it:AB-exc} Neither part applies to $\tri113{5}$, $\tri114{6}$ or $\tri135{9}$. \end{enumerate} Consequently no member of $\famA\cup\famB$ other than $\tri113{5}$, $\tri114{6}$ and $\tri135{9}$ is a lattice triangle; the first two are, by Veech's theorem; the third is not, by \cref{thm:finite}. \end{theorem} \begin{proof} \emph{A unit $a\equiv2\pmod {d}$ that is small on $p$ gives a holomorphic form.} Let $a$ be a unit with $a\equiv2\pmod {d}$ and $x=\res{ap}n1$, $r\nmid n$ and $n/s=d$, part \ref{it:TII} applies at the vertex $r$ with $b=a$. \ref{it:AB-A}, $n$ odd. Here $s$ is odd. Part \ref{it:TI} applies unless $q\mid n$, as $p=1$. Since $q=s(d-1)/2-1$ is prime to $s$, $q\mid n$ means $q\mid d$; as $d$ is odd, either $q=d$ or $q\le d/3$. If $q=d$, then $s(d-1)=2(d+1)$, $s=2+4/(d-1)$, and $s$ odd forces $d=5$, $s=3$: the triangle $\tri159{15}$. If $q\le d/3$, then $s(d-1)\le 2d/3+2$, which forces $s=1$ and $d\le9$, with $q=(d-3)/2$: $d=5$ gives $\tri113{5}$, $d=9$ gives $\tri135{9}$, $d=7$ gives $q=2\nmid7$, and $d=3$ gives no triangle. On $\tri159{15}$ part \ref{it:TII} applies at the vertex $9$ ($\gcd(9,15)=3$) with $b=7$: $\res7{15}+\res{35}{15}+\res{63}{15}=7+5+3=15$. \ref{it:AB-A}, $n$ even. Here $s$ is even. Take $b=2-kd$ with $1\le k\le s-1$; then $b\equiv2\pmod {d}$, $b\not\equiv2\pmod n$, and $\res b n=n+2-kd$, which is less than $n-s$ exactly when $kd>s+2$; then $(\ast)$ gives $H^{1,0}_b\ne0$, once $b$ is a unit. As $\gcd(kd-2,d)=\gcd(2,d)=1$, $b$ is a unit exactly when $\gcd(kd-2,s)=1$, that is, when $k$ is odd and $k\not\equiv2d^{-1}\pmod\ell$ for every odd prime $\ell\mid s$ with $\ell\nmid d$ (for $\ell\mid d$, $kd-2\equiv-2$). Call such $k$ \emph{valid}. Validity depends only on $k$ modulo $s^*$, the product of the primes dividing $s$ and not $d$ (so $2\mid s^*$); by the Chinese remainder theorem valid classes exist modulo every divisor of $s^*$, and $k=s$ is not valid, as $s$ is even. We need a valid $k$ with $(s+2)/d4$ is needed; every $d\ge5$ works, and $(s,d)=(2,3)$ is $\tri114{6}$. \item $s\ge4$ and $s^*s/2\ge(s+2)/3\ge(s+2)/d$. \item $s=s^*>2$, so $s$ is squarefree and prime to $d$: let $\ell$ be the least odd prime dividing $s$, and $k_1\in[1,s/\ell]$ a class valid for the primes of $s/\ell$. The $\ell$ numbers $k_1+js/\ell$, $0\le j<\ell$, are valid for the primes of $s/\ell$ and run over all classes modulo $\ell$, of which exactly one is not valid; so each class modulo $s/\ell$ valid for its primes contains $\ell-1$ valid classes modulo $s$, and $j=\ell-1$ or $j=\ell-2$ gives a valid $k>(\ell-2)s/\ell$, with $ks/3\ge(s+2)/5$; if $\ell\ge5$, then $k>3s/5\ge(s+2)/3$. \end{itemize} So part \ref{it:TII} applies at the vertex $r$ ($\gcd(r,n)=s>1$, $n/s=d$) to every member of $\famA$ with $n$ even except $\tri114{6}$. \ref{it:AB-exc} On $\tri113{5}$ both $p$ and $q$ divide $n$ and every $\gcd(x,n)$ is $1$. On $\tri114{6}$, $n$ is even, and at the vertex $4$ ($\gcd(4,6)=2$) the only unit $b\equiv2\pmod3$ with $b\not\equiv2\pmod6$ is $b=5$, for which $\res56+\res56+\res{20}6=12\ne6$. On $\tri135{9}$ the numerators $1$ and $3$ divide $9$, and $\gcd(5,9)=1$. \end{proof} \section{Cylinder families}\label{sec:cylinders} In this section we show by \cref{lem:cylinders} that several one-parameter families of obtuse triangles are not lattice triangles, exhibiting two parallel maximal cylinders whose moduli have an irrational ratio, or one maximal cylinder whose share of the area has a Galois conjugate outside $(0,1]$. The model of \cref{lem:model-branched} with $(p,q)=(4,1)$ is the unfolding drawn as a regular $n$-gon with four regular $(n/4)$-gons attached, in which Larsen, Norton and Zykoski find the two cylinders of family~(3) \cite[\S10]{LNZ}, after \cite[Figure~1]{Hooper}; the star of \cref{ssec:star} is the model in which Ward finds the two cylinders of \cite[Theorem~B]{Ward}. \Cref{thm:branched,thm:lnz5} use an irrational ratio of moduli in the first model for $p=2$, \cref{thm:star,thm:lnz4} one in the star, and \cref{thm:C4,thm:lnz6} a share of the area in the first model for $p=4$. \subsection{Two families with an angle \texorpdfstring{$2\pi/n$}{2pi/n}}\label{ssec:branched} In this subsection $\Delta=PQR$ has angles $p\psi$, $q\psi$, $r\psi$ at $P$, $Q$, $R$, where $\psi=\pi/n$ (the order of $p,q,r$ is not assumed). We use complex coordinates $P=0$, $Q=1$, $R=|PR|\,e^{ip\psi}$, where by the law of sines $|PR|=\sin q\psi/\sin r\psi$ and $|QR|=\sin p\psi/\sin r\psi$. Let $\Theta_k$ be the rotation by $2k\psi$ ($k\in\Z/n$) and $\kappa$ complex conjugation. The dihedral group $\mathcal D=\{\Theta_k,\Theta_k\kappa\}$ is generated by the linear parts of the reflections in the sides, $\kappa_{PQ}=\kappa$, $\kappa_{PR}=\Theta_p\kappa$, $\kappa_{QR}=\Theta_{-q}\kappa$ (the reflection in a line of direction $\phi$ is $z\mapsto e^{2i\phi}\bar z$). \needspace{4\baselineskip} \begin{lemma}\label{lem:model-branched} Assume $p\mid n$, $n/p\ge 3$, $p\chi\},\qquad \iota(i)=n-i . \] The map $\iota$ preserves $\mathcal I_{\chi}$. For an odd $i$ with $-n\chi>0$ already forces $0-n$. A value of $\chi$ is \emph{critical} if it equals $\sin(i\psi)$ for some integer $i$. A horizontal leaf that meets no vertex and enters the ears around an apex $O$ through a base, at height $|QR|\chi$ above $O$, turns clockwise about $O$, crossing one leg after another while the far end $v(j)$ of the next leg lies above it, that is, while $\sin(j\psi)>\chi$, and then leaves through a base. A chain of $\mathcal I_{\chi}$ is the set of indices of the legs crossed in one such passage (\cref{lem:passage}, \cref{fig:ears}). If the polygon the leaf crosses next is symmetric, it re-enters the ears at the next apex at the same height and crosses the legs of $\iota(\mathcal{W})$, and if the polygon after that is symmetric too, it is back at its starting point (\cref{lem:symmetric,lem:reentry}; after one passage if $\iota(\mathcal{W})=\mathcal{W}$). Since the chains stay the same while $\chi$ moves in an open interval, these closed leaves, all of one length, fill an annulus whose upper boundary leaf runs into a point over $R$ at which the chain changes, and whose lower boundary leaf runs into such a point or, when $a(\mathcal{W})=0$, into the apex $O$. It is a maximal cylinder when those points are singular (\cref{prop:chain}). \begin{lemma}\label{lem:passage} Let $y>0$ be a height with $\chi=y/|QR|$ not critical, and suppose a horizontal leaf crosses the base of $E_k$ from $G_{k\bmod p}$ into $E_k$ at height $y$ above $O=O_k$. Let $i_1$ be the representative of $2k+n-q$ in $(0,2n)$, so that the clockwise leg of $E_k$ has direction $i_1\psi$. Then $00\}$, along which the polar angle decreases strictly as $x$ increases and stays in $(0,\pi)$; since this range of angles is shorter than $\pi$, the development meets each ear at most once and is injective. Inside an ear (the triangle $O,v(j),v(j+2q)$, with the sector $[j\psi,(j+2q)\psi]$) the line enters through one side and leaves through another. In $E_k$ the leaf enters through the base; the other side it crosses has smaller polar angle, so it is the clockwise leg, of direction $i_1\psi$. The line separates $v(i_1)$ (above it) from $O$ (below), so $|QR|\sin(i_1\psi)>y$ and $i_1\psi\in(0,\pi)$; it also separates $v(i_1+2q)$ from $v(i_1)$, so $\sin((i_1+2q)\psi)<\chi$. After crossing a leg of direction $j\psi$ the leaf is in the next ear clockwise, with sector $[(j-2q)\psi,j\psi]$, entered through its counterclockwise leg; it leaves through the clockwise leg, of direction $(j-2q)\psi$, if $v(j-2q)$ lies strictly above the line, that is, if $j-2q\in \mathcal I_{\chi}$, and through the base otherwise ($\chi$ is not critical, so the leaf meets no vertex). The ear left through its base has sector $[(i_N-2q)\psi,i_N\psi]$, so $2k'+n+q\equiv i_N\pmod{2n}$. \end{proof} \needspace{4\baselineskip} \begin{lemma}\label{lem:symmetric} The polygon $G_j$ is symmetric under $(x,y)\mapsto(-x,y)$ about its centre if and only if $2j\equiv n/2\pmod p$; for $p=2$ this holds for both polygons if and only if $4\mid n$. If $G_j$ is symmetric, a horizontal leaf entering $G_j$ through side $k'$ leaves it through side $k''=n/2-k'$, its heights above $O_{k'}$ and above $O_{k''}$ are equal, and in the chart of $G_j$ we have $O_{k''}-O_{k'}=(-2\cos 2k'\psi,0)$. \end{lemma} \begin{proof} The reflection maps the direction $\phi$ to $\pi-\phi$, so the outward normal $2k\psi$ to $2(n/2-k)\psi$, and the set $\{k\equiv j\pmod p\}$ is preserved if and only if $n/2-j\equiv j\pmod p$. For $p=2$ and $j=0,1$ this says that $n/2$ is even. A horizontal line meets a symmetric convex polygon in a segment symmetric about the vertical axis, whose ends lie on mirror sides $k'$ and $n/2-k'$. Finally $e^{2ik''\psi}=e^{i(\pi-2k'\psi)}=-\cos2k'\psi+i\sin2k'\psi$. \end{proof} \needspace{4\baselineskip} \begin{lemma}\label{lem:reentry} In \cref{lem:passage}, if $G_{k'\bmod p}$ is symmetric, the leaf next enters $E_{k''}$, $k''=n/2-k'$, through its base at the same height $y$ above $O_{k''}$, and the first leg it crosses there has direction $(n-i_N)\psi$; so its next passage traverses $\iota(\mathcal{W})$, where $\mathcal{W}$ is the chain of the first passage. \end{lemma} \begin{proof} We have $2k''+n-q=2n-2k'-q\equiv 2n-(i_N-n-q)-q=3n-i_N\equiv n-i_N\pmod{2n}$, and $00$, and the apex $O$ if $a(\mathcal{W})=0$. If moreover $r\nmid n$, then $Z(\mathcal{W})$ is a maximal cylinder if $a(\mathcal{W})>0$, and also if $a(\mathcal{W})=0$ and $q\nmid n$. \end{proposition} \begin{proof} For $a(\mathcal{W})<\chi0$). An open annulus of closed leaves is a maximal cylinder if each of its boundary components contains a singular point. If $r\nmid n$, the $\gcd(r,n)$ points over $R$ have cone angle $2\pi r/\gcd(r,n)>2\pi$, and if $q\nmid n$ so have the points over $Q$; this gives the last sentence of the statement. \end{proof} \begin{lemma}\label{lem:chain-p2} If $p=2$ and $4\mid n$, then \eqref{eq:chain-sym} holds for every $\mathcal{W}$, and $r\nmid n$. \end{lemma} \begin{proof} Both sides of \eqref{eq:chain-sym} are even, as $4\mid n$. Next, $r=n-2-q$ is odd, as $n$ is even and $q$ is odd; $r>n/2-2$, as $q0$ the chain of $\mathcal I_{\chi}$ with top $2n-i_{\rm top}-2q$ is $\mathcal{W}$ or $\iota(\mathcal{W})$. Then $Z(\mathcal{W})\cup\gamma_0\cup\hat\kappa(Z(\mathcal{W}))$, where $\gamma_0$ is the leaf at height $0$ through $O$, is a maximal cylinder of circumference $\mathcal L(\mathcal{W})$ and height $2|QR|\,b(\mathcal{W})$. \end{proposition} \begin{proof} The base of $E_k$ has its ends at heights $|QR|\sin(i_{\rm top}+2q)\psi<0$ and $|QR|\sin(i_{\rm top}\psi)>0$, so it contains a point at height $0$ in its interior. The leaf $\gamma_0$ from that point is the limit of the leaves of $Z(\mathcal{W})$; it meets no point over $R$ (their heights above an apex are $|QR|\sin(j\psi)\ne0$ with $j$ odd, and polygon chords meet vertices only at ends of bases of ears), only regular points over $Q$, and possibly centres of polygons, points over $P$, which are regular since $p\mid n$; so it is a regular closed leaf, of length $\mathcal L(\mathcal{W})$. The leaves just below $\gamma_0$ cross the base of $E_k$ at small negative heights; they are the $\hat\kappa$-images of leaves entering $E_{-k}$ at small positive heights, whose first leg has index $2n-i_{\rm top}-2q$, so by hypothesis they are leaves of $Z(\mathcal{W})$, and the leaves just below $\gamma_0$ lie in $\hat\kappa(Z(\mathcal{W}))$. The annuli $Z(\mathcal{W})$ and $\hat\kappa(Z(\mathcal{W}))$ are disjoint: a leaf of $Z(\mathcal{W})$ enters each ear of its period at the same positive height above the apex, while its image enters the image ears at the negative height. The union is an annulus foliated by closed leaves of length $\mathcal L(\mathcal{W})$, of height $2|QR|\,b(\mathcal{W})$, whose two boundary components contain points over $R$, which are singular since $r\nmid n$. \end{proof} In the irrationality steps below we use the standard fact that, for $N\ge3$ and $\gcd(j,N)=1$, $\cos(2\pi j/N)$ has degree $\varphi(N)/2$ over $\Q$. \needspace{4\baselineskip} \begin{theorem}\label{thm:branched} For every even $m\ge 6$ the triangle $\tri2{m-3}{m+1}{2m}$ is not a lattice triangle. Let $\psi=\pi/2m$, so that $|QR|=\sin2\psi/\sin(m+1)\psi$. \begin{enumerate} \item If $m\ge 8$, in the direction of the side joining the vertices of angles $2\psi$ and $(m-3)\psi$, with that side of length $1$, the unfolding has two maximal cylinders, of circumference $4\sin5\psi\cos3\psi$ and height $|QR|(\sin5\psi-\sin\psi)$, and of circumference $4\cos3\psi\cos\psi$ and height $|QR|\cos\psi$; the ratio of their moduli is $1-\sin\psi/\sin5\psi$, which is irrational. \item If $m=6$, the triangle is $\tri237{12}$; in the same direction it has a maximal cylinder of circumference $2$ and height $2|QR|\sin3\psi$, and the first cylinder of (a); the ratio of their moduli is $(\sqrt3-1)/2$. \end{enumerate} \end{theorem} These triangles form the part of $\famC$ with $p=2$. For $m=4$ the triangle is $\tri125{8}$, a lattice triangle. For $m=6$ the statement is Vorobets's \cite[Theorem~4.5]{Vorobets}, proved there with two horizontal cylinders. For odd $m$ the integers $2$, $m-3$, $m+1$ have greatest common divisor $2$, and the triangle is $\tri1{(m-3)/2}{(m+1)/2}m$, which is treated in \cref{sec:list,sec:finite}. \begin{proof} Let $m\ge6$ be even, $n=2m$, $p=2$, $q=m-3$, $r=m+1$; then $4\mid n$, the hypotheses of \cref{lem:model-branched} hold, and by \cref{lem:chain-p2} so does \eqref{eq:chain-sym}, for every $\mathcal{W}$, and $r\nmid n$. The vertex $P$ gives two regular points, $R$ one point of cone angle $2\pi(m+1)$ (as $\gcd(m+1,2m)=1$), and $Q$ $\gcd(m-3,6)\in\{1,3\}$ points of cone angle $2\pi(m-3)/\gcd(m-3,6)$, which are regular exactly for $m\in\{4,6\}$. By \eqref{eq:genus}, $2\,\mathrm{genus}(X)-2=2m-3-\gcd(m-3,6)\ge6$. Since $m\psi=\pi/2$, $\sin(m+j)\psi=\cos j\psi$. The chain $\mathcal{W}_c=\{2m-5\}$: its neighbours in the progression are $2m-5+2q=2n-11$, with $\sin(2n-11)\psi=-\sin11\psi<0$ as $n\ge12$, and $2m-5-2q=1$. So $a(\mathcal{W}_c)=\sin\psi$ and $b(\mathcal{W}_c)=\sin(2m-5)\psi=\sin5\psi>\sin\psi$; $\iota(\mathcal{W}_c)=\{5\}\ne \mathcal{W}_c$, and $\mathcal L(\mathcal{W}_c)=2[\cos(m-2)\psi-\cos(3m-8)\psi]=2[\sin2\psi+\sin8\psi]=4\sin5\psi\cos3\psi$. By \cref{prop:chain}, $Z_c=Z(\mathcal{W}_c)$ is a maximal cylinder of height $|QR|(\sin5\psi-\sin\psi)$ and modulus $\mu_c=|QR|(\sin5\psi-\sin\psi)/(4\sin5\psi\cos3\psi)$. Part (a): let $m\ge8$ and $\mathcal{W}_d=\{m+1\}$. Its neighbours are $m+1+2q=3m-5$, with $\sin(3m-5)\psi=-\cos5\psi<0$, and $m+1-2q=7-m\in(-n,0)$. So $a(\mathcal{W}_d)=0$, $b(\mathcal{W}_d)=\cos\psi$, $\iota(\mathcal{W}_d)=\{m-1\}\ne \mathcal{W}_d$, and $\mathcal L(\mathcal{W}_d)=2[\cos4\psi-\cos(2m-2)\psi]=4\cos3\psi\cos\psi$. The points over $Q$ are singular, so $Z_d=Z(\mathcal{W}_d)$ is maximal, of modulus $\mu_d=|QR|/(4\cos3\psi)$. Then $\mu_c/\mu_d=1-\sin\psi/\sin5\psi$. With $x=2\cos2\psi=2\cos(\pi/m)$ we have $4\cos^2\psi=2+x$ and $\sin5\psi/\sin\psi=16\cos^4\psi-12\cos^2\psi+1=x^2+x-1$. If the ratio were rational, $x$ would have degree at most $2$ over $\Q$; but $[\Q(2\cos(2\pi/2m)):\Q]=\varphi(2m)/2$, and $\varphi(2m)\le4$ with $m\ge3$ only for $2m\in\{6,8,10,12\}$. So for even $m\ge8$ the ratio is irrational, and \cref{lem:cylinders}\ref{it:cyl-ratio} applies. Part (b): $m=6$, $n=12$, $q=3$, and the points over $Q$ are regular. Take $\mathcal{W}_a=\{3,9\}$: its neighbours $15$ and $-3$ have negative sines, so $a(\mathcal{W}_a)=0$, $b(\mathcal{W}_a)=\sin3\psi$, $\iota(\mathcal{W}_a)=\mathcal{W}_a$ and $\mathcal L(\mathcal{W}_a)=2$. The index $2n-i_{\rm top}-2q=24-9-6$ is $9=i_{\rm top}$, so \cref{prop:merge} gives a maximal cylinder $\hat Z_a$ of circumference $2$, height $2|QR|\sin3\psi$ and modulus $|QR|\sin3\psi$. Since $\sin6\psi=1$ and $\sin5\psi=\cos\psi$, \[ \frac{\mu_c}{\mu(\hat Z_a)}=\frac{1-\sin\psi/\sin5\psi}{4\sin3\psi\cos3\psi}=\frac{1-\tan(\pi/12)}2=\frac{\sqrt3-1}2, \] which is irrational, and \cref{lem:cylinders}\ref{it:cyl-ratio} applies. \end{proof} At $\tri125{8}$ ($m=4$) the chains $\{3\}$ and $\{5\}$ satisfy $a=b$ and give no cylinder. \needspace{4\baselineskip} \begin{theorem}\label{thm:lnz5} For every odd $q\ge3$ the triangle $\tri2q{3q+2}{(4q+4)}$ is not a lattice triangle. With $\psi=\pi/(4q+4)$ and $|QR|=\sin2\psi/\sin(3q+2)\psi$, in the direction of the side joining the vertices of angles $2\psi$ and $q\psi$, with that side of length $1$, the unfolding has two maximal cylinders, of circumference $4\cos\psi\sin(\pi/4+3\psi)$ and height $|QR|\sin\psi$, and of circumference $2\cos2\psi$ and height $|QR|\sin(\pi/4+\psi)$; the ratio of their moduli is $\tan\psi/(1+2\sin2\psi)$, which is irrational. \end{theorem} \begin{proof} Here $n=4q+4$, $m=n/2=2q+2$, so $4\mid n$ and $2q=m-2$; $q\psi=\pi/4-\psi$ and $(q+2)\psi=\pi/4+\psi$. The hypotheses of \cref{lem:model-branched} hold, and by \cref{lem:chain-p2} so do \eqref{eq:chain-sym} and $r\nmid n$. The points over $Q$ are singular, since $\gcd(q,4q+4)=1$ and $q\ge3$, and the point over $R$ is singular, since $\gcd(3q+2,4q+4)=\gcd(q+2,4)=1$. By \eqref{eq:genus}, $2\,\mathrm{genus}(X)-2=4q$. The chain $\mathcal{W}_T=\{1,2q+1,4q+1\}$: the sines are $\sin\psi$, $\cos\psi$ and $\sin3\psi$, so $b(\mathcal{W}_T)=\sin\psi$; the neighbours are $6q+1=3m-5$, with sine $-\cos5\psi<0$ as $n>10$, and $1-2q\in(-n,0)$; so $a(\mathcal{W}_T)=0$. Also $\iota(\mathcal{W}_T)=\{3,2q+3,4q+3\}\ne \mathcal{W}_T$, and $\mathcal L(\mathcal{W}_T)=2[\cos(q-1)\psi+\cos(q-3)\psi]=4\cos\psi\cos(q-2)\psi=4\cos\psi\sin(\pi/4+3\psi)$. So $Z_T=Z(\mathcal{W}_T)$ is maximal, with height $|QR|\sin\psi$. The chain $\mathcal{W}_S=\{q+2,3q+2\}$ satisfies $\iota(\mathcal{W}_S)=\mathcal{W}_S$; its neighbours $5q+2\in(n,2n)$ and $2-q<0$ have negative sines, so $a(\mathcal{W}_S)=0$, $b(\mathcal{W}_S)=\sin(q+2)\psi=\sin(\pi/4+\psi)$ and $\mathcal L(\mathcal{W}_S)=2\cos2\psi$. So $Z_S=Z(\mathcal{W}_S)$ is maximal, with height $|QR|\sin(\pi/4+\psi)$. Since $2\sin(\pi/4+3\psi)\sin(\pi/4+\psi)=\cos2\psi-\cos(\pi/2+4\psi)=\cos2\psi(1+2\sin2\psi)$, the ratio of the moduli is \[ \varrho=\frac{\tan\psi}{1+2\sin2\psi}=\frac{\sin2\psi}{(1+\cos2\psi)(1+2\sin2\psi)}. \] Let $\zeta=e^{2\pi i/n}$, so that $\cos2\psi,\sin2\psi\in\Q(\zeta)$ (as $i=\zeta^{n/4}$). Since $q$ is odd, $8\mid n$, and $n/2+1$ is odd and prime to $n$; so $\sigma_{n/2+1}:\zeta\mapsto\zeta^{n/2+1}=-\zeta$ is an automorphism of $\Q(\zeta)$. It fixes $i$ (as $n/4$ is even) and changes the signs of $\cos2\psi$ and $\sin2\psi$, so \[ \sigma_{n/2+1}(\varrho)=-\frac{\sin2\psi}{(1-\cos2\psi)(1-2\sin2\psi)}<0, \] because $\sin2\psi<1/2$ for $n>12$, while $\varrho>0$. So $\varrho$ is irrational, and \cref{lem:cylinders}\ref{it:cyl-ratio} applies. \end{proof} For $q=1$ the triangle is $\tri125{8}$ again. For odd $q\ge3$, $\tri2q{3q+2}{(4q+4)}$ is family~(5) of \cref{thm:lnz41}, and its obtuse angle exceeds $2\pi/3$. \subsection{Two families with an angle \texorpdfstring{$4\pi/n$}{4pi/n}: one cylinder and its share}\label{ssec:sign} We keep the notation of \cref{ssec:branched}, now with $p=4$; \cref{prop:chain,prop:merge} are stated for every $p$. \begin{theorem}\label{thm:C4} Let $m\equiv 2\pmod 4$ with $m\ge 10$, $n=2m$ and $\vartheta=2\pi/n=\pi/m$. The triangle $\tri4{m-5}{m+1}{2m}$ is not a lattice triangle. In the direction of the side joining its vertices of angles $4\pi/n$ and $(m-5)\pi/n$, its unfolding has a maximal cylinder $Z$ with \[ n\,\rho(Z)=4\cos\vartheta\,(\cos\vartheta-\cos4\vartheta)\quad(m\ge14),\qquad n\,\rho(Z)=8\cos\vartheta\,(\cos\vartheta-\cos4\vartheta)\quad(m=10), \] and $\sigma_\ell(\rho(Z))<0$ for $\ell=m/2-2$ if $m\ge 22$, and for $\ell=7,11,13$ if $m=10,14,18$. \end{theorem} These triangles form the part of $\famC$ with $p=4$. For $m=6$ the triangle is $\tri147{12}$, Hooper's lattice triangle; the proof below needs $n\ge 20$. \begin{proof} Here $p=4$, $q=m-5$, $r=m+1$, $\psi=\pi/n$, so that $m\psi=\pi/2$ and $\vartheta=2\psi$. The hypotheses of \cref{lem:model-branched} hold, $n$ is even and $q$ is odd; since $n/2=m\equiv2\pmod 4$, the polygon $G_j$ is symmetric if and only if $j$ is odd (\cref{lem:symmetric}). The vertex $R$ gives one point of cone angle $2\pi(m+1)$, since $\gcd(m+1,2m)=1$; it is singular, and $r\nmid n$, as \cref{prop:chain,prop:merge} require. The points over $Q$ are singular if and only if $q\nmid n$, that is, $(m-5)\nmid 10$, that is, $m\ne 6,10$. Let $\mathcal{W}=\{3,n-7\}$ (\cref{fig:model-p4}, left, for $m=14$), a progression with difference $n-10=2q$. Since $10\psi<\pi$ we have $\sin(n-7)\psi=\sin7\psi>\sin3\psi>0$. The neighbours are $(n-7)+2q=2n-17$, with $\sin(2n-17)\psi=-\sin17\psi<0$ because $17\psi<\pi$, and $3-2q=13-n<0$. So $\mathcal{W}$ is a chain of $\mathcal I_{\chi}$ for $0<\chi<\sin3\psi$, with $a(\mathcal{W})=0$ and $b(\mathcal{W})=\sin3\psi$, and $\iota(\mathcal{W})=\{7,n-3\}\ne \mathcal{W}$. The exit ears are given by $2k_1\equiv 3-n-q=8-3m$ and $2k_2\equiv 7-n-q=12-3m\pmod{4m}$, so $k_1\equiv m/2+4$ and $k_2\equiv m/2+6\pmod{n}$; both are odd, as $m/2$ is odd, so $2k_1\equiv2k_2\equiv2\equiv n/2\pmod 4$: this is \eqref{eq:chain-sym}, and \cref{prop:chain} applies, with \[ \mathcal L(\mathcal{W})=2\bigl[\cos(3-q)\psi-\cos(n-7+q)\psi\bigr]=2[\sin8\psi+\sin12\psi]=4\sin10\psi\cos2\psi, \] since $(3-q)\psi=8\psi-\pi/2$ and $(n-7+q)\psi=3\pi/2-12\psi$. \emph{The share.} The area of $X$ is $2n$ times the area $\frac12|PQ|\,|QR|\sin q\psi$ of $\Delta$, that is, $n|QR|\sin q\psi=n|QR|\cos5\psi$. If $m\ge 14$, the points over $Q$ are singular and $Z(\mathcal{W})$ is a maximal cylinder of height $|QR|\sin3\psi$, so \begin{multline*} n\,\rho=\frac{4\sin3\psi\,\sin10\psi\,\cos2\psi}{\cos5\psi}=8\sin3\psi\sin5\psi\cos2\psi\\ =4\cos2\psi\,(\cos2\psi-\cos8\psi)=4\cos\vartheta\,(\cos\vartheta-\cos4\vartheta). \end{multline*} If $m=10$, the points over $Q$ are regular, and we apply \cref{prop:merge} ($r\nmid n$ and $q\mid n$): $\sin(i_{\rm top}+2q)\psi=-\sin17\psi<0$, and $2n-i_{\rm top}-2q=17=n-3$ is the top of $\iota(\mathcal{W})$, which is the chain of $\mathcal I_{\chi}$ with top $17$ for $0<\chi<\sin3\psi$ (its neighbours $27$ and $-3$ have negative sines). So the maximal cylinder containing $Z(\mathcal{W})$ has the same circumference and height $2|QR|\sin3\psi$, and $n\rho=8\cos\vartheta(\cos\vartheta-\cos4\vartheta)$. \emph{The conjugates.} $n\rho$ is a polynomial in $\cos\vartheta$ with rational coefficients, so $\rho\in\Q(\zeta)^+$, and for a unit $\ell$ modulo $n$, $\sigma_\ell(n\rho)$ is the same expression with $\vartheta$ replaced by $\ell\vartheta$; it has the sign of $\sigma_\ell(\rho)$. (i) Let $m\ge 22$ and $\ell=m/2-2$. It is odd, and $\gcd(\ell,m/2)=\gcd(2,m/2)=1$, so $\ell$ is a unit modulo $2m$. As $\ell\vartheta=\pi/2-2\vartheta$, $\cos\ell\vartheta=\sin2\vartheta$ and $\cos4\ell\vartheta=\cos(2\pi-8\vartheta)=\cos8\vartheta$, so $\sigma_\ell(n\rho)=4\sin2\vartheta\,(\sin2\vartheta-\cos8\vartheta)$. Since $10\vartheta<\pi/2$, we have $0<2\vartheta<\pi/2-8\vartheta<\pi/2$, so $\sin2\vartheta<\cos8\vartheta$ and $\sigma_\ell(n\rho)<0$. (ii) $m=14$, $\ell=11$: $\cos(11\pi/14)=-\cos(3\pi/14)$ and $\cos(44\pi/14)=\cos(8\pi/7)=-\cos(\pi/7)$, so $\sigma_{11}(n\rho)=-4\cos(3\pi/14)\,(\cos(\pi/7)-\cos(3\pi/14))<0$. (iii) $m=18$, $\ell=13$: $\cos(13\pi/18)=-\cos(5\pi/18)$ and $\cos(52\pi/18)=\cos(8\pi/9)=-\cos(\pi/9)$, so $\sigma_{13}(n\rho)=-4\cos(5\pi/18)\,(\cos(\pi/9)-\cos(5\pi/18))<0$. (iv) $m=10$, $\ell=7$: $\cos(7\pi/10)=-\cos(3\pi/10)$ and $\cos(28\pi/10)=-\cos(\pi/5)$, so $\sigma_7(n\rho)=-8\cos(3\pi/10)\,(\cos(\pi/5)-\cos(3\pi/10))<0$. In each case \cref{lem:cylinders}\ref{it:cyl-sign} applies. \end{proof} \begin{figure} \centering \begin{tikzpicture}[x=1.55cm,y=1.55cm,line join=round,line cap=round] \fill[black!25,draw=black!25,line width=0.2pt] (3.2949,0.2225) -- (3.1507,0.3667) -- (3.0237,0.3667) -- (3.0567,0.2225) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (-0.3667,-0.8307) -- (-0.2225,-0.9749) -- (0.1896,-0.8307) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (1.9503,-1.5523) -- (1.5382,-1.6965) -- (1.8353,-1.6965) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (2.8047,-1.6965) -- (2.6897,-1.5523) -- (1.9503,-1.5523) -- (1.8353,-1.6965) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (3.1018,-1.6965) -- (2.6897,-1.5523) -- (2.8047,-1.6965) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (-0.1896,-3.1507) -- (0.2225,-3.2949) -- (0.3667,-3.1507) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (1.4893,0.3667) -- (1.3451,0.2225) -- (1.5833,0.2225) -- (1.6163,0.3667) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (1.6163,0.3667) -- (1.5833,0.2225) -- (3.0567,0.2225) -- (3.0237,0.3667) -- cycle; \draw[black!55,line width=0.3pt] (0.7677,0.3697) -- (0.1896,0.8307) -- (-0.5313,0.6662) -- (-0.8521,0.0000) -- (-0.5313,-0.6662) -- (0.1896,-0.8307) -- (0.7677,-0.3697) -- cycle; \draw[black!55,line width=0.3pt] (2.9862,0.5313) -- (2.3200,0.8521) -- (1.6538,0.5313) -- (1.4893,-0.1896) -- (1.9503,-0.7677) -- (2.6897,-0.7677) -- (3.1507,-0.1896) -- cycle; \draw[black!55,line width=0.3pt] (0.8521,-2.3200) -- (0.5313,-1.6538) -- (-0.1896,-1.4893) -- (-0.7677,-1.9503) -- (-0.7677,-2.6897) -- (-0.1896,-3.1507) -- (0.5313,-2.9862) -- cycle; \draw[black!55,line width=0.3pt] (3.1507,-2.1304) -- (2.6897,-1.5523) -- (1.9503,-1.5523) -- (1.4893,-2.1304) -- (1.6538,-2.8513) -- (2.3200,-3.1721) -- (2.9862,-2.8513) -- cycle; \draw[line width=0.5pt] (0.7677,-0.3697) -- (1.0000,0.0000) -- (0.7677,0.3697) -- (0.6235,0.7818) -- (0.1896,0.8307) -- (-0.2225,0.9749) -- (-0.5313,0.6662) -- (-0.9010,0.4339) -- (-0.8521,0.0000) -- (-0.9010,-0.4339) -- (-0.5313,-0.6662) -- (-0.2225,-0.9749) -- (0.1896,-0.8307) -- (0.6235,-0.7818) -- cycle; \draw[line width=0.5pt] (3.1507,-0.1896) -- (3.2949,0.2225) -- (2.9862,0.5313) -- (2.7539,0.9010) -- (2.3200,0.8521) -- (1.8861,0.9010) -- (1.6538,0.5313) -- (1.3451,0.2225) -- (1.4893,-0.1896) -- (1.5382,-0.6235) -- (1.9503,-0.7677) -- (2.3200,-1.0000) -- (2.6897,-0.7677) -- (3.1018,-0.6235) -- cycle; \draw[line width=0.5pt] (0.8521,-2.3200) -- (0.9010,-1.8861) -- (0.5313,-1.6538) -- (0.2225,-1.3451) -- (-0.1896,-1.4893) -- (-0.6235,-1.5382) -- (-0.7677,-1.9503) -- (-1.0000,-2.3200) -- (-0.7677,-2.6897) -- (-0.6235,-3.1018) -- (-0.1896,-3.1507) -- (0.2225,-3.2949) -- (0.5313,-2.9862) -- (0.9010,-2.7539) -- cycle; \draw[line width=0.5pt] (3.1507,-2.1304) -- (3.1018,-1.6965) -- (2.6897,-1.5523) -- (2.3200,-1.3200) -- (1.9503,-1.5523) -- (1.5382,-1.6965) -- (1.4893,-2.1304) -- (1.3451,-2.5425) -- (1.6538,-2.8513) -- (1.8861,-3.2210) -- (2.3200,-3.1721) -- (2.7539,-3.2210) -- (2.9862,-2.8513) -- (3.2949,-2.5425) -- cycle; \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-0.2946,-0.9028) -- (-0.0165,-0.9028); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (1.7442,-1.6244) -- (2.8958,-1.6244); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (0.0165,-3.2228) -- (0.2946,-3.2228); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (1.4172,0.2946) -- (3.2228,0.2946); \node at (0.8992,0.2583) {\scriptsize$19$}; \node at (0.8992,-0.2583) {\scriptsize$37$}; \node at (3.1392,0.4519) {\scriptsize\color{linkblue}$21$}; \node at (3.2541,-0.0517) {\scriptsize$39$}; \node at (0.6981,-1.6972) {\scriptsize$23$}; \node at (0.9222,-2.1625) {\scriptsize$41$}; \node at (2.8620,-1.5574) {\scriptsize\color{linkblue}$25$}; \node at (3.1841,-1.9613) {\scriptsize$43$}; \node at (0.3587,0.8641) {\scriptsize$27$}; \node at (0.7626,0.5420) {\scriptsize$45$}; \node at (2.4775,0.9222) {\scriptsize$29$}; \node at (2.9428,0.6981) {\scriptsize$47$}; \node at (-0.0517,-1.3859) {\scriptsize$31$}; \node at (0.4519,-1.5008) {\scriptsize$49$}; \node at (2.0617,-1.4208) {\scriptsize$33$}; \node at (2.5783,-1.4208) {\scriptsize$51$}; \node at (-0.4519,0.8192) {\scriptsize$35$}; \node at (0.0517,0.9341) {\scriptsize$53$}; \node at (1.6972,0.6981) {\scriptsize$37$}; \node at (2.1625,0.9222) {\scriptsize$55$}; \node at (-0.7626,-1.7780) {\scriptsize$39$}; \node at (-0.3587,-1.4559) {\scriptsize$1$}; \node at (1.4559,-1.9613) {\scriptsize$41$}; \node at (1.7780,-1.5574) {\scriptsize\color{linkblue}$3$}; \node at (-0.9222,0.1575) {\scriptsize$43$}; \node at (-0.6981,0.6228) {\scriptsize$5$}; \node at (1.3859,-0.0517) {\scriptsize$45$}; \node at (1.5008,0.4519) {\scriptsize\color{linkblue}$7$}; \node at (-0.8992,-2.5783) {\scriptsize$47$}; \node at (-0.8992,-2.0617) {\scriptsize$9$}; \node at (1.5008,-2.7719) {\scriptsize$49$}; \node at (1.3859,-2.2683) {\scriptsize$11$}; \node at (-0.6981,-0.6228) {\scriptsize$51$}; \node at (-0.9222,-0.1575) {\scriptsize$13$}; \node at (1.7780,-0.7626) {\scriptsize$53$}; \node at (1.4559,-0.3587) {\scriptsize$15$}; \node at (-0.3587,-3.1841) {\scriptsize$55$}; \node at (-0.7626,-2.8620) {\scriptsize$17$}; \node at (2.1625,-3.2422) {\scriptsize$1$}; \node at (1.6972,-3.0181) {\scriptsize$19$}; \node at (0.0517,-0.9341) {\scriptsize\color{linkblue}$3$}; \node at (-0.4519,-0.8192) {\scriptsize\color{linkblue}$21$}; \node at (2.5783,-0.8992) {\scriptsize$5$}; \node at (2.0617,-0.8992) {\scriptsize$23$}; \node at (0.4519,-3.1392) {\scriptsize\color{linkblue}$7$}; \node at (-0.0517,-3.2541) {\scriptsize\color{linkblue}$25$}; \node at (2.9428,-3.0181) {\scriptsize$9$}; \node at (2.4775,-3.2422) {\scriptsize$27$}; \node at (0.7626,-0.5420) {\scriptsize$11$}; \node at (0.3587,-0.8641) {\scriptsize$29$}; \node at (3.1841,-0.3587) {\scriptsize$13$}; \node at (2.8620,-0.7626) {\scriptsize$31$}; \node at (0.9222,-2.4775) {\scriptsize$15$}; \node at (0.6981,-2.9428) {\scriptsize$33$}; \node at (3.2541,-2.2683) {\scriptsize$17$}; \node at (3.1392,-2.7719) {\scriptsize$35$}; \node at (0.0000,-0.3000) {\small$G_0$}; \node at (2.3200,-0.3000) {\small$G_1$}; \node at (0.0000,-2.6200) {\small$G_2$}; \node at (2.3200,-2.6200) {\small$G_3$}; \fill (3.2949,0.2225) circle[radius=0.9pt]; \node[anchor=north west] at (3.2949,0.2125) {\small$O$}; \fill[black!25,draw=black!25,line width=0.2pt] (8.1760,-0.5000) -- (8.0384,-0.2048) -- (7.9516,-0.2048) -- (7.7811,-0.5000) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (4.6787,-3.0096) -- (4.8164,-3.3048) -- (5.1116,-3.0096) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (6.6204,-2.1984) -- (6.3252,-2.4936) -- (6.6725,-2.4936) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (7.9996,-2.1984) -- (6.6204,-2.1984) -- (6.6725,-2.4936) -- (7.9475,-2.4936) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (8.2948,-2.4936) -- (7.9996,-2.1984) -- (7.9475,-2.4936) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (4.8684,-0.6896) -- (5.1636,-0.9848) -- (5.3013,-0.6896) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (6.5816,-0.2048) -- (6.4440,-0.5000) -- (6.8389,-0.5000) -- (6.6684,-0.2048) -- cycle; \fill[black!25,draw=black!25,line width=0.2pt] (6.6684,-0.2048) -- (6.8389,-0.5000) -- (7.7811,-0.5000) -- (7.9516,-0.2048) -- cycle; \draw[black!55,line width=0.3pt] (5.6480,0.2395) -- (5.3401,0.6064) -- (4.8684,0.6896) -- (4.4536,0.4501) -- (4.2898,0.0000) -- (4.4536,-0.4501) -- (4.8684,-0.6896) -- (5.3401,-0.6064) -- (5.6480,-0.2395) -- cycle; \draw[black!55,line width=0.3pt] (7.9164,0.3501) -- (7.5495,0.6580) -- (7.0705,0.6580) -- (6.7036,0.3501) -- (6.6204,-0.1216) -- (6.8599,-0.5364) -- (7.3100,-0.7002) -- (7.7601,-0.5364) -- (7.9996,-0.1216) -- cycle; \draw[black!55,line width=0.3pt] (5.6902,-2.3200) -- (5.5264,-1.8699) -- (5.1116,-1.6304) -- (4.6399,-1.7136) -- (4.3320,-2.0805) -- (4.3320,-2.5595) -- (4.6399,-2.9264) -- (5.1116,-3.0096) -- (5.5264,-2.7701) -- cycle; \draw[black!55,line width=0.3pt] (7.9996,-2.1984) -- (7.7601,-1.7836) -- (7.3100,-1.6198) -- (6.8599,-1.7836) -- (6.6204,-2.1984) -- (6.7036,-2.6701) -- (7.0705,-2.9780) -- (7.5495,-2.9780) -- (7.9164,-2.6701) -- cycle; \draw[line width=0.5pt] (5.6480,-0.2395) -- (5.9900,0.0000) -- (5.6480,0.2395) -- (5.7560,0.6428) -- (5.3401,0.6064) -- (5.1636,0.9848) -- (4.8684,0.6896) -- (4.4900,0.8660) -- (4.4536,0.4501) -- (4.0503,0.3420) -- (4.2898,0.0000) -- (4.0503,-0.3420) -- (4.4536,-0.4501) -- (4.4900,-0.8660) -- (4.8684,-0.6896) -- (5.1636,-0.9848) -- (5.3401,-0.6064) -- (5.7560,-0.6428) -- cycle; \draw[line width=0.5pt] (7.9996,-0.1216) -- (8.2948,0.1736) -- (7.9164,0.3501) -- (7.9528,0.7660) -- (7.5495,0.6580) -- (7.3100,1.0000) -- (7.0705,0.6580) -- (6.6672,0.7660) -- (6.7036,0.3501) -- (6.3252,0.1736) -- (6.6204,-0.1216) -- (6.4440,-0.5000) -- (6.8599,-0.5364) -- (6.9680,-0.9397) -- (7.3100,-0.7002) -- (7.6520,-0.9397) -- (7.7601,-0.5364) -- (8.1760,-0.5000) -- cycle; \draw[line width=0.5pt] (5.6902,-2.3200) -- (5.9297,-1.9780) -- (5.5264,-1.8699) -- (5.4900,-1.4540) -- (5.1116,-1.6304) -- (4.8164,-1.3352) -- (4.6399,-1.7136) -- (4.2240,-1.6772) -- (4.3320,-2.0805) -- (3.9900,-2.3200) -- (4.3320,-2.5595) -- (4.2240,-2.9628) -- (4.6399,-2.9264) -- (4.8164,-3.3048) -- (5.1116,-3.0096) -- (5.4900,-3.1860) -- (5.5264,-2.7701) -- (5.9297,-2.6620) -- cycle; \draw[line width=0.5pt] (7.9996,-2.1984) -- (8.1760,-1.8200) -- (7.7601,-1.7836) -- (7.6520,-1.3803) -- (7.3100,-1.6198) -- (6.9680,-1.3803) -- (6.8599,-1.7836) -- (6.4440,-1.8200) -- (6.6204,-2.1984) -- (6.3252,-2.4936) -- (6.7036,-2.6701) -- (6.6672,-3.0860) -- (7.0705,-2.9780) -- (7.3100,-3.3200) -- (7.5495,-2.9780) -- (7.9528,-3.0860) -- (7.9164,-2.6701) -- (8.2948,-2.4936) -- cycle; \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (4.7475,-3.1572) -- (4.9640,-3.1572); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (6.4728,-2.3460) -- (8.1472,-2.3460); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (5.0160,-0.8372) -- (5.2325,-0.8372); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (6.5128,-0.3524) -- (8.1072,-0.3524); \node at (5.8512,0.1875) {\scriptsize$29$}; \node at (5.8512,-0.1875) {\scriptsize$43$}; \node at (8.1256,0.3342) {\scriptsize$31$}; \node at (8.1907,-0.0351) {\scriptsize$45$}; \node at (5.7352,-1.8493) {\scriptsize$33$}; \node at (5.8634,-2.2016) {\scriptsize$47$}; \node at (7.9621,-1.7270) {\scriptsize$35$}; \node at (8.1496,-2.0517) {\scriptsize$49$}; \node at (5.5292,0.6972) {\scriptsize$37$}; \node at (5.7702,0.4100) {\scriptsize$51$}; \node at (7.7200,0.7802) {\scriptsize$39$}; \node at (8.0072,0.5392) {\scriptsize$53$}; \node at (5.2583,-1.4804) {\scriptsize$41$}; \node at (5.5830,-1.6679) {\scriptsize$55$}; \node at (7.4284,-1.4466) {\scriptsize$43$}; \node at (7.7807,-1.5748) {\scriptsize$57$}; \node at (4.9549,0.8807) {\scriptsize$45$}; \node at (5.3242,0.8156) {\scriptsize$59$}; \node at (7.1225,0.8612) {\scriptsize$47$}; \node at (7.4975,0.8612) {\scriptsize$61$}; \node at (4.6558,-1.5044) {\scriptsize$49$}; \node at (5.0251,-1.4393) {\scriptsize$63$}; \node at (6.8393,-1.5748) {\scriptsize$51$}; \node at (7.1916,-1.4466) {\scriptsize$65$}; \node at (4.3970,0.6521) {\scriptsize$53$}; \node at (4.7217,0.8396) {\scriptsize$67$}; \node at (6.6128,0.5392) {\scriptsize$55$}; \node at (6.9000,0.7802) {\scriptsize$69$}; \node at (4.2098,-1.9100) {\scriptsize$57$}; \node at (4.4508,-1.6228) {\scriptsize$71$}; \node at (6.4704,-2.0517) {\scriptsize$59$}; \node at (6.6579,-1.7270) {\scriptsize$1$}; \node at (4.1166,0.1184) {\scriptsize$61$}; \node at (4.2448,0.4707) {\scriptsize$3$}; \node at (6.4293,-0.0351) {\scriptsize$63$}; \node at (6.4944,0.3342) {\scriptsize$5$}; \node at (4.1288,-2.5075) {\scriptsize$65$}; \node at (4.1288,-2.1325) {\scriptsize$7$}; \node at (6.4944,-2.6542) {\scriptsize$67$}; \node at (6.4293,-2.2849) {\scriptsize\color{linkblue}$9$}; \node at (4.2448,-0.4707) {\scriptsize$69$}; \node at (4.1166,-0.1184) {\scriptsize$11$}; \node at (6.6579,-0.5930) {\scriptsize$71$}; \node at (6.4704,-0.2683) {\scriptsize\color{linkblue}$13$}; \node at (4.4508,-3.0172) {\scriptsize$1$}; \node at (4.2098,-2.7300) {\scriptsize$15$}; \node at (6.9000,-3.1002) {\scriptsize$3$}; \node at (6.6128,-2.8592) {\scriptsize$17$}; \node at (4.7217,-0.8396) {\scriptsize$5$}; \node at (4.3970,-0.6521) {\scriptsize$19$}; \node at (7.1916,-0.8734) {\scriptsize$7$}; \node at (6.8393,-0.7452) {\scriptsize$21$}; \node at (5.0251,-3.2007) {\scriptsize\color{linkblue}$9$}; \node at (4.6558,-3.1356) {\scriptsize\color{linkblue}$23$}; \node at (7.4975,-3.1812) {\scriptsize$11$}; \node at (7.1225,-3.1812) {\scriptsize$25$}; \node at (5.3242,-0.8156) {\scriptsize\color{linkblue}$13$}; \node at (4.9549,-0.8807) {\scriptsize\color{linkblue}$27$}; \node at (7.7807,-0.7452) {\scriptsize$15$}; \node at (7.4284,-0.8734) {\scriptsize$29$}; \node at (5.5830,-2.9721) {\scriptsize$17$}; \node at (5.2583,-3.1596) {\scriptsize$31$}; \node at (8.0072,-2.8592) {\scriptsize$19$}; \node at (7.7200,-3.1002) {\scriptsize$33$}; \node at (5.7702,-0.4100) {\scriptsize$21$}; \node at (5.5292,-0.6972) {\scriptsize$35$}; \node at (8.1496,-0.2683) {\scriptsize\color{linkblue}$23$}; \node at (7.9621,-0.5930) {\scriptsize$37$}; \node at (5.8634,-2.4384) {\scriptsize$25$}; \node at (5.7352,-2.7907) {\scriptsize$39$}; \node at (8.1907,-2.2849) {\scriptsize\color{linkblue}$27$}; \node at (8.1256,-2.6542) {\scriptsize$41$}; \node at (4.9900,-0.3000) {\small$G_0$}; \node at (7.3100,0.3000) {\small$G_1$}; \node at (4.9900,-2.6200) {\small$G_2$}; \node at (7.3100,-2.6200) {\small$G_3$}; \fill (8.1760,-0.5000) circle[radius=0.9pt]; \node[anchor=north west] at (8.1760,-0.5100) {\small$O$}; \filldraw[fill=white,line width=0.4pt] (7.3100,-2.3200) circle[radius=1.1pt]; \end{tikzpicture} \caption{The model of \cref{lem:model-branched} with $p=4$: the regular $(n/4)$-gons $G_0,\dots,G_3$ with their ears; a leg is labelled $i$ when its direction from its apex is $i\psi$, and legs with the same label are glued. Left, $\tri49{15}{28}$ (\cref{thm:C4}, $m=14$); right, $\tri47{25}{36}$ (\cref{thm:lnz6}, $q=7$): shaded, the maximal cylinder $Z(\mathcal{W})$ of the chain $\mathcal{W}=\{3,21\}$, resp.\ $\{9,23\}$, whose share of the area has a negative conjugate; its lower boundary passes through the apex $O$, and on the right one of its leaves crosses the centre of $G_3$. For $\tri147{12}$ ($n=12$, $q=1$) the ears close up into a regular $12$-gon and the model is Hooper's Figure~1 \cite{Hooper}, turned through $13\pi/12$.}\label{fig:model-p4} \end{figure} In general $4\cos a\,(\cos a-\cos4a)=8\cos a\,\sin(5a/2)\sin(3a/2)$ is negative for $a\in(0,\pi)$ exactly when $a\in(2\pi/5,\pi/2)\cup(2\pi/3,4\pi/5)$; so $\sigma_\ell(\rho(Z))<0$ exactly when $\pm\ell/n$ is congruent modulo $1$ to a number in $(1/5,1/4)\cup(1/3,2/5)$. \needspace{4\baselineskip} \begin{theorem}\label{thm:lnz6} For every odd $q\ge3$ the triangle $\tri4q{3q+4}{(4q+8)}$ is not a lattice triangle. Let $n=4q+8$, $\psi=\pi/n$ and $\vartheta=2\psi$. \begin{enumerate} \item If $q\ge 7$, in the direction of the side joining the vertices of angles $4\psi$ and $q\psi$ the unfolding has a maximal cylinder $Z$ with $n\,\rho(Z)=2(1+\cos2\vartheta+\sin2\vartheta)$, and $\sigma_\ell(\rho(Z))<0$ for $\ell=q$ if $q\equiv3\pmod 4$ and for $\ell=q-2$ if $q\equiv1\pmod4$. \item If $q=5$, in the same direction the unfolding has a maximal cylinder $Z$ with $n\,\rho(Z)=8\cos^24\psi\,\cos2\psi\,\sin3\psi/\sin5\psi$, and $\sigma_{11}(\rho(Z))<0$. \item If $q=3$, the triangle is $\tri34{13}{20}$, which is not a lattice triangle by \cref{thm:finite}. \end{enumerate} \end{theorem} For $q\ge 5$ these triangles form family~(6) of \cref{thm:lnz41}: their obtuse angle exceeds $2\pi/3$ exactly when $q>4$. For $q=3$ the obtuse angle is $13\pi/20<2\pi/3$, and $\tri34{13}{20}$ is a member of $\listE$. For $q=1$ the triangle is $\tri147{12}$ once more. \begin{proof} Here $p=4$, $r=3q+4$, $(q+2)\psi=\pi/4$ and $q\psi=\pi/4-2\psi$. The hypotheses of \cref{lem:model-branched} hold, $n$ is even and $q$ is odd; $n/2=2q+4\equiv2\pmod4$, so $G_j$ is symmetric if and only if $j$ is odd. Since $\gcd(3q+4,4q+8)=\gcd(q+4,8)=1$, the vertex $R$ gives one singular point, and $r\nmid n$, as \cref{prop:chain} requires; since $\gcd(q,4q+8)=\gcd(q,8)=1$, the vertex $Q$ gives one point of cone angle $2\pi q$, singular for $q\ge3$. Let $q\ge5$ and $\mathcal{W}=\{q+2,3q+2\}$ (\cref{fig:model-p4}, right, for $q=7$). Its sines are $\sin(\pi/4)$ and $\sin(3q+2)\psi=\sin(q+6)\psi>\sin(\pi/4)$, because $\pi/4<(q+6)\psi<3\pi/4$. The neighbours are $(3q+2)+2q=n+(q-6)$, with sine $-\sin(q-6)\psi$, and $2-q<0$. So for $q\ge7$, $a(\mathcal{W})=0$; for $q=5$, $a(\mathcal{W})=\sin\psi$; in both cases $b(\mathcal{W})=\sin(\pi/4)>a(\mathcal{W})$, and $\iota(\mathcal{W})=\{q+6,3q+6\}\ne \mathcal{W}$. The exit ears are given by $2k_1\equiv(q+2)-n-q=2-n$ and $2k_2\equiv-(3q+2)-q=6-n\pmod{2n}$, so $k_1\equiv-2q-3$ and $k_2\equiv -2q-1\pmod n$, both odd, so $2k_1\equiv2k_2\equiv2\equiv n/2\pmod4$, which is \eqref{eq:chain-sym}. So \cref{prop:chain} applies: \[ \mathcal L(\mathcal{W})=2[\cos2\psi-\cos(4q+2)\psi]=2[\cos2\psi+\cos6\psi]=4\cos2\psi\cos4\psi, \] as $(4q+2)\psi=\pi-6\psi$, and $Z(\mathcal{W})$ is a maximal cylinder: for $q\ge7$ its lower boundary contains the apex $O$, a point over $Q$, and for $q=5$ a point over $R$. \emph{Part (a).} The height is $|QR|\sin(\pi/4)$ and the area of $X$ is $n|QR|\sin q\psi$, so \begin{multline*} n\rho=\frac{4\sin(\pi/4)\cos2\psi\cos4\psi}{\sin(\pi/4-2\psi)}=\frac{4\cos2\psi\cos4\psi}{\cos2\psi-\sin2\psi}\\ =4\cos2\psi(\cos2\psi+\sin2\psi)=2(1+\cos2\vartheta+\sin2\vartheta). \end{multline*} Since $(q+2)\vartheta=\pi/2$, $\sin2\vartheta=\cos q\vartheta$, so $n\rho=2+2\cos2\vartheta+2\cos q\vartheta\in\Q(\zeta)^+$ and $\sigma_\ell(n\rho)=2+2\cos2\ell\vartheta+2\cos q\ell\vartheta$. For odd $\ell$, $\cos q\ell\vartheta=\cos(\ell\pi/2-2\ell\vartheta)=\varepsilon_\ell\sin2\ell\vartheta$ with $\varepsilon_\ell=1$ if $\ell\equiv1$ and $\varepsilon_\ell=-1$ if $\ell\equiv3\pmod4$. If $q\equiv3\pmod4$, $\ell=q$ is a unit ($\gcd(q,8)=1$), $2q\vartheta=\pi-4\vartheta$, and $\sigma_q(n\rho)=2(1-\cos4\vartheta-\sin4\vartheta)=2-2\sqrt2\sin(4\vartheta+\pi/4)<0$, because $0<4\vartheta<\pi/2$. If $q\equiv1\pmod4$, then $q\ge9$ and $\ell=q-2\equiv3\pmod4$ is a unit (it is odd and $\gcd(q-2,q+2)$ divides $4$); $2(q-2)\vartheta=\pi-8\vartheta$, and $\sigma_{q-2}(n\rho)=2(1-\cos8\vartheta-\sin8\vartheta)<0$, because $0<8\vartheta<\pi/2$ for $n\ge44$. \emph{Part (b).} Now $\psi=\pi/28$ and $\pi/4=7\psi$; the height is $|QR|(\sin7\psi-\sin\psi)=2|QR|\cos4\psi\sin3\psi$ and $\sin q\psi=\sin5\psi$, so $n\rho=8\cos^24\psi\cos2\psi\sin3\psi/\sin5\psi$. With $\zeta=e^{i\vartheta}$ we have $\sin3\psi/\sin5\psi=\zeta(\zeta^3-1)/(\zeta^5-1)\in\Q(\zeta)$, so $\rho\in\Q(\zeta)^+$, and $\sigma_\ell$ maps $\sin3\psi/\sin5\psi$ to $\sin3\ell\psi/\sin5\ell\psi$. So $\sigma_{11}(n\rho)=8\cos^2(44\psi)\cos(22\psi)\sin(33\psi)/\sin(55\psi)$, where $\cos22\psi=\cos(11\pi/14)<0$, $\sin33\psi=\sin(33\pi/28)<0$ and $\sin55\psi=-\sin(\pi/28)<0$; so $\sigma_{11}(\rho)<0$. \emph{Part (c)} is \cref{thm:finite}. In parts (a) and (b), \cref{lem:cylinders}\ref{it:cyl-sign} applies. \end{proof} At $q=3$ the cylinder $Z(\mathcal{W})$ of the proof still exists, with $a(\mathcal{W})=\sin3\psi$, but its share $n\rho=4\cos^2(\pi/10)-2\cos(\pi/10)$ is totally positive, and we find that all eight maximal cylinders in the direction of $PQ$ have the same modulus, so that direction is consistent with \cref{lem:cylinders} and \cref{thm:finite} uses another one. \subsection{Two families with an angle \texorpdfstring{$\pi/n$}{pi/n}: the star}\label{ssec:star} In this subsection $\Delta=\tri1qrn$ has vertices $A$, $B$, $C$ with angles $\psi$, $q\psi$, $r\psi$, where $\psi=\pi/n$ as before, is obtuse at $C$, that is, $r>n/2$, and has $|BC|=1$. Place $A$ at the origin $O$ and, for a real number $\phi$, \begin{gather*} B_k=|AB|\,e^{i(2k\psi+\phi)},\quad C_k=|AC|\,e^{i((2k+1)\psi+\phi)}\quad(k\in\Z/n),\\ |AB|=\frac{\sin r\psi}{\sin\psi},\quad |AC|=\frac{\sin q\psi}{\sin\psi}. \end{gather*} The \emph{star} $\mathcal P$ is the union of the $2n$ triangles $\Delta_{2k}=OB_kC_k$ and $\Delta_{2k+1}=OC_kB_{k+1}$, each a copy of $\Delta$ with the angle $\psi$ at $O$. Its sides are $b_k=B_kC_k$ and $c_k=C_kB_{k+1}$, all of length $1$; its interior angles are $2q\psi$ at the $B_k$ and $2r\psi>\pi$ at the $C_k$, which we call the \emph{inner vertices}. For $\phi=0$ the star is $|AB|$ times the model of \cref{lem:model-branched} for $p=1$ and $(P,Q,R)=(A,B,C)$ (its hypotheses hold, as $q0$ and let $p_+,p_-\in\partial \mathcal P$ be points of height $h$, neither a vertex, lying on two different sides of $\mathcal P$. Suppose that every vertex of $\mathcal P$ met when going counterclockwise along $\partial \mathcal P$ from $p_+$ to $p_-$ has height greater than $h$. Then the open horizontal segment from $p_-$ to $p_+$ lies in the interior of $\mathcal P$, and $p_-$ is to the left of $p_+$. \end{lemma} \begin{proof} Height is affine along each side, equal to $h$ at $p_+$ and $p_-$ and greater than $h$ at the vertices between, so the open boundary arc from $p_+$ to $p_-$ lies in $\{y>h\}$; by (a) it is the set of $\beta(\alpha)$ with $\alpha$ in an interval $(\alpha_+,\alpha_-)\subset(0,\pi)$, where $\alpha_+$, $\alpha_-$ are the arguments of $p_+$, $p_-$. On the line $y=h$ the argument decreases strictly from left to right, so the open segment between $p_-$ and $p_+$ is $\{w: \operatorname{Im}w=h,\ \alpha_+<\arg w<\alpha_-\}$, and for such $w$, $|w|=h/\sin(\arg w)<|\beta(\arg w)|$, because $\beta(\arg w)$ has height greater than $h$. So $w$ lies on the open segment from $O$ to $\beta(\arg w)$, which is in the interior of $\mathcal P$ by (a). Finally $\alpha_->\alpha_+$ gives $x(p_-)=stealth] \fill[black!32,draw=black!32,line width=0.2pt] (0.0000,4.7357) -- (1.2728,4.7357) -- (0.9808,4.9307) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-3.8881,-2.9881) -- (3.8881,-2.9881) -- (4.1801,-2.7930) -- (-4.1801,-2.7930) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-1.2728,4.7357) -- (0.0000,4.7357) -- (-0.9808,4.9307) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-1.8123,-4.3752) -- (1.8123,-4.3752) -- (2.7930,-4.1801) -- (-2.7930,-4.1801) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-4.7357,0.0000) -- (4.7357,0.0000) -- (4.9307,0.9808) -- (-4.9307,0.9808) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-4.3752,1.8123) -- (4.3752,1.8123) -- (4.1801,2.7930) -- (-4.1801,2.7930) -- cycle; \draw[line width=0.5pt] (0.9808,4.9307) -- (0.0000,4.7357) -- (-0.9808,4.9307) -- (-1.8123,4.3752) -- (-2.7930,4.1801) -- (-3.3486,3.3486) -- (-4.1801,2.7930) -- (-4.3752,1.8123) -- (-4.9307,0.9808) -- (-4.7357,0.0000) -- (-4.9307,-0.9808) -- (-4.3752,-1.8123) -- (-4.1801,-2.7930) -- (-3.3486,-3.3486) -- (-2.7930,-4.1801) -- (-1.8123,-4.3752) -- (-0.9808,-4.9307) -- (0.0000,-4.7357) -- (0.9808,-4.9307) -- (1.8123,-4.3752) -- (2.7930,-4.1801) -- (3.3486,-3.3486) -- (4.1801,-2.7930) -- (4.3752,-1.8123) -- (4.9307,-0.9808) -- (4.7357,0.0000) -- (4.9307,0.9808) -- (4.3752,1.8123) -- (4.1801,2.7930) -- (3.3486,3.3486) -- (2.7930,4.1801) -- (1.8123,4.3752) -- cycle; \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (0.4904,4.8332) -- (1.1268,4.8332); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-4.0341,-2.8906) -- (4.0341,-2.8906); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-1.1268,4.8332) -- (-0.4904,4.8332); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-2.3026,-4.2776) -- (2.3026,-4.2776); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-4.2776,2.3026) -- (4.2776,2.3026); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-4.8332,0.4904) -- (4.8332,0.4904); \draw[<->,line width=0.4pt] (2.7000,0.0000) -- (2.7000,0.9808); \node[right=1pt] at (2.7000,0.7258) {\small$\cos\psi$}; \draw[->,line width=0.4pt] (1.0500,-3.7601) -- (1.0500,-4.1801); \draw[->,line width=0.4pt] (1.0500,-4.6752) -- (1.0500,-4.3752); \node[right=1pt] at (1.0500,-3.9101) {\small$\sin\psi$}; \node at (-2.7000,0.7258) {$Z$}; \node at (-2.7000,2.5380) {$Z$}; \node at (-2.7000,-2.5230) {$Z'$}; \node at (-2.7000,-3.9101) {$Z'$}; \node at (0.7512,4.4357) {$Z'$}; \node at (-0.7512,4.4357) {$Z'$}; \node at (-0.4280,5.1470) {\scriptsize$\Xi_{0}$}; \node at (0.4280,5.1470) {\scriptsize$\Xi_{1}$}; \node at (1.5743,4.9190) {\scriptsize$\Xi_{2}$}; \node at (-1.5743,4.9190) {\scriptsize$\Xi_{-1}$}; \node at (-3.9421,-3.3369) {\scriptsize$\Xi_{21}$}; \node at (3.9421,-3.3369) {\scriptsize$\Xi_{12}$}; \node at (-2.3651,-4.5915) {\scriptsize$\Xi_{19}$}; \node at (2.3651,-4.5915) {\scriptsize$\Xi_{14}$}; \node at (5.1470,0.4280) {\scriptsize$\Xi_{8}$}; \node at (4.5915,2.3651) {\scriptsize$\Xi_{6}$}; \node at (-5.1470,0.4280) {\scriptsize$\Xi_{-7}$}; \node at (-4.5915,2.3651) {\scriptsize$\Xi_{-5}$}; \node at (0.0000,5.0657) {\scriptsize$V_{0}$}; \node at (1.0452,5.2544) {\scriptsize$V_{1}$}; \node at (-1.0452,5.2544) {\scriptsize$V_{-1}$}; \node at (5.0657,0.0000) {\scriptsize$V_{8}$}; \node at (5.2544,1.0452) {\scriptsize$V_{7}$}; \node at (4.6801,1.9385) {\scriptsize$V_{6}$}; \node at (4.4545,2.9764) {\scriptsize$V_{5}$}; \node at (4.4545,-2.9764) {\scriptsize$V_{11}$}; \node at (2.9764,-4.4545) {\scriptsize$V_{13}$}; \node at (-2.9764,-4.4545) {\scriptsize$V_{19}$}; \node at (-4.4545,-2.9764) {\scriptsize$V_{21}$}; \fill (0.0000,0.0000) circle[radius=0.8pt]; \node[below left=0pt] at (0.0000,0.0000) {\scriptsize$O$}; \end{tikzpicture} \caption{Ward's star for $m=8$, that is for $\tri169{16}$, placed as in the proof of \cref{thm:star}, the notch $V_0$ on the positive $y$-axis; for odd $j$ the side $\Xi_j$ is glued to $\Xi_{j+2m-3}$. Shaded: the cylinders $Z'$ (dark), of height $\sin\psi$, and $Z$ (light), of height $\cos\psi$; dashed: their core curves.}\label{fig:star-ZT} \end{figure} At $m=6$, that is at $\tri147{12}$, the annuli $Z$ and $Z'$ exist with the moduli of the theorem, but the copies of $B$ are regular points and $Z$ and $Z'$ are not maximal; the maximal cylinders in that direction are four, with moduli $1/(5+3\sqrt3)$ and $1/(10+6\sqrt3)$ in pairs \cite[(5)]{Hooper}. \needspace{5\baselineskip} \begin{figure} \centering \begin{tikzpicture}[x=1.4cm,y=1.4cm,line join=round,line cap=round,>=stealth] \fill[black!32,draw=black!32,line width=0.2pt] (-2.7317,0.6235) -- (2.7317,0.6235) -- (2.8653,0.9010) -- (-2.8653,0.9010) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-1.2157,2.5245) -- (1.2157,2.5245) -- (1.0821,2.8019) -- (-1.0821,2.8019) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (0.0000,2.8019) -- (1.0821,2.8019) -- (0.7818,3.4254) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-2.8653,0.9010) -- (2.8653,0.9010) -- (3.1656,1.5245) -- (-3.1656,1.5245) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-1.0821,2.8019) -- (0.0000,2.8019) -- (-0.7818,3.4254) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-2.7317,-0.6235) -- (2.7317,-0.6235) -- (3.5135,0.0000) -- (-3.5135,0.0000) -- cycle; \draw[line width=0.5pt] (3.5135,0.0000) -- (2.7317,0.6235) -- (3.1656,1.5245) -- (2.1906,1.7470) -- (2.1906,2.7470) -- (1.2157,2.5245) -- (0.7818,3.4254) -- (0.0000,2.8019) -- (-0.7818,3.4254) -- (-1.2157,2.5245) -- (-2.1906,2.7470) -- (-2.1906,1.7470) -- (-3.1656,1.5245) -- (-2.7317,0.6235) -- (-3.5135,0.0000) -- (-2.7317,-0.6235) -- (-3.1656,-1.5245) -- (-2.1906,-1.7470) -- (-2.1906,-2.7470) -- (-1.2157,-2.5245) -- (-0.7818,-3.4254) -- (0.0000,-2.8019) -- (0.7818,-3.4254) -- (1.2157,-2.5245) -- (2.1906,-2.7470) -- (2.1906,-1.7470) -- (3.1656,-1.5245) -- (2.7317,-0.6235) -- cycle; \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-1.1489,2.6632) -- (1.1489,2.6632); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-2.7985,0.7622) -- (2.7985,0.7622); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (0.3909,3.1137) -- (0.9320,3.1137); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-3.0154,1.2127) -- (3.0154,1.2127); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-0.9320,3.1137) -- (-0.3909,3.1137); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-3.1226,-0.3117) -- (3.1226,-0.3117); \draw[->,line width=0.4pt] (1.5500,0.3035) -- (1.5500,0.6235); \draw[->,line width=0.4pt] (1.5500,1.2210) -- (1.5500,0.9010); \node[right=1pt] at (1.5500,0.4235) {\small$2\varepsilon\sin\Psi$}; \draw[<->,line width=0.4pt] (1.5500,-0.6235) -- (1.5500,0.0000); \node[right=1pt] at (1.5500,-0.1559) {\small$\sin\Psi$}; \node at (-1.5500,0.4235) {$U$}; \node at (-1.5500,1.3686) {$V$}; \node at (-1.5500,-0.1559) {$V$}; \node at (0.0000,2.3245) {$U$}; \node at (0.5988,2.9516) {$V$}; \node at (-0.5988,2.9516) {$V$}; \node at (3.1649,0.9698) {\scriptsize$c_0$}; \node at (-3.1649,0.9698) {\scriptsize$b_{6}$}; \node at (1.2150,3.0791) {\scriptsize$c_{2}$}; \node at (-1.2150,3.0791) {\scriptsize$b_{4}$}; \node at (0.2413,3.3013) {\scriptsize$b_{3}$}; \node at (-0.2413,3.3013) {\scriptsize$c_{3}$}; \node at (-3.2722,-0.4994) {\scriptsize$b_{7}$}; \node at (3.2722,-0.4994) {\scriptsize$c_{-1}$}; \node at (0.0000,3.0619) {\scriptsize$C_{3}$}; \node at (0.8397,3.6789) {\scriptsize$B_{3}$}; \node at (-0.8397,3.6789) {\scriptsize$B_{4}$}; \node at (2.9852,0.6813) {\scriptsize$C_0$}; \node at (3.3998,1.6373) {\scriptsize$B_1$}; \node at (3.7735,0.0000) {\scriptsize$B_0$}; \node at (1.3285,2.7587) {\scriptsize$C_{2}$}; \node at (-3.3998,1.6373) {\scriptsize$B_{6}$}; \node at (-3.7735,0.0000) {\scriptsize$B_{7}$}; \node at (2.9852,-0.6813) {\scriptsize$C_{-1}$}; \node at (-2.9852,0.6813) {\scriptsize$C_{6}$}; \fill (0.0000,0.0000) circle[radius=0.8pt]; \node[below left=0pt] at (0.0000,0.0000) {\scriptsize$O$}; \end{tikzpicture} \caption{The star of \cref{lem:star} for $\tri13{10}{14}$ ($q=3$ in \cref{thm:lnz4}), $B_0$ on the positive $x$-axis; $b_k$ is glued to $c_{k-4}$. Shaded: the cylinders $U$ (dark), of height $2\varepsilon\sin\Psi$, and $V$ (light), of height $\sin\Psi$; dashed: their core curves.}\label{fig:star-UV} \end{figure} \begin{theorem}\label{thm:lnz4} For every $q\ge3$ the triangle $\tri1q{3q+1}{(4q+2)}$ is not a lattice triangle. With $\varepsilon=\sin(\pi/(4q+2))$, in the direction of the side joining the vertices of angles $\pi/(4q+2)$ and $q\pi/(4q+2)$ the unfolding has two maximal cylinders $U$ and $V$ of moduli \[ \mu(U)=\frac{\varepsilon^2}{\cos\psi\,(1+2\varepsilon)},\qquad \mu(V)=\frac{\varepsilon}{4\cos\psi\,(1+\varepsilon)},\qquad\psi=\frac{\pi}{4q+2}, \] whose ratio $\mu(V)/\mu(U)=(1+2\varepsilon)/(4\varepsilon(1+\varepsilon))$ is irrational. \end{theorem} This is family (4) of \cref{thm:lnz41}; its obtuse angle exceeds $2\pi/3$ for $q\ge2$. For $q$ odd it is also a theorem of Ward \cite[Theorem~B]{Ward}. For $q=1$ and $q=2$ the triangle is $\tri114{6}$, respectively $\tri127{10}$, both lattice triangles. \begin{proof} Here $n=4q+2$, $\varepsilon=\sin\psi$, $\Psi:=q\psi=\pi/4-\psi/2$ and $r\psi=3\pi/4-\psi/2$; so $|AB|=\cos\Psi/\varepsilon$, $|AC|=\sin\Psi/\varepsilon$ and $\tan\Psi=\cos\psi/(1+\varepsilon)$. Take $\phi=0$: $B_k=|AB|e^{2ik\psi}$ and $C_k=|AC|e^{i(2k+1)\psi}$. Since $(2q+1)\psi=\pi/2$, the inner vertex $C_q$ lies on the positive $y$-axis, $B_0$ on the positive $x$-axis and $B_{2q+1}$ on the negative $x$-axis. The side $b_k$ rises by $\sin(q-2k)\psi$ and $c_k$ by $\sin(2k+q+2)\psi$ when traversed counterclockwise. By \cref{lem:star}, $c_k$ is glued to $b_{k+q+1}$, with $C_k\leftrightarrow C_{k+q+1}$ and $B_{k+1}\leftrightarrow B_{k+q+1}$. The mirror $x\mapsto -x$ maps $B_k\mapsto B_{2q+1-k}$, $C_k\mapsto C_{2q-k}$, and $y\mapsto-y$ maps $B_k\mapsto B_{-k}$, $C_k\mapsto C_{-k-1}$. The heights are $y(B_k)=|AB|\sin2k\psi$ and $y(C_k)=|AC|\sin(2k+1)\psi$. \emph{Heights used.} (i) $y(C_0)=\sin\Psi$ and $y(B_1)-y(C_0)=\sin(q+2)\psi=\sin\Psi+2\varepsilon\sin\Psi$, because $\sin(q+2)\psi-\sin q\psi=2\sin\psi\cos(q+1)\psi$ and $(q+1)\psi=\pi/2-\Psi$. (ii) $y(C_1)-y(B_1)=[\sin\Psi\sin3\psi-\cos\Psi\sin2\psi]/\varepsilon=\cos\Psi\cos\psi\,(1-2\varepsilon-4\varepsilon^2)/(1+\varepsilon)$, using $\sin\Psi(1+\varepsilon)=\cos\Psi\cos\psi$. The root of $1-2\varepsilon-4\varepsilon^2$ in $(0,1)$ is $\sin(\pi/10)$, so $y(C_1)>y(B_1)$ exactly when $\psi<\pi/10$, that is, $q\ge3$. (iii) For $1\le k\le 2q$, $y(B_k)\ge y(B_1)$; for $1\le k\le 2q-1$, $y(C_k)\ge y(C_1)>y(B_1)$ by (ii); for $0\le k\le 2q$, $y(C_k)\ge y(C_0)=\sin\Psi$. (iv) $y(B_q)-y(C_q)=\sin\Psi$ and $y(C_q)-y(C_{q-1})=|AC|(1-\cos2\psi)=2\varepsilon\sin\Psi>0$. \emph{Strips} (\cref{fig:star-UV}). $M$: heights $(\sin\Psi,y(B_1))$, right side $c_0$, left side $b_{2q}$, height $\sin\Psi+2\varepsilon\sin\Psi$; the arc from $B_1$ to $B_{2q}$ passes through $B_1,C_1,\dots,C_{2q-1},B_{2q}$, all of height at least $y(B_1)$ by (iii). $M_0$: heights $(0,\sin\Psi)$, right side $b_0$, left side $c_{2q}$; the vertices on the arc from $C_0$ to $C_{2q}$ have height at least $\sin\Psi$ by (iii). Its mirror image $M'$ in the $x$-axis has heights $(-\sin\Psi,0)$, left side $b_{2q+1}$ and right side $c_{-1}$. The \emph{slab}: heights $(y(C_{q-1}),y(C_q))$, right side $c_{q-1}$, left side $b_{q+1}$; the arc from $B_q$ to $B_{q+1}$ is $B_q,C_q,B_{q+1}$, of heights at least $y(C_q)$ by (iv). The \emph{tips}: heights $(y(C_q),y(B_q))$; the right tip has right side $c_{q-1}$ and left side $b_q$, the left tip right side $c_q$ and left side $b_{q+1}$. These strips are pairwise disjoint: the slab and the two tips lie at heights above $y(C_{q-1})\ge y(C_1)>y(B_1)$, the top of $M$; the tips lie in $x>0$ and $x<0$; $M'$ lies below the $x$-axis; and $U$ and $V$ below use complementary parts of $M$. \emph{The cylinder $U$} is the lower part of $M$, of heights $(\sin\Psi,\sin\Psi+2\varepsilon\sin\Psi)$, together with the slab. The gluing $c_0\leftrightarrow b_{q+1}$ takes $C_0$ to $C_{q+1}$, and $y(C_{q+1})=y(C_{q-1})$ by the mirror, so the bottom of $M$ goes to the bottom of the slab and the lower part of $M$ onto the slab (iv). The gluing $c_{q-1}\leftrightarrow b_{2q}$ takes $C_{q-1}$ to $C_{2q}$, of height $y(C_0)$: bottom to bottom, back into $M$ through its left side. By \cref{lem:annulus}, $U$ is an annulus of height $2\varepsilon\sin\Psi$; on its lower boundary its circumference is $2x(C_0)+2x(C_{q-1})=2|AC|(\cos\psi+\sin2\psi)=2\sin\Psi\cos\psi(1+2\varepsilon)/\varepsilon$, as $(2q-1)\psi=\pi/2-2\psi$. So $\mu(U)=\varepsilon^2/(\cos\psi(1+2\varepsilon))$. The lower boundary passes through $C_0$, the upper boundary through $C_q$. \emph{The cylinder $V$} is the right tip, the upper part of $M$ (heights $(y(B_1)-\sin\Psi,y(B_1))$), the left tip and $M'$. The gluing $c_{q-1}\leftrightarrow b_{2q}$ takes $B_q$ to $B_{2q}$, of height $y(B_1)$: the apex of the right tip goes to the top of $M$, and the tip (of height $\sin\Psi$) onto the upper part of $M$. The gluing $c_0\leftrightarrow b_{q+1}$ takes $B_1$ to $B_{q+1}$, of height $y(B_q)$: the upper part of $M$ goes onto the left tip. The gluing $c_q\leftrightarrow b_{2q+1}$ takes $B_{q+1}$ to $B_{2q+1}$ (height $0$, the top of $M'$) and $C_q$ to $C_{2q+1}$ (height $-\sin\Psi$): the left tip goes onto $M'$. The gluing $c_{-1}\leftrightarrow b_q$ takes $B_0$ to $B_q$: $M'$ goes onto the right tip. By \cref{lem:annulus}, $V$ has height $\sin\Psi$; on its upper boundary the tips have width zero and the chords are $B_{2q}B_1$ and $B_{2q+1}B_0$, so its circumference is $2x(B_1)+2x(B_0)=2|AB|(1+\cos2\psi)=4\cos\Psi\cos^2\psi/\varepsilon$, and $\mu(V)=\varepsilon/(4\cos\psi(1+\varepsilon))$. The upper boundary passes through $B_q$, the lower boundary through $C_q$. \emph{Maximality.} For $q$ even the $C_k$ form one point of cone angle $2\pi(3q+1)$ and the $B_k$ two points of cone angle $\pi q$; for $q$ odd the $B_k$ form one point of cone angle $2\pi q$ and the $C_k$ two points of cone angle $\pi(3q+1)$. For $q\ge3$ all exceed $2\pi$, so $U$ and $V$ are maximal cylinders, both horizontal; the horizontal direction is that of the side $AB$. \emph{Irrationality.} Let $\varrho=\mu(V)/\mu(U)=(1+2\varepsilon)/(4\varepsilon(1+\varepsilon))$. If $\varrho\in\Q$, then $\varepsilon$ is a root of $4\varrho x^2+(4\varrho-2)x-1$, a non-zero rational polynomial of degree $2$. But $\varepsilon=\cos(q\pi/(2q+1))$ is, up to sign, $\cos(2\pi j/(2q+1))$ with $\gcd(j,2q+1)=1$ ($j=q/2$ for $q$ even, and $\varepsilon=-\cos(2\pi\frac{q+1}2/(2q+1))$ for $q$ odd), so its degree is $\varphi(2q+1)/2\ge3$ for $q\ge3$. So $\varrho\notin\Q$, and \cref{lem:cylinders}\ref{it:cyl-ratio} applies. \end{proof} \section{Single triangles}\label{sec:finite} The list $\mathcal F$ of \cref{tab:F} consists of the triangles that no general theorem of this paper reaches: $\tri135{9}$ of $\famA$, to which \cref{thm:relative} does not apply (\cref{thm:AB}\ref{it:AB-exc}); the members of $\listE$ to which \cref{thm:relative} does not apply, among them $\tri34{13}{20}$, the case $q=3$ of \cref{thm:lnz6}; and the triangles $\tri14{11}{16}$, $\tri23{17}{22}$, $\tri38{29}{40}$ and $\tri2{11}{29}{42}$ of item~(7) of \cref{thm:lnz41}, to which it does not apply either. All of them have denominator at most $48$. \needspace{4\baselineskip} \begin{theorem}\label{thm:finite} No triangle of \cref{tab:F} is a lattice triangle. For each of them, in the direction $d$ given in the table, the unfolding has two maximal cylinders $Z_1$, $Z_2$ such that $\mu(Z_1)/\mu(Z_2)$ has the minimal polynomial given in the table, of degree between $2$ and $8$. The direction $d$, as an element of $\Z[z]$ with $z=e^{i\pi/n}$, is in the deposit \cite{Record}. \end{theorem} \begin{proof} The cylinders $Z_1$ and $Z_2$, their maximality, and the minimal polynomial of the ratio of their moduli are verified for each triangle by an exact computation, described in \cref{ssec:certificates}. A polynomial of degree at least $2$ that is irreducible over $\Q$ has no rational root, so the ratio is irrational, and \cref{lem:cylinders}\ref{it:cyl-ratio} applies. \end{proof} The certificates for the members of $\listE$ in \cref{tab:F} with even denominator are used in the window case of the proof of \cref{thm:main}, those for the triangles of item~(7) in \cref{thm:120}, and those for $\tri135{9}$ and $\tri3{7}{11}{21}$ in \cref{thm:odd}. \Cref{fig:cert-135} shows the certificate of $\tri135{9}$. The unfolding of $\tri135{9}$ has genus $3$, and $\Sigma$ is the one point over $C$. In its certificate, with $z=e^{i\pi/9}$, $d=-1+z+2z^3-2z^4$ is a positive multiple of $1+z$, at the angle $\pi/18$ to $AB$. The cylinder $Z_1$ consists of three strips, each with an itinerary of sixteen steps, which meet along a leaf through a point over $B$ and along a leaf through the point over $A$. The cylinder $Z_2$ consists of two strips, each with an itinerary of eight steps, which meet along a leaf through a point over $B$. The outer boundary leaves of both pass through the point over $C$. By \cref{lem:certificate}, \[ \mu(Z_1)=\sin\tfrac{\pi}{9}+\sin\tfrac{2\pi}{9}-\sin\tfrac{\pi}{3},\qquad\mu(Z_2)=\sin\tfrac{2\pi}{9}-\sin\tfrac{\pi}{9}, \] and $\mu(Z_1)/\mu(Z_2)\approx0.3949$ has the minimal polynomial $x^3-x+\tfrac13$, of degree $3$, as in row~1 of \cref{tab:F}. \begin{figure} \centering \begin{tikzpicture}[x=1.6cm,y=1.6cm,line join=round,line cap=round,>=stealth] \fill[black!32,draw=black!32,line width=0.2pt] (-2.0328,-1.7057) -- (2.0328,-1.7057) -- (2.1929,-1.2660) -- (-2.1929,-1.2660) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-2.6537,0.0000) -- (2.6537,0.0000) -- (2.4936,0.4397) -- (-2.4936,0.4397) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-1.8508,-2.2057) -- (1.8508,-2.2057) -- (2.0328,-1.7057) -- (-2.0328,-1.7057) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-2.8356,-0.5000) -- (2.8356,-0.5000) -- (2.6537,0.0000) -- (-2.6537,0.0000) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-0.8660,-2.3794) -- (0.8660,-2.3794) -- (1.8508,-2.2057) -- (-1.8508,-2.2057) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (0.0000,2.5321) -- (1.1305,2.5321) -- (0.9848,2.7057) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-2.6899,-0.6736) -- (2.6899,-0.6736) -- (2.8356,-0.5000) -- (-2.8356,-0.5000) -- cycle; \fill[black!32,draw=black!32,line width=0.2pt] (-1.1305,2.5321) -- (0.0000,2.5321) -- (-0.9848,2.7057) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (0.0000,-2.8794) -- (0.8660,-2.3794) -- (-0.8660,-2.3794) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-2.4936,1.4397) -- (2.4936,1.4397) -- (1.6276,1.9397) -- (-1.6276,1.9397) -- cycle; \fill[black!12,draw=black!12,line width=0.2pt] (-2.4936,0.4397) -- (2.4936,0.4397) -- (2.4936,1.4397) -- (-2.4936,1.4397) -- cycle; \draw[black!50,line width=0.3pt] (2.8356,-0.5000) -- (0,0) -- (2.4936,0.4397); \draw[line width=0.5pt] (2.8356,-0.5000) -- (2.4936,0.4397) -- (2.4936,1.4397) -- (1.6276,1.9397) -- (0.9848,2.7057) -- (0.0000,2.5321) -- (-0.9848,2.7057) -- (-1.6276,1.9397) -- (-2.4936,1.4397) -- (-2.4936,0.4397) -- (-2.8356,-0.5000) -- (-2.1929,-1.2660) -- (-1.8508,-2.2057) -- (-0.8660,-2.3794) -- (0.0000,-2.8794) -- (0.8660,-2.3794) -- (1.8508,-2.2057) -- (2.1929,-1.2660) -- cycle; \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-1.9902,-1.8227) -- (1.9902,-1.8227); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-2.6962,-0.1170) -- (2.6962,-0.1170); \draw[linkblue,line width=0.75pt,dash pattern=on 2.2pt off 1.4pt] (-2.4936,1.1897) -- (2.4936,1.1897); \node at (2.8244,0.0280) {\scriptsize$1$}; \node at (-2.1816,-1.7940) {\scriptsize$1$}; \node at (2.1456,1.8369) {\scriptsize$2$}; \node at (-0.5180,-2.7766) {\scriptsize$2$}; \node at (0.4629,2.7863) {\scriptsize$3$}; \node at (1.3879,-2.4600) {\scriptsize$3$}; \node at (-1.4364,2.4320) {\scriptsize$4$}; \node at (2.6445,-0.9923) {\scriptsize$4$}; \node at (-2.6636,0.9397) {\scriptsize$5$}; \node at (2.6636,0.9397) {\scriptsize$5$}; \node at (-2.6445,-0.9923) {\scriptsize$6$}; \node at (1.4364,2.4320) {\scriptsize$6$}; \node at (-1.3879,-2.4600) {\scriptsize$7$}; \node at (-0.4629,2.7863) {\scriptsize$7$}; \node at (0.5180,-2.7766) {\scriptsize$8$}; \node at (-2.1456,1.8369) {\scriptsize$8$}; \node at (2.1816,-1.7940) {\scriptsize$9$}; \node at (-2.8244,0.0280) {\scriptsize$9$}; \fill (2.4936,0.4397) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (2.8356,-0.5000) circle[radius=1.3pt]; \fill (1.6276,1.9397) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (2.4936,1.4397) circle[radius=1.3pt]; \fill (0.0000,2.5321) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (0.9848,2.7057) circle[radius=1.3pt]; \fill (-1.6276,1.9397) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (-0.9848,2.7057) circle[radius=1.3pt]; \fill (-2.4936,0.4397) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (-2.4936,1.4397) circle[radius=1.3pt]; \fill (-2.1929,-1.2660) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (-2.8356,-0.5000) circle[radius=1.3pt]; \fill (-0.8660,-2.3794) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (-1.8508,-2.2057) circle[radius=1.3pt]; \fill (0.8660,-2.3794) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (0.0000,-2.8794) circle[radius=1.3pt]; \fill (2.1929,-1.2660) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (1.8508,-2.2057) circle[radius=1.3pt]; \filldraw[fill=white,line width=0.4pt] (0,0) circle[radius=1.3pt]; \node[left=1pt] at (0,0) {\scriptsize$O$}; \node at (3.0031,-0.5295) {\scriptsize$B_0$}; \node at (-1.4178,-0.3085) {$Z_1$}; \node at (-1.2468,0.8147) {$Z_2$}; \draw[->,line width=0.5pt] (-2.8356,-2.6294) -- (-2.2356,-2.6294) node[right=1pt] {$d$}; \end{tikzpicture} \caption{The unfolding of $\tri135{9}$, the star of \cref{lem:star} turned so that the direction $d$ of \cref{tab:F}, row~1, is horizontal. Sides with the same numeral are glued. Filled dots: the copies of the vertex $C$, which form the one point of $\Sigma$, of cone angle $10\pi$. Hollow dots: the copies of $B$ and the point $O$ over $A$, which form four points of cone angle $2\pi$. Shaded: the cylinders $Z_1$ (dark) and $Z_2$ (light) of the certificate.}\label{fig:cert-135} \end{figure} That the triangles of \cref{tab:F} are not lattice triangles also follows from \cite[Theorem~4.1]{DelecroixRuth} for the nineteen of them other than $\tri14{11}{16}$ and $\tri34{13}{20}$. These two Delecroix and R\"uth list as orbit closures of rank two, and for $\tri34{13}{20}$ they prove that the orbit closure has complex dimension at least~$4$ \cite[Theorem~1.2]{DelecroixRuth}, so that this triangle is not a lattice triangle. That the triangles of item~(7) of \cref{thm:lnz41} are not lattice triangles was shown with the program of R\"uth, Delecroix and Eskin \cite[pp.~7120 and 7125]{LNZ}. \begin{table}[ht] \footnotesize\renewcommand{\arraystretch}{0.95} \caption{The triangles of \cref{thm:finite}, those that no general theorem of the paper reaches. Column~3: the list or family of the triangle ((7): item~(7) of \cref{thm:lnz41}). Column~4: the direction $d$, as the angle $\theta\in[0,\pi)$ from $AB$ to the line $\R d$, counterclockwise, in the coordinates of \cref{ssec:certificates}, with $A=0$ and $B$ on the positive real axis; in row~5, where $\theta/\pi$ is irrational, $d$ itself, as a polynomial in $z=e^{i\pi/n}$. Column~5: the minimal polynomial of $\mu(Z_1)/\mu(Z_2)$ over $\Q$, $Z_1$ and $Z_2$ being the first two cylinders of the certificate.}\label{tab:F} \centering \begin{tabular}{@{}r l l l l@{}} \toprule & triangle & in & direction & minimal polynomial\\ \midrule 1 & $(1,3,5)/9$ & $\famA$ & $\pi/18$ & $x^{3} - x + \tfrac{1}{3}$\\ 2 & $(1,4,11)/16$ & (7) & $3\pi/4$ & $x^{2} - 4x + 2$\\ 3 & $(3,4,9)/16$ & $\listE$ & $\pi/4$ & $x^{4} - 4x^{3} - 2x^{2} + 4x - 1$\\ 4 & $(1,6,11)/18$ & $\listE$ & $0$ & $x^{3} - 3x^{2} + 2x - \tfrac{1}{3}$\\ 5 & $(3,4,13)/20$ & $\listE$ & $2z^{2}+z^{4}-3z^{6}-2z^{8}+2z^{12}$ & $x^{2} - x + \tfrac{1}{5}$\\ 6 & $(3,7,11)/21$ & $\listE$ & $5\pi/14$ & $x^{6} - 14x^{5} + 63x^{4} - 126x^{3} + 119x^{2} - 49x + 7$\\ 7 & $(2,3,17)/22$ & (7) & $19\pi/22$ & $x^{5} - 6x^{4} - x^{3} + 10x^{2} - 6x + 1$\\ 8 & $(1,8,15)/24$ & $\listE$ & $0$ & $x^{2} - x + \tfrac{1}{6}$\\ 9 & $(3,8,13)/24$ & $\listE$ & $7\pi/8$ & $x^{4} - 4x^{2} + 1$\\ 10 & $(1,10,17)/28$ & $\listE$ & $5\pi/14$ & $x^{3} - 7x^{2} + 14x - 7$\\ 11 & $(3,10,15)/28$ & $\listE$ & $0$ & $x^{6} - 10x^{5} - 12x^{4} + 50x^{3} + 11x^{2} - 12x + 1$\\ 12 & $(2,9,19)/30$ & $\listE$ & $\pi/10$ & $x^{4} + 3x^{3} - 21x^{2} + 27x - 9$\\ 13 & $(3,10,17)/30$ & $\listE$ & $5\pi/6$ & $x^{2} - 4x - 1$\\ 14 & $(1,12,23)/36$ & $\listE$ & $5\pi/6$ & $x^{3} - 9x + 9$\\ 15 & $(5,12,19)/36$ & $\listE$ & $\pi/2$ & $x^{6} - 6x^{5} - 18x^{4} + 58x^{3} - 21x^{2} - 18x + 1$\\ 16 & $(3,8,29)/40$ & (7) & $13\pi/40$ & \\ \multicolumn{5}{r}{$x^{8} + 4x^{7} - 36x^{6} + 48x^{5} + 44x^{4} - 104x^{3} + 16x^{2} + 32x - 4$}\\ 17 & $(8,9,23)/40$ & $\listE$ & $11\pi/20$ & \\ \multicolumn{5}{r}{$x^{8} - 28x^{7} + 116x^{6} + 716x^{5} - 4711x^{4} + 2048x^{3} + 21334x^{2} - 12036x - 31679$}\\ 18 & $(2,11,29)/42$ & (7) & $31\pi/42$ & $x^{6} - 13x^{5} + 50x^{4} - 69x^{3} + 15x^{2} + 22x + 1$\\ 19 & $(5,14,23)/42$ & $\listE$ & $25\pi/42$ & $x^{6} + \tfrac{92}{41}x^{5} - \tfrac{43}{41}x^{4} - \tfrac{125}{41}x^{3} - \tfrac{22}{41}x^{2} + \tfrac{15}{41}x - \tfrac{1}{41}$\\ 20 & $(3,16,29)/48$ & $\listE$ & $0$ & \\ \multicolumn{5}{r}{$x^{8} + \tfrac{424}{191}x^{7} - \tfrac{2708}{191}x^{6} - \tfrac{1808}{191}x^{5} + \tfrac{8094}{191}x^{4} - \tfrac{3672}{191}x^{3} - \tfrac{132}{191}x^{2} + \tfrac{40}{191}x + \tfrac{2}{191}$}\\ 21 & $(5,16,27)/48$ & $\listE$ & $\pi/8$ & \\ \multicolumn{5}{r}{$x^{8} + 100x^{7} + 2362x^{6} + 12380x^{5} + 10851x^{4} - 5156x^{3} - 3758x^{2} - 52x + \tfrac{191}{2}$}\\ \bottomrule \end{tabular} \end{table} \section{The classification}\label{sec:classification} In this section we prove \cref{thm:odd}, \cref{cor:prime}, \cref{thm:120} and \cref{thm:main}, in this order. The members of $\listE$ and the triples of item~(7) of \cref{thm:lnz41} that are not in \cref{tab:F} are reached by \cref{thm:relative}: $\tri236{11}$, $\tri258{15}$, $\tri348{15}$ and $\tri27{12}{21}$ by part~\ref{it:TI}, and the triangles of the following lemma by part~\ref{it:TII}. \needspace{8\baselineskip} \begin{lemma}\label{lem:TII-single} \Cref{thm:relative}\ref{it:TII} applies to each of the thirteen triangles below at the vertex $x$ with the unit $b$, so none of them is a lattice triangle. Three of them, $\tri13{16}{20}$, $\tri14{21}{26}$ and $\tri18{19}{28}$, are triangles of item~(7) of \cref{thm:lnz41}, and the other ten are members of $\listE$ with even denominator. Here $g=\gcd(x,n)$, and the last column gives $\res{bp}n$, $\res{bq}n$, $\res{br}n$. \begin{center}\small \begin{tabular}{@{}lrrrrl@{\qquad\qquad}lrrrrl@{}} \toprule $\Delta$ & $x$ & $g$ & $n/g$ & $b$ & \textup{residues} & $\Delta$ & $x$ & $g$ & $n/g$ & $b$ & \textup{residues}\\ \midrule $\tri35{10}{18}$ & $10$ & $2$ & $9$ & $11$ & $15,1,2$ & $\tri18{19}{28}$ & $8$ & $4$ & $7$ & $9$ & $9,16,3$\\ $\tri13{16}{20}$ & $16$ & $4$ & $5$ & $7$ & $7,1,12$ & $\tri58{15}{28}$ & $8$ & $4$ & $7$ & $23$ & $3,16,9$\\ $\tri35{12}{20}$ & $12$ & $4$ & $5$ & $7$ & $1,15,4$ & $\tri3{13}{18}{34}$ & $18$ & $2$ & $17$ & $19$ & $23,9,2$\\ $\tri37{12}{22}$ & $12$ & $2$ & $11$ & $13$ & $17,3,2$ & $\tri1{16}{23}{40}$ & $16$ & $8$ & $5$ & $7$ & $7,32,1$\\ $\tri14{21}{26}$ & $4$ & $2$ & $13$ & $15$ & $15,8,3$ & $\tri3{16}{21}{40}$ & $16$ & $8$ & $5$ & $27$ & $1,32,7$\\ $\tri19{16}{26}$ & $16$ & $2$ & $13$ & $15$ & $15,5,6$ & $\tri3{17}{24}{44}$ & $24$ & $4$ & $11$ & $13$ & $39,1,4$\\ $\tri57{14}{26}$ & $14$ & $2$ & $13$ & $15$ & $23,1,2$ & & & & & & \\ \bottomrule \end{tabular} \end{center} \end{lemma} \begin{proof} In each row $11$ and $x\nmid n$. Also $\gcd(b,n)=1$, and $b-2$ is $n/g$, except for $\tri58{15}{28}$ and $\tri3{16}{21}{40}$, where it is $3n/g$ and $5n/g$; so $b\equiv2\pmod{n/g}$, and $02n/3$. For $n$ odd this is \cref{thm:odd} ($n-2>2n/3$ exactly when $n\ge7$); let $n$ be even. If $\Delta$ has a witness, it is not a lattice triangle (\cref{prop:test}). Otherwise, by \cref{thm:lnz41} in the direction ``no witness implies listed'', $\Delta$ lies in one of the families (1) to (6) or is one of the seven triangles of (7). \begin{itemize} \item Family (1): $\tri11{n-2}n$, with $n-2>2n/3$, that is $n\ge7$. It is a lattice triangle by Veech's theorem. \item Family (2): $\tri12{2k-3}{2k}$, with $2k-3>4k/3$, that is $k\ge5$. It is a lattice triangle by the theorem of Vorobets and Ward. \item Family (3): $\tri14{r}{(r+5)}$ with $r\equiv7\pmod8$. For $n=12$, $r=7<2n/3$; so $n\ge20$, and the triangle is not a lattice triangle by \cite[\S10]{LNZ}. \item Family (4): $\tri1q{3q+1}{(4q+2)}$, with $3q+1>2(4q+2)/3$, that is $q\ge2$. For $q=2$ it is $\tri127{10}$, of family (2). For $q\ge3$ it is not a lattice triangle by \cref{thm:lnz4}. \item Family (5): $\tri2q{3q+2}{(4q+4)}$. In lowest terms $q$ is odd (for even $q$ the triple has the common factor $2$ and the triangle lies in family (4)), and $p\le q$ gives $q\ge3$. It is not a lattice triangle by \cref{thm:lnz5}. \item Family (6): $\tri4q{3q+4}{(4q+8)}$, with $q$ odd in lowest terms (for $q\equiv2\pmod4$ the triangle lies in family (5), for $4\mid q$ in family (4)), and $3q+4>2(4q+8)/3$, that is $q\ge5$. It is not a lattice triangle by \cref{thm:lnz6}. \item The seven triangles of (7) are not lattice triangles: $\tri13{16}{20}$, $\tri14{21}{26}$ and $\tri18{19}{28}$ by \cref{lem:TII-single}, and $\tri14{11}{16}$, $\tri23{17}{22}$, $\tri38{29}{40}$ and $\tri2{11}{29}{42}$ by \cref{thm:finite}. \end{itemize} So $\Delta$ is a lattice triangle exactly when it is $\tri11{n-2}n$ with $n\ge7$ or $\tri12{2k-3}{2k}$ with $k\ge5$. \end{proof} \begin{proof}[Proof of \cref{thm:main}] Let $\Delta=\tri pqrn$ be obtuse. If $n$ is odd, the statement is \cref{thm:odd}, since $\tri12{2k-3}{2k}$ and $\tri147{12}$ have even denominator. If $n$ is even and $r>2n/3$, it is \cref{thm:120}, since the triangles of the statement with $n$ even and $r\le2n/3$ are $\tri114{6}$, $\tri125{8}$ and $\tri147{12}$. Let $n$ be even and $n/20$; \item exactly one strip has a corner in $\Sigma$ among the corners attaining its $\mathrm{lo}$, and exactly one has a corner in $\Sigma$ among those attaining its $\mathrm{hi}$; starting from the first, each next strip has no corner in $\Sigma$ attaining its $\mathrm{lo}$ and one of the corners attaining its $\mathrm{lo}$ is the same point of $X$ as one of the corners attaining the current $\mathrm{hi}$; the chain uses every strip once and ends at the second; \item $\mathcal{V}$ is the sum of $\mathrm{hi}-\mathrm{lo}$ over the strips. \end{enumerate} \needspace{5\baselineskip} \begin{lemma}\label{lem:certificate} Let $X$ have genus at least two. If a certificate in direction $d$ passes \textup{(v1)}--\textup{(v5)}, then the union of the leaves of its strips and of the leaves along which consecutive strips meet is a cylinder $Z$ of $X$ in direction $d$ in the sense of \cref{ssec:cylinders}, with core holonomy $c$, height $H=\mathcal{V}/(2i|d|)$, modulus $\mu=\mathcal{V}/(2i\,c\bar d)$, and $\Area(Z)/\Area(X)=\mathcal{V}\,c\bar d/(n\,d\bar d(\bar BC-B\bar C))$. \end{lemma} \begin{proof} (i) \emph{Each strip is an annulus of closed leaves.} Let $\eta$ lie strictly between $\mathrm{lo}$ and $\mathrm{hi}$. By (v2), at every step of the itinerary each corner of the copy lies strictly on the same side of $\eta$ as of $a_0+b\epsilon$, so the leaf at height $\eta$ leaves each copy through the same side as the leaf at $a_0+b\epsilon$ and follows the same itinerary. The height changes depend only on the steps, so after the whole itinerary the leaf at $\eta$ is back on the start side at height $\eta$, the total change being $0$ by (v1); its holonomy is $c$ by (v3), and it meets no corner, so it misses $\Sigma$. (ii) \emph{Embedded.} Let $\ell$ be the length of the itinerary, and suppose the leaf at $\eta$ enters the start copy through the start side at a step $k$ with $00$ by (v3), $|c|=c\bar d/|d|$, so $\mu=H/|c|=\mathcal{V}/(2i\,c\bar d)$ and $\Area(Z)=H|c|=\mathcal{V}\,c\bar d/(2i\,d\bar d)$. The copy of $\Delta$ has area $\operatorname{Im}(\bar BC)/2=(\bar BC-B\bar C)/(4i)$, and $X$ is made of $2n$ copies. \end{proof} A vertex of cone angle $2\pi$ on a boundary leaf does not stop the cylinder (\cref{ssec:cylinders}): the chain of (v4) continues through it. By \cref{lem:certificate}, if two cylinders in direction $d$ have certificates with holonomies $c_k$, $c_l$ and numbers $\mathcal{V}_k$, $\mathcal{V}_l$, the ratio of their moduli is $\mu_k/\mu_l=\mathcal{V}_k\,c_l\bar d/(\mathcal{V}_l\,c_k\bar d)\in\Q(z)$, and its minimal polynomial over $\Q$ is computed exactly. The certificates of the triangles of \cref{tab:F}, each with its direction and two or three cylinders, and the programs that verify (v1)--(v5) and compute the minimal polynomials are in the deposit \cite{Record}. The directions were found with the flatsurf software \cite{flatsurf}. \subsection{The denominators not covered by the general theorems}\label{ssec:exceptional} \Cref{thm:E-large,thm:O-large} leave finitely many denominators, which are determined here, with a proof that the lists are complete; the lists are in the deposit \cite{Record}. Let $\pi_i$ be the $i$-th odd prime, $\Omega_k=\pi_1\cdots\pi_k$ and $\beta_k=\prod_{i\le k}\pi_i/(\pi_i-1)$. If $N$ is odd and squarefree with $k$ prime factors, then $N\ge \Omega_k$ and $N/\varphi(N)\le \beta_k$, so by \cref{lem:jacobsthal} \[ J(N)\le\mathcal J_k,\qquad\mathcal J_1=2,\quad\mathcal J_2=3,\quad\mathcal J_k=\lfloor(2^k-1)\beta_k\rfloor+1\ (k\ge3). \] For $k=0$ we have $N=1$ and $J(1)=1$, and we put $\Omega_0=1$, $\mathcal J_0=1$. \emph{Odd $n$.} \Cref{thm:O-large} applies unless $5n\le324J(n)^2$. Let $n\ge5$ satisfy this inequality, with radical $N$ having $k\ge1$ primes. Then $5\Omega_k\le5N\le5n\le324\mathcal J_k^2$. For $k=8,9,10$ one has $\mathcal J_k=780$, $1{,}618$, $3{,}347$ and $5\Omega_k>324\mathcal J_k^2$; for $k=11$, $5\Omega_{11}>324\cdot9^{11}$, and since $\pi_{k+1}>9$ for $k\ge11$, $5\Omega_k>324\cdot9^k\ge324J(N)^2$ for every $k\ge11$ (\cref{lem:jacobsthal}, $J(N)\le3^k$). So $k\le7$ and $5N\le324\mathcal J_k^2$. Each odd squarefree $N$ with $k\le7$ primes and $5N\le324\mathcal J_k^2$ is discarded if $5N>324E(N)^2$, where $E(N)$ is the least of the bounds of \cref{lem:jacobsthal} at $N$ itself; otherwise $J(N)$ is computed exactly, as the largest gap between consecutive integers prime to $N$, and every odd $n$ with radical $N$ and $5n\le324J(N)^2$ is listed. The result is the list of $320$ values $n\ge5$ of \cref{thm:O-list}, the largest $3{,}135$, with $J=7$. \emph{Even $n=2m$.} \Cref{thm:E-large} applies unless $m\le N_0(m)=P_2(m)(31J(m^\circ)^2-1)$. Let $m$ satisfy this inequality, with $k\ge0$ odd primes and odd radical $N=m^\circ$. Then $P_2(m)\le\pi_{k+2}$ and $m\ge N\ge \Omega_k$, so $\Omega_k\le\pi_{k+2}(31\mathcal J_k^2-1)$. This fails for $9\le k\le12$, by exact computation; and for $k\ge13$, $J(N)\le3^k$ gives $m<31\cdot9^k\pi_{k+2}$, which fails at $k=13$, while from $k$ to $k+1$ the left side $\Omega_k$ is multiplied by $\pi_{k+1}\ge47$ and the right side by $9\pi_{k+3}/\pi_{k+2}<18$ (Bertrand's postulate). So $k\le8$ and $N\le\pi_{k+2}(31\mathcal J_k^2-1)$. One enumerates every set of $k\le8$ odd primes whose product satisfies this bound (for $k=2$, the larger range given by the third bound of \cref{lem:jacobsthal}), discards a set when its product exceeds $P_2(31E^2-1)$, with $P_2$ the second least odd prime outside the set and $E$ the bound of \cref{lem:jacobsthal} at the set itself (for the empty set, $E=J(1)=1$), computes $J$ exactly for the rest by an exhaustive search over the ways of covering an initial run of integers by one residue class per prime, and lists every $m\ge3$ with exactly those odd primes and $m\le N_0(m)$. The result is the list of $2{,}567$ values of \cref{thm:E-list}, the largest $69{,}615=3^2\cdot5\cdot7\cdot13\cdot17$. \subsection{Enumerations of the criterion}\label{ssec:enumerations} \emph{Odd $n$.} For the $320$ values of \cref{ssec:exceptional} an enumeration of the window triangles finds that those without witness lie in $\famA$ or $\famB$ or are the five odd members of $\listE$. \emph{Even $n$.} For the $2{,}567$ values of $m$ of \cref{ssec:exceptional} an enumeration of the window triangles of denominator $2m$ with $pq\le4mJ^2$ finds that those without witness lie in $\famA$ or $\famC$, have $s\mid n$, $n/s$ odd and $2\le p\le s/2$ (and so lie in $\famB$, by \cref{prop:region}\ref{it:region-B}), or are the twenty-five even members of $\listE$. \emph{The thirty members of $\listE$.} Each has no witness, by a direct test of all usable units, and none lies in $\famA$, $\famB$ or $\famC$. \section{Residues for every denominator, and the proof of \texorpdfstring{\cref{thm:E-large}}{the theorem for large even denominators}}\label{app:even} This appendix proves \cref{thm:E-large}: for $m>N_0(m)$, a window triangle of denominator $n=2m$ without witness lies in $\famA$, $\famB$ or $\famC$. \Cref{ssec:common} proves the lemmas on residues that hold for every denominator; \cref{ssec:list-odd} uses them as well. \Cref{ssec:even-proof} proves the lemmas special to even $n$ and the theorem. The plan of the proof follows the statement in \cref{ssec:list-even}, and a table of its cases follows \cref{prop:C-if}. \subsection{Lemmas for every denominator}\label{ssec:common} In this subsection $n$ is any integer and $\Delta=\tri pqrn$ is a window triangle (so $n\ge5$), $s=2r-n$ and $L=n-2q=2p+s$; the thresholds of $S_p$, $S_q$, $S_r$ are $2p$, $2q$, $s$, and $2p+2q+s=n$. Let $n^\circ$ be the product of the odd primes dividing $n$ ($n^\circ=1$ if $n$ is a power of $2$), and $J=J(n^\circ)$. For $n=2m$ this is the $J=J(m^\circ)$ of \cref{ssec:list-even}, and for odd $n$ it is $J(n)$, as $J$ depends only on the primes dividing its argument. \begin{lemma}\label{lem:G-usable} The units $b$ modulo $n$ with $2b\equiv2\pmod n$ are $b=1$ and, if $4\mid n$, $b=1+n/2$. Such a unit lying in $S_p$ has $\res{bp}n=p$. If one lies in $S_r$, then $4\mid n$, $r$ is odd, $b=1+n/2$ and $\res{br}n=s/2$. \end{lemma} \begin{proof} $2b\equiv2\pmod n$ means $b\equiv1$ modulo $n/\gcd(2,n)$, so $b=1$, or $n$ is even and $b=1+n/2$, which is prime to $n/2$, and so a unit exactly when it is odd, that is when $n/2$ is even. $\res{1\cdot x}n=x$ for $x\in\{p,r\}$, and $r\ge s$, so $1\notin S_r$. For $4\mid n$, $(1+n/2)x\equiv x$ if $x$ is even and $\equiv x+n/2$ if $x$ is odd. So $\res{(1+n/2)p}n$ is $p$ or $p+n/2\ge2p$ (as $pn-x>n-\bar u\ge\bar v$, and $z\ge n-x+1>\bar v+\bar w\ge\bar w$, so $a\notin S_v\cup S_w$; conversely $n-xn$, so the sum is $2n$. Finally $y$ in the first interval gives $x+y\le\bar u+\bar v0$ and $\tau \mathcal{Y}-L\mathcal{X}\le\tau L-L\mathcal{X}$. Below: $L\mathcal{X}-\tau \mathcal{Y}\ge L\mathcal{X}-\tau(\mathcal{X}-1)=\tau'\mathcal{X}+\tau>0$ and $L\mathcal{X}-\tau \mathcal{Y}\le L\mathcal{X}-\tau$. Both are multiples of $n$ by \ref{it:G-cong}. \end{proof} \begin{lemma}\label{lem:G-box} Let $\epsilon=\gcd(2,n)$. If $pq>\epsilon nJ^2$, there is a witness in $S_p\cap S_q$. \end{lemma} \begin{proof} Let $\Lambda=\{(x,z)\in\Z^2:(x,z)\equiv(bp,bq)\pmod n\text{ for some }b\equiv0\pmod\epsilon\}$. As $\gcd(p,q,n)$ divides $r=n-p-q$, it is $1$, so $b\mapsto(bp,bq)\bmod n$ is injective on $\Z/n$; the $b\equiv0\pmod\epsilon$ form a subgroup of order $n/\epsilon$, so $\Lambda$ is a lattice of index $n^2/(n/\epsilon)=\epsilon n$ in $\Z^2$. The open box $|x|

4\epsilon n$, so by Minkowski's theorem it contains a point $(x,z)\ne0$ of $\Lambda$; let $b\equiv0\pmod\epsilon$ be such that $(x,z)\equiv(bp,bq)$. For $1\le i\le J$ let $a_i=1+ib$. Then $a_ip\equiv p+ix$ and $a_iq\equiv q+iz$, with $0n$, $j$ being Jacobsthal's function, by the same progression through $a=1$. \begin{lemma}\label{lem:G-chain} Let $x\in\{p,r\}$ and let $a_0$ be a unit of $S_x$ with $\mathcal{X}_0=\res{a_0x}n\ge1$ and $\mathcal{X}_0+\tau'\le \mathcal{Y}_0=\res{-a_0q}n\le L$. Let $u\ge2$ be an integer prime to $n$ with $uLn-q$; they are $i=\pm q^{-1}\bmod g$, and $g\nmid2i$ as $g\ge3$ and $\gcd(i,g)=1$. If $g^*=1$ both are units; if $g^*=\ell$ is a prime, $i$ and $-i$ lie in different classes modulo $\ell$, as $\ell$ is odd and $\ell\nmid i$, so one of them avoids the excluded class. (b) $t=1,\dots,t_{\max}$ give elements of $S_q$; they are at least $J(g^*)$ consecutive values, so, as in \cref{lem:G-rows}(iii), one of them gives a unit; and $0<2t73.5J^2$. \end{proof} \needspace{4\baselineskip} \begin{lemma}\label{lem:E-smallsigma} If $2\varsigma(p+\varsigma)1-1/T>3/7$; so $h=0$ and $w(4/7)T>M$, and $\varsigma-L/w\ge\varsigma-(2p+2\varsigma)/(2p+1)\ge\varsigma-(2\varsigma+2)/3=(\varsigma-2)/3>M$. So it contains $M$ consecutive integers, one of them a unit $\mathcal{X}$ of $n$ with $1\le \mathcal{X}<\varsigma$, and $a_{\mathcal{X}}$ is a witness. \emph{$k=2$.} Then $h=1$ and $\alpha>1/2-1/(2T)>3/7$, which is impossible. \emph{$k\ge3$.} Choose an integer $e$ prime to $k$: if $k\alphaL$; so $\mathcal{Y}_j>m/2$ and $\mathcal{Y}_{j+1}=2\mathcal{Y}_j-m$ for $j=0,1,2$. Then $\mathcal{Y}_1>m/2$ gives $\mathcal{Y}_0>3m/4$, and $\mathcal{Y}_2=4\mathcal{Y}_0-3m>m/2$ gives $\mathcal{Y}_0>7m/8\ge L$, a contradiction. \end{proof} \begin{lemma}\label{lem:E-three} Let $\gcd(\varsigma,m)=3$, $9\nmid m$, $3L>n$, $\varsigma\ge9$, $\varsigma>6p$ and $18p0$, since $\varsigma=m/3$ would give $\gcd(\varsigma,m)=m/3=3$ and $\varsigma=3$; so $3sn$ means $f<3p$, so $0<6p-f<6p<\varsigma$, and $\mathcal{Y}$ is not an integer; a contradiction. \end{proof} \begin{proof}[Proof of \cref{thm:E-large}] As $P_2\ge5$, $m>155J^2-5\ge150$; and $J=1$ only for $m$ a power of $2$, when $m\ge256$. Suppose $\Delta$ has no witness. \emph{Step 1: $p$ is small.} $m\ge74J^2$, so $p\le7J^2-1$ by \cref{cor:E-small}. Hence $21pn$, then $\varsigma>m/3-p>6p$ (as $21p\varsigma$, and we show that there is a witness in $S_p\cap S_q$. Then $L=2p+2\varsigma\le4p-2\le28J^2-6$, so $10L2m/3$. The case $p=2$ does not occur: in (i), $\varsigma=1324J^2$, then $pn/6$, $pq\ge p^2>n^2/36$. So $nJ^2\ge5n^2/324$, which contradicts $5n>324J^2$. \end{proof} \needspace{4\baselineskip} \begin{theorem}\label{thm:O-large} Let $n$ be odd with $5n>324J(n)^2$. If a window triangle $\tri pqrn$ has no witness, then $s\mid n$; if moreover $p\ge2$, then $2p\le s$ and $\Delta$ lies in $\famB$. \end{theorem} \begin{proof} Suppose $\Delta$ has no witness. By \cref{cor:O-small}, $ps$. Then $L<4p<2n/9$, so $3Ln/2-2p>7n/18>n/3\ge n/p$ (as $s<2p7p/18-1$ in \cref{lem:G-rowone} with $g=p$. If $g^*$ has at most one prime, part (a) of that lemma applies; if two, $p\ge15$ and $t_{\max}\ge5\ge J(g^*)$; if $k\ge3$, $p\ge105\cdot11^{k-3}$ and $t_{\max}>39\cdot11^{k-3}\ge3^k\ge J(g^*)$; in these two cases part (b) applies. It gives a witness. In each case this is a contradiction. \emph{The unit condition.} As $n$ is odd, $n/s$ is odd; with $s\mid n$ and $2\le p\le s/2$, \cref{prop:region}\ref{it:region-B} gives $(\mathrm{U})$, and $\Delta$ lies in $\famB$. \end{proof} For $p=1$ the theorem says that a window triangle $\tri1qrn$ without a witness has $s\mid n$, that is, lies in $\famA$. \needspace{4\baselineskip} \begin{theorem}\label{thm:O-list} Let $n$ be odd. A window triangle of denominator $n$ has no witness if and only if it lies in $\famA$ or in $\famB$, or is one of $\tri236{11}$, $\tri258{15}$, $\tri348{15}$, $\tri27{12}{21}$, $\tri37{11}{21}$. \end{theorem} \begin{proof} \emph{Only if.} If $5n>324J(n)^2$, this is \cref{thm:O-large}. The odd $n\ge5$ with $5n\le324J(n)^2$ are $320$ values, the largest $3{,}135=3\cdot5\cdot11\cdot19$, with $J=7$ ($n=3$ also satisfies the inequality and has no window triangle); for them an enumeration of all window triangles finds no witness-free triangle other than those stated (\cref{ssec:enumerations}; the list of $320$ values is proved complete in \cref{ssec:exceptional}). \emph{If.} Members of $\famA$ and of $\famB$ have no witness by \cref{prop:region}. The five single triangles have no witness by a direct check of the units (\cref{sec:computations}). \end{proof} For prime $n$, \cref{thm:O-list} gives the prime case stated after \cref{thm:list}: $s\mid n$ forces $s=1$, which excludes $\famB$, and the only single triangle of prime denominator is $\tri236{11}$. \section*{Declarations}\phantomsection\label{sec:declarations} \subsection*{Use of AI} This paper and its results were produced by an AI research system that the author designed, built and directs, running on Anthropic's Claude Fable~5.2 in Claude Code and OpenAI's GPT-6 Astra in Codex. The author chooses the problems and prioritizes the work. Separate AI agents checked each result, and a second program confirmed the computations the proofs use, with any exceptions listed in the public deposit. The system also searched the literature and checked the citations against reliable sources. The author read the paper, reran the computations from the deposit and takes full responsibility for all of the paper's contents. \subsection*{Data availability} The programs and data used in this paper, with instructions for rerunning them, are deposited at Zenodo~\cite{Record}. \bibliographystyle{amsplain} \bibliography{refs} \end{document}