\documentclass[11pt]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern,amsmath,amssymb,amsthm,mathtools} \usepackage[margin=1.05in]{geometry} \usepackage{enumitem,booktabs,needspace,array} \usepackage{tikz} \usetikzlibrary{patterns} \usepackage[colorlinks=true,urlcolor=blue,linkcolor=black,citecolor=black]{hyperref} \renewcommand{\UrlBreaks}{\do\/} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{conjecture}[theorem]{Conjecture} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{remark}[theorem]{Remark} \newtheorem{computation}{Computation} \newcommand{\Z}{\mathbb Z} \newcommand{\Q}{\mathbb Q} \newcommand{\R}{\mathbb R} \newcommand{\W}{\mathcal O} \newcommand{\Lf}{\mathcal L} \newcommand{\A}{\mathcal S} \newcommand{\conv}{\operatorname{conv}} \newcommand{\inter}{\operatorname{int}} \newcommand{\vol}{\operatorname{vol}} \newcommand{\GL}{\mathrm{GL}} \setlist[enumerate]{leftmargin=*,label=\textup{(\roman*)}} \emergencystretch=2em \raggedbottom \hypersetup{pdftitle={Wills' (1981) conjecture holds if and only if the dimension is at most 4, and other miscellanea},pdfauthor={Dave Goel},pdfkeywords={Ehrhart polynomial, lattice polytope, centrally symmetric, interior lattice point, Wills' conjecture}} \title[Wills' (1981) conjecture]{Wills' (1981) conjecture holds if and only if the dimension is at most 4, and other miscellanea}\thanks{$^{*}$Additionally, Theorems~\ref{thm:four}, \ref{thm:five} and~\ref{thm:equality} and Corollaries~\ref{cor:exact} and~\ref{cor:g2} are derived in Lean~4 from six published results (Ehrhart's theorem, the properties (E1)--(E3) of Ehrhart polynomials, the boundary triangulation used in Lemma~\ref{lem:few}, and Minkowski's theorem on lattice tilings used in Lemma~\ref{lem:frame}) and five computer checks (Computations~\ref{comp:cube}, \ref{comp:index2}, \ref{comp:lattices} and~\ref{comp:frame}, and the volume of~$W$), each stated as an explicit hypothesis, and checked with \texttt{lake comparator}:\newline\url{https://github.com/david44hello/paper_wills-1981-conjecture/tree/v3.32}\nopunct} \author{Dave Goel} \email{david44hello@gmail.com} \subjclass[2020]{Primary 52B20; Secondary 11H06, 52-08, 52C07} \keywords{Ehrhart polynomial, lattice polytope, centrally symmetric polytope, interior lattice point, Wills' conjecture} \date{} \dedicatory{\normalfont\footnotesize First version: November 3, 2026. This version (3.32): November 16, 2026.\\[2pt]\itshape This is the expanded version of the paper D.~Goel, \emph{Wills' (1981) conjecture holds if and only if the dimension is at most 4}~\cite{GoelShort}.} \begin{document} \begin{abstract} Let $P$ be a lattice polytope in $\R^n$ that is symmetric about the origin and has the origin as its only interior lattice point. In 1981 Wills conjectured that each coefficient $g_i(P)$ of the Ehrhart polynomial of $P$ is at most the corresponding coefficient of the cube $[-1,1]^n$, that is, $g_i(P)\le 2^i\binom ni$. Wills proved this for $n\le3$, and Henze showed in 2014 that it fails for every $n\ge7$; the dimensions $4$, $5$ and $6$ remained open. We settle them. In dimension $4$ the conjecture holds; the new bounds are $g_1(P)\le8$ and $g_2(P)\le24$, and, up to unimodular equivalence, $g_1(P)=8$ only for the cube, one explicit $24$-cell and one explicit cross-polytope. In dimension $5$ we give a polytope with $50$ vertices in $[-1,1]^5$ whose linear Ehrhart coefficient is $101/10>10$; products with segments then give counterexamples in every dimension $n\ge6$. Hence Wills' conjecture holds in $\R^n$ if and only if $n\le4$. The proof in dimension $4$ combines Ehrhart reciprocity, the lower bound theorem for triangulated $3$-spheres and an exhaustive computer enumeration of about three million sets of lattice points, listed up to symmetry. This expanded version also gives a second proof of the case in which the integer functionals bounded by $1$ on $P$ do not span $\R^4$, determines the second largest value of $g_2$ in dimension $4$, discusses a width-two approach, describes the computer search for the counterexample, and describes all computations used in the proofs.$^{*}$ \end{abstract} \maketitle \section{Introduction}\label{sec:intro} A \emph{lattice polytope} in $\R^n$ is the convex hull of finitely many points of $\Z^n$. Ehrhart~\cite{Ehr62} proved that for a lattice polytope $P$ of dimension $n$ the number of lattice points in the dilate $kP$ is a polynomial in the integer $k\ge0$, \[ L_P(k)=\#(kP\cap\Z^n)=g_0(P)+g_1(P)\,k+\dots+g_n(P)\,k^n , \] the \emph{Ehrhart polynomial} of $P$. Here $g_0(P)=1$, the leading coefficient $g_n(P)$ is the volume of $P$, and $g_{n-1}(P)$ is half the lattice surface area of $P$ (Section~\ref{sec:prelim}); the remaining coefficients have no such simple description and may be negative \cite{BR15}. Minkowski's first theorem~\cite{Min96} states that a convex body $K\subset\R^n$ that is symmetric about the origin and has no lattice point other than $0$ in its interior has volume at most $2^n$, the volume of the cube $C_n=[-1,1]^n$. Wills~\cite{Wil81} looked for a discrete analog of this theorem, one that bounds every Ehrhart coefficient of a lattice polytope. Let $\W^n$ be the set of lattice polytopes $P\subset\R^n$ with $P=-P$ whose only interior lattice point is the origin. The cube $C_n$ belongs to $\W^n$, and $L_{C_n}(k)=(2k+1)^n$, so $g_i(C_n)=2^i\binom ni$. Wills called these numbers the expected values of the coefficients \cite[Remark~(b)]{Wil81}, and the following statement is known as Wills' conjecture. We state it as in~\cite[Conjecture~1.1]{Hen14}; see also~\cite{GW93}. \begin{conjecture}[Wills' conjecture]\label{conj:wills} For every $n\ge1$, every $P\in\W^n$ and every $i\in\{0,1,\dots,n\}$, \[ g_i(P)\le g_i(C_n)=2^i\binom ni . \] \end{conjecture} For $i=0$ there is equality, and for $i=n$ the inequality is Minkowski's theorem. Wills proved the case $i=n-1$ in every dimension \cite[Theorem~(i)]{Wil81}; hence the conjecture holds for $n\le2$. Wills also proved it for $n=3$ \cite[Theorem~(i), (ii)]{Wil81}, and, in dimension $4$, the bounds $g_1(P)<21$, $g_2(P)<31$ and $g_2(P)\le g_1(P)+16$ \cite[Theorem~(iii)]{Wil81}. A discrete version of Minkowski's first theorem by Betke, Henk and Wills~\cite[Theorem~2.1]{BHW93} implies that $L_P(k)\le L_{C_n}(k)$ for every $P\in\W^n$ and every integer $k\ge0$, which is the inequality of the conjecture for the values of the two polynomials instead of their coefficients. For $i=n-1$, a bound in terms of the successive minima that implies Wills' bound is due to Henk, Sch\"urmann and Wills~\cite{HSW05}. Henze~\cite{Hen14} constructed a family of polytopes $P_n$ and showed that $P_7\in\W^7$ has $g_1(P_7)=1534/105>14$; taking products with segments, he obtained counterexamples in every dimension $n\ge7$. On the positive side, Henze proved the conjecture for every $P\in\W^n$ whose Ehrhart roots all have real part $-1/2$ \cite[Corollary~3.3]{Hen14}; by a theorem of Bey, Henk and Wills~\cite{BHW07}, every such polytope is reflexive. About the remaining dimensions, Henze wrote \cite[Remark~2.2]{Hen14}: ``In dimensions 4, 5 and 6 we tested Wills' conjecture on lattice polytopes that have a similar structure to $P_n$. [\dots] None of these examples turned out to be a counterexample.'' We settle these three dimensions. \begin{table}[ht] \centering\small \caption{The best results before this paper and the results of this paper.}\label{tab:results} \begin{tabular}{>{\raggedright\arraybackslash}p{0.43\linewidth}>{\raggedright\arraybackslash}p{0.47\linewidth}} \toprule Best result before this paper & New result\\ \midrule $g_1(P)<21$ \cite[Theorem~(iii)]{Wil81} & $g_1(P)\le8$ for every $P\in\W^4$, with equality exactly for $C_4$, a $24$-cell $T$ and a cross-polytope $X$, up to unimodular equivalence (Theorems~\ref{thm:four} and~\ref{thm:equality})\\\addlinespace $g_2(P)<31$ and $g_2(P)\le g_1(P)+16$ \cite[Theorem~(iii)]{Wil81} & $g_2(P)\le24$ for every $P\in\W^4$, with equality exactly for $C_4$, up to unimodular equivalence (Theorem~\ref{thm:four} and Corollary~\ref{cor:g2})\\\addlinespace no counterexample known for $n\le6$; the tested families in dimensions $4$, $5$, $6$ contain none \cite[Remark~2.2]{Hen14} & the polytope $W$ of Table~\ref{tab:witness} is in $\W^5$ and has $g_1(W)=101/10>10$; counterexamples in every dimension $n\ge5$ (Theorem~\ref{thm:five})\\\addlinespace true for $n\le3$ \cite{Wil81}; false for $n\ge7$ \cite{Hen14} & Conjecture~\ref{conj:wills} holds in $\R^n$ if and only if $n\le4$ (Corollary~\ref{cor:exact})\\ \bottomrule \end{tabular} \end{table} \needspace{4\baselineskip} \medskip\noindent\textbf{Results.} Our main results are Theorems~\ref{thm:four} and~\ref{thm:five} and Corollary~\ref{cor:exact} below, and Theorem~\ref{thm:equality} in Section~\ref{sec:equality}, which determines the polytopes $P\in\W^4$ with $g_1(P)=8$. Table~\ref{tab:results} compares them with the best earlier results, and Figure~\ref{fig:status} shows the status of the conjecture in each dimension. Two of the three polytopes with $g_1(P)=8$ have Ehrhart roots off the line $\operatorname{Re}z=-\frac12$, so Henze's theorem does not cover them (Remark~\ref{rem:roots}). \begin{theorem}\label{thm:four} Every $P\in\W^4$ satisfies $g_1(P)\le8$ and $g_2(P)\le24$. \end{theorem} \begin{theorem}\label{thm:five} The polytope $W\subset\R^5$ of Table~\ref{tab:witness} belongs to $\W^5$, and \[ L_W(k)=1+\tfrac{101}{10}\,k+\tfrac{127}{12}\,k^2+\tfrac{139}{6}\,k^3+\tfrac{557}{12}\,k^4+\tfrac{386}{15}\,k^5 . \] In particular $g_1(W)=101/10>10=g_1(C_5)$. For every $n\ge6$ the polytope $W\times C_{n-5}$ belongs to $\W^n$ and has $g_1=2n+1/10>2n=g_1(C_n)$. \end{theorem} \needspace{4\baselineskip}\begin{corollary}\label{cor:exact} Wills' conjecture holds in $\R^n$ if and only if $n\le4$. \end{corollary} \begin{proof} For $n\le3$ this is Wills' theorem \cite{Wil81}. For $n=4$, $g_0=1$ by (E1) of Section~\ref{sec:prelim}, $g_3\le32$ by \cite[Theorem~(i)]{Wil81}, $g_4\le16$ by Minkowski's theorem, and $g_1$, $g_2$ are bounded by Theorem~\ref{thm:four}. For $n\ge5$, Theorem~\ref{thm:five} gives a counterexample. \end{proof} \begin{figure}[ht] \centering \begin{tikzpicture}[x=1.05cm,y=0.62cm,font=\small] \foreach \row/\lab in {1/{Known before this paper},0/{After this paper}}{ \node[anchor=east] at (0.3,\row*1.9) {\lab}; } \fill[green!18] (0.6,1.5) rectangle (3.6,2.3); \fill[pattern=north east lines] (3.6,1.5) rectangle (6.6,2.3); \fill[red!18] (6.6,1.5) rectangle (10.6,2.3); \node at (2.1,1.9) {true (Wills \cite{Wil81})}; \node[fill=white,inner sep=2pt] at (5.1,1.9) {open}; \node at (8.6,1.9) {false (Henze \cite{Hen14})}; \draw (0.6,1.5) rectangle (10.6,2.3); \fill[green!18] (0.6,-0.4) rectangle (3.6,0.4); \fill[green!55!black] (3.6,-0.4) rectangle (4.6,0.4); \fill[red!18] (6.6,-0.4) rectangle (10.6,0.4); \fill[red!70!black] (4.6,-0.4) rectangle (6.6,0.4); \draw (0.6,-0.4) rectangle (10.6,0.4); \draw[line width=1.6pt] (3.6,-0.4) rectangle (6.6,0.4); \node[text=white] at (4.1,0) {\textbf{true}}; \node[text=white] at (5.6,0) {\textbf{false}}; \node at (2.1,0) {true}; \node at (8.6,0) {false}; \foreach \n in {1,...,9}{ \node at (\n-0.4+0.5,-0.95) {$\n$}; } \node at (10.15,-0.95) {$\cdots$}; \node[anchor=east] at (0.3,-0.95) {dimension $n$}; \end{tikzpicture} \caption{The status of Wills' conjecture in each dimension $n$. Before this paper the conjecture was known to hold for $n\le3$ (Wills~\cite{Wil81}) and to fail for $n\ge7$ (Henze~\cite{Hen14}); the dimensions $4$, $5$ and $6$ (hatched) were open. This paper proves it for $n=4$ (Theorem~\ref{thm:four} and Corollary~\ref{cor:exact}) and disproves it for every $n\ge5$ (Theorem~\ref{thm:five}); the dimensions settled here are drawn in full colour, inside the thick frame.}\label{fig:status} \end{figure} \medskip\noindent\textbf{Idea of the proofs.} Theorem~\ref{thm:five} is a computation. The polytope $W$ is given by its $50$ vertices, and since $W$ has only one interior lattice point, its linear coefficient is determined by four numbers: the numbers of lattice points in $W$, in $2W$ and in the interior of $2W$, and the volume of $W$ (Proposition~\ref{prop:fournumbers}). Products with segments carry the example to every higher dimension. For Theorem~\ref{thm:four} we use Ehrhart reciprocity. Since $P\in\W^4$ has exactly one interior lattice point, $L_P(-1)=1$; comparing this with $L_P(1)=\#(P\cap\Z^4)$ shows that $g_1(P)\le8$ is equivalent to the inequality \[ \#(P\cap\Z^4)\le \sigma(P)+17 \] between the number of lattice points of $P$ and its lattice surface area $\sigma(P)=2g_3(P)$, and that this inequality also gives $g_2(P)\le24$ (Proposition~\ref{prop:reduction}). A triangulation of the boundary of $P$ that uses every boundary lattice point, together with Walkup's lower bound theorem for triangulated $3$-spheres~\cite{Wal70}, proves the inequality when $P$ has at most $29$ lattice points (Lemma~\ref{lem:few}). For $P$ with at least $31$ lattice points (their number is odd, since they come in pairs $\pm x$ together with $0$) we look at the integer linear functionals that take values in $[-1,1]$ on $P$. Either they contain a basis of $\Z^4$, and then $P$ is unimodularly equivalent to a polytope in the cube $C_4$; or they span $\R^4$ without containing a basis, and then a linear map carries $P$, together with its lattice, onto a polytope in $C_4$ with vertices in a sublattice of index $2$; or they span a proper subspace. In the last case we fix four lattice points of $P$ whose determinant is as large as possible; in their coordinates the lattice points of $P$ lie in $[-1,1]^4$, in one of $39$ lattices, up to signed permutations of the coordinates (Section~\ref{sec:cases}). This reduction, by four lattice points of largest determinant, follows Averkov and Schymura~\cite{AS24}, who used it to compute the largest number of columns of a $\Delta$-modular integer matrix for small parameters; their enumeration builds on that of Averkov, Borger and Soprunov~\cite{ABS21}. In each case only finitely many sets of lattice points remain, and they are enumerated by computer (Section~\ref{sec:computations}). Section~\ref{sec:prelim} collects the facts we use from Ehrhart theory. Section~\ref{sec:five} proves Theorem~\ref{thm:five}. Sections~\ref{sec:reduction}--\ref{sec:computations} prove Theorem~\ref{thm:four}, and Section~\ref{sec:equality} identifies the equality cases. Beyond the contents of the short paper~\cite{GoelShort}, Section~\ref{sec:rank} gives a second proof of Case~III, Section~\ref{sec:extremal} studies the extremal polytopes and the next values of $g_1$ and $g_2$, Section~\ref{sec:width} describes a width-two approach, Section~\ref{sec:search} describes how the counterexample was found, and Section~\ref{sec:programs} describes the computations in detail. \section{Preliminaries}\label{sec:prelim} \subsection{Ehrhart polynomials} Let $n\ge1$, let $\Lambda\subset\R^n$ be a lattice (a discrete subgroup of rank $n$), and let $P$ be a polytope of dimension $n$ whose vertices lie in $\Lambda$. Then $L_{P,\Lambda}(k)=\#(kP\cap\Lambda)=\sum_ig_i(P)k^i$ is a polynomial in $k$ of degree $n$ \cite{Ehr62}; its coefficients $g_i(P)$ depend on $\Lambda$, which will be clear from the context, and we write $L_P$ when $\Lambda=\Z^n$. We call two pairs $(P,\Lambda)$ and $(P',\Lambda')$ \emph{isomorphic} if a linear isomorphism $\phi$ of $\R^n$ satisfies $\phi(P)=P'$ and $\phi(\Lambda)=\Lambda'$; then $\phi(kP)\cap\Lambda'=\phi(kP\cap\Lambda)$ for every $k$, so $L_{P',\Lambda'}=L_{P,\Lambda}$. When $\Lambda=\Lambda'=\Z^n$, the map $\phi$ lies in $\GL_n(\Z)$, and the two lattice polytopes are called \emph{unimodularly equivalent}. We use four standard facts \cite[Chapters~3--5]{BR15}. \begin{enumerate} \item[(E1)] $g_0(P)=1$ and $g_n(P)=\vol(P)/\det\Lambda$. \item[(E2)] \emph{Reciprocity} (Ehrhart, Macdonald): $L_{P,\Lambda}(-k)=(-1)^n\,\#(\inter(kP)\cap\Lambda)$ for every integer $k\ge1$. \item[(E3)] For a facet $F$ of $P$ let $u_F$ be the primitive vector of the dual lattice $\Lambda^*$ that is an outer normal of $F$, and let $b_F$ be the integer with $F\subset\{x:\langle u_F,x\rangle=b_F\}$. Let $\operatorname{relvol}(F)$ be the $(n-1)$-dimensional volume of $F$, measured so that a fundamental domain of the lattice $\Lambda\cap\operatorname{aff}(F-F)$ has volume $1$. Then $g_{n-1}(P)=\frac12 \sigma(P)$, where $\sigma(P)=\sum_F\operatorname{relvol}(F)$ is the \emph{lattice surface area}. If $0\in\inter P$, then $\vol(P)/\det\Lambda=\frac1n\sum_F b_F\operatorname{relvol}(F)$, by cutting $P$ into the pyramids with apex $0$ over its facets. \item[(E4)] $L_{P\times Q}=L_P\,L_Q$ for lattice polytopes $P\subset\R^p$ and $Q\subset\R^q$, with the lattice $\Z^{p+q}$, since $k(P\times Q)\cap\Z^{p+q}=(kP\cap\Z^p)\times(kQ\cap\Z^q)$. \end{enumerate} In (E3), $b_F$ is an integer because $F$ contains points of $\Lambda$; when $0\in\inter P$ it is positive, and it is the lattice distance of $F$ from the origin. \subsection{The class \texorpdfstring{$\W^n$}{W\^{}n}} Let $\W^n_\Lambda$ be the set of polytopes $P$ with vertices in $\Lambda$, $P=-P$ and $\inter P\cap\Lambda=\{0\}$; so $\W^n=\W^n_{\Z^n}$. A polytope in $\W^n_\Lambda$ has dimension $n$, because $0$ is an interior point. We write $G(P)=\#(P\cap\Lambda)$. Since the lattice points of $P$ other than $0$ come in pairs $\pm x$, $G(P)$ is odd. \begin{lemma}\label{lem:member} Let $Y\subset\{-1,0,1\}^n$ be a set with $Y=-Y$ whose convex hull $P$ has dimension $n$. Then $P\in\W^n$. \end{lemma} \begin{proof} The polytope $P$ is a lattice polytope with $P=-P$. It has an interior point $x$, so $-x$ is interior too, and so is $0=\frac12x+\frac12(-x)$. Finally $\inter P\subset\inter C_n=(-1,1)^n$, whose only lattice point is $0$. \end{proof} The same proof applies to a sublattice $\Gamma\subseteq\Z^n$ in place of $\Z^n$: if $S\subseteq\Gamma\cap C_n$ is symmetric and spans $\R^n$, then $\conv S\in\W^n_\Gamma$. We use this in Sections~\ref{sec:cases} and~\ref{sec:computations}. \begin{lemma}[{cf.\ \cite[Proposition~2.3]{Hen14}}]\label{lem:product} If $P\in\W^p$ and $Q\in\W^q$, then $P\times Q\in\W^{p+q}$ and $g_1(P\times Q)=g_1(P)+g_1(Q)$. \end{lemma} \begin{proof} The product $P\times Q$ is a symmetric lattice polytope, and $\inter(P\times Q)=\inter P\times\inter Q$ contains only the lattice point $0$. By (E4) and (E1), the coefficient of $k$ in $L_PL_Q$ is $g_1(P)\cdot1+1\cdot g_1(Q)$. \end{proof} \section{A counterexample in dimension five}\label{sec:five} We compute four numbers for the polytope $W$ of Table~\ref{tab:witness}, three lattice-point counts and the volume; by Proposition~\ref{prop:fournumbers} they determine $g_1(W)$. \begin{table}[ht] \centering\small \caption{The polytope $W\subset\R^5$ is the convex hull of the $50$ points $\pm p$, where $p$ runs over the $25$ points below ($\bar1$ stands for $-1$). All $50$ points are vertices of $W$, and $W$ has $92$ facets.}\label{tab:witness} \begin{tabular}{ccccc} $(0,0,1,\bar1,\bar1)$ & $(0,0,1,1,\bar1)$ & $(0,1,\bar1,0,0)$ & $(0,1,0,1,\bar1)$ & $(0,1,0,1,1)$\\ $(0,1,1,\bar1,\bar1)$ & $(0,1,1,\bar1,0)$ & $(1,\bar1,\bar1,\bar1,0)$ & $(1,\bar1,0,1,0)$ & $(1,\bar1,0,1,1)$\\ $(1,\bar1,1,0,\bar1)$ & $(1,0,\bar1,0,\bar1)$ & $(1,0,\bar1,1,\bar1)$ & $(1,0,\bar1,1,1)$ & $(1,0,0,\bar1,1)$\\ $(1,0,0,1,1)$ & $(1,0,1,\bar1,0)$ & $(1,0,1,\bar1,1)$ & $(1,0,1,0,\bar1)$ & $(1,1,\bar1,\bar1,\bar1)$\\ $(1,1,\bar1,\bar1,0)$ & $(1,1,\bar1,1,0)$ & $(1,1,1,\bar1,0)$ & $(1,1,1,\bar1,1)$ & $(1,1,1,1,0)$\\ \end{tabular} \end{table} \begin{proposition}\label{prop:fournumbers} Let $P\in\W^5$. Then \[ g_1(P)=\frac{4\bigl(G(P)+1\bigr)+24\vol(P)-A(P)}{6},\qquad A(P)=\tfrac12\bigl(\#(2P\cap\Z^5)+\#(\inter(2P)\cap\Z^5)\bigr). \] \end{proposition} \begin{proof} Write $L=L_P$ and $g_i=g_i(P)$. By (E2) with $n=5$, $L(-1)=-1$ and $L(-2)=-\#(\inter(2P)\cap\Z^5)$. Hence \[ L(1)-L(-1)=2(g_1+g_3+g_5)=G(P)+1,\qquad L(2)-L(-2)=2(2g_1+8g_3+32g_5)=2A(P). \] Eliminating $g_3$ and using $g_5=\vol(P)$ (E1) gives $6g_1=4(G(P)+1)+24\vol(P)-A(P)$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:five}] The $50$ points of Table~\ref{tab:witness} lie in $\{-1,0,1\}^5$, form a symmetric set and span $\R^5$; by Lemma~\ref{lem:member}, $W\in\W^5$. Direct computation gives \[ G(W)=117,\qquad \#(2W\cap\Z^5)=1815,\qquad \#(\inter(2W)\cap\Z^5)=243,\qquad \vol(W)=\tfrac{386}{15}, \] so $A(W)=1029$ and, by Proposition~\ref{prop:fournumbers}, $g_1(W)=\bigl(4\cdot118+24\cdot\frac{386}{15}-1029\bigr)/6=\frac{101}{10}$. The whole polynomial is determined by $L_W(0)=1$ and the values $L_W(k)$ for $k=1,\dots,5$, namely $117$, $1815$, $10765$, $39927$ and $112639$. All numbers in this proof can be computed with Normaliz~\cite{Normaliz} or LattE~\cite{LattE}; the input files are in~\cite{Repo}. For $n\ge6$, Lemma~\ref{lem:product} gives $W\times C_{n-5}\in\W^n$ and $g_1(W\times C_{n-5})=\frac{101}{10}+2(n-5)=2n+\frac1{10}$, since $g_1(C_m)=2m$. \end{proof} The polytope $W$ was found by a computer search; the search and a first example, with $58$ vertices, are described in Section~\ref{sec:search}. \section{Dimension four: the key inequality\texorpdfstring{\\}{ } and polytopes with few lattice points}\label{sec:reduction} From now on $n=4$. For $P\in\W^4_\Lambda$ we write $G=G(P)$, $g_i=g_i(P)$ (with respect to $\Lambda$) and $\sigma=\sigma(P)$. \begin{lemma}\label{lem:identities} For $P\in\W^4_\Lambda$, $\;g_1+g_3=g_2+g_4=\frac{G-1}{2}$. \end{lemma} \begin{proof} We have $L(1)=G=1+g_1+g_2+g_3+g_4$, and by (E2) $L(-1)=\#(\inter P\cap\Lambda)=1=1-g_1+g_2-g_3+g_4$. Subtracting the second equation from the first, and adding the two, gives the two identities. \end{proof} \begin{proposition}\label{prop:reduction} Let $P\in\W^4_\Lambda$. Then $g_4\le16$ and $g_3\le 2g_4\le32$. Moreover \[ g_1\le8\iff G\le \sigma+17, \] and if $g_1\le8$ then $g_2\le g_4+8\le24$. \end{proposition} \begin{proof} The bound $g_4=\vol(P)/\det\Lambda\le16$ is Minkowski's theorem, applied in the coordinates of a basis of $\Lambda$. Since every $b_F\ge1$, (E3) gives $2g_3=\sigma\le\sum_Fb_F\operatorname{relvol}(F)=4g_4$; this is Wills' argument for $g_{n-1}$ \cite[Theorem~(i)]{Wil81}. By Lemma~\ref{lem:identities}, $g_1=\frac{G-1}2-g_3$, so $g_1\le8$ if and only if $G\le2g_3+17=\sigma+17$. If this holds, then $G\le4g_4+17$, and Lemma~\ref{lem:identities} gives $g_2=\frac{G-1}2-g_4\le g_4+8\le24$. \end{proof} So Theorem~\ref{thm:four} is equivalent to the key inequality $G\le \sigma+17$ on $\W^4$. Let $\operatorname{NVol}(F)=3!\operatorname{relvol}(F)$ be the normalized volume of a facet, so that a lattice tetrahedron in $\operatorname{aff}F$ has normalized volume a positive integer. The inequality reads \begin{equation}\label{eq:star} \sum_F\operatorname{NVol}(F)\ \ge\ 6G-102 . \end{equation} \needspace{4\baselineskip} \begin{lemma}\label{lem:few} If $P\in\W^4_\Lambda$ and $G\le29$, then~\eqref{eq:star} holds with strict inequality; hence $g_1<8$. \end{lemma} \begin{proof} Take a triangulation $\Delta$ of the boundary of $P$ whose vertex set is the set of all $G-1$ boundary lattice points and in which every tetrahedron lies in a facet. To construct one, take a triangulation of $P$ whose vertices are exactly the points of $P\cap\Lambda$; one exists, for example by strong pulling (\cite[Lemma~2.1(2)]{HPPS21}, applied in the coordinates of a basis of $\Lambda$). Its faces that lie in the boundary of $P$ form $\Delta$. Then $\Delta$ is a triangulated $3$-sphere with $f_0=G-1$ vertices. Each of its $f_3$ tetrahedra is a lattice tetrahedron in the affine hull of a facet, so it has normalized volume at least $1$, and $\sum_F\operatorname{NVol}(F)\ge f_3$. Walkup's lower bound theorem~\cite[Theorem~1]{Wal70} gives $f_1\ge4f_0-10$ for every triangulated $3$-sphere, and the Dehn--Sommerville relations $f_0-f_1+f_2-f_3=0$, $2f_2=4f_3$ give $f_3=f_1-f_0$. Hence $\sum_F\operatorname{NVol}(F)\ge3f_0-10=3G-13$, and $3G-13>6G-102$ because $G\le29$. \end{proof}\par The proof gives a bound on $g_1$ for every $P\in\W^4_\Lambda$: by Lemma~\ref{lem:identities} and (E3), $g_1=\frac{G-1}2-\frac\sigma2$, and $\sigma=\frac16\sum_F\operatorname{NVol}(F)\ge\frac{3G-13}6$, so $g_1\le\frac{3G+7}{12}$ (the solid line in Figure~\ref{fig:maxg1}). \begin{remark}\label{rem:three} A similar triangulation argument, with Euler's formula in place of Walkup's theorem, proves Wills' theorem in dimension $3$. For $P\in\W^3$, $L(1)=G$ and $L(-1)=-1$ give $g_2=\frac{G-3}2$ and $g_1+g_3=\frac{G+1}2$. A triangulation of the boundary by all $G-1$ boundary lattice points has $2G-6$ triangles, by Euler's formula, and the cones over them from $0$ are lattice tetrahedra of normalized volume at least $1$; hence $6g_3\ge2G-6$. By Minkowski's theorem, as in Proposition~\ref{prop:reduction}, $g_3=\vol(P)\le8$. This gives $G\le27$, hence $g_2\le12$, and $g_1=\frac{G+1}2-g_3\le\frac{G+1}2-\frac{G-3}3=\frac{G+9}6\le6$. In dimension $4$ the lower bound theorem proves~\eqref{eq:star} only when $G\le29$; this is why the computations of Section~\ref{sec:computations} are needed. \end{remark} \section{Dimension four: three cases}\label{sec:cases} By Lemma~\ref{lem:few} it remains to prove~\eqref{eq:star} for $P\in\W^4$ with $G\ge31$. For such $P$ let \[ M(P)=\{u\in\Z^4:\ |\langle u,x\rangle|\le1\ \text{for all }x\in P\},\qquad r(P)=\dim\operatorname{span}M(P). \] The set $M(P)$ is finite (since $P$ contains a ball around $0$) and symmetric, and $\langle u,x\rangle\in\{-1,0,1\}$ for $u\in M(P)$ and $x\in P\cap\Z^4$. We distinguish three cases: $M(P)$ contains a basis of $\Z^4$ (Case~I); $r(P)=4$ but $M(P)$ contains no basis (Case~II); $r(P)\le3$ (Case~III). Figure~\ref{fig:cases} shows how each case is settled. \begin{figure}[ht] \centering \begin{tikzpicture}[font=\fontsize{9}{11}\selectfont,box/.style={draw,rounded corners,align=flush center,inner sep=3pt,text width=3.1cm,anchor=north},hand/.style={fill=blue!12},comp/.style={fill=orange!18},>=stealth] \node[box,text width=1.6cm] (root) at (0,0) {$P\in\W^4$}; \node[box,hand,text width=2.2cm] (few) at (-6.0,-0.9) {$G\le29$:\\ Lemma~\ref{lem:few}}; \node[box,comp] (one) at (-2.4,-0.9) {$G\ge31$, Case~I:\\ $M(P)$ has a basis;\\ $P$ equivalent to a\\ polytope in $C_4$\\ $\to$ Computation~\ref{comp:cube}}; \node[box,comp] (two) at (1.4,-0.9) {$G\ge31$, Case~II:\\ $r(P)=4$, no basis;\\ index-$2$ sublattice\\ $\to$ Computations~\ref{comp:index}, \ref{comp:index2}}; \node[box,comp] (three) at (5.2,-0.9) {$G\ge31$, Case~III:\\ $r(P)\le3$;\\ $39$ classes of\\ frame lattices\\ $\to$ Computations~\ref{comp:lattices}, \ref{comp:frame}}; \foreach \t in {few,one,two,three} \draw[->] (root) -- (\t.north); \end{tikzpicture} \caption{The proof of Theorem~\ref{thm:four} splits $\W^4$ into four parts. Polytopes with at most $29$ lattice points are handled by Lemma~\ref{lem:few} (the blue box, on the left); for the others (the three orange boxes), the integer functionals bounded by $1$ in absolute value on $P$ decide which finite enumeration of Section~\ref{sec:computations} contains $P$, up to isomorphism (Section~\ref{sec:prelim}). The frame lattices are defined after Lemma~\ref{lem:frame}.}\label{fig:cases} \end{figure} \subsection{Case I} For a lattice $\Gamma$, a set $S\subseteq\Gamma$ is called \emph{$\Gamma$-saturated} if $\conv S\cap\Gamma=S$; when the lattice is clear, we say \emph{saturated}. If $M(P)$ contains a basis $u_1,\dots,u_4$ of $\Z^4$, let $U\in\GL_4(\Z)$ have rows $u_i$. Then $UP\subset C_4$, so $UP=\conv S$ for the set $S=UP\cap\Z^4\subset\{-1,0,1\}^4$, which is symmetric and saturated. \subsection{Case II} Let $r(P)=4$, and let $u_1,\dots,u_4\in M(P)$ be linearly independent, with $U$ the matrix of rows $u_i$ and $d=|\det U|$. Then $d\ge2$, since otherwise $M(P)$ would contain a basis. The map $x\mapsto Ux$ carries $\Z^4$ onto the sublattice $\Gamma=U\Z^4$ of index $d$ and carries $P$ into $C_4$; so $(P,\Z^4)$ is isomorphic to $(UP,\Gamma)$, $UP\in\W^4_\Gamma$, $L_{UP,\Gamma}=L_P$, and $UP=\conv S$ for the $\Gamma$-saturated symmetric set $S=UP\cap\Gamma\subset\Gamma\cap\{-1,0,1\}^4$. In particular \[ 31\le G(P)\le\#(\Gamma\cap\{-1,0,1\}^4). \] Moreover $d\le16$: the set $S$ spans $\R^4$, so it contains four linearly independent points of $\Gamma\cap\{-1,0,1\}^4$; their determinant is a nonzero multiple of $d$, and it is at most $16$ in absolute value by Hadamard's inequality. The sublattices of index $2$ are $\Gamma_I=\{x:\sum_{i\in I}x_i\text{ even}\}$ for nonempty $I\subseteq\{1,2,3,4\}$. A point of $\{-1,0,1\}^4$ lies in $\Gamma_I$ exactly when an even number of its coordinates in $I$ are nonzero, so \[ \#(\Gamma_I\cap\{-1,0,1\}^4)=3^{4-m}\cdot\frac{3^m+(-1)^m}{2}=27,\ 45,\ 39,\ 41\qquad(m=|I|=1,2,3,4). \] Computation~\ref{comp:index} shows that a sublattice of index $3\le d\le16$ contains at most $27$ points of $\{-1,0,1\}^4$ whenever these points span $\R^4$. Since $27<31$, in Case~II we have $d=2$ and $\Gamma=\Gamma_I$ with $|I|\in\{2,3,4\}$. \subsection{Case III: the frame} Now let $r(P)\le3$. Choose lattice points $v_1,\dots,v_4$ of $P$ for which $D=|\det(v_1,\dots,v_4)|$ is as large as possible, and let $V$ be the matrix with columns $v_i$. We call $(v_1,\dots,v_4)$ a \emph{frame} of $P$. Since $P$ spans $\R^4$, $D\ge1$. The map $x\mapsto V^{-1}x$ carries $\Z^4$ onto the lattice $\Lambda=V^{-1}\Z^4\supseteq\Z^4$, of index $D$ over $\Z^4$, and carries $P\cap\Z^4$ onto a set $S$. \begin{lemma}\label{lem:frame} In this situation: \begin{enumerate} \item $S\subset\Lambda\cap C_4$, $S=-S$, and $S$ contains the unit vectors $\pm e_1,\dots,\pm e_4$; \item $|\det(s_1,\dots,s_4)|\le1$ for all $s_1,\dots,s_4\in S$; \item $\conv S\cap\Lambda=S$ and $\inter(\conv S)\cap\Lambda=\{0\}$; \item $\Lambda$ contains no point $c\ne0$ with $|c_1|+\dots+|c_4|<1$; \item $D\le23$. \end{enumerate} \end{lemma} \begin{proof} For $x_1,\dots,x_4\in P\cap\Z^4$ we have $|\det(V^{-1}x_1,\dots,V^{-1}x_4)|=|\det(x_1,\dots,x_4)|/D\le1$ by the choice of $v_1,\dots,v_4$, which gives (ii). For (i), write a lattice point $x$ of $P$ as $x=\sum c_iv_i$; by Cramer's rule $|c_i|=|\det(v_1,\dots,x,\dots,v_4)|/D\le1$. Property (iii) holds because $x\mapsto V^{-1}x$ carries $\conv(P\cap\Z^4)\cap\Z^4=P\cap\Z^4$ and $\inter P\cap\Z^4=\{0\}$ to the corresponding statements for $S$ and $\Lambda$. For (iv), the open cross-polytope $\{c:|c_1|+\dots+|c_4|<1\}$ lies in the interior of $\conv\{\pm e_i\}\subseteq\conv S$, so by (iii) its only point of $\Lambda$ is $0$. For (v), $P$ contains the cross-polytope $\conv\{\pm v_i\}$ of volume $\frac{2^4}{4!}D=\frac23D$, and $\vol(P)\le16$, so $D\le24$. If $D=24$, then $P=\conv\{\pm v_i\}$ and $\vol P=16$. The translates $\frac12P+z$, $z\in\Z^4$, have disjoint interiors (an overlap would put a nonzero lattice point in $\inter P$) and total density $\vol(\frac12P)=1$, so they tile $\R^4$: the union of the translates is closed, and its complement is open, invariant under $\Z^4$ and of density $0$; a nonempty such set would contain a ball and all its $\Z^4$-translates, hence have positive density. A convex polytope that tiles space by lattice translations has centrally symmetric facets (Minkowski~\cite{Min97}; see Venkov~\cite{Ven54} and McMullen~\cite{McM80}), but the facets of a cross-polytope are tetrahedra. So $D\le23$. \end{proof} Let $\Lf$ be the set of lattices $\Lambda\supseteq\Z^4$ of index at most $23$ over $\Z^4$ that satisfy (iv); a lattice of index $D$ over $\Z^4$ lies in $\frac1D\Z^4$, so $\Lf$ is finite. We call the lattices in $\Lf$ the \emph{frame lattices}. For $\Lambda\in\Lf$ let $\A(\Lambda)$ be the family of sets $S$ satisfying (i)--(iii). The signed permutation matrices act on $\Lf$, and Computation~\ref{comp:lattices} shows that $\Lf$ consists of $39$ orbits; we fix one representative in each. Integer functionals transfer as well. For a set $S\subset\Lambda$ let $M(S)=\{w\in\Lambda^*:|\langle w,s\rangle|\le1\text{ for }s\in S\}$ and $r(S)=\dim\operatorname{span}M(S)$. Since $\langle u,Vs\rangle=\langle V^{\mathsf T}u,s\rangle$ and $V^{\mathsf T}\Z^4=\Lambda^*$, a vector $u\in\Z^4$ lies in $M(P)$ if and only if $V^{\mathsf T}u$ lies in $M(S)$; so $r(P)=r(S)$. In Case~III we therefore have a lattice $\Lambda$ among the $39$ representatives and a set $S\in\A(\Lambda)$ with $\#S\ge31$ and $r(S)\le3$ such that $(P,\Z^4)$ is isomorphic to $(\conv S,\Lambda)$ (by $x\mapsto V^{-1}x$, followed by a signed permutation of the coordinates); in particular $L_{\conv S,\Lambda}=L_P$. Lemma~\ref{lem:frame} does not use $r(P)\le3$, so this reduction applies to every $P\in\W^4$. We use it only in Case~III, where the condition $r(S)\le3$ shortens the search of Computation~\ref{comp:frame}; the polytopes with $r(P)=4$ are listed directly, in Cases~I and~II. \section{The computations}\label{sec:computations} All enumerations below are exhaustive and use exact integer and rational arithmetic; the repository~\cite{Repo} contains the programs and, for each computation, its output. Computations~\ref{comp:cube}, \ref{comp:index2} and~\ref{comp:frame} list sets of lattice points up to the symmetries named. For each set $S$ the Ehrhart polynomial of $\conv S$ is computed exactly: in Computations~\ref{comp:cube} and~\ref{comp:index2} from $G=\#S$ and the volume and the lattice surface area of $\conv S$ (taken from a triangulation of its boundary), by Lemma~\ref{lem:identities}, (E1) and (E3); in Computation~\ref{comp:frame} from the numbers of lattice points of $k\conv S$ for $k=1,2,3$ and of the interior of $k\conv S$ for $k=1,2$, which give its values at $k=-2,\dots,3$ by (E1) and (E2) and so determine it. Figure~\ref{fig:maxg1} shows, for each $G\ge31$, the largest value of $g_1$ in Case~I (Computation~\ref{comp:cube}) and the three equality cases of Theorem~\ref{thm:equality}, and, for $G\le29$, the bound on $g_1$ stated after the proof of Lemma~\ref{lem:few}. Computations~\ref{comp:cube} and~\ref{comp:index2} walk downward from the full set, $\{-1,0,1\}^4$ or $\Gamma_I\cap\{-1,0,1\}^4$, deleting one pair of opposite vertices at a time. By the next lemma, such a walk reaches every saturated symmetric subset of the full set. Sets that a symmetry carries onto each other form a class, and the walk keeps one set of each class; classes are recognized by a canonical form, the least image of the set under the symmetry group in a fixed total order. This loses no class: a symmetry $\gamma$ maps saturated symmetric sets to saturated symmetric sets and the vertices of $\conv Q$ to those of $\conv\gamma Q$, so the sets obtained from $\gamma Q$ by deleting a pair of opposite vertices are the images under $\gamma$ of those obtained from $Q$. Computation~\ref{comp:frame} uses a different method, a sandwich search, described after its statement; Section~\ref{sec:programs} describes a second search for it, which first finds the maximal members of $\A(\Lambda)$ and then walks downward from them. \begin{lemma}\label{lem:walk} Let $\Lambda$ be a lattice and $Q\subset\Lambda$ a finite symmetric set with $\conv Q\cap\Lambda=Q$. \begin{enumerate} \item If $v$ is a vertex of $\conv Q$, then $Q'=Q\setminus\{v,-v\}$ is symmetric and satisfies $\conv Q'\cap\Lambda=Q'$. \item If $S\subsetneq Q$ is symmetric with $\conv S\cap\Lambda=S$, then some vertex $v$ of $\conv Q$ is not in $S$, and $S\subseteq Q\setminus\{v,-v\}$. \end{enumerate} \end{lemma} \begin{proof} (i) Since $v$ is a vertex, $v\notin\conv(Q\setminus\{v\})\supseteq\conv Q'$, and by symmetry $-v\notin\conv Q'$; as $\conv Q'\cap\Lambda\subseteq Q$, it equals $Q'$. (ii) If every vertex of $\conv Q$ were in $S$, then $Q=\conv Q\cap\Lambda\subseteq\conv S\cap\Lambda=S$. A vertex $v\notin S$ has $-v\notin S$ by symmetry. \end{proof} So every saturated symmetric $S\subseteq Q$ is reached from $Q$ through saturated symmetric sets that contain $S$; in particular a search that stops at $31$ points misses no set with at least $31$ points. \begin{computation}\label{comp:cube} \emph{Case~I.} We list all symmetric sets $S\subseteq\{-1,0,1\}^4$ with $\conv S\cap\Z^4=S$ and $\#S\ge31$, up to the $384$ signed permutations of the coordinates. Each such $S$ spans $\R^4$ (a hyperplane through $0$ contains at most $27$ points of $\{-1,0,1\}^4$, since on it one coordinate is determined by the other three), so $\conv S\in\W^4$ by Lemma~\ref{lem:member}. There are $2{,}547{,}694$ classes; for all of them the linear coefficient $g_1$ of $\conv S$ is at most $8$, with equality only for $S=\{-1,0,1\}^4$. \end{computation} \begin{computation}\label{comp:index} \emph{Index bound.} We list every sublattice $\Gamma\subset\Z^4$ of index $d$ with $2\le d\le16$ once, as $\Gamma=B\Z^4$ with $B$ in Hermite normal form under column operations ($B\mapsto BU$); there are $40{,}406$ of them. For each $\Gamma$ whose points in $\{-1,0,1\}^4$ span $\R^4$ we count these points. The maximum is $45$ for $d=2$, and it is at most $27$ for every $d\ge3$. \end{computation} \begin{computation}\label{comp:index2} \emph{Case~II.} A permutation of coordinates carries $\Gamma_I$ onto $\Gamma_J$ when $|I|=|J|$, so it suffices to take $\Gamma_m=\Gamma_{\{1,\dots,m\}}=\{x:x_1+\dots+x_m\text{ even}\}$ for $m=2,3,4$. For each of these three lattices we list all $\Gamma_m$-saturated symmetric sets $S\subseteq\Gamma_m\cap\{-1,0,1\}^4$ that span $\R^4$ and have $\#S\ge31$, up to the signed permutations that preserve $\Gamma_m$; each such $\conv S$ lies in $\W^4_{\Gamma_m}$ (see the paragraph after Lemma~\ref{lem:member}). There are $531$, $58$ and $27$ classes for $m=2$, $3$, $4$, and for all of them the linear coefficient $g_1$ of $\conv S$ with respect to $\Gamma_m$ is at most $6$. \end{computation} \begin{computation}\label{comp:lattices} \emph{The frame lattices.} We list all lattices in $\Lf$. For $U\in\GL_4(\Z)$ the frame matrices $V$ and $UV$ give the same lattice $\Lambda=V^{-1}\Z^4$, so it suffices to take $V$ in Hermite normal form under row operations ($V\mapsto UV$), with determinant $D\le23$, and to keep the lattices $\Lambda$ that satisfy (iv). There are $1{,}040$ frame lattices, in $39$ orbits under signed permutations, with indices $D\in\{1,\dots,10,12,16,18\}$. \end{computation} \begin{computation}\label{comp:frame} \emph{Case~III.} For each of the $39$ representatives $\Lambda$ we list all $S\in\A(\Lambda)$ with $\#S\ge31$ and $r(S)\le3$, up to the signed permutations that preserve $\Lambda$. There are $588{,}592$ classes; for all of them the linear coefficient $g_1$ of $\conv S$ with respect to $\Lambda$ is at most $8$, with equality for exactly two of them, one with $G=49$ and one with $G=33$. The search follows \cite[Algorithm~1]{AS24}. It keeps a lower set $S^-\in\A(\Lambda)$ and an upper set $S^+\supseteq S^-$, and it starts from $S^-=\conv\{\pm e_1,\dots,\pm e_4\}\cap\Lambda$ and $S^+=\Lambda\cap C_4$. Every member of $\A(\Lambda)$ lies between these two sets: it lies in $S^+$ by (i), and it contains $S^-$ by (i) and (iii). The set $S^-$ is itself a member: its points $c$ have $|c_1|+\dots+|c_4|\le1$, so (ii) holds by Hadamard's inequality and (iii) by (iv). For a pair $\pm v$ in $S^+\setminus S^-$ the search branches: either $\pm v$ is excluded and $S^+$ is replaced by $S^+\setminus\{\pm v\}$, or $\pm v$ is included and $S^-$ is replaced by $\conv(S^-\cup\{\pm v\})\cap\Lambda$; every saturated set between $S^-$ and $S^+$ lies in exactly one branch. An include branch is abandoned when the new $S^-$ is not contained in $S^+$, violates (ii), or has a nonzero point of $\Lambda$ in the interior of its convex hull, since then no set in the branch satisfies (i)--(iii). A pair $\pm y$ in $S^+$ for which $S^-\cup\{\pm y\}$ violates (ii) is removed from $S^+$, since no member contains it together with $S^-$. A branch is abandoned when $\#S^+<31$ or $r(S^+)=4$; the second rule uses that $M$ reverses inclusion, that is, $M(Q')\subseteq M(Q)$ whenever $Q\subseteq Q'$, so that every subset of a set $Q'$ with $r(Q')=4$ also has $r=4$. The search stops at the pairs $(S^-,S^+)$ with $S^-=S^+$. Each include step keeps $S^-$ in $\A(\Lambda)$, so these sets are members, and they are the sets found. A pair $(S^-,S^+)$ that a signed permutation preserving $\Lambda$ carries onto a pair met earlier in the search is discarded, since the permutation carries the sets between the two pairs onto each other; the sets found are then reduced to one of each class by the canonical form described before Lemma~\ref{lem:walk}. \end{computation} \begin{proof}[Proof of Theorem~\ref{thm:four}] Let $P\in\W^4$. If $G(P)\le29$, Lemma~\ref{lem:few} gives $g_1(P)<8$. If $G(P)\ge31$, then by Section~\ref{sec:cases} the pair $(P,\Z^4)$ is isomorphic to $(\conv S,\Lambda_S)$ for a set $S$ listed in Computation~\ref{comp:cube} (Case~I), in Computation~\ref{comp:index2} (Case~II, using Computation~\ref{comp:index}), or in Computation~\ref{comp:frame} (Case~III, using Computation~\ref{comp:lattices}), where $\Lambda_S$ is the lattice of that computation ($\Z^4$, $\Gamma_m$ or a frame lattice); so $L_P=L_{\conv S,\Lambda_S}$. In each list $g_1\le8$. Hence $g_1(P)\le8$, and Proposition~\ref{prop:reduction} gives $g_2(P)\le24$. \end{proof} \section{The equality cases}\label{sec:equality} Let $T\subset\R^4$ be the convex hull of the $24$ points $\pm f_a\pm f_b$, $1\le a\rho$. For $r=3$ we use a different reduction: every $P\in\W^4$ with $r(P)=3$ is isomorphic to the convex hull of a set of lattice points of one of six explicit parallelepipeds, which we call boxes (Theorem~\ref{thm:boxes}). Each of the two proofs of Case~III is complete by itself. \subsection{Ranks zero, one and two} The rank is monotone: as noted in Section~\ref{sec:computations}, $M(Q')\subseteq M(Q)$ whenever $Q\subseteq Q'$; hence $r(Q')\le r(Q)$. So for every $\rho$, a set of $\A(\Lambda)$ that contains a set $S$ with $r(S)\le\rho$ also has $r\le\rho$. Computation~\ref{comp:frame} uses this property for $\rho=3$, and its two programs (Section~\ref{sec:programs}) remain exhaustive, by the same argument, when the bound $3$ is replaced by any $\rho$. \begin{computation}\label{comp:rank} \emph{Case~III with $r\le2$.} For each of the $39$ representatives $\Lambda$ of Computation~\ref{comp:lattices} and for $\rho=0,1,2$: all $S\in\A(\Lambda)$ with $\#S\ge31$ and $r(S)\le\rho$, up to the signed permutations that preserve $\Lambda$. The two programs are those of Computation~\ref{comp:frame}, described in Section~\ref{sec:programs}; the first abandons a branch as soon as its current set together with all its remaining candidates has $r>\rho$, and the second abandons a branch when its upper set $S^+$ has $r(S^+)>\rho$. Results: \begin{enumerate}[label=\textup{(\alph*)}] \item $\rho=0$: exactly three sets. Two of them are $T$ ($G=49$, in a frame lattice of index $D=8$) and $X$ ($G=33$, $D=18$). The third, which we call $R_{41}$, has $G=41$, lies in a lattice of index $16$, and has $L=1+\frac{20}3k+\frac{20}3k^2+\frac{40}3k^3+\frac{40}3k^4$. \item $\rho=1$: the three sets of~(a) and exactly two sets with $r=1$. One has $G=47$, lies in the lattice of~$T$, and has $L=1+7k+9k^2+16k^3+14k^4$; the other has $G=39$, lies in the lattice of $R_{41}$, and has $L=1+\frac{17}3k+7k^2+\frac{40}3k^3+12k^4$. The set $R_{41}$ and the set with $G=39$ are listed in~\cite{Repo}. \item $\rho=2$: the five sets of~(b) and $7{,}527$ sets with $r=2$. These $7{,}527$ sets satisfy $g_1\le\frac{19}3$ and $g_2\le14$. \end{enumerate} \end{computation} The set with $G=47$ in~(b) is obtained from $T$ by deleting one pair of opposite vertices. The $384$ signed permutations of the coordinates preserve the $24$-cell $\conv\{\pm e_a\pm e_b\}$ and the lattice $\Z^4\cup(\Z^4+\frac12(1,1,1,1))$, and they act transitively on the $24$ vertices; so it does not matter, up to isomorphism, which pair is deleted. We put \[ T'=\conv\bigl(\{\pm f_a\pm f_b:1\le a1$, the lattice point $(0,1)$ lies on the segment from the interior point $0$ of $Q$ to the point $(0,\tau(0))$ of $Q$, and is different from its endpoint $(0,\tau(0))$, so it is an interior point of $Q$; this is impossible. Hence $\tau(0)\le1$. The point $(0,\tau(0))$ lies on the boundary of $Q$, so $Q$ has a supporting hyperplane at it, $\{(z,t):\langle c,z\rangle+\kappa t=\kappa\tau(0)\}$ with $\langle c,z\rangle+\kappa t\le\kappa\tau(0)$ on $Q$. If $\kappa=0$, then $c\ne0$ and $\langle c,z\rangle\le0$ on $\pi(Q)$, which is impossible because $0$ is an interior point of $\pi(Q)$. If $\kappa<0$, the inequality fails at the origin, because $\tau(0)>0$. So $\kappa>0$, and with $a=-c/\kappa$ we get $t\le\tau(0)+\langle a,z\rangle\le\langle a,z\rangle+1$ on $Q$. Since $Q=-Q$, also $t\ge\langle a,z\rangle-1$ on $Q$. Therefore the set \[ \Omega=\bigl\{a\in\R^3:\ |t-\langle a,z\rangle|\le1\ \text{for all lattice points }(z,t)\text{ of }Q\bigr\} \] is not empty. It is a polyhedron, and it is bounded: if $a+se\in\Omega$ for all $s\ge0$, then $\langle e,z\rangle=0$ for the projections $z$ of all lattice points of $Q$, which span $\R^3$, so $e=0$. Hence $\Omega$ is a polytope. \emph{A vertex.} Let $a$ be a vertex of $\Omega$. Then $a$ satisfies three linearly independent equations $\langle a,z_j\rangle=t_j\mp1$, where $(z_j,t_j)$ are lattice points of $Q$; so $z_1,z_2,z_3$ are linearly independent points of $\Gamma\cap\{-1,0,1\}^3$ and $\langle a,z_j\rangle\in\Z$. In particular $a\in\Q^3$. Suppose that $a\in\Gamma^*$. Then the functional $h(z,t)=t-\langle a,z\rangle$ takes integer values on $\Gamma\times\Z$, so $h\circ\phi$ is an integer functional on $\Z^4$; it satisfies $|h\circ\phi|\le1$ on the lattice points of $P$, hence on $P$, so $h\circ\phi\in M(P)$. It does not lie in the span of $\lambda_1,\lambda_2,\lambda_3$, because it depends on $t$. Then $r(P)=4$, a contradiction. So $a\notin\Gamma^*$, which proves~(ii). Since $Q$ is the convex hull of its lattice points and $a\in\Omega$, we have $Q\subseteq\Pi(\Gamma,a)$. \emph{The interior.} If $(z,t)\in\Gamma\times\Z$ is an interior point of $\Pi(\Gamma,a)$, then $|z_i|<1$ and $z_i\in\Z$, so $z=0$; then $|t|<1$, so $t=0$. This is~(iii). Finally $\phi$ carries $P$ onto $Q=\conv S$ and $\Z^4$ onto $\Gamma\times\Z$, so the two pairs are isomorphic; $S$ is symmetric and saturated because $P\cap\Z^4$ is, and $r(S)=r(P)=3$ because $\phi$ carries integer functionals to integer functionals. \end{proof} Three kinds of maps carry boxes to boxes and are isomorphisms of the pairs $(\Pi(\Gamma,a),\Gamma\times\Z)$: a signed permutation of the coordinates $z$ (which replaces $\Gamma$ by its image and applies the same signed permutation to $a$), the map $(z,t)\mapsto(z,-t)$ (which replaces $a$ by $-a$), and, for $c\in\Gamma^*$, the map $(z,t)\mapsto(z,t+\langle c,z\rangle)$, which is unimodular on $\Gamma\times\Z$ and replaces $a$ by $a+c$. So only the class of $(\Gamma,a)$ under these maps matters. Conversely, by~(iii) and the argument of Lemma~\ref{lem:member}, every symmetric $(\Gamma\times\Z)$-saturated set of lattice points of a box that spans $\R^4$ gives a member of $\W^4_{\Gamma\times\Z}$; such a set may, however, have $r=4$. \begin{computation}\label{comp:boxes} \emph{The boxes.} All pairs $(\Gamma,a)$ with $\Gamma\subseteq\Z^3$ of index $d\le6$ that satisfy (i) and (ii) of Theorem~\ref{thm:boxes}, up to the three kinds of maps above. Each $\Gamma$ is listed once, by its Hermite normal form. For each triple of linearly independent points $z_1,z_2,z_3$ of $\Gamma\cap\{-1,0,1\}^3$, let $Z$ be the matrix with rows $z_j$; the vectors $a=Z^{-1}c$ with $c\in\{0,1,\dots,|\det Z|-1\}^3$ represent every class of $\{a:Za\in\Z^3\}$ modulo $\Gamma^*$, because $|\det Z|\,Z^{-1}\Z^3\subseteq\Z^3\subseteq\Gamma^*$; the class of $0$ is dropped. Result: exactly six classes, listed in Table~\ref{tab:boxes}. Four have $d=1$ and two have $d=2$; no class has $3\le d\le6$. \end{computation} The list was computed by two programs. The first identifies equivalent pairs by a canonical form, the least image under the signed permutations and $t\mapsto-t$ with $a$ reduced modulo $\Gamma^*$. The second uses no canonical form: it finds $33$ boxes before identification and identifies them by explicit unimodular maps between them, which gives the same six classes. The second program also reports the numbers of sublattices of $\Z^3$ of index $1,\dots,6$, namely $1$, $7$, $13$, $35$, $31$, $91$, of which $1$, $4$, $4$, $7$, $0$, $0$ have points in $\{-1,0,1\}^3$ that span $\R^3$. The six classes have different numbers of lattice points, so they are pairwise inequivalent. Both programs use exact rational arithmetic. \begin{table}[ht] \centering\footnotesize \caption{The six boxes $\Pi(\Gamma,a)$ of Computation~\ref{comp:boxes} and the sets of Computation~\ref{comp:boxwalk}. For $d=1$, $\Gamma=\Z^3$; for $d=2$, $\Gamma=\{z\in\Z^3:z_2\equiv z_3\pmod2\}$. $G$ is the number of lattice points of the box, and $(g_1,g_2,g_3,g_4)$ are the Ehrhart coefficients of their convex hull. The last three columns concern the sets $S$ with $\#S\ge31$ and $r(S)=3$: their number as point sets (in parentheses, up to the symmetries of the box), and the largest values of $g_1$ and $g_2$.}\label{tab:boxes} \begin{tabular}{ccccrcc} \toprule $d$ & $a$ & $G$ & $(g_1,g_2,g_3,g_4)$ & sets $S$ with $r(S)=3$ & largest $g_1$ & largest $g_2$\\ \midrule $1$ & $(0,\frac12,\frac12)$ & $69$ & $(6,18,28,16)$ & $2{,}495{,}033$ $(160{,}502)$ & $6$ & $18$\\ $1$ & $(\frac12,\frac12,\frac12)$ & $67$ & $(\frac{19}3,\frac{52}3,\frac{80}3,\frac{47}3)$ & $5{,}691{,}415$ $(122{,}559)$ & $\frac{19}3$ & $\frac{52}3$\\ $1$ & $(\frac13,\frac13,\frac13)$ & $63$ & $(6,\frac{31}2,25,\frac{31}2)$ & $389{,}891$ $(66{,}050)$ & $\frac{37}6$ & $\frac{31}2$\\ $1$ & $(\frac14,\frac14,\frac12)$ & $61$ & $(6,\frac{44}3,24,\frac{46}3)$ & $75{,}603$ $(19{,}804)$ & $6$ & $\frac{44}3$\\ $2$ & $(\frac12,0,\frac12)$ & $39$ & $(5,11,14,8)$ & $209$ & $5$ & $11$\\ $2$ & $(\frac12,\frac14,-\frac14)$ & $37$ & $(\frac{16}3,\frac{31}3,\frac{38}3,\frac{23}3)$ & $42$ & $\frac{16}3$ & $\frac{31}3$\\ \bottomrule \end{tabular} \end{table} \begin{computation}\label{comp:boxwalk} \emph{Rank three.} For each of the six boxes $\Pi=\Pi(\Gamma,a)$ of Table~\ref{tab:boxes}: all symmetric $(\Gamma\times\Z)$-saturated sets $S$ of lattice points of $\Pi$ that span $\R^4$ and have $\#S\ge31$ and $r(S)=3$. Results in Table~\ref{tab:boxes}. In particular $g_1\le\frac{19}3$ for all these sets, and the value $\frac{19}3$ is attained by the convex hull of all lattice points of the box with $G=67$. \end{computation} All six boxes have $r=3$. The searches walk downwards from the set of all lattice points of the box by Lemma~\ref{lem:walk}. A set with $r=4$ is recorded but not expanded: by the monotonicity of $r$, all its subsets have $r=4$, and they belong to Cases~I and~II, which are closed by Computations~\ref{comp:cube}--\ref{comp:index2}. No set with $r=3$ is lost by this cut, because by Lemma~\ref{lem:walk} it is reached through sets that contain it, and these have $r\le3$. For the four large boxes two programs walk downwards and reduce each level (the sets with a given $G$) by the symmetries of the box: the maps of the three kinds above that carry the box onto itself, which form groups of order $32$, $96$, $12$ and $8$ for $G=69$, $67$, $63$ and $61$. They compute $g_1$ by different formulas. By Lemma~\ref{lem:identities} and (E2), every $P\in\W^4_\Lambda$ satisfies \begin{equation}\label{eq:twoP} \#\bigl(\partial(2P)\cap\Lambda\bigr)=L_P(2)-L_P(-2)=8G-8-12g_1 ; \end{equation} the first program obtains $g_1$ from \eqref{eq:twoP} by counting the lattice points of $2\conv S$ and of its interior, while the second computes $g_3$ from a triangulation of the boundary of $\conv S$ and sets $g_1=\frac{G-1}2-g_3$. For every set, both programs verify exactly that the facet inequalities hold on $S$, that $S$ is saturated, that $0$ is the only interior lattice point, and that the computed vertex list is the exact one. Both report, for each level $G$, six numbers: the numbers of classes and of point sets with $r=3$, the largest $g_1$ and $g_2$ among them, and the numbers of classes and of point sets with $r=4$ that were cut. The two outputs agree in all six numbers on all $72$ levels of the four large boxes. For the two small boxes ($G=39$ and $G=37$), both programs and a third exact program that uses no symmetry find $209$ and $42$ sets with $r=3$; without the cut at $r=4$, both programs find all $483$ and $220$ symmetric saturated subsets with $\#S\ge31$. The largest values of $g_1$ on the small boxes are $5$ and $\frac{16}3$, attained by the boxes' own polytopes. \begin{corollary}\label{cor:caseIII} Let $P\in\W^4$ with $r(P)\le3$ and $G(P)\ge31$. Then $g_1(P)\le8$, with equality only if $P$ is unimodularly equivalent to $T$ or $X$. Moreover $g_1(P)\le7$ if $r(P)\in\{1,2\}$, and $g_1(P)\le\frac{19}3$ if $r(P)=3$. This gives a proof of Case~III that does not use Computation~\ref{comp:frame} itself (for $r\le2$ it reuses its programs with a different cut; see Remark~\ref{rem:tworoutes}). \end{corollary} \begin{proof} If $r(P)\le2$, Section~\ref{sec:cases} gives a pair $(\conv S,\Lambda)$ isomorphic to $(P,\Z^4)$ with $S\in\A(\Lambda)$, $\#S=G(P)$ and $r(S)=r(P)$, and $S$ is one of the sets of Computation~\ref{comp:rank}. If $r(P)=3$, Theorem~\ref{thm:boxes} gives a pair $(\conv S,\Gamma\times\Z)$ isomorphic to $(P,\Z^4)$ with $S$ one of the sets of Computation~\ref{comp:boxwalk}. The bounds are read off from the two computations. The only sets with $g_1=8$ are $T$ and $X$ in Computation~\ref{comp:rank}(a), and the argument at the end of the proof of Theorem~\ref{thm:equality} shows that $P$ is then unimodularly equivalent to $T$ or to $X$. \end{proof} \begin{remark}\label{rem:tworoutes} For $r\le2$ the two proofs of Case~III use the same frame reduction and the same two programs, with different cuts; for $r=3$ they use different reductions. Their counts are not comparable. Computation~\ref{comp:frame} counts $581{,}060$ sets with $r=3$, up to the symmetries of each frame lattice; Computation~\ref{comp:boxwalk} counts $8{,}651{,}942$ point sets in the four large boxes ($368{,}915$ up to the symmetries of the boxes) and $209+42$ in the two small ones. A polytope may have several frames and lie in several boxes. Both routes give the same largest value $g_1=\frac{19}3$ for $r=3$. \end{remark} \section{The extremal polytopes and the next values of \texorpdfstring{$g_1$}{g1} and \texorpdfstring{$g_2$}{g2}}\label{sec:extremal} This section studies the three polytopes with $g_1=8$ and the values of $g_1$ and $g_2$ just below $8$ and $24$. Each of $C_4$, $T$ and $X$ has all its facets at the same lattice distance from the origin (Proposition~\ref{prop:reflexive}), and no larger member of $\W^4$ contains $X$ (Proposition~\ref{prop:Xmax}). The cube with one pair of opposite vertices deleted has $g_1=\frac{15}2$ and $g_2=\frac{277}{12}$. The value $\frac{277}{12}$ is the second largest value of $g_2$ on $\W^4$, and $\frac{15}2$ is the second largest value of $g_1$ unless some $P\in\W^4$ with $29$ lattice points has $\frac{15}20$ with $bu\in\Z^4$ (then $bu$ is primitive: if $bu=cw$ with $w\in\Z^4$ and an integer $c\ge2$, then $(b/c)u=w\in\Z^4$, against the choice of $b$). The third coordinate has denominator $1$ or $3$. The number $\epsilon_4-\epsilon_1-3\epsilon_2-\epsilon_3$ is even, so the fourth coordinate is $m/3$ with $m=(\epsilon_4-\epsilon_1-3\epsilon_2-\epsilon_3)/2\in\Z$. Thus $b\in\{1,3\}$, and $b=1$ would need $3\mid\epsilon_3-\epsilon_1-\epsilon_2$, that is $\epsilon_1=\epsilon_2=-\epsilon_3=s$, and $3\mid m=(\epsilon_4-3s)/2$; but $(\epsilon_4-3s)/2\in\{\pm1,\pm2\}$. So $b=3$ for every facet. If $P\in\W^4$ is $\ell$-reflexive, (E3) gives $\vol(P)=\frac14\sum_Fb_F\operatorname{relvol}(F)=\frac{\ell}4\sigma(P)$, and by Proposition~\ref{prop:reduction} inequality~\eqref{eq:star} is $G\le\sigma+17$. Finally $(G,\vol)=(81,16)$, $(49,16)$ and $(33,12)$ for $C_4$, $T$ and $X$. \end{proof} Equivalently, $g_3(P)=2\vol(P)/\ell$ for $\ell$-reflexive $P\in\W^4$, which is the case $n=4$ of \cite[Proposition~3.6]{Hen14}. \subsection{The cross-polytope \texorpdfstring{$X$}{X} is maximal} \needspace{4\baselineskip} \begin{proposition}\label{prop:Xmax} If $P\in\W^4$ and $X\subseteq P$, then $P=X$. \end{proposition} The proof rests on the following computation. \begin{computation}\label{comp:Xmax} \emph{Enlargements of $X$.} Let $V$ be the matrix with columns $v_1,\dots,v_4$. The parallelepiped $V[-1,1]^4$ contains $417$ points of $\Z^4$: the $33$ lattice points of $X$ and $192$ further pairs $\pm x$. For each of these $192$ pairs, $\conv(X\cup\{\pm x\})$ has a lattice point other than $0$ in its interior. \end{computation} \begin{proof}[Proof of Proposition~\ref{prop:Xmax}] Since $X$ already realizes the largest possible frame index, every lattice point of $P$ lies in the frame parallelepiped of $X$, where Computation~\ref{comp:Xmax} applies. Since $X\subseteq P$, we have $M(P)\subseteq M(X)=\{0\}$, so $r(P)=0$, and $G(P)\ge G(X)=33$; so $P$ is in Case~III, and the frame construction of Section~\ref{sec:cases} applies to it. Let $D$ be the largest value of $|\det(x_1,\dots,x_4)|$ over lattice points $x_1,\dots,x_4$ of $P$. The vertices $v_1,\dots,v_4$ of $X$ give $D\ge|\det V|=18$. By Lemma~\ref{lem:frame} the frame lattice of a maximizing quadruple lies in $\Lf$ and has index $D$, and by Computation~\ref{comp:lattices} every lattice in $\Lf$ has index at most $18$. Hence $D=18$, and $v_1,\dots,v_4$ is a maximizing quadruple. By Lemma~\ref{lem:frame}(i) every lattice point of $P$ lies in $V[-1,1]^4$. If $P\ne X$, some vertex $x$ of $P$ is not in $X$; then $\conv(X\cup\{\pm x\})\subseteq P$, and by Computation~\ref{comp:Xmax} it has a nonzero interior lattice point, which is interior to $P$ as well. This contradicts $P\in\W^4$. \end{proof} \begin{figure}[ht] \centering \begin{tikzpicture}[x=0.13cm,y=0.85cm,font=\footnotesize] \fill[blue!12] (9,2.5) rectangle (29,8.7); \node at (19,8.45) {$G\le29$}; \draw[->] (9,2.5) -- (86,2.5) node[right] {$G$}; \draw[->] (9,2.5) -- (9,8.9) node[left] {$g_1$}; \foreach \g in {9,19,29,39,49,59,69,79} \draw (\g,2.5) -- (\g,2.4) node[below] {$\g$}; \foreach \v in {3,4,5,6,7,8} \draw (9,\v) -- (8.4,\v) node[left] {$\v$}; \draw[red!70!black,dashed] (9,8) -- (84,8); \draw[blue!70!black,thick] (9,{34/12}) -- (29,{94/12}); \node[blue!70!black] at (20,3.35) {$g_1\le\frac{3G+7}{12}$}; \foreach \g/\v in {81/8.0000, 79/7.5000, 77/7.0000, 75/7.3333, 73/6.8333, 71/7.0000, 69/6.8333, 67/6.6667, 65/6.6667, 63/7.0000, 61/6.5000, 59/6.5000, 57/6.6667, 55/6.3333, 53/6.0000, 51/6.3333, 49/6.0000, 47/5.8333, 45/6.0000, 43/6.0000, 41/5.8333, 39/5.6667, 37/6.0000, 35/5.8333, 33/5.6667, 31/5.6667} \fill (\g,{\v}) circle (1.3pt); \draw[thick] (49,8) circle (2.6pt) node[above=2pt] {$T$}; \draw[thick] (33,8) circle (2.6pt) node[above=2pt] {$X$}; \draw[thick] (47,7) circle (2.6pt) node[above=2pt] {$T'$}; \draw[thick] (41,{20/3}) circle (2.6pt) node[below=2pt] {$R_{41}$}; \node[above=2pt] at (81,8) {$C_4$}; \end{tikzpicture} \caption{Large values of $g_1$ on $\W^4$. The dots show, for each $G\ge31$, the largest $g_1$ in Case~I (Table~\ref{tab:caseI}); the dot at $G=81$ is the cube $C_4$. The circles show the polytopes of Case~III with $g_1>\frac{19}3$, namely $T$, $X$, $T'$ and $R_{41}$; all other polytopes of Cases~II and~III have $g_1\le6$ and $g_1\le\frac{19}3$, respectively (Proposition~\ref{prop:next}, Table~\ref{tab:rank}). The dashed line is $g_1=8$. In the shaded region $G\le29$, the solid line is the bound $g_1\le\frac{3G+7}{12}$ stated after Lemma~\ref{lem:few}, which comes from the triangulation in its proof. Every polytope with $G\ge31$ other than $C_4$, $T$ and $X$ has $g_1\le\frac{15}2$ (Proposition~\ref{prop:next}).}\label{fig:maxg1} \end{figure} \subsection{The next values} Table~\ref{tab:caseI} gives the result of Computation~\ref{comp:cube} level by level, and Figure~\ref{fig:maxg1} shows its largest values of $g_1$ together with the bound for $G\le29$. \begin{proposition}\label{prop:next} Let $P\in\W^4$ with $G(P)\ge31$. \begin{enumerate} \item In Case~I, if $P$ is not unimodularly equivalent to $C_4$, then $g_1(P)\le\frac{15}2$ and $g_2(P)\le\frac{277}{12}$. Both bounds are attained only by the cube with one pair of opposite vertices deleted, which has $G=79$ and $L=1+\frac{15}2k+\frac{277}{12}k^2+\frac{63}2k^3+\frac{191}{12}k^4$. \item In Case~II, $g_1(P)\le6$ and $g_2(P)\le14$. \item In Case~III, if $P$ is not unimodularly equivalent to $T$ or $X$, then $g_1(P)\le7$, with equality for $T'$; if moreover $P$ is not equivalent to $T'$, then $g_1(P)\le\frac{20}3$, with equality for $R_{41}$. In Case~III, $g_2(P)\le18$, with equality for the convex hull of all lattice points of the box with $G=69$ in Table~\ref{tab:boxes}, which has $L=1+6k+18k^2+28k^3+16k^4$. \end{enumerate} \end{proposition} \begin{proof} (i) is read off from Table~\ref{tab:caseI}: the level $G=79$ consists of one class, and every other level below $81$ has smaller maxima. All vertices of $C_4$ are equivalent under signed permutations, so this class is $C_4$ with a pair of opposite vertices deleted; its volume is $16-2\cdot\frac1{24}=\frac{191}{12}$, and with $g_1=\frac{15}2$ from Table~\ref{tab:caseI}, Lemma~\ref{lem:identities} gives $g_2=\frac{277}{12}$ and $g_3=\frac{63}2$. (ii) For $g_1$ this is Computation~\ref{comp:index2}; for $g_2$, the largest value over the classes of Computation~\ref{comp:index2} is $14$. (iii) follows from Table~\ref{tab:rank}, Computation~\ref{comp:rank}, Lemma~\ref{lem:Tprime} and Table~\ref{tab:boxes}: the values of $g_1$ above $\frac{19}3$ in Case~III are $8$ ($T$, $X$), $7$ ($T'$) and $\frac{20}3$ ($R_{41}$), and the largest $g_2$ is $18$. \end{proof} Since $g_2=\frac{G-1}2-g_4<14$ when $G\le29$, Proposition~\ref{prop:next} shows that $\frac{277}{12}$ is the second largest value of $g_2$ on $\W^4$, attained only by the cube with a pair of opposite vertices deleted. For $g_1$ we can say less: the bound $g_1\le\frac{3G+7}{12}$ stated after Lemma~\ref{lem:few} gives $g_1\le\frac{22}3<\frac{15}2$ for $G\le27$. So every $P\in\W^4$ other than $C_4$, $T$ and $X$ with $G(P)\ne29$ satisfies $g_1(P)\le\frac{15}2$. We do not know whether some $P\in\W^4$ with $G(P)=29$ has $\frac{15}20$, the part at heights $\pm1$ equals $2\mu$ times a point $\frac12(y-y')$ with $y,y'\in K_1$, and $\frac12(K_1-K_1)=\conv\frac12(S_1-S_1)$. Since also $\frac12(y,1)+\frac12(-y',-1)\in P$ for $y,y'\in S_1$, the slice at height $0$ is $\conv\bigl(S_0\cup\frac12(S_1-S_1)\bigr)=K_0$. This gives (i), and (iii) follows from $\frac12(S_1-S_1)\subseteq K_0$. Since $K_0\times\{0\}=P\cap\{x_4=0\}$, the lattice points of $K_0$ are the points of $S_0$. The slice $K_0$ is symmetric because $P$ is. Since $0$ is an interior point of $P$, the hyperplane $x_4=0$ meets the interior of $P$, so $K_0$ has dimension $3$ and its interior points are the interior points of $P$ at height $0$; the interior lattice points of $P$ lie at height $0$ because the heights $\pm1$ are faces. This gives (ii). \end{proof} \subsection{An exact identity and the shapes of \texorpdfstring{$K_1$}{K1}} \begin{proposition}\label{prop:split} In the situation of Lemma~\ref{lem:layers}, \[ 12\bigl(8-g_1(P)\bigr)=\delta_0+2\delta_1+2\delta_2, \] where \[ \delta_0=\#\partial(2K_0)-(4\#S_0-10),\qquad \delta_1=\#\partial(K_0+K_1)-(2\#S_0+3b-37),\qquad \delta_2=L_{K_1}(2)-(5b-10). \] \end{proposition} \begin{proof} The idea is to count the lattice points on the boundary of $2P$ by their heights. By~\eqref{eq:twoP}, $12(8-g_1)=\#(\partial(2P)\cap\Z^4)-8G+104$. By Lemma~\ref{lem:layers}(i), the slices of $2P$ at heights $0$, $\pm1$, $\pm2$ are $2K_0$, $\pm(K_0+K_1)$ and $\pm2K_1$. The slices at heights $\pm2$ are faces of $2P$, so all their $L_{K_1}(2)$ lattice points lie on $\partial(2P)$. The hyperplanes at heights $0$ and $\pm1$ meet the interior of $2P$, so a lattice point at these heights lies on $\partial(2P)$ if and only if it lies on the boundary of the slice. Hence $\#(\partial(2P)\cap\Z^4)=\#\partial(2K_0)+2\,\#\partial(K_0+K_1)+2L_{K_1}(2)$. Substituting this and $G=\#S_0+2b$ gives the identity. \end{proof} The number $12(8-g_1)$ does not depend on the direction $u$, so a proof may choose the direction. \begin{lemma}\label{lem:delta} \begin{enumerate} \item If $K_0$ is a lattice polytope, that is, $K_0=\conv S_0$, then $\delta_0=0$. \item If $K_1$ has dimension $3$, with $b_\partial$ lattice points on its boundary and $b_\circ\ge1$ in its interior, then $\operatorname{NVol}(K_1)=6\vol(K_1)\ge2b_\partial+3b_\circ-7$, and $\delta_2\ge0$. \end{enumerate} \end{lemma} \begin{proof} (i) For a lattice polytope $Q\subset\R^3$ of dimension $3$, reciprocity (E2) gives $\#\partial(kQ)=L_Q(k)+L_Q(-k)=2+2g_2(Q)k^2$. For $Q=K_0$, $\#\partial K_0=\#S_0-1$ by Lemma~\ref{lem:layers}(ii), so $2g_2(K_0)=\#S_0-3$ and $\#\partial(2K_0)=4\#S_0-10$. (ii) We count the tetrahedra of a triangulation of $K_1$ that uses all its lattice points. Triangulate $\partial K_1$ with vertex set the $b_\partial$ boundary lattice points, as in the proof of Lemma~\ref{lem:few}; by Euler's formula it has $2b_\partial-4$ triangles. Coning them from one interior lattice point triangulates $K_1$ into $2b_\partial-4$ tetrahedra. Insert the remaining $b_\circ-1$ interior lattice points one at a time, each time subdividing the cells of the current triangulation that contain the new point in their relative interior: a tetrahedron is replaced by $4$, two tetrahedra sharing a triangle by $6$, and the $k\ge3$ tetrahedra around an interior edge by $2k$. Each insertion adds at least $3$ tetrahedra, so the final triangulation has at least $2b_\partial+3b_\circ-7$ lattice tetrahedra, each of normalized volume at least $1$. For a lattice polytope of dimension $3$, $L_{K_1}(1)=b$ and $L_{K_1}(-1)=-b_\circ$ give $L_{K_1}(2)=6\vol(K_1)+3b_\partial+2b_\circ-3$, so $\delta_2=6\vol(K_1)-2b_\partial-3b_\circ+7\ge0$. \end{proof} \begin{lemma}\label{lem:hollow} If $K_1$ has dimension $3$ and no interior lattice point, then $6\vol(K_1)\ge b-3$, $L_{K_1}(2)=6\vol(K_1)+3b-3$, and $\delta_2\ge4-b$. \end{lemma} \begin{proof} Take a triangulation of $K_1$ whose vertex set is $S_1$; all its $f_0=b$ vertices lie on $\partial K_1$, and its boundary is a triangulated $2$-sphere with $3b-6$ edges and $2b-4$ triangles. Let $f_1,f_2,f_3$ be the numbers of edges, triangles and tetrahedra, and $e$, $t$ the numbers of interior edges and triangles. Euler's formula for the ball, $f_0-f_1+f_2-f_3=1$, gives $t=f_3+e-1$, and counting pairs (triangle, tetrahedron) gives $4f_3=2t+2b-4$. Hence $f_3=b-3+e\ge b-3$, and $6\vol(K_1)\ge f_3$. The formula for $L_{K_1}(2)$ is that of Lemma~\ref{lem:delta}(ii) with $b_\circ=0$, and $\delta_2=6\vol(K_1)-2b+7\ge4-b$. \end{proof} \begin{proposition}\label{prop:shapes} In the situation of Lemma~\ref{lem:layers}, $g_1(P)\le8$ holds whenever $\delta_0+2\delta_1$ is at least the number in the last column of Table~\ref{tab:shapes}, for the row that describes $K_1$. \end{proposition} \begin{table}[ht] \centering\small \caption{The bound for $\delta_2$ by the shape of $K_1=\conv S_1$, $b=\#S_1$, and the resulting sufficient condition for $g_1\le8$. For a polygon, $\beta$ is the number of lattice points on its boundary.}\label{tab:shapes} \begin{tabular}{lccc} \toprule $K_1$ & $L_{K_1}(2)$ & $\delta_2$ & sufficient: $\delta_0+2\delta_1\ge$\\ \midrule dimension $3$, $b_\circ\ge1$ & $6\vol(K_1)+3b_\partial+2b_\circ-3$ & $\ge0$ & $0$\\ dimension $3$, $b_\circ=0$ & $6\vol(K_1)+3b-3$ & $\ge4-b$ & $2b-8$\\ polygon & $4b-\beta-3$ & $=7-b-\beta$ & $2(b+\beta-7)$\\ segment & $2b-1$ & $=9-3b$ & $6b-18$\\ point ($b=1$) & $1$ & $=6$ & $-12$\\ \bottomrule \end{tabular} \end{table} \begin{proof} By Proposition~\ref{prop:split}, $g_1\le8$ if and only if $\delta_0+2\delta_1\ge-2\delta_2$, so it suffices that $\delta_0+2\delta_1$ is at least $-2$ times the lower bound for $\delta_2$. The first two rows are Lemmas~\ref{lem:delta}(ii) and~\ref{lem:hollow}. For a lattice polygon with $b$ lattice points, $\beta$ of them on the boundary, Pick's theorem \cite{BR15} gives $L(k)=(b-\frac\beta2-1)k^2+\frac\beta2k+1$, hence $L(2)=4b-\beta-3$. A lattice segment with $b$ lattice points has $L(2)=2b-1$, and a point has $L(2)=1$. \end{proof} So inequality~\eqref{eq:star} for all $P$ with $r(P)\ge1$ would follow from the inequality \begin{equation}\label{eq:J} \#\partial(2K_0)+2\,\#\partial(K_0+K_1)\ \ge\ 8\#S_0+6b-84\qquad(\iff\delta_0+2\delta_1\ge0) \end{equation} when $K_1$ has interior lattice points, from \begin{equation}\label{eq:J0} \#\partial(2K_0)+2\,\#\partial(K_0+K_1)\ \ge\ 8\#S_0+8b-92\qquad(\iff\delta_0+2\delta_1\ge2b-8) \end{equation} when $K_1$ is $3$-dimensional without interior lattice points, and from the inequalities of the last three rows of Table~\ref{tab:shapes} when $K_1$ is flat (of dimension at most $2$). Inequality~\eqref{eq:J} is sharp: if $K_0=\conv S_0$ and $K_1=K_0+v$ for some $v\in\Z^3$, then $\delta_0=0$ and $\delta_1=27-\#S_0$ by Lemma~\ref{lem:delta}(i), so~\eqref{eq:J} reads $\#S_0\le27$. This holds, because the lattice points of $K_0$ are pairwise incongruent modulo $3$: if $x\ne y$ were congruent, then $\frac13(x-y)=\frac23\cdot\frac12(x-y)+\frac13\cdot0$ would be a nonzero lattice point in the interior of $K_0$. Equality holds for $K_0=K_1-v=[-1,1]^3$, that is, for $P=C_4$. The next lemma shows which shapes of $K_1$ can occur. \begin{lemma}\label{lem:B1} In the situation of Lemma~\ref{lem:layers}, let $M_0=\{w\in\Z^3:|\langle w,y\rangle|\le1\text{ for all }y\in K_0\}$ and $r_0=\dim\operatorname{span}M_0$. \begin{enumerate} \item $r(P)=1+r_0$. \item If $r(P)=4$, then $K_1$ has at most one interior lattice point. \item The interior lattice points of $K_1$ are pairwise incongruent modulo $2$; hence there are at most $8$ of them. \end{enumerate} \end{lemma} \begin{proof} Write integer functionals on $\Z^4$ as $(w,s)\in\Z^3\times\Z$, acting by $(y,t)\mapsto\langle w,y\rangle+st$. Then $(w,s)\in M(P)$ if and only if $|\langle w,y\rangle|\le1$ for $y\in S_0$ and $|\langle w,y\rangle+s|\le1$ for $y\in S_1$. (i) If $(w,s)\in M(P)$, then $w\in M_0$: the bound holds on $S_0$, and for $y,y'\in S_1$, $\langle w,\frac12(y-y')\rangle=\frac12\bigl((\langle w,y\rangle+s)-(\langle w,y'\rangle+s)\bigr)\in[-1,1]$. Conversely, let $w\in M_0$. By Lemma~\ref{lem:layers}(iii), $\langle w,y-y'\rangle\le2$ for $y,y'\in S_1$, so the values $\langle w,y\rangle$, $y\in S_1$, lie in an interval $[m,m+2]$ with $m\in\Z$, and $(w,-m-1)\in M(P)$. Since also $(0,\pm1)\in M(P)$, the span of $M(P)$ is spanned by $(0,1)$ and the vectors $(w,-m_w-1)$, $w\in M_0$, and has dimension $1+r_0$. (ii) If $r(P)=4$, then $M_0$ spans $\R^3$ by~(i). Suppose that $p\ne q$ are interior lattice points of $K_1$, and choose $w\in M_0$ with $\langle w,p\rangle>\langle w,q\rangle$, so $\langle w,p\rangle\ge\langle w,q\rangle+1$. Since $p$ and $q$ are interior and the vertices of $K_1$ are lattice points, the maximum of $w$ on $K_1$ is an integer larger than $\langle w,p\rangle$, and the minimum is an integer smaller than $\langle w,q\rangle$; so the width of $K_1$ in the direction $w$ is at least $3$. But $K_1-K_1\subseteq2K_0$ and $|w|\le1$ on $K_0$ give width at most $2$. (iii) If $p\ne q$ are interior lattice points of $K_1$ with $p\equiv q\pmod2$, then $\frac12(p-q)$ is a nonzero lattice point in the interior of $\frac12(K_1-K_1)\subseteq K_0$, which contradicts Lemma~\ref{lem:layers}(ii). There are $8$ classes modulo $2$. \end{proof} \subsection{What is known about the two inequalities} Inequality~\eqref{eq:J} is not proved. It is an equality for the cube. Inequality~\eqref{eq:J0} is false. For $T'$ the direction $u=(1,1,0,1)$ of Lemma~\ref{lem:Tprime} gives $\#S_0=19$ and $b=14$; $K_1$ has dimension $3$ and no interior lattice point, and $K_0$ is not a lattice polytope. Here $\#\partial(2K_0)=50$, $\#\partial(K_0+K_1)=54$, $L_{K_1}(2)=63$ and $6\vol(K_1)=24$, so \[ \delta_0=-16,\qquad\delta_1=11,\qquad\delta_2=3,\qquad\delta_0+2\delta_1+2\delta_2=12=12\bigl(8-g_1(T')\bigr). \] Thus $\delta_0+2\delta_1=6<20=2b-8$. The loss comes from Lemma~\ref{lem:hollow}: the bound $6\vol(K_1)\ge b-3=11$ is far from the true value $24$. The bound $6\vol(K_1)\ge b-3$ of Lemma~\ref{lem:hollow} is attained for every $b\ge4$, by $\conv\{0,(b-3)e_1,e_2,e_3\}$, so a proof along these lines needs a lower bound for $\vol(K_1)$ that uses more than the absence of interior lattice points, for example the inclusion $K_1-K_1\subseteq2K_0$ of Lemma~\ref{lem:layers}(iii). By Lemma~\ref{lem:B1}(ii), in Cases~I and~II the first row of Table~\ref{tab:shapes} occurs only with $b_\circ=1$. In summary: the identity of Proposition~\ref{prop:split}, the bounds of Table~\ref{tab:shapes} and Lemma~\ref{lem:B1} are proved; \eqref{eq:J} is open; \eqref{eq:J0} is false, so the row of Table~\ref{tab:shapes} for $3$-dimensional $K_1$ without interior lattice points needs a different argument. Computations~\ref{comp:rank} and~\ref{comp:boxwalk} close the cases $r\le3$ anyway. A proof of~\eqref{eq:J}, of a correct replacement for~\eqref{eq:J0} and of the flat rows would prove~\eqref{eq:star} for every $P$ with $r(P)\ge1$ without enumeration; the remaining case $r(P)=0$ is Computation~\ref{comp:rank}(a). \section{The counterexample in dimension five and how it was found}\label{sec:search} The polytope $W$ of Theorem~\ref{thm:five} was obtained in two steps. A randomized local search over a class of layered polytopes found a polytope $W_5$ with $58$ vertices and $g_1(W_5)=\frac{301}{30}>10$; a greedy reduction then deleted four pairs of opposite vertices. This section describes both steps and shows that no single further pair of vertices can be deleted from $W$ without losing the property $g_1>10$. \subsection{Layered polytopes} For $n\ge2$, a symmetric set $Y_0\subseteq\{-1,0,1\}^{n-1}$ and an arbitrary set $Y_1\subseteq\{-1,0,1\}^{n-1}$ let \[ P(Y_0,Y_1)=\conv\bigl(Y_0\times\{0\}\ \cup\ Y_1\times\{1\}\ \cup\ (-Y_1)\times\{-1\}\bigr)\subset\R^n . \] This is a lattice polytope in $C_n$, and $P(Y_0,Y_1)=-P(Y_0,Y_1)$ whether or not $Y_1$ is symmetric. If it has dimension $n$, it belongs to $\W^n$ by Lemma~\ref{lem:member}. Henze's polytopes are of this form: $P_n=P(\{-1,1\}^{n-1},\{\pm e_1,\dots,\pm e_{n-1}\})$ is the convex hull of an $(n-1)$-dimensional cube at height $0$ and two $(n-1)$-dimensional cross-polytopes at heights $1$ and $-1$ \cite[proof of Theorem~2.1]{Hen14}. Henze showed $g_1(P_7)=\frac{1534}{105}>14$; in dimension~$5$, a direct count gives $g_1(P_5)=\frac{22}3<10$. \subsection{A randomized local search} Fix $n$ and the index $i$ of a coefficient, and let $m=n-1$. The search keeps a pair $(Y_0,Y_1)$ as above and changes it by one point or one pair at a time. \begin{enumerate} \item \emph{Start.} $Y_0=\{-1,1\}^m$ and $Y_1=\{\pm e_1,\dots,\pm e_m\}$, so that the first polytope is $P_n$. \item \emph{Move.} With probability $\frac34$, choose a point $y\in\{-1,0,1\}^m$ uniformly at random and add it to $Y_1$ if it is not in $Y_1$, or remove it if it is. Otherwise choose a pair $\{y,-y\}$ with $y\in\{-1,0,1\}^m\setminus\{0\}$ uniformly at random and add it to $Y_0$ or remove it from $Y_0$ in the same way. \item \emph{Evaluation.} If the new polytope $P(Y_0,Y_1)$ does not span $\R^n$, the move is rejected. Otherwise its Ehrhart polynomial is computed exactly: the program finds the facets with integer normals, counts the lattice points of $kP$ and of $\inter(kP)$ for $1\le k\le\lceil n/2\rceil+1$, checks that $0$ is the only interior lattice point of $P$, and interpolates by (E2); the extra values serve as a check. \item \emph{Acceptance.} The move is accepted if the new value of $g_i$ is at least the current one, and otherwise with probability $0.03$. The best polytope seen is recorded. \end{enumerate} With $n=5$ and $i=1$ the search produced the polytope $W_5$ below; it is a heuristic and need not reach $g_1>10$. \subsection{The polytope \texorpdfstring{$W_5$}{W5}} Let \[ q_1=(0,1,-1,-1,0),\quad q_2=(0,1,-1,0,1),\quad q_3=(1,-1,0,-1,0),\quad q_4=(1,0,1,0,1), \] and let $W_5=\conv\bigl(W\cup\{\pm q_1,\pm q_2,\pm q_3,\pm q_4\}\bigr)$, the convex hull of the $50$ points of Table~\ref{tab:witness} and the $8$ points $\pm q_j$. In the notation of the search, $W_5=P(Y_0,Y_1)$ for a set $Y_0$ of $18$ pairs and a set $Y_1$ of $25$ points; of these $86$ points, $58$ are vertices, $22$ at height $0$ and $18$ at each of the heights $\pm1$. \begin{proposition}\label{prop:W5} The polytope $W_5$ belongs to $\W^5$. All $58$ points $\pm p$, $\pm q_j$ are vertices of $W_5$, and $W_5$ has $102$ facets, at lattice distances $1$, $3$, $4$, $5$, $6$, $7$, $8$, $9$, $10$, $12$ and $17$ from the origin. Its Ehrhart polynomial is \[ L_{W_5}(k)=1+\tfrac{301}{30}\,k+\tfrac{47}{4}\,k^2+\tfrac{53}{2}\,k^3+\tfrac{197}{4}\,k^4+\tfrac{397}{15}\,k^5 ; \] in particular $g_1(W_5)=\frac{301}{30}>10$. \end{proposition} \begin{proof} As for $W$: the $58$ points lie in $\{-1,0,1\}^5$, form a symmetric set and span $\R^5$, so $W_5\in\W^5$ by Lemma~\ref{lem:member}. Counting lattice points gives $G(W_5)=125$, $\#(2W_5\cap\Z^5)=1915$, $\#(\inter(2W_5)\cap\Z^5)=243$ and $\vol(W_5)=\frac{397}{15}$, so $A(W_5)=1079$ and, by Proposition~\ref{prop:fournumbers}, $g_1(W_5)=\bigl(4\cdot126+24\cdot\frac{397}{15}-1079\bigr)/6=\frac{301}{30}$. The values $L_{W_5}(k)$ for $k=0,\dots,6$ are $1$, $125$, $1915$, $11273$, $41635$, $117147$, $275841$; those for $k\le5$ determine the polynomial, and the value for $k=6$ agrees with it. The polynomial can be computed with Normaliz~\cite{Normaliz} or LattE~\cite{LattE}. \end{proof} \begin{remark}\label{rem:margin} By Proposition~\ref{prop:fournumbers}, $g_1$ grows with the volume when $G(P)$ and $A(P)$ are fixed. Given $G(W_5)=125$ and $A(W_5)=1079$, the condition $g_1(W_5)>10$ is equivalent to $\vol(W_5)>\frac{635}{24}$, and $\vol(W_5)=\frac{397}{15}$ exceeds $\frac{635}{24}$ by exactly $\frac1{120}=\frac1{5!}$. Since the volume of a lattice polytope in $\R^5$ is an integer multiple of $\frac1{5!}$, this is the smallest possible excess. For $W$ the excess is $\frac{3}{120}=\frac1{40}$, and $g_1(W)-10=\frac1{10}$, three times $g_1(W_5)-10=\frac1{30}$. \end{remark} \begin{remark}\label{rem:outside} Henze tested polytopes of two kinds in dimensions $4$, $5$ and $6$ \cite[Remark~2.2]{Hen14}: convex hulls of lower-dimensional cubes or cross-polytopes at heights $0$, $-1$ and $1$, and Hanner polytopes. In dimension $5$ the former have at most $3\cdot16=48$ vertices and the latter at most $2^5=32$. Since $W_5$ and $W$ have $58$ and $50$ vertices, neither belongs to these families. Among the $3840$ signed permutations of the coordinates, only the identity and $x\mapsto-x$ map $W_5$ to itself, and the same holds for $W$. \end{remark} \subsection{From \texorpdfstring{$W_5$}{W5} to \texorpdfstring{$W$}{W}} Starting from $W_5$, a greedy reduction deletes pairs of opposite vertices one at a time. At each step it considers all pairs $\pm v$ of vertices whose deletion leaves a polytope with $g_1>10$, and deletes one for which this value of $g_1$ is largest; it stops when no deletion keeps $g_1>10$. It uses the exact Ehrhart computation of the search. It deleted the four pairs $\pm q_1,\dots,\pm q_4$ and stopped at $W$. Two of the deleted pairs lie at height $0$ and two at heights $\pm1$ (the height is the last coordinate); $W$ has $9$ vertex pairs at height $0$ and $16$ at heights $\pm1$. \begin{table}[ht] \centering\small \caption{The $25$ polytopes $W^{(p)}$ obtained from $W$ by deleting one pair $\pm p$ of opposite vertices ($\bar1$ stands for $-1$), and their linear Ehrhart coefficients. Each $W^{(p)}$ belongs to $\W^5$ and has $48$ vertices.}\label{tab:deletions} \begin{tabular}{crccr} \toprule $p$ & $g_1(W^{(p)})$ & \qquad & $p$ & $g_1(W^{(p)})$\\ \cmidrule{1-2}\cmidrule{4-5} $(0,0,1,\bar1,\bar1)$ & $99/10$ && $(1,0,\bar1,1,1)$ & $134/15$\\ $(0,0,1,1,\bar1)$ & $19/2$ && $(1,0,0,\bar1,1)$ & $149/15$\\ $(0,1,\bar1,0,0)$ & $49/5$ && $(1,0,0,1,1)$ & $10$\\ $(0,1,0,1,\bar1)$ & $271/30$ && $(1,0,1,\bar1,0)$ & $299/30$\\ $(0,1,0,1,1)$ & $281/30$ && $(1,0,1,\bar1,1)$ & $47/5$\\ $(0,1,1,\bar1,\bar1)$ & $131/15$ && $(1,0,1,0,\bar1)$ & $97/10$\\ $(0,1,1,\bar1,0)$ & $49/5$ && $(1,1,\bar1,\bar1,\bar1)$ & $43/6$\\ $(1,\bar1,\bar1,\bar1,0)$ & $127/15$ && $(1,1,\bar1,\bar1,0)$ & $146/15$\\ $(1,\bar1,0,1,0)$ & $49/5$ && $(1,1,\bar1,1,0)$ & $48/5$\\ $(1,\bar1,0,1,1)$ & $83/10$ && $(1,1,1,\bar1,0)$ & $143/15$\\ $(1,\bar1,1,0,\bar1)$ & $259/30$ && $(1,1,1,\bar1,1)$ & $271/30$\\ $(1,0,\bar1,0,\bar1)$ & $10$ && $(1,1,1,1,0)$ & $87/10$\\ $(1,0,\bar1,1,\bar1)$ & $93/10$ && & \\ \bottomrule \end{tabular} \end{table} \begin{proposition}\label{prop:deletions} For each of the $25$ points $p$ of Table~\ref{tab:witness}, the polytope $W^{(p)}$, the convex hull of the vertices of $W$ other than $\pm p$, belongs to $\W^5$, has $48$ vertices, and satisfies $g_1(W^{(p)})\le10$. Equality holds exactly for $p=(1,0,-1,0,-1)$ and $p=(1,0,0,1,1)$, and the smallest value, $\frac{43}{6}$, occurs for $p=(1,1,-1,-1,-1)$. Table~\ref{tab:deletions} lists all $25$ values. \end{proposition} \begin{proof} Each $W^{(p)}$ spans $\R^5$, so it belongs to $\W^5$ by Lemma~\ref{lem:member}. The values of $g_1$ were computed by an exact program that finds the facets by testing every set of five vertices and verifying the resulting inequalities in integers, counts the lattice points of $kW^{(p)}$ directly for $k=0,\dots,6$, interpolates from $k\le5$, checks the value at $k=6$, and checks (E2) for $k=1,2$. The inequality $g_1(W^{(p)})\le10$ for all $p$ is also the stopping condition of the greedy reduction, which used a different exact program. Normaliz gives the same $25$ values. \end{proof} The two polytopes with $g_1=10$ are members of $\W^5$, different from the cube, whose linear coefficient equals that of the cube. Proposition~\ref{prop:deletions} shows that $W$ is minimal for the deletion of one vertex pair. We do not know the least number of vertices of a polytope $P\in\W^5$ with $g_1(P)>10$. \section{The computations in detail}\label{sec:programs} Table~\ref{tab:programs} lists, for each computational claim that a proof in this paper uses, the two programs that establish it and how their outputs were compared. This section describes what the enumerations of Computations~\ref{comp:cube}--\ref{comp:rank} and~\ref{comp:boxwalk} have in common (exact arithmetic, walks, frame searches and lattice lists) and why each is exact. The programs for $W$, $W_5$ and the deletions from $W$ are described in Sections~\ref{sec:five} and~\ref{sec:search}, and those for the boxes in Section~\ref{sec:rank}. The programs, their outputs and one command per claim are in the repository~\cite{Repo}. \begin{table}[tp] \centering\footnotesize \caption{The computations, the two programs for each, and the comparison of their outputs.}\label{tab:programs} \begin{tabular}{>{\raggedright\arraybackslash}p{0.17\linewidth}>{\raggedright\arraybackslash}p{0.25\linewidth}>{\raggedright\arraybackslash}p{0.25\linewidth}>{\raggedright\arraybackslash}p{0.23\linewidth}} \toprule Claim & First program & Second program & Compared\\ \midrule $W\in\W^5$ and $L_W$ (Theorem~\ref{thm:five}) & facets from all sets of $5$ vertices, verified in integers; direct counts of $kW$, $k\le6$ & Normaliz~3.10.2; LattE integrale~1.7.6; the exact counter of the search & all six coefficients; the $92$ facets\\\addlinespace $W_5$ and $L_{W_5}$ (Proposition~\ref{prop:W5}) & the exact counter of the search & Normaliz; LattE; two further exact counters; the first program of the row above & all coefficients; $102$ facets\\\addlinespace deletions of $W$ (Proposition~\ref{prop:deletions}) & the first program of the first row & the greedy reduction (bound $g_1\le10$ only); Normaliz (all $25$ values) & the bound $g_1\le10$\\\addlinespace Case~I (Computation~\ref{comp:cube}) & top-down walk; canonical form under $384$ signed permutations; exact check of every class & top-down walk; canonical bit masks; Ehrhart polynomial by counting & per level $G=81,\dots,31$: classes, largest $g_1$, $g_2$ ($26$ levels)\\\addlinespace index bound (Computation~\ref{comp:index}) & all sublattices of index $2$ to $16$ by Hermite normal forms & an independent enumeration of the same & $40{,}406$ sublattices; maximum per index\\\addlinespace Case~II (Computation~\ref{comp:index2}) & top-down walk in all $11$ lattices $\Gamma_I$ with $|I|\ge2$ & the walker of Case~I, for $\Gamma_2,\Gamma_3,\Gamma_4$ & classes $531$, $58$, $27$; largest $g_1$, $g_2$\\\addlinespace frame lattices (Computation~\ref{comp:lattices}) & all lattices of index $D\le24$ (one more than Lemma~\ref{lem:frame}(v) needs) by Hermite normal forms & the same, independently written & number of lattices per $D$; $39$ orbits\\\addlinespace Case~III (Computation~\ref{comp:frame}) & maximal sets by branching, then a top-down walk; cut $r=4$ & sandwich search with include/exclude branching; cut $r(S^+)=4$ & per lattice: counts, ranks, largest $g_1$, $g_2$, sets with $g_1\ge8$\\\addlinespace $r\le\rho$, $\rho=0,1,2$ (Computation~\ref{comp:rank}) & as above, cut $r>\rho$ & as above, cut $r(S^+)>\rho$ & the lists for $\rho=0,1$; per lattice, the counts for $\rho=2$\\\addlinespace boxes (Computation~\ref{comp:boxes}) & canonical form & explicit unimodular maps & the six classes and their $G$\\\addlinespace rank three (Computation~\ref{comp:boxwalk}) & top-down walk reduced by the symmetries of the box; $g_1$ by~\eqref{eq:twoP} & independent top-down walk with its own symmetry computation; $g_3$ by triangulation & large boxes: six numbers per level ($72$ levels); small boxes: $209$ and $42$ sets\\\addlinespace $T$, $X$; $X$ maximal (Section~\ref{sec:equality}, Proposition~\ref{prop:Xmax}) & two exact counters; growth of $X$ by all $192$ pairs & Normaliz; the walker of Case~I; an independent check of the $192$ pairs & polynomials; none of the $192$ pairs gives a member of $\W^4$\\ \bottomrule \end{tabular} \end{table} \subsection{Exact arithmetic} All programs compute facets with integer normal vectors and integer right-hand sides, so that membership, interiority and saturation are decided by comparisons of integers, and they compute volumes and Ehrhart coefficients as rational numbers. Each enumeration program stops with an error if a set fails a check: in particular, if $\#(\conv S\cap\Lambda)\ne\#S$ or if $\conv S$ has an interior lattice point other than~$0$. The Ehrhart coefficients are computed in two ways: as described at the beginning of Section~\ref{sec:computations}, and by counting the lattice points of $k\conv S$ and of its interior for small $k$ and interpolating by (E2). \subsection{Walks} Computations~\ref{comp:cube}, \ref{comp:index2} and~\ref{comp:boxwalk} walk downwards from a largest set. A level is the set of classes with a given $G$; the children of a set $Q$ are the sets $Q\setminus\{v,-v\}$ for the vertices $v$ of $\conv Q$, and each level is reduced by a canonical form under a group of symmetries before the next level is formed. By Lemma~\ref{lem:walk} this reaches every saturated symmetric subset, through sets that contain it. Reducing a level by the group loses nothing: a symmetry $\gamma$ carries the vertices of $\conv Q$ to those of $\conv \gamma(Q)$, so it carries the children of $Q$ onto the children of $\gamma(Q)$; hence every set reached from $Q$ is, up to symmetry, reached from the representative of the class of $Q$, and since $\#S$ is invariant, each class lives on one level. Stopping below $G=31$ loses no set with $G\ge31$. The vertices must be found exactly: a missed vertex would lose children, and a non-vertex would produce an unsaturated child. In the first program of Computation~\ref{comp:cube}, every class is checked exactly: the facet inequalities hold on $S$, the cube points satisfying them are exactly $S$, the origin is the only interior lattice point, and the exact vertex list equals the list found by the convex hull routine. The second program determines the facets by an exact double description, declares a point a vertex when the normals of the facets through it have rank $4$, and asserts $L(1)=\#S$ and a single interior lattice point. Lower-dimensional children are skipped, since their subsets are lower-dimensional too; in Computation~\ref{comp:cube} no set with $G\ge31$ is lower-dimensional. \subsection{Frame searches} The first program of Computations~\ref{comp:frame} and~\ref{comp:rank} finds the maximal members of $\A(\Lambda)$ by a branching search. A member $S$ is maximal if and only if no single pair $\pm x$ can be added: if $S\subsetneq S'\in\A(\Lambda)$ and $\pm x\in S'\setminus S$, then $\conv(S\cup\{\pm x\})\cap\Lambda\subseteq S'$ satisfies (i)--(iii) of Lemma~\ref{lem:frame}, because $S'$ does. The search adds one pair at a time, prunes by a determinant test (a point $x$ with $|\det(x,s_1,s_2,s_3)|>1$ for points $s_j$ of the current set cannot be added), and breaks symmetry as follows. As in the Bron--Kerbosch algorithm~\cite{BK73}, it keeps, besides the current set $S$, a set of candidates and a set of excluded points; at each node the candidates are taken in a fixed order, and a candidate that is the image, under the stabilizer of $S$, of an earlier candidate that was branched on is not branched on but goes to the excluded set. No maximal set is lost. Without symmetry breaking, the search finds a maximal set $R$ along one path: at each node it enters the branch of the first candidate in $R$. (A set $R$ is found below a node only if $R$ contains the current set of that node and none of its excluded points; a set that contains an excluded point is found in an earlier sibling branch, at a smaller position.) Compare two such paths by the positions of these candidates, lexicographically. Among the images of a maximal set under the symmetries of $\Lambda$, take the one, $Q^*$, whose path is least. If the branch of this path at some node were skipped because its point $x_i$ equals $\gamma(x_j)$ with $j\rho$, by the monotonicity of $r$. Equivalent nodes are skipped by a canonical form under the symmetries of $\Lambda$; this loses no set, because the sets found below a node depend only on the node. The order of branching changes only the running time. \subsection{Lattice lists} In Computations~\ref{comp:index} and~\ref{comp:lattices} every lattice is listed exactly once by its Hermite normal form. Computation~\ref{comp:index} considers $40{,}406$ sublattices of index $2$ to $16$; the largest numbers of points of $\{-1,0,1\}^4$ in a sublattice of index $d$ whose cube points span $\R^4$ are $45$, $27$, $27$, $17$, $15$, $11$, $17$ for $d=2,\dots,8$ and $9$ for $d=9,10,12,16$, and no sublattice of index $11$, $13$, $14$ or $15$ has cube points that span $\R^4$. Computation~\ref{comp:lattices} finds $1$, $11$, $24$, $73$, $56$, $160$, $96$, $229$, $72$, $96$, $128$, $86$, $8$ lattices of index $D=1,\dots,10,12,16,18$ (in total $1{,}040$), forming $1$, $3$, $2$, $7$, $2$, $4$, $1$, $9$, $2$, $1$, $2$, $4$, $1$ orbits under signed permutations, and none of any other index $D\le24$. Both programs give these numbers. \subsection{Comparison} For each claim the two outputs agree: level by level in the walks, lattice by lattice in the frame searches (the lattices being matched by their Hermite normal forms), and, for the boxes, level by level in six numbers, including the cut sets. In Computation~\ref{comp:cube} both programs flag the cube, and only the cube, as a set with $g_1=8$; in Computation~\ref{comp:frame} both flag exactly $T$ and $X$. \subsection{Counting conventions} The counts in this paper are numbers of classes under the symmetry groups named in each computation: the $384$ signed permutations in Computation~\ref{comp:cube}, the stabilizer of $\Gamma_m$ in Computation~\ref{comp:index2}, the stabilizer of each frame lattice, summed over the $39$ lattices, in Computations~\ref{comp:frame} and~\ref{comp:rank}, and point sets or classes under the symmetries of each box in Computation~\ref{comp:boxwalk}. The same polytope may be counted in more than one computation, or under more than one frame or box. None of these totals is a number of unimodular equivalence classes of polytopes. \section*{Acknowledgments} This work was done overwhelmingly by AI models, steered by the author in a continuous critical dialogue, alternating between Anthropic's Claude and OpenAI's ChatGPT. The author takes full responsibility for the paper. \begin{thebibliography}{99} \bibitem{ABS21} G.~Averkov, C.~Borger and I.~Soprunov, \emph{Classification of triples of lattice polytopes with a given mixed volume}, Discrete Comput. Geom. \textbf{66} (2021), no.~1, 165--202. \bibitem{AS24} G.~Averkov and M.~Schymura, \emph{On the maximal number of columns of a $\Delta$-modular integer matrix: bounds and computations}, Math. Program. \textbf{206} (2024), 61--89. \bibitem{BR15} M.~Beck and S.~Robins, \emph{Computing the Continuous Discretely}, 2nd ed., Springer, New York, 2015. \bibitem{BHW93} U.~Betke, M.~Henk and J.~M.~Wills, \emph{Successive-minima-type inequalities}, Discrete Comput. Geom. \textbf{9} (1993), 165--175. \bibitem{BHW07} C.~Bey, M.~Henk and J.~M.~Wills, \emph{Notes on the roots of Ehrhart polynomials}, Discrete Comput. Geom. \textbf{38} (2007), 81--98. \bibitem{BK73} C.~Bron and J.~Kerbosch, \emph{Algorithm 457: finding all cliques of an undirected graph}, Commun. ACM \textbf{16} (1973), no.~9, 575--577. \bibitem{Ehr62} E.~Ehrhart, \emph{Sur les poly\`edres rationnels homoth\'etiques \`a $n$ dimensions}, C. R. Acad. Sci. Paris \textbf{254} (1962), 616--618. \bibitem{GoelShort} D.~Goel, \emph{Wills' (1981) conjecture holds if and only if the dimension is at most 4}, Zenodo, 2026,\newline\url{https://doi.org/10.5281/zenodo.23956180} \bibitem{GW93} P.~Gritzmann and J.~M.~Wills, \emph{Lattice points}, in: Handbook of Convex Geometry, Vol.~B, North-Holland, Amsterdam, 1993, 765--797. \bibitem{HPPS21} C.~Haase, A.~Paffenholz, L.~C.~Piechnik and F.~Santos, \emph{Existence of unimodular triangulations---positive results}, Mem. Amer. Math. Soc. \textbf{270} (2021), no.~1321; arXiv:1405.1687. \bibitem{HSW05} M.~Henk, A.~Sch\"urmann and J.~M.~Wills, \emph{Ehrhart polynomials and successive minima}, Mathematika \textbf{52} (2005), 1--16. \bibitem{Hen14} M.~Henze, \emph{On counterexamples to a conjecture of Wills and Ehrhart polynomials whose roots have equal real parts}, Electron. J. Combin. \textbf{21}(1) (2014), \#P1.28. \bibitem{KN12} A.~M.~Kasprzyk and B.~Nill, \emph{Reflexive polytopes of higher index and the number $12$}, Electron. J. Combin. \textbf{19} (2012), no.~3, Paper~P9. \bibitem{LattE} J.~A.~De~Loera, R.~Hemmecke, J.~Tauzer and R.~Yoshida, \emph{Effective lattice point counting in rational convex polytopes}, J. Symbolic Comput. \textbf{38} (2004), 1273--1302. \bibitem{McM80} P.~McMullen, \emph{Convex bodies which tile space by translation}, Mathematika \textbf{27} (1980), 113--121. \bibitem{Min96} H.~Minkowski, \emph{Geometrie der Zahlen}, Teubner, Leipzig, 1896. \bibitem{Min97} H.~Minkowski, \emph{Allgemeine Lehrs\"atze \"uber die convexen Polyeder}, Nachr. Ges. Wiss. G\"ottingen (1897), 198--219. \bibitem{Normaliz} W.~Bruns, B.~Ichim, C.~S\"oger and U.~von~der~Ohe, \emph{Normaliz. Algorithms for rational cones and affine monoids}, available at \url{https://www.normaliz.uni-osnabrueck.de} \bibitem{Repo} D.~Goel, \emph{Supporting materials for ``Wills' (1981) conjecture holds if and only if the dimension is at most 4''}, GitHub repository,\newline\url{https://github.com/david44hello/paper_wills-1981-conjecture} \bibitem{Ven54} B.~A.~Venkov, \emph{On a class of Euclidean polyhedra}, Vestnik Leningrad. Univ. Ser. Mat. Fiz. Him. \textbf{9} (1954), 11--31. \bibitem{Wal70} D.~W.~Walkup, \emph{The lower bound conjecture for $3$- and $4$-manifolds}, Acta Math. \textbf{125} (1970), 75--107. \bibitem{Wil81} J.~M.~Wills, \emph{On an analog to Minkowski's lattice point theorem}, in: The Geometric Vein (C.~Davis, B.~Gr\"unbaum, F.~A.~Sherk, eds.), Springer, New York, 1981, 285--288. \end{thebibliography} \end{document}