% Reconstructed from the authors' supplied PDF, November 11, 2026. % Revised November 16, 2026 to link the complete computational proof archive. \documentclass[11pt,letterpaper,reqno]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[margin=1in]{geometry} \usepackage{amsmath,amssymb,booktabs,array,tikz} \usepackage[hidelinks]{hyperref} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newcommand{\Z}{\mathbb Z} \newcommand{\R}{\mathbb R} \newcommand{\subsets}[2]{\binom{#1}{#2}} \newcommand{\binrow}[1]{\texttt{#1}} \newcommand{\smallmatrixblock}[9]{% \begin{minipage}[t]{.195\textwidth}\centering $#1$\par\vspace{2pt} $\left(\begin{array}{@{}c@{}} \binrow{#2}\\\binrow{#3}\\\binrow{#4}\\\binrow{#5}\\ \binrow{#6}\\\binrow{#7}\\\binrow{#8}\\\binrow{#9} \end{array}\right)$ \end{minipage}} \title[The Zarankiewicz number $z(13,15;3,4)$]{The Zarankiewicz number $z(13,15;3,4)$} \author[W. Howe]{Will Howe} \email{howew@purdue.edu} \author[G. Huffman]{Grant Huffman} \email{huffma97@purdue.edu} \author[J. Kohlmeyer]{Jensen Kohlmeyer} \email{jenwin@purdue.edu} \author[L. Kruer]{Liam Kruer} \email{kruerl@purdue.edu} \date{November 9, 2026} \subjclass[2020]{05C35, 05B20, 05C50} \keywords{Zarankiewicz number, forbidden submatrix, extremal bipartite graph, computer-assisted proof} \begin{document} \begin{abstract} We establish $z(13,15;3,4)=114$: a $13\times15$ binary matrix with $115$ ones necessarily contains an all-one submatrix on three rows and four columns. The lower bound is attained by a previously published construction of Rosero. For the upper bound, we exclude dense eight-row column neighborhoods by a finite classification and a complete extension check. The resulting local inequality reduces a hypothetical counterexample to $6{,}019$ pairs of degree sequences. Each pair is excluded by an exact integer certificate for a system of neighborhood-type equations. The proof is computer-assisted; its mathematical reductions, finite search domains, classification representatives, and certificate conditions are given explicitly. \end{abstract} \maketitle \section{Statement and notation} Place dots at intersections of thirteen horizontal rows and fifteen vertical columns. Require that every choice of three rows and four columns leaves at least one of their twelve intersections empty. How many dots can be placed? We prove that the answer is $114$. For positive integers $m,n,s,t$, let $z(m,n;s,t)$ denote the largest number of ones in an $m\times n$ binary matrix that contains no all-one submatrix on $s$ distinct rows and $t$ distinct columns. The selected rows and columns need not be consecutive. The order of the two parts matters: throughout this paper the forbidden submatrix has \emph{three rows and four columns}. \begin{theorem}\label{thm:main} The exact value of the oriented Zarankiewicz number is \[ z(13,15;3,4)=114. \] \end{theorem} The general matrix problem appears in Guy's treatment of the Zarankiewicz problem~\cite{guy}. In the notation of OEIS A006625~\cite{oeis}, Theorem~\ref{thm:main} gives $a(13)=115$. The construction establishing the lower bound is due to Rosero~\cite{rosero}; the upper bound is the result proved here. If $A$ is a binary matrix, write $|A|$ for its number of ones and $A_i\subseteq[n]$ for the support of its $i$th row, where $[n]=\{1,\ldots,n\}$. The avoidance condition is equivalent to \begin{equation} |A_i\cap A_j\cap A_k|\leq3\qquad(i,j,k\text{ distinct}).\label{eq:avoid} \end{equation} Two matrices are equivalent if one is obtained from the other by independent row and column permutations. This equivalence preserves all avoidance and extension questions considered below. A superscript in a degree sequence indicates multiplicity; for example, $(8^7,7)$ means seven entries equal to eight and one entry equal to seven. The proof has two finite components. First, a classification of $8\times14$ matrices with $63$ ones, together with extension checks, improves one local neighborhood bound from $64$ to $62$. Second, exact linear infeasibility certificates exclude every remaining pair of global degree sequences. These computations enter as explicit finite lemmas; the surrounding implications are proved below. \clearpage \section{A construction with 114 ones} The following matrix is Rosero's $114$-one construction~\cite{rosero}: \begin{equation} A_0=\left(\begin{array}{c} \binrow{011010110001111}\\ \binrow{101101111001100}\\ \binrow{010100101111110}\\ \binrow{101110000111101}\\ \binrow{100010110010100}\\ \binrow{111100111010001}\\ \binrow{011101001101011}\\ \binrow{110001011011101}\\ \binrow{111011100111000}\\ \binrow{000111110011011}\\ \binrow{110111100100101}\\ \binrow{011111010010110}\\ \binrow{001001101110111} \end{array}\right).\label{eq:witness} \end{equation} Here each string is one row and its digits are the matrix entries in column order. Twelve rows have weight nine and one has weight six, so $|A_0|=12\cdot9+6=114$. Direct intersection of the $\binom{13}{3}=286$ triples of displayed rows gives \[ \#\{\{i,j,k\}:|A_{0,i}\cap A_{0,j}\cap A_{0,k}|=r\} =\begin{cases} 11,&r=0,\\ 0,&r=1,\\ 72,&r=2,\\ 203,&r=3. \end{cases} \] Consequently $A_0$ satisfies~\eqref{eq:avoid}, proving $z(13,15;3,4)\geq114$. \begin{figure}[h] \centering \begin{tikzpicture}[x=.55cm,y=-.55cm] \foreach \r in {0,...,12}{\foreach \c in {0,...,14}{\draw[gray!55,line width=.25pt] (\c,\r) circle[radius=.045cm];}} % The filled cells are generated directly from the displayed witness. \fill (1,0) circle[radius=.06cm]; \fill (2,0) circle[radius=.06cm]; \fill (4,0) circle[radius=.06cm]; \fill (6,0) circle[radius=.06cm]; \fill (7,0) circle[radius=.06cm]; \fill (11,0) circle[radius=.06cm]; \fill (12,0) circle[radius=.06cm]; \fill (13,0) circle[radius=.06cm]; \fill (14,0) circle[radius=.06cm]; \fill (0,1) circle[radius=.06cm]; \fill (2,1) circle[radius=.06cm]; \fill (3,1) circle[radius=.06cm]; \fill (5,1) circle[radius=.06cm]; \fill (6,1) circle[radius=.06cm]; \fill (7,1) circle[radius=.06cm]; \fill (8,1) circle[radius=.06cm]; \fill (11,1) circle[radius=.06cm]; \fill (12,1) circle[radius=.06cm]; \fill (1,2) circle[radius=.06cm]; \fill (3,2) circle[radius=.06cm]; \fill (6,2) circle[radius=.06cm]; \fill (8,2) circle[radius=.06cm]; \fill (9,2) circle[radius=.06cm]; \fill (10,2) circle[radius=.06cm]; \fill (11,2) circle[radius=.06cm]; \fill (12,2) circle[radius=.06cm]; \fill (13,2) circle[radius=.06cm]; \fill (0,3) circle[radius=.06cm]; \fill (2,3) circle[radius=.06cm]; \fill (3,3) circle[radius=.06cm]; \fill (4,3) circle[radius=.06cm]; \fill (9,3) circle[radius=.06cm]; \fill (10,3) circle[radius=.06cm]; \fill (11,3) circle[radius=.06cm]; \fill (12,3) circle[radius=.06cm]; \fill (14,3) circle[radius=.06cm]; \fill (0,4) circle[radius=.06cm]; \fill (4,4) circle[radius=.06cm]; \fill (6,4) circle[radius=.06cm]; \fill (7,4) circle[radius=.06cm]; \fill (10,4) circle[radius=.06cm]; \fill (12,4) circle[radius=.06cm]; \fill (0,5) circle[radius=.06cm]; \fill (1,5) circle[radius=.06cm]; \fill (2,5) circle[radius=.06cm]; \fill (3,5) circle[radius=.06cm]; \fill (6,5) circle[radius=.06cm]; \fill (7,5) circle[radius=.06cm]; \fill (8,5) circle[radius=.06cm]; \fill (10,5) circle[radius=.06cm]; \fill (14,5) circle[radius=.06cm]; \fill (1,6) circle[radius=.06cm]; \fill (2,6) circle[radius=.06cm]; \fill (3,6) circle[radius=.06cm]; \fill (5,6) circle[radius=.06cm]; \fill (8,6) circle[radius=.06cm]; \fill (9,6) circle[radius=.06cm]; \fill (11,6) circle[radius=.06cm]; \fill (13,6) circle[radius=.06cm]; \fill (14,6) circle[radius=.06cm]; \fill (0,7) circle[radius=.06cm]; \fill (1,7) circle[radius=.06cm]; \fill (5,7) circle[radius=.06cm]; \fill (7,7) circle[radius=.06cm]; \fill (8,7) circle[radius=.06cm]; \fill (10,7) circle[radius=.06cm]; \fill (11,7) circle[radius=.06cm]; \fill (12,7) circle[radius=.06cm]; \fill (14,7) circle[radius=.06cm]; \fill (0,8) circle[radius=.06cm]; \fill (1,8) circle[radius=.06cm]; \fill (2,8) circle[radius=.06cm]; \fill (4,8) circle[radius=.06cm]; \fill (5,8) circle[radius=.06cm]; \fill (6,8) circle[radius=.06cm]; \fill (9,8) circle[radius=.06cm]; \fill (10,8) circle[radius=.06cm]; \fill (11,8) circle[radius=.06cm]; \fill (3,9) circle[radius=.06cm]; \fill (4,9) circle[radius=.06cm]; \fill (5,9) circle[radius=.06cm]; \fill (6,9) circle[radius=.06cm]; \fill (7,9) circle[radius=.06cm]; \fill (10,9) circle[radius=.06cm]; \fill (11,9) circle[radius=.06cm]; \fill (13,9) circle[radius=.06cm]; \fill (14,9) circle[radius=.06cm]; \fill (0,10) circle[radius=.06cm]; \fill (1,10) circle[radius=.06cm]; \fill (3,10) circle[radius=.06cm]; \fill (4,10) circle[radius=.06cm]; \fill (5,10) circle[radius=.06cm]; \fill (6,10) circle[radius=.06cm]; \fill (9,10) circle[radius=.06cm]; \fill (12,10) circle[radius=.06cm]; \fill (14,10) circle[radius=.06cm]; \fill (1,11) circle[radius=.06cm]; \fill (2,11) circle[radius=.06cm]; \fill (3,11) circle[radius=.06cm]; \fill (4,11) circle[radius=.06cm]; \fill (5,11) circle[radius=.06cm]; \fill (7,11) circle[radius=.06cm]; \fill (10,11) circle[radius=.06cm]; \fill (12,11) circle[radius=.06cm]; \fill (13,11) circle[radius=.06cm]; \fill (2,12) circle[radius=.06cm]; \fill (5,12) circle[radius=.06cm]; \fill (6,12) circle[radius=.06cm]; \fill (8,12) circle[radius=.06cm]; \fill (9,12) circle[radius=.06cm]; \fill (10,12) circle[radius=.06cm]; \fill (12,12) circle[radius=.06cm]; \fill (13,12) circle[radius=.06cm]; \fill (14,12) circle[radius=.06cm]; \end{tikzpicture} \caption{The $114$ occupied intersections of~\eqref{eq:witness}. Filled circles denote ones; open circles denote zeros. Every three rows have at most three filled columns in common. The construction is due to Rosero~\cite{rosero}.} \end{figure} \clearpage \section{Smaller matrices and column neighborhoods} We use the following bounds, denoted by $Z(d)$: \begin{equation} \begin{array}{c|rrrrrrrrrrrrrr} d&0&1&2&3&4&5&6&7&8&9&10&11&12&13\\\hline Z(d)&0&14&28&30&36&44&50&56&64&70&77&84&91&98 \end{array}.\label{eq:z} \end{equation} Thus $z(d,14;3,3)\leq Z(d)$. For $6\leq d\leq13$ these are the bounds in Table~4 of Collins, Riasanovsky, Wallace, and Radziszowski~\cite{collins}. The same table states uniqueness, up to equivalence, of the $8\times14$ matrix with $64$ ones and no all-one $3\times3$ submatrix. For completeness, the smaller values in~\eqref{eq:z} follow from elementary counting. If $q_1,\ldots,q_{14}$ are column degrees of a $d\times14$ matrix avoiding an all-one $3\times3$ submatrix, then \begin{equation} \sum_{j=1}^{14}\binom{q_j}{3}\leq2\binom{d}{3}.\label{eq:count} \end{equation} For $d=0,1,2$ the area bound suffices. For $d=3$, at most two columns have degree three, giving at most $30$ ones. For $d=4$, the inequality $q\leq2+\binom q3$ for $0\leq q\leq4$ gives at most $28+8=36$ ones. For $d=5$, use $3q\leq8+\binom q3$ for $0\leq q\leq5$ to obtain $3\sum q_j\leq112+20=132$. We next describe the unique $64$-one matrix. Index its rows by $\Z/8\Z$. Its first eight columns have zero sets \begin{equation} \{i,i+1,i+3\},\qquad i\in\Z/8\Z.\label{eq:zeros} \end{equation} Its other six columns have one sets \begin{equation} \{a,a+4,b,b+4\},\qquad0\leq a\ell D$. Accept an empty list precisely when $s=0$. For columns begin with $(15,115,0,13)$, accumulating $\binom v3$ and $W(v)$; discard as soon as the former exceeds $858$, and require the latter to be at least $904$ at completion. For each accepted column list, generate rows from $(13,115,0,15)$, accumulating $v(v-1)$ and discarding whenever the sum exceeds the corresponding column bound. All accumulated moment costs are nonnegative, so these prunings are valid. Evaluation gives the stated counts. The initial degree domains include zero; the narrower ranges in the statement are outputs, not assumptions. \end{proof} \section{Neighborhood types and exact certificates} Fix a pair of sequences in Lemma~\ref{lem:pairs}. Let $N_a$ be the number of rows of degree $a$ and $M_d$ the number of columns of degree $d$. Write \[ R=\{a:N_a>0\},\qquad D=\{d:M_d>0\}. \] These are sets of positive degrees by the finite enumeration. \clearpage \begin{table}[h] \caption{Degree enumeration grouped by the value $C=\sum_j W(d_j)$.}\label{tab:degrees} \centering \begin{tabular}{rrrr}\toprule $C$&Column sequences&Row sequences per column&Pairs\\\midrule 904&48&1&48\\905&41&1&41\\906&70&3&210\\907&64&3&192\\ 908&93&5&465\\909&78&5&390\\910&98&8&784\\911&64&8&512\\ 912&85&14&1190\\913&37&14&518\\914&43&21&903\\915&13&21&273\\ 916&12&30&360\\917&3&30&90\\918&1&43&43\\\midrule Total&750&&6019\\\bottomrule \end{tabular} \end{table} \begin{definition}\label{def:types} For $a\in R$, a \emph{row type} is an integer vector $t=(t_d)_{d\in D}$ satisfying \begin{equation} 0\leq t_d\leq M_d,\qquad\sum_{d\in D}t_d=a,\qquad \sum_{d\in D}t_d\binom{d-1}{2}\leq198.\label{eq:rowtype} \end{equation} Let $\mathcal T_a$ denote its finite set of possibilities. For $d\in D$, a \emph{column type} is an integer vector $u= (u_a)_{a\in R}$ satisfying \begin{equation} 0\leq u_a\leq N_a,\qquad\sum_{a\in R}u_a=d,\label{eq:coltype} \end{equation} with the following additional requirement. Sort the multiset containing $u_a$ copies of $a-1$ as $w_1\geq\cdots\geq w_d$. Require \begin{equation} \sum_{i=1}^{k}w_i\leq W(k)\qquad(1\leq k\leq d).\label{eq:prefix} \end{equation} Let $\mathcal U_d$ denote the resulting finite set. \end{definition} \begin{lemma}\label{lem:types} The neighborhood of every degree-$a$ row has a type in $\mathcal T_a$, and the neighborhood of every degree-$d$ column has a type in $\mathcal U_d$. \end{lemma} \begin{proof} For a fixed row, each pair of the other twelve rows shares at most three of its incident columns. Thus \[ \sum_{j:A_{ij}=1}\binom{d_j-1}{2}\leq3\binom{12}{2}=198, \] which gives~\eqref{eq:rowtype}; the availability constraints and degree sum are immediate. For a fixed column, Corollary~\ref{cor:local} bounds the sum of $a_i-1$ over each $k$-element subset of its neighbors. The largest such sum is the prefix of length $k$ in nonincreasing order. This is exactly~\eqref{eq:prefix}. \end{proof} \clearpage Associate a nonnegative variable $Y_{a,t}$ with every $t\in\mathcal T_a$ and a nonnegative variable $Q_{d,u}$ with every $u\in\mathcal U_d$. Counting vertices of each degree and counting incidences in two ways gives the equations \begin{align} \sum_{t\in\mathcal T_a}Y_{a,t}&=N_a&&(a\in R),\label{eq:y}\\ \sum_{u\in\mathcal U_d}Q_{d,u}&=M_d&&(d\in D),\label{eq:q}\\ \sum_{t\in\mathcal T_a}t_dY_{a,t}-\sum_{u\in\mathcal U_d}u_aQ_{d,u}&=0&&(a\in R,\ d\in D).\label{eq:balance} \end{align} An actual matrix provides nonnegative integer values of these variables. It therefore provides a nonnegative real solution as well. \begin{lemma}[Certificate criterion]\label{lem:certificate} Suppose numbers $\alpha_a,\beta_d,\gamma_{ad}$ satisfy \begin{align} \alpha_a+\sum_{d\in D}\gamma_{ad}t_d&\geq0&&(a\in R,\ t\in\mathcal T_a),\label{eq:dualrow}\\ \beta_d-\sum_{a\in R}\gamma_{ad}u_a&\geq0&&(d\in D,\ u\in\mathcal U_d),\label{eq:dualcol}\\ \sum_{a\in R}\alpha_aN_a+\sum_{d\in D}\beta_dM_d&<0.\label{eq:dualnegative} \end{align} Then~\eqref{eq:y}--\eqref{eq:balance} have no nonnegative solution. \end{lemma} \begin{proof} Multiply~\eqref{eq:y}, \eqref{eq:q}, and~\eqref{eq:balance} by $\alpha_a$, $\beta_d$, and $\gamma_{ad}$, respectively, and add. The left side is nonnegative by~\eqref{eq:dualrow} and~\eqref{eq:dualcol}, while the right side is negative by~\eqref{eq:dualnegative}. \end{proof} \begin{lemma}[Finite degree obstruction]\label{lem:obstruction} For each of the $6{,}019$ degree pairs in Lemma~\ref{lem:pairs}, there exist integer multipliers satisfying~\eqref{eq:dualrow}--\eqref{eq:dualnegative}. \end{lemma} \begin{proof}[Exact finite verification] For each degree pair, the sets in Definition~\ref{def:types} are enumerated over their stated integer boxes, retaining precisely the vectors that satisfy the displayed sums and prefix bounds. Integer multipliers are then checked by evaluating the left sides of~\eqref{eq:dualrow} and~\eqref{eq:dualcol} for every retained vector and the integer sum in~\eqref{eq:dualnegative}. All $6{,}019$ cases pass these tests. No approximate inequality is used: the checked conditions are coefficientwise nonnegativity and a strictly negative integer. The largest absolute multiplier in these certificates is $3334$. Table~\ref{tab:support} records their support sizes. \begin{table}[h] \caption{Exact dual certificates, grouped by number of nonzero multipliers.}\label{tab:support} \centering \begin{tabular}{lrrrrrrr}\toprule Support size&1&2&3--7&8--12&13--19&20--29&30--38\\\midrule Cases&3758&535&311&448&630&285&52\\\bottomrule \end{tabular} \end{table} \newpage Here is a nontrivial example with all its multipliers displayed. Take \[ (a_i)=(8^5,9^5,10^3),\qquad(d_j)=(6^5,7^2,8,9^7). \] In increasing degree order, let \begin{equation}\label{eq:example} \begin{aligned} (\alpha_8,\alpha_9,\alpha_{10})&=(6,0,-3),\\ (\beta_6,\beta_7,\beta_8,\beta_9)&=(0,0,-2,-3),\\ (\gamma_{ad})_{a\in R,d\in D}&=\begin{pmatrix}0&0&-1&-1\\0&0&0&0\\0&0&1&1\end{pmatrix}. \end{aligned} \end{equation} The nonnegative coefficients can also be seen directly. For a degree-eight row, seven neighbors of degrees eight or nine would contribute at least $21+6\cdot28+10=199$ to~\eqref{eq:rowtype}, since only one degree-eight column is available. Thus $t_8+t_9\leq6$, as required by its coefficient $6-t_8-t_9$. A degree-ten row has at most seven neighbors of degrees six or seven, so $t_8+t_9\geq3$, as required by its coefficient $-3+t_8+t_9$. Degree-nine rows have coefficient zero. For a degree-eight column, the full prefix bound gives $64-u_8+u_{10}\leq62$, so its coefficient $-2+u_8-u_{10}$ is nonnegative. For a degree-nine column, $u_8=0$ would make its largest eight weights at least $64$, which is impossible. The smallest weight is therefore seven, and the eight-element prefix bound gives $65-u_8+u_{10}\leq62$. Hence its coefficient $-3+u_8-u_{10}$ is nonnegative. Degree-six and degree-seven columns have coefficient zero. Finally, the right side is \[ 6\cdot5-3\cdot3-2\cdot1-3\cdot7=-2. \] Thus even this degree pair cannot occur. The same finite coefficient test, with the corresponding integer multipliers, excludes each of the enumerated pairs. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] Section~2 gives the lower bound $114$. Suppose an avoiding $13\times15$ matrix had at least $115$ ones. Delete ones, if needed, to obtain exactly $115$; deleting ones cannot create a forbidden submatrix. Corollary~\ref{cor:local} then applies, so its degree sequences belong to the list in Lemma~\ref{lem:pairs}. By Lemma~\ref{lem:types}, the matrix gives a nonnegative solution of~\eqref{eq:y}--\eqref{eq:balance}. Lemma~\ref{lem:obstruction} and the certificate criterion rule out such a solution. This contradiction proves the upper bound $114$. \end{proof} \clearpage \section*{Computational proof files}\label{sec:archive} The complete computational material for the upper bound is publicly available in \href{https://github.com/JENW1N/zarankiewicz-13x15-certificates/releases/tag/v1}{computational proof archive, release v1}. The corresponding \href{https://github.com/JENW1N/zarankiewicz-13x15-certificates/tree/10385847932cd2b61e25cb4e7d039633d108455b/computational-proof-v1}{version-specific source tree} is fixed at commit \texttt{10385847932cd2b61e25cb4e7d039633d108455b}. The archive includes all fifteen residual SAT inputs and their complete DRAT refutations, the five-row extension inputs and outputs for all $82$ bases in Table~\ref{tab:extensions}, and all $6{,}019$ integer dual certificates together with their degree sequences. It also supplies the input generators, exact verification scripts, dependency versions, and checking instructions. These files were regenerated on November 16, 2026 from the constraints and representatives displayed in this paper and the integer certificate data in the supplied companion Lean file. The regeneration reproduces the $1{,}585$ ordered forms in Table~\ref{tab:classes}, all fifteen unsatisfiable residual systems, all $82$ rejected extension tests, and the degree and certificate counts in Tables~\ref{tab:degrees} and~\ref{tab:support}. Each DRAT refutation was independently checked with \texttt{drat-trim}. The dual certificates were checked with exact Python integer arithmetic over every admissible type. The archive's \texttt{README.md} gives a single verification command that checks file hashes, regenerates the SAT inputs, checks the refutations, reruns the extension searches, and verifies every dual certificate. The included companion \texttt{Zarankiewicz1315.lean} retains its supplied toolchain and dependency information. The recorded archive checks concern the computational components listed above. The release ZIP has SHA-256 \begin{center}\small\path{07f3d6338b4597eb65bba0bd688fc29bf2919a91647d594ff2f8eff1b39979ff}.\end{center} \section*{Acknowledgment of AI assistance} OpenAI Codex was used for mathematical exploration, finite-case verification, and preparation of this manuscript. \begin{thebibliography}{4} \bibitem{collins} A. F. Collins, A. W. N. Riasanovsky, J. C. Wallace, and S. P. Radziszowski, \emph{Zarankiewicz numbers and bipartite Ramsey numbers}, arXiv:1604.01257 [math.CO], 2016. \href{https://doi.org/10.48550/arXiv.1604.01257}{doi:10.48550/arXiv.1604.01257}. \bibitem{guy} R. K. Guy, \emph{A many-facetted problem of Zarankiewicz}, in \emph{The Many Facets of Graph Theory} (G. Chartrand and S. F. Kapoor, eds.), Lecture Notes in Mathematics, vol.~110, Springer, Berlin, 1969, pp.~129--148. \href{https://doi.org/10.1007/BFb0060112}{doi:10.1007/BFb0060112}. \bibitem{oeis} OEIS Foundation Inc., \emph{The On-Line Encyclopedia of Integer Sequences}, entry A006625, \url{https://oeis.org/A006625}, accessed November 9, 2026. \bibitem{rosero} E. Rosero, \emph{Zarankiewicz numbers $z(m,n;3,4)$: SAT witnesses and DRAT-certified exact values (OEIS A006615 and A006625)}, Zenodo, 2026. \href{https://doi.org/10.5281/zenodo.21814869}{doi:10.5281/zenodo.21814869}. \end{thebibliography} \clearpage \appendix \section{The seventeen maximal representatives} The matrices below are the representatives in Lemma~\ref{lem:classification}. Each displayed binary string is one row, with columns read from left to right. The labels correspond to the families in Table~\ref{tab:extensions}. These matrices are used only as $8\times14$ bases; the all-one fifteenth column is adjoined when applying the extension test. \medskip \begingroup\scriptsize\setlength{\arraycolsep}{1pt}\renewcommand{\arraystretch}{.9} \noindent \smallmatrixblock{B_{1}}{11111111000000}{11111000111000}{11000110110110}{10100101101101}{10010011011011}{01001101011110}{00011110101011}{00110001110101} \smallmatrixblock{B_{2}}{11111111000000}{11110000111100}{11001110100011}{11000001011111}{00111100011011}{00101011110110}{00010111101101}{10011001110001} \smallmatrixblock{B_{3}}{11111111000000}{11110000111100}{11001100110011}{10100011101011}{01011010010111}{00111101100110}{00000111111101}{01100110001110} \smallmatrixblock{B_{4}}{11111111000000}{11110000111100}{11001100110011}{10100011101011}{01101010010111}{00111101100110}{00000111111101}{10010110011010} \smallmatrixblock{B_{5}}{11111111000000}{11110000111100}{11001100110011}{10100011101011}{01101010010111}{00111101100110}{00000111111101}{10010110001110} \par\medskip\noindent \smallmatrixblock{B_{6}}{11111111000000}{11110000111100}{11001000111011}{10101110100110}{01000111010111}{00110101101011}{00011011011101}{01101001001110} \smallmatrixblock{B_{7}}{11111111000000}{11110000111100}{11001100110011}{11000011001111}{00111010101011}{00110101010111}{00001111111100}{10100110100110} \smallmatrixblock{B_{8}}{11111111000000}{11110000111100}{11001100110011}{10001011101110}{01000111011101}{00111010011011}{00110101100111}{10100110010110} \smallmatrixblock{B_{9}}{11111111000000}{11110000111100}{11001100110011}{10001011101110}{01100011011011}{00110110100111}{00011101011101}{01011010101001} \smallmatrixblock{B_{10}}{11111111000000}{11110000111100}{11001100110011}{11000011001111}{00111100001111}{00110011110011}{00001111111100}{10101010101010} \par\medskip\noindent \smallmatrixblock{B_{11}}{11111111100000}{11110000011110}{11001110011001}{10101001010111}{01100101101101}{00111010101011}{01010011010011}{00011101011100} \smallmatrixblock{B_{12}}{11111111100000}{11111000011100}{11000111011010}{10100100110111}{01010010101111}{00011101111001}{00110011010011}{00101110001101} \smallmatrixblock{B_{13}}{11111111100000}{11111000011110}{11000110011101}{10100101110011}{01010011101011}{00101011010101}{00011101001101}{00001110111010} \smallmatrixblock{B_{14}}{11111111100000}{11111000011110}{11000111011100}{00110110011011}{00101101110101}{10010100100111}{01001010110011}{00011011001101} \smallmatrixblock{B_{15}}{11111111100000}{11111000011110}{11000111011100}{00110110011011}{00101101110101}{10010100100111}{10001010110011}{00011011001101} \par\medskip\noindent \smallmatrixblock{B_{16}}{11111111110000}{11111000001110}{11000111001101}{00110100111101}{00001011111110}{10100010101011}{10010001100111}{01010001011011} \smallmatrixblock{B_{17}}{11111111110000}{11111100001110}{11000011111110}{00110011001101}{00101010101011}{00100101100111}{00011010010111}{00010101011011} \endgroup \end{document}