\documentclass[11pt,reqno]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[a4paper,textwidth=155mm,textheight=244mm,centering,headsep=6mm,footskip=10mm]{geometry} \usepackage{microtype,mathtools,booktabs} \usepackage{tikz,flafter} \usepackage[hidelinks]{hyperref} \numberwithin{equation}{section} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \newcommand{\Z}{\mathbb Z} \setlength{\emergencystretch}{1.2em} \title[The unfair 0--1 conjecture for trinomial factors]{The unfair 0--1 polynomial conjecture\\ for trinomial factors} \date{18 November 2026} \hypersetup{pdftitle={The unfair 0-1 polynomial conjecture for trinomial factors}} \begin{document} \begin{abstract} Let two monic polynomials with nonnegative real coefficients have a product whose coefficients are all zero or one. If one factor has exactly three nonzero terms, we prove that every nonzero coefficient in both factors equals one. The exponents of the three terms and the degree of the other factor are arbitrary. Equivalently, if two independent, finitely supported integer-valued random variables have a sum uniform on its support, and one takes exactly three values, each is uniform on its own support. A fractional middle coefficient would have an algebraic conjugate of modulus greater than one. At this conjugate the recurrence forces certain differences between coefficients of the other factor to grow. We bound explicit linear combinations of these differences by counting walks on the integers and applying polynomial approximation on a real interval. \end{abstract} \maketitle \begin{center} \small The proof was generated by AI under the direction of Alessandro Baravelli. \end{center} \section{Statement and normalization} Call a polynomial \emph{binary} if all its coefficients are zero or one. A monic polynomial has leading coefficient one. A binary polynomial is \emph{fair} if, in every factorization into monic polynomials with nonnegative real coefficients, both factors are binary. The unfair 0--1 polynomial conjecture asserts that every monic binary polynomial is fair. We prove the conjectured conclusion whenever one factor has three nonzero terms. \begin{theorem}\label{thm:main} Let $A,B\in\mathbb R[X]$ be monic with nonnegative coefficients. If $AB$ is binary and $A$ has exactly three nonzero terms, then $A$ and $B$ are binary. \end{theorem} Ghidelli proved the case of exponents $0,2,5$ by studying a recurrence for the cofactor~\cite{ghidelli}. For exponents $0,2,m$, Dvorsky's combined theorem covers $m\ge7$~\cite[Theorem~1.1]{dvorsky}, using a companion manuscript for odd $7\le m\le339$~\cite{dvorsky-finite}. Hare tested patterns of zero, one, and fractional coefficients in candidate factors of degree at most fifteen~\cite{hare}. Hasanbegovi\'c gave a sufficient separation between two blocks of product coefficients for fairness of the lower block to imply fairness of the whole product~\cite{hasanbegovic}. For the probabilistic formulation, translate both supports to start at zero. By independence, the probability generating polynomials multiply to that of the sum. If the sum is uniform on its support, dividing each factor by its leading coefficient makes the product binary. Theorem~\ref{thm:main} then gives uniformity on each support. Conversely, given $A,B$ as in the theorem, take independent variables with generating polynomials $A/A(1)$ and $B/B(1)$. Their sum has generating polynomial $AB/(A(1)B(1))$, so its law is uniform on its support. The probabilistic assertion makes the positive coefficients of $A$ equal, and likewise for $B$. Monicity makes their common value one in each factor. Remove the largest power of $X$ dividing each factor. The resulting constant terms are positive and multiply to one. Each is at most one, since it contributes to a coefficient of the product with the opposite leading coefficient. Hence both constant terms equal one. Pairing with the opposite constant term now bounds every coefficient by one. Write \[ A=1+aX^r+X^m,\qquad B=\sum_n b_nX^n,\qquad C=AB=\sum_n c_nX^n \] with $01$. In the other cases we may reverse both polynomials, replacing each $P(X)$ by $X^{\deg P}P(1/X)$, so that $m>2r$. \begin{lemma}\label{lem:annulus} The middle coefficient $a$ has an algebraic conjugate $\lambda$ of modulus $1<|\lambda|\le2$. \end{lemma} \begin{proof} Every root of $A$ is a root of the monic integer polynomial $C$ and hence an algebraic integer. The coefficients of $A$ are signed elementary symmetric functions of those roots, so $a$ is an algebraic integer. Its conjugates, the roots of its minimal polynomial over $\mathbb Q$, have product a nonzero integer up to sign. Since $00$, the digit $c_{n+r}$ is one, so \[ 1-\lambda\sigma(b_n)=\sigma(b_{n+r})+\sigma(b_{n+r-m}). \] For $b_n=0$ the left side is $1=1+0$. Thus for every value $z$ in the set, $1-\lambda z$ is the sum of two values in it. Choose two points at the greatest distance $d$. Subtracting their two sums gives $|\lambda|d\le2d$ by the triangle inequality, and $d\ge1$. \end{proof} Fix such an embedding and put $R=|\lambda|$. We use the recurrence twice. At $a$, nonnegativity forces a stretch of at least five consecutive steps on which four neighboring equations all have right-hand side one. At $\lambda$, these ones survive the embedding. Subtracting successive equations along the stretch cancels their right-hand sides and forces growth in the cofactor differences. Taking a longest such stretch bounds the initial values inherited from its predecessor. Sections~\ref{sec:filter} and~\ref{sec:short} compare this growth with an upper bound on the same linear combinations of differences. The upper bound uses only the binary digits of the full product and is independent of its degree. \section{The lowest positive coefficient in each column}\label{sec:bottom} Place $b_{im+\ell r}$ in row $i$, column $\ell$, and denote it by $b_{i,\ell}$. Make the same array for the product digits $c_{i,\ell}$. For example, for $r=2$, $m=5$ the entries of row zero starting at column zero are $b_0,b_2,b_4,\ldots$, and the row immediately above starts $b_5,b_7,b_9,\ldots$. A step right adds $r$ to the exponent and a step up adds $m$. Thus \begin{equation}\label{eq:array} b_{i,\ell}+b_{i-1,\ell}+a b_{i,\ell-1}=c_{i,\ell}. \end{equation} The array may repeat coefficients or use only indices divisible by $\gcd(r,m)$. Every displayed equation still follows from~\eqref{eq:basic}. \begin{lemma}\label{lem:geometry} Every column has a lowest positive entry. Its row $s(\ell)$ satisfies \[ s(\ell)\le s(\ell-1),\qquad s(\ell+m)=s(\ell)-r . \] Call a strict decrease a \emph{drop}. If its depth is $d=s(\ell-1)-s(\ell)$, the $d$ entries from the new bottom up to just below the old bottom are binary. They start with one, end with zero, and have no adjacent ones. Over one period, the depths of the drops sum to $r$ and the distances between consecutive drops sum to $m$. \end{lemma} \begin{proof} Suppose $b_{i,\ell-1}>0$. Then $00. \] Thus the next column has a positive entry at or below row $i$. Starting with $b_{0,0}=1$ gives nonempty columns to the right. The identity $b_{i,\ell+m}=b_{i+r,\ell}$ gives nonemptiness for all integer $\ell$. Each column has finite support, so its bottom exists. Taking $i=s(\ell-1)$ proves monotonicity, and the same shift identity gives the formula for $s(\ell+m)$. At a drop write $s=s(\ell)$ and $p=s(\ell-1)$. In rows $s\le i0$, there is a drop. The drop positions repeat after $m$ steps, so the successive distances, including the distance to the first drop of the next period, sum to $m$. \end{proof} In particular $r=1$ is impossible, since even one drop has depth at least two. Call the columns from one drop to the next, including both, a \emph{strip}. Its width $w$ is the number of steps between the drops. If there are $D$ drops per period, then $D\le r/2$. A widest strip has width $W\ge m/D\ge 2m/r>4$. Let $v$ be the width immediately preceding it. Thus \begin{equation}\label{eq:max} W\ge5,\qquad 1\le v\le W . \end{equation} With a single drop per period, $v=W=m$. Use coordinates $(j,k)$ measured from the new bottom at the first drop of a strip. All entries with $j<0$ vanish for $0\le k0$, the row-zero digit at step $k$ must be one, and $p_k=1-a p_{k-1}>0$. This positive entry in turn forces the row-one digit to be one. On every strip we therefore have \begin{equation}\label{eq:two-bottom} p_k=1-a p_{k-1},\qquad q_k=1-p_k-aq_{k-1}\qquad(1\le k\le w). \end{equation} \begin{corollary}\label{cor:half} The middle coefficient satisfies $00$, which forces the row-three digit to be one. \end{proof} At the last column of any strip, a drop of depth $d$ changes the row label of each entry from $j$ to $j+d$. Figure~\ref{fig:shared-column} shows this change of coordinates. In the example $r=2$, $m=5$, there is one drop per period, so $w=5$ and the two rows $0,1$ become rows $2,3$. \begin{figure}[ht] \centering \begin{tikzpicture}[x=16mm,y=5.4mm,font=\small] \fill[black!6] (-.48,-3.45) rectangle (3.50,-.48); \draw[black!40] (-.48,1.56) -- (4.46,1.56); \draw[black!40] (4.58,-3.46) -- (4.58,1.56); \node at (-1.05,2) {row $j$}; \node at (5.62,2) {row in next strip}; \node[font=\footnotesize] at (2,2.82) {column}; \foreach \k/\lbl in {0/0,1/1,2/{\cdots},3/{w-1},4/w} \node at (\k,2) {$\lbl$}; \foreach \j/\old/\newrow in {1/1/{d+1},0/0/d,-1/{-1}/{d-1},-2/{\vdots}/{\vdots},-3/{-d}/0} { \node at (-1.05,\j) {$\old$}; \node at (5.62,\j) {$\newrow$}; } \node at (0,0) {$1$}; \node at (0,1) {$0$}; \foreach \k/\lbl in {1/1,3/{w-1},4/w} { \node at (\k,0) {$p_{\lbl}$}; \node at (\k,1) {$q_{\lbl}$}; } \foreach \j in {1,0,-1,-3} \node at (2,\j) {$\cdots$}; \foreach \k in {0,1,3} { \node[text=black!60] at (\k,-1) {$0$}; \node[text=black!60] at (\k,-3) {$0$}; } \node[text=black!60] at (1.5,-2) {all zero}; \node at (4,-1) {$0$}; \node at (4,-2) {$\vdots$}; \node at (4,-3) {$1$}; \draw[line width=.9pt] (-.48,-.48) -- (3.50,-.48) -- (3.50,-3.48) -- (4.46,-3.48); \draw[rounded corners=2pt] (3.65,-.43) rectangle (4.35,1.43); \end{tikzpicture} \caption{A strip of width $w$ ending in a drop of depth $d$. The heavy line follows the bottom. At column $w$ the new binary word occupies old rows $-d,\ldots,-1$ and the boxed entries pass to rows $d,d+1$ of the next strip.}\label{fig:shared-column} \end{figure} \section{Differences between consecutive columns}\label{sec:diff} Apply $\sigma$ entrywise to $B$ and write $B_\lambda=\sum_n\sigma(b_n)X^n$. In a fixed strip, write $V_{j,k}$ for its embedded cofactor entry at $(j,k)$. Encode rows zero through three as $V_k(y)=\sum_{j=0}^3 V_{j,k}y^j$, doing arithmetic modulo $y^4$. Set $S=1+y+y^2+y^3$. Lemma~\ref{lem:four} and the vanishing of row $-1$ give \[ (1+y)V_k+\lambda V_{k-1}=S \quad(1\le k\le w). \] Write $V_0=1+Hy^2+Gy^3$ and define $U,V$ by \[ ((1+y+\lambda)V_0-S)/\lambda=1+Uy^2+Vy^3\pmod{y^4}. \] For $k=1$, $(1+y)(V_1-V_0)=S-(1+y+\lambda)V_0$. Each subsequent difference is the previous one multiplied by $-\lambda/(1+y)$, by subtraction of consecutive recurrence equations. Hence \begin{equation}\label{eq:diff} V_k-V_{k-1}=(-\lambda)^k(1+y)^{-k}(1+Uy^2+Vy^3)\pmod{y^4}. \end{equation} In particular all dependence on the previous strip is confined to $U,V$. Let $p_v(\lambda),q_v(\lambda)$ be the sums in Lemma~\ref{lem:four} evaluated at $\lambda$ at the final step of the previous strip of width $v$. Its last column is the current strip's first. If the new drop has depth two, the word $10$ sits immediately below those two entries, so $H=p_v(\lambda)$ and $G=q_v(\lambda)$. These are the images under $\sigma$ of the boxed entries in Figure~\ref{fig:shared-column} with $w=v$. If the depth is three, the word is $100$, giving $H=0$, $G=p_v(\lambda)$. For a deeper drop, $H,G$ are adjacent digits in the new word and so form one of $(0,0),(1,0),(0,1)$. The finite sums satisfy \[ (1+\lambda)p_v(\lambda)=1+\lambda(-\lambda)^v,\qquad (1+\lambda)q_v(\lambda)+p_v(\lambda)-1=v(-\lambda)^{v+1}. \] Substitution in the definition of $U,V$ gives \begin{equation}\label{eq:start} (U,V)= \begin{cases} ((-\lambda)^v,-v(-\lambda)^v),&d=2,\\ (-1/\lambda,(-\lambda)^v),&d=3,\\ (-1/\lambda,-1/\lambda),\ (1,0),\ \text{or }(-1/\lambda,1),&d\ge4. \end{cases} \end{equation} In every case $|U|\le R^v$ and $|V|\le vR^v$. \section{A global bound for the column differences}\label{sec:filter} At a fixed step $1\le k\le w$ of a strip, we use the binary product digits to bound a linear combination of the four differences $V_{j,k}-V_{j,k-1}$ in~\eqref{eq:diff}. The common factor $(1+y)^{-k}$ has alternating coefficients. In its product with $(1-y)^2$, each coefficient therefore has magnitude equal to the sum of the adjacent magnitudes with weights $1,2,1$. We apply this same second difference to the full polynomial in~\eqref{eq:diff}. The other factor of the filter will control the global estimate. The endpoint terms in~\eqref{eq:array} act on rows by multiplication by $1+y$. Write $y=T^2$, so $1+y=TJ$, where $J=T+T^{-1}$ takes one step in either direction. After removing the shift $T$, the endpoint terms act by $J$ and the middle term shifts the column with coefficient $\lambda$. With a real variable $u$ in place of $J$, a homogeneous column solution satisfies $u h_s=-\lambda h_{s-1}$. It decays towards increasing column index $s$ for $|u|>R$ and towards decreasing $s$ for $|u|10$ and $u=26/z<3$, while $u=R^{k-2}>1$. Since $(u-1)(z-10)>0$, we have $10u+zR),\qquad f_{-j-1}(u)=p(u)\frac{(-u)^j}{R^{j+1}}\quad(|u|