\documentclass[12pt,a4paper]{amsart} \usepackage[margin=25mm]{geometry} \usepackage[T1,OT1]{fontenc} \usepackage{amssymb} \usepackage[hidelinks]{hyperref} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newcommand{\F}{\mathbb F} \newcommand{\sep}{\operatorname{sep}} \newcommand{\Disc}{\operatorname{Disc}} \newcommand{\Res}{\operatorname{Res}} \newtheorem{proposition}[theorem]{Proposition} \emergencystretch=1em \title[{Root separation over $\mathbb F_q[T]$}]{Root separation for polynomials over $\mathbb F_q[T]$} \author{} \thanks{\usefont{T1}{lmr}{m}{n}Tomislav Pejković proposed the topic and guided the research presented in this paper. Department of Mathematics, Faculty of Science, University of Zagreb, Bijenička cesta 30, 10000 Zagreb, Croatia (pejkovic@math.hr).} \subjclass[2020]{Primary 11C08; Secondary 11J61, 11T55} \keywords{Root separation, polynomial height, function fields of positive characteristic, Diophantine approximation} \date{} \hypersetup{ pdftitle={Root separation for polynomials over F\_q[T]}, pdfauthor={}, pdfsubject={Number Theory (math.NT)}, pdfkeywords={Root separation, polynomial height, function fields of positive characteristic, Diophantine approximation} } \begin{document} \begin{abstract} We study the minimum distance between distinct roots of a separable polynomial of degree $d\ge2$ over $\mathbb F_q[T]$, with respect to the absolute value $|T|=q$. For reducible polynomials of height $H$, the exponent $d-1$ in the lower bound $H^{-(d-1)}$ is best possible in every degree. In characteristic two we construct absolutely irreducible polynomials that attain this bound at every height $q^N$, $N\ge1$. In odd characteristic $p$, the exponent $d-1$ is best possible for absolutely irreducible polynomials whenever $d-1=p^a m$ with $p\nmid m$ and $m\mid p^a-1$. The separation exponents for irreducible and absolutely irreducible polynomials coincide for every $q$ and $d$. \end{abstract} \maketitle {\small\raggedright \noindent\textbf{Use of artificial intelligence.}\enspace The research was carried out under human direction from the choice of topic and research questions to the selection of results for the paper. OpenAI’s \mbox{ChatGPT 6.2 Astra} developed the constructions and proofs and drafted the manuscript. Anthropic’s \mbox{Claude Fable 5.2} reviewed the manuscript, checked examples, and suggested improvements to the proofs and to the comparison with known results. The human review did not include an independent check of every proof.\par} \section{Introduction} How close can two distinct roots of an integer polynomial of given degree and height be? For a separable integer polynomial of degree $d$ and height $H$, Mahler~\cite{Mahler1964} gave a lower bound of order $H^{-(d-1)}$ for the distance between any two distinct roots. In the other direction, Bugeaud and Dujella~\cite[Theorem~1]{BD2014} constructed, for each $d\ge4$, reducible integer polynomials of arbitrarily large height $H$ with two roots at distance $\ll_d H^{-(2d-1)/3}$. We refer to Bugeaud and Mignotte~\cite{BugeaudMignotte2010} for a survey. Here we consider polynomials with coefficients in $k[T]$, where $k=\F_q$ is the finite field with $q$ elements. Write $p$ for the characteristic of $k$ and put $K=k(T)$. For nonzero $a,b\in k[T]$ set $|a/b|=q^{\deg a-\deg b}$. The completion of $K$ for this absolute value is the field of Laurent series $K_\infty=k((T^{-1}))$. We extend this absolute value to an algebraic closure of $K_\infty$ and write $v=-\log_q|\cdot|$. Throughout the paper, $F=\sum_{i=0}^d a_iX^i\in k[T][X]$ is a primitive separable polynomial of degree $d\ge2$ in $X$, with roots $x_1,\ldots,x_d$. Thus the coefficients $a_i$ have no common factor of positive degree. We define its height $H(F)$ and its separation $\sep(F)$ by \[ \begin{aligned} h(F)&=\max_i\deg_T a_i, & H(F)&=q^{h(F)},\\ \sep(F)&=\min_{i1$ of $d$ and consider irreducible polynomials for which $K(x)$, with $x$ a root, has constant field $\F_{q^c}$. For this class the exponent is $2d/c^2$. It follows that $E_{q,\mathrm{irr}}(d)=E_{q,\mathrm{abs}}(d)$ for all $q$ and $d$. \section{Reducible polynomials}\label{sec:reducible} \subsection{Estimates from the discriminant and the resultant} The Gauss norm of a polynomial over a non-Archimedean valued field is the maximum of the absolute values of its coefficients. It is multiplicative. Indeed, after dividing each factor by a coefficient of maximal absolute value, both factors have nonzero reduction modulo the maximal ideal, and so does their product. In particular, $h(AB)=h(A)+h(B)$. For $F=a_d\prod_i(X-x_i)$, put $M_i=\max(1,|x_i|)$. Then $H(F)=|a_d|\prod_i M_i$. Choose two roots at distance $\sep(F)$. For every other pair use $|x_i-x_j|\le M_iM_j$ in the formula for the discriminant. Since $\Disc F$ is a nonzero polynomial in $T$, we obtain \[ \sep(F)\ge |\Disc F|^{1/2} H(F)^{-(d-1)}\ge H(F)^{-(d-1)}. \] For irreducible polynomials see also Ooto~\cite[Lemma~3.3]{Ooto2017}. Applying the same argument to the resultant of two coprime polynomials $A,B$ of degrees $i,j$, respectively, gives \begin{equation}\label{eq:res} v(x-y)\le jh(A)+ih(B)-\deg_T\Res_X(A,B) \end{equation} for every root $x$ of $A$ and every root $y$ of $B$. \subsection{A lower bound for \texorpdfstring{$\Delta$}{Delta} and examples of equality} Let $W$ be reducible of degree $n+1$. For $n=1$, \eqref{eq:bound} gives $\Delta(W)\ge0$. Suppose that $n\ge2$. If $s(W)\le(n-1)h(W)$, then $\Delta(W)\ge h(W)$. Otherwise, two roots at distance $\sep(W)$ must belong to different irreducible factors by~\eqref{eq:bound}. If both factors had degree at most $n-1$, \eqref{eq:res} would give $s(W)\le(n-1)h(W)$. Thus one factor has degree $n$ and the other is linear. Applying~\eqref{eq:res} to them gives \begin{equation}\label{eq:lossres} W(X)=F(X)(bX-a),\qquad \Delta(W)\ge(n-1)h(F)+\deg_T\Res_X(F,bX-a), \end{equation} where $F$ is irreducible of degree $n$ and $a,b\in k[T]$ are coprime. If $h(F)=0$, each root of $F$ is a constant outside $k$. Its distance from $a/b$ is at least one, since $a/b$ has residue in $k$ whenever $|a/b|\le1$. This contradicts $s(W)>(n-1)h(W)$, so $h(F)\ge1$. We conclude that $\Delta(W)\ge n-1$ whenever $h(W)\ge n-1$. Suppose first that $p\nmid n$. Let $r$ run through the powers of $p$ congruent to $1$ modulo $n$, and put $j=(r-1)/n$. Take $F=TX^n-(T+1)$ and $u=T^j/(T+1)^j$. Then $(T+1)^{nj}F(u)=-1$, so $v(F(u)/T)=r$. The reduction of $F/T$ modulo $T^{-1}$ is $X^n-1$, which has $n$ simple roots. Since $u$ reduces to $1$, exactly one root $x$ of $F$ satisfies $v(x-u)>0$, and for this root $v(x-u)=r$. The product $F(X)((T+1)^jX-T^j)$ therefore has logarithmic height $j+1$ and $\Delta=n-1$. Eisenstein's criterion at $T+1$ shows that $F$ is absolutely irreducible. Next suppose that $n=P=p^e$ with $e\ge1$, and take $F=X^P+TX-1$. It is separable since $F_X=T$. As a polynomial in $T$ it is linear with coprime coefficients $X$ and $X^P-1$, so it is absolutely irreducible. Put $B_j=(P^j-1)/(P-1)$ and \[ \alpha=\sum_{i\ge1}(-1)^{i-1}T^{-B_i},\qquad u_j=\sum_{i=1}^j(-1)^{i-1}T^{-B_i}. \] The recurrence $B_{i+1}=PB_i+1$ gives $T\alpha=1-\alpha^P$, so $F(\alpha)=0$. Moreover, $h(u_j)=B_j$ and $v(\alpha-u_j)=PB_j+1$. The difference $z$ between any two distinct roots of $F$ satisfies $z^{P-1}=-T$, whence $|z|>1$. Thus the product $W_j=F(X)T^{B_j}(X-u_j)$ has logarithmic height $B_j+1$ and satisfies $s(W_j)=PB_j+1$. It follows that $\Delta(W_j)=P-1$. \subsection{A construction for the remaining degrees}\label{subsec:mixed} Suppose that $n=mP$, where $P>1$ is a power of $p$, $m>1$ and $p\nmid m$. We construct a polynomial $Q(Z,X)$ of degree $n$ in $X$ and a polynomial $V(Z)$ such that $Q(Z,V(Z))$ has a zero of high order at $Z=1$. Take arbitrarily large powers $R=P^t>n$ with $R\equiv1\pmod m$. Put $M=(R-1)/m$ and \[ V(Z)=\sum_{i=0}^{t-1} Z^{MP^i}. \] Then $V^P-V+Z^M=Z^{MR}$. Raising both sides to the $m$th power and multiplying by $Z^R$ gives \[ Q(Z,X)=Z^R(X^P-X+Z^M)^m-1,\qquad Q(Z,V(Z))=Z^{R^2}-1=(Z-1)^{R^2}. \] Write $D=\deg_Z Q=2R-1$ and $E=\deg V=MR/P$. As a polynomial in $X$, $Q$ has leading coefficient $Z^R$ and constant term $Z^{2R-1}-1$, so it is primitive. The polynomial $Q(1,X)=(X^P-X+1)^m-1$ has $n$ simple roots, one of which is $V(1)$. By Hensel's lemma there is a unique root $x$ of $Q$ in $\bar k[[Z-1]]$ with $x(1)=V(1)$. The factorization of $Q(Z,V(Z))$ shows that $x-V$ has a zero of order $R^2$ at $Z=1$, since the leading coefficient of $Q$ and the other factors are units there. We now prove that $Q$ is absolutely irreducible. Since $V(1)$ is a simple root of $Q(1,X)$, exactly one irreducible factor $C$ of $Q$ over $\bar k[Z,X]$ vanishes at $(1,V(1))$, and it occurs only once. It follows that $C(Z,V(Z))$ is divisible by $(Z-1)^{R^2}$. Now $R+nE=R^2$ and $E=(R-1)R/n>R-1$, so \[ D+(n-1)E=R^2+R-1-E1$. Then $n\ge6$. Let $v_0$ be a valuation of the splitting field of $f(X)=T$ that is trivial on $k_0$. Since $k_0$ is infinite, a fractional linear transformation over $k_0$ makes the reductions of all the roots finite. If $x_1,x_2$ have the same reduction and $x_3,x_4$ both reduce to other points, then \[ v_0\!\left(\frac{(x_1-x_2)(x_3-x_4)}{(x_1-x_3)(x_2-x_4)}\right)>0. \] Indeed, the numerator vanishes upon reduction and the denominator is a unit. This is impossible because the cross-ratio is a nonzero constant. Thus, if two roots have the same reduction, at least $n-1$ roots have that reduction. Since fractional linear transformations preserve cross-ratios, this conclusion also holds after a change of coordinates over an extension field. We apply this observation to the fibers of $f$. Factor the homogeneous equation of a fiber into linear forms and normalize each form to have Gauss norm one. The scalar multiplying their product is a unit, since the homogeneous numerator and denominator of $f$ have no common zero. Reduction therefore gives the roots of the specialized equation with their multiplicities. Hence each ramified fiber consists either of a single point of multiplicity $n$ or of two points of multiplicities $n-1$ and $1$. We claim that $f$ is totally ramified at some point. Otherwise every ramification index greater than one would equal $n-1$, which is prime to $p$. Write $f=A/B$ in lowest terms and homogenize $A'B-AB'$ to degree $2n-2$. The zeros of this form are the critical points of $f$. At a point of ramification index $e$ prime to $p$, differentiation in local coordinates shows that the form vanishes to order $e-1$. Thus every zero would have multiplicity $n-2$, forcing $n-2$ to divide $2n-2$. This is impossible for $n\ge6$. Choose coordinates on the source and target so that a totally ramified point and its image are both zero. The numerator has a zero of order $n$ at $0$, so $f(X)=aX^n/B(X)$ with $aB(0)\ne0$. Substituting $T=t^n$ and $X=tY$ gives \[ aY^n-B(tY)=0 . \] The leading coefficient is a unit at $t=0$. The reduction $aY^n-B(0)$ has $m$ distinct roots, each of multiplicity $P$. Thus exactly $P$ of the roots reduce to each of these points. Since $11$ and $x_i^2+T^N=A(x_i)/B(x_i)$. Since $A$ and $B$ are monic and $\deg A\le\deg B$, the right-hand side has absolute value at most one. Hence $|x_1|=|x_2|=R$, where $R=q^{N/2}$. In characteristic two we have $BF'-B'F=(AB)'=1$, so $F$ is separable and $|F'(x_1)|=|B(x_1)|^{-1}=R^{-\ell}$. Each of the $\ell$ roots of absolute value at most one is at distance $R$ from $x_1$. Since $F$ is monic, the identity $F'(x_1)=\prod_{x\ne x_1}(x_1-x)$ gives $R^\ell|x_1-x_2|=R^{-\ell}$. The distance between the corresponding roots of $G$ is therefore $R^{-2\ell-2}=q^{-(d-1)N}$, as required. \end{proof} \section{Absolutely irreducible polynomials in odd characteristic}\label{sec:odd} In odd characteristic we can strengthen~\eqref{eq:bound} for absolutely irreducible polynomials of logarithmic height at least two. \begin{proposition}\label{prop:oddheight} Let $F$ be a separable absolutely irreducible polynomial of degree $d\ge2$ and logarithmic height $h\ge2$. If $p$ is odd, then \[ s(F)\le(d-1)(h-1)+\tfrac12, \qquad\text{or equivalently}\qquad \Delta(F)\ge d-\tfrac32. \] \end{proposition} \begin{proof} Put $t=1/T$. Choose $a\in\bar k$ different from the residue of every root $x_i$ of $F$ for which $|x_i|\le1$. Then $|x_i-a|=M_i=\max(1,|x_i|)$ for every $i$. The roots $y_i=1/(x_i-a)$ of the polynomial \[ G(t,Y)=t^hY^dF(1/t,a+1/Y) \] all have absolute value at most one. Moreover, $\deg_t G\le h$, and its leading coefficient as a polynomial in $Y$ is nonzero at $t=0$ because $|F(T,a)|=H(F)$. Substituting $|y_i-y_j|=|x_i-x_j|/(M_iM_j)$ in the formula for the discriminant gives \[ \sum_{im-\tfrac12$. As $h\ge2$, we have $2s(F)>(d-1)h$. At least one pair satisfies $v(y_i-y_j)\ge s(F)$, since $M_iM_j\ge1$. All terms in the sum above are nonnegative, so there is exactly one such pair, say $y_1,y_2$. Every automorphism of the splitting field of $G$ over $\bar k((t))$ preserves the set $\{y_1,y_2\}$. Hence $b=(y_1+y_2)/2$ belongs to $\bar k[[t]]$ and $(y_1-y_2)^2$ belongs to $\bar k((t))$. In particular $2v(y_1-y_2)$ is an integer greater than $2m-1$. It follows that $v(y_i-b)\ge m$ for $i=1,2$. We now choose a nonzero polynomial $Q\in\bar k[t,Y]$ of degree at most $h-1$ in $t$ and at most $d-1$ in $Y$. The condition $Q(t,b)\equiv0\pmod{t^m}$ gives at most $m$ homogeneous linear equations in the $hd$ coefficients of $Q$. Choose $h+d-3$ distinct nonzero elements $c\in\bar k$ at which the leading coefficient of $G$ in $Y$ does not vanish. For each $c$, choose a root $z$ of $G(c,Y)$ and impose the additional equation $Q(c,z)=0$. Since $m+h+d-3=hd-1$, the system has a nonzero solution. The resultant $D(t)=\Res_Y(G,Q)$ is nonzero since $G$ is irreducible over $\bar k(t)$ and $\deg_Y Q1.\end{cases} \] For $P>1$ the exponent of $U$ is an integer by hypothesis. \emph{A critical point of a rational function.} Put $u=L(x)$ and consider the rational function of degree $n+1$ \[ f(x)=x+\frac{au}{u^m-1}. \] We choose $a$ so that we can describe a zero of $f'$ explicitly. If $w=u^m-1\ne0$, the equation $f'(x)=0$ becomes $w^2=\epsilon a\{m+(m-1)w\}$. Completing the square gives \[ \left(\frac{2m}{w}+m-1\right)^2=(m-1)^2+\frac{4m\epsilon}{a}. \] In the notation of Lemma~\ref{lem:quadraticapprox}, let $z=\sigma-\sigma^{-1}$. Since $z^*=-z$ and $z^2=T^2-4$, we may take \[ a=\frac{4m\epsilon}{z^2-(m-1)^2}\in k(T),\qquad w=\frac{2m}{z-(m-1)}. \] At $\sigma=0$ we have $v(\sigma)=1$, since $1/T=\sigma/(1+\sigma^2)$. In particular $v(w)=1$. Modulo $\sigma$, the equation $L(x)^m=1+w$ has only simple roots. By Hensel's lemma it has a root $x_*$ in $\mathcal O=\bar k[[\sigma]]$. Put $t_*=f(x_*)$. Then \[ J=L(t_*)^m=(1+w)\{1+D_{1+w}(a/w)\}^m\in k(\sigma). \] Since $\epsilon a/w=w/(m+(m-1)w)$, the expansion at $w=0$ gives $D_{1+w}(a/w)=w/m+O(w^2)$ and hence $J=1+2w+O(w^2)$. Under $\sigma\mapsto\sigma^{-1}$, the function $w$ becomes $-mw/(m+(m-1)w)=-w+O(w^2)$. It follows that $v(J-J^*)=1$. \emph{Two conjugate roots.} For $P>1$ take $R=P^j$ and let $j$ tend to infinity. For $P=1$ let $R$ run through the powers of $p$ congruent to one modulo $m$. Put $\beta=B_R(J)$. Lemma~\ref{lem:quadraticapprox} gives \[ h(\beta)\le R/2+O_n(1),\qquad v(\beta-J)=R+1. \] Write $L^m(X)=L(X)^m$. Since $(L^m)'(t_*)$ is a unit, Hensel's lemma gives $t\in\mathcal O$ with $L(t)^m=\beta$ and $v(t-t_*)=R+1$. On clearing denominators in $f(X)=t$, we obtain the monic polynomial \[ N_t(X)=(X-t)(L(X)^m-1)+aL(X). \] As $v(a/w)>0$, we have $t\equiv t_*\equiv x_*\pmod{\sigma}$. Thus $N_t$ reduces to $(X-\bar x_*)(L(X)^m-1)$, which has $n-1$ simple roots and a double root at $\bar x_*$. The simple roots lift to roots in $\mathcal O$. Dividing out their linear factors and substituting $X=x_*+Z$, we obtain \[ N_t(x_*+Z)=U(Z)(Z^2+c_1Z+c_0), \qquad U\in\mathcal O[Z],\quad U(0)\in\mathcal O^\times,\quad c_0,c_1\in\mathcal O . \] Since $f(x_*)=t_*$ and $f'(x_*)=0$, we have $N_{t_*}(x_*)=N_{t_*}'(x_*)=0$. It follows that \[ N_t(x_*)=-(t-t_*)w,\qquad N_t'(x_*)=-(t-t_*) (L^m)'(x_*). \] Evaluating the factorization and its derivative at zero gives $v(c_0)=R+2$ and $v(c_1)\ge R+1$. Consequently $v(c_1^2-4c_0)=R+2$. This is odd, so the quadratic factor is irreducible over $\bar k((\sigma))$. The corresponding roots $x_1,x_2$ of $N_t$ satisfy $v(x_1-x_2)=(R+2)/2$, and all $n+1$ roots of $N_t$ are distinct. \emph{An equation over $k(T)$.} Put $b=L(t)$, so that $b^m=\beta$ is a unit. For each root $x$ of $N_t$, put $Y=(x^R-t)/b$. Then $L(x^R)=b(1+D_\beta(Y))$ and $(x-t)^R=b(Y-C)$, where $C=(t^R-t)/b$. If $P=1$, then $C=\beta^{(R-1)/m}-1$. If $P>1$, raising $t^P-t=b$ to the powers $1,P,\ldots,P^{j-1}$ and adding gives $C=\sum_{i=0}^{j-1}\beta^{(P^i-1)/m}$. In either case $C$ is a polynomial in $\beta$ of degree at most $(R/P-1)/m$. Raise $N_t(x)=0$ to the $R$th power and divide by $b$ to obtain \[ (Y-C)\bigl\{\beta(1+D_\beta(Y))^m-1\bigr\} +a^R(1+D_\beta(Y))=0 . \] This equation has degree $n+1$ over $k(T)$. Its roots are distinct because $x\mapsto x^R$ is injective. Let $F\in k[T][Y]$ be the primitive polynomial obtained by clearing denominators. The coefficients on the left are polynomials in $\beta$ and $a$ of degrees at most $P+(R/P-1)/m$ and $R$, respectively. It follows that \[ h(F)\le\bigl(P+(R/P-1)/m\bigr)h(\beta)+Rh(a) \le\frac{R^2}{2n}+O_n(R). \] The two roots $Y_i=(x_i^R-t)/b$, $i=1,2$, are conjugate over $\bar k((\sigma))$. They are therefore roots of the same irreducible factor $A$ of $F$ over $\bar k(T)$, taken primitive over $\bar k[T]$. Since $b$ is a unit, $v(Y_1-Y_2)=R(R+2)/2$. If $\deg_Y A\le n$, applying~\eqref{eq:bound} to $A$ gives \[ R(R+2)/2\le(n-1)h(A)\le(n-1)h(F) \le(n-1)R^2/(2n)+O_n(R), \] which is impossible for large $R$. Thus $F$ is absolutely irreducible. Applying~\eqref{eq:bound} to $F$ gives $nh(F)\ge R(R+2)/2$. Combining the two estimates for $h(F)$ gives $h(F)\asymp_n R^2$ and $\Delta(F)\le nh(F)-R(R+2)/2\ll_n R\ll_n\sqrt{h(F)}$. \end{proof} \section{Lower bounds in all degrees}\label{sec:uniform} We give a lower bound for the separation exponent in every degree. For odd degrees we also estimate the separation at finite height. \begin{theorem}\label{thm:universal} For every finite field $k=\mathbb F_q$ of characteristic $p$ and every $d\ge2$, \[ E_{q,\mathrm{abs}}(d)\ge \begin{cases} d-2,&d\text{ odd or }4\mid d,\\ d-4,& d\equiv2\pmod4,\ p\ne3,\\ d-8,& d\equiv2\pmod4,\ p=3 . \end{cases} \] If $d$ and $p$ are odd, put $j=1$ when $p\nmid d-1$ and $j=2$ otherwise. There are absolutely irreducible polynomials $F$ of degree $d$ and arbitrarily large logarithmic height $h$ for which $s(F)\ge(d-j)h-O_d(1)$. \end{theorem} Theorem~\ref{thm:abs} proves the assertion about the exponent if $p=2$ or $p\nmid d-1$. We first prove the estimate for odd $d$ and $p$. \subsection{Odd degrees}\label{subsec:odddegrees} Suppose that $d\ge3$ and $p$ are odd. Take $j=1$ if $p\nmid d-1$ and $j=2$ otherwise, and put $\nu=d-j$. Then $p\nmid j\nu$ and $\gcd(d,2\nu)=1$. We shall use a rational function $f$ of degree $d$ with a simple critical point. The two roots of $f(X)=t$ near this point will give the required estimate $s(F)\ge\nu h(F)-O_d(1)$. Let $c=\nu/j$ in $\mathbb F_p$ and put $A=(1+c)^2$, $B=(1-c)^2$. We choose a nonsquare $\lambda\in\mathbb F_p$ for which $(1-\lambda)(A-B\lambda)$ is also a nonsquare. To see that such a choice is possible, put $N=(p-1)/2$. For every nonsquare $v$, both $1-v$ and $A-Bv$ are nonzero, since $A$ and $B$ are squares and $A-B=4c\ne0$. If their product were always a square, then \[ P(X)=(A-BX)^N-(1-X)^N \] would vanish at all roots of $X^N+1$. As $\deg P\le N$, it would be a constant multiple of $X^N+1$. At least one of $A,B$ is a nonzero square, so either the constant coefficient or the coefficient of $X^N$ in $P$ is zero. This would give $P=0$, whereas $P(1)=(4c)^N\ne0$. Put $u=x^\nu$, $\rho(u)=(u-\lambda)/(1-u)$ and \[ t=f(x)=x^j\rho(u),\qquad z=t^\nu,\qquad D=\lambda^d,\qquad W=u+\lambda/u,\qquad U=z+D/z . \] The involution $u\mapsto\lambda/u$ sends $\rho$ to $\lambda/\rho$ and $z=u^j\rho^\nu$ to $D/z$. Thus $U$ belongs to the fixed field $k(W)$. Its degree in $W$ is $d$, since $z$ has degree $d$ in $u$ and both $u\mapsto W$ and $z\mapsto U$ have degree two. Away from the zeros and poles of $f$, the equation $f'(x)=0$ becomes \[ (u-\lambda)(1-u)+cu(1-\lambda)=0. \] Its discriminant is $(1-\lambda)(A-B\lambda)$, so its two roots $u_0$ and $\lambda/u_0$ are conjugate over $\mathbb F_p$. Neither is $0$, $1$ or $\lambda$. Write $z_0=u_0^j\rho(u_0)^\nu$. Then $z_0^p=D/z_0$ and $U_0=z_0+D/z_0$ belongs to $\mathbb F_p$. If $z_0^2=D$, we would have $z_0^p=z_0$, making $D$ a square in $\mathbb F_p$. Since $d$ is odd, $D=\lambda^d$ is a nonsquare. We therefore have $z_0^2\ne D$. Take $T=1/(U-U_0)$. Choose $1\le b\le2\nu$ with $db\equiv1\pmod{2\nu}$, and let $r$ run through the powers of $p$ congruent to $1$ modulo $2\nu$. Put \[ e=\lambda^{(r-db)/(2\nu)},\qquad w=\frac{x^r}{et^b},\qquad y=\frac{zw+D/z}{w+1}. \] Since $d\equiv j\pmod\nu$, we have $w=e^{-1}u^{(r-jb)/\nu}\rho(u)^{-b}$. Its product with its image under $u\mapsto\lambda/u$ is $\lambda^{(r-db)/\nu}/e^2=1$. Thus the involution fixes $y$, and $y\in k(W)$. Over $\bar k(t)$ the elements $y$ and $x^r$ are related by a fractional linear transformation. They generate $\bar k(x)$ over $\bar k(t)$, since $\bar k(x)/\bar k(t)$ is separable and $x$ is purely inseparable over $\bar k(t,x^r)$. Thus $y$ has degree $d$ over $\bar k(t)$ and at least $d$ over $\bar k(T)$. As $[k(W):k(T)]=d$, we have $k(T,y)=k(W)$, and the same equality holds over $\bar k$. The primitive polynomial $F(T,Y)$ associated with the minimal polynomial of $y$ is therefore absolutely irreducible. It is separable in $Y$ since both $f(x)$ and $U=t^\nu+D/t^\nu$ have nonzero derivative. Moreover, \[ h(F)=\deg_T F=[k(W):k(y)]=\deg_W y . \] After substituting $t=f(x)$, the coefficients of the fractional linear transformation from $x^r$ to $y$ and of its inverse have degrees in $x$ bounded in terms of $d$. Hence $\deg_x y=r+O_d(1)$. Since $\deg_x W=2\nu$, we obtain $h(F)=(r+O_d(1))/(2\nu)$. Choose $x_0$ with $x_0^\nu=u_0$ and set $t_0=f(x_0)$. The derivative of $U=t^\nu+D/t^\nu$ at $t_0$ is nonzero, since $p\nmid\nu$ and $z_0^2\ne D$. Consequently $v(t-t_0)=v(U-U_0)=1$ at $t=t_0$. The simple root $u_0$ of the critical equation gives a simple zero of $f'$ at $x_0$. Thus $f(X)-t_0=(X-x_0)^2g(X)$ with $g(x_0)\ne0$. As $p$ is odd, $g$ has a square root in $\bar k[[X-x_0]]$. The local parameter $(X-x_0)\sqrt{g(X)}$ takes the values $\pm\sqrt{t-t_0}$ at the two roots reducing to $x_0$. Their difference has valuation $1/2$. At $t=t_0$ the fractional linear transformation from $x^r$ to $y$ has regular coefficients and its determinant is a unit, since $t_0\ne0$ and $z_0^2\ne D$. At the common residue of the two $r$th powers, either this transformation or its reciprocal has a unit denominator. In that coordinate the difference of the two images has valuation $r/2$. Replacing $F$ by its reciprocal polynomial if necessary preserves its height and gives $s(F)\ge r/2\ge\nu h(F)-O_d(1)$. For example, $27\le E_{7^a,\mathrm{abs}}(29)\le28$ for every $a\ge1$. Whether $E_{7^a,\mathrm{abs}}(29)=28$ remains open. \subsection{A construction for even degrees} Suppose that $d=2\ell+b$, where $\ell,b\ge1$, $\gcd(\ell,b)=1$ and $p\nmid 2b\ell$. We shall construct absolutely irreducible polynomials $F$ of degree $d$ and arbitrarily large logarithmic height $h$ for which \begin{equation}\label{eq:uniform-even} s(F)\ge 2\ell h-O_{q,d}(\sqrt h). \end{equation} If $p\mid d$, then $p\nmid d-1$, and Theorem~\ref{thm:abs} gives the same estimate with $d-1$ in place of $2\ell$. Since $d-1\ge2\ell$, we may assume that $p\nmid d$. The construction uses two properties of a rational function $f$. Near a simple critical point of $f$, the equation $f(x)=t$ has two roots close to each other when $t$ is close to the critical value. At $T=2$, Eisenstein's criterion will give an irreducible equation of degree $d$. As in Lemma~\ref{lem:quadraticapprox}, write $T=\sigma+\sigma^{-1}$. Put \[ z=\frac{1+\sigma}{1-\sigma},\qquad a=\frac{bd}{(b-\ell)^2-z^2}. \] We have $z^2=(T+2)/(T-2)$, so $a\in k(T)$ has a simple zero at $T=2$. The substitution $\sigma\mapsto\sigma^{-1}$ sends $z$ to $-z$. Set \[ u=x^\ell,\qquad A=u^2-2au+a,\qquad B=au^2-2au+1,\qquad t=f(x)=x^b A/B. \] The two quadratics have the same nonzero discriminant $4a(a-1)$. They have no common zero. Indeed, $A-B=(1-a)(u^2-1)$, whereas $A(1)$ and $A(-1)$ are both nonzero. Since $B(0)=1$, it follows that $\deg f=d$. The identity $B(u)=u^2 A(1/u)$ explains our choice of $A$ and $B$. Write \[ W=u+u^{-1},\qquad \xi=t^\ell=u^b(A/B)^\ell,\qquad V=\xi+\xi^{-1}. \] Under $u\mapsto u^{-1}$, the function $\xi$ goes to $\xi^{-1}$, so $V=\Phi(W)$ for some rational function $\Phi$ over $k(T)$. The maps $x\mapsto W$ and $t\mapsto V$ both have degree $2\ell$. Taking degrees in $V=\Phi(W)$ therefore gives $\deg\Phi=d$. At a point where $f$ is finite and nonzero, the equation $f'(x)=0$ reduces to $bAB+\ell u(A'B-AB')=0$, with derivatives of $A$ and $B$ taken with respect to $u$. Divide by $u^2$ and complete the square to obtain \[ [bW-(b-\ell)a-b-\ell]^2=(a-1)^2\big((b-\ell)^2-bd/a\big). \] The two solutions are $W_\pm=((b-\ell)a+b+\ell\pm(a-1)z)/b$. Their images under $\Phi$ are $J(z)=\Phi(W_+)\in k(z)=k(\sigma)$ and $J(-z)=\Phi(W_-)$, since $\sigma\mapsto\sigma^{-1}$ interchanges $W_+$ and $W_-$. To apply Lemma~\ref{lem:quadraticapprox}, we must show that $J(z)\ne J(-z)$. At $z=\infty$ we have $a=-bd/z^2+O(z^{-4})$. Of the two solutions of $u+u^{-1}=W_+$, the one with a pole at $z=\infty$ has the expansion $u=-z/b+(b+\ell)/b+O(z^{-1})$. Hence \[ A/u^2=1+O(z^{-3}),\qquad B=-\frac{2\ell}{b}\bigl(1-d/z+O(z^{-2})\bigr). \] Using $\xi=u^b(A/B)^\ell$ in $J=\xi+\xi^{-1}$ gives \[ J(z)=Cz^d\bigl(1-bd/z+O(z^{-2})\bigr),\qquad C=(-1/b)^d(-2\ell/b)^{-\ell}\ne0. \] If $d$ is odd, the coefficient of $z^d$ changes sign under $z\mapsto-z$. If $d$ is even, the nonzero coefficient of $z^{d-1}$ changes sign. Thus $J(z)\ne J(-z)$. The only possible branch values of the maps $x\mapsto W$ and $t\mapsto V$ are $2,-2,\infty$. As elements of $k(z)$, $W_+$ and $J$ are nonconstant and hence distinct from these values. Since $W_+$ is a simple root of the equation for $W$, each $x_0$ lying over $W_+$ is a simple critical point of $f$. Let $r$ run through the sufficiently large powers of $p$ congruent to $1$ modulo $\ell$. By Lemma~\ref{lem:quadraticapprox}, there exists $V_r\in k(T)$ with a simple pole at $T=2$ such that $h(V_r)\le r/2+O(1)$ and $v(V_r-J)=r+O(1)$ at $\sigma=0$. We first work over $K_2=\bar k((T-2))$. Hensel's lemma, applied to $V_r^{-1}\eta^2-\eta+V_r^{-1}$ at $\eta=0$, gives a solution of $\eta+\eta^{-1}=V_r$ with a simple zero at $T=2$. By Eisenstein's criterion, a root $t$ of $t^\ell=\eta$ is a uniformizer of $K_2(t)$. The monic polynomial $x^b A-tB$ of degree $d$ is Eisenstein at $t$. Indeed, $a$ is divisible by $t$, and the constant term is $-t$. This polynomial is separable, since it is irreducible and $p\nmid d$. Choose $1\le j\le\ell$ with $bj\equiv1\pmod\ell$, and put \[ w=x^r/t^j=u^{(r-bj)/\ell}(A/B)^{-j},\qquad y=(\xi w+\xi^{-1})/(w+1). \] Replacing $u$ by $u^{-1}$ replaces both $w$ and $\xi$ by their inverses and leaves $y$ fixed. Thus $y$ is a rational function of $W$ over $k(T)$. For the value of $t$ just chosen, $y$ is an invertible fractional linear transformation of $x^r$, since $V_r\ne\pm2,\infty$. As $K_2(t,x)/K_2(t)$ is separable, $x^r$ generates this extension, and so does $y$. It follows that $y$ has degree at least $d$ over $\bar k(T)$. On the other hand, $y\in k(T,W)$ and $\Phi(W)=V_r$, so its degree over $k(T)$ is at most $d$. Consequently its primitive polynomial $F(T,Y)$ over $k[T]$ is separable and absolutely irreducible of degree $d$ in $Y$. To estimate $h(F)$, first regard $W$ as an indeterminate and write $y=R(W)$. The fractional linear transformation from $x^r$ to $y$ and its inverse have coefficients of bounded degree in $x$, since $t=f(x)$ is fixed. Hence $\deg_x y=r+O(1)$, and $\deg_x W=2\ell$ gives $\deg_W R=(r+O(1))/(2\ell)$. The degree of $R$ in $T$ is bounded independently of $r$. Indeed, extending the coefficient field $k(W)$ to $k(x)$ preserves the degree in $T$, and in the expression for $y$ in terms of $x$ only $t=f(x)$ depends on $T$. Clear denominators in $Y=R(W)$ and $\Phi(W)=V_r$. The resulting polynomials have degrees $O(1)$ and at most $h(V_r)+O(1)$ in $T$, respectively. Their resultant with respect to $W$ is divisible by $F$. For polynomials $P,Q$, the determinant of the Sylvester matrix gives $\deg_T\Res_W(P,Q)\le(\deg_W P)(\deg_T Q)+(\deg_W Q)(\deg_T P)$. Here the degrees in $W$ are $(r+O(1))/(2\ell)$ and $d$, whence \[ h(F)\le\frac{r+O(1)}{2\ell}(h(V_r)+O(1))+O(1) \le\frac{r^2}{4\ell}+O(r). \] We now work at $\sigma=0$. Fix a critical point $x_0$ above $W_+$ and put $t_0=f(x_0)$. Since $t\mapsto V=t^\ell+t^{-\ell}$ is unramified at $t_0$, the equation $t_r^\ell+t_r^{-\ell}=V_r$ has a solution with $v(t_r-t_0)=r+O(1)$ for all sufficiently large $r$. In the local coordinate $Z=x-x_0$ we have $f(x)-t_0=cZ^2(1+O(Z))$, with $c\ne0$. Since $p$ is odd, $1+O(Z)$ has an analytic square root near zero. Thus $f(x)=t_r$ has two roots $x_1,x_2$ tending to $x_0$ with $v(x_1-x_2)=r/2+O(1)$. For finite $a,b$, write $[a,b]=|a-b|/(\max(1,|a|)\max(1,|b|))$ for their chordal distance. Since $x_0$ is finite and independent of $r$, we have $[x_1,x_2]=q^{-r/2+O(1)}$ and hence $[x_1^r,x_2^r]=[x_1,x_2]^r=q^{-r^2/2+O(r)}$. At $t=t_r$ the fractional linear transformation from $x^r$ to $y$ has matrix $M(t_r)$, where $M(t)=\left(\begin{smallmatrix}t^\ell&t^{j-\ell}\\1&t^j\end{smallmatrix}\right)$. Since $t_0\ne0,\infty$ and $t_0^{2\ell}\ne1$, the limit $M(t_0)$ is invertible. Thus $M(t_r)$ and its inverse have bounded entries. The chordal distance between the two images is therefore $q^{-r^2/2+O(r)}$. The two images are roots of $F$. Indeed, the equation $\Phi(W)=V_r$ is irreducible over $k(T)$ because $k(T,W)$ contains $k(T,y)$ of degree $d$. Its roots therefore give conjugate values of $y$. Replacing $F$ by its reciprocal if necessary, we may assume that both roots have absolute value at most one, whence $s(F)\ge r^2/2-O(r)$. It follows from~\eqref{eq:bound} that $h(F)\gg r^2$. Together with the upper bound for $h(F)$ this proves~\eqref{eq:uniform-even}. For even $d$ it remains to consider odd $p$ dividing $d-1$, since Theorem~\ref{thm:abs} covers the other cases. For $4\mid d$ take $b=2$. For $d\equiv2\pmod4$ take $b=4$ if $p\ge5$ and $b=8$ if $p=3$. In the last case $d\ge10$. In each case $\ell=(d-b)/2$ is odd and $2\ell\equiv1-b\not\equiv0\pmod p$. Substitution in~\eqref{eq:uniform-even} gives the bounds in Theorem~\ref{thm:universal} for even $d$. \section{The constant field}\label{sec:constants} Let $x$ be a root of a separable irreducible polynomial $F$ of degree $d$. The full constant field of $K(x)$ is $K(x)\cap\bar k=\mathbb F_{q^c}$ for some divisor $c$ of $d$. We denote this field by $k_c$. The polynomial $F$ is absolutely irreducible if and only if $c=1$. \begin{theorem}\label{thm:constants} If $c>1$, then \[ s(F)\le \frac{2d}{c^2}\,h(F). \] For each divisor $c>1$ of $d$ and each integer $N\ge1$, there is such a polynomial of logarithmic height $cN$ for which equality holds. Hence the separation exponent for irreducible polynomials of degree $d$ that factor over $\bar k(T)$ is $2d/\ell^2$, where $\ell$ is the least prime divisor of $d$. \end{theorem} \begin{proof} Write $r=d/c$. The minimal polynomial of $x$ over $k_c(T)$ has degree $r$ and remains irreducible over $\bar k(T)$. To see this, let $\alpha\in\bar k$. The minimal polynomial of $\alpha$ over $K(x)$ has all its roots in $\bar k$, so its coefficients belong to $K(x)\cap\bar k=k_c$. Thus $[K(x,\alpha):K(x)]=[k_c(\alpha):k_c]$. By the tower law, $x$ has degree $r$ over $k_c(\alpha)(T)$, and $\bar k(T)$ is the union of these fields. Up to a constant factor, $F$ is the product of $c$ absolutely irreducible polynomials of degree $r$ in $k_c[T][X]$. Indeed, the Frobenius automorphism of $k_c/k$ permutes the monic irreducible factors transitively, since $F$ is irreducible over $K$. Their primitive representatives have the same logarithmic height, which by multiplicativity of the Gauss norm is $h(F)/c$. For distinct roots $u_1,u_2$ of the same factor, inequality~\eqref{eq:bound} gives $v(u_1-u_2)\le(r-1)h(F)/c$. If $u_1,u_2$ are roots of two different factors, \eqref{eq:res} gives $v(u_1-u_2)\le 2rh(F)/c$. This proves the upper bound. To obtain equality, we adapt a construction of Bugeaud and Mignotte~\cite[(3.1)]{BugeaudMignotte2004} for integer polynomials. Choose $\gamma\in k_c$ of degree $c$ over $k$, put $a=T^N$, and set \[ G_\eta(X)=X^r+(1+\eta X^{r-1})(1-aX),\qquad F_N=\prod_{i=0}^{c-1}G_{\gamma^{q^i}}. \] For each conjugate $\eta$ of $\gamma$, write $G_\eta=A-T^NB$, where $B=X(1+\eta X^{r-1})$ and $A=X^r+1+\eta X^{r-1}$. The polynomial $B$ has a simple zero at $0$, whereas $A(0)\ne0$. This also holds for $r=1$, since $\eta\notin k$. At each nonzero root of $B$ we have $A=X^r\ne0$. Hence $A$ and $B$ are coprime, and Eisenstein's criterion at $X$ shows that $AZ^N-B$ is irreducible in $\bar k[X,Z]$. Replacing $Z$ by $1/T$ and clearing denominators shows that $G_\eta$ is absolutely irreducible. For distinct conjugates $\eta,\theta$, the difference $G_\theta-G_\eta=(\theta-\eta)X^{r-1}(1-aX)$ is nonzero, while $G_\eta(1/a)=G_\theta(1/a)=a^{-r}$. Thus no two factors are scalar multiples of one another. Frobenius permutes the $c$ factors transitively, so $F_N$ is irreducible over $K$. Each factor has logarithmic height $N$, whence $h(F_N)=cN$. For $\eta\ne\theta$, the formula for $G_\theta-G_\eta$, together with $a^rG_\eta(1/a)=1$ and, for $r>1$, $G_\eta(0)=1$, gives $\Res_X(G_\eta,G_\theta)=(\theta-\eta)^r$. To find the distance between the closest roots, reduce $a^{-1}G_\eta$ modulo $T^{-1}$. Its reduction $-X(1+\eta X^{r-1})$ has degree $r$ and a simple root at $0$. In particular, the leading coefficient of $G_\eta$ has absolute value $|a|$, and all its roots have absolute value at most one. By Hensel's lemma it has a unique root $x_\eta$ with $|x_\eta|<1$. This root is simple, so $G_\eta$ is separable. The $c$ absolutely irreducible factors of $F_N$ therefore show, by the first part of the proof, that the full constant field is $k_c$. For $\eta\ne\theta$, the only common root of the two reductions is $0$. It follows that $|u-v|=1$ for every pair of roots $u$ of $G_\eta$ and $v$ of $G_\theta$, except for $(x_\eta,x_\theta)$. Taking absolute values in the product formula for the resultant gives $1=|a|^{2r}|x_\eta-x_\theta|$. Thus $s(F_N)=2rN$. \end{proof} We can now compare the two exponents $E_{q,\mathrm{irr}}(d)$ and $E_{q,\mathrm{abs}}(d)$. For $c>1$ the preceding theorem gives an exponent at most $d/2$. By Theorem~\ref{thm:universal}, $E_{q,\mathrm{abs}}(d)\ge d-8\ge d/2$ for $d\ge16$, and $E_{q,\mathrm{abs}}(d)\ge d-2\ge d/2$ for $d\ge4$ odd or divisible by four. For the remaining degrees $2,3,6,10,14$, the integer $d-1$ is either $1$ or a prime power, so Theorem~\ref{thm:abs} gives $E_{q,\mathrm{abs}}(d)=d-1$. Hence $E_{q,\mathrm{irr}}(d)=E_{q,\mathrm{abs}}(d)$ for every $q$ and $d$. \section{When the loss is bounded}\label{sec:loss} \begin{theorem}\label{thm:loss} Let $k=\mathbb F_q$ have characteristic $p$, and let $n\ge1$. There are reducible polynomials of degree $n+1$ with unbounded height and bounded $\Delta$ if and only if $n$ satisfies at least one of the following conditions. In the list, $P$ denotes a positive power of $p$. \begin{enumerate} \item $p\nmid n$. \item $n=Pm$ for some $m\mid P-1$. \item $p=2$ and $n\equiv2\pmod4$. \item $n=P(P^2-1)$ or $n=P(P^2-1)/\gcd(2,P-1)$. \item $p=3$ and $n=60$. \end{enumerate} In each case the polynomials can be taken to be products of a linear factor and a fixed absolutely irreducible factor of degree $n$. For every other $n$ there is a constant $c_n>0$ such that $\Delta(F)\ge c_n\sqrt{h(F)}$ for every reducible $F$ of degree $n+1$ and sufficiently large logarithmic height. \end{theorem} Theorem~\ref{thm:red} covers $p\nmid n$, so we may assume that $p\mid n$. Suppose that $W$ is reducible of degree $n+1$ and that $s(W)>(n-1)h(W)$. By~\eqref{eq:lossres}, a closest pair consists of a root $\alpha$ of an irreducible factor $F$ of degree $n$ and the root $u=a/b$ of the remaining linear factor. Put $B=h(F)$ and $r=b^nF(a/b)$. Then \begin{equation}\label{eq:lossclass} h(W)=B+h(u),\qquad (n-1)B+\deg r\le\Delta(W). \end{equation} Each conjugate of $\alpha$ over $K_\infty$ is just as close to $u$. Two distinct conjugates would therefore violate~\eqref{eq:bound} for $F$, since $s(W)>(n-1)B$. Thus $\alpha\in K_\infty$. The monic irreducible factor of $F$ over $\bar k(T)$ having root $\alpha$ is fixed by the automorphism that raises the constants to their $q$th powers and fixes $T$, since it fixes the Laurent series $\alpha$. It follows that $F$ is absolutely irreducible. Suppose that $n\ge4$ and that the cross-ratio of any four distinct roots of $F$ lies in $\bar k$. A fractional linear transformation $M$ sending three roots to $0,1,\infty$ then takes every root to a point of $\mathbb P^1(\bar k)$. For every automorphism $\tau$ of the splitting field over $\bar k(T)$, the transformation $M\tau(M)^{-1}$ permutes these $n$ points. Its coefficients can be taken in $\bar k$, since the images of three points determine it. These transformations form a finite group, because they induce the permutations $M(x)\mapsto M(\tau x)$ on these points. The action is transitive by absolute irreducibility. Thus $n$ is the size of an orbit of a finite subgroup of $\mathrm{PGL}_2(\bar k)$. \subsection{The possible orbit sizes} We determine the orbit sizes divisible by $p$, using the classical count of fixed points of fractional linear transformations (see Faber~\cite[Sections~5--6]{Faber}). Let $G$ be a finite subgroup of $\mathrm{PGL}_2(\bar k)$ of order $g$ divisible by $p$. Let $P$ be the order of a Sylow $p$-subgroup and $a$ the number of such subgroups. A nonidentity projectivity has two fixed points, except when it has order $p$, in which case it has one. A Sylow $p$-subgroup fixes the unique fixed point of any central element of order $p$. Sending this point to infinity identifies the subgroup with a group of translations. The translations in the full stabilizer form its unique Sylow $p$-subgroup. Thus the $a$ fixed points of the Sylow subgroups are distinct and form one orbit. A Sylow subgroup acts freely on the other $a-1$ points, so $a=1+tP$ for some $t\ge0$. Its fixed point has stabilizer of order $Pm$, where $m\mid P-1$, since each nonidentity multiplier permutes the $P-1$ nonzero translation parameters without fixing any of them. In particular $g=aPm$. An orbit of size $b$ accounts for $g-b$ pairs $(x,\gamma)$ with $1\ne\gamma\in G$ and $\gamma x=x$. There are $a(P-1)$ elements of order $p$. After subtracting the $g-a$ pairs from the orbit of the Sylow fixed points, the number of pairs left is \[ 2g-2-a(P-1)-(g-a)=g-2-a(P-2). \] Each further orbit of size less than $g$ accounts for at least $g/2$ pairs. There is consequently at most one such orbit, and its size, if it exists, is $b=2+a(P-2)$. For $t=0$ the only possible sizes are $1,P,g$. For $t>0$ the remaining count is at least $2a-2>0$. Since $b=P(1+t(P-2))$ divides $g$, we have \[ 1+t(P-2)\mid (1+tP)m. \] The two factors $1+t(P-2)$ and $1+tP$ have greatest common divisor dividing $2$. Hence $1+t(P-2)\le2m\le2(P-1)$. For $P\ge3$ this leaves $t=1,2$ and the additional possibility $P=t=3$. For $t=2$ the greatest common divisor is one, whereas $2P-3>m$. For $t=1$ we get $m=P-1$ or, for odd $P$, $m=(P-1)/2$. The possible sizes divisible by $p$ are then $g$ and $P(P-1)$. If $P=t=3$, then $m=2$, $g=60$ and $b=12=|\mathrm{PSL}_2(\mathbb F_3)|$. Finally, $P=2$ gives $m=1$, $g=2(1+2t)$ and $b=2$. This gives all the possibilities in Theorem~\ref{thm:loss}. Each size on the list occurs as the order of a group stable under the $q$th-power map on coefficients. For $p\nmid n$, take the rotations $z\mapsto uz$ with $u^n=1$. For $n=Pm$ take the maps $z\mapsto uz+v$ with $u^m=1$ and $v\in\mathbb F_P$. For $p=2$ and $n\equiv2\pmod4$ take the maps $z\mapsto uz$ with $u^{n/2}=1$ together with $z\mapsto 1/z$. The two orders $P(P^2-1)$ and $P(P^2-1)/\gcd(2,P-1)$ are those of $\mathrm{PGL}_2(\mathbb F_P)$ and $\mathrm{PSL}_2(\mathbb F_P)$. For $p=3$ and $n=60$ there is a small explicit example. Take $\eta^2=\eta+1$ in $\mathbb F_9$. Label $\infty$ by the pair $45$ and the nine finite points by the pairs in the table. \[ \begin{array}{c|ccc} &z&z+1&z+2\\ \hline z=0&12&23&13\\ z=\eta&25&35&15\\ z=-\eta&14&24&34 \end{array} \] On these labels $z\mapsto z+1$, $z\mapsto\eta^2/z$ and $z\mapsto z^3$ act as $(123)$, $(14)(25)$ and $(45)$, respectively. The first two generate $A_5$. Indeed, their product has order five, so the subgroup they generate has order divisible by $30$, and $A_5$ has no subgroup of index two since it is generated by $3$-cycles. The cubing map therefore normalizes the group of sixty projectivities. \subsection{Examples with a fixed factor} For each $n$ on the list with $p\mid n$, take one of the groups $G$ of order $n$ constructed above. The coefficients of $\prod_{\gamma\in G}(Y-\gamma(z))$ lie in $k(z)$ and generate a field $k(J)$ by L\"uroth's theorem. The polynomial gives $\deg_z J\le n$, while the $n$ automorphisms over $\bar k(J)$ give the reverse inequality. The fibers of $J$ are the $G$-orbits. The fixed-point count shows that at most two of them have fewer than $n$ points. Thus some point of $\mathbb P^1(k)$ has a fiber $\Omega$ of size $n$, even for $q=2$. Changing the coordinate on the target over $k$, take $T$ with pole at this fiber. It has degree $n$ in $z$ and a simple pole at each point of $\Omega$. Choose a power $Q$ of $q$ for which the points of $\Omega$ and the matrices of $G$ are defined over $\mathbb F_Q$. The transformation \[ M_z(X)=\frac{(X-z^Q)(z^{Q^2}-z)}{(X-z)(z^{Q^2}-z^Q)} \] sends $z,z^Q,z^{Q^2}$ to $\infty,0,1$. Hence $M_{\gamma z}(\gamma X)=M_z(X)$ for $\gamma\in G$. Fix a finite $c\in\Omega$ and put $\alpha=M_z(c)$. Its conjugates over $\bar k(T)$ are the $n$ distinct values $M_z(c')$, $c'\in\Omega$. Since the $q$th-power map on the constants permutes $\Omega$, their monic polynomial lies in $k(T)[X]$. Let $F\in k[T][X]$ be its primitive representative. For $R=Q^j$, $j\ge3$, the function $u=M_z(z^R)$ is invariant under $G$ and belongs to $k(z)$, so $u\in k(T)$. Its expression gives $nh(u)=\deg_z u\le R+Q^2$. At $z=c$ the difference \[ u-\alpha= \frac{(z^{Q^2}-z)(z^Q-z)(z^R-c)} {(z^{Q^2}-z^Q)(z^R-z)(c-z)} \] has order $R-Q$. Since $T$ has a simple pole there, $v(u-\alpha)=R-Q$. The limit $\alpha$ therefore belongs to $K_\infty$ by completeness. The heights $h(u)$ are unbounded because $\alpha$ has degree $n>1$. Multiplying $F$ by the primitive linear polynomial with root $u$ gives polynomials of unbounded height with $\Delta\le nh(F)+Q^2+Q$. For example, $n=6$ occurs on the list in characteristics two and three. In both cases the Laurent series $\alpha$ has degree six and irrationality exponent six. Indeed, $v(\alpha-u)\ge6h(u)-O(1)$, and Liouville's inequality gives the reverse bound for the exponent. Bugeaud singled out these two cases in his discussion of a question of Mahler \cite[Problem~1.1 and p.~193]{Bugeaud2022}. \subsection{A lower bound outside the list} If $n$ lies outside the list, then $n\ge4$. The argument at the beginning of this section shows that whenever $s(W)>(n-1)h(W)$, four roots of $F$ in~\eqref{eq:lossclass} have a nonconstant cross-ratio. For a fixed algebraic Laurent series, Voloch related rational approximation to the cross-ratios of its conjugates \cite{Voloch,LdM}. Here we need an estimate uniform in the factor $F$. \begin{lemma}\label{lem:lossfour} Let $F\in k[T][X]$ be separable of degree $n\ge4$, and suppose that four of its roots have a cross-ratio outside $\bar k$. Put $B=h(F)$. For coprime $a,b\in k[T]$ with $b\ne0$ and $r=b^nF(a/b)\ne0$, we have \[ h(a/b)\ll_n (B+1)(B+1+\deg r). \] \end{lemma} \begin{proof} Let $L$ be the field generated by the four roots $x_1,\ldots,x_4$ over $\bar k(T)$ and put $D=[L:\bar k(T)]\le n^4$. For $x\in L$ write $h_L(x)$ for the degree of its pole divisor. Each $x_i$ satisfies $h_L(x_i)\le DB$. Indeed, if $f_n$ is the leading coefficient of $F$, the homogeneous coordinates $[f_nx_i:f_n]$ are integral at every finite place. At a place of ramification index $e$ above infinity, each coordinate has pole order at most $eB$. For $f_nx_i$ this follows from its monic equation, whose coefficient in degree $n-j$ has degree at most $jB$ in $T$. The genus $g$ of $L$ satisfies $g\ll_n B+1$. To see this, choose a linear combination $z$ of the $x_i$ over $\bar k$ that separates the $D$ embeddings of $L$ over $\bar k(T)$. Then $L=\bar k(T,z)$ and $h_L(z)\le4DB$. For every $l\ge0$ the $D(l+1)$ functions $T^iz^j$ with $0\le i\le l$ and $0\le j