\documentclass[11pt]{amsart} \usepackage{amsmath,amssymb,indentfirst} \usepackage[letterpaper,margin=1.05in]{geometry} \setlength{\emergencystretch}{2em} \usepackage[colorlinks,citecolor=blue,urlcolor=blue,linkcolor=blue]{hyperref} \newtheorem{theorem}{Theorem} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \title{Density-one uniqueness for Markoff--Hurwitz solutions} \author{Qiyue Tang} \makeatletter \g@addto@macro\@setauthors{\par\vspace{0.55em}{\normalfont\centering 15 November 2026\par}} \makeatother \subjclass[2020]{Primary 11D45, Secondary 11D72} \keywords{Markoff--Hurwitz equation, maximum coordinate, density} \begin{document} \begin{abstract} Fix an integer $n\ge4$. We prove that, as $R\to\infty$, the proportion of sorted positive integer solutions of $\sum_{i=1}^n x_i^2=n\prod_{i=1}^n x_i$ with largest coordinate at most $R$ whose largest coordinate is shared by another sorted solution tends to zero. \end{abstract} \maketitle \section{Introduction} Markoff's work on indefinite binary quadratic forms in 1879 and 1880 led to the equation \[ x^2+y^2+z^2=3xyz. \] The sorted positive integer solutions are generated from $(1,1,1)$ by permutations and Vieta moves, such as $z\mapsto 3xy-z$. Markoff related these triples to forms whose minimum, divided by the square root of the discriminant, exceeds $1/3$ \cite{Markoff,Zagier}. In 1913 Frobenius asked whether the largest entry determines the sorted triple \cite{Frobenius,Baragar}. The generation of the triples gives a tree, and Frobenius's question asks whether the largest entries label its vertices without repetitions. Zagier counted the vertices of this tree by the largest entry. He showed that the number of sorted triples with largest entry at most $R$ is asymptotic to $C(\log R)^2$ for an explicit positive constant $C$ \cite{Zagier}. This count assigns one term to each triple, so a largest entry shared by several triples contributes several times. Baragar proved uniqueness for certain arithmetic classes of largest entries. For example, his result applies when either $3z-2$ or $3z+2$ is prime, where $z$ is the largest entry \cite{Baragar}. Chen's preprint proves density-one uniqueness in the classical equation, using the action on primitive torus curves and a $3$-adic zero-one law \cite{Chen}. Hurwitz studied positive integral tuples whose sum of squares is a multiple of their product \cite{Hurwitz}. The specialization in $n$ variables with coefficient $n$ has the all-one solution. The Vieta move again replaces one coordinate by the other root of a quadratic, and descent from the largest entry gives the tree used below. Baragar established the growth exponent for the higher-dimensional equation \cite{BaragarExponent}. Gamburd, Magee and Ronan subsequently obtained an asymptotic formula with positive leading coefficient and growth scale $(\log R)^{\beta_n}$ by comparing the Vieta action in logarithmic coordinates with a linear renewal problem \cite{GMR}. Their counting theorem provides the scale for asking how often a largest entry is repeated. Fix an integer $n\ge4$. For $R\ge1$, let $T_n(R)$ count the integer tuples satisfying \[ 1\le x_1\le\cdots\le x_n\le R, \qquad \sum_{i=1}^n x_i^2=n\prod_{i=1}^n x_i. \] Let $B_n(R)$ count the tuples in this set whose last coordinate equals the last coordinate of a different tuple in the set. \begin{theorem}\label{thm:main} For every fixed integer $n\ge4$, \[ \lim_{R\to\infty}\frac{B_n(R)}{T_n(R)}=0. \] \end{theorem} If $D_n(R)$ counts the distinct largest entries, then $T_n(R)-B_n(R)\le D_n(R)\le T_n(R)$. At most $B_n(R)$ of these entries are repeated. Thus the proportion of distinct largest entries that determine a unique sorted solution also tends to one. The proof assigns a primitive integer vector to each word of Vieta moves. On most pairs with equal largest entries, these vectors are related by a matrix from a fixed finite set of rational matrices. For a fixed matrix, a positive mass of pairs in the limiting angular distributions forces a local replacement of the initial word. This identifies a common suffix of moves from two fixed seeds. Counting those suffixes on the finite residue orbits reduces the collision estimate to $3$-adic separation of the seeds. \section{The Vieta tree and integer addresses} Write $d=n-1$. All vectors in $\mathbb R^d$ are columns unless explicitly written as rows. Superscript $t$ denotes transpose, $I$ denotes the identity matrix, and $\mathbf1$ is the column of ones. At coordinate $i$ the Vieta move replaces $x_i$ by $n\prod_{j\ne i}x_j-x_i$ and fixes the other coordinates. The new value is a positive integer because the product of the two roots of the quadratic for $x_i$ is the positive sum of the other squares. For a sorted positive solution write $y=x_{n-1}$, $z=x_n$, and $P=\prod_{i=1}^{n-2}x_i$. The quadratic $f$ for $z$ satisfies \[ f(y)=\sum_{i=1}^{n-2}x_i^2+(2-nP)y^2\le n(1-P)y^2\le0. \] Equality forces all entries preceding $z$ to equal one, and then $z$ is $1$ or $n-1$. Otherwise the replacement of $z$ is strictly less than $y$. Also, when $z$ is the unique maximum, the quadratic for any other entry takes the value $\sum_{j\ne i,n}x_j^2+(2-n\prod_{j\ne i,n}x_j)z^2$ at $z$. It is strictly negative because all entries preceding $z$ are strictly smaller. Thus every move at a nonmaximum entry increases the maximum. We keep the maximum in slot $n$ and start at $a_{\varnothing}=(1,\ldots,1,n-1)$. For $1\le i\le d$, let $\mu_i$ swap slots $i,n$ and then apply the Vieta move in slot $n$. For the finite word $w=i_1\cdots i_k$ put $a_w=\mu_{i_k}\cdots\mu_{i_1}a_{\varnothing}$, and let $M(w)$ be its maximum. Descent gives every nonroot sorted solution by sorting some $a_w$. Moreover, $a_w$ determines $w$: the largest entry among the first $d$ slots was the previous maximum and identifies $i_k$, after which the inverse move recovers $a_{i_1\cdots i_{k-1}}$. With $e_i$ the $i$th standard column, define \[ R_i=I+e_i(\mathbf1-e_i)^t,\quad A_i=R_i^t,\quad V=\mathbf1\mathbf1^t-(d-1)I,\quad P_w=A_{i_1}\cdots A_{i_k}. \] Empty products are identities. Direct multiplication gives $VA_i=R_iV$, $V\mathbf1=\mathbf1$, and $\mathbf1^tV=\mathbf1^t$. Set $R_w=R_{i_1}\cdots R_{i_k}$, $v_w=R_w\mathbf1=VP_w\mathbf1$, and $H(v)=\mathbf1^tv$. Each $v_w$ is positive and primitive because each $R_i$ is unimodular. For nonempty $w$, the coordinate indexed by its first letter exceeds the sum of the others. Inverting that letter and iterating proves that $v_w$ determines $w$. Assign degree zero to the parent entries of $a_{\varnothing}$ and degree one to the maximum. Every new maximum has assigned degree one plus the sum of the degrees of the retained entries. At each stage the maximum degree is one plus the sum of the parent degrees. The vector $q_w=P_w^t\mathbf1$ has entries $(n-2)r_i+1$, where $r_i$ is the degree in parent slot $i$: when the maximum becomes parent $i$, the corresponding entry of $q_w$ is replaced by the sum of all its entries. In particular, if $D(w)$ is the maximum degree, then $H(v_w)=(n-2)D(w)+1$. In any outward move the removed entry is bounded by the product of the retained entries. Hence the new entry lies between $n-1$ and $n$ times that product, with strict upper inequality. Induction gives \begin{equation}\label{eq:height} (n-1)^{D(w)}\le M(w)1$ denote the exponent of \cite[Theorems 2 and 13]{GMR}. For a positive column $q$ and $L>0$ let $N_q(L)$ count finite words $t$ with $\mathbf1^t P_t^tq\le L$. Write $r=\mathbf1^tq$ and $s=r-\max_i q_i$. The accelerated scalar renewal theorem of \cite[Theorem 13]{GMR}, followed by summation over the initial repetitions of the largest slot, gives \begin{equation}\label{eq:flow} N_q(L)\sim c(q)L^\beta,\quad N_q(L)\le C L^\beta c(q),\qquad c(q)\asymp (1+r/s)r^{-\beta},\qquad c(q)=\sum_i c(R_iq). \end{equation} The asymptotic is as $L\to\infty$. The relation $\asymp$ denotes upper and lower bounds by positive constant multiples. Constants may depend on the fixed dimension. To see the deacceleration, after sorting the slots the repeated largest-slot moves yield heights $r+As$ for integers $A\ge0$. Accelerated counts from these seeds are bounded by $C(L/(r+As))^\beta$, and their leading coefficients are bounded above and below by positive multiples of $(r+As)^{-\beta}$ by compactness of the normalized seed simplex and positivity of the GMR eigenfunction. Thus the sum converges and its coefficient is comparable to $\sum_{A\ge0}(r+As)^{-\beta}\asymp(1+r/s)r^{-\beta}$. The recursion in \eqref{eq:flow} follows by partitioning at the first letter. Ties in sorting the seed are resolved by a fixed rule on the slot labels during this formal count. For words $w,t$ we have $H(v_{wt})=q_w^tP_t\mathbf1$. Thus $c(q_w)/c(\mathbf1)$ is the limiting fraction of addresses with prefix $w$. These weights define a probability on infinite words. If a family of prefixes is prefix-free, the total number of finite descendants of height at most $L$ is bounded by $CL^\beta$ times the sum of its cylinder probabilities. This follows by applying \eqref{eq:flow} to each prefix. For a real column $z$ with $\mathbf1^tz\ne0$, write $[z]=z/(\mathbf1^tz)$. For a matrix $B$ and a point $z$ of the affine hyperplane $\mathbf1^tz=1$, write $[B]z=[Bz]$ when $\mathbf1^tBz\ne0$. All simplex interiors below are relative to this affine hyperplane. We use $\|z\|_1=\sum_i|z_i|$, and $\|B\|_1$ for its induced operator norm. Let $\Delta_0=\{z\ge0:\mathbf1^tz=1\}$. The simplices $[R_w]\Delta_0$ shrink uniformly with word length. The $i$th coordinate of $[R_i]z$ is $(2-z_i)^{-1}$, and its other coordinates are $z_j/(2-z_i)$. For a tangent vector $v$ with $\mathbf1^tv=0$, differentiating gives \[ \|D[R_i](z)v\|_1\le \frac{\sum_{j\ne i}|v_j|}{2-z_i} +\frac{(1+\sum_{j\ne i}z_j)|v_i|}{(2-z_i)^2} =\frac{\|v\|_1}{2-z_i}. \] For $i\ne j$, we have $([R_j]z)_i\le1/2$. Hence a pair of distinct successive letters contracts by at least a factor $2/3$. A constant word of length $k$ has image diameter $2/(k+1)$. The number of disjoint changing pairs or the length of the largest constant run must tend to infinity uniformly. Let $K=\bigcap_{k\ge0}\bigcup_{|w|=k}[R_w]\Delta_0$, where $|w|$ is the number of letters in $w$, and let $\kappa$ be the distribution of the limit under the cylinder law above. Let $\delta_z$ denote unit point mass at $z$. Testing against continuous functions at a fixed cylinder depth and then sending that depth to infinity gives weak convergence \begin{equation}\label{eq:weak} L^{-\beta}\sum_{H(v_w)\le L}\delta_{[v_w]} \ \Longrightarrow\ c(\mathbf1)\kappa. \end{equation} Indeed the finitely many shorter words have vanishing mass, and every other normalized address and its infinite continuations lie in the same shrinking simplex. Let $\Delta\subset\Delta_0$ be the sorted simplex. For $1\le j\le d$, let $S_j$ move slot $j$ to slot $d$ preserving the order of the other slots, and set $\gamma_j=S_jR_j$. Then $\gamma_j$ induces the sorted update, and the accelerated generators are $\gamma_d^A\gamma_j$ with $A\ge0$ and $j0\), where the matrix norm is induced by the vector \(\ell^1\) norm. \end{lemma} \begin{proof} Write \(p_a(z)=\|M_a z\|_1^{-\beta}h([M_a]z)/h(z)\). The roof is \(\log(1+(A+1)(1-z_j))\). Its derivative is uniformly bounded since \(z_j\le1/2\). Consequently \(p_a\) is bounded by \(C(A+2)^{-\beta}\) and \(\log p_a\) has a uniform Lipschitz constant. The conditional distribution of a letter given the next \(k\) letters is a positive weighted average of \(p_a\) on the corresponding cylinder. It follows that the conditional distribution given the infinite future is \(p_a\) at its coded point. Its log variation over completions of a \(k\)-letter future is \(O((2/3)^k)\). Apply the causal construction of Comets, Fern\'andez and Ferrari \cite{CFF} to this kernel in reversed time. In their notation let \(a_k\) be the infimum over recent histories of length \(k\) of the sum of the letterwise infima over their older completions. The log variation estimate gives \(a_k\ge \exp(-C(2/3)^k)\). The thresholds are nondecreasing and \(b=\prod_{k\ge0}a_k>0\). The resulting stationary process \(X_i\) is a function of independent uniforms \(U_j\), \(j\le i\), and its negative-time sequence \(X_{-1},X_{-2},\ldots\) has law \(m\) by uniqueness in the causal construction. For integers \(s0\), and therefore has infinitely many cuts in both directions. Conditional on \(\mathcal R_0\), any specified finite list of successive negative blocks has probability equal to the product of their finite-event probabilities: the underlying intervals and the uniforms defining \(\mathcal R_0\) are disjoint. These negative blocks are identically distributed. Also \(\mathcal R_0\) is independent of all negative-time symbols. Reversing each complete negative block and reading the blocks backward therefore gives the original law \(m\). A single-letter block with symbol \(a\) has probability equal to the letterwise infimum of the kernel, which is positive by the uniform log variation. For the gap \(L\) between cuts, put \(b_k=\prod_{i=0}^k a_i\) and \(b_{-1}=1\). On the cut event \(R_j\), the event \(R_0\) is \(H[0,j-1]\). Independence of its uniforms from those of \(R_j\), and \(\mathbb P(R_0)=\mathbb P(R_j)=b\), give \(\mathbb P(R_j\mid R_0)=b_{j-1}\). With \(Q(z)=\mathbb E[z^L]\) its generating function gives \[ \frac{1-z}{1-Q(z)} =1-\sum_{j\ge1}b_{j-2}(1-a_{j-1})z^j. \] The series on the right after the minus sign converges beyond \(1\), and at \(1\) equals \(1-b\). The geometric series in that sum converges at some \(z>1\), proving an exponential tail for \(L\) by the generating function of the tails. For a letter \(a\), \(c(a)=\log\max(e,\|M_a\|_1,\|M_a^{-1}\|_1)\) is \(O(1+\log(A+2))\), since \(\gamma_d\) is a unipotent matrix with square-zero nilpotent part. The conditional expectation of \(e^{t c(a)}\) is bounded uniformly by a function \(J(t)\) tending to \(1\) as \(t\downarrow0\). Iterated conditioning bounds the exponential reward of \(k\) consecutive letters by \(J(t)^k\). Conditioning on the cut at zero costs at most a factor \(b^{-1}\). Cauchy--Schwarz on the event \(L=k\) now bounds the conditional exponential reward of a whole block by \[ b^{-1/2}\sum_{k\ge1}J(2t)^{k/2}\mathbb P(L=k\mid\mathcal R_0)^{1/2}, \] finite for small positive \(t\). Submultiplicativity gives the claimed matrix moment. \end{proof} \begin{lemma}\label{geo:overlap} Let \(\mu_{\rm sym}\) be the average of the coordinate-permutation images of \(\mu\). Let \(G\in{\rm GL}_d(\mathbb Q)\) and let \(\rho\) be a finite measure on \(K\) such that \(\rho,[G]_*\rho\ll\mu_{\rm sym}\). For \(\rho\)-almost every \(x\) there are finite full words \(u,w\) and a coordinate permutation \(\Pi\) such that \[ G=cR_u\Pi R_w^{-1},\qquad x\in\operatorname{int}([R_w]\Delta_0),\qquad c\in\mathbb Q^\times . \] \end{lemma} \begin{proof} It suffices to prove the assertion with \(K_s\) and \(\mu\) in place of \(K\) and \(\mu_{\rm sym}\). Indeed, the hyperplane-nullity argument below gives zero mass to chamber boundaries. Split the source by the ordering chambers of \(x\) and \([G]x\), then conjugate the map by the two sorting permutations on its input and output. The resulting marginals are absolutely continuous with respect to \(\mu\), and moving the two permutations through the full words restores the claimed formula. The block law of Lemma \ref{geo:regen} makes \(\mu\) stationary. Its generated group contains \(\gamma_j\), \(\gamma_d\gamma_j\) for \(jK_0\). A target interval \([k,k+1)\) uses its chart at \(k\) and the cost at \(k\). Changing variables by \(h_*\) bounds the upper density of source-bad selected times by \[ \frac{\max\chi_i}{c_0\min\chi_i} \mathbb E[C_0\,\mathbf1_{\{\text{source interval at }0\text{ is bad}\}}]. \] The target bad density is its stationary probability. Both bounds follow by Birkhoff, with the fractional endpoint negligible since stationarity and integrability give \(C_l/l\to0\) by Borel--Cantelli. Compact exhaustion and integrability make both bounds arbitrarily small. Round the good times to \(m=\lceil h_*(t)\rceil\), \(k=\lfloor t\rfloor\). The carry \(N_k^{-1}GL_m\) in the moving charts is \(u\mapsto Bu/(1+\ell\cdot u)\), where \[ B_{ij}=A_{ij}a_j(L_m)/a_i(N_k),\quad \delta e^{-2K_0}\le\max_{ij}|B_{ij}|\le\delta,\qquad \ell_j=p_j a_j(L_m)\longrightarrow0. \] The upper bound holds at interpolated times for all entries and equality holds for an active entry. Rounding loses at most \(2K_0\) in that entry. On the compact chart set the coordinate ranges of \(\Delta\) are uniformly bounded and the target chart contains a uniform ball at zero in its interior. Choose \(\delta\) small enough for the carries to map all of \(\Delta\) inside \(\Delta\) with a uniform interior margin. In the moving homogeneous bases, normalize the carry to have top-left entry one, top row \((1,\ell)\) and bottom block \((0,B)\). The compact chart set bounds the basis matrices and their inverses. If \(x\in\Delta\) has homogeneous coordinates \(r(1,u)\) in the source basis, then \(r\) is bounded away from zero uniformly by the choice of charts. Its image in the target basis is \(r(1+\ell\cdot u,Bu)\), whose coordinate sum in the original basis is bounded below for small \(\ell\) and \(B\), since the target center has sum one. Thus the normalized matrices in the original basis range over a compact family with denominators bounded away from zero on \(\Delta\). For \(\ell=0\) each such map has positive oscillation on \(K_s\) by hyperplane nonconcentration. Compactness gives a uniform positive lower bound, which persists for sufficiently small \(\ell\). Denote the resulting compact family by \(\mathcal Q\). The union \(Y\) of the images of \(K_s\) under the singular members of \(\mathcal Q\) that map \(K_s\) into itself is compact by continuity, and has zero \(\mu\)-measure. For the latter assertion, every straight segment of positive length in \(K\) follows a rational edge line of a full cylinder. If it crosses from one corner cell to another, their intersection is the shared midpoint of a simplex edge. A line through that point lying on both sides in the simplex must follow the edge. Otherwise descend through cells until a crossing occurs, which must happen by shrinking. Every outer edge belongs to \(K\), since its two corner maps subdivide it at its midpoint. Choose one \(W=R_w\) with \([W]\Delta_0\subset\operatorname{int}\Delta\), possible by a typical sorted point in the chamber interior and full-cylinder shrinking. Then \([W]K\subset K_s\). For any admissible singular \(H\), the image of each outer edge under \([HW]\) is a point or a segment of \(K\), and all nonconstant connecting lines of its column images are rational. In rank two this is the entire range. In higher rank choose three noncollinear column images, which are rational as intersections of two rational lines. Each other column image is one of them or is determined by two rational connecting lines to them. Thus the range is a rational proper subspace. Only countably many ranges are possible, so hyperplane nonconcentration gives \(\mu(Y)=0\). Delete target times whose tail is in a sufficiently small fixed neighborhood of \(Y\). This costs arbitrarily little density by stationarity, ergodicity and compactness. The compact family was fixed before this deletion, so it still contains the normalized carries on the remaining positive-density set. Were a subsequence to have a singular limit \(H\), the density of \([L_m]^{-1}E\) would imply \([H]K_s\subset K_s\). Indeed the whole-domain interior bound places each image in the interior of the target prefix simplex, where unique sorted decoding implies the target tail belongs to \(K_s\). If a limit image left \(K_s\), continuity and full support would give a relatively open set of positive measure missing those source densities eventually. The actual target tail along this subsequence would consequently approach \(Y\), a contradiction. Thus the normalized carries are bounded away from singular matrices. If an actual carry is \(aH\) for a normalized representative \(H\), then \(|a|^d=|\det G|/|\det H|\). Thus the actual carries have bounded entries. All entries lie in one fixed rational lattice since both prefixes are unimodular. An actual carry \(D\) repeats infinitely often. On the open domain \(U=\operatorname{int}\Delta\cap[D]^{-1}\operatorname{int}\Delta\), the same density-to-inclusion argument gives \([D](K_s\cap U)\subset K_s\). Choose a repetition for which the actual tail is in \(U\), and take a full cylinder containing that tail, with its simplex contained in \(U\). The coding relation for the full gasket is precisely \(wi j^\infty\sim wj i^\infty\), \(i\ne j\): at the first differing letter the two cells meet only at the midpoint, and the two inverse images are the uniquely coded opposite vertices. This is also the coding quotient of the affine Sierpi\'nski gasket, equivariantly for all prefixes and permutations. The self-embedding theorem of Bandt and Retta \cite[Proposition~1.2, Theorem~5.1, Corollary~5.2 and Remark~5.7]{BR} therefore makes the restriction of \([D]\) to the chosen full cylinder a full prefix replacement followed by a permutation. Two projective transformations that agree on \(K\) agree everywhere. Their quotient fixes the vertices and is therefore represented by a diagonal matrix. Fixing each edge midpoint then forces its diagonal entries to be equal. Composing with the original sorted prefixes and moving permutations through the full words gives \(G=c R_u\Pi R_w^{-1}\). The chosen input cylinder contains \(x\), and it contains \(x\) in its simplex interior after deleting the countable union of boundary hyperplanes. The scalar is rational because the matrices are rational. \end{proof} \section{Comparison of the two measures} Let $\nu_{\mathrm{sym}}$ be the average of the coordinate-permutation images of $\nu$. We compare it with $\kappa$. For $1\le j\le d$ write \[ (T_j f)(z)=\|\gamma_j z\|_1^{-\beta} f([\gamma_j]z). \] The accelerated operator is $\sum_{j0$ such that for $L\ge2$, every nonzero real linear form $\ell$ and $0<\delta\le1$, \[ \#\{w:H(w)\le L, |\ell(z_w)|\le\delta\|\ell\|\|z_w\|\} \ll L^\beta\delta^a+L^{\beta-b}. \] For every sublattice $\Lambda\le\mathbb Z^d$ of finite index $I$, \[ \#\{w:H(w)\le L, z_w\in\Lambda\} \ll L^\beta\min(I,L^d)^{-c}. \] \end{lemma} \begin{proof} The vectors $v_w$ and $q_w$ range over the same set, since $q_w$ is the $v$-address of the reversed word. Sorting this set gives the orbit of the maps $\gamma_j$ from $\mathbf1$. Each word in the sorted maps corresponds to exactly one word in the unsorted maps given the stable order of the initial coordinates and the current labels. Unsorting contributes only finitely many permutations to the estimates. In the lattice estimate the change of variables takes a column lattice to the integral lattice $V\Lambda$, of index $|\det V|I$. We therefore prove the estimates for the sorted orbit. Use the accelerated branches $B_{A,j}=\gamma_d^A\gamma_j$ for $A\ge0$ and $j0$ be the Perron vector of $\gamma_1$. The span of its images under the accelerated semigroup is invariant under $\gamma_1$ and $\gamma_d\gamma_1$, hence under $\gamma_d$. It contains $(\gamma_d-I)v$, a nonzero multiple of $e_d$. For $0\le k0$, by summing the killed norm moments over the number of trials. Take $T$ to be a sufficiently small constant times $\delta^{-1/d}$. Any successful trial starting below $T$ excludes the slab in the lemma. The remaining words either have no completed successful trial, including words terminating inside a trial, or extend an unsuccessful crossing prefix. The former words are covered by the subcritical sum. Each crossing prefix $B$ has at most $O((L/(r\|B\|_1))^\beta)$ inner continuations from a seed of norm $r$, by the scalar count and the norm comparison. Summing yields $O((L/r)^\beta\delta^{a_0/d}+(L/r)^s)$. For the lattice assertion take $T=\tfrac12 I^{1/d}$ and $R=\sqrt T$. For bounded $I$ the scalar count suffices. We may replace $T$ by $\min(T,L)$ and increase constants. Given a sorted seed of norm $r_0T/R-2$. There are $O(1+\log R)$ steps before crossing $R$, since every step multiplies the norm by at least $3/2$. Under the forward Gibbs probabilities \[ \|B_{A,j}x\|_1^{-\beta}\frac{h_\beta([B_{A,j}]x)}{h_\beta(x)} \] the total mass of such stops is $O((1+\log R)(T/R)^{1-\beta})$. A prefix ending at norm $r'$ has probability comparable to $(r'/r_0)^{-\beta}$ and at most $O((L/r')^\beta)$ completions. Thus the overshoot contribution is bounded by its mass times $O((L/r_0)^\beta)$. Words ending below $R$ number $O((R/r_0)^\beta)$. Since $T\le L$ and $R^2=T$, this is bounded by $O((L/r_0)^\beta R^{-\beta})$. In all the preceding estimates, remove the initial unaccelerated run by summing over the seeds $\gamma_d^k\mathbf1$, whose norms are comparable to $k+1$. Choose $s>1$, so both seed sums with exponents $s$ and $\beta$ converge. In the lattice estimate seeds of norm at least $R$ contribute $O(L^\beta R^{1-\beta})$ by the scalar count. Dyadic summation over $r\ge R$ and absorption of the logarithm into a smaller power of $R$ give the desired bound. \end{proof} \section{Logarithmic interpolation}\label{ar:interpolation} Set $\sigma=n^{1/(n-2)}$. For the frame $P=P_w$, let $y_i=\sigma x_i$ be its parent labels and let $s=\sigma M(w)$ be its maximum. Define the linear form $\ell_P$ by $\ell_P(Pe_i)=\log y_i$. The exact Vieta recurrence in these coordinates is $s=\prod_i y_i-a$, where $a$ is the scaled coordinate removed in the preceding Vieta move. At the first child $a=\sigma$. Thus, putting $\Delta_P=\sum_i\log y_i-\log s$, \begin{equation}\label{ar:increment} \ell_{PA_i}-\ell_P=-\Delta_P e_i^tP^{-1}, \qquad 0\le\Delta_P\ll\exp(-c_0\min_i q_i), \qquad q=P^t\mathbf1 . \end{equation} Indeed the new parent in slot $i$ is $s$ and its column is the sum of the old columns. Also $a\le\max_i y_i$, and the degree bounds give $\log y_i\asymp q_i$. Cancelling a largest $y_i$ bounds $a/\prod_i y_i$ by $\sigma^{-(d-1)}<1$ and by $\exp(-c\min_i q_i)$. On this interval $-\log(1-t)\le t/(1-\sigma^{-(d-1)})$, which gives the defect bound. All constants below may depend on the fixed dimension. We record the growth gap needed to control this variation. Let $\xi_m$ denote the counting exponent for the $m$-coordinate row system, with the matrices $R_i$ defined in dimension $m$. In the notation of the counting theorem this exponent is $\beta_{m+1}$. The two-coordinate system has exponent $2$ by the Euclidean algorithm and the elementary count of primitive pairs. For $m\ge3$ use the scalar count of \cite{GMR}. Then $\beta=\xi_d>\xi_{d-1}$. In fact a word in the first $m=d-1$ letters, lifted to dimension $d$, has operator norm at most a fixed multiple of its original norm. Its upper-right column in the lift is $(B\mathbf1-\mathbf1)/(m-1)$ if $B$ is the original matrix, by multiplication of the block matrices. Appending the extra letter as a delimiter preserves this norm comparison. The scalar lower asymptotic implies divergence of the sum of inverse $\xi_m$ powers of the word norms. Select finitely many delimited blocks with that sum greater than one. The inequality persists at an exponent $\xi_m+\varepsilon$. All concatenations are distinct words and their norms are bounded by the products of block norms. Thus the corresponding Dirichlet series in dimension $d$ diverges at $\xi_m+\varepsilon$, and the scalar upper bound gives $\xi_d\ge\xi_m+\varepsilon$. From a seed $q$ of height $H=\mathbf1^tq$ with all $q_i\ge\theta H$, where $0<\theta\le1/d$, let $a_*=\xi_{d-1}$ and $g=\beta-a_*>0$. The number of continuations of height at most $T$ whose smallest coordinate is at most $h$ is \begin{equation}\label{ar:small} \ll \theta^{-\beta} H^{-\beta}h^g T^{a_*}\qquad(h\ge\theta H). \end{equation} For a final coordinate of value $b\le h$, cut at its last update. Its update makes $b$ the previous total, and the new total lies between $b$ and $2b$. The other $d-1$ coordinates, translated by $b/(d-2)$, thereafter follow the lower-dimensional row system and have initial values at least $b/(d-2)$. In a dyadic bin for $b$, there are $O(\theta^{-1}(b/H)^\beta)$ cut prefixes and $O((T/b)^{a_*})$ suffixes, by scalar counting and monotonicity from a constant lower seed. Summing uses $g>0$. If a coordinate is never updated, the bound $O((T/(\theta H))^{a_*})$ suffices, since $h\ge\theta H$. For a row state $q$ of total $r$ put $\omega(q)=1+r/(r-\max q_i)$. The same count weighted by $\omega(q)$ is bounded by \[ O\bigl(\theta^{-\beta-2}H^{-\beta}h^g T^{a_*}(1+\log(2T/h))\bigr) \quad(\theta H\le h\le T). \] To see this, partition by the start of the terminal run of maximum updates. At the start of any noninitial run the ratio $r/(r-\max q_i)$ is at most $3$: after updating a nonmaximum the latest total is less than twice the preceding total and the new sum of the other coordinates is at least half that preceding total. At the initial start the ratio is $O(\theta^{-1})$. The sum of weights for run lengths with final height at most $T$ is $O(\theta^{-2}(T/r)^2)$. The possible run starts in a dyadic bin at height $r$ are bounded by $O(\theta^{-\beta}H^{-\beta}\min(r^\beta,h^g r^{a_*}))$. Since $a_*\ge2$ and $\beta>2$, the weighted dyadic sum gives the displayed estimate, with a logarithm only if $a_*=2$. The chronological probability of a prefix extending the current seed is $c(q_{\rm new})/c(q)$, where $c(q)\asymp \omega(q)r^{-\beta}$. Denote this probability by $\mathbb P_P$. At the first crossing of $T\ge H$, the total lies between $T$ and $2T$. By the weighted bound, for any fixed $0<\kappaT$ as well. Choose $00). \] Multiply \eqref{ar:increment} on the right by any fixed real matrix $U$. Subadditivity of the $p$th power and summation of $2^{j(dp-\kappa)}$ give, for fixed constants $C,B,c_1>0$, \begin{equation}\label{ar:variation} \mathbb E_P\left(\sum_t\| (\ell_{P_{t+1}}-\ell_{P_t})U\|\right)^p \le C\theta^{-B}\|P^{-1}U\|^p e^{-c_1\theta H}. \end{equation} Thus the row forms converge almost surely to a measurable row form $\ell_\infty$. The bound $\ell_P(P\mathbf1)\gg\|P\mathbf1\|$ also gives $\|\ell_\infty\|\gg1$. We next prove a uniform small-value estimate. A return state means a frame just after updating a nonmaximum parent, so $r/(r-\max q_i)\le3$. From any such state, a return block consists of $A$ maximum updates followed by one nonmaximum update. Its probability is comparable to $(A+1)^{-\beta}$: the ratio of final to initial height is comparable to $A+1$, and both return states have bounded $\omega$. Therefore the height ratio in any fixed number of return blocks has a uniformly bounded moment of every order strictly below $\beta-1$. We show that from every return state there is a uniformly positive-probability escape for every nonzero form argument $k$. In the frame $P$, fix the current maximum $s$ and $d-2$ of the old parents, including a largest old parent, leaving two old parents with labels $b,c$ to move. Let $f_1=s,f_2,\ldots,f_{d-1}$ be the frozen labels. Use $f_{i,v},b_v,c_v$ for their vectors expressed in the original parent coordinates. Put \[ C=\prod_i f_i,\quad S=\sum_i f_i^2,\quad \lambda=(C+\sqrt{C^2-4})/2,\quad L_0=\log\lambda,\quad \delta_0=\log(C/\lambda),\quad u=\sum_i f_{i,v}. \] The alternating mutations of the two moving coordinates satisfy \[ y_{j+1}=Cy_j-y_{j-1},\qquad y_0=b,\quad y_1=c,\qquad y_j=A\lambda^j+B'\lambda^{-j}, \qquad AB'=S/(C^2-4). \] The product identity follows by inserting the solution into $y_0^2+y_1^2-Cy_0y_1=-S$, and $A+B'=b$ proves that both coefficients are positive. The corresponding vectors satisfy $v_j=c_v+(j-1)u$ for $j\ge1$ and $v_{-j}=b_v+ju$ for $j\ge0$, by the parent updates. The vectors $f_{i,v}$ and $c_v$ form a unimodular basis with bounded matrix and inverse. After a long alternating segment in either direction, update every frozen label, choosing $f_i$ first. The limiting linear forms in the original parent coordinates, denoted by $G_{i,+},G_{i,-}$, satisfy \[ G_{i,\pm}(f_{j,v})=\log f_j-\delta_0\mathbf1_{i=j},\quad G_{i,\pm}(u)=L_0,\quad G_{i,+}(c_v)=L_0+\log A,\quad G_{i,-}(b_v)=\log B'. \] The difference of consecutive moving vectors is $u$ in the direction of travel, and their logarithmic difference tends to $L_0$. Thus the first removed frozen vector absorbs the difference $\log C-L_0$. Later updates add parent vectors and, in the limit, add their logs. Write $r=d-1$ for the number of frozen labels. The estimates $C\ge2s$, $\lambda\ge s$, $s^{1-2r}\le A,B'\le s$, and $s^{-2r}\le\delta_0\le s^{-2}$ follow from $f_i\in[\sigma,s]$ and the product identity. Differences of positive-arm forms give $\delta_0$ times all coordinate differences on the frozen span. For $i\ne1$ the difference of the opposite arms vanishes on that span and evaluates on $c_v$ to \[ \log\biggl(\frac{\lambda^2 S}{(C^2-4)s^2}\biggr)\ge\delta_0/2 . \] Here $b_v+c_v=2f_{1,v}-u$, and the bound follows from $S\ge s^2+(r-1)$, $\lambda^2>C^2-4$ and $\delta_0\le s^{-2}$. The common value on $u$ is $L_0>0$. It follows quantitatively that $\max_{i,\pm}|G_{i,\pm}(x)|\gg\delta_0\|x\|$: the differences bound the moving coordinate and all frozen coordinate differences, and a single form bounds the remaining coefficient of $u$, since $\|G_{i,\pm}\|\ll L_0$. At the end of $m$ consecutive alternating mutations, the absolute indices of the two moving labels are $m+O(1)$, with the smaller absolute index at least $m-1$. The relative errors from keeping only the corresponding dominant Pell terms are $O(s^{2r}\lambda^{-2m+2})$, since each of $A/B'$ and $B'/A$ is at most $s^{2r}$. As $\lambda\ge s$ and $r\ge2$, the errors in the two logs are $O(s^{4r-2m+1})$. Both moving values are $\gg s^{m-2r}$. Each of the next $d-1$ mutations removes a frozen entry at most $s$ while retaining two entries at least as large as the smaller moving value. Hence each nonlinear log defect is $O(s^{4r-2m+1})$. The new maximum column is the sum of the surviving columns. The target agrees with the ideal logs on those columns and therefore agrees with their sum on the new maximum column. Induction over the activations gives terminal residuals $O_d(s^{4r-2m+1})$ on the parent columns. The terminal relative parent matrix has entries $O_d(m)$ and determinant one, so its inverse has norm $O_d(m^{d-1})$ by cofactors. The terminal row error in the original parent coordinates is consequently $O_d(m^{d-1}s^{4r-2m+1})$. Choose a fixed large $m$ so that this error is smaller than a fixed fraction of the spanning bound. The two moving columns at the cusp have masses comparable to $mH$, since $H\le\mathbf1^tPu\le2H$. Each activated parent is replaced by the preceding maximum column, which also has mass comparable to $mH$. Thus the terminal row entries lie between $c_d mH$ and $C_d mH$ for some $c_d>0$. Apply \eqref{ar:variation} at the terminal frame with $U=P$. Its exponential decreases with $mH$ while $\delta_0\ge e^{-C_dH}$ by the degree bounds. Increasing $m$ once more makes the probability that subsequent variation exceeds the remaining spanning margin $O(e^{-cH})$. In particular, for each $k\ne0$, one of these words ensures \[ |\ell_\infty(k)|\ge e^{-C_2H}\|k\| quad\text{outside conditional mass }O(e^{-cH}). \] We used $\|P^{-1}k\|\ge\|k\|/\|P\|$. Each chosen word comprises exactly $d+1$ return blocks: two initial nonmaximum updates, a maximum run, then the $d-1$ activations. Its probability is bounded below uniformly, because its finitely many prescribed waiting times are bounded and both endpoints are return states. At any return trial select the corresponding escape word for the current frame and fixed $k$. Kill the trial process when the selected word occurs. H\"older's inequality, the positive escape probability and the uniform height ratio moment show that, for some $\alpha>0$, the conditional expectation of the $\alpha$th power of the height ratio times the failure indicator is at most $\rho<1$. Stop upon crossing height $U_0=t/C_2$. For a failed stop at trial $j$, condition successively on the preceding trials. The expected $\alpha$th power of the stopped height divided by $H$, restricted to failure through trial $j$, is at most $\rho^j$. Summing bounds the failed stopping mass by $(H/U_0)^\alpha\sum_{j\ge1}\rho^j$. If a success occurs before crossing, the probability of exceptional later variation summed over the disjoint first-success cylinders is $O(e^{-cH})$. This proves, from return states of height $H$, \begin{equation}\label{ar:sv} \mathbb P_P\bigl(|\ell_\infty(k)|1),\\ F=C_4\eta^{-s},\qquad b_j=F^2e_j,qquad T_j=Fe_j\log M,\end{gathered} \] where $s>3/\kappa+2/\alpha+3$ and $c_d,C_4$ are fixed sufficiently large. At the return cut for $T_j$, apply \eqref{ar:variation} to discard points with $\|\ell_\infty-\ell_P\|>c M^{-e_j}$. Its conditional cost is at most \[ C\varepsilon^{-3B}(C\varepsilon^{-2}Fe_j\log M)^{p(d-1)}M^{p e_j} \exp(-c\varepsilon Fe_j\log M), \] using the cofactor bound on $P^{-1}$. This is at most $\eta/(100d)$ since $e_j=O(F^{2(j-1)})$ and $\varepsilon F$ is a sufficiently large fixed power of $\eta^{-1}$. The remaining directions $\ell_\infty/\|\ell_\infty\|$ in each cylinder lie in a patch of diameter at most $M^{-e_j}$, by the lower bound on the limiting norm. Their norms are bounded above by $U_j=C(\varepsilon^{-2}Fe_j\log M)^d$, again by the cofactor bound and the parent logarithm estimates. In a patch containing a direction retained at all previous stages there can be at most one allowed $j$-plane whose neighborhood meets the patch. Indeed otherwise that retained direction would be within $O(M^{-e_j}+M^{-b_j})$ of two planes. By the separation inequality it would be within $O(M^{2d a_j-e_j})$ of their intersection. A zero intersection is impossible for a unit vector. A positive-dimensional intersection has smaller dimension $l$ and height at most $M^{a_l}$ when $M_d$ is large, and its radius $M^{-b_l}$ covers that direction. If the single allowed plane exists, choose a real unit vector $k$ perpendicular to it. A direction in its neighborhood has $|\ell_\infty(k)|\le U_j M^{-b_j}$. Apply \eqref{ar:sv} on the cut cylinder at $t=b_j\log M-\log(2U_j)$. The height is at most $C\varepsilon^{-2}Fe_j\log M$ and $t\ge b_j\log M/2$ for large fixed $C_4$. Thus the conditional loss is $O(e^{-cT_j}+(\varepsilon^{-2}/F)^\alpha)\le\eta/(100d)$. Sum the cylinder masses and then the stage losses. The scales $e_j,b_j$ are bounded by a fixed power of $\eta^{-1}$. Taking $D=b_{d-1}$ and using the hyperplanes orthogonal to primitive integer vectors of norm at most $M$ shows that the total boundary mass where $|\ell_\infty(k)|<4c M^{-D}\|k\|$ for some such vector is $O(\eta)$ for a fixed $c>0$. Now take a deeper return cut with \[ T=C_5 F((D+1)\log M+\log(1/E)+1). \] At each retained cylinder, \eqref{ar:variation} gives $\|\ell_\infty-\ell_P\|\le c M^{-D}$ with conditional probability at least $3/4$. If $|\ell_P(k)|0$ small. These first exits are prefix-free. By \eqref{ar:variation} their conditional mass is at most $C\theta^{-B}\tau^{-p}e^{-c_1\theta H}\le\eta$ when $C_5$ is large. Every remaining endpoint has $z_w=P z'$ with $z'\ge0$ and $\|z'\|_1\le\|z_w\|_1/(\theta H)$. Its partial row variation yields an error at most $C\tau\|z'\|$. The terminal Vieta defect is $O(e^{-c_0\theta H})$, which is at most $E\|z_w\|/2$ upon increasing $C_5$. This gives the required estimate at the actual endpoint. The uniform scalar continuation bound charges every prefix discard by $O(L^\beta)$ times its boundary mass. The endpoints stopping before the cut and the number of faces are both $O(H_*^\beta)$ by scalar counting. The cut estimates give the required polynomial bounds. \end{proof} \begin{proposition}\label{ar:exhaustion} For each $\epsilon>0$ there is a finite set $\mathcal F\subset\mathrm{GL}_d(\mathbb Q)$ with the following property. In any collection of pairs of words of height at most $L$ with equal maxima and both projections injective, at most $\epsilon L^\beta+o(L^\beta)$ pairs lie off the graphs $z_{w'}=Gz_w$, $G\in\mathcal F$. The remainder is uniform in the collection. \end{proposition} \begin{proof} Write a rational matrix uniquely as $A/q$ with $q>0$ and the entries of $A$ and $q$ having joint gcd one. Its height is $Q=\max(q,\max_{i,j}|A_{ij}|)$. For a graph of height $Q\le C L^d$, Lemma~\ref{ar:population} gives the uniform bound \begin{equation}\label{ar:matrix-count} \#\{z:Gz\text{ and }z\text{ are addresses of height at most }L\} \ll L^\beta Q^{-b_1}+L^{\beta-b_2} \end{equation} for fixed $b_1,b_2>0$. If $q\ge Q^{1/2}$, the integrality lattice of $G$ has index divisible by $q$: the image of $\mathbb Z^d$ in $(\mathbb Q/\mathbb Z)^d$ under $G$ has exponent $q$. The lattice bound applies. Otherwise a row of $G$ has norm at least $Q^{1/2}$. Remove the ball of radius $LQ^{-1/4}$ by scalar counting and apply the slab bound to that row on the remaining addresses, using $\|Gz\|\ll L$. Fix a small loss $\eta$ and take $M=L^{d+2}$ in Lemma~\ref{ar:faces}. Choose $E=L^{-A-1}$ with $A>(d+2)D+d+2$, where $D$ is independent of $E$. Outside $O(\eta L^\beta)+o(L^\beta)$ pairs, both endpoints are in retained faces and have absolute interpolation errors $O(L^{-A})$. For each face pair let $W$ be the span of its actual paired vectors. If the first projection has rank $r<\dim W$, select $r+1$ independent pairs, the first $r$ of whose left vectors are independent. Signed $r$-minors give a nonzero integral relation among the left vectors, with coefficients $O(L^r)$. The same combination on the right gives a nonzero integer vector $k$ by independence of the pairs. It satisfies $\|k\|=O(L^{r+1})$ and $|\ell_{P'}(k)|=O(L^{r-A})$ by equality of the maxima. This contradicts the nonresonance bound. The reverse projection is also injective on $W$. Thus $\dim W\le d$. If it is less than $d$, the left endpoints lie in a real hyperplane, so the slab bound with $\delta\downarrow0$ makes the piece negligible. Otherwise $W$ is the graph of an invertible rational matrix of height $O(L^d)$ by Cramer's rule. There are $O_\eta((\log L)^{2\beta})$ face pairs. By \eqref{ar:matrix-count}, all graphs with height above $(\log L)^K$ have total negligible count for sufficiently large fixed $K$. Partition the remaining graphs into dyadic bands $Q_j/2d^2+2$. The slab bound and scalar small-ball bound imply that any subset of at least $L^\beta/Q_j^B$ addresses has $d$ vectors with smallest singular value at least $cLQ_j^{-K_1d}$ for a fixed large $K_1$. In fact each hyperplane has at most $CL^\beta t^a+CL^{\beta-b}$ points in the absolute slab of width $tL$ after possibly decreasing $a$ by splitting off the radius $\sqrt t L$. For $t=Q_j^{-K_1}$ and $j$ sufficiently large this bound is smaller than the subset size uniformly over $Q_j\le2(\log L)^K$. Choose the vectors successively at distance at least $tL$ from the previous span. Their determinant has absolute value at least $(tL)^d$, and all other singular values are $O(L)$. Use losses $\eta_j=j^{-2}$, normal bound $M_j=C_d Q_j^2$, nonresonance threshold $\omega_j=c M_j^{-D_j}$, and relative error $E_j=c'\omega_j Q_j^{-K_1d-4}$ in Lemma~\ref{ar:faces}. The heights and numbers of faces are polynomial in $j$ and independent of $L$. Delete endpoints failing their band cut, at total cost $O(L^\beta\sum_{j\ge j_0}j^{-2})+o(L^\beta)$ by injectivity and $j=O(\log\log L)$. For each graph and new face pair merge all occurrences inherited from the initial cut. Call the resulting piece large if its cardinality is at least $L^\beta/Q_j^B$. On a large piece, the singular value bound and the two relative error bounds yield $\|\ell_P-\ell_{P'}G\|\ll E_j Q_j^{K_1d}$. Two distinct large graphs $G=A/q$ and $G'=A'/q'$ in one face pair would give a nonzero integer column $k=(q'A-qA')e_i$ of norm at most $M_j$, with \[ |\ell_{P'}(k)|\ll E_j Q_j^{K_1d+2}<\omega_j \] for sufficiently small fixed $c'$. This contradicts nonresonance. There is thus at most one large graph per face pair. There are $O(Q_j^{d^2+1})$ possible matrices in the band. Hence the small pieces contribute at most $L^\beta Q_j^{d^2+1-B}$ times a polynomial in $j$. The large pieces contribute at most a polynomial in $j$ times the bound in \eqref{ar:matrix-count}. The exponential band tails in both estimates are summable, and the $L^{\beta-b_2}$ error remains negligible after summation over $j=O(\log\log L)$. First choose $\eta$ and then $j_0$ so that the total normalized loss is less than $\epsilon$. The finite set consists of all invertible rational matrices below the resulting fixed height threshold. Its choice is independent of $L$ and of the paired collection. \end{proof} \section{Equal maxima from two seeds}\label{ar:paired-section} \begin{lemma}\label{ar:paired} Let $a=a_w$ be the framed solution of a finite word $w$, and let $b$ be a distinct positive integral solution with its unique maximum in the last slot. Apply the same word of fixed-slot outward moves to both. The number of words with equal terminal maxima at most $R$ is $o((\log R)^\beta)$. \end{lemma} \begin{proof} Let $\Gamma$ be the group of polynomial automorphisms generated by the Vieta involutions and coordinate permutations. Put \[ J_9=\ker\bigl(\Gamma\longrightarrow\operatorname{Sym}((\mathbb Z/9\mathbb Z)^n)\bigr). \] On an ambient ball $z+3\mathbb Z_3^n$, with $z\in\{0,1,2\}^n$, the map $g\in J_9$ becomes $(g(z+3u)-z)/3$. This is congruent to the identity coefficientwise modulo $3$. The constant term is divisible by $3$, evaluation at $u=e_i$ determines the linear congruences, and every term of degree at least two has a factor of $3$. Poonen's interpolation theorem \cite{P} gives a restricted analytic parameterization of its iterates for times in $\mathbb Z_3$. On the closure of a diagonal $J_9$ orbit of a pair, let $\mu_{\rm orb}$ be the inverse limit of the uniform probabilities on its finite residue orbits. This measure has full support and is ergodic: the projection of any invariant function onto the functions on a finite quotient is constant by transitivity, and the finite quotient projections converge in $L^2$. The interpolated maps preserve the measure and satisfy the addition law for their time parameters, by continuity from integer iterates. These facts hold on each invariant residue ball and hence on the orbit closure. Let $E$ be the set on the closure where last coordinates agree. On each interpolated $g$ orbit its indicator is almost everywhere constant in the time variable, since a nonzero restricted analytic function has only finitely many zeros in $\mathbb Z_3$. Write $\mathbf1_E$ for its indicator and $dt$ for normalized Haar measure on $\mathbb Z_3$. Then \[ \int_{\mathbb Z_3}\!\int |\mathbf1_E(g^t x)-\mathbf1_E(g^{t+1}x)|\,d\mu_{\rm orb}(x)\,dt=0 . \] By invariance the inner integral equals $\int |\mathbf1_E(x)-\mathbf1_E(gx)|\,d\mu_{\rm orb}(x)$ for every $t$. Applying this to each $g\in J_9$ proves that $E$ has measure zero or one. To rule out measure one, write $m_i$ for the fixed-slot outward move on slots $i,n$. It acts on these two coordinates by the matrix \[ U(C_i)=\begin{pmatrix}0&1\\-1&C_i\end{pmatrix}, \qquad C_i=n\prod_{j\ne i,n}x_j . \] The value $C_i$ is unchanged by iteration. For a direct check modulo $9$, set $F_0=0$, $F_1=1$ and $F_{m+1}=CF_m-F_{m-1}$. Cayley--Hamilton gives $U(C)^m=F_m U(C)-F_{m-1}I$ for $m\ge1$. Iterating the scalar recurrence gives $(F_{35},F_{36})=(8,0)$ modulo $9$ for each $C=0,\ldots,8$. Thus $m_i^{36}\in J_9$. If the two pairs have matching last coordinates at times $0,36,72$ under $m_i$, the Cayley--Hamilton identity and positivity of the middle value imply equality of the two traces of $U(C_i)^{36}$. This trace equals $r^{36}+r^{-36}$ where $r=(C_i+\sqrt{C_i^2-4})/2$. Since $C_i>2$, equality forces equality of $C_i$. The nonzero lower-left entry of $U(C_i)^{36}$ then forces equality of coordinate $i$. Doing this for all $i$ would make the seeds equal. This argument holds at any translated pair because the maps are invertible and preserve positivity of integral solutions. Since $E$ is closed, full support excludes measure one. The $\Gamma$ orbit closure is the union of finitely many $J_9$ orbit closures. For each component, the proportion of its reduction modulo $3^k$ on which the last coordinates agree tends to zero by continuity from above of the compatible measure. The fraction for the full orbit is bounded by the sum of these finitely many component fractions, since the size of each component reduction is at most the size of the full orbit reduction. Fix the modulus $3^k$. Order the parent slots of both seeds by increasing entries of $q_w$, breaking ties by their slot indices. During a common fixed-slot move, each tuple replaces the selected parent by its old maximum and places the new maximum last. The corresponding row state replaces the selected entry by the sum of the row entries. This sum is strictly larger than every old row entry, so the order of the row state is maintained by appending the replaced parent at the end. Keep the parents of both tuples in this same row-state order, whether or not their entries increase in that order. With this convention the slot-$j$ update of the row state is $\gamma_j$, and the simultaneous slot-$j$ update of the tuples is $\lambda_j$, which moves slot $j$ last and then applies the Vieta involution there. The maxima remain in the last slot of each tuple. These updates give a bijection between the formal fixed-slot suffixes and the formal ordered suffixes, by tracking the parent-slot permutation at every step. The identities \[ \lambda_j^{-1}\lambda_{j+1}=(j,j+1)\quad (j1$ by the bound $O((A+2)^{-\operatorname{Re}s})$ and its logarithmic derivatives. The compact inclusion of $C^1$ into $C^0$ for the finite tuple of functions gives, by the two-norm argument of \cite[Proposition 45 and Lemma 46]{GMR}, an essential spectral radius strictly below $\lambda_u$, the leading eigenvalue of $L_u$. The branch orbit from every $(x,c)$ is dense in the scalar attractor times $\mathcal C$: use an outer block placing the simplex in a desired cylinder, and use an inner block to adjust the color to its preimage under that outer block. For an eigenfunction with eigenvalue $1$ at $s=\beta+it$, normalize by $h_\beta$. The maximum-modulus inequality, density of these orbits and uniform contraction imply that its modulus is constant and nonzero on the attractor times $\mathcal C$. Indeed iteration places all projective orbits uniformly near the attractor, and on the attractor equality propagates the maximum along every branch. Equality in the triangle inequality then gives a phase relation for each branch. For $a_A=\gamma_d^A\gamma_{d-1}$ the subspace where the first $d-2$ entries vanish is invariant. On the last two coordinates the matrix is $\bigl(\begin{smallmatrix}0&1\\1&A+1\end{smallmatrix}\bigr)$. Its positive Perron eigenvector is sorted and projects to the branch fixed point. The roof at this point is \[ \log\left(\frac{A+1+\sqrt{(A+1)^2+4}}2\right). \] Along any progression in $A$ with step equal to the order of the color permutation induced by $\lambda_d$, the branch color permutation is constant. Raising the branch to the order $m$ of that permutation shows that $mt$ times every such roof belongs to $2\pi\mathbb Z$. Successive roof differences are positive and tend to zero, so $t=0$. At $t=0$ the same branch relation and density show that the eigenfunction is constant after normalization on the attractor. Uniform contraction and iteration then show constancy everywhere. The normalized operator contracts the sup norm, which excludes a Jordan block at $1$. For the scalar leading eigenvalue, differentiation of the eigenvalue equation and application of the dual eigenmeasure give \[ \lambda'_\beta =-\int_{\Delta}\sum_{A\ge0,\,j0$. The color after the run remains in the same finite orbit. The scalar upper estimate bounds each normalized colored summand by $C(r+As)^{-\beta}$, independently of the modulus. Summation and dominated convergence therefore give the limit for ordinary words as $|E|/|\mathcal C|$ times their scalar coefficient $c(q)$ from \eqref{eq:flow}. For the two seeds, take $q$ to be the sorted version of $q_w$ and take $E$ to be the set of colors where the final coordinates agree. For a suffix $t$, its final row-state norm equals $H(v_{wt})$ by the continuation identity, since sorting preserves the sum of the entries. By \eqref{eq:height}, every suffix counted in the lemma has row-state norm at most $L_R=1+(n-2)\log R/\log(n-1)$ and terminal color in $E$. For this fixed modulus, its count divided by $(\log R)^\beta$ thus has upper limit at most \[ c(q)\left(\frac{n-2}{\log(n-1)}\right)^\beta\frac{|E|}{|\mathcal C|}. \] The factor preceding the residue fraction depends on $w$ and $n$ only. Sending $k\to\infty$ makes the residue fraction vanish and proves the lemma. \end{proof} \section{Proof of Theorem~\ref{thm:main}} \begin{proof}[Proof of Theorem~\ref{thm:main}] Fix an invertible rational matrix $G$, and count pairs of addresses $v_w,v_u$ of height at most $L$ with $v_u=Gv_w$, equal maxima, and distinct sorted tuples. If this count is not $o(L^\beta)$, along a sequence of $L$ the measures assigning mass $L^{-\beta}$ to each pair $([v_w],[v_u])$ have a nonzero weak limit. Each address occurs at most once in each marginal because $G$ is invertible. By \eqref{eq:weak} the limiting marginals are dominated by $c(\mathbf1)\kappa$. The limit is supported on the projective graph of $G$ with positive scale: for simplex points $x,y$ on that graph, $Gx=ay$ with $a=\|Gx\|_1$ bounded below by a positive constant. Thus for the first marginal $\rho$, both $\rho$ and $[G]_\ast\rho$ are absolutely continuous with respect to $\nu_{\mathrm{sym}}$ by \eqref{eq:comparison}. Lemma~\ref{geo:overlap} covers almost every such source point by open prefix simplices on each of which \[ G=cR_{u_0}\Pi R_{w_0}^{-1}, \] with $\Pi$ a coordinate permutation and $c\ne0$. Any actual finite address in the interior of $[R_{w_0}]\Delta_0$ has prefix $w_0$: at each inverse step the strict dominant-coordinate inequality forces the next letter, and $[\mathbf1]$ cannot lie in a first-level simplex. If the patch contains an actual collision pair, primitivity of both integral vectors forces $|c|=1$, and positivity of the two vectors forces $c=1$. For the remaining tail $t$ the two addresses are therefore $v_{w_0 t}$ and $v_{u_0\Pi(t)}$, where $\Pi(t)$ acts on the letters of $t$. Their maxima are obtained by applying the common tail $t$ to the framed seeds $a_{w_0}$ and $\Pi^{-1}a_{u_0}$, with the permutation acting on the parent slots. If the seeds coincide, word recovery gives $u_0=\Pi(w_0)$, so $G=\Pi$ and the sorted tuples are equal. If the seeds are distinct, Lemma~\ref{ar:paired} and \eqref{eq:height} bound the patch count by $o(L^\beta)$, since the maxima of its addresses are at most $\exp(L\log n/(n-2))$. The open-set inequality for weak convergence gives zero $\rho$-mass to each patch. The countable cover contradicts the choice of $\rho$. This proves the required $o(L^\beta)$ estimate for every fixed $G$. For each maximum occurring at least twice among the sorted tuples of largest entry at most $R$, form a directed cycle through its distinct tuples. It has one pair per tuple and both projections are injective after choosing one framed address per sorted tuple. By \eqref{eq:height} both address heights are at most \[ L_R=1+\frac{n-2}{\log(n-1)}\log R . \] Given $\varepsilon>0$, Proposition~\ref{ar:exhaustion}, conjugated by $V$ from column addresses $z_w=P_w\mathbf1$ to normal addresses $v_w$, puts all but $\varepsilon L_R^\beta+o_\varepsilon(L_R^\beta)$ of these pairs on a fixed finite set of rational graphs. Each such graph contributes $o(L_R^\beta)$ by the preceding argument. Letting $R\to\infty$ and then $\varepsilon\downarrow0$ shows $B_n(R)=o((\log R)^\beta)$. The scalar word count in \eqref{eq:flow} gives the required lower bound for $T_n(R)$. For $H(v_w)\le(n-2)\log R/\log n$, equation \eqref{eq:height} gives $M(w)