\documentclass[11pt]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[margin=1in]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools} \usepackage{booktabs,longtable,array,enumitem,microtype,xcolor,fancyvrb} \usepackage{mathrsfs,xurl} \usepackage{float} \usepackage[colorlinks=false, linkbordercolor=red!55!, citebordercolor=green!55!, urlbordercolor=blue!55!, breaklinks=true]{hyperref} \usepackage[nameinlink,noabbrev]{cleveref} \usepackage{cite} \numberwithin{equation}{section} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{claim}[theorem]{Claim} \newtheorem{assumption}[theorem]{Assumption} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{construction}[theorem]{Construction} \newtheorem{algorithm}[theorem]{Algorithm} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newtheorem{example}[theorem]{Example} \newtheorem{convention}[theorem]{Convention} \DeclareMathOperator{\Gal}{Gal} \DeclareMathOperator{\Cl}{Cl} \DeclareMathOperator{\rd}{rd} \DeclareMathOperator{\Tr}{Tr} \DeclareMathOperator{\tr}{tr} \DeclareMathOperator{\rank}{rank} \DeclareMathOperator{\Res}{Res} \DeclareMathOperator{\Frob}{Frob} \DeclareMathOperator{\covol}{covol} \DeclareMathOperator{\vol}{vol} \DeclareMathOperator{\supp}{supp} \DeclareMathOperator{\ord}{ord} \DeclareMathOperator{\Span}{span} \DeclareMathOperator{\Hom}{Hom} \DeclareMathOperator{\coker}{coker} \DeclareMathOperator{\disc}{disc} \DeclareMathOperator{\Ind}{Ind} \DeclareMathOperator{\End}{End} \DeclareMathOperator{\artanh}{artanh} \DeclareMathOperator{\diam}{diam} \newcommand{\Q}{\mathbb Q} \newcommand{\R}{\mathbb R} \newcommand{\C}{\mathbb C} \newcommand{\Z}{\mathbb Z} \newcommand{\N}{\mathbb N} \newcommand{\F}{\mathbb F} \newcommand{\Ftwo}{\mathbb F_2} \newcommand{\OO}{\mathcal O} \newcommand{\E}{\mathbb E} \newcommand{\U}{\mathcal U} \newcommand{\one}{\mathbf 1} \newcommand{\eps}{\varepsilon} \newcommand{\Norm}{\mathrm N} \newcommand{\id}{\mathrm{id}} \newcommand{\Sha}{\mathrm{SHA256}} \newcommand{\code}[1]{\texttt{\detokenize{#1}}} \setcounter{tocdepth}{1} \setlength{\emergencystretch}{2em} \allowdisplaybreaks[2] \newcommand{\hgt}{\operatorname{h}} \newcommand{\inc}{\operatorname{I}} \DeclareMathOperator{\trdeg}{trdeg} \DeclareMathOperator{\Var}{Var} \DeclareMathOperator{\TV}{TV} \hypersetup{pdftitle={Constructions and explicit power savings for point-unit-circle incidences},pdfauthor={}} \setlist{itemsep=3pt,topsep=5pt} \begin{document} \title[Point--unit-circle incidences]{Constructions and explicit power savings for point--unit-circle incidences} \subjclass[2020]{Primary 52C10; Secondary 52C35, 11R37, 11R42} \keywords{Point--unit-circle incidences, unit distances, incidence bounds, number-field lattices, class field towers, polynomial partitioning, arithmetic heights} \author{Yuda Chen} \email{Yuda Chen: \textcolor{blue}{\href{mailto:math.yuda.chen@gmail.com}{math.yuda.chen@gmail.com}}} \address{School of Mathematical Sciences and LPMC, Nankai University, Tianjin 300071, China} \author{Yifei Wang} \email{Yifei Wang: \textcolor{blue}{\href{mailto:yifeiwang.math@gmail.com}{yifeiwang.math@gmail.com}}} \address{School of Mathematical Sciences and LPMC, Nankai University, Tianjin 300071, China} %\date{13 November 2026} \begin{abstract} Let $\mathcal U(n,m)$ be the maximum number of incidences between $n$ distinct points and $m$ distinct unit circles in the plane. We give two constructions and an explicit upper bound for this function. Writing $s=\min(n,m)$ and $M=\max(n,m)$, the first construction gives a four-piece monomial lower bound for all cardinalities, with balanced exponent $1.041169920$. The second gives balanced configurations along an unbounded sequence with exponent $1.0431722305539701$, and further unbalanced configurations. The upper bound is $\mathcal U(n,m)\ll M+s^{2/3-2\delta}M^{2/3+\delta}$, where $\delta=2^{-2^{664}}/514560000000>2^{-2^{665}}$. The constructions use two windows in a relative-norm-one number-field lattice, strengthened arithmetic estimates, and interpolation between local profiles. The upper bound uses finite incidence regularization, prediction, arithmetic heights, and a six-by-five rigidity obstruction. The resulting bounds for the unit-distance problem follow as corollaries. \end{abstract} \maketitle \section{Introduction}\label{sec:intro} \raggedbottom For a finite point set $P\subset\mathbb R^2$ and a finite family $\mathcal C$ of circles of radius one, write \[ I(P,\mathcal C)=\#\{(p,C)\in P\times\mathcal C:p\in C\},\qquad \mathcal U(n,m)=\max_{|P|=n,\ |\mathcal C|=m}I(P,\mathcal C). \] The points are distinct, and the circles are distinct. If $Q$ is the set of circle centers, then \[ I(P,\mathcal C)=I(P,Q):=\#\{(p,q)\in P\times Q:|p-q|=1\}. \] The sets $P$ and $Q$ may intersect. Interchanging them gives $\mathcal U(n,m)=\mathcal U(m,n)$. We use $s=\min(n,m)$ and $M=\max(n,m)$ throughout. The main question is how large $\mathcal U(n,m)$ can be as the two cardinalities vary independently. We study it from two directions: explicit configurations give lower bounds, and restrictions on all configurations give upper bounds. The lower-bound results below also distinguish configurations available at every sufficiently large scale from configurations along an unbounded sequence of scales. For comparison with the classical problem, we use the following normalization. Let $u(N)$ denote the maximum number of unordered unit-distance pairs among $N$ distinct planar points. The incidence formulation gives \begin{equation}\label{eq:unit-distance-comparison} 2u(N)\le\mathcal U(N,N)\le2u(2N). \end{equation} Indeed, using a point set as its own set of centers counts every pair twice. Conversely, the union of the points and centers has at most $2N$ elements, and each unordered pair contributes at most two incidences. \subsection{Lower bounds: constructions} \label{intro:lower} \paragraph{\textbf{Bounds along an unbounded sequence.}} Erd\H{o}s's lattice construction~\cite{Erdos1946} gives, for some $c>0$, \[ u(N)\ge N\exp(c\log N/\log\log N) \] along an unbounded sequence. The extra factor grows more slowly than every positive power of $N$. The later number-field constructions produce a fixed power gain, and Table~\ref{tab:comparison} records the principal steps relevant to our construction. The lower-bound exponent changed in May 2026. The construction of OpenAI~\cite{OpenAI2026} gives $u(N)\ge N^{1+\eta}$ for some fixed $\eta>0$ and arbitrarily large $N$. Its central idea is to use an ideal lattice in a number field of increasing degree and bounded root discriminant. An embedding into $\mathbb C$ turns suitable algebraic norm equations into Euclidean unit distances. Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang and Wood~\cite{AlonEtAl2026} gave a streamlined account, including an explicit positive value of $\eta$. Sawin~\cite{Sawin2026} sharpened the arithmetic and geometric estimates and obtained $1+\eta=1.014114$. The passage from many small split primes to many elements of one ideal class has a useful antecedent in the work of Ellenberg and Venkatesh~\cite{EllenbergVenkatesh2007} on class-group torsion. In the present setting, choosing one class turns many ideal identities into elements with a prescribed relative norm. Those elements become the unit directions of the geometric construction. The subsequent improvements came from several parts of the argument. Emmerich~\cite{Emmerich2026} studied the optimization of Sawin's integer parameters and the verification of explicit lower-bound certificates. Naslund's earlier manuscript~\cite{NaslundEarlier2026} reached the sequence exponent $1.0358324$. His mixed-signature construction~\cite{Naslund} then reached $1.043171$. Its tower is built over $\mathbb Q(\sqrt{241})$. The relative norm-one group, the local conditions at selected primes and a relative zeta estimate determine the final exponent. Our construction uses this arithmetic framework, with a stronger finite quotient, a sharper arithmetic estimate and separate local profiles for points and circle centers. \begin{table}[H] \centering\small \caption{Sequence lower bounds for $u(N)$ and, by \eqref{eq:unit-distance-comparison}, for $\mathcal U(N,N)$. Each power bound holds along an unbounded sequence of cardinalities.} \label{tab:comparison} \begin{tabular}{@{}p{0.49\textwidth}p{0.43\textwidth}@{}} \toprule Construction & Lower bound\\ \midrule Erd\H{o}s~\cite{Erdos1946} & $N\exp(c\log N/\log\log N)$\\ OpenAI; Alon et al.~\cite{OpenAI2026,AlonEtAl2026} & $N^{1+\eta}$, for a fixed $\eta>0$\\ Sawin~\cite{Sawin2026} & $N^{1.014114}$\\ Naslund's earlier manuscript~\cite{NaslundEarlier2026} & $N^{1.0358324}$\\ Naslund~\cite{Naslund} & $N^{1.043171}$\\ Theorem~\ref{thm:sequences} & $N^{1.0431722305539701}$\\ \bottomrule \end{tabular} \end{table} \paragraph{\textbf{Bounds at every sufficiently large scale.}} The distinction between a sequence of sizes and every sufficiently large size matters here. The degree of a field is discrete, and passing to a larger field can make the resulting configuration much larger. Adding isolated points fills a gap in cardinality but can lose the exponent one is trying to preserve. Tseng~\cite{Tseng2026} treated this issue directly and obtained the exponent $1.03158935$ for every sufficiently large $N$. Theorem~\ref{thm:uniform} uses two phases of local profiles and interpolates between them within one field. This produces the exponent $\beta_*$ at every scale and, more generally, the four terms of $\mathcal L(s,M)$. Theorem~\ref{thm:sequences} instead chooses the strongest balanced profile and retains the larger exponent $\alpha_*$ along a sequence. The two constructions therefore optimize different questions. \begin{table}[H] \centering\small \caption{Lower bounds valid for every sufficiently large $N$. The displayed estimates apply to both $u(N)$ and $\mathcal U(N,N)$, with absolute positive implicit constants.} \label{tab:uniform-lower} \begin{tabular}{@{}p{0.49\textwidth}p{0.43\textwidth}@{}} \toprule Construction & Lower bound at every large scale\\ \midrule Tseng~\cite{Tseng2026} & $\gg N^{1.03158935}$\\ Theorem~\ref{thm:uniform} & $\gg N^{1.041169920}$\\ \bottomrule \end{tabular} \end{table} \paragraph{\textbf{Two-parameter lower bounds.}} Theorems~\ref{thm:uniform} and~\ref{thm:sequences} give the incidence versions of these two conclusions. In the first, a piecewise monomial bound describes all pairs $(n,m)$. In the second, selecting particular scales gives a stronger balanced exponent and further unbalanced configurations. Set \begin{gather} \alpha_*=1.0431722305539701,\qquad \beta_*=1.041169920, \label{eq:lower-exponents}\\ r_*=\frac{10603071}{10000000},\quad a_*=\frac{39597848504393}{50767750000000},\quad b_*=\frac{1073769}{4061420},\notag\\ q_*=\frac{60307177}{1060307100},\quad R_*=\frac{1060307100}{999999923}. \label{eq:lower-parameters} \end{gather} Terminating decimals denote exact rational numbers. \begin{theorem}[Construction at every scale]\label{thm:uniform} There is an absolute constant $c>0$ such that, for all positive integers $n,m$, \begin{equation}\label{eq:main-uniform} \mathcal U(n,m)\ge c\bigl(n+m+\mathcal L(s,M)\bigr), \end{equation} where \begin{equation}\label{eq:lower-envelope} \mathcal L(s,M)=\min\left\{ s^{0.65015512}M^{0.3910148},\; s^{0.7456363}M^{0.2974058},\; s^{a_*}M^{b_*},\; sM^{q_*}\right\}. \end{equation} The active terms change at $M=s^{1.02},s^{1.04},s^{r_*}$. In particular, $\mathcal U(N,N)\gg N^{\beta_*}$. For every fixed $1\le r0$, there is an unbounded sequence of integers $s$ for which \[ \mathcal U(s,\lceil s^r\rceil) \gg_{r,\epsilon}s^{1+r(\alpha_*-1)-\epsilon}. \] \end{theorem} \paragraph{\textbf{Arithmetic mechanism of the constructions.}} The field tower supplies many algebraic directions that become unit vectors in the plane. Golod--Shafarevich theory~\cite{GolodShafarevich,Ershov} keeps the tower infinite, while prescribed local conditions control ramification and the degrees of selected primes. Cutting towers in this way was developed by Hajir, Maire and Ramakrishna~\cite{HajirMaireRamakrishna2021}. We use the class field theory and pro-$2$ group conventions of \cite{MilneCFT,NSW}; the relation between the group filtration and the associated graded algebra is the one studied by Jennings and Quillen~\cite{Jennings,Quillen}. Choose a quadratic extension $K/F$ in the tower and a complex embedding $\phi$ above a real place of $F$. The nontrivial involution $\iota$ of $K/F$ acts there as complex conjugation. Thus \[ N_{K/F}(\gamma-\eta)=1 \quad\Longrightarrow\quad |\phi(\gamma)-\phi(\eta)|=1. \] We choose finite subsets $\Gamma_L,\Gamma_R$ of ideal cosets by intersecting a common translate of a lattice with two windows. The points and circles are then given explicitly by \begin{equation}\label{eq:coordinate-output} \begin{split} P&=\{(\Re\phi(\gamma),\Im\phi(\gamma)):\gamma\in\Gamma_L\},\\ C_\eta&:\ (X-\Re\phi(\eta))^2+(Y-\Im\phi(\eta))^2=1, \qquad\eta\in\Gamma_R. \end{split} \end{equation} The embedding is injective, so points and centers remain distinct. Ideal-class selection and a mass formula count the useful norm-one differences. Poisson summation controls the sizes of the two windows. The class-number and Fourier-analytic normalizations are those of \cite{MilneANT,Tate1967}. Three logarithmic quantities organize the transfer: the size of the point window, the size of the center window, and the number of incidences. Keeping all three quantities separate is what makes the unequal-cardinality estimates possible. At a complex place, the two coordinates of a relative norm-one element can have different absolute values. Averaging over the relative units accounts for these values, and the regulator cancels against the class-number formula. The finite places contribute independent choices of valuation shells. Mixing the local functions changes the three quantities in controlled proportions without forming a Cartesian product of planar configurations. The arithmetic loss is governed by a relative zeta value. Character decompositions~\cite{Serre} reduce its estimation to $L$-functions, with functional equations coming from Hecke theory~\cite{Hecke1920}. The limiting prime inequalities of Tsfasman and Vl\u{a}du\c{t}~\cite{TsfasmanVladut2002} allow a comparison near $s=1$; explicit prime estimates \cite{RosserSchoenfeld1962} control the tails. The analytic evaluation follows the approximate-functional-equation approach of~\cite{Dokchitser}, with explicit error bounds at each stage. Together with the improved signature bound, this gives \[ \theta\ge\frac12-2^{-35},\qquad C=0.04226263259973019243081. \] The balanced profile and the two-window profiles use this same arithmetic input. \subsection{Upper bounds: incidence estimates} \label{intro:upper} An upper bound must hold for every configuration. The classical exponent is $4/3$ in the balanced problem; an improvement of its constant and an improvement of its exponent are different results. Table~\ref{tab:upper-history} separates these two kinds of progress. The basic geometric estimate is \begin{equation}\label{eq:classical-incidence} \mathcal U(n,m)\ll n^{2/3}m^{2/3}+n+m. \end{equation} It has the same form as the point--line theorem of Szemer\'edi and Trotter~\cite{SzemerediTrotter1983}. Spencer, Szemer\'edi and Trotter~\cite{SpencerSzemerediTrotter1984} established the corresponding $4/3$ exponent for unit distances, and Sz\'ekely~\cite{Szekely1997} gave a short proof using crossing numbers. The more general point--curve theorem of Pach and Sharir~\cite{PachSharir1998} explains the role of two geometric properties: two distinct unit circles meet at at most two points, and two distinct points belong to at most two unit circles. We recall a crossing-number proof of \eqref{eq:classical-incidence} in Lemma~\ref{ub:ip:incidence}. The incidence viewpoint also connects the problem with arrangements of curves. Clarkson, Edelsbrunner, Guibas, Sharir and Welzl~\cite{ClarksonEtAl1990} developed decomposition methods that bound the complexity of curve arrangements and yield incidence estimates. Cutting curves into arcs with at most one intersection per pair gives another route; Sharir and Zahl~\cite{SharirZahl2017} extended this approach to algebraic curves of bounded degree. Pach and Agarwal~\cite{PachAgarwal1995} give a broader account of these links between arrangements, repeated distances and geometric graphs. For unit circles, the common radius provides the additional structure used in the arithmetic stage below. For unit distances, \'Agoston and P\'alv\"olgyi~\cite{AgostonPalvolgyi2021} sharpened the constant to $u(N)\le1.94N^{4/3}$. OpenAI~\cite{ub:OAI2026} subsequently proved $u(N)\ll N^{4/3-\delta}$ for an absolute $\delta>0$, improving the exponent. Theorem~\ref{thm:upper} gives a concrete value of the exponent difference and a bound for unequal cardinalities. Polynomial partitioning supplies the passage to the two-parameter problem. Guth and Katz~\cite{ub:GK2015} used this method in their distinct-distance theorem, and Kaplan, Matou\v{s}ek and Sharir~\cite{ub:KMSh2012} developed direct applications to classical incidence estimates. \begin{table}[H] \centering\small \caption{Upper bounds for $u(N)$. By \eqref{eq:unit-distance-comparison}, the same exponents give upper bounds for $\mathcal U(N,N)$. The constant $1.94$ refers to $u(N)$.} \label{tab:upper-history} \begin{tabular}{@{}p{0.40\textwidth}p{0.28\textwidth}p{0.24\textwidth}@{}} \toprule Source & Upper bound & Exponent difference\\ \midrule Spencer--Szemer\'edi--Trotter~\cite{SpencerSzemerediTrotter1984} & $\ll N^{4/3}$ & $0$\\ \'Agoston--P\'alv\"olgyi~\cite{AgostonPalvolgyi2021} & $\le1.94N^{4/3}$ & $0$\\ OpenAI~\cite{ub:OAI2026} & $\ll N^{4/3-\delta}$ & Some $\delta>0$\\ Theorem~\ref{thm:upper} & $\ll N^{4/3-\delta_{\mathrm{up}}}$ & $\displaystyle\frac{2^{-2^{664}}}{514560000000}$\\ \bottomrule \end{tabular} \end{table} \paragraph{\textbf{An explicit two-parameter upper bound.}} Writing $M=s^r$ makes the role of the imbalance transparent: the gain over the classical nonlinear term is a factor $s^{(2-r)\delta_{\mathrm{up}}}$. It is positive for $1\le r<2$; at the endpoint the linear term takes over. Define \begin{equation}\label{eq:delta-explicit} \delta_{\mathrm{up}}= \frac{2^{-2^{664}}}{514560000000},\qquad \delta_{\mathrm{simple}}=2^{-2^{665}}. \end{equation} \begin{theorem}[Explicit incidence upper bound]\label{thm:upper} For all positive integers $n,m$, \begin{equation}\label{eq:main-upper} \mathcal U(n,m)\ll M+s^{2/3-2\delta_{\mathrm{up}}}M^{2/3+\delta_{\mathrm{up}}}. \end{equation} In particular, \[ \mathcal U(N,N)\ll N^{4/3-\delta_{\mathrm{up}}}. \] The exponent difference satisfies $\delta_{\mathrm{up}}>\delta_{\mathrm{simple}}>0$. The same bounds hold with $\delta_{\mathrm{simple}}$ in place of $\delta_{\mathrm{up}}$. \end{theorem} For $M=s^r$, $1\le r\le2$, the nontrivial term in \eqref{eq:main-upper} is \[ s^{\frac23(1+r)-(2-r)\delta_{\mathrm{up}}}. \] Thus the saving persists throughout $1\le r<2$. At the far end the elementary geometry gives an exact answer. \begin{proposition}[The terminal range]\label{thm:extreme} For positive integers $s,M$ with $M\ge s(s-1)$, \[ \mathcal U(s,M)=M+s(s-1). \] \end{proposition} \paragraph{\textbf{Mechanism of the upper-bound proof.}} The upper-bound proof starts with an incidence graph having too many edges. Regularization produces a graph with controlled degrees and expansion. A finite prediction argument then reduces the variation of certain conditional distributions. In the arithmetic stage, an edge is represented by an algebraic ratio: these ratios have large height, but suitable products around rectangles have small height. This is the probabilistic and arithmetic strategy of \cite{ub:OAI2026}. Two quantitative principles support the geometric and algebraic steps. The Borsuk--Ulam theorem supplies simultaneous polynomial bisectors; Matou\v{s}ek~\cite{Matousek2003} explains this topological mechanism and its combinatorial applications. Effective forms of the Nullstellensatz, including Koll\'ar's degree bounds~\cite{Kollar1988}, control polynomial identities certifying that a system has no common zero. Our argument also needs bounds on the heights of the coefficients in these identities. To obtain a numerical exponent difference, we keep the parameters finite throughout the argument. Heights are normalized as in Bombieri and Gubler~\cite{ub:BG2006}. Effective algebraic estimates come from the arithmetic Nullstellensatz of Krick, Pardo and Sombra~\cite{ub:KPS2001}, the ideal-membership bounds of Aschenbrenner~\cite{ub:Aschen2004}, and real algebraic geometry \cite{BasuPollackRoy}. These estimates bound the exceptional polynomial conditions before the size of the configuration tends to infinity. The final obstruction is a six-by-five grid. The grid reductions of~\cite{ub:GridRefinement2026,ub:GridJoint2026} suggest the small configuration on which to concentrate. We use the full logarithmic-Jacobian rank pattern and prove the required arbitrary-flow statement. Independent radicals and rigidity then force a height bound incompatible with the selected edge heights. The resulting parameters are $K=2^{68}$, $S_0=2^{649}$ and $a=2^{-2^{664}}$, yielding \eqref{eq:delta-explicit}. Finally, polynomial partitioning \cite{ub:GK2015,ub:KMSh2012} transfers the balanced estimate to arbitrary $(n,m)$. \subsection{Consequences for unit distances} The two lower-bound theorems and the upper-bound theorem now transfer to $u(N)$ through~\eqref{eq:unit-distance-comparison}. The quantifiers on the two lower bounds remain separate. \begin{corollary}\label{cor:unit-distances} For every sufficiently large integer $N$, \[ N^{\beta_*}\ll u(N)\ll N^{4/3-\delta_{\mathrm{up}}}. \] There is a sequence $N_j\to\infty$ such that \[ \frac{u(N_j)}{N_j^{\alpha_*}}\longrightarrow\infty. \] Consequently, for every $\epsilon>0$, the sequence lower bound may be written as $u(N_j)\ge N_j^{1.0431722305539701-\epsilon}$ for all sufficiently large $j$. \end{corollary} \begin{proof} For the lower bound at every scale, apply Theorem~\ref{thm:uniform} with $s=M=\lfloor N/2\rfloor$. The union of the points and centers has at most $N$ elements and at least half as many unordered unit pairs as cross-incidences. Add distinct points to obtain exactly $N$ elements. The upper bound follows from the left inequality in \eqref{eq:unit-distance-comparison} and Theorem~\ref{thm:upper}. For the sequence, use the coincident point and center sets in Theorem~\ref{thm:sequences}. \end{proof} \subsection{Organization and conventions} The construction proof occupies Part~\ref{part:constructions}. Sections~\ref{fi:section} and~\ref{ar:section} establish the field and arithmetic inputs; Section~\ref{tr:section} proves the two-window transfer. The balanced margin in Section~\ref{ba:section} proves \eqref{eq:main-balanced}, and Section~\ref{as:section} proves the all-scale envelope and~\eqref{eq:main-asym-sequence}. The elementary transfers in Section~\ref{el:section} give the other sequence statements and the terminal range. Part~\ref{part:upper} proves Theorem~\ref{thm:upper}; its final parameter calculation is in Section~\ref{ub:opt:completion}. All logarithms are natural unless a base is indicated. The construction proof uses exact finite algebra and outward-rounded interval arithmetic. The defining coefficients, parameter tables and finite recurrences are given in Appendix~\ref{sc:profiles}. The numerical sections specify their algorithms, error bounds and intermediate values. Standard results from class field theory, height theory, and real algebraic geometry are stated with the normalizations needed in their applications. \subsection{AI model disclosure} We acknowledge the use of GPT‑6 Astra in the discovery and exploration of the theorems presented in this work. \newpage \flushbottom \tableofcontents \clearpage \part{Incidence constructions}\label{part:constructions} \section{Elementary geometric transfers and an exact terminal range} \label{el:section} For finite sets $A,B\subset\mathbb R^2$, write \[ I(A,B)=\#\{(a,b)\in A\times B:|a-b|=1\}, \qquad \mathcal U(n,m)=\max_{|A|=n,\ |B|=m} I(A,B). \] Each element $b\in B$ represents the unit circle with centre $b$. The points within each set are distinct; the two sets are allowed to intersect. Thus this definition is exactly the point--unit-circle incidence problem, rather than a weighted version of it. Interchanging the two sets gives \begin{equation}\label{el:symmetry} \mathcal U(n,m)=\mathcal U(m,n). \end{equation} Throughout this section, $n,m\ge1$, $s=\min(n,m)$ and $M=\max(n,m)$. A \emph{balanced seed} is an actual pair $(A,B)$ with $|A|=|B|=N$ and $I(A,B)=E$. If the starting construction is a single set $U$ with $e(U)$ unordered unit-distance pairs, take $A=B=U$; then $E=2e(U)$. No disjointness between the two sides is needed for the arguments below. \subsection{Deterministic rectangular selection} \label{el:selection-section} \begin{lemma}[Two degree selections]\label{el:selection} Let $(A,B)$ be a balanced seed with parameters $(N,E)$. For integers $0\le a,b\le N$, put $G_{N,E}(a,b)=0$ if $ab=0$, and otherwise put \begin{equation}\label{el:rounded-rectangle} G_{N,E}(a,b)= \max\left\{ \left\lceil\frac bN\left\lceil\frac{aE}{N}\right\rceil\right\rceil, \left\lceil\frac aN\left\lceil\frac{bE}{N}\right\rceil\right\rceil \right\}. \end{equation} There are subsets $A'\subseteq A$, $B'\subseteq B$ with $|A'|=a$, $|B'|=b$ and \begin{equation}\label{el:rectangle-bound} I(A',B')\ge G_{N,E}(a,b) \ge \left\lceil\frac{abE}{N^2}\right\rceil. \end{equation} They can be obtained by two degree rankings, performed in each of the two possible orders, and retaining the better selection. \end{lemma} \begin{proof} The sum of the $a$ largest degrees on the $A$ side is at least $aE/N$. It is an integer, so retaining these $a$ vertices leaves at least $\lceil aE/N\rceil$ edges. In the resulting bipartite graph, the sum of the $b$ largest degrees on the $B$ side is at least $b/N$ times the total number of edges. This proves the first term in \eqref{el:rounded-rectangle}. Reversing the two rankings proves the second. Either term is an integer at least $abE/N^2$. Ties may be resolved by any fixed ordering of the seed vertices. \end{proof} The selected coordinates and circle equations are unchanged: a selected point is still a seed point, and a selected circle still has equation $|x-b|=1$ for its original centre $b$. Thus Lemma~\ref{el:selection} is a deterministic geometric construction, not only an averaging statement. Repeating the two rankings whenever the edge count strictly increases is optional; it can only improve the certified count. \subsection{Exact sizes, separated copies, and bridges} \label{el:copies-section} \begin{lemma}[Padding and joining]\label{el:padding-joining} Suppose a unit-distance configuration has two nonempty sides of sizes $a\le n$ and $b\le m$, with $F$ edges. It can be enlarged to sizes exactly $(n,m)$ with at least \begin{equation}\label{el:padding} F+(n-a)+(m-b) \end{equation} edges. More generally, let finitely many nonempty configurations have parameters $(a_i,b_i,F_i)$, where $\sum_i a_i\le n$ and $\sum_i b_i\le m$. They can be placed, joined and padded to give \begin{equation}\label{el:bridge-budget} \mathcal U(n,m)\ge n+m-1+ \sum_i(F_i-a_i-b_i+1). \end{equation} For an empty list of configurations, the right-hand side is $n+m-1$ and is still attainable. \end{lemma} \begin{proof} To add a point on the first side, put it on the unit circle about any existing point on the second side, avoiding the finitely many already used coordinates. It contributes at least one new edge. Apply the same procedure with the sides reversed for each remaining point on the second side. This proves \eqref{el:padding}. In particular, starting with a single unit-distance pair gives $n+m-1$ edges. To join a new configuration to those already placed, choose a point $a$ on its first side and an existing point $b$ on the second side. Translate the whole new configuration so that $a$ becomes $b+u$, where $|u|=1$. A collision between any new and old coordinate excludes at most one choice of $u$. There are only finitely many such exclusions, so a permitted $u$ exists. This translation preserves every internal edge and guarantees a new bridge. Other cross-configuration unit pairs only increase the count. If there are $k\ge1$ configurations, this procedure gives at least $\sum_iF_i+k-1$ edges on the two sums of their vertex counts. Padding then gives \eqref{el:bridge-budget}. \end{proof} Configurations may instead be kept completely separated when bridges are unnecessary. For example, after enclosing a seed in a disc of radius $R$, translating successive copies by a fixed vector of length greater than $2R+2$ keeps all inter-copy distances greater than one. \begin{proposition}[Integer transfer from one seed]\label{el:integer-transfer} Fix a balanced seed $(N,E)$ and use $G=G_{N,E}$ from \eqref{el:rounded-rectangle}. For every integer $1\le k\le s$, set \[ a_k=\min\{N,\lfloor s/k\rfloor\},\qquad b_k=\min\{N,\lfloor M/k\rfloor\}. \] Then \begin{align} \mathcal U(s,M)&\ge kG(a_k,b_k),\label{el:equal-copy-bound}\\ \mathcal U(s,M)&\ge s+M-1+ k\bigl(G(a_k,b_k)-a_k-b_k+1\bigr). \label{el:equal-copy-bridge} \end{align} One may take the maximum of these bounds, the bound $s+M-1$, and the bounds obtained from any other available seeds. For a useful allocation without equal-sized remainder blocks, write \[ q=\lfloor s/N\rfloor,\qquad v=s-qN,\qquad t=\min\{N,M-qN\}. \] Then \begin{equation}\label{el:full-plus-remainder} \mathcal U(s,M)\ge qE+G(v,t). \end{equation} Writing $x_+=\max\{x,0\}$, the stronger bridge-and-padding guarantee \begin{equation}\label{el:surplus-remainder} \mathcal U(s,M)\ge s+M-1+ q(E-2N+1)_++H(v,t) \end{equation} also holds, where $H(0,t)=0$ and $H(v,t)=(G(v,t)-v-t+1)_+$ for $v>0$. \end{proposition} \begin{proof} For \eqref{el:equal-copy-bound}, select an $(a_k,b_k)$ rectangle by Lemma~\ref{el:selection}, take $k$ separated copies, and add any missing vertices. Their sizes do not exceed $(s,M)$ by the definitions of $a_k,b_k$. Joining these copies and using Lemma~\ref{el:padding-joining} proves \eqref{el:equal-copy-bridge}. For \eqref{el:full-plus-remainder}, use $q$ full copies and, when $v>0$, one $(v,t)$ selected rectangle. For \eqref{el:surplus-remainder}, retain the full copies only when $E-2N+1>0$, and retain the remainder only when its analogous surplus is positive; then apply \eqref{el:bridge-budget}. Omitting a block frees vertices and does not violate either size budget. \end{proof} These are finite formulas for a given initial configuration and given integer targets. They do not assert that a seed is available at every requested size. They also distinguish the surplus in \eqref{el:bridge-budget} from a gain that would require counting an extra incidence for an isolated padding point. \begin{proposition}[A common-root construction for diagonal seeds] \label{el:diagonal-root} Suppose the seed has $A=B=U$, $|U|=N\ge2$, and $I(U,U)=E$. If \[ k=\left\lfloor\frac{s-1}{N-1}\right\rfloor\ge1, \] then \begin{equation}\label{el:diagonal-root-bound} \mathcal U(s,M)\ge s+M-2+k(E-2N+2). \end{equation} \end{proposition} \begin{proof} Choose a root $u_0\in U$. Take $k$ rigid copies of $U$ with their roots at the same point. Choose the rotations successively so that every non-root point of a new copy avoids all previously used coordinates; only finitely many angles are excluded at each stage. The resulting set $V$ has exactly $1+k(N-1)$ points. Different copies share only the root, so they share no unordered unit-distance edge. Hence $I(V,V)\ge kE$. Pad the two copies of $V$ to sizes $s$ and $M$ using \eqref{el:padding}. The number of added vertices is $s+M-2(1+k(N-1))$, which gives \eqref{el:diagonal-root-bound}. \end{proof} Both copies of the root, as a point and as a circle centre, are shared in this construction. Counting one as new would give incorrect sizes. The proposition makes no optimality claim about other gluing patterns. \subsection{The exponent transferred from a balanced sequence} \label{el:sequence-section} \begin{theorem}[A specified rectangular sequence]\label{el:sparse-transfer} Let $1<\alpha<2$. Suppose that, for every $\eta>0$, there is an unbounded sequence of actual balanced seeds $(N_j,E_j)$ satisfying $E_j\ge N_j^{\alpha-\eta}$. For every fixed $r\ge1$ and every $\varepsilon>0$, there is an unbounded sequence $s_j$ for which \begin{equation}\label{el:sparse-exponent} \mathcal U(s_j,\lceil s_j^r\rceil) \ge s_j^{\max\{r,\,1+r(\alpha-1)\}-\varepsilon}. \end{equation} The sizes and the construction can be chosen as follows: \begin{equation}\label{el:sparse-sizes} s_j=\lfloor N_j^{1/r}\rfloor, \qquad M_j=\lceil s_j^r\rceil, \end{equation} followed by the two degree selections of Lemma~\ref{el:selection}. \end{theorem} \begin{proof} Choose $0<\eta<\min\{\alpha-1,\varepsilon/(2r)\}$ and use its seed sequence. The integers in \eqref{el:sparse-sizes} satisfy $1\le s_j\le M_j\le N_j$ for all sufficiently large $j$, and $M_j/N_j\to1$. Lemma~\ref{el:selection} therefore gives \[ \mathcal U(s_j,M_j) \ge\frac{s_jM_j}{N_j^2}E_j \ge\frac{s_jM_j}{N_j^2}N_j^{\alpha-\eta} =s_j^{1+r(\alpha-1-\eta)+o(1)}. \] For sufficiently large $j$, this is at least $s_j^{1+r(\alpha-1)-\varepsilon}$. Independently, $\mathcal U(s_j,M_j)\ge M_j\ge s_j^r$, by the elementary padding construction. Taking the better construction proves \eqref{el:sparse-exponent}. \end{proof} With the certified balanced exponent used in this paper, \begin{equation}\label{el:alpha-value} \alpha_0=1.0431722305539701, \end{equation} the transferred exponent $1+r(\alpha_0-1)$ exceeds the linear exponent $r$ precisely when \begin{equation}\label{el:sparse-threshold} 1\le r<\frac1{2-\alpha_0} =1.045120168887829545\ldots. \end{equation} For instance, at $r=1.04$ the exponent before the arbitrary $\varepsilon$ loss is $1.044899119776128904$. Theorem~\ref{el:sparse-transfer} retains the sequence quantifier of its hypothesis. Its sequence may depend on $r$ and $\varepsilon$; it is not an assertion for every sufficiently large $s$. The all-size constructions elsewhere in the paper require their separate uniform counting and scale-coverage arguments. \subsection{An elementary construction at large imbalance} \label{el:terminal-section} \begin{theorem}[Exact terminal range]\label{el:terminal} For $s\ge2$, put $S=s(s-1)$. Then \begin{equation}\label{el:terminal-lower} \mathcal U(s,M)\ge M+\min\{M,S\}, \qquad M\ge s, \end{equation} and \begin{equation}\label{el:terminal-exact} \mathcal U(s,M)=M+s(s-1),\qquad M\ge s(s-1). \end{equation} For $s=1$ one has $\mathcal U(1,M)=M$. \end{theorem} \begin{proof} Choose $s$ distinct points $a_1,\ldots,a_s$ in a disc of radius $1/4$, with no three lying on a common unit circle. This choice can be made recursively with rational coordinates: after finitely many points have been selected, exclude the unit circles through each pair of old points, as well as the old points themselves. This is a finite union of circles and points, so its complement in the disc contains a rational point. Every pair $a_i,a_j$ has distance strictly between zero and two, and therefore lies on exactly two unit circles. Explicitly, writing $d_{ij}=|a_i-a_j|$ and letting $J$ denote rotation through $\pi/2$, their centres are \[ b_{ij}^{\pm}= \frac{a_i+a_j}{2} \ \pm\ \frac{\sqrt{1-d_{ij}^2/4}}{d_{ij}}\,J(a_i-a_j). \] The no-three condition makes all $2\binom s2=S$ circles distinct, and each contains exactly two of the chosen points. If $M\le S$, select any $M$ of these circles, obtaining $2M$ incidences. If $M\ge S$, select all of them and then add $M-S$ distinct unit circles through $a_1$ and through none of $a_2,\ldots,a_s$. Their centres can be chosen on the unit circle about $a_1$, avoiding finitely many forbidden centres. This gives $2S+(M-S)=M+S$ incidences. Thus \eqref{el:terminal-lower} and the lower half of \eqref{el:terminal-exact} hold. For the matching upper bound, consider arbitrary $s$ points and $M$ distinct unit circles, and let $d_C$ be the number of selected points on a circle $C$. Two distinct points lie on at most two unit circles, so \[ \sum_C\binom{d_C}{2}\le2\binom s2=S. \] The integer inequality $d\le1+\binom d2$ holds for every $d\ge0$. Consequently \[ I\le M+\sum_C\binom{d_C}{2}\le M+S, \] which proves the upper half of \eqref{el:terminal-exact}. For $s=1$, at most one incidence is possible per circle, and any number of distinct unit circles can be chosen through the single point. \end{proof} The bounds from the terminal construction and the seed constructions can always be combined by taking their maximum. None of the elementary transfers above purports to determine the extremal order throughout the intermediate imbalance range. \subsection{Continuous triple spectra and a sharper exact interval} \label{el:triple-section} Call a set of $s$ circle centres \emph{admissible} if its diameter is less than two, no four of the corresponding unit circles concur, and every intersection has multiplicity two or three. Distinct centres then give exactly two transverse intersections per pair of circles. A \emph{continuous triple spectrum up to $T$} means that, for each integer $0\le t\le T$, an admissible $s$-centre configuration with exactly $t$ triple intersections is available. Different values of $t$ may use different configurations. \begin{theorem}[Incidences from a continuous spectrum] \label{el:spectrum-incidences} Let $s\ge2$, $S=s(s-1)$, and suppose a continuous triple spectrum up to $T$ is available with exactly $s$ centres. For $s\le M\le S$, \begin{equation}\label{el:triple-lower} \mathcal U(s,M)\ge 2M+\min\left\{T,M,\left\lfloor\frac{S-M}{2}\right\rfloor\right\}. \end{equation} If, in addition, $3T0$ and $v_2(\sqrt{241})<0$. The following eight elements will also fix the coordinates of every elementary Galois image: \begin{equation}\label{fi:alpha} \begin{split} &\alpha_0=-1,\qquad \alpha_1=-71011068+4574225\sqrt{241},\\ &\alpha_2=(-6101-393\sqrt{241})/2,\quad \alpha_3=(6101-393\sqrt{241})/2,\\ &\alpha_4=31-2\sqrt{241},\quad\alpha_5=31+2\sqrt{241},\\ &\alpha_6=326-21\sqrt{241},\quad\alpha_7=326+21\sqrt{241}. \end{split} \end{equation} Their respective norms are $1,-1,-2,-2,-3,-3,-5,-5$. Let $\mathfrak p_j=(\alpha_{1+j})$, $\mathfrak q_j=(\alpha_{3+j})$, and $\mathfrak r_j=(\alpha_{5+j})$ for $j=1,2$, and set \[ S=\{\mathfrak p_1,\mathfrak p_2,\mathfrak q_1,\mathfrak q_2, \mathfrak r_1,\mathfrak r_2\}. \] We additionally distinguish \[ \mathfrak t_1=(-14127+910\sqrt{241}),\quad \mathfrak t_2=(-14127-910\sqrt{241}),\quad \mathfrak t_3=(621+40\sqrt{241}). \] The first two primes have norm $29$, the third norm $41$. The conjugate prime $\mathfrak t_4=(621-40\sqrt{241})$ is not capped. \begin{lemma}\label{fi:base-lemma} The ring of integers is $\mathbb Z[(1+\sqrt{241})/2]$, its discriminant is $241$, and its class number is one. The group of square classes with even valuations outside $S$ is an eight-dimensional $\mathbb F_2$-space, with basis the classes of \eqref{fi:alpha}. \end{lemma} \begin{proof} The formula for the integers and discriminant is the quadratic-field formula. The displayed norms and $\alpha_2\alpha_3=2$, $\alpha_4\alpha_5=-3$, $\alpha_6\alpha_7=-5$ show that the primes over $2,3,5$ are principal. The prime $7$ is inert, since $241\equiv3\pmod7$ is a nonsquare. Every ideal class has an integral ideal of norm at most $\sqrt{241}/2<8$ by Minkowski's theorem. Its prime factors are therefore among the principal primes just listed; this proves class number one. Consequently an element having even valuations outside $S$, modulo a square, is a unit times a product of the six specified prime generators. Their valuation vectors make those six square classes independent. The unit theorem gives two unit square classes as a vector-space dimension; $-1$ and the norm-$-1$ unit $\alpha_1$ are independent, as their two real sign vectors show. This proves the assertion without needing a claim that $\alpha_1$ is the smallest fundamental unit. \end{proof} Let $B_S$ be the maximal pro-$2$ extension of $B$ unramified at all finite primes outside $S$, without a condition at infinity, and let $G_S=\operatorname{Gal}(B_S/B)$. Kummer theory and Lemma~\ref{fi:base-lemma} identify its maximal elementary quotient with \[ \operatorname{Gal}(E/B)=\mathbb F_2^8,\qquad E=B(\sqrt{\alpha_0},\ldots,\sqrt{\alpha_7}). \] For a Galois element $g$ its elementary image is the string $a_0\cdots a_7$ defined by $g\sqrt{\alpha_i}=(-1)^{a_i}\sqrt{\alpha_i}$. Throughout, $[g,h]=g^{-1}h^{-1}gh$. Here is the complete local-vector input. At $3$ and $5$, $t$ denotes inertia and $f$ a Frobenius lift fixing the square root of the chosen uniformizer; $F_j$ denotes the Frobenius at $\mathfrak t_j$. \begin{equation}\label{fi:vectors} \begin{array}{c|c@{\qquad}c|c} \iota_1&10111010&\iota_2&11000101\\ x_1&00100000&x_2&00010000\\ y_1&11110010&y_2&10000001\\ z_1&10000100&z_2&11111000\\ t_{q1}&00001000&f_{q1}&10100111\\ t_{q2}&00000100&f_{q2}&11101011\\ t_{r1}&00000010&f_{r1}&01100101\\ t_{r2}&00000001&f_{r2}&01011010\\ F_1&00100111&F_2&00011011\\ F_3&00111000&& \end{array} \end{equation} For reproducibility, these vectors can be calculated directly from \eqref{fi:alpha}, using the roots of $241$ congruent respectively to $7,25\pmod {32}$ at the two dyadic places, $2,1\pmod3$, $1,4\pmod5$, $3,26\pmod {29}$, and $6\pmod {41}$. At an odd prime, record the parity of the valuation and whether the unit part is a quadratic residue. At $2$, record the valuation modulo two and the odd part modulo eight, in the square-class basis $-1,2,5$. The dyadic generators $x,y,z$ act on $(\sqrt2,\sqrt{-1},\sqrt5)$ by $(-,+,+),(+,-,+),(+,-,-)$. Thus they evaluate a class $(-1)^a2^b5^c$ as $b,a,a+c$, respectively. These rules give every entry of \eqref{fi:vectors}; for example at $\mathfrak p_1$, $\alpha_1\equiv3\equiv-5\pmod8$. The real rows are simply the signs of \eqref{fi:alpha}. \subsection{The prescribed dyadic extension}\label{fi:dyadic} Let $i=\sqrt{-1}$, $k=\mathbb Q_2(i)$, $\pi_5=1+2i$, and let $U_4/\mathbb Q_2$ be unramified of degree four. Define \[ M_8=\mathbb Q_2(i,\sqrt5,\sqrt{\pi_5}),\quad M_{16}=M_8(\sqrt2),\quad L_2=M_{16}U_4. \] The finite group needed at both dyadic places is \begin{equation}\label{fi:D} D=\langle x,y,z\mid x^2=y^2=z^4=[y,z]^2=[x,y]=[x,z] =[[y,z],z]=1\rangle. \end{equation} \begin{lemma}\label{fi:dyadic-lemma} The extension $L_2/\mathbb Q_2$ is Galois with group $D$, degree $32$, ramification index $8$, residue degree $4$, and normalized different exponent $18/8=9/4$. Its Zassenhaus layers have dimensions $3,2$ in degrees one and two, with all later layers zero. Therefore \[ h_D(t):=h_{\mathbb F_2[D]}(t)=(1+t)^3(1+t^2)^2. \] \end{lemma} \begin{proof} The identity $\pi_5\overline{\pi_5}=5$ supplies both conjugate radicals. There are automorphisms $y_0,z_0$ of $M_8$ which both send $i$ to $-i$ and $\sqrt{\pi_5}$ to $\sqrt5/\sqrt{\pi_5}$, and send $\sqrt5$ respectively to $\sqrt5$ and $-\sqrt5$. They satisfy $y_0^2=1$, $z_0^4=1$, $y_0z_0y_0=z_0^{-1}$, with $z_0^2\ne1$. Over $k$, the extension by $\sqrt5$ is unramified whereas that by $\sqrt{\pi_5}$ is ramified, so $[M_8:\mathbb Q_2]=8$. Its three quadratic subfields have square classes $-1,5,-5$; hence $\sqrt2\notin M_8$ and $[M_{16}:\mathbb Q_2]=16$. The largest unramified subextension of $M_{16}$ is $\mathbb Q_2(\sqrt5)$: its Galois group has elementary abelianization, so every cyclic quotient has order at most two. Consequently $M_{16}\cap U_4=\mathbb Q_2(\sqrt5)$ and $[L_2:\mathbb Q_2]=32$. Lift $y_0$ trivially on $U_4$, lift $z_0$ by the Frobenius on $U_4$, and take $x$ to negate only $\sqrt2$. These elements satisfy \eqref{fi:D}. Writing $w=[y,z]$, its normal forms are $x^a y^b z^c w^e$, with $a,b,e\in\{0,1\}$ and $0\le c<4$; their actions on the radicals and $U_4$ distinguish all $32$ forms. The inertia group is $\langle x,y,w\rangle$ of order eight. Here is also a direct different calculation. In $k$, put $\varpi=1+i$. The conductor exponents of the quadratic characters with square classes \[ 1,\quad5,\quad\pi_5,\quad5\pi_5,\quad2,\quad10,\quad 2\pi_5,\quad10\pi_5 \] are $0,0,2,2,4,4,4,4$. To check the nonzero entries, use the Eisenstein polynomial $X^2+2(1+\varpi)X/\varpi+2/\varpi$ for the extension by $\sqrt{\pi_5}$, and $X^2+2X+1-u$ for $u=2-i$ and $u=i$, corresponding to $2\pi_5$ and $2$ modulo squares. The derivative valuations give $2,4,4$; an unramified quadratic twist by $5$ preserves these conductors. The conductor--discriminant formula gives relative discriminant exponent $20$ for $M_{16}/k$. Since the different exponent of $k/\mathbb Q_2$ is two, $\operatorname{ord}_2\mathfrak d_{M_{16}/\mathbb Q_2}=8\cdot2+20=36$. Its residue degree is two, so its different exponent is $18$; passage to the unramified compositum does not change this. Finally $D_2D$ is the central elementary group generated by $z^2,w$, and $D_3D=1$. Jennings's basis theorem yields the displayed Hilbert polynomial. \end{proof} We use the following standard local presentation statements. At a real place the maximal pro-$2$ group is $C_2$. At an odd prime $p$, it is generated by $t,f$ with relation $ftf^{-1}t^{-p}$, whose quadratic initial over $\mathbb F_2$ is $[t,f]+((p-1)/2)t^{[2]}$. The maximal pro-$2$ group of $\mathbb Q_2$ is a rank-three Demushkin group with one relation. In the dyadic coordinates above its quadratic initial is \begin{equation}\label{fi:dyadic-initial} y^{[2]}+[x,y]+[x,z]. \end{equation} For the last assertion, local duality identifies the relation pairing with the Hilbert symbol. In the dual square-class basis $2,-5,5$, that symbol has indicator matrix $\left(\begin{smallmatrix}0&1&1\\1&1&0\\1&0&0\end{smallmatrix}\right)$; off-diagonal entries give brackets and diagonal entries restricted squares. This proves \eqref{fi:dyadic-initial}, including its diagonal term, rather than choosing a convenient but unspecified dyadic lift. These local presentation and duality statements are the standard local Galois theorems used here \cite{NSW,MilneCFT}. \subsection{The global presentation and the tower group}\label{fi:presentation} Let $\mathfrak F$ be the free pro-$2$ group on eight generators mapping minimally onto $G_S$. For a local place choose lifts of its generators to $\mathfrak F$, and denote its lifted defining relation by $\rho_\nu$. Let $\Sigma=S\cup\{v_1,v_2\}$. \begin{lemma}\label{fi:seven-relations} The group $G_S$ has relation rank seven. Any seven of the eight local relations $\rho_\nu$ normally generate the kernel of $\mathfrak F\longrightarrow G_S$. \end{lemma} \begin{proof} There are eight generators by Kummer theory. To bound the relation rank, let $X=\operatorname{Spec}\mathcal O_B[1/30]$. The Kummer exact sequence, $\operatorname{Pic}(X)=0$, and the Brauer--Hasse--Noether theorem \cite{MilneCFT,Poonen} give \[ H^2(X,\mu_2)=\operatorname{Br}(X)[2] =\ker\left(\bigoplus_{\nu\in\Sigma}\operatorname{Br}(B_\nu)[2] \longrightarrow\tfrac12\mathbb Z/\mathbb Z\right). \] Each local term has dimension one and the sum map is onto, so the right side has dimension seven. The Hochschild--Serre sequence for finite Galois $2$-covers of $X$ \cite{MilneEtale} injects $H^2(G_S,\mathbb F_2)$ into this space: the intervening direct limit of $H^1$ vanishes, since a double cover of a finite $2$-cover becomes trivial on a further $2$-cover (take its Galois closure, whose group embeds in a wreath product of $2$-groups). Thus the relation rank is at most seven. For completeness, the seven local quadratic initials are independent. Under the relation--cup pairing their classes are the local Hilbert forms on the span of \eqref{fi:alpha}. The eight forms sum to zero by Hilbert reciprocity. Evaluate them on \[ (-1,\alpha_2),\ (-1,\alpha_1),\ (-1,\alpha_7),\ (-1,\alpha_4),\ (-1,\alpha_5),\ (\alpha_2,\alpha_6),\ (\alpha_1,\alpha_7). \] Their nontrivial local supports, in that order, are \[ \{\mathfrak p_1,v_1\},\ \{\mathfrak p_1,v_2\},\ \{\mathfrak p_2,v_2\},\ \{\mathfrak q_1,v_1\},\ \{\mathfrak q_2,v_2\},\ \{\mathfrak r_1,v_1\},\ \{\mathfrak r_2,v_2\}. \] These are obtained by the elementary Hilbert formulas. For odd $p$, \[ (p^a u,p^b v)_p=(-1)^{ab(p-1)/2}(u/p)^b(v/p)^a. \] For odd $u,v$ at $2$, the exponent is $((u-1)/2)((v-1)/2)+a(v^2-1)/8+b(u^2-1)/8$ modulo two. Together with the residue rules following \eqref{fi:vectors}, this is a finite direct check of the supports. They form a spanning tree on the eight places, so the sole dependence among the eight forms is their sum. In particular, omitting $\mathfrak p_1$ leaves seven independent quadratic initials. Their images are independent in the relation module $R/(R^2[R,\mathfrak F])$. Its dimension is at most seven by the cohomological bound. Pro-$2$ Nakayama now proves normal generation; the same argument applies to any omitted place. \end{proof} Define $G$ as the quotient of $G_S$ obtained by imposing the local quotient $D$ at each dyadic place, $C_2\times C_2$ at each place over $3,5$, and $F_1^4=F_2^4=F_3^4=1$. Equivalently, quotient by the closed normal subgroup generated by the kernels of those local maps and the three fourth powers. Let $B_G$ be its fixed field in $B_S$. All those kernels lie in the Frattini subgroup, so $E\subset B_G$ and $G/D_2G=\mathbb F_2^8$ with coordinates \eqref{fi:vectors}. An explicit normal presentation of $G$ over $\mathfrak F$ has thirty relations: the seven $\rho_\nu$ with $\nu\ne\mathfrak p_1$; for each $j=1,2$ the six words \[ x_j^2,\ [x_j,y_j],\ [x_j,z_j],\ [[y_j,z_j],z_j],\ z_j^4,\ [y_j,z_j]^2; \] the eight squares $t_\nu^2,f_\nu^2$ at $3,5$; and the three $F_k^4$. To justify the dyadic assertion without replacing an unknown relation, let $N_D$ be the kernel of the free three-generator presentation \eqref{fi:D}. Its seven displayed relators span $N_D/(N_D^2[N_D,\mathfrak F_3])$. The true local arithmetic relation $r$ belongs to $N_D$. In its expression in this relation module, the coefficient of $y^2$ is one: the four quadratic initials $x^2,y^2,[x,y],[x,z]$ are independent, whereas the other three relators start in degree at least three, and $r$ has initial \eqref{fi:dyadic-initial}. Replacing $y^2$ by $r$ therefore still spans the relation module and normally generates $N_D$ by pro-$2$ Nakayama. In the local arithmetic group $r=1$, leaving exactly the six additional relations stated above. At odd places the original tame relation together with the two squares makes the two involutions commute and therefore presents $C_2\times C_2$. The omitted dyadic arithmetic relation already follows from the other seven global relations, by Lemma~\ref{fi:seven-relations}. We do not replace its true higher coefficients by a guessed word. \subsection{Finite restricted-Lie calculations}\label{fi:finite-algebra} We specify the finite calculations sufficiently to determine all their defining matrices. Work in $\mathbb F_2\langle X_0,\ldots,X_7\rangle$, with $[a,b]=ab+ba$ and $a^{[2]}=a^2$, and let $L_i$ be the degree-$i$ part of its free restricted Lie algebra. Identify every vector in \eqref{fi:vectors} with the corresponding linear polynomial. Bases are obtained by row reduction of the following lists: \[ L_2=\langle X_i^2,[X_i,X_j]\rangle,\quad L_3=[L_2,L_1],\quad L_4=[L_3,L_1]+[L_2,L_2]+\langle u^2:u\in L_2\rangle. \] For the final square list it suffices to square a basis, since the cross terms are already in $[L_2,L_2]$. Let $q_1,\ldots,q_{21}$ be, in order, the following quadratic initials: the squares of $\iota_1,\iota_2$; the four expressions $[t_\nu,f_\nu]+t_\nu^2$ at $3$ and $[t_\nu,f_\nu]$ at $5$; $y_2^2+[x_2,y_2]+[x_2,z_2]$; the six expressions $x_j^2,[x_j,y_j],[x_j,z_j]$; and the eight $t_\nu^2,f_\nu^2$. Put \begin{equation}\label{fi:relation-lists} \begin{split} R_2&=\langle q_i\rangle,\\ R_3&=\langle[q_i,X_a]:1\le i\le21,0\le a<8\rangle +\langle\omega_1,\omega_2\rangle, \quad\omega_j=[[y_j,z_j],z_j],\\ I_4&=[R_3,L_1]+[R_2,L_2]+\langle q_i^2:1\le i\le21\rangle\\ &\hspace{14mm}+\langle z_1^4,z_2^4,[y_1,z_1]^2,[y_2,z_2]^2, F_1^4,F_2^4,F_3^4\rangle. \end{split} \end{equation} \begin{algorithm}[Row reduction for the restricted-Lie spaces] \label{fi:row-reduction-algorithm} \begin{enumerate}[label=(\roman*)] \item In each degree $r\le4$, order the words $X_{i_1}\cdots X_{i_r}$ lexicographically. Write a polynomial as its coefficient vector in this order, with coefficients in $\mathbb F_2$. There are at most $8^4=4096$ columns. \item Form the rows in the displayed spanning lists. In a product, multiply words by juxtaposition and collect equal words; every coefficient is reduced modulo two. \item Apply Gaussian elimination, beginning with the last nonzero coefficient of each row. The surviving independent rows give a basis; their number is the required dimension or rank. \item To find the relations among an ordered list of rows, append the corresponding identity matrix before elimination and perform the same row operations on both matrices. A zero row in the first matrix gives its linear relation in the second. \end{enumerate} \end{algorithm} The resulting dimensions are \begin{equation}\label{fi:ranks} \begin{array}{c|rrrrrr} &\dim L_1&\dim L_2&\dim L_3&\dim L_4&\dim R_2&\dim R_3\\ \hline &8&36&168&1044&21&142 \end{array} \qquad \dim I_4=963. \end{equation} The $170$-row list defining $R_3$ has $28$ dependencies; the coefficients on its last two rows are zero in every dependency. In addition, \begin{equation}\label{fi:ad-lower} \operatorname{rank}(\operatorname{ad}\iota_1:L_1\to L_2/R_2)=7, \qquad \operatorname{rank}(\operatorname{ad}\iota_1:L_2\to L_3/R_3)=8. \end{equation} These equalities follow by exact finite linear algebra on the displayed lists, using the elimination algorithm above. The same elimination returns a row-dependency vector for each relation. \begin{lemma}\label{fi:lower-quotient} The group $H=G/D_4G$ has order $2^{49}$; its conjugacy class of $\iota_1$ has size $2^{15}$. The prescribed local groups embed in $G$, and already embed in $G/D_3G$; the odd local elementary groups and real involutions embed in $G/D_2G$. \end{lemma} \begin{proof} The $21$ quadratic initials are independent. Modulo $D_4\mathfrak F$, the subgroup $D_2\mathfrak F$ is elementary abelian and its third layer is central. Normal generation therefore says that the entire third relation space is exactly the list $R_3$ in \eqref{fi:relation-lists}: a product of the $21$ quadratic relations can enter degree three only when every quadratic coefficient vanishes. Hence the first three dimensions of $G$ are $8,36-21,168-142=8,15,26$. The local injection assertions reduce to additional small tests on the same data: $x_j,y_j,z_j$ are independent, and $z_j^2,[y_j,z_j]$ are independent modulo $R_2$; each odd pair $t_\nu,f_\nu$ is independent; and $F_k^2\notin R_2$. These tests can be performed by adding the specified vectors to the matrices and repeating the row reduction. They give ranks $3,2,2,1$, respectively. Since $D_3D=1$, the two graded injections imply the injection of the entire dyadic group. The capped Frobenius order is therefore exactly four, as it is at most four by definition. In $H$, the map $h\mapsto[h,\iota_1]$ on the abelian group $D_2H$ is a homomorphism into $D_3H$; its rank is eight by \eqref{fi:ad-lower}. Its kernel has order $2^{15+26-8}=2^{33}$. The degree-one kernel in \eqref{fi:ad-lower} is the line through $\iota_1$, and $\iota_1$ itself centralizes $\iota_1$. Thus its full centralizer has order $2^{34}$ and its class has size $2^{49-34}$. \end{proof} \subsection{Why the tower is infinite}\label{fi:infinite} We include the filtered relation-module argument because the ordinary count of thirty defining relations does not by itself prove infinitude. For a finite $2$-group $\Delta$, write $\Lambda=\mathbb F_2[\Delta]$ with augmentation ideal $\mathfrak I$, and let $h_\Lambda(t)$ be its augmentation Hilbert polynomial. If $\Delta$ were a finite quotient equal to $G$, its local subgroups are embedded by Lemma~\ref{fi:lower-quotient}. The filtered Fox exact sequence for a free presentation with eight generators identifies the module $\mathcal M$ of relation rows with the kernel of $\Lambda^8(-1)\to\mathfrak I$, $(b_i)\mapsto\sum b_i(g_i-1)$. This map is strict and onto: the augmentation ideal is generated by $g_i-1$, and its $r$th power by products of $r$ such elements. Consequently \begin{equation}\label{fi:fox-exact} h_{\mathcal M}=1-(1-8t)h_\Lambda. \end{equation} The relevant Fox identities are the product rule and the fundamental identity $r-1=\sum_i(\partial_i r)(g_i-1)$; the chain rule sends the local relation rows to the global ones. Equivalently, exactness follows from the cellular chain complex of the covering of an eight-petal graph corresponding to the presentation kernel. These are the standard Fox statements used in \eqref{fi:fox-exact}. For a local group $\Gamma$ with $k$ generators put \[ \psi_\Gamma(t)=kt-1+h_{\mathbb F_2[\Gamma]}(t)^{-1}. \] Its global relation-row module has Hilbert polynomial at most $\psi_\Gamma h_\Lambda$ for $00$. Thus $G$ is infinite. \subsection{The degree-four enclosure and the stronger finite quotient} \label{fi:degree-four} The unknown higher coefficients of arithmetic local lifts must be accounted for. In particular, using $I_4$ alone as the degree-four relation space would not be justified. We now describe the containing space $T$ used to control every possible lift. \begin{algorithm}[The degree-four containing space] \label{fi:degree-four-algorithm} \begin{enumerate}[label=(\roman*)] \item For $v=\sum v_iX_i$, take the group word $\prod_{i=0}^7 f_i^{v_i}$ in increasing order. Use the literal squares for the two real relations, the words $ftf^{-1}t^{-p}$ at odd places, and the listed squares and commutators for the quadratic quotient relations. For the remaining dyadic initial, use \[ W=y_2^2[x_2,y_2][x_2,z_2] \] when finding the cubic coefficients. \item Expand each resulting $\rho_i-1$ through degree four, writing its first homogeneous terms as $q_i+t_i+\cdots$. Products are products of words in the free associative algebra. To invert $1+a$, use $1+a+a^2+a^3+a^4$ and discard terms of degree above four. \item Apply Algorithm~\ref{fi:row-reduction-algorithm} to the first $168$ rows of $R_3$. For each of their $28$ relations $k=(k_{ia})$, form \begin{equation}\label{fi:corrections} c(k)=\sum_{i,a}k_{ia}\bigl([t_i,X_a]+X_a[q_i,X_a]\bigr). \end{equation} \item In turn, add each basis element $U$ of $L_2$ to the quadratic coefficient of each of the sixteen local generators other than $F_1,F_2,F_3$. Recompute \eqref{fi:corrections} and retain its change. The required changes in the quadratic expressions are \[ D_v(v^2)[U]=[v,U],\qquad D_v[v,w][U]=[U,w],\qquad D_w[v,w][U]=[v,U]. \] \item Let $T$ be the span of $I_4$, all the vectors $c(k)$, and all these changes. Apply Gaussian elimination to this list and to the images of $[L_3,\iota_j]$ modulo $I_4$ and $T$. \end{enumerate} \end{algorithm} The calculation gives The same Gaussian elimination, now entirely on degree-four words, gives \begin{equation}\label{fi:rank18} \begin{gathered} \dim T=996,\qquad \dim(T/I_4)\text{ before variations}=2,\\ \operatorname{rank}([L_3,\iota_j]\bmod I_4) =\operatorname{rank}([L_3,\iota_j]\bmod T)=18 \qquad(j=1,2). \end{gathered} \end{equation} There are precisely $16\cdot36=576$ perturbation directions. The definition of $T$ and the row-reduction rule specify a reproducible finite proof of \eqref{fi:rank18}; no floating-point arithmetic or unlisted arithmetic relator coefficients enter it. \begin{lemma}\label{fi:enclosure} If $R_4$ is the true degree-four relation space of the arithmetic group $G$, then $I_4\subseteq R_4\subseteq T$. In particular $\operatorname{ad}\iota_1:\operatorname{gr}_3G\to \operatorname{gr}_4G$ has rank $18$. \end{lemma} \begin{proof} The first inclusion follows from squares, brackets, and the seven explicit quartic relators. Work modulo $D_5\mathfrak F$ and $I_4$. The quadratic relations commute and square to one; their commutators with the eight free generators, and the two cubic relations, are central involutions. Indeed, the possible obstructions are $[R_2,L_2]$, $[R_3,L_1]$, and squares of $R_2$, all included in $I_4$. This accounts also for arbitrary conjugating words, not just the displayed conjugations. Independence of the $q_i$ forces their coefficients to vanish in any normal-relation product of degree four. The remaining coefficients give a dependency among the $170$ third degree rows. Every such dependency has zero cubic-relation coordinates. For a relation with initial terms $1+q+t+\cdots$, direct multiplication in our commutator convention gives degree-four coefficient $[t,X]+X[q,X]$ for its commutator with $1+X$. Thus every surviving correction has the form \eqref{fi:corrections}. Any true lift of a local generator differs from its canonical lift by an element of $D_2\mathfrak F$. The cubic part of a quadratic relator depends affinely on the sixteen possible quadratic lift coefficients: two quadratic perturbations first interact in degree four, and cubic lift coefficients cancel because the exponent sum of each generator in each quadratic relation is even. The derivative rules above therefore exhaust the changes in \eqref{fi:corrections}. There remains the dyadic arithmetic relation, whose full word has not been specified. Both it and $W$ belong to the true normal relation subgroup, and their quadratic initials agree. Their quotient is therefore in that subgroup intersected with $D_3\mathfrak F$; their cubic coefficients differ by an element of $R_3$. Replacing one by the other in \eqref{fi:corrections} changes it only by $[R_3,L_1]$, already in $I_4$. This justifies the substitution for this purpose without claiming a replacement of the full presentation. We have now accounted for every possible degree-four relation, proving the second inclusion. The equal endpoint ranks in \eqref{fi:rank18} give rank $18$ modulo the intermediate space $R_4$ as well. \end{proof} \begin{proposition}\label{fi:quotient67} There is a quotient $Q$ of $G$ of order $2^{67}$, mapping onto $H=G/D_4G$, in which the conjugacy class of $\iota_1$ has at least $2^{33}$ elements. \end{proposition} \begin{proof} The group $D_4G/D_5G$ is central elementary abelian in $G/D_5G$. Choose eighteen vectors in $[L_3,\iota_1]$ independent modulo $T$. Extend them to a basis of $L_4/T$, and project onto their eighteen coordinates. This linear map vanishes on $R_4\subseteq T$ and is an isomorphism on the specified adjoint image. Quotient the central fourth layer by its kernel. The result $Q$ has order $|H|2^{18}=2^{67}$. Let $C_0=C_H(\iota_1)$ and let $C=C_Q(\iota_1)$. Although $Q\to H$ is onto, its restriction $C\to C_0$ need not be onto. In fact $D_3H\subset C_0$, and for $u\in D_3H$ the commutator of a lift with $\iota_1$ is the rank-eighteen adjoint map just retained. Changing the lift by the central kernel leaves this commutator unchanged. Hence $[C_0:\operatorname{image}C]\ge2^{18}$. Because the entire kernel is central, $|C|=2^{18}|\operatorname{image}C|$, and thus \[ [Q:C]=[H:C_0]\,[C_0:\operatorname{image}C] \ge2^{15}2^{18}=2^{33}. \] This argument explicitly uses the obstruction to lifting a centralizer; it does not assume centralizer surjectivity. \end{proof} \subsection{Fields, signature, discriminant, and prime densities} \begin{theorem}\label{fi:fields} For every power of two $d\ge2^{67}$ there is a finite Galois extension $K/B$ inside $B_G$ with $[K:B]=d$ whose Galois group maps onto $Q$. For any such field choose the complex conjugation $\iota$ above $v_1$ and set $F=K^{\langle\iota\rangle}$. Then $[F:\mathbb Q]=d$, $K/F$ is quadratic and unramified at every finite place, $K$ is totally imaginary, and the signature $(b,c)$ of $F$ satisfies \begin{equation}\label{fi:signature} b\ge1,\qquad \theta_*:=\frac{17179869183}{34359738368} =\frac12-2^{-35}\le\frac cd<\frac12. \end{equation} Both root discriminants equal \begin{equation}\label{fi:lambda} \lambda=2^{9/4}\sqrt{3615},\qquad \ell=\log\lambda=\tfrac94\log2+\tfrac12\log3615. \end{equation} Every prime of $F$ above a selected prime of $B$ splits in $K$. Their absolute types and counts are \begin{equation}\label{fi:prime-densities} \begin{array}{c|c|c|c} \text{selected rational prime or branch}&(e,f)&N\mathfrak p &\#\{\mathfrak p\subset F\}/d\\\hline 2&(8,4)&2^4&1/32\\ 3&(2,2)&3^2&1/4\\ 5&(2,2)&5^2&1/4\\ 29\text{ (both branches)}&(1,4)&29^4&1/4\\ 41\text{ (the branch }\mathfrak t_3\text{ only)}&(1,4)&41^4&1/8 \end{array} \end{equation} \end{theorem} \begin{proof} Let $V=\ker(G\to Q)$. Since $G$ is infinite and finitely generated, it has open normal subgroups $V'\subset V$ of arbitrarily large index. For a desired dyadic $d\ge2^{67}$ choose $[G:V']\ge d$. A normal subgroup of a finite $2$-group has ambient-normal subgroups of every intermediate power-of-two order: its intersection with the center contains an involution, and induction applies after quotienting by that central subgroup. Apply this to $V/V'$ to get an open normal $U\subset V$ with $[G:U]=d$. Its fixed field is the required $K$. In particular, this is a statement for every sufficiently large dyadic degree, not just a sparse sequence chosen after the analytic estimates. Since $i=\sqrt{\alpha_0}\in K$, the field $K$ is totally imaginary. Let $J=\operatorname{Gal}(K/B)$ and let $N_\iota$ be the size of the conjugacy class of $\iota$. Its image in $Q$ gives $N_\iota\ge2^{33}$. An embedding of $F$ over $v_1$ is real exactly when its corresponding element of $J$ centralizes $\iota$; quotienting by $\langle\iota\rangle$ gives $b=|C_J(\iota)|/2$. There are no real embeddings over $v_2$, because the two real involutions have different elementary images in \eqref{fi:vectors}, and conjugation preserves elementary images. Since $d=|J|$, \[ \frac bd=\frac1{2N_\iota},\qquad \frac cd=\frac12-\frac1{4N_\iota}, \] which proves \eqref{fi:signature} and also $b\ge1$. The elementary image of neither real involution lies in the span of the local vectors of any selected finite place. This is a direct test on \eqref{fi:vectors}. Since the local decomposition groups are generated by these vectors' lifts and the local groups embed, no conjugate of $\iota$ belongs to any such decomposition group. Its intersection with $\operatorname{Gal}(K/F)=\langle\iota\rangle$ is therefore trivial. This proves splitting in $K/F$ and preserves the prescribed local types. Outside $S$, $K/B$ is unramified by its definition; at $S$, splitting in $K/F$ also kills inertia. Thus $K/F$ is unramified at every finite prime and $|\Delta_K|=|\Delta_F|^2$. For the common root discriminant, the quadratic base contributes $241^{1/2}$. Each dyadic branch has relative different exponent divided by ramification index $9/4$, so the two branches together contribute $2^{9/4}$ to the absolute root discriminant. At $3,5$ the ramification is tame of index two, hence the two branches contribute $3^{1/2},5^{1/2}$. The capped primes are unramified and no other prime ramifies. This proves \eqref{fi:lambda}. For a rational prime having both degree-one branches in $B$, the sum of $ef$ over primes of $F$ is $d$, giving $d/(ef)$ selected primes. For the single selected branch above $41$ it is $d/2$, giving $d/8$. The norms are $p^f$. This proves \eqref{fi:prime-densities}; no uniform type for the unselected branch $\mathfrak t_4$ is asserted. \end{proof} \subsection{The exact analytic input needed later}\label{fi:analytic-interface} For the fields just constructed write $L_F(s)=\zeta_K(s)/\zeta_F(s)$, with its value at one defined as the ratio of residues. The analytic class-number formula gives \[ L_F(1)=2^c\pi^{d-c}\lambda^{-d/2} \frac{h_K\operatorname{Reg}_K w_F} {h_F\operatorname{Reg}_F w_K}>0, \qquad w_F=2. \] Let $h_{\rm rel}=h_K/h_F$, let $h_{\rm cap}$ be the order of the kernel of $\operatorname{Cl}(F)\to\operatorname{Cl}(K)$, and let $\operatorname{Reg}^1$ be the covolume of the relative-unit logarithms $\tfrac12(\log|\phi_j^+\varepsilon|-\log|\phi_j^-\varepsilon|)$. The normalization used in the geometric transfer is \begin{equation}\label{fi:mass} \frac{w_K}{h_{\rm rel}h_{\rm cap}\operatorname{Reg}^1} =\frac{2^{d-1}\pi^{b+c}}{\lambda^{d/2}L_F(1)}. \end{equation} Its full unit/cohomology and regulator proof is given in Proposition~\ref{tr:mass}; the factors of two in \eqref{fi:mass} correspond to the logarithmic normalization just stated. \begin{proposition}[Quantifier required of the arithmetic bound] \label{fi:arithmetic-input} The numerical arithmetic result used in this paper is the following uniform statement: there is $d_0$ such that \emph{every} admissible $K/B$ in the above tower with $[K:B]=d\ge d_0$ satisfies \begin{equation}\label{fi:C-input} \frac1d\log L_F(1)2^{d-1}\pi^{(1-c/d)d}\lambda^{-d/2}e^{-dC}. \] \end{proposition} The uniform quantifier is essential: a bound along a selected sequence would not suffice for choosing the degree from a prescribed pair of cardinalities in the later all-scale construction. The arithmetic proof obtains the displayed eventual bound by ruling out every infinite sequence of violating fields, using a convergent subsequence of normalized prime counts and a strict limiting inequality. This compactness argument does not require a nested tower and does not produce an effective numerical threshold. The stronger quotient $Q$ only restricts the original family; therefore a uniform estimate for the original family remains valid for all fields used here. \subsection{Explicit comparison fields and residue-degree detectors} \label{ar:comparison-construction} We now construct the comparison field used in the analytic estimate. The choice of forms and primitive radicals follows Naslund~\cite{Naslund}; the rational witness checks and the proof organisation below are explicit. This also specifies the fields underlying every one of its $576$ auxiliary $L$-functions. The construction uses the group $G$, its Zassenhaus filtration, and the eight elements $\alpha_0,\ldots,\alpha_7$ of Section~\ref{fi:section}; put $E_0=E=B(\sqrt{\alpha_0},\ldots, \sqrt{\alpha_7})$. In particular $[E_0:\mathbb Q]=512$. Every field $K$ used here retains the quotient $G/D_4G$ and therefore contains the fixed field of $D_3G$. Write $V=\mathbb F_2^8$, and interpret a string $r_0\cdots r_7$ in the order of the Kummer basis in \eqref{fi:alpha}. Set $\alpha_r=\prod_{i=0}^7\alpha_i^{r_i}$ and $\chi_r(g)=(-1)^{r\cdot\overline g}$. The $36$ coordinates on the free degree-two restricted Lie algebra are, in order, the eight squares $e_i^{[2]}$ and the $28$ brackets $[e_i,e_j]$, $i41^4$ the required inequality is automatic. The exact enumeration can be shortened: test all degree-one free primes below $1681$, and all vector-zero degree-one primes up to $41^4$. For larger norms a nonzero elementary vector already gives a sufficient floor of at least two. For split rational primes the vector-zero screen is exactly $p\equiv1,49\pmod{120}$ together with the four residue tests for $\alpha_1,\alpha_2,\alpha_4,\alpha_6$; this follows from $\alpha_0=-1$, $\alpha_2\alpha_3=2$, $\alpha_4\alpha_5=-3$, $\alpha_6\alpha_7=-5$ and $\mathrm N(\alpha_1)=-1$. The enumeration gives $248$ small degree-one free primes, finds $658$ degree-one vector-zero primes up to $41^4$, all detected, and independently verifies the thirteen inert primes above. The least vector-zero norm is $65809$. The same residue tests define the full finite prime sums used in Section~\ref{ar:section}. \subsection{Rational data for the comparison fields}\label{sc:radical-tables} Put $D=\sqrt{241}$. For each row $o$ in the next table set $a_o=\alpha_{r_{1,o}}$, $b_o=\alpha_{r_{2,o}}$, $A_o^2=a_o$ and $B_o^2=b_o$, with the convention for $\alpha_r$ in Section~\ref{ar:comparison-construction}. The element defining the quadratic extension is \[ \beta_o=c_{o,0}+c_{o,1}D+(c_{o,2}+c_{o,3}D)A_o +(c_{o,4}+c_{o,5}D)B_o+(c_{o,6}+c_{o,7}D)A_oB_o. \] All the entries below are rational numbers. The label $o$ is used only to match the field, its quadratic form, and its prime detector. \begin{center}\small \begin{tabular}{rccc} \toprule $o$&$r_{1,o}$&$r_{2,o}$&$\varrho_o$\\\midrule $0$&$00000100$&$00010000$&$(1,1)$\\ $3$&$00000101$&$01001000$&$(1,1)$\\ $4$&$00000111$&$01000001$&$(0,1)$\\ $5$&$00001000$&$10100000$&$(1,1)$\\ $6$&$00001010$&$11000001$&$(1,1)$\\ $7$&$00001010$&$11000100$&$(1,1)$\\ $9$&$00010001$&$11000001$&$(0,1)$\\ $10$&$00010011$&$01001100$&$(1,1)$\\ $12$&$00100010$&$11001100$&$(1,1)$\\ $13$&$00100100$&$01001011$&$(1,1)$\\ $14$&$00110011$&$01010001$&$(0,1)$\\ $15$&$01000010$&$10100010$&$(1,0)$\\ $17$&$01000110$&$10001000$&$(1,1)$\\ $19$&$01001100$&$10000010$&$(1,1)$\\ $20$&$01001100$&$10010001$&$(1,1)$\\ $22$&$01001101$&$10000100$&$(0,1)$\\ $23$&$01011111$&$10001011$&$(0,1)$\\ $24$&$01011111$&$10011000$&$(0,1)$\\ \bottomrule\end{tabular}\end{center} \begin{center}\small\setlength{\tabcolsep}{5pt} \begin{tabular}{rrrrr} \toprule $o$&$c_{o,0}$&$c_{o,1}$&$c_{o,2}$&$c_{o,3}$\\\midrule $0$&$-8$&$-\frac{1}{2}$&$\frac{5}{2}$&$0$\\ $3$&$-\frac{65217}{2}$&$-\frac{4201}{2}$&$-\frac{323}{2}$&$-\frac{21}{2}$\\ $4$&$-\frac{74249}{2}$&$\frac{4783}{2}$&$0$&$0$\\ $5$&$-8$&$\frac{1}{2}$&$\frac{5}{2}$&$0$\\ $6$&$\frac{112303}{2}$&$-3617$&$293$&$-17$\\ $7$&$-\frac{65217}{2}$&$\frac{4201}{2}$&$-\frac{323}{2}$&$\frac{21}{2}$\\ $9$&$-\frac{31}{2}$&$1$&$0$&$0$\\ $12$&$-\frac{426347}{4}$&$\frac{27459}{4}$&$231045$&$-14883$\\ $13$&$\frac{10995005}{4}$&$\frac{708183}{4}$&$0$&$0$\\ $14$&$\frac{6753}{4}$&$-\frac{435}{4}$&$0$&$0$\\ $15$&$-\frac{3}{2}$&$0$&$83000$&$\frac{10693}{2}$\\ $17$&$-39$&$\frac{5}{2}$&$-\frac{1925}{2}$&$-62$\\ $19$&$-\frac{7886559}{4}$&$\frac{508001}{4}$&$-\frac{144775}{4}$&$-\frac{9375}{4}$\\ $20$&$-\frac{279699}{4}$&$\frac{18017}{4}$&$-\frac{41}{4}$&$-\frac{1}{4}$\\ $22$&$-\frac{15}{4}$&$-\frac{1}{4}$&$0$&$0$\\ $24$&$-\frac{91332223}{4}$&$\frac{5883225}{4}$&$0$&$0$\\ \bottomrule\end{tabular}\end{center} \begin{center}\small\setlength{\tabcolsep}{5pt} \begin{tabular}{rrrrr} \toprule $o$&$c_{o,4}$&$c_{o,5}$&$c_{o,6}$&$c_{o,7}$\\\midrule $0$&$0$&$0$&$0$&$0$\\ $3$&$0$&$0$&$0$&$0$\\ $4$&$-\frac{699}{2}$&$-\frac{49}{2}$&$0$&$0$\\ $5$&$0$&$0$&$0$&$0$\\ $6$&$-\frac{1661}{2}$&$-65$&$-47767$&$-3077$\\ $7$&$0$&$0$&$0$&$0$\\ $9$&$\frac{31}{2}$&$1$&$0$&$0$\\ $12$&$-\frac{236491}{4}$&$-\frac{15209}{4}$&$-4341$&$-293$\\ $13$&$-\frac{266484672787}{4}$&$-\frac{17165786781}{4}$&$0$&$0$\\ $14$&$0$&$0$&$-\frac{761}{40}$&$-\frac{49}{40}$\\ $15$&$0$&$0$&$0$&$0$\\ $17$&$0$&$0$&$0$&$0$\\ $19$&$\frac{92871}{2}$&$-\frac{6079}{2}$&$-414500$&$-26700$\\ $20$&$-4797$&$309$&$15$&$1$\\ $22$&$0$&$0$&$\frac{2251}{12}$&$\frac{145}{12}$\\ $24$&$0$&$0$&$-\frac{1321960925}{24}$&$-\frac{85154989}{24}$\\ \bottomrule\end{tabular}\end{center} The remaining two coefficient vectors are as follows: \begin{align*} (c_{10,0},c_{10,1})&=\left(-\frac{2892966757041403369}{4},-\frac{186352370650554249}{4}\right),\\ (c_{10,2},c_{10,3})&=\left(17869643568757778063,1151084931623631393\right),\\ (c_{10,4},c_{10,5})&=\left(-\frac{794087129829920621481}{2},-25575824330978012432\right),\\ (c_{10,6},c_{10,7})&=\left(\frac{39240143878073548164507}{4},\frac{2527679869998307278733}{4}\right),\\ (c_{23,0},\ldots,c_{23,7})&=(C_0,-C_1/2,C_2/2,-C_3/2,C_4/2,-C_5,C_6/10,-C_7/2). \end{align*} The positive integers $C_j$ are given by their base-$10^8$ digits. A row $(d_0,\ldots,d_k)$ denotes $\sum_{i=0}^k d_i10^{8(k-i)}$; leading zeroes inside a digit are only typographical. Within each block labelled $j$, read the digits from left to right and then from top to bottom. \begin{longtable}{r l} \toprule $j$ & Base-$10^8$ digits of $C_j$\\\midrule\endhead 0 & $1,\ 74510253,\ 60717598,\ 75908937,\ 16911066$\\ & $39456000,\ 90234466,\ 55894358,\ 87867505,\ 86638758$\\ & $17982229,\ 76772239,\ 12695036,\ 95162104,\ 21143576$\\ & $27456833,\ 18370036,\ 34063252,\ 02922730,\ 86767929$\\ & $09860055,\ 94974870,\ 20362575$\\ \addlinespace 1 & $22482387,\ 24719038,\ 17215355,\ 51947364,\ 98058428$\\ & $56713531,\ 00247991,\ 74208515,\ 23159475,\ 64610967$\\ & $25772018,\ 46869563,\ 15956800,\ 44527910,\ 70294152$\\ & $90055354,\ 13384992,\ 95080980,\ 51277486,\ 24137160$\\ & $73881828,\ 57274345$\\ \addlinespace 2 & $91,\ 59029286,\ 29453336,\ 12211730,\ 48598714$\\ & $32070187,\ 25846677,\ 83448236,\ 86981542,\ 07191611$\\ & $70993215,\ 14394504,\ 18874875,\ 27262368,\ 50555749$\\ & $35646452,\ 64442163,\ 96813687,\ 18619874,\ 70246292$\\ & $81871762,\ 89008821$\\ \addlinespace 3 & $5,\ 89984940,\ 61658962,\ 38887151,\ 81555788$\\ & $82412309,\ 92506419,\ 51981627,\ 39424367,\ 11802506$\\ & $89500430,\ 98702276,\ 76479766,\ 26671906,\ 79133801$\\ & $64568107,\ 79889859,\ 76753457,\ 48667474,\ 40359780$\\ & $36314435,\ 22186127$\\ \addlinespace 4 & $19815307,\ 97740989,\ 17041772,\ 62334033,\ 89720818$\\ & $97955614,\ 95504304,\ 79925327,\ 69857602,\ 62420292$\\ & $04820133,\ 45180980,\ 41479110,\ 28140821,\ 90999693$\\ & $94939042,\ 16228504,\ 98817114,\ 66783624,\ 56635009$\\ & $10103644,\ 83583425$\\ \addlinespace 5 & $638208,\ 09689109,\ 13499645,\ 91657229,\ 94433351$\\ & $90643553,\ 31608707,\ 34432674,\ 77108328,\ 18070756$\\ & $02404113,\ 34330615,\ 13264175,\ 09610688,\ 14906045$\\ & $01313072,\ 48188240,\ 73197257,\ 80643329,\ 93753240$\\ & $95985823,\ 70173255$\\ \addlinespace 6 & $25,\ 99975966,\ 03372215,\ 40103718,\ 02733514$\\ & $09760623,\ 78821630,\ 87248335,\ 90659344,\ 63093977$\\ & $86652900,\ 77092944,\ 26508322,\ 62532508,\ 39111414$\\ & $99761832,\ 50753640,\ 98407725,\ 18653042,\ 99398517$\\ & $17387267,\ 63638587$\\ \addlinespace 7 & $33495834,\ 94007047,\ 33540442,\ 31150225,\ 90196915$\\ & $32201182,\ 79844008,\ 86119143,\ 50511082,\ 26120257$\\ & $23259866,\ 66480608,\ 17335618,\ 95066745,\ 20312236$\\ & $13669879,\ 60461526,\ 77155044,\ 41490265,\ 17835603$\\ & $27026885$\\ \addlinespace \bottomrule\end{longtable} The positive sign of every norm and the three exponents of $2,3,5$ are listed in Section~\ref{ar:comparison-construction}. They follow by successively applying \[ (x+y\sqrt d)(x'+y'\sqrt d)=(xx'+dyy')+(xy'+yx')\sqrt d, \qquad N(x+y\sqrt d)=x^2-dy^2 \] with $d=b_o,a_o,241$. The same formulas give a short recursive check of the Galois identities. For $x+y\sqrt d$, $y\ne0$, first take either square root $n$ of $x^2-dy^2$, then a nonzero square root $r$ of $(x+n)/2$; the candidate square root is $r+(y/(2r))\sqrt d$. When $y=0$, use a square root of $x$ in the smaller field or $\sqrt d$ times a square root of $x/d$. At the rational level, numerator and denominator must both be squares. Trying the two signs at each step exhausts all possibilities. Applied to $\sigma(\beta_o)/\beta_o$, this gives the three identities \eqref{ar:D1-D2} for every row. Changing the sign of a returned square root does not change $u\sigma(u)$. \subsection{Intrinsic coefficients and prime sums}\label{sc:intrinsic-arithmetic} For each quadratic extension $A'/A$ in \eqref{ar:quartic-radicals}--\eqref{ar:octic-radicals}, define \[ D_T(p,X)=\prod_{\mathfrak q\mid p\text{ in }T}(1-X^{f(\mathfrak q/p)}), \qquad P_p(X)=\frac{D_{A'}(p,X)}{D_A(p,X)}. \] The quotient $P_p$ is a polynomial with integer coefficients. Its reciprocal coefficients are uniquely specified by \[ a_1=1,\qquad a_{p^k}=-\sum_{j=1}^{\min(k,\deg P_p)}[X^j]P_p(X)\,a_{p^{k-j}}, \qquad a_n=\prod_{p^k\Vert n}a_{p^k}. \] Thus the rational radicals above, ordinary ideal factorisation and this recursion determine each coefficient of all $576$ factors. Their conductors are $|\Delta_{A'}|/|\Delta_A|$, and their gamma factors follow from the signatures. The analogous definitions apply to the $255$ quadratic factors of $\zeta_{E_0}/\zeta_B$. The prime correction can equally be defined without bins. Let $X_C=1125899906842000$ and let $\mathcal P_2$ be the set in \eqref{ar:floor-definition-explicit}, with the stated cutoff $X_C$. For $s>1$ put \[ \mathcal C_2(s)=\sum_{\mathfrak p\in\mathcal P_2} \frac12\operatorname{arctanh}((N\mathfrak p)^{-s}),\qquad \mathcal C_2=\sum_i c_i\mathcal C_2(\sigma_i). \] Every summation condition is an exact residue-field test specified by \eqref{ar:pattern-definition}. If a norm $N$ is assigned an interval $[L,U]$ and $\tau(x)=\sum_i c_i\operatorname{arctanh}(x^{-\sigma_i})/2$ is decreasing there, then $\tau(N)\ge\tau(U)$. Consequently replacing any enlarged-bin lower bound by the exact sum $\mathcal C_2$ preserves the direction of the relative-zeta upper bound. This comparison applies to the signed kernel as a whole. For a factor with gamma data $(a,\mu_j,\kappa)$, conductor $Q$, $t=\kappa/\sqrt Q$ and cutoff $N$, its intrinsic finite value is \[ F_N(s)=\frac{t^s}{\prod_j\Gamma(s/a+\mu_j)} \sum_{n=1}^{N}a_n\left\{I_{s/a}((tn)^a)+I_{(1-s)/a}((tn)^a)\right\}. \] The series for $I_b$ and their remainder bounds in Section~\ref{ar:kd-section}, together with the tail estimate \eqref{ar:kd-Dirichlet-tail}, give two-sided rational enclosures for this expression and for $L(s)$. This formulation specifies the finite sum through its coefficients and kernel, independently of a choice of floating-point accumulation or stored moments. \section{The arithmetic constant}\label{ar:section} This section proves the arithmetic input \[ C_{\mathrm{ar}}=0.04226263259973019243081 \] used in the incidence construction. It bounds a \emph{relative zeta value}. It is neither a bound for an individual Dedekind zeta value at its pole nor a direct bound for an unnormalised class number or regulator. All decimal constants defining the witness are exact rational numbers. Numerical optimisation is used only to find the witness; none of its termination messages is used in the proof. \subsection{Field input and the quantity being bounded}\label{ar:field-input} We use the following data from the tower construction of Theorem~\ref{fi:fields}, following the arithmetic framework of \cite{Naslund}. The base is $B=\mathbb Q(\sqrt{241})$. The fields $K/B$ under consideration are finite Galois extensions in the fixed pro-$2$ tower, of unbounded degree, containing the fixed comparison field $E_*$ of degree $8192$ over $\mathbb Q$ constructed in Section~\ref{ar:comparison-construction}. They are totally imaginary. If $\iota$ is the specified complex conjugation, then $F=K^{\langle\iota\rangle}$, $[F:\mathbb Q]=d$, and $[K:\mathbb Q]=2d$. The quadratic extension $K/F$ is unramified at finite places. Both fields have root discriminant \begin{equation}\label{ar:discriminant} \lambda=2^{9/4}\sqrt{3615},\qquad \ell=\log\lambda=\tfrac94\log2+\tfrac12(\log3+\log5+\log241). \end{equation} Write $(r_1,r_2)$ for the signature of $F$ and $\theta=r_2/d$; thus $r_1/d=1-2\theta$ and $0\leq\theta\leq1/2$. There are nine selected primes of $B$: the two primes over each of $2,3,5$, two specified primes of norm $29$, and one specified prime of norm $41$. Their types $(e,f)$ in $K/B$ are respectively $(8,4)$, $(2,2)$, $(2,2)$, $(1,4)$ and $(1,4)$. Every other prime of $B$ is unramified in $K/B$. The residue-degree floors used below divide the actual residue degrees in every field of the family. These field-theoretic inputs are established in Section~\ref{fi:section}. Put $L_F(s)=\zeta_K(s)/\zeta_F(s)$. The class-number formula gives the exact identity \begin{equation}\label{ar:class-number} L_F(1)=2^{r_2}\pi^{d-r_2}\lambda^{-d/2} \frac{h_K\operatorname{Reg}_K w_F} {h_F\operatorname{Reg}_F w_K}>0, \end{equation} where $w_A$ is the number of roots of unity in $A$. \begin{proposition}\label{ar:arithmetic-ceiling} For all sufficiently large fields in this family, \begin{equation}\label{ar:ceiling-value} \frac1d\log L_F(1)1$. Consequently, for $0<\varepsilon\leq1$, \begin{align} \frac1d\log L_F(1) &\leq\frac{\log\zeta_K(1+\varepsilon)}{2d} +\Upsilon(\varepsilon,\theta),\label{ar:shift}\\ \Upsilon(\varepsilon,\theta) &=\frac\varepsilon2(\ell-\log\pi)-\varepsilon\theta\log2 +(1-2\theta)\log\Gamma(1+\varepsilon/2) +\theta\log\Gamma(1+\varepsilon). \end{align} The last expression tends uniformly to zero for $0\leq\theta\leq1/2$. For a prime power $q$ and a sequence of fields of increasing degree, pass to a subsequence on which $\beta_q=\lim N_q(K)/(2d)$ exists for every $q$. This is possible by a diagonal argument, since $N_{p^f}(K)/(2d)\leq1/f$. Put \begin{equation}\label{ar:g-w} g_s(q)=-\log(1-q^{-s}),\qquad w(q)=2\log q\sum_{m\geq1}\frac1{q^m+1},\qquad Z(s)=\sum_q\beta_qg_s(q). \end{equation} For fixed $s>1$, dominated convergence identifies $Z(s)$ with the limit of $(2d)^{-1}\log\zeta_K(s)$. The unconditional basic inequality of Tsfasman--Vl\u{a}du\c{t}~\cite{TsfasmanVladut2002}, applied with $\log|\Delta_K|^{1/2}=d\ell$ and $K$ totally imaginary, is \begin{equation}\label{ar:TV} \sum_q\beta_q w(q)\leq \frac{\ell-\gamma-\log(4\pi)}2=2\kappa_\infty. \end{equation} Here $\gamma$ is Euler's constant. The convergence of $Z(1)$ follows, for example, from $g_1(q)\leq3w(q)/(2\log2)$. First passing to the field limit in \eqref{ar:shift} for a fixed $\varepsilon$, and then letting $\varepsilon\downarrow0$, shows that every subsequential lower bound for $d^{-1}\log L_F(1)$ is at most $Z(1)$. This order of limits is essential. The selected local types give \begin{equation}\label{ar:selected} Z_{\rm sel}(s)=\tfrac1{32}g_s(16)+\tfrac14g_s(9)+\tfrac14g_s(25) +\tfrac14g_s(29^4)+\tfrac18g_s(41^4). \end{equation} Define $T_{\rm sel}$ by replacing each $g_s$ in this expression by $w$, and let $B_r=2\kappa_\infty-T_{\rm sel}$. For a free prime $\mathfrak p$ of $B$, let $f^0_{\mathfrak p}\mid f_K(\mathfrak p)$ be its certified floor and suppose $(\mathrm N\mathfrak p)^{f^0_{\mathfrak p}}\geq q_0=41^4$. The associated floor sum is \[ Y_{\rm fl}(s)=\sum_{\mathfrak p\ \text{free}} \frac{g_s((\mathrm N\mathfrak p)^{f^0_{\mathfrak p}})}{2f^0_{\mathfrak p}}. \] \begin{lemma}[Signed comparison]\label{ar:signed-comparison} Let $\mathsf K(q)=\sum_i c_i g_{\sigma_i}(q)$, where $1<\sigma_i<1.01$, $c_i\in\mathbb R$, and $b\geq0$. Suppose, for all $q\geq q_0$, \begin{equation}\label{ar:kernel-conditions} \mathsf K(q)\geq g_1(q)-bw(q),\qquad \mathsf K(q)\geq0,\qquad \mathsf K'(q)\leq0. \end{equation} Then the eventual relative-zeta ceiling is every rational number strictly larger than \begin{equation}\label{ar:objective} Z_{\rm sel}(1)+\sum_i c_iY_{\rm fl}(\sigma_i)+bB_r. \end{equation} \end{lemma} \begin{proof} A free prime of norm $N$ and actual residue degree $f$ contributes $\mathsf K(N^f)/(2f)$. Nonnegativity and monotonicity give $\mathsf K(N^f)/(2f)\leq\mathsf K(N^{f^0})/(2f^0)$. Summing, then applying Fatou's lemma to the nonnegative kernel, gives $\sum_q\beta_q^{\rm free}\mathsf K(q)\leq\sum_i c_iY_{\rm fl}(\sigma_i)$. Interchanging the finite signed sum with the floor sum is valid because $|\mathsf K|\leq\sum_i|c_i|g_{\sigma_i}$, whose floor sums converge. Apply the first inequality in \eqref{ar:kernel-conditions} and the remaining budget $\sum\beta_q^{\rm free}w(q)\leq B_r$ from \eqref{ar:TV}. This bounds $Z(1)$ by \eqref{ar:objective}. A hypothetical sequence of exceptions of increasing degree contradicts the preceding limit argument. \end{proof} \subsection{The rational kernel and its global verification}\label{ar:kernel} The witness has $36$ nonzero coefficients, selected from a grid of $71$ abscissae. Table~\ref{ar:kernel-table} specifies all of them exactly by $\sigma_i=1+1/n_i$, $c_i=A_i/10^{45}$. Also \begin{equation}\label{ar:kernel-b} b=\frac{112270274070135453110396603100681372994522}{10^{45}}. \end{equation} In particular $\sum_i|c_i|<3000000$. This last number is a property of the witness, not a restriction on all admissible kernels. \begin{table}[p] \centering\scriptsize \caption{Exact signed kernel: $\sigma=1+1/n$, $c=A/10^{45}$.} \label{ar:kernel-table} \begin{tabular}{r r} $n$ & $A$\\\hline 50000 & 5946470685733009828965940757034543337754041927 \\ 46739 & -45166657905922747968921708062913488702245586149 \\ 43691 & 127084102048074452625041278806433514634006630511 \\ 40842 & -137293173282986199475430720762233828387795664061 \\ 33361 & 210776912025226374525398623731477264052374104277 \\ 29151 & -426953231727198846981824595906055369346107852903 \\ 25472 & 460701764287886209947399471819820891553679915500 \\ 19450 & -899185684252966364971024651013439896144640188994 \\ 16996 & 1387831044284906909979741573472722541579947236318 \\ 13882 & -1502379641586598700820639903635130094610535867225 \\ 11340 & 1652905611328496553138767707716725339685314749416 \\ 8659 & -3166500043655977290019012978379165789535242025977 \\ 7566 & 3672537631644523588764562693655479082075333995330 \\ 5777 & -3043454406107411306990740713909430008326866619476 \\ 4411 & 5002046755015031547562388836159262517477306458881 \\ 3854 & -381863355688516570704111777920138477636774 \\ 3368 & -16200194533448661474438768629192832987384268242585 \\ 2943 & 22697787623332014355304850430278768962645583046129 \\ 2248 & -29880027710674424722863821958732498498273331827071 \\ 1836 & 49468850188032721736372616322944656457907169703407 \\ 1499 & -60882223293988663534267525953997994183154575396766 \\ 1145 & 101233596903595918930464196761819823806386520389946 \\ 935 & -178108835467012795768824517719514498090444787181107 \\ 764 & 246333642053729500396942402714377088519015388480023 \\ 624 & -281307633428907770267508973357966242210820040328588 \\ 510 & 245429857684989000446942965370000583709171695434253 \\ 389 & -246020362605256592340622677229908086307363475695318 \\ 318 & 286251049017181073158354847353193405384490123937748 \\ 260 & -262068723428071040721956094019939525412996574095064 \\ 212 & 222351741642143260494663261466668301271320309420661 \\ 173 & -232843049809678650855210372539318829538798071660459 \\ 151 & 218940340781997199472809078316262512220594679972544 \\ 132 & -133435419194267873226791889907398009929071274648813 \\ 115 & 86573284906192695497138732121014432151352869446345 \\ 108 & -50032097264347342407127955807531256085215217123001 \\ 101 & 8200518951450504399278019918041552184006415396450 \\ \end{tabular} \end{table} Here are full interval-check rules for \eqref{ar:kernel-conditions}. Set $x=\log q$ and \[ \Phi_1=e^x(\mathsf K-g_1+bw)(e^x),\quad \Phi_2=e^x\mathsf K(e^x),\quad \Phi_3=e^x\frac{d}{dx}\mathsf K(e^x). \] Their required signs are respectively $+,+,-$. Use \[ e^xg_s(e^x)=e^{-(s-1)x}\sum_{k\geq0}\frac{e^{-ksx}}{k+1},\quad e^x\frac{d}{dx}g_s(e^x)=-s e^{-(s-1)x}\sum_{k\geq0}e^{-ksx}, \] and $e^xw(e^x)=2x\sum_{m\geq1}e^{-(m-1)x}/(1+e^{-mx})$. Retain $k=0,\ldots,9$ in the first two series. For this particular witness the common lower endpoint satisfies $x>14.85$, so $e^{-sx}<3.56\cdot10^{-7}$. The omitted series and their Taylor coefficients through order three, after summing all signed coefficients, are bounded in absolute value by \[ 2(3000000\cdot1.01+1)\frac{10.11^3}{6}(3.56\cdot10^{-7})^{10} <3.413\cdot10^{-56}<10^{-55}. \] The geometric ratio of successive derivative majorants is less than $2\cdot10^{-6}$; the factor $2$ above covers its sum. Thus a radius $10^{-55}$ on each of the four Taylor coefficients is sufficient. Only $m\leq6$ is retained in $w$, which gives a pointwise lower function for $\Phi_1$ because $b\geq0$; its derivatives are then evaluated exactly as derivatives of this finite lower function. \begin{algorithm}[Verification of the kernel signs] \label{ar:kernel-sign-algorithm} \begin{enumerate}[label=(\roman*)] \item Put $x_*=10^{-6}\lfloor10^6\log q_0\rfloor$ and $X=2000000$. Partition $[x_*,X]$ into intervals of length $1/20$ below $60$; thereafter use length $x/100$ at a left endpoint $x$, shortening the last interval if necessary. \item On an interval with midpoint $a$ and half-length $h$, enclose each of $\Phi_1,\Phi_2,\Phi_3$ by \[ \Phi(a)+\Phi'(a)t+\tfrac12\Phi''(a)t^2 +\tfrac16\Phi'''([a-h,a+h])t^3,\qquad |t|\le h. \] Use the finite series and remainder bounds above for the coefficients. \item If these intervals establish the required signs $+,+,-$, retain the lower bounds for $\Phi_1,\Phi_2,-\Phi_3$. Otherwise divide the interval into two equal parts and repeat the preceding step. \item Take the smallest of the retained lower bounds. Check the three displayed inequalities below to cover $x\ge X$. \end{enumerate} \end{algorithm} The exact rational coefficients establish the signs on all $4428$ resulting cells, with lower bounds \[ \Phi_1>2.9510\cdot10^{-11},\quad \Phi_2>1.4846\cdot10^{-17},\quad -\Phi_3>1.4846\cdot10^{-17}. \] For $x\geq X$ the first abscissa is smallest and $c_1>0$. The following three strictly positive scalar checks cover the entire remaining ray: \begin{align*} c_1&>\sum_{j>1}|c_j|(1+e^{-X})e^{-(\sigma_j-\sigma_1)X},\\ c_1\sigma_1&>\sum_{j>1}|c_j|\sigma_j(1+2e^{-X}) e^{-(\sigma_j-\sigma_1)X},\\ 2bX&>1+3e^{-X}. \end{align*} The first two bound $\Phi_2$ and $-\Phi_3$ after multiplication by $e^{(\sigma_1-1)x}$; the third makes $g_1-bw<0$. This proves the global kernel conditions, including the infinite tail. \subsection{Comparison-field factorisation and prime corrections}\label{ar:factors} Use the exact forms, radicals, twist characters and residue-degree detectors of Section~\ref{ar:comparison-construction}; in particular the integer family labels below refer to Table~\ref{ar:seven-form-table} and the pairs in \eqref{ar:rank-four-pairs}. Let $E_0\subset E_*$ be the genus field of degree $512$. The relevant class-two quotient over $B$ has order $4096$, elementary abelian quotient of dimension eight, and a four-dimensional central character space. Its fifteen nonzero central characters have seven polar forms of rank two and eight of rank four. For polar rank $r$, an irreducible representation with this central character has dimension $2^{r/2}$ and there are $256/2^r$ such representations. One way to see the dimension formula is to pass to the quotient in which the central character has order two: the character vanishes off the inverse image of the radical, and character orthogonality gives $(\dim\pi)^2=2^r$. The degree check is $256+7\cdot64\cdot4+8\cdot16\cdot16=4096$. The regular representation and Artin formalism therefore give the \emph{factor-by-factor} identity \begin{equation}\label{ar:factorization} \frac{\log\zeta_{E_*}(s)}{8192} =\frac1{16}\frac{\log\zeta_{E_0}(s)}{512} +\frac1{4096}\sum_{\nu=1}^{448}\log L_{4,\nu}(s) +\frac1{2048}\sum_{\nu=1}^{128}\log L_{8,\nu}(s). \end{equation} The seven degree-four families are indexed by $10,23,24,19,20,17,7$, each with $64$ twists. The eight degree-eight families are $(19,23),(19,24),(7,10),(17,10),(17,24),(7,24),(7,23),(17,23)$, each with $16$ twists. Every factor here is a quotient $\zeta_{A'}/\zeta_A$ for a specified quadratic extension $A'/A$; hence it is a quadratic Hecke $L$-function with root number $+1$, not an assumed instance of Artin holomorphy. The defining radicals and their exact finite verification are given in Section~\ref{ar:comparison-construction}. Exact local arithmetic in these fields determines their conductors, signatures and all Euler factors. The floor of a free prime starts with its residue degree in $E_*/B$. For the certified set $\mathcal P_4$ it is raised from $2$ to $4$ when needed, and for the detected set $\mathcal P_2$, split in $E_*$, it is raised from $1$ to $2$. These sets are disjoint. Write $Y_*(s)$ for the left side of \eqref{ar:factorization}, and define \begin{align} T_s(N)&=\tfrac12g_s(N)-\tfrac14g_s(N^2) =\tfrac12\operatorname{arctanh}(N^{-s}),\label{ar:T}\\ \Delta_4(s)&=\sum_{\substack{\mathfrak p\in\mathcal P_4\\f_{E_*}(\mathfrak p)=2}} \left(\tfrac14g_s((\mathrm N\mathfrak p)^2) -\tfrac18g_s((\mathrm N\mathfrak p)^4)\right). \end{align} Let $\Delta_{\rm sel}$ be the difference between the selected-prime contribution in $E_*$ and the actual selected-prime contribution in $K$. It is evaluated from the bad Euler factors at $2$, together with $g_s(N^{f_{E_*}})/(2f_{E_*})-g_s(N^4)/8$ at the three capped primes. The primes over $3,5$ give zero difference because their types agree. Termwise comparison of Euler products proves \begin{equation}\label{ar:floor-identity} Y_{\rm fl}(s)=Y_*(s)-\Delta_{\rm sel}(s)-Z_{\rm sel}(s)-\Delta_4(s) -\sum_{\mathfrak p\in\mathcal P_2}T_s(\mathrm N\mathfrak p). \end{equation} In particular the signed coefficients cannot be applied to one-sided upper estimates for $Y_*(\sigma_i)$ separately. The prime sum is defined directly by its residue tests and the endpoint $X_C=1125899906842000$. Put \[ \mathcal C_2(s)=\sum_{\mathfrak p\in\mathcal P_2}T_s(N\mathfrak p), \qquad \tau(N)=\sum_i c_iT_{\sigma_i}(N). \] Exact enumeration therefore determines every term, without a choice of binning. Grouping norms into intervals can accelerate the evaluation. Indeed \[ -2\frac{d}{dx}\tau(e^x)= \sum_i\frac{c_i\sigma_i e^{-\sigma_i x}}{1-e^{-2\sigma_i x}}, \] and directed interval bisection verifies positivity after multiplication by $e^x$ on $16\le x\le36$; the finite check uses $327600$ cells. If $n$ norms belong to an interval $[L,U]$ in this range, their total contribution is at least $n\tau(U)$. Thus an exact prime sum is at least any such lower endpoint estimate. The signed kernel is combined before this comparison, so its direction is unaffected by negative coefficients. \subsection{Finite evaluation of the arithmetic constant} \label{ar:original-L}\label{ar:replay}\label{ar:summation} The number to be evaluated is the exact real scalar \begin{equation}\label{sc:arithmetic-scalar} \Phi=Z_{\rm sel}(1)+bB_r+ \sum_i c_i\left\{Y_*(\sigma_i)-\Delta_{\rm sel}(\sigma_i) -Z_{\rm sel}(\sigma_i)-\Delta_4(\sigma_i) -\mathcal C_2(\sigma_i)\right\}, \end{equation} where \[ \mathcal C_2(s)=\frac12\sum_{\mathfrak p\in\mathcal P_2} \operatorname{arctanh}((N\mathfrak p)^{-s}). \] The set $\mathcal P_2$ consists of the degree-one free primes of norm at most $1125899906842000$ satisfying the residue tests in Section~\ref{ar:comparison-construction}. The exact field coefficients are listed in Section~\ref{sc:radical-tables}. Thus every operand in \eqref{sc:arithmetic-scalar} is determined by rational data and finite residue-field calculations, together with Dedekind zeta values at the specified rational arguments. Here is a direct verification algorithm using only finite Euler products. It also gives an independent specification of the numerical inequality, while the factorisation \eqref{ar:factorization} and the approximate functional equations below make its evaluation much faster. For an integer $P\ge2$ put \[ Y_{*,P}(s)=\frac1{8192} \sum_{p\le P}\sum_{\mathfrak q\mid p\ \text{in }E_*} -\log(1-p^{-f(\mathfrak q/p)s}). \] Every local factor includes all prime ideals over $p$, including those at ramified primes. The field can be constructed directly as \[ E_*=E_0(\sqrt{\beta_{10}},\sqrt{\beta_{23}}, \sqrt{\beta_{19}},\sqrt{\beta_{17}}). \] Its degree and its identification with the comparison field follow from Section~\ref{ar:comparison-construction}. \begin{lemma}\label{sc:euler-truncation} For $s>1$ and $P\ge2$, \[ 0\le Y_*(s)-Y_{*,P}(s) \le R_P(s):=\frac{P^{1-s}}{(s-1)(1-(P+1)^{-s})}. \] Consequently, replacing $Y_*$ by $Y_{*,P}$ in \eqref{sc:arithmetic-scalar} changes the scalar by at most \[ \sum_i|c_i|R_P(\sigma_i) <\frac{300000000000}{P^{1/50000}}. \] \end{lemma} \begin{proof} For each rational prime $p$, the number of prime ideals over $p$ is at most $8192$, and every residue degree is at least one. The normalised local logarithm is therefore at most $-\log(1-p^{-s})$. Hence \[ 0\le Y_*(s)-Y_{*,P}(s) \le\sum_{n>P}\frac{n^{-s}}{1-n^{-s}} \le\frac1{1-(P+1)^{-s}}\int_P^\infty x^{-s}\,dx. \] The final bound uses $\sigma_i\ge1+1/50000$, $\sum_i|c_i|<3000000$ and $1-(P+1)^{-\sigma_i}>1/2$. \end{proof} \begin{algorithm}[An upper bound for the arithmetic scalar] \label{sc:arithmetic-algorithm} Fix a rational error allowance $\eta>0$. \begin{enumerate}[label=(\roman*)] \item Construct the eighteen listed quadratic extensions and verify their identities by Algorithm~\ref{ar:radical-verification-algorithm}. Form $E_*$ from the four displayed radicals. \item Use Algorithm~\ref{ar:residue-floor-algorithm} to enumerate $\mathcal P_2$ and $\mathcal P_4$. For each rational prime that splits in $B$, include both prime ideals. Compute $\mathcal C_2$, adding the signed kernel at each norm before adding the contributions from different norms. Compute the selected-prime corrections by their stated local factors. \item Enclose the required zeta values. One choice is to factor primes in $E_*$ through a bound $P$ with $300000000000P^{-1/50000}<\eta$ and use Lemma~\ref{sc:euler-truncation}. The shorter choice is Algorithm~\ref{ar:short-sum-algorithm} applied to the factors in \eqref{ar:factorization}. In either case retain the proved remainder intervals. \item Evaluate the finite expressions with rational upper and lower bounds. Increase the number of terms and the arithmetic precision until the propagated interval width is less than $\eta$. Add the remainder intervals with the sign of each $c_i$, obtaining $\Phi\in[L,U]$. \item Verify the rational inequality $U0$, put $z=(x-1)/(x+1)$ and use \[ \log x=2\sum_{k=0}^{m}\frac{z^{2k+1}}{2k+1}+\varepsilon_m, \qquad |\varepsilon_m|\le \frac{2|z|^{2m+3}}{(2m+3)(1-|z|^2)}. \] Range reduction by powers of two speeds convergence. Exponentials are enclosed by their Taylor series with the usual positive remainder majorant; rational powers are evaluated as $\exp(r\log x)$. The alternating arctangent series and $\pi=16\arctan(1/5)-4\arctan(1/239)$ give rational intervals for $\pi$. For Euler's constant one may use $H_n-\log(n+1)<\gamma2$ at every abscissa used here, so the tail bound tends to zero as $N$ increases. The kernel remainders also tend to zero. Thus the procedure gives enclosures of arbitrarily small width. Positivity of the real $L$-values follows from their expression as quotients of Dedekind zeta functions for $s>1$. The kernels may also be evaluated on short intervals. They are completely monotone, since $H$ is a positive multiplicative convolution of $x^{\mu_j}e^{-x}$. About the right endpoint $\xi$ of a cell $[\xi',\xi]$, the degree-$D$ Taylor remainder is at most \begin{equation}\label{ar:cell-remainder} \frac{|I_b^{(D+1)}(\xi')|}{(D+1)!}(\xi-\xi')^{D+1}. \end{equation} The identity $xI_b'(x)+bI_b(x)=-H(x)$ supplies its derivatives. Exact integer moments can then replace term-by-term summation within a cell; this changes neither the finite sum nor its remainder bound. The genus-field factorisation is \[ \zeta_{E_0}(s)=\zeta_B(s) \prod_{0\ne w\in\mathbb F_2^8} \frac{\zeta_{B(\sqrt{\alpha_w})}(s)}{\zeta_B(s)}, \qquad \zeta_B(s)=\zeta(s)\,241^{-s} \sum_{a=1}^{240}\left(\frac a{241}\right)\zeta(s,a/241). \] The $255$ quotients are quadratic Hecke factors over $B$. Their exact coefficients come from the same local definition, and their degree-two kernel formulas and explicit tails are given below. The remaining Hurwitz zeta values have elementary Euler--Maclaurin enclosures. For $x>0$, let $y=N+x$. Then \begin{align*} \zeta(s,x)={}&\sum_{n=0}^{N-1}(n+x)^{-s} +\frac{y^{1-s}}{s-1}+\frac12y^{-s}\\ &+\sum_{k=1}^{M}\frac{B_{2k}}{(2k)!}(s)_{2k-1} y^{-s-2k+1}+R_{M,N},\\ |R_{M,N}|\le{}&\frac{4(s)_{2M}}{6^{2M}(s+2M-1)} y^{-s-2M+1}. \end{align*} Here $(s)_j=s(s+1)\cdots(s+j-1)$ and the Bernoulli numbers are exact rationals. The remainder follows by integrating the periodic Bernoulli polynomial and using $|B_{2M}(\{u\})|\le2(2M)!\zeta(2M)/(2\pi)^{2M}$, $\zeta(2M)<2$ and $\pi>3$. Increasing $N$ with $M$ fixed makes this bound tend to zero. These formulas specify the entire genus-field calculation, including its pole contribution at $s=1$ evaluated at the arguments $s>1$ used here. \subsection{Explicit kernel series and their error bounds} \label{ar:kd-section} This subsection supplies the kernel expansions and remainder estimates used in Section~\ref{sc:short-Dirichlet}. It follows the Mellin construction of Naslund~\cite{Naslund}. All truncation bounds are inequalities between exact real numbers, so they can be used directly in rational interval evaluation. \paragraph{Residues for the degree-four and degree-eight kernels.} Use the notation $H,I_b$ of Section~\ref{sc:short-Dirichlet}; in particular $I_b(x)=\int_1^\infty v^{b-1}H(xv)\,dv$. Group the shifts $\mu_j$ into the distinct values $\alpha\in\mathcal A\subseteq\{0,1/2\}$, with multiplicities $m_\alpha$. Put $m=\sum m_\alpha$. For $n\ge0$ define the function analytic at $e=0$ \begin{equation}\label{ar:kd-F} F_{\alpha,n}(e)= \left\{\frac{(-1)^n}{n!}\Gamma(1+e) \prod_{j=1}^n(1-e/j)^{-1}\right\}^{m_\alpha} \prod_{\beta\in\mathcal A\setminus\{\alpha\}} \Gamma(-n+\beta-\alpha+e)^{m_\beta}. \end{equation} Here an empty product is one, and $[e^j]$ means Taylor coefficient at zero. At the pole $z=-n-\alpha$, the product of gamma factors is $e^{-m_\alpha}F_{\alpha,n}(e)$. Taking residues yields \begin{align} H(x)&=\sum_{\alpha\in\mathcal A}\sum_{n\ge0} [e^{m_\alpha-1}]F_{\alpha,n}(e)x^{n+\alpha-e}, \label{ar:kd-H-series}\\ x^b I_b(x)&=\prod_j\Gamma(b+\mu_j) -\sum_{\alpha\in\mathcal A}\sum_{n\ge0} [e^{m_\alpha-1}] \frac{F_{\alpha,n}(e)x^{n+\alpha+b-e}} {n+\alpha+b-e}. \label{ar:kd-I-series} \end{align} The second formula is used when none of $n+\alpha+b$ is zero, as is the case at all the rational abscissae here. The additional term comes from the pole of $1/z$ in the Mellin integral for $x^bI_b$. When $b<0$, its initial contour lies to the right of the positive pole $z=-b$ as well. Thus no pole is accidentally lost near $b=0$. These formulas follow either by shifting the inverse Mellin contour, or first on a compact interval of positive $x$ and then by analytic continuation. The reflection formula and Stirling's estimate make the left-contour remainder tend to zero; the factorial bounds below also prove locally uniform absolute convergence of the residue series. Here are explicit bounds for the omitted residues. Choose $r=1/2$ if $|\mathcal A|=1$, and $r=1/4$ otherwise. For all sufficiently large $n$ (take $n>k+3$ for a derivative of order $k$), put \begin{align} C_{\alpha,n}(x)={}&r^{1-m_\alpha}e^{r|\log x|}x^{n+\alpha} \left\{\frac{\Gamma(1-r)}{n!} \prod_{j=1}^n(1-r/j)^{-1}\right\}^{m_\alpha}\notag\\ &\quad\times \prod_{\beta\ne\alpha} \left\{\frac{\pi} {(19/20)\cos(\pi r)\Gamma(n+1-(\beta-\alpha)-r)}\right\}^{m_\beta}. \label{ar:kd-Cauchy-majorant} \end{align} Then $C_{\alpha,n}(x)$ bounds the absolute value of the corresponding term in \eqref{ar:kd-H-series}. To check this, on $|e|=r$ use $|x^{-e}|\le e^{r|\log x|}$ and $|\Gamma(1+e)|\le\Gamma(1-r)$. For the other shift class the reflection formula gives \[ |\Gamma(-n+\beta-\alpha+e)| \le\frac{\pi} {\cos(\pi r)|\Gamma(n+1-(\beta-\alpha)-e)|}. \] If $\sigma\ge1$ and $|y|\le1/4$, the gamma product formula implies \[ \frac{|\Gamma(\sigma+iy)|}{\Gamma(\sigma)} =\prod_{j\ge0}\left(1+\frac{y^2}{(\sigma+j)^2}\right)^{-1/2} \ge\left(\frac{\pi|y|}{\sinh(\pi|y|)}\right)^{1/2}>\frac{19}{20}. \] At zero the ratio on the right is interpreted as one. The last rational bound follows, for instance, by using $\pi/4<11/14$ and the positive power series for $\sinh x/x$ to verify $\sinh(11/14)/(11/14)<400/361$. The relevant real parts are at least $3/2$, where $\Gamma$ is increasing. This proves the cross-shift factor in \eqref{ar:kd-Cauchy-majorant}. Cauchy's coefficient estimate supplies the factor $r^{1-m_\alpha}$. For $H^{(k)}(x)/k!$, multiply $C_{\alpha,n}(x)$ by \begin{equation}\label{ar:kd-derivative-factor} D_{\alpha,n,k}(x)= \frac{\prod_{j=0}^{k-1}(|n+\alpha-j|+r)}{k!x^k}. \end{equation} This follows by differentiating $x^{n+\alpha-e}$ before taking its coefficient. For $I_b(x)$ itself, the omitted term is at most $C_{\alpha,n}(x)/(n+\alpha+b-r)$ whenever the denominator is positive; the external factor $x^b$ in \eqref{ar:kd-I-series} cancels. There is a finite geometric bound for each infinite residue tail. The ratio of two consecutive $C$ bounds is exactly \begin{equation}\label{ar:kd-ratio} q_{\alpha,n}(x)= \frac{x}{(n+1-r)^{m_\alpha}} \prod_{\beta\ne\alpha} (n+1-(\beta-\alpha)-r)^{-m_\beta}. \end{equation} It decreases with $n$. For the derivative bound include the ratio $D_{\alpha,n+1,k}/D_{\alpha,n,k}$; this also decreases once $n>k$. Choose $n_0>k+3$ so that the appropriate ratio at $n_0$ is at most $1/2$. The tail beginning at $n_0$ is bounded by its first majorant divided by one minus that ratio. For $I_b$, the extra factor $(n+\alpha+b-r)^{-1}$ decreases and can be held at its value at $n_0$ while using \eqref{ar:kd-ratio}. Sum over the one or two shift classes. This is a completely specified omitted-pole bound. The choice of $n_0$ may be enlarged arbitrarily; the inequalities, not a floating estimate of an optimal cutoff, decide whether it is acceptable. Only the first $m_\alpha$ Taylor coefficients in $e$ are needed for each retained residue in \eqref{ar:kd-H-series}--\eqref{ar:kd-I-series}; the products, gamma functions and exponential $x^{-e}$ are evaluated as outward power-series balls. If $h_j=H^{(j)}(c)/j!$ and $p_j=I_b^{(j)}(c)/j!$, the differential identity in Section~\ref{sc:short-Dirichlet} gives the finite recurrence \begin{equation}\label{ar:kd-recurrence} p_{j+1}=-\frac{h_j+(j+b)p_j}{c(j+1)}. \end{equation} It carries both interval rounding and the explicit residue-tail bounds. The complete-monotonicity bound \eqref{ar:cell-remainder} then bounds the Taylor remainder on every cell. The Dirichlet tail has an explicit bound. If $\eta>0$ and $v_j=b+\mu_j+\eta>0$, set $U=\max_jv_j$ and let $m_0=2$ for $m=2$, $m_0=4$ for $m=4$. Mellin inversion and the gamma product inequality give \begin{equation}\label{ar:kd-kernel-power-bound} 0\beta$, the omitted terms after $N$ satisfy \begin{equation}\label{ar:kd-Dirichlet-tail} \sum_{n>N}|a_n|I_b((tn)^a) \le E_\beta N^\beta \frac{U}{m_0\eta}\prod_j\Gamma(v_j)\,(tN)^{-a(b+\eta)}. \end{equation} This is Rankin's inequality applied term by term. Multiplying by the positive AFE prefactor gives the omitted-tail interval used in Section~\ref{sc:short-Dirichlet}. \paragraph{Degree-two kernels for the genus-field factors.} The degree-two factors admit the following shorter kernels. Define, for real $\mu$, \begin{align} H_\mu(x)&=\int_1^\infty u^{\mu-1}e^{-xu}\,du, & I_\mu^{\rm B}(x)&=\int_1^\infty u^\mu K_0(xu)\,du, \label{ar:kd-base-kernels}\\ V_1(x)&=e^{-x}(1+1/x)/x, & V_2(x)&=\sqrt{\pi/2}\,e^{-x}(1+1/x)/x^{3/2}. \label{ar:kd-base-majorants} \end{align} Here $K_0(x)=\int_0^\infty e^{-x\cosh v}\,dv$ and $K_1(x)=\int_0^\infty e^{-x\cosh v}\cosh v\,dv$. These definitions give $K_0'=-K_1$ and, by integration by parts, $x^2K_0''+xK_0'-x^2K_0=0$. They also give $K_0(x)\le\sqrt{\pi/(2x)}e^{-x}$, by $\cosh v\ge1+v^2/2$. It follows directly that \[ 0c_J$ ensure no term is omitted or assigned twice. The choices $D=20$ and $N=8\lfloor\sqrt Q\rfloor+1$ are convenient initial values. Increase the cutoff, top endpoint and degree until the combined interval and tail errors meet the prescribed tolerance. Finally, for any $\beta>1$ and $X=t(N+1)$, the omitted Dirichlet tails satisfy the explicit bounds \begin{align} \sum_{n>N}|a_n|H_\mu(tn) &\le\zeta(\beta)^2(N+1)^\beta V_1(X), &&X\ge\beta-1,\label{ar:kd-base-tail-mixed}\\ \sum_{n>N}|a_n|n^\nu I_\mu^{\rm B}(tn) &\le\zeta(\beta)^2(N+1)^{\beta+\nu}V_2(X), &&X\ge\beta+\nu-3/2.\label{ar:kd-base-tail-pure} \end{align} The required exponents are $\mu=\sigma,1-\sigma$ in the first line, and $\mu=\sigma+\nu-1,\nu-\sigma$ in the second. For $1<\sigma<1.01$ their envelope conditions are satisfied. To prove the bounds, note that $x^\beta V_1(x)$ decreases for $x\ge\beta-1$, and $x^{\beta+\nu}V_2(x)$ decreases for $x\ge\beta+\nu-3/2$, as direct logarithmic differentiation shows. Multiply by $|a_n|n^{-\beta}$ and sum, using $\sum d_2(n)n^{-\beta}=\zeta(\beta)^2$. One may minimise the upper endpoint over $\beta=101/100,21/20,11/10,9/8,6/5,5/4,4/3,3/2,2$, or simply use $\beta=2$. Every candidate satisfying the displayed conditions gives a valid upper bound. Multiply these one-sided errors and the interpolation errors by the positive prefactor in the appropriate AFE, and add the errors for its two kernels. This proves the two-sided genus-field enclosures. All coefficients are the exact coefficients defined in Section~\ref{sc:intrinsic-arithmetic}. \section{A two-window transfer to planar incidences}\label{tr:section} This section proves the geometric part of the construction. In particular, the two sides of the incidence graph are allowed to have different local profiles, different counting powers, and different cardinalities. All averages below concern the \emph{absolute number of cross edges}; no symmetry of the endpoint distributions is assumed. \subsection{Number-field coordinates and the relative-unit mass} Let $K/F$ be a quadratic extension, with nontrivial automorphism $\iota$, such that $K$ is totally imaginary, $F$ has a real place, and $K/F$ is unramified at every finite place. Write \[ [F:\mathbb Q]=d=b+2c,\qquad \theta=c/d,\qquad \operatorname{rd}(F)=\operatorname{rd}(K)=\lambda,\qquad \ell=\log\lambda. \] Here $(b,c)$ is the signature of $F$; in particular $b\ge1$. The fields constructed in the arithmetic part satisfy these hypotheses. Put \[ L_F(1)=\frac{\operatorname*{res}_{s=1}\zeta_K(s)} {\operatorname*{res}_{s=1}\zeta_F(s)}. \] The arithmetic estimate used in this section is \begin{equation}\label{tr:arithmetic-input} \log L_F(1)\le Cd. \end{equation} Neither its numerical value nor the existence of the field family is deduced from a successful geometric computation. For each real place $v$ of $F$ choose an extension $\phi_v:K\to\mathbb C$. For each complex place choose extensions $\phi_j^+$ and $\phi_j^-=\phi_j^+\circ\iota$ inducing the same embedding of $F$. These choices identify $K\otimes_{\mathbb Q}\mathbb R$ with $\mathbb C^d$; write $x_\infty$ for the image of $x\in K$. There are $b$ single, \emph{compact} coordinates and $c$ pairs of coordinates. Define \[ l_j(x)=\tfrac12\bigl(\log|\phi_j^+(x)|-\log|\phi_j^-(x)|\bigr), \qquad l(x)=(l_j(x))_{j=1}^c. \] If $N_{K/F}x=x\iota(x)=1$, then \begin{equation}\label{tr:norm-coordinates} |\phi_v(x)|=1,\qquad (|\phi_j^+(x)|,|\phi_j^-(x)|)=(e^{l_j(x)},e^{-l_j(x)}). \end{equation} Indeed, complex conjugation over a real place is $\iota$, and the product of the two embeddings over a complex place is the corresponding embedding of $N_{K/F}x$. For $h\in\mathbb R^c$, let $W_h$ fix the compact coordinates and send $(z_j,w_j)$ to $(e^{-h_j}z_j,e^{h_j}w_j)$. It has real determinant one. Let $\mathcal U^1=\{u\in\mathcal O_K^\times:N_{K/F}u=1\}$, let $w_K$ be the number of roots of unity of $K$, and set \[ h_{\rm cap}=\left|\ker\bigl(\operatorname{Cl}(F)\to \operatorname{Cl}(K)\bigr)\right|, \quad h_{\rm rel}=h_K/h_F, \quad q_{\rm cl}=\left|\operatorname{Cl}(K)/ \operatorname{im}\operatorname{Cl}(F)\right| =h_{\rm cap}h_{\rm rel}. \] \begin{lemma}[Relative-unit mass]\label{tr:mass} The image $l(\mathcal U^1)$ is a full lattice in $\mathbb R^c$, and its kernel has order $w_K$. If $R^1$ denotes the covolume of this lattice ($R^1=1$ when $c=0$), then \begin{equation}\label{tr:mass-identity} \frac{w_K}{q_{\rm cl}R^1} =\frac{2^{d-1}\pi^{b+c}}{\lambda^{d/2}L_F(1)}. \end{equation} \end{lemma} \begin{proof} We give the normalization details because an omitted power of two would change the final exponent. Write $U_K=\mathcal O_K^\times$, $U_F=\mathcal O_F^\times$, and $u_N=[U_F:N_{K/F}U_K]$. The norm image contains $U_F^2$, so this index is finite. The rank form of Dirichlet's unit theorem~\cite{MilneANT} gives \[ \operatorname{rank}U_K=d-1,\quad \operatorname{rank}U_F=b+c-1,\quad \operatorname{rank}\mathcal U^1=c. \] Under the ordinary logarithmic embedding of $K$, a norm-one unit has coordinates $(0,(2l_j,-2l_j)_j)$. Thus its $l$-image is discrete and has rank $c$. Its kernel consists of algebraic integers all of whose conjugates have modulus one; these are roots of unity. One can see the last assertion directly: the monic integral polynomials of their powers have bounded coefficients and bounded degree, so only finitely many powers are possible. Every root of unity of $K$ has norm one, since its norm is a root of unity of $F$, hence $\pm1$, and is positive at a real place. This proves the lattice and kernel assertions. The same positivity shows that $-1$ is not a norm from $K^\times$. We next compute the capitulation factor. For the involution on $U_K$, the degree-zero Tate quotient has order $u_N$, and \[ H^1(\langle\iota\rangle,U_K) =\mathcal U^1/\{u/\iota(u):u\in U_K\}. \] The Herbrand quotient of this module is $2^{b-1}$. For completeness, its rational plus and minus ranks are $b+c-1$ and $c$, respectively. A trivial infinite cyclic module has quotient $2$, a sign module has quotient $1/2$, and a finite module has quotient $1$ (by counting the kernels and images of $1-\iota$ and $1+\iota$). Passing to a finite-index sum of plus and minus lattices therefore gives $2^{b-1}$. It follows that $|H^1|=u_N/2^{b-1}$. There is an isomorphism from the capitulation kernel to this $H^1$: if $\mathfrak a\mathcal O_K=x\mathcal O_K$, send $[\mathfrak a]$ to the class of $x/\iota(x)$. It is well defined. If this class is trivial, then $x/u\in F$ for a unit $u\in U_K$, and contraction of ideals shows that $\mathfrak a$ is principal. Conversely, Hilbert's theorem~90\cite{MilneCFT} writes any norm-one unit as $x/\iota(x)$. Its principal ideal is then $\iota$-invariant. Because no finite prime ramifies, every invariant fractional ideal of $K$ is extended from $F$: the exponents at a split pair agree, and an inert prime has ramification index one. This proves surjectivity, and hence \begin{equation}\label{tr:cap-factor} h_{\rm cap}=u_N/2^{b-1}. \end{equation} Use ordinary regulators, with logarithm $2\log|\cdot|$ at complex places. Delete the coordinate of one real place of $F$ in both unit logarithmic lattices. The norm map keeps each remaining compact coordinate and adds the two coordinates in each pair. The change $(x_j,y_j)\mapsto(x_j,x_j+y_j)$ has determinant one. Its kernel lattice is $2l(\mathcal U^1)$, of covolume $2^cR^1$. Its image has index $u_N/2$ in the logarithmic lattice of $U_F$, since logarithms kill $\{\pm1\}$ and $-1$ is not a norm. Taking covolumes gives \begin{equation}\label{tr:regulator-factor} R_K/R_F=2^{c-1}u_NR^1. \end{equation} The analytic class-number formula~\cite{MilneCFT}, with $w_F=2$, now gives \[ L_F(1)=\frac{2^{c+1}\pi^{b+c}h_{ m rel}}{w_K\lambda^{d/2}} \frac{R_K}{R_F}. \] Substitution of \eqref{tr:cap-factor} and \eqref{tr:regulator-factor}, together with $d=b+2c$, proves \eqref{tr:mass-identity}. Only the standard unit theorem, Hilbert~90 and analytic class-number formula are used here; the displayed calculation fixes their conventions for the present construction. \end{proof} \subsection{Finite coordinates, periods, and uniform point counts} Let $\mathcal P$ be a finite set of primes of $F$ that split in $K$. Choose $\mathfrak P_{\mathfrak p}$ over each $\mathfrak p$, so that $\mathfrak p\mathcal O_K=\mathfrak P_{\mathfrak p}\iota\mathfrak P_{\mathfrak p}$. Write $Q_{\mathfrak p}$ for the residue cardinality, $\mathcal O_{\mathfrak p}$ for the local valuation ring, and $\varpi_{\mathfrak p}$ for a uniformizer. Identify $K_{\mathfrak P_{\mathfrak p}}$ with $F_{\mathfrak p}$. Put \[ \mathbb A_{\mathcal P}=\prod_{\mathfrak p\in\mathcal P}F_{\mathfrak p}^2, \qquad \mathcal E=\mathbb C^d\times\mathbb A_{\mathcal P}. \] Finite additive Haar measures give $\mathcal O_{\mathfrak p}$ mass one; at infinity we use ordinary real Lebesgue measure. Let $\mathcal O_{\mathcal P}$ be the ring of elements of $K$ integral away from the chosen split pairs, embedded by \[ \gamma\longmapsto(\gamma_\infty,\gamma_f),\qquad \gamma_f=(\gamma_{\mathfrak p},(\iota\gamma)_{\mathfrak p})_{\mathfrak p}. \] For integers $a_{\mathfrak p},b_{\mathfrak p}$ define a rectangular group \[ \mathcal C_f=\prod_{\mathfrak p} (\varpi_{\mathfrak p}^{a_{\mathfrak p}}\mathcal O_{\mathfrak p} \times\varpi_{\mathfrak p}^{b_{\mathfrak p}}\mathcal O_{\mathfrak p}), \qquad H=d^{-1}\log|\mathcal C_f|. \] \begin{lemma}[The period lattice]\label{tr:period-lattice} Set \[ \mathfrak a_{\mathcal C_f}=\prod_{\mathfrak p} \mathfrak P_{\mathfrak p}^{a_{\mathfrak p}} (\iota\mathfrak P_{\mathfrak p})^{b_{\mathfrak p}}, \qquad \Delta_0=(\lambda/2)^d. \] The elements of $\mathcal O_{\mathcal P}$ with finite part in $\mathcal C_f$ are precisely $\mathfrak a_{\mathcal C_f}$, and $N\mathfrak a_{\mathcal C_f}=|\mathcal C_f|^{-1}$. Every coset of $\mathcal C_f$ in $\mathbb A_{\mathcal P}$ contains such a finite part. The group \[ \Lambda_h=\{(W_h\gamma_\infty,\gamma_f):\gamma\in\mathcal O_{\mathcal P}\} \] is a lattice in $\mathcal E$ of covolume $\Delta_0$. A fundamental domain is $W_hP_{\mathcal C_f}\times\mathcal C_f$, where $P_{\mathcal C_f}$ is an archimedean fundamental parallelepiped of $\mathfrak a_{\mathcal C_f}$. Every nonzero vector in the Euclidean dual of $W_h(\mathfrak a_{\mathcal C_f})_\infty$ satisfies \begin{equation}\label{tr:dual-product} \prod_{w=1}^d|\xi_w|\ge\mu^d, \qquad \mu=2e^{H/2-\ell}. \end{equation} \end{lemma} \begin{proof} The valuation conditions give the first assertion and the norm formula. For the coset assertion, enclose a given coset in a larger rectangular group $\widetilde{\mathcal C}_f$. The natural map \[ \mathfrak a_{\widetilde{\mathcal C}_f}/\mathfrak a_{\mathcal C_f} \longrightarrow \widetilde{\mathcal C}_f/\mathcal C_f \] is injective by those valuation conditions. Both groups have order $|\widetilde{\mathcal C}_f|/|\mathcal C_f|$, so it is surjective. First reducing the finite coordinates by this assertion and then reducing the infinite coordinates modulo $\mathfrak a_{\mathcal C_f}$ proves the fundamental-domain assertion, with uniqueness off the chosen half-open boundaries. The standard ideal covolume formula is \[ \operatorname{covol}(\mathfrak a_\infty) =2^{-d}|\Delta_K|^{1/2}N\mathfrak a =\Delta_0N\mathfrak a. \] Thus the product domain has volume $\Delta_0$. Discreteness follows locally from the discreteness of the ideal lattice. The map $W_h$ has determinant one and hence preserves this covolume. For the last assertion, use the real inner product $\langle x,y\rangle=\operatorname{Re}\sum_wx_w\overline{y_w}$. If $\mathfrak a^\vee$ is the trace-dual ideal, the Euclidean dual of $\mathfrak a_\infty$ is $\overline{(2\mathfrak a^\vee)_\infty}$, since $\operatorname{Re}\sum_w\phi_w(xy)=\tfrac12\operatorname{Tr}_{K/\mathbb Q}(xy)$. The discriminant formula gives $N\mathfrak a^\vee=\lambda^{-2d}(N\mathfrak a)^{-1}$. For nonzero $y\in\mathfrak a^\vee$ its principal ideal is contained in $\mathfrak a^\vee$, so $|N_{K/\mathbb Q}y|\ge N\mathfrak a^\vee$. Consequently the product of the moduli of the corresponding dual vector is at least $2^d\lambda^{-d}(N\mathfrak a)^{-1/2}$. Applying this to $\mathfrak a_{\mathcal C_f}$ gives \eqref{tr:dual-product}. Finally the dual of $W_h\mathfrak a_\infty$ is $W_{-h}$ applied to the old dual, and this map preserves the product of the moduli. \end{proof} \begin{lemma}[Uniform Poisson bound]\label{tr:poisson} Put $\|\xi\|_* =\sum_w|\xi_w|$. Suppose a Euclidean lattice $\Lambda\subset\mathbb C^d$ obeys \eqref{tr:dual-product}, and a nonnegative, smooth, integrable function $\Psi$ has bounded derivatives of every order and satisfies \[ |\widehat\Psi(\xi)|\le M_F^d e^{-\sigma\|\xi\|_*}\int\Psi, \qquad M_F\ge1. \] If \begin{equation}\label{tr:fourier-condition} \sigma\mu\ge\log M_F+2\log5+2, \end{equation} then, for every translate $y$, \[ \sum_{x\in\Lambda+y}\Psi(x)\le \frac{2}{\operatorname{covol}\Lambda}\int\Psi. \] Consequently, if $\Psi_f$ is nonnegative, constant on $\mathcal C_f$-cosets and supported on finitely many such cosets, then for every $h$ and $z_0\in\mathcal E$, \begin{equation}\label{tr:adelic-count} \sum_{x\in\Lambda_h+z_0}\Psi_\infty(x_\infty)\Psi_f(x_f) \le\frac2{\Delta_0}\int\Psi_\infty\int\Psi_f, \end{equation} provided $\Psi_\infty$ satisfies the preceding Fourier bound and \eqref{tr:fourier-condition} uses $H=d^{-1}\log|\mathcal C_f|$. \end{lemma} \begin{proof} Arithmetic--geometric mean gives $\|\xi\|_*\ge d\mu$ for nonzero dual vectors; the same bound separates distinct dual vectors. Disjoint balls of radius $d\mu/2$ show that the shell $jd\mu\le\|\xi\|_*<(j+1)d\mu$ contains at most $(2j+3)^{2d}$ points. Since $(2j+3)\le5^j$ for $j\ge1$, condition \eqref{tr:fourier-condition} gives \[ M_F^d\sum_{\xi\in\Lambda^*\setminus\{0\}}e^{-\sigma\|\xi\|_*} \le\sum_{j\ge1}e^{-2dj}\le1. \] Poisson summation now proves the assertion for a Schwartz function. For the stated hypotheses apply it to $\Psi(x)e^{-\varepsilon|x|^2}$. The Fourier transform is convolved with a Gaussian probability density; the exponential bound is multiplied by $m_\varepsilon=\int e^{\sigma\|u\|_*}G_\varepsilon(u)\,du\to1$. The zero-frequency term is $\int\Psi e^{-\varepsilon|x|^2}$. Monotone convergence as $\varepsilon\downarrow0$ gives the claimed pointwise bound for every translate. To prove \eqref{tr:adelic-count}, partition by the finite $\mathcal C_f$-coset. By Lemma~\ref{tr:period-lattice}, the infinite coordinates in each nonempty part are a translate of $W_h(\mathfrak a_{\mathcal C_f})_\infty$. Its covolume is $\Delta_0/|\mathcal C_f|$, so the first assertion bounds each part by $2|\mathcal C_f|\int\Psi_\infty/\Delta_0$. The sum of the constant finite weights is $(\int\Psi_f)/|\mathcal C_f|$. \end{proof} \subsection{Bilinear shell resources} For a local field with residue size $Q$, put $\mathcal B_i=\varpi^{-i}\mathcal O$, $S_0=\mathcal O$, $S_i=\mathcal B_i\setminus\mathcal B_{i-1}$ for $i\ge1$, and $m_0=1$, $m_i=(Q-1)Q^{i-1}$. Fix $k\ge0$ and two finite nonnegative matrices $w^L,w^R$. The corresponding edge profiles are \[ f_a(x,y)=\sum_{i,j}w^a_{ij} \mathbf1_{S_i\times\varpi^{-k}S_j}(x,y),\qquad a\in\{L,R\}. \] They are invariant under separate multiplication by local units and under the common additive period $\mathcal O\times\varpi^{-k}\mathcal O$. The latter follows because adding an integral element preserves every shell; finite support also makes the profiles compactly supported. \begin{lemma}[Exact bilinear shell formulas]\label{tr:shell-resources} Define the symmetric matrix $D$ by \[ D_{ii'}=m_{\min(i,i')}\ (i\ne i'),\quad D_{00}=0,\quad D_{ii}=im_i-Q^{i-1}\ (i\ge1), \] and set \[ G_{(i,j),(i',j')}=(k+1)m_im_j\mathbf1_{i=i',j=j'} +m_jD_{ii'}\mathbf1_{j=j'}+m_iD_{jj'}\mathbf1_{i=i'}. \] For $s(n,\omega)=(\varpi^n\omega,\varpi^{-n}\omega^{-1})$, with Haar probability measure on $\mathcal O^\times$, the cross overlap is \begin{equation}\label{tr:shell-cross} I=\sum_{n\in\mathbb Z}\int_{\mathcal O^\times}\int f_L(x)f_R(x+s(n,\omega))\,dx\,d\omega =Q^k(w^L)^TGw^R. \end{equation} If $I>0$, the expectations of $-\log f_L$ at the left endpoint and $-\log f_R$ at the right endpoint under this normalized edge law are \[ E_L=-\frac{\sum_{i,j}w^L_{ij}(Gw^R)_{ij}\log w^L_{ij}} {(w^L)^TGw^R},\qquad E_R=-\frac{\sum_{i,j}w^R_{ij}(Gw^L)_{ij}\log w^R_{ij}} {(w^L)^TGw^R}. \] Terms with zero weight are omitted. For counting profiles $\psi_a=f_a^{v_a}$, $v_a>0$, the exact local resources are \begin{align} A_a&=\int\psi_a=Q^k\sum_{i,j}m_im_j(w^a_{ij})^{v_a},\notag\\ X_a^{\rm loc}&=\log A_a+v_aE_a,\label{tr:shell-count-resource}\\ Z^{\rm loc}&=k\log Q+\log((w^L)^TGw^R)+E_L+E_R.\notag \end{align} The period contribution is $k\log Q$. It must not be added a second time to either resource in \eqref{tr:shell-count-resource}. \end{lemma} \begin{proof} Two nonarchimedean balls have intersection volume equal to the smaller volume when the displacement belongs to the larger ball, and otherwise have empty intersection. Therefore the cross overlap between the rectangles $\mathcal B_i\times\mathcal B_{j+k}$ and $\mathcal B_{i'}\times\mathcal B_{j'+k}$ is \[ Q^{k+\min(i,i')+\min(j,j')} \bigl(k+1+\max(i,i')+\max(j,j')\bigr). \] Indeed the possible valuations are exactly the integers from $-\max(i,i')$ to $k+\max(j,j')$, and unit factors do not affect this test. Replace each shell indicator by the difference of its two ball indicators, omitting the lower ball when the index is zero. The double difference of $Q^{\min(i,i')}$ is $m_i\mathbf1_{i=i'}$, while that of $\max(i,i')Q^{\min(i,i')}$ is $D_{ii'}$; direct subtraction gives the displayed diagonal and off-diagonal values. Applying these differences in both coordinates gives $Q^kG$. Bilinear expansion proves \eqref{tr:shell-cross}. This argument does not require either matrix to be symmetric or to factor as a product of two vectors. The contribution of a fixed left shell to the overlap is $Q^kw^L_{ij}(Gw^R)_{ij}$. Dividing by $I$ gives its marginal probability, and hence $E_L$. Symmetry of $G$ gives the right marginal and $E_R$. Finally the shell rectangles are disjoint and have volumes $Q^km_im_j$, which gives $A_a$ and the remaining formulas. All $G$-entries are nonnegative, since $im_i-Q^{i-1}=Q^{i-1}(iQ-i-1)\ge0$. \end{proof} \subsection{The four energies and the two windows} At each compact place, each paired complex place, and each selected finite place choose two edge profiles $f_{L,\tau},f_{R,\tau}$ and two counting profiles $\psi_{L,\tau},\psi_{R,\tau}$. Here $\tau$ indexes blocks, not individual real coordinates. At a compact block the shift is $1$; at a pair block it is $(e^u,e^{-u})$, integrated over $u\in\mathbb R$; at a finite block it is $s(n,\omega)$ as above, summed over $n$ and integrated over the unit group. Let $I_\tau>0$ be the cross integral of $f_{L,\tau}f_{R,\tau}$ with these shifts. We state precisely the hypotheses on the local profiles. Both edge and counting profiles are invariant under separate coordinate rotations at infinity and separate local-unit multiplications at finite places. Finite profiles have a common rectangular additive period and compact support. A counting profile is positive exactly where its edge profile is positive. Each infinite counting energy $-\log\psi_{a,\tau}$ is bounded below and coercive. The infinite counting profiles are smooth, integrable, have bounded derivatives, and satisfy Fourier bounds with common constants $(M_F,\sigma)$, with prefactor $M_F$ on a compact block and $M_F^2$ on a paired block. Finally the following four local energies have finite means and variances under the \emph{cross-edge} law: \[ -\log\psi_{L,\tau}(x),\quad -\log\psi_{R,\tau}(x+e),\quad -\log f_{L,\tau}(x),\quad -\log f_{R,\tau}(x+e). \] The analytic part verifies these conditions for the profiles used here. Counting powers may differ between sides and between blocks. No restriction such as $\psi_L=f_L^p$ with one common $p$ is imposed. Write $f_a=\prod_\tau f_{a,\tau}$, $\psi_a=\prod_\tau\psi_{a,\tau}$, $A_a=\int_{\mathcal E}\psi_a$, and $I=\prod_\tau I_\tau$. Let $e(u,n,\omega)$ denote the product shift: its compact coordinates are $1$, its paired coordinates are $(e^{u_j},e^{-u_j})$, and its finite coordinates are $(\varpi_{\mathfrak p}^{n_{\mathfrak p}}\omega_{\mathfrak p}, \varpi_{\mathfrak p}^{-n_{\mathfrak p}}\omega_{\mathfrak p}^{-1})$. Tonelli's theorem shows that \begin{equation}\label{tr:cross-law} \frac{f_L(x)f_R(x+e(u,n,\omega))}{I} \,dx\,du\,d\omega\,d\#n \end{equation} is a probability measure, the product of the local edge laws. Define \[ \begin{aligned} W_a&=-\log\psi_a, & V_a&=-\log f_a,\\ \mu_L&=\mathbb E W_L(x), &\mu_R&=\mathbb E W_R(x+e),\\ \nu_L&=\mathbb E V_L(x), &\nu_R&=\mathbb E V_R(x+e). \end{aligned} \] These are total means, not means divided by $d$. \begin{lemma}[Four-energy truncation]\label{tr:concentration} Suppose each of the four total energies has variance at most $Vd$. For $\eta>0$ and $d\ge8V/\eta^2$, define \[ \Omega_a=\{x:f_a(x)>0,\ W_a(x)\le\mu_a+\eta d\}. \] Then $\Omega_L,\Omega_R$ are bounded Borel sets, invariant under the common finite period and the rotations and units specified above. Put \[ \Phi(u,n)=|\Omega_L\cap(\Omega_R-e(u,n,1))|, \qquad \mathcal W(n)=\int_{\mathbb R^c}\Phi(u,n)\,du. \] Only finitely many $n$ have nonzero $\mathcal W(n)$, and \begin{equation}\label{tr:unweighted-overlap} \sum_n\mathcal W(n)\ge\tfrac12 I \exp(\nu_L+\nu_R-2\eta d). \end{equation} Also $\mathbf1_{\Omega_a}\le e^{\mu_a+\eta d}\psi_a$. \end{lemma} \begin{proof} Coercivity and lower boundedness of the infinite count energies, together with finite support and bounded energies at the finite places, make each sublevel set bounded. Measurability and all invariances follow from those of its factors. Let $Y$ be the event in \eqref{tr:cross-law} where the two count energies are at most their means plus $\eta d$, and the two edge energies are at least their means minus $\eta d$. Chebyshev and the union bound give $\mathbb P(Y)\ge1-4V/(\eta^2d)\ge1/2$. Independence between these four energies is neither asserted nor needed. On $Y$ the endpoints lie in the appropriate two windows, and \[ f_L(x)f_R(x+e)\le e^{-\nu_L-\nu_R+2\eta d}. \] Multiplying by this upper bound and integrating the event proves \eqref{tr:unweighted-overlap}. The unit integral can be removed from the unweighted overlap: multiplication of the two finite coordinates by $(\omega,\omega^{-1})$ preserves both windows and measure. Coordinate rotations similarly remove arbitrary archimedean phases. For finiteness, the support of the difference of two bounded windows has bounded coordinates. Thus both $\varpi^n$ and $\varpi^{-n}$ must lie in fixed local balls, restricting $n$ to a finite interval at each prime. Likewise $e^{u_j}$ and $e^{-u_j}$ are bounded whenever the overlap is nonempty; hence the relevant $u$ lie in a bounded box. The last inequality is immediate from the definition of $W_a$. \end{proof} For a fixed finite catalog of profiles and $O(d)$ finite places, the variance hypothesis holds with one constant $V$: each total energy is a sum of independent block contributions, and the maximum of finitely many local variances is finite. This statement remains uniform under arbitrary mixtures, including proportions tending to zero. It does not require each block population to tend to infinity. \subsection{Class selection, common translations, and projection} \begin{lemma}[Weighted class selection]\label{tr:selection} For nonnegative finitely supported weights $\mathcal W(n)$ with positive sum, there exist a subset $\mathcal N'$ of valuation indices, an index $n_0\in\mathcal N'$, and elements $\beta_n\in\mathcal O_{\mathcal P}$ such that \begin{align} \sum_{n\in\mathcal N'}\mathcal W(n)&\ge q_{\rm cl}^{-1}\sum_n\mathcal W(n), \label{tr:selected-mass}\\ N_{K/F}\beta_n&=1,\qquad (\beta_n)=\prod_{\mathfrak p}\mathfrak P_{\mathfrak p}^{n_{\mathfrak p}-n_{0,\mathfrak p}} (\iota\mathfrak P_{\mathfrak p})^{-n_{\mathfrak p}+n_{0,\mathfrak p}}. \notag \end{align} The cosets $\beta_n\mathcal U^1$ are pairwise disjoint. \end{lemma} \begin{proof} Map $n$ to the class of $\mathfrak A(n)=\prod_{\mathfrak p}\mathfrak P_{\mathfrak p}^{n_{\mathfrak p}}$ in $\operatorname{Cl}(K)/\operatorname{im}\operatorname{Cl}(F)$ and choose a fiber of largest total weight. It has positive weight, so choose $n_0$ in it with positive weight. For $n$ in this fiber there are $x_n\in K^\times$ and a fractional ideal $\mathfrak a_n$ of $F$ with $\mathfrak A(n)\mathfrak A(n_0)^{-1}=x_n\mathfrak a_n\mathcal O_K$. Put $\beta_n=x_n/\iota(x_n)$. Dividing the ideal identity by its $\iota$-image cancels the extended ideal and gives the stated ideal of $\beta_n$, while its relative norm is one. Its valuations vanish away from the selected primes, so it belongs to $\mathcal O_{\mathcal P}$. Distinct $n$ give distinct ideals. Two norm-one generators of the same ideal differ by a norm-one unit, which proves the coset assertion. \end{proof} \begin{lemma}[Absolute cross-edge average]\label{tr:average} Let $\Omega_L,\Omega_R$ be as in Lemma~\ref{tr:concentration}. Suppose that each side has a fixed finite cardinality cap for every common translate, uniformly under reciprocal finite-coordinate rescaling of the two windows. There exist finite sets of field elements $\Gamma_L,\Gamma_R$ within those respective caps whose cross norm-one edges number at least \begin{equation}\label{tr:absolute-average} \frac{w_K}{q_{\rm cl}R^1\Delta_0}\sum_n\mathcal W(n). \end{equation} \end{lemma} \begin{proof} Apply Lemma~\ref{tr:selection}. Let $\alpha_0$ act only on the finite coordinates, with components \[ (\varpi_{\mathfrak p}^{n_{0,\mathfrak p}}, \varpi_{\mathfrak p}^{-n_{0,\mathfrak p}}). \] Set $\Omega'_a=\alpha_0^{-1}\Omega_a$, $\mathcal C'_f=\alpha_0^{-1}\mathcal C_f$. This reciprocal scaling preserves Haar measure, the period measure, and all infinite coordinates. It is not a change of the planar radius. For $n\in\mathcal N'$ and $\beta\in\beta_n\mathcal U^1$, its finite coordinates after multiplication by $\alpha_0$ have the moduli prescribed by $n$ and inverse unit parts. Thus rotation and unit invariance give \begin{equation}\label{tr:overlap-beta} |\Omega'_L\cap(\Omega'_R-(W_h\beta_\infty,\beta_f))| =\Phi(l(\beta)-h,n). \end{equation} For fixed $h$, average a single common translation $z_0$ uniformly over $W_hP_{\mathcal C'_f}\times\mathcal C'_f$, of volume $\Delta_0$. Let $\Gamma_a(h,z_0)$ consist of the $\gamma\in\mathcal O_{\mathcal P}$ whose point $(W_h\gamma_\infty,\gamma_f)+z_0$ lies in $\Omega'_a$. Let $E(h,z_0)$ count pairs $(\gamma,\beta)$ with $\gamma\in\Gamma_L(h,z_0)$, $\beta\in\bigcup_{n\in\mathcal N'}\beta_n\mathcal U^1$, and $\gamma+\beta\in\Gamma_R(h,z_0)$. Tiling by the lattice and Tonelli give the exact conditional average \[ \mathbb E_{z_0}E(h,z_0)=\Delta_0^{-1} \sum_{n\in\mathcal N'}\sum_{\beta\in\beta_n\mathcal U^1} \Phi(l(\beta)-h,n). \] Now average $h$ uniformly over a fundamental parallelepiped $\Pi^1$ of $l(\mathcal U^1)$. Each point of $l(\beta_n)+l(\mathcal U^1)$ has exactly $w_K$ preimages in $\beta_n\mathcal U^1$. The translates of $\Pi^1$ tile $\mathbb R^c$, and therefore \[ \frac1{R^1}\int_{\Pi^1}\sum_{\beta\in\beta_n\mathcal U^1} \Phi(l(\beta)-h,n)\,dh =\frac{w_K}{R^1}\int_{\mathbb R^c}\Phi(u,n)\,du. \] This is an unfolding identity, not an equidistribution assertion about the arguments or finite-unit parts of algebraic units. Combining it with \eqref{tr:selected-mass} proves that the average of $E$ is at least \eqref{tr:absolute-average}. The functions are measurable countable sums of indicators. No edge is counted twice: the unit cosets are disjoint, and the ordered pair of field elements determines its difference $\beta$. Hence $E$ is bounded by the product of the two uniform cardinality caps. It is integrable, and some common choice of $h,z_0$ attains at least its average. Both cardinality caps hold for this choice because they hold for every choice. The same translation must be used on the two sides. \end{proof} \begin{theorem}[Two-window geometric transfer]\label{tr:transfer} Assume \eqref{tr:arithmetic-input} and the local-profile hypotheses above. Suppose the common period satisfies \eqref{tr:fourier-condition}, and the four energy variances are at most $Vd$. Define \begin{align} X_L&=\log2-\ell+\frac{\log A_L+\mu_L}{d},\qquad X_R=\log2-\ell+\frac{\log A_R+\mu_R}{d},\label{tr:resources}\\ Z&=2\log2-\tfrac32\ell-C+(1-\theta)\log\pi +\frac{\log I+\nu_L+\nu_R}{d}.\notag \end{align} For every $\eta>0$ with $d\ge8V/\eta^2$, there exist finite sets of distinct planar points $P,Q$ such that \begin{equation}\label{tr:transfer-bounds} |P|\le2e^{d(X_L+\eta)},\quad |Q|\le2e^{d(X_R+\eta)},\quad \#\{(p,q)\in P\times Q:|p-q|=1\} \ge\tfrac14e^{d(Z-2\eta)}. \end{equation} Putting a unit circle at every center in $Q$ gives the same number of point--unit-circle incidences. The sets $P$ and $Q$ may overlap, but each set separately has no repeated point, and the circles are distinct. \end{theorem} \begin{proof} Take the two windows of Lemma~\ref{tr:concentration} and perform the common finite rescaling in Lemma~\ref{tr:average}. The rescaled counting majorant is $\psi_a\circ\alpha_0$ in the finite coordinates. Its integral is unchanged, and it is periodic under $\mathcal C'_f$, whose measure is unchanged. Its infinite part has Fourier bound $M_F^{b+2c}e^{-\sigma\|\xi\|_*}$ times its mass, by multiplication of the block Fourier bounds. Lemma~\ref{tr:poisson} and the pointwise majorization in Lemma~\ref{tr:concentration} therefore give, simultaneously for both sides and for every $h,z_0$, \[ |\Gamma_a(h,z_0)|\le\frac{2A_a e^{\mu_a+\eta d}}{\Delta_0} =2e^{d(X_a+\eta)}. \] These are absolute caps, not bounds holding only for some favorable translate on each side. Lemmas~\ref{tr:average}, \ref{tr:concentration} and \ref{tr:mass} now give a common translate with at least \[ \frac{w_K I e^{\nu_L+\nu_R-2\eta d}} {2q_{\rm cl}R^1\Delta_0} =\frac{2^{2d-2}\pi^{b+c}}{\lambda^{3d/2}L_F(1)} I e^{\nu_L+\nu_R-2\eta d} \ge\tfrac14e^{d(Z-2\eta)} \] cross norm-one edges. This calculation accounts for both factors of two: one from the four-energy event and one from the class/unit mass identity. There is no additional division by two for ordered bipartite incidences. Choose a real place $v_0$ of $F$ and project both sets to its compact coordinate. Equivalently, discard their common translated coordinate and take $P=\{\phi_{v_0}(\gamma):\gamma\in\Gamma_L\}$ and $Q=\{\phi_{v_0}(\gamma):\gamma\in\Gamma_R\}$. The field embedding is injective, so neither cardinality decreases. For every counted difference $\beta$, \[ |\phi_{v_0}(\beta)|^2 =v_0(\beta\iota(\beta))=1. \] Distinct ordered pairs remain distinct ordered planar pairs. Assigning the circle of radius one centered at each element of $Q$ proves the incidence assertion. A coordinate common to $P$ and $Q$ causes no problem: its distance from its own center is zero and was never counted. \end{proof} \begin{corollary}[Strict resource inequalities and exact sizes] \label{tr:strict-caps} Consider a family with degrees tending to infinity and a fixed finite catalog of local profiles. Suppose the preceding hypotheses hold uniformly, and, for some $\kappa>0$, prescribed $L,r,p$ satisfy \[ X_L\le L-\kappa,\qquad X_R\le rL-\kappa,\qquad Z\ge pL+\kappa. \] For all sufficiently large degrees there are exactly $\lceil e^{dL}\rceil$ points and $\lceil e^{drL}\rceil$ unit circles with at least $e^{dpL}$ incidences. The threshold is uniform if $L,r,p$ range over a compact set and the same $\kappa$, Fourier margin and variance bound work throughout. \end{corollary} \begin{proof} Choose $0<\eta<\kappa/4$ in Theorem~\ref{tr:transfer}. For large $d$, the strict exponential gaps absorb the constants $2$ and $1/4$ in \eqref{tr:transfer-bounds}. Add distinct points and circle centers to reach the prescribed cardinalities; this cannot remove an incidence. The proof uses only the stated uniform constants. If one sets $L=\log s/d$ for an integer $s$, the same argument gives exactly $s$ points and $\lceil s^r\rceil$ circles. The separate all-size argument must supply an admissible mixture at each such $L$; a mere unbounded sequence of field degrees does not by itself supply every cardinality. \end{proof} \begin{corollary}[Balanced positive-margin criterion] \label{tr:balanced-corollary} Let $0<\delta<1$ and $p_0=2/(1+\delta)$. At every block take equal left and right edge profiles $f$, with counting profile $f^{p_0}$. Suppose the hypotheses of Theorem~\ref{tr:transfer} hold uniformly along a field family with $d\to\infty$. Write \[ J_{\mathbb R}=\log I_{\mathbb R}-(1+\delta)\log A_{\mathbb R},\quad J_{\mathbb C}=\log I_{\mathbb C}-(1+\delta)\log A_{\mathbb C},\quad F_{\rm fin}=\frac1d\sum_{\mathfrak p\in\mathcal P} \bigl(\log I_{\mathfrak p}-(1+\delta)\log A_{\mathfrak p}\bigr), \] when the compact blocks share a profile and the paired blocks share a profile. If the margin \begin{equation}\label{tr:balanced-margin} \mathcal M=F_{\rm fin}-(\tfrac12-\delta)\ell-C +(1-\delta)\log2+(1-\theta)\log\pi +(1-2\theta)J_{\mathbb R}+\theta J_{\mathbb C} \end{equation} has a positive uniform lower bound, there are finite planar sets $U_j$ such that \[ |U_j|\longrightarrow\infty,\qquad \frac{u(U_j)}{|U_j|^{1+\delta}}\longrightarrow\infty, \] where $u(U)$ counts unordered unit distances. Different profiles within either archimedean type are handled by replacing the corresponding terms by their normalized sums. \end{corollary} \begin{proof} Every profile is even. For each fixed local displacement $e$, the map $x\mapsto-x-e$ preserves the symmetric density $f(x)f(x+e)$ and exchanges the two endpoint energies. Thus $\nu_L=\nu_R=\nu$, $\mu_L=\mu_R=p_0\nu$, and the two windows, their rescalings, and their selected field sets coincide. Write $X=X_L=X_R$. Substituting the resource formulas and using $p_0(1+\delta)=2$ gives \[ Z-(1+\delta)X=\mathcal M; \] all energy means cancel exactly. There is no extra finite-period term: the factors $Q_{\mathfrak p}^{k_{\mathfrak p}}$ are already in both $I_{\mathfrak p}$ and $A_{\mathfrak p}$. Theorem~\ref{tr:transfer}, with its now identical projected sets $U$, gives \[ \frac{u(U)}{|U|^{1+\delta}} \ge 2^{-4-\delta} \exp\bigl(d[\mathcal M-(3+\delta)\eta]\bigr). \] Here we used $u(U)\ge E/2$ for the counted ordered edges and $|U|\le2e^{d(X+\eta)}$. Choose $4\eta<\inf\mathcal M$. The right side diverges. Since $u(U)\le|U|^2/2$ and $\delta<1$, the cardinalities also tend to infinity. This gives the sequence of cardinalities used in Theorem~\ref{thm:sequences}. \end{proof} \begin{algorithm}[Coordinates from two lattice windows]\label{tr:enumeration} Fix the field data and the common choices $h,z_0$ in the construction. \begin{enumerate}[label=(\roman*)] \item List the finitely many $\mathcal C'_f$-cosets meeting the finite support. For each coset choose the representative $\gamma_r$ supplied by Lemma~\ref{tr:period-lattice}; its fiber is $\gamma_r+\mathfrak a_{\mathcal C'_f}$. \item Choose a $\mathbb Z$-basis of $\mathfrak a_{\mathcal C'_f}$. The bounded infinite window bounds every coefficient in this basis. List the lattice elements with integer coefficients in these bounds. \item Select the elements satisfying the left window inequalities and, separately, those satisfying the right window inequalities. Project both sets by the embedding in Theorem~\ref{tr:transfer}. Use the first set as points and the second as centres of circles of radius one. \end{enumerate} The selected elements lie in the number field, so the resulting point coordinates and circle centres are real algebraic numbers. \end{algorithm} \begin{lemma}[Effective extraction of coordinates]\label{tr:effective} Given integers $N,M,E\ge0$, there is a terminating algorithm which decides whether $N$ distinct points and $M$ distinct unit-circle centers can have at least $E$ incidences, and, if so, returns such a configuration with real algebraic coordinates. \end{lemma} \begin{algorithm}[Extraction of an incidence configuration] \label{tr:coordinate-algorithm} Fix integers $N,M,E\ge0$. \begin{enumerate}[label=(\roman*)] \item If $E>NM$, there is no such configuration. If $E=0$, choose $N$ distinct rational points and $M$ distinct rational centres. \item For $1\le E\le NM$, list the bipartite graphs with exactly $E$ edges on the labelled sides $\{1,\ldots,N\}$ and $\{1,\ldots,M\}$. For each graph, introduce real coordinates $(x_i,y_i)$ and $(u_j,v_j)$ and impose \[ (x_i-u_j)^2+(y_i-v_j)^2=1 \quad\text{for each edge }(i,j). \] Also impose \[ \begin{split} (x_i-x_k)^2+(y_i-y_k)^2&>0\quad(i\ne k),\\ (u_j-u_l)^2+(v_j-v_l)^2&>0\quad(j\ne l). \end{split} \] \item Apply effective quantifier elimination over real closed fields to each of these finitely many systems. If one is nonempty, choose a real algebraic point in it by algebraic sample-point extraction \cite{BasuPollackRoy}. Its coordinates give the desired points and centres. If all the systems are empty, no required configuration exists. \end{enumerate} \end{algorithm} \begin{proof}[Proof of Lemma~\ref{tr:effective}] Each system in Algorithm~\ref{tr:coordinate-algorithm} has integer coefficients. Effective quantifier elimination and algebraic sample-point extraction terminate for every such system \cite{BasuPollackRoy}; there are only finitely many graphs to consider. Any configuration with at least $E$ incidences contains one of their edge sets. Conversely, a solution realizes all $E$ edges of its graph. Extra incidences and points common to the two sides are allowed. This proves termination and correctness. \end{proof} Whenever the construction guarantees a specified integer target $E$, Lemma~\ref{tr:effective} gives an alternative finite procedure for obtaining algebraic coordinates attaining that target. \section{A certified balanced exponent} \label{ba:section} This section proves a lower bound along an unbounded sequence of cardinalities. The assertions for every sufficiently large cardinality are separate. All terminating decimals occurring below denote exact rational numbers. \begin{theorem}\label{ba:main} Put \[ \alpha_* = \frac{10431722305539701}{10^{16}} =1.0431722305539701. \] For every $\varepsilon>0$ there are arbitrarily large integers $N$ and sets of $N$ points and $N$ distinct unit circles in the plane with at least $N^{\alpha_*-\varepsilon}$ incidences. \end{theorem} The proof uses the fields of Theorem~\ref{fi:fields}, the arithmetic bound of Proposition~\ref{ar:arithmetic-ceiling}, and the balanced transfer criterion of Corollary~\ref{tr:balanced-corollary}. The analytic and finite numerical inequalities needed by that criterion are proved here by explicit finite expressions with rigorous remainder bounds. \subsection{The fixed data and the margin} \label{ba:data} Write \begin{equation}\label{ba:constants} \ell=\frac94\log2+\frac12\log3615,\qquad \theta_0=\frac{17179869183}{34359738368},\qquad C_* = \frac{4226263259973019243081}{10^{23}}. \end{equation} Here $\ell$ is the logarithmic root discriminant from \eqref{fi:lambda}, $\theta_0$ is the lower bound for the proportion of complex places of the underlying field, and $C_*$ is an upper bound for the normalized logarithm of the relative zeta value. Let $m=96$, $s=57/50$, and \begin{equation}\label{ba:a} a=\frac{76524029038425969}{10^{22}} =0.0000076524029038425969. \end{equation} Appendix~\ref{ba:witness} specifies the polynomial $P$ by an integer Chebyshev coefficient table. The rational change of basis given there produces its symmetric Bernstein matrix $B=(\beta_{ij})_{0\le i,j\le m}$: \begin{equation}\label{ba:polynomial} P(t,u)=\sum_{i,j=0}^{m}\beta_{ij}b_i^m(t)b_j^m(u), \qquad b_i^m(t)=\binom mi t^i(1-t)^{m-i}. \end{equation} The polynomial has degree at most $96$ in each variable. Its Bernstein matrix may have negative entries; positivity follows from subdivision on the square. For $0<\delta<1$ put $p=2/(1+\delta)$. The complex-place profile is \begin{equation}\label{ba:profile} t=(1+a|z|^2)^{-1},\quad u=(1+a|w|^2)^{-1},\qquad g(z,w)=t^su^sP(t,u),\qquad (z,w)\in\mathbb C^2. \end{equation} Its mass, overlap, and local functional are \begin{align} A_{\mathbb C}(p)&=\int_{\mathbb C^2}g(z,w)^p\,dz\,dw,\label{ba:mass-def}\\ I_{\mathbb C}&=\int_{\mathbb R}\int_{\mathbb C^2} g(z,w)g(z+e^v,w+e^{-v})\,dz\,dw\,dv,\label{ba:overlap-def}\\ J_{\mathbb C}(\delta)&=\log I_{\mathbb C} -(1+\delta)\log A_{\mathbb C}(p). \label{ba:JC-def} \end{align} Lebesgue measure is used on each copy of $\mathbb C=\mathbb R^2$. At a real place use the radial profile on $\mathbb C$ \[ f(z)=\exp(-a_{\mathbb R}|z|^2/p),\qquad a_{\mathbb R}=2p\delta=\frac{4\delta}{1+\delta}. \] Completing a square gives \[ A_{\mathbb R}=\frac{\pi}{a_{\mathbb R}},\qquad I_{\mathbb R}=\frac{\pi p}{2a_{\mathbb R}} e^{-a_{\mathbb R}/(2p)}, \] and hence \begin{equation}\label{ba:JR} J_{\mathbb R}(\delta)=\log(p/2) +\delta\left(\log\frac{2p\delta}{\pi}-1\right), \qquad J_{\mathbb R}''(\delta)=\frac1{\delta(1+\delta)}. \end{equation} The five finite-place choices and their densities per degree are \[ \begin{array}{c|r|c|r|r} \text{rational prime}&Q&\nu_Q&k&\text{weight-matrix size}\\\hline 2&16&1/32&7&64\times64\\ 3&9&1/4&9&64\times64\\ 5&25&1/4&6&64\times64\\ 29&707281&1/4&1&12\times12\\ 41&2825761&1/8&1&12\times12 \end{array} \] At $41$, the local ramification index and residue degree are $(1,4)$. Selecting only one of the two primes of $B$ gives the density $1/8$. The finite recurrence in Appendix~\ref{ba:finite-witness} specifies nonnegative rational weight matrices $w$. These matrices are fixed as $\delta$ varies. Here is an explicit formula for their contribution, in the notation of Lemma~\ref{tr:shell-resources}. Set \[ \mu_0=1,\qquad \mu_i=(Q-1)Q^{i-1}\quad(i\ge1), \] and let $D_{ij}=\mu_{\min(i,j)}$ for $i\ne j$, $D_{00}=0$, and $D_{ii}=i\mu_i-Q^{i-1}$ for $i\ge1$. On pairs of indices define \[ G_{(i,j),(a,b)}=(k+1)\mu_i\mu_j\mathbf1_{i=a}\mathbf1_{j=b} +\mu_jD_{ia}\mathbf1_{j=b} +\mu_iD_{jb}\mathbf1_{i=a}. \] With zero weights beyond the stated matrix sizes, put \begin{equation}\label{ba:finite} A_Q(p)=Q^k\sum_{i,j}\mu_i\mu_jw_{ij}^{p},\qquad I_Q=Q^k\sum_{i,j,a,b}w_{ij}G_{(i,j),(a,b)}w_{ab}, \end{equation} \begin{equation}\label{ba:F-H} \mathcal F(\delta)=\sum_{\text{five types}} \nu_Q\bigl(\log I_Q-(1+\delta)\log A_Q(p)\bigr), \qquad H=\sum_{\text{five types}}\nu_Q k\log Q. \end{equation} Thus the factor $Q^k$ has already been included; there is no additional term $-\delta H$ in the following margin: \begin{align} \Phi(\delta;\theta,C)={}&\mathcal F(\delta) -(\tfrac12-\delta)\ell-C+(1-\delta)\log2 +(1-\theta)\log\pi \notag\\ &+(1-2\theta)J_{\mathbb R}(\delta) +\theta J_{\mathbb C}(\delta). \label{ba:margin} \end{align} The balanced transfer criterion applies when this margin is positive, the profiles are admissible, and the Fourier separation condition holds. The remaining subsections establish these statements with exact data. \subsection{Exact subdivision, positivity, and a Fourier tube} \label{ba:tube-section} We describe the subdivision calculation explicitly. Let $P(t,u)=\sum d_{ij}t^iu^j$, and write $D=(d_{ij})$. The change of basis matrices are \[ T_{ri}=\begin{cases} \binom mi\binom{m-i}{r-i}(-1)^{r-i},&r\ge i,\\0,&ri. \end{cases} \] Then $D=TBT^{\mathsf T}$. For a rational interval $[l,l+h]$, let $U_{rj}=\binom jr l^{j-r}h^r$ for $j\ge r$, and $U_{rj}=0$ otherwise. On a rectangle $[l,l+h]\times[l',l'+h']$, the Bernstein coefficient matrix in the two rescaled variables is \begin{equation}\label{ba:restriction} B_{\square}=(VUT)B(VU'T)^{\mathsf T}. \end{equation} Every operation in this formula is rational. Since Bernstein basis functions are nonnegative and sum to one, the smallest and largest entries of $B_{\square}$ bound $P$ on that rectangle. \begin{lemma}\label{ba:tube} The polynomial in \eqref{ba:polynomial} is positive on $[0,1]^2$. For every \begin{equation}\label{ba:anchor-interval} \delta\in[\delta_0,\delta_1] =[0.04317218,0.04317228] \end{equation} the profile \eqref{ba:profile} satisfies \[ |\widehat{g^p}(\xi_1,\xi_2)| \le K_{\mathbb C}A_{\mathbb C} e^{-\sigma_{\mathbb C}(|\xi_1|+|\xi_2|)}, \qquad K_{\mathbb C}<4,\quad \sigma_{\mathbb C}>1. \] The Fourier transform has phase $e^{-2\pi i x\cdot\xi}$. \end{lemma} \begin{proof} Apply \eqref{ba:restriction} to the uniform $8\times8$ grid. The exact minimum $L$ of all cell coefficients and their maximum $U$ satisfy the exact rational inequalities \[ \frac{8305722}{10^7}0$, form the rational number \begin{equation}\label{ba:q-cell} q_{\square}=\frac1{L_{\square}} \sum_{r+k>0}\binom mr\binom mk8^{r+k} \max_{i,j}|(\Delta_1^r\Delta_2^kB_{\square})_{ij}| \omega^{r+k},\qquad q_0=\max_{\square}q_{\square}. \end{equation} The finite sum has $0\le r,k\le m$. The Bernstein derivative identity bounds $|\partial_t^r\partial_u^kP|/(r!k!)$ by the coefficient in this sum without $\omega^{r+k}$. Taylor's identity for a polynomial is exact also at complex arguments. Therefore, at a real point in $\square$, the complexified polynomial differs from its real value by at most $q_{\square}L_{\square}\le q_0P$. Derivatives need only be bounded on the cell containing the real Taylor centre; the complex point need not stay in that cell. Exact rational evaluation of \eqref{ba:q-cell} gives \[ q_0<0.976994<1. \] Consequently $P$ stays in the open right half-plane throughout the specified tube, and its principal real powers are holomorphic there. Pointwise in that tube, \[ |g(x+iy,x'+iy')^p| \le (1-\rho^2)^{-2sp}(1+q_0)^p g(x,x')^p. \] Here the variables have first been scaled by $\sqrt a$. At the largest $p$, namely $p_0=2/(1+\delta_0)$, outward evaluation gives \begin{equation}\label{ba:tube-numbers} \log K_{\mathbb C} =-2sp_0\log(1-\rho^2)+p_0\log(1+q_0) <1.306742<\log4, \qquad \sigma_{\mathbb C}=\frac{2\pi\rho}{\sqrt a}>1.022100>1. \end{equation} Both upper factors decrease as $p$ decreases. Shift each of the two Fourier integration contours in the direction of its frequency by $\rho/\sqrt a$. The bound just proved supplies an integrable majorant because $sp>1$. Cauchy's theorem, followed by passage from finite rectangles to the full plane, gives the asserted Fourier bound. \end{proof} The real-place Gaussian has Fourier transform $A_{\mathbb R}\exp(-\pi^2|\xi|^2/a_{\mathbb R})$. Since $a_{\mathbb R}\le1$ on \eqref{ba:anchor-interval}, $r-\pi^2r^2/a_{\mathbb R}\le a_{\mathbb R}/(4\pi^2)<\log2$. Thus both profiles are admissible with the common constants \begin{equation}\label{ba:common-fourier} M=2,\qquad\sigma=1. \end{equation} Here the convention in the transfer criterion assigns prefactor $M$ to a real-place block and $M^2$ to a complex-place block; thus the complex bound $K_{\mathbb C}<4$ has exactly the required normalization. The remaining admissibility conditions also follow directly: the profiles are smooth with bounded derivatives; their energies are bounded below and coercive; their masses are finite because $sp>1$. Finiteness of the overlap follows from the reduction below. Finally, $\log^2(1+x)\le C_\eta(1+x)^\eta$ and any $0<\eta0$ define \[ Q_{A,B}(y)=\int_{\mathbb R}(1+ye^{2v})^{-A} (1+ye^{-2v})^{-B}\,dv. \] For $00$, $\alpha+\alpha'>1$, with $T\sim\operatorname{Beta}(\alpha,\alpha')$. Indeed, insert the two Gamma integrals, complete the square in $z$, and put the two Laplace variables equal to $RT$ and $R(1-T)$; the Jacobian is $R$. Tonelli's theorem applies to the nonnegative integrand. Set $\eta=2s-1$. Applying \eqref{ba:laplace} termwise to the finite monomial expansion $D=(d_{ij})$ yields \begin{equation}\label{ba:cross-reduction} Z:=\frac{a^2I_{\mathbb C}}{\pi^2} =\sum_{i,j,k,l}\frac{d_{ij}d_{kl}}{(\eta+i+k)(\eta+j+l)} \mathbb E Q_{\eta+i+k,\eta+j+l}(y), \end{equation} where the independent variables and their product parameter are \[ T\sim\operatorname{Beta}(s+i,s+k),\qquad T'\sim\operatorname{Beta}(s+j,s+l),\qquad y=a\sqrt{T(1-T)T'(1-T')}\le a/4. \] Every expectation is finite: $Q_{A,B}(y)$ is bounded by $\log(1/y)$ plus a constant, and logarithms are integrable under beta laws. Hence the finite signed sum is legitimate. Here is the fully rational coefficient calculation for the truncated sum. Define $D_x(j)=\sum_{h=0}^{j-1}(x+h)^{-1}$, with $D_x(0)=0$, and \begin{align*} K_n(i,k)&=\frac{(s+i)_n(s+k)_n} {n!(\eta+i+k+n)_{n+1}},\\ r_n(i,k)&=-\eta^{-1}+\frac{H_n}2 -\frac{D_\eta(i+k+n)}2+D_\eta(i+k+2n+1) -\frac{D_s(i+n)+D_s(k+n)}2,\\ R_n(i,k)&=K_n(i,k)r_n(i,k),\qquad c_s=-2\psi(s)+\psi(2s)+\eta^{-1}-\gamma,\\ G_n&=\sum_{i,j,k,l}d_{ij}d_{kl}K_n(i,k)K_n(j,l),\\ \widetilde G_n&=\sum_{i,j,k,l}d_{ij}d_{kl} \{R_n(i,k)K_n(j,l)+K_n(i,k)R_n(j,l)\}. \end{align*} The finite expression is \begin{equation}\label{ba:ZN} Z_N=\sum_{n=0}^Na^{2n} \{G_n[\log(1/a)+c_s]+\widetilde G_n\}. \end{equation} To verify it, multiply the beta densities by $(T(1-T))^n$. Their normalizing ratios are $(s+i)_n(s+k)_n/(2s+i+k)_{2n}$, and the tilted logarithmic moments are $\psi(s+i+n)+\psi(s+k+n)-2\psi(2s+i+k+2n)$. Substitution in \eqref{ba:hyperbola-series}, followed by $\psi(x+j)=\psi(x)+D_x(j)$, gives \eqref{ba:ZN}. Put \[ \rho_0=\frac{(\eta+2m)^2a^2}{16},\qquad S_D=\sum_{i,j}|d_{ij}|. \] For the present data, $\rho_0<1$, $\log(4/a)>1/2$, and $\log(4/a)>H_{\lceil\eta+2m\rceil-1}$. The function $y^{2N+2}\log(1/y)$ is increasing on $(0,a/4]$. Consequently \eqref{ba:hyperbola-tail} and \eqref{ba:cross-reduction} give the two-sided, signed-coefficient bound \begin{equation}\label{ba:overlap-error} |Z-Z_N|\le E_N:=\frac{\log(4/a)S_D^2}{\eta^2} \frac{\rho_0^{N+1}}{1-\rho_0}. \end{equation} The absolute values in $S_D$ are essential. Discarding positive cross-remainders alone would prove only a lower bound and would not justify the upper root endpoint below. At $N=45$, exact rational evaluation of the coefficients and outward evaluation of the displayed logarithm and digamma values give \begin{equation}\label{ba:overlap-numbers} \begin{aligned} E_{45}&<1.439\cdot10^{-39},\\ Z&>20687.62742665686375548155206277886824132345,\\ Z&<20687.62742665686375548155206277886824132347. \end{aligned} \end{equation} \subsection{Mass integration with a proved remainder} \label{ba:mass-section} Polar coordinates in each complex variable give, for $q=sp-1>0$, \begin{equation}\label{ba:mass-beta} A_{\mathbb C}=\pi^2a^{-2}q^{-2}\mathcal A(p),\qquad \mathcal A(p)=\mathbb E P(T,U)^p, \quad T,U\text{ independent with density }qt^{q-1}. \end{equation} The partition consists of all $i/64$, $0\le i\le64$, together with \[ 1/512,1/256,1/128,\quad 127/128,255/256,511/512. \] There are $70$ intervals and $2485$ rectangles after symmetry is used. For each rectangle, \eqref{ba:restriction} supplies positive rational bounds $l_\square\le P\le u_\square$. Write \[ c_\square=(l_\square+u_\square)/2,\quad r_\square=(u_\square-l_\square)/(u_\square+l_\square)<1. \] For $10$. These series and their remainder bounds reduce every numerical comparison to finitely many rational inequalities. \subsection{Curvature, arithmetic binding, and an exact root bound} \label{ba:root-section} We need more than two sign tests, because the final decimal lies much closer to the root than the anchor spacing. The requisite curvature bound is independent of numerical differentiation. \begin{lemma}\label{ba:variance} For a fixed positive profile $h$ with finite logarithmic second moment, let $L(p)=\log\int h^p$ and $J(\delta)=\log I-(1+\delta)L(2/(1+\delta))$, with $I$ independent of $\delta$. Then \[ J''(\delta)=-\frac{p^2}{1+\delta} \operatorname{Var}_p(\log h)\le0, \] where the probability density is $h^p/\int h^p$. \end{lemma} \begin{proof} Differentiation under the integral gives $L'=\mathbb E_p\log h$ and $L''=\operatorname{Var}_p(\log h)$. Since $p'=-p/(1+\delta)$, one has $J'=-L+pL'$ and the displayed second derivative. The moment bounds already proved justify these derivatives uniformly on the compact anchor interval. \end{proof} The finite-place variances use the probabilities proportional to $\mu_i\mu_jw_{ij}^p$, ignoring zero weights. Their contribution $K_f$ to $-\mathcal F''$ is enclosed by the finite sums \[ K_f=\sum\frac{\nu_Qp^2}{1+\delta} \left\{\frac{\sum\mu_i\mu_jw_{ij}^p(\log w_{ij})^2} {\sum\mu_i\mu_jw_{ij}^p} -\left(\frac{\sum\mu_i\mu_jw_{ij}^p\log w_{ij}} {\sum\mu_i\mu_jw_{ij}^p}\right)^2\right\}. \] For the complex-place variance, put $X=-s(\log t+\log u)\ge0$ and $Y=\log(P/L)\ge0$. Then $\log g-\log L=-X+Y$, and $(-X+Y)^2\le X^2+Y^2$. On a mass rectangle replace $P^p$ in this numerator by $u_\square^p$ and $Y$ by $\log(u_\square/L)$, and bound its denominator below by the sum of $l_\square^p$ times the rectangle probabilities. The required one-dimensional logarithmic moments have primitives \[ t^q(\log t-1/q),\qquad t^q\{(\log t)^2-2\log t/q+2/q^2\}, \] with limiting value $0$ at $t=0$. The mixed term in $X^2$ factors into the two first moments. Thus this variance bound is again a finite sum with explicit interval endpoints. Uniformly on \eqref{ba:anchor-interval}, these calculations give \begin{equation}\label{ba:curvature-numbers} K_f<2,\qquad K_{\mathbb C}:= \frac{p^2}{1+\delta}\operatorname{Var}_p(\log g)<100. \end{equation} More precisely, the enclosing interval for $K_f$ is \[ [1.5538279509314,1.5538298808880]. \] The explicit upper-bounding expression for $K_{\mathbb C}$ just described lies in \[ [96.937586626011038,96.939887401158926]; \] the upper endpoint therefore bounds $K_{\mathbb C}$. Writing $\Phi(\delta)=\Phi(\delta;\theta_0,C_*)$, equations \eqref{ba:JR} and Lemma~\ref{ba:variance} imply \begin{equation}\label{ba:curvature} -52<\Phi''(\delta)\le \frac{1-2\theta_0}{\delta_0(1+\delta_0)}<10^{-8}. \end{equation} Here are outward rational anchor enclosures using the final $C_*$: \begin{equation}\label{ba:anchor-margins} \begin{array}{c|r|r} &\text{lower endpoint}&\text{upper endpoint}\\ \hline \Phi(\delta_0)&0.000001205125462729712620&0.000001205126954786316853\\ \Phi(\delta_1)&-0.000001178713945287339534&-0.000001178712453226317178 \end{array} \end{equation} The displayed endpoints enclose the finite expressions above and suffice for the root comparison below. Further checks give \begin{equation}\label{ba:other-checks} 15.60265468263737410, \end{equation} and, throughout the anchor interval, \begin{equation}\label{ba:signature} -\log\pi-2J_{\mathbb R}(\delta)+J_{\mathbb C}(\delta)>0.6247407>0. \end{equation} For the last inequality, $J_{\mathbb C}$ is concave by Lemma~\ref{ba:variance}, so its minimum on the interval is bounded below by the smaller enclosed endpoint value; the elementary Gaussian term is enclosed on the whole interval. It follows that increasing $\theta$ from $\theta_0$ can only improve the margin. The anchor bounds use the fixed coefficient table and finite recurrences in the appendices, together with the arithmetic constant $C_*$ of Proposition~\ref{ar:arithmetic-ceiling}. The constant enters \eqref{ba:margin} through the single term $-C_*$. \begin{lemma}\label{ba:exact-root} The function $\Phi$ has a unique zero $\delta_*$ in $[\delta_0,\delta_1]$, and \[ 1.04317223055397014<1+\delta_*<1.04317223055403600. \] In particular, $1+\delta_*>\alpha_*+10^{-17}$. \end{lemma} \begin{proof} Put $h=\delta_1-\delta_0=10^{-7}$ and let $a_i^-,a_i^+$ be the two endpoints in row $i$ of \eqref{ba:anchor-margins}. Their signs give existence. By \eqref{ba:curvature}, \[ \Phi'(\delta)\le\frac{a_1^+-a_0^-}{h}+52h<-20, \] so the zero is unique. A function with second derivative between $-K$ and $k$ differs from its endpoint chord by an amount between $-kh^2/8$ and $Kh^2/8$. This follows by subtracting the corresponding quadratic with zero endpoint values and using convexity or concavity. The zero of the line through $a_0>0>a_1$ is $\delta_0+ha_0/(a_0-a_1)$, increasing in both endpoint values. The inverse-slope bound above therefore gives \[ \delta_0+\frac{ha_0^-}{a_0^--a_1^-}-\frac{10^{-8}h^2}{160} \le\delta_*\le \delta_0+\frac{ha_0^+}{a_0^+-a_1^+}+\frac{52h^2}{160}. \] Substituting the four rational endpoints in \eqref{ba:anchor-margins} and clearing denominators yields the displayed bounds. More precisely, one plus the lower endpoint is greater than $1.04317223055397014902826$, while one plus the upper endpoint is less than $1.04317223055403598960517$. \end{proof} \begin{proof}[Proof of Theorem~\ref{ba:main}] Take $\delta=\alpha_*-1+10^{-17}$. Lemma~\ref{ba:exact-root} and strict decrease give $\Phi(\delta;\theta_0,C_*)>0$. Theorem~\ref{fi:fields} and Proposition~\ref{ar:arithmetic-ceiling} supply arbitrarily large admissible field degrees with $\theta\ge\theta_0$ and the required arithmetic bound. Equation~\eqref{ba:signature} preserves the positive margin for their actual signatures. The exact finite weights, the admissibility proof, and \eqref{ba:other-checks} verify every remaining hypothesis of Corollary~\ref{tr:balanced-corollary}. Its concentration and lattice selection argument, with loss $10^{-17}/2$, produces an unbounded sequence with incidence count at least $N^{\alpha_*+5\cdot10^{-18}}$. Hence the quotient by $N^{\alpha_*}$ tends to infinity, proving \eqref{eq:main-balanced} as well as the stated $\varepsilon$ form. Replacing each point of such a unit-distance configuration by the unit circle centred there gives the required balanced incidence configuration; the factor between ordered distances and unordered edges does not change the exponent. Adding isolated points or circles, if necessary, makes both cardinalities exactly $N$. \end{proof} This proves the sequence construction from explicit profiles and finite expressions with stated remainder bounds. \section{Asymmetric profiles and two independent cardinality bounds} \label{as:section} All logarithms in this section are natural. We write $U(n,m)$ for the largest number of unit-distance pairs between sets $P,Q\subset\mathbb C$ with $|P|=n$ and $|Q|=m$. Equivalently, $Q$ is the set of centres of $m$ distinct unit circles and $P$ is a set of $n$ distinct points. An element of $P\cap Q$ is permitted. A pair is ordered by its two sides, so that no factor $1/2$ is introduced when passing from this bipartite graph to incidences. In particular, $U(n,m)=U(m,n)$. This section uses the fields of Theorem~\ref{fi:fields}, the relative-unit mass identity~\ref{fi:mass}, and the arithmetic input stated in \ref{fi:arithmetic-input}. The numerical values used below are \begin{equation}\label{as:constants} \begin{aligned} \ell&=\frac94\log2+\frac12\log3615,\qquad C=\frac{4226263259973019243081}{10^{23}},\\ \theta_*&=\frac{17179869183}{34359738368},\qquad (\sigma,M_F)=(30,4). \end{aligned} \end{equation} The degrees are $d=[F:\mathbb Q]=2^k$, for every sufficiently large integer $k$. Preserving the specified finite quotient requires $k\ge67$; this is not an effective threshold for all the subsequent asymptotic estimates. The integer $96$ occurring in the balanced polynomial certificate is a polynomial degree, not a factor in the field degrees. \subsection{Local masses, overlaps and true endpoint energies} \label{as:resources} At a complex-place pair, choose $a,b>0$, $s,t>0$, $s+t>1$, counting powers $p_A,p_B>0$, two symmetric polynomials $P_A,P_B$ with rational coefficients which are strictly positive on $[0,1]^2$, and $\lambda_B\ge0$. Put \[ T_a(z)=\frac1{1+a|z|^2},\qquad \phi(x,y)=(1-x)(1-y), \] and define the two edge densities \begin{align} g_A(z,w)&=T_a(z)^sT_a(w)^sP_A(T_a(z),T_a(w)),\\ g_B(z,w)&=T_b(z)^tT_b(w)^tP_B(T_b(z),T_b(w)) e^{\lambda_B\phi(T_b(z),T_b(w))}. \label{as:profile} \end{align} The counting densities are $g_A^{p_A}$ and $g_B^{p_B}$. All examples have $sp_A>1$ and $tp_B>1$, so their counting masses are finite. The condition $s+t>1$ gives finiteness of the overlap, also in an open neighbourhood of the chosen exponents. Positivity and boundedness of the polynomial and exponential factors reduce these assertions to the corresponding product-Student assertions proved below. Let $x=(z,w)$, $e(u)=(e^u,e^{-u})$, and set \begin{align} I&=\int_{\mathbb R}\int_{\mathbb C^2}g_A(x)g_B(x+e(u))\,dx\,du, & A_A&=\int_{\mathbb C^2}g_A^{p_A}, & A_B&=\int_{\mathbb C^2}g_B^{p_B}.\label{as:overlap-masses} \end{align} All endpoint expectations in this section refer to the probability measure $I^{-1}g_A(x)g_B(x+e(u))\,dx\,du$. In particular, put \begin{align} E_A&=\mathbb E\{\log(1+a|z|^2)+\log(1+a|w|^2)\},\\ E_B&=\mathbb E\{\log(1+b|z+e^u|^2)+\log(1+b|w+e^{-u}|^2)\},\\ v_A&=\mathbb E[-\log g_A(x)] =sE_A-\mathbb E\log P_A,\\ v_B&=\mathbb E[-\log g_B(x+e(u))] =tE_B-\mathbb E\log P_B-\lambda_B\mathbb E\phi_B. \label{as:true-energies} \end{align} The complex-pair resources are \begin{equation}\label{as:pair-resources} y_A=\log A_A+p_Av_A,\qquad y_B=\log A_B+p_Bv_B,\qquad z=\log I+v_A+v_B. \end{equation} Different counting powers and different endpoint energies are essential. For example, replacing $E_A,E_B$ by expectations under the counting mass laws would not give the cardinality bounds in Theorem~\ref{tr:transfer}. Neither equality of the endpoint energies nor an Euler--Lagrange stationarity assumption is used. At each real place of $F$, which gives one complex coordinate of $K$, use the edge density $f(z)=e^{-\alpha|z|^2}$ on both sides and the counting density $h(z)=e^{-|z|^2/1000}$, where $\alpha=2/3999$. The edge law has $\mathbb E|z|^2=1/(2\alpha)+1/4=1000$. Direct Gaussian integration gives \begin{equation}\label{as:real-resources} y_R=\log(1000\pi)+1,\qquad z_R=\log(3999\pi/4)+1. \end{equation} Indeed $\int f(z)f(z+1)\,dz=\pi e^{-\alpha/2}/(2\alpha)$ and the sum of the two edge energies is $2000\alpha$. For clarity, the finite-place formula used in the computation is recorded here as well. For a shell pair with residue cardinality $Q$, common period parameter $k$, shell measures $m_0=1$ and $m_i=Q^i-Q^{i-1}$, let $G$ be the Gram matrix of Lemma~\ref{tr:shell-resources}. For nonnegative rational matrices $w_A,w_B$ define \begin{align} J&=w_A^{\mathsf T}G w_B,&I_f&=Q^kJ,\\ e_A&=-J^{-1}\sum_{i,j}(w_A)_{ij}(Gw_B)_{ij}\log(w_A)_{ij},& e_B&=-J^{-1}\sum_{i,j}(w_B)_{ij}(Gw_A)_{ij}\log(w_B)_{ij},\\ A_{f,A}&=Q^k\sum_{i,j}m_im_j(w_A)_{ij}^{q_A},& A_{f,B}&=Q^k\sum_{i,j}m_im_j(w_B)_{ij}^{q_B},\\ X_{f,A}&=\log A_{f,A}+q_Ae_A,& X_{f,B}&=\log A_{f,B}+q_Be_B,\\ Z_f&=\log I_f+e_A+e_B,&H_f&=k\log Q. \label{as:finite-resources} \end{align} Here matrices are vectorized in the quadratic form, $q_A,q_B>0$, and $0\log0=0$. The cross-Gram numerator and the two marginal numerators are computed with exact integer arithmetic after clearing rational denominators. Thus~\eqref{as:finite-resources} is valid also when the matrices do not solve any optimization problem. The selected finite-place densities per degree are \begin{equation}\label{as:finite-densities} \begin{array}{c|ccccc} \text{rational prime}&2&3&5&29&41\\ \hline Q&16&9&25&29^4&41^4\\ \text{number of selected places}/d&1/32&1/4&1/4&1/4&1/8. \end{array} \end{equation} At $41$ only the places above the designated prime of $\mathbb Q(\sqrt{241})$ are selected. Their actual absolute local type is $(1,4)$, and their density is $1/8$ as displayed in \eqref{as:finite-densities}. For one complete local-profile specification, let $X_{f,A},X_{f,B},Z_f,H$ now denote the sums of the local quantities with the weights in~\eqref{as:finite-densities}. A signature-safe triple is \begin{align} X_A&=\log2-\ell+X_{f,A} +\max_{\vartheta\in\{\theta_*,1/2\}} \{(1-2\vartheta)y_R+\vartheta y_A\},\\ X_B&=\log2-\ell+X_{f,B} +\max_{\vartheta\in\{\theta_*,1/2\}} \{(1-2\vartheta)y_R+\vartheta y_B\},\\ Z&=2\log2-\frac32\ell-C+Z_f +\min_{\vartheta\in\{\theta_*,1/2\}} \{(1-\vartheta)\log\pi+(1-2\vartheta)z_R+\vartheta z\}. \label{as:global-resources} \end{align} These are affine signature expressions before taking maxima or minima. For every profile in Appendix~\ref{sc:profiles} $y_A,y_B>2y_R$ and $z-2z_R-\log\pi>0$, both checked with intervals. Consequently the point caps use $y_A/2,y_B/2$, while the edge lower bound uses $\theta_*$. Thus the point bounds and the edge bound use the indicated signature endpoints. The term $-C$ occurs once, and only once, in $Z$; changing a certified arithmetic upper bound therefore changes $Z$ affinely without requiring another local integral evaluation. By Theorem~\ref{tr:transfer}, an admissible specification with these resources gives, for the field of degree $d$, a common-translate configuration with \begin{equation}\label{as:resource-interpretation} |P|\le e^{dX_A+o(d)},\qquad |Q|\le e^{dX_B+o(d)},\qquad I(P,Q)\ge e^{dZ-o(d)}. \end{equation} There is no division of the edge lower bound by a random endpoint count. The geometric Fourier condition is \begin{equation}\label{as:fourier-global} 30\cdot2e^{H/2-\ell}>\log4+2\log5+2. \end{equation} Since $H$ is the normalized logarithm of the \emph{whole} finite period, this condition is imposed on a complete mixture, not separately on every individual finite-place factor. \subsection{Exact analytic evaluation of the asymmetric profiles} \label{as:analytic} We use $\Gamma$ for Euler's gamma function, $\psi=\Gamma'/\Gamma$, $(x)_n=\Gamma(x+n)/\Gamma(x)$, and $H_m=\sum_{j=1}^m j^{-1}$ with $H_0=0$. We give formulas and explicit remainder bounds, so the numerical assertions below are finite checks rather than appeals to numerical quadrature. The same formulas cover the product, finite-factor, exponential and general-polynomial profiles in all the stated witnesses. \begin{lemma}[One-coordinate Feynman integral]\label{as:feynman} Let $s,t>0$, $\nu=s+t-1>0$, and $a,b>0$. For $h\ge0$ put \[ L(x)=ax+b(1-x),\quad D(x)=\frac{abx(1-x)}{L(x)},\quad w(x)=\frac{x^{s-1}(1-x)^{t-1}}{L(x)}. \] Then \[ \int_{\mathbb C}(1+a|z|^2)^{-s}(1+b|z+h|^2)^{-t}\,dz =\frac{\pi\Gamma(\nu)}{\Gamma(s)\Gamma(t)} \int_0^1 w(x)(1+D(x)h^2)^{-\nu}\,dx. \] Moreover $D(x)\le D_*:=ab/(\sqrt a+\sqrt b)^2$ and $D(x)\ge\min(a,b)x(1-x)$. \end{lemma} \begin{proof} Insert the beta identity for the product of the two negative powers. Completing the square leaves $L(x)|z+b(1-x)h/L(x)|^2+1+abx(1-x)h^2/L(x)$. The planar radial integral is $\pi L(x)^{-1}(1+D(x)h^2)^{1-s-t}/(s+t-1)$. The beta prefactor gives the displayed formula. Maximizing $abx(1-x)/(ax+b(1-x))$ gives $D_*$; bounding $L(x)$ by $\max(a,b)$ gives the lower bound. All interchanges are justified by nonnegativity. \end{proof} For two possibly different coordinate exponent pairs $(s_i,t_i)$, $i=1,2$, put $\nu_i=s_i+t_i-1>0$ and use the corresponding functions $w_i,D_i$. Integration in the hyperbolic variable reduces to \begin{align} Q_{\nu_1,\nu_2}(y) &=\frac12\int_0^\infty(1+yx)^{-\nu_1}(1+y/x)^{-\nu_2}\frac{dx}{x}\\ &=\sum_{n=0}^\infty\frac{(\nu_1)_n(\nu_2)_n}{(n!)^2}y^{2n} \left[-\log y+\psi(n+1) -\frac{\psi(\nu_1+n)+\psi(\nu_2+n)}2\right],\quad 0 H_{\lceil\nu\rceil-1}$ is explicitly performed. This also explains why $0<\nu<1$ is permitted; it is not silently excluded. For polynomial factors, write $P_A(x,y)=\sum A_{ij}x^iy^j$ and $P_B(x,y)=\sum B_{kl}x^ky^l$. Their overlap is the finite signed sum \begin{equation}\label{as:polynomial-contraction} \sum_{i,j,k,l}A_{ij}B_{kl} \mathcal J(s+i,t+k;s+j,t+l). \end{equation} Finite summation is exact; a common absolute error for a scalar overlap is multiplied by $\|A\|_1\|B\|_1$. No favourable cancellation is used to shrink a remainder. Equivalently, the same finite sum can be grouped as $A^{\mathsf T}KB$ and $A^{\mathsf T}FB$. All shifted gamma and digamma values may be obtained from $\Gamma(x+1)=x\Gamma(x)$ and $\psi(x+1)=\psi(x)+1/x$. For the exponential factor let $E_L(v)=\sum_{j=0}^L\lambda_B^jv^j/j!$. On $0\le v\le1$, \begin{equation}\label{as:exp-tail} 0\le e^{\lambda_Bv}-E_L(v) \le R_L:=e^{\lambda_B}\lambda_B^{L+1}/(L+1)!. \end{equation} Since the unmodified polynomial densities are positive, if $I_0$ is their overlap and $I_L$ is the finite polynomial contraction with $E_L(\phi)$, then $I_L\le I\le I_L+R_LI_0$. This inequality is applied also at the perturbed exponent values needed for the next lemma. \begin{lemma}[Rigorous endpoint expectations]\label{as:expectations} Holding the polynomial coefficients, scales and $\lambda_B$ fixed, \[ E_A=-\partial_s\log I,\qquad E_B=-\partial_t\log I. \] For any sufficiently small exact rational $h>0$, $-\partial_s\log I$ is enclosed by \[ -\frac{\log I(s+h,t)-\log I(s,t)}h \le E_A\le -\frac{\log I(s,t)-\log I(s-h,t)}h, \] and likewise for $E_B$. \end{lemma} \begin{proof} The overlap is finite in an open neighbourhood, so differentiation under the integral is justified by exponential moment domination of the logarithmic energies. Its logarithm is convex separately in $s,t$ by H\"older's inequality. The derivative identity and the two secant bounds follow. There is no factor $1/2$: the two exponents are varied separately. \end{proof} To evaluate $\mathbb E\log P$, suppose exact Bernstein subdivision proves $01$, polar coordinates followed by $x=T_a(z),y=T_a(w)$ give \begin{equation}\label{as:mass-beta} \int_{\mathbb C^2}g^p =\frac{\pi^2}{a^2(V-1)^2} \mathbb E_{\mu_V\otimes\mu_V} [P(x,y)^pe^{p\lambda(1-x)(1-y)}], \qquad d\mu_V(x)=(V-1)x^{V-2}\,dx. \end{equation} Thus all polynomial integrals are explicit moments. On a cell $[a_0,b_0]\times[c_0,d_0]$, \[ \int_{a_0}^{b_0}x^j\,d\mu_V(x) =\frac{V-1}{V-1+j} (b_0^{V-1+j}-a_0^{V-1+j}), \] with the analogous formula in $y$. If Bernstein coefficients on the cell give $0p$, \begin{equation}\label{as:mass-binomial} P^p=c_C^p\sum_{j=0}^N\binom pj(P/c_C-1)^j+\mathcal E_C, \qquad |\mathcal E_C|\le c_C^p\frac{|\binom p{N+1}|r_C^{N+1}}{1-r_C}. \end{equation} After the first omitted term the ratio of absolute binomial terms is $r_C(j-p)/(j+1)p$. For arbitrary $q<1$, including the negative values below $-1$ in our fourth-degree examples, use instead \begin{equation}\label{as:centered-factor} 1-q\phi=(1-q/2)\left[1-\frac{q}{2-q}(2\phi-1)\right], \qquad \left|\frac q{2-q}\right|<1. \end{equation} The same uniform binomial remainder now applies on $|2\phi-1|\le1$. Convolving the finite expansions proves the factorized counting certificates without any restriction $q>-1$. Explicitly, if factor $i$ is bounded by $B_i=\max(1,(1-q_i)^p)$ and its approximation has uniform error $\varepsilon_i$, the total product error is at most $\prod_i(B_i+\varepsilon_i)-\prod_iB_i$, by expanding the product of the true factors plus their errors. For logarithmic expectations one may sum the individual log-factor expansions and their uniform remainders. We next justify the complex contour shifts used for the count densities. For a shift of a real two-vector by $ih$, with $|h|\le\sigma/(2\pi)$, put $\rho=a\sigma^2/(4\pi^2)<1$. If $R=a|z|^2$ and $R_c=a\sum(z_j+ih_j)^2$, then \begin{equation}\label{as:contour-elementary} |1+R_c|\ge(1-\rho)(1+R),\qquad |T_c-T|\le \frac{2\sqrt\rho\,T^{3/2}\sqrt{1-T}+\rho T^2}{1-\rho}. \end{equation} The first follows from $\operatorname{Re}(1+R_c)=1+R-a|h|^2$; the second follows by subtracting the two reciprocals. For completeness we describe an exact finite Fourier certificate for any polynomial used in~\eqref{as:profile}. Express $P$ in a Bernstein basis on each rational cell $C=[a_0,b_0]\times[c_0,d_0]$. Its minimum and maximum Bernstein coefficients bound $P$ there. Finite differences of the coefficients give upper bounds \[ D_{ij,C}\ge\sup_C\frac{|\partial_x^i\partial_y^jP|}{i!j!}. \] This follows by differentiating the Bernstein basis; for bidegree $(m,n)$ the factor is $\binom mi\binom nj$ divided by the side lengths to the powers $i,j$. Let $\delta_x,\delta_y$ be cellwise upper bounds from \eqref{as:contour-elementary}, and set \begin{align} R_C&=L_C^{-1}\sum_{i+j>0}D_{ij,C}\delta_x^i\delta_y^j,\\ D_C&=(1-c_0)\delta_x+(1-a_0)\delta_y+\delta_x\delta_y. \end{align} The polynomial Taylor expansion is finite, so these bounds hold even though the displaced arguments are complex. If every $R_C<1$, then $|P(T_c,U_c)-P(T,U)|0$ and \[ 1-q\phi(T,U)=\frac{1+R+S+cRS}{(1+R)(1+S)}. \] Require $1-2\rho+c\rho^2>0$ and $1-c\rho>0$. On writing $R_c=U-\rho,S_c=V-\rho$, the numerator is $A+B(U+V)+cUV$ with $A,B>0$ and $\operatorname{Re}U,\operatorname{Re}V\ge0$. The real part of $(A+BU)/(B+cU)$ is positive, since its numerator after multiplication by the conjugate denominator is $AB+(Ac+B^2)\operatorname{Re}U+Bc|U|^2$. Hence the numerator cannot vanish. Moreover $|R_c|\le R+\rho$, giving a per-coordinate logarithmic cost at most \[ p\{\log(1+\max(1,c)\rho)-\log(1-\rho)\}. \] Alternatively let $\delta=((3\sqrt3/8)\sqrt\rho+\rho)/(1-\rho)$ and $D=2\delta+\delta^2$. Since $|\phi_c-\phi|\le D$, the cost is at most \[ \frac p2\log\left(1+\frac{|q|D}{\min(1,1-q)}\right). \] The smaller of the two bounds may be used, retaining the zero-free test. Sum these costs over the factors, add $-sp\log(1-\rho)$ for the base Student density, and add $p\lambda D/2$ for an exponential factor. These inequalities give the stated Fourier bounds. All these contour bounds dominate the displaced count density by a fixed multiple of its integrable real counterpart, uniformly throughout the tube. For each complex coordinate block $\mathbb C\simeq\mathbb R^2$ with $\xi_j\ne0$, rotate its real plane so that the first axis is parallel to $\xi_j$, and shift this one real variable by $-i\sigma/(2\pi)$. Equivalently the real two-vector is shifted by $ih_j$, where $h_j=-\sigma\xi_j/(2\pi|\xi_j|)$; it is not necessary or correct to shift both real scalar coordinates by that length. For $\xi_j=0$ no shift is needed. Cauchy's theorem applies to each truncated rectangle because the chosen logarithms are holomorphic in the tube. To remove its vertical sides, integrate their absolute values over the other real variables. The uniform domination bounds them by a fixed constant times an integrable real slice function at the two rectangle endpoints. There is a sequence of endpoints tending to infinity on which the sum of these two slice functions tends to zero. Along that sequence the vertical integrals vanish; dominated convergence handles the horizontal integrals. Fubini's theorem permits the shifts to be made successively in all blocks. The Fourier exponential supplies $e^{-\sigma|\xi_j|}$ in block $j$. Consequently \[ |\widehat\Psi(\xi)|\le M_F^d e^{-\sigma\sum_j|\xi_j|}\int\Psi. \] The real count densities are smooth and have bounded derivatives, as is also immediate from their positive polynomial factors on the compact $T$-square and their Student decay. Together with \eqref{as:fourier-global}, these are exactly the uniform lattice-count hypotheses of Theorem~\ref{tr:transfer}. \subsection{Frozen witnesses and numerical inequalities} \label{as:witnesses} A complete primitive consists of one archimedean specification and one finite shell specification for each column of~\eqref{as:finite-densities}. A phase is a finite list of primitives with nonnegative rational weights summing exactly to one. A phase resource is the weighted sum of the primitive resources~\eqref{as:global-resources}; its $H$ is the same weighted sum of the finite period logarithms. Independent mixtures of archimedean and finite choices are represented by their Cartesian product, so this definition imposes no artificial coupling of those choices. Appendix~\ref{sc:profiles} gives the thirteen archimedean choices $A_1,\ldots,A_{13}$ and every exact phase proportion. The finite-place choices $F_1,\ldots,F_{27}$ are defined by the parameter table and fixed recurrence in Appendix~\ref{sc:finite-profiles}. Together these specify all local functions in the four all-scale rows, the row with $r=1.0603071$, and the sequence row with $r=1.0629395$. All entries outside the stated finite matrix sizes are zero. \begin{algorithm}[Evaluating an asymmetric phase]\label{sc:asymmetric-evaluation} Fix one of the phases listed in Appendix~\ref{sc:profiles}. \begin{enumerate} \item For each pair $(A_i,F_j)$ in that phase, take the archimedean parameters from $A_i$ and construct the five finite-place matrices from $F_j$ by Algorithm~\ref{sc:finite-generation}. \item Compute the finite Gram sums in \eqref{as:finite-resources}. The matrices are rational, so the Gram sums are exact. Bound the logarithms and powers by the series described in Section~\ref{ba:verification}. \item Calculate the archimedean overlap from \eqref{as:inner-series}--\eqref{as:outer-tail}. At each truncation add the displayed upper bound for the omitted terms. For an exponential factor, also include the error in \eqref{as:exp-tail}. \item Calculate the two endpoint expectations using the secant bounds of Lemma~\ref{as:expectations} and the polynomial-logarithm formula \eqref{as:log-chebyshev}. Calculate the counting masses by the beta-moment formulas in Section~\ref{as:count-fourier}. \item Average the local quantities with the exact phase proportions. Insert $\ell,C,\theta_*$ in \eqref{as:global-resources} to obtain intervals for $X_A,X_B,Z$. \item For phase number $j$, bound the three differences $jA-X_A$, $jrA-X_B$, and $Z-jpA$ from below. Finally check the Fourier inequality \eqref{as:fourier-global} for the whole phase. \end{enumerate} \end{algorithm} The tables below give the resulting intermediate values and lower bounds for the three differences. \begin{table}[htbp] \centering\small \begin{tabular}{c|c|c|c|c} $r$&$p$&$A$&certified slack&Fourier gap\\\hline $1$&$1.041169920$&$16.788682174888262$&$7.98\,10^{-8}$&$30.91608$\\ $1.02$&$1.048990216$&$15.354937578311516$&$6.91\,10^{-8}$&$13.59661$\\ $1.04$&$1.054938332$&$13.957152859335954$&$8.05\,10^{-8}$&$3.50131$\\ $1.06$&$1.060225408$&$12.77577761760308$&$5.19\,10^{-8}$&$0.16350$\\ $1.0603071$&$1.060307177$&$12.759652001254373$&$7.12\,10^{-8}$&$0.16350$ \end{tabular} \caption{Two-phase witnesses. At phase $j=1,2$ the three resource slacks are $jA-X_A$, $jrA-X_B$, and $Z-jpA$. The last column bounds $30\cdot2e^{H_{\min}/2-\ell}-\log4-2\log5-2$ from below.} \label{as:table} \end{table} The resource enclosures for the explicitly generated shell matrices at the final arithmetic constant are as follows. Every coordinate in the next table has absolute error less than $10^{-14}$. The two triples in each row correspond to the low and high phases. \begin{equation}\label{as:displayed-resources} \begin{array}{c|rrr} r&X_A&X_B&Z\\\hline 1&16.78868158843516031&16.78868158843516311&17.47987129259717701\\ &33.57736376332853005&33.57736376332851334&34.95974183376258108\\[2pt] 1.02&15.35493718282736379&15.66203593439365147&16.10717960172946562\\ &30.70987476114307411&31.32407226427460953&32.21435864299422203\\[2pt] 1.04&13.95715205046627547&14.51543816483972122&14.72393608412320808\\ &27.91430490980325812&29.03087713854989154&29.44787119439221925\\[2pt] 1.06&12.77577752622962604&13.54232418328576750&13.54520419129381841\\ &25.55155514383284223&27.08464845794506936&27.09040812622815656\\[2pt] \end{array} \end{equation} For $r=1.0603071$, each coordinate below has absolute error less than $10^{-10}$: \begin{align} &(12.7596514059164,\ 13.5291490151536,\ 13.5291510087870),\\ &(25.5193034072574,\ 27.0582986256356,\ 27.0583012571363).\label{as:critical-resources} \end{align} The corresponding low/high values of $H$, each with absolute error less than $10^{-8}$, are \begin{equation}\label{as:H-table} \begin{array}{c|rr} 1&10.37313535827153&20.54980773330726\\ 1.02&9.13486236690564&18.48399167887957\\ 1.04&7.74967445154824&17.04332318075847\\ 1.06&6.94792967872249&15.60265468263737\\ 1.0603071&6.94792967872249&15.60265468263737. \end{array} \end{equation} Substituting the exact $A,r,p$ values of Table~\ref{as:table} into the displayed intervals verifies all three strict inequalities. The logarithmic counting and edge exponents are normalized by $d$. The one-phase witness has \begin{equation}\label{as:sequence-data} r_s=1.0629395,\qquad p_s=1.062939573,\qquad A_s=17.884565041128177, \end{equation} and its resource triple is enclosed by \[ (17.8845647411425,\ 19.0102103225485,\ 19.0102119817337)+[-3.1\cdot10^{-9},3.1\cdot10^{-9}]^3. \] Its three resource slacks exceed $5.23\cdot10^{-8}$, and its Fourier gap exceeds $45.23154$. It uses the general-polynomial profile \eqref{as:profile}, with \begin{align*} s&=1.2643704818394643,&t&=0.3569370831364012,\\ \log a&=-6.960105659675447,&\log b&=-6.8719742698834,\\ \lambda_B&=0.7958371960711847,& p_A&=\frac{2621439999847412109375}{2093974328671816122368},\\ &&p_B&=\frac{13107199999237060546875}{3256468905406373298176}. \end{align*} The exact polynomial coefficient matrices are listed as profile $A_{13}$ in Appendix~\ref{sc:profiles}. Its two finite specifications have period vectors $(5,7,5,1,0)$ and $(5,7,4,1,0)$, and exact proportions $192596700618323/10^{15}$ and $807403299381677/10^{15}$. The polynomial counting and cell-Fourier bounds above apply directly to these matrices. Algorithm~\ref{sc:asymmetric-evaluation} gives all the inequalities in Table~\ref{as:table} directly from the parameter tables and the stated series bounds. \subsection{Dyadic degree interpolation and all integer sizes} \label{as:dyadic} \begin{theorem}[Two phases give all sufficiently large sizes] \label{as:allsize} Suppose two finite mixtures of admissible primitives from one fixed finite list have a common Fourier pair $(\sigma,M_F)$ and satisfy, for some $A>0,r\ge1,p>0$, \begin{equation}\label{as:two-phases} X_{A,j}jpA, \qquad j=1,2. \end{equation} Suppose also that their whole-period Fourier inequalities are strict. Then, for all sufficiently large integers $N$, there are exactly $N$ points and $\lceil N^r\rceil$ distinct unit circles with $\Omega(N^p)$ incidences. \end{theorem} \begin{proof} Let $\delta>0$ be smaller than every slack in~\eqref{as:two-phases}. For $0\le t\le1$, assign proportions $1-t,t$ to the low and high phase profiles, respectively. Since all logarithmic local resources are additive in the numbers of places, the mixed resources satisfy \[ X_A\le(1+t)A-\delta,\quad X_B\le r(1+t)A-\delta,\quad Z\ge p(1+t)A+\delta. \] The whole-period condition is preserved: it is equivalent to the linear inequality $H>2\ell+2\log((\log M_F+2\log5+2)/(2\sigma))$. Choose a sufficiently large available dyadic degree $d$ such that $Ad\le\log N<2Ad$, and put $t=\log N/(Ad)-1$. Allocate the actual complex-place pairs and the selected finite places according to these proportions, rounding to integers and correcting the last part to keep the total fixed. The number of profile types is fixed, so this changes the sum of local logarithmic quantities by $O(1)$, uniformly in $t$. There is no need for every part to contain a growing number of places. The four total energies in Theorem~\ref{tr:transfer} are sums of independent variables, each drawn from a fixed finite family with finite second moment. Consequently their variances are bounded by $Kd$ with one $K$ for all $t$. Joint concentration therefore holds uniformly. The arithmetic estimate is uniform over this field family as well. Choose the energy-band tolerance small compared with $\delta$, and then choose $d$ large. The uniform $o(d)$ terms and the $O(1)$ allocation errors are absorbed by the strict reserve. Formula~\eqref{as:resource-interpretation} gives $|P|\le N$, $|Q|\le\lceil N^r\rceil$, and at least $cN^p$ cross unit distances, with $c>0$ independent of $N$. Append distinct points and centres far from the finite existing configuration until the prescribed cardinalities are exact. Existing edges remain. Making one unit circle per centre completes the construction. \end{proof} A fixed one-phase witness gives a different quantifier. Taking $N_j=\lceil e^{A_sd_j}\rceil$ and $M_j=\lceil N_j^{r_s}\rceil$, its strict slacks give exact sets of those sizes by the same padding argument. Thus~\eqref{as:sequence-data} proves \begin{equation}\label{as:sequence-theorem} U(N_j,\lceil N_j^{1.0629395}\rceil) =\Omega(N_j^{1.062939573}),\qquad N_j\longrightarrow\infty. \end{equation} This one-phase construction gives the second sequence in Theorem~\ref{thm:sequences}; the two-phase hypothesis of Theorem~\ref{as:allsize} gives the all-scale conclusion. \subsection{Physical mixing of different ratios and the exact envelope} \label{as:concave} The next interpolation is at the level of local functions in the same number field; it would not follow from a list of unrelated incidence lower bounds. \begin{theorem}[Uniform interpolation in the size ratio] \label{as:mixing} Let finitely many two-phase witnesses satisfy Theorem~\ref{as:allsize}, with parameters $(A_i,r_i,p_i)$, one common finite primitive list, and common strict Fourier bounds. For any $\beta_i\ge0$ with $\sum_i\beta_i=1$ put \begin{equation}\label{as:physical-mix} A=\left(\sum_i\frac{\beta_i}{A_i}\right)^{-1},\qquad \lambda_i=\frac{\beta_i/A_i}{\sum_j\beta_j/A_j},\qquad r=\sum_i\beta_i r_i,\qquad p=\sum_i\beta_i p_i. \end{equation} Then $U(N,\lceil N^r\rceil)\ge cN^p$ for all sufficiently large $N$, with $c>0$ and the size threshold uniform over the entire mixture simplex. In particular, on a closed interval between two successive upper-concave-hull vertices $(r_i,p_i)$ and $(r_{i+1},p_{i+1})$, \begin{equation}\label{as:monomial-segment} U(N,M)\ge cN^{a_i}M^{b_i},\qquad b_i=\frac{p_{i+1}-p_i}{r_{i+1}-r_i},\quad a_i=p_i-b_ir_i, \quad N^{r_i}\le M\le N^{r_{i+1}}. \end{equation} \end{theorem} \begin{proof} At both the low and high ends mix witness $i$ with physical proportion $\lambda_i$. The identity $\lambda_iA_i=A\beta_i$ gives, for $j=1,2$, \[ \sum_i\lambda_iX_{A,i,j}jpA. \] If $\delta$ is the minimum of all original positive raw slacks, each mixed slack is at least $\sum_i\lambda_i\delta=\delta$. The Fourier threshold is linear in $H$ and hence also survives mixing. Next perform the low/high interpolation of Theorem~\ref{as:allsize}. The quantities $A$ and the primitive variances remain in fixed compact ranges, the number of possible primitives is finite, and every operation has the same positive reserve. The proof of that theorem is therefore uniform in $\beta$, including boundary points where some parts vanish. It applies even if $r$ depends on $N$. For two adjacent vertices choose $\beta_{i+1}=(r-r_i)/(r_{i+1}-r_i)$, $\beta_i=1-\beta_{i+1}$. Then $p=a_i+b_ir$. Substituting $r=\log M/\log N$ gives $N^p=N^{a_i}M^{b_i}$. Exact ceilings or floors change neither the uniform constant nor the argument, since the strict exponential reserves absorb the bounded cardinality corrections. \end{proof} The upper concave hull of the five rows of Table~\ref{as:table} has exactly four vertices, at \[ r=1,\quad 1.02,\quad 1.04,\quad 1.0603071. \] The $r=1.06$ row is below the chord joining the last two vertices: that chord takes the exact value \[ \frac{53825287754393}{50767750000000} >1.060225408 \quad\hbox{at }r=1.06. \] All these comparisons use rational arithmetic on the terminating witness targets. The three segment pairs $(a_i,b_i)$ are exactly \begin{equation}\label{as:three-segments} \left(\frac{8126939}{12500000},\frac{977537}{2500000}\right),\quad \left(\frac{7456363}{10000000},\frac{1487029}{5000000}\right),\quad \left(\frac{39597848504393}{50767750000000}, \frac{1073769}{4061420}\right). \end{equation} Their slopes are strictly decreasing. To extend past the last ratio, let $(r_*,p_*)=(1.0603071,1.060307177)$. For $M\ge s^{r_*}$ construct an all-size seed with $x=\lceil M^{1/r_*}\rceil$ left vertices and $\lceil x^{r_*}\rceil$ right vertices. Retain $M$ right vertices of largest degree, then retain $s$ left vertices of largest degree. Each step keeps at least the corresponding fraction of all current edges, since the mean of the largest degrees is at least the overall mean. As $\lceil x^{r_*}\rceil=O(M)$, this yields \begin{equation}\label{as:tail-ray} U(s,M)\ge c sM^{(p_*-1)/r_*},\qquad \frac{p_*-1}{r_*}=\frac{60307177}{1060307100}. \end{equation} This is an actual deletion operation in a planar configuration; it does not tensor or formally multiply incidence graphs. \begin{corollary}[A global four-monomial lower bound] \label{as:global-bound} Let $s=\min(n,m)$ and $M=\max(n,m)$. There is an absolute $c>0$ such that for all positive integers $n,m$, \begin{align} U(n,m)\ge c\biggl[n+m+\min\biggl\{& s^{8126939/12500000}M^{977537/2500000},\\ &s^{7456363/10000000}M^{1487029/5000000},\\ &s^{39597848504393/50767750000000}M^{1073769/4061420},\\ &sM^{60307177/1060307100}\biggr\}\biggr]. \label{as:four-monomials} \end{align} The nonlinear term is strictly larger in power order than $M=s^r$ for \[1\le r<\frac{1060307100}{999999923}=1.060307181643\ldots.\] This is a constructive lower bound, not a matching upper bound for arbitrary unit-circle incidences. \end{corollary} \begin{proof} The first three lines in logarithmic coordinates describe the concave polygonal chain through the four hull vertices. Appending the line $1+((p_*-1)/r_*)r$ preserves strict decrease of slopes. Such a concave polygonal function is the minimum of its extended affine pieces, giving the four monomials by Theorem~\ref{as:mixing} and~\eqref{as:tail-ray}. The tail line meets $p=r$ at $r=r_*/(1+r_*-p_*)$, the stated rational number. One point and arbitrarily many distinct unit circles through it give $U(n,m)\ge M$ after padding. Taking the better of the two constructions and decreasing the constant gives their sum. For bounded $s$, the nonlinear minimum is at most $sM^{(p_*-1)/r_*}=O(M)$, so a further constant adjustment covers all remaining small sizes. \end{proof} \subsection{A fully specified \texorpdfstring{$N$ versus $N^{1.04}$}{N versus N to the power 1.04} recipe} \label{as:worked-example} Use the $r=1.04$ row, with $A=13.957152859335954$ and $p=1.054938332$. For an arbitrary sufficiently large integer $N$ choose \begin{equation}\label{as:recipe-degree} d=2^{\lfloor\log_2(\log N/A)\rfloor},\qquad t=\frac{\log N}{Ad}-1. \end{equation} Assign low and high phases with proportions $1-t,t$. Their archimedean parameters are as follows; all displayed decimals are exact inputs. \begin{center}\small \begin{tabular}{c|rr} &low phase&high phase\\\hline $s$&$1.3325274080489524$&$1.2128575878994745$\\ $t$ in the profile exponent&$0.53875810170025862$&$0.30695909643943042$\\ $\log a$&$-7.2898212608618529$&$-10.185246038705765$\\ $\log b$&$-7.1737906334714499$&$-10.105937104613458$\\ $p_A$&$1.3678970737119656$&$1.2281456888253364$\\ $p_B$&$3.1221335967204542$&$4.5512896290149945$ \end{tabular} \end{center} Here the symbol $t$ for an endpoint exponent is distinct from the interpolation proportion in~\eqref{as:recipe-degree}. Both phases use $\lambda_B=0$ and $P=\prod_{q\in\mathbf q}(1-q\phi)$. The exact factor vectors are \begin{align*} \mathbf q_{A,\mathrm{low}}={}&( .88036964413350938,-1.2599682068584213,\\ &\quad .21830494292067165,.13926299502740697),\\ \mathbf q_{B,\mathrm{low}}={}&( .72924190769296271,-.24792205322667021,\\ &\quad -.26836424936085079,-.28090971334204096),\\ \mathbf q_{A,\mathrm{high}}={}&( .85224580275921558,-2.0550667048347195,\\ &\quad .28466325079340971,.18710522486521236),\\ \mathbf q_{B,\mathrm{high}}={}&( .63561360296940816,-.30103276870604978,\\ &\quad -.29459829242192803,-.19660203638447701). \end{align*} The three low-phase finite specifications have exact proportions \[ \left(\frac{222724515886789}{500000000000000}, \frac{313697877428409}{10^{15}}, \frac{240853090798013}{10^{15}}\right) \] and period vectors $(5,7,4,0,0)$, $(5,7,5,0,0)$, $(5,7,4,0,0)$, in the prime order of~\eqref{as:finite-densities}. The high-phase proportions are \[ \left(\frac{16116346494661}{250000000000000}, \frac{6607013992789}{50000000000000}, \frac{100424291770697}{125000000000000}\right), \] with period vector $(8,10,7,1,1)$ for all three specifications. These are the finite choices $F_{12},F_{13},F_{14}$ in the low phase and $F_{15},F_{16},F_{17}$ in the high phase. Their counting powers and finite recurrences are given in Appendix~\ref{sc:finite-profiles}. The archimedean choices are $A_6$ and $A_7$, respectively, with the exact phase mixtures displayed in Appendix~\ref{sc:profiles}. Choose the field $K/F$ of degree $d$; allocate its complex-place pairs and each selected finite-place type with the above proportions, multiplied by $1-t$ or $t$. Use the two counting/edge-energy windows of Theorem~\ref{tr:transfer} on the \emph{same} $S$-integer lattice, the same relative-unit parameter and the same translate. A common translate with enough norm-one differences exists by that theorem, while the two point-count caps hold uniformly for every translate. Project both sets to a compact complex coordinate, and take the right-hand images as unit circle centres. Padding gives exactly $(N,\lceil N^{1.04}\rceil)$, with $\Omega(N^{1.054938332})$ incidences. For example, at the symbolic scales $N_k=\lceil\exp(20.935729289003931\,2^k)\rceil$ the two phases have proportions $1/2+o(1)$ and $1/2+o(1)$. This is only an illustrative subsequence; formula~\eqref{as:recipe-degree} covers every sufficiently large $N$. For an intermediate ratio, such as $r=1.03$, first apply \eqref{as:physical-mix} to the $1.02$ and $1.04$ witnesses, then apply \eqref{as:recipe-degree} to their common $A$. This gives $p(1.03)=1.051964274$ and the monomial $N^{0.7456363}M^{0.2974058}$ on the whole closed range $N^{1.02}\le M\le N^{1.04}$. The logarithmic formulas above determine the allocation from $(n,m)$. The condition $k\ge67$ gives the required field quotient; the asymptotic threshold absorbs concentration and rounding errors. The field and common-translate construction in Section~\ref{tr:section} then specifies the coordinates. Finally, the balanced degree-$96$ construction has a direct, weaker asymmetric consequence requiring no new local analysis. If its sets $U_j$ have $N_j$ vertices and at least $N_j^{\alpha-\varepsilon}$ unordered unit pairs, fix $r\ge1$, put $n_j=\lceil N_j^{1/r}\rceil$, and choose the $n_j$ vertices of largest unit-distance degree as the left side. Keep all $N_j$ vertices as centres and pad to $m_j=\lceil n_j^r\rceil$. Since the sum of the degrees is twice the number of unordered edges, this gives \[ U(n_j,m_j)\ge\frac{2n_j}{N_j}u(U_j) =\Omega(n_j^{\,1+r(\alpha-1)-\varepsilon}), \] after renaming $\varepsilon$. This gives the stated sequence of asymmetric configurations. The all-scale bounds use the independent left and right profiles constructed above. \section{An arbitrarily small twenty-circle configuration} \label{core:section} The construction in this section supplies the geometric primitive used in the exact-tail results. A zero in a limiting determinant will only be used to exclude additional contacts. Every contact that is claimed to exist is separately forced by an exact distance identity. \begin{proposition}\label{core:twenty} For every $\eta>0$ there are twenty distinct centres, of diameter less than $\eta$, whose unit circles have exactly sixty triple points and no point on four circles. Every pair of circles meets in two distinct points. The centres admit an order such that the number of triple points among the first $h$ circles, for $0\le h\le20$, is \begin{equation}\label{core:tau} (\tau_0,\ldots,\tau_{20})= (0,0,0,1,2,4,8,11,14,17,20,24,28,32,36,40,44,48,52,56,60). \end{equation} In particular every prefix increment is at most $4$. \end{proposition} We identify the plane with $\mathbb C$. All subscripts of centres in this appendix start at $0$. Opposite centres are always taken relative to the common origin of the construction. \subsection{A nonsingular ten-centre seed} \label{core:seed} Let $T_0$ be the real root of \begin{equation}\label{core:cubic} H(T)=T^3+54T+108=0, \qquad A_0=-T_0^2/18+2T_0/3+4. \end{equation} The derivative $H'(T)>0$ and $H(-2)<00,\qquad \det\begin{pmatrix} 2(e_1-e_0)&2(u_1-u_0)\\ 2(e_2-e_0)&2(u_2-u_0) \end{pmatrix} =40+\frac83T_0+\frac49T_0^2>0. \end{equation} These equalities, including the rational derivatives in \eqref{core:IFT-certificate}, follow by rational substitution and reduction as detailed in Section~\ref{core:program}. No decimal approximation to the root is used in that calculation. We justify that this limiting calculation gives actual circles. For a pair of centres with distinct $b$-coefficients, their difference has the form $\epsilon h(\epsilon)$ with $h(0)\ne0$. Replace the quotient of that difference by its norm by $h(\epsilon)/|h(\epsilon)|$, choosing the sign that gives the desired branch. The ordinary formula for intersections of equal circles then shows that the selected branch is real analytic at $\epsilon=0$. Conjugating $\epsilon$ to $-\epsilon$ conjugates all centres and branches, so the squared-distance equations are even in $\epsilon$. Their constant terms vanish; division by $\epsilon^2$ is therefore analytic. The six resulting equations in $X,Y,U,V,A,T$ have nonsingular Jacobian at zero: elimination of the two blocks in \eqref{core:blocks} leaves exactly the determinant in \eqref{core:IFT-certificate}. The real analytic implicit-function theorem supplies exact parameter functions for all sufficiently small $\epsilon$. They are even by uniqueness. Thus the coefficient pairs computed above remain the leading pairs of actual centres, with the prescribed six contacts for $w,-w$ and six for $z,-z$. \subsection{Excluding all unprescribed contacts} \label{core:exclusion} A third circle is incident to the branch in \eqref{core:point-jet} only if its coefficient pair lies on the line through the first two transformed pairs \[ (b_j,c_j-e b_j^2/2). \] Indeed the coefficient of $\epsilon^2$ in $|p-d_j|^2-1$ is $2e(v-c_j)+(u-b_j)^2$. Consequently every actual triple with branch sign $e$ satisfies the necessary limiting equation \begin{equation}\label{core:determinant} \det\begin{pmatrix} 1&b_i&c_i-e b_i^2/2\\ 1&b_j&c_j-e b_j^2/2\\ 1&b_k&c_k-e b_k^2/2 \end{pmatrix}=0. \end{equation} If the determinant is nonzero, the corresponding actual contact is absent for every sufficiently small positive $\epsilon$. Here are all the limiting triples for the first ten centres. An entry $ijk$ denotes the unordered set $\{i,j,k\}$. \begin{equation}\label{core:seed-inventory} \begin{array}{c|lllll} e=-1&012&037&045&068&135\\ &146&234&256&478&679\\ \hline e=1 &015&024&036&123&147\\ &169&257&345&379&678 \end{array} \end{equation} The exact rational enumeration in Section~\ref{core:program} gives precisely this list, and no collinear group of four transformed pairs. Every listed triple is an exact contact already prescribed: eight are the original unit-vector contacts, six come from $w,-w$, and six from $z,-z$. Since all other signed determinants are nonzero, these are all the actual triple points for sufficiently small $\epsilon$, and there are no fourfold points. The exact check also gives $b_i\ne b_j$ for all distinct indices; in particular all centres are distinct. \subsection{An exact four-contact identity and five additions} \label{core:additions} Let $d,-d,z,-z$ be four centres of sufficiently small norm. Let $n_1$ be perpendicular to $z-d$, and $n_2$ perpendicular to $z+d$, with positive real parts and with lengths \[ |n_1|^2=1-|z-d|^2/4,\qquad |n_2|^2=1-|z+d|^2/4. \] Their selected orientations give $n_1,n_2\longrightarrow1$ as $\epsilon\longrightarrow0$. Put $D=n_1-n_2$. The circle centred at $D$ contains the following four existing pair-intersection points: \begin{equation}\label{core:four-points} \frac{d+z}{2}+n_1,\quad -\frac{d+z}{2}+n_1,\quad \frac{d-z}{2}-n_2,\quad -\frac{d-z}{2}-n_2. \end{equation} For example, subtracting the first point from $D$ gives $-(d+z)/2-n_2$, of norm $1$ by perpendicularity. The other three identities are identical applications of the two displayed length relations. The reflected centre $-D$ gives the reflected contacts. If all eight cross-pair intersections of $\{d,-d\}$ and $\{z,-z\}$ are ordinary, adding $D,-D$ therefore forces eight distinct new triple points. This assertion is an exact norm identity, not an inference from a limiting determinant. For coefficient pairs $(b,c)$ of $d$ and $(V,U)$ of $z$, set \[ R_1=\frac{U-c}{V-b},\qquad R_2=\frac{U+c}{V+b}. \] Expansion of the two oriented normals gives the coefficient pair of $D$: \begin{equation}\label{core:augmentation} b_D=-R_1+R_2,\qquad c_D=-\frac{(V-b)^2}{8}-\frac{R_1^2}{2} +\frac{(V+b)^2}{8}+\frac{R_2^2}{2}. \end{equation} Thus $D=O(\epsilon)$, and both the construction and its exact algebraic check can be iterated. Perform the following five additions: \[ \begin{array}{c|c|r} \text{parent indices }(i,j)&\text{new indices }D,-D &\text{triple points after addition}\\ \hline (2,8)&10,11&28\\ (0,10)&12,13&36\\ (1,10)&14,15&44\\ (6,10)&16,17&52\\ (0,14)&18,19&60 \end{array} \] At each step the rational check establishes three assertions. The four cross-pairs of parents belong to no old triple, for either sign; all the new $b$-coefficients remain distinct; and the complete signed inventory consists of the old inventory together with exactly the eight forced triples. It also excludes every fourfold group. The signed list of newly forced triples is particularly simple. If $i',j'$ are the opposite indices and $k,k'$ are the new opposite indices, it is \[ \begin{array}{c|llll} \text{centre }k&(+,i,j,k)&(+,i',j',k)&(-,i,j',k)&(-,i',j,k)\\ \text{centre }k'&(-,i',j',k')&(-,i,j,k')& (+,i',j,k')& (+,i,j',k'). \end{array} \] This list follows directly from \eqref{core:four-points} and reflection. Section~\ref{core:program} excludes every signed triple outside this forced list at each stage. \begin{proof}[Proof of Proposition~\ref{core:twenty}] Choose a positive $\epsilon$ small enough for the seed implicit functions and all five additions. There are finitely many nonzero limiting determinants, so a single sufficiently small choice excludes every unprescribed triple and fourfold point. All sixty prescribed triples hold by the exact identities above. Distinct leading $b$-coefficients give distinct centres, and all centres have norm $O(\epsilon)$. Taking $\epsilon$ smaller makes their diameter less than $\min\{\eta,2\}$. Hence every pair of unit circles crosses in two distinct points. Deleting centres cannot create a new triple; the prefix counts are therefore obtained by counting which of the sixty forced triples lie in the initial segment. The exact enumeration below gives \eqref{core:tau}. For example the full arrangement has $20\cdot19-3\cdot60=200$ ordinary double points, since a triple point uses three of the $20\cdot19$ pairwise intersection occurrences. \end{proof} \subsection{A short modular verification} \label{core:program} All nonvanishing statements needed for the twenty-centre construction can be checked in a small residue ring. This avoids large rational coefficients while keeping every input in the paper. Set \[ q=43550027,\qquad t=10001. \] Then $H(t)=103\cdot223\cdot q$. Reduce the coefficient pairs $(b_j,c_j)$ from the preceding formulas modulo $q$, sending $T_0$ to $t$. The ten pairs below determine all twenty: the opposite indices are $(0,3),(1,4),(2,5),(6,7),(8,9),(10,11),(12,13),(14,15), (16,17),(18,19)$, and opposite coefficient pairs have opposite residues. \begin{equation}\label{core:modular-jets} \begin{array}{r|r|r@{\qquad}r|r|r} j&b_j&c_j&j&b_j&c_j\\\hline 0&33013073&8157915&10&15899737&14658938\\ 1&10526953&20077072&12&18988445&3023102\\ 2&33023071&1697946&14&18737466&18438607\\ 6&33023077&23427946&16&42220276&15928316\\ 8&22138069&35882666&18&12631489&31859312 \end{array} \end{equation} The denominator checks are equally short. The two elimination quantities in \eqref{core:blocks} reduce to $43520030$ and $5768908$. Each is relatively prime to $q$, as are $2,3$, and the product $B_{20}$ defined below. The remaining divisors in the coefficient formulas are powers of $2$ and $3$, individual $b$-coefficients, and differences or sums of two $b$-coefficients. An individual $b$-coefficient is half its difference from its opposite, and a sum is a difference from an opposite coefficient. The formulas therefore reduce unambiguously modulo $q$. Define $\mathcal T_{10}$ to be the twenty signed triples in \eqref{core:seed-inventory}. At each of the five additions, append the eight signed triples displayed after \eqref{core:augmentation}; let $\mathcal T_n$ denote the resulting list at size $n=10,12,14,16,18,20$. For $i0\quad(i\ne j). \] This finite system over $\Q$ has a real solution. The transfer theorem for real closed fields says that an existential formula over $\Q$ which holds in $\R$ holds in the real closure of $\Q$. Applying that theorem gives the required solution. Additional unit pairs cause no difficulty, because the graph $F$ may be retained as a subgraph. \end{proof} \begin{lemma}[A point--unit-circle bound]\label{ub:ip:incidence} For $m$ distinct points and $l$ distinct unit circles, \[ I\le 16m^{2/3}l^{2/3}+16m+l. \] Only the existence of an absolute constant in this inequality will be needed later. \end{lemma} \begin{proof} We first recall a short proof of the crossing estimate needed here. For a simple graph with $v$ vertices and $e\ge4v$ edges, take a drawing with the minimum number $c$ of crossings. Such a drawing can be chosen with no crossings of adjacent edges: a crossing of two edges with a common endpoint is removed by interchanging their initial portions and then separating the coincident portions locally. It can also be chosen with no three edge interiors meeting at one point. Removing one edge at each crossing leaves a planar simple graph, so $c\ge e-3v$ by the planar Euler bound. Independently retain each vertex with probability $p=4v/e$. Apply the same planar bound to the induced drawing and take expectations. The four endpoints of every crossing are distinct, whence \[ p^4c\ge p^2e-3pv, \qquad c\ge \frac{e^3}{64v^2}. \] Discard circles incident to at most one of the specified points; their total contribution is at most $l$. On every remaining circle draw the cyclic arcs between successive incident points. A circle with two incident points contributes its two complementary arcs. The resulting multigraph has $m$ vertices and $E$ edges, where $E$ equals the number of incidences on the retained circles. A given pair of points lies on at most two unit circles, and each of these contributes at most two arcs with that pair of endpoints. Thus edge multiplicity is at most four. Retain one arc for each unordered pair of endpoints. The resulting simple graph has $e\ge E/4$ edges and a drawing with at most $2\binom l2\le l^2$ crossings. Multiple circle intersections may be separated by local perturbations, with at most one crossing for each pair of arcs at each former intersection. Intersections which are graph vertices do not contribute crossings. If $e<4m$, then $E<16m$. Otherwise the crossing estimate gives $e^3\le64m^2l^2$, and consequently $E\le16m^{2/3}l^{2/3}$. Adding back the discarded incidences proves the assertion. Empty point or circle sets are immediate. \end{proof} \begin{lemma}[A finite cutting]\label{ub:ip:cutting} Let $P$ be a finite point set and let $\mathcal C$ consist of $l\ge1$ distinct unit circles. For $R\ge1$ there is a multiset of $s=O(R\log(2+l))$ sampled circles such that, after deleting points on the sampled circles, the remaining points can be assigned to $O(1+s^2)$ open regions, each meeting at most $l/R$ circles of $\mathcal C$. Each assigned point belongs to its assigned region. All constants here are absolute. \end{lemma} \begin{proof} Fix a horizontal direction satisfying the following finite list of genericity conditions. A point of $P$ and a circle-intersection point have different horizontal coordinates unless they are the same point. Also a point of $P$ has an extremal horizontal coordinate of a circle only if it is that extremal point. For the second condition the excluded directions satisfy $u\cdot(p-c)=\pm1$, where $u$ is a unit direction and $c$ is a circle centre; there are finitely many such directions. The first condition also excludes only finitely many. Sample $s=\lceil C R\log(2+l)\rceil$ circles independently and uniformly. Repetitions may be ignored when forming their arrangement. Write each sampled circle as its upper and lower semicircle graphs. Their endpoints and intersections are the events of a vertical sweep. Between events, the gaps between consecutive graphs are open vertical regions. Continue a gap across an event precisely when the same two boundary graphs remain consecutive and its closed vertical cross section contains no event point. Otherwise terminate it and begin a new gap on the other side. This procedure produces open regions bounded by at most two semicircle graphs and two vertical event lines, with unbounded boundaries allowed. They avoid all sampled circles. The direction conditions ensure that an input point on an event line which is not in one of these open regions is itself an event point, and hence is already deleted. There are $O(1+s^2)$ regions. Indeed extrema create $O(s)$ events. If $c_x\ge2$ sampled circles meet at an intersection $x$, only $O(c_x)$ new gaps can start there. Since $c_x\le2\binom{c_x}{2}$ and each circle pair intersects at most twice, the sum of these contributions is $O(s^2)$. This argument includes multiple intersections and coincidences of extrema and intersections. Every possible region has a description involving at most six original circles: two boundary graphs and at most two circles for each vertical endpoint. Semicircle signs and choices of events contribute only an absolute factor, so there are at most $C_1(1+l)^6$ candidate regions. A fixed candidate meeting more than $l/R$ circles is avoided by all samples with probability at most $\exp(-s/R)$. An actual region necessarily avoids every sampled circle. A union bound therefore shows that the probability of any actual region violating the conflict bound is at most $C_1(1+l)^6\exp(-s/R)$. Choose the absolute $C$ so this is less than $1/10$. Some sample has the required property. \end{proof} \subsection{Fixed-parameter regularization} \begin{proposition}\label{ub:ip:regular} Fix $02^{-112}$, the bound $1\le2^{112}s$ suffices. This proves the asserted global bound, including $s=0$. \end{proof} The \emph{wedge law} chooses $q$ with law $\kappa$ and, conditional on $q$, chooses two neighbours $p,r$ independently and uniformly. Thus \begin{equation}\label{ub:ip:wedge} \Omega(p,q,r)=\frac{\mathbf1_{pq\in E}\mathbf1_{rq\in E}} {|E|d(q)}. \end{equation} Its two point marginals are $\mu$, and each point--centre marginal is $\omega$. \begin{lemma}[Finite small-set estimates]\label{ub:ip:rare} Let $0<\alpha\le1$. If $F\subseteq P$ has mass $h\le\alpha$ and $G\subseteq Q$ has arbitrary mass $j$, then \[ \omega(F,G)\le2(\alpha^{1/25}+\xi)\sqrt{hj}. \] If also $j\le\alpha$, then \[ \omega(F,G)\le(\alpha^{2/25}+\xi)(h+j). \] For every $0d\}$ has mass at most $h/d$. Splitting according to $G_d$ gives \[ \E_\kappa g^2\le dh+\E_\kappa[g\mathbf1_{G_d}] =dh+\omega(F,G_d). \] Apply \eqref{ub:ip:marginal-expansion} with $hj\le h^2/d\le\alpha^2/d$ and $\sqrt{hj}\le h/\sqrt d$. The case $h=0$ is immediate. \end{proof} \begin{lemma}[Variation of an edge bit]\label{ub:ip:bitvariation} Give every edge a bit. Let $V_P$ and $V_Q$ be its disagreement probabilities in independent two-edge stars based at a point and at a centre, respectively. If $f$ is the smaller of the two global bit probabilities, then, for $\xi\le2^{-112}$, \[ f\le2^{112}(V_P+V_Q). \] \end{lemma} \begin{proof} Label each vertex by a conditional majority bit, breaking ties by a fixed rule. If the conditional bit parameter is $\rho$, the chance of disagreement with this label is $\min(\rho,1-\rho)$, at most $2\rho(1-\rho)$. Thus the two endpoint labels differ on an edge with probability at most $v=V_P+V_Q$. If $v\le2^{-112}$, the local part of Lemma~\ref{ub:ip:labels} gives a common label whose total marginal exceptional mass is at most $2v$. The edge bit differs from that label with probability at most $3v$, which bounds $f$. If $v>2^{-112}$, use $f\le1/2\le2^{112}v$. \end{proof} \begin{lemma}[Information in a wedge]\label{ub:ip:entropy} Let $\pi(p,r)=\sum_q\Omega(p,q,r)$, and let \[ D=\sum_{p,r}\pi(p,r)\log\frac{\pi(p,r)}{\mu(p)\mu(r)}. \] For the graph of Proposition~\ref{ub:ip:regular}, \begin{align} \pi(p,r)&\le\frac2{|E|d_{Q,\min}}\quad(p\ne r), &\Pr(p=r)&=\frac{|Q|}{|E|},\label{ub:ip:wedge-atoms}\\ 0&\le D\le2a\log t\quad\text{for sufficiently large }t. \label{ub:ip:entropy-bound} \end{align} \end{lemma} \begin{proof} Two distinct real points have at most two common unit-circle centres. Summing \eqref{ub:ip:wedge} therefore gives the off-diagonal estimate. On the diagonal, \[ \sum_p\pi(p,p)=\sum_q\sum_{p:pq\in E}\frac1{|E|d(q)} =\frac{|Q|}{|E|}. \] Put $e_*=30b$. The bounds of Proposition~\ref{ub:ip:regular} imply \[ |P|\le t^{1+2a+e_*},\quad |E|d_{Q,\min}\ge t^{2+3a-2e_*},\quad |Q|/|E|\le t^{-a+2e_*}. \] For a finite random variable $X$ let $\mathcal H(X)=-\sum_x\Pr(X=x)\log\Pr(X=x)$. Dropping the nonnegative diagonal contribution to $\mathcal H(p,r)$ and using the off-diagonal atom bound gives \[ \mathcal H(p,r)\ge(1-t^{-a+2e_*}) \bigl((2+3a-2e_*)\log t-\log2\bigr). \] On the other hand $\mathcal H(p),\mathcal H(r)\le\log|P|$. Expanding the definition gives $D=\mathcal H(p)+\mathcal H(r) -\mathcal H(p,r)$. Hence \[ \frac D{\log t} \le a+4e_*+(2+3a)t^{-a+2e_*}+\frac{\log2}{\log t} \le2a \] for sufficiently large $t$, since $4e_*=120bv(x)\}$ and set $p=u(A)$, $q=v(A)$. Then $\operatorname{TV}(u,v)=p-q$. Data processing gives a lower bound by \[ B(p,q)=p\log(p/q)+(1-p)\log((1-p)/(1-q)). \] For fixed $00. \] Histories of zero mass are irrelevant. \begin{lemma}[Prediction from a strict prefix]\label{ub:ip:prediction} Suppose each test also specifies vector maps $f_T:[M]\to\R^d$ and $g_T:Q\to\R^d$ with norms at most $B$ and $B'$ respectively. Let $H\ge1$ be an integer, take $H$ independent tests of the same law, and choose a step uniformly from $\{1,\ldots,H\}$. Use only its strict prefix as the point history. For a fresh independent test put $m_{T,h}=\sum_x\mu_h(x)f_T(c_T(x))$. The average, over prefixes and the fresh test, of \[ \left|\E_{(p,q)\sim\omega} \left\langle f_T(c_T(p))-m_{T,\mathscr H(p)},g_T(q)\right\rangle\right| \] is at most $4BB'(D/H)^{1/4}$. In particular one may fix a history, before the fresh test is drawn, with at most $M^{H-1}$ values for which the same bound holds after averaging only the fresh test. \end{lemma} \begin{proof} Freeze all $H$ tests temporarily. Apply the map recording the complete sequence of labels to both coordinates of $(p,r)$. The actual law is its image under $\pi$; the reference law is its image under $\mu\otimes\mu$. Their relative entropy is at most $D$. At step $j$ let $\mathscr H_j$ be the strict prefix. Given prefix values $(h,h')$, denote the actual law of the next two labels by $A_{j,h,h'}$. The reference conditional law is \[ B_{j,h,h'}=(c_{T_j})_*\mu_h\otimes(c_{T_j})_*\mu_{h'}. \] Every actual prefix of positive probability also has positive reference probability. The chain rule in Lemma~\ref{ub:ip:information} gives \[ \sum_{j=1}^H\E_\pi D_{\rm KL}(A_{j,\mathscr H_j(p),\mathscr H_j(r)} \Vert B_{j,\mathscr H_j(p),\mathscr H_j(r)})\le D. \] Consequently Pinsker and Cauchy--Schwarz give \begin{equation}\label{ub:ip:average-tv} \frac1H\sum_{j=1}^H\E_\pi \operatorname{TV}(A_{j,\mathscr H_j(p),\mathscr H_j(r)}, B_{j,\mathscr H_j(p),\mathscr H_j(r)})\le\sqrt{\frac D{2H}}. \end{equation} For each step define the residual \[ U_j(p)=f_{T_j}(c_{T_j}(p))- \sum_x\mu_{\mathscr H_j(p)}(x)f_{T_j}(c_{T_j}(x)). \] Its norm is at most $2B$. Under the reference conditional product, the two residuals have mean zero and are independent, so their inner product has expectation zero. Its absolute value is bounded by $4B^2$. The bounded-expectation estimate and \eqref{ub:ip:average-tv} therefore yield \[ \frac1H\sum_j\E_\pi\langle U_j(p),U_j(r)\rangle \le8B^2\sqrt{\frac D{2H}}. \] Each summand is nonnegative. More precisely, conditional independence of the two endpoints of a wedge proves the identity \[ \E_\pi\langle U_j(p),U_j(r)\rangle =\E_{q\sim\kappa}\left\|\E_{p\mid q}U_j(p)\right\|^2. \] It follows that \[ \left|\E_\omega\langle U_j(p),g_{T_j}(q)\rangle\right| \le B'\left(\E_\pi\langle U_j(p),U_j(r)\rangle\right)^{1/2}. \] Average over $j$ and use Cauchy--Schwarz again. The result is at most $B'(8B^2\sqrt{D/(2H)})^{1/2}\le4BB'(D/H)^{1/4}$. Now average over the previously frozen tests. The residual at step $j$ depends only on the strict prefix and the current test; the latter is independent and identically distributed with a fresh test. Thus it may be replaced by such a fresh test without changing the law. Finally choose one prefix whose conditional average discrepancy is no greater than the overall average. Its number of values is at most $M^{H-1}$. \end{proof} \begin{corollary}[Acceptance of finite labels]\label{ub:ip:acceptance} Suppose a test assigns a set of accepted labels $A_T(q)\subseteq[M]$ to each centre, and let $s_0=\E_T\Pr_\omega(c_T(p)\notin A_T(q))$. There is a fixed history with at most $M^{H-1}$ values such that a prediction $\widehat p\sim\mu_{\mathscr H(p)}$ satisfies \[ \E_T\Pr(c_T(\widehat p)\notin A_T(q)) \le s_0+4\sqrt M(D/H)^{1/4}. \] The probability on the left first samples $(p,q)$ with law $\omega$ and then samples the prediction conditional only on $\mathscr H(p)$. The test is independent of this entire experiment. \end{corollary} \begin{proof} Use $f_T(c)=e_c\in\R^M$ and let $g_T(q)$ be the indicator vector of $A_T(q)$. Their norms are at most $1$ and $\sqrt M$. The difference between actual and predicted acceptance probabilities is exactly the inner product discrepancy in Lemma~\ref{ub:ip:prediction}. \end{proof} \begin{lemma}[The prediction coupling]\label{ub:ip:coupling} For any fixed history and integer $r\ge1$, define the joint law \[ \Pr(p,\widehat p,q_1,\ldots,q_r) =\mu(p)\mu_{\mathscr H(p)}(\widehat p) \prod_{j=1}^r\omega(q_j\mid p). \] Both point marginals are $\mu$; every $(p,q_j)$ has law $\omega$; every $q_j$ has law $\kappa$; and \[ \Pr(\widehat p=p)\le |\mathscr H(P)|\max_p\mu(p). \] With a fresh independent test, the probability of any of the $r$ acceptance failures in Corollary~\ref{ub:ip:acceptance} is bounded by $r$ times that corollary's bound. \end{lemma} \begin{proof} All marginal claims follow by summation. In particular, if $\mathscr H(x)=h$, then \[ \sum_{p:\mathscr H(p)=h}\mu(p)\mu_h(x)=\mu(x). \] The collision probability equals \[ \sum_{h:\mu(\mathscr H=h)>0} \frac{\sum_{p:\mathscr H(p)=h}\mu(p)^2}{\mu(\mathscr H=h)} \le |\mathscr H(P)|\max_p\mu(p). \] Every pair $(\widehat p,q_j)$ has the required prediction law, so the last assertion is a union bound; independence of failures is not asserted or needed. \end{proof} \begin{corollary}[Prediction of a strict cross]\label{ub:ip:crossprediction} A test now assigns each vertex a state $y_T(x)=(A_T(x),B_T(x))$ in a finite product of label sets. For a state $y=(u,v)$ put \[ \mathcal R_y=\{(u',v'):\text{exactly one of }u'=u,\ v'=v \text{ holds}\}. \] For integers $H\ge1$, $L\ge2$, there is a fixed history with at most $L^{2H}$ values such that, writing $\nu_{h,T}=(y_T)_*\mu_h$, \begin{equation}\label{ub:ip:cross-error} \E_T\E_\omega[1-\nu_{\mathscr H(p),T}(\mathcal R_{y_T(q)})] \le \E_T\Pr_\omega(y_T(p)\notin\mathcal R_{y_T(q)}) +\frac4L+18(D/H)^{1/4}. \end{equation} If the tests instead have a finite measure of total mass $W$, and $m(T)$ bounds their actual failure probability, the integrated bound is \begin{equation}\label{ub:ip:cross-integrated} \int\E_\omega[1-\nu_{\mathscr H(p),T}(\mathcal R_{y_T(q)})]\,dT \le\int m(T)\,dT+W\left(\frac4L+18(D/H)^{1/4}\right). \end{equation} \end{corollary} \begin{proof} For each attained first-coordinate label $u$, choose $h_1(u)$ independently and uniformly from $[L]$. Independently make the same choices $h_2(v)$ for the second-coordinate labels. Use these two random maps for both points and centres. Only finitely many labels occur; ordering their first occurrences among a fixed ordering of the vertices makes these choices measurable even when the test space is infinite. The new label of a state $(u,v)$ is the pair $(a,b)=(h_1(u),h_2(v))$. Set \[ f(a,b)=(e_a,e_b,e_{(a,b)}),\qquad g(q)=(e_{a_q},e_{b_q},-2e_{(a_q,b_q)}). \] Their norms are $\sqrt3$ and $\sqrt6$. The inner product is \[ \mathbf1_{a=a_q}+\mathbf1_{b=b_q} -2\mathbf1_{(a,b)=(a_q,b_q)}, \] which is the indicator of the strict cross after relabelling. Lemma~\ref{ub:ip:prediction} therefore gives error at most $4\sqrt{18}(D/H)^{1/4}$. For two fixed original states, these random maps preserve equalities and create either of the two previously false coordinate equalities with probability at most $1/L$. The probability that the original and relabelled strict-cross indicators differ is at most $2/L$. This applies to both the actual and the predicted pair: the predicted point uses previous random maps only, and the current random maps are fresh. The two comparisons therefore add $4/L$. Average the new random maps and test before choosing the history, and use $4\sqrt{18}<18$. The set of labels has size $L^2$, so the history count is at most $(L^2)^{H-1}\le L^{2H}$. For $W>0$, apply the probability statement to the normalized test measure and multiply by $W$. For $W=0$, use the empty history. \end{proof} \section{Heights, finite tests, and coordinate states} \label{ub:ht:section} This section supplies the arithmetic part of the argument, including the quantitative treatment of the bounded-height alternative. The underlying height and prediction strategy is that of \cite{ub:OAI2026}; all constants needed below are specified here. Write $\omega$ for the uniform law on the incidence edges of the preceding section, and $\mu,\kappa$ for its point and centre marginals. We use the following conclusions already proved there: \begin{equation}\label{ub:ht:inputs} \begin{gathered} D\le 2a\log t,\qquad \max_p\mu(p)\le t^{-1+11b},\\ d_{P,\min}\ge t^{1-11b},\qquad d_{Q,\min}\ge t^{a-11b},\qquad \xi\le t^{-a/100}. \end{gathered} \end{equation} Here $D$ is the mutual information of the two point endpoints of a wedge, and $b=a/10^6$. The matching-label inequality is available with $C_0=2^{112}$: for every finite labelling of the two vertex sets, the probability that two independent $\mu$-points have different labels is at most $C_0$ times the edge mismatch probability. We also use the finite-label acceptance and strict-cross prediction lemmas proved in the preceding section. Throughout, sufficiently large $t$ means a threshold depending only on the fixed numerical parameters, never on the configuration. For a point or centre $x=(x_1,x_2)$, set \[ z_x=x_1+ix_2,\qquad w_x=x_1-ix_2. \] The algebraic specialization already made ensures that all these coordinates are algebraic. Choose a finite normal number field $F$ containing them and $i$. On an incidence edge $e=(p,q)$ put \begin{equation}\label{ub:ht:jumps} d_e=z_p-z_q,\qquad (z_p-z_q)(w_p-w_q)=1,\qquad w_p-w_q=d_e^{-1}. \end{equation} In particular every $d_e$ is nonzero. Distinct original real points have distinct $z$-coordinates and distinct $w$-coordinates. Their nonzero differences remain nonzero under every field embedding, although the two embedded coordinates need not remain complex conjugates. \subsection{Normalizations and elementary height estimates} We recall the place normalizations explicitly. If $d=[F:\Q]$, include each embedding $\sigma:F\longrightarrow\C$ as an archimedean entry of weight $1/d$, with absolute value $|u|_\sigma=|\sigma(u)|$. Thus conjugate embeddings are included separately. At a finite place $v$ above the rational prime $\ell$, use the absolute value extending $|\ell|_\ell=\ell^{-1}$ and the weight $w_v=[F_v:\Q_\ell]/d$. Write $w_\sigma=1/d$ at infinity. The standard product formula, with these normalizations, states \begin{equation}\label{ub:ht:product-formula} \sum_v w_v\log|u|_v=0\qquad(u\in F^\times), \qquad \sum_{v\mid\infty}w_v=1. \end{equation} We use $\sum_v^*g(v)=\sum_v w_vg(v)$. Expectations denoted by $\E_E$ or $\E_\omega$ both use the uniform law on incidence edges. Only finitely many finite places contribute for any fixed $u\ne0$. These statements follow from the prime factorization of a principal fractional ideal and the norm formula; they are the usual number-field product formula, rather than an additional geometric hypothesis \cite{ub:BG2006}. Define $\log^+r=\max(0,\log r)$ for $r>0$, and set $\log^+0=0$. The absolute logarithmic height is \[ \hgt(u)=\sum_v w_v\log^+|u|_v,\qquad \hgt(0)=0. \] \begin{lemma}[Height identities and embedding tails]\label{ub:ht:height} This height is unchanged by enlarging $F$. For $u\ne0$, \begin{equation}\label{ub:ht:inverse} \hgt(u^{-1})=\hgt(u),\qquad \sum_vw_v\bigl|\log|u|_v\bigr|=2\hgt(u). \end{equation} For arbitrary $u,v$, \begin{equation}\label{ub:ht:basic} \hgt(uv)\le\hgt(u)+\hgt(v),\qquad \hgt(u+v)\le\hgt(u)+\hgt(v)+\log2. \end{equation} Suppose that $X$ is an $F$-valued random variable and that $\sigma$ is an independent uniformly random embedding of $F$ into $\C$. For $B>1$ and $0B)\le\frac{\E\hgt(X)}{\log B},\qquad \Pr(X\ne0,\ |\sigma(X)|s\}},\qquad \pi(v,s)=\Pr_\omega(Y_e=1),\\ f(v,s)=\min\{\pi(v,s),1-\pi(v,s)\},\qquad S=\int f. \end{gathered} \end{equation} Further enlarging the intervals does not change $S$: outside the range of the edge thresholds all edge bits coincide. For independent uniform edges $e,e'$, the threshold representation of the distance between two real numbers and \eqref{ub:ht:inverse} give \begin{align} 2\E\hgt(d_e/d_{e'}) &=\E\sum_vw_v|n_e(v)-n_{e'}(v)|\notag\\ &=2\int\pi(1-\pi)\le2S. \label{ub:ht:mean-ratio-height} \end{align} All terms are nonnegative, so the exchanges of integration and expectation are justified by Tonelli's theorem. \subsection{Transferring edge heights to point heights} \begin{lemma}[A quantitative height transfer]\label{ub:ht:transfer} Let $U$ assign an element of $F$ to every point and centre vertex. The matching-label inequality with constant $C_0$ implies \begin{equation}\label{ub:ht:transfer-estimate} \E_{p,p'\sim\mu}\hgt(U_p-U_{p'}) \le C_0\E_{(p,q)\sim\omega}\hgt(U_p-U_q)+2C_0+1. \end{equation} The two point variables on the left are independent. \end{lemma} \begin{proof} At a finite place and a level $s\ge0$, label vertices by the equivalence classes of $|U_x-U_{x'}|_v\le e^s$. The ultrametric inequality makes this an equivalence relation. Its mismatch indicator integrates over $s\ge0$ to $\log^+|U_x-U_{x'}|_v$. At an infinite place, identify the embedded coordinates with complex numbers and use square cells of side $e^s$. Shift the grid by a uniformly random fractional vector in $[0,1)^2$. Two coordinates at Euclidean distance $A$ have different cells with probability at most \[ \min\{1,\sqrt2 A e^{-s}\}. \] Indeed, in each real coordinate the mismatch probability is at most the coordinate separation divided by the side length; summing the two coordinate separations costs at most $\sqrt2 A$. Direct integration gives \[ \int_0^\infty\min\{1,\sqrt2 A e^{-s}\}\,ds \le \log^+A+2. \] Conversely, two points at distance greater than $\sqrt2 e^s$ must occupy different cells for every shift. Consequently their positive logarithmic distance is at most the integral of their mismatch indicator plus $\log\sqrt2$. At each fixed place, level, and shift, apply the matching-label inequality to the finite set of labels that occurs. Integrate it, average the archimedean shifts, and then sum the place weights. Finite places cause no additive error. The total weight at infinity is one, so the edge error is at most $2C_0$, and the point error is less than one. This is \eqref{ub:ht:transfer-estimate}. \end{proof} \subsection{Excluding a prescribed bounded scale} For the rest of this section the numerical parameters are \begin{equation}\label{ub:ht:parameters} K=10^{100004},\qquad k=(2^{1024}K)^{1024},\qquad S_0=k^{100},\qquad a=2^{-2^{200}S_0},\qquad b=a/10^6. \end{equation} The proof below also works for any fixed $K\ge10$ with these definitions. \begin{proposition}[The bounded-scale alternative]\label{ub:ht:bounded} For every sufficiently large $t$, the incidence piece under consideration satisfies $S>S_0$. \end{proposition} \begin{proof} Assume $S\le S_0$. By \eqref{ub:ht:mean-ratio-height}, some edge value $d_0$ satisfies $\E_\omega\hgt(d_e/d_0)\le S_0$. Initially rescale the two coordinates to $z/d_0,d_0w$, and write $D_e=d_e/d_0$ for the first rescaled jump. Both rescaled edge jumps, $D_e$ and $D_e^{-1}$, have mean height at most $S_0$. Apply Lemma~\ref{ub:ht:transfer} to both coordinates and average over one of the two point variables. A point $p_0\in P$ can be chosen such that, after putting \[ Z_x=(z_x-z_{p_0})/d_0,\qquad W_x=d_0(w_x-w_{p_0}), \] one has \[ A_0:=\E_\mu(\hgt(Z_p)+\hgt(W_p)) \le 2C_0S_0+4C_0+2. \] The identities $Z_q=Z_p-D_{pq}$ and $W_q=W_p-D_{pq}^{-1}$ on an edge and \eqref{ub:ht:basic} give \[ B_0:=\E_\kappa(\hgt(Z_q)+\hgt(W_q)) \le A_0+2S_0+2\log2. \] Also $C_0':=\E_\omega(\hgt(D_e)+\hgt(D_e^{-1}))\le2S_0$. Thus, with the fixed numerical bound \begin{equation}\label{ub:ht:height-budget} C_h=2^{117}S_0, \qquad A_0+B_0+C_0'\le16C_0(S_0+1)\le C_h. \end{equation} This normalization preserves the exact identity $(Z_p-Z_q)(W_p-W_q)=1$ on every edge. Given any finite history map $\mathcal H$ on $P$, sample a point $p\sim\mu$, a prediction $\widehat p$ independently from $\mu$ conditional on $\mathcal H(\widehat p)=\mathcal H(p)$, and three independent neighbours $q_1,q_2,q_3$ of $p$ conditional on $p$. Choose a fresh uniform embedding $\sigma:F\longrightarrow\C$, independently of this entire experiment. Both point marginals are $\mu$, every centre marginal is $\kappa$, and each $(p,q_j)$ has law $\omega$. These facts hold for every history map, including one that will be selected below. Put $D_j=D_{pq_j}$. Choose fixed geometric tolerances \begin{equation}\label{ub:ht:geometric-tolerances} \log B=2^{10}C_h,\qquad \log(1/r)=2^{12}C_h,\qquad \tau=\frac{r^4}{256B^5}. \end{equation} In particular $B\ge2$, $0s$ says that the first jump has absolute value less than $e^{-s}$ and its inverse has absolute value greater than $e^s$. Thus exactly the $A$-labels agree. When $n_e(v)0$. The integer $k$ makes $L=k^{100}$. For large enough $t$ we have $1\le (a/2)k^{400}\log t$, so $H\le(5a/2)k^{400}\log t$. Consequently \[ \frac{\log L^{2Hk}}{\log t} \le500ak^{401}\log k<\tfrac12. \] For the final numerical inequality, take logarithms to base two: the negative term $-2^{200}k^{100}$ from $a$ exceeds in absolute value $\log_2 500+401\log_2 k+\log_2\log k+1$. This holds already for $k\ge2$, using $\log_2 k\le k$; our $k\ge2^{1048576}$ is much larger. Thus the first inequality in \eqref{ub:ht:history-counts} holds. Dividing by $d_{P,\min}\ge t^{1-11b}$ proves the second for large $t$. Since $L^{2H}\le L^{2Hk}\le t^{1/2}$, the last follows from $\max_p\mu(p)\le t^{-1+11b}$. The relevant power of $t$ is strictly negative, which proves the final assertion as well. \end{proof} The strict-cross prediction lemma can therefore be applied to any finite-measure family of the coordinate-state tests. Its history has at most $L^{2H}$ values, and a $k$-tuple of histories has at most $L^{2Hk}$ possible values. The next section selects the test family and keeps track of its integrated error; no asymptotic choice of $k$ or of $a$ remains. \section{Scores on common levels and a quantitative lower bound for rare mass} \label{ub:sc:section} This section gives finite-parameter versions of the level and score arguments. The organization follows the method introduced in \cite{ub:OAI2026}, but every estimate used below is stated with a numerical error. In particular, no passage to a subsequence, and no unspecified rate in a little-oh term, is used. Write $\omega$ for the uniform probability measure on the retained incidence edges, and write $\mu$ and $\kappa$ for its point and center marginals. All integrals over levels use the weighted measure of the preceding section. For a level, $Y_e\in\{0,1\}$ is its edge bit, $y_x=(A_x,B_x)$ its vertex state, and $m$ its probability of inconsistency. An edge is consistent when exactly its $A$ labels agree if $Y_e=1$, and exactly its $B$ labels agree if $Y_e=0$. Put \[ \begin{gathered} \alpha=k^{-1/16},\qquad \vartheta=k^{-1/400},\\ \mathcal C=\{f\ge\alpha\},\quad \mathcal B=\{f<\alpha,\ m>k^{-1}\},\quad \mathcal G=\{f<\alpha,\ m\le k^{-1}\}. \end{gathered} \] The letters $\mathcal C,\mathcal B,\mathcal G$ denote common, damaged rare, and good rare levels, respectively. Denote their first two widths by $W_c=\int_{\mathcal C}1$ and $W_b=\int_{\mathcal B}1$. We have \begin{equation}\label{ub:sc:basic-input} S\ge k^{100},\qquad \int m\le4,\qquad W_c\le k^{1/16}S,\qquad W_b\le4k. \end{equation} We shall also use the marginal expansion estimate \begin{equation}\label{ub:sc:expansion} \Pr_\omega(F,G) \le 2\sqrt{\mu(F)\kappa(G)} \bigl((\mu(F)\kappa(G))^{1/25}+\xi\bigr), \qquad \xi\le k^{-100}. \end{equation} All these statements hold for sufficiently large $t$ with the fixed numerical parameters chosen earlier. \subsection{Rare exceptions and the history map} At a level of $\mathcal G$, let $r$ be the minority bit. The coordinate which agrees on consistent majority-bit edges is $B$ when $r=1$, and $A$ when $r=0$. Its mismatch probability is at most $f+m$. The quantitative matching-label lemma, with its local constant two, supplies one label whose point and center exception probabilities have sum at most $2(f+m)$. Select such a label by a fixed order. Let $I_x$ indicate that the selected coordinate of $x$ is not this label, and define \[ h_P=\E_\mu I_p,\quad h_Q=\E_\kappa I_q,\quad H_x^*=\int_{\mathcal G} I_x,\qquad \mathcal A=\E_\mu H_p^*,\quad \mathcal B_*=\E_\kappa H_q^*. \] The label selection is measurable: at every level only the labels on the finite vertex set are relevant, and ties are settled by their first vertex. \begin{lemma}\label{ub:sc:exceptions} The exception means and the integrated probabilities of joint exceptions satisfy \begin{align} h_P+h_Q&\le2(f+m)\le4\alpha &&\text{on }\mathcal G, &\mathcal A+\mathcal B_*&\le3S,\label{ub:sc:exception-means}\\ J_e:=\int_{\mathcal G}\Pr_\omega(I_p=I_q=1) &\le9\vartheta^2S,\label{ub:sc:joint-edge}\\ J_w:=\int_{\mathcal G}\Pr(I_p=I_{p'}=1) &\le21\vartheta S.\label{ub:sc:joint-wedge} \end{align} In the last line the law samples a center with marginal $\kappa$ and two independent neighbors of that center. Moreover $f\le h_P+h_Q+m$ on $\mathcal G$. \end{lemma} \begin{proof} The local matching lemma applies because $\alpha+k^{-1}\le2^{-112}$. Integrating its conclusion gives $\mathcal A+\mathcal B_*\le2(S+4)\le3S$. If both endpoints have the selected coordinate label and the edge is consistent, its bit is the majority bit. Every minority edge therefore has an exceptional endpoint or is inconsistent, proving the last assertion. For the edge estimate, write $h=h_P$ and $u=h_Q$. Since $2\sqrt{hu}\le h+u$, \eqref{ub:sc:expansion} gives \[ \Pr_\omega(I_p=I_q=1) \le(h+u)\bigl((hu)^{1/25}+\xi\bigr) \le3\vartheta^2(h+u). \] Here $h+u\le4\alpha$ implies $hu\le4\alpha^2$, and $\alpha^{2/25}=\vartheta^2$. Integration proves \eqref{ub:sc:joint-edge}. For the wedge estimate, fix a point subset $F$ of marginal mass $h\le4\alpha$, and put $g(q)=\Pr(p\in F\mid q)$. Its wedge self-intersection probability is $\E_\kappa g^2$. For $0d\}$ has marginal mass at most $h/d$. Consequently \[ \E_\kappa g^2 \le dh+\Pr_\omega(F,G_d) \le h\left[d+2d^{-1/2} \left(\left(\frac{(4\alpha)^2}{d}\right)^{1/25}+\xi\right)\right]. \] Choose $d=(4\alpha)^{1/25}$. Then $d\le2\vartheta$, and the first nonconstant term inside the second summand is $2(4\alpha)^{73/1250}\le2d$; the term containing $\xi$ is at most $\vartheta$. The bracket is therefore at most $7\vartheta$. Apply this estimate to the point exceptions at each good rare level and integrate to obtain $J_w\le7\vartheta\mathcal A\le21\vartheta S$. \end{proof} For a state $y=(a,b)$ define its strict cross by \[ R_y=\{(u,v):(u=a,\ v\ne b)\ \text{or}\ (u\ne a,\ v=b)\}. \] The prediction lemma applied to the normalized measure on $\mathcal C$ provides one fixed history map $\mathsf H$ with at most $L^{2H}$ values. At a level, let $\nu_h$ be the distribution of the true state of a $\mu$-point conditional on its history being $h$. We use true states, not the auxiliary relabelled states used in constructing the history. Put \begin{equation}\label{ub:sc:prediction-error} \mathcal R=\int_{\mathcal C} \E_\omega[1-\nu_{\mathsf H(p)}(R_{y_q})]. \end{equation} Consistency implies strict-cross membership. The prediction bound thus gives \begin{equation}\label{ub:sc:R-bound} \mathcal R \le4+W_c\bigl(4/L+18(D/H)^{1/4}\bigr) \le26 k^{-99}S. \end{equation} When $W_c=0$, take the empty history and set $\mathcal R=0$. Throughout, the history is fixed before the fresh level is sampled. For each center $q$, sample $k$ independent neighbors and record their histories in an ordered template $\theta=(h_1,\ldots,h_k)$. Write $a(q,\theta)=\Pr(\theta\mid q)$ and $M$ for the number of possible templates. The joint edge--template measure is \begin{equation}\label{ub:sc:template-law} \omega(p,q,\theta)=\frac{\boldsymbol1_{(p,q)\in E}}{|E|}a(q,\theta), \qquad M\le L^{2Hk}. \end{equation} In this law the original point and the template are independent conditional on $q$. The parameter choice ensures \begin{equation}\label{ub:sc:collision-input} d_{Q,\min}^{-1}\le k^{-100},\qquad d_{P,\min}^{-1}\le k^{-100},\qquad \frac{M}{d_{P,\min}}\le k^{-100},\qquad |\mathsf H(P)|\max_p\mu(p)\le k^{-100-1/16}. \end{equation} \subsection{Three pair laws and an upper score bound} We use the following three experiments, each fixed across the level space: \begin{enumerate} \item $Q$: sample $q$ with law $\kappa$ and then two independent neighbors $p,p'$ of $q$; \item $P$: sample $p$ with law $\mu$ and then two independent neighbors $q,q'$ of $p$; \item $P^\theta$: sample $(p,\theta)$ with its marginal in \eqref{ub:sc:template-law}, and then sample $q,q'$ independently conditional on $(p,\theta)$. \end{enumerate} For any one of these laws write $x$ for the base, $y_1,y_2$ for the opposite endpoints, and $Y_1,Y_2$ for the two bits. The base includes $\theta$ in the third law. At a level let $\rho$ be the probability of bit one conditional on this full base, and define \[ C_i=\int_{\mathcal C}\E_{\mathrm{base}}[2\rho(1-\rho)], \qquad i\in\{Q,P,P^\theta\}. \] \begin{lemma}\label{ub:sc:slot-laws} Every edge slot, after its base is averaged, has law $\omega$. In the $P^\theta$ experiment, for every fixed $p$, the probability of $q=q'$, averaged over the conditional template law, is at most $M/d(p)$. \end{lemma} \begin{proof} Summing \eqref{ub:sc:template-law} over $\theta$ gives the original edge law; conditioning and then resampling one slot preserves that marginal. For the collision claim put $a_p(\theta)=\Pr(\theta\mid p)$. Whenever this is positive, \[ \Pr(q\mid p,\theta) =\frac{\boldsymbol1_{(p,q)\in E}a(q,\theta)}{d(p)a_p(\theta)} \le\frac1{d(p)a_p(\theta)}. \] The sum of the squares of a probability vector is at most its largest entry. Multiplication by $a_p(\theta)$ and summation over at most $M$ templates proves the bound. This is a pointwise assertion in $p$, so it can be multiplied by an arbitrary nonnegative function of $p$ before averaging. \end{proof} Define the score constants \[ d_*:=\frac1{100},\qquad \ell_*:=\frac1{20000}=\frac{d_*^2}{2},\qquad c(\rho)= \begin{cases} 1+\ell_*,&|\rho-\tfrac12|\le d_*,\\ 1-\ell_*,&|\rho-\tfrac12|>d_*. \end{cases} \] If $Y_1=Y_2=1$, let $N$ mean $B_{y_1}=B_{y_2}$; if $Y_1=Y_2=0$, let $N$ mean $A_{y_1}=A_{y_2}$. Set $N$ false when the bits differ. Thus $N$ tests the coordinate not shared with the base on a consistent edge. The score is \begin{equation}\label{ub:sc:score-definition} \mathrm{sc}= \begin{cases} c(\rho)\boldsymbol1_{\{Y_1=Y_2,\,N\}},&\text{on }\mathcal C,\\ 2I_x\boldsymbol1_{\{Y_1=Y_2=r,\,N\}},&\text{on }\mathcal G,\\ 0,&\text{on }\mathcal B \end{cases} -\boldsymbol1_{\{Y_1\ne Y_2\}}. \end{equation} Let $D_i$ denote its expected integral in experiment $i$. Superscripts $c$ and $nc$ will denote its contributions on $\mathcal C$ and on the complement of $\mathcal C$. \begin{lemma}[An interval calculation]\label{ub:sc:interval} Let $b\le b'$, $T\ge0$, $I_-=[b-T,b]$, and $I_+=[b',b'+T]$. For any interval $J$ and $u=|J\cap I_-|+|J\cap I_+|$, \[ [(1-\ell_*)2T+2\ell_*u-2T]_+ \le\ell_*|J|\boldsymbol1_{\{b,b'\in\operatorname{int}J\}}. \] If a real random variable $X$ satisfies $|\Pr(X>s)-1/2|\le d_*$ throughout $J$, then $\Pr(X\in\operatorname{int}J)\le2d_*$. \end{lemma} \begin{proof} The positive part on the left is $2\ell_*(u-T)_+$. Positivity requires both interval intersections to have positive length, which places $b,b'$ in the interior of $J$. Also $u\le2T$ and $u\le|J|$, whence $2(u-T)\le u\le|J|$. For the probability assertion, if $ss)-\Pr(X>s')\le2d_*$. Exhaustion of the interior by such intervals gives the result, including possible atoms at its endpoints. \end{proof} \begin{proposition}[Upper and rare lower budgets]\label{ub:sc:budgets} Put $v_*=10d_*^2\ell_*=1/20000000$. For all three experiments, \begin{equation}\label{ub:sc:upper-budget} D_i\le v_*C_i+\vartheta S. \end{equation} If $x$ is the base side and $y$ the opposite side, then \begin{equation}\label{ub:sc:rare-lower} D_i^{nc}\ge2\E H_x^*-2\E H_y^*-40\vartheta^2S, \end{equation} where the expectations use the original marginals, also for $P^\theta$. \end{proposition} \begin{proof} We first prove the upper bound for distinct opposite endpoints. Their two unit-edge jumps, oriented from point to center in both cases, are nonzero and distinct; call them $d,d'$. At a place $v$, set \[ \begin{gathered} b=\min(-\log|d|_v,-\log|d'|_v),\quad b'=\max(-\log|d|_v,-\log|d'|_v),\\ u=-\log|d-d'|_v,\quad \beta=u-b. \end{gathered} \] The two opposite-endpoint coordinate differences are, up to sign, $d-d'$ and $1/d-1/d'$. Consequently \begin{equation}\label{ub:sc:product-identity} -\log|(d-d')(1/d-1/d')|_v=2\beta-(b'-b). \end{equation} Let $c_v=\log2$ at the archimedean places and zero otherwise. The triangle inequality gives $\beta\ge-c_v$; put $T_v=\beta+c_v\ge0$. A positive score with both bits one must lie in $[b-T_v,b]$, and a positive score with both bits zero must lie in $[b',b'+T_v]$. Indeed the matching nonshared coordinate bounds its difference by the corresponding grid scale, with factor at most $\sqrt2$ at infinity; substitution of the definitions gives these two inclusions. The integrated negative score at this place is exactly $b'-b$. There are two possibilities. If a rare positive score with minority bit one occurs at a threshold $s_0$, then $s_0\eta$. Chebyshev's inequality bounds the probability of fewer than $k\eta/2$ misses by $4/(k\beta)\le1/2$. It remains to bound twice the probability of a cross which contains one half of $2k$ samples and misses at least $k\eta/2$ members of the other half. Condition on the $2k$ indexed samples and split their indices uniformly into two groups of size $k$. A strict cross produces at most $(2k+1)^2$ different subsets of these indices: each of its two coordinate labels can be one of the at most $2k$ labels observed in that coordinate, or a label not observed. If a pattern has $m$ misses, the probability that all misses belong to the second group is zero for $m>k$, and otherwise at most \[ \frac{k(k-1)\cdots(k-m+1)}{(2k)(2k-1)\cdots(2k-m+1)}\le2^{-m}. \] The union bound proves the first assertion. This argument allows repeated states, since the sampling indices remain distinct. For the second assertion, after selecting the histories in $\theta$, draw one synthetic state independently from each $\nu_{h_j}$. Before the histories are revealed these are independent of law $\widetilde\nu_q$. If the template admits a violating candidate, choose one using only the template. Conditional on this choice, all synthetic states are in its cross with probability at least $1-k\cdot k^{-2}$. Apply the finite framework with $\eta=k^{-1/4}$ and divide by $1-1/k\ge1/2$. \end{proof} \begin{lemma}[Restricted-support transfer]\label{ub:sc:support-transfer} Fix a level. Under the law \eqref{ub:sc:template-law}, let $\mathcal U(q,\theta)$ be a selected row, a selected column, or a selected set of at most two states. On an event $G$ determined by $(q,\theta)$ suppose that $\widetilde\nu_q(\mathcal U(q,\theta))\ge1-3\eta^2$, where $0<\eta<1/2$. Then \[ \Pr\bigl(G,\ y_p\notin \mathcal U(q,\theta)\bigr)\le8\eta. \] \end{lemma} \begin{proof} The conditional history law of the original point, given $(q,\theta)$, is the ordinary history law of a neighbor of $q$. Consequently, on $G$, \[ \E[\nu_{\mathsf H(p)}(\mathcal U(q,\theta)^c)\mid q,\theta]\le3\eta^2. \] The event that this conditional measure is greater than $\eta$ has probability at most $3\eta$. For each history $h$, define exceptional states using only $h$ and the level. If there is a row of $\nu_h$-mass at least $1-\eta$, that row is unique; include its complement as the first exceptional set. Use the analogous column complement as the second set. Each has mass at most $\eta$, with an empty set used when the respective concentrated support does not exist. If there is a set of at most two states of mass at least $1-\eta$, put $M_h=\{x:\nu_h(x)>\eta\}$. Every such concentrated two-state support contains $M_h$, and one such support shows that $\nu_h(M_h^c)\le3\eta$. Use $M_h^c$ as the third exceptional set; again use the empty set if no such support exists. The union has $\nu_h$-mass at most $5\eta$. The original point has marginal $\mu$, and conditional on its history its state has exactly law $\nu_h$. Its probability of belonging to this union is therefore at most $5\eta$, under the full edge--template measure. Outside this event and the preceding event of probability at most $3\eta$, the actual state belongs to whichever permitted concentrated support has been selected. The assertion follows. No assertion about the point law after conditioning on the selected support is required. \end{proof} Two distinct states $(a,b),(a,b')$ in the same row have intersecting crosses contained in that row, namely \begin{equation}\label{ub:sc:row-cross} R_{(a,b)}\cap R_{(a,b')}=\{(a,v):v\notin\{b,b'\}\}. \end{equation} A third state outside the row reduces this intersection to at most one state. The column statements are identical with the coordinates exchanged. When both coordinates differ, \begin{equation}\label{ub:sc:corner-cross} R_{(a,b)}\cap R_{(a',b')}=\{(a,b'),(a',b)\}. \end{equation} Three states with pairwise distinct first and pairwise distinct second coordinates have empty common cross. These identities prove the following exhaustive alternatives for the candidate set: it is empty; it is a singleton; it has a repeated coordinate, with either all candidates in that row or column or a third outside it; or it has pairwise distinct coordinates, with either two candidates or at least three. Select the two or three witnesses used in these alternatives by a fixed order. On the complement of the failure event in \eqref{ub:sc:candidate-mixture}, their common cross has $\widetilde\nu_q$-mass at least $1-3k^{-1/4}$. The containing row, column, or at most two states furnished by \eqref{ub:sc:row-cross}--\eqref{ub:sc:corner-cross} is therefore a permitted support in Lemma~\ref{ub:sc:support-transfer}, with $\eta=k^{-1/8}$. Its transfer cost, integrated over common levels, is at most $8k^{-1/8}W_c\le8\vartheta^{25}S$. Empty common crosses cannot have the asserted mass and yield no good edge. Declare an edge with its template bad at a common level if it is inconsistent, if its center state is not a candidate, if the event in \eqref{ub:sc:candidate-mixture} occurs, or if the selected support fails to contain the actual point. By \eqref{ub:sc:actual-candidate}, the two lemmas, and \eqref{ub:sc:basic-input}, its integrated probability satisfies \begin{equation}\label{ub:sc:bad-budget} B_{\mathrm{bad}}\le 4+26k^{-96}S+36k^2 2^{-k^{3/4}/2}W_c+8\vartheta^{25}S \le10\vartheta^2S. \end{equation} For example, the exponential term times $W_c/S$ is at most $\vartheta^2$ for $k\ge2^{1048576}$: the logarithm of its polynomial prefactor is smaller than $10\log_2 k$, whereas $k^{3/4}/2$ exceeds this and any required fixed multiple of $\log_2 k$ by that threshold. The remaining terms are bounded directly by powers of $k$. \subsection{Suppression of common center variation} The following elementary inequality is useful for every $0\le\rho\le1$: \begin{equation}\label{ub:sc:two-corner-numeric} c(\rho)\bigl(\rho^2+(1-\rho)^2\bigr)-2\rho(1-\rho) \ge\ell_*/2. \end{equation} To verify it, subtract $2\rho(1-\rho)$ from the bracket with coefficient one, obtaining $(2\rho-1)^2$. In the central window the extra term is at least $\ell_*/2$. Outside the window the result is at least $4d_*^2-\ell_*>\ell_*/2$. \begin{proposition}\label{ub:sc:common-variation} We have \begin{equation}\label{ub:sc:common-result} C_Q+C_{P^\theta}\le2^{19}\vartheta S, \qquad \mathcal B_*\le\mathcal A+2^{20}\vartheta S. \end{equation} \end{proposition} \begin{proof} Fix a level and a positive-mass template slice. Let $b_\theta$ be the single-edge bad probability in this slice. In the center experiment, sample its center using the conditional law of $q$ given $\theta$; the two points still have their ordinary independent law conditional on $q$, because the template was independent of the original point given $q$. In the point--template experiment sample the point conditional on $\theta$, then its two centers independently conditional on $(p,\theta)$. Both experiments have the same conditional single-edge law. Thus a pair has a bad slot with probability at most $2b_\theta$ in each experiment. Let $V_Q,V_{P^\theta}$ be the average full-law variations in this slice, and let $d_Q,d_{P^\theta}$ be its common expected scores. We claim \begin{equation}\label{ub:sc:slice-inequality} d_Q+d_{P^\theta} \ge\frac{\ell_*}{4}(V_Q+V_{P^\theta})-16b_\theta. \end{equation} Each variation is at most $1/2$, and each score is at least minus its variation. We check all the candidate alternatives explicitly. If there are no candidates, every edge is bad, so $b_\theta=1$, and the bound follows from $d_Q+d_{P^\theta}\ge-1$. For a single candidate state, two good edges of a point--template star have the same opposite state and hence the same bit and the required coordinate match. Its mean score is at least $1-\ell_*-4b_\theta$. The other score is at least $-1/2$. Their sum is more than sufficient for \eqref{ub:sc:slice-inequality}. If all candidates belong to one row or one column and at least two are distinct, support transfer puts all good point states in the same row or column. Every good edge therefore has the same bit. A bit difference requires a bad slot, so $V_Q+V_{P^\theta}\le4b_\theta$. The general lower bound by minus variation proves the claim. If there is instead a third candidate outside the shared row or column, the selected intersection has at most one state. If it is empty all edges are bad; otherwise all good points have the same state, and the center-star score is at least $1-\ell_*-4b_\theta$. The other score is at least $-1/2$, as in the singleton case. Finally suppose distinct candidates share no coordinate. Three candidates have empty common cross and no good edges. With exactly two candidates, good point states are their opposite corners. At either kind of base, the bit of a good edge determines its opposite state. Thus equal bits on a good pair imply the required coordinate match. The full-law expected score is at least the left side of \eqref{ub:sc:two-corner-numeric} minus $4b_\theta$, after averaging the bases. Each orientation is at least $\ell_*/2-4b_\theta$, which proves \eqref{ub:sc:slice-inequality}. The penalty $4b_\theta$ used here follows by bounding the missing positive reward by $(1+\ell_*)$ times the probability $2b_\theta$ of a bad slot. All these comparisons use the full conditional $\rho$ in $c(\rho)$. In particular, no deletion and renormalization of the bit law is made near the discontinuities of $c$. Multiply \eqref{ub:sc:slice-inequality} by the actual probability of its template, sum the slices, and integrate. By \eqref{ub:sc:bad-budget}, \[ D_Q^c+D_{P^\theta}^c \ge\frac{\ell_*}{4}(C_Q+C_{P^\theta})-160\vartheta^2S. \] Adding the two rare lower budgets cancels their main exception means. The upper budgets then give \[ g_*(C_Q+C_{P^\theta}) \le(2\vartheta+240\vartheta^2)S, \qquad g_*:=\ell_*(1/4-10d_*^2)=0.00001245>10^{-5}. \] Since $\vartheta$ is smaller than $1/240$, the first assertion of \eqref{ub:sc:common-result} follows. The common center score alone is at least $-C_Q$. Combining this with both center budgets gives \[ 2\mathcal B_*-2\mathcal A \le(1+v_*)C_Q+\vartheta S+40\vartheta^2S. \] The first assertion now implies the second, with the displayed constant. \end{proof} \subsection{A second score and positive rare point mass} For a state measure $\nu$, let $\nu_A,\nu_B$ be its coordinate marginals. Define \[ \tau(\nu)=\|\nu_A\|_2^2+\|\nu_B\|_2^2-1, \qquad \mathcal T=\int_{\mathcal C}\E_\mu\tau(\nu_{\mathsf H(p)}). \] The squared norms are sums of squares of the atom probabilities. \begin{lemma}\label{ub:sc:two-predictions} The two-prediction score satisfies \begin{equation}\label{ub:sc:T-bound} \mathcal T\le2\mathcal A+2\vartheta S. \end{equation} \end{lemma} \begin{proof} Choose a history according to its $\mu$-marginal, and then choose two independent points $p',p''$ from the conditional point law of that history. This coupling is fixed across the levels, and both its marginals are $\mu$. The integrand of $\mathcal T$ is the mean of \[ F(p',p'')=\boldsymbol1_{\{A_{p'}=A_{p''}\}} +\boldsymbol1_{\{B_{p'}=B_{p''}\}}-1. \] For a distinct pair put $Z=z_{p'}-z_{p''}$ and $W=w_{p'}-w_{p''}$. Both are nonzero. At a place let the level interval be $[l,u]$. Its endpoint convention gives $l\le-\log|Z|_v$ and $u\ge\log|W|_v$. An $A$ match can occur only below $-\log|Z|_v+c_v'$, and a $B$ match only above $\log|W|_v-c_v'$, where $c_v'=\log\sqrt2$ at infinity and zero otherwise. Bounding the lengths of these two threshold ranges gives \[ \int_l^u F(p',p'')\,ds \le-\log|Z|_v-\log|W|_v+2c_v'. \] The product formula yields an all-level integral at most $\log2$. On a good rare level where both points are normal their dominant labels match, so $F\ge0$. On every good rare level $F\ge-I_{p'}-I_{p''}$, and on damaged levels $F\ge-1$. Thus a distinct pair has common-level integral at most $\log2+H_{p'}^*+H_{p''}^*+W_b$. For a diagonal pair the common-level integral is $W_c$. Its probability is \[ \sum_{p\in P}\frac{\mu(p)^2} {\mu(\mathsf H^{-1}(\mathsf H(p)))} \le |\mathsf H(P)|\max_p\mu(p). \] By \eqref{ub:sc:collision-input} its contribution is at most $k^{-100}S\le\vartheta S$. The two preserved marginals bound the mean of the distinct-pair height terms by $2\mathcal A$, while $\log2+4k\le\vartheta S$. This proves \eqref{ub:sc:T-bound}. \end{proof} \begin{proposition}[A numerical rare-mass lower bound]\label{ub:sc:positive-mass} The good rare point mass satisfies \begin{equation}\label{ub:sc:positive-result} \mathcal A\ge2^{-132}S. \end{equation} Furthermore \begin{equation}\label{ub:sc:ordinary-variation} C_P\le2^{18}\mathcal A+2^{38}\vartheta S. \end{equation} \end{proposition} \begin{proof} At a common level call an ordinary edge bad when it is inconsistent or when $\nu_{\mathsf H(p)}(R_{y_q})<1-\zeta$, where $\zeta=k^{-2}$. Its integrated bad probability is at most $4+k^2\mathcal R\le\vartheta^2S$. Fix a point $p$ and write $b_p$ for its conditional bad probability, $\nu=\nu_{\mathsf H(p)}$, and $\rho_p$ for its full conditional bit probability. Let $\mathcal Y_p$ be the states of its good neighbors. For every $y\in\mathcal Y_p$, $\nu(R_y)\ge1-\zeta$, and consistency places the actual point state in $R_y$. We prove \begin{equation}\label{ub:sc:point-classification} d_P(p)+\tau(\nu) \ge\frac{\ell_*}{4}\,2\rho_p(1-\rho_p)-16(b_p+\zeta), \end{equation} where $d_P(p)$ is the conditional common score. If $\mathcal Y_p$ is empty then $b_p=1$, while $d_P(p)\ge-1/2$ and $\tau(\nu)\ge-1$, which suffices. If it is a singleton $(a,b)$, set $r=\nu_A(a)$, $s=\nu_B(b)$. Then $r+s\ge1-\zeta$, and hence $\tau(\nu)\ge r^2+s^2-1\ge-1/2-\zeta$. Two good neighbors have the same state and bit, giving $d_P(p)\ge1-\ell_*-4b_p$. This proves the claim in that case. If two states share a row, their common cross has $\nu$-mass at least $1-2\zeta$ and lies in that row, so $\tau(\nu)\ge(1-2\zeta)^2-1\ge-4\zeta$. If all good neighbor states belong to this row, consistency puts the actual point in it and makes every good bit one. Therefore $2\rho_p(1-\rho_p)\le2b_p$ and $d_P(p)\ge-2b_p$, proving \eqref{ub:sc:point-classification}. Columns work identically with bit zero. If a third state is outside the shared row or column, the three crosses intersect in at most one state and their intersection has mass at least $1-3\zeta$. It must therefore be a singleton. Both coordinate marginals have an atom of at least that mass, and $\tau(\nu)\ge2(1-3\zeta)^2-1\ge1-12\zeta$. The lower bound $d_P(p)\ge-1/2$ suffices. In the remaining nontrivial case, the distinct neighbor states share neither coordinate. There cannot be three, because the actual point belongs to all their crosses. For two, their opposite corners carry $\nu$-mass at least $1-2\zeta$. Each coordinate marginal consequently has at least this mass on two labels, and $\tau(\nu)\ge(1-2\zeta)^2-1\ge-4\zeta$. The actual point is an opposite corner. Its good-edge bit determines the neighbor state, so \eqref{ub:sc:two-corner-numeric} gives $d_P(p)\ge\ell_*/2-4b_p$. This finishes the proof of \eqref{ub:sc:point-classification}; the coefficient sixteen bounds each of the displayed penalties without changing the full bit law. Average and integrate \eqref{ub:sc:point-classification}. Since $\zeta W_c\le\vartheta^2S$, it follows that \[ D_P^c+\mathcal T\ge\frac{\ell_*}{4}C_P-32\vartheta^2S. \] Combining this with the rare lower budget, the upper budget, and \eqref{ub:sc:T-bound} gives \[ g_*C_P\le2\mathcal B_*+3\vartheta S+72\vartheta^2S \le2\mathcal B_*+4\vartheta S. \] Use \eqref{ub:sc:common-result}, $2/g_*<2^{18}$, and $(2^{21}+4)/g_*<2^{38}$ to obtain \eqref{ub:sc:ordinary-variation}. In particular, \begin{equation}\label{ub:sc:sum-variation} C_P+C_Q\le2^{18}\mathcal A+2^{39}\vartheta S. \end{equation} It remains to compare the edge minority probability to ordinary star variation. At any level give each point and center its conditional majority-bit label. If $V_P,V_Q$ are the two ordinary star variations, the bit differs from these labels with probabilities at most $V_P,V_Q$, because $\min(\rho,1-\rho)\le2\rho(1-\rho)$. Thus endpoint labels disagree with probability at most $v=V_P+V_Q$. When $v\le2^{-112}$, the local matching-label bound supplies a common label with total marginal exceptions at most $2v$; the edge bit differs from this label with probability at most $3v$. When $v>2^{-112}$, the trivial bound $f\le1/2$ is enough. Uniformly, \begin{equation}\label{ub:sc:global-bit-variation} f\le2^{112}(V_P+V_Q). \end{equation} Integrating \eqref{ub:sc:global-bit-variation} over common levels and using \eqref{ub:sc:sum-variation} bounds their minority mass by $2^{130}\mathcal A+2^{151}\vartheta S$. On good rare levels Lemma~\ref{ub:sc:exceptions} bounds it by $\mathcal A+\mathcal B_*+4$, and damaged levels contribute at most $2k$. Consequently \[ S\le(2^{130}+2)\mathcal A +(2^{151}+2^{20})\vartheta S+4+2k \le2^{131}\mathcal A+2^{152}\vartheta S. \] Since $\vartheta\le2^{-154}$, this implies $\mathcal A\ge2^{-132}S$, as asserted. \end{proof} \section{Weighted extraction of a graph of ratios}\label{ub:ex:section} We retain the probability laws and notation of the preceding sections. In particular, $\mu,\kappa$ are the two edge marginals, $\mathcal C,\mathcal B, \mathcal G$ are the common, damaged, and good rare levels, and $I_x$ is the exception indicator on $\mathcal G$. Write \[ A=\E_\mu H_p^*,\quad B=\E_\kappa H_q^*,\quad \vartheta=k^{-1/400},\quad R_0=k^{1/8},\quad c_0=2^{-133}. \] The quantitative conclusions already proved are \begin{equation}\label{ub:ex:inputs} A+B\le3S,\quad A\ge2^{-132}S,\quad C_Q\le2^{19}\vartheta S, \quad W_c\le k^{1/16}S,\quad W_b\le4k,\quad \E_E m_e\le4. \end{equation} Here $C_Q$ is the integral, over common levels, of the probability that the two bits in an ordinary center star differ. We prove the extraction rather than assuming that a small exceptional probability controls lost height. \begin{lemma}[A pointwise inequality]\label{ub:ex:pointwise} If $q,q'$ are distinct neighbors of a point $p$, then \begin{equation}\label{ub:ex:pointwiseeq} 2H_p^*\le3(H_q^*+H_{q'}^*+m_{pq}+m_{pq'})+W_c+W_b+2\log2. \end{equation} \end{lemma} \begin{proof} At a good rare level set $X=I_p$, $U=I_q$, $V=I_{q'}$ and let $M,M'$ record the two inconsistencies. The rare positive score $P_+$ equals two when $p$ is exceptional, both bits are minority bits, and the two other endpoints share the nonmatched coordinate; otherwise it is zero. Let $N=|Y_{pq}-Y_{pq'}|$. When $U=V=M=M'=0$, both bits are the same function of $X$, and $P_+=2X$. Therefore, without restrictions, \[ P_+\ge2X-2(U+V+M+M'),\qquad N\le U+V+M+M', \] and $P_+-N\ge2X-3(U+V+M+M')$. We also need a placewise upper bound for this score. Put $d=z_p-z_q$, $d'=z_p-z_{q'}$, and, at a place $v$, let \[ b=\min(-\log|d|_v,-\log|d'|_v),\quad b'=\max(-\log|d|_v,-\log|d'|_v),\quad u=-\log|d-d'|_v. \] Let $c_v=\log2$ at an archimedean entry and zero otherwise, and put $T_v=u-b+c_v\ge0$. Coordinate agreement permits positive scores only in $[b-T_v,b]$ or $[b',b'+T_v]$. Since the fraction of edges with bit one decreases with the threshold, a rare positive score can occur on only one of these two intervals. Its integrated positive contribution is at most $2T_v$. The all-level negative contribution is $b'-b$. Moreover, \[ -\log|(d-d')(d^{-1}-{d'}^{-1})|_v=2(u-b)-(b'-b). \] The element inside the absolute value is nonzero. Summing over places annuls its logarithm by the product formula and leaves at most $2\log2$. Outside $\mathcal G$, the negative score has absolute value at most one, and its total cost is at most $W_c+W_b$. Integrating the lower bound and comparing with this upper bound proves \eqref{ub:ex:pointwiseeq}. \end{proof} \begin{lemma}[A uniform tail bound]\label{ub:ex:tail} For all sufficiently large $t$, with the fixed parameters of the paper, \begin{equation}\label{ub:ex:tailbound} \E_\mu\bigl[H_p^*\mathbf1_{H_p^*>R_0S}\bigr] \le2^{13}\vartheta^4S. \end{equation} \end{lemma} \begin{proof} Denote the expectation on the left by $A_T$. On the event $H_p^*>R_0S$, the last three terms of \eqref{ub:ex:pointwiseeq} have sum at most $H_p^*$: their ratio to $R_0S$ is at most $k^{-1/16}+(4k+2\log2)/(R_0S)<1$. After absorption, cap either neighbor height at $H_p^*$. If a cap changes a term, that term alone makes the resulting inequality true. Thus \[ H_p^*\le3\bigl(\min(H_p^*,H_q^*)+\min(H_p^*,H_{q'}^*) +m_{pq}+m_{pq'}\bigr). \] Sampling two neighbors independently and retaining distinct neighbors on the left gives \begin{equation}\label{ub:ex:tailineq} (1-d_{P,\min}^{-1})A_T \le6\E_E[\min(H_p^*,H_q^*)\mathbf1_{H_p^*>R_0S}]+24. \end{equation} For $u\ge0$ put $F_u=\{p:H_p^*>\max(u,R_0S)\}$ and $G_u=\{q:H_q^*>u\}$. The first marginal mass is at most $3/R_0$. The marginal expansion inequality consequently gives, uniformly in $u$, \[ \Pr_E(F_u,G_u)\le\theta\sqrt{\mu(F_u)\kappa(G_u)},\qquad \theta=2(3/R_0)^{1/25}+2\xi\le4\vartheta^2. \] Layer cake and Cauchy--Schwarz imply \[ \E_E[\min(H_p^*,H_q^*)\mathbf1_{H_p^*>R_0S}] \le\theta\sqrt{A_TB}\le\theta\sqrt{3A_TS}. \] For $d_{P,\min}\ge2$, \eqref{ub:ex:tailineq} and Young's inequality yield \[ \tfrac12 A_T\le6\theta\sqrt{3A_TS}+24 \le\tfrac14 A_T+108\theta^2S+24. \] Hence $A_T/S\le432\theta^2+96/S\le2^{13}\vartheta^4$, since $S\ge k^{100}$. \end{proof} For each point define a profile on the set of places by \[ a_p(v)=\int_{\{s:(v,s)\in\mathcal G\}}\sigma(v,s)I_p(v,s)\,ds, \] where $\sigma=1$ when the minority bit is one and $\sigma=-1$ otherwise. For a place profile $b$ use the weighted norm $\|b\|_1=\sum_v^*|b(v)|$. Let $n_{pq}(v)=-\log|z_p-z_q|_v$. \begin{lemma}[Wedge approximation]\label{ub:ex:profiles} For an ordinary wedge $(p,q,r)$ there are nonnegative errors $E_1,E_2$ such that \begin{align} \|(n_{pq}-n_{rq})-(a_p-a_r)\|_1&\le E_1,\label{ub:ex:profile1}\\ \|n_{pq}-n_{rq}\|_1&\ge H_p^*+H_r^*-E_2,\label{ub:ex:profile2}\\ \E(E_1+E_2)&\le2^{20}\vartheta S.\label{ub:ex:profilemean} \end{align} \end{lemma} \begin{proof} At a good rare level let $b$ be the majority bit and $M_{pq}$ the inconsistency indicator. The identity $Y_{pq}=b+\sigma(I_p+I_q)$ has error at most $2(I_pI_q+M_{pq})$. Subtract its two versions in the wedge; the center term cancels. Threshold integration gives \eqref{ub:ex:profile1} with \[ E_1=\int_{\mathcal C}|Y_{pq}-Y_{rq}|+W_b +2\int_{\mathcal G}(I_pI_q+I_rI_q+M_{pq}+M_{rq}). \] Pointwise one also has \[ |Y_{pq}-Y_{rq}|\ge I_p+I_r -2(I_pI_r+I_pI_q+I_rI_q+M_{pq}+M_{rq}). \] If no deducted term occurs, exactly one exceptional point forces opposite bits; whenever a deducted term occurs the displayed lower bound is nonpositive or follows immediately. Integrating proves \eqref{ub:ex:profile2} with \[ E_2=2\int_{\mathcal G}(I_pI_r+I_pI_q+I_rI_q+M_{pq}+M_{rq}). \] The preceding joint-exception estimates bound the edge and wedge integrals by $9\vartheta^2S$ and $21\vartheta S$, respectively. Thus \[ \E(E_1+E_2)\le C_Q+W_b+72\vartheta^2S+42\vartheta S+32 \le2^{20}\vartheta S. \] \end{proof} \begin{lemma}[Selecting height before probability]\label{ub:ex:band} Put $\theta=1/(4K)$ and $B_k=1+\lceil\log_2(2R_0/c_0)\rceil$. There is $J\ge K$ and a set of ordinary wedges of probability at least $c_0/(R_0B_k)$ such that \begin{equation}\label{ub:ex:bandconditions} J\le H_p^*+H_r^*\le2J,\qquad E_1+E_2\le\theta(H_p^*+H_r^*). \end{equation} All these wedges have $p\ne r$. \end{lemma} \begin{proof} Set $W=H_p^*+H_r^*$ and $Z=E_1+E_2$. Then $\E W=2A\ge4c_0S$. Remove $Z>\theta W$. Its lost $W$-mass is at most \[ \E Z/\theta\le2^{22}K\vartheta S\le c_0S/4. \] The inequality uses $\vartheta\le2^{-157}/K$, which follows from the chosen $k$. Removing $W2R_0S$, one summand exceeds $R_0S$ and is at least $W/2$. Thus \[ \E[W\mathbf1_{W>2R_0S}] \le4\E_\mu[H_p^*\mathbf1_{H_p^*>R_0S}] \le2^{15}\vartheta^4S\le c_0S/4. \] At least $2c_0S$ weighted mass remains, and every remaining weight is at most $2R_0S$. The remaining probability is at least $c_0/R_0$. Divide $[c_0S,2R_0S]$ into at most $B_k$ dyadic intervals, assigning each endpoint to one interval. One interval has the stated probability. Its lower endpoint is $J\ge c_0S\ge K$. If $p=r$, \eqref{ub:ex:profile2} would imply $0\ge W-E_2\ge(1-\theta)W>0$, a contradiction. \end{proof} \begin{lemma}[Deleting all popular unit circles]\label{ub:ex:popular} Call any real unit circle popular if it contains at least $t^{9/10}$ points of $P$, whether or not its center belongs to $Q$. The wedge probability that $p,r$ share such a circle tends to zero as $t\to\infty$ with $K$ fixed. \end{lemma} \begin{proof} Let $v=|P|\le t^{1+2a+\eps}$. Pair counting bounds the number $M_c$ of popular circles by $O(v^2t^{-9/5})=O(t^{1/5+4a+2\eps})$. At most $M_c(M_c-1)$ points lie on two of them, since each pair of circles intersects in at most two points. By the marginal atom bound their total point probability is \[ O(t^{-3/5+8a+4\eps+11b})=o(1). \] Centers whose own unit circles are popular have probability at most \[ M_c\max_q\kappa(q)=O(t^{-9/5+3a+2\eps+11b})=o(1). \] Outside these events, fix $p,q$. There is at most one popular circle through $p$, and it differs from the circle centered at $q$. At most two neighbors $r$ of $q$ lie on it. The remaining probability is at most $2/d_{Q,\min}\le2t^{-a+11b}$, which tends to zero. All exponents are strictly negative for the fixed numerical parameters. \end{proof} \begin{proposition}[Finite organized pair graph]\label{ub:ex:organized} For all sufficiently large $t$ in a putative incidence counterexample, there are $J\ge K$ and a graph $G\subset P\times P$, using two copies of $P$, such that \[ |P|\le t^{1+1/K},\qquad |G|\ge t^{2-1/K}. \] Each edge has different endpoints and a nonzero algebraic ratio $\lambda_{pr}$ satisfying \begin{align} (z_p-z_r)(w_p-w_r)&=2-\lambda_{pr}-\lambda_{pr}^{-1},\label{ub:ex:dist}\\ \hgt(\lambda_{pr})&\ge J/3,\label{ub:ex:edgeheight}\\ \hgt\left(\frac{\lambda_{pr}\lambda_{p'r'}} {\lambda_{pr'}\lambda_{p'r}}\right)&\le J/K \quad\text{on every complete rectangle}.\label{ub:ex:defect} \end{align} No edge has both endpoints on a popular unit circle. \end{proposition} \begin{proof} The probability in Lemma~\ref{ub:ex:band} is a fixed positive number once $K$ is fixed. Lemma~\ref{ub:ex:popular} therefore removes less than half of it for sufficiently large $t$. For a fixed off-diagonal pair $(p,r)$, there are at most two possible centers, so its original wedge probability is at most $2/(|E|d_{Q,\min})$. Thus the number of retained pairs is at least \[ \frac{c_0}{4R_0B_k}|E|d_{Q,\min} \ge c(K)t^{2+3a-21b}\ge t^{2-1/K}. \] Also $|P|\le t^{1+2a+\eps}\le t^{1+1/K}$. Select one retained center $q=q(p,r)$ for every pair, and define $\lambda_{pr}=(z_p-z_q)/(z_r-z_q)$. Writing the two jumps as $d,d'$, the unit equations give \[ (z_p-z_r)(w_p-w_r)=(d-d')(d^{-1}-{d'}^{-1}) =2-d/d'-d'/d. \] The product formula and \eqref{ub:ex:profile2} give $2\hgt(\lambda_{pr})\ge(1-\theta)J$, proving \eqref{ub:ex:edgeheight}. Put $b_{pr}(v)=\log|\lambda_{pr}|_v+a_p(v)-a_r(v)$. Equation~\eqref{ub:ex:profile1} and \eqref{ub:ex:bandconditions} give $\|b_{pr}\|_1\le2\theta J$. On a rectangle, the unary profiles cancel. The norm of the resulting logarithmic defect is at most $8\theta J$, and its height is half that norm. Hence it is at most $4\theta J=J/K$. \end{proof} \section{A finite algebraic obstruction}\label{ub:alg:section} This section replaces the limiting height field by a finite algebraic argument. The geometric square-root obstruction follows the mechanism of \cite{ub:OAI2026}; we give its proof, including the field-enlargement step and the exceptional curve case. The additional ingredient is a critical-value construction with a uniform numerical height bound. All algebraic extensions in this section have characteristic zero. For an ordinary coordinate pair $x=(x_1,x_2)$ put $z_x=x_1+\mathrm{i}x_2$ and $w_x=x_1-\mathrm{i}x_2$. We use the complexified quadratic form, not a Hermitian form: \[ L(p,r)=(p_1-r_1)^2+(p_2-r_2)^2 =(z_p-z_r)(w_p-w_r). \] There are always $20$ left positions $p_1,\ldots,p_{20}$ and $5$ right positions $r_1,\ldots,r_5$. A grid therefore has $50$ position coordinates and $100$ pairs. We abbreviate $L(p_i,r_j)$ to $L_{ij}$. \subsection{A fixed set of exceptional position tuples} \begin{definition}\label{ub:alg:bad-set} Let $\mathcal Z\subseteq\mathbb A^{50}_{\C}$ be the set of position tuples for which there are scalars $R_{ij},A_i,A'_j$ and vectors $H_i,U_j\in\C^2$ satisfying \begin{align} R_{ij}^2&=L_{ij}(L_{ij}-4),\label{ub:alg:radical-equations}\\ 2(p_i-r_j)\mathbin{\cdot}(H_i-U_j) &=R_{ij}(A_i-A'_j),\label{ub:alg:flow-equations}\\ A_i-A'_j&\ne0\qquad(1\le i\le20,\ 1\le j\le5). \label{ub:alg:nonzero-flow} \end{align} Here the dot product is bilinear. Membership in $\mathcal Z$ means solvability in an algebraically closed overfield when the positions are themselves given over a smaller field. \end{definition} \begin{lemma}[Finite polynomial tests]\label{ub:alg:boolean} There is a fixed finite list of polynomials with integer coefficients in the $50$ position coordinates whose zero/nonzero pattern determines membership in $\mathcal Z$. The list is independent of every point configuration and of every rectangle-defect parameter introduced below. \end{lemma} \begin{proof} Introduce one further variable $v$ and replace \eqref{ub:alg:nonzero-flow} by \[ v\prod_{i=1}^{20}\prod_{j=1}^{5}(A_i-A'_j)-1=0. \] We now have $201$ equations in $176$ auxiliary variables: $100$ variables $R_{ij}$, $50$ vector coordinates, $25$ variables $A_i,A'_j$, and $v$. Their degree in the auxiliary variables is at most $101$. Their coefficients are integer polynomials of degree at most $4$ in the positions. The finite list is obtained by linear algebra. We use the standard effective ideal-membership theorem: if polynomials of degree at most $d$ in $n$ variables over a field generate $1$, then they have a representation of $1$ with multiplier degrees at most $(2d)^{2^n}$. This is Hermann's bound in the form stated and proved in \cite[Theorem~3.4]{ub:Aschen2004}. Thus put \[ B=202^{2^{176}}. \] The required list is specified by the following finite procedure. \begin{algorithm}[Polynomial conditions for solvability] \begin{enumerate}[label=(\roman*)] \item For each auxiliary equation, form an arbitrary polynomial multiplier of degree at most $B$, with one unknown coefficient for each monomial. \item Set the sum of the multiplied equations equal to $1$. Equate coefficients of all monomials of degree at most $B+101$, obtaining \[ M(p_1,\ldots,r_5)c=b, \] where $b$ is the coefficient vector of the constant polynomial $1$. \item Form all minors of $M$ and $[M\mid b]$. These are the desired integer polynomials in the position coordinates. \end{enumerate} \end{algorithm} For any specialization of the positions, the degree bound and Hilbert's Nullstellensatz identify nonsolvability of the auxiliary equations with solvability of this linear system. The latter is equivalent to $\operatorname{rank}M=\operatorname{rank}[M\mid b]$, which is determined by the vanishing or nonvanishing of the listed minors. Thus the list determines membership in $\mathcal Z$ itself. \end{proof} \subsection{Generic grids and independent square roots} We call a plane curve \emph{real-compatible} if its zero set is preserved by \[ (z,w)\longmapsto(\overline w,\overline z). \] This is the real structure corresponding to ordinary real $x_1,x_2$. A curve containing more real points than its degree squared is real-compatible: its image under this involution has the same degree, and B\'ezout forces the two irreducible curves to coincide. The curves to which we apply the next proposition have nonconstant $z$- and $w$-projections. A curve with constant $z$ or $w$ contains at most one point of the original real plane, so it cannot occur after the large-set restrictions used in the grid-selection argument. We use a generic point in the algebraic sense. A point of an irreducible variety $V$ over an algebraically closed field $B$ is generic over $B$ if its coordinates have transcendence degree $\dim V$ over $B$. A jointly generic tuple is generic on the product of its specified varieties. Generic evaluation identifies the function field $B(V)$ with the field generated by the generic point: the kernel is a prime ideal of the same dimension as $V$, and so is its defining prime ideal. The same statement holds after enlarging the coefficient field provided genericity is established again over that enlarged field. \begin{lemma}[Independent quadratic radicals]\label{ub:alg:kummer} Let $F$ be a field of characteristic different from $2$. Suppose $D_1,\ldots,D_m\in F^\times$ have independent classes in $F^\times/(F^\times)^2$. If $R_j^2=D_j$, then \[ [F(R_1,\ldots,R_m):F]=2^m, \] and every independent choice of signs for the $R_j$ is induced by an $F$-automorphism of this extension. In particular the single radicals $R_1,\ldots,R_m$ are linearly independent over $F$. \end{lemma} \begin{proof} We prove the degree and automorphism assertions by induction. They are immediate for one nonsquare. Assume them for the first $m-1$ radicals and write $E=F(R_1,\ldots,R_{m-1})$. If $D_m$ were a square in $E$, a root $u\in E$ would transform under each sign-change automorphism $\sigma$ as $\sigma u=\pm u$. The signs define a character of the group $(\mathbb Z/2\mathbb Z)^{m-1}$. The products $\prod_{j\in J}R_j$ form an $F$-basis of $E$ by the induction hypothesis and have distinct sign characters. Expanding $u$ in this basis shows that it is an $F$-multiple of one such product. Squaring contradicts independence of the classes of the $D_j$. Thus adjoining $R_m$ doubles the degree. Each preceding automorphism extends by either sign on $R_m$, because it fixes $D_m$; the assertions follow. Finally, in a linear relation between the single radicals, changing just one sign and subtracting kills every other term. \end{proof} \begin{proposition}[Generic avoidance]\label{ub:alg:generic-avoidance} Let $V_1,V_2$ each be either $\mathbb A^2_{\C}$ or an irreducible real-compatible plane curve with nonconstant coordinate projections. Suppose they are not both the same real unit circle. A jointly generic point of $V_1^{20}\times V_2^5$ does not belong to $\mathcal Z$. \end{proposition} \begin{proof} Take a generic tuple over $B=\C$ in an algebraically closed extension $\Omega$ of $B$ and suppose a solution of \eqref{ub:alg:radical-equations}--\eqref{ub:alg:nonzero-flow} exists in $\Omega$, enlarging $\Omega$ if necessary. Put $d=\dim V_1$. Let $K$ be the algebraic closure, inside $\Omega$, of \[ B(r_1,\ldots,r_5,U_1,\ldots,U_5,A'_1,\ldots,A'_5). \] The last fifteen coordinates are arbitrary and may depend on all left positions. Adjoining them nevertheless reduces transcendence degree by at most fifteen. Joint genericity gives \[ \operatorname{trdeg}_{K}K(p_1,\ldots,p_{20})\ge20d-15. \] If every $p_i$ had transcendence degree at most $d-1$ over $K$, subadditivity would give an upper bound $20(d-1)$, a contradiction. Choose a surviving point $p=p_i$ generic on $V_1$ over $K$. Hence \[ F=K(V_1)=K(p). \] This argument is why independence over the enlarged field is valid; it has not been assumed automatically. Write $a_j=z_{r_j}$, $b_j=w_{r_j}$ and consider the functions \[ L_j(z,w)=(z-a_j)(w-b_j),\qquad D_j=L_j(L_j-4)\in F. \] We first prove the following stronger assertion: for every $j$ there is an integer-valued valuation $v_j$ of $F$ with \begin{equation}\label{ub:alg:private-valuation} v_j(D_j)=1,\qquad v_j(D_l)=0\quad(l\ne j). \end{equation} The nonvanishing assertions in its construction will be established using the original joint genericity over $B$, before any flow data are adjoined. \emph{Plane case.} If $V_1=\mathbb A^2$, the polynomial \[ Q_j=(z-a_j)(w-b_j)-4 \] is irreducible in $K[z,w]$. Indeed its two coefficients as a linear polynomial in $w$ are coprime in $K[z]$, and it is linear over $K(z)$. The polynomials $Q_j$ are pairwise distinct: their normalized coefficients recover $a_j,b_j$. None divides any $L_l$, a product of two linear polynomials. The exponent of $Q_j$ in a rational function is therefore the valuation in \eqref{ub:alg:private-valuation}. \emph{A curve with a different anchor locus or a multiple projection.} Let $V_1$ have irreducible equation $f(z,w)=0$. Suppose either $V_1\ne V_2$ or $\deg_w f\ge2$. Fix $j$ and put $a=a_j$. The coordinate $a$ is transcendental over $B$. Since $f$ is irreducible and has positive $w$-degree, characteristic zero implies that $f$ and $f_w$ are coprime in $B(z)[w]$. A B\'ezout identity in this polynomial ring, with denominators cleared, shows that $f(a,w)$ has no repeated root; its degree is unchanged because its leading coefficient cannot vanish at the transcendental $a$. If $V_1\ne V_2$, genericity puts $r_j$ outside $V_1$, so any root $b$ of $f(a,w)$ differs from $b_j$. If $V_1=V_2$ and $\deg_w f\ge2$, choose a root other than $b_j$. Thus in both cases \[ s_j=(a,b)\in V_1, \qquad f_w(a,b)\ne0, \qquad b-b_j\ne0. \] The element $b$ is algebraic over $B(a)$ and belongs to $K$ because $K$ is algebraically closed and contains $a$. For $l\ne j$, the anchor $r_l$ remains generic on $V_2$ over $B(s_j)$: $s_j$ is algebraic over one coordinate of $r_j$, and the anchors were jointly generic. The two projections of $V_2$ are nonconstant, so neither coordinate of $r_l$ equals its counterpart in $s_j$. Consequently $L_l(s_j)\ne0$. Also $L_l(s_j)\ne4$. Otherwise genericity would force $V_2$ to be contained in the curve \[ (z-a)(w-b)=4. \] This is impossible for a plane. For a curve, irreducibility would make its equation proportional to this one. After normalizing the $zw$ coefficient, the $w$ coefficient would put $a$ in $B$, contrary to its transcendence. These inequalities are assertions about the chosen elements of $\Omega$ and remain true after passing to $K$. There is a unique formal solution \[ w(T)=b+c_1T+c_2T^2+\cdots\in K[[T]],\qquad f(a+T,w(T))=0. \] At the $m$th step the new coefficient occurs with the nonzero coefficient $f_w(a,b)$, which proves existence recursively. The substitutions $z\mapsto a+T$, $w\mapsto w(T)$ embed $F$ in $K((T))$: $a+T$ is transcendental over $K$ and the irreducible equation of $V_1$ is the only polynomial relation. Now \[ L_j(a+T,w(T))=T\bigl(b-b_j+O(T)\bigr) \] has a simple zero. The functions $L_j-4$ and $D_l$ for $l\ne j$ have nonzero constant terms. Order in $T$ proves \eqref{ub:alg:private-valuation} in this case. \emph{A common curve of projection degree one.} The only remaining case has $V_1=V_2$ and $\deg_w f=1$. Real compatibility gives $\deg_z f=1$ as well. Thus the common curve has one of the forms \begin{equation}\label{ub:alg:degree-one-forms} w=ez+e',\quad e\ne0, \qquad\hbox{or}\qquad (z-u)(w-v)=b,\quad b\ne0. \end{equation} To see this directly, write its equation as $c_{11}zw+c_{10}z+c_{01}w+c_{00}=0$. If $c_{11}=0$, nonconstant projections give the first form. Otherwise completion of the product gives the second form, with $b\ne0$ by irreducibility. In the second form, real compatibility implies $v=\overline u$ and $b\in\R$. If $b=1$, the real locus is the unit circle centred at $u$, excluded by the hypothesis. Hence \begin{equation}\label{ub:alg:radius-not-one} b\ne1. \end{equation} For a line in \eqref{ub:alg:degree-one-forms}, $L_j=e(z-a_j)^2$. Choose $c$ with $c^2=4/e$ and take the point $s_j$ of the line with $z$-coordinate $a_j+c$. The function $L_j-4$ has a simple zero there, and $L_j=4$. For $l\ne j$, either equation $L_l(s_j)=0$ or $L_l(s_j)=4$ would make $a_l$ algebraic over $B(s_j)$, contradicting the independence of the anchors. Order at this point of the rational function field $K(z)$ gives \eqref{ub:alg:private-valuation}. For the second curve in \eqref{ub:alg:degree-one-forms}, put $\zeta=(z-u)/(a_j-u)$. Since $a_j$ is generic, $a_j-u\ne0$. Substitution gives \[ L_j=-b\frac{(\zeta-1)^2}{\zeta}. \] Zeros of $L_j-4$ correspond to roots of \[ b\zeta^2+(4-2b)\zeta+b=0. \] Its discriminant is $16(1-b)\ne0$; its roots are nonzero since $b\ne0$. Either root gives a finite point $s_j$ with $z_{s_j}\ne u$ at which $L_j-4$ has a simple zero and $L_j=4$. This point is algebraic over $B(a_j)$. As a function of the independent coordinate $a_l$, for $l\ne j$, the value $L_l(s_j)$ is \[ -b\frac{(z_{s_j}-a_l)^2} {(z_{s_j}-u)(a_l-u)}. \] It is nonconstant, since its numerator has degree two, its denominator has degree one, and $z_{s_j}\ne u$. It therefore cannot take the values $0$ or $4$ at the transcendental $a_l$ over $B(s_j)$. Order at $s_j$ again proves \eqref{ub:alg:private-valuation}. We have exhausted all cases. The valuations show that no nonempty product of the $D_j$ is a square: a valuation indexed by one factor of that product has odd value on it. Thus the $D_j$ have independent square classes in $F$, and Lemma~\ref{ub:alg:kummer} applies to the actual roots $R_j=R_{ij}$ for the surviving point $p_i=p$. We finish with a linear system. Set $H=H_i$, $A=A_i$ and rewrite \eqref{ub:alg:flow-equations} as \begin{equation}\label{ub:alg:five-linear-equations} 2(p-r_j)\mathbin{\cdot}H-R_jA =2(p-r_j)\mathbin{\cdot}U_j-R_jA'_j \qquad(1\le j\le5). \end{equation} If the first three rows $p-r_1,p-r_2,p-r_3$ have rank two, the determinant of their three equations in $H_1,H_2,A$ is an $F$-linear combination of $R_1,R_2,R_3$ with at least one nonzero coefficient. Indeed its coefficients, up to signs and factors of two, are the two-by-two minors of the position rows. The determinant is nonzero by Lemma~\ref{ub:alg:kummer}. Cramer's rule places $H_1,H_2,A$ in $F(R_1,R_2,R_3)$. Apply to the fourth equation the automorphism that changes only $R_4$ and subtract the original equation. Since $R_4\ne0$, this gives $A=A'_4$, contradicting \eqref{ub:alg:nonzero-flow}. Here is the complete rank check. If $V_2$ is a plane or a curve other than a line, three jointly generic anchors are noncollinear: after fixing two distinct anchors, containment of a generic third on their line would force its entire irreducible locus to be that line. The three position rows then have rank two for every $p$. If $V_2$ is a line, the first two rows already have rank two unless $p$ is on that line. As $p$ is generic on $V_1$ over $K$, the latter possibility forces $V_1=V_2$ to be that same line. In this last case write $p-r_j=s_jv$ for a nonzero direction $v\in K^2$ and $s_j\in F^\times$. Only the unknown $h=v\cdot H$ is needed. The first two equations in $h,A$ have determinant $2(s_2R_1-s_1R_2)\ne0$. Thus $h,A\in F(R_1,R_2)$, and changing the sign of $R_3$ in the third equation gives $A=A'_3$, again a contradiction. The surviving left point is distinct from the anchors by its genericity over $K$, which justifies $s_j\ne0$. This proves generic avoidance. Notice that the flow variables have nowhere been required to be derivatives of the positions. \end{proof} \subsection{An effective height bound for finite nonexceptional grids} We recall the precise effective algebraic input used in the height calculation. If $F_1,\ldots,F_s\in L[X_1,\ldots,X_n]$, where $L$ is a number field, have degree at most $d$ and no common zero over an algebraic closure, then there are $G_i\in L[X]$ with \begin{equation}\label{ub:alg:kps-degree} 1=\sum_iG_iF_i, \qquad \deg G_i\le4nd^n. \end{equation} This is the degree assertion of \cite[Theorem~3.6]{ub:KPS2001}. Although that theorem is stated with algebraic-integer coefficients and a nonzero constant on the right, clearing denominators and dividing by that constant gives exactly \eqref{ub:alg:kps-degree}. We do not use the height assertion of the quoted theorem. For a nonzero polynomial with algebraic coefficients, its coefficient height means the absolute projective height of its coefficient vector. If $F(\alpha)=0$ and $F\ne0$, the local Cauchy root bound gives \begin{equation}\label{ub:alg:root-height} \hgt(\alpha)\le\hgt(F)+\log2. \end{equation} Indeed, after dividing by the leading coefficient, at a nonarchimedean place a root has absolute value at most the maximum of $1$ and the absolute values of the remaining coefficients. At an archimedean place the same bound with a factor $2$ follows from the geometric-series estimate. Sum these local inequalities; the leading coefficient disappears by the product formula. \begin{theorem}[Numerical critical-value bound]\label{ub:alg:height-bound} Suppose a grid has algebraic position coordinates and nonzero algebraic numbers $\lambda_{ij}$ such that \[ L_{ij}=2-\lambda_{ij}-\lambda_{ij}^{-1}. \] For $2\le i\le20$, $2\le j\le5$ set \[ \chi_{ij}=\frac{\lambda_{ij}\lambda_{11}} {\lambda_{i1}\lambda_{1j}}. \] If its position tuple is outside $\mathcal Z$, at least one pair $e=(i,j)$ satisfies \begin{equation}\label{ub:alg:numerical-height} \hgt(\lambda_e)\le C_{\rm alg} \left(1+\sum_{i=2}^{20}\sum_{j=2}^{5}\hgt(\chi_{ij})\right), \qquad C_{\rm alg}=10^{100000}. \end{equation} \end{theorem} \begin{proof} Fix the actual values of the $76$ numbers $\chi_{ij}$ and put $L=\Q(\chi_{ij}:i\ge2,j\ge2)$. Consider the following affine scheme in $150$ variables over $L$: its variables are the $50$ position coordinates and the $100$ quantities $\lambda_{ij}$, and its $176$ defining equations are \begin{align} f_{ij}&=\lambda_{ij}^2+(L_{ij}-2)\lambda_{ij}+1=0, &&1\le i\le20,\quad1\le j\le5,\label{ub:alg:fiber-distance}\\ g_{ij}&=\lambda_{ij}\lambda_{11} -\chi_{ij}\lambda_{i1}\lambda_{1j}=0, &&2\le i\le20,\quad2\le j\le5.\label{ub:alg:fiber-rectangle} \end{align} Write these equations as $F_1=\cdots=F_{176}=0$ and the variables as $X_1,\ldots,X_{150}$. Their total degree is at most three. Every $\lambda_{ij}$ is nonzero at a solution, by the constant term in \eqref{ub:alg:fiber-distance}. Let $x$ denote the actual solution. We use the literal kernel of the Jacobian matrix at $x$, so the argument does not presuppose smoothness or a radical presentation. For a vector in this kernel write $t_{ij}=d\lambda_{ij}/\lambda_{ij}$. Differentiating \eqref{ub:alg:fiber-rectangle} gives \[ t_{ij}+t_{11}-t_{i1}-t_{1j}=0. \] Consequently $t_{ij}=A_i-A'_j$ on the full grid, with $A_i=t_{i1}$ and $A'_j=t_{11}-t_{1j}$. Differentiating the distance relation gives \[ dL_{ij}=(\lambda_{ij}^{-1}-\lambda_{ij})t_{ij}. \] The position part of the tangent vector supplies $H_i,U_j$, so it satisfies \eqref{ub:alg:flow-equations} with $R_{ij}=\lambda_{ij}^{-1}-\lambda_{ij}$. The identity $R_{ij}^2=L_{ij}(L_{ij}-4)$ is immediate by expanding both sides. Thus a tangent vector for which all $d\lambda_{ij}$ are nonzero would put the actual position tuple in $\mathcal Z$. A finite-dimensional vector space over an infinite field cannot be a union of finitely many proper linear subspaces. For example, the product of their nonzero defining linear forms is a nonzero polynomial and cannot vanish on the entire vector space. It follows that for some fixed edge $e$, the linear functional $d\lambda_e$ vanishes on the entire Jacobian kernel at $x$. By finite-dimensional linear algebra its row vector is in the row span of that Jacobian. Choose $176$ multipliers $a_1,\ldots,a_{176}$ realizing this row-span relation. Introduce these multipliers as new variables and define a closed affine algebraic set $C_e$ by \begin{align} F_l(X)&=0 &&(1\le l\le176),\label{ub:alg:critical-first}\\ \frac{\partial\lambda_e}{\partial X_v} &=\sum_{l=1}^{176}a_l\frac{\partial F_l}{\partial X_v} &&(1\le v\le150).\label{ub:alg:critical-second} \end{align} The chosen point lifts to $C_e$. There are $326$ variables and $326$ equations, all of degree at most $3$. The combined vector of input coefficients, together with $1$, has height at most \begin{equation}\label{ub:alg:input-height} H\le\sum_{i,j}\hgt(\chi_{ij})+\log4. \end{equation} In fact every coefficient is $0$, an integer of absolute value at most $2$, or $\pm\chi_{ij}$; the factor $4$ safely allows all displayed expansions. The projection of $C_e$ to its $\lambda_e$ coordinate is finite. To prove this, take an irreducible component over $\overline L$ and its function field $T$. Equations \eqref{ub:alg:critical-first} and \eqref{ub:alg:critical-second} imply, for every $\overline L$-derivation $D:T\to T$, \[ D\lambda_e=\sum_l a_lDF_l=0. \] If $\lambda_e$ were transcendental over $\overline L$, it could be included in a transcendence basis. Assigning it derivative $1$ and the other basis elements derivative $0$ gives a derivation of the corresponding rational-function field. It extends uniquely through the finite separable algebraic extension to $T$: for an algebraic element with minimal polynomial $P$, implicit differentiation defines its derivative since $P'$ is nonzero. This contradicts the preceding identity. Thus $\lambda_e$ is algebraic over the algebraically closed field $\overline L$, and is constant on that component. There are finitely many components, which proves the assertion. This reasoning applies at singular points as well; it asserts constancy on components of $C_e$, not on the original distance--rectangle fibre. We next bound these finitely many critical values. Put $n=326$ and \[ D_0=4n3^n. \] Adjoin to the equations of $C_e$ the equation $\lambda_e-T=0$, where $T$ is one new parameter, not an additional unknown for the degree calculation. For every rational specialization $T=t$ outside the finite projection, the resulting equations have no common zero. By \eqref{ub:alg:kps-degree} there is a representation of $1$ with all multiplier degrees in the $n$ unknowns at most $D_0$. Expand those multipliers and equate monomial coefficients. This is a finite linear system \[ M(T)c=b \] over $L[T]$. Its entries have degree at most one in $T$, and its number of rows is at most \begin{equation}\label{ub:alg:row-bound} R=\binom{n+D_0+3}{n}\le(8n3^n)^n. \end{equation} The system is compatible for all but finitely many rational $t$. Ranks of a univariate polynomial matrix and its augmentation are constant away from finitely many values, as follows by inspecting their nonzero maximal minors. Therefore it is compatible over $L(T)$. Choose a nonzero maximal pivot minor $\Delta(T)$ of $M(T)$. Using its pivot columns and rows, Cramer's rule gives a solution whose only denominator is $\Delta(T)$; compatibility ensures that the other rows are satisfied as well. Clearing this denominator produces the polynomial identity \begin{equation}\label{ub:alg:eliminant-identity} \sum_l G_l(X,T)F_l^{\rm crit}(X) +G_*(X,T)(\lambda_e-T)=\Delta(T), \end{equation} where the $F_l^{\rm crit}$ are the $326$ critical equations. At any point of $C_e$, substitute $T=\lambda_e$ into \eqref{ub:alg:eliminant-identity}. It follows that $\Delta(\lambda_e)=0$, and $\Delta$ is a nonzero polynomial. All steps were performed after fixing the actual $\chi_{ij}$. In particular this is not a specialization of a generic eliminant whose coefficients might all disappear at the actual parameter. We finish the arithmetic estimate explicitly. At a place $v$ of $L$, let $A_v$ be the maximum of $1$ and the absolute values of all input coefficients. For a matrix entry linear in $T$, its archimedean coefficient $\ell^1$ norm is at most $2A_v$ and its nonarchimedean Gauss norm is at most $A_v$. A determinant of order $r\le R$ consequently has local coefficient norm at most $r!(2A_v)^r$ at infinity and $A_v^r$ at a finite place. Summing with normalized place weights and using \eqref{ub:alg:input-height} gives \[ \deg\Delta\le R, \qquad \hgt(\Delta)\le R\bigl(H+\log(2R)\bigr). \] Combining with \eqref{ub:alg:root-height} yields \begin{equation}\label{ub:alg:height-before-rounding} \hgt(\lambda_e) \le R\left(\sum_{i,j}\hgt(\chi_{ij})+\log4+\log(2R)\right)+\log2. \end{equation} No number-field degree occurs in this estimate. For an entirely elementary rounding of \eqref{ub:alg:row-bound}, use $\log_{10}(8\cdot326)<4$ and $\log_{10}3<1/2$ to obtain \[ \log_{10}R<326(4+326/2)=54442<55000. \] Thus the right side of \eqref{ub:alg:height-before-rounding} is less than $10^{55006}(1+\sum_{i,j}\hgt(\chi_{ij}))$, which is certainly bounded by \eqref{ub:alg:numerical-height}. This proves the theorem. \end{proof} \begin{corollary}[A numerical contradiction on one grid] \label{ub:alg:height-contradiction} Put $K=10^{100004}$. There is no algebraic grid outside $\mathcal Z$ for which \[ J\ge K,\qquad \hgt(\lambda_{ij})\ge J/3\quad\hbox{for all }i,j, \qquad \hgt(\chi_{ij})\le J/K\quad(i,j\ge2), \] and $L_{ij}=2-\lambda_{ij}-\lambda_{ij}^{-1}$. \end{corollary} \begin{proof} There are $76$ displayed defects. Theorem~\ref{ub:alg:height-bound} would give, for some $e$, \[ \hgt(\lambda_e) \le10^{100000}\left(1+\frac{76J}{K}\right) \le\frac{77}{10000}J0\), \(|B|=b>0\), then for positive integers \(r,s\), \[ \operatorname{hom}(K_{r,s},G) \ge \frac{e^{rs}}{a^{rs-r}b^{rs-s}}. \] Here a homomorphism is an ordered grid; repetitions are allowed. \end{lemma} \begin{proof} For an ordered \(r\)-tuple \(\boldsymbol p\in A^r\), let \(c(\boldsymbol p)\) be its number of common neighbors. Then \[ \sum_{\boldsymbol p\in A^r}c(\boldsymbol p) =\sum_{q\in B}\deg(q)^r\ge b(e/b)^r. \] Convexity, now applied to the \(s\)-th power, gives \[ \sum_{\boldsymbol p\in A^r}c(\boldsymbol p)^s \ge a^r\left(\frac{e^r}{a^rb^{r-1}}\right)^s, \] which is the claimed bound. \end{proof} \begin{theorem}[Selection of a generic polynomial pattern] \label{ub:gs:selection} Fix a finite collection \(F_1,\ldots,F_L\) of complex polynomials in the coordinates of twenty left points and five right points. After discarding zero polynomials, put \[ D=\max\left\{1,\sum_{\nu=1}^L\deg F_\nu\right\}. \] Fix \(0<\varepsilon\le10^{-8}\). For all sufficiently large \(t\), depending only on \(D\) and \(\varepsilon\), the following holds. Let \(A,B\) be finite subsets of the physical real plane and \(G\subset A\times B\), with \[ |A|,|B|\le t^{1+\varepsilon},\qquad |G|\ge t^{2-\varepsilon}. \] There are loci \(V_A,V_B\), each a plane or a curve of degree at most \(D\), and a complete ordered \(20\)-by-\(5\) grid in \(G\), such that its vertices lie on these loci and, for every \(\nu\), \begin{equation}\label{ub:gs:pattern} F_\nu(\text{grid})=0 \quad\Longleftrightarrow\quad F_\nu|_{V_A^{20}\times V_B^5}\equiv0. \end{equation} Every curve locus is invariant under \(\iota\), and both its coordinate projections are nonconstant. Distinctness of all twenty-five vertices may be required by adding the pairwise differences of their \(z\) coordinates to the fixed polynomial collection. If, additionally, \(A,B\subset P\) and no edge of \(G\) has both endpoints on a real unit circle containing at least \(t^{9/10}\) points of \(P\), the loci \(V_A,V_B\) cannot both be that same unit circle. \end{theorem} \begin{proof} We keep a subgraph, a locus for each side, and a parameter \(\gamma\) such that \begin{equation}\label{ub:gs:invariants} e\ge t^{2-\gamma},\qquad d_{\min}\ge t^{1-\gamma-2\varepsilon},\qquad |A'|,|B'|\le t^{1+\varepsilon}. \end{equation} Initially take both loci to be planes. Repeatedly delete a vertex whose current degree is below \(t^{1-3\varepsilon}\). The total deletion cost is at most \(2t^{2-2\varepsilon}\). If \(t^\varepsilon\ge4\), at least half of the original edges survive, and their number is at least \(t^{2-2\varepsilon}\). Thus \eqref{ub:gs:invariants} holds initially with \(\gamma=2\varepsilon\). Suppose a side whose current locus is a plane has at least \(t^{1-300\gamma}\) vertices on the zero set of a nonzero polynomial of degree at most \(D\). One of its at most \(D\) irreducible factors contains at least \(D^{-1}t^{1-300\gamma}\) such vertices. Restrict that side to this irreducible curve. The old minimum degree gives at least \[ D^{-1}t^{2-301\gamma-2\varepsilon} \ge t^{2-301\gamma-3\varepsilon} \] remaining edges, provided \(t^\varepsilon\ge D\). Prune again, now at degree threshold \(t^{1-301\gamma-5\varepsilon}\). The number of deleted edges is at most \(2t^{2-301\gamma-4\varepsilon}\), so at least \(t^{2-301\gamma-4\varepsilon}\) edges remain for \(t^\varepsilon\ge4\). Reset \(\gamma\) to \(305\gamma\). The inequalities \[ 301\gamma+4\varepsilon\le305\gamma, \qquad 301\gamma+5\varepsilon\le305\gamma+2\varepsilon \] show that \eqref{ub:gs:invariants} is preserved. Each such step replaces one plane locus by a curve. There can therefore be at most two steps. After each step we repeat the concentration test on every side still assigned the plane, using the new value of \(\gamma\). At termination, \begin{equation}\label{ub:gs:gamma} 2\varepsilon\le\gamma\le 2\cdot305^2\varepsilon =186050\varepsilon. \end{equation} On a plane side, a nonzero polynomial of degree at most \(D\) has fewer than \(t^{1-300\gamma}\) zeros among the retained vertices. On a curve side, a polynomial of degree at most \(D\) not vanishing identically on that curve has at most \(D^2\) zeros, by the plane-curve B\'ezout bound. Thus the one-block zero bound is \[ Z=\max\{t^{1-300\gamma},D^2\}. \] Multiply together those \(F_\nu\) which are not identically zero on \(V_A^{20}\times V_B^5\); denote their product by \(F\), with \(F=1\) if the collection is empty. The product of irreducible varieties over \(\C\) is irreducible, so \(F\) is not identically zero, and its degree is at most \(D\). Lemma~\ref{ub:gs:zero-count} bounds the number of ordered vertex tuples on which \(F\) vanishes by \begin{equation}\label{ub:gs:bad-tuples} 25\max\{t^{1-300\gamma},D^2\}\,t^{24(1+\varepsilon)}. \end{equation} On the other hand, Lemma~\ref{ub:gs:hom-count} gives at least \begin{equation}\label{ub:gs:good-grids} \frac{e^{100}}{|A'|^{80}|B'|^{95}} \ge t^{25-100\gamma-175\varepsilon} \end{equation} complete ordered grids. The exponent in \eqref{ub:gs:good-grids} exceeds the two exponents in \eqref{ub:gs:bad-tuples}, respectively, by \[ 200\gamma-199\varepsilon>0, \qquad 1-100\gamma-199\varepsilon \ge1-18605199\cdot10^{-8}=0.81394801>0. \] For sufficiently large \(t\), the fixed factors \(25\) and \(D^2\) are also absorbed. A complete grid with \(F\ne0\) exists. The polynomials omitted from the product vanish on every tuple of the product loci, so this grid satisfies \eqref{ub:gs:pattern}. By the minimum-degree condition, each retained side contains at least \(t^{1-\gamma-2\varepsilon}\) vertices. For sufficiently large \(t\) this exceeds \(D^2\). A retained curve and its image under \(\iota\) share all these physical real points; B\'ezout therefore makes the two irreducible curves equal. A curve with constant \(z\), or constant \(w\), contains at most one physical real point, so both projections are nonconstant. In particular, the coordinate differences added to enforce distinctness are not identically zero on the relevant product of loci. Condition \eqref{ub:gs:pattern} then makes these differences nonzero on the selected grid. Finally, \[ \gamma+2\varepsilon\le0.00186052<1/10. \] If both loci were the same real unit circle, that circle would contain more than \(t^{9/10}\) points of \(P\). Every retained edge would be forbidden by the last hypothesis. This proves the final assertion. \end{proof} \begin{remark}\label{ub:gs:threshold} All size thresholds in the proof are uniform in the polynomial coefficients. For example, it is enough to make \[ t^\varepsilon\ge\max\{100,D\},\qquad t^{0.81394801}>100D^2. \] The exponent \(10^{-8}\) in the theorem is independent of \(D\). It is neither asserted nor needed that this threshold is smaller than the arithmetic parameter used elsewhere in the proof. \end{remark} \begin{corollary}[Avoidance of a constructible set]\label{ub:gs:avoid} Let \(Z\) be a fixed constructible subset of the space of \(20\)-by-\(5\) position grids, described by a finite Boolean combination of polynomial equations over \(\Q\). Suppose the generic point of \(V_A^{20}\times V_B^5\) lies outside \(Z\) for every pair of loci allowed in Theorem~\ref{ub:gs:selection}, apart from two copies of the same real unit circle. Under the hypotheses of that theorem, including the popular-circle exclusion, the graph contains a complete grid outside \(Z\). \end{corollary} \begin{proof} Use the polynomials in a Boolean description of \(Z\) as the fixed collection in Theorem~\ref{ub:gs:selection}, adding the coordinate differences for distinctness. Membership in \(Z\) depends solely on which of these polynomials vanish. The selected grid has the same pattern as the generic point of the product of its loci. The common unit circle is excluded by the theorem, so that generic point and the selected grid both lie outside \(Z\). \end{proof} \begin{theorem}[Explicit finite obstruction]\label{ub:gs:finite-obstruction} Set \(K=10^{100004}\). There exists an absolute \(t_0\) such that no data of the following kind exist when \(t\ge t_0\): a finite algebraic point set \(P\) in the physical real plane, a graph \(G\subset P\times P\), a real number \(J\ge K\), and nonzero algebraic edge ratios \(\lambda_{pr}\), satisfying \begin{align} |P|&\le t^{1+1/K},& |G|&\ge t^{2-1/K},\label{ub:gs:obstruction-size}\\ (z_p-z_r)(w_p-w_r)&=2-\lambda_{pr}-\lambda_{pr}^{-1}, &\hgt(\lambda_{pr})&\ge J/3,\label{ub:gs:obstruction-height}\\ \hgt\left(\frac{\lambda_{pr}\lambda_{p'r'}} {\lambda_{pr'}\lambda_{p'r}}\right)&\le J/K &&\text{on every complete rectangle.}\label{ub:gs:obstruction-defect} \end{align} Moreover, no edge may have both endpoints on a real unit circle containing at least \(t^{9/10}\) points of \(P\). \end{theorem} \begin{proof} Apply Corollary~\ref{ub:gs:avoid} with \(\varepsilon=10^{-8}\), the constructible set of Definition~\ref{ub:alg:bad-set}, its finite Boolean description from Lemma~\ref{ub:alg:boolean}, and the generic avoidance statement of Proposition~\ref{ub:alg:generic-avoidance}. This is permitted because \(1/K<10^{-8}\). The selected \(20\)-by-\(5\) grid is outside that set. Its seventy-six normalized rectangle defects \(\chi_{ij}\), \(2\le i\le20\), \(2\le j\le5\), all satisfy \(\hgt(\chi_{ij})\le J/K\). Theorem~\ref{ub:alg:height-bound} gives an edge for which \[ \hgt(\lambda_e) \le 10^{100000}\left(1+\sum_{i=2}^{20}\sum_{j=2}^{5} \hgt(\chi_{ij})\right) \le10^{100000}+\frac{76J}{10000}. \] Since \(J\ge K=10^4\cdot10^{100000}\), the right side is at most \(77J/10000=0.0077J0\}\le |A_i|/2, \qquad \#\{p\in A_i:f(p)<0\}\le |A_i|/2 \] for every \(i\). Points at which \(f\) vanishes are omitted from both inequalities. \end{lemma} \begin{proof} For a positive \(\rho\), replace each point of \(A_i\) by a smooth nonnegative bump of integral one supported in its radius-\(\rho\) disc. Let \(\mu_{i,\rho}\) be the sum of these measures. The real space of polynomials of degree at most \(d\) has dimension \(m=\binom{d+2}{2}\). On its coefficient unit sphere define \[ \Phi_\rho(f)_i =\int_{\R^2}\operatorname{sgn}(f(x))\,d\mu_{i,\rho}(x). \] This map is odd and continuous. To see continuity, a nonzero polynomial has a planar zero set of Lebesgue measure zero, by Fubini applied to its one-variable fibers, and coefficient convergence gives pointwise sign convergence away from that zero set. Dominated convergence then applies. Pad the map with zero coordinates if necessary and apply Borsuk--Ulam on \(S^{m-1}\). There is a unit-coefficient polynomial \(f_\rho\) for which all integrals vanish. Choose \(\rho\downarrow0\) along a sequence and take a convergent subsequence of the coefficient vectors. Its limit \(f\) has coefficient norm one, so is nonzero. For each original point with \(f(p)>0\), all the mass of its bump eventually lies in the positive set of \(f_\rho\): coefficient convergence is uniform on a fixed compact neighborhood, and \(f\) is positive in a sufficiently small neighborhood of \(p\). The positive mass of every \(\mu_{i,\rho}\) is exactly \(|A_i|/2\), because its signed integral is zero and a polynomial zero set has measure zero. Thus there can be at most \(|A_i|/2\) original points with \(f(p)>0\). The argument for negative values is identical. \end{proof} \begin{lemma}[Planar partition into sign regions]\label{ub:ub:partition} For a finite \(P\subset\R^2\), \(|P|=M\), and an integer \(D\ge1\), there is a nonzero real polynomial \(f\) of degree at most \(D\), and a partition of \(\R^2\setminus Z(f)\) into at most \(D^2\) open sign regions, each containing at most \(256M/D^2\) points of \(P\). A unit circle not contained in \(Z(f)\) meets at most \(2D+1\) of these regions. The sign regions need not be connected. \end{lemma} \begin{proof} For \(D<16\), take \(f=1\) and the single region \(\R^2\). Suppose now that \(D\ge16\), and put \[ j=\left\lfloor2\log_2(D/8)\right\rfloor. \] At step \(\ell=1,\ldots,j\), apply Lemma~\ref{ub:ub:bisection} to the at most \(2^{\ell-1}\) point classes specified by the nonzero signs of the previously chosen polynomials. We may choose a nonzero bisecting polynomial \(f_\ell\) of degree \[ d_\ell=\lceil2^{\ell/2}\rceil, \] since \(\binom{d_\ell+2}{2}-1\ge d_\ell^2/2\ge2^{\ell-1}\). Each final sign class has at most \(M/2^j\) points. Put \(f=\prod_{\ell=1}^j f_\ell\). The nonempty sets \[ \{x:\operatorname{sgn}(f_\ell(x))=\sigma_\ell, \ 1\le\ell\le j\},\qquad \sigma_\ell\in\{-1,1\}, \] form the required open sign regions. The estimates \[ \deg f\le\sum_{\ell=1}^j(2^{\ell/2}+1) \le8\,2^{j/2}\le D, \qquad D^2/128<2^j\le D^2 \] give both the number of regions and their point bound, with room in the stated constant \(256\). A circle not contained in \(Z(f)\) has at most \(2D\) intersection points with that zero set. One can use the degree-two B\'ezout bound here, or substitute the rational parametrization of a circle into \(f\): after multiplying by the denominator to the power \(D\), one obtains a nonzero polynomial of degree at most \(2D\), with the possible point at infinity counted projectively. On each connected arc left after removing these intersections, all the signs of the \(f_\ell\) are constant. There are at most \(2D+1\) such arcs (including the case of no intersections), so the circle meets at most that many sign regions. \end{proof} \begin{remark} The usual polynomial partition theorem describes connected components of a complement. The sign-region version above, sufficient here, avoids any separate bound on their number of connected components. The polynomial ham-sandwich construction is the standard mechanism behind polynomial partitioning; see \cite{ub:GK2015,ub:KMSh2012}. \end{remark} \begin{theorem}[Balanced-to-unbalanced transfer]\label{ub:ub:transfer} Let \(1<\beta\le4/3\), and suppose that, for every positive integer \(r\), \[ \mathcal U(r,r)\le C r^\beta. \] For positive integers \(n,m\), set \(s=\min\{n,m\}\) and \(M=\max\{n,m\}\). Then \begin{equation}\label{ub:ub:global} \mathcal U(n,m)\le C'\bigl(M+s^{2\beta-2}M^{2-\beta}\bigr), \end{equation} where \(C'\) depends only on \(C\) and \(\beta\). More precisely, \[ \begin{cases} \mathcal U(n,m)=O_{C,\beta}\bigl(s^{2\beta-2}M^{2-\beta}\bigr), &s\le M\le s^2,\\ \mathcal U(n,m)\le M+s(s-1)\le2M, &M\ge s^2. \end{cases} \] \end{theorem} \begin{proof} Replacing a unit circle by its center interprets an incidence as a cross-pair of points at distance one. Exchanging the two point sets gives \(\mathcal U(n,m)=\mathcal U(m,n)\). It is therefore enough to consider a set \(P\) of \(M\) points and a family \(\mathcal C\) of \(s\) distinct unit circles. First, for \(d(p)=\#\{C\in\mathcal C:p\in C\}\), pair counting gives \[ \sum_{p\in P}\binom{d(p)}2\le2\binom s2=s(s-1). \] Since \(d\le1+\binom d2\) for every nonnegative integer \(d\), \begin{equation}\label{ub:ub:pair} I(P,\mathcal C)\le M+s(s-1). \end{equation} This proves the stated bound for \(M\ge s^2\) and deals with \(s=1\). Assume \(s\le M\le s^2\). Set \[ D=\lfloor M/s\rfloor,\qquad T=\left\lceil\frac{256M}{D^2}\right\rceil. \] Then \(M/(2s)\le D\le M/s\), and \begin{equation}\label{ub:ub:scales} T=O(s^2/M),\qquad sD\le M, \qquad D^2T=O(M),\qquad D^2\le M. \end{equation} Use Lemma~\ref{ub:ub:partition}. Let \(\mathcal C_0\) consist of the circles contained in \(Z(f)\), and put \(\mathcal C_1=\mathcal C\setminus\mathcal C_0\). For each sign region \(\Omega_i\), let \(b_i\) be the number of circles of \(\mathcal C_1\) meeting it. The circle-crossing bound implies \[ \sum_i b_i\le(2D+1)s\le3sD. \] Each region contains at most \(T\) points. Split its \(b_i\) circles into batches containing at most \(T\) circles. Pad the points and each batch with distinct dummy points and distinct unit circles to obtain \(T\) points and \(T\) circles. Padding only adds incidences, so the balanced hypothesis bounds each batch by \(CT^\beta\). Thus the incidences off \(Z(f)\) are at most \begin{align*} \sum_i CT^\beta\left\lceil\frac{b_i}{T}\right\rceil &\le C\left(T^{\beta-1}\sum_i b_i+D^2T^\beta\right)\\ &\le O(C)MT^{\beta-1} \le O_{C,\beta}\bigl(s^{2\beta-2}M^{2-\beta}\bigr). \end{align*} An empty batch is omitted; counting an additional \(CT^\beta\) for each region only increases this upper bound. In particular, no balance between the point and circle counts of an individual region has been assumed. For every circle of \(\mathcal C_1\), at most \(2D\) points lie on \(Z(f)\), so these boundary incidences total at most \(2sD\le2M\). If \(k_0=|\mathcal C_0|\), then \(k_0\le D/2\): the distinct irreducible quadratic equations of these circles divide \(f\). For completeness, a circle having infinitely many common zeros with \(f\) must be a component by B\'ezout, which proves this divisibility claim. Applying the same pair count as in \eqref{ub:ub:pair} to \(\mathcal C_0\) yields \[ I(P,\mathcal C_0)\le M+2\binom{k_0}{2}\le M+D^2\le2M. \] This includes the possibility that every point of \(P\) lies on \(Z(f)\), as well as arbitrary concurrence of component circles. All boundary incidences have therefore been included. Combining the bounds proves \eqref{ub:ub:global}. In the middle range, \[ \frac{s^{2\beta-2}M^{2-\beta}}M =\left(\frac{s^2}{M}\right)^{\beta-1}\ge1, \] so the separate linear term can be absorbed there. \end{proof} \begin{corollary}[The explicit unbalanced exponent]\label{ub:ub:explicit} If \(\mathcal U(r,r)=O(r^{4/3-\delta})\) for a fixed \(0<\delta<1/3\), then \[ \mathcal U(n,m)=O\left(M+s^{2/3-2\delta}M^{2/3+\delta}\right). \] For \(s\le M\le s^2\), the non-linear term is \[ (sM)^{2/3}\left(\frac{s^2}{M}\right)^{-\delta}. \] In particular, writing \(M=s^\alpha\) with \(1\le\alpha\le2\) gives the exponent \(\frac23(1+\alpha)-(2-\alpha)\delta\) in \(s\). \end{corollary} \begin{proof} Substitute \(\beta=4/3-\delta\) into Theorem~\ref{ub:ub:transfer}; the displayed identities are immediate. \end{proof} \section{Completion and the explicit exponent}\label{ub:sec:completion} \begin{proof}[Proof of Proposition~\ref{ub:thm:main}] We first show that arbitrarily large point sets with at least $N^{4/3-\delta_1}$ ordered unit pairs cannot exist. Write $t=N^{1/3}$; the supposed number of ordered pairs is at least $t^{4-\eps}$. The incidence regularization gives the graph, degree bounds, marginals, and expansion proved in the preceding sections. Algebraic specialization preserves its required edges and distinct vertices. The bounded-height argument gives $S>S_0$ for all sufficiently large $t$. The explicit history and score estimates then yield the hypotheses of Proposition~\ref{ub:ex:organized}. Consequently, for every sufficiently large alleged counterexample there is a graph of ratios with \[ |P|\le t^{1+1/K},\quad |G|\ge t^{2-1/K},\quad J\ge K, \] all edge heights at least $J/3$, all rectangle defect heights at most $J/K$, and no edge on a popular unit circle. The finite obstruction proved in Theorem~\ref{ub:gs:finite-obstruction} excludes precisely this data for $t\ge t_0(K)$. The two lower bounds on $t$ can be imposed simultaneously. This contradiction proves a uniform bound for all sufficiently large $N$. Increasing its finite multiplicative constant handles the remaining positive integers. Since unordered pairs number half the ordered ones, this proves the claim for $u(N)$. Lemma~\ref{ub:lem:symmetry} gives the bound for $\mathcal U(N,N)$. It remains to justify the simpler numerical difference. Since $10<2^{10/3}$, \[ \log_2\!\left(2^{201}k^{100}\right) =201+102400(1024+100004\log_2 10) <35\cdot10^9<10^{11}-1. \] It follows that $2^{201}k^{100}+\log_2 3<2^{10^{11}}$, and therefore \[ \delta_1=\frac13\,2^{-2^{201}k^{100}} >2^{-2^{10^{11}}}=\delta_0. \] A bound with the larger difference implies the bound with the smaller one. This completes the proof. \end{proof} \begin{proof}[Proof of Proposition~\ref{ub:thm:unbalanced}] Apply the balanced-to-unbalanced transfer proved in the preceding section with $\beta=4/3-\delta$. Its mixed exponents are $2\beta-2=2/3-2\delta$ and $2-\beta=2/3+\delta$. If $M\le s^2$, the mixed term is at least $M$, because their ratio is $(s^2/M)^{\beta-1}\ge1$. The direct pair count proves the other regime. \end{proof} For illustration, if $M=s^\alpha$ with $1\le\alpha\le2$, the bound is \[ \mathcal U(s,s^\alpha)\ll s^{\frac23(1+\alpha)-(2-\alpha)\delta}. \] In particular, \begin{center} \begin{tabular}{c c} \toprule $M$ & Upper bound in terms of $s$\\ \midrule $s$ & $O(s^{4/3-\delta})$\\ $s^{5/4}$ & $O(s^{3/2-3\delta/4})$\\ $s^{3/2}$ & $O(s^{5/3-\delta/2})$\\ $s^{7/4}$ & $O(s^{11/6-\delta/4})$\\ $s^2$ & $O(s^2)$\\ \bottomrule \end{tabular} \end{center} These formulas describe the upper bound in each cardinality range. For $M\ge s^2$, the order $M$ is attained by placing $M$ distinct points on one unit circle and adding the other distinct circles. \clearpage \section{The final parameter regime}\label{ub:opt:statement} The finite framework supplies the geometric, arithmetic and probabilistic notation used below. In this parameter regime, $K,k,a,b,\varepsilon,S_0$ are given by~\eqref{ub:opt:parameter-table}. Each estimate in the following sections is proved with these choices. \begin{proposition}[The final quantitative estimate]\label{ub:opt:main} Put \[ \delta_{\mathrm{up}} = \frac{2^{-2^{664}}}{514560000000}, \qquad \delta_{\rm simple}=2^{-2^{665}}. \] There is an absolute finite constant $C$ such that for every positive integer $N$, \[ u(N)\le C N^{4/3-\delta_{\mathrm{up}}},\qquad \mathcal U(N,N)\le C N^{4/3-\delta_{\mathrm{up}}}. \] For $s=\min(n,m)$ and $M=\max(n,m)$, \[ \mathcal U(n,m)\ll M+s^{2/3-2\delta_{\mathrm{up}}}M^{2/3+\delta_{\mathrm{up}}}. \] All these conclusions also hold with $\delta_{\rm simple}$ in place of $\delta_{\mathrm{up}}$. \end{proposition} The proof uses four improvements: a larger expansion exponent and a smaller matching-label constant; a second-moment small-set estimate; one prediction history controlling all subsequent bounded comparison vectors, which makes restricted-support transfer linear; and a rank-sensitive arithmetic height bound for a six-by-five grid. The final refinement keeps the logarithmic Jacobian and the full rigidity-rank pattern, so that only the actual geometric rank cases enter the height calculation. The intermediate eight-by-four calculation is included to make the algebraic reductions and their numerical consequences separately checkable. The final obstruction uses $K=2^{68}$. \paragraph{Relation to public results.} The power saving of \cite{ub:OAI2026} asserts the existence of an absolute positive exponent difference without specifying its size. The public grid-refinement drafts \cite{ub:GridRefinement2026,ub:GridJoint2026} contain eight-by-four and six-by-five reductions, respectively, but do not supply an explicit final power saving. We prove their required finite, arbitrary-flow versions below, including the line and conic cases, and combine them with explicit height and probability bounds. The additional standard arithmetic inputs in the finite framework are the arbitrary-intersection and projection height inequalities of \cite[Corollary 2.11 and Lemma 2.6]{ub:KPS2001}. Their precise numerical forms and treatment of singular or positive-dimensional fibers are given below. All new problem-specific arguments are included. \section{A stronger finite incidence normalization} \label{ub:opt:front-section} We now replace the expansion exponent and the matching-label constant used in the finite framework. The change is quantitative: the cutting and the elementary incidence theorem remain the same. The estimates in this section are proved with the new parameters, so their use does not rely on substituting numbers into the old regularization statement. Let $r_{\mathrm{ht}}\ge17$ and define \begin{equation}\label{ub:opt:front-parameters} \rho=\frac{r_{\mathrm{ht}}+16}{6r_{\mathrm{ht}}+98},\qquad \gamma=\frac12+\rho,\qquad R_{\rm rect}=128(r_{\mathrm{ht}}+17). \end{equation} Thus \[ \frac{33}{200}\le\rho<\frac16,\qquad \gamma\ge\frac{133}{200},\qquad \frac23-\gamma=\frac1{18r_{\mathrm{ht}}+294},\qquad R_{\rm rect}\left(\frac23-\gamma\right)>7. \] Fix $01$. Let $h,j$ be the corresponding complementary marginal masses and $J$ their joint edge probability. Lower marginals imply $h\ge\gamma u$, $j\ge\gamma v$, so \begin{equation}\label{ub:opt:front-joint-complement} J\le\frac{h+j}{2\gamma}\bigl((uv)^\rho+\xi\bigr). \end{equation} If $s\le1/16$, then \eqref{ub:opt:front-complement} gives $\sqrt{uv}<19/100$. Hence \[ (uv)^\rho+\xi <(19/100)^{33/100}+2^{-20} <93/160+2^{-20}<931/1600\le7\gamma/8. \] For clarity the only nonrational comparison here is equivalent to the integer inequality $19^{33}160^{100}<100^{33}93^{100}$. It follows that $J\le7(h+j)/16$. The edges with exactly one endpoint carrying the selected label are mismatches, so $s\ge h+j-2J\ge(h+j)/8$. This proves the first assertion. For $s\le1/128$ we may use the weaker hypotheses $\rho\ge4/25$ and $\xi\le2^{-14}$ in \eqref{ub:opt:front-complement}, to obtain \[ \sqrt{uv}\le\frac{3225}{131072}<\frac1{40}. \] Therefore \[ (uv)^\rho+\xi <40^{-8/25}+2^{-14}<5/16+2^{-14}<33/100\le\gamma/2. \] The power comparison is equivalent to $5^{33}>2^{76}$. Now \eqref{ub:opt:front-joint-complement} gives $J\le(h+j)/4$, and $s\ge(h+j)/2$. This proves the second assertion. Finally, for $s\le1/16$ two independent point labels disagree with probability at most $2h\le16s$, using the first assertion. For $s>1/16$, use the bound $1\le16s$. Centers are treated identically. \end{proof} \begin{lemma}[Bit variation with constant sixteen]\label{ub:opt:front-bits} Under the hypotheses of Lemma~\ref{ub:opt:front-labels}, assign a bit to each edge. If $V_P,V_Q$ are the disagreement probabilities in independent two-edge stars based at a point and a center, and $f$ is the smaller global bit probability, then \[ f\le16(V_P+V_Q). \] \end{lemma} \begin{proof} Give each vertex its conditional majority-bit label, breaking ties by a fixed rule. If its conditional bit probability is $\pi$, disagreement with this label has probability $\min(\pi,1-\pi)\le2\pi(1-\pi)$. The probability of different endpoint labels on an edge is thus at most $v:=V_P+V_Q$. If $v\le1/16$, the first assertion of Lemma~\ref{ub:opt:front-labels} gives a common label with total marginal exceptions at most $8v$. The edge bit differs from that label with probability at most $V_P+8v\le9v$, which bounds $f$. If $v>1/16$, $f\le1/2\le8v$. Both estimates imply the stated bound. \end{proof} \begin{remark}\label{ub:opt:front-transfer} Lemma~\ref{ub:ht:transfer} now applies with $C_0=16$. The rare-level exception construction still uses the stronger local conclusion with constant two, at its explicitly smaller threshold $1/128$. For the later extraction, the positive degree exponents remain $1-3h_0$ and $a-3h_0$, and $|E|d_{Q,\min}\ge t^{2+3a-5h_0}$. The popular-circle deletion proof has respective error bounds \[ O(t^{-3/5+8a+4\varepsilon+3h_0}),\quad O(t^{-9/5+3a+2\varepsilon+3h_0}),\quad 2t^{-a+3h_0}, \] all tending to zero. These are direct substitutions in the already proved counting argument, so it applies uniformly with the optimized parameters. \end{remark} \section{Uniform prediction and linear support transfer}\label{ub:opt:uniform-section} \subsection{One history for all bounded comparison vectors} The following strengthening makes the order of selection explicit. The history is chosen before any of the comparison vectors used later in the template argument. \begin{lemma}[Uniform prediction of bounded vectors]\label{ub:opt:uniform-prediction} Use the finite incidence law $\omega$, its marginals $\mu,\kappa$, the wedge endpoint law $\pi$, and $D=D_{\rm KL}(\pi\Vert\mu\otimes\mu)$. A random test $T$, independent of the vertices, specifies a labelling $c_T:P\to[M]$ and a vector map $f_T:[M]\to\mathbb R^d$ satisfying $\|f_T(c)\|\le B$. Let $H\ge1$ be an integer. There is a fixed strict prefix of fewer than $H$ independent tests, giving a history map $\mathsf H$ with at most $M^{H-1}$ values, such that, on putting \[ U_T(p)=f_T(c_T(p))- \sum_x\mu_{\mathsf H(p)}(x)f_T(c_T(x)), \] one has for a fresh independent test \begin{equation}\label{ub:opt:uniform-energy} \mathbb E_T\mathbb E_{q\sim\kappa} \left\|\mathbb E_{p\mid q}U_T(p)\right\|^2 \le8B^2\sqrt{\frac D{2H}}. \end{equation} The same history consequently satisfies, simultaneously for every measurable family $g_T:Q\to\mathbb R^d$ with $\|g_T(q)\|\le B'$, \begin{equation}\label{ub:opt:uniform-comparison} \mathbb E_T\left| \mathbb E_\omega\langle U_T(p),g_T(q)\rangle\right| \le4BB'(D/H)^{1/4}. \end{equation} In particular, $g$ may be defined after $\mathsf H$ has been fixed. \end{lemma} \begin{proof} Temporarily freeze independent tests $T_1,\ldots,T_H$. Let $\mathsf H_j$ be the history of the first $j-1$ tests. The image of $\pi$ under the two full sequences of labels has relative entropy at most $D$ with respect to the corresponding image of $\mu\otimes\mu$. Given two prefix values $(h,h')$, the reference law of the next two labels is $(c_{T_j})_*\mu_h\otimes(c_{T_j})_*\mu_{h'}$. The entropy chain rule and Pinsker's inequality therefore give \[ \frac1H\sum_{j=1}^H \mathbb E_\pi\operatorname{TV}(A_{j,h,h'},B_{j,h,h'}) \le\sqrt{\frac D{2H}}, \] where $A$ denotes the actual conditional law of the next labels, $B$ the displayed reference law, and $(h,h')$ the actual random prefixes. This is exactly the calculation for these sequences of labels in Lemma~\ref{ub:ip:prediction}; it uses no comparison vector. Let $U_j$ be the residual defined above with $T_j,\mathsf H_j$. Its norm is at most $2B$. Under the conditional reference product, the two residuals have zero means and are independent. Their inner product has absolute value at most $4B^2$, so the bounded-expectation estimate for total variation yields \[ \frac1H\sum_{j=1}^H \mathbb E_\pi\langle U_j(p),U_j(r)\rangle \le8B^2\sqrt{\frac D{2H}}. \] Conditional independence of the two point endpoints of a wedge gives the identity \[ \mathbb E_\pi\langle U_j(p),U_j(r)\rangle =\mathbb E_{q\sim\kappa} \left\|\mathbb E_{p\mid q}U_j(p)\right\|^2\ge0. \] Average over the tests and over a uniformly chosen index $j$. Because the residual uses a strict prefix, $T_j$ is independent of that prefix and may be replaced by a fresh test of the same law. One can therefore fix a prefix for which the fresh-test squared norm satisfies \eqref{ub:opt:uniform-energy}. This selects the history without reference to $g$. For any subsequently specified $g$, conditional expectation and Cauchy--Schwarz give \[ \left|\mathbb E_\omega\langle U_T(p),g_T(q)\rangle\right| \le B'\left( \mathbb E_\kappa\left\|\mathbb E_{p\mid q}U_T(p)\right\|^2 \right)^{1/2}. \] Averaging over $T$ and using \eqref{ub:opt:uniform-energy} proves \eqref{ub:opt:uniform-comparison}, since $\sqrt8\,2^{-1/4}<4$. The inequality holds for every $g$ with this one fixed history, which proves the asserted simultaneous conclusion. \end{proof} \begin{lemma}[Linear restricted-support transfer]\label{ub:opt:linear-support} Consider any finite-measure family of coordinate-state levels, of measure $W>0$, sampled with its normalized measure. Include in each test two independent random maps $h_1,h_2$ from the coordinate-label sets to $[L]$, assigning each attained label an independent uniform value as in Corollary~\ref{ub:ip:crossprediction}. Use the label $(h_1(A),h_2(B))$ and the vector \[ f(a,b)=(e_a,e_b,e_{(a,b)}),\qquad \|f(a,b)\|=\sqrt3. \] Fix the single history supplied by Lemma~\ref{ub:opt:uniform-prediction} for this vector-valued map and test law. Let $\nu_h$ be the true state law under $\mu_h$ at the current level, and put $\widetilde\nu_q=\mathbb E_{p\mid q}\nu_{\mathsf H(p)}$. Conditional on $q$, sample a template $\theta$ independently of the original point $p$. At each level let $G=G(q,\theta)$ be an event and let $\mathcal U(q,\theta)$ be a selected row, a selected column, or a set of at most two true states. These selections may depend on the fixed history and the current level, but not on the new random maps. Suppose that on $G$ \[ \widetilde\nu_q(\mathcal U(q,\theta))\ge1-3\eta. \] Then the integral over this family of levels satisfies \begin{equation}\label{ub:opt:linear-support-bound} \int\Pr(G,\ y_p\notin \mathcal U(q,\theta)) \le W\left(3\eta+\frac4L+4\sqrt6(D/H)^{1/4}\right). \end{equation} The same fixed history simultaneously gives the strict-cross prediction bound and this support-transfer bound for every such selected support. \end{lemma} \begin{proof} At a fixed fresh test, membership after relabelling in $U$ has a representing vector $s(q,\theta)$ in the Euclidean space of these vectors. For a row or column use its single basis vector in the corresponding block. For at most two true states use the sum of the distinct basis vectors of their relabelled pairs in the third block. Counting each distinct relabelled pair only once makes this exactly an indicator and gives $\|s(q,\theta)\|\le\sqrt2$. Set \[ g_T(q)=\mathbb E_{\theta\mid q} [\mathbf1_G s(q,\theta)]. \] Its norm is at most $\sqrt2$. This comparison vector may be defined after the history has been chosen, by Lemma~\ref{ub:opt:uniform-prediction}. Because $\theta$ and $p$ are conditionally independent given $q$, its inner product with the residual is the difference between the actual membership after relabelling probability restricted to $G$ and the corresponding probability for $\widehat p\sim\mu_{\mathsf H(p)}$. Thus their difference, averaged over normalized levels and new random maps, has absolute value at most $4\sqrt6(D/H)^{1/4}$. For a state outside a selected row or column, false membership after relabelling has probability at most $1/L$. For a state outside a selected set of at most two states, each of the two possible matches requires at least one collision between unequal coordinate labels. Its total probability is at most $2/L$. The selected support and all sampled vertices are independent of the new random maps conditional on the level and fixed history, so these estimates remain valid for their selections here. Actual and predicted probabilities after relabelling can each be replaced by true probabilities at cost at most $2/L$. For the conditionally resampled point, averaging first over the original neighbor of $q$ gives exactly $\widetilde\nu_q$. Hence \[ \Pr(G,\,y_{\widehat p}\notin \mathcal U(q,\theta)) =\mathbb E_{q,\theta} [\mathbf1_G\widetilde\nu_q(\mathcal U(q,\theta)^c)]\le3\eta. \] Combining the three bounds and multiplying by $W$ proves \eqref{ub:opt:linear-support-bound}. For completeness, a strict cross of a relabelled state $(a,b)$ has comparison vector $(e_a,e_b,-2e_{(a,b)})$, of norm $\sqrt6$. Its inner product with the same vector-valued map is precisely the strict-cross indicator. Consequently the uniform comparison lemma gives that estimate with the same history; the usual error from random relabelling is unchanged. No separate history is selected for support transfer, and no conditional point law after conditioning on a selected support has been assumed. \end{proof} \section{Sharper joint-exception and averaged-template estimates} \label{ub:opt:probability} The estimates in this section replace the initial parameter choices of the finite framework. All probability laws retain their earlier definitions; only the explicitly specified thresholds change. \begin{lemma}[An $L^2$ consequence of marginal expansion]\label{ub:opt:hardy} Suppose, for some $0<\rho<1/2$ and $\gamma=1/2+\rho$, that \[ \omega(F,G)\le2\sqrt{\mu(F)\kappa(G)} ((\mu(F)\kappa(G))^\rho+\xi). \] For a point set $F$ of marginal mass $h$ put $g(q)=\Pr(p\in F\mid q)$, and let $\kappa_{\min}$ be the smallest positive center mass. Then \[ \E_\kappa g^2\le\frac4\rho h^{1+2\rho} +8\xi^2h(1+\log(1/\kappa_{\min})). \] \end{lemma} \begin{proof} Order the centers so that $g_1\ge g_2\ge\cdots$, and write $j_r=\sum_{i\le r}\kappa_i$. Apply expansion to the first $r$ centers: \[ g_r\le2h^\gamma j_r^{\gamma-1}+2\xi\sqrt h\,j_r^{-1/2}. \] Squaring and using $(u+v)^2\le2u^2+2v^2$ bounds the second moment by \[ 8h^{2\gamma}\sum_r\kappa_rj_r^{2\gamma-2} +8\xi^2h\sum_r\kappa_r/j_r. \] The first sum is a right-endpoint Riemann sum for a decreasing integrable function, and is at most $\int_0^1x^{2\gamma-2}\,dx=1/(2\rho)$. The first term of the second sum is one, and its remaining terms total at most $\int_{\kappa_1}^1dx/x\le\log(1/\kappa_{\min})$. These facts prove the assertion, including zero-mass sets by continuity. \end{proof} In our incidence piece $\kappa_{\min}\ge\gamma/|Q|$ and $|Q|\le t^{2+a}$. Thus $\xi^2(1+\log(1/\kappa_{\min}))\to0$ uniformly as $t\to\infty$. This is a consequence of the explicit negative power of $t$ in $\xi$; it introduces no loss in the final exponent. \begin{lemma}[Few-miss sampling]\label{ub:opt:few-misses} Let $X_1,\dots,X_k$ be independent label pairs with law $\nu$. If $0<\eta<1$ and $k\eta\ge8$, then \[ \Pr\left(\exists y:\nu(R_y)<1-\eta,\quad \#\{j:X_j\notin R_y\}\le k\eta/8\right) \le2(2k+1)^2 2^{-k\eta/16}. \] \end{lemma} \begin{proof} On the event in question choose a witness using only the first sample. Draw a second independent sample of size $k$. Its number of misses has mean $k\beta$ and variance at most $k\beta$, with $\beta>\eta$. Chebyshev's inequality bounds the probability of fewer than $k\eta/2$ misses by $4/(k\beta)\le1/2$. It suffices to bound twice the probability of a cross with at most $k\eta/8$ misses in the first half and at least $k\eta/2$ in the second half. Condition on the $2k$ indexed observations and split their indices into uniform random halves. There are at most $(2k+1)^2$ strict-cross patterns. For one pattern let $m$ be its total number of misses and $X$ its misses in the first half. The event requires $m\ge k\eta/2$ and $X\le m/4$. Any fixed $j$ indices all lie in the second half with probability $(k)_j/(2k)_j\le2^{-j}$, interpreted as zero for $j>k$. Expanding the product defining $2^{m-X}$ consequently gives $\E2^{-X}\le(3/4)^m$. Therefore \[ \Pr(X\le m/4)\le[2^{1/4}(3/4)]^m \le2^{-m/8}\le2^{-k\eta/16}. \] The middle inequality follows from $4(3/4)^8=26244/65536<1/2$. The union bound proves the lemma. \end{proof} Replace the earlier candidate set by an averaged template: \begin{equation}\label{ub:opt:candidates} \mathcal Y_\theta= \left\{y:\frac1k\sum_{j=1}^k\nu_{h_j}(R_y)\ge1-\eta/32\right\}. \end{equation} Then the two required candidate bounds are \begin{align} \int_{\mathcal C}\Pr(y_q\notin\mathcal Y_\theta) &\le32\eta^{-1}\mathcal R,\label{ub:opt:actual-candidate}\\ \Pr\left(\exists y\in\mathcal Y_\theta: \widetilde\nu_q(R_y)<1-\eta\mid q\right) &\le4(2k+1)^2 2^{-k\eta/16}.\label{ub:opt:mixture-candidate} \end{align} For the first, apply Markov to the average template miss probability. Its conditional expectation is $1-\widetilde\nu_q(R_{y_q})$. For the second, select a violating candidate from the template and draw one independent synthetic state from each $\nu_{h_j}$. Conditional on the template, the expected number of misses is at most $k\eta/32$. With probability at least $3/4$, hence at least $1/2$, it is at most $k\eta/8$. Before conditioning on the template, the synthetic states are independent with common law $\widetilde\nu_q$. Lemma~\ref{ub:opt:few-misses} and a factor two prove \eqref{ub:opt:mixture-candidate}. \section{The improved explicit parameter chain}\label{ub:opt:parameters} Let $K\ge128$ be an integer for which the finite algebraic obstruction holds. Define \begin{equation}\label{ub:opt:parameter-table} \begin{gathered} r_{\mathrm{ht}}=49+\lceil\log_2K\rceil,\quad q=2^{-r_{\mathrm{ht}}},\quad \alpha=2^{-(3r_{\mathrm{ht}}+49)},\quad \rho=\frac{r_{\mathrm{ht}}+16}{6r_{\mathrm{ht}}+98},\quad\gamma=\frac12+\rho,\\ R_{\rm rect}=128(r_{\mathrm{ht}}+17),\quad S_0=2^{64}q^{-5},\quad a=2^{-2^{15}S_0},\quad b=\frac{a}{10^6R_{\rm rect}},\quad\varepsilon=b/10,\\ k=q^{-8},\quad L=q^{-32},\quad H=\lceil2a q^{-128}\log t\rceil,\quad \eta=2^{-53}q^4,\quad \zeta=2^{-56}q^4. \end{gathered} \end{equation} All parameter formulas are fixed before $t$ tends to infinity; the history length $H$ depends on $t$ through the displayed formula. We use the strengthened regularization and matching-label estimates proved in the preceding optimization section: their global matching constant is $C_0=16$; their local constant is still two when mismatch is at most $1/128$. In particular $D\le2a\log t$, and all earlier positive degree exponents remain positive with the displayed $R_{\rm rect}$ and $b$. \subsection{Heights and histories} We retain the bounded-height experiment, but tighten its geometric tolerances. With $C_0=16$, its three height means satisfy \[ A_0\le32S_0+66,\qquad B_0\le34S_0+68,\qquad C'_0\le2S_0. \] Since $S_0\ge336$, the total height of its sixteen quantities is at most $2A_0+3B_0+3C'_0\le173S_0$; the total height of its four nonzero differences is at most $2A_0+3C'_0+4\log2\le71S_0$. Choose \[ \log B=300S_0,\quad\log(1/r)=220S_0,\quad \tau=r^4/(256B^5). \] The two geometric failure probabilities total at most $173/300+71/220<0.9$. The three acceptance failures total at most $3/256$, and the collision probability tends to zero; their sum is therefore strictly less than one. The determinant contradiction is unchanged. The label count satisfies \[ \log M\le24\log B+16\log(1/r)+31\le10721S_0, \qquad\log_2M<15500S_0. \] The logarithm to base two of the history compatibility expression $2^{42}aM^2\log M$ is at most \[ 42-32768S_0+31000S_0+\log_2(15500S_0)<-1. \] Thus the full bounded-scale proof applies and yields $S>S_0$. The large-scale histories obey \[ (D/H)^{1/4}\le q^{32},\qquad L^{-1}=q^{32},\qquad \frac{\log L^{2Hk}}{\log t} \le160a q^{-136}\log(1/q)<1/2. \] For the last inequality, take base-two logarithms and use $-\log_2a=2^{79+5r_{\mathrm{ht}}}$, which exceeds $136r_{\mathrm{ht}}+\log_2(160r_{\mathrm{ht}}\log2)+1$ for $r_{\mathrm{ht}}\ge56$. All the collision probabilities used in the score arguments can therefore be made at most $q^{100}$ by increasing the starting scale. The history is selected by the energy criterion in Lemma~\ref{ub:opt:uniform-prediction}, so the same history also permits the linear support estimate of Lemma~\ref{ub:opt:linear-support}. \subsection{Joint exceptions and bad edges} Declare rare levels good when $f<\alpha$ and $m\le\alpha$. Then \[ W_c\le2^{49}q^{-3}S,\qquad W_b\le2^{51}q^{-3},\qquad \sigma:=\alpha^{2\rho}=2^{-16}q. \] The local matching lemma gives $h_P+h_Q\le4\alpha$ and $\mathcal A+\mathcal B_*\le3S$. Taking $t$ large enough that $\xi\le\sigma$ and $8\xi^2(1+\log(1/\kappa_{\min}))\le\sigma$, expansion yields \[ J_e\le9\sigma S\le qS/4096. \] Lemma~\ref{ub:opt:hardy}, $\rho\ge1/7$, and $4^{2\rho}<2$ give $\E g^2\le57\sigma h$. After integration, \begin{equation}\label{ub:opt:joint-budgets} J_w\le171\sigma S\le qS/256. \end{equation} The history prediction error satisfies \begin{equation}\label{ub:opt:prediction-budget} \mathcal R\le4+22q^{32}W_c \le4+22\cdot2^{49}q^{29}S. \end{equation} Use \eqref{ub:opt:candidates} with the displayed $\eta$. The actual-candidate failure divided by $S$ is at most \[ 32\eta^{-1}\mathcal R/S \le q/16+22\cdot2^{107}q^{25}\le q/16+q/4096. \] The integrated mixture failure is also at most $qS/4096$: its ratio to $S$ is at most $36\cdot2^{49}q^{-19}2^{-2^{-64}q^{-4}}$, and $2^{4r_{\mathrm{ht}}-64}\ge20r_{\mathrm{ht}}+67$ for $r_{\mathrm{ht}}\ge56$. For every good mixture, the containing row, column, or at most two states has predicted mass at least $1-3\eta$. Lemma~\ref{ub:opt:linear-support} bounds the integrated support failure by \[ 3\eta W_c+(4/L+4\sqrt6(D/H)^{1/4})W_c \le\frac3{16}qS+14\cdot2^{49}q^{29}S \le3qS/16+qS/4096. \] Finally, inconsistency contributes at most $4\le qS/4096$. Thus the combined bad-edge budget satisfies \begin{equation}\label{ub:opt:bad-budget} B_{\rm bad}\le(q/4+q/1024)S<17qS/64. \end{equation} \subsection{Scores and positive rare mass} The upper score remains $D_i\le v_*C_i+qS$: its error is bounded by $2\log2+2q^{100}(W_c+3S)$. The rare lower score is now \begin{equation}\label{ub:opt:rare-budget} D_i^{nc}\ge2\E H_x^*-2\E H_y^*-qS/512, \end{equation} since its error is at most $4J_e+24+W_b$, and $24+W_b\le qS/1024$. The two common-score inequalities consequently give \[ g_*(C_Q+C_{P^\theta}) \le(2+1/256)qS+16B_{\rm bad}<\frac{13}{2}qS. \] The exact value $g_*=0.00001245$ satisfies $2^{19}g_*=6.5273856>13/2$. Therefore, as before, \begin{equation}\label{ub:opt:common-budget} C_Q+C_{P^\theta}\le2^{19}qS,\qquad \mathcal B_*\le\mathcal A+2^{20}qS. \end{equation} The two-prediction proof still gives $\mathcal T\le2\mathcal A+2qS$, because $W_b+\log2\le qS$ and the diagonal contribution is at most $qS$. For the ordinary-point classification replace its threshold by the displayed $\zeta$. Then \[ 4+\mathcal R/\zeta\le(q/64+q/4096)S,\qquad \zeta W_c\le qS/128. \] The integrated classification penalty is at most $97qS/256R_0S}]\le2^{13}q^4S$. The wedge approximation error is bounded by \[ C_Q+W_b+8J_e+2J_w+32\le2^{20}qS. \] In the height-band selection the weighted error deletion consequently costs at most $2^{22}KqS\le c_0S/4$, because $q\le2^{-49}/K$. The tail deletion costs at most $2^{15}q^4S\le c_0S/4$, and $c_0S_0\ge K$. All other steps of the organized graph extraction and finite polynomial selection are unchanged. Thus, once the finite algebraic obstruction has been established with this $K$, the improved exponent difference is \begin{equation}\label{ub:opt:delta-final} \boxed{\displaystyle \delta=\frac{2^{-2^{79+5r_{\mathrm{ht}}}}}{3\cdot10^7R_{\rm rect}},\qquad r_{\mathrm{ht}}=49+\lceil\log_2K\rceil,\quad R_{\rm rect}=128(r_{\mathrm{ht}}+17).} \end{equation} The finite algebraic obstruction below supplies the value of $K$; together with the preceding incidence normalization, it fixes the exponent difference in this formula. \section{An eight-by-four finite algebraic obstruction} \label{ub:optalg:section} This section replaces the twenty-by-five algebraic grid and its Macaulay height estimate. The eight-by-four normalization has a public antecedent in the manuscript \emph{A smaller generic grid in the unit-distance argument}~\cite{ub:GridRefinement2026}; the argument below is given independently because its application here permits arbitrary flow variables, not only derivatives. Its new quantitative use is through the finite, specialized height estimate proved below. There are eight positions $p_i$ and four positions $r_j$. Write $L_{ij}=(p_i-r_j)\cdot(p_i-r_j)$. Define the constructible set $Z_{8,4}$ by the existence of $R_{ij},H_i,U_j,A_i,A'_j$ satisfying \begin{equation}\label{ub:optalg:bad} R_{ij}^2=L_{ij}(L_{ij}-4),\qquad 2(p_i-r_j)\cdot(H_i-U_j)=R_{ij}(A_i-A'_j),\qquad A_i-A'_j\ne0. \end{equation} The positions comprise $24$ coordinates. This set has a fixed finite Boolean description by integer polynomial tests: adjoin one inverse to $\prod_{i,j}(A_i-A'_j)$ and use effective ideal membership and matrix ranks as in the proof of Lemma~\ref{ub:alg:boolean}. More explicitly, there are $69$ auxiliary variables, $65$ equations, and degree at most $33$ in the auxiliary variables, so the multiplier bound $66^{2^{69}}$ suffices. The size of this description affects the eventual starting scale only. \begin{proposition}\label{ub:optalg:avoid} Let $V_1,V_2$ be either the plane or irreducible real-compatible curves with nonconstant $z$- and $w$-projections. If they are not the same real unit circle, a jointly generic point of $V_1^8\times V_2^4$ lies outside $Z_{8,4}$. \end{proposition} \begin{proof} Suppose a solution of \eqref{ub:optalg:bad} exists. Let $J(x,y)=(-y,x)$. First replace all $H_i,U_j$ by $H_i-U_1,U_j-U_1$, and all $A_i,A'_j$ by $A_i-A'_1,A'_j-A'_1$. Thus $U_1=0$ and $A'_1=0$. These operations preserve all the equations and nonzero differences. The transformation \[ H_i\longmapsto H_i+\tau J(p_i-r_1),\qquad U_j\longmapsto U_j+\tau J(r_j-r_1) \] also preserves the equations: the added difference is $\tau J(p_i-r_j)$, whose scalar product with $p_i-r_j$ is zero. It preserves $U_1=0$. Since $r_2\ne r_1$, one coordinate of $J(r_2-r_1)$ is nonzero. Choose $\tau$ so that the corresponding coordinate of $U_2$ is zero. The choice of this coordinate is one of two finite possibilities. Of the original twelve right-hand flow coordinates, only eight can now be nonzero. If all are zero, they generate no transcendence. Otherwise multiply every $H_i,U_j,A_i,A'_j$ by the reciprocal of one nonzero remaining right-hand coordinate. All of \eqref{ub:optalg:bad} still holds, and that coordinate is now $1$. In either case the normalized right-hand flow data have transcendence degree at most seven over the field generated by the right-hand positions. Let $B$ be the algebraically closed field defining the loci and let $K$ be the algebraic closure after adjoining all four right-hand positions and the normalized right-hand flow data. For $d=\dim V_1$, joint genericity therefore gives \[ \operatorname{trdeg}_K K(p_1,\ldots,p_8)\ge8d-7>8(d-1). \] Some $p_i$ remains generic on $V_1$ over $K$. The private-valuation proof in Proposition~\ref{ub:alg:generic-avoidance} applies verbatim to these four anchors: it is proved one anchor at a time and is independent of the number of other anchors. It gives independent square classes for all four $D_j=L(p_i,r_j)(L(p_i,r_j)-4)$ over $K(V_1)$. In the noncollinear case the first three equations determine $H_i,A_i$ over the field with the first three radicals; flipping the fourth gives $A_i=A'_4$, a contradiction. When both loci are the same line, the first two equations determine the two effective unknowns, and flipping the third radical gives the same contradiction. The rank check and all private-valuation cases are exactly those already proved there. Hence four anchors suffice, and \eqref{ub:optalg:bad} is impossible. \end{proof} We use the following standard consequence of the arithmetic B\'ezout and projection inequalities of Krick--Pardo--Sombra. Here the coefficient height $H$ includes the constant $1$ in the joint coefficient vector. If a variety $V\subset\mathbb A^n_{\overline{\mathbb Q}}$ is defined by polynomials of degree at most $d$ and a coordinate projection $\pi(V)$ is finite, then \begin{equation}\label{ub:optalg:finiteimage} h(\alpha)\le d^n\left(\frac ndH+(5n+3)\log(n+1)\right) \qquad(\alpha\in\pi(V)). \end{equation} Indeed, Corollary~2.11 of \cite{ub:KPS2001} bounds its arithmetic height by $nd^{n-1}H+2nd^n\log(n+1)$, with $\deg V\le d^n$. Lemma~2.6 of that paper adds at most $3(n+1)\log(n+1)d^n$ on projection. Arithmetic height is nonnegative and additive over components; for a point its value is at least the ordinary absolute logarithmic height. This proves \eqref{ub:optalg:finiteimage}. These quoted results explicitly permit non-equidimensional and singular varieties, arbitrary intersections, and projections with positive-dimensional fibers. No complete-intersection or generic-coefficient assumption is being made. \begin{theorem}\label{ub:optalg:height} Suppose an algebraic real-position grid lies outside $Z_{8,4}$ and its nonzero algebraic ratios satisfy $L_{ij}=2-\lambda_{ij}-\lambda_{ij}^{-1}$. For $i\ge2,j\ge2$ put $\chi_{ij}=\lambda_{ij}\lambda_{11}/(\lambda_{i1}\lambda_{1j})$ and $\Sigma=\sum_{i=2}^8\sum_{j=2}^4h(\chi_{ij})$. Some edge $e$ satisfies \begin{equation}\label{ub:optalg:affineheight} h(\lambda_e)\le 4^{65}\left[\frac{65}{4}(2\Sigma+\log8)+328\log66\right] <1.4\cdot10^{39}(32.5\Sigma+1500). \end{equation} \end{theorem} \begin{proof} Translate and rotate the actual real grid so that $r_1=(0,0)$ and $r_2=(s,0)$ with $s\ne0$. This is possible because the anchors are distinct real points. The transformation has algebraic coefficients. It preserves distances, the ratios, and membership in $Z_{8,4}$, since flows transform by the same orthogonal linear map. There are now $21$ position variables: sixteen for the left positions, one for $r_2$, and four for $r_3,r_4$. Set $\chi_{i1}=\chi_{1j}=1$, $b_1=1$, and introduce eleven variables $a_1,\ldots,a_8,b_2,b_3,b_4$. Put \[ \Lambda_{ij}=\chi_{ij}a_ib_j,\qquad F_{ij}=\Lambda_{ij}^2+(L_{ij}-2)\Lambda_{ij}+1. \] These are $32$ polynomials of degree at most four in $32$ variables $X$. The actual point is obtained by taking $a_i=\lambda_{i1}$ and $b_j=\lambda_{1j}/\lambda_{11}$. Every solution has all $a_i,b_j$ and $\Lambda_{ij}$ nonzero, as follows from the constant term of $F_{ij}$. If a vector in the literal Jacobian kernel of all $F_{ij}$ had every $d\Lambda_{ij}$ nonzero, logarithmic differentiation of the factorization would give $d\Lambda_{ij}/\Lambda_{ij}=A_i-A'_j$. Differentiating the distance identities would then give \eqref{ub:optalg:bad}, with $R_{ij}=\Lambda_{ij}^{-1}-\Lambda_{ij}$ and the position components of the vector as flows. This is impossible at the actual position tuple. The finite-union-of-subspaces argument therefore gives an edge $e$ such that $d\Lambda_e$ annihilates the whole Jacobian kernel. Its gradient is in the row span of the $32$ gradients $\nabla F_{ij}$. Adjoin $32$ multipliers $c_{ij}$ and a coordinate $y$. Define $C_e$ by \[ F_{ij}(X)=0,\qquad \frac{\partial\Lambda_e}{\partial X_v} =\sum_{i,j}c_{ij}\frac{\partial F_{ij}}{\partial X_v} \quad(1\le v\le32),\qquad y-\Lambda_e(X)=0. \] This system has $65$ variables and $65$ equations, all of degree at most four; the actual point lifts to it. It is defined over the number field of the actual defects. Its combined coefficient height, together with $1$, is at most $2\Sigma+\log8$: all coefficients are bounded integers times $1,\chi_{ij},\chi_{ij}^2$, with integer absolute value at most four. The factor eight is a harmless uniform allowance. On the function field of every irreducible component of $C_e$, for every derivation over the algebraically closed coefficient field, the defining equations imply \[ Dy=D\Lambda_e=\sum_{i,j}c_{ij}DF_{ij}=0. \] Characteristic zero implies that $y$ is constant on that component: if it were transcendental, extend it to a transcendence basis and take the derivation sending it to $1$. Consequently the projection of $C_e$ to $y$ is finite. This applies without any smoothness assumption. Apply \eqref{ub:optalg:finiteimage} with $n=65,d=4$. It gives the first inequality in \eqref{ub:optalg:affineheight}. Finally $4^{65}<1.4\cdot10^{39}$, $\log8<2.1$, and $\log66<4.2$ give the displayed strict bound. \end{proof} \begin{corollary}\label{ub:optalg:contradiction} Take $K=10^{43}$. No grid outside $Z_{8,4}$ can satisfy \[ J\ge K,\qquad h(\lambda_{ij})\ge J/3,\qquad h\!\left(\frac{\lambda_{ij}\lambda_{i'j'}} {\lambda_{ij'}\lambda_{i'j}}\right)\le J/K \] on every complete rectangle, along with the distance identities. \end{corollary} \begin{proof} There are $21$ normalized defects, hence $\Sigma\le21J/K$. Theorem~\ref{ub:optalg:height} gives for some edge \[ \frac{h(\lambda_e)}J <\frac{1.4\cdot10^{39}}K(32.5\cdot21+1500) =0.30555<\frac13, \] a contradiction. One may instead take the convenient power of two $K=2^{143}>10^{43}$. \end{proof} The finite polynomial-pattern selection proof requires no new geometric argument. One may apply its existing twenty-by-five statement to the tests for $Z_{8,4}$, interpreted as polynomials in the first eight left and first four right blocks. Projection of a jointly generic twenty-by-five product tuple is jointly generic on the required eight-by-four product. If the two loci are the same unit circle, the popular-circle exclusion applies exactly as before. Thus the new numerical contradiction replaces the old one in the finite obstruction theorem without altering its fixed exponent $10^{-8}$. \section{A six-by-five obstruction for arbitrary flows} \label{ub:opt6:section} The joint linear-system and spanning-tree approach in this section has a public antecedent in PR~3 of \emph{unit-distance-grid-refinement}~\cite{ub:GridJoint2026}. We prove the needed geometric statements here, including all line cases for arbitrary flow variables. In particular the argument does not require the flows to be values of a derivation, and it does not import a rigidity theorem. This distinction is needed for a finite constructible bad set. Take six positions $p_i$ and five positions $r_j$. Let $Z_{6,5}$ be the set where there exist scalars $R_{ij},A_i,A'_j$ and vectors $H_i,U_j$ such that \begin{equation}\label{ub:opt6:bad} \begin{gathered} L_{ij}=(p_i-r_j)\cdot(p_i-r_j),\qquad R_{ij}^2=L_{ij}(L_{ij}-4),\\ 2(p_i-r_j)\cdot(H_i-U_j)=R_{ij}(A_i-A'_j),\qquad A_i-A'_j\ne0. \end{gathered} \end{equation} As before this is a fixed constructible set with a finite Boolean description by integer polynomial tests. The same inverse-product and effective ideal-membership proof applies, now with $64$ auxiliary variables, $61$ equations, and auxiliary degree at most $31$. \begin{theorem}\label{ub:opt6:avoid} Let each of $V_1,V_2$ be the plane or an irreducible real-compatible curve with nonconstant complex coordinate projections. Suppose they are not the same real unit circle. A jointly generic point of $V_1^6\times V_2^5$ lies outside $Z_{6,5}$. \end{theorem} \subsection{Independent radicals on the whole product} Let $B$ be an algebraically closed coefficient field for the two loci and define, at a jointly generic tuple, \[ F=B(p_1,\ldots,p_6,r_1,\ldots,r_5). \] All thirty elements $D_{ij}=L_{ij}(L_{ij}-4)$ have independent square classes in $F^\times/(F^\times)^2$. Indeed, fix $i$ and take the algebraic closure $K_i$ after adjoining every position except $p_i$. Joint genericity implies that $p_i$ remains generic on $V_1$ over $K_i$. The private-valuation construction already proved in Proposition~\ref{ub:alg:generic-avoidance}, applied over $K_i$, gives a valuation with value one on $D_{ij}$ and zero on the other four radicands in that row. Every radicand in another row is a nonzero element of $K_i$ and consequently has value zero. Thus every nonempty product of distinct $D_{ij}$ has odd valuation somewhere and cannot be a square. By Lemma~\ref{ub:alg:kummer}, all square-free products of the $R_{ij}$ are linearly independent over $F$. \subsection{An elementary complete-bipartite rigidity calculation} For positions on a graph $G$, write $\operatorname{Rig}(G)$ for the matrix whose edge row at $pr$ has $p-r$ in the two columns of $p$, $r-p$ in the two columns of $r$, and zero elsewhere. Thus its kernel consists of solutions of $(p-r)\cdot(H-U)=0$. Its row rank is unchanged by an invertible affine change of positions $x\mapsto Mx+t$, provided the velocity vectors are changed by $H\mapsto M^{-\mathsf T}H$. This is only a rank calculation, not a claim that an arbitrary affine map preserves Euclidean distances. \begin{lemma}\label{ub:opt6:spanning-parts} If both vertex sets of a complete bipartite framework affinely span the plane, its infinitesimal motion space has dimension $3+q$, where $q$ is the dimension of the vector space of quadratic polynomials vanishing on all its vertices. \end{lemma} \begin{proof} Choose three affinely independent right vertices. For a left vertex $p$, write $c=p\cdot H$. The three edge equations are $c-r_j\cdot H=p\cdot U_j-r_j\cdot U_j$. Their coefficient matrix in the three unknowns $(H_1,H_2,c)$ is invertible. Hence there are a matrix $A$, vectors $b,c_0$, and scalar $d$ such that every left velocity has $H=Ap+b$ and $p\cdot H=c_0\cdot p+d$. In particular \[ Q(p):=p^{\mathsf T}Ap+(b-c_0)\cdot p-d=0. \] The remaining edge equations read \[ p\cdot(c_0-A^{\mathsf T}r-U)+d-r\cdot b+r\cdot U=0. \] Since the left vertices affinely span the plane, they imply $U=c_0-A^{\mathsf T}r$ and $Q(r)=0$. Conversely these formulas produce every motion whenever $Q$ vanishes at all vertices. The nine parameters in $(A,b,c_0,d)$ determine the six coefficients of $Q$ surjectively. The kernel is precisely $A=\omega J$, $b=c_0=T$, $d=0$, the three-dimensional space of rigid translations and rotations. The parameterization of motions is injective, as three left vertices determine $A,b$, the right velocities then determine $c_0$, and any one equality determines $d$. This proves the stated dimension. \end{proof} \begin{lemma}\label{ub:opt6:conics} If neither locus is a line, four generic points of $V_1$ and three generic points of $V_2$ lie on a conic exactly when the two loci are the same irreducible conic. In that case the space of quadratics vanishing on these seven points has dimension one. \end{lemma} \begin{proof} The conic condition is the vanishing of the six-by-six minors of the seven-by-six quadratic evaluation matrix, hence a polynomial condition on the product. It suffices to exhibit a tuple violating it whenever the loci are not the same conic. If one locus is the plane, first choose the three or four points on the other locus in general position. They impose three or four independent conditions on quadratics. Successive plane points can each be chosen off the zero set of a surviving nonzero quadratic, reducing the remaining dimension by one, until no quadratic remains. There are enough plane points on the prescribed side. If $V_1=V_2$ is a curve of degree at least three, the six quadratic monomials are linearly independent when restricted to that curve; successive evaluation points therefore give rank six. If $V_1\ne V_2$ are curves, choose four points on $V_1$ off $V_2$, with no three collinear. These impose four independent conditions on conics. The pencil through them has base locus exactly the four chosen points: two distinct unions of opposite joining lines already have exactly those four common points. Choose a point on $V_2$ outside this base locus. One conic of the pencil passes through it. That conic cannot contain $V_2$: a nonlinear irreducible curve contained in a conic must be the conic itself, whereas the four first points were off $V_2$. Choose a second point of $V_2$ off the surviving conic; then no conic remains. A third point does not change this. Finally, if the loci are the same irreducible conic, its defining quadratic vanishes on all points, and any independent quadratic meets it in at most four points by B\'ezout. Seven distinct points therefore leave precisely a one-dimensional quadratic space. \end{proof} \begin{lemma}\label{ub:opt6:line-ranks} At a jointly generic complete bipartite framework with at least three vertices on each side, the motion dimension is three when precisely one locus is a line, four when the two loci are different lines, and $v+1$ when both loci are the same line and $v$ is the total vertex count. \end{lemma} \begin{proof} For exactly one line, normalize it to $r_j=(t_j,0)$ and write $U_j=(u_j,v_j)$, with distinct $t_j$. The other vertices $p_i=(x_i,y_i)$ affinely span the plane. Work modulo the span of the two vectors $\mathbf1,(t_j)_j$ in the vector space of right-indexed lists, using brackets for quotient classes. The edge equations give \[ x_i[u]+y_i[v]-[tu]=0 \quad\hbox{for every }i. \] Affine spanning forces $[u]=[v]=[tu]=0$. Hence $u_j=A+Bt_j$ and $v_j=C+Dt_j$; since there are at least three distinct $t_j$, the condition $[tu]=0$ forces $B=0$. Thus the right velocities have just the three parameters $A,C,D$. Every left point is off the line and two right neighbors determine its velocity uniquely. The three rigid motions supply all these solutions, so the dimension is three. If the line is on the other side, interchange the sides. For two different lines, an invertible affine change sends them either to $p_i=(s_i,0),r_j=(t_j,1)$ or to $p_i=(s_i,0),r_j=(0,t_j)$. In the second case all generic $s_i,t_j$ are nonzero. Write the velocities as $(h_i,k_i)$ and $(u_j,v_j)$. In the parallel case set \[ F_i=s_ih_i-k_i,\quad B_i=h_i,\quad C_j=u_j,\quad G_j=t_ju_j+v_j; \] in the intersecting case set \[ F_i=s_ih_i,\quad B_i=k_i,\quad C_j=u_j,\quad G_j=t_jv_j. \] Both changes of velocity variables are invertible, and the edge system becomes $F_i+G_j-t_jB_i-s_iC_j=0$. Subtracting two rows with distinct $s_i$ shows $C_j=D-Bt_j$; comparing any other row, using two distinct $t_j$, gives \[ B_i=A+Bs_i,\quad F_i=C+Ds_i,\quad C_j=D-Bt_j,\quad G_j=At_j-C. \] These are exactly all solutions, with four free parameters. For the same line, put it on the horizontal axis. All edge differences are nonzero horizontal vectors. The horizontal velocity is common on the connected graph, and the $v$ vertical velocities are unrestricted. The dimension is $v+1$, as asserted. \end{proof} \subsection{A tree whose removal preserves rigidity rank} Let $E$ be the thirty edges of $K_{6,5}$ and put \[ T=\{p_1r_4,p_2r_4,p_3r_5,p_4r_5,p_5r_1,p_5r_2, p_6r_3,p_6r_4,p_6r_5,p_5r_5\}. \] This is a connected graph on eleven vertices with ten edges, hence a spanning tree. Its complement consists of the core $K_{4,3}$ on $p_1,\ldots,p_4$ and $r_1,r_2,r_3$, followed by four attachments: $r_5$ to $p_1,p_2$; $r_4$ to $p_3,p_4$; $p_6$ to $r_1,r_2$; and $p_5$ to $r_3,r_4$. Except when both loci are the same line, each attachment increases rigidity rank by two. The two new rows have an invertible two-by-two minor in the new vertex columns exactly when the three points are not collinear. Genericity gives this noncollinearity: a generic point cannot lie on the line through the other two unless its entire locus is that line, and the latter possibility with two generic points of the other locus forces both loci to be that same line. If neither locus is a line and they are not the same conic, the core has rank $14-3=11$ by Lemmas~\ref{ub:opt6:spanning-parts}--\ref{ub:opt6:conics}. Its four attachments give rank $19$, the maximum possible rank of the full framework. For a common irreducible conic, the core rank is $14-4=10$ and the full rank is $22-4=18$, again agreeing after the attachments. Exactly one line gives the ranks $11$ and $19$ by Lemma~\ref{ub:opt6:line-ranks}; two different lines give $10$ and $18$. Finally, for one common line both the full graph and the complement of $T$ are connected, so both rigidity ranks equal ten. In all cases \begin{equation}\label{ub:opt6:tree-rank} \operatorname{rank}\operatorname{Rig}(E\setminus T) =\operatorname{rank}\operatorname{Rig}(E). \end{equation} \subsection{The full flow system and its kernel} The thirty equations in \eqref{ub:opt6:bad} form a matrix with $33$ columns: $22$ velocity columns and eleven scalar columns. Write its edge row as $a_e+R_eb_e$, where $a_e$ is twice the rigidity row on the velocity columns, and $b_e$ is the signed incidence row on the scalar columns. Put $r=\operatorname{rank}\operatorname{Rig}(E)$. Choose ten scalar columns on which the tree rows $b_e$, $e\in T$, have nonzero determinant, and choose $r$ independent rigidity rows from $E\setminus T$ and $r$ velocity columns with nonzero determinant. The determinant of the resulting $(10+r)$-row square minor is a linear combination over $F$ of square-free products of the radicals. The coefficient of $\prod_{e\in T}R_e$ is, up to sign, the product of the two nonzero determinants just chosen. Independence of all the square-free radical products makes this minor nonzero. The full matrix therefore has rank at least $10+r$. Its kernel already contains every rigidity motion together with a common scalar value on all eleven vertices, a space of dimension $22-r+1=23-r$. The rank is consequently at most $10+r$. Equality holds, and every kernel vector has a constant scalar part. This remains true after any extension of the coefficient field, since matrix rank is unchanged. In particular no solution can have all $A_i-A'_j\ne0$. This proves Theorem~\ref{ub:opt6:avoid}. \begin{remark}[A twenty-nine-edge version]\label{ub:opt6:sparse} One may delete the fixed core edge $p_4r_3$. At a generic core, all twelve edges are symmetric under permutations within the two parts. If deleting one row decreased the rank, every core row would be a coloop; then all twelve rows would be independent. This is impossible, as the ranks just computed are $11$, $10$, or $6$. Deleting $p_4r_3$ thus leaves the core rank unchanged. The attachments and tree minor argument still apply. Consequently the generic-avoidance theorem also holds for the sparse graph $E\setminus\{p_4r_3\}$, with the nonzero flow constraints required only on its twenty-nine edges. \end{remark} \subsection{The direct finite height bound} For the full grid, translate and rotate actual algebraic real positions so that $r_1=(0,0)$ and $r_2=(s,0)$. There are $19$ position variables. Set $b_1=1$ and $\chi_{i1}=\chi_{1j}=1$, and write \[ \Lambda_{ij}=\chi_{ij}a_ib_j, \qquad F_{ij}=\Lambda_{ij}^2+(L_{ij}-2)\Lambda_{ij}+1. \] The six $a_i$ and four $b_j$ give ten further variables, for a total of $29$ variables $X$. There are thirty distance equations of degree at most four. There are twenty defects, and their sum of heights is denoted by $\Sigma$. Outside $Z_{6,5}$, the same literal-Jacobian-kernel argument as in Theorem~\ref{ub:optalg:height} gives an edge $e$ with $d\Lambda_e$ in the span of the gradients of the $F_{ij}$. Adjoin thirty multipliers and $y=\Lambda_e$. The critical-value variety has sixty variables and sixty equations of degree at most four, with coefficient height at most $2\Sigma+\log8$. Its $y$-projection is finite, by the derivation argument already given. Thus \[ h(\lambda_e)\le4^{60}\left[15(2\Sigma+\log8)+303\log61\right] <1.33\cdot10^{36}(30\Sigma+1280). \] With $\Sigma\le20J/K$ and $J\ge K$, $K=10^{40}$ is already sufficient to contradict $h(\lambda_e)\ge J/3$. For completeness the sparse version gives a better direct numerical bound without any rank-stratification argument. Delete $F_{43}$ and its defect. There remain $29$ variables, $29$ equations, and $19$ defects. Adjoining $29$ multipliers and $y$ gives $59$ variables. Of the $59$ equations, six have degree at most three, $52$ have degree at most four, and one has degree at most two. The product and reciprocal sum of this upper degree list are \[ \Delta=2\cdot3^6\cdot4^{52},\qquad \sigma=\frac{31}{2}. \] Using Corollary~2.11 of \cite{ub:KPS2001} with these individual degrees, followed by its Lemma~2.6, gives \begin{equation}\label{ub:opt6:sparseheight} h(\lambda_e)\le\Delta\left[\frac{31}{2}(2\Sigma+\log8) +298\log60\right]. \end{equation} The harmless upper degree bounds can be substituted since each term in the product-times-reciprocal sum is increasing in each degree. Here $\Sigma\le19J/K$. Since \[ 3\Delta\left(31\cdot19+\frac{31}{2}\log8+298\log60\right) <1.64\cdot10^{38}<2^{127}, \] the choice $K=2^{127}$ gives the required strict contradiction. The stronger rank-stratified bound below instead uses the full grid; this avoids any ambiguity when its distinguished edge is chosen as the base of a new ratio chart. \subsection{A rank-sensitive logarithmic-Jacobian height estimate} \label{ub:opt:rank-height} We use the arbitrary-intersection and projection estimates of Krick--Pardo--Sombra, Corollary~2.11 and Lemma~2.6, with their absolute number-field height conventions. We also record explicitly a useful individual-equation version of the intersection estimate. \begin{lemma}[Keeping the individual equation heights] \label{ub:opt:individual-heights} Let $f_1,\ldots,f_s$ be polynomials on $\mathbb A^n$, with degree bounds $d_i\ge1$ and absolute affine coefficient heights at most $h_i$. Put $D=\prod_i d_i$, and let $Z$ be their common zero set. Then \[ \deg Z\le D,\qquad h_{\rm KPS}(Z)\le D\left[h_{\rm KPS}(\mathbb A^n) +\sum_{i=1}^s\frac{h_i}{d_i}+s\log(n+1)\right]. \] Degrees and heights here include all irreducible components, regardless of their dimensions; no proper-intersection hypothesis is imposed. \end{lemma} \begin{proof} For a component $Y$ on which $f_i$ is not identically zero, the single-hypersurface arithmetic Bezout estimate used in the proof of KPS Corollary~2.11 gives \[ h_{\rm KPS}(Y\cap V(f_i)) \le d_i h_{\rm KPS}(Y) +(h_i+d_i\log(n+1))\deg Y, \qquad \deg(Y\cap V(f_i))\le d_i\deg Y. \] If $f_i$ vanishes identically on $Y$, the intersection retains $Y$. The same two inequalities still hold, since $d_i\ge1$, $h_i\ge0$, and $h_{\rm KPS}(Y)\ge0$. Summing over components and deleting any duplicates or components contained in larger ones only decreases the degree and height sums. Thus these recurrences hold for an arbitrary intermediate zero set. Iterating and dividing the height recurrence at stage $i$ by $\prod_{j\le i}d_j$ gives the assertion. \end{proof} The ambient height in this convention is the explicit Stoll number \[ s_n:=h_{\rm KPS}(\mathbb A^n) =\frac12\sum_{i=1}^n\sum_{j=1}^i\frac1j =\frac12\left((n+1)\sum_{j=1}^n\frac1j-n\right). \] It increases with $n$. We will use only $s_{29}<45$ and $s_{50}<90$. For completeness their exact rational values are \[ s_{29}=\frac{6975593267347}{155272637520}<45, \qquad s_{50}=\frac{10904958651492685640759} {121531157029254772800}<90. \] \paragraph{Coordinates and a fixed coefficient tuple.} The grid has six left vertices and five right vertices. Choose the critical edge before imposing a Euclidean gauge, and relabel its endpoints as $(1,1)$. Normalize the rectangle defects relative to it. All twenty new defects are complete rectangle defects from the hypothesis. The old and new normalized families are Laurent monomials in one another, so fixing either family fixes the other; their literal tangent conditions agree on the locus of nonzero ratios. Use ten ratio variables $a_1,\ldots,a_6,b_2,\ldots,b_5$, with $b_1=1$, so that \[ \Lambda_{ij}=\lambda_{ij}=\chi_{ij}a_i b_j, \qquad \chi_{i1}=\chi_{1j}=1. \] In particular the critical ratio is the coordinate $a_1$. An algebraic Euclidean isometry puts $r_1=(0,0)$ and $r_2=(s,0)$; the two centers are distinct real points, so this is permissible. The twelve left-point coordinates remain free. There are $N=29$ variables $X$: nineteen position coordinates and ten ratio variables. The thirty equations are \[ F_{ij}(X)=\Lambda_{ij}^2+(L_{ij}-2)\Lambda_{ij}+1=0. \] Six equations $F_{i1}$ have degree at most three and all others at most four. The gauged tangent vectors induce ungauged tangent vectors, so $da_1$ still annihilates the literal Jacobian kernel at the actual point. The actual values of all defects are fixed before any of the following varieties is formed. \paragraph{A basis containing ten coefficient-free equations.} Let $c$ be the Jacobian rank at the actual point. The six rows of $F_{i1}$ are independent, since \[ F_{i1}=a_i^2+(|p_i|^2-2)a_i+1, \qquad \nabla_{p_i}F_{i1}=2a_i p_i\ne0. \] Here $a_i\ne0$ and $p_i\ne r_1$, and the six rows have disjoint supports in the corresponding triples $(p_i,a_i)$. Under the contradictory lower-height hypothesis, no $\lambda_{ij}$ equals $1$ or $-1$, since either number has height zero. Moreover, at a solution, \[ \frac{\partial F_{1j}}{\partial b_j} =a_1(\lambda_{1j}-\lambda_{1j}^{-1})\ne0 \qquad(2\le j\le5). \] Each $b_j$ is absent from the other nine equations in the base star $\{F_{i1}\}_{i=1}^6\cup\{F_{1j}\}_{j=2}^5$. These four columns, followed by the preceding six-row independence, show that all ten star gradients are independent. Thus $10\le c\le29$. Extend them to a basis of $c$ rows, corresponding to $F_1,\ldots,F_c$. Let $\mathcal I$ be the $c-10$ selected nonstar indices and set \[ \Sigma=\sum_{(i,j)\in\mathcal I}h(\chi_{ij}). \] Only these defects occur in the selected equations, and $\Sigma\le(c-10)J/K$ under the assumed rectangle-height bounds. \paragraph{The logarithmic Jacobian.} Let $D_X$ be diagonal, with entry one in each position column and entry $a_i$ or $b_j$ in the corresponding ratio column. Divide row $(i,j)$ of the ordinary Jacobian by $\Lambda_{ij}$ and multiply its columns by $D_X$. The result $B$ is a polynomial matrix: \[ \begin{aligned} B_{ij,x}&=\frac{\partial L_{ij}}{\partial x} &&\text{in position columns},\\ B_{ij,a_i}=B_{ij,b_j}&=T_{ij}:=2\Lambda_{ij}+L_{ij}-2 &&\text{in ratio columns}. \end{aligned} \] with all other ratio entries zero and the absent $b_1$ column omitted. The position entries have degree at most one and the ratio entries degree at most two. On the nonzero-ratio locus this row and column scaling is invertible. The criticality of $a_1$ therefore says that the coordinate row $e_{a_1}$ is in the span of the selected $c$ rows of $B$ at the actual point. All common zeros of the selected equations are on this nonzero-ratio locus: $F_{i1}=0$ forces every $a_i\ne0$, and $F_{1j}=0$ then forces every $b_j\ne0$. Thus no inverse variables are needed. For any field derivation $\partial$, form the vector consisting of $\partial x$ in the position columns and $\partial a_i/a_i,\partial b_j/b_j$ in the ratio columns. Its product with row $B_i$ equals $\partial F_i/\Lambda_i$. \paragraph{Ranks at most twenty-one.} In $n=N+c\le50$ variables $(X,A)$ impose \[ F_i(X)=0\quad(1\le i\le c),\qquad e_{a_1}=\sum_{i=1}^c A_i B_i(X). \] The nineteen position-column equations have degree at most two and the ten ratio-column equations degree at most three. Together with the six cubic and $c-6$ quartic distance equations, their degree product is bounded by \[ \Delta_c=2^{19}3^{16}4^{c-6}. \] The individual coefficient heights give a sharper sum than their joint maximum. Each selected distance equation satisfies $h(F_{ij})\le2h(\chi_{ij})+\log2$. The nineteen position-column equations have integer coefficients of absolute value at most two. In a ratio-column equation the coefficients involve only defects of selected incident edges, each to the first power, so its height is at most the sum of those defect heights plus $\log2$. Each selected nonstar defect occurs in exactly two ratio columns. Consequently the sum of equation heights divided by their assigned degree bounds is at most \[ \frac{\Sigma}{2}+\frac{2\Sigma}{3} +\left(\frac c4+\frac12+\frac{19}{2}+\frac{10}{3}\right)\log2 =\frac76\Sigma+\left(\frac c4+\frac{40}{3}\right)\log2. \] Lemma~\ref{ub:opt:individual-heights} bounds the height of the whole critical zero set by \[ \Delta_c\left[s_n+\frac76\Sigma +\left(\frac c4+\frac{40}{3}\right)\log2+n\log(n+1)\right]. \] Choose an irreducible component $W$ through the lifted actual point. On the nonempty open set where the selected rows have rank $c$, the multipliers are uniquely determined by $X$. Thus $\dim W\le N-c$. Applying a derivation to the selected equations, the logarithmic-Jacobian identity gives $\partial a_1/a_1=0$, and hence $a_1$ is constant on $W$. By the KPS coordinate-projection bound, \[ h(a_1)\le\Delta_c\left[ \frac76\Sigma+\left(\frac c4+\frac{40}{3}\right)\log2+s_n +(119-2c)\log(n+1)\right]. \] Using $s_n<90$, $\log2<0.7$, $\log51<4$, $10\le c\le21$, and $\Sigma\le11J/K$, this yields \begin{equation}\label{ub:opt:lowrank} h(a_1)<2^{49}3^{16}\left(\frac76\Sigma+500\right) \le513\cdot2^{49}3^{16}\frac JK\qquad(J\ge K). \end{equation} \paragraph{Ranks at least twenty-two.} Let $V$ be the irreducible component through the actual point of the selected distance zero set. The selected gradients are independent there, so $\dim V=r=N-c$. Put $D_0=3^6 4^{c-6}$. The preceding individual-height lemma gives \[ \deg V\le D_0,\qquad h_{\rm KPS}(V)\le D_0\left[ s_{29}+\frac12\Sigma +\left(\frac c4+\frac12\right)\log2+c\log30\right]. \] Delete the $a_1$ column of the selected $c$-by-$N$ matrix $B$ and take all its $c$-by-$c$ minors. At most nine remaining ratio columns have degree two, while all position columns have degree one, so every minor has degree at most $d=c+9$. For $c=29$ there are no such minors and $V$ is already zero-dimensional. Their joint affine coefficient height is at most $\Sigma+6c$. Indeed row $i$ involves only its own defect $\chi_i$, with $\chi_i=1$ on a star edge. At an infinite place, each entry in that row has coefficient $\ell^1$ norm at most $12\max(1,|\chi_i|_v)$: this follows by expanding $T_i=2\Lambda_i+L_i-2$, whose squared-distance part has coefficient $\ell^1$ norm at most eight. A determinant therefore has coefficient norm at most $c!12^c\prod_i\max(1,|\chi_i|_v)$. At finite places the bound is the same product without $c!12^c$. Summing over places gives \[ H_{\rm min}\le\Sigma+\log(c!)+c\log12<\Sigma+6c, \] because $c\le29$ and $12c\le348S_0$; its uniform history, exception estimates, scores, and weighted extraction produce a graph with at least $t^{2-1/K}$ edges on at most $t^{1+1/K}$ positions, edge heights at least $J/3$, rectangle-defect heights at most $J/K$, and $J\ge K$. All of its edges avoid popular unit circles. For sufficiently large $t$ this contradicts the finite obstruction just established. Therefore the asserted upper bound holds for all large $N$, and an absolute multiplicative constant covers the finitely many smaller integers. Unordered unit pairs and balanced point-circle incidences are related by Lemma~\ref{ub:lem:symmetry}. The balanced-to-unbalanced transfer in the finite framework gives the asserted two-parameter bound with the same $\delta_{\mathrm{up}}$. Finally $514560000000<2^{40}$, so \[ -\log_2\delta_{\mathrm{up}} = 2^{664}+\log_2(514560000000) <2^{664}+40<2^{665}. \] Thus $\delta_{\mathrm{up}}>\delta_{\rm simple}$, proving the simpler displayed bound as well. \end{proof} This also strictly improves the stronger baseline value $\delta_1$, not only its simplified lower bound $\delta_0$. Let $k_{\rm I}$ denote the parameter in \eqref{ub:eq:globalparams}. Since $k_{\rm I}>2^{1024}$, \[ -\log_2\delta_1 =2^{201}k_{\rm I}^{100}+\log_2 3 >2^{102601}>2^{665}>-\log_2\delta_{\mathrm{up}}. \] Hence $\delta_{\mathrm{up}}>\delta_1$. The resulting piecewise form is \[ \mathcal U(n,m)\ll \begin{cases} s^{2/3-2\delta_{\mathrm{up}}}M^{2/3+\delta_{\mathrm{up}}},&s\le M\le s^2,\\ M,&M\ge s^2. \end{cases} \] For $M=s^\alpha$ with $1\le\alpha\le2$, the first line is $O(s^{\frac23(1+\alpha)-(2-\alpha)\delta_{\mathrm{up}}})$. For $M\ge s^2$, Proposition~\ref{thm:extreme} gives the exact value. \appendix \section{Explicit local profiles} \label{sc:profiles} All numbers in this appendix are exact. A terminating decimal denotes its rational value. The exponential of a displayed logarithmic scale is taken as a real number, so no rounding of a scale is part of the definition. \subsection{The asymmetric archimedean profiles} We use thirteen archimedean choices. In the notation of \eqref{as:profile}, put \[ \phi(x,y)=(1-x)(1-y),\qquad P_A(x,y)=\prod_{q\in\boldsymbol q_A}(1-q\phi(x,y)),\qquad P_B(x,y)=\prod_{q\in\boldsymbol q_B}(1-q\phi(x,y)). \] A one-entry vector gives a single factor. Unless specified otherwise, $\lambda_B=0$. Profile $A_{13}$ uses the two displayed coefficient matrices instead of these product formulas. In every row $a=\exp(\log a)$ and $b=\exp(\log b)$; the counting powers are $p_A,p_B$. \begingroup\small \paragraph{Profile $A_{1}$.} \begin{align*} s&=1.2043989997038962,&t&=1.2043886824255563,\\ \log a&=-8.5510686631045161,&\log b&=-8.550651149265045,\\ p_A&=1.9021620032341802,&p_B&=1.9020662133751431,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.76887061009101054,\,-0.10783506631385534,\,-0.22871734724239584,\,-0.07264260580682555),\\ \boldsymbol q_B=(0.77117724356602846,\,-0.14696642002230928,\,-0.1534442371149321,\,-0.11105762495305768). \end{gather*} \paragraph{Profile $A_{2}$.} \begin{align*} s&=1.2043886824255563,&t&=1.2043989997038962,\\ \log a&=-8.550651149265045,&\log b&=-8.5510686631045161,\\ p_A&=1.9020662133751431,&p_B&=1.9021620032341802,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.77117724356602846,\,-0.14696642002230928,\,-0.1534442371149321,\,-0.11105762495305768),\\ \boldsymbol q_B=(0.76887061009101054,\,-0.10783506631385534,\,-0.22871734724239584,\,-0.07264260580682555). \end{gather*} \paragraph{Profile $A_{3}$.} \begin{align*} s&=1.1245092973972035,&t&=1.1247709010373244,\\ \log a&=-14.911457126049363,&\log b&=-14.912506062422008,\\ p_A&=1.9407041990130223,&p_B&=1.9405879030029978,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.65417664278633447,\,-0.15486713748163783,\,-0.10075053127190532,\,-0.24170651732520421),\\ \boldsymbol q_B=(0.65379078025747095,\,-0.21555353527088272,\,-0.10638723412874324,\,-0.17180448022262906). \end{gather*} \paragraph{Profile $A_{4}$.} \begin{align*} s&=1.3594315662614589,&t&=0.70699635212401812,\\ \log a&=-7.9465391424177376,&\log b&=-7.9447708032757562,\\ p_A&=1.4740864760107992,&p_B&=2.6806106845531392,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.76872077289684793),\\ \boldsymbol q_B=(0.50877380881889245). \end{gather*} \paragraph{Profile $A_{5}$.} \begin{align*} s&=1.2255540979059307,&t&=0.40460385905497714,\\ \log a&=-12.220791783262671,&\log b&=-12.285429182578341,\\ p_A&=1.3037203671086752,&p_B&=3.7748411700555744,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.7215142797140347),\\ \boldsymbol q_B=(0.35785243603442674). \end{gather*} \paragraph{Profile $A_{6}$.} \begin{align*} s&=1.3325274080489524,&t&=0.53875810170025862,\\ \log a&=-7.2898212608618529,&\log b&=-7.1737906334714499,\\ p_A&=1.3678970737119656,&p_B&=3.1221335967204542,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.88036964413350938,\,-1.2599682068584213,\,0.21830494292067165,\,0.13926299502740697),\\ \boldsymbol q_B=(0.72924190769296271,\,-0.24792205322667021,\,-0.26836424936085079,\,-0.28090971334204096). \end{gather*} \paragraph{Profile $A_{7}$.} \begin{align*} s&=1.2128575878994745,&t&=0.30695909643943042,\\ \log a&=-10.185246038705765,&\log b&=-10.105937104613458,\\ p_A&=1.2281456888253364,&p_B&=4.5512896290149945,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.85224580275921558,\,-2.0550667048347195,\,0.28466325079340971,\,0.18710522486521236),\\ \boldsymbol q_B=(0.63561360296940816,\,-0.30103276870604978,\,-0.29459829242192803,\,-0.19660203638447701). \end{gather*} \paragraph{Profile $A_{8}$.} \begin{align*} s&=1.3305301253599091,&t&=0.47060916132191954,\\ \log a&=-6.5625552610604423,&\log b&=-6.4379422235146553,\\ p_A&=1.3173941176848287,&p_B&=3.395835141488464,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.87714813827272919,\,-1.5244741451477446,\,0.47906719965973621),\\ \boldsymbol q_B=(0.71467078140860318,\,-0.41308437981809476,\,-0.41307579827320773). \end{gather*} \paragraph{Profile $A_{9}$.} \begin{align*} s&=1.2118504781572903,&t&=0.27097683672088069,\\ \log a&=-9.1090438455467684,&\log b&=-9.3167870504534669,\\ p_A&=1.1959224755575879,&p_B&=5.0110364991429845,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.84409554192296543),\\ \boldsymbol q_B=(0.36503284547727038). \end{gather*} \paragraph{Profile $A_{10}$.} \begin{align*} s&=1.3317397837887492,&t&=0.4892702297723529,\\ \log a&=-6.7376369289230027,&\log b&=-6.612451176882594,\\ p_A&=1.3310896776022887,&p_B&=3.31273523536685,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.90327587325239911,\,-1.1886509103101082,\,0.29319880855158598,\,0.00019948162526266087),\\ \boldsymbol q_B=(0.7515748826447175). \end{gather*} Here $\lambda_B=0.79381059197794501$. \paragraph{Profile $A_{11}$.} \begin{align*} s&=1.216127878632149,&t&=0.26791281537822215,\\ \log a&=-8.9171753192348788,&\log b&=-8.8407237731055606,\\ p_A&=1.1946063213291391,&p_B&=5.0308042073269617,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.83140364256019872,\,0.59452433678117078,\,-1.9884905298678195),\\ \boldsymbol q_B=(0.61188917676187526,\,-0.43810531231284111,\,-0.2919158644300448). \end{gather*} \paragraph{Profile $A_{12}$.} \begin{align*} s&=1.2070837456977095,&t&=0.27270146505219034,\\ \log a&=-9.6653965853318624,&\log b&=-9.5865452608277746,\\ p_A&=1.2011561014509533,&p_B&=4.9645738745963817,\\ \end{align*} \begin{gather*} \boldsymbol q_A=(0.87989553320102365,\,0.3371138266507665,\,-1.9840895161416183,\,0.00078042641678250225),\\ \boldsymbol q_B=(0.66490582684150423). \end{gather*} Here $\lambda_B=0.79887701051583659$. \paragraph{Profile $A_{13}$.} \begin{align*} s&=1.2643704818394643,&t&=0.3569370831364012,\\ \log a&=-6.960105659675447,&\log b&=-6.8719742698834,\\ p_A&=\frac{2621439999847412109375}{2093974328671816122368},&p_B&=\frac{13107199999237060546875}{3256468905406373298176}. \end{align*} Here $\lambda_B=0.7958371960711847$. Writing $P_A(x,y)=\sum_{i,j=0}^4a_{ij}x^iy^j$ and $P_B(x,y)=\sum_{i,j=0}^3b_{ij}x^iy^j$, the symmetric coefficient matrices are as follows: \begin{longtable}{rrr}\toprule $i$&$j$&$a_{ij}$\\\midrule\endhead 0&0&$0.08986490242942435$\\ 0&1&$1.6882926023090434$\\ 0&2&$-1.0295317237726267$\\ 0&3&$-0.9419136560076167$\\ 0&4&$1.1932878750417757$\\ 1&1&$3.6568888170936775$\\ 1&2&$-11.964534678648683$\\ 1&3&$11.576781135858283$\\ 1&4&$-4.957427876612321$\\ 2&2&$27.567384358283995$\\ 2&3&$-16.845106163973167$\\ 2&4&$2.2717882081104745$\\ 3&3&$-13.967139715783283$\\ 3&4&$20.17737839990578$\\ 4&4&$-18.68502660644571$\\ \bottomrule\end{longtable} \begin{longtable}{rrr}\toprule $i$&$j$&$b_{ij}$\\\midrule\endhead 0&0&$0.24276175946342138$\\ 0&1&$1.003335297720344$\\ 0&2&$-0.9926933584947837$\\ 0&3&$0.7465963013110183$\\ 1&1&$-1.5068996810513218$\\ 1&2&$3.6068291735844076$\\ 1&3&$-3.10326479025343$\\ 2&2&$-9.954310501105798$\\ 2&3&$7.340174686016173$\\ 3&3&$-4.983506197073761$\\ \bottomrule\end{longtable} \endgroup \subsection{Phase mixtures} Write $F_1,\ldots,F_{27}$ for the finite-place choices in the finite parameter table. A pair $(A_i,F_j)$ means that all archimedean places use $A_i$ and all selected finite places use $F_j$. The following formulas give the exact proportions. A phase is the physical mixture described in the proof of the transfer theorem; every list of proportions sums to one. The first and second phases correspond to the low and high rows of Table~\ref{as:table}. \paragraph{Ratio $r=1$.} Phase 1 has the following nonzero proportions: \begin{align*} \lambda(A_{1},F_{1})&=\frac{16958907852215594713173329267}{62500000000000000000000000000},\\ \lambda(A_{1},F_{2})&=\frac{24339448698597155286826670733}{62500000000000000000000000000},\\ \lambda(A_{2},F_{1})&=\frac{8706320241288655286826670733}{62500000000000000000000000000},\\ \lambda(A_{2},F_{2})&=\frac{12495323207898594713173329267}{62500000000000000000000000000}. \end{align*} Phase 2 has the following nonzero proportions: \begin{align*} \lambda(A_{3},F_{3})&=\frac{24934353685121}{200000000000000},\\ \lambda(A_{3},F_{4})&=\frac{196019949657233}{500000000000000},\\ \lambda(A_{3},F_{5})&=\frac{483288332259929}{1000000000000000}. \end{align*} \paragraph{Ratio $r=1.02$.} Phase 1 has the following nonzero proportions: \begin{align*} \lambda(A_{4},F_{6})&=\frac{1798339241187}{62500000000000},\\ \lambda(A_{4},F_{7})&=\frac{96423893118511}{500000000000000},\\ \lambda(A_{4},F_{8})&=\frac{389189392951993}{500000000000000}. \end{align*} Phase 2 has the following nonzero proportions: \begin{align*} \lambda(A_{5},F_{9})&=\frac{1571914426989}{2000000000000},\\ \lambda(A_{5},F_{10})&=\frac{208994562167819}{1000000000000000},\\ \lambda(A_{5},F_{11})&=\frac{5048224337681}{1000000000000000}. \end{align*} \paragraph{Ratio $r=1.04$.} Phase 1 has the following nonzero proportions: \begin{align*} \lambda(A_{6},F_{12})&=\frac{222724515886789}{500000000000000},\\ \lambda(A_{6},F_{13})&=\frac{313697877428409}{1000000000000000},\\ \lambda(A_{6},F_{14})&=\frac{240853090798013}{1000000000000000}. \end{align*} Phase 2 has the following nonzero proportions: \begin{align*} \lambda(A_{7},F_{15})&=\frac{16116346494661}{250000000000000},\\ \lambda(A_{7},F_{16})&=\frac{6607013992789}{50000000000000},\\ \lambda(A_{7},F_{17})&=\frac{100424291770697}{125000000000000}. \end{align*} \paragraph{Ratio $r=1.06$.} Phase 1 has the following nonzero proportions: \begin{align*} \lambda(A_{8},F_{18})&=\frac{3993394610163}{12500000000000},\\ \lambda(A_{8},F_{19})&=\frac{275259845593321}{500000000000000},\\ \lambda(A_{8},F_{20})&=\frac{65004370000159}{500000000000000}. \end{align*} Phase 2 has the following nonzero proportions: \begin{align*} \lambda(A_{9},F_{21})&=\frac{91393322620347}{1000000000000000},\\ \lambda(A_{9},F_{22})&=\frac{90545392677349}{1000000000000000},\\ \lambda(A_{9},F_{23})&=\frac{25564415146947}{31250000000000}. \end{align*} \paragraph{Ratio $r=1.0603071$.} Phase 1 has the following nonzero proportions: \begin{align*} \lambda(A_{8},F_{24})&=\frac{3298340691235568359442235627}{50000000000000000000000000000},\\ \lambda(A_{8},F_{20})&=\frac{32526359599855431640557764373}{50000000000000000000000000000},\\ \lambda(A_{10},F_{24})&=\frac{1305104228659281640557764373}{50000000000000000000000000000},\\ \lambda(A_{10},F_{20})&=\frac{12870195480249718359442235627}{50000000000000000000000000000}. \end{align*} Phase 2 has the following nonzero proportions: \begin{align*} \lambda(A_{11},F_{25})&=\frac{16105327022701024995863770053}{125000000000000000000000000000},\\ \lambda(A_{11},F_{23})&=\frac{98157173024743100004136229947}{125000000000000000000000000000},\\ \lambda(A_{12},F_{25})&=\frac{1513453215799975004136229947}{125000000000000000000000000000},\\ \lambda(A_{12},F_{23})&=\frac{9224046736755899995863770053}{125000000000000000000000000000}. \end{align*} \paragraph{Ratio $r=1.0629395$.} Phase 1 has the following nonzero proportions: \begin{align*} \lambda(A_{13},F_{26})&=\frac{192596700618323}{1000000000000000},\\ \lambda(A_{13},F_{27})&=\frac{807403299381677}{1000000000000000}. \end{align*} \subsection{The finite-place profiles}\label{sc:finite-profiles} The shell matrices are specified by a finite recurrence. This gives their entries directly from the following parameter table. For $x>0$, write $e(x)=\lfloor\log_2x\rfloor$ and set \[ \mathcal R(x)=2^{e(x)-96}\bigl\lfloor2^{96-e(x)}x\bigr\rfloor, \qquad \mathcal R(0)=0. \] The number $\mathcal R(x)$ is rational and satisfies \[ 0\le x-\mathcal R(x)<2^{-96}x\qquad(x>0). \] Thus the error is bounded relative to $x$, including when $x$ is very small. For a residue cardinality $Q$, put $\mu_0=1$ and $\mu_i=(Q-1)Q^{i-1}$ for $i\ge1$. Define the $n\times n$ matrix \[ K_{ia}=\begin{cases} \mu_{\min(i,a)}/\mu_i,&i\ne a,\\ i-1/(Q-1),&i=a\ge1,\\ 0,&i=a=0, \end{cases} \qquad T_k(W)=(k+1)W+KW+WK^{\mathsf T}. \] For $v>1$ let \[ [\mathcal P_{k,v}(W)]_{ij}= \mathcal R\!\left( \left(\frac{T_k(W)_{ij}}{T_k(W)_{00}}\right)^{1/(v-1)} \right). \] \begin{algorithm}[Constructing the finite-place matrices]\label{sc:finite-generation} Fix $Q,n,k,v_L,v_R$ from the parameter table. \begin{enumerate} \item Form the matrix $K$ above. Let $E_{00}$ have entry $1$ in position $(0,0)$ and entry $0$ elsewhere, and set $L_0=R_0=E_{00}$. \item For each $j=0,1,\ldots,511$, calculate \[ L_{j+1}=\mathcal P_{k,v_L}(R_j),\qquad R_{j+1}=\mathcal P_{k,v_R}(L_{j+1}). \] In each application of $\mathcal P$, first form $T_k(W)$, divide its entries by $T_k(W)_{00}$, take the indicated positive real powers, and then apply $\mathcal R$ to each entry. \item Set $w_A=L_{512}$ and $w_B=R_{512}$. These rational matrices are the finite-place weights. \end{enumerate} \end{algorithm} In every finite choice below, $n=32$ at $Q=16,9,25$, and $n=8$ at $Q=29^4,41^4$. The period vector is ordered as $(16,9,25,29^4,41^4)$. The same two counting powers are used at all five places. \begin{longtable}{rlll} \toprule Choice&Period vector&$v_L$&$v_R$\\\midrule\endhead $F_{1}$&$(6,7,5,0,0)$&$1.9022689819024641$&$1.9019590028041919$\\ $F_{2}$&$(6,7,5,1,0)$&$1.9019716183498516$&$1.9022566182314828$\\ $F_{3}$&$(10,13,9,1,1)$&$1.9406729483332157$&$1.9401892938290983$\\ $F_{4}$&$(11,13,9,1,1)$&$1.9396621452441349$&$1.9415133796297677$\\ $F_{5}$&$(11,13,9,1,1)$&$1.9411333073005879$&$1.9401684191430082$\\ $F_{6}$&$(6,7,5,1,0)$&$1.5150990483651603$&$2.5549211308995434$\\ $F_{7}$&$(6,7,5,1,0)$&$1.4846999478560836$&$2.6463310741257056$\\ $F_{8}$&$(6,7,5,0,0)$&$1.4756691623397213$&$2.6753750220619135$\\ $F_{9}$&$(9,11,8,1,1)$&$1.3058560497722596$&$3.7576821673361782$\\ $F_{10}$&$(9,11,8,1,1)$&$1.3025106190128546$&$3.7837028164053099$\\ $F_{11}$&$(9,11,8,1,1)$&$1.3037447922823469$&$3.774609208029104$\\ $F_{12}$&$(5,7,4,0,0)$&$1.3728917907209091$&$3.0943782316699675$\\ $F_{13}$&$(5,7,5,0,0)$&$1.3699424065203709$&$3.1115949392473694$\\ $F_{14}$&$(5,7,4,0,0)$&$1.3684277759573105$&$3.1192389794694209$\\ $F_{15}$&$(8,10,7,1,1)$&$1.2351336586754948$&$4.4451277217711773$\\ $F_{16}$&$(8,10,7,1,1)$&$1.2322835046253935$&$4.5088452416972107$\\ $F_{17}$&$(8,10,7,1,1)$&$1.228835306227319$&$4.5421194493594781$\\ $F_{18}$&$(5,6,4,0,0)$&$1.3373525036196185$&$3.2859074942080362$\\ $F_{19}$&$(5,6,4,0,0)$&$1.3214178043370663$&$3.375419719293832$\\ $F_{20}$&$(5,6,4,0,0)$&$1.3209948616341927$&$3.3752128297683086$\\ $F_{21}$&$(7,9,6,1,1)$&$1.2083788305014338$&$4.8126854566529449$\\ $F_{22}$&$(7,9,6,1,1)$&$1.19326286612058$&$5.0651385979551264$\\ $F_{23}$&$(7,9,6,1,1)$&$1.1967000740450726$&$4.9971089399496451$\\ $F_{24}$&$(5,6,4,0,0)$&$1.320295962717883$&$3.3767412090880851$\\ $F_{25}$&$(7,9,6,1,1)$&$1.2003100925483199$&$4.9285262048455696$\\ $F_{26}$&$(5,7,5,1,0)$&$1.2522867984705557$&$4.0212303103124363$\\ $F_{27}$&$(5,7,4,1,0)$&$1.2521690761030486$&$4.0223722356362588$\\ \bottomrule\end{longtable} The recurrence has nonnegative entries and preserves symmetry in the shell indices. Substitution in \eqref{as:finite-resources}, followed by the five densities in \eqref{as:finite-densities}, gives the following intermediate values. Each displayed number has absolute error less than $10^{-13}$. Here $X_{f,A},X_{f,B},Z_f$ denote the complete weighted finite-place resources, including the terms $k\log Q$. \begin{longtable}{rrrr} \toprule Choice&$X_{f,A}$&$X_{f,B}$&$Z_f$\\\midrule\endhead $F_{1}$&$8.75655867279179$&$8.75653190899294$&$9.80998821817131$\\ $F_{2}$&$12.12381511637379$&$12.12383970217037$&$13.35055616439864$\\ $F_{3}$&$20.87357373220748$&$20.87354800397013$&$22.45992559462880$\\ $F_{4}$&$20.95753542234338$&$20.95763312879784$&$22.54652329134110$\\ $F_{5}$&$20.96076454682457$&$20.96071354888252$&$22.54977460504412$\\ $F_{6}$&$12.18372151632316$&$12.29655447963901$&$13.46428342347475$\\ $F_{7}$&$12.19723738444273$&$12.32877451515930$&$13.48568608819373$\\ $F_{8}$&$8.83498016883761$&$8.97562085819919$&$9.95390224718216$\\ $F_{9}$&$19.64772635987888$&$19.96825713051418$&$21.30449246106980$\\ $F_{10}$&$19.61889322789329$&$19.94648271230000$&$21.27660951529022$\\ $F_{11}$&$19.63825352508882$&$19.96341232488492$&$21.29594604619528$\\ $F_{12}$&$8.22068639405639$&$8.50099747958038$&$9.35576499184704$\\ $F_{13}$&$8.83495059431152$&$9.09508618656232$&$9.99511903512623$\\ $F_{14}$&$8.24042867766068$&$8.52910213089902$&$9.37921474920714$\\ $F_{15}$&$18.20015023528243$&$18.77892344481823$&$19.87221182260070$\\ $F_{16}$&$18.47222574993064$&$19.07432174978183$&$20.15872231569113$\\ $F_{17}$&$18.32241478232529$&$18.93838228245181$&$20.00694235178725$\\ $F_{18}$&$7.88966666910419$&$8.25309877304806$&$9.03350114477615$\\ $F_{19}$&$7.79308079667491$&$8.19172596255829$&$8.94241331291362$\\ $F_{20}$&$7.76971901690643$&$8.16827188209009$&$8.91778247016684$\\ $F_{21}$&$17.18308072187814$&$17.96054470990158$&$18.86281575745856$\\ $F_{22}$&$17.09844631489943$&$17.99425409756351$&$18.79914495427625$\\ $F_{23}$&$17.04815117598609$&$17.91285807347277$&$18.74087730931423$\\ $F_{24}$&$7.74685058717637$&$8.14599529909959$&$8.89386783140356$\\ $F_{25}$&$17.00174429332277$&$17.83542657797112$&$18.68655383965078$\\ $F_{26}$&$12.50471278568225$&$13.07528224987849$&$13.93875913297122$\\ $F_{27}$&$12.01985330510428$&$12.62605306785589$&$13.43986132053947$\\ \bottomrule\end{longtable} \begin{algorithm}[Determining the rational rounding]\label{sc:rational-rounding} The values $\mathcal R(0)=0$ and $\mathcal R(1)=1$ are immediate. For every other positive number $x$ occurring in Algorithm~\ref{sc:finite-generation}, proceed as follows. \begin{enumerate} \item Enclose $x$ between rational numbers, using the series bounds for $\log$ and $\exp$ in Section~\ref{ba:verification}. \item Refine these bounds until they determine an integer $e$ with $2^e\le x<2^{e+1}$. \item Multiply the bounds by $2^{96-e}$ and refine them until they determine the integer $h=\lfloor2^{96-e}x\rfloor$. Then $\mathcal R(x)=2^{e-96}h$. \end{enumerate} \end{algorithm} The finite calculations here use precision $224$ in base $2$ and determine every displayed integer $e,h$. Once the rational matrices are obtained, their Gram sums are exact; the remaining logarithms and powers are bounded by the same series. \subsection{The balanced finite-place profiles}\label{ba:finite-witness} Use Algorithm~\ref{sc:finite-generation} with the common power \[ v_L=v_R=v_0=\frac{2}{1+0.04317220235645} =\frac{40000000000000}{20863444047129}. \] The five parameter choices are \[ \begin{array}{c|rrrrr} Q&16&9&25&29^4&41^4\\\hline n&64&64&64&12&12\\ k&7&9&6&1&1 \end{array} \] For each choice take $w=R_{512}$. Equivalently, apply the one-matrix map $\mathcal P_{k,v_0}$ exactly $1024$ times to $E_{00}$. The generation power $v_0$ is fixed; the exponent $p=2/(1+\delta)$ in \eqref{ba:finite} varies with $\delta$. Substitution in \eqref{ba:F-H} gives the following finite-place intermediate values, each with absolute error less than $10^{-34}$: \[ \begin{array}{c|l} \delta&\mathcal F(\delta)\\\hline 0.04317218&0.71826540946064369003155693342528624\\ 0.04317228&0.71826381501041550320423926484819084\\ 0.0431722305539701&0.71826460340275401512882615794339374\\ \end{array} \] On the full interval $[0.04317218,0.04317228]$, the finite-place curvature expression satisfies \[ 1.5538279509314