% Mechanical adaptation of https://doi.org/10.5281/zenodo.24220727. % The article and its technical companion form one source-only project. % Mathematical source files and bibliography contents are unchanged. \PassOptionsToPackage{expansion=false,nopatch=footnote}{microtype} \documentclass[11pt,a4paper]{article} % Shared typesetting and notation for the two associated documents. \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage[english]{babel} \usepackage{lmodern,microtype} \usepackage[a4paper,margin=24mm,headheight=14pt]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools,mathrsfs} \usepackage{booktabs,longtable,array,tabularx} \usepackage{xcolor,enumitem,fancyhdr,needspace,etoolbox,placeins} \usepackage{xr-hyper} \usepackage[colorlinks=true,linkcolor=blue!40!black,citecolor=blue!40!black,urlcolor=blue!40!black,filecolor=blue!40!black]{hyperref} \usepackage{bookmark,xurl} \newcommand{\E}{\mathbb E} \renewcommand{\P}{\mathbb P} \newcommand{\N}{\mathbb N} \newcommand{\Z}{\mathbb Z} \newcommand{\F}{\mathbb F} \newcommand{\Law}{\mathcal L} \newcommand{\Pois}{\operatorname{Pois}} \newcommand{\TV}{d_{\mathrm{TV}}} \newcommand{\ind}{\mathbf 1} \newcommand{\eps}{\varepsilon} \DeclareMathOperator{\Var}{Var} \DeclareMathOperator{\Cov}{Cov} \DeclareMathOperator{\rank}{rank} \DeclareMathOperator{\codim}{codim} \renewcommand{\sf}{\operatorname{sf}} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newtheorem{example}[theorem]{Example} \numberwithin{equation}{section} \setlength{\parindent}{1.2em} \setlength{\parskip}{3pt} \setlength{\emergencystretch}{2em} \setlist{itemsep=3pt,topsep=5pt,leftmargin=20pt} \renewcommand{\arraystretch}{1.12} \pagestyle{fancy} \fancyhf{} \fancyfoot[C]{\thepage} \renewcommand{\headrulewidth}{0pt} \setcounter{tocdepth}{1} \hypersetup{ pdftitle={Sharp Poisson approximation and second-order correction for long multiplicative runs}, pdfauthor={Brice Pouly}, pdfsubject={Random multiplicative functions, Poisson approximation, hereditary stopping, arithmetic dependencies}, pdfkeywords={Random multiplicative functions; long runs; sharp Poisson approximation; factorial cumulants}, bookmarksdepth=2} \allowdisplaybreaks[1] \widowpenalty=10000 \clubpenalty=10000 \usepackage{xparse} \hypersetup{hypertexnames=false} \setlength{\emergencystretch}{3em} % Keep each original contents list while making all references internal. \makeatletter \newcommand{\HXtocfile}{articlecontents} \AddToHook{begindocument/end}[HXcontents]{% \let\HXoriginaladdtocontents\addtocontents \let\HXoriginalstarttoc\@starttoc \renewcommand{\addtocontents}[2]{% \ifstrequal{#1}{toc}{\HXoriginaladdtocontents{\HXtocfile}{#2}}{\HXoriginaladdtocontents{#1}{#2}}} \def\@starttoc#1{% \ifstrequal{#1}{toc}{\HXoriginalstarttoc{\HXtocfile}}{\HXoriginalstarttoc{#1}}}} % The article class clears title commands after the first title page. \AddToHook{begindocument/end}[HXtitles]{% \let\HXkernelmaketitle\maketitle \let\HXkernelatmaketitle\@maketitle \let\HXkerneltitle\title \let\HXkernelauthor\author \let\HXkerneldate\date \let\HXkernelthanks\thanks \let\HXkerneland\and \def\HXrepeatmaketitle{% \HXkernelmaketitle \global\let\maketitle\HXrepeatmaketitle \global\let\@maketitle\HXkernelatmaketitle \global\let\title\HXkerneltitle \global\let\author\HXkernelauthor \global\let\date\HXkerneldate \global\let\thanks\HXkernelthanks \global\let\and\HXkerneland} \let\maketitle\HXrepeatmaketitle} % Preserve both original bibliography orders and citation numbers without % duplicating citation destinations. Source citation commands are unchanged. \newcommand{\HXbibprefix}{article} \newcommand{\HXciteone}[1]{% \edef\HXkey{\zap@space#1 \@empty}% \HXciteseparator \ifcsname HXbib@\HXbibprefix @\HXkey\endcsname \hyperlink{HXbib.\HXbibprefix.\HXkey}{\csname HXbib@\HXbibprefix @\HXkey\endcsname}% \else \PackageError{hexagon-source}{Unknown citation \HXkey}{Missing frozen bibliography key}% \fi \def\HXciteseparator{,\penalty10000\space}} \AddToHook{begindocument/end}[HXbibliographies]{% \RenewDocumentCommand{\cite}{o m}{% \begingroup\upshape[\def\HXciteseparator{}% \forcsvlist{\HXciteone}{#2}% \IfNoValueF{#1}{,\space#1}% ]% \endgroup} \RenewDocumentCommand{\bibitem}{o m}{% \IfNoValueTF{#1}{\item}{\item[{#1}]}% \hypertarget{HXbib.\HXbibprefix.#2}{}}} \makeatother \expandafter\def\csname HXbib@article@agg1989\endcsname{1} \expandafter\def\csname HXbib@article@agw1990\endcsname{2} \expandafter\def\csname HXbib@article@barbour1987\endcsname{3} \expandafter\def\csname HXbib@article@barboureagleson1987\endcsname{4} \expandafter\def\csname HXbib@article@bkn2014\endcsname{5} \expandafter\def\csname HXbib@article@bretwngxu2026\endcsname{6} \expandafter\def\csname HXbib@article@dohmen1999\endcsname{7} \expandafter\def\csname HXbib@article@godbole1991\endcsname{8} \expandafter\def\csname HXbib@article@naimanwynn1997\endcsname{9} \expandafter\def\csname HXbib@article@najnudel2020\endcsname{10} \expandafter\def\csname HXbib@article@pouly2026c\endcsname{11} \expandafter\def\csname HXbib@article@zeilberger1997\endcsname{12} \expandafter\def\csname HXbib@article@cekanaviciusvellaisamy2015\endcsname{13} \expandafter\def\csname HXbib@article@chhaibietal2020\endcsname{14} \expandafter\def\csname HXbib@article@cekanaviciusvellaisamy2020\endcsname{15} \expandafter\def\csname HXbib@article@basaknathzaharescu2024\endcsname{16} \expandafter\def\csname HXbib@article@shparlinskisleiman2026\endcsname{17} \expandafter\def\csname HXbib@article@kempkemp1965\endcsname{18} \expandafter\def\csname HXbib@companion@bkn2014\endcsname{1} \expandafter\def\csname HXbib@companion@binyaminicluckersnovikov2024\endcsname{2} \expandafter\def\csname HXbib@companion@bretwngxu2026\endcsname{3} \expandafter\def\csname HXbib@companion@erdosturk1984\endcsname{4} \expandafter\def\csname HXbib@companion@heathbrown2012\endcsname{5} \expandafter\def\csname HXbib@companion@munsch2026\endcsname{6} \expandafter\def\csname HXbib@companion@pila1995\endcsname{7} \expandafter\def\csname HXbib@companion@reuss2014\endcsname{8} \expandafter\def\csname HXbib@companion@salberger2023\endcsname{9} \expandafter\def\csname HXbib@companion@zacharovashwang2010\endcsname{10} \expandafter\def\csname HXbib@companion@griesmer1960\endcsname{11} \expandafter\def\csname HXbib@companion@cilleruelogaraev2011\endcsname{12} \expandafter\def\csname HXbib@companion@tao2026anatomy\endcsname{13} \expandafter\def\csname HXbib@companion@shparlinskisleiman2026\endcsname{14} \expandafter\def\csname HXbib@companion@binyaminicluckerskato2025\endcsname{15} \expandafter\def\csname HXbib@companion@bretkurlbergshparlinski2021\endcsname{16} \title{\bfseries Sharp Poisson approximation\\and second-order correction\\for long multiplicative runs} \author{Brice Pouly\\[5pt] \small With mathematical and editorial assistance from ChatGPT (GPT-6 Astra)} \date{} \begin{document} \pdfbookmark[0]{Main article}{HXarticle} \maketitle \thispagestyle{plain} \begin{abstract} Let \(f\) be completely multiplicative with independent fair signs at primes. We count starts of constant runs of length at least \(L\) in \([N,2N)\), with \(L=\log_2N+O(1)\) and both signs included. The total variation distance to the Poisson law with nominal or exact mean has an explicit equivalent of order \(LN^{-2/3}\), with a coefficient depending on \(N2^{-L}\) and \(L\bmod3\). After retaining the exact mean and second factorial cumulant, the quadratic Poisson--Charlier approximation has optimal error power \(4/3\): for every fixed \(\varepsilon>0\), its \(\ell^1\) error lies between constant multiples of \(N^{-4/3-\varepsilon}\) and \(N^{-4/3+\varepsilon}\), uniformly in each fixed critical band. The leading correction comes from rational pairs. Keeping all channels up to a fixed height \(h\) gives an explicit distance equivalent with power \(h/(h+1)\), determined by the first omitted height. The third factorial cumulant has optimal power \(41/30\). The proof expands around the exact Bernoulli marginals and stops at the first excluded dependent configuration. Arithmetic bounds for fixed configurations and a separate uniform tail estimate control the resulting positive boundary mass. Short square relations and a common algebraic curve bound the triple interactions; a fourth Charlier test detects the two-pair contribution that makes the quadratic exponent optimal. \end{abstract} \clearpage \begingroup \small \setlength{\parskip}{0pt} \tableofcontents \endgroup \clearpage \section{Introduction} \label{p1:intro} A long constant run in independent fair signs starts with probability \(2^{-L}\). Thus \(N\) possible starts have a bounded expected count when \(L=\log_2N+O(1)\). A completely multiplicative sign sequence has the same rare-event scale, but its distant windows share prime signs. The resulting dependencies leave an explicit signature in the Poisson approximation. The first signature comes from pairs of windows in the ratio \(2:3\). Their total contribution has size \(LN^{-2/3}\). We determine its coefficient and then ask how accurately the whole counting law can be recovered after the exact pair correction is retained. The answer has optimal power \(4/3\): two pair interactions survive at fourth order, whereas the total triple activity decays with power \(41/30\). The gap \(41/30-4/3=1/30\) is enough, provided that the expansion also controls configurations whose cardinality grows with \(N\). \subsection{Model and principal results} \label{p1:intro:model} Choose independent fair signs \(f(\ell)\), indexed by primes, and set \[ f(n)=\prod_{\ell}f(\ell)^{v_\ell(n)}\qquad(n\ge1). \] The function \(f\) is completely multiplicative; in particular, \(f(n^2)=1\). Fix \(C>0\). Throughout, \(N\ge2\) and \(L\ge1\) are integers satisfying \begin{equation} I_N=[N,2N)\cap\Z,\qquad |L-\log_2N|\le C,\qquad B=L+1,\qquad p=2^{-L},\qquad \lambda_N=Np. \label{p1:eq:critical} \end{equation} All logarithms without a subscript are natural. The nominal mean \(\lambda_N\) remains in a compact subinterval of \((0,\infty)\), but need not converge. Define \begin{equation} J_x=\ind_{\{f(x-1)=-f(x),\ f(x)=f(x+1)=\cdots=f(x+L-1)\}}, \qquad Z_N=\sum_{x\in I_N}J_x. \label{p1:eq:model} \end{equation} Both signs of the run are included. The left boundary condition counts a long run once, at its start. A retained run may continue beyond \(2N\), and a run starting before \(N\) is not counted. Write \[ p_x=\E J_x,\qquad \mu_N=\E Z_N,\qquad \alpha_N=\Var(Z_N)-\mu_N. \] The exact marginals \(p_x\) need not equal \(p\). For a signed measure \(\nu\) on \(\N=\{0,1,\ldots\}\), let \(\|\nu\|_1=\sum_{k\ge0}|\nu\{k\}|\); total variation between probabilities is \(\TV(P,Q)=\frac12\|P-Q\|_1\). Set \(\Pi_\mu=\Pois(\mu)\) and \(\Delta=\delta_1-\delta_0\). Products of measures are convolutions. For $r\ge1$, write $(z)_r=z(z-1)\cdots(z-r+1)$, with $(z)_0=1$, and set $\Delta^{*0}=\delta_0$. The quadratic reference is \begin{equation} Q_{\mu,\alpha} =\Pi_\mu+\frac{\alpha}{2}\Pi_\mu*\Delta^{*2}. \label{p1:eq:Q} \end{equation} It has total mass one and, in the range of the theorem below, is also nonnegative. \begin{theorem}[Optimal quadratic approximation with exact parameters] \label{p1:thm:main} Fix \(C>0\). For every \(\varepsilon>0\), there are positive constants \(c_{C,\varepsilon}\), \(C_{C,\varepsilon}\) and a threshold \(N_0(C,\varepsilon)\) such that, for all \(N\ge N_0\) and every \(L\) satisfying \eqref{p1:eq:critical}, \begin{equation} c_{C,\varepsilon}N^{-4/3-\varepsilon} \le \left\|\Law(Z_N)-Q_{\mu_N,\alpha_N}\right\|_1 \le C_{C,\varepsilon}N^{-4/3+\varepsilon}. \label{p1:eq:main} \end{equation} Moreover, \(\mu_N-\lambda_N=O_{C,\varepsilon}(N^{-1+\varepsilon})\), and \(Q_{\mu_N,\alpha_N}\) is a probability measure for all sufficiently large \(N\), uniformly in the same band. \end{theorem} The theorem determines the optimal power of the remainder, with an arbitrary fixed power margin. It does not determine its logarithmic factor or leading constant. The exact mean matters: replacing it by \(\lambda_N\) can have a larger cost of order \(N^{-1+o(1)}\). Proposition~\ref{p13:optimal:nominal} makes this distinction precise. The pair calculation gives a full leading equivalent. If \(r\in\{0,1,2\}\) is the residue of \(L\) modulo \(3\), then \begin{equation} \alpha_N\sim \frac{r+2}{18\,2^{r/3}}\lambda_N^{5/3}L N^{-2/3}. \label{p1:eq:alpha-leading} \end{equation} Here and below, an equivalent with moving parameters means that the ratio tends to one uniformly over the fixed critical band. \begin{corollary}[Exact Poisson total variation scale] \label{p1:cor:tv} For every \(\varepsilon>0\), uniformly under \eqref{p1:eq:critical}, \begin{equation} \TV(\Law(Z_N),\Pi_{\lambda_N}) =\frac{\alpha_N}{4}\|\Pi_{\lambda_N}*\Delta^{*2}\|_1 +O_{C,\varepsilon}(N^{-1+\varepsilon}). \label{p1:eq:tv-main} \end{equation} Consequently, \begin{equation} \TV(\Law(Z_N),\Pi_{\lambda_N}) \sim\frac{r+2}{72\,2^{r/3}}\lambda_N^{5/3} \|\Pi_{\lambda_N}*\Delta^{*2}\|_1\,L N^{-2/3}. \label{p1:eq:tv-equiv} \end{equation} With \(\Pi_{\mu_N}\) as the reference, the first formula holds with \(\mu_N\) in the difference norm and error \(O_{C,\varepsilon}(N^{-4/3+\varepsilon})\); the equivalent is unchanged. \end{corollary} Thus the power \(2/3\), the factor \(L\), and the leading coefficient are determined for both specified Poisson means. No minimization over all possible means is asserted. \begin{example}[A fixed phase] \label{p1:ex:unit} Take \(N=2^{3k}\) and \(L=3k\). Then \(\lambda_N=1\), \(r=0\), and \(\|\Pi_1*\Delta^{*2}\|_1=3/e\). Hence \begin{equation} \TV(\Law(Z_N),\Pi_1) \sim\frac{1}{12e}\,L N^{-2/3}. \label{p1:eq:unit-example} \end{equation} The constant combines the arithmetic coefficient \(1/9\), the second-difference norm \(3/e\), and the factor \(1/4\) from the cumulant and total variation conventions. \end{example} A rational channel $a:b$ has coprime integers $a0$ are fixed before $N\to\infty$. \begin{center} \small \begin{tabularx}{\linewidth} {@{}p{.20\linewidth}p{.25\linewidth}X@{}} \toprule Reference & Parameters retained & Precision established here\\ \midrule Poisson & Nominal mean $\lambda_N$ or exact mean $\mu_N$ & Explicit equivalent of order $LN^{-2/3}$ (Corollary~\ref{p1:cor:tv}).\\[3pt] Exact quadratic & $\mu_N,\alpha_N$ & Optimal power $4/3$, with arbitrary fixed power margins (Theorem~\ref{p1:thm:main}).\\[3pt] Explicit quadratic & Nominal mean and complete channel sums of heights $\le h$ & Explicit equivalent of order $LN^{-h/(h+1)}$ (Theorem~\ref{p13:channels:hierarchy}).\\[3pt] Exact Hermite & $\mu_N,\alpha_N$ & Upper bound $O_{C,\varepsilon}(N^{-4/3+\varepsilon})$; optimality is not established (Corollary~\ref{p1:cor:compound}).\\ \bottomrule \end{tabularx} \end{center} For $h=4$, the explicit quadratic reference retains the $2:3$ and $3:4$ channels and has optimal power $4/5$. Theorem~\ref{p13:third:main} also determines the optimal power $41/30$ of the third factorial cumulant; its positive witness consists of triples in the ratio $10:12:15$. \subsection{How the first channel produces its scale} \label{p1:intro:mechanism} The interval selects the first interior ratio. If \(x\in I_N\), then \(2x\notin I_N\), so the ratio \(1:2\) has no macroscopic family of interior placements. The next ratio, \(2:3\), does. For \(L=3m\), the windows at \(2t\) and \(3t\) contain the matched values \(2(t+j)\) and \(3(t+j)\), \(0\le j0\}\), where \(p_x=\E J_x\). An event outside \(\mathcal I_N\) is identically zero and can be removed without changing \(Z_N\). For every \(x\in I_N\), let \(S_x\) be the image of the even selections of \(V_x=\{x-1,\ldots,x+L-1\}\) in \(\mathbb Q^\times/\mathbb Q^{\times2}\), regarded as a vector space over \(\F_2\). Write \(d_x=\dim S_x\). The even-selection space has dimension $L$. Its kernel consists of local even square relations. The following bound makes the possible effective ranks explicit. Lemma~\ref{p1:limits:local-rigidity} in the companion proves that the local even square-relation code has dimension at most one in the critical band. Its proof combines a short-interval square-product bound with the binary Griesmer bound. For an admissible window the affine prescription is consistent. Its probability is $2^{-d_x}$, as the character expansion in Section~\ref{p1:alg:section} also shows. The lemma therefore gives \begin{equation} d_x\in\{L-1,L\},\qquad p_x=2^{-d_x}\in\{p,2p\}, \qquad p_*:=\max_{x\in\mathcal I_N}p_x\ll_C N^{-1}. \label{p13:exact:marginals} \end{equation} The assertion about \(d_x\) holds for all sufficiently large \(N\) in a fixed critical band. A local square relation may instead contradict the run prescription; that is precisely the case already removed by \(p_x=0\). For a finite \(I\subseteq\mathcal I_N\), put \begin{align} \rho_{\mathrm{rel}}(I) &=\sum_{x\in I}d_x-\dim\sum_{x\in I}S_x, & d_{xy}&=\dim(S_x\cap S_y),\label{p13:exact:rank}\\ q(I)&=\E\prod_{x\in I}J_x, & w(I)&=2^{\rho_{\mathrm{rel}}(I)} \prod_{x\in I}p_x,\label{p13:exact:weights}\\ c(I)&=\E\prod_{x\in I}(J_x-p_x). \label{p13:exact:center} \end{align} The empty products are one. We have \(q(I)=w(I)\) when the affine prescriptions are jointly compatible, and \(q(I)=0\) otherwise. Write $G_N=\{x\in I_N:d_x=L\}$ for the regular set. On $G_N$, these definitions agree with the homogeneous-marginal notation used in the companion. Outside \(G_N\), \(\rho_{\mathrm{rel}}\) counts dependencies between effective spaces; it does not count local relations a second time. The notions of directness, connectedness, and primitive family refer to the direct-sum decomposition of the effective spaces proved below. A family is direct when its effective spaces are in direct sum, and connected when it has just one canonical direct component. In particular, a primitive family has a nonzero relation and every proper subfamily is direct. Every nonzero relation of a primitive family has a nonzero projection in each of its blocks. A connected family need not have a relation with this property on its entire support. This distinction matters in the conditional count below. For fixed \(r\ge2\), define \begin{equation} A_r=\sum_{\substack{I\subseteq\mathcal I_N\\|I|=r}}|c(I)|, \qquad K_3^+=\sum_{\substack{|T|=3,\ T\text{ connected}\\q(T)>0}}c(T). \label{p13:exact:activities} \end{equation} The nonnegativity in the definition of \(K_3^+\) will be proved in Proposition~\ref{p13:activity:kappa3}. The exact second and third factorial cumulants satisfy \begin{align} \alpha_N&=2\sum_{|e|=2}c(e)-\sum_xp_x^2, \label{p13:exact:alpha}\\ \kappa_{3,N} &=6\sum_{|T|=3}c(T) -6\sum_{e=\{x,y\}}(p_x+p_y)c(e) +2\sum_xp_x^3. \label{p13:exact:kappa3} \end{align} These identities follow by expanding the first three factorial moments. They display the repeated-index terms that an independent Bernoulli reference will retain exactly. \paragraph{Fixed sources on effective spaces.} The fixed-source estimates stated in Propositions~\ref{p1:fixed:targets} and~\ref{p1:fixed:propagation} apply on \(\mathcal I_N\), with the new conditional rank \[ \delta_z=\dim\bigl(S_z\cap\textstyle\sum_{x\in I}S_x\bigr) \] and weight \(p_z2^{\delta_z}\). To see this, choose in each even-selection space a linear complement to its local kernel. The valuation map is injective on that complement and maps it onto \(S_x\). Every label regularization, prime row, and disjoint square equation in the proofs is applied to this linear realization. Restriction to a complement restricts the allowed binary selections within the same physical coordinates; it does not enlarge any prime support. The factor \(p_z\) is at most \(2p\), and the number of sources and targets is fixed in those propositions. Thus their constants change by a fixed factor. The same reasoning applies to homogeneous prescriptions. In particular, for fixed \(s,m\) and \(\varepsilon>0\), \begin{equation} \sum_{\substack{T\subseteq\mathcal I_N\setminus I\\|T|=m}} q(I\cup T)\ll_{s,m,C,\varepsilon}N^\varepsilon q(I) \quad (|I|=s), \label{p13:exact:propagation} \end{equation} and the analogous homogeneous bound holds with \(q\) replaced by \(w\). If \(q(I)=0\), every summand on the left is zero. No estimate here is asserted for \(s\) or \(m\) growing with \(N\). \subsection{Affine structure and centered coefficients} \label{p1:alg:section} The arithmetic estimates naturally count relations, whereas the approximation problem concerns a signed difference of laws. We introduce the exact identities linking these two objects. They will be used with a stopping rule, so that large connected families outside the required nullity range never need to be summed in absolute value. Throughout this section $I\subseteq\mathcal I_N$, and $\rho_{\mathrm{rel}}(I)$ is the effective relative nullity. Write $b_x=\delta_0+p_x\Delta$. The identities below hold for heterogeneous marginals. On the regular set $G_N$, one has $p_x=p$ and $d_x=L$. The same identities therefore cover both regular and irregular windows. \subsubsection{An exact expansion around independent Bernoulli variables} Set \[ Y_x=J_x-p_x,\qquad c(I)=\E\prod_{x\in I}Y_x, \qquad c(\varnothing)=1. \] \begin{proposition}[Centered Bernoulli expansion] \label{p1:alg:centred} For every \(N\), \begin{equation} \Law(Z_N)=\sum_{I\subseteq\mathcal I_N} c(I)\Delta^{*|I|}*\mathop{*}_{x\notin I}b_x . \label{p1:alg:centred-identity} \end{equation} The complementary convolution products are deterministic probabilities. The terms of sizes zero and two are \[ \mathop{*}_{x}b_x+ \sum_{|e|=2}c(e)\Delta^{*2}*\mathop{*}_{x\notin e}b_x. \] Terms of size one vanish. \end{proposition} \begin{proof} For an indicator, \(\delta_{J_x}=b_x+Y_x\Delta\). Expand the finite convolution product over \(\mathcal I_N\) and take expectations. No independence assumption is used in this identity. \end{proof} Each \(s\in S_x\) is a square-class vector. Let \(\chi_s\) be its character on the prime signs, and let \(\eta_x(s)\in\{\pm1\}\) be the affine sign prescribed by the run-start event. The local prescription is consistent on \(\mathcal I_N\), so its affine character \(\eta_x\) is well defined on the effective space. Thus \[ J_x=p_x\sum_{s\in S_x}\eta_x(s)\chi_s,\qquad \eta_x(0)=1. \] \begin{lemma}[Full-support Fourier formula] \label{p1:alg:fourier} For nonempty \(I\subseteq\mathcal I_N\), \begin{equation} c(I)=\Bigl(\prod_{x\in I}p_x\Bigr) \sum_{\substack{s_x\in S_x\setminus\{0\}\ (x\in I)\\ \sum_xs_x=0}} \prod_{x\in I}\eta_x(s_x). \label{p1:alg:fourier-identity} \end{equation} Consequently, \[ |c(I)|\le\Bigl(\prod_{x\in I}p_x\Bigr)(2^{\rho_{\mathrm{rel}}(I)}-1) \le\Bigl(\prod_{x\in I}p_x\Bigr)2^{\rho_{\mathrm{rel}}(I)}. \] If \(c(I)\ne0\), there is a relative relation meeting every block of \(I\). \end{lemma} \begin{proof} Subtracting \(p_x\) removes the zero character from each factor. The expectation of a character product is zero unless the sum of its indices vanishes. The kernel of \(\bigoplus_{x\in I}S_x\to\sum_{x\in I}S_x\) has \(2^{\rho_{\mathrm{rel}}(I)}\) elements. The displayed sum omits its zero element and possibly others. Every affine coefficient has modulus one. \end{proof} We extend the same formulas for $q(I)$, $c(I)$ and relative nullity to arbitrary original window families when stating the next lemma. \begin{lemma}[A common homogeneous envelope] \label{p2:homogeneous:envelope} For a nonempty family $I\subseteq I_N$ of original windows, set \[ q_0(I)=2^{-\dim(\sum_{x\in I}S_x)},\qquad w(I)=\Bigl(\prod_{x\in I}p_x\Bigr)2^{\rho_{\mathrm{rel}}(I)}. \] Then \begin{equation} q_0(I)=2^{\rho_{\mathrm{rel}}(I)}\prod_{x\in I}2^{-d_x}, \qquad w(I),\ q(I),\ |c(I)|\le q_0(I). \label{p2:homogeneous:domination} \end{equation} These assertions allow locally inconsistent start prescriptions and zero marginals. In the critical band, for all sufficiently large $N$, \begin{equation} q_0(I)\le 2^{|I|}p^{|I|}2^{\rho_{\mathrm{rel}}(I)}. \label{p2:homogeneous:nominal} \end{equation} \end{lemma} \begin{proof} Rank--nullity gives the displayed factorization. A consistent affine prescription has the same probability as the corresponding homogeneous system; an inconsistent one has probability zero. Thus $p_x\le2^{-d_x}$ and $q(I)\le q_0(I)$, proving the assertions for $w(I)$ and $q(I)$. If some $p_x=0$, then $J_x=0$ almost surely and $c(I)=0$. Otherwise the full-support Fourier formula in Lemma~\ref{p1:alg:fourier} gives $|c(I)|\le w(I)$. Finally, local rigidity gives $d_x\ge L-1$ uniformly on $I_N$ in the critical band, so $2^{-d_x}\le2p$ and the last inequality follows. \end{proof} The homogeneous quantity is an actual probability: it prescribes zero for every relative bit. It can be positive even when an original start is impossible. Its finite domination is valid at every cardinality; absorbing the factor $2^{|I|}$ into a constant requires a fixed cardinality. \subsubsection{The nullity that survives affine compatibility} Define \begin{equation} \rho_*(I)=\max\{\rho_{\mathrm{rel}}(A):A\subseteq I,\ q(A)>0\}, \qquad q(\varnothing)=1,\quad\rho_{\mathrm{rel}}(\varnothing)=0 . \label{p1:alg:compatible-nullity} \end{equation} This parameter records the largest rank loss in a compatible subfamily. It can be smaller than the nullity of the whole family. \begin{lemma}[Compatible-nullity majorant] \label{p1:alg:compatible} For every \(I\subseteq\mathcal I_N\), \begin{equation} |c(I)|\le 2^{|I|}\Bigl(\prod_{x\in I}p_x\Bigr)2^{\rho_*(I)},\qquad \rho_*(I)\le\rho_{\mathrm{rel}}(I). \label{p1:alg:compatible-bound} \end{equation} Equality holds in the second inequality when \(q(I)>0\). \end{lemma} \begin{proof} The exact inclusion-exclusion formula is \[ c(I)=\sum_{A\subseteq I}(-1)^{|I|-|A|}q(A) \prod_{x\in I\setminus A}p_x. \] A compatible term has absolute value \(\bigl(\prod_{x\in I}p_x\bigr)2^{\rho_{\mathrm{rel}}(A)}\), and an incompatible term is zero. There are at most \(2^{|I|}\) terms. Nullity increases under inclusion, because adding the block \(S_x\) increases the total rank by at most \(d_x\). This proves the other assertions. \end{proof} \subsubsection{Canonical direct components} A partition \(\mathcal P\) of \(I\) is called direct if the spaces \(V_C=\sum_{x\in C}S_x\), \(C\in\mathcal P\), are in direct sum. Its finest such partition is the canonical direct decomposition of \(I\). We call a nonempty family \emph{connected} if this partition has only one part. \begin{lemma}[Existence, uniqueness, and factorization] \label{p1:alg:components} Every finite family \(I\) has a unique finest direct partition \(\mathcal P(I)\). For it, \begin{equation} q(I)=\prod_{C\in\mathcal P(I)}q(C),\qquad c(I)=\prod_{C\in\mathcal P(I)}c(C),\qquad \rho_{\mathrm{rel}}(I)=\sum_{C\in\mathcal P(I)}\rho_{\mathrm{rel}}(C). \label{p1:alg:factorization} \end{equation} \end{lemma} \begin{proof} If \(\mathcal P,\mathcal Q\) are direct partitions, their nonempty intersections form a direct partition. Indeed, inside each \(P\in\mathcal P\), the spaces \(\sum_{x\in P\cap Q}S_x\) lie in the mutually direct spaces \(V_Q\), and sum to \(V_P\). The direct sum over \(P\) proves the claim. Intersecting the finitely many direct partitions gives the unique finest one. Characters indexed by mutually direct subspaces of the finite prime-sign vector space are independent uniform character families. Events, and centered variables, depending on different \(V_C\) are therefore independent. This proves both probability identities; rank additivity proves the last one. \end{proof} Connectedness here is defined by all relative relations. A minimal dependent family of original blocks is primitive, but primitive block supports alone do not determine the canonical components: an additional relation can join a third block to an already dependent pair. Exact pair compression will retain this possibility. \section{Pair and triple scales} \label{p2:scales} The first deviation from Poisson is determined by rational coincidences between two separated windows. We keep the complete finite coefficient: its lower-order terms matter before the next channel becomes visible. The counting estimates behind this calculation are proved in Sections~\ref{p1:arith:appendix} and~\ref{p1:arith:section}. For the pair profiles write $\rho(I)=|I|L-\dim\sum_{x\in I}S_x$ and let $\eta(I)$ indicate affine compatibility, so that $q(I)=p^{|I|}\eta(I)2^{\rho(I)}$. This raw nullity includes local kernels and is distinct from the effective relative nullity used in Section~\ref{p13:exact:section}. \subsection{The two leading rational channels} An exact unit of the reduced channel $a:b$, $aL}(2^{\rho(x,y)}-1) &=\frac N3\{U_{2,3}(L)+U_{3,4}(L)\} +O_{C,\varepsilon}(N^{6/5+\varepsilon}),\label{p1:arith:positiveprofile}\\ \sum_{|x-y|>L}(\eta(x,y)2^{\rho(x,y)}-1) &=\frac N3\{S_{2,3}(L)+S_{3,4}(L)\} +O_{C,\varepsilon}(N^{6/5+\varepsilon}).\label{p1:arith:signedprofile} \end{align} Writing $\mu_N=\E Z_N$ and $\alpha_N=\operatorname{Var}(Z_N)-\mu_N$, one has \begin{equation}\label{p1:arith:cumulantexpansion} \alpha_N=\frac{\lambda_N^2}{3N}\{S_{2,3}(L)+S_{3,4}(L)\} +O_{C,\varepsilon}(N^{-4/5+\varepsilon}). \end{equation} Thus, with $r_3=L\bmod3$, \begin{equation}\label{p1:arith:leadingphase} \alpha_N\sim \frac{r_3+2}{18\,2^{r_3/3}}\lambda_N^{5/3}L N^{-2/3}. \end{equation} After the entire first-channel term is removed, the next term is \begin{equation}\label{p1:arith:secondphase} \alpha_N-\frac{\lambda_N^2}{3N}S_{2,3}(L) \sim\frac{r_4+3}{24\,2^{r_4/4}}\lambda_N^{7/4}L N^{-3/4}, \qquad r_4=L\bmod4. \end{equation} \end{theorem} The exact-unit calculation gives the displayed channel sums. Both orientations of each of the channels $2:3$ and $3:4$ have a translation interval of total length $N/3+O(B)$. Private prime coordinates identify their rational kernels; the exceptional normalizations and all remaining channels have the stated residual mass. Section~\ref{p14:comp:complete-profile} gives the complete proof, including the endpoint, contact, and mean-transfer estimates. The subtraction in \eqref{p1:arith:secondphase} must retain the whole finite expression $S_{2,3}(L)$. Its own term of size $N^{-2/3}$ without the factor $L$ is larger than $LN^{-3/4}$. The two exact channel polynomials together depend on $L\bmod12$. \subsection{A computable two-channel reference} An entirely explicit reference has a different optimal rate. Retain the complete \(2:3\) and \(3:4\) channel sums, and set \begin{equation} \widehat\alpha_N=\frac{\lambda_N^2}{3N} \{S_{2,3}(L)+S_{3,4}(L)\}. \label{p1:eq:alpha-explicit} \end{equation} The sums have been defined above and are evaluated by \eqref{p1:arith:S23}--\eqref{p1:arith:S34}. \begin{corollary}[An explicit probability reference] \label{p1:cor:explicit} For every \(\varepsilon>0\), uniformly in the critical band, \begin{equation} \left\|\Law(Z_N)-\Pi_{\lambda_N} -\frac{\widehat\alpha_N}{2} \Pi_{\lambda_N}*\Delta^{*2}\right\|_1 \ll_{C,\varepsilon}N^{-4/5+\varepsilon}. \label{p1:eq:explicit-law} \end{equation} For all sufficiently large \(N\), the reference \(Q_{\lambda_N,\widehat\alpha_N}\) is a probability measure. Its mass at \(k\ge0\) is \begin{equation} e^{-\lambda_N}\frac{\lambda_N^k}{k!} \left[ 1+\frac{\widehat\alpha_N}{2} \left\{\frac{k(k-1)}{\lambda_N^2} -\frac{2k}{\lambda_N}+1\right\} \right]. \label{p1:eq:explicit-masses} \end{equation} All parameters in this reference are determined by \(N,L\). \end{corollary} \begin{proof} Theorem~\ref{p1:arith:pairs} gives \(\alpha_N-\widehat\alpha_N=O_{C,\varepsilon}(N^{-4/5+\varepsilon})\). Since \(\|\Pi_{\lambda_N}*\Delta^{*2}\|_1\le4\), replacing the coefficient costs at most \(2|\alpha_N-\widehat\alpha_N|\). Apply Theorem~\ref{p1:thm:main}. The replacement of \(\mu_N\) by \(\lambda_N\) costs \(O(N^{-1+\varepsilon})\), which is absorbed here. The finite formulas give \(\widehat\alpha_N\sim\alpha_N>0\) and \(\widehat\alpha_N\to0\), uniformly in the band. The positivity criterion in Lemma~\ref{p1:close:positivity} therefore applies, since \(\lambda_N\) remains in a compact subinterval of \((0,\infty)\). \end{proof} The exponent \(4/5\) in \eqref{p1:eq:explicit-law} is optimal for this two-channel reference. Its leading error is half the height-five contribution times \(\|\Pi_{\lambda_N}*\Delta^{*2}\|_1\), as proved in Corollary~\ref{p13:channels:height-five}. The explicit parameters do not supply a numerical value of the constant in the remainder or a numerical threshold for \(N\). The same two-channel coefficient also gives a positive Hermite reference; see Corollary~\ref{p1:cor:compound}. \paragraph{The probability of no retained start.} Evaluating \eqref{p1:eq:explicit-law} at zero gives the immediate consequence \begin{equation} \P(Z_N=0)=e^{-\lambda_N} \left(1+\frac{\widehat\alpha_N}{2}\right) +O_{C,\varepsilon}(N^{-4/5+\varepsilon}), \label{p1:eq:no-start} \end{equation} because \((\Pi_{\lambda_N}*\Delta^{*2})\{0\}=e^{-\lambda_N}\). The positive correction shows an effect of the arithmetic pairs that is absent from the mean alone. Here \(Z_N=0\) means that no retained run start lies in \(I_N\). A run starting before \(N\) may still intersect the interval, and a run starting in \(I_N\) may continue beyond \(2N\). The exact-parameter refinement is Corollary~\ref{p13:optimal:zero}. \subsection{Why triple interactions are smaller} The third factorial cumulant is \begin{equation} \kappa_{3,N}=\E(Z_N)_3-3\mu_N\E(Z_N)_2+2\mu_N^3. \label{p13:third:definition} \end{equation} It is distinct from the ordinary third cumulant, which equals the mean for a Poisson variable. \begin{theorem}[Optimal power of the third factorial cumulant] \label{p13:third:main} For each fixed critical band $C>0$, there is a constant $c_C>0$ such that, for all sufficiently large $N$, \begin{equation} c_C N^{-41/30}\leq\kappa_{3,N} \ll_{C,\varepsilon}N^{-41/30+\varepsilon} \qquad(\varepsilon>0). \label{p13:third:two-sided} \end{equation} Both bounds are unconditional. Thus $41/30$ is the optimal power exponent for this factorial cumulant. The theorem does not assert an asymptotic equivalent or identify a logarithmic factor for the full cumulant. \end{theorem} The complete proof is given in Section~\ref{p13:third:section}. It bounds all admissible triples, with the exact marginals retained. The following form is the input to the reconstruction. \begin{corollary}[Absolute third-order activity] \label{p13:third:activity} For every $\varepsilon>0$, \begin{equation} A_3:=\sum_{|T|=3}|c(T)| \ll_{C,\varepsilon}N^{-41/30+\varepsilon}. \label{p13:third:absolute} \end{equation} \end{corollary} \begin{proof} Equation~\eqref{p13:activity:kappa-comparison} gives $A_3=K_3^++O(N^{-3/2+\varepsilon'})$. Use the upper bound of Theorem~\ref{p13:third:main} and reduce the internal margin. \end{proof} The geometric step in the upper bound changes the scale at which points are counted. Write $d_{xy}=\dim(S_x\cap S_y)$ and suppose $d_{xy}>((2/3)\log2+\delta)B$ for a fixed $\delta>0$. Short-relation extraction yields a space of dimension $\gg_\delta B$ and a common curve through $(x,y)$, of degree $D\le B^{a+o(1)}$ for some fixed $a<1/2$. For the relations in a subspace of codimension at most one, the associated products are squares in the curve's function field. At a selected noncritical offset $u$, the simple zeros of $X+u$ must therefore be matched by zeros of factors $Y+v$ from the other window. After $O(D^2)$ exceptional offsets are removed, these fibres give $\gg_\delta B$ integer points in $[-B,B]^2$. Their coordinates are logarithmic-size offsets, although the original centres have size $N$. For $D\ge2$, the degree-uniform point bound gives only $o(B)$ points, a contradiction. For $D=1$, matching the roots gives a small rational slope and an alignment. Thus nonaligned pairs obey the rank bound of Theorem~\ref{p13:rig:nonaligned}; the full argument is in Section~\ref{p13:rig:section}. The configuration $10t,12t,15t$ explains the third-order power. Multiplication by $6,5,4$ gives a common center $60t$ and offset steps $6,5,4$. The canonical relative nullity is \[ \rho_L=\left\lfloor\frac{L-1}{6}\right\rfloor+ \left\lfloor\frac{L-1}{3}\right\rfloor+ \left\lfloor\frac{L-1}{5}\right\rfloor- \left\lfloor\frac{L-1}{15}\right\rfloor =\frac{19}{30}L+O(1). \] A compatible triple with no additional square relations therefore has weight $N^{-71/30+o(1)}$. For a concrete phase, take $L=30m$ and $N=2^L$. Then \[ \begin{aligned} \rho_L&=(5m-1)+(10m-1)+(6m-1)-(2m-1)=19m-2,\\ q(\{10t,12t,15t\})&=p^3 2^{\rho_L}=\frac14N^{-71/30}. \end{aligned} \] There are $N/30+O(1)$ translations in the dyadic interval. Proposition~\ref{p13:third:lower} shows that all but $N^{1/2+o(1)}$ of them have private prime coordinates and compatible predecessor signs. On this population, the largest pair exponent subtracted in the centered coefficient is $10m-1$, so $c(T)=q(T)(1+o(1))$. This family thus contributes $(1/120+o(1))N^{-41/30}$ to $K_3^+$; the coefficient is specific to this family, not to the full cumulant. The upper bound also needs pair rigidity and an exhaustive partition of the other triples. It is not a consequence of this example alone. The distinction between pairs and triples is decisive: $\alpha_N^2=N^{-4/3+o(1)}$, whereas $A_3\ll N^{-41/30+\varepsilon}$ and $41/30-4/3=1/30$. The reconstruction will retain the first pair term and allow its square in the error. The fourth-order test will prove that this remaining scale cannot be improved in power for the quadratic reference. \section{Stopping the expansion at its first excluded prefix} \label{p1:stop:section} The leading correction comes from retaining one dependent pair. An expansion of the entire counting law would also retain every possible interaction after that pair. Taking absolute values at that stage can be expensive, even when the interactions that first leave the retained class have small total probability. We therefore stop each term at its first excluded prefix and leave all subsequent indicators inside a probability kernel. The first identity is finite and does not require independence. Its exact-marginal application will reduce the approximation problem to families whose effective nullity is at most $2L$. The first-excluded-prefix identity is a specialization of Zeilberger's abstract lace expansion \cite[formula~(*)]{zeilberger1997}. The broader inclusion--exclusion context includes the abstract-tube identities \cite{naimanwynn1997,dohmen1999}. We give the prefix proof and its convolution form; the model-specific estimates concern the retained arithmetic class and its full boundary. \subsection{An identity for an arbitrary hereditary class} Let $G$ be a finite totally ordered set. A family $\mathscr D\subseteq 2^G$ is hereditary if $I\in\mathscr D$ and $J\subseteq I$ imply $J\in\mathscr D$. Assume throughout that $\varnothing\in\mathscr D$. Its \emph{ordered boundary} is \begin{equation} \partial\mathscr D =\{I\subseteq G:I\notin\mathscr D,\ I\setminus\{\max I\}\in\mathscr D\}. \label{p1:stop:boundary} \end{equation} This definition records the first excluded prefix in the prescribed order. It is not the family of all inclusion-minimal excluded sets: an element of $\partial\mathscr D$ can contain a smaller excluded subset which is not a prefix. \begin{lemma}[Expansion at the first excluded prefix] \label{p1:stop:polynomial} Let $(u_x)_{x\in G}$ belong to a commutative unital algebra. Then, as a polynomial identity, \begin{equation} \prod_{x\in G}(1+t u_x) =\sum_{I\in\mathscr D}t^{|I|}\prod_{x\in I}u_x +\sum_{I\in\partial\mathscr D} t^{|I|}\prod_{x\in I}u_x \prod_{\substack{y\in G\\y>\max I}}(1+t u_y). \label{p1:stop:polynomial-identity} \end{equation} \end{lemma} \begin{proof} Expand the product on the left and index its monomials by $S\subseteq G$. If $S\notin\mathscr D$, read the elements of $S$ in their prescribed order. There is a unique first prefix $I\notin\mathscr D$. Its preceding prefix belongs to $\mathscr D$, so $I\in\partial\mathscr D$, and every element of $S\setminus I$ is larger than $\max I$. Conversely, take $I\in\partial\mathscr D$ and $T\subseteq\{y\in G:y>\max I\}$. Every proper prefix of $I$ belongs to $\mathscr D$, by heredity. The set $I\cup T$ is excluded, since it contains $I$, and its first excluded prefix is $I$. Thus the pairs $(I,T)$ partition the monomials indexed by $S\notin\mathscr D$. Summing these monomials proves the identity. \end{proof} To identify the specialization explicitly, define \(\ell(S)=S\) for \(S\in\mathscr D\), and otherwise let \(\ell(S)\) be its first excluded prefix. This is a lace map: \(\ell(S)\subseteq S\); \(\ell(T)=\ell(S)\) whenever \(\ell(S)\subseteq T\subseteq S\); and equal images of two sets remain unchanged under their union. Its fixed points are \(\mathscr D\cup\partial\mathscr D\). The compatible-extension set is empty for \(I\in\mathscr D\), and equals \(\{y\in G:y>\max I\}\) for \(I\in\partial\mathscr D\). Zeilberger's formula~(*) with \(Y_x=t u_x\) is therefore exactly \eqref{p1:stop:polynomial-identity}. For indicators $(J_x)_{x\in G}$, write \[ Z=\sum_{x\in G}J_x,\qquad q(I)=\E\prod_{x\in I}J_x,\qquad q(\varnothing)=1. \] Let $\Delta=\delta_1-\delta_0$, and use the convention $\Delta^{*0}=\delta_0$. The measure retained by $\mathscr D$ is \begin{equation} T_{\mathscr D} =\sum_{I\in\mathscr D}q(I)\Delta^{*|I|}. \label{p1:stop:retained} \end{equation} It has total mass one, although its individual masses need not be nonnegative. \begin{proposition}[A positive remainder kernel] \label{p1:stop:measure} For arbitrary dependence among the indicators, \begin{align} \Law Z-T_{\mathscr D} &=\sum_{I\in\partial\mathscr D}\Delta^{*|I|}*\nu_I, \label{p1:stop:measure-identity}\\ \nu_I &=\E\left[ \left(\prod_{x\in I}J_x\right) \delta_{\sum_{y>\max I}J_y} \right]. \notag \end{align} The measure $\nu_I$ is positive and has mass $q(I)$. Consequently, \begin{equation} \|\Law Z-T_{\mathscr D}\|_1 \le \sum_{I\in\partial\mathscr D}2^{|I|}q(I). \label{p1:stop:norm} \end{equation} \end{proposition} \begin{proof} Apply Lemma~\ref{p1:stop:polynomial} in the convolution algebra of finite signed measures, with unit $\delta_0$ and $t=\Delta$. For an indicator, \[ \delta_0+J_x\Delta=\delta_{J_x}, \qquad \mathop{*}_{y>\max I}(\delta_0+J_y\Delta) =\delta_{\sum_{y>\max I}J_y}. \] The expectation of the complete product is $\Law Z$. Taking expectations in \eqref{p1:stop:polynomial-identity} therefore gives \eqref{p1:stop:measure-identity}. Positivity and the mass formula for $\nu_I$ follow directly from its definition. Finally, $\|\Delta^{*|I|}\|_1=2^{|I|}$ and the convolution norm is submultiplicative, which proves \eqref{p1:stop:norm}. \end{proof} If $q(I)>0$, the normalized measure $\nu_I/q(I)$ is the law of the suffix count conditioned on the activity of every member of $I$. This conditional law can differ substantially from its unconditional version. The proof uses only its total mass, so no independence between the prefix and suffix is required. In particular, it incurs no factor exponential in the number of suffix indicators. The use of the ordinary $\ell^1$ norm matters: an exponential weight would also charge the suffix count. \begin{example}[A suffix that remains fully dependent] \label{p1:stop:example} Take $G=\{1,2,3\}$ in its natural order, and retain only the empty set and the singletons. Then \[ \partial\mathscr D=\{\{1,2\},\{1,3\},\{2,3\}\}, \qquad T_{\mathscr D}=\delta_0+(q_1+q_2+q_3)\Delta, \] and the exact remainder is \[ \Law Z-T_{\mathscr D} =\Delta^{*2}* \left(\E[J_1J_2\delta_{J_3}] +(q_{13}+q_{23})\delta_0\right). \] For $J_1=J_2=J_3=H$, where $\P(H=1)=\theta$, the conditional suffix following $\{1,2\}$ is identically one. The formula becomes \[ \Law(3H)-(\delta_0+3\theta\Delta) =\theta\Delta^{*2}*(2\delta_0+\delta_1). \] The remainder has norm $6\theta$, while \eqref{p1:stop:norm} gives the valid bound $12\theta$. The mechanism is the positive kernel of mass $3\theta$, not a comparison of the conditional suffix with an independent one. This finite example explains the stopping identity; the remaining task in our model is to prove that the appropriate boundary has small weighted mass. \end{example} \section{Reconstructing the law with every marginal retained} \label{p13:transfer:section} At a precision finer than $N^{-1}$, removing the exceptional windows is too expensive. We retain their exact marginals instead. This also keeps the repeated occurrence of a single Bernoulli variable inside the reference law, where it can be evaluated exactly. The remaining task is to control interactions between distinct windows and to sum the expansion without applying a fixed-order estimate at an order growing with $N$. We use the full admissible set $\mathcal I_N=\{x\in I_N:p_x>0\}$, the exact marginals $p_x=\E J_x$, the centered coefficients $c(I)=\E\prod_{x\in I}(J_x-p_x)$, and the relative nullity defined in Section~\ref{p13:exact:section}. Throughout this section we abbreviate $\rho_{\mathrm{rel}}(I)$ to $\rho(I)$; this convention uses the effective local ranks and does not add the local rank defects again. Put \[ A_2=\sum_{|e|=2}|c(e)|,\qquad A_3=\sum_{\substack{|T|=3\\T\text{ connected}}}|c(T)|, \qquad p_* =\max_{x\in\mathcal I_N}p_x. \] The marginal and pair estimates give $p_*\leq2p\ll_C N^{-1}$, $\mu=\sum_xp_x=O_C(1)$, and $A_2\ll_{C,\varepsilon}N^{-2/3+\varepsilon}$. The connection between the arithmetic estimates and the distribution is the following bound: \begin{equation} \|\Law(Z_N)-Q_{\mu_N,\alpha_N}\|_1 \ll_{C,\varepsilon}N^\varepsilon A_3+N^{-4/3+\varepsilon}. \label{p13:transfer:announced-bound} \end{equation} We prove it in Theorem~\ref{p13:transfer:reconstruction}. The two terms have different origins: the first records a genuinely connected three-window interaction, while the second permits two pair interactions. \paragraph{Three errors in the reconstruction.} For $e=\{x,y\}$ write $S_e=S_x+S_y$. The retained class \(\mathscr D_2\) consists of direct families and families \(e\cup U\) with a dependent pair \(e\) for which \(S_e\) and the spaces \((S_z)_{z\in U}\) form a direct sum. The dependent pair is then unique. Write \(T_{\mathscr D_2}=\sum_{I\in\mathscr D_2}q(I)\Delta^{*|I|}\). With the covariance-corrected Bernoulli measure \(B_2\) defined in \eqref{p13:transfer:B2}, the proof separates \[ \begin{aligned} \|\Law(Z_N)-Q_{\mu_N,\alpha_N}\|_1 &\le \|\Law(Z_N)-T_{\mathscr D_2}\|_1 +\|T_{\mathscr D_2}-B_2\|_1\\ &\qquad +\|B_2-Q_{\mu_N,\alpha_N}\|_1. \end{aligned} \] The first term is the stopping boundary, with the unexpanded suffix kept in a positive conditional measure. The second consists of the independent and pair deficits in \eqref{p13:transfer:retained-difference}; the covariance signs are retained before taking norms. In particular, an incompatible pair is allowed in \(\mathscr D_2\): its joint probability is zero, but its covariance is still present in \(B_2\). The last term costs \(O(N^{-5/3+\varepsilon})\) by Lemma~\ref{p13:transfer:conversion}; that rate concerns the full measure \(B_2\), including every covariance. Theorem~\ref{p13:transfer:finite} assembles the three estimates, using fixed orders and the separately uniform tail of Proposition~\ref{p13:transfer:uniform-tail} for the boundary. \subsection{The exact Bernoulli reference and its conversion} For any finite family of Bernoulli variables, without an independence assumption, define $b_i=\delta_0+p_i\Delta$ and \begin{equation} B_2=\mathop{*}_{i\in\mathcal I_N}b_i +\sum_{|e|=2}c(e)\Delta^{*2}* \mathop{*}_{i\notin e}b_i. \label{p13:transfer:B2} \end{equation} This is a signed measure. Its first term is the law of a sum of independent Bernoulli variables with the exact individual marginals; its second term restores every covariance once. \begin{lemma}[Conversion with heterogeneous marginals] \label{p13:transfer:conversion} Let $s_j=\sum_i p_i^j$, $C_2=\sum_{|e|=2}c(e)$, and $\alpha=2C_2-s_2$. If $\mu\leq M$ and $p_*\leq1/4$, then \begin{equation} \|B_2-Q_{\mu,\alpha}\|_1 \leq C_M\{s_3+s_2^2+A_2(s_2+p_*)\}. \label{p13:transfer:conversion-bound} \end{equation} For the present model this is $O_{C,\varepsilon}(N^{-5/3+\varepsilon})$. \end{lemma} \begin{proof} In the convolution algebra of finite signed measures, the series for $\log_*(\delta_0+p_i\Delta)$ converges absolutely, since $\|p_i\Delta\|_1=2p_i\leq1/2$. Summing it gives \[ \log_*\left(\mathop{*}_ib_i\right) =\mu\Delta-\frac{s_2}{2}\Delta^{*2}+E, \qquad \|E\|_1\leq\sum_{j\geq3}\frac{2^js_j}{j} \leq\frac{16}{3}s_3. \] Exponentiation and $\exp_*(\mu\Delta)=\Pi_\mu$ imply \[ \left\|\mathop{*}_ib_i-\Pi_\mu +\frac{s_2}{2}\Pi_\mu*\Delta^{*2}\right\|_1 \leq C_M(s_3+s_2^2). \] The product omitting a pair $e=\{i,j\}$ differs from $\Pi_{\mu-p_i-p_j}$ by $O_M(s_2)$ in norm, and \[ \|\Pi_{\mu-p_i-p_j}-\Pi_\mu\|_1\leq2(p_i+p_j). \] Replacing each such product in \eqref{p13:transfer:B2} by $\Pi_\mu$ therefore costs $O_M(A_2(s_2+p_*))$. The exact identity $\Var Z-\E Z=2C_2-s_2$ proves the lemma. Here $s_2\leq p_*\mu=O_C(N^{-1})$ and $s_3\leq p_*^2\mu=O_C(N^{-2})$, giving the stated application. \end{proof} \paragraph{Algebraic form of the conversion.} The estimate in Lemma~\ref{p13:transfer:conversion} also holds for arbitrary real pair weights $c(e)$ in \eqref{p13:transfer:B2}, with $C_2=\sum_e c(e)$, $A_2=\sum_e|c(e)|$ and $\alpha=2C_2-s_2$. The mean and marginal bounds are unchanged. The two approximations in its proof and the triangle inequality give this algebraic form; only the interpretation $\alpha=\Var Z-\E Z$ uses actual covariances. \subsection{Arithmetic estimates used in the assembly} The proof below uses three estimates proved in Section~\ref{p2:technical-transfer} of the companion. They separate fixed geometry from the growing number of windows. For each fixed $r\ge4$, the total connected activity is \[ \sum_{\substack{|I|=r\\I\text{ connected}}}|c(I)| \ll_{r,C,\varepsilon}N^\varepsilon A_3+N^{-4/3+\varepsilon}, \] and its incompatible part is $O(N^{-3/2+\varepsilon})$. The independent deficit of all pair extensions is $O(N^{-3/2+\varepsilon})$ at every cardinality. Finally, the connected activity beyond the fixed size $256$, up to the logarithmic boundary cutoff and with compatible nullity at most $2L$, is $O(N^{-4/3+\varepsilon})$; the corresponding primitive tail has the same bound without a cardinality cutoff. These are Proposition~\ref{p13:transfer:fixed-activities}, Proposition~\ref{p13:transfer:pair-deficit} and Proposition~\ref{p13:transfer:uniform-tail}, respectively. The last estimate has a separate proof. It uses small-kernel counts, moderate congruence certificates and a graph with small incoming and outgoing degrees. It does not let the order in the first estimate grow with $N$. The choice $256$ is sufficient for these estimates; no optimality is attached to that cutoff. \paragraph{Dependencies of the reconstruction.} The table separates the arithmetic inputs by the quantity they control. The fixed-order estimates and the uniform tails enter different parts of the stopped expansion; the lower bound uses the independently proved moment estimate. \begin{center} \small \begin{tabularx}{\linewidth}{@{}>{\raggedright\arraybackslash}p{0.26\linewidth}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{0.24\linewidth}@{}} \toprule Quantity & Main inputs in the companion & Use in the article\\ \midrule Complete pair coefficient & Sections~\ref{p1:arith:appendix}--\ref{p1:arith:supplement} & Theorem~\ref{p1:arith:pairs}\\[5pt] Rank of a nonaligned pair & \ref{p13:rig:short-basis}--\ref{p13:rig:nonaligned}; norm estimate in \ref{p13:rig:two-witnesses} & Theorem~\ref{p13:third:main}\\[5pt] Absolute triple activity & Section~\ref{p13:activity:section}; \ref{p13:third:rational-rank}--\ref{p13:third:lower} & Corollary~\ref{p13:third:activity}\\[5pt] Connected activity at fixed orders & \ref{p13:transfer:dense-targets}, \ref{p13:transfer:primitive}, \ref{p13:transfer:fixed-deficit}, \ref{p13:transfer:fixed-activities} & Theorem~\ref{p13:transfer:reconstruction}\\[5pt] Pair deficit at every cardinality & \ref{p13:transfer:independent-deficit}--\ref{p13:transfer:pair-deficit} & Equation~\eqref{p13:transfer:retained-error}\\[5pt] Uniform tails & Section~\ref{p14:cert:section}; \ref{p13:transfer:independent-deficit}, \ref{p13:transfer:independent-tail}--\ref{p13:transfer:uniform-tail} & Theorems~\ref{p13:transfer:finite}--\ref{p13:transfer:reconstruction}\\[5pt] Quadratic lower bound & Theorem~\ref{p1:fixed:moments}; incompatible part of Proposition~\ref{p13:transfer:fixed-activities} & \ref{p13:lower:quartic}--\ref{p13:lower:optimal}\\ \bottomrule \end{tabularx} \end{center} \subsection{The stopped expansion and its finite input} Let $\mathscr D_2$ consist of the direct families and the families $e\cup U$ for which $e$ is any dependent pair and $S_e,(S_z)_{z\in U}$ are direct. Affine incompatibility of $e$ is allowed. This class is hereditary; every non-direct member contains exactly one dependent pair. For a fixed ordering of $\mathcal I_N$, its boundary contains the first excluded prefixes. A boundary family has relative nullity at most $2L$: its retained predecessor has at most $L$, and adding one block adds at most $L$. In particular every subfamily of a boundary family has compatible nullity at most $2L$. \begin{theorem}[Finite transfer to the complete law] \label{p13:transfer:finite} Fix $1<\beta\leq4/3$. Suppose that, for each of the fixed integers $3\leq r\leq256$ and every $\varepsilon>0$, \begin{equation} \sum_{\substack{|I|=r,\ I\text{ connected}\\\rho_*(I)\leq2L}} |c(I)|\ll_{r,C,\varepsilon}N^{-\beta+\varepsilon}. \label{p13:transfer:finite-hypothesis} \end{equation} Then \begin{equation} \|\Law(Z_N)-Q_{\mu_N,\alpha_N}\|_1 \ll_{C,\varepsilon}N^{-\beta+\varepsilon}. \label{p13:transfer:finite-conclusion} \end{equation} Only these $254$ fixed orders are assumed in \eqref{p13:transfer:finite-hypothesis}; the larger orders are supplied by Proposition~\ref{p13:transfer:uniform-tail}. \end{theorem} \begin{proof} Fix a weight $w>0$. The hypothesis at sizes at most $256$ and \eqref{p13:transfer:connected-tail} give, for $K_N=o(L)$, \begin{equation} T:=\sum_{\substack{I\text{ connected},\ 3\leq|I|\leq K_N\\ \rho_*(I)\leq2L}} w^{|I|}|c(I)|\ll N^{-\beta+\eta}. \label{p13:transfer:connected-range} \end{equation} The finitely many factors $w^r$ at $r\leq256$ only change constants. For a primitive $P$, $\rho(P)\leq L$ and $\prod_{x\in P}p_x\leq|c(P)|$, because $c(P)=q(P)-\prod_Pp_x$ and $q(P)$ is either zero or at least twice the independent product. Together with \eqref{p13:transfer:primitive-tail}, this gives \begin{equation} \sum_{\substack{P\text{ primitive}\\|P|\geq3}} w^{|P|}|c(P)|\ll N^{-\beta+\eta} \label{p13:transfer:all-primitives} \end{equation} over all cardinalities. Put $T_{\mathscr D_2}=\sum_{I\in\mathscr D_2}q(I)\Delta^{*|I|}$. The finite expansion of $B_2$ yields the identity \begin{equation} T_{\mathscr D_2}-B_2 =-\sum_{I\notin\mathscr D_2}\Delta^{*|I|} \left[\prod_{i\in I}p_i+ \sum_{\substack{e\subset I\\|e|=2}} c(e)\prod_{i\in I\setminus e}p_i\right]. \label{p13:transfer:retained-difference} \end{equation} For the independent term, choose a primitive subfamily of a discarded family. If it has at least three members, use \eqref{p13:transfer:all-primitives}; if it is a pair $e$, the complement must be a non-direct extension of $S_e$, and $\prod_{i\in e}p_i\leq|c(e)|$. The pair terms in \eqref{p13:transfer:retained-difference} are bounded by the same all-cardinality pair deficit. Free complements have total weight at most $e^{2\mu}$. Consequently \begin{equation} \|T_{\mathscr D_2}-B_2\|_1\ll N^{-\beta+\eta}. \label{p13:transfer:retained-error} \end{equation} Take $K_N=(5+o(1))\log N/\log\log N$. The exact stopping identity from Section~\ref{p1:stop:section} gives \[ \|\Law(Z_N)-T_{\mathscr D_2}\|_1 \leq\sum_{I\in\partial\mathscr D_2}2^{|I|}q(I). \] For $|I|\leq K_N$, expand $q(I)$ in centered coefficients. The terms of size zero and two are the independent and pair deficits already bounded. Every remaining nonzero centered term has either a connected component of size at least three or at least two pair components. Canonical factorization, \eqref{p13:transfer:connected-range}, and the free independent complements bound their sum by a constant times \[ e^{4A_2+T}-1-4A_2=O(T+A_2^2). \] Here $T=o(1)$ and $A_2^2\ll N^{-4/3+\eta}$; the centered coefficient of a singleton is zero. This is where the linear contribution of one pair is subtracted before taking an absolute bound. On the larger boundary, $q(I)\leq2^{2L}\prod_Ip_x$, so \[ \sum_{\substack{I\in\partial\mathscr D_2\\|I|>K_N}}2^{|I|}q(I) \ll_C N^2\sum_{r>K_N}\frac{(2\mu)^r}{r!}=N^{-3+o(1)}. \] Combining these bounds, \eqref{p13:transfer:retained-error}, and Lemma~\ref{p13:transfer:conversion} proves \eqref{p13:transfer:finite-conclusion}. All internal margins are chosen after the final $\varepsilon$; none is made to vary with $N$. \end{proof} \begin{theorem}[Reconstruction from triple activity] \label{p13:transfer:reconstruction} With the exact mean and second factorial cumulant, \begin{equation} \|\Law(Z_N)-Q_{\mu_N,\alpha_N}\|_1 \ll_{C,\varepsilon}N^\varepsilon A_3+N^{-4/3+\varepsilon}. \label{p13:transfer:reconstruction-bound} \end{equation} In particular the arithmetic triple estimate gives the upper bound $O_{C,\varepsilon}(N^{-4/3+\varepsilon})$ for the full model. \end{theorem} \begin{proof} In the proof of Theorem~\ref{p13:transfer:finite}, replace the finite activity input by Proposition~\ref{p13:transfer:fixed-activities}. At sizes $3,\ldots,256$ its total cost is $N^{\varepsilon'}A_3+N^{-4/3+\varepsilon'}$; the larger sizes have the second cost by Proposition~\ref{p13:transfer:uniform-tail}. The primitive deficits, pair deficits, and boundary expansion obey the same bounds. In the component generator, \[ T\ll N^{\varepsilon'}A_3+N^{-4/3+\varepsilon'},\qquad e^{4A_2+T}-1-4A_2=O(T+A_2^2). \] The established arithmetic estimate makes $T=o(1)$, so this passage loses no new fixed power of $N$. The large boundary costs $N^{-3+o(1)}$ and conversion costs $N^{-5/3+\varepsilon'}$. To make the choices explicit, denote the requested final margin by $\varepsilon_{\mathrm{out}}=\varepsilon$ and choose $0<\tau<\min(\varepsilon_{\mathrm{out}}/4,1/960)$. Invoke each completed fixed-order estimate with sufficiently small fixed input margins that its total displayed loss is at most $\tau$. For example, use the pair estimate with loss $\tau/2$ before squaring $A_2$. Only the fixed orders $3,\ldots,256$ occur in this choice; the larger orders use the uniform certificate bounds, not constants from a fixed-order theorem. With $\varepsilon'=\tau$ the preceding argument proves \eqref{p13:transfer:reconstruction-bound}, since $\tau<\varepsilon_{\mathrm{out}}$. Finally invoke Corollary~\ref{p13:third:activity} with loss $\tau$. The triple term is then at most $N^{-41/30+2\tau}$, and \[ \frac{41}{30}-2\tau-\frac43=\frac1{30}-2\tau>0. \] The conversion and large-boundary exponents exceed $4/3$ by $1/3$ and $5/3$, respectively. These comparisons give the announced upper bound. All accuracy choices, source counts and fixed activity orders used here are fixed before $N$; the growing cardinality cutoff $K_N$ is controlled separately. \end{proof} \paragraph{Power margins.} The strict power savings used in the last stages are recorded below. Each margin compares complete contributions, after coefficient choices, propagation and rank factors have been included. \begin{center} \small \begin{tabularx}{\textwidth}{@{}Xcl@{}} \toprule Comparison & Exact margin & Location\\ \midrule Nonaligned target above a rational pair versus dominant triples & $11/8-41/30=1/120$ & \eqref{p13:third:unaligned-target-cost}\\[4pt] Incompatible triples versus the third factorial cumulant & $3/2-41/30=2/15$ & \eqref{p13:activity:kappa-comparison}\\[4pt] Third activity versus the quadratic remainder & $41/30-4/3=1/30$ & \eqref{p13:transfer:reconstruction-bound}\\[4pt] Low-nullity uniform tail versus the quadratic remainder & $27/20-4/3=1/60$ & \eqref{p13:transfer:low-rank-tail}\\[4pt] Quartic signed error versus $\alpha_N^2$ & $3/2-4/3=1/6$ & \eqref{p13:lower:positive-quartic}\\ \bottomrule \end{tabularx} \end{center} The smallest displayed margin is $1/120$. In contrast, two pair activities have size $A_2^2\ll N^{-4/3+\eta}$ and are retained at the target scale. They have no strict power saving to spend. For a prescribed final accuracy $\varepsilon_0>0$, one may choose $0<\tau<\min(\varepsilon_0/4,1/960)$ and invoke completed auxiliary estimates with total loss at most $\tau$ each. Their own input margins are chosen smaller as needed; in a product, divide the allotted loss between the factors. Thus the same symbol $\varepsilon$ in successive statements never requires the same numerical value at every stage. Rank thresholds and fixed geometric degrees have separate roles. The fixed connected-activity estimates in reconstruction are used only through order $256$; larger cardinalities have the uniform certificate estimate and a factorial boundary bound. \section{Why the quadratic approximation stops at the fourth-order scale} \label{p13:lower:section} The quadratic reference retains the second factorial cumulant linearly. Two independent pair contributions therefore leave a fourth-order term of size $\alpha_N^2$. To turn this observation into a lower bound, one must check that incompatible arithmetic configurations cannot cancel that term. The half-exponent count supplies this missing sign information. The formal term \(\alpha_N^2\Pi_{\mu_N}*\Delta^{*4}/8\) is classical in Poisson--Charlier expansions \cite{bkn2014,cekanaviciusvellaisamy2015}. Lower bounds with a squared second-cumulant contribution are known for locally dependent sums \cite[Theorem~2.1, (2.3)]{cekanaviciusvellaisamy2020}. The issue here is its arithmetic size, the sign of the other fourth-order contributions, and the passage from an unbounded polynomial test to the full norm. Throughout this section $\mu=\mu_N$, $\alpha=\alpha_N$, and all marginals and centered coefficients are those of the full admissible set. \subsection{Cancellation of the disconnected pair terms} Write \[ C_2^{\mathrm s}=\sum_{|e|=2}c(e),\qquad K_4=\sum_{\substack{|I|=4\\I\text{ connected}}}c(I),\qquad K_4^+=\sum_{\substack{|I|=4\\I\text{ connected},\ q(I)>0}}c(I). \] Compatibility makes every character on the relation space positive, so $K_4^+\geq0$. \begin{lemma}[The fourth factorial cumulant] \label{p13:lower:quartic} For every fixed critical band and every $\varepsilon>0$, \begin{equation} \kappa_4=24K_4+O_{C,\varepsilon}(N^{-5/3+\varepsilon}),\qquad K_4=K_4^++O_{C,\varepsilon}(N^{-3/2+\varepsilon}). \label{p13:lower:quartic-identities} \end{equation} Consequently \begin{equation} \kappa_4+3\alpha^2 =3\alpha^2+24K_4^++O_{C,\varepsilon}(N^{-3/2+\varepsilon}). \label{p13:lower:positive-quartic} \end{equation} \end{lemma} \begin{proof} Expanding the first four factorial moments with the individual $p_i$ gives the exact identity \begin{align} \kappa_4={}&24\sum_{|I|=4}c(I)-12(C_2^{\mathrm s})^2 -24\sum_{|T|=3}\Bigl(\sum_{i\in T}p_i\Bigr)c(T)\notag\\ &+24\sum_{e=\{i,j\}}(p_i^2+p_ip_j+p_j^2)c(e) -6\sum_i p_i^4. \label{p13:lower:exact-cumulant} \end{align} For example, first write \[ \E(1+t)^Z=\prod_i(1+p_it) \left[1+\sum_{2\leq|I|\leq4} \frac{c(I)t^{|I|}}{\prod_{i\in I}(1+p_it)}+O(t^5)\right], \] then take $24$ times the coefficient of $t^4$ in its formal logarithm. This is a finite coefficient calculation; no analytic logarithm of the true generating function is required. Canonical direct decomposition gives $\sum_{|I|=4}c(I)=K_4+D_{22}$, where $D_{22}$ sums $c(e)c(f)$ over unordered disjoint pairs of edges whose spaces are direct. In $(C_2^{\mathrm s})^2/2-D_{22}$ only repeated edges, edges sharing a vertex, and disjoint edges with a nonzero intersection remain. The positive propagation estimate, together with the product of the marginals, implies \[ \sum_{y\ne x}|c(x,y)|\ll_{C,\eta}p_xN^\eta. \] Repeated edges and shared vertices consequently have total absolute weight $O(p_*A_2N^\eta)=O(N^{-5/3+2\eta})$. Apply the pair partition of Lemma~\ref{p13:activity:pair-partition} with a small fixed parameter $\eta$. A product involving a weak or exceptional edge has total weight at most $A_2(A_{\mathrm w}+A_{\mathrm s}) \ll N^{-5/3+O(\eta)}$. For two good edges from the remaining finite family of channels, fix the first edge and the offset pattern of the second compression. There are $NB^{O(1)}$ choices for the first edge. A nonzero intersection imposes an exceptional target of fixed width, giving $N^{1/2+o(1)}$ choices for the second compressed center by the elementary exceptional-target count in Proposition~\ref{p1:arith:fixed-branching}. The proof of that count uses only a fixed width and applies to this compression as well. Each product has weight at most $N^{-10/3+o(1)}$. Hence this last cost is $N^{-11/6+o(1)}$, and \[ D_{22}=(C_2^{\mathrm s})^2/2+O(N^{-5/3+\varepsilon}). \] The remaining terms of \eqref{p13:lower:exact-cumulant} are bounded by $O(p_*A_3+p_*^2A_2+p_*^3\mu)$. The triple estimate in Corollary~\ref{p13:third:activity} gives $A_3\ll N^{-41/30+\eta}$, so each is smaller than the stated error. This proves the first identity. The incompatible part of Proposition~\ref{p13:transfer:fixed-activities} at the fixed order \(r=4\) gives \[ |K_4-K_4^+| \le\sum_{\substack{|I|=4,\ I\text{ connected}\\q(I)=0}}|c(I)| \ll_{C,\varepsilon}N^{-3/2+\varepsilon}. \] Compatible coefficients are nonnegative. This proves the second identity and then \eqref{p13:lower:positive-quartic}. \end{proof} \subsection{A Charlier polynomial detects the remaining mass} \begin{theorem}[Optimal power for the exact quadratic reference] \label{p13:lower:optimal} For every $C>0$ and $\varepsilon>0$, there are $c_{C,\varepsilon}>0$ and $N_0(C,\varepsilon)$ such that \begin{equation} \|\Law(Z_N)-Q_{\mu_N,\alpha_N}\|_1 \geq c_{C,\varepsilon}N^{-4/3-\varepsilon} \label{p13:lower:norm} \end{equation} whenever $N\geq N_0$ and $|L-\log_2N|\leq C$. Together with Theorem~\ref{p13:transfer:reconstruction} and the triple activity bound, this identifies the optimal power as $4/3$. It does not assert an asymptotic constant or a logarithmic factor for the remainder. \end{theorem} \begin{proof} The arithmetic coefficient estimate gives $\alpha\asymp_C LN^{-2/3}>0$ uniformly in the critical band. The sign argument has the margin $3/2-4/3=1/6$. Choose a fixed total error loss $\tau<1/6$ in \eqref{p13:lower:positive-quartic}; then \[ \frac{N^{-3/2+\tau}}{\alpha^2} \ll_C L^{-2}N^{-1/6+\tau}\longrightarrow0. \] Positivity therefore yields \[ \kappa_4+3\alpha^2\geq (3-o_C(1))\alpha^2. \] The fourth Charlier polynomial is \[ H_4(z;\mu)=(z)_4-4\mu(z)_3+6\mu^2(z)_2-4\mu^3z+\mu^4 =4![t^4]e^{-\mu t}(1+t)^z. \] Its integral against $Q_{\mu,\alpha}$ vanishes, because the factorial generating function of that reference is $e^{\mu t}(1+\alpha t^2/2)$. On the true law its expectation equals $\kappa_4+3\alpha^2$. Thus, with $R=\Law(Z_N)-Q_{\mu,\alpha}$, \[ \left|\int H_4\,dR\right|\gg_C L^2N^{-4/3}. \] The test is unbounded, so a moment argument is needed before it gives a norm estimate. For every fixed integer $k>1$, the fixed moments from Theorem~\ref{p1:fixed:moments} give $\int |H_4(z;\mu)|^k\,d\Law(Z_N)(z)\ll_{C,k}1$. The same bound holds for $|Q_{\mu,\alpha}|$, since $\mu$ stays in a compact subset of $(0,\infty)$, $\alpha=O_C(1)$, and \[ |Q_{\mu,\alpha}| \leq\Pi_\mu+\frac{|\alpha|}{2} \Pi_\mu*(\delta_2+2\delta_1+\delta_0). \] Consequently H\"older's inequality for the positive measure $|R|$ gives \[ \left|\int H_4\,dR\right| \ll_{C,k}\|R\|_1^{1-1/k},\qquad \|R\|_1\gg_{C,k}(L^2N^{-4/3})^{k/(k-1)}. \] After $\varepsilon$ has been prescribed, choose the integer $k$ so that $4/[3(k-1)]\leq\varepsilon$. It remains fixed as $N$ tends to infinity, and proves \eqref{p13:lower:norm}. The upper bound follows from the reconstruction and $A_3\ll N^{-41/30+\eta}$, since $41/30>4/3$. \end{proof} The obstruction is specific to the quadratic reference. The Hermite law with the same first two factorial cumulants already contains the term $\alpha^2\Pi_\mu*\Delta^{*4}/8$. The fourth Charlier test against that law sees $\kappa_4$ instead of $\kappa_4+3\alpha^2$, so the lower bound above does not establish an optimal Hermite approximation rate. \section{Completing the approximation and its consequences} \label{p14:reference:section} The reconstruction is carried out in the algebra of signed measures. For the parameters of the model its final reference is a probability. The following positivity and continuity facts are elementary and do not depend on the law approximation. The deduction of Corollary~\ref{p1:cor:tv} from the main theorem is given below. \subsection{Continuity of the reference} For nonnegative $\mu,\mu'$, \begin{equation} \|\Pi_\mu-\Pi_{\mu'}\|_1\le2|\mu-\mu'|. \label{p1:close:poisson-lipschitz} \end{equation} For example, if $\mu'=\mu+h$, couple by adding an independent Poisson variable of mean $h$. Its probability of being nonzero is $1-e^{-h}\le h$, proving the inequality. It follows that \begin{equation} \|Q_{\mu,\alpha}-Q_{\mu',\alpha'}\|_1 \le2(1+2|\alpha|)|\mu-\mu'|+2|\alpha-\alpha'|. \label{p1:close:reference-lipschitz} \end{equation} Here we have expanded the difference with coefficient $\alpha$ on $\Pi_\mu-\Pi_{\mu'}$ and used $\|\Pi_{\mu'}*\Delta^{*2}\|_1\le4$. \subsection{Positivity and the coefficient of total variation} \label{p1:close:positivity-section} The signed representation of the reference is useful in the proof, but for the parameters of the theorem it is eventually a probability. The following exact criterion also specifies its domain for general parameters. \begin{lemma}[Positivity of the quadratic Poisson correction] \label{p1:close:positivity} Let $\mu>0$, $\alpha\in\mathbb R$, and $\theta=\mu-\lfloor\mu\rfloor$. Then $Q_{\mu,\alpha}$ is a probability measure if and only if \begin{equation} 0\le\alpha\le \frac{2\mu^2}{\mu+\theta(1-\theta)}. \label{p1:close:positivity-range} \end{equation} Its masses are \begin{equation} Q_{\mu,\alpha}\{k\} =e^{-\mu}\frac{\mu^k}{k!} \left[1+\frac{\alpha}{2\mu^2} \bigl((k-\mu)^2-k\bigr)\right], \qquad k\ge0. \label{p1:close:charlier-masses} \end{equation} \end{lemma} \begin{proof} Formula \eqref{p1:close:charlier-masses} follows by expressing the Poisson masses at $k-1$ and $k-2$ relative to the mass at $k$; negative-index masses are taken to be zero. The total mass is one because $\Delta^{*2}$ has total mass zero. If $\alpha<0$, the bracket is negative for all sufficiently large $k$, so positivity is impossible. For $\alpha\ge0$, minimize the quadratic $f(k)=(k-\mu)^2-k$ over nonnegative integers. Its successive difference is $f(k+1)-f(k)=2(k-\mu)$. If $\mu$ is not an integer, the minimum is attained at $k=\lceil\mu\rceil$; if $\mu$ is an integer, it is attained at both $k=\mu$ and $k=\mu+1$. In either case its value is $-\mu-\theta(1-\theta)$. Nonnegativity of the bracket is therefore exactly \eqref{p1:close:positivity-range}. \end{proof} In the theorem, $\alpha_N>0$ for large $N$ and $\alpha_N\to0$, while $\mu_N$ and $\lambda_N$ stay in a compact subinterval of $(0,\infty)$. Both corrected references are therefore probabilities for all sufficiently large $N$. The norm estimate simultaneously controls all their masses; it is not only a pointwise assertion at each fixed count. Define, for $\lambda>0$, \begin{equation} \mathcal C(\lambda) =\frac14\|\Pi_\lambda*\Delta^{*2}\|_1 =\frac14\E\left| 1-\frac{2P_\lambda}{\lambda} +\frac{P_\lambda(P_\lambda-1)}{\lambda^2} \right|, \qquad P_\lambda\sim\operatorname{Pois}(\lambda). \label{p1:close:charlier-constant} \end{equation} The coefficient is explicit even when $\lambda_N$ does not converge. A summable signed sequence of total mass zero has $\ell^1$ norm at least twice any positive coefficient. The mass $e^{-\lambda}$ at zero of $\Pi_\lambda*\Delta^{*2}$ and its convolution norm therefore give \begin{equation} \frac{e^{-\lambda}}2\le\mathcal C(\lambda)\le1. \label{p1:close:constant-bounds} \end{equation} \subsection{Completion of the main theorem} \label{p13:optimal:section} The arithmetic and probabilistic estimates now meet at a single quantity, the absolute third-order activity. We complete the main theorem, then record what its precision allows for unbounded scalar observables and for the choice of mean. \begin{proof}[Proof of Theorem~\ref{p1:thm:main}] Write $\varepsilon_{\mathrm{out}}=\varepsilon$ for the requested final margin and choose $0<\tau<\min(\varepsilon_{\mathrm{out}}/4,1/960)$. Invoke Theorem~\ref{p13:transfer:reconstruction} with accuracy $\tau$ and Corollary~\ref{p13:third:activity} with accuracy $\tau$. The triple contribution is \[ N^\tau A_3\ll N^{-41/30+2\tau} =N^{-4/3-(1/30-2\tau)}. \] Since $1/30-2\tau>0$ and the other reconstruction term is $N^{-4/3+\tau}$, this proves the upper bound. Theorem~\ref{p13:lower:optimal} gives the lower bound for the same exact parameters, using its own fixed Charlier interpolation order. The mean estimate is Proposition~\ref{p1:arith:core}. The pair asymptotic \eqref{p1:eq:alpha-leading} shows that \(\alpha_N>0\), \(\alpha_N\to0\), and \(\mu_N\) stays in a compact subset of \((0,\infty)\). Lemma~\ref{p1:close:positivity} therefore makes the reference a probability measure for sufficiently large \(N\). Every auxiliary accuracy parameter and fixed order was chosen after \(C,\varepsilon\) and before \(N\). The conclusion is uniform on the entire fixed critical band. \end{proof} \subsection{Polynomial weights and fixed moments} \begin{corollary}[Every fixed polynomial weight] \label{p13:optimal:weighted} For every fixed integer \(m\ge0\) and \(\varepsilon>0\), \begin{equation} \sum_{k\ge0}(1+k)^m \left|\P(Z_N=k)-Q_{\mu_N,\alpha_N}\{k\}\right| \ll_{C,m,\varepsilon}N^{-4/3+\varepsilon}. \label{p13:optimal:weighted-bound} \end{equation} For every fixed \(r\ge2\), consequently, \begin{equation} \E(Z_N)_r =\mu_N^r+\binom r2\mu_N^{r-2}\alpha_N +O_{C,r,\varepsilon}(N^{-4/3+\varepsilon}). \label{p13:optimal:moments} \end{equation} The equalities for \(r=0,1,2\) hold without a remainder when the right side is interpreted in its natural form. \end{corollary} \begin{proof} Let \(R=\Law Z_N-Q_{\mu_N,\alpha_N}\). For every fixed integer \(v>1\), H\"older gives \[ \sum_k(1+k)^m|R\{k\}| \le\|R\|_1^{1-1/v} \left(\sum_k(1+k)^{mv}|R\{k\}|\right)^{1/v}. \] The second factor is uniformly bounded. For the true law this follows from the independent fixed-moment proof in Theorem~\ref{p1:fixed:moments}; for the reference it follows by dominating its variation by a fixed sum of shifted Poisson laws, as in Theorem~\ref{p13:lower:optimal}. Choose \(v\) sufficiently large after \(\varepsilon\) and then a sufficiently small margin in the main theorem. This proves \eqref{p13:optimal:weighted-bound}. The factorial generating function of \(Q_{\mu,\alpha}\) is \(e^{\mu t}(1+\alpha t^2/2)\); its coefficient of \(t^r\), multiplied by \(r!\), gives the moment formula. \end{proof} This interpolation concerns the scalar count and fixed polynomial degrees. \begin{corollary}[No retained start] \label{p13:optimal:zero} Uniformly in a fixed critical band, \[ \P(Z_N=0) =e^{-\mu_N}\left(1+\frac{\alpha_N}{2}\right) +O_{C,\varepsilon}(N^{-4/3+\varepsilon}). \] \end{corollary} \begin{proof} Evaluate the approximation at zero, using \((\Pi_\mu*\Delta^{*2})\{0\}=e^{-\mu}\). \end{proof} The event is the absence of a retained start in \(I_N\). It does not exclude a long run that started before \(N\). \subsection{Why the mean must be specified at this precision} \begin{proposition}[The quantized mean bias] \label{p13:optimal:nominal} Let \(\eta_x\) indicate one-window compatibility and put \(h_x=L-\dim S_x\). Then \begin{equation} \mu_N-\lambda_N=pM_{N,L},\qquad M_{N,L}=\sum_{x\in I_N}(\eta_x2^{h_x}-1)\in\Z. \label{p13:optimal:integer-mean} \end{equation} Thus either \(\mu_N=\lambda_N\) or \(|\mu_N-\lambda_N|\ge p\asymp_C N^{-1}\). Keeping the same exact \(\alpha_N\), one has \[ \|\Law Z_N-Q_{\lambda_N,\alpha_N}\|_1 \ll_{C,\varepsilon}N^{-1+\varepsilon}. \] On every sufficiently large admissible pair \((N,L)\) for which \(M_{N,L}\ne0\), \begin{equation} \|\Law Z_N-Q_{\lambda_N,\alpha_N}\|_1 \asymp_C|\mu_N-\lambda_N|. \label{p13:optimal:nominal-distance} \end{equation} No assertion that \(M_{N,L}\ne0\) at infinitely many scales is included. \end{proposition} \begin{proof} Sum the one-window formula \(\E J_x=\eta_xp2^{h_x}\) over the exactly \(N\) centres to obtain \eqref{p13:optimal:integer-mean}. For two means in the compact range in question, \eqref{p1:close:reference-lipschitz} gives \[ \|Q_{\mu,\alpha_N}-Q_{\lambda,\alpha_N}\|_1 \ll_C|\mu-\lambda|. \] The reverse bound follows from the mass at zero, \(e^{-\theta}(1+\alpha_N/2)\), by the mean value theorem. The upper bound for the nominal reference follows from the main theorem and \(|\mu_N-\lambda_N|\ll N^{-1+\varepsilon}\). If \(M_{N,L}\ne0\), the mean difference is at least \(p\). Choose a fixed margin less than \(1/3\) in the main theorem; its remainder is then \(o(p)\). The triangle and reverse triangle inequalities transfer the two-sided reference comparison to \eqref{p13:optimal:nominal-distance}. \end{proof} Irregular compatible and incompatible windows contribute with opposite signs to \(M_{N,L}\). Their presence alone does not prove a nonzero total bias. \subsection{The total-variation coefficient} \label{p14:comp:tv} \begin{proof}[Proof of Corollary~\ref{p1:cor:tv}] Theorem~\ref{p1:thm:main} and the reverse triangle inequality first give \[ \TV(\Law Z_N,\Pi_{\mu_N}) =\alpha_N\mathcal C(\mu_N) +O_{C,\varepsilon}(N^{-4/3+\varepsilon}). \] Use \(|\mu_N-\lambda_N|\ll N^{-1+\varepsilon}\) and \eqref{p1:close:reference-lipschitz} to obtain \[ \Law Z_N-\Pi_{\lambda_N} =\frac{\alpha_N}{2}\Pi_{\lambda_N}*\Delta^{*2} +O_{\ell^1,C,\varepsilon}(N^{-1+\varepsilon}). \] The reverse triangle inequality, followed by the factor $1/2$ converting the $\ell^1$ norm to total variation, yields \begin{equation} \TV(\Law Z_N,\Pi_{\lambda_N}) =\alpha_N\mathcal C(\lambda_N) +O_{C,\varepsilon}(N^{-1+\varepsilon}). \label{p1:close:tv-exact-parameter} \end{equation} The pair asymptotic from Theorem~\ref{p1:arith:pairs} now gives \begin{equation} \TV(\Law Z_N,\Pi_{\lambda_N}) \sim \mathcal C(\lambda_N) \frac{r_N+2}{18\,2^{r_N/3}}\, \lambda_N^{5/3}L N^{-2/3}, \qquad r_N=L\bmod3. \label{p1:close:tv-asymptotic} \end{equation} The lower bound in \eqref{p1:close:constant-bounds} is uniform on the fixed critical band, and the remainder is $o(LN^{-2/3})$. This proves the sharp order asserted in the corollary. The equivalence is relative to the displayed oscillating expression; no convergence of its phase or intensity factors is required. \end{proof} Replacing $\alpha_N$ by its leading equivalent proves \eqref{p1:close:tv-asymptotic}. It does not by itself preserve the error $O(N^{-1+\varepsilon})$ in \eqref{p1:close:tv-exact-parameter}, which uses the exact second factorial cumulant. Proposition~\ref{p1:discussion:charlier-norm} gives $\|\Pi_1*\Delta^{*2}\|_1=3/e$, hence $\mathcal C(1)=3/(4e)$. Along $N=2^L$ with $3\mid L$, the pair coefficient therefore gives \begin{equation} \TV(\Law Z_N,\Pi_1) \sim \frac1{12e}\,L N^{-2/3}. \label{p1:close:dyadic-example} \end{equation} This is the coefficient in the example announced in the introduction. The total-variation asymptotic uses the exact-marginal reconstruction and the explicit pair coefficient. The independent positivity and reference-continuity estimates do not use the reconstruction. \section{Explicit references and the first omitted channel} \label{p13:channels:section} The exact second factorial cumulant retains every pair interaction. An explicit reference can instead retain the complete contributions of all rational channels up to a chosen height. The first omitted height then determines its error, including the leading coefficient. This gives a different optimal exponent for each finite truncation, even though all these references approximate the same random variable. For coprime integers $b/20$, and write $b=h+1$ and $r=L\bmod b$. For every $\varepsilon>0$, uniformly in the critical band, \begin{align} \alpha_N-\widehat\alpha_N^{(\leq h)} &=\frac{\lambda_N^2}{N}T_b(L) +O_{h,C,\varepsilon}(N^{-b/(b+1)+\varepsilon}), \label{p13:channels:cumulant-remainder}\\ \Law(Z_N)-Q_{\lambda_N,\widehat\alpha_N^{(\leq h)}} &=\frac{\lambda_N^2T_b(L)}{2N}\Pi_{\lambda_N}*\Delta^{*2} +O_{\ell^1,h,C,\varepsilon}(N^{-b/(b+1)+\varepsilon}). \label{p13:channels:law-remainder} \end{align} Define the positive number \begin{equation} c_b(r)=\frac{b+r-1}{b^2\,2^{r/b}} \sum_{\substack{b/20$, \begin{equation} \sum_{\substack{x,y\in G_N,\ |x-y|>L\\(x,y)\ {\rm nonaligned}}} (2^{\rho(x,y)}-1) \ll_{C,\varepsilon}N^{1+\varepsilon}. \label{p13:channels:nonaligned-weight} \end{equation} Consequently the sum of the absolute covariances on these pairs is $O_{C,\varepsilon}(N^{-1+\varepsilon})$. \end{lemma} \begin{proof} Use the large-prime component decomposition \eqref{p1:arith:graphbound}, so that $\rho(x,y)\leq D(x,y)+c(x,y)$ and $D(x,y)=o(B)$ uniformly. Fix $\eta>0$ temporarily. On $c(x,y)<\eta B$, the weight is at most $2^{\eta B+o(B)}$. A pair in $G_N$ with $\rho(x,y)>0$ has a nonzero intersection between its two effective spaces. The pair count of Corollary~\ref{p13:half:pairs} therefore bounds the contribution of this sector by $N^{1+\eta+\varepsilon'+o(1)}$. On $c(x,y)\geq\eta B$, Lemma~\ref{p1:arith:dense} already bounds the entire weighted sum by $N^{1+\varepsilon'}$. Choose $\eta$ and then $\varepsilon'$ inside the prescribed margin. On $G_N$ both marginals are $p$. The affine identity \eqref{p1:arith:affine} gives $|\Cov(J_x,J_y)|\leq p^2(2^{\rho(x,y)}-1)$, including incompatible pairs. Since $p\asymp_C N^{-1}$, this proves the last assertion. \end{proof} \begin{proof}[Proof of Theorem~\ref{p13:channels:hierarchy}] We first prove, for every fixed integer $m\geq3$, \begin{equation} \alpha_N-\widehat\alpha_N^{(\leq m)} =O_{m,C,\varepsilon}(N^{-m/(m+1)+\varepsilon}). \label{p13:channels:coarse-truncation} \end{equation} The regular-set transfer in Proposition~\ref{p1:arith:core}, justified by the independent fixed-moment bound of Theorem~\ref{p1:fixed:moments}, changes the second factorial cumulant by $O(N^{-1+\varepsilon})$. On $G_N$, that cumulant is the sum of covariances over ordered distinct pairs minus $|G_N|p^2$. The last term is $O(N^{-1})$. Strict contacts are incompatible and number $O(NB)$, so their covariance mass is $O(B/N)$. At separation exactly $L$, the union of the two windows lies in an interval of length $O(B)$ and has formal rank $2L-o(L)$, by Lemma~\ref{p1:arith:defects}; this band has the same $N^{-1+o(1)}$ cost. Lemma~\ref{p13:channels:weighted-nonaligned} treats the separated nonaligned pairs. For aligned separated pairs, reduced slopes are unique for sufficiently large $N$: their separation is at least $B^{-6}$, whereas the allowed offset changes the ratio of the starts by only $B^{O(1)}/N$. There are $NB^{O(1)}$ aligned hosts altogether. A slope of height one has no exact unit for separated windows and hence only residual nullity $o(L)$ by Lemma~\ref{p1:arith:alignedrank}. The slope $1:2$ is confined to $B^{O(1)}$ pairs at the dyadic endpoints; its total mass is at most $N^{-3/2+o(1)}$. Every remaining reduced slope is an interior channel $b/20$ and tends to zero. The positivity criterion of Lemma~\ref{p1:close:positivity} therefore applies uniformly on the compact range of $\lambda_N$. The leading coefficient in \eqref{p13:channels:tv-equivalent} is bounded above and away from zero, since $\|\Pi_{\lambda_N}*\Delta^{*2}\|_1\geq e^{-\lambda_N}$. This proves optimality of the stated power for the specified reference. \end{proof} \subsection{What the first omitted height contributes} The four channels of heights at most five are $2:3$, $3:4$, $3:5$, and $4:5$. Thus the next explicit correction after \eqref{p1:eq:alpha-explicit} has a short closed formula. \begin{corollary}[The error of the two-channel reference] \label{p13:channels:height-five} Write $L=5q+r$, $0\leq r<5$, and, for $q\geq1$, set \begin{equation} \begin{split} M_5(N,L)&=\frac{\lambda_N^2}{N} \left\{\frac2{15}S_{3,5}(L)+\frac3{10}S_{4,5}(L)\right\}\\ &=\frac{\lambda_N^2}{30N} \left\{2^{q-1}\bigl[17(r+4)q+13r^2+87r+620\bigr] -113L-127\right\}. \end{split} \label{p13:channels:M5} \end{equation} Then \begin{align} \Law(Z_N)-Q_{\lambda_N,\widehat\alpha_N} &=\frac{M_5(N,L)}2\Pi_{\lambda_N}*\Delta^{*2} +O_{\ell^1,C,\varepsilon}(N^{-5/6+\varepsilon}), \label{p13:channels:two-channel-law}\\ \TV(\Law(Z_N),Q_{\lambda_N,\widehat\alpha_N}) &\sim\frac{17(r+4)}{1200\,2^{r/5}}\lambda_N^{9/5} \|\Pi_{\lambda_N}*\Delta^{*2}\|_1\,L N^{-4/5}. \label{p13:channels:two-channel-tv} \end{align} The four-channel coefficient is $\widehat\alpha_N^{(\leq5)}=\widehat\alpha_N+M_5(N,L)$. Its reference has optimal exponent $5/6$, with leading phase constant $c_6(r)=(r+5)/(45\,2^{r/6})$, where now $r=L\bmod6$. \end{corollary} \begin{proof} The finite identity gives \begin{align*} S_{3,5}(L)&=2^{q-1}\bigl[2(r+4)q+r^2+6r+38\bigr]-8L-7,\\ S_{4,5}(L)&=2^{q-1}\bigl[(r+4)q+r^2+7r+52\bigr]-9L-11. \end{align*} Their weighted sum proves \eqref{p13:channels:M5}. Theorem~\ref{p13:channels:hierarchy} at $h=4$ gives the first two conclusions. At height six the only reduced interior channel is $5:6$; substitution in \eqref{p13:channels:phase-constant} gives $c_6$ and the assertion for the four-channel reference. \end{proof} For example, if $N=2^L$ and $5\mid L$, the error of the two-channel reference is asymptotic in total variation to $17LN^{-4/5}/(100e)$, because $\|\Pi_1*\Delta^{*2}\|_1=3/e$. The table records how successive complete channels change the error. \begin{center} \begin{tabular}{@{}clcl@{}} \toprule Height $h$ & Retained channels & Optimal power & First omitted channels\\ \midrule 3 & $2:3$ & $3/4$ & $3:4$\\ 4 & $2:3,\ 3:4$ & $4/5$ & $3:5,\ 4:5$\\ 5 & $2:3,\ 3:4,\ 3:5,\ 4:5$ & $5/6$ & $5:6$\\ $h$ fixed & all reduced interior heights $\leq h$ & $h/(h+1)$ & height $h+1$\\ \bottomrule \end{tabular} \end{center} The full coefficient in \eqref{p13:channels:definitions} has a phase dividing $\operatorname{lcm}(3,\ldots,h)$, whereas its leading omitted term uses only $L\bmod(h+1)$, together with $\lambda_N$. The theorem permits any fixed height, with constants depending on that height. It makes no assertion for a height growing with $N$. In particular it does not identify an obstruction of power one for other explicit references: the stated optimality concerns precisely the finite truncations defined in \eqref{p13:channels:definitions}. \subsection{Positive Hermite references} The classical Hermite distribution \cite{kempkemp1965} has generating function \[ \exp\{(\mu-\alpha)(z-1)+(\alpha/2)(z^2-1)\}. \] For \(0\le\alpha\le\mu\), it is the law of a Poisson variable of mean \(\mu-\alpha\) plus twice an independent Poisson variable of mean \(\alpha/2\). \begin{corollary}[Hermite approximation with single and double jumps] \label{p1:cor:compound} For all sufficiently large \(N\), let \(U_N,V_N\) be independent Poisson variables with means \(\mu_N-\alpha_N\) and \(\alpha_N/2\), respectively. Then \[ \TV\bigl(\Law(Z_N),\Law(U_N+2V_N)\bigr) \ll_{C,\varepsilon}N^{-4/3+\varepsilon}. \] For every fixed \(h\ge3\), the entirely explicit reference \[ \widehat U_N+2\widehat V_N,\qquad \widehat U_N\sim\Pois(\lambda_N-\widehat\alpha_N^{(\le h)}),\qquad \widehat V_N\sim\Pois(\widehat\alpha_N^{(\le h)}/2), \] with independent summands has optimal power \(h/(h+1)\) and the same leading distance equivalent as the corresponding quadratic reference in Theorem~\ref{p13:channels:hierarchy}. In particular the two-channel choice \(h=4\) has error \(O(N^{-4/5+\varepsilon})\). No matching lower bound is asserted here for the exact-parameter Hermite reference. \end{corollary} \begin{proof} Both means are nonnegative for large \(N\). In the convolution algebra, the law of \(U_N+2V_N\) is \[ \exp_*\left(\mu_N\Delta+\frac{\alpha_N}{2}\Delta^{*2}\right). \] Expanding the second exponential to first order gives \(Q_{\mu_N,\alpha_N}\) with an \(\ell^1\) error \(O_C(\alpha_N^2)\). Indeed \(\|\Delta^{*2}\|_1=4\), and the exponential remainder is bounded by \(\frac12(2\alpha_N)^2e^{2\alpha_N}\). Since \(\alpha_N^2=N^{-4/3+o(1)}\), the main theorem proves the claim. For each fixed-height explicit reference, the coefficient is positive and tends to zero uniformly in the critical band, while \(\lambda_N\) stays bounded away from zero. Thus both proposed Poisson means are nonnegative for all sufficiently large \(N\). Repeat the same expansion with parameters \((\lambda_N,\widehat\alpha_N^{(\le h)})\), and apply Theorem~\ref{p13:channels:hierarchy}. The additional norm error is \(O_C((\widehat\alpha_N^{(\le h)})^2) =N^{-4/3+o(1)}\), smaller than the positive first omitted channel term of order \(LN^{-h/(h+1)}\). The reverse triangle inequality preserves its equivalent and hence its optimal exponent. \end{proof} The double jump represents the leading pair interaction at the level of the scalar law. The contribution here is the precision of the Hermite approximation for \(Z_N\). Its double jump intensity uses either the exact second factorial cumulant or a complete finite channel sum. This is a representation of the reference law; it does not decompose the actual run starts into independent components. Keeping only the leading equivalent of that coefficient does not preserve the precision of a fixed-height truncation. The exact Hermite law already contains the quartic term \(\alpha_N^2\Pi_{\mu_N}*\Delta^{*4}/8\), so the lower bound for the quadratic reference does not transfer to it. This distinction is already present in the locally dependent setting of \v{C}ekanavi\v{c}ius and Vellaisamy \cite[Theorem~3.1, (3.5), and Theorem~3.2, (3.6)]{cekanaviciusvellaisamy2015}. Their quadratic reference is \(\Pi_{\Gamma_1}*(\delta_0+\Gamma_2\Delta^{*2})\), whereas the exponential reference is \(\exp_*(\Gamma_1\Delta+\Gamma_2\Delta^{*2})\), with \(\Gamma_2=(\Var Z-\E Z)/2\). The quadratic estimate contains an additional squared error term that is absent from the exponential estimate. These results concern sums reduced to one-dependence under specified moment conditions. We use this comparison to place the references in context; the arithmetic estimates here provide their approximation to the multiplicative run count. \section{What the proof makes reusable} \label{p1:discussion} The exponent \(4/3\) results from the relation between two scales: the leading pair mass is \(N^{-2/3+o(1)}\), and the quadratic reference omits a contribution formed from two such interactions. The arithmetic work shows that all other terms relevant to the reconstruction can be controlled at this precision. The proof also separates several choices that can be made in other rare-event problems. \subsection{Choose a reference before estimating the error} The independent Bernoulli reference retains every exact marginal. Its centered expansion has no one-site term, and repeated indices are already accounted for. The retained pair correction then removes the complete second-order contribution. This makes the quantity to be bounded a residual interaction, with a specified sign and normalization. A homogeneous envelope provides a second useful reduction. The joint probability of the homogeneous constraints bounds the actual joint probability, the exact marginal rank weight and the absolute centered activity. Geometric counting can therefore be performed once and used for each of these weights. At a fixed arity \(r\), the comparison with the nominal weight costs at most \(2^r\). That fixed factor can be absorbed into the constant of a fixed-order estimate. The estimates that sum over unbounded arities keep their own weights and quantifiers. Resultants and degree-uniform curve bounds also appear in Shparlinski and Sleiman's count of consecutive multiplicatively dependent triples \cite[Theorem~1.1 and Section~3]{shparlinskisleiman2026}. The additional feature used here is the simultaneous placement of a large space of square-class relations on one common curve, together with the integer fibres that those relations force in the offset grid. This converts geometric information about logarithmic windows into a bound for their shared rank. \subsection{Stop where the remaining dependence is still positive} The specialization of the abstract lace expansion \cite{zeilberger1997} in Section~\ref{p1:stop:section} retains a hereditary class of subsets with computable joint probabilities. At the first excluded prefix, all later variables remain in a positive conditional measure. Its total mass is the joint probability of the prefix. The boundary therefore has a direct probabilistic weight even though the retained expansion is signed. Incompatible pair and triple terms must still be combined before absolute values are taken. To use this identity in another model, one needs a retained class that matches the chosen reference and quantitative control of the first excluded configurations. Here compression makes a dependent pair one affine block. Conditional arithmetic counts control a fixed number of blocks, and congruence certificates control the large-cardinality part. The reconstruction estimate records exactly what these inputs accomplish: \[ \|\Law(Z_N)-Q_{\mu_N,\alpha_N}\|_1 \ll N^\varepsilon A_3+N^{-4/3+\varepsilon}. \] The finite identity itself does not give the arithmetic bounds needed on its right-hand side. \subsection{Separate fixed orders from a uniform tail} A fixed-order theorem allows its constant and scale threshold to depend on the order. It cannot be summed up to an order growing with \(N\) without further information. The proof instead chooses a finite cutoff after the desired accuracy, treats the finitely many lower orders by their individual estimates, and uses one uniform weighted tail above that cutoff. Proposition~\ref{p1:uniform:full-support} illustrates this arrangement. Finite layers give a half power for each effective block, while the independent tail gives every power strictly below \(2\). Their combination yields the exponent \(\min(s/2,2)\), with an arbitrary fixed margin and a common constant for all \(s\ge2\). This is a bound on a deterministic sum of independent counting weights. It is not an exponential-tail statement for the law of \(Z_N\). The independent fixed-moment proof has a different use: it turns a polynomial test into a norm lower bound and gives fixed polynomially weighted consequences of the main theorem. The finite-difference example in Section~\ref{p1:moment-obstruction} shows why even very strong information at each fixed moment order does not by itself provide a polynomial rate for the complete law. \subsection{Test the approximation that was actually retained} The fourth Charlier polynomial detects \(\kappa_{4,N}+3\alpha_N^2\) after subtraction of the quadratic reference. The nonnegative compatible part of the quartic identity prevents cancellation at the scale of \(\alpha_N^2\). For the exact Hermite reference the same test instead detects \(\kappa_{4,N}\), because its exponential already contains the two-pair term. A lower bound is therefore tied to a specified reference, not just to the underlying dependent count. The full equivalent for the exact quadratic remainder, the full constant for the third factorial cumulant, and the sharp error of the exact Hermite reference remain open in this work. So do the errors after cubic or quartic correction and the optimization of total variation over all Poisson means. The explicit channel coefficients determine reference laws, but the constants in the asymptotic remainders and the thresholds in \(N\) are not supplied numerically. Applying the method elsewhere requires counterparts of the conditional counts, compatibility cancellations and uniform boundary estimate; the stopping identity alone does not impose a universal approximation exponent. \section*{Formal proofs and reproducibility} \addcontentsline{toc}{section}{Formal proofs and reproducibility} The accompanying \href{https://github.com/Vulkin-prog/paper-p1-lean}{Lean development} provides proofs for the 118 numbered results in the article and companion (26 and 92, respectively). Its \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/e564207ab543c7c0967e34b52f3d456d7a8f302c/docs/manuscript-results-publication-2026-11-12.json}{statement index} records the correspondence of clauses, definitions and inherited hypotheses, including the conditional premise of Theorem~\ref{p13:transfer:finite}. The documentation of the recorded replay is fixed at commit \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/a1c8092f7e25acf7f3b78283039cfbd69491f88e}{\texttt{a1c8092f7e25}}. The linked \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/a1c8092f7e25acf7f3b78283039cfbd69491f88e/docs/validation/v2.0/proof-source-manifest.json}{source manifest} and \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/a1c8092f7e25acf7f3b78283039cfbd69491f88e/docs/validation/v2.0/validation.json}{validation record} identify the proof source used for the recorded full replay as \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/aa12e797ce895f6b7aef6e4f1f0c6e5e9a58aab4}{\texttt{aa12e797ce89}}. That record specifies Lean~4.36.0-rc2 and Mathlib commit \href{https://github.com/leanprover-community/mathlib4/tree/9d581dfff2e41de989072ea63abf3daf69311520}{\texttt{9d581dfff2e4}}, and includes compilation results, extracted statement types and the kernel-axiom audit. The documentation snapshot and the replayed proof source have distinct roles. The correspondence between the written and formal statements was examined with AI assistance. It is not a machine translation of the written proofs or a substitute for independent human review. The conditional-sum and propagation clauses of Proposition~\ref{p1:fixed:propagation} allow homogeneous prescriptions; its small-contact clause uses the original run-start events. The formal interfaces distinguish these scopes as well. \Needspace{14\baselineskip} The final five-dossier allocation selects 42 formal statements covering every clause of the article's 26 numbered results. It preserves the four issued registrations and merges the two previously unregistered remainder scopes. The allocation is as follows, with numbers referring to the article: \begin{center} \small \begin{tabularx}{\linewidth}{@{}Xlr@{}} \toprule Dossier & Main results & Statements\\ \midrule Quadratic approximation and Poisson scales & 1.1--1.2, 7.2--7.4 & 10\\ Cumulants and optimality & 3.3--3.4, 6.1--6.2 & 8\\ Pair profiles and channel hierarchy & 3.1--3.2, 8.1--8.3 & 13\\ Stopping, conversion and reconstruction & 4.1--4.2, 5.1--5.3 & 8\\ Uniform support and positive reference limits & 2.1--2.5, 7.1, 8.4 & 3\\ \bottomrule \end{tabularx} \end{center} The last dossier uses three compound statements, retaining all thirteen remaining main-result clauses and directly selecting companion Propositions~T13.1 and~T15.1, with the latter's displayed sequence corollary. The former controls the weighted full-support sum at all cardinalities with one threshold before every subset and minimum size. The latter shows that fixed factorial moments and uniform exponential-moment control cannot replace growing-order estimates for a polynomial total-variation rate; no multiplicative realization of that counterexample is claimed. The other companion results supply proof dependencies, without being claimed as all directly registered. The \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/eeda0e055bc16afa6bb6402727afe733ace9961e/palomar/REGISTRATION_PLAN_FINAL.json}{exact clause allocation} fixes these scopes. \Needspace{6\baselineskip} As of 13 November 2026, all five entries are published, in table order: \par\smallskip\noindent \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-12-000013\&version=1}{\texttt{PALOMAR-2026-11-12-000013}};\quad \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000001\&version=1}{\texttt{PALOMAR-2026-11-13-000001}};\\ \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000002\&version=1}{\texttt{PALOMAR-2026-11-13-000002}};\quad \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000008\&version=1}{\texttt{PALOMAR-2026-11-13-000008}};\\ \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000020\&version=1}{\texttt{PALOMAR-2026-11-13-000020}}. The first entry fixes source \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/f9a2d213265a6833318745dd7f2d74c02090ab19}{\texttt{f9a2d213265a}}; entries 2--4 fix \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/d6a0f6c5109886127526677603eb6b74e8c0cd82}{\texttt{d6a0f6c51098}}; the final dossier fixes \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/eeda0e055bc16afa6bb6402727afe733ace9961e}{\texttt{eeda0e055bc1}}. The final interface has a separate audit of 111 declarations and fresh statement comparison under the same Lean and Mathlib pins. Individual public entries provide the actual mechanical reports, rendered statements, preserved sources and published warnings. \Needspace{11\baselineskip} \section*{Statement on AI assistance} \addcontentsline{toc}{section}{Statement on AI assistance} ChatGPT (GPT-6 Astra) was used to organize the research material, develop and rewrite mathematical arguments, inspect references, and prepare the article and its technical companion. Parallel AI-assisted reviews examined selected identities, constants, proof dependencies and the exposition. The associated Lean development also used AI assistance for proof construction and statement comparison. AI-assisted review and machine checking are distinct from independent human verification. The author remains responsible for the mathematical claims, their correspondence with the formal statements, and any submission. \Needspace{6\baselineskip} % Generated from bibliography.tex by build.py. \begin{thebibliography}{99} \addcontentsline{toc}{section}{References} \bibitem{agg1989} R.~Arratia, L.~Goldstein, and L.~Gordon, \emph{Two moments suffice for Poisson approximations: the Chen--Stein method}, Annals of Probability \textbf{17} (1989), no.~1, 9--25. \href{https://doi.org/10.1214/aop/1176991491} {doi:10.1214/aop/1176991491}. \bibitem{agw1990} R.~Arratia, L.~Gordon, and M.~S.~Waterman, \emph{The Erd\H{o}s--R\'enyi law in distribution, for coin tossing and sequence matching}, Annals of Statistics \textbf{18} (1990), no.~2, 539--570. \href{https://doi.org/10.1214/aos/1176347615} {doi:10.1214/aos/1176347615}. \href{https://dornsife.usc.edu/msw/wp-content/uploads/sites/236/2023/09/msw-093.pdf} {Article text}. \bibitem{barbour1987} A.~D.~Barbour, \emph{Asymptotic expansions in the Poisson limit theorem}, Annals of Probability \textbf{15} (1987), no.~2, 748--766. \href{https://doi.org/10.1214/aop/1176992169} {doi:10.1214/aop/1176992169}. \href{https://user.math.uzh.ch/barbour/pub/Barbour/AsympExpanPois.pdf} {Article text}. \bibitem{barboureagleson1987} A.~D.~Barbour and G.~K.~Eagleson, \emph{An improved Poisson limit theorem for sums of dissociated random variables}, Journal of Applied Probability \textbf{24} (1987), no.~3, 586--599. \href{https://doi.org/10.2307/3214091} {doi:10.2307/3214091}. \href{https://user.math.uzh.ch/barbour/pub/Barbour/BEagleson_ImprovedPoisson.pdf} {Article text}, Theorem~1, p.~588. \bibitem{bkn2014} A.~D.~Barbour, E.~Kowalski, and A.~Nikeghbali, \emph{Mod-discrete expansions}, Probability Theory and Related Fields \textbf{158} (2014), 859--893. \href{https://doi.org/10.1007/s00440-013-0498-8} {doi:10.1007/s00440-013-0498-8}. Section numbers refer to the 26-page author text, Sections~5 and~8.1: \url{https://people.math.ethz.ch/~kowalski/mod-discrete.pdf}. \bibitem{bretwngxu2026} R.~de~la~Bret\`eche, V.~Y.~Wang, and M.~W.~Xu, \emph{Random multiplicative functions and making squares from polynomial values}, \href{https://arxiv.org/abs/2607.06398v1} {arXiv:2607.06398v1} (2026), Theorem~1.2 and Lemmas~6.2--6.3. \bibitem{dohmen1999} K.~Dohmen, \emph{Improved inclusion-exclusion identities and inequalities based on a particular class of abstract tubes}, Electronic Journal of Probability \textbf{4} (1999), paper~5, 1--12. \href{https://doi.org/10.1214/EJP.v4-42} {doi:10.1214/EJP.v4-42}. \href{https://www.maths.tcd.ie/EMIS/journals/EJP-ECP/article/download/42/42-83-1-PB.pdf} {Sections~2--3}. \bibitem{godbole1991} A.~P.~Godbole, \emph{Poisson approximations for runs and patterns of rare events}, Advances in Applied Probability \textbf{23} (1991), no.~4, 851--865. \href{https://doi.org/10.2307/1427680} {doi:10.2307/1427680}. \bibitem{naimanwynn1997} D.~Q.~Naiman and H.~P.~Wynn, \emph{Abstract tubes, improved inclusion-exclusion identities and inequalities and importance sampling}, Annals of Statistics \textbf{25} (1997), no.~5, 1954--1983. \href{https://doi.org/10.1214/aos/1069362380} {doi:10.1214/aos/1069362380}. \bibitem{najnudel2020} J.~Najnudel, \emph{On consecutive values of random completely multiplicative functions}, Electronic Journal of Probability \textbf{25} (2020), paper~59, 1--28. \href{https://doi.org/10.1214/20-EJP456} {doi:10.1214/20-EJP456}. \href{https://arxiv.org/abs/1702.01470v2} {arXiv:1702.01470v2}, Lemma~3.1 and Proposition~6.3. \bibitem{pouly2026c} B.~Pouly, \emph{Long runs and rare patterns of a random completely multiplicative function}, with technical companion, Zenodo (2026). \href{https://doi.org/10.5281/zenodo.22872154} {doi:10.5281/zenodo.22872154}. Theorem~1.1 and Section~3.6, (3.22). \bibitem{zeilberger1997} D.~Zeilberger, \emph{The abstract lace expansion}, Advances in Applied Mathematics \textbf{19} (1997), no.~3, 355--359. \href{https://doi.org/10.1006/aama.1997.0552} {doi:10.1006/aama.1997.0552}. \href{https://sites.math.rutgers.edu/~zeilberg/mamarim/mamarimPDF/lace.pdf} {Author version}, formula~(*). \bibitem{cekanaviciusvellaisamy2015} V.~\v{C}ekanavi\v{c}ius and P.~Vellaisamy, \emph{Discrete approximations for sums of $m$-dependent random variables}, ALEA, Latin American Journal of Probability and Mathematical Statistics \textbf{12} (2015), 765--792. \href{https://alea.impa.br/articles/v12/12-29.pdf}{Journal text}, Theorem~3.1, equation~(3.5), and Theorem~3.2, equation~(3.6). \href{https://arxiv.org/abs/1301.7196v3}{arXiv:1301.7196v3}. \bibitem{chhaibietal2020} R.~Chhaibi, F.~Delbaen, P.-L.~M\'eliot, and A.~Nikeghbali, \emph{Mod-$\phi$ convergence: Approximation of discrete measures and harmonic analysis on the torus}, Annales de l'Institut Fourier \textbf{70} (2020), no.~3, 1115--1197. \href{https://doi.org/10.5802/aif.3332}{doi:10.5802/aif.3332}. \href{https://numdam.org/item/10.5802/aif.3332.pdf}{Journal text}, Theorem~3.11. \bibitem{cekanaviciusvellaisamy2020} V.~\v{C}ekanavi\v{c}ius and P.~Vellaisamy, \emph{Lower bounds for discrete approximations to sums of $m$-dependent random variables}, Probability and Mathematical Statistics \textbf{40} (2020), no.~1, 23--35. \href{https://doi.org/10.37190/0208-4147.40.1.2} {doi:10.37190/0208-4147.40.1.2}. \href{https://www.math.uni.wroc.pl/~pms/files_doi/f1648/f1648.pdf} {Journal text}, Theorem~2.1, equation~(2.3). \bibitem{basaknathzaharescu2024} D.~Basak, K.~Nath, and A.~Zaharescu, \emph{Poisson behavior for chains of small quadratic non-residues}, International Mathematics Research Notices \textbf{2024} (2024), no.~4, 3356--3390. Published online 24 July 2023. \href{https://doi.org/10.1093/imrn/rnad168} {doi:10.1093/imrn/rnad168}. Theorem~1. \bibitem{shparlinskisleiman2026} I.~E.~Shparlinski and N.~Sleiman, \emph{Counting consecutive multiplicatively dependent triples}, \href{https://arxiv.org/abs/2609.29408v1} {arXiv:2609.29408v1} (24 September 2026), Theorem~1.1, Lemmas~2.1--2.2, and Section~3. \bibitem{kempkemp1965} C.~D.~Kemp and A.~W.~Kemp, \emph{Some properties of the `Hermite' distribution}, Biometrika \textbf{52} (1965), nos.~3--4, 381--394. \href{https://doi.org/10.1093/biomet/52.3-4.381} {doi:10.1093/biomet/52.3-4.381}. \end{thebibliography} \clearpage \renewcommand{\HXtocfile}{companioncontents} \renewcommand{\HXbibprefix}{companion} \pagenumbering{arabic} \setcounter{section}{0} \setcounter{subsection}{0} \setcounter{subsubsection}{0} \setcounter{theorem}{0} \setcounter{equation}{0} \setcounter{table}{0} \setcounter{figure}{0} \setcounter{footnote}{0} \newcommand{\PoneTechnicalCompanion}{} \makeatletter \patchcmd{\l@section}{1.5em}{2.3em}{}{\PackageError{p1}{Could not widen contents numbers}{}} \makeatother \renewcommand{\thesection}{T\arabic{section}} \renewcommand{\theHsection}{T.\arabic{section}} \title{\bfseries Sharp Poisson approximation\\and second-order correction\\for long multiplicative runs\\[9pt]\large Technical companion} \author{Brice Pouly\\[5pt] \small With mathematical and editorial assistance from ChatGPT (GPT-6 Astra)} \date{} \pdfbookmark[0]{Technical companion}{HXcompanion} \maketitle \thispagestyle{plain} \begin{abstract} This companion supplies the arithmetic and uniform counting estimates for the article \emph{Sharp Poisson approximation and second-order correction for long multiplicative runs}. The two texts prove the explicit Poisson distance equivalent, the optimal quadratic error power $4/3$, the third factorial cumulant power $41/30$, and the hierarchy of errors for rational channels of fixed height. The proofs are organized by the estimates they supply: exact pair profiles, fixed-source and half-power counts, confinement and pair rigidity, positive triple masses, and independent certificate tails. A common homogeneous probability bounds the several weights used in these arguments. The resulting consequences include positive Palm costs for arbitrary target windows and a support count uniform in its lower cardinality. Fixed-order and growing-cardinality estimates retain distinct hypotheses throughout. \end{abstract} \clearpage \begingroup \small \setlength{\parskip}{0pt} \tableofcontents \endgroup \clearpage \section*{How the estimates fit together} \addcontentsline{toc}{section}{How the estimates fit together} The model, exact marginals, effective relation spaces and centered coefficients are defined in Sections~\ref{p1:intro:model} and~\ref{p13:exact:section} of the article. References without a T prefix refer to that article. Both documents are needed for the complete proof. \paragraph{Pair coefficients and independent fixed moments.} Sections~\ref{p1:arith:appendix}--\ref{p1:arith:supplement} give the arithmetic pair estimates. The elementary residual power $6/5$ suffices for the first pair profile; its later quasi-linear strengthening is not used to prove itself. Section~\ref{p1:fixed:section} establishes the fixed moments independently of the approximation of the law. The good-pair compression estimate remains necessary in this moment proof. \paragraph{The gain below the quadratic scale.} Short-selection counts, conditional half-counts, confinement and pair rigidity culminate in the triple analysis of Section~\ref{p13:third:section}. Its positive and signed estimates give $A_3\ll N^{-41/30+\varepsilon}$. The all-target Palm estimate is a consequence of the positive connected-triple mass and the absolute marginal bias. It is not used as a premise for that mass. \paragraph{Summing the entire expansion.} The elementary certificates of Section~\ref{p14:cert:section} are used in Section~\ref{p2:technical-transfer}. The latter proves separate finite-order estimates and a uniform tail beyond order $256$. Its homogeneous primitive estimate rests on the geometric dense-host construction and the half-count, not on a law approximation. The article applies these estimates to its finite stopped expansion. \paragraph{Two useful consequences.} The uniform support estimate in Section~\ref{p1:uniform:appendix} combines finitely many half-power layers with one independent tail. Its threshold and constant are uniform in the lower support size. The final finite-difference example explains why estimates at all fixed moment orders alone cannot supply a polynomial distance rate. In every asymptotic assertion, the critical-band constant, accuracy parameters, fixed source counts and fixed channel heights precede the scale threshold. Growing degrees and cardinalities occur only in estimates that state their uniform range explicitly. \section{Uniform arithmetic estimates} \label{p1:arith:appendix} This section supplies the uniformity needed when the window length grows. The degree parameters in the determinant arguments are fixed before $N$ tends to infinity. Dependence on such fixed parameters is permitted; dependence on a squarefree coefficient of size $N^{O(1)}$ is not. \subsection{Split products, defects, and the regular set} Products in short intervals have a long history; see Erd\H{o}s and Turk~\cite{erdosturk1984}. The polynomial square-root argument for a fixed even number of distinct factors appears in the proof of Munsch, Shparlinski, Sun, and Xiao~\cite[Lemma~3.3]{munsch2026}. We record an explicit bound uniform in the degree, as required when the number of selected factors grows. No constant depending implicitly on that degree is used in the following lemma. \begin{lemma}[A split-product bound uniform in the degree]\label{p1:arith:runge} Let $d=2k\geq2$ and let $\gamma_1,\ldots,\gamma_d$ be distinct integers with $|\gamma_i|\leq R$, $R\geq1$. If $U\geq2R$ is an integer and $\prod_i(U+\gamma_i)$ is a square, then \[ U\leq(64R)^d. \] \end{lemma} \begin{proof} Write $G(T)=\prod_i(T+\gamma_i)$ and \[ \phi(z)=\prod_i(1+\gamma_i z)^{1/2}=\sum_{j\geq0}c_jz^j, \qquad P(T)=\sum_{j=0}^k c_jT^{k-j}. \] The branches equal one at the origin. Cauchy's bound on $|z|=(2R)^{-1}$ gives $|c_j|\leq(3/2)^{d/2}(2R)^j$ and, for $j\leq k$, $|c_j|\leq2^dR^j$. Comparing coefficients in $\phi^2$ inductively gives $2^{2j-1}c_j\in\mathbb Z$ for $j\geq1$, so $2^dP\in\mathbb Z[T]$. For $U>4R$ the tail satisfies \[ |\sqrt{G(U)}-P(U)|\leq4\,3^{d/2}R^{k+1}/U. \] If this difference is nonzero, it is at least $2^{-d}$, whence $U\leq(8R)^d$. If it vanishes, the nonzero integer polynomial $Q=2^{2d}(P^2-G)$ has root $U$. It is nonzero because $G$ has simple roots. Keeping the homogeneous powers of $R$ in the coefficient products gives \[ |[T^{d-j}](P^2-G)|\leq(d+2)4^dR^j\leq8^dR^j, \qquad H(Q)\leq(32R)^d. \] The integral-root bound gives $U\leq1+H(Q)\leq(64R)^d$. The range $2R\leq U\leq4R$ already satisfies the conclusion. \end{proof} The following coefficient-uniform count is classical; see Cilleruelo and Garaev~\cite[Section~3, Lemma~4 and Proposition~1]{cilleruelogaraev2011} and the directly comparable formulation of Tao~\cite[Lemma~2.10]{tao2026anatomy}. When $ab$ is not a square, the substitution $(au)^2-abv^2=ah$ reduces it to the former lemma; the square case is a divisor count. We include a proof to keep the polynomial coefficient range explicit. \begin{lemma}[Uniform quadratic equation count]\label{p1:arith:pell} Fix $A_0>0$. Uniformly for positive integers $a,b$ and nonzero integers $h$ with $a,b,|h|\leq N^{A_0}$, the number of positive solutions to \[ au^2-bv^2=h,\qquad u,v\leq N^{A_0}, \] is $N^{o_{A_0}(1)}$. \end{lemma} \begin{proof} If $ab$ is a square, multiply the equation by $a$ and factor the resulting difference of squares; the divisor bound suffices. Otherwise write $ab=Ds^2$ with $D$ squarefree and nonsquare. Each solution gives the algebraic integer $\xi=au+sv\sqrt D$ of norm $ah$ in the maximal order of $\mathbb Q(\sqrt D)$. Its principal ideal has norm $|ah|$. The number of ideals of this norm is at most $\tau(|ah|)$: at a split prime there are at most $e+1$ ideals of norm $\ell^e$, and at an inert or ramified prime there are no more. Elements generating the same ideal differ by a unit. A real quadratic unit greater than one is at least $(1+\sqrt5)/2$, as follows from its integral trace and norm $\pm1$. The positive values of $\xi$ arising here lie in $[1,N^{O(A_0)}]$; there are therefore $O_{A_0}(\log N)$ possibilities in each ideal class of generators. Finally $\tau(|ah|)=N^{o_{A_0}(1)}$. This argument makes the coefficient uniformity explicit and requires no bound for a fundamental unit or a class number. \end{proof} \begin{lemma}[Local and global defect counts]\label{p1:arith:defects} Suppose $B\asymp\log N$ and $H\asymp B$, with fixed comparison constants. Any interval of diameter $O(B)$ in $[cN,CN]$ contains $O(B/\log B)$ $H$-defective integers. Moreover \begin{align} \#\{n\leq CN:K_H(n)=1\}&\leq N^{1/2+o(1)},\label{p1:arith:onedfect}\\ \#\{x\in I_N:V_x\text{ contains two }H\text{-defective integers}\} &\leq N^{o(1)}.\label{p1:arith:twodefect} \end{align} All statements remain valid after fixed dilations and translations of the intervals. \end{lemma} \begin{proof} For $D$ defective integers in a short interval, form the binary matrix with their small-prime valuation columns and an additional parity row. Its kernel has redundancy at most $\pi(H)+1$. A nonzero word selects an even square product, so Lemma~\ref{p1:arith:runge} gives minimum distance $\delta\gg\log N/\log B$. Choose $t\asymp\log N/\log B$ with $2t<\delta$. The Hamming bound gives \[ \sum_{j=0}^t\binom Dj\leq2^{\pi(H)+1}. \] If $D\geq2t$, use $\binom Dt\geq(D/t)^t$. Since $(\pi(H)+1)/t=O(1)$, this proves $D=O(B/\log B)$; the other case is immediate. Write a defective integer uniquely as $n=a^2c$, where $c$ is squarefree and $H$-smooth. Then \[ \#\{n\leq CN:K_H(n)=1\} \leq\sqrt{CN}\prod_{\ell\leq H}(1+\ell^{-1/2}) \leq\sqrt{CN}\exp(O(\sqrt H)). \] For two distinct defective values in one window, write $x+i=c_1a^2$, $x+j=c_2b^2$. The offsets and the squarefree smooth parts have $B^2 4^{\pi(H)}=N^{o(1)}$ choices. All their heights are polynomial in $N$, and $c_1a^2-c_2b^2=i-j\ne0$. Lemma~\ref{p1:arith:pell} supplies $N^{o(1)}$ possibilities. This proves \eqref{p1:arith:twodefect} and the stated uniform versions. \end{proof} \subsection{Two small kernels with their smooth coefficients retained} The following form of the approximate determinant method is needed because the small-prime parts are fixed separately and may be polynomially large. It is an extension of the proof of \cite[Sections~2.1--2.3]{reuss2014}, rather than an application of its fixed-coefficient statement. \begin{lemma}[Uniform determinant estimate]\label{p1:arith:reuss} Fix $A_0,\delta,\varepsilon>0$. Suppose $1\leq c_1,c_2\leq3N$, $0<|h|\leq(\log N)^{A_0}$, and \[ c_1D^2U\asymp c_2E^2V\asymp N,\quad D,E\ll\sqrt N,\quad c_1c_2(DE)^2\geq N, \quad\log(DE),\log(UV)\geq\delta\log N. \] The comparison constants are fixed. The number of positive solutions $d\asymp D,e\asymp E,u\asymp U,v\asymp V$ to $c_2e^2v-c_1d^2u=h$ is \begin{equation}\label{p1:arith:reussbound} \ll_{A_0,\delta,\varepsilon}N^\varepsilon \min\{(DE\mathcal M)^{1/2}+D+E, (UV\mathcal M)^{1/2}+U+V\}, \qquad \log\mathcal M=\frac98\frac{\log(DE)\log(UV)}{\log N}. \end{equation} Fixed or polylogarithmic factors in the lower bound for $c_1c_2(DE)^2$ can be absorbed by decreasing the error margin. \end{lemma} \paragraph{Parameters in the determinant argument.} Here and in the proof, height of a polynomial means the maximum absolute value of its integer coefficients. The ledger below distinguishes the fixed auxiliary degrees from the coefficients that vary with $N$. The symbols $A,B_0,\kappa,L_1,T_1$ are specified in Steps~2--3. \begin{center} \small \begin{tabularx}{\textwidth}{@{}>{\raggedright\arraybackslash}p{.23\textwidth}>{\raggedright\arraybackslash}p{.19\textwidth}>{\raggedright\arraybackslash}p{.22\textwidth}>{\raggedright\arraybackslash}X@{}} \toprule Object and dimension & Degree & Point or box scale & Coefficient height\\ \midrule Input hypersurface in $\mathbb A^4$, dimension $3$ & Total degree $3$ & $(D,E,U,V)$, each $O(N)$ & $c_1,c_2\leq3N$; $|h|\leq(\log N)^{A_0}$\\ Auxiliary curve in $\mathbb P^1\!\times\!\mathbb P^1$, dimension $1$ & Bidegree $(a,b)\leq(A,B_0)$, fixed & Primitive pairs of heights $L_1,T_1\leq N^C$ & $\|G\|\leq N^\kappa$ after the lattice changes\\ Eliminant curve in $\mathbb A^2$, when $a=1$ or $b=1$ & Bounded in terms of $A,B_0$ & The same primitive boxes & Resultants and nonzero right-hand sides $N^{O(\kappa)}$\\ \bottomrule \end{tabularx} \end{center} All displayed height exponents, including $C$, are fixed, uniformly over the input boxes. The relative chart has errors $\alpha=O(M^{-1})$ and $\beta=O(|h|/N)$, with $M=N^\zeta\mathcal M$; no coefficient ratio is charged to the latter error. Choose the final $\varepsilon$ first, then a sufficiently small $\zeta$, then $A,B_0$ and $\kappa$, and finally the curve-count exponent $\epsilon'$ so that $(C+\kappa)\epsilon'\leq\zeta$. Only after these choices does $N$ tend to infinity. In particular, the factor $\|G\|^{\epsilon'}$ in Heath-Brown's bound is explicitly included in this budget. The coefficient uniformity in this ledger is established below; Reuss's stated theorem allows its constant to depend on the fixed shift $h$ and does not itself assert this extension. \begin{proof} The proof has four steps. They identify separately the determinant scale, the uniform count on each auxiliary curve, and the two lattice summations. Only the last step sums over the short intervals. \medskip\noindent \textbf{Step 1: primitive pairs and relative normalization.} If $g=\gcd(d,e)$ and $q=\gcd(u,v)$, then $g^2q\mid h$. Divide by $g^2q$ and replace the four dyadic scales accordingly. Both pairs are now primitive. The effective size $N/(g^2q)$ and the hypotheses on the scales change only by polylogarithmic factors; in particular the lower bound for $c_1c_2(DE)^2$ has the permitted polylogarithmic loss. There are only polylogarithmically many reductions, which will be summed at the end. We keep the notation $N,D,E,U,V,h$ for a reduced box. Set $s=d/e$, $t=v/u$. The exact curve approximation is \[ t=(c_1/c_2)s^2+h/(c_2e^2u). \] On a relative interval of length $M^{-1}$ write \[ s=s_0(1+\alpha),\qquad t=t_0(1+2\alpha+\alpha^2+\beta),\qquad t_0=(c_1/c_2)s_0^2\asymp V/U. \] Then $|\alpha|\ll M^{-1}$ and $|\beta|\ll|h|/N$, uniformly in $c_1,c_2$. The relevant vertical error is relative to $t_0$; its bound therefore contains no additional factor $c_1/c_2$. \medskip\noindent \textbf{Step 2: the determinant with variable coefficients.} For the auxiliary monomials $s^at^b$, $0\leq a\leq A$, $0\leq b\leq B_0$, the row denominator is $e^Au^{B_0}$. Extracting $s_0^at_0^b$ from the columns leaves the same logarithmic denominator cost as in the normalized calculation: \[ A\log E+B_0\log U+\frac A2\log(D/E)+\frac{B_0}2\log(V/U) =\frac A2\log(DE)+\frac{B_0}2\log(UV). \] Take $M=N^\zeta\mathcal M$, with a small fixed $\zeta>0$, and use $N/|h|$ for the vertical approximation scale. The strict determinant inequality in \cite[Section~2.1, Lemma~7]{reuss2014} then holds, with degrees fixed in terms of $\zeta,\delta$. The resulting integer polynomial has coefficient height $N^{\kappa(\zeta,\delta)}$. The exponent is uniform in $c_1,c_2$: their sizes, and the rational column scalings, are polynomial in $N$, while the degrees are fixed. To check the allowed scale, put $z=\log(DE)/\log N$ and $w=1-\log(c_1c_2)/(2\log N)$. Then \[ \frac{\log(UV)}{\log N}=2w-2z+O(1/\log N),\qquad \frac{\log\mathcal M}{\log N}=\frac94z(w-z)+o(1) \leq\frac9{16}+o(1). \] Thus $M\leq N/|h|$ for sufficiently small $\zeta$. The two logarithmic size ratios stay in a compact subset of the positive quadrant, so the degree choices are uniform over all boxes. The parameter order is: choose $\zeta$ from the final error margin, then the auxiliary degrees and the height exponent $\kappa$, and only then the internal exponent in the curve count below. \medskip\noindent \textbf{Step 3: fibres and the count on one auxiliary curve.} Apply the two unimodular lattice changes of \cite[Section~2.2]{reuss2014} to $(d,e)$ and $(v,u)$. Write the resulting primitive pairs as $\lambda=(\lambda_1,\lambda_2)$ and $\tau=(\tau_1,\tau_2)$, in boxes of side lengths $L_1\geq L_2$ and $T_1\geq T_2$, respectively. The changes of coordinates have polynomial height; enlarging $\kappa$ absorbs their contribution to the heights of the original and auxiliary equations. The original equation still has bidegree $(2,1)$. An auxiliary factor of degree zero in $(d,e)$ fixes only boundedly many primitive $(u,v)$; Lemma~\ref{p1:arith:pell} counts the remaining quadratic equation uniformly. For degree zero in $(u,v)$, fix $(d,e)$ and set $g_f=\gcd(c_1d^2,c_2e^2)$. Solvability requires $g_f\mid h$. The step of the resulting one-dimensional lattice in $u$ is $c_2e^2/g_f$; the step in $v$ is $c_1d^2/g_f$. Its intersection with the box has at most \[ O\left(1+\frac{Ng_f}{c_1c_2(DE)^2}\right) \leq O\left(1+\frac{N|h|}{c_1c_2(DE)^2}\right) \] points. This is polylogarithmic by the stated lower bound. For an absolutely irreducible factor of bidegree $(a,b)$, $a,b\geq1$, \cite[Lemma~2]{heathbrown2012} bounds primitive solutions by $O_{a,b,\epsilon'}(X^{2/b+\epsilon'}\|G\|^{\epsilon'})$ when the first pair has height at most $X$; the second pair has no height restriction. Thus $b\geq2$ gives the required bound in $L_1$, and $a\geq2$ gives the bound in $T_1$ after exchanging the pairs. Choose $\epsilon'$ after $\kappa$ so that the coefficient-height factor costs at most the chosen power $N^{O(\zeta)}$. The missing bound when one bidegree is one follows directly by elimination. In bidegree $(1,b)$ write $\lambda_1F_1(\tau)+\lambda_2F_2(\tau)=0$. Primitivity gives $q\lambda_1=F_2(\tau)$, $q\lambda_2=-F_1(\tau)$ for a nonzero integer $q$. The two dehomogenizations of the coprime forms $F_1,F_2$ give nonzero resultants $R,R'$ of height $N^{O(\kappa)}$ and \[ q\mid R\tau_1^m,\qquad q\mid R'\tau_2^{m'}. \] Since $(\tau_1,\tau_2)=1$, prime by prime $q$ divides $RR'$. There are $N^{o(1)}$ possibilities, uniformly in the coefficients. Substitution into the original equation gives $F_3(\tau_1,\tau_2)=hq^2$, where $F_3$ is homogeneous of fixed degree $2b+1$. If $F_3$ is zero there is no solution. Otherwise $F_3-hq^2$ is a nonzero polynomial. Fixing the first coordinate leaves boundedly many second coordinates, except for boundedly many full fibres, each of length $O(T_2)$. As $T_2\ll T_1$, the count is $O(T_1)$. This includes powers of a linear form and needs no uniform height-free bound for a Thue equation. Exchanging the pairs handles bidegree $(a,1)$ with count $O(L_1)$ and right-hand side $hq$. These cases prove, with arbitrarily small internal losses, the local bound of \cite[Lemma~8]{reuss2014} in both orientations. Reducible auxiliary polynomials have boundedly many factors. Rational points on an irreducible factor that is not geometrically integral lie in its bounded-degree conjugate intersections. The resulting local count is therefore uniform in all the variable coefficients. \medskip\noindent \textbf{Step 4: the two short-vector summations.} We give the scale check behind \cite[Section~2.3]{reuss2014}. For the first lattice, use intervals with $s_0=(D/E)x_3/M$, $x_3\asymp M$. A shortest primitive vector $(x_1,x_2)$ in a box with $L0$, \[ G_2(N,N^\beta):= \#\{x\in I_N:\#\{n\in V_x:K_B(n)\leq N^\beta\}\geq2\} \ll_{\beta,C,\varepsilon}N^{\max(\beta,7/12)+\varepsilon}. \] For $7/12\leq\beta<1$ this exponent is sharp in the sense $G_2(N,N^\beta)=N^{\beta+o(1)}$. \end{lemma} \begin{proof} Two selected integers have the unique forms $n=c_1d^2u$ and $n+h=c_2e^2v$, where $1\leq h1/2$. Both remaining cofactors are at most $T$. The total number of incidences is $\gg D^2N/D^4\asymp T$. A fixed $n$ is counted at most $\tau(n)\tau(n+1)=N^{o(1)}$ times. The two neighboring integers lie in a common window, proving the claim. \end{proof} \begin{lemma}[A large selected kernel]\label{p1:arith:anchor} For any fixed $0<\tau<1/2$ and $1\leq T\leq3N$, the number of separated pairs possessing a mixed square relation with a selected vertex whose $B$-kernel exceeds $T$ is at most \[ N^{3/2+o(1)}T^{-1/2+\tau}. \] Also the number of windows containing one vertex with kernel at most $T$ is at most $N^{1/2+o(1)}T^{1/2}$. \end{lemma} \begin{proof} Write $n=a^2cr$, with $c$ squarefree and $B$-smooth and $r=K_B(n)$. The one-kernel assertion follows from \[ \sum_{\substack{n\leq3N\\K_B(n)\leq T}}1 \leq\sqrt{3N}\prod_{\ell\leq B}(1+\ell^{-1/2}) \sum_{r\leq T}r^{-1/2} \leq N^{1/2+o(1)}T^{1/2}, \] with a factor $B$ for its placements. For the first assertion anchor a selected value $n=x+i$ with $K_B(n)>T$. Every prime of its kernel must occur at one of the $B$ offsets of the other window. There are at most $B^{\omega(r)}$ assignments; each specifies a residue class modulo $r$. Since $r\leq3N$, there are $O(N/r)$ possible other starts, including $r>N$. Summation over anchors is bounded by \[ B N^{3/2}\prod_{\ell\leq B}(1+\ell^{-1/2}) \sum_{\substack{r>T\ {\rm squarefree}\\\ell\mid r\Rightarrow\ell>B}} B^{\omega(r)}r^{-3/2}. \] Rankin's inequality bounds the last sum by \[ T^{-1/2+\tau}\prod_{\ell>B}(1+B\ell^{-1-\tau}) =T^{-1/2+\tau}\exp(O_\tau(B^{1-\tau})). \] Both products and the placement factor are $N^{o(1)}$. \end{proof} \subsection{Multiquadratic surfaces and the curves that can survive} We use two uniform determinant estimates. A geometrically integral affine surface of degree $d$ in $\mathbb A^3$ has, outside $O_{d,\varepsilon}(H^{2/\sqrt d+\varepsilon})$ integral points of height at most $H$, a covering by $O_{d,\varepsilon}(H^{1/\sqrt d+\varepsilon})$ curves of bounded degree \cite[Lemma~2]{salberger2023}. An integral curve of degree $e$ in fixed affine space has $O_{e,n,\varepsilon}(H^{1/e+\varepsilon})$ integral points in the same box \cite[Theorem~A]{pila1995}. These constants do not depend on the defining coefficients. For a curve irreducible over $\mathbb Q$ but not geometrically integral, its rational points lie in the intersection of distinct conjugate components and have bounded cardinality. Consider equations \begin{equation}\label{p1:arith:surfaceequations} w_\nu^2=d_\nu P_\nu(X)Q_\nu(Y),\quad P_\nu(X)=\prod_{i\in I_\nu}(X+i),\quad Q_\nu(Y)=\prod_{j\in J_\nu}(Y+j),\quad1\leq\nu\leq k, \end{equation} where $I_\nu,J_\nu$ are nonempty, disjoint within each list, and $|I_\nu|+|J_\nu|\leq K$. The offsets are $O(B)$ and the nonzero coefficients $d_\nu$ are squarefree and $O(B)$-smooth. Constants $k,K$ are fixed. \begin{lemma}[The uniform geometric counts]\label{p1:arith:geometric} Fix $A_0,K,\varepsilon>0$. For sufficiently large fixed $k$, a system of the form \eqref{p1:arith:surfaceequations}, with arbitrary nonzero integer coefficients $|d_\nu|\leq N^{A_0}$, has $O_{A_0,K,\varepsilon}(N^{1/2+\varepsilon})$ integral hosts $X,Y\asymp N$ at which the even-selection square-class maps on the two actual sets of selected values are injective and their images are in direct sum. If every $Q_\nu$ has degree one, the bound is $O_{A_0,K,\varepsilon}(N^\varepsilon)$. If the degrees of the $P_\nu$ vary, select sufficiently many equations of one of their finitely many possible degrees before applying the assertion. These estimates are uniform in the coefficients and do not require their smoothness; any enumeration of such coefficients is a separate cost. They remain valid after other fixed centers have been specialized into the $d_\nu$. For one variable, a bounded-degree system of split-product square equations with polynomial-height coefficients has $N^{o(1)}$ possible integer values as soon as its disjoint root sets include at least two distinct selected offsets. This includes either one equation of degree at least two or two equations with disjoint nonempty root sets. The following graph consequences hold, with the number of component equations chosen sufficiently large after the stated fixed parameters. \begin{enumerate} \item For any $\eta,\varepsilon>0$, pairs of separated nonaligned logarithmic windows with at least $\eta B$ nontrivial unpinned components number $O_{\eta,C,\varepsilon}(N^{1/2+\varepsilon})$. \item For every fixed integer $a\geq1$ and $\varepsilon>0$, there is $k_a$ such that $k_a$ components having $a$ vertices in one window and one in the other have $O_{a,C,\varepsilon}(N^\varepsilon)$ dyadic hosts. For $a=1$ aligned hosts are excluded; for $a\geq2$ no such exclusion is required. \item After one start is fixed, a component with at least two vertices in the other window has $N^{o(1)}$ possible partners, uniformly for a fixed bound on the component size and polynomial coefficient heights. \end{enumerate} \end{lemma} \paragraph{Parameters in the geometric counts.} The external estimates are applied in the following embeddings. Write $m=\binom{D+2}{2}+k-1$ for the ambient dimension of the affine Veronese chart, and $q=(a+1+A_0)/2$ in the degree-$(a,1)$ case. \begin{center} \small \begin{tabularx}{\textwidth}{@{}>{\raggedright\arraybackslash}p{.23\textwidth}>{\raggedright\arraybackslash}p{.23\textwidth}>{\raggedright\arraybackslash}p{.15\textwidth}>{\raggedright\arraybackslash}X@{}} \toprule External count & Dimension and degree & Point height & Coefficient height\\ \midrule Surface covering, Steps~1--4 & Surface in $\mathbb A^3$; $2^k\leq d_{\rm im}$, $d_{\rm im}=O_{k,K}(1)$ & $H\ll N^D$ & Polynomial in $N$; no coefficient-height dependence in the covering constant\\ Lifted curves, Steps~2--4 & Curves in $\mathbb A^m$; $2D\leq e=O_{D,k,K}(1)$ & $H\ll N^D$ & No bound on the coefficients of the covering curves is required\\ Degree-$(a,1)$ surface covering, Steps~5--6 & Surface in $\mathbb A^3$; $2^k\leq d_{\rm im}$, $d_{\rm im}=O_{k,a}(1)$ & $H\ll N^q$ & Polynomial in $N$; the same coefficient-uniform covering estimate\\ Degree-$(a,1)$ lifted curves, Steps~5--6 & Curves in $\mathbb A^{k+2}$; $E+1\leq e=O_{k,a}(1)$ & $H\ll N^q$ & No coefficient-height bound is required by the curve estimate\\ \bottomrule \end{tabularx} \end{center} The curve rows concern components that can contain the hosts being counted; Steps~3 and~5--6 exclude the lower degrees. The input coefficients satisfy $|d_\nu|\leq N^{A_0}$, and the offsets are $O(\log N)$. Thus the defining equations and the projected surface equations have polynomial height with fixed exponents, while the external constants depend only on the fixed dimensions, degrees and internal error margins. For the first bound choose $A_0,K,\varepsilon$, then $D\geq(K+A_0)/2$, then $k$, and finally the internal margin in the surface and curve counts. For the degree-$(a,1)$ bound choose $q$, then $E$ with $q/(E+1)<\varepsilon/8$, then $k$ satisfying the two conditions in Step~6, and finally that internal margin. All these choices precede $N\to\infty$. The count for one system is uniform over arbitrary coefficients in the stated height range; the $N^{o(1)}$ enumeration in the graph application uses smoothness separately. \begin{proof} We first count one fixed system, keeping the coefficient height visible; only afterwards do we sum the patterns arising from the graph. We will retain a sufficiently large fixed number \(k\) of equations, chosen after the height parameter \(D\) below. All degrees and ambient dimensions below depend on that number and on the fixed parameters, not on the values of the coefficients. \medskip\noindent \textbf{Step 1: geometric integrality and the finite affine projection.} Over an algebraically closed field of characteristic zero, the square classes on the right of \eqref{p1:arith:surfaceequations} are independent: the valuation along a factor $X+i$ detects exactly its own equation. The coordinate ring is free of rank $2^k$ over the polynomial ring in $X,Y$, and its generic localization is a field. It is therefore integral. The affine projection $(X,Y,w)\mapsto(X,Y,\sum_\nu w_\nu)$ is finite and birational onto its image. Finiteness follows from integrality over $\mathbb Q[X,Y]$. Birationality follows because no nontrivial sign change fixes the sum: the individual radicals are independent basis elements of the multiquadratic extension. The image surface in $\mathbb A^3$ has degree at least $2^k$ and degree bounded in terms of $k,K$. The projection need not extend across the projective boundary; only its affine finiteness is used when curves are lifted. \medskip\noindent \textbf{Step 2: projective degrees and the height of integral points.} Choose an integer $D\geq(K+A_0)/2$ before choosing $k$, and homogenize in the weighted ring \[ \mathbb Q[X,Y,Z,W_1,\ldots,W_k]/ \left(W_\nu^2-d_\nu Z^{2D-|I_\nu|-|J_\nu|} \prod_{i\in I_\nu}(X+iZ)\prod_{j\in J_\nu}(Y+jZ)\right), \] with $\deg X=\deg Y=\deg Z=1$, $\deg W_\nu=D$. Its $D$th Veronese ring is generated in degree one by the degree-$D$ monomials in $X,Y,Z$ and the $W_\nu$. This embeds an integral projective surface finite of degree $2^k$ over $\mathbb P^2$. The hyperplane bundle is the pullback of $\mathcal O_{\mathbb P^2}(D)$, so its degree is $2^kD^2$ and every integral curve $C$ satisfies \begin{equation}\label{p1:arith:curvedegree} \deg C=D\deg\sigma(C)\,[\mathbb Q(C):\mathbb Q(\sigma(C))]. \end{equation} Here $\sigma$ is the finite map to $\mathbb P^2$. At an integral point in the dyadic region, $|w_\nu|\ll N^{(K+A_0)/2}$. In the chart $Z=1$, the Veronese coordinates are integers of height $H\ll N^D$, including the $W_\nu$. The same bound holds for the affine projection in Step~1. Thus both applications of the determinant estimates use one common height bound. \medskip\noindent \textbf{Step 3: excluding the curves of smallest possible degree.} A curve of degree $D$ maps with degree one to a line. A vertical or horizontal line meeting the positive region is impossible: a nonconstant product of distinct linear factors in the other variable would have to be a rational square. On an oblique line, square products force every root on one side to match exactly one root on the other. Two component equations give two distinct exact units, hence a positive rational alignment of height $O(B)$. Thus a nonaligned point is on no curve of degree $D$; by \eqref{p1:arith:curvedegree}, its curves have degree at least $2D$. In the direct-sum formulation the exceptional degree-$D$ curves are excluded for a more intrinsic reason: their two exact units give nonempty even selections, two values on each side, with equal square classes. Injectivity makes both classes nonzero, contradicting the direct sum. Thus the argument does not require absence of geometric alignment when the actual spaces are already in direct sum. \medskip\noindent \textbf{Step 4: the surface covering and the first graph consequence.} Apply the surface covering, lift its curves by the finite affine map, and take their closures in this embedding. Their numbers of components and degrees are bounded in terms of the fixed parameters by B\'ezout. Pila's estimate and the degree lower bound $2D$ give, per pattern, \[ O\big(N^{2D2^{-k/2}+\varepsilon'} +N^{1/2+D2^{-k/2}+\varepsilon'}\big). \] Choose $k$ large enough after $D,\varepsilon$ and reduce the internal margin $\varepsilon'$. This proves the per-system bound. For the graph consequence, smoothness gives $|d_\nu|\leq N^{A_0}$ for a fixed $A_0$. At least half the components in a population $c\geq\eta B$ have size at most $K=\lceil4/\eta\rceil$. Select $k$ of them. The offsets and smooth coefficients have $B^{O_K(k)}2^{k\pi(O(B))}=N^{o(1)}$ patterns. Choosing $k$ large after $D,\varepsilon$, and summing these patterns, proves the first assertion. \medskip\noindent \textbf{Step 5: reduced fibres for equations of type $(a,1)$.} For the stronger degree-one assertion we exclude all curves of a fixed degree at most $E$, before making $k$ large. For this part use the ordinary affine embedding with coordinates $X,Y,w_1,\ldots,w_k$. Write the equations as $w_\nu^2=d_\nu P_\nu(X)(Y+j_\nu)$ with $\deg P_\nu=a$. On the normalization of an integral curve of degree $e\leq E$, the coordinate functions $X,Y$ have degrees $m_X,m_Y\leq e$ and the genus is at most $(e-1)(e-2)/2$. Neither function can be constant: otherwise the many distinct roots in the other variable would all be branch values, contradicting Riemann--Hurwitz. Discard equations having a branch value among their selected roots, or a finite fibre meeting the pole divisor of the other coordinate. At most \[ (2g-2+2m_X)+(2g-2+2m_Y)+m_X+m_Y\leq2e^2 \] equations are lost. For each remaining equation all finite fibres are reduced and avoid the poles. Parity of divisors gives \[ F_Y(-j_\nu)=\sum_{i\in I_\nu}F_X(-i),\qquad m_Y=am_X. \] In fact \[ \operatorname{div}\!\left(\frac{P_\nu(X)}{Y+j_\nu}\right) =F_Y(\infty)-aF_X(\infty), \] independently of $\nu$. For two good equations the quotients therefore differ by a nonzero constant. Eliminating this common quotient gives $Y=R(X)$ with $\deg R\leq a$. Since $m_Y=(\deg R)m_X=am_X$, the degree is exactly $a$. \medskip\noindent \textbf{Step 6: interpolation on the offsets and exclusion of low-degree curves.} If $k\geq4(E+1)^2$, there are at least three good equations. Their $3a$ distinct grid points $(-i,-j_\nu)$ lie on the graph of $R$: the reduced-fibre identity shows that $R$ is finite and equals $-j_\nu$ at each selected $-i$. These grid points determine $R$. Choose $2a+1$ of them. Rational interpolation, using maximal minors of the integer evaluation matrix, gives coprime integer polynomials $P,Q$ of degrees at most $a$, coefficients $B^{O_a(1)}$, and $R=P/Q$. The kernel of this interpolation system is one-dimensional: two solutions would have cross-product of degree at most $2a$ vanishing at $2a+1$ distinct abscissae. Since the reduced rational function has degree exactly $a$, no nonconstant polynomial multiple of its coprime numerator and denominator still has both degrees at most $a$. This makes the coefficient uniformity explicit: the interpolation data are the offsets, so their height bound is independent of the $d_\nu$. If $Q$ is nonconstant, the nonzero integer resultant $\operatorname{Res}(P,Q)$ has size $B^{O_a(1)}$ and $Q(X)$ divides it at an integral point. Thus $|X|\leq B^{O_a(1)}$. If $Q$ is constant and $a\geq2$, then $|Y|\gg |X|^a/B^{O_a(1)}$, impossible for $X,Y\asymp N$. If $a=1$ and $Q$ is constant, the graph is a rational line; its grid points give an exact positive alignment. Thus every remaining dyadic nonaligned point lies on a curve of degree greater than $E$. In the ordinary affine embedding the height is $H\ll N^q$ for a fixed $q=(a+1+A_0)/2$. Choose $E$ so that $q/(E+1)<\varepsilon/8$, then $k\geq4(E+1)^2$ so that $2q2^{-k/2}<\varepsilon/8$. The same finite projection, surface covering and curve estimate now give $N^\varepsilon$ per fixed coefficient pattern. For $a=1$, its exceptional line is also excluded under the direct-sum hypothesis: two grid matches give the forbidden even selections on the two actual blocks. Hence the degree-one per-system estimate is proved with arbitrary polynomial-height coefficients. In the graph application one sums the $N^{o(1)}$ smooth patterns, proving the second graph assertion, including $a=1$. \medskip\noindent \textbf{Step 7: fixing one center and reducing to a quadratic equation.} For the last assertion fix the first start and the component data. The other start satisfies $\prod_{j\in J}(Y+j)=ev^2$, with fixed squarefree $e\leq N^{O(1)}$ and $|J|\geq2$. Outside primes dividing $e\prod_{i\ne j}(i-j)$, every selected factor has even valuation. Fixing the squarefree parts of two factors costs $2^{O(\omega(e\prod_{i\ne j}(i-j)))}=N^{o(1)}$ for fixed $|J|$. Their difference is a nonzero fixed offset, so Lemma~\ref{p1:arith:pell} gives $N^{o(1)}$ starts. This is a bounded-height argument; no height-free theorem for all integral points is being invoked. The one-variable assertion with two separate equations follows in exactly the same localization: take one selected factor from each equation and put the prime divisors of both coefficients and both within-equation discriminants into the exceptional prime set. Outside it each chosen factor has even valuation. The two distinct offsets again give one quadratic equation counted by Lemma~\ref{p1:arith:pell}. \end{proof} \begin{lemma}[The dense weighted sector]\label{p1:arith:dense} For every fixed $\eta,\varepsilon>0$, \[ \sum_{\substack{x,y\in I_N,\ |x-y|>L\\ (x,y)\ {\rm nonaligned},\ c(x,y)\geq\eta B}} (2^{\rho(x,y)}-1) \ll_{C,\eta,\varepsilon}N^{1+\varepsilon}. \] The same bound holds for the full value kernel, without block-parity rows. \end{lemma} \begin{proof} Let $t_2,t_3$ count nontrivial unpinned components with two and three vertices. If either exceeds a sufficiently large fixed threshold, Lemma~\ref{p1:arith:geometric}(2), in both orientations for size three, gives $N^\varepsilon$ hosts. Their weight is at most $2^{B+o(B)}=N^{1+o(1)}$, since each other component needs two vertices and the number of defective vertices is $o(B)$. Otherwise $t_2,t_3=O(1)$ and the vertex budget gives $4c-2t_2-t_3\leq2B$, hence $c\leq B/2+O(1)$. Lemma~\ref{p1:arith:geometric}(1) gives $N^{1/2+\varepsilon}$ hosts, each of weight at most $2^{B/2+o(B)}=N^{1/2+o(1)}$. Reduce the internal margins in both estimates. \end{proof} \begin{lemma}[Uniform residual rank on an aligned pair]\label{p1:arith:alignedrank} For separated aligned windows with reduced height $b=\max(a,b)\leq B^3$, let $u$ be the number of exact units. Then \[ \rho(x,y)\leq(u-1)_++o(B)\leq L/b+o(B), \] uniformly over the channels. For fixed channel height $b$ and fixed $D\ge6b$, an enlarged prime cutoff $6bB\le H\le DB$ leaves residual $O_{C,b,D}(B/\log B)$. \end{lemma} \begin{proof} Normalize the two windows to $bV_x$ and $aV_y$, and merge equal values. All remaining distinct integers lie in an interval of diameter $O(B^4)$ centered at an integer at least $N$. Remove prime rows $\ell\leq B$ and rows with $\ell\mid ab$. Every other prime row meets at most two distinct normalized vertices, since it occurred at most once in each original window. Its equations again leave one variable per unpinned component. Let $c_K$ count these components of size at most a fixed $K$. For these components, the removed prime rows and one total-parity row form a binary matrix of redundancy at most $s=\pi(B)+\omega(ab)+1=O(B/\log B)$. A nonzero kernel word gives an even square product of distinct normalized integers. By Lemma~\ref{p1:arith:runge}, it uses at least $\delta\gg\log N/\log B$ vertices, and hence at least $\delta/K$ components. The Hamming argument in Lemma~\ref{p1:arith:defects} yields $c_K=O_{C,K}(B/\log B)$. Components larger than $K$ number at most $2B/K$. Since $K$ can be chosen arbitrarily large before $N$, the dimension of the value kernel on the distinct normalized set is $o(B)$. The kernel of the merging map on the original relative selections is exactly the even-selection space of the $u$ exact units, of dimension $(u-1)_+$. Its image lies in the value kernel just bounded, proving the first assertion. For fixed height the normalized interval has diameter $O(B)$. Private primes above a sufficiently large fixed multiple of $B$ leave only $O(B/\log B)$ defective normalized vertices, directly by Lemma~\ref{p1:arith:defects}; this gives the last assertion. \end{proof} \subsection{Asymmetric extensions and homogeneous seeds} \begin{lemma}[An asymmetric pair profile]\label{p1:arith:asymmetric} Fix integers $D,Q\geq1$ and $D_{\rm cut}\geq2D+1$. Let $B=L+1$ and $|L-\log N/\log2|\leq C$. For each $N\leq X5B^4$ but aligned at radius two; its dyadic mass of order $N^{4/3}$ refutes a radius-one estimate for arbitrary $D$. Compressed-source applications use the same enlarged radius on both sides of the alignment partition. \begin{lemma}[Aligned loss above a rational pair]\label{p1:arith:palmrank} Fix a geometrically good pair in channel $a:b=2:3$ or $3:4$, regardless of its affine compatibility. Among targets $z$ with $|z-x|,|z-y|\geq L$ and distinct from the two sources, there are $B^{O(1)}$ aligned targets per source. Except for $B^{O(1)}$ source pairs at dyadic endpoints, their additional relative rank loss modulo the old value space is at most \[ L/3+o(L)\quad(a:b=2:3),\qquad L/2+o(L)\quad(a:b=3:4). \] The result allows the prime bits of $a,b$ in the old information. \end{lemma} \begin{proof} Represent the source by $x=at-i_0$, $y=bt-j_0$ and put $X=abt$. The true old normalized values form a subset of $X+[-bB,bB]$. Applying the residual argument of Lemma~\ref{p1:arith:alignedrank} to a fixed-width enclosing interval and the new interval leaves $o(B)$ dimensions after exact units are removed. If an exact unit of the enlarged graph uses an old value absent from the true set, its new nondefective vertex is pinned in the true graph. Consequently the remaining loss is at most the number of new vertices matched to actual old values, plus $o(B)$. Write an aligned target as $z=(ct+h)/d$, with $c,d$ positive coprime, $c,d\leq B^{O(1)}$ and $h/d=O(B)$. A match to the first old family requires $h+dk\equiv0\pmod{c/(c,a)}$; the second family requires the same condition modulo $c/(c,b)$. Their intersection is the class modulo $c$, because $(a,b)=1$. Thus the union has at most \begin{equation}\label{p1:arith:matchdensity} B\frac{(c,a)+(c,b)-1}{c}+O(1) \end{equation} new offsets. Conditions on the true intervals can only reduce this count. For $2:3$, the source range is $N/2-O(B)\leq t\leq2N/3+O(B)$, so $3/2\leq c/d\leq4$ apart from endpoint errors of order $B/N$. A distinct rational of polynomial height cannot cross either endpoint by such a small amount. For $c\geq7$ the density in \eqref{p1:arith:matchdensity} is at most $1/3$ (check $7\leq c\leq12$, then bound the numerator by four). For $c=5$ it is $1/5$. The cases $c/d=2,3$ physically reproduce one old family. Their supports intersect the new support in at most one integer, because the starts are at distance at least $L$. The other family supplies at most $B/3+O(1)$ units, counted in its actual window of width $B$. The only remaining possibilities with $c\leq6$ are $c/d=4,3/2$; these force $t=N/2+O(B)$ or $t=2N/3+O(B)$. They account for only $B^{O(1)}$ source pairs. For $3:4$, one has $N/3-O(B)\leq t\leq N/2+O(B)$ and $2\leq c/d\leq6$. The density is at most $1/2$ for $c\geq9$ (check $9\leq c\leq12$, then use numerator at most six), and also for $c=5,7,8$. The physical cases $c/d=3,4$ leave at most one unit in one old family and $B/4+O(1)$ in the other. The remaining fractions $2,6$ force $t=N/2+O(B)$ or $N/3+O(B)$ and are endpoint sources. Finally the canonical slopes and displacements have polynomially many choices in $B$, proving the target count. \end{proof} \begin{proposition}[Homogeneous seeds for fixed extensions] \label{p1:arith:fixed-homogeneous-seeds} Let $\mathcal A_{\rm raw}$ be the geometric good-edge family obtained by omitting the compatibility condition in the definition of $\mathcal E_N$. For a finite set $I\subset G_N$ put \[ W(I)=p^{|I|}2^{\rho(I)}. \] All configurations in the following sums have distinct, pairwise separated starts. For every $\varepsilon>0$: \begin{enumerate} \item The sum of $W(\{x,y\})$ over dependent pairs outside $\mathcal A_{\rm raw}$ is $O_{C,\varepsilon}(N^{-4/5+\varepsilon})$. Moreover $\sum_{e\in\mathcal A_{\rm raw}}W(e)\ll_C LN^{-2/3}$. \item For a geometric good source $e\in\mathcal A_{\rm raw}$ in channel $a:b\in\{2:3,3:4\}$, sum over targets $z$ for which $S_z\cap\sum_{x\in e}S_x\ne\{0\}$. Then \[ \sum_{e,z}W(e\cup\{z\})\ll_{C,\varepsilon}N^{-41/30+\varepsilon}. \] \item For two distinct geometric good edges, whose union has three or four starts, \[ \sum_{e\ne f}W(e\cup f) \ll_C L^2N^{-4/3}. \] \end{enumerate} These are rank bounds without any affine compatibility restriction. \end{proposition} \begin{proof} The homogeneous indicators $H_x=\ind_{\{f(x-1)=\cdots=f(x+L-1)\}}$ have zero affine target, so $\E\prod_{x\in I}H_x=W(I)$. The first assertion follows from the positive profile and the exceptional-channel estimates in Theorem~\ref{p1:arith:pairs}; on a dependent pair, $2^\rho\le2(2^\rho-1)$. For (2), the good source is dependent, so positive crossing gives a connected triple. The separated homogeneous variant in Section~\ref{p14:comp:third:aligned-family}, with at most three marked source pairs per triple, gives $O(N^{-41/30+\varepsilon})$. \paragraph{Two interacting raw edges.} Write $W_e=\sum_{x\in e}S_x$ and sum over ordered pairs: \[ \mathfrak I_N= \sum_{\substack{e,f\in\mathcal A_{\rm raw},\ e\ne f\\ e\cup f\ {\rm separated},\ W_e\cap W_f\ne\{0\}}}W(e\cup f). \] We prove $\mathfrak I_N\ll_{C,\varepsilon}N^{-41/30+\varepsilon}$. If the two edges share a vertex, their union is a connected triple; the same homogeneous estimate, with bounded marking multiplicity, applies. For disjoint edges first treat nonaligned compressed blocks, using a fixed radius enclosing both true value sets in the alignment test. Its complementary enlarged aligned sector is retained below. Write $r_f=|U_f|-1\geq5L/3-O(1)$ and let $\delta^*$ be the loss of the even-selection rank on $U_f$ after the old normalized values and small primes have been exposed. The excess above $W(e)W(f)$ is at most $W(e)W(f)(2^{\delta^*}-1)$. The weak-component support proof of Lemma~\ref{p1:arith:asymmetric} still applies to two subsets of fixed-width logarithmic intervals: it needs two selected new values and one selected old value. It gives $N^{31/24+\varepsilon}$ hosts. In this sector $W(e)W(f)\leq N^{-10/3+o(1)}$, so its mass is $N^{-49/24+\varepsilon}$. In the dense sector, the multiquadratic argument gives $N^{1/2+\varepsilon}$ hosts. A fixed large population of components with one new vertex and boundedly many old vertices has only $N^\varepsilon$ hosts, by Lemma~\ref{p1:arith:geometric}(2). On that exceptional population use $W(f)2^{\delta^*}\leq1$ and $W(e)\leq N^{-5/3+o(1)}$. Otherwise all but $O(B/K)+O_K(1)$ components use two new vertices. Hence \[ \delta^*\leq|U_f|/2+O(B/K)+o(B),\qquad W(f)2^{\delta^*}\leq N^{-5/6+O(1/K)+o(1)}. \] The remaining dense mass is $N^{-2+\varepsilon}$. The entire nonaligned excess is therefore $N^{-5/3+\varepsilon}$. On an interacting pair let $\delta=\dim(W_e\cap W_f)>0$. Since $2^\delta\le2(2^\delta-1)\le2(2^{\delta^*}-1)$, the same nonaligned excess bound controls its full homogeneous mass. For disjoint edges with aligned compressed centres, the fixed channel, offset and alignment codes leave $NB^{O(1)}$ quadruples. Choose $z\in f$. If $S_z\cap W_e\ne0$, the triple $e\cup\{z\}$ is connected. For each such triple the mate of $z$ in $f$ has $O(B)$ choices, because the raw good-edge graph has degree $O(B)$. Since $W(e\cup f)\le W(e\cup\{z\})$, the homogeneous triple bound applies. If $S_z\cap W_e=0$, instead $W(e\cup f)\le W(e\cup\{z\})=W(e)p\le N^{-\kappa_b-1+o(1)}$. The host count gives $N^{-\kappa_b+o(1)}$ in this case. This includes a positive intersection between the two pair spaces even when both individual target crossings vanish. Thus all sectors satisfy the asserted bound for $\mathfrak I_N$. Finally, pairs with disjoint effective spaces factor exactly. Hence \[ \sum_{\substack{e,f\in\mathcal A_{\rm raw},\ e\ne f\\e\cup f\ {\rm separated}}} W(e\cup f) \le \Bigl(\sum_{e\in\mathcal A_{\rm raw}}W(e)\Bigr)^2+\mathfrak I_N. \] The first assertion bounds the squared sum by $O_C(L^2N^{-4/3})$. Fix an internal margin below $1/30$; since $41/30>4/3$, the interaction term is absorbed into that scale with a fixed constant. This proves (3). \end{proof} \subsection{Support and branching at a fixed number of blocks} \begin{lemma}[Small-kernel counts for fixed-width blocks] \label{p1:arith:small-kernels} Let $U_X\subset X+[-DB,DB]$ with $D$ fixed and $X\asymp N$, and let $H\asymp B$ exceed the diameter. Uniformly for $1\leq T\leq CN$, the number of translations containing one value with $K_H(n)\leq T$ is at most $N^{1/2+o(1)}T^{1/2}$. For fixed $0<\theta<1$, the number containing two such distinct values with $T=N^\theta$ is $O_{C,D,\theta,\varepsilon}(N^{\max(\theta,7/12)+\varepsilon})$. In particular it is $O(N^{\theta+\varepsilon})$ for $\theta\geq7/12$. \end{lemma} \begin{proof} Enlarge the selected set to its interval. The proofs of Lemmas~\ref{p1:arith:anchor} and~\ref{p1:arith:twokernel} use only a fixed logarithmic width and a prime threshold of that order. In particular $2^{\pi(H)}=N^{o(1)}$, the smooth parts have fixed polynomial height, and all changes of offset count are powers of $B$. \end{proof} Here a projection is an element of the effective even-selection image in the square-class space, rather than a formal coordinate vector. \begin{proposition}[A relative relation meeting several blocks] \label{p1:arith:fixed-support} Fix integers $r\geq s\geq2$. For $r$ separated logarithmic windows, the number of tuples having a relative square relation whose projection is nonzero in at least $s$ blocks is \[ O_{C,r,\varepsilon}\left(N^{r-s/2+\varepsilon}\right). \] The same assertion holds for effective even-selection images on subsets of fixed-width logarithmic intervals, including the compressed good edges, with only $B^{O_r(1)}$ choices of fixed offsets and fixed dilation factors. \end{proposition} \begin{proof} Choose $s$ blocks on which a witnessing relation has nonzero effective projection, and fix the other $r-s$ centres. Their values generate the old space in Theorem~\ref{p13:half:conditional}. The selected $s$ nonzero projections sum into that old space; they need not remain nonzero modulo it. The theorem therefore gives $O(N^{s/2+\varepsilon'})$ choices for the selected centres. Summing the $O_r(1)$ choices of blocks and $O(N^{r-s})$ other centres gives $O(N^{r-s/2+\varepsilon'})$. The same theorem permits the effective images of all fixed-width blocks in its stated height range. Fixed dilations preserve that range, and the $B^{O_r(1)}$ offset and motif choices are absorbed by reducing $\varepsilon'$. Injectivity of the local selection maps is unnecessary: the projections must be nonzero in their own effective images, even when they vanish modulo the old space. Formal vectors with zero local image are not counted. \end{proof} \begin{proposition}[A fixed number of exposed source windows] \label{p1:arith:fixed-branching} Fix $j\geq1$ and source starts $x_1,\ldots,x_j\in I_N$. Outside a deterministic set of $N^{1/2+o_{C,j}(1)}$ possible starts, the new relative vector is uniform and independent of every value in the old windows. An explicit bound for the exceptional set is \[ j(2L+1)+B\sqrt{3N}\exp(2\sqrt{m_{j,N}}),\qquad m_{j,N}=\left\lceil\pi(B)+\frac{jB\log(3N)}{\log B}\right\rceil. \] Fixed dilations, subsets, and fixed enlarged logarithmic windows satisfy the same $N^{1/2+o(1)}$ conclusion. \end{proposition} \begin{proof} Let $\mathcal P$ contain all primes at most $B$ and all primes dividing an old value. It has at most $m_{j,N}$ members. The number of integers at most $3N$ whose squarefree part is supported on $\mathcal P$ is at most \[ \sqrt{3N}\prod_{\ell\in\mathcal P}(1+\ell^{-1/2}) \leq\sqrt{3N}\exp(2\sqrt{|\mathcal P|}), \] because the $i$th distinct prime is at least $i$. Discard overlapping starts and windows containing two such values; the stated estimate follows even by charging each such window to one of its values and multiplying by $B$. If a nonzero new relative combination belonged to the old value space, its even selection would contain at least two integers. Every large odd prime of each selected integer is private inside the new window, so it would have to divide an old value. Both selected integers would have squarefree part supported on $\mathcal P$, contrary to the retained set. The new rank therefore increases by exactly $L$ over the old value space, which proves the asserted conditional uniformity. For fixed $j$, $m_{j,N}=O_{C,j}((\log N)^2/\log\log N)$. \end{proof} \subsection{Products grouped by their ternary support} \begin{proposition}[A ternary group budget]\label{p1:arith:ternary-budget} Let $U_i\subset X_i+[-DB,DB]$, $X_i\asymp N$, be three sets with injective even-selection square-class maps whose images are pairwise in direct sum. The width constant and the number of offset patterns are fixed. Choose $H$ greater than the three diameters. For an offset triple $u=(u_1,u_2,u_3)$ let \[ G_u=\prod_{\substack{\ell>H\ {\rm prime}\\ v_\ell(X_i+u_i)\ {\rm odd}\ (i=1,2,3)}}\ell. \] Retain the groups greater than one and write their decreasing values $G_{(1)}>G_{(2)}>\cdots$. Put $\mathcal G_k=\prod_{i=1}^kG_{(i)}$, or zero if fewer than $k$ exist. For every fixed $k\geq3$, $\varepsilon>0$, and $1\leq T\leq N$, \begin{equation}\label{p1:arith:ternarycount} \#\{\text{tuples}:\mathcal G_k>T\} \ll_{C,D,k,\varepsilon}N^\varepsilon (N^3/T^2+N^{6/5}). \end{equation} \end{proposition} \begin{proof} A prime larger than $H$ has only one occurrence in each block, so the groups are pairwise coprime and $G_u\mid\gcd(X_1+u_1,X_2+u_2,X_3+u_3)$. Two groups satisfy $G_uG_v\ll NB$. If they share an occurrence this follows from divisibility of that integer. Otherwise choose two blocks in which the offsets differ. Their product divides a determinant $nm'-n'm$ of size $O(NB)$. Its vanishing would give a nonzero relative relation on those two blocks, contrary to their direct sum. Three groups with product greater than a sufficiently large constant times $NB^2$ force their three value vectors to have zero determinant. The corresponding offset triples cannot be collinear. Indeed, writing $u_j=u_1+t_jd$, with $d$ a primitive integer vector and $t_j$ integers, in the collinear case, all three coprime groups divide $d\times(X+u_1)$. If this vector is nonzero, its size $O(NB)$ gives a contradiction. If it is zero, two positive value vectors are proportional with distinct coordinates in every block, again giving a forbidden pair relation. Thus the centers lie on a genuine plane \[ ((u_2-u_1)\times(u_3-u_1))\cdot X =-((u_2-u_1)\times(u_3-u_1))\cdot u_1. \] There are only $B^{O(1)}$ such planes, determined by the offsets alone. For fixed $k$ group first by the squarefree product $Q=\mathcal G_k$. The ordered factorizations cost $d_k(Q)\ll_{k,\varepsilon'}Q^{\varepsilon'}$, and the supports cost $B^{3k}$. CRT fixes one residue class modulo $Q$ for each center. Summing $O((N/Q+1)^3)$ gives, for $1\leq T\leq N\leq R\leq N^{O(1)}$, \begin{equation}\label{p1:arith:groupcrt} \#\{T<\mathcal G_k\leq R\} \ll_{k,\varepsilon}N^\varepsilon(N^3/T^2+R). \end{equation} For $\mathcal G_3>R\gg NB^2$, use its forced plane. The two largest groups have product $P>R^{2/3}$ and $P\ll NB$. Two free coordinates are fixed by CRT modulo $P$; the plane determines the third. Summing over the two positive group sizes and their offsets gives \begin{equation}\label{p1:arith:planecrt} \#\{\mathcal G_3>R\} \ll_\varepsilon N^\varepsilon(N^2R^{-2/3}+N). \end{equation} Here one uses $\sum_{ab>A}(ab)^{-2}\ll\log(2A)/A$, $\sum_{ab\leq U}(ab)^{-1}\ll\log^2(2U)$, and $\#\{ab\leq U\}\ll U\log(2U)$. The third numerical group is not enumerated again: it certifies the plane and the lower bound for $P$. For $k=3$, take $R=N^{6/5}$ in \eqref{p1:arith:groupcrt}--\eqref{p1:arith:planecrt}. For induction on fixed $k$, the region $T<\mathcal G_k\leq N^{6/5}$ is already covered by \eqref{p1:arith:groupcrt}. Beyond it, $\mathcal G_{k-1}>N^{6(k-1)/(5k)}$. At $k=4$ the preceding $N^3/T^2$ term is $N^{6/5}$; at $k=5$ it is smaller; for $k\geq6$ the threshold is at least $N$ and the preceding tail at $N$ suffices. Only a fixed number of induction steps is taken. This proves \eqref{p1:arith:ternarycount}. \end{proof} \subsection{The probability of two active good edges} \begin{lemma}[A positive second moment for good edges] \label{p1:arith:twogoodpositive} Let $\Xi_N=\sum_{e\in\mathcal E_N}\prod_{x\in e}J_x$. Then \[ \E(\Xi_N)_2=\beta_N^2+o_C(\beta_N^2) \ll_C L^2N^{-4/3}. \] In particular $0\leq\beta_N-\P(\Xi_N\geq1)\leq\tfrac12\E(\Xi_N)_2$. \end{lemma} \begin{proof} For compatible good edges put $q_e=q(e)=W(e)$ and $W_e=\sum_{x\in e}S_x$. On pairs with $W_e\cap W_f=\{0\}$, their original affine events factor exactly. On an interacting separated pair, \[ |q(e\cup f)-q_eq_f|\le W(e\cup f): \] the joint is zero or $W(e\cup f)$, while $q_eq_f\le W(e\cup f)$. The interaction estimate proved in Proposition~\ref{p1:arith:fixed-homogeneous-seeds} therefore bounds the total separated covariance by $O(N^{-41/30+\varepsilon})$. The diagonal and strict-contact baseline products have only $NB^{O(1)}$ hosts, using the geometric degree $O(B)$. Each product is at most $N^{-10/3+o(1)}$, giving a total $O(N^{-7/3+o(1)})$; strict contacts have zero original joint probability. Thus \[ \E(\Xi_N)_2=\beta_N^2+O_{C,\varepsilon}(N^{-41/30+\varepsilon}). \] Choose $\varepsilon<1/30$ and use $\beta_N\asymp_C LN^{-2/3}$. This gives the relative asymptotic and its fixed-constant upper bound. Finally $(m-1)_+\le m(m-1)/2$ for integers $m\ge0$ gives the stated Bonferroni inequality after expectation. \end{proof} \subsection{Local rank at critical length} We apply the binary Griesmer bound \cite[Section~2, equation~(2.10)]{griesmer1960} to the code of local even square relations. The arithmetic input is its minimum nonzero weight; the residual-code proof below keeps the dependence on the parameters explicit. \begin{lemma}[Local rigidity] \label{p1:limits:local-rigidity} Let $V=\{U,\ldots,U+B-1\}$, where $U\asymp N$ and $B=O(\log N)$. Put \[ d_{N,B}=\left\lceil\frac{\log U}{\log(64B)}\right\rceil. \] If the relative square-relation code has dimension $r\ge t$, then \begin{equation} \pi(B)+1\ge \sum_{i=0}^{t-1}\left\lceil d_{N,B}/2^i\right\rceil-t. \label{p1:limits:griesmer} \end{equation} Consequently $r\le1$ when $B=\log N/\log2+o(\log N)$. More generally, if $B\sim c\log N$ and $c<2-2^{1-t}$, then $r0$, the relative signs are exactly independent in all windows with $B\le(1-\varepsilon)\log N$. \end{lemma} \begin{proof} Every square relation is supported on the $D$ integers whose odd prime valuations are all at primes at most $B$: a larger prime is private in the window. The even square-relation code on these coordinates has redundancy at most $\pi(B)+1$. Its minimum nonzero weight is at least $d_{N,B}$ by Lemma~\ref{p1:arith:runge}. For completeness, a binary linear code of length $D$, dimension $r$ and minimum distance $d$ has redundancy at least $\sum_{iH\ {\rm prime}\\v_\ell(n)\ {\rm odd}}}\ell. \] An integer with $K_H(n)=1$ will be called $H$-defective. When the diameter of an interval is less than $H$, every other integer in it has a prime coordinate which occurs nowhere else in the interval. Such a coordinate excludes that integer from every square relation. \begin{proposition}[Regular set]\label{p1:arith:core} Let \[ G_N=\{x\in I_N:\dim S_x=L\},\qquad Z_N^\circ=\sum_{x\in G_N}J_x. \] Then $|I_N\setminus G_N|=N^{o(1)}$, and \begin{equation}\label{p1:arith:coretransfer} d_{\rm TV}(\Law(Z_N),\Law(Z_N^\circ))+ |\E Z_N-\lambda|+|\E Z_N^\circ-\lambda| \ll_C N^{-1+o(1)}. \end{equation} More generally, for every deterministic $A\subset I_N$ and every choice of affine targets on its windows, \begin{equation}\label{p1:arith:marginalsum} \sum_{x\in A}|\P(A_xX=b_x)-p|\leq pN^{o(1)}. \end{equation} If the fixed moments of $Z_N$ are uniformly bounded, then, for every fixed integer $j\geq1$ and every $\varepsilon>0$, \begin{equation}\label{p1:arith:momenttransfer} \big|\E Z_N^j-\E(Z_N^\circ)^j\big| \ll_{C,j,\varepsilon}N^{-1+\varepsilon}. \end{equation} In particular the second factorial cumulants have the same transfer bound. \end{proposition} The proof is given in Section~\ref{p14:comp:regular-transfer}. \subsection{Sparse hosts and weighted nonaligned pairs} For separated windows, $|x-y|>L$, a prime larger than $B$ occurs with odd valuation at most once in either window. Its equation either pins one vertex to zero or joins two vertices. After the pinned components are removed, each remaining component has one binary degree of freedom. If $D(x,y)$ counts the defective vertices and $c(x,y)$ the other unpinned components, then \begin{equation}\label{p1:arith:graphbound} \rho(x,y)\leq D(x,y)+c(x,y),\qquad D(x,y)=o(B). \end{equation} The small prime equations and the two block parities can only reduce this dimension. A pair is called aligned if, in one orientation, it satisfies $bx-ay=h$ for coprime positive $a,b\leq B^3$ and $|h|\leq(a+b)B$. Using this slightly enlarged definition includes every alignment produced by two exact units. Two different reduced slopes cannot occur for the same dyadic pair once $N$ is large: distinct slopes are separated by at least $B^{-6}$, whereas the allowed displacement changes the ratio of the starts by $O(B^{O(1)}/N)$. \begin{proposition}[Support and residual mass]\label{p1:arith:hosts} For every $\varepsilon>0$, \begin{align} \#\{(x,y)\in I_N^2:|x-y|>L,\ \rho(x,y)>0\} &\ll_{C,\varepsilon}N^{1+\varepsilon}, \label{p1:arith:hostbound}\\ \sum_{\substack{x,y\in I_N,\ |x-y|>L\\(x,y)\ {\rm nonaligned}}} (2^{\rho(x,y)}-1) &\ll_{C,\varepsilon}N^{1+\varepsilon}. \label{p1:arith:nonaligned} \end{align} \end{proposition} The elementary proof in Section~\ref{p14:comp:sparse-hosts} first gives the sufficient bound $N^{6/5+\varepsilon}$ in both displays. The stronger bounds stated here are completed in Section~\ref{p14:comp:confinement}, using the conditional half-count and dense-pair confinement. That later strengthening is not an input to either of those arguments; the elementary bounds suffice for the earlier pair-profile calculation. \subsection{Good edges, compression, and one additional window} Let $\mathcal E_N$ consist of the unordered compatible dependent pairs in the channels $2:3$ and $3:4$ for which the distinct normalized values have private prime coordinates above the fixed threshold $H$. Write $q_e=\P(J_x=J_y=1)$ for $e=\{x,y\}$ and $\beta_N=\sum_{e\in\mathcal E_N}q_e$. The family is deterministic. \begin{corollary}[The pair input for the expansion]\label{p1:arith:edgefamily} One has $|\mathcal E_N|=O_C(NB)$ and, for an edge in channel $a:b$, \[ \rho(e)\leq L/b+O(1),\qquad q_e\leq p^2 2^{L/b+O(1)}. \] Moreover, \begin{align} 2\beta_N&=\alpha_N+O_{C,\varepsilon}(N^{-4/5+\varepsilon}), \label{p1:arith:edgemass}\\ \sum_{\substack{x,y\in G_N,\ x\ne y\\\{x,y\}\notin\mathcal E_N}} |\E[(J_x-p)(J_y-p)]| &\ll_{C,\varepsilon}N^{-4/5+\varepsilon}+B/N. \label{p1:arith:badpairs} \end{align} In particular $\beta_N\asymp_C LN^{-2/3}$. \end{corollary} The proof is given in Section~\ref{p14:comp:good-edge-mass}. \begin{lemma}[Exact compression of a good edge]\label{p1:arith:compression} For a good edge in channel $a:b$, the joint homogeneous event space is the injective square-class image of the even-selection space of \[ U_e=bV_x\cup aV_y, \] where repeated values are included only once. In particular its rank is \begin{equation}\label{p1:arith:compressedrank} \dim(S_x+S_y)=|U_e|-1\geq(2-1/b)L-O(1). \end{equation} After the endpoint offsets are fixed, $U_e$ is a subset of an interval of length $O(B)$ depending on one integer translation parameter. There are only $B^{O(1)}$ offset patterns. \end{lemma} \begin{proof} Multiplying every integer in a window by the same constant does not change the square class of an even selection. The two even-selection spaces on $bV_x$ and $aV_y$ therefore map to the original event spaces. They intersect on at least one normalized value: dependence in the good family requires at least two exact units. The sum of the even-selection spaces of two intersecting finite sets is the even-selection space of their union. Its square-class map is injective, because every distinct value in $U_e$ has a private prime. There are at most $L/b+O(1)$ repeated values, which proves \eqref{p1:arith:compressedrank}. \end{proof} \subsection{Pair profiles and the regular-set estimates} \label{p14:comp:pair-profiles} This section supplies the estimates used in Section~\ref{p1:arith:section}. The regular-set transfer is used for the explicit channel formulas and for the independent fixed-moment argument. The scalar reconstruction with exact marginals does not discard these exceptional windows. \subsubsection{Regular-set transfer} \label{p14:comp:regular-transfer} \begin{proof}[Proof of Proposition~\ref{p1:arith:core}] The one-window estimates proved in Lemma~\ref{p1:arith:defects} give $m_B(x)=O_C(B/\log B)$ uniformly, and only $N^{o(1)}$ windows contain two $B$-defective integers. A nonzero relative relation is an even nonempty selection. Private prime coordinates therefore force it to use at least two defects. Its dimension is at most $m_B(x)$. Consequently \[ \sum_{x\in I_N}(2^{L-\dim S_x}-1)=N^{o(1)}. \] Fourier inversion proves \eqref{p1:arith:marginalsum}. The sum of the probabilities of active discarded starts is at most $pN^{o(1)}$; the coupling using the original prime signs proves the total-variation assertion. On $G_N$ the marginal is exactly $p$, giving both assertions about the mean. For the last statement, let $D=\{Z_N\ne Z_N^\circ\}$. Since $0\leq Z_N^\circ\leq Z_N$, H\"older's inequality gives, for any fixed $s>1$, \[ \E\big[Z_N^j\ind_D\big] \leq(\E Z_N^{js})^{1/s}\P(D)^{1-1/s}. \] Choose a sufficiently large integer $s$, and then use the fixed-moment bound. This last step is used only after that bound has been established; a small coupling error alone would not justify a variance transfer. \end{proof} \subsubsection{Sparse hosts and weighted residual mass} \label{p14:comp:sparse-hosts} \begin{proof}[The elementary bounds preceding Proposition~\ref{p1:arith:hosts}] A relative relation that is internal to one window contributes at most $N^{1+o(1)}$ hosts, by Proposition~\ref{p1:arith:core}. Otherwise it selects at least two vertices in each window. Put $T=N^{3/5}$. If all selected kernels are at most $T$, each window contains two small kernels. The estimate \begin{equation}\label{p1:arith:twokernelused} \#\{x\in I_N:\#\{n\in V_x:K_B(n)\leq N^\beta\}\geq2\} \ll_{C,\beta,\varepsilon}N^{\max(\beta,7/12)+\varepsilon} \end{equation} proved in Lemma~\ref{p1:arith:twokernel} gives $N^{6/5+\varepsilon}$ choices. If one selected kernel exceeds $T$, Lemma~\ref{p1:arith:anchor} gives $N^{3/2+o(1)}T^{-1/2+\tau}$, for every fixed $\tau>0$. Choose $\tau$ sufficiently small. For the weighted assertion, first restrict to $c(x,y)<\eta B$. Equation~\eqref{p1:arith:graphbound} bounds the weight by $2^{\eta B+o(B)}$; choose $\eta$ after the desired error exponent. The complementary sector has total weight $O_{C,\eta,\varepsilon}(N^{1+\varepsilon})$ by Lemma~\ref{p1:arith:dense}. The latter estimate uses the component equations themselves, rather than just their support: many small components put the pair on a multiquadratic surface, and its low-degree curves are either aligned or absent from the dyadic region. This proves both bounds with exponent $6/5+\varepsilon$, which is all that the pair-profile proof below uses. The exponent $1+\varepsilon$ in the revised statement is proved later in Section~\ref{p14:comp:confinement}; no late confinement estimate is being inferred from this elementary calculation. \end{proof} \subsubsection{The complete pair profile} \label{p14:comp:complete-profile} \begin{proof}[Proof of Theorem~\ref{p1:arith:pairs}] For the channel $2:3$, either orientation permits $N/2\leq t<2N/3$, up to $O(B)$ endpoint changes. For $3:4$ the corresponding interval is $N/3\leq t0\), at most \[ \bigl(A^{-1}+O_A(1/\log B)\bigr)B \] of its values satisfy \(u_B(n)>B^A\). \end{lemma} \begin{proof} Dropping the parity condition gives \[ \sum_{n\in V}\log u_B(n) \leq\sum_{\ell\leq B} \left(\left\lfloor\frac B\ell\right\rfloor+1\right)\log\ell \leq\log(B!)+\sum_{\ell\leq B}\log\ell. \] Here the sums are over primes. Legendre's formula gives the second inequality, and the elementary Chebyshev bound makes the last sum \(O(B)\). Finally \(\log(B!)\leq B\log B\). Each value counted in the last assertion contributes more than \(A\log B\) to the nonnegative sum. \end{proof} \subsection{Regularising coordinate forms by independent restrictions} \begin{lemma}[Coordinate regularisation] \label{p1:arith:coordinate-regularisation} Let \(R\subset\F_2^V\) have dimension \(r\), put \(n=|V|\), and let \(\ell_v\in R^*\) be the coordinate forms. Suppose that \(s\) distinct nonzero forms occur among them. For integers \(K,d\geq1\), there is \(D\subset V\) such that \[ R'=\{x\in R:x_v=0\text{ for every }v\in D\},\qquad \operatorname{codim}_R R'=|D| \leq\frac{n}{K+1}+\frac{s-r}{d+1}, \] and the restricted coordinate forms have the following properties: \begin{enumerate} \item every nonzero form has multiplicity at most \(K\); \item every nonzero form \(\ell\) has at most \(d\) unordered decompositions \(\ell=a+b\), where \(a,b\) are distinct nonzero coordinate forms present on \(R'\). \end{enumerate} Coordinates whose restricted forms vanish may then be removed without any additional loss of dimension. \end{lemma} \begin{proof} The coordinate forms span \(R^*\), and their restrictions continue to span the dual of every subspace encountered below. At an intermediate stage let \(S\) be the set of distinct nonzero forms and let \(r_0\) be their rank. Setting a present nonzero form \(\ell\) equal to zero reduces the dimension by one and replaces the forms by their images modulo \(\langle\ell\rangle\). The form \(\ell\) disappears. All other collisions are exactly the disjoint unordered pairs \(\{a,a+\ell\}\subset S\). If there are \(t(\ell)\) such pairs, then \[ |S_{\rm new}|=|S|-1-t(\ell),\qquad (|S|-r_0)_{\rm new}=|S|-r_0-t(\ell). \] Thus the nonnegative potential \(|S|-r_0\) never increases. If a form has multiplicity greater than \(K\), set it equal to zero. This permanently kills at least \(K+1\) previously nonzero coordinates. There are at most \(n/(K+1)\) such steps, since these sets of killed coordinates are disjoint. When the multiplicity condition holds but a form has more than \(d\) decompositions, set that form equal to zero instead. Such a step decreases the potential by at least \(d+1\), so there are at most \((s-r)/(d+1)\) of these steps. Each step decreases the dimension, and the procedure therefore ends with both required properties. At every step select one coordinate carrying the form just killed. Its restriction was nonzero before that step, so its original coordinate form is independent of the previously selected forms. The selected coordinates constitute \(D\), and consequently \(\operatorname{codim}_R R'=|D|\). \end{proof} \begin{corollary}[Constraints meeting three blocks] \label{p1:arith:ternary-degree} Suppose \(V\) is partitioned into three blocks and each row of a family of binary constraints annihilating \(R\) has at most one occurrence in each block. Apply Lemma~\ref{p1:arith:coordinate-regularisation} and delete the coordinates whose forms on \(R'\) are zero. Every remaining row has weight zero, two, or three. A row of weight two joins equal nonzero forms. A row of weight three has three distinct nonzero forms \(a,b,c\) with \(a+b+c=0\). The connected components formed by the rows of weight two have at most \(K\) vertices. After repeated rows with the same support are identified, the number of ternary rows incident to any vertex is at most \(2dK^2\). \end{corollary} \begin{proof} The sum of the forms on a row is zero, since that row annihilates \(R'\). A row of weight one is therefore impossible. At weight two its forms are equal. At weight three a repeated form would force the third form to be zero, which has already been excluded. Binary components stay inside a class of equal forms, of size at most \(K\). At a vertex of form \(\ell\), a ternary row chooses one of at most \(d\) decompositions \(\ell=a+b\), then the other two coordinates from classes of size at most \(K\). Allowing their two assignments to the remaining blocks gives the asserted \(2dK^2\) bound. \end{proof} \begin{remark}[Order of restrictions] If some coordinates must first be killed for an arithmetic reason, do this before applying the regularisation lemma. A further restriction after regularisation can identify previously distinct forms and destroy both bounds. The lemma is an existence statement about a subspace; its proof does not enumerate possible relation spaces or histories of restrictions. \end{remark} \subsection{A finite rank bound for every fixed channel} Fix coprime integers \(a,b\) with \[ b\geq3,\qquad b/2b(a+b-1)\), the density is less than \(1/b\). All other nonphysical fractions are included in \eqref{p1:arith:gamma-fixed-channel}. If \(c\mid a\), a proper divisor has \(c\leq a/20\), \begin{equation} \sup_{x_1,\ldots,x_s\ {\rm distinct}} \#\{z\in I_N\setminus\{x_1,\ldots,x_s\}: \delta_z\ge\eta L\} \ll_{s,C,\eta,\varepsilon}N^\varepsilon . \label{p1:fixed:targets-bound} \end{equation} The sources may be dependent or overlapping. \end{proposition} \noindent The complete proof is given in Section~\ref{p14:comp:fixed-targets}. \begin{proposition}[Conditional sums and propagation] \label{p1:fixed:propagation} For every fixed \(s\ge1\) and \(\varepsilon>0\), uniformly over compatible prescribed affine events at \(s\) distinct sources, \begin{equation} \sum_{z\notin\{x_1,\ldots,x_s\}} \P(J_z=1\mid J_{x_1}=\cdots=J_{x_s}=1) \ll_{s,C,\varepsilon}N^\varepsilon . \label{p1:fixed:conditional} \end{equation} If a family of ordered \(s\)-tuples has total positive joint mass \(W_{s,N}\), the mass of its extensions to ordered distinct \(r\)-tuples, for fixed \(r\ge s\), is \begin{equation} W_{s\to r,N}\ll_{r,C,\varepsilon}N^\varepsilon W_{s,N}. \label{p1:fixed:propagate} \end{equation} Both \eqref{p1:fixed:conditional} and \eqref{p1:fixed:propagate} also hold for homogeneous prescriptions. For the original run-start events, the mass of fixed \(r\)-tuples containing a local defect or a compatible contact is \(O_{r,C,\varepsilon}(N^{-1+\varepsilon})\). \end{proposition} \begin{proof} For \(\delta_z<\eta L\), the conditional sum is at most \(2Np2^{\eta L}\ll_C N^\eta\). For \(\delta_z\ge\eta L\), Proposition~\ref{p1:fixed:targets} counts \(O(N^{\varepsilon/2})\) targets, each of conditional probability at most one. Take \(\eta=\varepsilon/2\). The proof uses only the rank loss and compatibility, so it applies unchanged to homogeneous prescriptions. For \(r=s\), \eqref{p1:fixed:propagate} is immediate. For \(r>s\), iterating with margin \(\varepsilon/(r-s)\) proves the bound. Allowing a seed at any positions in the ordered tuple multiplies the estimate by a constant depending on \(r\). For the original run-start events, the local and contact seeds have positive masses \(N^{-1+o(1)}\), by the regular-core estimates in Section~\ref{p1:arith:section}. Contacts at distance strictly less than \(L\) have zero joint probability. Propagate the remaining seeds to obtain the last claim. \end{proof} \subsection{The correction to every fixed moment} \begin{theorem}[Fixed factorial moments] \label{p1:fixed:moments} For every fixed integer \(r\ge2\) and every \(\varepsilon>0\), \begin{equation} \E(Z_N)_r =\mu_N^r+\binom r2\mu_N^{r-2}\alpha_N +O_{r,C,\varepsilon}(N^{-4/5+\varepsilon}). \label{p1:fixed:moment-expansion} \end{equation} Equivalently, one may use \(\lambda_N\) and \(a_N=\E(Z_N)_2-\lambda_N^2\). All fixed ordinary moments are bounded. More precisely, with \(T_r(u)=\E[\Pois(u)^r]\), \begin{equation} \E Z_N^r =T_r(\mu_N)+\frac{\alpha_N}{2}T_r''(\mu_N) +O_{r,C,\varepsilon}(N^{-4/5+\varepsilon}). \label{p1:fixed:ordinary} \end{equation} For every fixed \(r\ge3\), the \(r\)-th factorial cumulant is \(O_{r,C,\varepsilon}(N^{-4/5+\varepsilon})\). \end{theorem} This first estimate is proved independently of the law approximation. Its bounded-moment conclusion is used in the sharp lower bound of Theorem~\ref{p13:lower:optimal} in the article. After that reconstruction, Corollary~\ref{p13:optimal:weighted} improves the scalar moment remainders to \(O_{r,C,\varepsilon}(N^{-4/3+\varepsilon})\) when the exact parameters are retained. \noindent The complete proof is given in Section~\ref{p14:comp:fixed-moments}. For \(r\ge2\), the first discrepancy in \eqref{p1:fixed:moment-expansion} has the exact scale \(L N^{-2/3}\), by \eqref{p1:eq:alpha-leading}. The theorem fixes \(r\) first. Its proof does not allow \(r\) to grow like \(\log N/\log\log N\). The uniform certificate tools in Section~\ref{p14:cert:section} and the large-order reconstruction in Section~\ref{p13:transfer:section} supply that separate input. \subsection{Fixed-source geometry and the independent moment proof} \label{p14:comp:fixed} The statements of the fixed-source estimates and the fixed-moment theorem are in Section~\ref{p1:fixed:section}. This section supplies the geometric count and the positive classification used in their proofs. The original window estimates below use the regular-set convention $p=2^{-L}$; Section~\ref{p13:exact:section} gives their extension to effective spaces. The moment proof is independent of the reconstruction of the complete law. \subsubsection{Uniformly few targets of large rank loss} We first record the one-parameter geometric estimate in the precise form needed here. \begin{lemma}[Several disjoint square equations] \label{p1:fixed:curve} Fix \(K,A,D_0>0\) and \(\varepsilon>0\). There is a fixed integer \(j\) such that the following holds. Let \(P_1,\ldots,P_j\) be nonconstant products of distinct linear factors \(Z+u\), of degrees at most \(K\), with \(|u|\le D_0B\). Suppose their root lists are pairwise disjoint, and let \(0<|D_\nu|\le N^A\) be integers. Then \[ \#\{z\asymp N: w_\nu^2=D_\nu P_\nu(z)\ (1\le\nu\le j) \text{ for some integers }w_\nu\} \ll_{K,A,D_0,\varepsilon}N^\varepsilon . \] The constants are uniform in the displayed coefficients and offsets. \end{lemma} \begin{proof} The square classes of the functions \(D_\nu P_\nu(Z)\) are geometrically independent: a root of \(P_\nu\) has odd valuation in that class and even valuation in every other class. Thus the resulting curve is geometrically integral and has degree \(2^j\) over the \(Z\)-line. Choose an integer \(D\ge(K+A)/2\). Give \(T,S\) degree one and each \(W_\nu\) degree \(D\), and use the equations \[ W_\nu^2 =D_\nu S^{2D-\deg P_\nu} \prod_{u\in P_\nu}(T+uS). \] The graded ring is free of rank \(2^j\) over \(\mathbb Q[T,S]\); its generic localization is a field. Its \(D\)-th Veronese ring is generated in degree one by the monomials in \(T,S\) of degree \(D\) and the \(W_\nu\). The corresponding projective curve has degree \(D2^j\), since its hyperplane bundle is the pullback of \(\mathcal O_{\mathbb P^1}(D)\). On \(S=1\), the coordinates are integral and have height \(O_{K,D}(N^D)\). Pila's uniform curve estimate \cite[Theorem~A]{pila1995} bounds their number by \[ O_{K,A,j,\varepsilon'} \bigl((N^D)^{1/(D2^j)+\varepsilon'}\bigr) =O(N^{2^{-j}+D\varepsilon'}). \] Choose \(j\) and then \(\varepsilon'\) to put this exponent below \(\varepsilon\). Both choices precede \(N\). \end{proof} \begin{lemma}[A product certificate with fixed sources] \label{p1:fixed:CRT} Fix \(s,k\ge1\). Uniformly over sources \(x_1,\ldots,x_s\in I_N\), the number of centres \(z\in I_N\) for which there are pairwise coprime \(G_1,\ldots,G_k>1\) satisfying \[ G_i\mid x_{a_i}+u_i,\qquad G_i\mid z+w_i,\qquad \prod_iG_i>N, \quad u_i,w_i\in\{-1,\ldots,L-1\}, \] is \(O_{s,k,C,\varepsilon}(N^\varepsilon)\) for every \(\varepsilon>0\). \end{lemma} \begin{proof} There are \((sB)^kB^k=N^{o(1)}\) choices of positions. For fixed positions, the product \(Q=\prod_iG_i\) divides \(\prod_i(x_{a_i}+u_i)\le(3N)^k\). Its divisors and their ordered factorizations into \(k\) factors have \(N^{o(1)}\) possibilities. Coprimality fixes one residue class for \(z\) modulo \(Q>N\), hence at most one centre in \(I_N\). This counts the certificate without requiring the groups to be fixed in advance. \end{proof} \subsubsection{Proof of Proposition~\ref{p1:fixed:targets}} \label{p14:comp:fixed-targets} \begin{proof} Take \(H=D_0B\), with fixed \(D_0>1\). A prime exceeding \(H\) has at most one occurrence of odd valuation in each window. Apply Lemma~\ref{p1:fixed:recovery} to \(R(\boldsymbol x,z)\), with \(d=s\), choosing \(t\) so that the loss of dimension is at most \(\eta L/4\). The surviving target projection has dimension at least \(3\eta L/4\). Nonzero labels have multiplicity at most \(K=t-1\). For each prime \(q>H\) incident to an active target coordinate whose label has odd multiplicity in the \(q\)-row, choose a minimal circuit of distinct labels containing this label. Such a circuit exists: a minimal zero-sum subset containing the label cannot contain a proper zero-sum subset, because it could be removed. Choose incident coordinates for its other labels. They all belong to sources, because \(q\) has only one target occurrence. Group primes according to the selected physical support. These groups are pairwise coprime. The number of supports through a fixed target coordinate is at most \[ D=\max\left(1,\sum_{j=2}^{s}j!t^{j-1}K^j\right). \] Choose the integer \(j\) in Lemma~\ref{p1:fixed:curve} for degree at most \(K\) and coefficient height \(N^{A_0+2K+2}\), where \(\prod_{q\le H}q\le N^{A_0}\). Choose \(k\) large enough that \(\theta=KD/k\le1\) and \(j\theta<\varepsilon/4\). If the product of the \(k\) largest nontrivial groups exceeds \(N\), Lemma~\ref{p1:fixed:CRT} applies. It remains to count the complementary case. Discard from selection the label classes meeting the target positions of these \(k\) groups. At most \(k\) classes are discarded; no further restriction of the relation space is made. Every group incident to a remaining target class has size at most \(N^{1/k}\). For such a class \(A\), write \[ M_A=\prod_{v\in A\ {\rm at\ sources}} n_v,\qquad P_A(Z)=\prod_{v\in A\ {\rm at\ target}}(Z+w_v). \] Its squarefree product \(d_A=\sf(M_AP_A(z))\) can be factored as \[ d_A=d_{0,A}d_{1,A}e_A,\qquad d_{0,A}\mid\prod_{q\le H}q,\quad d_{1,A}\mid M_A,\quad 1\le e_A\le N^\theta . \] Indeed, every remaining prime with odd multiplicity and not dividing \(M_A\) lies in one of at most \(KD\) incident groups. The choices of \(d_{0,A}\) and \(d_{1,A}\) are \(N^{o(1)}\), uniformly for fixed source positions. There are at least \(3\eta L/4-k\) available classes meeting the target. Select \(j\) of them. Their target polynomials have simple disjoint roots. After choosing their positions and coefficients, the equations \[ w_\nu^2=d_{A_\nu}M_{A_\nu}P_{A_\nu}(z) \] are counted by Lemma~\ref{p1:fixed:curve}. The positions cost \(B^{O(1)}\), and the coefficients cost \(N^{j\theta+o(1)}\). The stated choices, with the error margin split between these factors, give \eqref{p1:fixed:targets-bound}. Only the selected positions and coefficients have been enumerated, not the possible relation spaces or their regularizations. \end{proof} \subsubsection{Primitive families at a fixed order} A family is \emph{primitive} if it has a nonzero relative relation and every proper subfamily has none. Its relation space projects injectively to each block. For original windows its nullity is consequently at most \(L\). For the homogeneous prescription define \[ W(I)=p^{|I|}2^{\rho(I)}. \] This is an actual joint probability, regardless of the compatibility of the run-start prescriptions. Those have joint probability either \(W(I)\) or zero. \begin{corollary}[Primitive families at a fixed order] \label{p1:fixed:primitive} For every fixed $r\ge3$ and $\varepsilon>0$, the sum of the homogeneous weights $p^r2^{\rho(I)}$ over ordered primitive $r$-tuples of separated regular windows is $O_{r,C,\varepsilon}(N^{-r/2+\varepsilon})$. The same bound holds for their actual joint mass and absolute centered mass. \end{corollary} \begin{proof} Apply Proposition~\ref{p13:transfer:primitive} on $G_N$, where $q_0(I)=p^r2^{\rho_{\mathrm{rel}}(I)}$, and multiply by the fixed ordering factor $r!$. The proposition uses the geometric half-count and dense-host construction, independently of the moment proof below. \end{proof} The proof is homogeneous throughout. It therefore bounds incompatible configurations as well as compatible ones. This feature will be needed for centered Fourier coefficients. \begin{proposition}[Fixed primitive mass after pair compression] \label{p1:fixed:compressed} Fix a channel \(a:b\in\{2:3,3:4\}\), and put \(\kappa_b=2-1/b\). All original starts lie in \(G_N\) and are distinct and pairwise separated. For every fixed \(r\ge3\) and \(\varepsilon>0\), the total homogeneous mass of primitive families consisting of one exactly compressed pair in this channel and \(r-1\) further original windows is \begin{equation} O_{r,C,\varepsilon} (N^{-\kappa_b+\varepsilon}). \label{p1:fixed:compressed-bound} \end{equation} For a compressed geometric good pair and just one further original window with positive relative rank loss, the mass is \(O_{C,\varepsilon}(N^{-41/30+\varepsilon})\) in either channel. The affine positive masses satisfy the same bounds. \end{proposition} \begin{proof} Put $t=r-1\ge2$ and fix a compression motif. The compressed block is the even-selection image of a genuine set of $O(B)$ values of width $O(B)$; write $r_e$ for its rank. Use its normalized centre $bx$; it lies above the original target range and determines the source pair once the motif is fixed. There are $B^{O(1)}$ motifs and $NB^{O(1)}$ source pairs. For its homogeneous mass, Lemma~\ref{p1:arith:alignedrank} gives $q_0(e)=2^{-r_e}\ll N^{-\kappa_b+\varepsilon'}$, including affine-incompatible sources. Primitivity gives $\rho\le L$ and the exact homogeneous weight $q_0(e)p^t2^\rho$. Every primitive family has a nonzero effective projection in each of its $t+1$ blocks. Theorem~\ref{p13:half:conditional}, with no old sources, counts at most $N^{(t+1)/2+\varepsilon'}$ hosts, including the compressed centre. For $t\ge3$, their total mass is \[ N^{(t+1)/2+\varepsilon'}N^{-\kappa_b-t+1+\varepsilon'} =N^{-\kappa_b+(3-t)/2+2\varepsilon'} \le N^{-\kappa_b+2\varepsilon'}. \] For $t=2$, the same count on $\rho\le L/4$ gives $N^{-\kappa_b-1/4+2\varepsilon'}$. On $\rho\ge L/4$, fix the compressed source image $W$. Every proper subfamily is direct, so each target is disjoint from $W$ and the two targets are direct. Lemma~\ref{p13:transfer:dense-targets} therefore gives conditional homogeneous target mass $O(N^{-1+\varepsilon'})$; the original targets are regular, so their homogeneous marginals are exactly $p$. The source sum is $\sum_e q_0(e)\ll N^{1-\kappa_b+2\varepsilon'}$, giving total $O(N^{-\kappa_b+3\varepsilon'})$. Reduce the margins and absorb the motif choices to obtain \eqref{p1:fixed:compressed-bound}. The one-target clause is part (2) of Proposition~\ref{p1:arith:fixed-homogeneous-seeds}. Original affine joint probabilities are bounded by the homogeneous ones. \end{proof} \subsubsection{Proof of Theorem~\ref{p1:fixed:moments}} \label{p14:comp:fixed-moments} \begin{proof} Expand \(\E(Z_N)_r\) over ordered distinct centres. We describe all configurations removed before evaluating the two principal classes. Local defects and contacts have mass \(N^{-1+\varepsilon}\) after propagation. A dependent separated pair outside the two retained channels has mass \(N^{-4/5+\varepsilon}\), also after propagation. Tuples containing a primitive original subfamily of size at least three have that same bound, by Corollary~\ref{p1:fixed:primitive}. Two distinct good pairs, including pairs sharing a vertex, have positive seed mass $O_C(L^2N^{-4/3})$ by Proposition~\ref{p1:arith:fixed-homogeneous-seeds}(3): their union belongs to the separated raw-edge population counted there, and its affine mass is bounded by its homogeneous mass. Fixed-order propagation preserves the required precision. There remains the case of one good pair with an additional relation. Compress this pair exactly. A minimal dependent subfamily of the compressed system has at least two blocks and contains the compressed block, unless an original primitive family has already been removed. Apply Proposition~\ref{p1:fixed:compressed} to this seed and then propagate. Each of its exponents exceeds \(4/5\). This exhausts the dependent residual configurations: after their removal the spaces are either all direct, or one good pair is direct from all remaining singletons. The support estimate of Proposition~\ref{p1:arith:fixed-support}, together with the \(O_r(N^{r-1}B)\) contact tuples, shows that the number of full-rank tuples is \[ N^r+O_{r,C,\varepsilon}(N^{r-4/5+\varepsilon}). \] Their joint probability is exactly \(p^r\), and their contribution is \(\lambda_N^r+O(N^{-4/5+\varepsilon})\). Let \(A_N\) be the ordered positive mass of good pairs. The pair calculation gives \[ A_N=a_N+O_{C,\varepsilon}(N^{-4/5+\varepsilon}). \] Fix one such pair. After any fixed number of already chosen windows, at most \(N^{1/2+o(1)}\) target centres are dependent on their relative space, by Proposition~\ref{p1:arith:fixed-branching}. Thus the number of ordered direct extensions by \(r-2\) windows is \[ N^{r-2}\bigl(1+O_{r,C}(N^{-1/2+o(1)})\bigr), \] uniformly in the pair. Their joint weight is \(q(e)p^{r-2}\). Placing the ordered pair in two of the \(r\) positions gives contribution \[ \binom r2\lambda_N^{r-2}A_N +O_{r,C}(A_NN^{-1/2+o(1)}). \] Since \(A_N=O_C(LN^{-2/3})\), the last error is smaller than the stated remainder. The two principal classes give the formula with \((\lambda_N,a_N)\). The mean transfer \(\mu_N-\lambda_N=O(N^{-1+\varepsilon})\) and \(\alpha_N-a_N=O(N^{-1+\varepsilon})\) give \eqref{p1:fixed:moment-expansion}. The Stirling-number identity \(Z^r=\sum_{j=0}^r\left\{\begin{smallmatrix}r\\j\end{smallmatrix}\right\}(Z)_j\) proves \eqref{p1:fixed:ordinary} and boundedness. Finally, to any fixed formal order, the factorial generating function equals \[ e^{\mu_N t}\left(1+\frac{\alpha_Nt^2}{2}\right) +O_{\rm coeff}(N^{-4/5+\varepsilon}). \] Taking its formal logarithm is legitimate at that fixed order. Terms quadratic in \(\alpha_N\) are \(O(N^{-4/3+o(1)})\), which proves the cumulant claim. \end{proof} \section{Square products with short selections} \label{p13:short:section} The conditional count in the next section must handle a relation even when none of its selected integers has a large prime kernel. For selections of at most a small multiple of \(\log\log N\), we instead count square products through moments. The constants in this argument are explicit in the selection size; a fixed-degree result alone would not give the required range. For a fixed number of unrestricted coordinates, global counts of tuples whose product is a square were obtained by de~la~Bret\`eche, Kurlberg, and Shparlinski \cite[Theorem~2.2 and Corollary~2.4]{bretkurlbergshparlinski2021}. Here the selections lie in short additive windows, and the ensuing conditional estimates are uniform in arbitrary fixed source windows even when the number of selected factors grows with $N$. The quadratic kernel argument is related to the decomposition in de la Bret\`eche, Wang, and Xu \cite[Lemmas~6.2--6.3]{bretwngxu2026}. We give the estimates needed here, including their dependence on the number and location of the roots. Write \[ P_A(X)=\prod_{a\in A}(X+a),\qquad A\subset[-DB,DB]\cap\Z, \qquad 2\le d=|A|\le c\log\log N, \] where $d$ is even and $D,c>0$ are fixed. The interval of summation has height and length comparable to $N$. \begin{proposition}[Uniform moments for short split products] \label{p13:short:moments} For every fixed even integer \(v\ge2\), with the preceding notation, \begin{equation} \E\left|\sum_{n\asymp N}f(P_A(n))\right|^v \le N^{v/2+c[v(v-1)+2]\log2+o(1)}. \label{p13:short:moment-bound} \end{equation} In particular, for each fixed \(v\) and \(\varepsilon>0\), one may choose \(c>0\) so that the \(L^v\) norm is \(O(N^{1/2+\varepsilon})\), uniformly in all the indicated motifs. The summation interval may have any fixed comparable length and positive height. \end{proposition} The proof, in Section~\ref{p14:comp:short:moments}, pairs the linear factors into quadratics and controls the common kernel by their resultant. In the moment expansion, primes are assigned to the even subsets of indices on which they occur. Averaging the exposure order then gives the exponent $v/2$. The explicit term $c[v(v-1)+2]\log2$ is what permits a growing selection size. \begin{corollary}[Short projections in an old space] \label{p13:short:old-space} Fix \(r,s,D\) and \(\varepsilon>0\). Let \(V_{\rm old}\) be generated by the square classes of all values in \(s\) fixed blocks of diameter and cardinality \(O_D(B)\). For a sufficiently small fixed \(c>0\), the number of lists of \(r\) new centres admitting nonempty even selections \(A_i\), \(|A_i|\le c\log\log N\), whose square classes sum to an element of \(V_{\rm old}\), is \(O_{r,s,D,C,\varepsilon}(N^{r/2+\varepsilon})\), uniformly in the old blocks. \end{corollary} \begin{proof} For a square class \(u\), let \(\chi_u(f)\) denote its character, and put \(\Phi_V=\sum_{u\in V}\chi_u\), summing over distinct classes. Character orthogonality gives \[ \Phi_V\ge0,\qquad \E\Phi_V=1,\qquad \|\Phi_V\|_a=2^{\dim V(1-1/a)}\quad(a\ge1). \] For fixed motifs, with \(X_i=\sum_z f(P_{A_i}(z))\), the expectation \(\E(\Phi_{V_{\rm old}}\prod_iX_i)\) counts exactly the required lists. For each list, a unique old character cancels the new sum if it belongs to the old space. All terms after expectation are zero or one. Since \(\dim V_{\rm old}=O_{s,D}(B)\), choose a fixed \(a>1\) sufficiently close to one that \(\|\Phi_{V_{\rm old}}\|_a\le N^{\varepsilon'}\). Next choose a fixed even \(v\ge ra/(a-1)\), and then choose \(c\) by Proposition~\ref{p13:short:moments}. H\"older gives \[ \left|\E\left(\Phi_{V_{\rm old}}\prod_iX_i\right)\right| \le N^{\varepsilon'}\prod_i \|X_i\|_{ra/(a-1)} \ll N^{r/2+(r+1)\varepsilon'}. \] The motifs have \(\exp(O_{r,D}(c(\log\log N)^2))=N^{o(1)}\) choices. Choose the internal margins in the stated \(\varepsilon\). There is no enumeration of the \(2^{\dim V_{\rm old}}\) old classes. \end{proof} \subsection{Short split products: kernels and moment enumeration} \label{p14:comp:short} This section proves the short-support moment input from Section~\ref{p13:short:section}. The degree is allowed to grow as a sufficiently small multiple of $\log\log N$; all estimates below retain that dependence. \begin{lemma}[A quadratic kernel with a prescribed divisor] \label{p13:short:quadratic} Let \(1\le h\le2DB\), where \(D\) is fixed, and \(N\le n\le2N\). Suppose \(N^\beta\le q=\sf(n(n+h))<2N^\beta\) and a fixed squarefree integer \(e\) divides \(q\). Put \(\delta=\log e/\log N\). Uniformly in these data, the number of \(n\) is at most \[ N^{c_2(\beta,\delta)+o(1)},\qquad c_2(\beta,\delta)= \min\{1,\beta-\delta,\max(\beta/2-\delta,1-\beta/2)\} \le \tfrac12+\tfrac\beta4-\tfrac\delta2. \] Every \(o(1)\) in this lemma is uniform and may be replaced by an arbitrary fixed positive exponent. The uncapped minimum without the entry \(1\) also gives a valid counting exponent. Its exact comparison with the affine expression requires \(\beta-2\delta\le2\); the displayed capped comparison is valid without this additional range restriction. \end{lemma} \begin{proof} For a fixed \(q\), the equation \(n(n+h)=qv^2\) becomes \((2n+h)^2-4qv^2=h^2\). Lemma~\ref{p1:arith:pell} gives \(N^{o(1)}\) solutions, uniformly for the polynomial heights in question. When \(q=1\), factor the difference of squares instead. Summing over the \(O(1+N^\beta/e)\) possible kernels gives the first exponent. For the other estimate, \(v\asymp N^{1-\beta/2}\) and \(ev^2\mid n(n+h)\). If \(\rho_h(M)=\#\{a\bmod M:a(a+h)\equiv0\pmod M\}\), then \[ \rho_h(\ell^a) \le2\ell^{\min(v_\ell(h),\lfloor a/2\rfloor)},\qquad \rho_h(M)\le h\,2^{\omega(M)}. \] For \(a\le2v_\ell(h)\), a root is divisible by \(\ell^{\lceil a/2\rceil}\); for larger \(a\), the two separated roots give the displayed bound. The Chinese remainder theorem proves the second inequality. Here \(M=ev^2\le N^{2+o(1)}\), so summing residue classes gives \[ N^{o(1)} \sum_{v\asymp N^{1-\beta/2}}\left(\frac{N}{ev^2}+1\right) \le N^{o(1)} \{N^{\beta/2-\delta}+N^{1-\beta/2}\}. \] The trivial \(O(N)\) count supplies the entry \(1\) in the minimum. If \(\beta-\delta\ge1\), the maximum is \(\beta/2-\delta\), and the capped exponent is at most \[ \min(1,\beta/2-\delta) \le \frac{1+\beta/2-\delta}{2} =\frac12+\frac\beta4-\frac\delta2. \] Otherwise the maximum is \(1-\beta/2\), and the minimum of this number and \(\beta-\delta\) is at most their average, with the same affine value. This proves the exact comparison, including admissible cases with \(\beta>2\). \end{proof} \paragraph{Necessity of the exponent cap.} In Lemma~\ref{p13:short:quadratic}, the counting exponent is capped: \[ c_2(\beta,\delta) =\min\{1,\beta-\delta,\max(\beta/2-\delta,1-\beta/2)\}. \] Its affine upper bound is then valid for every real $\beta,\delta$. The uncapped expression still gives a valid counting estimate, but its exact affine comparison requires $\beta-2\delta\le2$ when $\delta\ge0$. This restriction is necessary on actual positive-divisor data: squarefree consecutive $N,N+1$ with $h=e=1$ give $\beta=\log(N(N+1))/\log N>2$ and $\delta=0$. The extra entry $1$ is supplied by the trivial $O(N)$ count. \begin{lemma}[Kernels of a split product] \label{p13:short:split} Let \[ P_A(X)=\prod_{a\in A}(X+a),\qquad A\subset[-DB,DB]\cap\Z,\qquad 2\le d=|A|\le c\log\log N, \] where \(d\) is even and \(D,c>0\) are fixed. For a fixed squarefree \(e\), the number of \(n\asymp N\) with \(q=\sf(P_A(n))\asymp N^\beta\) and \(e\mid q\) is at most \begin{equation} N^{1/2+\beta/(2d)-\delta/d+c\log2+o(1)}, \qquad \delta=\frac{\log e}{\log N}. \label{p13:short:split-count} \end{equation} For a fixed kernel \(q\), the fibre has at most \(N^{2c\log2+o(1)}\) elements. All estimates are uniform in \(A\). \end{lemma} \begin{proof} Pair the \(d\) linear factors into \(m=d/2\) quadratics \(Q_j\), put \(q_j=\sf(Q_j(n))\), and write \[ D_A=\prod_{\substack{a,b\in A\\a0\), and \(\varepsilon>0\). Let \(V_{\rm old}\) be the space generated by all values in \(s\) fixed source blocks. Uniformly in those sources, the number of lists of distinct new centres \(z_1,\ldots,z_r\) such that \begin{equation} u_i\in S_{z_i}\setminus\{0\}\quad(1\le i\le r), \qquad \sum_i u_i\in V_{\rm old} \label{p13:half:condition} \end{equation} for some \(u_i\), is \(O_{r,s,D,C,\varepsilon}(N^{r/2+\varepsilon})\). The \(u_i\) need not remain nonzero modulo \(V_{\rm old}\). The source blocks may overlap or be dependent. \end{theorem} The condition asks for a nonzero projection in each new effective space. It imposes neither affine compatibility nor a separation condition on physical values. For $H=C_DB$ larger than every block diameter, write \[ K_H(n)=\prod_{\substack{\ell>H\\v_\ell(n)\ {\rm odd}}}\ell. \] This is the odd prime kernel after the small primes have been removed. \paragraph{Why the half power appears.} For a fixed witness, short selections are counted by Corollary~\ref{p13:short:old-space}. A long selection either has many large odd kernels, which determine another centre by congruences, or has many large smooth parts, whose least common multiple exceeds the length of the centre interval. In the second case the centre lies in a sparse deterministic set. The complete proof in Section~\ref{p14:comp:half:iteration} iterates these two alternatives. At a stage indexed by $x_j\le1/2$, a surviving source has at most $N^{1-x_{j-1}+o(1)}$ possibilities and its receiver degree is at most $N^{x_j+o(1)}$. Removing this source--receiver pair and applying induction to the other $r-2$ blocks costs $N^{r/2+x_j-x_{j-1}+o(1)}$. Small fixed increments and a final $N^{1/2+o(1)}$ exceptional source set give the theorem. The induction is simultaneous in every fixed number of old sources, so adding exposed blocks to that space introduces no dependence assumption. \begin{corollary}[Connected and primitive families at fixed order] \label{p13:half:connected} For every fixed \(r\ge2\), connected families of \(r\) original windows have \(O(N^{r/2+\varepsilon})\) hosts. Their independent mass, and in particular the independent mass of primitive families, is \(O(N^{-r/2+\varepsilon})\). \end{corollary} \begin{proof} A primitive family has a relation with nonzero projection in each block, so Theorem~\ref{p13:half:conditional} applies with no sources. A connected family contains a minimal dependent subfamily; expose it first. If \(A\) is the set already exposed, connectedness implies \[ \left(\sum_{i\in A}S_i\right) \cap\left(\sum_{j\notin A}S_j\right)\ne\{0\}. \] Choose a minimal nonempty subset \(U\) of the complement whose sum meets the old sum nontrivially. A nonzero vector in that intersection has a representation with all projections on \(U\) nonzero, by minimality. The theorem counts these new blocks by \(N^{|U|/2+\varepsilon'}\). Continue until the whole family is exposed. The group sizes sum to \(r\), and the number of possible ordered partitions of the index set depends only on \(r\). Their product count is \(N^{r/2+O_r(\varepsilon')}\). Multiplying by \(\prod p_x\ll_C N^{-r}\) gives the mass. \end{proof} \begin{corollary}[Dependent pairs] \label{p13:half:pairs} The number of pairs of blocks with nonzero intersection of their effective even-selection spaces is \(O(N^{1+\varepsilon})\), uniformly for fixed logarithmic widths and fixed normalizations. \end{corollary} \begin{proof} Use Theorem~\ref{p13:half:conditional} with \(r=2,s=0\). The two equal nonzero classes give the required relation. \end{proof} \paragraph{Sharpness and range.} For \(r=2k\), take \(k\) pairs of adjacent windows, with the pairs far apart. In each pair select the same two consecutive values in the overlap. Their square class is nonzero, since \(n(n+1)\) lies strictly between \(n^2\) and \((n+1)^2\), and the equal projections cancel. There are \(\gg N^k\) such lists. Thus \(r/2\) is optimal for the general full-support count at even orders. For \(k>1\) these examples can be decomposable; they make no optimality claim for connected or primitive subclasses. Every order and width in this section remains fixed. \subsection{Conditional half-counts and the certificate iteration} \label{p14:comp:half} We prove Theorem~\ref{p13:half:conditional} with its stated uniformity in arbitrary fixed source blocks. Blocks and effective spaces have the meaning specified in Section~\ref{p13:half:section}; in particular, the nonzero projections are required in the individual new spaces, not in their quotients by the old space. \subsubsection{Small kernels and smooth moduli} Choose \(H=C_DB\) larger than every block diameter, and set \[ K_H(n)=\prod_{\substack{\ell>H\\v_\ell(n)\ {\rm odd}}}\ell. \] For \(q>0\), let \(\mathcal X_q\) be the set of centres containing distinct values \(n_1,n_2\) with \(K_H(n_1)K_H(n_2)\le N^q\). \begin{lemma}[Two small kernels] \label{p13:half:small-kernels} For every fixed \(q>0\) and \(\varepsilon>0\), \(\#\mathcal X_q\ll_{q,D,C,\varepsilon}N^{q+\varepsilon}\). \end{lemma} \begin{proof} Write \(n_i=c_i k_i a_i^2\), where \(k_i=K_H(n_i)\) and \(c_i\) is squarefree and supported on primes at most \(H\). The \(c_i\) have \(2^{2\pi(H)}=N^{o(1)}\) choices, while the pairs \(k_1k_2\le N^q\) have \(N^{q+o(1)}\) choices. Fix the two distinct offsets. Their nonzero difference \(h\) gives \[ c_1k_1a_1^2-c_2k_2a_2^2=h. \] The effective coefficients are \(O_D(N)\). Lemma~\ref{p1:arith:pell} gives \(N^{o(1)}\) solutions, including the factorizable case, and each solution determines the centre. The \(O_D(B^2)\) offsets are absorbed. \end{proof} For \(Y=(\log N)^A\), with fixed \(A\ge2\), write \[ n=s_Y(n)a_Y(n)^2K_Y(n). \] Here \(s_Y(n)\) is the entire part supported on primes at most \(Y\); \(a_Y(n)\) contains the halves of the valuations at larger primes; and \(K_Y(n)\) is their odd squarefree part. \begin{lemma}[A large smooth modulus] \label{p13:half:smooth} Fix \(t\), \(c>0\), and \(A\ge2\). There is a deterministic set \(\mathcal M_Y\) of centres, of cardinality \(O(N^{1-1/A+\varepsilon})\), with the following property. If a selection has more than \(c\log\log N\) values, its centre lies outside \(\mathcal X_{1/2}\), and it contains fewer than \(t\) values with \(K_Y(n)>N^{1/5}\), then its centre belongs to \(\mathcal M_Y\). \end{lemma} \begin{proof} Outside \(\mathcal X_{1/2}\), at most one value satisfies \(a_Y(n)^2>N^{3/4+1/100}\): two such values would give \[ K_H(n_1)K_H(n_2) \le\frac{O_D(N^2)}{a_Y(n_1)^2a_Y(n_2)^2}C_1N, \] where \(C_1N\) exceeds the length of the centre interval. Multiply its prime factors, one at a time and with multiplicity, until the first passage above \(C_1N\). The resulting smooth modulus satisfies \[ C_1NC_1N\). The offsets themselves have \(B^{26}=N^{o(1)}\) choices. Choose \(\sigma\) sufficiently close to \(1-1/A\). \end{proof} \subsubsection{Several anchors force a congruence certificate} \begin{lemma}[Incoming and outgoing certificate counts] \label{p13:half:saturation} Suppose a relation in \eqref{p13:half:condition} selects \(t\) values in one new block whose kernels exceed \(N^b\), where \(b=1/5\) and \(tb>r+s-1\). Some other old or new block receives a congruence certificate whose modulus exceeds a fixed multiple of \(NB\). For fixed \(r,s,t,A\), the directed graph of these source--receiver certificates has both incoming and outgoing degrees at most \(N^{1/A+o(1)}\). \end{lemma} \begin{proof} Represent the old vector by a selection of actual old values; parity in the old blocks is not required. Together with the new selections this is a square product. Primes above \(Y\) cannot occur in two distinct values of one block. The \(t\) anchor kernels are therefore pairwise coprime, and their product \(Q\) satisfies \(Q>N^{tb}\), \(Q\ll_D N^t\). Assign each prime of \(Q\) to one of its other odd occurrences. The \(r+s-1\) receiver moduli have product \(Q\). For sufficiently large \(N\), at least one exceeds \(C_2NB\). In that modulus keep the primes only up to its first passage above \(C_2NB\). The retained list has at most \[ k=\left\lfloor\frac{\log(C_2NB)}{\log Y}\right\rfloor+1 \] members. For a fixed source, the \(t\) offsets and a subset of the \(O_t(\log N/\log Y)\) available primes cost \(N^{o(1)}\). Assigning their destinations among \(O_D(B)\) receiver positions costs at most \(B^k=N^{1/A+o(1)}\). The congruences determine the receiver. For a fixed receiver, its values contain \(R=O_D(B\log N/\log Y)\) primes above \(Y\). Choose at most \(k\) of them and assign each to an anchor: the number of codes is at most \[ B^t\sum_{j\le k}\binom Rj t^j=N^{1/A+o(1)}. \] When \(R\ge2k\), use \((k+1)(etR/k)^k\); when \(R<2k\), use \((1+t)^R=N^{o(1)}\). The retained modulus again exceeds the source interval. Forgetting all other primes enlarges the count and does not remove the original full-support relation from the remaining blocks. \end{proof} \subsubsection{The iteration} \label{p14:comp:half:iteration} \begin{proof}[Proof of Theorem~\ref{p13:half:conditional}] We induct strongly on \(r\), simultaneously for every fixed number \(s\) of sources. Put \(T_0=1\). The case \(r=1,s=0\) is empty, since its sole projection is required to be nonzero. For all other cases, choose \[ t=5(r+s-1)+1. \] First choose a small fixed step \(\delta>0\) within the final error margin. Let \[ x_0=0,\qquad x_j=\min\{\tfrac12,x_{j-1}+\delta\}, \qquad A_j=x_j^{-1},\qquad Y_j=(\log N)^{A_j}. \] There are \(J=\lceil1/(2\delta)\rceil\) stages. All these numbers are fixed before \(N\); choose the internal margins after \(J\). Corollary~\ref{p13:short:old-space} counts lists possessing a witness with all new selections short. For every remaining list choose one witness, without enumerating it. It has at least one long new selection. At the start of stage \(j\), every surviving long new source outside \(\mathcal X_{1/2}\) belongs to a set of size \(N^{1-x_{j-1}+\varepsilon'}\). This is trivial for \(j=1\). If such a source selects \(t\) large \(K_{Y_j}\) kernels, Lemma~\ref{p13:half:saturation} supplies a receiver. When the receiver is new, count the source--receiver pair by \(N^{1-x_{j-1}+x_j+O(\varepsilon')}\). The other \(r-2\) projections are still nonzero in their own spaces; their sum belongs to the old space enlarged by these two fixed blocks. The induction hypothesis counts them by \(N^{(r-2)/2+\varepsilon'}\). The total is \[ N^{r/2+x_j-x_{j-1}+O(\varepsilon')} \le N^{r/2+\delta+O(\varepsilon')}. \] If the receiver is old, it is already fixed. The incoming degree counts the new source by \(N^{x_j+O(\varepsilon')}\). Adjoin that source to the old space and apply induction to the \(r-1\) remaining targets. The cost is at most \[ N^{x_j+(r-1)/2+O(\varepsilon')} \le N^{r/2+O(\varepsilon')}. \] Only a fixed number of choices of roles is involved. Remove all these saturated cases. Every surviving long selection outside \(\mathcal X_{1/2}\) has fewer than \(t\) large kernels, and Lemma~\ref{p13:half:smooth} places its centre in a set of size \(N^{1-x_j+\varepsilon'}\). This proves the stage invariant. At the final stage every long source lies in \(\mathcal X_{1/2}\cup\mathcal M_{Y_J}\), whose size is \(N^{1/2+\varepsilon'}\). Fix one such centre and use the induction hypothesis for the other \(r-1\) targets. This costs \(N^{1/2+\varepsilon'}N^{(r-1)/2+\varepsilon'}\). Every application of induction uses fewer targets, although the fixed number of old sources can increase. Thus the simultaneous induction is legitimate. Choosing \(\delta\) and then all internal margins within \(\varepsilon\) proves the theorem. \end{proof} \section{Confining large common divisors} \label{p13:conf:section} At small relative rank, the host count already controls triple probabilities. At macroscopic rank, many large prime factors must be shared between the windows. An interpolation argument confines the offsets of these common divisors to a fixed-degree curve. Unless the pair has an exact rational alignment, that curve contains too few offsets to sustain many large groups. Fix \(H=C_1B\) larger than the diameters and use the odd kernel \(K_H\) from Section~\ref{p13:half:section}. For this section, a pair is called aligned if two distinct offset pairs give the same ratio of true values. Equivalently, two points \((i,j)\) lie on a line through \((-x,-y)\). Reducing the ratio bounds its height by \(O(B)\); the enlarged height-\(B^3\) convention used elsewhere in the paper includes these alignments. The confinement lemma, proved in Section~\ref{p14:comp:confinement}, starts from pairwise coprime groups $g_a>N^\tau$ dividing $x+i_a$ and $y+j_a$. Determinants of bounded-degree monomials force their offset points onto an algebraic curve. Nonlinear components have $o(B)$ grid points, while a line containing many groups must be an exact alignment. After removing these few groups, the surviving coordinates have small odd kernels and the half-count machinery makes their centres rare. The argument never enumerates the auxiliary curve or a chosen relation subspace. For each fixed arity, we also use an anchored version: after coordinate regularization, each active anchor value meets only a fixed number of minimal prime supports. Confinement removes the large supports at a cost of $o(B)$ anchor coordinates. Two remaining small kernels make that centre rare. On a primitive family this works at every window, giving $N^\varepsilon$ dense hosts and the fixed-size weights in Proposition~\ref{p13:transfer:primitive}. The arity and density proportion are fixed before $N$. The same section completes the stronger pair profile of Proposition~\ref{p1:arith:hosts} without using that profile as an input. The two forms needed to partition compatible triples are as follows. The argument also proves the primitive-triple and dense-pair versions, Theorem~\ref{p13:conf:dense-primitive} and Corollary~\ref{p13:conf:dense-pairs}. \begin{proposition}[Dense triples with weak proper intersections] \label{p13:conf:near-primitive} Fix \(\eta>0\) and \(0<\theta<\eta/100\). Admissible triples satisfying \[ \rho_{\mathrm{rel}}(T)\ge\eta B,\qquad d_{xy}\le\theta B\quad\text{for every pair} \] have \(O_{\eta,\theta,C,\varepsilon}(N^\varepsilon)\) hosts and total homogeneous mass \(O(N^{-2+\theta+\varepsilon})\). \end{proposition} Its proof is given in Section~\ref{p14:comp:confinement:near}. \begin{proposition}[A strong nonaligned pair in a triple] \label{p13:conf:strong-profile} Fix \(\theta,\sigma,\varepsilon>0\). Compatible connected triples with a nonaligned pair of rank at least \(\theta L\), and \(\rho_{\mathrm{rel}}(T)\le\sigma L\), have total joint mass \[ O_{\theta,\sigma,C,\varepsilon} \bigl(N^{-\min\{3/2,\,3-\sigma\}+\varepsilon}\bigr). \] \end{proposition} Its proof is given in Section~\ref{p14:comp:confinement:strong}. The last proposition preserves the actual conditional probabilities. Replacing them by independent weights would lose the large-rank case. What remains is a pointwise bound on the possible pair ranks and a separate calculation for rationally aligned triples. \subsection{Polynomial confinement and dense arithmetic profiles} \label{p14:comp:confinement} We use the block and alignment conventions of Section~\ref{p13:conf:section}. The group argument treats nonaligned pairs and primitive families at every fixed arity, then proves the two triple interfaces stated there. \begin{lemma}[Polynomial confinement of common divisors] \label{p13:conf:groups} Fix \(\tau>0\), \(0<\delta<1/2\), and \(C_0>0\). Let \(x,y\asymp N\), and let pairwise coprime integers \(g_a>N^\tau\) satisfy \[ g_a\mid x+i_a,\quad g_a\mid y+j_a,\qquad |i_a|,|j_a|\le C_0B. \] Suppose no line through \((-x,-y)\) contains two distinct points among the \((i_a,j_a)\). Then their number, with multiplicity, is \(O_{\tau,\delta,C_0}(B^{1/2+\delta})=o(B)\). The same estimate holds after points on exact alignments have been removed. \end{lemma} \begin{proof} Choose a fixed degree \(d\) such that, for \(M=(d+1)(d+2)/2\), \((M-1)\tau>d\). Let \(v_d(s,t)\) list the \(M\) monomials of total degree at most \(d\), including \(1\). If \(M\) of the vectors \(v_d(i_a,j_a)\) form a basis, replace each row of that basis in turn by \(v_d(-x,-y)\). Subtracting the new row from the other rows shows that the resulting determinant is divisible by the product of their \(M-1\) pairwise coprime divisors. Its absolute value is at most \[ C_dN^dB^{d(M-1)}0$. Let $M$ be a binary space of physical square relations on $r$ true windows with centres in $I_N$: its words select an even set in each window and their total product has even prime valuations. Write $\ell_v\in M^*$ for the coordinate forms and call a position active when $\ell_v\ne0$. Designate one window as the anchor. Assume that: \begin{itemize} \item $\dim M\ge\eta L$ and the physical projection into the anchor is injective; \item no nonzero form on $M$ has $t$ representations by pairwise disjoint coordinate sets of size between $1$ and $r-1$; \item between the anchor and each other block, no two distinct pairs of active offsets have the same ratio of true values. \end{itemize} For every cutoff $H\ge B=L+1$, the anchor then contains two distinct values $n_1,n_2$ with $K_H(n_i)\le N^{D\tau}$, where \[ D=\sum_{j=1}^{r-1}j!(t-1)^{j-1}2^{j(t-1)}. \] One threshold, chosen after $r,t,C,\eta,\tau$, precedes the centres, $M$ and $H$. No injectivity or nonalignment between the other blocks is assumed. \begin{proof} The recovery-exclusion hypothesis gives multiplicity at most $t-1$ for every nonzero coordinate label, by considering singleton recovery sets. The final counting argument of Lemma~\ref{p1:fixed:recovery} also bounds the independent $j$-label representations of a nonzero label by $j!(t-1)^{j-1}$: a maximal disjoint family of their supports has fewer than $t$ members; its union, of size at most $j(t-1)$, meets every representation, and induction removes a label. A zero remaining sum is excluded by independence. Permitting every subset of the $j$ label fibres gives at most $2^{j(t-1)}$ physical lifts. Thus at most $D$ physical supports of size at most $r-1$ can represent a fixed active anchor label. A prime $\ell>H$ has at most one odd occurrence in each true window. If it occurs at an active anchor position $v$, its parity row on $M$ expresses $\ell_v$ as a sum of at most $r-1$ other labels. Remove zero labels and choose a minimal subrepresentation. Its labels are distinct, nonzero and independent: otherwise a nonempty zero-sum subset could be removed without changing the represented nonzero form. Choose their actual positions in the prime row. The resulting support $A$ consists of $v$ and positions in other blocks and uses only active coordinates. Assign every prime with an active anchor occurrence to one such support, and let $g_A$ be the product assigned to $A$. Distinct nontrivial groups are coprime and divide every true value selected by their support. Every anchor position meets at most $D$ groups, and its entire odd kernel above $H$ divides their product. Assign each group to one other block it meets. For each of the $r-1$ choices, Lemma~\ref{p13:conf:groups} gives $o(B)$ groups larger than $N^\tau$, with their full multiplicities. Its no-line hypothesis is exactly the stated active-offset condition. Equal projected points are covered by the multiplicity bound in that lemma. Impose zero only on the anchor positions touched by these large groups. Each group has a unique anchor position, so the dimension cost is at most the number of such groups, namely $o(B)$. A nonzero space remains for large $N$ and its anchor projection is still injective. A nonzero even projected word uses at least two positions. Each of those positions was active on the original space $M$ and met only groups at most $N^\tau$, so its odd kernel is at most $N^{D\tau}$. The incidence bound is used on $M$ before erasure; no recovery condition or prime-row classification on the erased space is assumed. \end{proof} \paragraph{Dense primitive hosts and their exact weights.} \label{p18:conf:fixed-primitive} For each fixed $r\ge3$, $\eta>0$ and $\varepsilon>0$, primitive families of $r$ true windows with $\rho_{\mathrm{rel}}\ge\eta L$ have $O(N^\varepsilon)$ hosts. Indeed, every effective relation-space projection is injective. Choose linear sections of the local even-selection maps to lift that space without increasing its dimension or adding local square kernels. Lemma~\ref{p1:fixed:recovery}, with $d=r-1$ and $t$ chosen after $r,\eta$, loses at most $2rB\,t^{-1/(r-1)}\le\eta L/4$ dimensions. The surviving physical space has dimension at least $3\eta L/4$, retains all injective projections, and has the required recovery exclusion. Two exact units between any pair would give equal nonzero even square classes in those two blocks: the selected two-value products are nonsquares by Lemma~\ref{p1:arith:runge}. This contradicts primitivity, since the pair is a proper subfamily. Thus the active-offset condition holds for each choice of anchor. Apply the witness with $H=B$ and density parameter $\eta/2$ at each of the $r$ windows, with the same fixed constant $D$. Every centre lies in $\mathcal X_{2D\tau}$. Lemma~\ref{p13:half:small-kernels} therefore counts at most $N^{2rD\tau+r\varepsilon'}$ unordered families. Choose $\tau$ and $\varepsilon'$ after $r,\eta,\varepsilon$ so that this exponent is at most $\varepsilon$. No auxiliary support or chosen subspace is enumerated in this last count. The conversion of this host count to homogeneous probabilities and the three original weights is given in Proposition~\ref{p13:transfer:primitive}. That proposition uses the host count proved here, together with the conditional half-count, and treats the weight comparison at fixed arity. \begin{theorem}[Dense primitive triples] \label{p13:conf:dense-primitive} For fixed \(\eta,\varepsilon>0\), primitive triples of true windows with \(\rho_{\mathrm{rel}}(T)\ge\eta B\) have \(O_{\eta,C,\varepsilon}(N^\varepsilon)\) hosts. Their homogeneous mass is \(O(N^{-2+\varepsilon})\). The total absolute centered activity of all primitive triples is \(O(N^{-3/2+\varepsilon})\). \end{theorem} \begin{proof} Apply Proposition~\ref{p13:transfer:primitive} with $r=3$. The dense condition $\rho_{\mathrm{rel}}(T)\ge\eta B$ implies $\rho_{\mathrm{rel}}(T)\ge\eta L$, giving the host count and the homogeneous mass $N^{-2+\varepsilon}$. The global absolute-centered bound is $N^{-3/2+\varepsilon}$. \end{proof} Here homogeneous mass is the sum of the probabilities of the homogeneous events, with zero right-hand side in place of the affine start prescription. It must be distinguished from a weight using the original marginals $p_x$, which can vanish. For instance, a start at $x=3$ of length $L=7$ is impossible: the square $4$ forces the common bit on $3,\ldots,9$ to be zero, and $6=2\cdot3$ then forces the bit at $2$ to be zero, contradicting the left-boundary equation. The corresponding homogeneous event is compatible and has positive probability. \begin{corollary}[Dense nonaligned pairs] \label{p13:conf:dense-pairs} For fixed \(\eta,\varepsilon>0\), nonaligned pairs with \(d_{xy}\ge\eta B\) have \(O(N^\varepsilon)\) hosts. \end{corollary} \begin{proof} Realize their effective relation space with injective projections. Regularize with Lemma~\ref{p1:fixed:recovery}, now with \(d=1\), losing at most \(\eta B/4\). Every active prime row has two equal labels. A coordinate belongs to at most \(K=t-1\) active supports. Confinement bounds the groups above \(N^\tau\) by \(o(B)\). Kill both endpoints of every such group. The remaining positive-dimensional space has two active positions in each block, with kernels at most \(N^{K\tau}\). The centres lie in \(\mathcal X_{2K\tau}\). Their count is at most \(N^{4K\tau+2\varepsilon'}\); choose \(\tau,\varepsilon'\) sufficiently small. \end{proof} \paragraph{Completion of the quasi-linear pair profile.} \label{p18:conf:pair-profile} We now prove the stronger bounds in Proposition~\ref{p1:arith:hosts}, retaining its full arithmetic nullity $\rho(x,y)=2L-\dim(S_x+S_y)$. If both windows are regular, positive $\rho$ is exactly a nonzero effective intersection; Corollary~\ref{p13:half:pairs} counts these pairs by $N^{1+\varepsilon'}$. A locally deficient coordinate gives at most $2N\,|I_N\setminus G_N|=N^{1+o(1)}$ further ordered pairs. This proves the support bound, even before imposing separation. For the weighted nonaligned sum, split at $d_{xy}=\eta L$ with fixed $\eta>0$. Local rigidity gives, uniformly, \[ \rho(x,y)=d_{xy}+(L-d_x)+(L-d_y)\le d_{xy}+2, \qquad \rho(x,y)\le L+2. \] In the low sector, the just proved support bound and $2^{\eta L+2}\ll_C N^\eta$ give $N^{1+\eta+\varepsilon'}$. In the high sector, Corollary~\ref{p13:conf:dense-pairs}, with a smaller fixed density parameter if necessary to replace $L$ by $B$, gives $N^{\varepsilon'}$ hosts, each of weight $O_C(N)$. Choose $\eta,\varepsilon'$ after the final $\varepsilon$. This proves both assertions with $N^{1+\varepsilon}$ and with a common threshold before all pairs. Only conditional half-counts, local rigidity and confinement were used; the pair-profile theorem and its elementary $6/5$ proof are not inputs to this completion. \subsubsection{Proof of Proposition~\ref{p13:conf:near-primitive}} \label{p14:comp:confinement:near} \begin{proof} The kernel of each projection of the relation space has dimension at most \(\theta B\). Intersect three complements to these kernels. The resulting subspace loses at most \(3\theta B\) dimensions and has all projections injective. Use linear sections of the local even-selection maps to realize this subspace in physical coordinates, preserving its dimension and the injectivity of its block projections. In a given pair, two exact units impose a reduced rational alignment of height \(O(B)\). Two distinct such alignments would bound the centres by \(O(B^3)\), which is impossible at the present scale. The alignment is therefore unique. If it has \(u\) units, their even-selection space has dimension \(u-1\); its image loses at most one more dimension by Lemma~\ref{p1:limits:local-rigidity}. Thus \(u\le d_{xy}+2\le\theta B+2\). Kill all the positions of these units in the three pairs, at a cost of at most \(6\theta B+12\) dimensions. The remaining space still has macroscopic dimension and injective projections. Apply Lemma~\ref{p1:fixed:recovery} with \(d=2\), choosing a fixed integer \(t\ge2\) such that \(6t^{-1/2}\le\eta/8\). There are at most \(3B\) physical coordinates, so this regularization loses at most \(\eta B/8\) further dimensions. Denote the resulting physical relation space by \(M\). For all sufficiently large \(N\), \[ \dim M\ge \left(\frac{7\eta}{8}-9\theta\right)B-12 \ge\frac{\eta L}{2}, \] since \(\theta<\eta/100\). Apply the \hyperref[p18:conf:anchored]{anchored small-kernel witness} of Section~\ref{p14:comp:confinement} separately to each of the three windows, using the same space \(M\), the fixed cutoff \(H\), and density parameter \(\eta/2\). Its block projections are injective, its coordinate labels satisfy the recovery exclusion, and its active offsets have no remaining exact alignments. Let \(D_*=D(3,t)\) be the support-incidence constant in that witness. Each centre therefore belongs to \(\mathcal X_{2D_*\tau}\), and Lemma~\ref{p13:half:small-kernels} bounds the number of triples by \[ O\!\left(N^{6D_*\tau+3\varepsilon'}\right). \] Choose \(\tau,\varepsilon'>0\) after the requested \(\varepsilon\) so that \(6D_*\tau+3\varepsilon'\le\varepsilon/2\). In each application, the incidence bound is used before the final anchor-coordinate erasure. No auxiliary subspace or deletion history is enumerated. Finally, \(\rho_{\mathrm{rel}}(T)\le d_z+d_{xy}\le L+\theta B\). The homogeneous weight is at most \(N^{-2+\theta+o(1)}\). Absorbing the \(o(1)\) in the remaining exponent margin proves the mass estimate. \end{proof} \subsubsection{Proof of Proposition~\ref{p13:conf:strong-profile}} \label{p14:comp:confinement:strong} \begin{proof} Choose the strong pair \(e=\{x,y\}\). Put \(\delta_z=\dim(S_z\cap(S_x+S_y)) =\rho_{\mathrm{rel}}(T)-d_{xy}\). Connectedness gives \(\delta_z>0\), and compatibility gives \(q(T)=q(e)p_z2^{\delta_z}\). For \(0<\delta_z<\eta'L\), the conditional count in Theorem~\ref{p13:half:conditional} gives \(N^{1/2+\varepsilon'}\) targets of weight at most \(N^{-1+\eta'+o(1)}\). Lemma~\ref{p1:arith:dense}, together with Corollary~\ref{p13:half:pairs} for the independent part, bounds the total joint mass of strong nonaligned pairs by \(N^{-1+\varepsilon'}\). This branch costs \(N^{-3/2+\eta'+O(\varepsilon')}\). For \(\delta_z\ge\eta'L\), the effective-space form of Proposition~\ref{p1:fixed:targets}, explained after \eqref{p13:exact:kappa3}, gives \(N^{\varepsilon'}\) targets after the fixed pair. Corollary~\ref{p13:conf:dense-pairs} also gives \(N^{\varepsilon'}\) source pairs. Each triple has \(q(T)\ll N^{-3+\sigma+o(1)}\). Choose \(\eta'\) and the internal margins after \(\varepsilon\). \end{proof} \section{Short square relations and the rank of a pair} \label{p13:rig:section} A large intersection of the square-class spaces of two windows contains many relations, but those relations need not initially have short support. We first extract a large subspace with a short basis. Two independent basis vectors then produce an algebraic curve through the two window centres. A resultant argument puts the whole subspace on one curve, while the relations force many integer points in the small grid of offsets. A point count uniform in the degree rules out this configuration unless the centres lie on a rational alignment. The distinction between a fixed degree and a growing degree is essential: the basis supports grow with $L$ throughout the argument. The combination of resultant elimination and degree-uniform curve bounds also appears in Shparlinski and Sleiman's work on consecutive multiplicatively dependent triples \cite[Theorem~1.1 and Section~3]{shparlinskisleiman2026}. Here the common curve must accommodate an entire space of square-class relations and the integer fibres forced by that space in the offset grid. All estimates in this section are uniform in $x,y\in I_N$, $|L-\log_2N|\le C$, for fixed $C$, and we put $B=L+1$. A logarithm without a subscript is natural. The conclusion used in the probability argument is Corollary~\ref{p13:rig:triple}. \subsection{Square classes on all windows} \label{p13:rig:notation} We use the true window $V_x=\{x-1,x,\ldots,x+L-1\}$, including its left endpoint. Let $E_x$ be the space of even selections from $V_x$, and let \[ \phi_x:E_x\longrightarrow \mathbb Q_{>0}^{\times}/\mathbb Q_{>0}^{\times2}, \qquad I\longmapsto\left[\prod_{n\in I}n\right] \] be the square-class map. Its image is the space $S_x$ already used above. Put $d_{xy}=\dim(S_x\cap S_y)$ and, for any finite set of centres $T$, define the \emph{relative nullity} \begin{equation} \rho_{\rm rel}(T)=\sum_{x\in T}\dim S_x- \dim\left(\sum_{x\in T}S_x\right). \label{p13:rig:relative-nullity} \end{equation} On the regular set, where every $S_x$ has dimension $L$, this is the earlier nullity $\rho(T)$. Formula \eqref{p13:rig:relative-nullity} also applies to the remaining windows and separates their local rank defects from dependence between distinct windows. A \emph{complete relation} of a pair is a pair of even selections with the same square class. We represent its selections by offsets $I,J\subset\{-1,0,\ldots,L-1\}$ and attach the polynomial \begin{equation} F_w(X,Y)=\prod_{u\in I}(X+u)\prod_{v\in J}(Y+v), \qquad w=(I,J). \label{p13:rig:word-polynomial} \end{equation} Thus $F_w(x,y)$ is an integer square. The word $w$ is a vector over $\F_2$; addition means symmetric difference in each block. The two blocks remain disjoint formal copies even when the actual windows overlap. Consequently, independent words give independent classes of their polynomials in $\mathbb Q(X,Y)^\times/\mathbb Q(X,Y)^{\times2}$: the factors $X+u$ and $Y+v$ are distinct irreducible polynomials. We call $x,y$ \emph{aligned} if there are coprime positive integers $a,b\le B^3$ such that \begin{equation} |bx-ay|\le(a+b)B. \label{p13:rig:alignment} \end{equation} This definition includes nearby centres. Otherwise they are nonaligned. Alignment is a property of the centres; compatibility of their run events is a separate condition. \subsection{Extracting short relations} Primes larger than $10B$ meet each physical block at most once. Their parity equations therefore join positions into components; a component meeting a prime only once is forced to zero. Local square-product separation shows that there are only $O_C(B/\log B)$ defective positions. Thus the number of free components meeting both blocks is $d_{xy}+O_C(B/\log B)$; see Lemmas~\ref{p13:rig:defective} and~\ref{p13:rig:components}. For medium primes, columns of weight two can be kept as edges of a graph. Its short cycles cancel their equations. Discarding only columns of weight at least three costs one third of their total weight, explaining the factor $2/3$ in the following extraction. \begin{theorem}[A short basis with a controlled loss of dimension] \label{p13:rig:short-basis} For every fixed $a\in(0,1)$ there is a subspace $\mathcal K_a$ of complete relations of the pair, with injective projections onto each physical block and onto the common square class, such that \begin{equation} \dim\mathcal K_a\ge d_{xy}-\frac23B\log(1/a) -O_{a,C}\!\left(\frac B{\sqrt{\log B}}\right). \label{p13:rig:short-dimension} \end{equation} It has a basis whose vectors have total physical support at most \begin{equation} O_{a,C}\!\left( \frac{B^a\log B}{\log\log B}\right). \label{p13:rig:short-support} \end{equation} Every nonzero vector selects a nonempty even set in each block and has nonzero common square class. Neither compatibility nor nonalignment is assumed. \end{theorem} The full extraction proof, including the final small-prime equations and projection injectivity, is in Section~\ref{p14:comp:rigidity:short-basis}. \subsection{One curve for a whole relation space} Two independent short words first yield a nonzero polynomial through the window centres. Proposition~\ref{p13:rig:two-witnesses} constructs it as the norm of a biquadratic auxiliary function, with degree and logarithmic height linear in the support size. Polynomial gcd removal is performed before specialization, so vanishing at the integer point does not destroy the norm argument. The simultaneous step below is stronger than constructing a separate curve for each word. Pairwise resultants have quadratic height cost in the support, independent of the number of basis vectors. \begin{theorem}[A common curve] \label{p13:rig:common-curve} There is an absolute constant $c>0$ with the following property. Let $V$ be a space of even formal words of dimension $t\ge2$, with a basis $w_1,\ldots,w_t$ satisfying $|I_{w_i}|,|J_{w_i}|\le2q_0$, where $1\le q_0\le B$. Suppose all $F_{w_i}(x,y)$ are integer squares and $x,y\in[N,2N)$. If $N>(B+1)^{cq_0^2}$, there is a primitive irreducible polynomial $C_0\in\Z[X,Y]$, depending on both variables, such that \begin{equation} C_0(x,y)=0,\qquad \deg C_0\le48q_0, \qquad H(C_0)\le(B+1)^{cq_0}. \label{p13:rig:common-degree-height} \end{equation} In the function field $K_0=\mathbb Q(C_0=0)$, the image of \[ V\longrightarrow K_0^\times/K_0^{\times2}, \qquad w\longmapsto[F_w] \] has dimension at most one. The constant is independent of $t$. \end{theorem} The proof is given in Section~\ref{p14:comp:rigidity:common-curve}. Its key point is that every norm polynomial must share the same irreducible factor through $(x,y)$: otherwise a resultant would give a nonzero polynomial of height $(B+1)^{O(q_0^2)}$ with integer root $x\ge N$. Primitivity of the auxiliary coefficients then forces the functional square-class image to have dimension at most one. \subsection{Functional squares force integer fibres} The centres have size $N$, but the next step counts points in a box of side $O(B)$. The bridge is that a square function has even valuation at every point of the normalized curve. A simple zero belonging to a selected first-block offset must therefore be matched by a selected second-block offset. \begin{proposition}[Integer fibres in the offset grid] \label{p13:rig:grid-fibres} Under the hypotheses of Theorem~\ref{p13:rig:common-curve}, assume both physical projections of $V$ are injective. Put $D=\deg C_0$, $d_X=\deg_XC_0$, $d_Y=\deg_YC_0$. For an absolute constant $c$, \begin{equation} \#\{(u,v)\in[-B,B]^2\cap\Z^2:C_0(-u,-v)=0\} \ge\max(d_X,d_Y)(t-1-cD^2). \label{p13:rig:grid-count} \end{equation} A negative right side may be replaced by zero. \end{proposition} \begin{proof} The kernel $V_0$ of the map to functional square classes has dimension at least $t-1$. The leading coefficients and discriminants of $C_0$ in each variable are nonzero polynomials, of total degree $O(D^2)$ in the remaining variable. Exclude the offsets at which one of these polynomials vanishes. Requiring the corresponding coordinates of a word to be zero leaves a subspace $V_1$ of dimension at least $t-1-cD^2$. Let $\mathcal U_X$ be the union of first-block offsets used by $V_1$. Projection injectivity gives $|\mathcal U_X|\ge\dim V_1$. For $u\in\mathcal U_X$, select a word containing $X+u$. The noncritical fibre $X=-u$ has exactly $d_Y$ distinct affine roots on the geometric normalization. At each, $X+u$ has a simple zero and no other first-block factor vanishes. The product for the selected word is a square function, so a selected factor $Y+v$ must vanish there. Every root is thus an integer $-v$ in the offset interval. Distinct $u$ give disjoint fibres, proving at least $d_Y\dim V_1$ grid points. Interchanging $X,Y$ proves the bound with $d_X$ as well. The argument may be performed after extension to $\mathbb C$ and normalization of the components; geometric integrality is not required. \end{proof} \subsection{The point count at growing degree} We use the integer-point part of Binyamini--Cluckers--Novikov \cite[Theorem~2, equation~(27), in the specified author version]{binyaminicluckersnovikov2024}. \begin{theorem}[Binyamini--Cluckers--Novikov] \label{p13:rig:bcn} For an integral algebraic curve $\mathcal C\subset\mathbb A^n_K$ of degree $D$, where $n>1$ and $K\subset\mathbb C$, there are a constant $c(n)$ and an absolute constant $\kappa$ such that \[ \#\{z\in\mathcal C\cap\Z^n:\max_i|z_i|\le H\} \le c(n)D^2H^{1/D}(\log H)^\kappa, \qquad H>2. \] The constants are independent of the curve's coefficients and degree. Integrality over $K$ is sufficient. \end{theorem} For curves over $\mathbb Q$, a corresponding affine estimate is also given by Binyamini, Cluckers, and Kato \cite[Corollary~8, equation~(23)]{binyaminicluckerskato2025}. The external theorem above asserts the existence of an absolute logarithmic exponent. The admissible value $\kappa=17$ is supplied by the pinned Lean declaration \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/a1c8092f7e25acf7f3b78283039cfbd69491f88e/PaperP1/Arithmetic/IntegralAffineCurvePointCount.lean} {\texttt{article\_13\_4}} in the namespace \texttt{PaperP1.IntegralAffineCurvePointCount}. Only the existence of a fixed exponent is used below. \begin{corollary}[Sublinear counts in the offset box] \label{p13:rig:sublinear-grid} Fix $0\le a<1/2$ and $c_0>0$. Let $D_{\max}=D_{\max}(B)$ be any fixed degree envelope such that, for every $\delta>0$, $D_{\max}(B)\le B^{a+\delta}$ for all sufficiently large $B$. Uniformly for integral plane curves of degree $2\le D\le D_{\max}(B)$, \[ \#(\mathcal C\cap[-c_0B,c_0B]^2\cap\Z^2) \ll_{c_0}(B^{1/2}+D_{\max}^2)(\log B)^\kappa=o(B). \] \end{corollary} \begin{proof} Set $H=\max(3,c_0B)$. The logarithmic derivative of $u^2H^{1/u}$ is $(2u-\log H)/u^2$, so its maximum on $[2,M]$, $M\ge2$, is attained at an endpoint. Put $M'=\max(M,\log H)$. Applying this observation on $[2,M']$, and using $H^{1/M'}\le e$, gives, for every $2\le D\le M$, \[ D^2H^{1/D} \le 4H^{1/2}+e(M')^2 \le4H^{1/2}+e\{M^2+(\log H)^2\}. \] The last logarithmic square is absorbed into $B^{1/2}$. Apply Theorem~\ref{p13:rig:bcn} with $n=2$ and $M=D_{\max}(B)$. Choose a fixed $\delta>0$ with $2(a+\delta)<1$. Both $B^{1/2}(\log B)^\kappa$ and $B^{2(a+\delta)}(\log B)^\kappa$ are $o(B)$. The threshold is chosen after $c_0,a$, the envelope and the requested error, and before the curve and its coefficients. No lower growth of $D_{\max}$ is needed; bounded and polylogarithmic envelopes are included. \end{proof} This application uses the theorem's degree-uniform constant. A bound with an unspecified constant depending on $D$ would not give the conclusion of Corollary~\ref{p13:rig:sublinear-grid}. \subsection{Nonaligned pairs have rank less than half a window} \begin{theorem}[A uniform nonaligned rank bound] \label{p13:rig:nonaligned} For every fixed critical band, \begin{equation} \limsup_{N\to\infty}\ \sup_{\substack{x,y\in I_N\\x,y\ {\rm nonaligned}}} \frac{d_{xy}}B \le\frac23\log2<\frac12. \label{p13:rig:nonaligned-bound} \end{equation} Compatibility of the two events is not needed. \end{theorem} \begin{proof} Suppose along a sequence that $d_{xy}\ge((2/3)\log2+\delta)B$ for some fixed $\delta>0$. Choose a fixed $a<1/2$, sufficiently close to $1/2$, that $(2/3)\log(1/a)<(2/3)\log2+\delta/2$. Theorem~\ref{p13:rig:short-basis} gives a space $V$ of dimension at least $\delta B/2-o(B)$ with injective physical projections and a basis whose half-cardinalities are at most $q_0=B^{a+o(1)}$. The quantitative bridge to the common-curve theorem is \begin{equation} q_0^2\log(B+1)=B^{2a+o(1)}=o(B)=o(\log N). \label{p13:rig:threshold-bridge} \end{equation} Thus Theorem~\ref{p13:rig:common-curve} supplies an irreducible $C_0$ through $(x,y)$ of degree $D\le B^{a+o(1)}$ and logarithmic height $o(B)$, with functional image of dimension at most one. If $D\ge2$, Proposition~\ref{p13:rig:grid-fibres} gives at least $\delta B/2-o(B)$ integer points in the offset box, because $D^2=o(B)$. Corollary~\ref{p13:rig:sublinear-grid} gives $o(B)$ points, a contradiction. If $D=1$, write $C_0=\alpha X+\beta Y+\gamma$. Both $\alpha,\beta$ are nonzero. Since $H(C_0)=\exp(o(B))=N^{o(1)}$ and $x,y\ge N$, they have opposite signs; otherwise $\gamma$ could not cancel the two terms. The line therefore has the form $X=\lambda Y+\nu$ with $\lambda>0$. The kernel of the functional-class map has positive dimension. Choose a nonzero word in it. Each selection is nonempty and even by projection injectivity. On the line, the distinct simple roots from the two blocks must pair, each pairing giving $\nu+u=\lambda v$. There are at least two pairings, hence \[ \lambda=\frac{u_1-u_2}{v_1-v_2}=\frac pq, \qquad (p,q)=1,\qquad 1\le p,q\le2B. \] Evaluating one pairing at the centres gives $qx-py=pv-qu$, of absolute value at most $(p+q)B$. This is an alignment as in \eqref{p13:rig:alignment}, contrary to the assumption. Both degree cases are impossible, and $\delta$ was arbitrary. \end{proof} The number $(2/3)\log2$ is a sufficient upper threshold arising from the extraction and degree bounds. No matching rank construction is asserted for this constant. \subsection{Aligned pairs and the exact contact distance} For an aligned pair the line is already known. Substitution into it costs a height linear in the support, so the extraction exponent may approach one without the quadratic resultant threshold. \begin{proposition}[Rank on a known alignment] \label{p13:rig:aligned} Suppose $bx-ay=h$, where $(a,b)=1$, $1\le a,b\le B^3$, and $|h|\le(a+b)B$. For every fixed $\tau\in(0,1)$, \begin{equation} d_{xy}\le\frac B{\max(a,b)}+\frac23B\log(1/\tau) +O_{\tau,C}(B/\sqrt{\log B}). \label{p13:rig:aligned-quantitative} \end{equation} For distinct compatible centres this implies, uniformly over aligned pairs, $d_{xy}\le B/2+o_C(B)$. If such a pair has an alignment with $a=b=1$, then $d_{xy}=o_C(B)$. \end{proposition} The proof in Section~\ref{p14:comp:rigidity:aligned} counts the matched linear factors on that line. At height one, compatibility imposes $|x-y|\ge L$; distance exactly $L=B-1$ leaves only one shared true position. A nonzero even relation needs at least two matches, so this contact contributes no macroscopic intersection rank. Thus the bound includes the minimum compatible separation exactly. \subsection{Uniform triple nullity and its probabilistic cost} \begin{corollary}[Triple rank and maximum joint probability] \label{p13:rig:triple} Every compatible triple $T$ of distinct centres in the fixed critical band satisfies \begin{equation} \rho_{\rm rel}(T)\le\frac32L+o_C(L). \label{p13:rig:triple-nullity} \end{equation} If any of its pairs is nonaligned, the stronger bound $\rho_{\rm rel}(T)\le(1+(2/3)\log2)L+o_C(L)$ holds. In particular, with $q(T)=\E\prod_{x\in T}J_x$, \begin{equation} \max_{\substack{T\subset I_N\\|T|=3}}q(T) \le N^{-3/2+o_C(1)}. \label{p13:rig:max-triple} \end{equation} \end{corollary} \begin{proof} For three spaces of dimensions at most $L$, the dimension identity and the intersection bound give \[ \begin{split} \rho_{\rm rel}(T) &=\dim(S_1\cap S_2)+\dim((S_1+S_2)\cap S_3)\\ &\le L+\dim(S_1\cap S_2). \end{split} \] Permuting the indices gives $\rho_{\rm rel}(T)\le L+\min_{i(B+1)^{5m}$, then $\prod_{a\in I}(x+a)$ is not an integer square. \end{lemma} \begin{proof} The assumed threshold implies $x\ge2B$. If the product were a square, Lemma~\ref{p1:arith:runge}, with $R=B$ and $d=m$, would give $x\le(64B)^m$. For $B\ge2$, we have $64B\le(B+1)^5$. Thus $x\le(B+1)^{5m}$, contrary to the hypothesis. \end{proof} \begin{lemma}[Few defective positions in every critical window] \label{p13:rig:defective} The number of $n\in V_x$ for which \[ \prod_{\substack{p>10B\\v_p(n)\ {\rm odd}}}p=1 \] is $O_C(B/\log B)$, uniformly over all critical windows. \end{lemma} \begin{proof} Apply the local assertion of Lemma~\ref{p1:arith:defects} with $H=10B$ to $V_x$. In the critical band its diameter is $O(B)$ and its values lie in $[N/2,3N]$ for all sufficiently large $N$. The comparison constants are uniform in $x$. \end{proof} We use the elementary bound $\pi(t)\ll t/\log t$. For completeness, the product of the primes in $(u,2u]$ divides $\binom{2u}{u}\le4^u$ for integer $u$; dyadic summation gives $\vartheta(t)=\sum_{p\le t}\log p\ll t$. Splitting at $\sqrt t$ gives the stated bound for $\pi(t)$. We will also need the elementary prime harmonic estimate \begin{equation} \sum_{B^a10B$ has at most one occurrence of odd valuation in either block. If it has two occurrences, join their vertices; if it has only one, mark that vertex as forced to zero. The corresponding binary equation says that the endpoints of an edge have equal coordinates, or that a marked coordinate is zero. A component is free when it contains no marked vertex. \begin{lemma}[Free components] \label{p13:rig:components} Let $c$ be the number of free components containing an edge. Then \[ c=d_{xy}+O_C(B/\log B). \] Each such component meets both blocks and is selected in its entirety or omitted in any solution to the large-prime equations. \end{lemma} \begin{proof} The isolated unmarked vertices are precisely the defective positions. Their total number $D_0$ is $O_C(B/\log B)$ by Lemma~\ref{p13:rig:defective}. The kernel of the large-prime equations has dimension $c+D_0$. Imposing the remaining prime equations and the two block parities costs at most $\pi(10B)+2$ dimensions. If $\mathcal R$ is the space of complete relations and $k_x=\dim\ker\phi_x$, then the map taking a relation to its common class gives an exact sequence with kernel $\ker\phi_x\oplus\ker\phi_y$. Hence \[ c+D_0-\pi(10B)-2\le\dim\mathcal R=k_x+k_y+d_{xy} \le c+D_0. \] A local relation can use only defective positions, so $k_x+k_y\le D_0$. These inequalities prove the assertion. \end{proof} \subsubsection{Proof of Theorem~\ref{p13:rig:short-basis}} \label{p14:comp:rigidity:short-basis} \begin{proof} \emph{A graph estimate.} A multigraph with $n$ vertices and $e/n\to\infty$ has a cycle of length $O(1+\log(n+1)/\log(e/n))$. Loops and parallel edges give cycles of length one and two. Otherwise delete vertices of degree less than $t=\lfloor e/(2n)\rfloor$. Not all vertices can disappear, since fewer than $nt\le e/2$ edges would then be deleted. The remaining graph has minimum degree at least $t$. If its girth exceeds $2r+1$, a breadth first search contains at least $1+t\sum_{j0$ with the following property. Let $u,v$ be independent formal words, even in each block, with offsets in $[-B,B]$. For $w=u,v,u+v$, write $|I_w|=2r_w$, $|J_w|=2s_w$, and assume $\max(r_w,s_w)\le q$, where $1\le q\le B$. Suppose that $F_u(x,y)$ and $F_v(x,y)$ are integer squares, with $x,y\in[N,2N)$. If $N>(B+1)^{cq}$, there is a nonzero $R\in\Z[X,Y]$ such that \begin{equation} R(x,y)=0,\qquad \deg_XR,\deg_YR\le12q,\qquad \deg R\le24q, \qquad H(R)\le(B+1)^{cq}. \label{p13:rig:two-polynomial} \end{equation} Moreover, $R$ is the norm of a biquadratic auxiliary function whose four polynomial coefficients have no common nonconstant factor. \end{proposition} \begin{proof} Set $F_1=F_u$, $F_2=F_v$, $G=\gcd(F_1,F_2)$, with monic gcd, and $F_3=F_{u+v}=F_1F_2/G^2$. Formal independence makes \[ K=\mathbb Q(X,Y)(W_1,W_2),\qquad W_i^2=F_i, \] a degree-four extension. With $W_0=1$ and $W_3=W_1W_2/G$, the four $W_\chi$ form a basis. At the given integer point use positive square roots. Then all $W_\chi(x,y)$ are integers: $F_3(x,y)$ is an integer which is a rational square, hence an integer square, and $W_3(x,y)=\sqrt{F_1(x,y)}\sqrt{F_2(x,y)}/G(x,y)$. \emph{The linear system at infinity.} Put $r_0=s_0=0$, $k=3q$, and consider \[ \mathcal A=\sum_{\chi=0}^3A_\chi(X,Y)W_\chi, \qquad \deg_XA_\chi\le k-r_\chi, \quad \deg_YA_\chi\le k-s_\chi. \] There are \[ n=\sum_{\chi=0}^3(k-r_\chi+1)(k-s_\chi+1) \ge21q^2+18q+4 \] unknown coefficients, and $n\le4(3q+1)^2$. Writing $\xi=1/X$, $\eta=1/Y$, the positive formal branches are \[ W_\chi=X^{r_\chi}Y^{s_\chi} \prod_{u\in I_\chi}(1+u\xi)^{1/2} \prod_{v\in J_\chi}(1+v\eta)^{1/2}. \] Thus $f(\xi,\eta)=\xi^k\eta^k\mathcal A$ is a formal power series. Require its coefficients of total degree at most $2k$ to vanish. This gives at most \[ t=\frac{(2k+1)(2k+2)}2=18q^2+9q+1, \qquad n-t\ge3q^2+9q+3,\qquad \frac t{n-t}<6 \] homogeneous linear equations. All coefficients required from the binomial series have denominator dividing $4^{2k}$. Cauchy's estimate on circles of radius $(2B)^{-1}$ bounds their numerators, after this common denominator is cleared, by $(B+1)^{c_0q}$ for an absolute $c_0$. An elementary pigeonhole form of Siegel's lemma yields a nonzero integer solution of height $(B+1)^{c_1q}$. Indeed, if the integer matrix has height $A$, the images of $\{0,\ldots,M\}^n$ take at most $(2nAM+1)^t$ values. Choosing $M=\lceil(3nA)^7\rceil$ makes two images equal, because $t/(n-t)<6$. Their difference is a nonzero kernel vector; $\log n=O(\log B)$ is absorbed in the stated height. \emph{Primitive coefficients.} Let $D$ be the primitive polynomial gcd of the four $A_\chi$, omitting zero coefficients, and write $A_\chi=DA'_\chi$. The factor bound of Lemma~\ref{p13:rig:factor-height} keeps $H(A'_\chi)\le(B+1)^{O(q)}$. Write $a=\deg_XD$, $b=\deg_YD$, $e=\deg D$. Then \[ f=\widetilde D f',\qquad \widetilde D=\xi^a\eta^bD(\xi^{-1},\eta^{-1}),\qquad f'=\xi^{k-a}\eta^{k-b}\sum_\chi A'_\chi W_\chi. \] These are power series, and $\operatorname{ord}\widetilde D=a+b-e$ for total order. Since order is additive in the power-series domain, \begin{equation} \operatorname{ord}f'\ge2k+1-a-b+e =(k-a)+(k-b)+e+1. \label{p13:rig:primitive-order} \end{equation} Thus removing $D$ preserves vanishing beyond the remaining pole weight. This is a formal division, never a division by the possibly zero value $D(x,y)$. Replace the auxiliary by its primitive version from now on. \emph{Integer separation and the norm.} On the bidisc $|\xi|,|\eta|\le(4B)^{-1}$, the analytic numerator is bounded by $(B+1)^{O(q)}$. Cauchy's estimate and the order of vanishing give, in the comparable range $x,y\in[N,2N)$, \begin{equation} |\mathcal A(x,y)|\le \frac{(B+1)^{c_2q}}N. \label{p13:rig:integer-gap} \end{equation} To see the power of $N$ explicitly before removing $D$, the tail of the numerator is bounded by its coefficient bound times \[ \sum_{\ell\ge2k+1}(\ell+1)(4B/N)^\ell \ll(k+1)(4B/N)^{2k+1}. \] Multiplication by $x^ky^k\le(2N)^{2k}$ gives \eqref{p13:rig:integer-gap}. After removing $D$, both pole weights decrease, and \eqref{p13:rig:primitive-order} gives the same estimate with the new height bound. The specialization $\mathcal A(x,y)$ is an integer, so it vanishes for the stated threshold. The auxiliary is nonzero in $K$ by independence of the four characters. Its norm is therefore nonzero. An explicit polynomial expression is \[ \begin{split} U&=A_0^2+A_1^2F_1-A_2^2F_2-A_3^2F_3,\\ V&=A_0A_1-A_2A_3(F_2/G),\\ R&=U^2-4F_1V^2. \end{split} \] Here $F_2/G$ is a polynomial. The partial degrees of $U$ are at most $2k$, and those of $V$ are at most $2k-r_1$, $2k-s_1$. For the second term in $V$, use $r_3=r_1+r_2-\deg_XG$ and its $Y$ analogue. It follows that the partial degrees of $R$ are at most $4k=12q$. The coefficient $\ell^1$ norms of all factors in this formula are $(B+1)^{O(q)}$, which proves the height bound and completes \eqref{p13:rig:two-polynomial}. \end{proof} \subsubsection{Proof of Theorem~\ref{p13:rig:common-curve}} \label{p14:comp:rigidity:common-curve} \begin{proof} For $i0}} |c(e)| \sum_{\substack{z\in\mathcal I_N\setminus e\\ S_z\cap(S_x+S_y)\ne\{0\}}}p_z. \] Then \(D_{2,1}\ll_{C,\varepsilon}N^{-3/2+\varepsilon}\). \end{proposition} The proof in Section~\ref{p14:comp:activity:deficit} partitions pairs into finitely many fixed rational channels, weak pairs, and an exceptional class. A fixed-source half-count handles the last two classes. For a good channel, exact compression turns the pair and its target into two logarithmic blocks; their total host count is $N^{1+\varepsilon}$ and their individual weights give an even smaller $N^{-5/3+\varepsilon}$ contribution. This bound uses independent weight only for the new target, while preserving the pair activity. \begin{theorem}[Absolute activity of incompatible triples] \label{p13:activity:incompatible} For every \(\varepsilon>0\), \[ \sum_{\substack{T\subseteq\mathcal I_N,\ |T|=3\\q(T)=0}} |c(T)|\ll_{C,\varepsilon}N^{-3/2+\varepsilon}. \] Both signs of the centered coefficient are included. \end{theorem} \begin{proof} A decomposable triple has a singleton component, hence centered coefficient zero. For a connected incompatible triple, expansion of its product gives exactly \[ c(T)=-\prod_{i\in T}p_i -\sum_{\substack{e\subset T\\|e|=2}} c(e)\prod_{i\in T\setminus e}p_i. \] The first terms have total magnitude \(O(N^{-3/2+\varepsilon'})\) by Corollary~\ref{p13:half:connected}. A direct pair has \(c(e)=0\); for any other pair in a connected triple the remaining space meets its sum nontrivially. The pair terms are therefore bounded by three copies of \(D_{2,1}\). Proposition~\ref{p13:activity:deficit} concludes. \end{proof} \begin{proposition}[The third cumulant and the positive activity] \label{p13:activity:kappa3} Every compatible triple has \(c(T)\ge0\), and \begin{equation} \kappa_{3,N}=6K_3^++O_{C,\varepsilon}(N^{-3/2+\varepsilon}), \qquad A_3=K_3^++O_{C,\varepsilon}(N^{-3/2+\varepsilon}). \label{p13:activity:kappa-comparison} \end{equation} In the second equality the difference \(A_3-K_3^+\) is nonnegative. \end{proposition} \begin{proof} Order the pair ranks of a compatible triple as \(a\ge b\ge c\). Their relation spaces show \(\rho_{\mathrm{rel}}(T)\ge a+b\), so \[ \frac{c(T)}{\prod_{x\in T}p_x} =2^{\rho_{\mathrm{rel}}(T)}-2^a-2^b-2^c+2\ge0. \] If \(b=0\), this is immediate. Otherwise use \((2^a-1)(2^b-1)\ge2^c-1\). The same formula includes the primitive case \(a=b=c=0\). Decomposable triples again have zero coefficient. In \eqref{p13:exact:kappa3}, the repeated-index terms are bounded by \(O(p_*A_2+p_*^2\mu_N) =O(N^{-5/3+\varepsilon'})\). Apply Theorem~\ref{p13:activity:incompatible} to the other triples. This proves both formulas. \end{proof} For example, three mutually exclusive events have \(c(T)=2p_xp_yp_z>0\) despite being incompatible. Thus it is the absolute estimate for incompatible triples, rather than an assumption on their sign, that permits the reduction to \(K_3^+\). \subsection{Pair partitions and the cost of one dependent target} \label{p14:comp:activity} The signed triple identities in Section~\ref{p13:activity:section} use one quantitative input: the independent target deficit following a pair. We first prove the finite partition needed for that estimate. \begin{lemma}[A finite partition of the pair activity] \label{p13:activity:pair-partition} For each fixed \(\eta>0\), dependent pairs can be partitioned into good pairs from a finite fixed set of rational channels, a weak class, and an exceptional class, with \[ A_{\rm w}\ll_{C,\eta,\varepsilon}N^{-1+\eta+\varepsilon}, \qquad A_{\rm s}\ll_{C,\eta,\varepsilon}N^{-1+\varepsilon}. \] A good pair in a channel \(a:b\), \(b\ge3\), has \[ |c(e)|\ll_{a,b,C}N^{-(2-1/b)+o(1)} \] and an exact compression into a block of fixed logarithmic width. Moreover \(A_2\ll_{C,\varepsilon}N^{-2/3+\varepsilon}\). \end{lemma} \begin{proof} Choose \(Q\) with \(1/(Q+1)<\eta/2\), and retain the channels \(3\le b\le Q\), \(b/20\\ \text{some pair of }T\text{ is aligned in }\mathcal F}}q(T) \ll_{\mathcal F,C,\varepsilon}N^{-41/30+\varepsilon}. \label{p13:third:aligned-family-bound} \end{equation} The sum allows irregular admissible windows, bad normalization of the source pair, and contacts at distance exactly $L$. \end{proposition} The complete normalization argument is in Section~\ref{p14:comp:third:aligned-family}. After exact merging, the residual kernel has dimension $o(L)$ uniformly even when the last rational coefficients have polynomial height in $B$. Distinct physical slopes then lose at most $19L/30+O(1)$ dimensions through canonical coincidences, so each triple has probability at most $N^{-71/30+o(1)}$. Their $NB^{O(1)}$ hosts give $N^{-41/30+\varepsilon}$. The proof separately controls nonaligned targets, repeated physical slopes, and dyadic endpoint populations; none of these cases requires good normalization of the source pair. \begin{corollary}[Strong aligned pairs] \label{p13:third:aligned} For every fixed $\theta>0$ and every $\varepsilon>0$, \begin{equation} \sum_{\substack{|T|=3,\ T\ {\rm connected},\ q(T)>0\\ \exists\{x,y\}\subset T\ {\rm aligned}:\ d_{xy}\geq\theta B}} q(T)\ll_{C,\theta,\varepsilon}N^{-41/30+\varepsilon}. \label{p13:third:strong-aligned-bound} \end{equation} \end{corollary} \begin{proof} The uniform bound $d_{xy}\leq L/b+o(L)$ forces $b\leq2/\theta$ for sufficiently large $N$. Thus only a finite collection of heights is relevant. Heights at least three satisfy Proposition~\ref{p13:third:aligned-family}. Height two forces the source pair to the dyadic endpoints: there are $B^{O(1)}$ pairs of weight at most $N^{-3/2+o(1)}$. Proposition~\ref{p1:fixed:propagation} adds a third window with a loss $N^{\varepsilon'}$, which is harmless. At height one, compatible distinct windows have disjoint interiors and at most one exact unit, so their relative intersection is $o(L)$. They cannot reach the threshold $\theta B$. \end{proof} \subsection{A positive family that survives in the total cumulant} The equality case $(10,12,15)$ of the overlap inequality also supplies a population of actual compatible run starts. Its role is stronger than merely maximizing a geometric formula: its centered activity is positive, and there are linearly many admissible translations. \begin{proposition}[The rational lower bound] \label{p13:third:lower} For all sufficiently large $N$ in the critical band, \begin{equation} K_3^+:=\sum_{\substack{|T|=3,\ T\ {\rm connected}\\q(T)>0}}c(T) \gg_C N^{-41/30}. \label{p13:third:K3-lower} \end{equation} In particular $\kappa_{3,N}\geq c_C N^{-41/30}$ for a suitable $c_C>0$. \end{proposition} \begin{proof} Take integers $t$ with $N\leq10t<12t<15t<2N$. There are $N/30+O(1)$ choices. Multiplication of the three windows by $6,5,4$ respectively gives the true normalized sets \[ U_1=60t+6\{-1,0,\ldots,L-1\},\quad U_2=60t+5\{-1,0,\ldots,L-1\},\quad U_3=60t+4\{-1,0,\ldots,L-1\}. \] Choose $H=10B$, larger than the diameter of their union. By \eqref{p1:arith:onedfect}, only $N^{1/2+o(1)}$ integers in the relevant range have $K_H(n)=1$. For each of the $O(B)$ possible normalized offsets there is at most one parameter $t$ per such integer. Hence, outside $N^{1/2+o(1)}$ choices of $t$, every distinct value in $U_1\cup U_2\cup U_3$ has a private prime above $H$. Its square-class map is then injective. The three sets meet at $60t$. The sum of their even-selection spaces is consequently the even-selection space of their union. The pairwise intersection sizes are \[ 1+\left\lfloor\frac{L-1}{6}\right\rfloor,\qquad 1+\left\lfloor\frac{L-1}{3}\right\rfloor,\qquad 1+\left\lfloor\frac{L-1}{5}\right\rfloor, \] and the triple intersection has size $1+\lfloor(L-1)/15\rfloor$. Inclusion--exclusion therefore gives the exact relative nullity \begin{equation} \rho_L=\left\lfloor\frac{L-1}{6}\right\rfloor +\left\lfloor\frac{L-1}{3}\right\rfloor +\left\lfloor\frac{L-1}{5}\right\rfloor -\left\lfloor\frac{L-1}{15}\right\rfloor =\frac{19}{30}L+O(1). \label{p13:third:witness-nullity} \end{equation} The normalized predecessors have offsets $-6,-5,-4$, each belonging to just one set. Assign sign $-1$ to a private prime of each of these three values and sign $+1$ to every other prime occurring in the windows and multipliers. The chosen primes are distinct and occur in no other normalized value. Each original predecessor then has sign $-1$ and every internal value has sign $+1$. This verifies affine compatibility explicitly. For all these parameters, \begin{equation} q(\{10t,12t,15t\})=p^3 2^{\rho_L}. \label{p13:third:witness-probability} \end{equation} The three pair nullities are the first three floor terms in \eqref{p13:third:witness-nullity}. Since all marginals are $p$ on this private-prime population, the centered identity is exact: \begin{equation} \begin{split} c(\{10t,12t,15t\})=p^3\biggl\{2^{\rho_L} &-2^{\lfloor(L-1)/6\rfloor}-2^{\lfloor(L-1)/3\rfloor}\\ &-2^{\lfloor(L-1)/5\rfloor}+2\biggr\}. \end{split} \label{p13:third:witness-centered} \end{equation} The largest subtracted exponent is $L/3+O(1)$, whereas the first exponent is $19L/30+O(1)$. Thus $c(T)\geq q(T)/2>0$ for large $N$. This also ensures connectedness: a direct singleton split would make the centered coefficient vanish. All compatible connected triple coefficients are nonnegative, and the triples obtained from distinct $t$ are different. Consequently \[ K_3^+\gg Np^3 2^{19L/30} \gg_C N^{-41/30}. \] Finally \eqref{p13:activity:kappa-comparison} gives $\kappa_{3,N}=6K_3^++O(N^{-3/2+\varepsilon'})$. Choose $\varepsilon'<3/2-41/30=2/15$ to make the error negligible. \end{proof} When $N=2^L$ and $L=30m$, the contribution of the private-prime population in this proof to $K_3^+$ is $(1/120+o(1))N^{-41/30}$, since $\rho_L=19m-2$. This coefficient belongs to the displayed subfamily; it is not a coefficient for the full cumulant. \subsection{Partitioning every compatible triple} \begin{proof}[Proof of Theorem~\ref{p13:third:main}] Proposition~\ref{p13:third:lower} supplies the lower bound. For the upper bound, fix $\eta=1/30$ and then a positive $\theta<\eta/200$ for which Proposition~\ref{p13:conf:near-primitive} applies. All these parameters are fixed before $N$. We bound $K_3^+$ by partitioning its compatible connected triples. If $\rho_{\mathrm{rel}}(T)<\eta B$, the host bound of Corollary~\ref{p13:half:connected} and $p_x\leq2p$ give \[ \sum q(T)\ll N^{3/2+\varepsilon'}N^{-3+\eta+o(1)} =N^{-22/15+\varepsilon'+o(1)}. \] If $\rho_{\mathrm{rel}}(T)\geq\eta B$ and every proper pair has $d_{xy}<\theta B$, Proposition~\ref{p13:conf:near-primitive} gives total homogeneous mass $O(N^{-2+\theta+\varepsilon'})$, which bounds the compatible mass as well. It remains to consider a pair with $d_{xy}\geq\theta B$. If any such pair is aligned, Corollary~\ref{p13:third:aligned} gives $O(N^{-41/30+\varepsilon'})$. Otherwise choose a strong nonaligned pair. Theorem~\ref{p13:rig:nonaligned} gives \[ d_{xy}\leq\left(\frac23\log2+o(1)\right)L. \] For the third space, \[ \rho_{\mathrm{rel}}(T) =d_{xy}+\dim\bigl(S_z\cap(S_x+S_y)\bigr) \leq d_{xy}+L \leq\left(1+\frac23\log2+o(1)\right)L. \] Since $(2/3)\log2<1/2$, this is eventually at most $3L/2$. Proposition~\ref{p13:conf:strong-profile}, with $\sigma=3/2$, therefore bounds this sector by $O(N^{-3/2+\varepsilon'})$. This is where the unconditional nonaligned-pair rigidity enters the third-cumulant theorem. Each exponent other than $41/30$ is strictly larger than $41/30$. Shrink the internal error margins to obtain $K_3^+\ll N^{-41/30+\varepsilon}$. The identity \eqref{p13:activity:kappa-comparison} transfers the estimate to $\kappa_{3,N}$, completing the upper bound on the full critical band. \end{proof} The gap $41/30-4/3=1/30$ explains why third-order activity is below the optimal quadratic remainder scale. This conclusion does not identify the optimal error for a reference that also retains cubic or quartic terms; such a reference requires its own reconstruction and lower-bound argument. The rank threshold $\eta=1/30$ and the weak-pair threshold $\theta<\eta/200$ are distinct from the accuracy parameters. Using the convention in Section~\ref{p1:intro:dependencies}, prescribe $\varepsilon_{\mathrm{out}}>0$ and choose $0<\tau<\min(\varepsilon_{\mathrm{out}}/4,1/960)$. Allocate smaller internal losses so that each completed sector loses at most $\tau$. The nonaligned target above a rational pair retains $1/120-\tau>0$, and the incompatible-triple error retains $2/15-\tau>0$. These budgets include offset choices and residual-rank losses before $N$ is taken large. The latter error is therefore negligible in the positive rational lower bound as well. \subsection{All triples above finitely many rational pair channels} \label{p14:comp:third} The canonical overlap inequality and the positive $10:12:15$ witness are proved in Section~\ref{p13:third:section}. The remaining technical input is uniformity of the upper bound across every triple above a fixed finite set of aligned channels, including arithmetically exceptional source pairs. We prove that input here. \subsubsection{Proof of Proposition~\ref{p13:third:aligned-family}} \label{p14:comp:third:aligned-family} \begin{proof} Fix one channel $a:b$ and orient the source pair $e=\{x,y\}$. Choosing $B^{O_{a,b}(1)}$ offsets, write $x=at-i_0$, $y=bt-j_0$ and $X=abt$. Merge repeated values in $bV_x\cup aV_y$ to obtain $U_X\subset X+[-D B,D B]$, where $D$ depends only on the fixed channel. Use radius $DB$ for alignment with this compressed source, and choose the large-prime cutoff above $(2D+1)B$. The old centre lies in a fixed expanded interval starting at $N$, while the new centre remains in $I_N$, as required by the expanded version of Lemma~\ref{p1:arith:asymmetric}. Multiplying an even selection by a constant does not change its square class, so \begin{equation} S_x+S_y\subseteq \operatorname{span}\{v(n):n\in U_X\}. \label{p13:third:true-compression} \end{equation} This inclusion holds regardless of injectivity on $U_X$. The uniform aligned rank bound of Lemma~\ref{p1:arith:alignedrank} gives \begin{equation} q(e)\leq N^{-\kappa_b+o(1)},\qquad \kappa_b=2-\frac1b\geq\frac53. \label{p13:third:source-weight} \end{equation} For fixed height the residual defect is uniform, so the same bound includes admissible local rank defects. \paragraph{Targets not aligned with the compressed source.} For $z\notin e$, put $\delta=\dim(S_z\cap(S_x+S_y))$. Connectedness of $T=e\cup\{z\}$ implies $\delta>0$. The inclusion \eqref{p13:third:true-compression} bounds $\delta$ by the asymmetric rank loss $\delta^*(X,z)$ of Lemma~\ref{p1:arith:asymmetric}. That lemma permits an arbitrary true set $U_X$ of this fixed logarithmic width; it does not require injectivity. It gives \[ \sum_{X,z\ {\rm nonaligned}}(2^{\delta^*(X,z)}-1) \ll N^{31/24+\varepsilon'}. \] Since $q(T)=q(e)p_z2^\delta$, $p_z\leq2p$, and $2^\delta\leq2(2^{\delta^*}-1)$, these targets have total mass at most \begin{equation} N^{-\kappa_b-1+31/24+\varepsilon'+o(1)} \ll N^{-11/8+\varepsilon}. \label{p13:third:unaligned-target-cost} \end{equation} The number of source-offset patterns is polynomial in $B$ and is absorbed into the exponent margin. \paragraph{The residual kernel for an aligned target.} This sector includes every target aligned at the same enlarged radius $DB$. Enlarging that fixed radius changes only the constants in the polynomial offset and height bounds below. Write an aligned target as $z=(ct+h)/d$, with $(c,d)=1$ and $c,d,|h|\leq B^{O_{a,b}(1)}$, as in the proof of Lemma~\ref{p1:arith:fixed-channel-rank}. Before this final normalization, $U_X$ and $V_z$ each have fixed logarithmic width. Choose $H>C_D B$ larger than both diameters and remove the prime rows at most $H$ and those dividing the normalization multipliers. Each remaining prime row meets at most one distinct old value and at most one new value. This uses the fact that repeated old values were merged before the target was added. After the common normalization, merge equal values once more. There are at most $3B$ distinct integers, all at height at least a constant multiple of $N$ and in an interval of diameter $B^{O_{a,b}(1)}$. The remaining prime rows have weight at most two, and leave one variable per unpinned component. We give the uniform residual estimate, since the coefficients $c,d$ need not be fixed. For a fixed integer $K$, restrict to components of at most $K$ vertices. The removed prime rows and a total-parity row form a code on these components of redundancy \[ s\leq\pi(H)+O_{a,b}(\log B)+1=O_{a,b}(B/\log B). \] A nonzero word in its kernel gives an even square product of distinct normalized integers. Lemma~\ref{p1:arith:runge}, applied with the polynomial-in-$B$ diameter just specified, forces it to contain $\gg_{a,b}\log N/\log B$ physical vertices and hence $\gg_{a,b,K}\log N/\log B$ components. If $c_K$ is the number of these components, the Hamming inequality with $t\asymp_{a,b,K}\log N/\log B$ gives \[ \sum_{j=0}^{t}\binom{c_K}{j}\leq2^s, \qquad c_K=O_{a,b,K}(B/\log B). \] Here $t$ is chosen below half the component minimum distance. Components larger than $K$ number at most $3B/K$. Choosing $K$ arbitrarily large before $N$ therefore bounds the entire value kernel on the distinct normalized set by $o_{a,b}(L)$, uniformly over $c,d,h$. The relevant map factors as \[ \bigoplus_{v\in T}E_v \longrightarrow\{\text{merged normalized coordinates}\} \longrightarrow\{\text{square classes}\}. \] Its nullity is at most the nullity of merging plus the residual $o(L)$ just proved. Internal defects of the old windows are included in this composition and do not require any assumption of good normalization. \paragraph{Counting the exact coincidences.} In the common parameter the physical coefficients are $ad,bd,c$. Suppose first that they are distinct. Divide by their common gcd and order them as $u0$, uniformly in $G$ and in the stated range of $H$, for all sufficiently large $N$, \begin{equation} \begin{split} &\beta_{a,b}>0,\qquad M_{\mathrm{ex}}(G)+M_{\mathrm{nom}}(G) \ll_{a,b,C,D,\varepsilon}N^{-41/30+\varepsilon},\\ &\frac{M_{\mathrm{ex}}(G)+M_{\mathrm{nom}}(G)}{\beta_{a,b}} \ll_{a,b,C,D,\varepsilon}N^{-11/30-1/b+\varepsilon}. \end{split} \label{p2:palm:all-targets-bound} \end{equation} No regularity condition is imposed on the target windows. \end{proposition} \begin{proof} For $z\notin e$, put $d_e(z)=\dim(S_z\cap\sum_{x\in e}S_x)$. The finite affine rank bound gives \[ q(e\cup\{z\})\le q_ep_z\,2^{d_e(z)}. \] The inequality includes incompatible joint prescriptions and zero marginals. Thus $p_z=0$ forces $q(e\cup\{z\})=0$, and $d_e(z)=0$ makes the exact positive excess vanish. A positive summand of $M_{\mathrm{ex}}$ therefore has $d_e(z)>0$ and $q(e\cup\{z\})>0$. Compatibility forces all three pairwise distances to be at least $L$, including possible equality. The source is a dependent pair, so neither source singleton can split off as a direct component. Positive crossing prevents the target singleton from splitting off. The triple is connected. Each summand is at most $q(e\cup\{z\})$, and a triple has at most three choices of a marked source pair. Proposition~\ref{p13:third:aligned-family}, which includes irregular windows and exact contacts, gives \[ M_{\mathrm{ex}}(G) \le3\sum_{\substack{|T|=3,\ T\ \mathrm{connected},\ q(T)>0\\ T\text{ contains a pair in the fixed channel}}}q(T) \ll N^{-41/30+\eta} \] for every fixed $\eta>0$. All conditioned sources have $q_e>0$. The positive-part inequality gives \[ \bigl(q(e\cup\{z\})-q_ep\bigr)_+ \le\bigl(q(e\cup\{z\})-q_ep_z\bigr)_+ +q_e(p_z-p)_+. \] Using the marginal estimate~\eqref{p1:arith:marginalsum}, \[ M_{\mathrm{nom}}(G) \le M_{\mathrm{ex}}(G)+\beta_{a,b}\sum_{z\in I_N}(p_z-p)_+ \ll N^{-41/30+\eta}+\beta_{a,b}N^{-1+\eta}. \] Lemma~\ref{p1:arith:fixed-channel-mass} gives, in the stated cutoff range, $\beta_{a,b}\asymp_{a,b,C,D}LN^{-1+1/b}>0$. Each unordered compatible edge has two ordered orientations, so its mass is half the ordered mass in that lemma. Monotonicity of the private-prime condition sandwiches the mass between the phases at $H=6bB$ and $H=DB$, proving uniformity throughout the stated range. Since $b\ge3$, the second raw term is $O(LN^{-2+1/b+\eta})$, smaller than the $41/30$ scale. Division by the positive phase bounds the normalized sum by a constant times \[ L^{-1}N^{-11/30-1/b+\eta}+N^{-1+\eta}. \] The second exponent is smaller because $11/30+1/b\le7/10<1$. Choose the internal margin within $\varepsilon$ and enlarge the scale threshold to absorb fixed factors. All parameters precede $N,L,H,G$. \end{proof} For the two principal channels, one common choice $24B\le H\le DB$ with fixed $D\ge24$ gives \begin{align} \frac{M_{\mathrm{nom},2:3}(G)}{\beta_{2,3}} &\ll_{C,D,\varepsilon}N^{-7/10+\varepsilon}, \label{p1:arith:palm23}\\ \frac{M_{\mathrm{nom},3:4}(G)}{\beta_{3,4}} &\ll_{C,D,\varepsilon}N^{-37/60+\varepsilon}. \label{p1:arith:palm34} \end{align} These are bounds for positive increases in the target probabilities; no optimality or absolute-discrepancy statement is asserted. \section{Uniform certificate tools} \label{p14:cert:section} Fixed-order estimates leave a separate problem: summing orders that grow with $N$. The arithmetic certificates below solve it by counting congruences and graph components with explicit cardinality weights. Their definitions concern the actual values in the windows and apply before any affine prescription is imposed. In particular, the geometric counts also permit irregular windows; replacing $p$ by $p_x\le2p$ only changes a fixed weight. The final parameter choices for the optimal reconstruction are made in Section~\ref{p13:transfer:section}. The complete elementary proofs are collected in Section~\ref{p14:comp:certificate-tools}. \subsection{Small kernels and moderate congruence certificates} For $Y\ge B$, write \[ K_Y(v)=\prod_{\substack{\ell>Y\ \mathrm{prime}\\v_\ell(v)\ \mathrm{odd}}}\ell. \] The following elementary form of the Rankin argument will be used with its exponent fixed. \begin{lemma}[Counting a small truncated squarefree kernel] \label{p1:core:rankin} For fixed $1/2<\sigma<1$, fixed $C_1>0$, and $T\ge1$, \begin{equation} \#\{v\le C_1N:K_Y(v)\le T\} \ll_{\sigma,C_1}N^\sigma T^{1-\sigma} \exp(C_\sigma Y^{1-\sigma}). \label{p1:core:rankin-bound} \end{equation} \end{lemma} \noindent The complete proof is given in Section~\ref{p14:comp:certificate-rankin}. Fix \begin{equation} A_0>1,\quad 0<\beta<1,\quad \tfrac12<\sigma<1, \quad A_0(1-\sigma)<1, \label{p1:core:parameter-conditions} \end{equation} and put $H=\log N$, $Y=H^{A_0}$, $T=N^\beta$ and $\delta=(1-\sigma)(1-\beta)>0$. Mark a window if one of its values has $K_Y(v)\le T$. Each value lies in at most $B$ windows, so the number $M$ of marked centres satisfies \begin{equation} M\le N^{1-\delta+o(1)}. \label{p1:core:marks} \end{equation} Here and below the $o(1)$ depends on the fixed parameters, but not on the size of a family subsequently summed over. Choose two distinct values in an unmarked anchor window. Their kernels are coprime, because a prime exceeding $Y\ge B$ cannot divide their nonzero difference. Thus their product $Q$ is squarefree and $T^20$, \begin{equation} \sum_{r\ge1}\frac{(up)^r}{r!}|\mathcal U_r| \le N^{-2\beta+4/A_0+o(1)}. \label{p1:core:moderate-baseline} \end{equation} \end{lemma} \noindent The complete proof is given in Section~\ref{p14:comp:certificate-moderate}. \subsection{Saturation has small degree in both directions} Put a directed edge $x\to y$ when two values in window $x$ have a kernel product possessing a divisor $D>6NB$ all of whose primes divide values in window $y$. We impose no direct-sum condition on $S_x,S_y$. Let $\Gamma$ be the underlying simple undirected graph. \begin{lemma}[Incoming and outgoing saturation degrees] \label{p1:core:degrees} Uniformly in the fixed vertex, \begin{equation} \deg^+(x),\deg^-(x)\le N^{2/A_0+o(1)}, \qquad \Delta(\Gamma)\le N^{2/A_0+o(1)}. \label{p1:core:degree-bound} \end{equation} \end{lemma} \noindent The complete proof is given in Section~\ref{p14:comp:certificate-degrees}. If a primitive family has no moderate certificate, every unmarked vertex has a neighbour in its induced saturation graph. To see this, choose a nonzero relation, select two of its values at that vertex and assign the primes to exterior occurrences. Some receiver must have $D_j>6NB$. The choices made at different vertices need not agree. Their existence in one deterministic graph is sufficient. \begin{lemma}[A generating function with marked isolated vertices] \label{p1:core:generator} Let a simple graph have $n$ vertices, maximum degree $\Delta$ and $M$ marked vertices. If $\mathcal F$ consists of the vertex sets whose induced isolated vertices are all marked, then, for $z\ge0$, \begin{equation} \sum_{D\in\mathcal F}z^{|D|} \le\exp\left\{Mz+nz\bigl((1+z)^\Delta-1\bigr)\right\}. \label{p1:core:generator-bound} \end{equation} More generally, for nonnegative vertex weights $w_v$ and marked set $\mathcal M$, \[ \sum_{D\in\mathcal F}\prod_{v\in D}w_v \le \exp\left\{\sum_{v\in\mathcal M}w_v+ \sum_vw_v\left(\prod_{u\sim v}(1+w_u)-1\right)\right\}. \] \end{lemma} \noindent The complete proof is given in Section~\ref{p14:comp:certificate-generator}. \subsection{Proofs of the uniform certificate tools} \label{p14:comp:certificate-tools} All definitions and parameter conditions are those of Section~\ref{p14:cert:section}. The counts are geometric, and all constants are uniform in the size of the family being summed. \subsubsection{Proof of Lemma~\ref{p1:core:rankin}} \label{p14:comp:certificate-rankin} \begin{proof} Write uniquely $v=k s a^2$, where $k=K_Y(v)$ and $s$ is the squarefree product of the primes at most $Y$ occurring to odd order. The inequality $1\le (C_1N/(ksa^2))^\sigma$ on the range being counted gives an upper bound \[ (C_1N)^\sigma \left(\sum_{k\le T}k^{-\sigma}\right) \left(\sum_{a\ge1}a^{-2\sigma}\right) \prod_{\ell\le Y}(1+\ell^{-\sigma}). \] Enlarging the sum over $k$ to all positive integers only increases it. The three factors are respectively $O_\sigma(T^{1-\sigma})$, $\zeta(2\sigma)$, and at most $\exp(\sum_{j\le Y}j^{-\sigma})$. This proves the bound. \end{proof} \subsubsection{Proof of Lemma~\ref{p1:core:moderate}} \label{p14:comp:certificate-moderate} \begin{proof} There are at most $rNB^2$ choices for the anchor and its two offsets. The product $Q$ has at most $m$ prime factors, and each has at most $rB$ destinations. Once these assignments are fixed, the Chinese remainder theorem leaves at most $1+N/D_j$ possibilities for centre $j$. At most two of the $D_j$ exceed $N$, since $Q\le9N^2$. For those two modules, replacing $1+N/D_j$ by $2N/D_j$ costs at most $(6B)^2$. Therefore \[ \prod_{j\ne\mathrm{anchor}}(1+N/D_j) \le 2^{r-1}(6B)^2N^{r-1}T^{-2}. \] This proves \eqref{p1:core:moderate-count}; allowing collisions and incompatible assignments only enlarges the count. After multiplication by $(up)^r/r!$, the remaining series is, up to a fixed factor, \[ N^{-2\beta}B^{m+4}\sum_{r\ge0}\frac{z^r r^{m+1}}{r!}, \qquad z=2uNp=O_u(1). \] Since $r^{m+1}\le(m+1)!e^r$, this series is at most $N^{-2\beta}B^{m+4}(m+1)!\exp(ez)$. Finally \[ m=\left(\frac2{A_0}+o(1)\right)\frac{H}{\log H},\qquad \log(B^m(m+1)!)=\left(\frac4{A_0}+o(1)\right)H. \] This yields \eqref{p1:core:moderate-baseline}. The factorial here divides the orderings of a set of distinct windows; it is not an assumption on the growth of a fixed-order arithmetic constant. \end{proof} \subsubsection{Proof of Lemma~\ref{p1:core:degrees}} \label{p14:comp:certificate-degrees} \begin{proof} For an outgoing edge, choose two anchor offsets in $B^2$ ways. Each of the at most $m$ primes is either unused or assigned to one of $B$ offsets in the receiver. A qualifying code has modulus $D>N$ and therefore determines at most one receiving centre in $I_N$. This gives $B^2(B+1)^m=N^{2/A_0+o(1)}$ possibilities. For the incoming degree, fix the receiving window $y$ and let $\Omega_y$ be the primes greater than $Y$ dividing any of its values. Writing $R=\lceil B\log(3N)/\log Y\rceil$, we have $|\Omega_y|\le R$. A code chooses two anchor offsets, at most $m$ primes from $\Omega_y$, and one of the two offsets for each chosen prime. For a qualifying code their product exceeds $N$, so its congruences determine at most one anchor centre. Thus the number of possible incoming neighbours is at most \[ B^2\sum_{j=0}^m\binom Rj2^j \le B^2(m+1)(2eR/m)^m =N^{2/A_0+o(1)}. \] The displayed binomial bound holds for large $N$, since $m\le R/2$. Nonqualifying codes may be included in the number of codes; the one-centre conclusion is applied only to qualifying ones. No receiving offset needs to be chosen in this incoming count: the receiver has already been fixed. Symmetrisation adds the two degree bounds. \end{proof} \subsubsection{Proof of Lemma~\ref{p1:core:generator}} \label{p14:comp:certificate-generator} \begin{proof} Fix an admissible set $D$. Its isolated vertices are marked singletons. On the other vertices choose an inclusion-minimal edge cover of the induced graph. Every edge of this cover has an endpoint of degree one in the cover, since otherwise that edge could be removed. Its components are therefore nontrivial stars. Thus $D$ has a partition into marked singletons and stars, each with one centre and a nonempty subset of its neighbours as leaves. Write $w(A)=\prod_{v\in A}w_v$. Choose one such partition for each $D$. The product over all possible blocks of $1+w(A)$ contains a distinct term of weight $w(D)$ for this chosen partition: its union recovers $D$. Dropping disjointness only adds nonnegative terms. Since $1+u\le e^u$ for $u\ge0$, the weighted sum is at most the exponential of the total block weight. The singleton contribution is $\sum_{v\in\mathcal M}w_v$. For a fixed star centre $v$, its leaf sum is exactly \[ w_v\sum_{\varnothing\ne A\subseteq N(v)}\prod_{u\in A}w_u =w_v\left(\prod_{u\sim v}(1+w_u)-1\right). \] Counting the same two-vertex star from both centres is harmless. This proves the heterogeneous assertion. Set $w_v=z$ and use $\deg(v)\le\Delta$ to obtain \eqref{p1:core:generator-bound}. There is no small-activity hypothesis. When $\Delta z<2$, the elementary bound $e^u-1\le2u/(2-u)$ also gives the convenient majorant $\exp\{Mz+2n\Delta z^2/(2-\Delta z)\}$. \end{proof} \section{Arithmetic estimates for the exact reconstruction} \label{p2:technical-transfer} We use the notation of Section~\ref{p13:transfer:section} of the article. All original windows with $p_x>0$ remain present. In this section, $\rho$ means the effective relative nullity $\rho_{\mathrm{rel}}$. The estimates culminate in separate bounds for each fixed connected order and for the entire remaining range of cardinalities. The article then assembles them using the positive stopped expansion. \subsection{Two dense targets above a prescribed space} The independent counts do not control a large rank gain: a family of nullity $d$ carries an additional factor $2^d$. We now isolate the conditional geometric estimate that pays for this factor. Only two centers remain free. Every prime circuit meeting at least three physical blocks must then meet one of the fixed sources, which makes its coefficient enumerable by a divisor bound. \begin{lemma}[Conditional dense pair estimate] \label{p13:transfer:dense-targets} Fix a number $s$ of source windows and an effective subspace $W$ of their value space. Suppose $S_y\cap W=S_z\cap W=0$, and write \[ d=\dim W+\dim S_y+\dim S_z-\dim(W+S_y+S_z). \] For every fixed $\eta,\varepsilon>0$ there exists a fixed $\theta>0$ such that, uniformly in the sources and $W$, \begin{equation} \sum_{\substack{y\ne z,\ y,z\text{ outside the sources}\\ d\geq\eta L\\ (y,z)\text{ nonaligned or }\dim(S_y\cap S_z)\leq\theta L}} p_yp_z2^d\ll_{s,C,\eta,\varepsilon}N^{-1+\varepsilon}. \label{p13:transfer:dense-target-bound} \end{equation} In particular the hypothesis on the pair is satisfied when $S_y,S_z$ are direct. For a primitive family of fixed size $r\geq3$, with a prescribed collection of $r-2$ sources having at most $N^{a+\varepsilon}$ possibilities, its dense homogeneous mass is at most \begin{equation} \sum_{\rho\geq\eta L}\Bigl(\prod_{i\in I}p_i\Bigr)2^{\rho(I)} \ll N^{-(r-1)+a+\varepsilon}. \label{p13:transfer:restricted-dense} \end{equation} \end{lemma} The kernel defining $d$ projects injectively onto each target, so its coordinate labels force many binary components in each target separately. After regularization and removal of the weak aligned components, let $m_t$ count the remaining components meeting target $t$, and let $n_{1,t}$ count those using exactly one of its $B=L+1$ coordinates. Then $n_{1,t}\geq2m_t-B$. Above $d=(1/2+\delta)L$, this forces a positive proportion of singleton occurrences in both targets, allowing the stronger Pell or degree-one curve count. The resulting bounds are $N^{1/2+\varepsilon'}$ target pairs in the lower stratum and $N^{\varepsilon'}$ in the upper stratum; multiplying by $p_yp_z2^d$ gives $N^{-1+\varepsilon}$ after choosing the internal margins. Section~\ref{p14:comp:transfer-dense-targets} supplies the separate component counts, the coefficient enumeration from fixed sources, and the removal of small aligned curves, including effective local sections. \begin{lemma}[A dense compressed target] \label{p13:transfer:compressed-target} For fixed sources and a fixed channel of good pairs, the number of compressed centers $X$ for which $\dim(S(U_X)\cap W)\geq\eta L$ is $O_{s,C,\eta,\varepsilon}(N^\varepsilon)$. \end{lemma} \begin{proof} The compression has fixed width $O(B)$ and an injective even-selection map. In Proposition~\ref{p1:fixed:targets}, take $H$ larger than this width. Its regularization loses a fixed small fraction of $\eta L$. A prime above $H$ has at most one target occurrence, so all other occurrences in its circuit are source occurrences. The circuit products are divisors of fixed source values, and the fixed-source CRT and split-product steps in that proposition are unchanged. The proof uses the width bound and the target projection rank, not a target space of dimension exactly $L$. The finitely many channel dilations and $B^{O(1)}$ offset patterns are absorbed in the margins. \end{proof} \subsection{The exact weights of primitive families} The primitive estimate below combines the dense-host construction of Section~\ref{p14:comp:confinement}, the homogeneous envelope of Lemma~\ref{p2:homogeneous:envelope}, and the conditional half-count through Corollary~\ref{p13:half:connected}. We first record a separate small-kernel estimate for related counting arguments; it is not used in the proof of Proposition~\ref{p13:transfer:primitive}. For $Y\geq H$, put \[ K_Y(n)=\prod_{\substack{\ell>Y\\v_\ell(n)\ \mathrm{odd}}}\ell,\qquad a_Y(n)=\prod_{\ell>Y}\ell^{\lfloor v_\ell(n)/2\rfloor}. \] Then $n=s_Y(n)a_Y(n)^2K_Y(n)$, where $s_Y(n)$ is $Y$-smooth. \begin{lemma}[Several small kernels in one window] \label{p13:transfer:small-kernels} Fix $A>2$, $01$. For $Y=(\log N)^A$, the number of windows containing $m$ distinct values with $K_Y(n)\leq N^b$ and $a_Y(n)^2\leq N^u$ is \begin{equation} \ll N^{\max(1-1/A,(1+u)/2)+\varepsilon}. \label{p13:transfer:small-kernels-bound} \end{equation} Also, with the previously fixed cutoff $H=C_DB$, the set $\mathcal X_q$ of windows containing distinct $n_1,n_2$ with $K_H(n_1)K_H(n_2)\leq N^q$ has cardinality $O(N^{q+\varepsilon})$, for fixed $q\geq0$. \end{lemma} \noindent The complete proof is given in Section~\ref{p14:comp:transfer-small-kernels}. \begin{proposition}[Homogeneous mass at fixed primitive arity] \label{p13:transfer:primitive} \label{p2:homogeneous:primitive} Fix an integer $r\ge3$, a critical band, and $\varepsilon>0$. Let $\mathcal P_{r,N}$ be the primitive $r$-element families of original windows in $I_N$. With the four weights of Lemma~\ref{p2:homogeneous:envelope}, \begin{equation} \sum_{I\in\mathcal P_{r,N}}q_0(I) \ll_{r,C,\varepsilon}N^{-r/2+\varepsilon}, \qquad \sum_{I\in\mathcal P_{r,N}}w(I),\quad \sum_{I\in\mathcal P_{r,N}}q(I),\quad \sum_{I\in\mathcal P_{r,N}}|c(I)| \ll_{r,C,\varepsilon}N^{-r/2+\varepsilon}. \label{p13:transfer:primitive-bound} \end{equation} For each fixed $\eta>0$, the subfamily \[ \mathcal P_{r,N}(\eta) =\{I\in\mathcal P_{r,N}:\rho_{\mathrm{rel}}(I)\ge\eta L\} \] has $O_{r,C,\eta,\varepsilon}(N^\varepsilon)$ members, and each of its four total weights is $O_{r,C,\eta,\varepsilon}(N^{1-r+\varepsilon})$. One scale threshold controls these conclusions and every deterministic subfamily. No regularity, separation, affine compatibility, or positivity of the original marginals is required. The arity is fixed before that threshold is chosen. \end{proposition} \begin{proof} The \hyperref[p18:conf:fixed-primitive]{dense-host construction} in Section~\ref{p14:comp:confinement} proves the required count. This part of the construction is geometric: it counts centres before assigning probability weights, and applies to all true windows. Primitivity makes each relation-space projection injective, whence \[ \rho_{\mathrm{rel}}(I)\le\min_{x\in I}d_x\le L. \] Lemma~\ref{p2:homogeneous:envelope} therefore gives \[ q_0(I)\le(2p)^r2^L\ll_{r,C}N^{1-r}. \] Multiplying by the dense host count proves the dense homogeneous bound. The other three bounds follow pointwise from the same lemma. For the full homogeneous sum, choose fixed small $\xi,\delta>0$ and split at $\rho_{\mathrm{rel}}=\xi L$. Corollary~\ref{p13:half:connected} counts at most $O_{r,C,\delta}(N^{r/2+\delta})$ primitive hosts. On the lower part, $q_0(I)\ll_{r,C}N^{-r+\xi}$, so their contribution is $O(N^{-r/2+\xi+\delta})$. On the upper part, the dense bound just proved with parameter $\xi$ gives $O(N^{1-r+\delta})$. Because $r\ge3$, $1-r<-r/2$. Choose $\xi,\delta$ within the requested margin and take the maximum of the finitely many scale thresholds involved. Pointwise domination again gives the other three sums, and restriction to any deterministic subfamily preserves every estimate. \end{proof} On the regular set, $q_0(I)=p^r2^{\rho_{\mathrm{rel}}(I)}$. For ordered lists of distinct starts, the same bounds acquire only the fixed factor $r!$. Estimates at growing cardinalities will instead use the independent certificate tail below. The exact centered expansion is \eqref{p1:alg:centred-identity}. The retained measure $B_2$ uses its terms of cardinalities zero and two; the stopping identity controls the remainder through a positive suffix. \subsection{An independent extension estimate at all cardinalities} For a fixed number $s$ of source windows and an effective subspace $W$ of their value space, let \[ E_W(u)=\sum_{\substack{U\subset\mathcal I_N\setminus\{\text{sources}\}\\ W,(S_x)_{x\in U}\text{ not direct}}} u^{|U|}\prod_{x\in U}p_x, \qquad u>0. \] The weights here are products of marginal probabilities. They are not conditional probabilities of dependent extensions. \begin{lemma}[Uniform independent deficit] \label{p13:transfer:independent-deficit} For fixed $s,u,C$ and every $\varepsilon>0$, \begin{equation} E_W(u)\ll_{s,u,C,\varepsilon}N^{-1/2+\varepsilon}, \label{p13:transfer:independent-bound} \end{equation} uniformly in the sources and $W$. The sum includes all cardinalities. \end{lemma} \begin{proof} Set $Y=B^2$ and \[ a(n)=\prod_{\substack{\ell>Y\\\ell\mid n}}(1+\ell^{-1/2}),\qquad A_x=\prod_{\substack{\ell>Y\\\ell\mid n\text{ for some }n\in V_x}} (1+\ell^{-1/2}). \] H\"older applied to the $B$ positions in a window gives \[ \sum_{N\leq x<2N}A_x \leq\prod_{i=-1}^{L-1} \left(\sum_{N\leq x<2N}a(x+i)^B\right)^{1/B}. \] Write $a(n)^B=\sum_{d\mid n}h_B(d)$, where $h_B$ is multiplicative, supported on squarefree integers with primes exceeding $Y$, and $h_B(\ell)=(1+\ell^{-1/2})^B-1\ll B\ell^{-1/2}$. The coefficients are nonnegative, so \[ \sum_{n\leq3N}a(n)^B \leq3N\prod_{Y<\ell\leq3N}\left(1+\frac{h_B(\ell)}\ell\right) \leq3N\exp\left(O\left(B\sum_{m>Y}m^{-3/2}\right)\right) \ll N. \] Thus $\sum_xA_x\ll N$, with a constant independent of $B$. The small-prime contribution satisfies \[ A_{\leq Y}=\prod_{\ell\leq Y}(1+\ell^{-1/2}) \leq\exp(O(B/\log B))=N^{o(1)}. \] Chebyshev's bound and partial summation suffice for this estimate. Let $P_0$ contain the small primes and the prime factors of the fixed sources. Their total logarithmic size is $O_s(B\log N)$. Using the same $Y=B^2$, we have \[ \sum_{\substack{\ell\in P_0\\\ell>Y}}\ell^{-1/2} \leq\frac{\sum_{\ell\in P_0}\log\ell}{\sqrt Y\log Y} =O_s(B/\log B). \] The primes up to $Y$ have the same bound by partial summation. Consequently $A_0=\prod_{\ell\in P_0}(1+\ell^{-1/2}) =\exp(O_s(B/\log B))=N^{o(1)}$. If $P_k$ also includes the primes of an arbitrary list $x_1,\ldots,x_k$, then \[ \prod_{\ell\in P_k}(1+\ell^{-1/2}) \leq A_0A_{\leq Y}\prod_{j=1}^k A_{x_j}. \] The elementary proof of the exceptional-target estimate in Proposition~\ref{p1:arith:fixed-branching}, before restricting the number of sources, gives \begin{equation} D(x_1,\ldots,x_k) \ll B(s+k)+B\sqrt N\prod_{\ell\in P_k}(1+\ell^{-1/2}). \label{p13:transfer:explicit-exception} \end{equation} Contacts account for the first term. Outside them, a selected target value whose odd prime factors all belong to $P_k$ has the form $a v^2$, with $a$ squarefree and supported on $P_k$. Summing $\sqrt{N/a}$ over these $a$ gives the second term, with $B$ possible placements. This upper bound also includes local defects. Put $w_x=up_x$, $w_*\ll N^{-1}$, $M_0=\sum_xw_x=O_{u,C}(1)$, and $M_1=\sum_xw_xA_x=O_{u,C}(1)$. Every non-direct extension of size $h$ has an exceptional exposure in each of its $h!$ orders. Sum over the possible first exceptional position, allowing repeated centers for an upper bound. The first term of \eqref{p13:transfer:explicit-exception} contributes at most \[ O(Bw_*)\sum_{h\geq1}\frac1{h!} \sum_{i=1}^h(s+i-1)M_0^{h-1}=O(B/N). \] The second contributes at most \[ O(B\sqrt Nw_*A_0A_{\leq Y}) \sum_{h\geq1}\frac1{h!}\sum_{i=1}^h M_1^{i-1}M_0^{h-i} =N^{-1/2+o(1)}. \] The double sum is bounded by $e^{\max(1,M_0,M_1)}$. The varying number of sources has therefore been averaged explicitly; no constant from a fixed-order theorem has been used at a growing order. \end{proof} \begin{lemma}[A fixed number of extensions of a pair] \label{p13:transfer:fixed-deficit} For every fixed $k\geq1$, \begin{equation} \sum_{|e|=2}|c(e)| \sum_{\substack{U\cap e=\varnothing,\ |U|=k\\ S_e,(S_z)_{z\in U}\text{ not direct}}} \prod_{z\in U}p_z \ll_{k,C,\varepsilon}N^{-3/2+\varepsilon}, \qquad S_e=\sum_{i\in e}S_i. \label{p13:transfer:fixed-pair-bound} \end{equation} \end{lemma} \begin{proof} If the new windows are not direct among themselves, choose a primitive subfamily. A primitive of fixed size $h\geq2$ has independent mass $O(N^{-h/2+\eta})$ by Theorem~\ref{p13:half:conditional} with no prescribed sources. The largest such cost is $N^{-1+\eta}$. After multiplication by $A_2$ and summation of the remaining independent windows this is $N^{-5/3+O(\eta)}$. Otherwise choose a minimal nonempty $A\subseteq U$ with $S_e\cap\sum_{a\in A}S_a\ne0$. If $|A|=1$, the total cost is the one-target deficit of Proposition~\ref{p13:activity:deficit}. If $h=|A|\geq2$, a relation meeting $S_e$ has a nonzero projection in each of the $h$ target spaces, by minimality. Conditional on $e$, Theorem~\ref{p13:half:conditional} gives $N^{h/2+\eta}$ possibilities. The product of the $h$ target marginals costs $O_h(N^{-h})$; summing the pair weights therefore gives $N^{-2/3-h/2+O(\eta)}$. The remaining $k-h$ independent weights have bounded sum. There are only finitely many choices of $h$ and roles because $k$ is fixed. \end{proof} \subsection{Uniform certificate tails and the deficit of all pairs} We record the part of the prime-certificate argument that is uniform in the number of blocks. A block may be an original window or one of a fixed finite collection of geometric compressions of width $O(B)$. Its center ranges over $O(N)$ values and its independent weight is bounded by a fixed multiple of $p$. A relation of full support selects a nonzero even vector in every block. \begin{lemma}[Independent full-support tails] \label{p13:transfer:independent-tail} Fix a critical band, fixed bounds on the block widths and centre ranges, a weight $v>0$, and an exponent $a<2$. There exist a fixed integer $s\ge3$ and a scale threshold such that, for every $N$ above that threshold, \begin{equation} \sum_{\substack{|F|\ge s\\F\text{ has a full-support relation}}} \prod_{x\in F}(vp_x)\le N^{-a}. \label{p13:transfer:tail-independent} \end{equation} The same conclusion holds with one block of any of the fixed compressed types, weighted by $vp$, and with the remaining factors $vp_x$. The cutoff and threshold depend only on the fixed parameters, not on $N,L$ or the chosen blocks. In particular the cutoff is fixed before any later exponent margin. \end{lemma} \begin{proof} Here are the parameter calculations in the certificate argument of Lemmas~\ref{p1:core:moderate}--\ref{p1:core:generator}. Choose $Y=(\log N)^{A_0}$ and a kernel threshold $T=N^b$, with $A_0>1$ and $00$. Every unmarked block in a full-support relation has two selected values of kernel exceeding $T$. Their kernels are coprime, since $Y$ exceeds the fixed block diameter. For a certificate in which every receiving modulus is at most $O(NB)$, the CRT and the exponential sum over cardinalities give independent mass \[ N^{-2b+4/A_0+o(1)}. \] Here $m=\lceil2\log(O(N))/\log Y\rceil$ and \[ \log(B^m(m+1)!)=(4/A_0+o(1))\log N. \] If no such certificate exists, each unmarked block has a saturated neighbor. Its directed degree is $N^{2/A_0+o(1)}$ in both directions, by Lemma~\ref{p1:core:degrees}. When $2/A_0+\delta<1$, the component generator of Lemma~\ref{p1:core:generator}, tilted by $N^{\delta/2}$, stays bounded. The mass at cardinality at least $s$ is consequently $N^{-s\delta/2+o(1)}$. Choose $A_0$ large and then $b$ close enough to one that $2b-4/A_0>a$; choose an allowed $\sigma$ and then $s$ with $s\delta/2>a$. All these choices are made before $N$. The proof uses the actual prime divisors of selected values and a fixed bound on the block width. Changing one block to a fixed geometric type changes only constant factors and $B^{O(1)}$ offset choices. Its two anchors, receiving congruences, and degree bounds obey the same estimates. Thus the same parameter choices prove the compressed assertion. The size cutoff can be chosen before any final $\varepsilon$. Indeed, fix $a<2$, put $a'=(a+2)/2$ and $\varepsilon'=(2-a)/4$, and apply the established estimate with these fixed parameters. Since $-a'+\varepsilon'<-a$, it yields the pure bound $N^{-a}$ above a suitable threshold, for both versions with the same fixed cutoff. This argument retains $a<2$ and a fixed bound on the compressed block width. \end{proof} For $e=\{x,y\}$ put $E_e(u)=E_{S_e}(u)$ and $D_2(u)=\sum_{|e|=2}|c(e)|E_e(u)$. \begin{proposition}[The pair deficit over all cardinalities] \label{p13:transfer:pair-deficit} For fixed $u>0$ and every $\varepsilon>0$, \begin{equation} D_2(u)\ll_{u,C,\varepsilon}N^{-3/2+\varepsilon}. \label{p13:transfer:pair-deficit-bound} \end{equation} In particular it is smaller than the $N^{-4/3}$ scale required below. \end{proposition} \begin{proof} First consider extensions with an internal primitive family of new windows. At each fixed size their independent mass is $O(N^{-1+\eta})$ by the half-exponent count; at all larger sizes apply Lemma~\ref{p13:transfer:independent-tail} with any $a>1$. Summing free complements costs at most $e^{u\mu}$. Their contribution to $D_2$ is $O(A_2N^{-1+\eta})=O(N^{-5/3+2\eta})$. For a direct extension, choose its minimal subfamily $A$ crossing $S_e$. Use Lemma~\ref{p13:activity:pair-partition}. Weak and exceptional pairs have total activity $O(N^{-1+\eta})$ and $O(N^{-1+\eta})$, respectively; Lemma~\ref{p13:transfer:independent-deficit} bounds their complete contribution by $N^{-3/2+O(\eta)}$. There remain good pairs in a finite set of fixed channels. The cases $|A|0$, \begin{equation} \sum_{\substack{|I|=r\\I\text{ connected}}}|c(I)| \ll_{r,C,\varepsilon}N^\varepsilon A_3+N^{-4/3+\varepsilon}. \label{p13:transfer:fixed-activity-bound} \end{equation} The incompatible part, for every fixed $r\geq4$, is $O_{r,C,\varepsilon}(N^{-3/2+\varepsilon})$. \end{proposition} \begin{proof} We first bound compatible families by their positive probabilities. Let $T_3^+$ denote the sum of $q(T)$ over connected compatible triples. For a primitive triple, $q(T)\leq2c(T)$. For a nonprimitive compatible triple, \[ q(T)=c(T)+\sum_{e\subset T,\ |e|=2}c(e)p_{T\setminus e} +\prod_{i\in T}p_i. \] The pair terms sum to the one-target deficit, and the independent term has mass $O(N^{-3/2+\eta})$ by the half-exponent count. Therefore \begin{equation} T_3^+\leq2A_3+O(N^{-3/2+\eta}). \label{p13:transfer:positive-triples} \end{equation} Families containing such a triple have mass $N^\eta T_3^+$ after positive propagation~\eqref{p13:exact:propagation}. Families containing a primitive of size at least four have mass $N^{-4/3+\eta}$, by Proposition~\ref{p13:transfer:primitive} and the same propagation. All seed sizes here are bounded by the fixed $r$. Consider a remaining compatible connected family. It contains a dependent pair $e$; otherwise a minimal dependent subfamily would be one of the seeds just treated. Put $U=S_e$ and choose a nonempty $A\subset I\setminus e$ minimal with $U\cap\sum_{a\in A}S_a\ne0$. A single target gives a connected triple, so $h=|A|\geq2$. If $A$ contains a proper primitive $P$, then $U\cap\sum_{i\in P}S_i=0$ by minimality. Thus $q(e\cup P)=q(e)q(P)$. Only $|P|=2$ remains, since larger primitives were already treated. Compatible dependent pairs satisfy $q(e)\leq2c(e)$, so the total seed mass is at most $4A_2^2$. Propagate it to $I$. Suppose next that $A$ is direct. The nullity $d$ of the compressed family $U,(S_a)_{a\in A}$ is at most $L$, and \[ q(e\cup A)=q(e)\Bigl(\prod_{a\in A}p_a\Bigr)2^d. \] For $d<\eta L$, the conditional half-exponent count of the $h$ targets gives a cost at most $A_2N^{-h/2+O(\eta)}$, hence $N^{-5/3+O(\eta)}$. For $d\geq\eta L$, fix $h-2$ members of $A$. The last two targets are each direct from the sum of $U$ and these fixed targets, and are direct from one another. Otherwise there would be an internal dependence in $A$ or an earlier crossing of $U$. Lemma~\ref{p13:transfer:dense-targets} gives conditional weighted cost $N^{-1+\eta}$. The independent weights of the $h-2$ additional sources sum to a constant; multiplication by the pair mass gives $N^{-5/3+O(\eta)}$ again. If $A$ is itself primitive of size at least three, it was already one of the excluded seeds. It remains that $A=f=\{y,z\}$ is a dependent pair. By minimality, $U\cap S_y=U\cap S_z=0$. Put \[ d_e=\dim(S_x\cap S_{x'}),\quad d_f=\dim(S_y\cap S_z),\quad \delta=\dim(U\cap S_f),\qquad e=\{x,x'\}. \] If $\delta<\eta L$, then $q(e\cup f)=q(e)q(f)2^\delta$, and the total is $O(N^\eta A_2^2)$. For $\delta\geq\eta L$, choose $\theta$ from the dense-target lemma. If $f$ is nonaligned or $d_f\leq\theta L$, that lemma gives $N^{-5/3+O(\eta)}$ after summing $e$. If $f$ is a good pair of a fixed channel, fix $e$ and use Lemma~\ref{p13:transfer:compressed-target}; there are only $N^\eta$ choices for the compressed center of $f$. Linear algebra gives \begin{equation} d_f+\delta\leq L, \qquad q(e\cup f)\leq p_*q(e). \label{p13:transfer:two-pair-rank} \end{equation} Indeed the projection of $U\cap(S_y+S_z)$ modulo $S_y$ is injective, and its image has dimension at most $\dim S_z-d_f$. This case also costs $N^{-5/3+O(\eta)}$. Every aligned pair with rank greater than $\theta L$ has bounded channel height, by Lemma~\ref{p1:arith:alignedrank}. Outside the good pairs at those finitely many heights, the normalization estimates in Theorem~\ref{p1:arith:pairs} and Lemma~\ref{p1:arith:fixed-channel-mass} give at most $N^{1/2+\eta}$ exceptional aligned pairs. The endpoint pairs of height two and the pairs meeting an irregular window satisfy the same bound; height one has rank $o(L)$ when compatible. Call this exceptional collection $\mathcal B$. If only $f$ belongs to $\mathcal B$, exchange $e$ and $f$ and use one of the preceding cases. This exchange is valid because a target in $e$ intersecting $S_f$ would form a connected triple, already removed. If both belong to $\mathcal B$, there are at most $N^{1+2\eta}$ hosts. The symmetric forms of \eqref{p13:transfer:two-pair-rank} give \[ d_e+d_f+\delta\leq L+\min(d_e,d_f)\leq\tfrac32L+o(L), \] where the final inequality uses Theorems~\ref{p13:rig:nonaligned} and~\ref{p13:rig:aligned}. The joint probability is at most $N^{-5/2+o(1)}$, so this contribution is $N^{-3/2+O(\eta)}$. These cases exhaust the possible minimal sets $A$. Positive propagation adds the remaining windows, proving the compatible part of \eqref{p13:transfer:fixed-activity-bound}. For incompatible families we use induction on $r$, starting with the triple estimate. The exact identity is \begin{equation} c(I)=-\prod_{i\in I}p_i- \sum_{\substack{J\subsetneq I\\|J|\geq2}} c(J)\prod_{i\in I\setminus J}p_i. \label{p13:transfer:incompatible-expansion} \end{equation} The independent term over connected $r$-tuples is $O(N^{-r/2+\eta})$. The pair terms cost $O(N^{-3/2+\eta})$ by Lemma~\ref{p13:transfer:fixed-deficit}. For a fixed proper $J$ of size at least three, connectedness of $I$ requires a non-direct extension of its total space; its independent cost is $N^{-1/2+\eta}$ by Lemma~\ref{p13:transfer:independent-deficit}. Canonical factorization of $c(J)$ has no singleton factors. By induction a factor of size at least three costs $N^\eta A_3+N^{-4/3+\eta}$; a decomposition consisting only of pairs has at least two of them and costs $O(A_2^2)$. There are finitely many partitions at each fixed size. Multiplying by the independent extension cost proves \eqref{p13:transfer:fixed-activity-bound} at order $r$. Inserting the arithmetic bound $A_3\ll N^{-41/30+\eta}$ from Corollary~\ref{p13:third:activity} into the same induction gives the stated $N^{-3/2+\varepsilon}$ incompatible activity. No law approximation or lower bound has been used in this induction. \end{proof} \subsection{A uniform tail after 256 blocks} The preceding proposition has constants depending on its fixed order. We now use the certificate generator to handle all larger orders in the stopped expansion. The number $256$ is a sufficient fixed cutoff, not an optimal threshold. \begin{lemma}[Certificate budgets with several anchors] \label{p13:transfer:anchor-budgets} Choose an even integer $t\geq2$, $A_0>1$, $00$, every $\varepsilon>0$, and every $K_N=o(L)$, \begin{align} \sum_{\substack{I\text{ connected},\ 256<|I|\leq K_N\\ \rho_*(I)\leq2L}} w^{|I|}|c(I)|&\ll N^{-4/3+\varepsilon}, \label{p13:transfer:connected-tail}\\ \sum_{\substack{P\text{ primitive}\\|P|>256}} w^{|P|}|c(P)|&\ll N^{-4/3+\varepsilon}. \label{p13:transfer:primitive-tail} \end{align} The second sum includes all cardinalities. Here $\rho_*(I)=\max\{\rho(A):A\subseteq I,\ q(A)>0\}$. \end{proposition} \begin{proof} For compatible $10$-cores of nullity at most $2L$, choose \[ t=10,\quad A_0=12,\quad b=31/60,\quad \sigma=23/25,\quad\delta=29/750,\quad\theta=5/6. \] The condition $\theta+\delta<1$ is strict. Multiplying the independent budgets by $2^{2L}=O_C(N^2)$ gives, at sizes at least $256$, \begin{equation} \sum_D w^{|D|}q(D) \ll N^{-3/2+o(1)}+N^{-1106/375+o(1)}. \label{p13:transfer:large-core} \end{equation} Indeed $2-10(31/60)+20/12=-3/2$ and $2-256(29/1500)=-1106/375$. If $\rho_*(I)\leq L/60$, the centered expansion gives \[ |c(I)|\leq2^{|I|}\Bigl(\prod_{x\in I}p_x\Bigr)2^{L/60}. \] A nonzero centered coefficient has a full-support Fourier relation, so two anchors are available in every block. Use \[ t=2,\quad A_0=12,\quad b=17/20,\quad \sigma=23/25,\quad\delta=3/250,\quad\theta=1/6. \] Starting at cardinality $256$, the two costs after the rank factor are \begin{equation} N^{-27/20+o(1)}+N^{-2279/1500+o(1)}. \label{p13:transfer:low-rank-tail} \end{equation} The four costs in \eqref{p13:transfer:large-core} and \eqref{p13:transfer:low-rank-tail} leave the exact margins \[ \begin{aligned} \frac32-\frac43&=\frac16,& \frac{1106}{375}-\frac43&=\frac{202}{125},\\ \frac{27}{20}-\frac43&=\frac1{60},& \frac{2279}{1500}-\frac43&=\frac{93}{500}. \end{aligned} \] Thus the smallest margin is $1/60$. For the fixed certificate parameters and fixed weight $w$, every displayed $o(1)$ loss is less than $1/120$ for sufficiently large $N$; no certificate parameter is allowed to vary with $N$. For the remaining connected $I$, choose deterministically a compatible $A\subseteq I$ attaining $\rho_*(I)>L/60$. Delete blocks of projection dimension at most eight until a $10$-core $D$ remains. Since $8K_N=o(L)$, its nullity exceeds $L/120$ for large $N$, so $|D|\geq2$. With $r=|I|$ and $h=|D|$, the centered expansion and maximality of $\rho_*(I)$ imply \begin{equation} |c(I)|\leq2^r2^{8(r-h)} \Bigl(\prod_{x\in I\setminus D}p_x\Bigr)q(D). \label{p13:transfer:peeling} \end{equation} The elements outside $A$ are already in this product; there is no enumeration of deletion orders. All fixed powers of two can be absorbed in the independent weight of the complements. If $h\geq256$, sum the complements freely and apply \eqref{p13:transfer:large-core}. If $3\leq h<256$, then $q(D)\leq2c(D)$ by \eqref{p13:transfer:core-ratio}. Proposition~\ref{p13:transfer:fixed-activities}, the arithmetic triple bound, and canonical factorization show that the sum of $|c(D)|$ at each such fixed size is $O(N^{-1+\eta})$. Because $D$ is proper in the connected $I$, its complement is a non-direct independent extension. Lemma~\ref{p13:transfer:independent-deficit} adds $N^{-1/2+\eta}$. If $h=2$, then $D=e$ is a compatible dependent pair and $q(e)\leq2c(e)$; the complement is exactly the pair deficit of Proposition~\ref{p13:transfer:pair-deficit}. These bounds prove \eqref{p13:transfer:connected-tail}. A primitive family has nullity at most $L$. If it is incompatible, $|c(P)|=\prod_Pp_x$ and the independent two-anchor budgets suffice. If it is compatible with nullity at most $L/60$, use \eqref{p13:transfer:low-rank-tail}. Otherwise each projection has dimension $\rho(P)>L/60$, so it is itself a $10$-core for large $N$. Apply \eqref{p13:transfer:large-core} directly, without an upper cardinality cutoff. This proves \eqref{p13:transfer:primitive-tail}. \end{proof} \subsection{Conditional dense targets and small-kernel counts} \label{p14:comp:transfer} We prove the conditional dense-pair estimate of Lemma~\ref{p13:transfer:dense-targets} and then the separate auxiliary small-kernel estimate of Lemma~\ref{p13:transfer:small-kernels}. The first supplies a fixed-source input to the reconstruction; the second completes the auxiliary statement recorded above. All source counts, widths, and rank thresholds are fixed before $N$ tends to infinity. Neither proof uses the finite-transfer theorem or the law approximation. \subsubsection{Proof of Lemma~\ref{p13:transfer:dense-targets}} \label{p14:comp:transfer-dense-targets} \begin{proof} We first count target pairs, keeping the number of binary components in \emph{each} target explicit. All component counts below include isolated active coordinates. \paragraph{The two injective projections.} Put \[ R=\ker\bigl(W\oplus S_y\oplus S_z\longrightarrow W+S_y+S_z\bigr), \qquad \dim R=d. \] If an element of $R$ has zero $y$-projection, its other two entries belong to $W\cap S_z=0$. Hence the projection $R\to S_y$ is injective; the other projection is injective by $W\cap S_y=0$. In particular $d\leq\min(\dim S_y,\dim S_z)\leq L$. Represent $W$ by a linear section in the physical source coordinates, and each target space by a section in its formal even-selection space. These sections preserve both injective target projections and all actual prime-valuation equations, including on an irregular window. Fix $0<\delta<1/2$, to be chosen inside the final exponent margin, and choose fixed positive $\xi,\theta$ with \[ \xi\leq\tfrac14\min(\eta,\delta),\qquad \theta\leq\tfrac18\min(\eta,\delta). \] Apply Lemma~\ref{p1:fixed:recovery} to obtain $R'\subseteq R$ with $d':=\dim R'\geq d-\xi L$. Choose its recovery parameter large enough to give this loss and to control circuits in the $s+2$ physical blocks. Every nonzero coordinate label on $R'$ then has a fixed bounded multiplicity $K$. For each prime $\ell>H\asymp B$, pair equal active labels in its row. All edges of the resulting binary graph join equal nonzero labels, so every component lies within one label class and has at most $K$ vertices. The coordinate labels in either target span $(R')^*$, since its projection is injective. A component contributes only one label to this span, even if it meets that target several times. Consequently each target meets at least $d'$ binary components. This is a separate lower bound for each target, not just for their union. \paragraph{Enumerating coefficients from the fixed sources.} After the equal labels have been paired, the remaining distinct labels in a prime row split into minimal circuits. Each such circuit uses at least three physical blocks: $\ell>H$ has at most one occurrence in any block. Since only two blocks are free, every circuit contains a source occurrence. The product of primes assigned to a circuit support therefore divides a fixed source value. The recovery bound gives a fixed bound on the number of circuit supports incident to any selected coordinate. For a fixed number of selected binary components, enumerate their positions, these circuit supports, one source occurrence on each support, and divisors of the corresponding source values. The cost is $B^{O(1)}N^{o(1)}=N^{o(1)}$. Small-prime parities cost $2^{O(\pi(H))}=N^{o(1)}$ as well. Thus the squarefree coefficients of their component equations have $N^{o(1)}$ choices and height $N^{A_0}$ for a fixed $A_0$. The height bound per component is fixed once $s$, the recovery parameters and the width multiplier are fixed; it does not increase when more disjoint component equations are selected. The number of equations required by Lemma~\ref{p1:arith:geometric} is chosen afterwards. No subspace or full deletion pattern is enumerated. \paragraph{Removing the small aligned curves.} Two exact units between the targets determine a reduced alignment of height $O(B)$. Two distinct alignments of this height would force $y,z=O(B^3)$, impossible in the critical range. If there are fewer than two exact units, no removal is needed. Otherwise let $u\geq2$ be the number of units on the unique alignment. Their even selections have dimension $u-1$ before passage to square classes, and a target has at most one local rank defect. Their common image therefore gives \[ u-2\leq\dim(S_y\cap S_z)\leq\theta L \] in the weak aligned case. Keep at most one of these units and exclude from selection every binary component meeting an endpoint of another unit. At most $2(u-1)\leq2\theta L+2$ components are excluded. This deletes components from the selectable collection; it does not restrict $R'$ again. The surviving positions contain no two exact units on an alignment, which excludes the small-curve alternatives in the proof of Lemma~\ref{p1:arith:geometric}. No removal is needed for a nonaligned pair. For $t=y,z$, write $m_t$ for the number of selectable components meeting target $t$. The separate projection bounds now give \begin{equation*} m_t\geq d'-2\theta L-2 \geq d-(\xi+2\theta)L-2. \end{equation*} Thus $m_t\geq\eta L/3$ for large $N$ whenever $d\geq\eta L$; if $d>(1/2+\delta)L$, then $m_t>(1/2+\delta/3)L$ for both targets. \paragraph{The first stratum: a positive rank proportion.} A component type records its number of coordinates in each physical block. There are finitely many types, because the component size is at most $K$. If a type meeting both targets occurs sufficiently often, select the fixed number of its components required by the two-variable geometric count. With the small curves excluded above and coefficients of height $N^{A_0}$, the proof of Lemma~\ref{p1:arith:geometric} gives $N^{1/2+\varepsilon'}$ target pairs, including the $N^{o(1)}$ pattern and coefficient choices. Otherwise only a bounded number of components meet both targets. Each target still has a positive proportion of $L$ components avoiding the other target. In each target choose a univariate type occurring with positive frequency, and then the fixed number of its components needed by the one-variable part of Lemma~\ref{p1:arith:geometric}, taking at least two. Their component equations form a one-variable system with fixed coefficients and at least two distinct selected offsets; different components have disjoint vertices. In the degree-one case, these are precisely the two linear factors for Lemma~\ref{p1:arith:pell}. The one-variable count gives $N^{\varepsilon'}$ target pairs after allocating the margin between the two centers and the coefficient choices. Every selected set has fixed size, so its even-selection square-class map is injective for large $N$ by Lemma~\ref{p1:arith:runge}; irregularity of the entire window is irrelevant to this local use. We have proved \[ \#\{(y,z):d\geq\eta L\}\ll N^{1/2+\varepsilon'}. \] Both counts here and below impose the pair restriction in the lemma. \paragraph{The second stratum: rank greater than half a window.} Suppose $d>(1/2+\delta)L$. Let $n_{1,t}$ be the number of selectable components using exactly one coordinate in target $t$. Every other component meeting that target uses at least two of its $B=L+1$ coordinates. Hence, separately for $t=y,z$, \[ B\geq n_{1,t}+2(m_t-n_{1,t}),\qquad n_{1,t}\geq2m_t-B>\frac{2\delta}{3}L-1. \] If a mixed type with one coordinate in either target occurs sufficiently often, the degree-one case of the geometric count gives $N^{\varepsilon'}$ pairs directly. Otherwise only boundedly many of the $n_{1,t}$ components are mixed. The displayed inequality then provides two components for each target, each using one coordinate of that target and no coordinate of the other. The resulting two distinct linear factors give the uniform two-value Pell count for each center. Thus \[ \#\{(y,z):d>(1/2+\delta)L\}\ll N^{\varepsilon'}. \] This explains the threshold $L/2$: above it, the coordinate budget forces singleton occurrences in each target. Finally $p_y,p_z\ll N^{-1}$ and $2^L\asymp_C N$. The stratum $\eta L\leq d\leq(1/2+\delta)L$ has weighted cost $N^{1/2+\varepsilon'}N^{-2+1/2+\delta}$, and the stratum above it has cost $N^{\varepsilon'}N^{-2+1}$. Choose $\delta>0$ and then $\varepsilon'>0$ so that $\delta+\varepsilon'<\varepsilon$; all preceding choices are fixed before $N$. This proves \eqref{p13:transfer:dense-target-bound}. In a primitive family, each target is direct from the prescribed $r-2$ sources separately, as well as from the other target; the full nullity is at most $L$. Multiplication by the $r-2$ source marginals and summation over the prescribed source collection gives \eqref{p13:transfer:restricted-dense}, after dividing the final exponent margin between the source count and the conditional estimate. \end{proof} \subsubsection{Proof of Lemma~\ref{p13:transfer:small-kernels}} \label{p14:comp:transfer-small-kernels} \begin{proof} For the first assertion fix the $m$ offsets, at a polylogarithmic cost, and write $d_i=n_i/K_Y(n_i)=s_i a_i^2$. Since $d_i\gg N^{1-b}$ and $\gcd(d_i,d_j)\ll B$, \[ \operatorname{lcm}(d_1,\ldots,d_m) \geq\frac{\prod_i d_i}{\prod_{iN. \] The $a_i$ are pairwise coprime. If $\prod_i a_i^2>N$, multiply the square factors in a fixed order until their product first exceeds $N$. The resulting square $D$ lies between $N$ and $N^{1+u}$, so it has at most $N^{(1+u)/2}$ possible roots. Assign its complete prime powers to the $m$ offsets; the number of assignments is $m^{\omega(D)}=N^{o(1)}$. The CRT leaves at most one center modulo $D>N$ in the interval. If $\prod_i a_i^2\leq N$, start with this product and add the small prime factors from the least common multiple until the first crossing of $N$. Now $N\max(1/2,1-1/A)$, the number of such modules is at most \[ (NY)^\sigma\sum_{a\geq1}a^{-2\sigma} \prod_{\ell\leq Y}(1-\ell^{-\sigma})^{-1} =N^{\sigma+o(1)}. \] Choose $\sigma$ just above $1-1/A$. Assigning the complete prime powers again costs $N^{o(1)}$ and the CRT gives at most one center. This proves \eqref{p13:transfer:small-kernels-bound}. For the second assertion, use $H=C_DB$ as in Lemma~\ref{p13:half:small-kernels}. That lemma gives the result for $q>0$. For $q=0$, use $\mathcal X_0\subseteq\mathcal X_{\varepsilon/2}$ and apply the lemma with error margin $\varepsilon/2$. \end{proof} \section{Uniform independent support counts} \label{p1:uniform:appendix} \subsection{Full-support relations at all cardinalities} For $G\subseteq G_N$ and $r\ge2$, let $\mathcal H_{r,N}(G)$ consist of the ordered lists of $r$ distinct starts in $G$, with pairwise distances greater than $L$, for which there exist \[ s_i\in S_{x_i}\setminus\{0\},\qquad s_1+\cdots+s_r=0. \] The projections are nonzero in their effective square-class spaces. Minimality and affine compatibility are not required. \begin{proposition}[A uniform weighted count of full supports] \label{p1:uniform:full-support} For a fixed critical band, fixed $u>0$ and $\varepsilon>0$, uniformly in deterministic $G\subseteq G_N$ and integers $s\ge2$, \begin{equation} \sum_{r=s}^{|G|}\frac{u^r}{r!} \sum_{\boldsymbol x\in\mathcal H_{r,N}(G)}p^r \ll_{u,C,\varepsilon}N^{-\min\{s/2,2\}+\varepsilon}. \label{p2:uniform:full-support-bound} \end{equation} The scale threshold and implicit constant are independent of $s$ and $G$. Empty sums are zero. \end{proposition} \begin{proof} Fix $0<\eta<\min\{1,\varepsilon\}$. Lemma~\ref{p13:transfer:independent-tail}, applied with $v=u$ and $a=2-\eta$, gives a fixed cutoff. Increase it to $R\ge4$, still independent of $s,G$. On $G_N$ the exact marginals equal $p$, and each unordered family of $r$ distinct starts has exactly $r!$ orderings. Thus all terms with $r\ge R$, and any positive subfamily of them, have total mass at most $N^{-2+\eta}$. For each of the finitely many $2\le r0$, put $\pi_\mu(k)=e^{-\mu}\mu^k/k!$ when $k\ge0$ and $\pi_\mu(k)=0$ when $k<0$. Define \[ u_\pm=\mu+\frac12\pm\frac{\sqrt{4\mu+1}}2,\qquad a=\lfloor u_-\rfloor+1,\qquad b=\lceil u_+\rceil-1. \] Then \begin{equation} \|\Pi_\mu*\Delta^{*2}\|_1 =2\bigl\{\pi_\mu(a-1)-\pi_\mu(a-2) +\pi_\mu(b-1)-\pi_\mu(b)\bigr\}. \label{p1:discussion:finite-norm} \end{equation} More generally, for every integer $r\ge0$, \begin{equation} \|\Pi_\mu*\Delta^{*r}\|_1 \le\min\left\{2^r,\frac{\sqrt{r!}}{\mu^{r/2}}\right\}. \label{p1:discussion:general-norm} \end{equation} In particular, \begin{equation} \|\Pi_2*\Delta^{*2}\|_1=\frac{10}{3e^2}, \qquad \mathcal C(2)=\frac{5}{6e^2}. \label{p1:discussion:lambda-two} \end{equation} \end{proposition} The proof, including the normalization of the Charlier polynomials and the telescoping at integer roots, is given in Section~\ref{p14:comp:poisson-differences}. The function $\mathcal C$ is continuous and strictly positive on $(0,\infty)$. Indeed, coupling two Poisson variables by an independent increment gives $\|\Pi_\mu-\Pi_\nu\|_1\le2(1-e^{-|\mu-\nu|})$. Convolving with $\Delta^{*2}$ proves continuity; the mass at zero gives positivity, as in \eqref{p1:close:constant-bounds}. Thus the finite formula remains continuous when one of its two roots crosses an integer. For completeness, let $H_{\mu,\theta}=\Law(\Pois(\mu-\theta)+2\Pois(\theta/2))$, with independent summands and $0\le\theta<\mu$. The expansion of its convolution exponential gives the parameter-uniform estimate \begin{equation} \left|\TV(H_{\mu,\theta},\Pi_\mu) -\theta\mathcal C(\mu)\right| \le\theta^2e^{2\theta}. \label{p1:discussion:compound-coefficient} \end{equation} To verify the constant, write $H_{\mu,\theta} =\Pi_\mu*\exp_*(\theta\Delta^{*2}/2)$. The $\ell^1$ norm of the exponent is at most $2\theta$, so the remainder after its linear term has norm at most $2\theta^2e^{2\theta}$. The reverse triangle inequality and $\TV=\|\cdot\|_1/2$ prove \eqref{p1:discussion:compound-coefficient}. In particular, the factor $1/4$ in $\mathcal C$ consists of the factor $1/2$ in the quadratic correction and the factor $1/2$ in the total variation convention. Neither \eqref{p1:discussion:general-norm} nor the remainder constant in \eqref{p1:discussion:compound-coefficient} is asserted to be optimal. \subsection{Norms of Poisson differences} \label{p14:comp:poisson-differences} The following calculation proves Proposition~\ref{p1:discussion:charlier-norm}. It fixes both the signed-difference convention and the factor of two between $\ell^1$ distance and total variation. \begin{proof}[Proof of Proposition~\ref{p1:discussion:charlier-norm}] Write $v_k=(\Pi_\mu*\Delta^{*2})\{k\}$. Direct calculation gives \[ v_k=\pi_\mu(k)-2\pi_\mu(k-1)+\pi_\mu(k-2) =\frac{\pi_\mu(k)}{\mu^2}\bigl((k-\mu)^2-k\bigr). \] The quadratic has roots $u_-,u_+$ and is strictly negative exactly at the integers $a,\ldots,b$. These form a nonempty interval; integer roots contribute zero and are excluded by the definitions of $a,b$. Since $\sum_kv_k=0$, \[ \begin{aligned} \sum_k|v_k| &=-2\sum_{k=a}^b v_k\\ &=2\bigl\{\pi_\mu(a-1)-\pi_\mu(a-2) +\pi_\mu(b-1)-\pi_\mu(b)\bigr\}, \end{aligned} \] by telescoping. This proves \eqref{p1:discussion:finite-norm}. For $\mu=2$, the quadratic is $(k-1)(k-4)$: the negative masses are $v_2=-e^{-2}$ and $v_3=-2e^{-2}/3$. Their sum, multiplied by $-2$, gives \eqref{p1:discussion:lambda-two}. The same calculation at $\mu=1$ gives the value $3/e$ used in \eqref{p1:close:dyadic-example}. For the general bound, define the normalized Charlier polynomials \[ C_r(k;\mu)=\sum_{j=0}^r \binom rj(-1)^{r-j}\frac{(k)_j}{\mu^j}. \] Then $(\Pi_\mu*\Delta^{*r})\{k\}=\pi_\mu(k)C_r(k;\mu)$. Their exponential generating function is \[ \sum_{r\ge0}C_r(k;\mu)\frac{t^r}{r!} =e^{-t}(1+t/\mu)^k. \] For $P\sim\Pois(\mu)$, the expectation of the product of the generating functions with variables $s,t$ is \[ \E\!\left[e^{-s-t}(1+s/\mu)^P(1+t/\mu)^P\right] =e^{st/\mu}. \] Coefficient comparison, justified in a neighborhood of $(s,t)=(0,0)$ by the Poisson exponential moments, yields \[ \E[C_r(P;\mu)C_\ell(P;\mu)] =\ind_{\{r=\ell\}}\,\frac{r!}{\mu^r}. \] Cauchy--Schwarz now gives the second bound in \eqref{p1:discussion:general-norm}. The first follows from $\|\Pi_\mu\|_1=1$, $\|\Delta\|_1=2$, and submultiplicativity of the convolution norm. \end{proof} \section{Fixed moments do not give a polynomial rate} \label{p1:moment-obstruction} This example isolates the logical role of the growing-order estimate. It is a family of probability laws, with no claimed multiplicative realization. \begin{proposition}[A perturbation invisible at each fixed order] \label{p1:obstruction:fixed} For \(0\le a\le1\) put \[ Q_a=\Pi_1+\frac a2\Pi_1*\Delta^{*2}. \] For every integer \(m\ge3\), define \[ \epsilon_{m,a}=\frac{(1-a/2)e^{-1}}{2m!}, \qquad P_{m,a}=Q_a+\epsilon_{m,a}\Delta^{*m}. \] Then \(P_{m,a}\) is a probability measure satisfying \(\frac12Q_a\le P_{m,a}\le\frac32Q_a\) pointwise. Its factorial moments agree exactly with those of \(Q_a\) at every order other than \(m\), and \[ \TV(P_{m,a},Q_a) =\frac{(1-a/2)2^{m-2}}{e\,m!}. \] For every fixed \(t\ge0\), its exponential moment \(\E_{P_{m,a}}e^{tX}\) is bounded uniformly in \(m,a\). \end{proposition} \begin{proof} The second Charlier polynomial at mean one is \(k^2-3k+1\), whose minimum on \(\N\) is \(-1\). Thus \[ Q_a\{k\}\ge(1-a/2)e^{-1}/k!. \] For \(0\le k\le m\), \[ \frac{\epsilon_{m,a}\binom mk}{Q_a\{k\}} \le\frac1{2(m-k)!}\le\frac12 . \] Outside this range the perturbation is zero. Positivity and the two pointwise bounds follow. The perturbation has total mass zero, so \(P_{m,a}\) has mass one. The probability generating function of \(\Delta^{*m}\) is \((z-1)^m\). Its derivatives at \(z=1\) vanish at every order other than \(m\), where the derivative is \(m!\). This proves the factorial-moment statement. Its \(\ell^1\) norm is \(2^m\), yielding the exact total variation distance. Finally, \[ \E_{P_{m,a}}e^{tX} \le\frac32 \exp(e^t-1)\left(1+\frac12(e^t-1)^2\right), \] which is independent of \(m,a\). \end{proof} Take \(m_N=\lfloor\sqrt{\log N}\rfloor\) and \(a_N^\circ=(\log N)N^{-2/3}\), for \(N\) sufficiently large. At every fixed order \(r\), the factorial moments of \(P_{m_N,a_N^\circ}\) eventually equal \(1+\binom r2a_N^\circ\) exactly. Nevertheless, Stirling's formula gives \[ \log\TV(P_{m_N,a_N^\circ},Q_{a_N^\circ}) =-\tfrac12\sqrt{\log N}\log\log N +O(\sqrt{\log N}). \] This distance exceeds \(N^{-\gamma}\) for every fixed \(\gamma>0\), eventually. Fixed-order information, even with the displayed exponential-moment control, therefore cannot replace the uniform estimate on the boundary of the stopped expansion. \section*{Formal proofs and reproducibility} \addcontentsline{toc}{section}{Formal proofs and reproducibility} The accompanying \href{https://github.com/Vulkin-prog/paper-p1-lean}{Lean development} provides proofs for the 118 numbered results in the article and companion (26 and 92, respectively). Its \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/e564207ab543c7c0967e34b52f3d456d7a8f302c/docs/manuscript-results-publication-2026-11-12.json}{statement index} records the correspondence of clauses, definitions and inherited hypotheses, including the conditional premise of Theorem~\ref{p13:transfer:finite}. The documentation of the recorded replay is fixed at commit \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/a1c8092f7e25acf7f3b78283039cfbd69491f88e}{\texttt{a1c8092f7e25}}. The linked \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/a1c8092f7e25acf7f3b78283039cfbd69491f88e/docs/validation/v2.0/proof-source-manifest.json}{source manifest} and \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/a1c8092f7e25acf7f3b78283039cfbd69491f88e/docs/validation/v2.0/validation.json}{validation record} identify the proof source used for the recorded full replay as \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/aa12e797ce895f6b7aef6e4f1f0c6e5e9a58aab4}{\texttt{aa12e797ce89}}. That record specifies Lean~4.36.0-rc2 and Mathlib commit \href{https://github.com/leanprover-community/mathlib4/tree/9d581dfff2e41de989072ea63abf3daf69311520}{\texttt{9d581dfff2e4}}, and includes compilation results, extracted statement types and the kernel-axiom audit. The documentation snapshot and the replayed proof source have distinct roles. The correspondence between the written and formal statements was examined with AI assistance. It is not a machine translation of the written proofs or a substitute for independent human review. The conditional-sum and propagation clauses of Proposition~\ref{p1:fixed:propagation} allow homogeneous prescriptions; its small-contact clause uses the original run-start events. The formal interfaces distinguish these scopes as well. \Needspace{14\baselineskip} The final five-dossier allocation selects 42 formal statements covering every clause of the article's 26 numbered results. It preserves the four issued registrations and merges the two previously unregistered remainder scopes. The allocation is as follows, with numbers referring to the article: \begin{center} \small \begin{tabularx}{\linewidth}{@{}Xlr@{}} \toprule Dossier & Main results & Statements\\ \midrule Quadratic approximation and Poisson scales & 1.1--1.2, 7.2--7.4 & 10\\ Cumulants and optimality & 3.3--3.4, 6.1--6.2 & 8\\ Pair profiles and channel hierarchy & 3.1--3.2, 8.1--8.3 & 13\\ Stopping, conversion and reconstruction & 4.1--4.2, 5.1--5.3 & 8\\ Uniform support and positive reference limits & 2.1--2.5, 7.1, 8.4 & 3\\ \bottomrule \end{tabularx} \end{center} The last dossier uses three compound statements, retaining all thirteen remaining main-result clauses and directly selecting companion Propositions~T13.1 and~T15.1, with the latter's displayed sequence corollary. The former controls the weighted full-support sum at all cardinalities with one threshold before every subset and minimum size. The latter shows that fixed factorial moments and uniform exponential-moment control cannot replace growing-order estimates for a polynomial total-variation rate; no multiplicative realization of that counterexample is claimed. The other companion results supply proof dependencies, without being claimed as all directly registered. The \href{https://github.com/Vulkin-prog/paper-p1-lean/blob/eeda0e055bc16afa6bb6402727afe733ace9961e/palomar/REGISTRATION_PLAN_FINAL.json}{exact clause allocation} fixes these scopes. \Needspace{6\baselineskip} As of 13 November 2026, all five entries are published, in table order: \par\smallskip\noindent \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-12-000013\&version=1}{\texttt{PALOMAR-2026-11-12-000013}};\quad \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000001\&version=1}{\texttt{PALOMAR-2026-11-13-000001}};\\ \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000002\&version=1}{\texttt{PALOMAR-2026-11-13-000002}};\quad \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000008\&version=1}{\texttt{PALOMAR-2026-11-13-000008}};\\ \href{https://palomar-registry.org/entry.html?id=PALOMAR-2026-11-13-000020\&version=1}{\texttt{PALOMAR-2026-11-13-000020}}. The first entry fixes source \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/f9a2d213265a6833318745dd7f2d74c02090ab19}{\texttt{f9a2d213265a}}; entries 2--4 fix \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/d6a0f6c5109886127526677603eb6b74e8c0cd82}{\texttt{d6a0f6c51098}}; the final dossier fixes \href{https://github.com/Vulkin-prog/paper-p1-lean/tree/eeda0e055bc16afa6bb6402727afe733ace9961e}{\texttt{eeda0e055bc1}}. The final interface has a separate audit of 111 declarations and fresh statement comparison under the same Lean and Mathlib pins. Individual public entries provide the actual mechanical reports, rendered statements, preserved sources and published warnings. \Needspace{11\baselineskip} \section*{Statement on AI assistance} \addcontentsline{toc}{section}{Statement on AI assistance} ChatGPT (GPT-6 Astra) was used to organize the research material, develop and rewrite mathematical arguments, inspect references, and prepare the article and its technical companion. Parallel AI-assisted reviews examined selected identities, constants, proof dependencies and the exposition. The associated Lean development also used AI assistance for proof construction and statement comparison. AI-assisted review and machine checking are distinct from independent human verification. The author remains responsible for the mathematical claims, their correspondence with the formal statements, and any submission. \Needspace{6\baselineskip} % Generated from bibliography.tex by build.py. \begin{thebibliography}{99} \addcontentsline{toc}{section}{References} \bibitem{bkn2014} A.~D.~Barbour, E.~Kowalski, and A.~Nikeghbali, \emph{Mod-discrete expansions}, Probability Theory and Related Fields \textbf{158} (2014), 859--893. \href{https://doi.org/10.1007/s00440-013-0498-8} {doi:10.1007/s00440-013-0498-8}. Section numbers refer to the 26-page author text, Sections~5 and~8.1: \url{https://people.math.ethz.ch/~kowalski/mod-discrete.pdf}. \bibitem{binyaminicluckersnovikov2024} G.~Binyamini, R.~Cluckers, and D.~Novikov, \emph{Bounds for rational points on algebraic curves, optimal in the degree, and dimension growth}, International Mathematics Research Notices \textbf{2024} (2024), no.~11, 9256--9265. \href{https://doi.org/10.1093/imrn/rnae034} {doi:10.1093/imrn/rnae034}. The degree-uniform integer-point estimate is cited from \href{https://arxiv.org/abs/2302.04209v2} {arXiv:2302.04209v2} (20 September 2023), Theorem~2, equation~(27). \bibitem{bretwngxu2026} R.~de~la~Bret\`eche, V.~Y.~Wang, and M.~W.~Xu, \emph{Random multiplicative functions and making squares from polynomial values}, \href{https://arxiv.org/abs/2607.06398v1} {arXiv:2607.06398v1} (2026), Theorem~1.2 and Lemmas~6.2--6.3. \bibitem{erdosturk1984} P.~Erd\H{o}s and J.~Turk, \emph{Products of integers in short intervals}, Acta Arithmetica \textbf{44} (1984), no.~2, 147--174. \href{https://doi.org/10.4064/aa-44-2-147-174} {doi:10.4064/aa-44-2-147-174}. \href{https://www.renyi.hu/~p_erdos/1984-17.pdf} {Article text}. \bibitem{heathbrown2012} D.~R.~Heath-Brown, \emph{Square-free values of \(n^2+1\)}, Acta Arithmetica \textbf{155} (2012), no.~1, 1--13. \href{https://doi.org/10.4064/aa155-1-1} {doi:10.4064/aa155-1-1}. Preprint version: \href{https://arxiv.org/abs/1010.6217v3} {arXiv:1010.6217v3}, Lemma~2. \bibitem{munsch2026} M.~Munsch, I.~E.~Shparlinski, Y.-C.~Sun, and Y.~Xiao, \emph{A large sieve inequality for sums of Legendre symbols over short intervals}, \href{https://arxiv.org/abs/2604.23661v3} {arXiv:2604.23661v3} (5 August 2026), Lemma~3.3 and its proof. \bibitem{pila1995} J.~Pila, \emph{Density of integral and rational points on varieties}, in \emph{Columbia University Number Theory Seminar (New York, 1992)}, Ast\'erisque \textbf{228} (1995), 183--187. \href{https://www.numdam.org/item/AST_1995__228__183_0/} {Numdam: AST\_1995\_\_228\_\_183\_0}. \bibitem{reuss2014} T.~Reuss, \emph{Pairs of \(k\)-free numbers, consecutive square-full numbers}, \href{https://arxiv.org/abs/1212.3150v2} {arXiv:1212.3150v2} (19 March 2014), Sections~2.1--2.3. \bibitem{salberger2023} P.~Salberger, \emph{Counting integral points of affine hypersurfaces}, \href{https://arxiv.org/abs/2311.05561v1} {arXiv:2311.05561v1} (9 November 2023), Lemma~2. \bibitem{zacharovashwang2010} V.~Zacharovas and H.-K.~Hwang, \emph{A Charlier--Parseval approach to Poisson approximation and its applications}, Lithuanian Mathematical Journal \textbf{50} (2010), no.~1, 88--119. \href{https://doi.org/10.1007/s10986-010-9073-5} {doi:10.1007/s10986-010-9073-5}. Preprint version used for equation and remark numbers: \href{https://arxiv.org/abs/0810.4756v1} {arXiv:0810.4756v1}, (2.13) and Remark~4.4. \bibitem{griesmer1960} J.~H.~Griesmer, \emph{A bound for error-correcting codes}, IBM Journal of Research and Development \textbf{4} (1960), no.~5, 532--542. \href{https://doi.org/10.1147/rd.45.0532} {doi:10.1147/rd.45.0532}. \href{https://bitsavers.trailing-edge.com/pdf/ibm/IBM_Journal_of_Research_and_Development/045/ibmrd0405M.pdf} {Original article scan}, Section~2, equation~(2.10). \bibitem{cilleruelogaraev2011} J.~Cilleruelo and M.~Z.~Garaev, \emph{Concentration of points on two and three dimensional modular hyperbolas and applications}, Geometric and Functional Analysis \textbf{21} (2011), no.~4, 892--904. \href{https://doi.org/10.1007/s00039-011-0127-6} {doi:10.1007/s00039-011-0127-6}. Numbering refers to \href{https://arxiv.org/abs/1007.1526v2}{arXiv:1007.1526v2}, Section~3, Lemma~4 and Proposition~1. \bibitem{tao2026anatomy} T.~Tao, \emph{Products of consecutive integers with unusual anatomy}, \href{https://arxiv.org/abs/2603.27990v3}{arXiv:2603.27990v3} (26 September 2026), Lemma~2.10. \bibitem{shparlinskisleiman2026} I.~E.~Shparlinski and N.~Sleiman, \emph{Counting consecutive multiplicatively dependent triples}, \href{https://arxiv.org/abs/2609.29408v1} {arXiv:2609.29408v1} (24 September 2026), Theorem~1.1, Lemmas~2.1--2.2, and Section~3. \bibitem{binyaminicluckerskato2025} G.~Binyamini, R.~Cluckers, and F.~Kato, \emph{Sharp bounds for the number of rational points on algebraic curves and dimension growth, over all global fields}, Proceedings of the London Mathematical Society \textbf{130} (2025), no.~1, e70016. Published online 16 December 2024. \href{https://doi.org/10.1112/plms.70016} {doi:10.1112/plms.70016}. Corollary~8, equation~(23). \bibitem{bretkurlbergshparlinski2021} R.~de~la~Bret\`eche, P.~Kurlberg, and I.~E.~Shparlinski, \emph{On the number of products which form perfect powers and discriminants of multiquadratic extensions}, International Mathematics Research Notices \textbf{2021} (2021), no.~22, 17140--17169. Published online 30 December 2019. \href{https://doi.org/10.1093/imrn/rnz316} {doi:10.1093/imrn/rnz316}. \href{https://kurlberg.github.io/eprints/SquareProduct.pdf} {Author text}, Theorem~2.2 and Corollary~2.4. \end{thebibliography} \end{document}