\documentclass[11pt]{amsart} \usepackage{amsmath,amssymb,amsthm,array} \usepackage[hidelinks]{hyperref} \hypersetup{ pdftitle={Erdos problem 706 and uniform multiscale refinements}, pdfauthor={Will Howe, Grant Huffman, Jensen Kohlmeyer, Liam Kruer}, pdfsubject={An attributed proof of the OpenAI-Galan argument and uniform distance-density refinements}, pdfkeywords={Erdos problem 706, character rigidity, forbidden distances, independence density} } \setlength{\textwidth}{6.3in} \setlength{\oddsidemargin}{0.1in} \setlength{\evensidemargin}{0.1in} \setlength{\textheight}{8.7in} \setlength{\topmargin}{-0.1in} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newcommand{\R}{\mathbb R} \newcommand{\C}{\mathbb C} \newcommand{\Q}{\mathbb Q} \newcommand{\Z}{\mathbb Z} \newcommand{\N}{\mathbb N} \newcommand{\E}{\mathbb E} \newcommand{\cK}{\mathcal K} \newcommand{\cP}{\mathcal P} \newcommand{\dd}{\,d} \newcommand{\vol}{\operatorname{vol}} \newcommand{\diam}{\operatorname{diam}} \newcommand{\dist}{\operatorname{dist}} \newcommand{\supp}{\operatorname{supp}} \title[Erd\H{o}s 706 and multiscale refinements]{Erd\H{o}s problem 706 and uniform multiscale refinements} \author{Will Howe} \email{howew@purdue.edu} \author{Grant Huffman} \email{huffma97@purdue.edu} \author{Jensen Kohlmeyer} \email{jenwin@purdue.edu} \author{Liam Kruer} \email{kruerl@purdue.edu} \date{November 4, 2026. Revised author-review draft} \subjclass[2020]{52C10, 05C15, 37A30, 43A35, 90C25} \keywords{Forbidden distances, chromatic number, character rigidity, compact extensions, complete positivity} \begin{document} \begin{abstract} We present a detailed, attributed exposition of the OpenAI--Gal\'an argument giving exponential growth in Erd\H{o}s problem 706. The specialized character-rigidity proof is included, together with its compact-tower and radial arguments, the transfer from arbitrary colorings to weak measurable colorings for algebraic distances, and Gal\'an's deduction using Bukh's separated-distance theorem. This yields finite planar graphs with $r$ forbidden distances and chromatic number greater than $4.0485^r$. The full negative answer to problem 706 is credited to that prior work. We then prove a uniform product estimate for arbitrary nonempty compact distance clusters in a fixed normalization band, using one separation ratio for every number and every choice of clusters. For a fixed finite weighted unit-distance graph of independence ratio at most $a$, we also obtain all-length exponential bounds of every base $b<1/a$ with sufficient separation $O((1-ab)^{-5/2})$. The latter results are presented as refinements, with their relationship to earlier Fourier, complete-positivity, and finite-graph methods made explicit. \end{abstract} \maketitle \raggedbottom \setlength{\emergencystretch}{1em} \begingroup\small\tableofcontents\endgroup \section{The problem, attribution, and organization}\label{in:intro} For a finite set $P\subset\R^2$ and a finite set $D\subset(0,\infty)$, let $G(P,D)$ join distinct points $x,y\in P$ when $|x-y|\in D$. Put \[ L(r)=\sup_{\substack{P\subset\R^2\ { \rm finite}\\ D\subset(0,\infty),\ |D|=r}} \chi(G(P,D)),\qquad r\ge1. \] Erd\H{o}s problem 706 asks whether the chromatic numbers of these graphs have a polynomial upper bound in the number $r$ of permitted edge lengths. We use the formulation recorded in \cite{Erdos706,Erdos1981}. The argument in Part~I establishes \begin{equation}\label{in:full706} 4.0485^r{\raggedright\arraybackslash}p{0.58\textwidth}>{\raggedright\arraybackslash}p{0.32\textwidth}} \hline Mathematical component & Attribution\\ \hline Character rigidity, relative singularity, compact towers, conditional averages, and radial identities & OpenAI \cite{OpenAI2026}\\ Continuous-factor transfer for the unit distance & OpenAI \cite{OpenAI2026}\\ Algebraic-distance extension and exponential solution of problem 706 & Sergi Gal\'an \cite{Galan2026}\\ Separated-scale density product limit & Bukh \cite{Bukh2008}\\ The bound $m(\{1\})\le0.2470$ & Ambrus et al.\ \cite{ACMVZ2023}\\ Uniform compact-cluster estimate and graph-dependent separation bounds & Refinements developed in Part~II\\ \hline \end{tabular} \end{center} The complete solution of problem 706 is prior work in the attribution above. Our contribution is the exposition of that proof chain and the refinements in Part~II. The proof of the prior result has been rewritten with explicit hypotheses and intermediate estimates. The original sources retain credit for the mathematical ideas. \subsection{The additional uniformity} For $M\ge1$, let $\cK_M$ be the nonempty compact subsets of $[1,M]$. Theorem~\ref{thm:clusters} has the order of quantifiers \[ \forall M\ge1\ \forall\varepsilon\in(0,1/2)\ \exists Q>M \quad\forall r\ge1\quad\forall K_1,\ldots,K_r\in\cK_M. \] When $t_{i+1}/t_i\ge Q$, it gives \[ \prod_i m(K_i)\le m\!\left(\bigcup_i t_iK_i\right) <\prod_i(m(K_i)+\varepsilon). \] Thus one ratio is chosen before both the length and the cluster shapes. The proof uses the complete-positivity description of independence density \cite{DOV2022}, finite certificate extraction \cite{DOV2022,Davies2026}, and a uniform finite cover of $\cK_M$. Theorem~\ref{thm:graph} supplies an effective companion from a finite weighted graph, with separation $O((1-ab)^{-5/2})$ for each fixed weighted realization. It uses established weighted quadratic inequalities \cite{GHPR1997,LaurentVargas2022}. Earlier separated-scale arguments of Bukh, de Oliveira Filho and Vallentin, DeCorte--de Oliveira Filho--Vallentin, and Castro-Silva are discussed at the start of Part~II \cite{Bukh2008,OFV2010,DOV2022,CastroSilva2023}. Gal\'an's explicit two-distance estimates are an additional quantitative predecessor \cite{Galan2026}. The exact all-length and all-cluster uniformity is the refinement asserted here. Historical priority for that refinement remains provisional because the full texts of \cite{Szekely1984} and \cite{SzekelyWormald1989} were unavailable in the literature audit. Their bibliographic records and later descriptions were checked. \part{The OpenAI--Gal\'an argument for problem 706} % BEGIN attributed rewritten module rigidity_spectral.tex \section{Character laws and relative singularity} \label{sp:section} The results and the proof strategy in this section are due to OpenAI~\cite{OpenAI2026}. We give their spectral argument in full, including the conditional-measure constructions and the order of the finite choices in the relative-singularity argument. The additional explanations concern versions of conditional probabilities and quantitative estimates. \subsection{The algebraic plane and its character group} Identify the real plane with \(\C\), and put \[ F=\overline{\Q}\cap\R,\qquad E=F(i),\qquad K=\{u\in E:u\overline u=1\}. \] Both the additive group \(E\) and the multiplicative group \(K\) will carry the discrete topology. Write \(\mathbb T=\{z\in\C:|z|=1\}\). The compact character group of the discrete additive group \(E\) is \[ \mathbb D=\widehat{E_{\mathrm{disc}}} =\{d:E\to\mathbb T:d(z+w)=d(z)d(w)\}. \] Its operation is written additively, so \[ (d+d')(z)=d(z)d'(z),\qquad (-d)(z)=\overline{d(z)}. \] For \(a,z\in E\), define \[ (ad)(z)=d(az),\qquad \psi_z(d)=d(z). \] In particular, \(K\) acts on \(\mathbb D\) by continuous automorphisms. The ordinary continuous characters of the real plane give the subgroup \begin{equation} \label{sp:ordinary} C=j(\R^2),\qquad j(\xi)(z)=\exp(i\xi\cdot z), \end{equation} where \(z\) is viewed as a real coordinate vector in the dot product. \begin{lemma}[OpenAI~\cite{OpenAI2026}] \label{sp:borel} The group \(K\) is countable and divisible, and \(\mathbb D\) is compact metrizable. The map \(j\) is a continuous injection. Its image \(C\) is a Borel subgroup, and \(j^{-1}:C\to\R^2\) is Borel. Every nonzero scalar \(a\in E\) maps \(C\) onto itself. On the frequency parameters this action is the transpose of the real-linear map of multiplication by \(a\), which is orthogonal when \(a\in K\). \end{lemma} \begin{proof} The fields \(F\) and \(E\) are countable. If \(n\) is a positive integer and \(u\in K\), a complex root \(v^n=u\) is algebraic and has modulus one. Its real and imaginary parts belong to \(F\), so \(v\in K\). Inversion handles negative powers. Every nonzero power map on \(K\) is therefore surjective. The group \(\mathbb D\) is a closed subspace of the countable product \(\mathbb T^E\), defined by the character equations. It is consequently compact metrizable. Continuity of \(j\) follows coordinate by coordinate. If \(j(\xi)=j(\eta)\), the continuous plane character with frequency \(\xi-\eta\) is trivial on the dense subgroup \(\Q+i\Q\subset E\). It is trivial on the whole plane, so \(\xi=\eta\). For each positive integer \(n\), the restriction of \(j\) to the closed ball \(\overline B(0,n)\subset\R^2\) is a homeomorphism onto its compact image. Hence \[ C=\bigcup_{n\geq1}j(\overline B(0,n)) \] is Borel. The inverse is continuous on each of these compact images, which proves its Borel measurability. Let \(M_a\) be the real matrix of multiplication by \(a\). Directly from the definitions, \[ a\,j(\xi)=j(M_a^{\mathsf T}\xi). \] For \(a\ne0\), the inverse scalar \(a^{-1}\) belongs to \(E\). This proves the assertion about \(C\). If \(|a|=1\), the matrix \(M_a\), and therefore its transpose, is orthogonal. \end{proof} \begin{definition}[OpenAI~\cite{OpenAI2026}] \label{sp:wild} A \(K\)-invariant Borel probability measure \(\nu\) on \(\mathbb D\) is \emph{wild} if \(\nu(C)=0\). Its Fourier coefficients are \[ f_\nu(z)=\int_{\mathbb D}\psi_z(d)\,\dd\nu(d), \qquad z\in E. \] \end{definition} The theorem whose proof is developed here and in the following compact-factor arguments is the following result of OpenAI. \begin{theorem}[Haar rigidity, OpenAI~\cite{OpenAI2026}] \label{sp:rigidity} Every wild probability measure on \(\mathbb D\) is the normalized Haar measure \(m_{\mathbb D}\). Equivalently, if a \(K\)-invariant probability measure \(\nu\) satisfies \(\nu(C)=0\), then \[ f_\nu(z)=0\qquad(z\in E\setminus\{0\}). \] \end{theorem} The equivalence uses uniqueness of finite measures from their Fourier coefficients on \(\mathbb D\). Indeed, the complex linear span of the coordinate characters \(\psi_z\) is a unital conjugation-closed algebra separating points, so Stone--Weierstrass gives uniform density in the continuous functions. Haar measure has coefficient zero at every nonzero \(z\). To see that \(\psi_z\) is nontrivial there, choose an ordinary plane frequency \(\xi\) with \(\exp(i\xi\cdot z)\ne1\). \subsection{Cosets and line-continuity events} \begin{lemma}[OpenAI~\cite{OpenAI2026}] \label{sp:null} Let \(\nu\) be wild. Every coset of \(C\) has \(\nu\)-measure zero. The difference of two independent samples of law \(\nu\) also gives \(C\) measure zero. For every \(a\in E\setminus\{0\}\), the set \[ L_a=\{d\in\mathbb D:t\mapsto d(at) \text{ is continuous on }F\text{ in its real topology}\} \] is Borel and has \(\nu\)-measure zero. \end{lemma} \begin{proof} Suppose \(\nu(d+C)>0\). Distinct cosets in its \(K\)-orbit are disjoint and have equal positive mass, so this orbit is finite. The action on the orbit has a finite abelian image \(A\), since \(K\) is abelian. Divisibility of \(K\) makes every power map on \(A\) surjective. Taking the power \(|A|\) shows \(A\) is trivial. Thus \((u-1)d\in C\) for every \(u\in K\). Choose \(u\ne1\). Multiplication by \((u-1)^{-1}\in E\) preserves \(C\), so \(d\in C\), contradicting \(\nu(C)=0\). The relation \(d'-d\in C\) is Borel. Its sections are cosets of \(C\) of measure zero. Integrating the sections proves the independent difference assertion. Restriction to \(aF\) is a continuous map \[ R_a:\mathbb D\to\widehat{F_{\mathrm{disc}}}, \qquad R_a(d)(t)=d(at). \] A continuous character on \(F\) is uniformly continuous and extends uniquely to a continuous character on \(\R\). It therefore has the form \(t\mapsto e^{i\lambda t}\). To obtain this form, choose a continuous argument near zero. The character equation makes that argument additive on a smaller interval, hence linear by continuity. Subdividing an arbitrary real input extends the formula to all of \(\R\). Conversely each such character restricts continuously to \(F\). The map \(\lambda\mapsto(t\mapsto e^{i\lambda t})\) is continuous and injective into \(\widehat{F_{\mathrm{disc}}}\). Compact exhaustion, as in Lemma~\ref{sp:borel}, makes its image Borel. Consequently \(L_a\) is Borel. If \(aF\) and \(bF\) are nonparallel, \(a,b\) form an \(F\)-basis of \(E\). A character continuous on both lines satisfies \[ d(at+bs)=e^{i(\lambda t+\eta s)} \qquad(t,s\in F) \] for some real \(\lambda,\eta\). The corresponding real-linear coordinate change is invertible, so this character belongs to \(C\). Hence \(L_a\cap L_b\subset C\). There are infinitely many distinct unoriented lines \(uaF\), \(u\in K\). For example, the elements \((1+in)/(1-in)\), for positive integers \(n\), have distinct arguments modulo \(\pi\). The corresponding events all have the same \(\nu\)-measure by invariance. Their pairwise intersections have measure zero. Summing their measures over an arbitrarily large finite collection forces the common measure to be zero. \end{proof} \subsection{Integration along an algebraic line} Fix a positive normalized translation-invariant mean \(\mathfrak m\) on the bounded functions on the discrete group \(F/\Z\). Such a mean exists because the group is abelian \cite[Theorem~G.2.1]{BHV2008}. For a bounded function \(h:F\to\C\) supported in a bounded interval, define \begin{equation} \label{sp:integral} I(h)=\mathfrak m\left(t+\Z\longmapsto \sum_{k\in\Z}h(t+k)\right). \end{equation} \begin{lemma}[Invariant line integration, OpenAI~\cite{OpenAI2026}] \label{sp:integral-properties} The functional \(I\) is well defined, positive, linear, and invariant under translations by \(F\). For \(\phi\in C_c(\R)\), \[ I(\phi|_F)=\int_\R\phi(t)\,\dd t. \] If \(V:F\to\mathcal H\) is bounded and supported in a bounded interval, where \(\mathcal H\) is a Hilbert space, a weak integral \(I(V)\) is well defined and satisfies \begin{equation} \label{sp:vector-integral-bound} \|I(V)\|_{\mathcal H}\leq I(\|V(t)\|_{\mathcal H}). \end{equation} Bounded linear operators commute with this weak integral. \end{lemma} \begin{proof} Changing a representative of \(t+\Z\) merely reindexes the sum in \eqref{sp:integral}. If \(h\) is supported in \([-R,R]\), at most \(2R+2\) summands are nonzero at any point. The periodization is therefore bounded. Positivity and linearity follow from the mean. Periodizing \(h(\cdot+s)\) translates the periodization by \(s+\Z\), so \(I\) is translation invariant. For a nonzero integer \(n\), the ordinary circle character \(t+\Z\mapsto e^{2\pi int}\) has mean zero. Indeed a rational translation can multiply it by a scalar different from one. The constant character has mean one. Uniform approximation by trigonometric polynomials shows that \(\mathfrak m\) agrees with circle integration on continuous periodic functions restricted to \(F\). The periodization of \(\phi\in C_c(\R)\) is continuous and periodic. Thus \[ I(\phi|_F) =\int_0^1\sum_{k\in\Z}\phi(t+k)\,\dd t =\int_\R\phi(t)\,\dd t. \] For the Hilbert-valued assertion, use inner products linear in the first variable. Positivity of \(I\) gives, for every \(v\in\mathcal H\), \[ \left|I(\langle V(t),v\rangle)\right| \leq I(|\langle V(t),v\rangle|) \leq \|v\|_{\mathcal H}I(\|V(t)\|_{\mathcal H}). \] Hilbert-space duality gives the vector \(I(V)\) and \eqref{sp:vector-integral-bound}. Scalar pairing also proves commutation with bounded linear operators and translation invariance for these integrals. \end{proof} This construction applies to all the bounded functions in question. Ordinary real-variable measurability is unnecessary. \begin{lemma}[Vanishing line average, OpenAI~\cite{OpenAI2026}] \label{sp:line} If \(\nu\) is wild, then \begin{equation} \label{sp:line-vanishing} I\bigl(\phi(t)f_\nu(at)\bigr)=0 \qquad(a\in E\setminus\{0\},\ \phi\in C_c(\R)). \end{equation} \end{lemma} \begin{proof} On \(L^2(\nu)\), let \(T_tv=\psi_{at}v\), \(t\in F\). We first show that the only vector whose orbit is norm-continuous in the real topology on \(F\) is zero. Suppose \(v\) is such a vector and put \[ c_v(t)=\int_{\mathbb D}|v|^2\psi_{at}\,\dd\nu. \] This function is positive definite. Moreover, \[ |c_v(t+s)-c_v(t)| \leq\|v\|_2\|T_sv-v\|_2, \] so it is uniformly continuous on \(F\). It extends to a continuous function on \(\R\). Approximating any finite list of real arguments by elements of \(F\) shows that the extension remains positive definite. Bochner's theorem on \(\R\) gives a finite positive frequency measure representing this extension \cite[Theorem~D.2.2]{BHV2008}. Push that measure to \(\widehat{F_{\mathrm{disc}}}\) using \(\lambda\mapsto(t\mapsto e^{i\lambda t})\). Its Fourier coefficients agree with those of \(R_{a*}(|v|^2\nu)\). Coordinate characters separate points, and their complex linear span is a self-conjugate algebra, so Stone--Weierstrass and uniqueness of finite measures identify these two measures. The first is concentrated on the Borel set of continuous characters. Consequently \(|v|^2\nu\) is concentrated on \(L_a\). Lemma~\ref{sp:null} gives \(\nu(L_a)=0\), so \(v=0\). Now form the weak integral \[ V_\phi=I\bigl(\phi(t)T_t1\bigr)\in L^2(\nu). \] Commutation with \(T_s\) and translation invariance give \[ T_sV_\phi-V_\phi =I\bigl((\phi(t-s)-\phi(t))T_t1\bigr). \] Since \(\|T_t1\|_2=1\), Lemma~\ref{sp:integral-properties} yields \[ \|T_sV_\phi-V_\phi\|_2 \leq I(|\phi(\cdot-s)-\phi|) =\int_\R|\phi(t-s)-\phi(t)|\,\dd t \longrightarrow0 \] as \(s\in F\) tends to zero. Thus \(V_\phi=0\) by the preceding argument. Pairing with the constant function proves \eqref{sp:line-vanishing}. \end{proof} \subsection{A measurable center without moment assumptions} The subgroup \(C\) need not be closed. A probability measure on one of its cosets nevertheless allows an equivariant choice of a center. \begin{lemma}[Measurable center, OpenAI~\cite{OpenAI2026}] \label{sp:center} For every Borel probability measure \(\lambda\) on \(\R^2\), the function \begin{equation} \label{sp:center-objective} J_\lambda(s)=\int_{\R^2} \left(\sqrt{1+|s-t|^2}-\sqrt{1+|t|^2}\right)\,\dd\lambda(t) \end{equation} has a unique minimizer \(L(\lambda)\). The map \(L\) is Borel on the space of probability measures with its usual weak Borel structure. It is equivariant under translations and orthogonal maps. \end{lemma} \begin{proof} The function \(g(u)=\sqrt{1+|u|^2}\) is 1-Lipschitz, so the integrand in \eqref{sp:center-objective} has absolute value at most \(|s|\). The integral is finite for every \(s\), without any moment assumption, and is Lipschitz in \(s\). The Hessian of \(g\) is positive definite. Its eigenvalues are \((1+|u|^2)^{-1/2}\) and \((1+|u|^2)^{-3/2}\). Hence \(J_\lambda\) is strictly convex. Choose \(R>0\) with \(p=\lambda(\overline B(0,R))>1/2\). On this ball the integrand is at least \(|s|-R-\sqrt{1+R^2}\), and on its complement it is at least \(-|s|\). It follows that \[ J_\lambda(s)\geq (2p-1)|s|-p(R+\sqrt{1+R^2})\longrightarrow\infty. \] Continuity and coercivity give a minimum, and strict convexity gives uniqueness. For fixed \(s\), the integrand is bounded and continuous in \(t\). Thus \(\lambda\mapsto J_\lambda(s)\) is Borel. By continuity in \(s\), the minimum equals the infimum over \(\Q^2\). Enumerate \(\Q^2\) and let \(q_n(\lambda)\) be its first element with objective value less than this infimum plus \(1/n\). The maps \(q_n\) are Borel. For fixed \(\lambda\), their tail lies in a compact sublevel set. Every cluster point is a minimizer, so the entire sequence converges to \(L(\lambda)\). This proves that \(L\) is Borel. If \(\lambda'\) is the law of \(t-r\) for \(t\) of law \(\lambda\), an identity between bounded integrands gives \[ J_{\lambda'}(s)=J_\lambda(s+r)-J_\lambda(r). \] Therefore \(L(\lambda')=L(\lambda)-r\). For an orthogonal map \(O\), \[ J_{O_*\lambda}(Os)=J_\lambda(s), \] which gives \(L(O_*\lambda)=OL(\lambda)\). \end{proof} \subsection{Conditional measures and compact approximation} All systems in the rest of this section are standard probability spaces with probability-preserving actions of the countable group \(K\). Factors are countably generated modulo null sets. We use standard Borel models for them, identify factor functions with their pullbacks, and write \(U_kH(x)=H(kx)\). Countability allows all group laws and equivariance identities used for these models to hold on common invariant conull sets. For an equivariant factor map \(\pi:X\to Y\), disintegration gives a measurable kernel \(y\mapsto\mu_y\), concentrated on \(\pi^{-1}(y)\) for almost every \(y\), with \[ \mu=\int_Y\mu_y\,\dd\mu_Y(y). \] We use existence and almost-everywhere uniqueness of disintegration in this standard Borel setting \cite[Theorem~4 and Remark~5]{Tao2008}. Uniqueness implies \[ \mu_{ky}=k_*\mu_y \] for almost every \(y\), for each \(k\in K\). Equality here is equality of probability measures, obtained by testing a countable generating algebra. Countability of \(K\) gives a common invariant conull set. On this set the equality holds for all Borel tests at once. We also use the composition rule for conditional laws. If \(W\subset Y\subset X\), first conditioning on \(Y\) and then integrating over \(Y\) conditioned on \(W\) gives the conditional law over \(W\). This follows from the tower property, with equality of probability measures fixed on a common conull set by a countable generating algebra. \begin{definition}[Compact approximation property, OpenAI~\cite{OpenAI2026}] \label{sp:tower} Let \((Y_\alpha)_{\alpha\leq\theta}\) be a tower of factors of an ergodic system, where \(\theta\) is a countable ordinal. The initial factor \(Y_0\) is trivial. At a limit ordinal, take the factor generated by all earlier factors. At every successor, with \(Y=Y_{\alpha+1}\) and \(W=Y_\alpha\), require that for every bounded complex-valued \(H\) on \(Y\) and every \(\varepsilon>0\), there are bounded functions \(H_1,\ldots,H_q\) on \(Y\) and a finite \(B\) such that, for every \(k\in K\), there are \(b_1^k,\ldots,b_q^k\in L^\infty(W)\) satisfying \begin{equation} \label{sp:tower-property} \left\|U_kH-\sum_{r=1}^q b_r^kH_r\right\|_{L^2(Y)} <\varepsilon, \qquad \|b_r^k\|_\infty\leq B. \end{equation} The functions \(H_r\) and the bound \(B\) are chosen before \(k\). \end{definition} The compact-factor construction later supplies a tower with precisely this property. The next lemma uses only \eqref{sp:tower-property} and the preceding spectral results. \begin{lemma}[Relative singularity, OpenAI~\cite{OpenAI2026}] \label{sp:relative} Let \(X\) be an ergodic standard probability-preserving \(K\)-system, and let \(d:X\to\mathbb D\) be an equivariant measurable map with wild law. For every tower from Definition~\ref{sp:tower} and every factor \(Y\) in that tower, let \(\nu_Y\) be the law of \[ d(x)-d(x') \] for two conditionally independent samples of \(X\) over \(Y\). Then \(\nu_Y(C)=0\). \end{lemma} \begin{proof} All the relative difference laws are \(K\)-invariant by equivariance of their conditional kernels. We prove the assertion by transfinite induction in a universal form. At each ordinal it is asserted for all ambient ergodic standard systems, all equivariant maps with wild law, and all towers satisfying \eqref{sp:tower-property}. This permits the earlier induction assertion to be used on a new factor system. At the trivial factor, Lemma~\ref{sp:null} proves the assertion. Fix a noninitial ordinal, assume this universal assertion at every earlier ordinal, and suppose it fails at its stage \(Y\) in some \(X\). Let \(\kappa_y\) denote the conditional law of \(d\) over \(Y\). \paragraph{Finite cosets and equivariant centers.} The function \[ r(y,z)=\kappa_y(z+C) =\int_{\mathbb D}\mathbf1_C(d'-z)\,\dd\kappa_y(d') \] is Borel in \((y,z)\) by Borel parameter integration. Kernel equivariance gives \(r(ky,kz)=r(y,z)\) for every \(z\) on a common conull set of \(y\). Thus \(r(y(x),d(x))\) is an invariant function on \(X\), with integral \(\nu_Y(C)>0\). Ergodicity makes it a constant \(a>0\) almost surely. For almost every \(y\), therefore, \(\kappa_y\)-almost every \(z\) belongs to a coset of mass \(a\). Any positive-mass coset meets this full-measure set and consequently has mass \(a\). There are finitely many such cosets and they exhaust the measure. Their number \(m\) satisfies \(ma=1\), so it is independent of \(y\). Also \(\kappa_y(C)=0\) almost surely because the marginal law of \(d\) is wild. On the Borel set \(r(y,z)>0\), define a probability kernel on \(\R^2\) by \[ \lambda_{y,z}(A)= \frac1{r(y,z)} \int_{\mathbb D} \mathbf1_C(d'-z)\, \mathbf1_A\bigl(j^{-1}(d'-z)\bigr)\,\dd\kappa_y(d'). \] The integrand is set to zero off \(d'-z\in C\). Lemma~\ref{sp:borel} and parameter integration make this a Borel kernel. Set \[ b(y,z)=z+j\bigl(L(\lambda_{y,z})\bigr). \] If \(z\) is replaced by \(z+j(r_0)\), the kernel is translated by \(-r_0\). Lemma~\ref{sp:center} shows that \(b(y,z)\) is unchanged. Thus it is constant on each positive-mass coset and lies in that coset. Let \(O_k\) be the orthogonal action of \(k\) on frequency parameters. Then \(\lambda_{ky,kz}=(O_k)_*\lambda_{y,z}\), which gives \[ b(ky,kz)=k\,b(y,z). \] Pull this function back along \(x\mapsto(y(x),d(x))\), which lies in its domain almost surely, and extend it arbitrarily on the null complement. Let \(Z\) be the factor generated by \(Y\) and this map \(b\). It is standard and ergodic. Over \(Y\), the conditional law \(\beta_y\) of \(b\) is uniform on the \(m\) distinct centers of the positive-mass cosets. Since \(b(x)\in d(x)+C\), its marginal law is wild. It is equivariant by construction. \paragraph{Encoding the finite phase lists.} For each \(y\), take the unordered list of \(m\) values of \(\psi_1\) at the centers, with repetitions when phases coincide. Encode the list by its elementary symmetric functions, giving \(H(y)\in\C^m\). This is a bounded measurable vector. Indeed, its power sums are \[ m\int_{\mathbb D}\psi_1(b)^n\,\dd\beta_y(b), \qquad 1\leq n\leq m, \] and Newton's identities recover the elementary symmetric functions. The encoding is a continuous injection of the compact space \(\mathbb T^m/S_m\). Give this space the matching metric \[ d_{\mathrm{match}}([z_1,\ldots,z_m],[z'_1,\ldots,z'_m]) =\min_{\pi\in S_m}\max_r|z_r-z'_{\pi(r)}|. \] Its inverse encoding is uniformly continuous. Equivariance of \(\beta_y\) shows that \(U_kH\) encodes the list at frequency \(k\). Choose \(\rho>0\) such that the closed arc \(\{|z-1|\leq5\rho\}\) has circle Haar probability less than \(1/(16m^2)\). There are \(N\geq1\) and \(c>0\) such that every probability \(\omega\) on \(\mathbb T\) satisfying \[ \left|\int z^n\,\dd\omega(z)\right|0\) such that encoding vectors at distance less than \(s\) have phase lists matching within \(\rho\). Put \begin{equation} \label{sp:error-parameters} \tau=s/100,\qquad \gamma>0,\qquad \frac{2\gamma^2}{\tau^2}<10^{-4}. \end{equation} \paragraph{An approximation over an earlier factor.} There is an earlier factor \(W\) and fixed bounded vector-valued functions \(H_1,\ldots,H_q\) on \(Y\) such that, for every \(k\in K\), \begin{equation} \label{sp:vector-approx} \left\|U_kH-\sum_{r=1}^q b_r^kH_r\right\|_2<\gamma, \qquad b_r^k\in L^\infty(W),\quad \|b_r^k\|_\infty\leq B, \end{equation} for one fixed finite \(B\). At a successor stage, apply \eqref{sp:tower-property} to each coordinate and combine the functions, taking coordinate errors whose squares sum to less than \(\gamma^2\). At a limit stage, first approximate \(H\) within \(\gamma/2\) by a bounded vector on an earlier successor factor. Such approximation follows from convergence in \(L^2\) of conditional expectations onto an increasing cofinal family of earlier factors. The approximation error is unchanged by rotation. Apply \eqref{sp:tower-property} at that successor with error below \(\gamma/2\), and let \(W\) be its predecessor. In either case all data in \eqref{sp:vector-approx} are fixed before \(k\) is chosen. The tower through \(W\) is also a tower of factors of \(Z\). Its measures, joins, and approximation properties are the same. The earlier universal induction assertion, applied in \(Z\) to \(b\), therefore says that the relative difference law of \(b\) over \(W\) is wild. Its Fourier coefficient at \(a\in E\) is \begin{equation} \label{sp:conditional-coefficient} p(a)=\left\|\E_Z(\psi_a(b)\mid W)\right\|_2^2\geq0, \qquad p(0)=1. \end{equation} This follows by conditional independence. Lemma~\ref{sp:line} applies to \(p\). \paragraph{Finite bins before a small algebraic parameter.} The two coefficient arrays for two instances of \eqref{sp:vector-approx} lie in a fixed bounded subset of \(\C^{2q}\). Partition it into finitely many Borel bins, say \(L\) bins, with coordinate diameters smaller than \[ \frac{\tau}{1+\sum_{r=1}^q\|H_r\|_\infty}. \] At any fiber where the essential bounds hold, two pairs of arrays in one bin give approximating vectors differing pointwise by at most \(\tau\) in each of the two instances. Choose an integer \(M\geq2\) satisfying \begin{equation} \label{sp:many-tests} \frac{0.9M}{L}>10m^2. \end{equation} All approximation data and this number of tests have now been fixed. Set \(Q=N\binom M2\), and choose \[ \varepsilon_0=\min\left\{\frac12,\frac{c^2}{200Q}\right\}. \] There is \(t\in F\), \(0<|t|<1/M\), such that \begin{equation} \label{sp:small-tests} p(n(j-l)t)<\varepsilon_0 \qquad(1\leq n\leq N,\ 1\leq j0\), positivity would contradict \(I(\phi)=\int\phi>0\). Some such \(t\) makes the sum less than \(\varepsilon_0\), and hence satisfies \eqref{sp:small-tests}. It is nonzero because the sum at zero is \(Q\geq1\). Put \(x_j=jt\). By Markov's inequality and \eqref{sp:conditional-coefficient}, the probability that any one of \[ \left|\E_Z\bigl(\psi_{n(x_j-x_l)}(b)\mid W=w\bigr)\right|10m^2\). Fix \(j_0\in J\). For each \(j\in J\), the conditional \(L^2\) distance from \(U_{k_j}H\) to \(U_{k_{j_0}}H\) is less than \(3\tau\). The two approximation errors each contribute less than \(\tau\), and the bin contributes at most \(\tau\). The same holds with \(k_j,k_{j_0}\) replaced by \(k'_j,k'_{j_0}\). For uniform \(j\in J\) and \(y\) sampled over \(W=w\), Markov's inequality shows that the probability that either encoding distance is at least \(s\) is at most \[ \frac{18\tau^2}{s^2}=0.0018. \] Let \(q(y)\) be the fraction of \(j\in J\) for which both phase lists match their \(j_0\) lists within \(\rho\). The matching metric is measurable, and the choice of \(s\) gives \begin{equation} \label{sp:matching-fraction} \E(q\mid W=w)\geq1-0.0018>0.95. \end{equation} \paragraph{The collision contradiction.} Fix such a \(y\), and let \(b_0\) be any center in the finite support of \(\beta_y\). For every matching \(j\), each of \(\psi_{k_j}(b_0)\) and \(\psi_{k'_j}(b_0)\) is within \(\rho\) of an entry in its representative list for \(j_0\). It follows that \[ \psi_{x_j}(b_0)=\psi_{k_j}(b_0)\psi_{k'_j}(b_0) \] is within \(2\rho\) of one of at most \(m^2\) products of entries from those two lists. The two matchings may select different centers, and all \(m^2\) products are allowed. Assign the \(q(y)|J|\) matching indices to such products. If their class sizes are \(n_1,\ldots,n_{m^2}\), with zero classes included as necessary, then \[ \sum_r n_r^2\geq\frac{q(y)^2|J|^2}{m^2}. \] These count ordered pairs in common classes, including diagonals. Removing diagonals loses at most \(|J|\) pairs. Hence independent uniform \(j,l\in J\) satisfy \[ \mathbb P_{j,l}\bigl( j\ne l,\ |\psi_{x_j}(b_0)-\psi_{x_l}(b_0)|\leq4\rho \bigr) \geq\frac{q(y)^2}{m^2}-\frac1{|J|}. \] This bound holds for every \(b_0\) in the finite support. Integrate first over \(\beta_y\), then over the conditional law of \(y\) given \(w\). The composition rule for conditional laws identifies this with conditioning in \(Z\) over \(W=w\). Jensen's inequality and \eqref{sp:matching-fraction} give a lower bound strictly larger than \[ \frac{0.95^2}{m^2}-\frac1{|J|} >\frac{0.9025-0.1}{m^2} >\frac1{4m^2}. \] On the other hand, every distinct pair in \(J\) obeys \eqref{sp:collision-upper}. Averaging those simultaneous bounds over the ordered pairs in \(J\times J\), with the diagonal excluded, gives an upper bound less than \(1/(4m^2)\). The choices of \(J\) and \(j_0\) after fixing \(w\) are legitimate because all original pair bounds already hold at that fiber. This is the required contradiction. The induction step holds for every ambient system and every tower at the current ordinal. Universal transfinite induction therefore proves the lemma for all countable towers. \end{proof} % END module % BEGIN attributed rewritten module rigidity_tower.tex \section{Compact towers and conditional averages} \label{ct:section} This section develops the compact-factor construction and the conditional multiple-average argument of OpenAI~\cite{OpenAI2026}. Their role in the rigidity proof is to produce a factor that satisfies the approximation hypothesis of Lemma~\ref{sp:relative}, while retaining enough mixing above that factor to separate conditional correlations. The underlying structure theory originates with Furstenberg and Zimmer. The finite-dimensional fiber argument below is closely related to \cite[Lemmas~6.6 and~7.2]{Furstenberg1977}, and the module formulation is discussed, with corrections, in \cite{JamneshanCorr}. We give the details needed here. All actions in this section are probability-preserving actions of the discrete group \(K\). As established in the spectral preliminaries, \(K\) is countable and abelian, and each integer power homomorphism \[ K\longrightarrow K,\qquad u\longmapsto u^d, \qquad d\in\Z\setminus\{0\}, \] is surjective. The operator convention is \(U_kf(x)=f(kx)\). Factors and functions are identified modulo null sets. \subsection{Conditional products} Suppose that \(\pi:X\to Y\) is a factor map between standard probability spaces, with measures \(\mu\) and \(\nu\). Disintegration gives a measurable family of probabilities \(\mu_y\), concentrated on \(\pi^{-1}(y)\), such that \[ \mu=\int_Y\mu_y\,d\nu(y). \] The standard Borel disintegration theorem and uniqueness of its kernel are sufficient here \cite[Theorem~4 and Remark~5]{Tao2008}. Indeed, the conditional integration identity for a countable family of Borel sets separating the points of \(Y\) gives the asserted concentration on a fiber. For an equivariant map, uniqueness gives \(k_*\mu_y=\mu_{ky}\) for almost every \(y\), for each \(k\). Intersecting these conull sets over the countable group, and then over their translates, makes the kernel equivariant on a common invariant conull set. The relative square \(\mathcal R=X\times_YX\) carries the probability \begin{equation} \label{ct:relative-product} \mu_{\mathcal R}=\int_Y\mu_y\otimes\mu_y\,d\nu(y) \end{equation} and the diagonal \(K\)-action. Its two coordinate marginals are \(\mu\). We write \(P=\E(\,\cdot\mid Y)\) on \(X\), and \(P_{\mathcal R}\) for conditional expectation from \(\mathcal R\) onto \(Y\). For bounded \(a,b\) on \(X\), set \((a\otimes b)(x,x')=a(x)b(x')\). The defining conditional independence then says \begin{equation} \label{ct:tensor-expectation} P_{\mathcal R}(a\otimes b)=(Pa)(Pb). \end{equation} Conditional expectation commutes with every \(U_k\), is a contraction in both \(L^2\) and \(L^\infty\), and satisfies \(P(cf)=cPf\) for bounded \(Y\)-measurable \(c\). \subsection{Construction of the tower} \begin{lemma}[Compact tower, OpenAI~\cite{OpenAI2026}] \label{ct:tower} For an ergodic probability-preserving action of \(K\) on a standard space \(X\), one can find a countable ordinal \(\eta\) and a tower of invariant factors \((Y_\alpha)_{\alpha\leq\eta}\) satisfying the following conditions. \begin{enumerate} \item \(Y_0\) is trivial. At each limit ordinal, the factor is generated by the preceding factors. \item For every \(\alpha<\eta\), every \(H\in L^\infty(Y_{\alpha+1})\), and every \(\varepsilon>0\), there are bounded functions \(H_1,\ldots,H_q\) on \(Y_{\alpha+1}\) and \(B<\infty\) such that for each \(k\in K\) one can choose \(b_1^k,\ldots,b_q^k\in L^\infty(Y_\alpha)\) with \begin{equation} \label{ct:tower-approximation} \left\|U_kH-\sum_{r=1}^q b_r^kH_r\right\|_2<\varepsilon, \qquad \|b_r^k\|_\infty\leq B. \end{equation} The functions \(H_r\) and the bound \(B\) are independent of \(k\). \item \(X\times_{Y_\eta}X\) is ergodic. \end{enumerate} Thus every successor has precisely the approximation property \eqref{sp:tower-property}. \end{lemma} \begin{proof} We first show how to enlarge an invariant factor \(Y\) whenever \(X\times_YX\) is not ergodic. \paragraph{An invariant kernel.} Choose a bounded nonconstant invariant function \(L\) on \(X\times_YX\), for example the indicator of a nontrivial invariant set. The conditional expectation of \(L\) onto either coordinate is invariant on \(X\). Ergodicity makes both expectations equal to \(\int L\,d\mu_{\mathcal R}\). Subtracting this constant preserves nonzeroness and invariance and gives \[ \int L(x,x')\,d\mu_{\pi(x)}(x')=0, \qquad \int L(x,x')\,d\mu_{\pi(x')}(x)=0 \] almost everywhere in the respective coordinates. On \(\mathcal H_y=L^2(\mu_y)\), define \[ (A_yf)(x)=\int L(x,x')f(x')\,d\mu_y(x'). \] Put \(C=\|L\|_\infty\). Each \(A_y\) is Hilbert--Schmidt and \begin{equation} \label{ct:kernel-bounds} \|A_y\|_{\mathrm{HS}}^2\leq C^2, \qquad A_y1=A_y^*1=0. \end{equation} The kernel is nonzero on a positive-measure set of fibers, so the same is true of the operator field. Kernel invariance identifies the operators at \(y\) and \(ky\) by unitary conjugacy. More explicitly, pullback by \(x\mapsto kx\) maps \(\mathcal H_{ky}\) unitarily onto \(\mathcal H_y\) and intertwines \(A_{ky}\) with \(A_y\). \paragraph{Measurability of a finite-dimensional spectral space.} For \(\delta>0\), consider \[ Q_y=\mathbf1_{(\delta,\infty)}(A_y^*A_y). \] We justify the required measurability before choosing \(\delta\). Take a countable algebra of Borel sets generating \(X\), and enumerate the associated simple functions with complex rational coefficients. Their restrictions are dense in every \(\mathcal H_y\). Parameter integration against \(L\) and its adjoint kernel shows that applying \(A_y\), \(A_y^*\), or any polynomial in \(A_y^*A_y\) to these sections gives measurable sections. All spectra of \(A_y^*A_y\) lie in \([0,C^2]\). On this interval the continuous functions \[ h_n(t)=\min\{1,\max\{0,n(t-\delta)\}\} \] converge pointwise to \(\mathbf1_{(\delta,\infty)}(t)\), including the value zero at \(t=\delta\). Choose polynomials \(p_n\) with \(\sup_{[0,C^2]}|p_n-h_n|\to0\). They may be chosen with a common supremum bound. The spectral theorem on each fiber gives \[ p_n(A_y^*A_y)s_y\longrightarrow Q_ys_y \quad\hbox{in }\mathcal H_y \] for every test section \(s\) from the countable family. The common bound and dominated convergence show that the measurable sections on the left are Cauchy in global \(L^2(X)\). A subsequence converges in fiber norm for almost every \(y\), identifying the global limit with \(Q_ys_y\). Hence every projected test section is measurable. Their Gram matrices show that the rank of \(Q_y\) is measurable. For each fiber with \(A_y\ne0\), some positive spectral cutoff has nonzero range. By countability of the cutoffs \(1/n\), one can now fix \(\delta>0\) for which \(Q_y\ne0\) on a set of positive measure. The trace bound gives \[ \delta\,\operatorname{rank}Q_y \leq\operatorname{tr}(A_y^*A_y)\leq C^2. \] Unitary conjugacy makes this integer-valued rank invariant on \(Y\). Every factor of the ergodic system \(X\) is ergodic, so \(\operatorname{rank}Q_y=d\) almost everywhere for one integer \(d\geq1\). By \eqref{ct:kernel-bounds}, the range is perpendicular to the fiber constants. The projected test sections span a dense subspace of each range. Select the first available independent section at each step and apply Gram--Schmidt in the fibers. The selection tests and normalizing factors are measurable because they are expressed by the Gram matrices. This constructs measurable sections \(e_1,\ldots,e_d\) satisfying \begin{equation} \label{ct:fiber-basis} \int e_i\overline{e_j}\,d\mu_y=\delta_{ij}, \qquad \int e_i\,d\mu_y=0. \end{equation} They belong to global \(L^2(X)\), since their fiber norms are one. \paragraph{Boundedness and the new factor.} For each \(k\), the functions \(U_ke_1,\ldots,U_ke_d\) form another orthonormal basis of the same fiber range. Thus \[ U_ke_i=\sum_{j=1}^d m_{ij}^k e_j \] with a measurable unitary matrix \(M^k=(m_{ij}^k)\) over \(Y\). In particular \(|m_{ij}^k|\leq1\) almost everywhere. These identities hold for every \(k\) on a common conull set, because \(K\) is countable. They imply that \[ x\longmapsto \sum_{i=1}^d |e_i(x)|^2 \] is \(K\)-invariant. Its integral is \(d\), and ergodicity therefore gives \begin{equation} \label{ct:basis-bound} \sum_{i=1}^d |e_i(x)|^2=d \quad\hbox{almost everywhere}. \end{equation} In particular each \(e_i\) is bounded by \(\sqrt d\). Let \(Y'\) be the factor generated by \(Y,e_1,\ldots,e_d\). The transformation formula above, also applied to \(k^{-1}\), makes \(Y'\) invariant. It is strict over \(Y\). Indeed, a \(Y\)-measurable \(e_i\) would equal its conditional mean, which is zero by \eqref{ct:fiber-basis}, although its conditional squared norm is one. \paragraph{Uniform approximation over \(Y\).} Finite polynomials in \(e_1,\ldots,e_d\) and their conjugates, with coefficients in \(L^\infty(Y)\), form a bounded unital algebra closed under conjugation and generating \(Y'\). The functional monotone class theorem, followed by \(L^2\) approximation, shows that this algebra is dense in \(L^2(Y')\). Choose a polynomial \(p\) with \(\|H-p\|_2<\varepsilon\). Write its terms as \[ p=\sum_{a=1}^m c_a M_a(e_1,\ldots,e_d,\overline e_1,\ldots,\overline e_d), \qquad c_a\in L^\infty(Y), \] where \(M_a\) is a monomial of total degree \(D_a\). Substituting the unitary basis transformations into \(U_kp\) gives only monomials of degrees at most \(\max_aD_a\). The \(a\)-th original monomial contributes at most \(d^{D_a}\) terms. Every resulting coefficient is a \(Y\)-measurable function, and after collecting equal monomials its supremum norm is at most \[ B=\sum_{a=1}^m d^{D_a}\|c_a\|_\infty. \] This estimate holds for every \(k\), since unitary matrix entries have modulus at most one and \(\|U_kc_a\|_\infty=\|c_a\|_\infty\). Taking the finitely many possible monomials as \(H_r\) gives \(U_kp=\sum_r b_r^kH_r\) with the required coefficient bound. Finally, \(\|U_kH-U_kp\|_2=\|H-p\|_2<\varepsilon\). This proves \eqref{ct:tower-approximation} for \(Y'\) over \(Y\). The argument uses bounded coefficients for the transformations of the chosen basis and of a fixed polynomial. For precision, the algebraic module \(\sum_iL^\infty(Y)e_i\) can have an \(L^2\) closure containing vectors with merely \(L^2(Y)\) coefficients. The construction above takes its basis from the measurable fiber ranges and establishes all coefficient bounds directly. \paragraph{Iteration and termination.} Begin with the trivial factor. Whenever the relative square over the current factor is nonergodic, apply the preceding construction. At a countable limit ordinal take the generated factor. There are countably many preceding factors and each is countably generated modulo null sets, so their join has the same property and admits a standard probability model. If the process continued through every countable ordinal, then for each \(\alpha<\omega_1\) there would be a unit vector \[ f_\alpha\in L^2(Y_{\alpha+1}) \quad\hbox{perpendicular to }L^2(Y_\alpha). \] One of the newly adjoined \(e_i\) supplies such a vector. For \(\alpha<\beta\), nestedness places \(f_\alpha\) in \(L^2(Y_\beta)\), so \(f_\alpha\) and \(f_\beta\) are orthogonal. This would be an uncountable orthonormal family in the separable space \(L^2(X)\). Consequently the procedure stops at a countable ordinal \(\eta\). At that stage the relative square is ergodic, while every earlier successor satisfies \eqref{ct:tower-approximation}. \end{proof} If \(d:X\to\mathbb D\) is equivariant and its law gives \(C\) measure zero, Lemma~\ref{sp:relative} now applies to every factor in this tower. Its hypotheses are exactly the countable construction from the trivial factor, the generated joins at limit stages, and the uniform approximation \eqref{ct:tower-approximation}. It follows that the law of \(d(x)-d(x')\) for independent conditional samples over any \(Y_\alpha\) also gives \(C\) measure zero. \subsection{Averages along the discrete rotation group} A sequence of nonempty finite sets \((\Phi_N)\) in \(K\) is a F{\o}lner sequence if \begin{equation} \label{ct:folner} \frac{|k\Phi_N\mathbin{\triangle}\Phi_N|}{|\Phi_N|} \longrightarrow0 \qquad\hbox{for every }k\in K. \end{equation} Such a sequence exists. To see this, enumerate \(K\) and at step \(N\) consider the subgroup generated by its first \(N\) elements. It is a product of a finite abelian group and a finitely generated free abelian group. Take the whole finite factor and a sufficiently large box in the free factor so that translations by the first \(N\) elements have relative boundary less than \(1/N\). These finite sets satisfy \eqref{ct:folner}. For any family \(b_u\) in a vector space, write \[ \operatorname{Av}_{u\in\Phi}b_u =\frac1{|\Phi|}\sum_{u\in\Phi}b_u. \] \begin{lemma}[Mean averages] \label{ct:mean} Let \(V\) be a unitary representation of \(K\) on a Hilbert space. For every F{\o}lner sequence, the operators \(\operatorname{Av}_{u\in\Phi_N}V_u\) converge strongly to orthogonal projection onto the invariant vectors. \end{lemma} \begin{proof} The orthogonal complement of the invariant vectors is the closed linear span of the vectors \(V_kv-v\). Indeed, a vector perpendicular to every such difference is fixed by every \(V_k^*\), hence by every \(V_k\). On an individual difference, commutativity and \eqref{ct:folner} give \[ \left\|\operatorname{Av}_{u\in\Phi_N}V_u(V_kv-v)\right\| \leq \frac{|k\Phi_N\mathbin{\triangle}\Phi_N|}{|\Phi_N|}\|v\| \longrightarrow0. \] The averaging operators have norm at most one. Density proves convergence to zero on the stated orthogonal complement, and they act as the identity on invariant vectors. \end{proof} For every nonzero integer \(d\), the operators \(u\mapsto U_{u^d}\) also form a unitary representation of \(K\). Surjectivity of \(u\mapsto u^d\) shows that its invariant vectors are the invariant vectors of the original action. In particular, on any ergodic \(K\)-system, \begin{equation} \label{ct:power-mean} \operatorname{Av}_{u\in\Phi_N}U_{u^d}f \longrightarrow \int f \quad\hbox{in }L^2 \qquad(d\ne0). \end{equation} This applies the mean theorem to a representation indexed by the original averaging sets \(\Phi_N\). \begin{lemma}[Conditional multiple averages, OpenAI~\cite{OpenAI2026}] \label{ct:multiple} Let \(X\) be an ergodic standard probability-preserving \(K\)-system, let \(Y\) be an invariant factor, and suppose that \(X\times_YX\) is ergodic. Put \(P=\E(\,\cdot\mid Y)\). For \(l\geq1\), pairwise distinct integers \(n_1,\ldots,n_l\), bounded functions \(v_1,\ldots,v_l\), and every F{\o}lner sequence in \(K\), \begin{equation} \label{ct:multiple-limit} \operatorname{Av}_{u\in\Phi_N} \left\| P\left(\prod_{j=1}^lU_{u^{n_j}}v_j\right) -\prod_{j=1}^lU_{u^{n_j}}Pv_j \right\|_2^2 \longrightarrow0. \end{equation} \end{lemma} \begin{proof} Keep the notation \(\mathcal R=X\times_YX\) from \eqref{ct:relative-product}. We first establish a two-function estimate on \(X\), transfer that estimate to \(\mathcal R\), and then prove the multiple-average assertion. \paragraph{A conditional two-function estimate.} For bounded \(v,w\) and \(d\in\Z\setminus\{0\}\), we claim that \begin{equation} \label{ct:pair} \operatorname{Av}_{u\in\Phi_N} \|P(vU_{u^d}w)-Pv\,U_{u^d}Pw\|_2^2 \longrightarrow0. \end{equation} Let \(v^\circ=v-Pv\) and \(w^\circ=w-Pw\). The conditional module identity and commutation with the action show that the difference inside the norm is \(P(v^\circ U_{u^d}w^\circ)\). By conditional independence, \begin{equation} \label{ct:pair-tensor} \|P(v^\circ U_{u^d}w^\circ)\|_2^2 = \int_{\mathcal R} (v^\circ\otimes\overline{v^\circ})\, U_{u^d}(w^\circ\otimes\overline{w^\circ})\,d\mu_{\mathcal R}. \end{equation} The second tensor has integral \(\int_Y|Pw^\circ|^2\,d\nu=0\). The action on \(\mathcal R\) is ergodic by hypothesis. Apply \eqref{ct:power-mean} there and pair the resulting \(L^2\) convergence with \(v^\circ\otimes\overline{v^\circ}\). This proves \eqref{ct:pair} along every F{\o}lner sequence. The same estimate holds for bounded functions on \(\mathcal R\), with \(P_{\mathcal R}\) in place of \(P\). For elementary tensors \(v=a_1\otimes a_2\) and \(w=b_1\otimes b_2\), formula \eqref{ct:tensor-expectation} gives \[ P_{\mathcal R}(vU_{u^d}w) =P(a_1U_{u^d}b_1)P(a_2U_{u^d}b_2). \] If the two factors on the right are \(A_1,A_2\), and their respective factor approximations are \(B_1,B_2\), use \[ A_1A_2-B_1B_2=(A_1-B_1)A_2+B_1(A_2-B_2). \] Every companion factor is uniformly bounded. The estimate on \(X\) therefore proves the assertion for elementary tensors, and bilinearity proves it for finite sums. For completeness, finite sums of bounded tensors approximate every bounded function on \(\mathcal R\) in \(L^2\), with a common supremum bound. Choose increasing finite partitions generated by countable Borel generators in each coordinate. Conditional expectations onto their rectangular grids converge in \(L^2\), preserve the supremum bound, and are finite sums of coordinate tensors restricted to \(\mathcal R\). If \(D_u(v,w)\) denotes the pair difference for either of the conditional expectation operators, contraction gives \[ \|D_u(v,w)-D_u(v',w)\|_2 \leq2\|w\|_\infty\|v-v'\|_2. \] The analogous estimate for changing \(w\) has right side \(2\|v'\|_\infty\|w-w'\|_2\). These bounds are uniform in \(u\). Approximating the two arguments successively extends the pair estimate to all bounded functions on \(\mathcal R\). \paragraph{A vector assertion with one centered function.} Let \(Z\) denote either \(X\) or \(\mathcal R\), let \(\lambda\) be its probability measure, and let \(Q=\E_Z(\,\cdot\mid Y)\). Both systems are ergodic, and both now satisfy the pair estimate. We prove that, whenever the integers \(n_1,\ldots,n_l\) are pairwise distinct and nonzero and \(Qv_{j_*}=0\) for some \(j_*\), \begin{equation} \label{ct:centered-vector} \left\| \operatorname{Av}_{u\in\Phi_N} \prod_{j=1}^lU_{u^{n_j}}v_j \right\|_2\longrightarrow0 \end{equation} for every F{\o}lner sequence. Rescaling reduces to \(\|v_j\|_\infty\leq1\). For \(l=1\), the integral of \(v_1\) is zero, so \eqref{ct:power-mean} proves the assertion. Assume the assertion for \(l-1\). Put \[ H_u=\prod_{j=1}^lU_{u^{n_j}}v_j, \qquad b_{j,h}=(U_{h^{n_j}}v_j)\overline{v_j}. \] For each fixed \(h\), commutativity and invariance of \(\lambda\) give \begin{align} \label{ct:correlation-reduction} \langle H_{hu},H_u\rangle &=\int_Z\prod_{j=1}^l U_{u^{n_j}}b_{j,h}\,d\lambda \notag\\ &=\int_Z b_{l,h} \prod_{j0\), increasing indices \(M_i\), and \(t_i\in K\) for which the averages over \(\Psi_{M_i}t_i\) exceed \(\varepsilon\). Those translated sets are themselves a F{\o}lner sequence, since \[ \frac{|k(\Psi_{M_i}t_i)\mathbin{\triangle}(\Psi_{M_i}t_i)|} {|\Psi_{M_i}t_i|} = \frac{|k\Psi_{M_i}\mathbin{\triangle}\Psi_{M_i}|}{|\Psi_{M_i}|} \longrightarrow0 \] for each \(k\). This contradicts the vanishing already obtained for every F{\o}lner sequence. \paragraph{The Hilbert-space shift estimate.} Fix a finite nonempty subset \(S\subset K\). Because \(\|H_u\|_2\leq1\), the difference between \(\operatorname{Av}_{u\in\Phi_N}H_u\) and \(\operatorname{Av}_{u\in\Phi_N}\operatorname{Av}_{h\in S}H_{hu}\) has norm at most \[ \frac1{|S|}\sum_{h\in S} \frac{|h\Phi_N\mathbin{\triangle}\Phi_N|}{|\Phi_N|}, \] which tends to zero. Jensen's inequality in \(L^2(Z)\), followed by expansion of the finite square, yields \begin{align} \label{ct:shift-estimate} \limsup_{N\to\infty} \left\|\operatorname{Av}_{u\in\Phi_N}H_u\right\|_2^2 &\leq \frac1{|S|^2}\sum_{h,h'\in S} \limsup_{N\to\infty} \left| \operatorname{Av}_{u\in\Phi_N} \langle H_{hu},H_{h'u}\rangle \right| \notag\\ &\leq \frac1{|S|^2}\sum_{h,h'\in S}a(h(h')^{-1}). \end{align} For the last inequality substitute \(u'=h'u\). The set of new averaging indices is \(h'\Phi_N\), again a F{\o}lner sequence, so \eqref{ct:correlation-bound} applies. Now choose \(S=\Psi_M\). The last line of \eqref{ct:shift-estimate} is at most \(\sup_t\operatorname{Av}_{h\in\Psi_M}a(ht)\), which tends to zero by \eqref{ct:uniform-translates}. Here the limit superior in \(N\) is taken with \(S\) fixed, and only afterwards is \(M\) allowed to grow. This proves \eqref{ct:centered-vector} and completes the induction on both \(X\) and \(\mathcal R\). \paragraph{Return to conditional averages.} The integers in the lemma may include zero. Expand every \(v_j\) as \(Pv_j+(v_j-Pv_j)\). Consider one expanded product \[ W_u=\prod_{j=1}^lU_{u^{n_j}}w_j \] with \(Pw_j=0\) for at least one index. A further use of conditional independence gives \begin{equation} \label{ct:multiple-tensor} \|PW_u\|_2^2 = \int_{\mathcal R} \prod_{j=1}^lU_{u^{n_j}} (w_j\otimes\overline{w_j})\,d\mu_{\mathcal R}. \end{equation} For the centered index, \[ P_{\mathcal R}(w_j\otimes\overline{w_j})=|Pw_j|^2=0. \] Choose one integer \(c\) such that \(n_j+c\ne0\) for every \(j\). Applying \(U_{u^c}\) to the whole integrand in \eqref{ct:multiple-tensor} leaves its scalar integral unchanged and replaces each exponent \(n_j\) by \(n_j+c\). They remain distinct, and are now all nonzero. The vector assertion \eqref{ct:centered-vector} on \(\mathcal R\) therefore makes the averaged integrand converge to zero in \(L^2\). Its scalar integral tends to zero as well. Thus \[ \operatorname{Av}_{u\in\Phi_N}\|PW_u\|_2^2\longrightarrow0. \] The expansion contains finitely many such centered products. For \(q\) of them, the squared norm of their sum is at most \(q\) times the sum of their squared norms, so their total projected contribution has vanishing averaged squared norm. The remaining product is already \(Y\)-measurable and equals \(\prod_jU_{u^{n_j}}Pv_j\). This establishes \eqref{ct:multiple-limit}. \end{proof} % END module % BEGIN attributed rewritten module rigidity_radial.tex \section{Radial identities and completion of character rigidity} \label{rd:section} The radial identities, the construction inside a number field, and the argument with bounded residues in this section follow OpenAI~\cite{OpenAI2026}. \subsection{Conditional Fourier data} \label{rd:conditional-data} Retain the groups \[ F=\overline{\Q}\cap\R,\qquad E=F(i),\qquad K=\{u\in E:|u|=1\},\qquad \mathbb D=\widehat{E_{\mathrm{disc}}}, \] and the Borel subgroup $C\subset\mathbb D$ of ordinary continuous characters of the plane. For $z\in E$ write $\psi_z(d)=d(z)$. The action of $u\in K$ on $\mathbb D$ is given by $(ud)(z)=d(uz)$, and its action on functions is $U_u H(d)=H(ud)$. Let $\mu$ be an ergodic $K$-invariant Borel probability on $\mathbb D$ such that $\mu(C)=0$. Since $E$ and $K$ are countable and $\mathbb D$ is compact metrizable, $(\mathbb D,\mu)$ is a standard probability-preserving $K$-system. Apply Lemma~\ref{ct:tower} to this system and denote the terminal factor by $Y$, with probability measure $\eta$. The relative product $\mathbb D\times_Y\mathbb D$ is ergodic. The identity map $\operatorname{id}_{\mathbb D}$ from $(\mathbb D,\mu)$ to $\mathbb D$ is equivariant and has wild law, so Lemma~\ref{sp:relative} applies to the same tower. Consequently the law $\nu_Y$ of $d-d'$, for two conditionally independent points over $Y$, satisfies $\nu_Y(C)=0$. Equivariance of the conditional measures also makes $\nu_Y$ $K$-invariant. Let $\mu_y$ denote conditional probability over $y\in Y$, and set \begin{equation} \label{rd:fourier-data} P=\E(\,\cdot\mid Y),\qquad g_z=P\psi_z,\qquad m(z)=\lVert g_z\rVert_2^2\quad(z\in E). \end{equation} Conditional independence gives \begin{align} \int_{\mathbb D}\psi_z(d)\,d\nu_Y(d) &= \int_Y \left(\int_{\mathbb D}\psi_z(d)\,d\mu_y(d)\right) \overline{\left(\int_{\mathbb D}\psi_z(d')\,d\mu_y(d')\right)} \,d\eta(y) \notag\\ &=\int_Y |g_z|^2\,d\eta=m(z). \label{rd:difference-transform} \end{align} In particular, $0\le m(z)\le1$ and $m(0)=1$. Conditional expectation commutes with the action on this invariant factor. Hence \[ g_{uz}=U_u g_z,\qquad m(uz)=m(z),\qquad g_{-z}=\overline{g_z},\qquad \lVert g_z\rVert_\infty\le1. \] For every nonzero $z\in E$, its modulus belongs to $F_{>0}$ and $z/|z|\in K$. Thus $m(z)$ depends only on $|z|$. We also use $m(r)$ for this function on $F_{\ge0}$, and put \[ w(r)=-\log m(r)\in[0,\infty],\qquad -\log0=\infty. \] Then $w(0)=0$. Write $I$ for the positive functional in Lemma~\ref{sp:line}. Its domain contains every bounded, boundedly supported function on $F$, and $I(\phi|_F)=\int_{\R}\phi(t)\,dt$ for $\phi\in C_c(\R)$. Applying that lemma to $\nu_Y$, with the nonzero direction $1\in E$, and using \eqref{rd:difference-transform}, gives \begin{equation} \label{rd:line-vanishes} I\bigl(t\longmapsto\phi(t)m(t)\bigr)=0 \qquad\bigl(\phi\in C_c(\R)\bigr). \end{equation} For a finite nonempty set $\Phi\subset K$, use the notation \[ \operatorname{Avg}_{u\in\Phi}b_u =\frac{1}{|\Phi|}\sum_{u\in\Phi}b_u. \] We will use mean ergodicity along every F{\o}lner sequence. For a unitary representation $V$ of $K$, the orthogonal complement of its invariant vectors is the closed span of vectors $V_h f-f$. Indeed, orthogonality to all these differences is equivalent to invariance under each $V_h$. The average of such a difference has norm at most \[ \frac{|h\Phi\mathbin{\triangle}\Phi|}{|\Phi|}\lVert f\rVert. \] This tends to zero along a F{\o}lner sequence, and the averaging operators have norm at most one. Approximation by finite linear combinations of differences proves convergence to the invariant projection. For each nonzero integer $q$, the power map $u\mapsto u^q$ is onto $K$. The representation $u\mapsto U_{u^q}$ therefore has the same invariant vectors as $u\mapsto U_u$. The factor $Y$ is ergodic because the original system is ergodic, so its invariant functions are constant. \subsection{The two radial identities} \begin{lemma} \label{rd:radial-identities} For all $A,B\in F_{\ge0}$, \begin{equation} \label{rd:difference-inequality} w(|A-B|)\le w(A)+w(B). \end{equation} For every $r\in F_{>0}$ and every F{\o}lner sequence $(\Phi_n)$ in $K$, \begin{equation} \label{rd:two-direction-average} \operatorname{Avg}_{u\in\Phi_n} m\bigl(2r|\operatorname{Re}u|\bigr) \longrightarrow m(r)^2. \end{equation} \end{lemma} \begin{proof} If $v_u$ and $h_u$ have $L^2$ norm at most one, then \begin{equation} \label{rd:norm-error} \bigl|\lVert v_u\rVert_2^2-\lVert h_u\rVert_2^2\bigr| \le 2\lVert v_u-h_u\rVert_2. \end{equation} If the average of $\lVert v_u-h_u\rVert_2^2$ tends to zero, Cauchy--Schwarz in the finite average shows that the average of the left side of \eqref{rd:norm-error} tends to zero as well. Suppose $A,B>0$ and write $L=\sqrt{AB}\in F$. For every $u\in K$, \begin{equation} \label{rd:equal-moduli} |A-Bu^2|^2 =(A-B)^2+4AB(\operatorname{Im}u)^2 =|-L+(A-B)u+Lu^2|^2. \end{equation} The first equality follows from $\operatorname{Re}(u^2)=1-2(\operatorname{Im}u)^2$. For the second, factor the last expression inside the modulus as \[ u\bigl((A-B)+L(u-u^{-1})\bigr) =u\bigl((A-B)+2iL\operatorname{Im}u\bigr). \] The character on the left of \eqref{rd:equal-moduli} factors as \[ \psi_{A-Bu^2}=\psi_A\,U_{u^2}\psi_{-B}. \] Apply Lemma~\ref{ct:multiple} with exponents $0,2$ and the bounded functions $\psi_A,\psi_{-B}$. Its hypotheses hold because $\mathbb D\times_Y\mathbb D$ is ergodic. Along any F{\o}lner sequence, \[ \operatorname{Avg}_{u\in\Phi_n} \bigl\lVert P\psi_{A-Bu^2} -g_A U_{u^2}g_{-B}\bigr\rVert_2^2 \longrightarrow0. \] Both functions in the difference have norm at most one, so \eqref{rd:norm-error} transfers this to squared norms. Moreover, \[ \bigl\lVert g_A U_{u^2}g_{-B}\bigr\rVert_2^2 =\int_Y |g_A|^2 U_{u^2}|g_{-B}|^2\,d\eta. \] Mean ergodicity on $Y$ for the representation $u\mapsto U_{u^2}$ therefore yields \begin{equation} \label{rd:two-term-limit} \operatorname{Avg}_{u\in\Phi_n}m(A-Bu^2) \longrightarrow \left(\int_Y|g_A|^2\,d\eta\right) \left(\int_Y|g_{-B}|^2\,d\eta\right) =m(A)m(B). \end{equation} For the other character in \eqref{rd:equal-moduli}, apply Lemma~\ref{ct:multiple} with exponents $0,1,2$. The resulting approximant to $P\psi_{-L+(A-B)u+Lu^2}$ is \[ h_u=g_{-L}\,U_u g_{A-B}\,U_{u^2}g_L, \] and its averaged squared $L^2$ error tends to zero. Its two outer factors have absolute value at most one, whence \[ \lVert h_u\rVert_2^2 \le\lVert U_u g_{A-B}\rVert_2^2 =m(|A-B|). \] Radiality and \eqref{rd:equal-moduli} identify the squared norms of the two projected characters. Apply \eqref{rd:norm-error} once more and use \eqref{rd:two-term-limit} to obtain \[ m(A)m(B)\le m(|A-B|). \] If $m(A)$ and $m(B)$ are positive, the last quantity is positive too, and taking negative logarithms proves \eqref{rd:difference-inequality}. If either factor vanishes, its $w$-value is infinite and the same inequality holds in the extended-real order. The computation also permits $A=B$, when the middle character in the second application is $\psi_0=1$. If $A=0$ or $B=0$, \eqref{rd:difference-inequality} is an equality because $w(0)=0$. For \eqref{rd:two-direction-average}, use $u^{-1}=\overline u$ and \[ ru+ru^{-1}=2r\operatorname{Re}u. \] Lemma~\ref{ct:multiple}, with exponents $1,-1$ and the two functions $\psi_r,\psi_r$, gives \[ \operatorname{Avg}_{u\in\Phi_n} \bigl\lVert P\psi_{ru+ru^{-1}}-U_u g_r\,U_{u^{-1}}g_r \bigr\rVert_2^2\longrightarrow0. \] The squared norm of the approximant is \begin{align*} \bigl\lVert U_u g_r\,U_{u^{-1}}g_r\bigr\rVert_2^2 &=\int_Y U_u|g_r|^2\,U_{u^{-1}}|g_r|^2\,d\eta\\ &=\int_Y |g_r|^2\,U_{u^{-2}}|g_r|^2\,d\eta. \end{align*} Mean ergodicity for $u\mapsto U_{u^{-2}}$ makes its average converge to $m(r)^2$. Applying \eqref{rd:norm-error} and radiality proves \eqref{rd:two-direction-average}. All uses of Lemma~\ref{ct:multiple} and mean ergodicity hold along every F{\o}lner sequence, which proves the stated quantifier. \end{proof} \subsection{Bases of unit directions in a number field} \begin{lemma} \label{rd:general-position} Let $N_0\subset F$ be a number field, put $N=N_0(i)$, and write $d_0=[N:\Q]$. For every positive integer $L$ there is a finite set $S\subset K\cap N$ with $|S|=L$ such that every subset of $S$ with at most $d_0$ elements is linearly independent over $\Q$. In particular, every $d_0$ distinct elements form a $\Q$-basis of $N$. \end{lemma} \begin{proof} For $t\in N_0$ consider \[ u(t)=\frac{1+it}{1-it}. \] Its denominator is nonzero, since $t$ is real. Complex conjugation interchanges numerator and denominator, so $|u(t)|=1$, and $u(t)\in K\cap N$. Choose $\Q$-bases of $N_0$ and $N$. Expressing $t$ in the first basis identifies its parameters with a point of $\Q^{[N_0:\Q]}$. Multiplication by $1-it$ on $N$ has a matrix with entries affine in these parameters and rational coefficients. The inverse matrix expresses the coordinates of $u(t)$ as rational functions over $\Q$. At every rational parameter point, this matrix is invertible because $1-it$ is a nonzero element of the field $N$. Suppose a nonzero $\Q$-linear functional $\ell:N\to\Q$ annihilates all the $u(t)$, and fix $a\in N_0$. As a function of a rational variable $h$, $\ell(u(ha))$ is a rational function over $\Q$ that vanishes for every $h\in\Q$. Its denominator is nonzero at these points, so its numerator has infinitely many roots and the rational function is identically zero. The formal expansion in $N[[h]]$ is \[ u(ha)=1+2iha-2h^2a^2+O(h^3). \] Its constant, linear, and quadratic coefficients imply \[ \ell(1)=0,\qquad \ell(ia)=0,\qquad \ell(a^2)=0. \] This holds for every $a\in N_0$. Every element of $N_0$ is a rational linear combination of squares, since \[ 4a=(a+1)^2-(a-1)^2. \] Thus $\ell$ vanishes on both $N_0$ and $iN_0$. These two subspaces span $N$, contradicting $\ell\ne0$. Now let $W_1,\ldots,W_b$ be finitely many proper rational linear subspaces of $N$. For each one choose a nonzero rational linear functional $\ell_j$ that vanishes on $W_j$. The rational function $\ell_j(u(t))$ is not identically zero by the preceding paragraph, so it has a nonzero numerator polynomial in the rational coordinates of $t$. The product of these finitely many numerator polynomials is a nonzero polynomial over $\Q$. Such a polynomial has a rational point where it is nonzero, as follows by induction on the number of variables from the corresponding assertion for one variable. At that parameter point all denominators are nonzero as well. Consequently some $u(t)$ avoids every $W_j$. Construct $S$ successively. At each step exclude the spans of all subsets of at most $d_0-1$ previously selected points. There are finitely many such spans, and each has dimension less than $d_0$. The preceding avoidance property supplies another unit direction outside all of them. Every subset with at most $d_0$ elements remains independent after this addition. Repeating the procedure gives a set of any prescribed finite size. \end{proof} \subsection{Integer combinations below a fixed cap} \begin{lemma} \label{rd:capped-residues} Suppose $w:F_{\ge0}\to[0,\infty]$ satisfies $w(0)=0$ and \eqref{rd:difference-inequality}. Let $p_1,\ldots,p_q\in F_{\ge0}$ and $H\in[0,\infty)$ satisfy \[ p=\max_{1\le j\le q}p_j>0,\qquad w(p_j)\le H. \] For an integer $M\ge0$, suppose integers $a_1,\ldots,a_q$ obey \[ z=\sum_{j=1}^q a_jp_j\in[0,p),\qquad \sum_{j=1}^q|a_j|\le M, \] then $w(z)\le4MH$. \end{lemma} \begin{proof} For $x\in F$ let \[ \rho_p(x)=x-p\lfloor x/p\rfloor\in F\cap[0,p). \] This is the unique representative of $x$ modulo $p\Z$ in $[0,p)$, so $\rho_p(x+y)=\rho_p(\rho_p(x)+\rho_p(y))$. Since $p$ equals one of the $p_j$, we have $w(p)\le H$. If $0p$, then $\rho_p(a+b)=a-(p-b)>0$ and \[ w(\rho_p(a+b)) \le w(a)+w(p-b) \le w(a)+w(p)+w(b). \] If $a+b=p$, the new residue is zero and its $w$-value is zero. In all three cases, \[ w(\rho_p(a+b))\le w(a)+w(b)+2H. \] Adding the residue of one signed input therefore increases the current bound by at most $4H$. Expand the integer combination into $\sum_j|a_j|$ signed inputs, starting with residue zero. After all inputs the resulting residue has $w$-value at most $4H\sum_j|a_j|\le4MH$. The assumed range $0\le z0$ for some fixed $r\in F_{>0}$. Choose, for example, \[ H=1-2\log m(r)>0, \qquad e^{-H}=e^{-1}m(r)^2H$, and hence $m(2r|\operatorname{Re}u|)J(d_0-1)$, some index occurs at least $d_0$ times. Choosing $d_0$ of those points produces a basis $B\in\mathcal B$ satisfying \begin{equation} \label{rd:good-basis} kv\in A_0\qquad(k\in B). \end{equation} Thus the basis may depend on $v$, while the family $\mathcal B$ has already been fixed. For each $B\in\mathcal B$ write the rational basis expansions \[ 1=\sum_{k\in B}b_{B,k}k,\qquad i=\sum_{k\in B}c_{B,k}k. \] Take a positive integer $D_0$ divisible by the denominators of all the $b_{B,k}$, over every $B\in\mathcal B$. Set $a_{B,k}=D_0b_{B,k}\in\Z$ and \[ M_0=\max_{B\in\mathcal B}\sum_{k\in B}|a_{B,k}|. \] This is a finite positive integer, and every candidate basis obeys \begin{equation} \label{rd:integer-selection} D_0\,1=\sum_{k\in B}a_{B,k}k,\qquad \sum_{k\in B}|a_{B,k}|\le M_0. \end{equation} For the second uniform choice put \[ \beta_B=\sum_{k\in B}|b_{B,k}|,\qquad \gamma_B=\sum_{k\in B}|c_{B,k}|,\qquad c_0=\min\left\{1,\, \min_{B\in\mathcal B} \frac{1}{\sqrt{\beta_B^2+\gamma_B^2}}\right\}>0. \] Both basis expansions represent nonzero elements, so the denominators in this expression are positive. For any complex $v$ with $|v|=1$ let $f_B(v)=\max_{k\in B}|\operatorname{Re}(kv)|$. Taking real parts in the two expansions after multiplying by $v$ gives \[ |\operatorname{Re}v|\le\beta_B f_B(v),\qquad |\operatorname{Im}v|=|\operatorname{Re}(iv)| \le\gamma_B f_B(v). \] Since $|v|^2=1$, it follows that \begin{equation} \label{rd:uniform-transversality} \max_{k\in B}|\operatorname{Re}(kv)| \ge \frac{1}{\sqrt{\beta_B^2+\gamma_B^2}} \ge c_0 \qquad(B\in\mathcal B,\ |v|=1). \end{equation} The constants $D_0,M_0,c_0$ have all been fixed before the next choice of $v$. The elements $k$ of each candidate basis lie in the fixed field $N$, and $kv$ lies in $K$ for every $v\in K$. Now let $v\in K$ satisfy $|\operatorname{Re}v|0. \end{equation} Multiply \eqref{rd:integer-selection} by $v$, take real parts, and then multiply by $2r$. This gives \[ 2rD_0\operatorname{Re}v =\sum_{k\in B}a_{B,k}\,2r\operatorname{Re}(kv). \] Choose $\varepsilon_0,\varepsilon_k\in\{1,-1\}$ so that \[ \varepsilon_0\operatorname{Re}v=|\operatorname{Re}v|, \qquad \operatorname{Re}(kv)=\varepsilon_k|\operatorname{Re}(kv)|. \] At a zero real part take the corresponding sign to be $1$. Then the nonnegative number \[ z=2rD_0|\operatorname{Re}v| =\sum_{k\in B} (\varepsilon_0a_{B,k}\varepsilon_k)p_k \] is an integer combination whose coefficient absolute-value sum is at most $M_0$. Also \[ 0\le z<2rc_0\le p. \] Lemma~\ref{rd:capped-residues}, with the data in \eqref{rd:cap-data}, now proves \begin{equation} \label{rd:small-radius} w\bigl(2rD_0|\operatorname{Re}v|\bigr)\le4M_0H \qquad\left(v\in K,\ |\operatorname{Re}v|<\frac{c_0}{D_0}\right). \end{equation} This includes $\operatorname{Re}v=0$ and any zero values among the individual $p_k$. Take any $s\in F$ with $00 \qquad\bigl(s\in F\cap(0,2rc_0)\bigr). \end{equation} Choose a nonnegative $\phi\in C_c(\R)$ with support contained in $(0,2rc_0)$ and with positive integral. The functions $\phi|_F$ and $t\mapsto\phi(t)m(t)$ belong to the domain of $I$. By \eqref{rd:interval-lower-bound} and positivity, \[ I\bigl(t\longmapsto\phi(t)m(t)\bigr) \ge e^{-4M_0H}I(\phi|_F) =e^{-4M_0H}\int_{\R}\phi(t)\,dt >0. \] This contradicts \eqref{rd:line-vanishes}. The functional acts on all the bounded, boundedly supported functions on $F$ used here, so the argument applies directly to $m$ with its original algebraic domain. It follows that $m(r)=0$ for every $r\in F_{>0}$. Radiality then gives $m(z)=0$ for every nonzero $z\in E$. For such $z$, \[ \left|\int_{\mathbb D}\psi_z\,d\mu\right| =\left|\int_Yg_z\,d\eta\right| \le\lVert g_z\rVert_2=0. \] Let $m_{\mathbb D}$ be Haar probability on the compact abelian group $\mathbb D$. For $z\ne0$, an ordinary continuous character $j(\xi)\in\mathbb D$ can be chosen with $\psi_z(j(\xi))=e^{i\xi\cdot z}\ne1$. Thus $\psi_z$ is a nontrivial character of $\mathbb D$, and translation invariance of $m_{\mathbb D}$ forces $\int\psi_z\,dm_{\mathbb D}=0$. The characters $\psi_z$ contain the constants, are closed under multiplication and conjugation, and separate points of $\mathbb D$. The Stone--Weierstrass theorem makes their complex linear span dense in the continuous functions on this compact space. The two probabilities have the same integral on this dense span, and hence on every continuous function. Uniqueness of Borel probability measures on a compact metrizable space gives $\mu=m_{\mathbb D}$. Finally let $\mu$ be any wild $K$-invariant probability. The Farrell--Varadarajan ergodic decomposition theorem applies to the probability-preserving action of the countable group $K$ on the standard Borel space $\mathbb D$ \cite[Theorem~3.1]{Chen2018}. It writes \[ \mu=\int \mu_\omega\,d\lambda(\omega), \] where almost every $\mu_\omega$ is $K$-invariant and ergodic. The set $C$ is Borel and $K$-invariant, so \[ 0=\mu(C)=\int\mu_\omega(C)\,d\lambda(\omega). \] Nonnegativity gives $\mu_\omega(C)=0$ for almost every $\omega$. The ergodic case proved above identifies each of these $\mu_\omega$ with $m_{\mathbb D}$. Integrating this equality gives $\mu=m_{\mathbb D}$, which proves the theorem. \end{proof} % END module % BEGIN attributed rewritten module transfer_and_706.tex \section{Transfer to ordinary colorings}\label{tr:section} The proper-to-weak transfer for the unit distance, including the continuous-factor construction below, is due to OpenAI \cite{OpenAI2026}. Sergi Gal\'an extended the argument to algebraic forbidden distances and combined it with the separated-distance theorem of Bukh to answer Erd\H{o}s's exponential-growth question \cite{Galan2026}. We give the finite-distance transfer and that deduction in full, using the rigidity theorem proved above. Retain the notation \[ F=\overline{\Q}\cap\R,\qquad E=F(i),\qquad K=\{u\in E:u\overline u=1\},\qquad \mathbb D=\widehat{E_{\mathrm{disc}}}. \] The group $E$ is additive and $K$ is multiplicative. For $\chi\in\mathbb D$ and $z\in E$, write \[ \psi_z(\chi)=\chi(z),\qquad (u\chi)(z)=\chi(uz),\qquad C=j(\R^2),\quad j(\xi)(z)=e^{i\xi\cdot z}. \] The ordinary-character subgroup $C$ is Borel in $\mathbb D$. Indeed, $j$ is continuous and injective, and its image is the countable union of the compact images of closed frequency balls. Its inverse on $C$ is Borel, since each compact restriction is a homeomorphism onto its image. Theorem~\ref{sp:rigidity} states that a $K$-invariant Borel probability measure on $\mathbb D$ assigning zero mass to $C$ is Haar probability. For $X\subseteq\R^2$ and $D\subset(0,\infty)$, let $G(X,D)$ have vertex set $X$, with an edge between distinct $x,y$ precisely when $|x-y|\in D$. Write $\chi(\R^2,D)=\chi(G(\R^2,D))$. Let $\sigma$ denote normalized arc length on $S^1$. A measurable set $A\subseteq\R^2$ \emph{weakly avoids} $D$ if \begin{equation}\label{tr:weakavoidance} \int_{B(0,R)}\int_{S^1} 1_A(x)1_A(x+du)\dd\sigma(u)\dd x=0 \qquad(d\in D,\ R>0). \end{equation} A weak measurable $k$-coloring is a measurable partition into $k$ classes satisfying \eqref{tr:weakavoidance}. As throughout the paper, $m(D)$ uses the upper densities of measurable sets that avoid every distance in $D$ pointwise. \begin{theorem}[OpenAI and Gal\'an]\label{tr:transfer} Let $D\subset F\cap(0,\infty)$ be finite and nonempty, and let $k\ge1$ be an integer. The following assertions are equivalent. \begin{enumerate} \item $G(E,D)$ has a proper $k$-coloring. \item $G(\R^2,D)$ has a proper $k$-coloring. \item There is a weak measurable $k$-coloring of $\R^2$ avoiding $D$. \end{enumerate} Consequently, \begin{equation}\label{tr:chromaticdensity} \chi(\R^2,D)\ge\frac1{m(D)}. \end{equation} \end{theorem} \subsection{Continuous vectors form a positive factor} In the next lemma, the acting group $E$ initially has the discrete topology. Continuity of a vector refers to the usual subspace topology on $E\subset\R^2$. We use the inner product $\langle v,w\rangle=\int\overline v w$, which is linear in the second variable. \begin{lemma}\label{tr:continuousfactor} Suppose a standard probability space $(\Omega,\mu)$ carries a measure-preserving action of $E$, with Koopman operators $T_z$. The subspace \[ \mathcal H_c= \left\{v\in L^2(\mu): \lim_{\substack{z\to0\\z\in E}}\|T_zv-v\|_2=0\right\} \] is $L^2(\mathcal F_c)$ for a completed sub-$\sigma$-algebra $\mathcal F_c$. Its orthogonal projection is conditional expectation onto $\mathcal F_c$. It is also the range of the spectral projection on $C\subset\mathbb D$. On $\mathcal H_c$, the translations extend uniquely to a strongly continuous unitary representation $(T_x)_{x\in\R^2}$ that preserves nonnegative functions and fixes the constant one. If unitary Koopman operators $V_u$, $u\in K$, satisfy \[ V_uT_zV_u^{-1}=T_{uz}, \] then the orthogonal projection onto $\mathcal H_c$ commutes with every $V_u$. \end{lemma} \begin{proof} The subspace $\mathcal H_c$ is closed, because \[ \|T_zv-v\|_2\le2\|v-w\|_2+\|T_zw-w\|_2. \] It contains constants, is closed under complex conjugation, and is invariant under $T_z$ and $T_z^{-1}$. The spectral theorem for the discrete abelian group $E$ gives a projection-valued measure $\mathsf P$ on $\mathbb D$. For $v\in L^2(\mu)$, its finite positive scalar measure $\lambda_v(A)=\|\mathsf P(A)v\|_2^2$ satisfies \begin{equation}\label{tr:spectralidentities} \langle v,T_zv\rangle=\int_{\mathbb D}\psi_z(\chi)\dd\lambda_v(\chi), \qquad \|T_zv-v\|_2^2= \int_{\mathbb D}|\psi_z(\chi)-1|^2\dd\lambda_v(\chi). \end{equation} The applicable spectral theorem and Bochner's theorem are stated in \cite[Theorems D.3.1 and D.2.2]{BHV2008}. If $\lambda_v(\mathbb D\setminus C)=0$, the second identity and dominated convergence imply $v\in\mathcal H_c$. Conversely, suppose $v\in\mathcal H_c$. The function $F_v(z)=\langle v,T_zv\rangle$ is uniformly continuous on $E$, since \[ |F_v(z+h)-F_v(z)|\le\|v\|_2\|T_hv-v\|_2. \] It extends uniquely to a continuous function on $\R^2$. For any finite list of points in $\R^2$, approximation by points of $E$ passes the positive-definite quadratic inequality to the limit. The extension is therefore positive definite. Bochner's theorem gives a finite positive measure $\rho_v$ on $\R^2$ whose Fourier transform is this extension. The pushforward $j_*\rho_v$ and $\lambda_v$ have identical integrals against every $\psi_z$. The complex linear span of the functions $\psi_z$ is a unital self-adjoint algebra separating points of $\mathbb D$. Stone--Weierstrass and uniqueness of finite measures on a compact space give \[ \lambda_v=j_*\rho_v. \] Thus $\lambda_v$ is carried by $C$. This proves that $\mathcal H_c$ is the range of $\mathsf P(C)$. We next establish the factor assertion, which will preserve the color probabilities. If bounded $v,w$ belong to $\mathcal H_c$, then \[ \|T_z(vw)-vw\|_2 \le\|v\|_\infty\|T_zw-w\|_2+ \|w\|_\infty\|T_zv-v\|_2. \] Hence the bounded members form a unital self-adjoint algebra $\mathcal A$. For a Lipschitz map $\Phi:\C\to\C$, the Koopman property also gives \[ \|T_z\Phi(v)-\Phi(v)\|_2 \le\operatorname{Lip}(\Phi)\|T_zv-v\|_2. \] In particular, projection of $v$ onto a closed complex disk produces bounded members of $\mathcal H_c$ converging to $v$ in $L^2$. Thus the $L^2$ closure of $\mathcal A$ is $\mathcal H_c$. Let $\mathcal F_c$ be the completed $\sigma$-algebra generated by $\mathcal A$. For any finite family from $\mathcal A$, polynomials in their coordinates and conjugates approximate continuous functions on the closure of their joint range. The joint range is bounded, and its closure is compact. By regularity of the pushforward probability, these continuous functions are dense in the $L^2$ space of the joint distribution. The monotone-class theorem then shows that finite-coordinate functions are dense in $L^2(\mathcal F_c)$. Consequently $L^2(\mathcal F_c)\subseteq\mathcal H_c$. The reverse inclusion follows from the bounded truncations just constructed, so equality holds. The orthogonal projection is therefore conditional expectation. In particular, it sends nonnegative functions to nonnegative functions and fixes one. For completeness, if the conditional expectation of $f\ge0$ were negative on a positive-measure $\mathcal F_c$-measurable set, its integral on that set would be negative, contradicting its defining equality with the integral of $f$. For $x\in\R^2$, choose $z_n\in E$ with $z_n\to x$. For $v\in\mathcal H_c$, \[ \|T_{z_n}v-T_{z_m}v\|_2= \|T_{z_n-z_m}v-v\|_2\longrightarrow0. \] Define $T_xv$ to be the limit. The same identity gives independence of the approximating sequence and strong continuity in $x$. Approximating sums and negatives gives the group law, inverses, and unitarity. The positive cone of $L^2$ is closed, so these extensions preserve positivity and fix one. Finally, covariance implies \[ \|T_zV_uv-V_uv\|_2= \|T_{u^{-1}z}v-v\|_2\longrightarrow0. \] Thus $V_u$ and $V_u^{-1}$ preserve $\mathcal H_c$. The orthogonal projection onto this subspace commutes with $V_u$. \end{proof} \subsection{An invariant probability on proper labelings} We give the averaging argument explicitly. The elementary amenability facts used here can also be found in \cite[Appendix G]{BHV2008}. \begin{lemma}\label{tr:labelinglaw} If $G(E,D)$ has a proper $k$-coloring, there is a standard probability space of proper labelings on which translations by $E$ and rotations by $K$ act measure preservingly. The origin-label indicators $f_1,\ldots,f_k$ satisfy \begin{equation}\label{tr:originlabels} 0\le f_i\le1,\qquad \sum_i f_i=1,\qquad V_uf_i=f_i, \qquad\langle f_i,T_zf_i\rangle=0 \quad(z\in dK,\ d\in D). \end{equation} \end{lemma} \begin{proof} The proper labelings form a nonempty closed subset $\Omega$ of $\{1,\ldots,k\}^E$. Each edge imposes one closed cylinder condition. Since $E$ is countable, $\Omega$ is compact metrizable. The group $\Gamma=E\rtimes K$, with multiplication \[ (z,u)(w,v)=(z+uw,uv), \] acts continuously by \[ ((z,u)\omega)(x)=\omega\bigl(u^{-1}(x-z)\bigr). \] This action preserves the proper-labeling conditions. We first construct finite left F\o lner sets in $\Gamma$. An abelian group admits a finite set approximately invariant under any prescribed finite set of translations. Indeed, the prescribed elements generate a finitely generated abelian group, which is a product of a finite group and finitely many infinite cyclic groups. Products of the finite group with sufficiently large integer boxes have the required property. Given a finite $S\subset\Gamma$ and $\eta>0$, choose a finite nonempty $B\subset K$ such that \[ |uB\mathbin\triangle B|\le\eta|B| \qquad((z,u)\in S). \] Choose a finite nonempty $A\subset E$ such that \[ |(A+v^{-1}z)\mathbin\triangle A|\le\eta|A| \] for every $(z,u)\in S$ and every $v\in B\cup uB$. Only finitely many additive translations occur here. Put \[ F_S=\{(va,v):v\in B,\ a\in A\}. \] Left multiplication by $(z,u)$ sends the fiber above $v$ to the fiber above $uv$, with additive coordinate translated by $(uv)^{-1}z$. Fibers with rotation coordinate in $B\mathbin\triangle uB$ contribute at most $\eta|A||B|$ to the symmetric difference. The common fibers contribute at most the same quantity. Hence \[ |(z,u)F_S\mathbin\triangle F_S|\le2\eta|F_S| \qquad((z,u)\in S). \] Enumerating $\Gamma$ and taking increasing finite $S$ and $\eta\downarrow0$ gives finite nonempty sets $F_n$ with $|gF_n\mathbin\triangle F_n|/|F_n|\to0$ for every $g\in\Gamma$. Fix one proper labeling $\omega_0$, and let \[ \mu_n=\frac1{|F_n|}\sum_{g\in F_n}\delta_{g\omega_0}. \] Compactness of the probability measures on $\Omega$ gives a weakly convergent subsequence with limit $\mu$. For continuous $h$ on $\Omega$ and fixed $g\in\Gamma$, the difference between the integrals of $h$ and $h\circ g$ against $\mu_n$ is bounded by \[ 2\|h\|_\infty\frac{|gF_n\mathbin\triangle F_n|}{|F_n|}. \] It tends to zero, so $\mu$ is $\Gamma$-invariant. Use the Koopman convention $U_gf(\omega)=f(g^{-1}\omega)$, and write $T_z=U_{(z,1)}$ and $V_u=U_{(0,u)}$. For $f_i(\omega)=1_{\{\omega(0)=i\}}$, the function $T_zf_i$ indicates that $\omega(z)=i$. Rotations fix the origin, and properness forbids equal labels at $0$ and every $z\in dK$. This proves \eqref{tr:originlabels} and the covariance used in Lemma~\ref{tr:continuousfactor}. \end{proof} \subsection{Rigidity removes the residual correlations} Apply Lemma~\ref{tr:continuousfactor} to the invariant labeling law, and write \[ p_i=\mathsf P(C)f_i=\E(f_i\mid\mathcal F_c),\qquad b_i=f_i-p_i. \] Then $p_i,b_i$ are real and rotation invariant, and \begin{equation}\label{tr:positivepartition} 0\le p_i\le1,\qquad\sum_i p_i=1. \end{equation} For the scalar spectral measure of $b_i$, covariance gives \[ \int\psi_{uz}\dd\lambda_{b_i} =\langle b_i,T_{uz}b_i\rangle =\langle b_i,T_zb_i\rangle =\int\psi_z\dd\lambda_{b_i}. \] Fourier uniqueness on $\mathbb D$ shows that $\lambda_{b_i}$ is $K$-invariant. Since $\mathsf P(C)b_i=0$, it assigns zero mass to $C$. If its mass is positive, normalize it and apply Theorem~\ref{sp:rigidity}. Including the zero-mass case, the result is \begin{equation}\label{tr:residualhaar} \lambda_{b_i}=\|b_i\|_2^2m_{\mathbb D},\qquad \langle b_i,T_zb_i\rangle=0\quad(z\in E\setminus\{0\}), \end{equation} where $m_{\mathbb D}$ is Haar probability. The last equality follows from Haar orthogonality. In fact, for $z\ne0$ one can choose an ordinary character $\chi_0$ with $\chi_0(z)\ne1$. Translation of Haar measure by $\chi_0$ then multiplies its $z$th Fourier coefficient by $\chi_0(z)$ while preserving that coefficient, so it is zero. Both $\mathcal H_c$ and its orthogonal complement reduce every $T_z$. Thus the cross terms in $\langle f_i,T_zf_i\rangle$ vanish. Equations \eqref{tr:originlabels} and \eqref{tr:residualhaar} imply \[ \langle p_i,T_zp_i\rangle=0\qquad(z\in dK,\ d\in D). \] The algebraic unit directions are dense in $S^1$. For example, \[ \frac{1-t^2+2it}{1+t^2}\in K\qquad(t\in\Q) \] forms a dense subset as rational $t$ varies. Since $d\in F$, the set $dK$ consists of algebraic points and is dense in the circle of radius $d$. Strong continuity consequently gives \begin{equation}\label{tr:continuouszero} \langle p_i,T_vp_i\rangle=0 \qquad(v\in\R^2,\ |v|\in D). \end{equation} Unitarity and the group law give, for every $x\in\R^2$, \begin{equation}\label{tr:translatedzero} \langle T_xp_i,T_{x+v}p_i\rangle=0 \qquad(|v|\in D). \end{equation} \subsection{Jointly measurable fields and a single sample} We now realize the continuous $L^2$-valued maps by measurable fields and keep track of both translated endpoints. Choose measurable rounding maps $q_n:\R^2\to\Q^2\subset E$ with countable range and $|q_n(x)-x|\le2^{-n}$. For each rational point in the union of these ranges, choose Borel representatives of all $T_qp_i$ simultaneously. They may be chosen with values in $[0,1]$ and sum one everywhere, by modifying a common null set for this countable family. Let $H_i^{(n)}(x,\omega)$ be the representative at $q_n(x)$. These functions are jointly measurable and take values in the probability simplex. Moreover, \[ \sup_x\|H_i^{(n)}(x,\cdot)-T_xp_i\|_2 \le\sup_{|h|\le2^{-n}}\|T_hp_i-p_i\|_2\longrightarrow0. \] Pass to a subsequence for which this supremum is at most $2^{-n}$, simultaneously for the finitely many $i$. On each bounded spatial domain, the successive differences then have summable product-space $L^2$ norms and hence summable $L^1$ norms. Their pointwise absolute sums are finite almost everywhere, by Tonelli. Consequently $H_i^{(n)}$ converges almost everywhere on $\R^2\times\Omega$. Define $H_i$ by this limit where it exists, and assign the fixed simplex vector $(1,0,\ldots,0)$ elsewhere. Then the $H_i$ are jointly measurable, belong to $[0,1]$, and sum to one everywhere. For almost every $x$, bounded convergence in $\omega$ and the displayed $L^2$ convergence imply \begin{equation}\label{tr:fieldrepresentative} H_i(x,\cdot)=T_xp_i\quad\text{in }L^2(\mu). \end{equation} Let $N\subset\R^2$ be a common null set outside which \eqref{tr:fieldrepresentative} holds for all $i$. For each $d\in D$ and positive integer $m$, \begin{equation}\label{tr:endpointnull} \int_{B(0,m)}\int_{S^1} \bigl(1_N(x)+1_N(x+du)\bigr)\dd\sigma(u)\dd x=0. \end{equation} For the second term, fix $u$, use translation invariance of Lebesgue measure, and then integrate in $u$. Thus \eqref{tr:translatedzero} applies to the selected representatives for almost every pair $(x,u)$. Since the fields are real and nonnegative, Tonelli gives \[ \int_\Omega\int_{B(0,m)}\int_{S^1} H_i(x,\omega)H_i(x+du,\omega) \dd\sigma(u)\dd x\dd\mu(\omega)=0. \] There are only finitely many $i,d$ and countably many $m$. Choose one $\omega$ for which all the inner integrals vanish. Assign $x$ the least index $i$ with $H_i(x,\omega)>0$. This gives a measurable coloring, because the fields sum to one. For each $i,d,m$, the positive-product locus has zero $dx\dd\sigma(u)$ measure. A pair assigned color $i$ has positive $H_i$ at both endpoints. The corresponding color class therefore satisfies \eqref{tr:weakavoidance}. This proves the implication from assertion 1 to assertion 3 of Theorem~\ref{tr:transfer}. \subsection{Density points and finite-color compactness} \begin{lemma}\label{tr:density} If a measurable set $A$ weakly avoids $D$, then its set $A^*$ of Lebesgue density-one points avoids $D$ pointwise and satisfies $|A\mathbin\triangle A^*|=0$. For a measurable partition into finitely many classes, their density-one sets are disjoint and have conull union. \end{lemma} \begin{proof} Suppose $x,y\in A^*$ with $v=y-x$ and $|v|=d\in D$. For small $h>0$, take $X$ uniformly in $B(x,h)$ and, independently, take $U$ uniformly on the positive-length arc \[ \{u\in S^1:|du-v|1, \] because $\sqrt{21}>9/2$. Assign each boundary point to one of its incident cells. The strict bounds remain valid, proving the seven-coloring. For $D=\{d_1,\ldots,d_r\}$, take the product of the seven-colorings obtained by scaling this construction by each $d_i$. If two points have distance $d_i$, their $i$th coordinates differ. The resulting coloring has at most $7^r$ colors. \end{proof} \subsection{Integer power lists and the limiting density} Put $m_1=m(\{1\})$. One has $m_1>0$, for example from sufficiently small squares on $2\Z^2$. Also \begin{equation}\label{tr:half} m_1\le\frac12. \end{equation} Indeed, a unit-distance-avoiding set $A$ and its translate by a unit vector are disjoint. Their areas in $[-R,R]^2$ differ by $O(R)$, since a fixed translation changes the square by a region of area $O(R)$. Dividing the disjointness inequality by $4R^2$ and taking the upper limit proves \eqref{tr:half}. Bukh's separated-distance product theorem, applied to the fixed list of $r$ singleton distance sets, gives \begin{equation}\label{tr:bukhlimit} m(D_{r,q})\longrightarrow m_1^r \quad\text{as }q\to\infty,\qquad D_{r,q}=\{1,q,q^2,\ldots,q^{r-1}\}. \end{equation} This is \cite[Theorem 1]{Bukh2008}. The parameter $r$ is fixed when the limit is taken. The limit remains valid along integers $q\ge2$, when $D_{r,q}$ has exactly $r$ distinct positive algebraic distances. For $r=1$, the statement is the identity $m(D_{1,q})=m_1$. Let \[ n_r=\lceil m_1^{-r}\rceil. \] By \eqref{tr:bukhlimit}, for every sufficiently large integer $q$, \begin{equation}\label{tr:rounding} \frac1{m(D_{r,q})}>n_r-1. \end{equation} This also holds when $m_1^{-r}$ is an integer, because $n_r-1$ is strictly smaller than that limit. If $G(E,D_{r,q})$ admitted a proper $(n_r-1)$-coloring, Theorem~\ref{tr:transfer} would give $m(D_{r,q})\ge1/(n_r-1)$, contradicting \eqref{tr:rounding}. Here $n_r\ge2$ by \eqref{tr:half}. Finite-color compactness therefore supplies a finite $P\subset E$ with \begin{equation}\label{tr:finitegrowth} \chi(G(P,D_{r,q}))\ge n_r\ge m_1^{-r}. \end{equation} Equivalently, apply the finite intersection property to the proper-labeling constraints on the countable vertex set $E$. If every finite induced graph were $(n_r-1)$-colorable, those constraints would give such a coloring of all of $E$. The elementary input \eqref{tr:half} already implies \begin{equation}\label{tr:numericalfree} L(r)\ge2^r. \end{equation} Thus this exponential lower bound requires no numerical optimization of the unit-distance density. \subsection{The numerical bound and Erd\H{o}s 706} The theorem of Ambrus, Csisz\'arik, Matolcsi, Varga and Zs\'amboki gives \begin{equation}\label{tr:acmvz} m_1\le0.2470=\frac{247}{1000} \end{equation} for measurable planar sets avoiding the unit distance \cite{ACMVZ2023}. This is the published density bound used for the numerical conclusion below. The exact comparison \[ \frac{1000}{247}-4.0485 =\frac{1000}{247}-\frac{8097}{2000} =\frac{41}{494000}>0 \] makes the strict numerical inequality independent of decimal rounding. \begin{theorem}[Gal\'an, resolving Erd\H{o}s 706]\label{tr:706} For every integer $r\ge1$, there are an integer $q\ge2$ and a finite set of algebraic planar points $P\subset E$ such that \[ \chi\bigl(G(P,\{1,q,\ldots,q^{r-1}\})\bigr) \ge\lceil m_1^{-r}\rceil>4.0485^r. \] Consequently \begin{equation}\label{tr:finalbounds} 4.0485^r0$ and every $c\ge0$, all sufficiently large $r$ admit a finite distance graph with $r$ forbidden distances and chromatic number greater than $Ar^c$. \end{theorem} \begin{proof} Equation \eqref{tr:finitegrowth} gives the finite graph, with integer power distances. By \eqref{tr:acmvz}, \[ m_1^{-r}\ge(1000/247)^r>4.0485^r. \] Lemma~\ref{tr:seven} gives the upper bound in \eqref{tr:finalbounds}. The numerical-free statement follows from \eqref{tr:numericalfree}. Finally, $2^r/(Ar^c)\to\infty$, since $r\log2-c\log r-\log A\to\infty$. This proves the stated failure of every polynomial upper bound. \end{proof} % END module \part{Uniform cluster bounds and quantitative amplification} \section{Uniform refinements and their predecessors} We retain the independence density $m(D)$ defined in Section~\ref{in:intro}. Dilation gives $m(tD)=m(D)$ for every $t>0$. Density estimates for sets with forbidden distances have a long history. Early work includes Sz\'ekely \cite{Szekely1984} and Sz\'ekely and Wormald \cite{SzekelyWormald1989}. Bukh proved a separated-scale product limit for a fixed finite family of finite distance sets, and more generally for a fixed family of supersaturable properties \cite[Theorems 1 and 8]{Bukh2008}. The corresponding translation-product lower bound appears in \cite[Lemma 6]{Bukh2008}. Castro-Silva obtained product limits for fixed families of admissible configurations, together with continuity results for their independence densities \cite[Theorems 4 and 5]{CastroSilva2023}. Our first result allows both the length of the family and all of its cluster shapes to vary after a single separation ratio has been chosen. Fix $M\ge1$, and let $\cK_M$ be the Hausdorff metric space of nonempty compact subsets of $[1,M]$. In particular, cluster cardinalities and internal spacings are unrestricted. \begin{theorem}\label{thm:clusters} Let $M\ge1$ and $0<\varepsilon<1/2$. There is $Q=Q(M,\varepsilon)>M$ such that, for every integer $r\ge1$, every $K_1,\ldots,K_r\in\cK_M$, and every sequence of positive scales satisfying \[ \frac{t_{i+1}}{t_i}\ge Q\qquad(1\le iM$, define \[ \mathcal R_M(Q)= \sup\left\{ \left(\frac{m(\bigcup_{i=1}^r t_iK_i)} {\prod_{i=1}^r m(K_i)}\right)^{1/r}: \begin{array}{l} r\ge1,\ K_i\in\cK_M,\ t_i>0,\\ t_{i+1}/t_i\ge Q\ (1\le i1$ and $0<\kappa\le1/2$ such that, for every $r\ge1$ and every positive distance list $d_1<\cdots\sup D$. For a $D$-avoiding set $A$, take restrictions $B_L=A\cap[-L,L]^2$ along a sequence realizing its upper density, and form \[ A_L=B_L+(2L+R)\mathbb Z^2. \] Points in different translates of $B_L$ are at distance at least $R$. Thus $A_L$ avoids $D$, and its density is \[ \frac{|B_L|}{(2L+R)^2}, \] which approaches $\overline\delta(A)$ along that sequence. If centers are optimized as well, translate each witnessing restriction before periodizing it. Periodic sets have the same density under either averaging convention, proving the last assertion. \end{proof} \begin{lemma}\label{lem:product} For bounded sets $D_1,\ldots,D_r\subset(0,\infty)$, \[ m\!\left(\bigcup_{i=1}^rD_i\right)\ge\prod_{i=1}^r m(D_i). \] \end{lemma} \begin{proof} Choose periodic sets $A_i$ avoiding $D_i$, with densities $\delta_i$ arbitrarily close to $m(D_i)$. Let $X_i$ be independent uniform translations in fundamental domains for their respective periods. For every fixed square $S$, Fubini's theorem gives \[ \E\frac{|S\cap\bigcap_i(A_i+X_i)|}{|S|} =\prod_i\delta_i. \] The random variables on the left lie in $[0,1]$. Reverse Fatou along the squares $[-n,n]^2$ gives a choice of translations for which the upper density of the intersection is at least $\prod_i\delta_i$. This intersection avoids every $D_i$. Letting the $\delta_i$ approach their suprema proves the claim. This argument also accommodates incommensurate period lattices. \end{proof} To prove \eqref{eq:elementary}, place copies of $[0,1/4]^2$ on the lattice $(M+1)\mathbb Z^2$. The diameter of each copy is below one, and points in different copies are at distance at least $M+3/4$. The resulting density is $a_M$. For the upper bound, choose $d\in K$. If $A$ avoids $K$, then $A$ and $A+de_1$ are disjoint. Their areas in a square of side $2L$ differ by $O(dL+d^2)$, so averaging and taking the upper limit gives $2\overline\delta(A)\le1$. \subsection{The radial spectral measure} We use the Fourier phase $e^{-i\xi\cdot x}$. Let $\Lambda\subset\R^2$ be a full-rank lattice, and equip the torus $\R^2/\Lambda$ with probability Haar measure. Its dual frequencies are \[ \Lambda^*=\{\xi\in\R^2:\xi\cdot\lambda\in2\pi\mathbb Z \text{ for all }\lambda\in\Lambda\}. \] For real $f\in L^2(\R^2/\Lambda)$, set \[ u=\int f,\qquad v=\int f^2,\qquad C_f(x)=\int f(y)f(y+x)\dd y. \] Parseval's identity and $L^2$ continuity of translation give \[ C_f(x)=\sum_{\xi\in\Lambda^*}|\widehat f(\xi)|^2e^{i\xi\cdot x}. \] The series is absolutely and uniformly convergent. Consequently its rotational mean has the representation \begin{equation}\label{eq:spectral} p_f(d):=\frac1{2\pi}\int_0^{2\pi}C_f(d\omega_\theta)\dd\theta =\int_0^\infty J_0(ds)\dd\mu_f(s), \qquad \mu_f=\sum_{\xi\in\Lambda^*}|\widehat f(\xi)|^2\delta_{|\xi|}, \end{equation} where $\omega_\theta=(\cos\theta,\sin\theta)$. The measure $\mu_f$ has total mass $v$ and zero atom $u^2$. If $f=1_A$ for a periodic set of positive density $\delta$, then \begin{equation}\label{eq:nu} \nu_A:=\delta^{-1}\mu_{1_A} \quad\text{is a probability measure with}\quad \nu_A(\{0\})=\delta. \end{equation} If $A$ avoids $d$, then \begin{equation}\label{eq:zero} \int J_0(ds)\dd\nu_A(s)=0. \end{equation} For $f\ge0$, one also has $p_f(d)\ge0$ for every $d$. \subsection{Bessel estimates and thick copositive kernels} We will use the following elementary bounds throughout. \begin{lemma}\label{lem:bessel} For $t>0$, \begin{equation}\label{eq:bessel} |J_0(t)|\le\min\left\{1,\frac4{5\sqrt t}\right\}, \qquad |J_1(t)|\le1, \qquad 0\le1-J_0(t)\le\frac{t^2}{4}. \end{equation} \end{lemma} \begin{proof} The integral representations of $J_0$ and $J_1$ give their bounds by one. The representation $J_0(t)=\E\cos(t\cos\Theta)$ for uniform $\Theta$ also gives $1-J_0(t)\le t^2\E\cos^2\Theta/2=t^2/4$. For the decay estimate, put $w(t)=\sqrt tJ_0(t)$ and $q(t)=1+1/(4t^2)$. Bessel's differential equation gives $w''+qw=0$. Thus \[ S(t)=w(t)^2+\frac{w'(t)^2}{q(t)}, \qquad S'(t)=-\frac{q'(t)w'(t)^2}{q(t)^2}\ge0. \] The usual large-argument expansions of $J_0$ and its derivative yield $S(t)\to2/\pi$ as $t\to\infty$. It follows that $tJ_0(t)^2\le2/\pi<16/25$. The differential equation, integral representations, and asymptotic expansions used here are recorded in \cite[Sections 10.2, 10.9 and 10.17]{DLMF}. \end{proof} A real symmetric matrix $Z=(Z_{ij})$ is copositive if $x^{\mathsf T}Zx\ge0$ for every vector $x$ with nonnegative entries. Given a nonempty finite set $U\subset\R^2$, a copositive matrix $Z$ indexed by $U$, and $\rho>0$ small enough that the balls $B(u,\rho)$ are disjoint, define \begin{equation}\label{eq:thick} r(s)=\sum_{u,v\in U}Z_{uv} \int_{B(u,\rho)}\int_{B(v,\rho)}J_0(s|x-y|)\dd y\dd x. \end{equation} We call these functions thick copositive kernels. For a singleton $U$, any $\rho>0$ is allowed. They satisfy \begin{equation}\label{eq:thickfactor} r(s)=\left|\widehat{1_{B(0,\rho)}}(se_1)\right|^2 \sum_{u,v\in U}Z_{uv}J_0(s|u-v|), \qquad \widehat{1_{B(0,\rho)}}(se_1)=\frac{2\pi\rho J_1(\rho s)}s. \end{equation} The value at $s=0$ is understood by continuity. Rotational averaging proves the factorization. In particular $r$ is continuous, tends to zero at infinity, and \begin{equation}\label{eq:thickconstants} r(0)\ge0,\qquad |r(s)|\le D_r s^{-2}\ (s>0), \qquad |r(s)-r(0)|\le C_r s^2, \end{equation} where one may take \begin{align*} D_r&=4\pi^2\rho^2\sum_{u,v}|Z_{uv}|,\\ C_r&=\frac{\pi^2\rho^4}{4} \sum_{u,v}|Z_{uv}|(|u-v|+2\rho)^2. \end{align*} The first inequality uses copositivity on the all-ones vector. The other two follow from \eqref{eq:thickfactor} and Lemma~\ref{lem:bessel}. For $f\ge0$ on a torus and any $t>0$, copositivity also gives \begin{equation}\label{eq:thickmoment} \int r(ts)\dd\mu_f(s)\ge0. \end{equation} Indeed, for each translation $z$ and angle $\theta$, insert the nonnegative vector \[ \left(\int_{B(u,\rho)} f(z+tR_\theta x)\dd x\right)_{u\in U} \] into $Z$, then integrate over $z$ and $\theta$. Fubini and \eqref{eq:spectral} identify the result with the left side of \eqref{eq:thickmoment}. The calculation is justified by Cauchy--Schwarz and the finite measures of the balls. \subsection{The completely positive input} A bounded continuous real function $F$ on $\R^2$ is completely positive if, for every finite $U\subset\R^2$, the matrix $(F(u-v))_{u,v\in U}$ is completely positive. A finite matrix is completely positive when it is a sum of matrices $xx^{\mathsf T}$ with $x\ge0$. Such an $F$ is positive definite. For radial $F$ with $F(0)=1$, the radial Bochner representation is \[ F(x)=\int_0^\infty J_0(s|x|)\dd\sigma(s) \] for a probability measure $\sigma$. Its mean over expanding squares is $\sigma(\{0\})$. This last statement follows by first writing $J_0$ as a circle average of characters and then applying dominated convergence. We use two results of DeCorte, de Oliveira Filho and Vallentin. For closed $K\subset(0,\infty)$, $m(K)$ equals the supremal mean of continuous completely positive functions $F$ with $F(0)=1$ and $F(x)=0$ whenever $|x|\in K$. Rotation averaging allows $F$ to be radial. Moreover, a bounded continuous real $F$ is completely positive exactly when every thick copositive test obtained from \eqref{eq:thick} is nonnegative with $J_0(s|x-y|)$ replaced by $F(x-y)$. These are \cite[Theorem 6.3 and Lemma 10.1]{DOV2022}. Lemma~\ref{lem:periodic} identifies the density conventions in that characterization with ours. \section{Finite certificates and uniformity in the cluster}\label{rf:certificates} \subsection{A finite certificate for a compact cluster} The finite-distance form of the next lemma is stated in \cite[Appendix A, Lemma A.1]{Davies2026}. We give a compact-set version of the weak-* compactness and separation argument, retaining escaped spectral mass and the upper-semicontinuous zero-atom objective explicitly. \begin{lemma}\label{lem:finite} Let $K\in\cK_M$ and $m(K)<\lambda<1$. There are finitely many $a_j\in K$, real coefficients $z_j$, thick copositive kernels $r_k$, and coefficients $y_k\le0$ such that \begin{equation}\label{eq:certificate} g(s)=\sum_{j=1}^n z_jJ_0(a_js)+\sum_{k=1}^N y_kr_k(s) \end{equation} satisfies \begin{equation}\label{eq:certificatebounds} g(0)\ge1-\lambda>0,\qquad g(s)\ge-\lambda\quad(s\ge0). \end{equation} \end{lemma} \begin{proof} Let $X=[0,\infty]$ be the one-point compactification of the frequency half-line. Every $J_0(as)$ with $a\in K$, and every thick kernel, extends continuously to $X$ with value zero at infinity. Consider probability measures $\mu\in\cP(X)$ satisfying \begin{equation}\label{eq:allconstraints} \int J_0(as)\dd\mu(s)=0\quad(a\in K),\qquad \int r(s)\dd\mu(s)\ge0\quad\text{for every thick kernel }r. \end{equation} The maximum of $\mu(\{0\})$ over these measures is $m(K)$. To prove the upper bound, put $\gamma=\mu([0,\infty))$. If $\gamma>0$, the function \[ F(x)=\gamma^{-1}\int_{[0,\infty)}J_0(s|x|)\dd\mu(s) \] is continuous, has $F(0)=1$, and vanishes at distances in $K$. The thick tests in \eqref{eq:allconstraints} make it completely positive by the cited characterization. Its mean is $\mu(\{0\})/\gamma$, so \[ \mu(\{0\})\le\gamma m(K)\le m(K). \] If $\gamma=0$, the atom at zero vanishes. For the reverse bound, the measures \eqref{eq:nu} of periodic $K$-avoiding sets satisfy every constraint and have zero atoms approaching $m(K)$, by Lemma~\ref{lem:periodic}. The space $\cP(X)$ is weakly compact. Each moment constraint in \eqref{eq:allconstraints} is closed. The function $\mu\mapsto\mu(\{0\})$ is upper semicontinuous, so its superlevel set at $\lambda$ is also closed. Since the intersection of all these closed sets is empty, compactness gives finitely many distance constraints and finitely many thick constraints for which no feasible measure has zero atom at least $\lambda$. For these selected constraints, let $\mathcal C$ consist of the triples $(A,B,c)$ for which some $\mu\in\cP(X)$ satisfies \[ A_j=\int J_0(a_js)\dd\mu(s),\qquad B_k\le\int r_k(s)\dd\mu(s),\qquad c\le\mu(\{0\}). \] This is a closed convex subset of a finite-dimensional real vector space. For closedness, take a convergent sequence of triples and a weakly convergent subsequence of witnessing measures. Continuity of the moments passes the first two conditions to the limit. For the final condition, upper semicontinuity gives \[ c\le\limsup_n\mu_n(\{0\})\le\mu(\{0\}). \] The point $(0,0,\lambda)$ lies outside $\mathcal C$. Strict separation therefore gives $(\alpha,\beta,\tau)$ with \begin{equation}\label{eq:separation} \tau\lambda> \sup_{(A,B,c)\in\mathcal C} (\alpha\cdot A+\beta\cdot B+\tau c). \end{equation} Because $\mathcal C$ is downward closed in the $B$ and $c$ coordinates, $\beta_k\ge0$ and $\tau\ge0$. The measure $\delta_\infty$ places $(0,0,0)$ in $\mathcal C$. Thus \eqref{eq:separation} forces $\tau>0$. Testing the measure $\delta_s$ in that inequality yields \[ \lambda-\sum_j\frac{\alpha_j}{\tau}J_0(a_js) -\sum_k\frac{\beta_k}{\tau}r_k(s) >1_{\{0\}}(s). \] Taking $z_j=-\alpha_j/\tau$ and $y_k=-\beta_k/\tau$ proves the lemma. \end{proof} \subsection{Hausdorff perturbations} \begin{lemma}\label{lem:lsc} The function $K\mapsto m(K)$ is lower semicontinuous on $\cK_M$. \end{lemma} \begin{proof} Fix $K\in\cK_M$ and $\eta>0$. Restrict a nearly optimal $K$-avoiding set to a sufficiently large square and then take a compact subset $B$ losing sufficiently little area. Periodizing $B$ with gaps wider than $M$ gives density at least $m(K)-\eta$. The compact set \[ S_B=\{|x-y|:x,y\in B\} \] is disjoint from $K$. Thus $\dist(S_B,K)>0$. Every sufficiently Hausdorff-close $K'\in\cK_M$ remains disjoint from $S_B$. The same periodic set then avoids $K'$, since all cross-copy distances exceed $M$. Therefore $m(K')\ge m(K)-\eta$ throughout a neighborhood of $K$. \end{proof} Fix $0<\varepsilon<1/2$. At each $K\in\cK_M$, choose Lemma~\ref{lem:finite} with \[ \lambda_K=m(K)+\frac{\varepsilon}{8}, \] and write $g_K$ for the certificate and $G_K=g_K(0)$. If $K'$ is close to $K$, choose $a'_j\in K'$ close to each certificate distance $a_j\in K$ and define \[ g_{K,K'}(s)=\sum_jz_jJ_0(a'_js)+\sum_ky_kr_k(s). \] Its value at zero remains $G_K$. The map $a\mapsto J_0(a\,\cdot)$ from $[1,M]$ to $C_0([0,\infty))$ is continuous in supremum norm. Indeed, Lemma~\ref{lem:bessel} bounds the tail uniformly, and uniform continuity applies on each remaining compact rectangle. After shrinking the neighborhood, this fact and Lemma~\ref{lem:lsc} give \begin{equation}\label{eq:perturb} \|g_{K,K'}-g_K\|_\infty\le\frac{\varepsilon}{8}, \qquad m(K')\ge m(K)-\frac{\varepsilon}{4}. \end{equation} Set \[ h_{K,K'}=G_K^{-1}g_{K,K'},\qquad \beta_K=\frac{m(K)+\varepsilon/4}{G_K},\qquad b(K')=(m(K')+\varepsilon)^{-1}. \] Then \begin{equation}\label{eq:uniformslack} h_{K,K'}(0)=1,\qquad h_{K,K'}\ge-\beta_K, \qquad (b(K')-1)\beta_K\le1-\Delta, \qquad \Delta=\frac{\varepsilon}{2}. \end{equation} To verify the last inequality, write $A=m(K)$ and $e=\varepsilon$. Using \eqref{eq:perturb} and $G_K\ge1-A-e/8$, the product on its left is at most \[ \frac{(1-A-3e/4)(A+e/4)}{(A+3e/4)(1-A-e/8)}. \] All factors are positive. The denominator minus the numerator equals \[ \frac e2+\frac{Ae}{8}+\frac{3e^2}{32}\ge\frac e2, \] and the denominator is at most one. This proves \eqref{eq:uniformslack}. The Hausdorff space $\cK_M$ is compact. Choose a finite subcover of the preceding neighborhoods, indexed by $\ell$. Write its unperturbed certificates as \[ g_\ell(s)=\sum_jz_{\ell j}J_0(a_{\ell j}s) +\sum_ky_{\ell k}r_{\ell k}(s), \qquad G_\ell=g_\ell(0),\qquad S_\ell=\sum_j|z_{\ell j}|. \] Only the finitely many distance locations are perturbed when a certificate is assigned to a member of its neighborhood. Define \begin{align} B&=\max_\ell\frac{(4/5)S_\ell+ \sum_k|y_{\ell k}|D_{r_{\ell k}}}{G_\ell}, \label{eq:Buniform}\\ C&=\max_\ell\frac{(M^2/4)S_\ell+ \sum_k|y_{\ell k}|C_{r_{\ell k}}}{G_\ell}, \qquad \eta=\min\left\{1,\sqrt{\frac{\varepsilon}{8C}}\right\}. \label{eq:Cuniform} \end{align} Here $C>0$. In fact $y_{\ell k}\le0$ and $r_{\ell k}(0)\ge0$ imply \begin{equation}\label{eq:zsum} \sum_j z_{\ell j}=G_\ell-\sum_k y_{\ell k}r_{\ell k}(0) \ge G_\ell>0. \end{equation} Every assigned normalized certificate $h$ satisfies \begin{equation}\label{eq:uniformmodulus} |h(s)|\le Bs^{-1/2}\quad(s\ge1),\qquad h(s)\ge1-Cs^2\ge1-\Delta/4\quad(0\le s\le\eta). \end{equation} These estimates follow from \eqref{eq:thickconstants} and Lemma~\ref{lem:bessel}. All constants have been chosen before any family length or cluster list. \section{Variable weights and the product estimate} \begin{lemma}\label{lem:variable} Suppose $h_1,\ldots,h_r$ satisfy \eqref{eq:uniformmodulus}, with $h_i(0)=1$ and $h_i\ge-\beta_i$. Suppose also that \[ 11$ and $\kappa>0$, it is strictly smaller than $P_r^{-1}$. The lower bound follows from Lemma~\ref{lem:product} and scale invariance. \end{proof} \begin{proof}[Proof of Corollary~\ref{cor:rootlimit}] Theorem~\ref{thm:clusters} and \eqref{eq:elementary} imply \[ 1\le\mathcal R_M(Q(M,\varepsilon))\le1+\frac{\varepsilon}{a_M}. \] The function $\mathcal R_M$ is nonincreasing. Letting $\varepsilon$ decrease to zero gives the result. \end{proof} \subsection{A pair-density consequence} The same finite cover also yields a uniform stability estimate for small pair densities. The torus and pair-density notation is that of \eqref{eq:spectral}. \begin{corollary}\label{cor:robust} For fixed $M$ and $\varepsilon$ as in Theorem~\ref{thm:clusters}, there is $L_0\ge1$ such that the following holds with the same $Q$ and $\kappa$. Let $r$, $K_i$, and $t_i$ satisfy the hypotheses of that theorem, and let $f\ge0$ be nonzero in $L^2$ of any flat planar torus. Write \[ \begin{gathered} u=\int f,\qquad v=\int f^2,\qquad q_i=\sup_{d\in K_i}p_f(t_id),\\ b_i=(m(K_i)+\varepsilon)^{-1},\qquad P_0=1,\qquad P_i=\prod_{j=1}^i b_j,\qquad w_i=(b_i-1)P_{i-1}. \end{gathered} \] Then \begin{equation}\label{eq:robust} L_0\sum_{i=1}^r w_iq_i \ge(P_r-\kappa)u^2-(1-\kappa)v. \end{equation} In particular, if $0\le\tau<1/L_0$ and $q_i\le\tau u^2$ for every $i$, then \[ \frac{u^2}{v}\le \frac{1-\kappa}{(1-L_0\tau)P_r+L_0\tau-\kappa}. \] \end{corollary} \begin{proof} For the finite cover in Section~\ref{rf:certificates}, set \[ L_0=\max_\ell G_\ell^{-1}\sum_j\max\{z_{\ell j},0\}. \] Equation \eqref{eq:zsum} gives $L_0\ge1$. By \eqref{eq:thickmoment} and $p_f\ge0$, each selected certificate satisfies \[ \int h_i(t_is)\dd\mu_f(s) \le G_{\ell(i)}^{-1}\sum_jz_{\ell(i)j}p_f(t_ia'_{ij}) \le L_0q_i. \] The pointwise lower bound for $H$, the mass $v$ of $\mu_f$, and its zero atom $u^2$ give \[ (P_r-\kappa)u^2+\kappa v \le\int H\dd\mu_f \le v+L_0\sum_iw_iq_i. \] This proves \eqref{eq:robust}. The final estimate uses $\sum_iw_i=P_r-1$ and rearranges. \end{proof} \section{Effective certificates from weighted graphs} Fix a finite unit-distance realization $(u_i)_{i\in V}$ and weights as in Theorem~\ref{thm:graph}. Throughout this section put \begin{equation}\label{eq:graphdata} D_0=\max_{i,j}|u_i-u_j|,\qquad z=a\sum_{\{i,j\}\in E}(p_i+p_j) =a\sum_i\deg(i)p_i. \end{equation} The assumptions imply $00$. Indeed, if the support of $p$ contained no edge, it would be an independent set of weight one. \begin{lemma}[Weighted Motzkin--Straus inequality]\label{lem:MS} For every $x_i\ge0$, \begin{equation}\label{eq:MS} a\left(\sum_i p_ix_i^2+ \sum_{\{i,j\}\in E}(p_i+p_j)x_ix_j\right) \ge\left(\sum_i p_ix_i\right)^2. \end{equation} Consequently the matrix \begin{equation}\label{eq:Z} Z_{ii}=ap_i-p_i^2,\qquad Z_{ij}=\frac a2(p_i+p_j)1_{\{i,j\}\in E}-p_ip_j\quad(i\ne j) \end{equation} is copositive. \end{lemma} \begin{proof} Denote the quadratic expression in parentheses by $q(x)$. Set every zero-weight coordinate to zero, which preserves $\sum_i p_ix_i$ and can only decrease $q$. If adjacent positive-weight coordinates $x_i,x_j$ are both positive, keep $S=p_ix_i+p_jx_j$ fixed and write $v=p_ix_i$. Their contribution to $q$ is \[ \frac{v^2}{p_i}+\frac{(S-v)^2}{p_j} +\left(\frac1{p_i}+\frac1{p_j}\right)v(S-v) =\frac{Sv}{p_i}+\frac{S(S-v)}{p_j}. \] It is affine for $0\le v\le S$. All terms involving other coordinates are affine as well. An endpoint therefore does not increase $q$, and removes a positive coordinate. After finitely many such steps, the support is an independent set $I$. On that support, weighted Cauchy--Schwarz gives \[ \left(\sum_i p_ix_i\right)^2 \le\left(\sum_{i\in I}p_i\right) \left(\sum_{i\in I}p_ix_i^2\right) \le a q(x). \] This proves \eqref{eq:MS}, and expansion proves the assertion about $Z$. \end{proof} Define \begin{equation}\label{eq:PR} P(t)=\sum_{i,j}p_ip_jJ_0(t|u_i-u_j|),\qquad R(t)=a+zJ_0(t)-P(t). \end{equation} Rotational averaging of $|\sum_i p_i e^{it\omega\cdot u_i}|^2$ gives \begin{equation}\label{eq:Pbounds} 0\le P(t)\le1,\qquad P(0)=1,\qquad P(t)\ge1-\frac{D_0^2t^2}{4}. \end{equation} For a smoothing radius $s>0$, let \begin{equation}\label{eq:smooth} \phi_s(t)=\left(\frac{2J_1(st)}{st}\right)^2,\quad\phi_s(0)=1, \qquad g_s(t)=zJ_0(t)-\phi_s(t)R(t). \end{equation} Equivalently, \begin{equation}\label{eq:decomposition} g_s(t)=\phi_s(t)(P(t)-a)+(1-\phi_s(t))zJ_0(t). \end{equation} \begin{lemma}\label{lem:graphmoment} For the spectral probability measure of any periodic set avoiding $d>0$, \[ \int g_s(dt)\dd\nu_A(t)\le0. \] \end{lemma} \begin{proof} Smooth $1_A$ by convolution with uniform probability on the disk of radius $sd$. The result is nonnegative. Insert its values at $x+dR_\theta u_i$ into the copositive matrix \eqref{eq:Z}, and integrate over $x$ and $\theta$. The disk Fourier multiplier is $2J_1(sd|\xi|)/(sd|\xi|)$. Its square is $\phi_s(d|\xi|)$, while the angular average of the matrix contribution is $R(d|\xi|)$ by \eqref{eq:PR}. It follows that \[ \int\phi_s(dt)R(dt)\dd\nu_A(t)\ge0. \] All Fourier sums are absolutely convergent after the products are integrated, by Parseval and boundedness of the disk multiplier. The forbidden-distance moment \eqref{eq:zero} then proves the claim. \end{proof} \begin{lemma}\label{lem:smoothedbounds} Set \begin{align} K_0&=\frac{27(4z/5)^4}{1024a^3},\qquad E_s=K_0s^2, \label{eq:Es}\\ \beta_s&=\frac{a+E_s}{1-a},\qquad B_s=\frac{(4/5)z+(1+a+z)\sqrt{2/s}}{1-a}, \label{eq:Bs}\\ C_s&=\frac{D_0^2+(1-a+z)s^2}{4(1-a)}. \label{eq:Cs} \end{align} The normalized function $h_s=g_s/(1-a)$ satisfies \begin{equation}\label{eq:hgraphbounds} h_s(0)=1,\qquad h_s(t)\ge-\beta_s,\qquad |h_s(t)|\le B_s t^{-1/2}\ (t>0),\qquad h_s(t)\ge1-C_st^2. \end{equation} \end{lemma} \begin{proof} The Fourier transform of uniform probability on the radius-$s$ disk gives \begin{equation}\label{eq:phibounds} 0\le\phi_s(t)\le1,\qquad 1-\phi_s(t)\le\min\{1,s^2t^2/4\},\qquad \phi_s(t)\le\min\{1,4/(s^2t^2)\}. \end{equation} For the middle inequality, take independent uniform disk points $X,Y$. Then $\phi_s(t)=\E\cos(t(X_1-Y_1))$, and $\E(X_1-Y_1)^2=s^2/2$. The last bound follows from $|J_1|\le1$. By \eqref{eq:decomposition}, $P\ge0$, and Lemma~\ref{lem:bessel}, \[ g_s(t)\ge-a+(1-\phi_s(t)) \left(a-\frac{4z}{5\sqrt t}\right). \] The possible deficit below $-a$ is therefore bounded by \[ \frac{s^2}{4}\sup_{t>0} t^2\left(\frac{4z}{5\sqrt t}-a\right)_+. \] Writing $y=\sqrt t$ and $K=4z/5$, the polynomial to maximize is $Ky^3-ay^4$. Its positive maximum is $27K^4/(256a^3)$, attained at $y=3K/(4a)$. This proves $g_s\ge-a-E_s$. Also $g_s(0)=1-a$ by \eqref{eq:PR} and \eqref{eq:smooth}. Since $|R|\le1+a+z$, the decay estimate follows from \[ |g_s(t)|\sqrt t\le\frac45z+(1+a+z) \sqrt t\min\{1,4/(s^2t^2)\} \le\frac45z+(1+a+z)\sqrt{2/s}. \] The maximum in the last step occurs at $t=2/s$. Finally, the identity \[ g_s(t)-(1-a)=\phi_s(t)(P(t)-1) +(1-\phi_s(t))(zJ_0(t)-(1-a)) \] and \eqref{eq:Pbounds} give \[ g_s(t)-(1-a)\ge -\frac{D_0^2t^2}{4}-\frac{(1-a+z)s^2t^2}{4}. \] Dividing by $1-a$ completes the proof. \end{proof} \section{A two-range estimate and the separation rate} The following estimate uses a fixed initial frequency threshold. This is useful when the lower bound for a certificate approaches its limiting value. \begin{lemma}\label{lem:tworange} Suppose \[ h(0)=1,\quad h(t)\ge-\beta,\quad h(t)\ge1-Ct^2, \quad |h(t)|\le Bt^{-1/2}\quad(t>0), \] where $B,C>0$. Fix $b>1$, put $c=b-1$ and $\Delta=1-c\beta>0$, and choose $u>0$ with $Cu^2\le1/8$. Let $T=(4cB)^2\ge u$ and $Q>\max\{1,b^2\}$. Define \begin{equation}\label{eq:losses} \mathcal L=\frac{cBb}{\sqrt u(\sqrt Q-b)},\qquad \mathcal A=\frac{cCT^2}{bQ^2-1}. \end{equation} If \begin{equation}\label{eq:losscondition} \mathcal L\le1/8,\qquad \Delta-\mathcal A-\mathcal L\ge\kappa>0, \qquad\kappa\le1/2, \end{equation} then, for every $r\ge1$ and every positive list with $d_{i+1}/d_i\ge Q$, \begin{equation}\label{eq:Hfixed} H_r(t)=1+c\sum_{i=1}^r b^{i-1}h(d_it)\ge\kappa \quad(t\ge0),\qquad H_r(0)=b^r. \end{equation} \end{lemma} \begin{proof} If all $d_it1$ and $b-c=1$. The later loss is at most $\mathcal L\le1/8$. Hence the right side of \eqref{eq:crossing} is at least $1/2\ge\kappa$. Multiplying by $b^{j-1}\ge1$ proves the lower bound. At zero, the geometric sum gives $H_r(0)=b^r$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:graph}] Use $h_s$ from Lemma~\ref{lem:smoothedbounds} and write $c=b-1$. Since $b<1/a$, choose $s>0$ small enough that \[ \Delta=1-c\beta_s =\frac{1-ab-cK_0s^2}{1-a}>0. \] Choose $u>0$ with $C_su^2\le1/8$, and enlarge $B_s$ if needed so that $(4cB_s)^2\ge u$. As $Q$ tends to infinity, both losses in \eqref{eq:losses} tend to zero. Thus an integer $Q$ can be chosen to satisfy \eqref{eq:losscondition} with $\kappa=\min\{\Delta/2,1/4\}$. For any periodic set avoiding $d_1,\ldots,d_r$, Lemma~\ref{lem:graphmoment} gives $\int h_s(d_it)\dd\nu_A(t)\le0$ for every $i$. Integrating \eqref{eq:Hfixed} therefore yields \[ 1\ge\int H_r\dd\nu_A \ge b^r\delta+\kappa(1-\delta). \] Lemma~\ref{lem:periodic} proves \eqref{eq:graphbound} for all measurable sets. For the rate, let $\epsilon=1-ab$ decrease to zero and set \[ s=\sqrt{\frac{\epsilon}{2cK_0}},\qquad C_0=\frac{D_0^2+1-a+z}{4(1-a)},\qquad u=(8C_0)^{-1/2}. \] For sufficiently small $\epsilon$, one has $s\le1$, $C_s\le C_0$, and \[ \Delta=\frac{\epsilon}{2(1-a)}\le\frac12. \] Take $B=\max\{B_s,\sqrt u/(4c)\}$ and $T=(4cB)^2$. Choose an integer $Q$ strictly above \begin{equation}\label{eq:explicitQrate} \max\left\{ 1,b^2, \sqrt{\frac{1+4cC_0T^2/\Delta}{b}}, \left(b+\frac{4cBb}{\sqrt u\,\Delta}\right)^2 \right\}. \end{equation} Then $\mathcal A\le\Delta/4$ and $\mathcal L\le\Delta/4\le1/8$. Lemma~\ref{lem:tworange} applies with $\kappa=\Delta/2$. As $b\uparrow1/a$, $c$ stays bounded above and away from zero, and \[ s\asymp\epsilon^{1/2},\qquad B=O(\epsilon^{-1/4}),\qquad T=O(\epsilon^{-1/2}). \] The third entry of \eqref{eq:explicitQrate} is $O(\epsilon^{-1})$, and the fourth is $O(\epsilon^{-5/2})$. Taking the next integer preserves this order, proving \eqref{eq:rate}. \end{proof} \section{Explicit weighted-graph examples}\label{sec:examples} Bellitto, P\^echer and S\'edillot give a $607$-vertex unit-distance graph and nonnegative weights whose maximum independent weight divided by total weight is \begin{equation}\label{eq:BPSa} a=\frac{512933}{1999983}. \end{equation} We use this published finite result from \cite[Section 4.4]{BPS2021}. Its component ratios average to the whole-graph ratio, with the component total weights as averaging weights. Consequently some positive-weight connected component has independence ratio at most $a$. Normalize its weights to sum to one. Any two vertices of this component are joined by a simple path of at most $606$ unit edges. Thus \begin{equation}\label{eq:BPSgeometry} D_0\le606,\qquad z\le606a=:z_*. \end{equation} Zero-weight vertices may be kept when estimating path length and then omitted from the quadratic form. For these data, Lemma~\ref{lem:smoothedbounds} remains valid with the larger error \[ E_s=\frac{27}{2500}a606^4s^2 \] and the bounds \[ B_s\le\frac{(4/5)z_*+(1+a+z_*)\sqrt{2/s}}{1-a}, \qquad C_s\le\frac{606^2+(1-a+z_*)s^2}{4(1-a)}. \] The following table gives sufficient choices for \eqref{eq:graphbound}. All entries are exact terminating decimals or integers. \begin{center} \begin{tabular}{c|c|r|c|c} $b$ & $s$ & $B$ & $Q$ & $\kappa$\\ \hline $3.89$ & $10^{-6}$ & $320000$ & $10^{24}$ & $10^{-3}$\\ $3.899$ & $10^{-7}$ & $1000000$ & $10^{28}$ & $10^{-5}$\\ $3.8991$ & $10^{-8}$ & $3200000$ & $10^{32}$ & $10^{-6}$\\ $3.89911$ & $10^{-9}$ & $10000000$ & $10^{34}$ & $10^{-7}$ \end{tabular} \end{center} Here are rational bounds verifying every condition without a frequency discretization. In all four rows take $C=125000$ and $u=1/1000$, so $Cu^2=1/8$ and $\sqrt u>3/100$. Upper bounds for $\sqrt{2/s}$ are, in row order, \[ 1500,\quad4500,\quad15000,\quad45000. \] Substitution gives $B_s\kappa$ and $\mathcal L<1/8$ in every row. The remaining inequalities $T\ge u$ and $Q>\max\{1,b^2\}$ are immediate from the tabulated values. Lemma~\ref{lem:tworange} and the integration argument prove each stated bound for every $r$. \section{Fixed separation for ordinary finite graphs}\label{rf:ordinary} The transfer proved in Part~I converts the density refinements into statements about arbitrary colorings whenever all forbidden distances are algebraic. The finite graphs in the next result follow by the same compactness step used by Gal\'an. The additional assertion is the choice of a single separation ratio before the length $r$. \begin{corollary}\label{rf:fixedq} Let $m_1=m(\{1\})$ and $1b^r. \] If a fixed finite weighted unit-distance graph has independence ratio at most $a<1$, then for $1b^r$ and also $\chi(G(E,D))>b^r$. If every finite induced graph on $E$ had a proper $\lfloor b^r\rfloor$-coloring, finite-color compactness would give such a coloring of $G(E,D)$. A finite $P$ with the required chromatic number therefore exists. For the effective assertion, Theorem~\ref{thm:graph} gives \[ \frac1{m(D)}\ge\frac{b^r-\kappa}{1-\kappa}>b^r, \] since $0<\kappa<1$ and $b^r>1$. Apply the same transfer and compactness argument. Section~\ref{sec:examples} gives the stated numerical pair. \end{proof} This corollary supplies finite witnesses by compactness. The ratio in its first assertion is existential. Its effective assertion gives an explicit separation bound from a supplied weighted realization and independence-ratio certificate. \appendix \section{Sources, versions, and verification scope} \subsection{Frozen primary sources} The OpenAI source used for Part~I is the September 23, 2026 manuscript \emph{The Euclidean plane is not five-colorable}, frozen at Git commit \begin{center} \texttt{4eaee87548250b2682affda7d05d947c05554bae}. \end{center} The spectral, rigidity, and transfer source files are recorded in the accompanying manifest \cite{OpenAI2026}. The Gal\'an source is version 8, October 11, 2026, frozen at Git commit \begin{center} \texttt{cb660f322dce47a773f347d39818ffc46e6ca72c}. \end{center} Both are cited as manuscripts. The attribution concerns their mathematical arguments and is separate from prize evaluation. The source package includes the applicable Apache 2.0 license for the OpenAI repository and the MIT license carrying \textcopyright\ 2026 Sergi Gal\'an for his repository. The frozen \href{https://github.com/openai/math/blob/4eaee87548250b2682affda7d05d947c05554bae/LICENSE}{OpenAI Apache 2.0 license} and \href{https://github.com/decalion89/chromatic-number-of-the-plane/blob/cb660f322dce47a773f347d39818ffc46e6ca72c/LICENSE}{Gal\'an MIT license} retain their terms for their respective upstream materials. This deposit contains the rewritten manuscript. The separate proof and supporting source bundle are available in the \href{https://github.com/Ghuffy1/erdos-706-lean/tree/2a22f736eea81c04ec696c1b1d7f770b4417f5f7/paper}{pinned repository}. The exposition in Sections~\ref{sp:section}--\ref{tr:706section} rewrites and expands the arguments of those sources. These changes include the explicit conditional-measure conventions, auxiliary estimates, and a unified notation. \subsection{Mathematical and numerical checks} Independent AI-agent reviews checked the rigidity proof, the transfer, and the full finite-graph deduction, and a separate review checked the uniform compact-cluster and graph-amplification arguments. These reviews are supporting audit records. The mathematical evidence is the proofs and the identified published inputs. The separate formal verification described below establishes the negative answer to the original polynomial-bound question with an elementary exponential lower bound. The base $4.0485$ uses the published bound $m_1\le247/1000$. The graph examples in Section~\ref{sec:examples} use the published $607$-vertex weighted-graph result of Bellitto, P\^echer and S\'edillot \cite{BPS2021}. The underlying large optimization and independent-set computations were not rerun in this work. The displayed separation inequalities were checked in exact rational arithmetic. The proof of the negative answer with base $2$ uses no such numerical optimization. \subsection{Formal verification of the original negative answer} A separate Lean development, frozen at commit \begin{center} \texttt{331157628c3937f5fd992c1b85c06f146184c465}, \end{center} proves the negation of the eventual polynomial upper bound in the original problem. The principal declaration is \texttt{Erdos706.erdos\_706\_negative}. Its statement is also written out independently in \texttt{Erdos706.Audit.original\_problem\_negative}, with the finite point set, the $r$ positive distances, and the chromatic-number quantifiers explicit. The development establishes \[ 2^r\le L(r)\le7^r\qquad(r\ge1). \] It uses the OpenAI rigidity formalization, Gal\'an's algebraic-distance formalization, and an elementary separated-scale spectral certificate. The ordinary-character spectral measure has an atom of mass at least $1/k$ at zero for a proper $k$-coloring. A positive certificate made from the circular Fourier kernel contradicts this atom bound at $k=2^r$. Finite-subgraph compactness then gives the finite witnesses required by the original statement. The ratio in this formal construction may depend on $r$. The verification used Lean 4.34.1 and Mathlib commit \begin{center} \texttt{d13f23b723b8a846827a245b89c10fc7d3f11612}. \end{center} All 56 local modules were rebuilt into an empty output directory, and their source hashes were checked against the frozen proof commit. The axiom reports for the final declarations contain exactly \texttt{propext}, \texttt{Classical.choice}, and \texttt{Quot.sound}. A target-directed replay into an empty Lean kernel accepted all 73,146 transitive declarations for the negative answer, its explicit-statement bridge, the exponential lower bound, and the seven-color upper bound. The accompanying verification files record the commands, hashes, axiom reports, and replay output. A separate attempt to replay the entire imported environment reached its 30-minute resource limit before completion. This formal result verifies the original negative answer through the base-two argument. The stronger numerical base $4.0485$, the compact-cluster theorem, the quantitative amplification theorem, and the fixed-ratio refinements are supported here by their written proofs and cited inputs. Their formalization is outside the scope of the frozen Lean development. The formal source retains the OpenAI and Gal\'an attribution and licenses, with provenance recorded for each imported module. Codex-assisted additions are identified separately in the source documentation. \subsection{Remaining historical comparison} The literature audit compared the relevant full arguments in \cite{Bukh2008,OFV2010,DOV2022,CastroSilva2023,Davies2026,Galan2026} and the weighted form in \cite{LaurentVargas2022}. The source of the earlier weighted characterization is credited through the latter's attribution to \cite{GHPR1997}. The two historical originals \cite{Szekely1984,SzekelyWormald1989} remain to be compared in full. Accordingly, the paper asserts and proves the stated uniform refinements while leaving first-discovery priority for them open to that comparison. \section*{Acknowledgment of AI assistance} OpenAI's Codex was used for mathematical exploration, proof development, exact arithmetic checks, literature searches, independent-agent review, and drafting this manuscript. Attribution of OpenAI's pre-existing rigidity and transfer results is given separately throughout Part~I. This acknowledgment records assistance in preparing the present text. \begin{thebibliography}{99} \bibitem{ACMVZ2023} G. Ambrus, A. Csisz\'arik, M. Matolcsi, D. Varga and P. Zs\'amboki, \emph{The density of planar sets avoiding unit distances}, Math. Program. \textbf{207} (2024), 303--327, \url{https://arxiv.org/abs/2207.14179v3}. \bibitem{BHV2008} B. Bekka, P. de la Harpe and A. Valette, \emph{Kazhdan's Property (T)}, New Mathematical Monographs 11, Cambridge University Press, 2008, DOI 10.1017/CBO9780511542749. Theorem locators refer to the author manuscript of February 23, 2007. \bibitem{BPS2021} T. Bellitto, A. P\^echer and A. S\'edillot, \emph{On the density of sets of the Euclidean plane avoiding distance $1$}, Discrete Math. Theor. Comput. Sci. \textbf{23} (2021), no. 1, DOI 10.46298/dmtcs.5153. \bibitem{Bukh2008} B. Bukh, \emph{Measurable sets with excluded distances}, Geom. Funct. Anal. \textbf{18} (2008), 668--697. \bibitem{CastroSilva2023} D. Castro-Silva, \emph{Geometrical sets with forbidden configurations}, Forum Math. Sigma \textbf{11} (2023), e44, DOI 10.1017/fms.2023.43. \bibitem{Chen2018} R. Chen, \emph{Decompositions and measures on countable Borel equivalence relations}, arXiv:1801.02767v2, Section 3, Theorem 3.1 (Farrell--Varadarajan), \url{https://arxiv.org/abs/1801.02767v2}. \bibitem{Davies2026} J. Davies, \emph{Measurable obstructions for unmeasurable colourings}, preprint, arXiv:2610.12301v1 (2026). \bibitem{deBruijnErdos1951} N. G. de Bruijn and P. Erd\H{o}s, \emph{A colour problem for infinite graphs and a problem in the theory of relations}, Indag. Math. \textbf{13} (1951), 369--373. \bibitem{DLMF} F. W. J. Olver and L. C. Maximon, \emph{Bessel functions}, Chapter 10 of the \emph{NIST Digital Library of Mathematical Functions}, \texttt{https://dlmf.nist.gov/10}, accessed November 4, 2026. \bibitem{DOV2022} E. DeCorte, F. M. de Oliveira Filho and F. Vallentin, \emph{Complete positivity and distance-avoiding sets}, Math. Program. \textbf{191} (2022), 487--558, DOI 10.1007/s10107-020-01562-6. \bibitem{Erdos1981} P. Erd\H{o}s, \emph{On the combinatorial problems which I would most like to see solved}, Combinatorica \textbf{1} (1981), 25--42. \bibitem{Erdos706} T. F. Bloom, \emph{Erd\H{o}s Problem \#706}, \url{https://www.erdosproblems.com/706}. \bibitem{Furstenberg1977} H. Furstenberg, \emph{Ergodic behavior of diagonal measures and a theorem of Szemer\'edi on arithmetic progressions}, J. Analyse Math. \textbf{31} (1977), 204--256, DOI 10.1007/BF02813304. \bibitem{Galan2026} S. Gal\'an, \emph{Measurable density bounds hold for all colourings of the plane with algebraic forbidden distances}, manuscript, version 8, October 11, 2026. Frozen source, \href{https://github.com/decalion89/chromatic-number-of-the-plane/blob/cb660f322dce47a773f347d39818ffc46e6ca72c/papers/forbidden-distances/forbidden-distances.tex}{commit cb660f322dce47a773f347d39818ffc46e6ca72c}. \bibitem{GHPR1997} L. E. Gibbons, D. W. Hearn, P. M. Pardalos and M. V. Ramana, \emph{Continuous characterizations of the maximum clique problem}, Math. Oper. Res. \textbf{22} (1997), no. 3, 754--768. \bibitem{JamneshanCorr} A. Jamneshan, \emph{An uncountable Furstenberg--Zimmer structure theory---Corrigendum}, Ergodic Theory Dynam. Systems \textbf{46} (2026), no. 1, 211--217, DOI 10.1017/etds.2025.10214. \bibitem{LaurentVargas2022} M. Laurent and L. F. Vargas, \emph{Finite convergence of sum-of-squares hierarchies for the stability number of a graph}, SIAM J. Optim. \textbf{32} (2022), no. 2, 491--518, DOI 10.1137/21M140345X. \bibitem{OFV2010} F. M. de Oliveira Filho and F. Vallentin, \emph{Fourier analysis, linear programming, and densities of distance-avoiding sets in $\R^n$}, J. Eur. Math. Soc. \textbf{12} (2010), 1417--1428, DOI 10.4171/JEMS/236. \bibitem{OpenAI2026} OpenAI, \emph{The Euclidean plane is not five-colorable}, OpenAI Math Release manuscript, September 23, 2026. Frozen source, \href{https://github.com/openai/math/tree/4eaee87548250b2682affda7d05d947c05554bae/preprints/The-Euclidean-plane-is-not-five-colorable-September-23-2026}{commit 4eaee87548250b2682affda7d05d947c05554bae}. \bibitem{Szekely1984} L. A. Sz\'ekely, \emph{Measurable chromatic number of geometric graphs and sets without some distances in Euclidean space}, Combinatorica \textbf{4} (1984), 213--218, DOI 10.1007/BF02579223. \bibitem{SzekelyWormald1989} L. A. Sz\'ekely and N. C. Wormald, \emph{Bounds on the measurable chromatic number of $\R^n$}, Discrete Math. \textbf{75} (1989), 343--372, DOI 10.1016/0012-365X(89)90099-X. \bibitem{Tao2008} T. Tao, \emph{254A, Lecture 9. Ergodicity}, February 4, 2008, Theorem 4 and Remark 5, \url{https://terrytao.wordpress.com/2008/02/04/254a-lecture-9-ergodicity/}. \end{thebibliography} \bigskip \begingroup\footnotesize \noindent Purdue University\par\smallskip \noindent\begin{tabular}{@{}ll@{}} Will Howe & \texttt{howew@purdue.edu}\\ Grant Huffman & \texttt{huffma97@purdue.edu}\\ Jensen Kohlmeyer & \texttt{jenwin@purdue.edu}\\ Liam Kruer & \texttt{kruerl@purdue.edu} \end{tabular}\par \endgroup \makeatletter\let\@setaddresses\relax\makeatother \end{document}