\documentclass[11pt,reqno]{amsart} \usepackage{amssymb,mathtools} \usepackage{newtxtext,newtxmath} \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue,pdfencoding=auto]{hyperref} \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} \numberwithin{equation}{section} \newcommand{\Z}{\mathbb Z} \newcommand{\R}{\mathbb R} \newcommand{\C}{\mathbb C} \newcommand{\dd}{\,\mathrm d} \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 17, 2026} \thanks{AI tools assisted with developing and checking arguments and with drafting this manuscript.} \emergencystretch=2em \setlength{\textwidth}{446.4bp} \setlength{\oddsidemargin}{\dimexpr82.8bp-1in\relax} \setlength{\evensidemargin}{\oddsidemargin} \addtolength{\topmargin}{-33.3bp} \addtolength{\textheight}{44.6bp} \hypersetup{pdfauthor={Will Howe, Grant Huffman, Jensen Kohlmeyer, Liam Kruer},pdftitle={The Sharp Constant in the Erdos Maximum-Term Problem}} \title[The Erd\H{o}s maximum-term problem]{The sharp constant in the Erd\H{o}s maximum-term problem} \subjclass[2020]{Primary 30D15. Secondary 30D20, 65G30.} \keywords{Maximum term, entire function, theta series, exposed indices} \begin{document} \begin{abstract} For a transcendental entire function \(f\), let \(\mu(r,f)\) be its maximum term and \(M(r,f)\) its maximum modulus. We determine the supremum of \(\liminf_{r\to\infty}\mu(r,f)/M(r,f)\) as the reciprocal of the minimum circle norm in the He--Tang theta family. The minimum is characterized by a unique solution of three real equations, and the supremum is attained. The proof combines coefficient transport at radii where two terms tie, a classification of the resulting gap sequences, and a uniform quadratic comparison for arithmetic gaps. The local comparisons use exact finite arithmetic. \end{abstract} \maketitle \section{The maximum-term problem} Given \(f(z)=\sum_{n\ge0} a_n z^n\), set \begin{gather*} \mu(r,f)=\max_{n\ge0}|a_n|r^n,\qquad M(r,f)=\max_{|z|=r}|f(z)|,\\ Q(f)=\limsup_{r\to\infty} M(r,f)/\mu(r,f). \end{gather*} Erd\H{o}s Problem~513 asks for the supremum \(B=\sup_f Q(f)^{-1}\) over transcendental entire functions \cite{Bloom}. Here the reciprocal of \(+\infty\) is zero. We construct an extremizer in the two-parameter family of He and Tang \cite{HT}, which extends the scaling construction of Clunie and Hayman \cite{CH}. The upper bound requires comparisons valid for arbitrary entire functions. \begin{theorem}\label{thm:main} Put \(T_n=n(n+1)/2\), \(e(t)=(t^2-1+2it)/(1+t^2)\), and define \(S_0=1\), \(S_1(y)=4y-3\), and \(S_{n+1}=(4y-2)S_n-S_{n-1}\). For \[ R(q,t,y)=4y\left|\sum_{n\ge0}(qe(t))^{T_n}S_n(y)\right|^2 \] let \((q_*,t_*,y_*)\) denote the unique solution of \begin{equation}\label{eq:system} \begin{aligned} R_y(q,t,y)&=0,\\ R(q,t,1)&=R(q,t,y),\\ R_q(q,t,1)R_t(q,t,y)&=R_t(q,t,1)R_q(q,t,y) \end{aligned} \end{equation} in the box \begin{equation}\label{eq:optbox} \mathcal B=[0.28010,0.28040]\times[0.43450,0.43472]\times[0.633,0.643]. \end{equation} Then \(B=1/A_*\approx0.58507882\), where \(A_*=\sqrt{R(q_*,t_*,1)}\) and \(1.709171421.70917142\) and place every minimizer in the interior of the parameter part of \eqref{eq:optbox}. For the angular and derivative inequalities, keep \(0\le n\le8\) and differentiate the recurrences for \(S_n\) and \((qe(t))^{T_n}\). For example the recursion for \((S_n,S_n',S_n'')\) is multiplication by \(4y-2\), subtraction of the preceding triple, and addition of \((0,4S_n,8S_n')\). The absolute coefficient sum of \(S_n\) is bounded by \(7^n\). On the parameter rectangle, all needed partial derivatives of the omitted complex sum are bounded by \[ \sum_{n\ge9}16(T_n+1)^2 n^2 7^n(3/10)^{T_n}<10^{-8}. \] Indeed the ratio of consecutive terms is at most \(112(3/10)^{10}<1/2\), and twice the first term is smaller than \(10^{-8}\). The bounding factor \(16(T_n+1)^2\) follows by twice differentiating with \(q\ge1/4\) and the arc derivative bounds. Add \([-10^{-8},10^{-8}]\) to each real and imaginary sum before differentiating \(R=4y|S|^2\) by the product rule on the enclosed sums. The following are lower bounds using interval operations at scale \(2^{60}\). Bisect the displayed \(y\)-interval until the signed quantity has positive lower endpoint. The \(q,t\) rectangle stays fixed. \[ \begin{array}{c|c|r|c} y&\text{quantity}&\text{leaves}&\text{lower bound}\\ \hline \relax[0,.62]&R_y&8&.00032\\ \relax[.62,.66]&-R_{yy}&1&3.93\\ \relax[.66,.86]&-R_y&16&.0152\\ \relax[.86,.90]&R_{yy}&1&4.03\\ \relax[.90,1]&R_y&5&.0389\\ \{.633\}&R_y&1&.0205\\ \{.643\}&-R_y&1&.0188 \end{array} \] It follows that only the endpoint \(1\) and a unique interior maximum \(Y(q,t)\) in \((.633,.643)\) can attain the global angular maximum. The implicit function theorem gives smoothness because \(R_{yy}<0\) there. Write \(F_0=R(q,t,1)\), \(F_1=R(q,t,Y(q,t))\), \(D_0=F_0-F_1\). The same enclosure with \(y\in[.633,.643]\) gives \((F_0)_t>0>(F_1)_t\), \((D_0)_q<0<(D_0)_t\). In this interval evaluation the first derivatives of \(F_1\) are \(R_q,R_t\), and its Hessian entries are \(R_{ij}-R_{iy}R_{jy}/R_{yy}\), \(i,j\in\{q,t\}\). Set \(s=-(D_0)_q/(D_0)_t\) and \(\lambda=-(F_1)_t/(D_0)_t\). Substitution gives \(0<\lambda<1\) and \begin{equation}\label{eq:crosscurv} (1,s)\bigl(\lambda\nabla^2F_0+(1-\lambda)\nabla^2F_1\bigr)(1,s)^{\mathsf T}>18.75 . \end{equation} For the interval evaluation of the \(F_1\) term, subtract \((R_{qy}+sR_{ty})^2/R_{yy}\) from \(R_{qq}+2sR_{qt}+s^2R_{tt}\). The lower endpoint for \eqref{eq:crosscurv} exceeds \(18.7552\). At an interior minimizer both branches must be active, since moving \(t\) slightly in the decreasing direction of a sole active branch would lower the maximum. Between any two points of \(D_0=0\) the locus is a graph \(t=t(q)\) entirely in the rectangle. In fact \(t'(q)=s>0\), and at every intermediate \(q\) the two \(t\)-values bracket the unique zero by the signs of the derivatives of \(D_0\). Differentiating the common value on this graph twice gives precisely \eqref{eq:crosscurv}, as the choice of \(\lambda\) cancels the \(t''\) term. The first derivative vanishes precisely at the last equation of \eqref{eq:system}. Existence follows from the interior global minimizer. Strict convexity along the graph gives uniqueness, and the first equation specifies its unique interior angular maximum. This proves the restricted minimax and its enclosure. \section{Reduction for arbitrary entire functions}\label{sec:reduction} Assume \(Q(f)d_{k+1}\) and \(d_{k-1}=d_k>d_{k+1}\). \item By Theorem~\ref{thm:plateau}, for a decrease \(d,E\) with \(E(43/25)\|T_k\|_1\). When \(d_{k-1}10000\), the same test has \(N_k>(12/7)\|T_k\|_1\). \item By Theorem~\ref{thm:arithmetic}, on arbitrarily late runs of equal exposed gaps of unbounded length, one has \(Q(f)\ge A_*\). \end{enumerate} We prove that these comparisons yield the unrestricted upper bound. The first two comparisons are used also in reversed finite windows. Indeed, a long Taylor polynomial retains every stipulated maximum term at finitely many selected radii and approaches the norms there uniformly. The strict cap allows this truncation. For a Taylor polynomial \(p\) of degree \(m\), \(z^m p(1/z)\) reverses the exponents, and its \(M/\mu\) at \(r^{-1}\) equals that of \(p\) at \(r\). The local comparisons then apply in the reverse order. A late decrease must be preceded by a decrease by the first comparison. Positive integer gaps cannot decrease indefinitely. The gaps are consequently eventually nondecreasing. Once a late strict increase occurs, the reflected equal-pair exclusion makes the next comparison strict as well. Thus the gaps are eventually constant or eventually strictly increasing. In the increasing case, comparison 2 applied in reverse to \(d_{k-1}0\) there are arbitrarily late blocks of that length in which every successive ratio is at most \(1+\varepsilon\). Otherwise every sufficiently late block of that length contains a ratio above \(1+\varepsilon\), giving geometric growth and a convergent reciprocal sum. Take \(224\) interior knots in such a block with \(\varepsilon=10^{-4}\). At one such knot write \(P=d_{k-1}10000\). Thus comparison 3 gives \(N_k>(12/7)\|T_k\|_1\) at each interior knot. For clarity, let \(H_k(\theta)\) denote the circle function at \(\tau_k\) divided by the lower tied complex term. Write \(H_k=\sum_j h_k(j)e^{ij\theta}\), so \(h_k(0)=1\), \(h_k(d_k)=A_k\) for a unit complex \(A_k\). Set \(z_k=A_k e^{id_k\theta}\), \(x_k=e^{-d_{k-1}\Delta_{k-1}}\), and \(y_k=e^{-d_{k+1}\Delta_k}\). The tests and their known contributions are \begin{align*} T_k={}&\tfrac13 z_k^{-1}(1+z_k)^3 +\tfrac13 x_k(z_{k-1}^{-1}-z_{k-1}) +\tfrac13 y_k z_k(z_{k+1}-z_{k+1}^{-1}),\\ N_k={}&2+\tfrac13(x_k^2+y_k^2). \end{align*} The coefficient transport is \(h_{k+1}(j)=\overline{A_k}e^{j\Delta_k}h_k(j+d_k)\). It cancels \(z_k^2/3\) at knot \(k\) against \(-x_{k+1}z_k/3\) at knot \(k+1\) in the real pairings. It also cancels \(z_{k+1}^{-1}/3\) at knot \(k+1\) against \(-y_k z_k z_{k+1}^{-1}/3\) at knot \(k\). The four unmatched boundary terms contribute in absolute value at most \(4/3\). H\"older's inequality and \(N_k\ge2\) now require \[ 2\cdot224\bigl(1-C_0/(12/7)\bigr)\le \tfrac43 . \] The left side is \(800576/600000>4/3\), a contradiction. The remaining case of an eventually constant gap gives \(Q(f)\ge A_*\) by comparison 4, which contradicts the starting assumption. Together with the equality for \(f_{q_*,e(t_*)}\), this proves Theorem~\ref{thm:main}. \section{A decrease preceded by a long plateau}\label{sec:plateau} \begin{theorem}\label{thm:plateau} Let \(t1/2500\). More generally, with any fixed cap strictly below \(C_1\), the same quantity is bounded away from zero uniformly in \(d,E\). \end{theorem} \begin{proof} Project at \(r,s\) to the two residue classes \(N\) and \(N+E\) modulo \(d\). At a fixed value of the high variable \(z=e^{id\theta}\), orthogonality on the \(d\) root-of-unity shifts bounds the Euclidean norm of the component vector by the original cap. Use the Taylor coefficients divided by the positive maximum terms in each profile, with independent unit phases for the two components and a rotation of \(z\) fixing the first three displayed ties. Take real parts of the resulting coefficients by averaging the vector with its conjugate at \(\bar z\). These operations commute with positive coefficient transport and preserve the anchors. Delete the terms of high degree at most \(-1\) in the first component and at most \(-2\) in the second. Each deleted coefficient has normalized size at \(r\) at most \((t/r)^{N-j}\), where \(j\) is its original index. The same bound holds at \(s\) because the maximum term moves to \(N+d\) and \(jP\). The factor \(p^3/(1-p)\) also increases. These give the necessary inequalities with \(p\) replaced by \(P\). Here is the resulting two-variable arithmetic check. Start with \((c,u)\in[-1,0]\times[-1,1]\). For a rectangle use \(J\) for the interval of \(\xi^2\), \([A,B]\) for the interval of \(c_2(c)\), \(b=r_0(1-\inf J)\), \(U=C^2-2-P-(\inf c^2)/P-(\inf u^2)/P^2\), and \(l=1/20-(4/5)\sup u\). Reject if \(\inf J>1\), if \([A-b,B+b]\) misses the \(u\)-interval, if \(U<0\), or if \(l>0\) and \(P^3U/(1-P).41\). Since \(x^2+c^2/x^2\le C^2-2\) and \(p+\ell^2/p\) increases for \(p>\ell\), its value at \(p=.91^2\) implies \(x<.91\). At \(r\) the two ties and the next exposed term give \(\beta^2\le C^2-2<.97^2\). For either analytic vector, compress multiplication by that vector from the input frequencies \(0,1,2\) to the output frequencies \(0,1,2,3\). The norm is at most \(C\) by integration of the pointwise bound and orthogonal projection. Its columns are \((D,A,B,F),(0,D,A,B),(0,0,D,A)\). Their squared norms and pairwise inner products give the Gram matrix \(G_{\rm v}\). Thus \(C^2I-G_{\rm v}\) has nonnegative principal minors. The following interval recurrence verifies that both sets of minors cannot be nonnegative at any point in the six-parameter box. Use scale \(2^{24}\), ten slabs \(x\in[(41+5j)/100,(46+5j)/100]\), \(0\le j<10\), \(\beta\in[.41,.97]\), \(c\in[-2A_0+\ell,-\ell]\), and \(a,e,g\in[-1,1]\). Discard a box if \(\sup\beta<\inf x\). In each remaining box intersect the upper \(x\) bound with \(\sup\beta\) and the lower \(\beta\) bound with \(\inf x\) for evaluation. For a row \(D=(0,g), A=(a_0,a_1), B=(b_0,b_1), F=(f_0,f_1)\), form the diagonal intervals \[ d_2=C^2-g^2-|A|^2,\qquad d_1=d_2-|B|^2,\qquad d_0=d_1-|F|^2 \] and the positive-sign Gram entries \[ o_{12}=ga_1+A\cdot B,\quad o_{01}=o_{12}+B\cdot F,\quad o_{02}=gb_1+A\cdot F. \] A negative upper endpoint of a diagonal rejects the box. In the remaining test intersect the diagonals with \([0,\infty)\) and reject on a negative upper endpoint of any \(d_i d_j-o_{ij}^2\) or of \[ d_0d_1d_2-2o_{01}o_{02}o_{12}-d_0o_{12}^2-d_1o_{02}^2-d_2o_{01}^2 . \] Use the first row before the second. For the second, intersect each transformed coefficient with \([-1,1]\) and reject if any intersection is empty. Apply interval operations in the displayed order, using \(o_{12}=ga_1+a_0b_0+a_1b_1\), \(o_{01}=o_{12}+b_0f_0+b_1f_1\), and \(o_{02}=gb_1+a_0f_0+a_1f_1\). Bisect an unrejected original box on its widest coordinate, using the first such coordinate in the order \(x,\beta,a,c,e,g\). Use the lower midpoint on the fixed-point grid. \[ \begin{array}{c|rrrr|r} j&\sup\beta<\inf x&\text{first matrix}&\text{empty}&\text{second matrix}&\text{depth}\\ \hline 0&1247&110516&67&58309&36\\ 1&3157&125088&64&59044&36\\ 2&3537&87613&0&38368&35\\ 3&1707&37405&0&16684&35\\ 4&808&13589&0&6284&32\\ 5&416&4953&0&1968&29\\ 6&12&1687&0&540&26\\ 7&11&535&0&107&23\\ 8&4&105&0&6&15\\ 9&6&38&0&0&13 \end{array} \] Each branch of the subdivision ends in a rejection by the displayed inequalities. The two matrices are therefore incompatible. This excludes the modified profiles and proves the theorem. \end{proof} \section{Two-sided tests for increasing gaps}\label{sec:mix} \begin{lemma}\label{lem:mix} Use \(T_k,N_k\) from Section~\ref{sec:reduction} at a double knot of norm at most \(\sqrt3\), and write \(P,D,R\) for its left, central and right gaps. If \(P,-P,R,-R\bmod D\) are distinct and nonzero, then \(N_k>(43/25)\|T_k\|_1\). If \(P+R=2D\) and \(D/\gcd(D,P)\ge2048\), then \(N_k>(12/7)\|T_k\|_1\). \end{lemma} \begin{proof} Write \(x=x_k,\ y=y_k,\ S=x^2+y^2\le1\) by Parseval. Put \(P_0(z)=z^{-1}(1+z)^3/3\). Averaging over the shifts by \(2\pi/D\) fixes the central variable. In the separated case the other four terms are orthogonal and have zero mean on this fiber. Taking the square root of the mean square bounds their contribution together with the central polynomial by \[ I(S)=\int_0^{2\pi}\sqrt{|P_0(e^{i\theta})|^2+2S/9}\,\frac{\dd\theta}{2\pi}. \] Use \(S\in[j/256,(j+1)/256]\), the lower endpoint in \(2+S/3\) and the upper endpoint in \(I(S)\). For \(z=(1+it)/(1-it)\) and \(-\bar z\), \(0\le t\le1\), the squared moduli of \(P_0\) are \(64/(9(1+t^2)^3)\) and \(64t^6/(9(1+t^2)^3)\). Therefore \(I(S)\) is \(1/\pi\) times the integral over \([0,1]\) of the sum of the two roots, each with \(2S/9\) added under the root, times \(2/(1+t^2)\). Each root is 1-Lipschitz in \(|P_0|\), is bounded by \(19/6\), and the angular derivative of \(P_0\) has modulus at most \(2\). With \(2/(1+t^2)\le2\) and the sharper bound \(|(2/(1+t^2))'|\le3\), the integrand is \(124/3\)-Lipschitz. At the \(4096\) midpoints \((j+1/2)/4096\), take upward roots at scale \(2^{32}\), multiply by the rational weights and round upward. Add \((124/3)/(4\cdot4096)\) to the mean and divide by a lower bound for \(\pi\) from Machin's identity. Substitution in \(2+j/768-(43/25)I((j+1)/256)\) gives a lower margin exceeding \(0.0051476\) in all \(256\) tests. For \(P+R=2D\), write \(z=z_k,\ w=z_{k-1}\). The third phase is \(z_{k+1}=\gamma z^2w^{-1}/y\) for \(|\gamma|=y>0\), and the test equals \[ T(z,w)=P_0(z)+\tfrac13 x(w^{-1}-w) +\tfrac13(\gamma z^3w^{-1}-\bar\gamma z^{-1}w). \] Let \(D_0(x,\gamma)\) be the two-circle mean of \(|T|\), convex in \((x,\Re\gamma,\Im\gamma)\). By conjugation consider \(\Im\gamma\ge0\). Take all cubes of side \(1/5\) in \([0,1]\times[-1,1]\times[0,1]\) whose points of shortest distance to the origin have squared distance at most one. There are 192 such cubes with 333 distinct vertices. At each vertex use the rectangular equiangular mesh \(128\times64\), including \(z=w=1\). Bound the mean above by the mean of the sampled moduli plus \[ \frac{10}{3}\left(\frac{(8+10|\gamma|)/3}{128^2} +\frac{2(x+|\gamma|)/3}{64^2}\right). \] Indeed, interpolating linearly between endpoints of an angular segment of length \(h\) costs at most \(h^2\sup|T''|/12\) in the mean by the linear interpolation remainder. The two derivative bounds are the sums of coefficient magnitudes times squared frequency. Use \(\pi^2<10\). For the vertex arithmetic, enclose the real and imaginary root coordinates to within \(2^{-30}\) by Machin's identity and the degree-30 cosine and degree-31 sine polynomials on \([0,\pi/2]\), with quadrant rotations. For each monomial use the root table for the grid of order 128, since the order 64 divides it. Add to each sample coordinate the interval error bounded by \(2^{-30}\) times the sum of the real and imaginary absolute coefficient parts. The upper modulus is the upward square root of the sum of the larger squared endpoints in the two coordinates. For \(|\gamma|\) use an upward root as well. These substitutions in \(2+(x^2+|\gamma|^2)/3-(12/7)D_0(x,\gamma)\) give a least lower bound greater than \(0.0133108\). Multilinear interpolation of the vertex inequalities loses at most \(3(1/5)^2/12=1/100\) from the squared parameter term. Thus the margin exceeds \(3/1000\) throughout the ball. At fixed \(z\), the actual \(w\)'s are a rotated regular polygon of order \(D/\gcd(D,P)\). The angular Lipschitz constant of \(|T|\) in \(w\) is at most \(2(x+|\gamma|)/3<1\). Averaging this bound on the equal intervals centered at the polygon points gives error at most \(\pi/(2D/\gcd(D,P))<2/(D/\gcd(D,P))\) between the polygon mean and the continuous mean. Multiplying the error at order \(2048\) by \(12/7\) gives \(3/1792<3/1000\). The strict inequality survives, proving the second claim. \end{proof} \section{Fourier comparisons at three knots}\label{sec:duality} Use brackets \(\langle\cdot\rangle\) for normalized angular average. Let \(n_0h. \end{cases} \end{equation} The pivot is \(0,1,2\), respectively. A table of test weights lists three raw entries in every selected coefficient column. For an unknown column replace its pivot entry by minus the sum of the other two entries times their factors. Its total pairing then vanishes. On a known unit anchor, leave the three entries unchanged and sum their transported real parts into a numerator \(N\), multiplying all weights in that column by the anchor phase in the test. This convention gives the correct real pairing for complex weights since the weights in the pairing are conjugated. If the normalized norms are at most \(C\), the resulting three polynomial vectors \(P_k\) satisfy \begin{equation}\label{eq:testmargin} N-C\sum_{k=0}^2\langle\Psi(P_k)\rangle\le0 . \end{equation} \subsection{Cyclic mean estimates and norm properties}\label{subsec:angular} Let \(S=\sum_i|p_i|^2\). In formulas below each norm is extended by zero at \(S=0\). For an edge set \(\mathcal G\) of unordered pairs, use \[ \Psi_{\eta,\mathcal G}=\sqrt S-\eta\sum_{\{i,j\}\in\mathcal G}|p_i p_j|^2/S^{3/2}. \] For an arithmetic progression of three residues with middle magnitude \(x\) and endpoint magnitudes \(y,z\), use \[ \Psi_{\rm AP}=\sqrt S-\bigl(y^2z^2/4+x^2(y-z)^2/16\bigr)/S^{3/2}. \] A two-component norm with an extra correction is \[ \Psi_{\eta,\zeta}^{(2)}=\sqrt S-\eta|p_0p_1|^2/S^{3/2} -\zeta|p_0p_1|^4/S^{7/2}. \] For the cyclic bounds, normalize \(S=1\) and let \(Z\) be the squared scalar test minus one. Then \(\langle Z\rangle=0\). Write \(v_2=\langle Z^2\rangle\), \(m_3=\langle Z^3\rangle\), and \(H(t)=1/[2(1+\sqrt{1+t})^2]\) on \(t\ge-1\). The tangent identity and Jensen with the probability weight \(Z^2/v_2\) give \begin{equation}\label{eq:jens} 1-\langle\sqrt{1+Z}\rangle =\langle Z^2 H(Z)\rangle\ge v_2H(m_3/v_2),\qquad 1-\langle\sqrt{1+Z}\rangle\ge v_2/8-m_3/16 . \end{equation} Here \(H\) is decreasing and convex by differentiation. The second inequality uses its tangent at zero. If \(v_2=0\) the correction is zero. With two residues of relative order \(o\), write \(X,Y\) for their squared magnitudes. The fluctuation is \(2\sqrt{XY}\cos\theta\) on an \(o\)-gon. Order two uses the Euclidean bound. For \(o\ge3\) its second moment is \(2XY\) and its third at most \(2(XY)^{3/2}\), yielding \(\eta=1/5\) from \(m_3/v_2\le1/2\) and \eqref{eq:jens}. If \(o\ge4\) the third moment vanishes, giving \(\eta=1/4\). If \(o\ge6\), all odd moments through order five vanish and \(\langle Z^4\rangle=6X^2Y^2\). The degree-five Taylor polynomial of \(\sqrt{1+Z}\) at zero is an upper bound on \([-1,1]\) because the sixth derivative is negative on \((-1,1]\), with the left endpoint by continuity. Its even coefficients give \((\eta,\zeta)=(1/4,15/64)\). For three residues \(0,u,v\), let \(A,B,C\) be the normalized squared magnitudes in that order. In any moment of degree \(j\), the character products have frequency \(mu+nv\), where \(\max(|m|,|n|,|m+n|)\le j\). If no nonzero relation of this size occurs through degree three, the only surviving products are opposite edges and oriented triangles. Thus \(v_2=2(AB+AC+BC)\), \(m_3=12ABC\). The ratio \(m_3/v_2\le2/3\) follows from \(1/A+1/B+1/C\ge9\), with zero cases by continuity. This yields the all-pair loss \(21(AB+AC+BC)/125\) from \eqref{eq:jens}. For progression \(0,j,2j\) of step order at least seven, the positive harmonics of \(Z\) are \(U=p_1\bar p_0+p_2\bar p_1\), \(W=p_2\bar p_0\), so \(v_2=2(|U|^2+|W|^2)\), \(m_3=6\Re(U^2\bar W)\). The norm \(\Psi_{\rm AP}\) follows from \(|W|\le1/2\), \(|U|\ge x|y-z|\), and the second inequality of \eqref{eq:jens}. For \(0,j,3j\) of order at least ten, distinguish \(U=p_1\bar p_0,\ V=p_2\bar p_1\) and the third harmonic \(p_2\bar p_0\). The only extra cubic products arise from \(1+1-2=0\), and \(m_3\le12ABC+6AB\sqrt{BC}\). Putting \(u_0=\sqrt{B/C}\) and \(v_0=C/A\), the ratio is bounded by \[ \frac{3u_0^2(u_0+2)v_0}{(1+(1+u_0^2)v_0)(1+u_0^2+u_0^2v_0)} \le \frac{3u_0^2(u_0+2)}{(1+u_0+u_0^2)^2}\le\frac{17}{16}. \] The first inequality is the arithmetic--geometric mean inequality on the constant and quadratic terms of the denominator divided by \(v_0\). For the second, the cleared difference is \(17s^4+54s^3+15s^2-30s+9\) at \(s=u_0-1\). For \(s\in[-1,0]\) it is bounded below by \((1+s)(9-39s)+17s^4\). For \(s\ge0\), the cubic part is convex, has derivative \(-2\) at \(s=1/3\) and positive derivative at \(s=1/2\). Its value at \(1/3\) is \(8/3\), so its tangent is positive on \([0,1/2]\) and the cubic is increasing thereafter. The resulting loss in \eqref{eq:jens} again exceeds \(21(AB+AC+BC)/125\). A third estimate has edge set consisting of the two pairs containing a distinguished coefficient of squared magnitude \(A\), and weight \(1/5\). Suppose the samples, after a unit rotation, have the form \(p_0+(p_1+\delta p_2)\chi\) where \(\delta\) ranges over the roots of order two or three, and conditional on \(\delta\) the character \(\chi\) ranges over a rotated polygon of order at least four. Put \(\Delta(\delta)=2\Re(p_1\bar p_2\bar\delta)\). Conditional second and third moments of the first harmonic arise from its two opposite frequencies, and its odd moments through degree three vanish. For order two put \(J=|\Delta(1)|^2\). Then \(v_2=2A(1-A)+J\), \(m_3=6AJ\). With \(t=J/(2A(1-A))\) and \(D=(1+t)(1+2t)\), the bound \(J\le(1-A)^2\) gives \(m_3/v_2\le 6t/D\). The identity \[ (1+5t)D^2-12tD-9t^2=1-t-2t^2+53t^3+64t^4+20t^5\ge0 \] holds for \(t\ge0\), by using \(1-t-2t^2\ge0\) on \([0,1/2]\) and \(53t^3\ge t+2t^2\) thereafter. It implies \((2+3t/D)^2\le5(1+t)\). Then \(1+\sqrt{1+6t/D}\le2+3t/D\) and \eqref{eq:jens} imply loss at least \(A(1-A)/5\). The cases \(A=0,1\) follow from Jensen without a correction. For order three, \(\langle\Delta^2\rangle=2BC\) and \(\langle\Delta^3\rangle=2\Re((p_1\bar p_2)^3)\le(B+C)BC\). Thus \(v_2=2(A(B+C)+BC)\) and \(m_3\le12ABC+(B+C)BC\). The tangent lower bound for the loss minus \(A(B+C)/5\) is \[ A(1-A)/20+BC(3-11A)/16. \] If \(3-11A<0\), replace \(BC\) by \((1-A)^2/4\), giving \((1-A)(55A^2-54A+15)/320\ge0\). This proves the stated bound for order three as well. For use in interpolation, each formula is a norm, invariant under independent phases and increasing in component magnitudes. For \(\Psi_{\eta,\mathcal G}\), at \(S=1\) a unit tangent direction \(v\) satisfies \(\sum u_i=0\) for \(u_i=\Re(\bar p_i v_i)\). With \(P=\sum_{\mathcal G}|p_ip_j|^2\), \[ D^2P[v,v]=2\sum_i |v_i|^2\sum_{j:\{i,j\}\in\mathcal G}|p_j|^2 +8\sum_{\mathcal G}u_iu_j\le4 . \] In fact the sum of the positive \(u_i\) equals the sum of absolute negative \(u_i\), at most \(1/2\), so the positive pair products total at most \(1/4\). The tangent Hessian of the norm is at least \(1+3\eta P-4\eta\ge0\). Homogeneity gives zero radial Hessian. The partial derivative in a magnitude \(p\ge0\) is at least \((1-2\eta)p/\sqrt S\) since the adjacent magnitudes have total squared sum at most \(S\). For \(\Psi_{\rm AP}\) use \(P=y^2z^2/4+x^2(y-z)^2/16\). The largest Hessian eigenvalue of \(u^2v^2\) is at most \(3(u^2+v^2)\), by its two-by-two matrix. Apply this to \((y,z)\) and to \((x,(y-z)/\sqrt2)\) to obtain \(D^2P\le15/16\) in unit directions on \(S=1\). For coordinate derivatives of the correction use \((y-z)^2\le S\) in the \(x\)-derivative and \((y-z)_+\le y\) in the \(y\)-derivative, and interchange \(y,z\) for its last derivative. These yield nonnegative derivatives of the norm. Hence this formula is convex and increasing on the nonnegative orthant. The triangle inequality in complex coordinates follows from coordinatewise magnitude domination and convexity. Continuity gives the assertions at zero. For the extra two-component norm on real magnitudes \(x,y\), it equals \[ \left\langle |x\cos t+y\sin t|\, \frac{\pi}{16384}(7891-4740\cos4t+945\cos8t)\right\rangle_t. \] Indeed the three integrals of \(|\cos t|\) times \(1,\cos4t,\cos8t\) over a full period are \(4,-4/15,-4/63\). In polar coordinates this gives \(\sqrt S(7891+316\cos4\theta-15\cos8\theta)/8192\), which is the claimed formula. The density is positive and invariant under coordinate reflections. This gives convexity and coordinatewise monotonicity on magnitudes, and again the triangle inequality in complex coordinates. For all norms used here \(0\le\Psi\le\sqrt S\). Consequently \(|\Psi(p)-\Psi(q)|\le\|p-q\|_2\). \subsection{Directed integration and interpolation}\label{subsec:quadrature} If \(P=(\sum_n c_{jn}z^n)_j\), use an equiangular grid of \(J\) nodes with \(J\) larger than each component's frequency span. Independently shift each component by a unit monomial of integer degree \(-s_j\). The exact angular bound is \begin{equation}\label{eq:grid} \langle\Psi(P)\rangle\le J^{-1}\sum_{v=0}^{J-1}\Psi(P(e^{2\pi iv/J})) +\frac{(2\pi/J)^2}{12}\left(\sum_{j,n}|c_{jn}|^2(n-s_j)^4\right)^{1/2}. \end{equation} To see this, interpolate \(P\) linearly over each angular step of length \(h_0=2\pi/J\). Convexity controls the norm of the interpolant. At relative position \(s\), its difference from \(P\) is bounded by integrating \(|P''|\) against the absolute Green kernel \(h_0^2\min(s,v)(1-\max(s,v))\) for \(0\le v\le1\). For each \(v\) the mean over all steps of the derivative magnitude is at most its root mean square, which is the displayed coefficient sum by discrete orthogonality. Integrating over \(s,v\) gives \(h_0^2/12\). The directed version uses coefficient real and imaginary parts rounded down at scale \(S_0=2^{30}\), with the total absolute rounding error added. Grid roots are approximated with complex error \(<3/S_0\) and the root-error allowance is \(3/S_0\) times the sum of the absolute rounded real and imaginary parts of the coefficients. The same root bound with rational turn argument applies to \(e^{i\pi t}\) at a corner. One construction uses \(\pi=16\arctan(1/5)-4\arctan(1/239)\), alternating sums of 24 and 8 terms and directed truncations at \(2^{64}\), with errors bounded by \(16\cdot25+4\cdot9<500\) units at that scale. After quadrant reduction, cosine and sine on \([0,8/5]\) use the terms of degree at most 32 at that scale. The argument error is below 251 units. Term truncation and the Taylor remainder add at most 17 units per component. Truncating the result at scale \(2^{30}\) therefore gives the stated root error. Evaluate the sample norms by their squared-component formulas. In \(\Psi_{\rm AP}\), use the lower bound \(\max(0,y^2+z^2-2\sqrt{y^2z^2}_{\,\rm up})\) for \((y-z)^2\). For any rational \(t\ge0\) an upper root at scale \(S_1\) is \((\lfloor\sqrt{\lfloor tS_1^2\rfloor}\rfloor+1)/S_1\). For exact integer sample energies combine the rational correction factor with the upper integer root and round upward in the sample units. Use \(44/7\) for \(2\pi\) and upper roots at scale \(2^{50}\) in the curvature term, minimizing each component's sum over \(s_j\) in its mode range. These bounds also apply to the zero sample. The three-knot tables use rectangular parameter boxes. For a fixed raw test row, let \(D_{N,i}\) bound the absolute pure second derivative in coordinate \(i\) of \(N\), and let \(D_{k,i}\) bound the sum of squares of the magnitudes of the corresponding coefficient derivatives in \(P_k\). Successive linear interpolation contributes at most \begin{equation}\label{eq:interp} \sum_i\frac{w_i^2}{8}\left(D_{N,i}+C\sum_k\sqrt{D_{k,i}}\right) \end{equation} to the loss of the minimum corner margin, where \(w_i\) are the widths. This follows from the interval interpolation kernel, convexity of the means, and the Euclidean coefficient error bound by Parseval. For a column, sum the absolute raw weights times the pure derivative bounds of \eqref{eq:transport}. Sum over known columns for \(D_{N,i}\), and square and sum at the replaced pivots for \(D_{k,i}\). A known phase column with phase \(e^{i\pi t}\) contributes at most \((22/7)^2\) times its raw weight magnitude in the phase derivative and is independent of the phase in \(N\). All other columns are phase independent. For the fractional monomial factors \(a^{p(u)}b^{q(u)}\) used below, \(p,q\) are nonnegative affine functions of \(u\). The \(a\)-second derivative is \(p(p-1)a^{p-2}b^q\), and analogously for \(b\). Bound the factor \(|p(p-1)|\) by its largest value at the exponent endpoints or \(p=1/2\) if present. For the remaining power use the lower exponent, taking the lower base endpoint if that exponent is negative and the upper endpoint otherwise. For \(u\) the derivative bound is \(a_{\rm hi}^{p_{\min}}b_{\rm hi}^{q_{\min}}(|p'|\Lambda_a+|q'|\Lambda_b)^2\), where \(\Lambda_a\ge-\log a_{\rm lo}\). With \(z=(1-a_{\rm lo})/(1+a_{\rm lo})\) take \(2(\sum_{j=0}^5 z^{2j+1}/(2j+1)+z^{13}/(13(1-z^2)))\); likewise for \(b\). Endpoint powers with rational exponent \(j/k\) are bracketed by comparing \(k\)-th powers to the \(j\)-th power of the base. Equivalently enclose the exponent by dyadic rationals and bracket the powers by repeated directed square roots. At a corner the enclosure midpoint can be used for a factor, charging its radius times the absolute raw weights in the numerator or the row error. In all cases the full corner and interpolation allowances must be subtracted before testing positivity. For tables using vertex-dependent raw rows, interpolate the raw entries by the tensor weights before doing coefficient transport. For one coordinate of width \(w\), let \(M\) bound the endpoint magnitudes and \(D\) the edge differences of the raw entry. With first and second derivative bounds \(B_1,B_2\) of its factor, the pure interpolation error for the product is at most \(w^2 M B_2/8+w D B_1/4\). This is the second derivative estimate using \(|w_{\rm raw}'|\le D/w\). Use \((22/7),(22/7)^2\) for the phase factor. Sum over coordinates and send these errors through the same linear numerator and Euclidean coefficient bounds as before. If \(F_v\) are directed lower bounds on the corner margins, the margin throughout the box is at least \(\min_v F_v\) minus these errors. Fixed rows are the special case \(D=0\). Finally a finite set of closed positive-width boxes with disjoint interiors lying in an ambient box and having total volume equal to that ambient volume covers every point of it. Otherwise a relatively open portion of its complement would have positive volume. We use this check for the listed tables, including shared faces and corners. \section{Exclusion of a strict local maximum of the gaps}\label{sec:peak} \begin{theorem}\label{thm:peak} If all three knot norms are strictly below \(C_0\), then the four selected exposed indices cannot have consecutive gaps \(d,E,D\) with \(d1/2500\). Applying it in reverse gives \(\beta>1/2500\), using finite reversal as in Section~\ref{sec:reduction}. At the middle knot, project to the common residue of the tied terms and gauge the two coefficients to 1. If the projection is \(F(z)\), take the real part of \(e^{-i\theta/2}F(e^{i\theta})/2+e^{i\theta/2}F(e^{-i\theta})/2\). It is at most \(C_0\). The ties contribute \(2c\), \(c=\cos(\theta/2)\), and the weights on the two tails at displacements \(j\ge1\) are \(\cos((j+1/2)\theta)\). For \(0\le Y<1\), set \[ \begin{aligned} S(Y,c) &=\sum_{j\ge1}Y^{2j}\cos^2((j+1/2)\theta)\\ &=\frac{Y^2}{2(1-Y^2)}+\frac{Y^2(4h^3-3h-Y^2h)}{2(1-2Y^2(2h^2-1)+Y^4)}, \quad h=2c^2-1 . \end{aligned} \] The identity follows by summing the doubled-angle geometric series. The left tail comes from distinct coefficients at the left knot outside its two ties and the middle knot's upper tie, whose normalized magnitude there is \(y\). Transport weights are \(y^j\); the right side has the analogous weights \(\beta^j\). Parseval and Cauchy--Schwarz bound the adverse tails by \(B(U,c;T_i)=\sqrt{(T_i-U^2)S(U,c)}\) at \(U=y,\beta\), with either \(T_i=T_0\) or \(T_i=T\). If \(2Y^2\le T_i\), this bound is increasing in \(U\in[0,Y]\) since each \((T_i-U^2)U^{2j}\) is increasing there. We also need a refined right tail bound when \(D/g\ge6\) and \(E/g\ge7\) for \(g=\gcd(D,E)\). Reflect the right knot and index from its lower tied term in the reflected order. Its two ties now have indices \(0,D\), with phases made 1 by a rotation. The known coefficient at index \(D+E\) has magnitude \(\beta\). The coefficients transporting to the tail have reflected indices \(D-jE\), with factors \(\beta^j\). Let \(G(\theta)\) be the normalized full profile at this reflected knot. Take \(P(\theta)=3(1+e^{iD\theta})\), \(W=\max(1,|P|/C_1)\), \(V=P/W\). Expanding squared moduli pointwise gives \[ |G-P|^2=|P-V|^2+W|G-V|^2+(W-1)(C_1^2-|G|^2). \] The two ties give \(\langle|G-P|^2\rangle\le C_1^2+6\). Set \(\alpha=C_1/6=57/200\), \(t_0=\arccos\alpha\). Direct integration gives \[ \begin{aligned} \langle|P-V|^2\rangle &=\tfrac2\pi((18+C_1^2)t_0-9C_1\sin t_0)>7.679,\\ g_0:=\langle W^{-1}\rangle &=1-\tfrac2\pi\bigl(t_0-\alpha\log((1+\sin t_0)/\alpha)\bigr) \in(7/25,107/200). \end{aligned} \] For the rational bounds use \(t_0>64089/50000\) from the degree-ten cosine lower polynomial, \(\sin t_0<47927/50000\) by squaring, and \(\log((1+\sin t_0)/\alpha)<193/100\) from the degree-eight exponential sum. Use \(\pi<22/7\) and \(g_0\ge\alpha\). Thus \(\langle W|G-V|^2\rangle\le B_0=12451/10000\). In this weighted inner product the representative of a coefficient at \(k\) is \(e^{ik\theta}/W\). Its Gram entry against the representative at \(l\) is the \(k-l\) coefficient of \(W^{-1}\), and the diagonal is \(g_0\). Off-diagonal entries vanish unless \(D\mid k-l\). Writing \(m=(k-l)/D\), two integrations by parts on \(0\le\phi\le 2t_0\) give the off-diagonal bound \(\tan(t_0)/(\pi m^2)\). Indeed the loss \(1-W^{-1}=1-\alpha\sec(\phi/2)\) on that arc vanishes at \(2t_0\), has derivative zero at \(0\), and its decreasing derivative ends at \(-\tan(t_0)/2\). The endpoint and derivative variation have combined magnitude \(\tan(t_0)\). We use \(\tan t_0<7/2\). For the frequencies \(D-jE\), the first five are in distinct residue classes modulo \(D\). The off-diagonal row sum is at most \(2\cdot (7/2)/(\pi (E/g)^2)\sum_{k\ge1}k^{-2}\le\pi/42<2/21\). The row at \(D+E\) has nonzero entries against these tail frequencies only for \(j+1\) a multiple of \(D/g\), hence \(j\ge5\). Its cross entries have squared sum at most \((7/(2\pi \cdot 7^2))^2\sum_{k\ge1}k^{-4}<1/40^2\). Projecting \(G-V\) orthogonally off the representative at \(D+E\) leaves squared weighted norm at most \(B_0-\beta^2/g_0\) because \(V\) vanishes at this frequency. The projected-off part contributes at most \((25\beta/(7\cdot40))\) times the test's Euclidean tail from \(j=5\). Finally \(V\) has negative coefficients at these positions only from \(j= D/g,2D/g,\ldots\), thus from \(j\ge6\). For \(m=jE/D\) at such a position, its coefficient is the negative of the \((1-m)\)-th coefficient of \(P-V\) in the variable \(e^{iD\theta}\). Integrate \(3(1+e^{i\phi})-C_1 e^{i\phi/2}\) over \(|\phi|\le 2t_0\) against \(e^{i(m-1)\phi}\), using \(C_1=6\cos t_0\). This gives the coefficient of \(P-V\) as \[ \frac{3}{2\pi(m-1/2)} \left(\frac{\sin(2(m-1)t_0)}{m-1}-\frac{\sin(2mt_0)}{m}\right). \] For \(m\ge7\) its absolute value is at most \(7/(2\pi m^2)\). Consequently this contribution is at most \(1/40\) times the test's Euclidean tail from \(j=6\). Put \(G_0=107/200\), \(L_0=S(Y,c)\), \(L_2=Y^{12}/(1-Y^2)\), \(L_5=Y^{10}/(1-Y^2)\), and \(J_0=(2/21)(2\sqrt{L_0 L_2}+L_2)\). The squared weighted norm of the sum of the tail representatives with weights \(\beta^j\cos((j+1/2)\theta)\) is at most \((\beta/Y)^2(g_0 L_0+J_0)\). In fact the first five have diagonal Gram block, and the rest and the cross block are bounded by the row-sum estimate and Cauchy--Schwarz. Combining the three contributions just bounded, for \(\beta\le Y\), \(2Y^2\le B_0G_0\), gives the upper bound \begin{equation}\label{eq:weighted} B_{\rm w}(Y,c)=\sqrt{(B_0G_0-Y^2)(L_0+J_0/G_0)} +\frac{25Y}{280}\sqrt{L_5}+\frac{\sqrt{L_2}}{40}. \end{equation} For the first root, \((B_0g_0-\beta^2)(L_0+J_0/g_0)\) is increasing in \(g_0\), and \(\beta^2(B_0G_0-\beta^2)\) is increasing for \(\beta^2\le Y^2\). The other terms can use \(\beta\le Y\) directly. All estimates hold first for finite coefficient sums and then for the full sums by geometric decay. Arbitrary phases are allowed in the bound because the Hilbert norm estimates use only test magnitudes and Gram absolute values. In the coincident case \(u+v=1\) with \(u\le1/46\), the reduced orders for the right side satisfy \(E/g\ge46\) and \(D/g\ge45\). The middle Parseval inequality is \(y^{2u}+\beta^{2v}\le T_0\). In particular \(y\le T_0^{23}<.15177\). Whenever \(y\le Y\), test an upper ceiling \(b\) on \(\beta\) using \(2c-B(Y,c;T_0)-B_{\rm w}(b,c)>C_0\). This forces \(\beta>b\), and then \(y<(T_0-b^2)^{23}\) since \(v<1\). The three rational tests, with upward roots in the subtracted terms, are \[ \begin{array}{c|ccc|c} Y& b&c&10^9\,\text{lower excess over }C_0&\text{next upper }y\\ \hline .15177&.27&.915&162141&.0228\\ .0228&.425&.95&309529&.00101\\ .00101&.46&.955&169068&.00038 \end{array} \] The final bound contradicts \(y>1/2500\). Reflection shows that \(v>1/46\) as well. \subsection{Models for the strict peak} If \(d+D\ne E\), the three residues of \(n_1,n_0,n_3\) modulo \(E\) are distinct. In units \(E\) relative to \(n_1\) they have frequencies \(n,n-u,n+1+v\), \(n\in\mathbb Z\). Write \(F,H,G\) for the components, with known pivot anchors \(F_0,F_1,H_0,G_0\). An angular rotation fixes \(F_1\) relative to \(F_0\), and independent unit component factors fix \(H_0,G_0\). Let \(s\) denote time measured in units of \(E\) from the first knot. For artificial ceilings \(\bar u=r/h\ge u\), \(\bar v=s_1/h\ge v\), multiply \(H\) by \(e^{-(\bar u-u)s}\) and \(G\) by \(e^{(\bar v-v)(s-L-R)}\). These factors are at most one, preserve the unit anchors, and replace the transport frequencies by \(n,n-r/h,n+1+s_1/h\). By the norm properties in Section~\ref{subsec:angular} this operation bounds the test mean by that for the transported row at the artificial ceilings. The angular fiber still uses the actual residue characters. Let the individual orders of the nonzero residue characters be \(E/\gcd(E,d)\) and \(E/\gcd(E,D)\). If both are at most nine, both ratios are at most \(8/9\) and the Euclidean norm applies. If one has order two or three and the other at least ten, reverse if needed so the smaller order is on the left. Factor out the character of the right component in the scalar mean, leaving \(G+(F+\delta H)\chi\). Conditional on \(\delta\), \(\chi\) has order at least four because taking its second or third power reduces order by at most that divisor. Apply the two-edge norm centered at \(G\), with ceilings \(2/3,1\). It remains that both orders are at least four and at least one is at least ten. If the difference of the two exterior residues has order two or three, the same conditioning with distinguished component \(F\) applies. Otherwise classify the relations \(m u_0+n v_0=0\bmod E\) for the two nonzero residues \(u_0,v_0\), with \(\max(|m|,|n|,|m+n|)\le3\). Neither multiplier can be zero because both individual orders are at least four. Up to overall sign and interchange, the same-sign pairs are \((1,1),(1,2)\), giving an arithmetic progression. Opposite-sign magnitudes are \((1,1),(1,2),(1,3),(2,2),(2,3),(3,3)\). The first is impossible by distinctness, the second gives an arithmetic progression, the first-equal \((2,2),(3,3)\) give the difference cases just treated, and \((1,3),(2,3)\) give translates or reflections of \(0,j,3j\). For example \(2u_0=3v_0\) makes \(u_0=3j,v_0=2j\) with \(j=u_0-v_0\). The step in either configuration has order at least the larger individual order, since all three residues differ by multiples of the step. The progression bound works for any midpoint, and reversal exchanges \(H\) and \(G\). The sparse and relation-free cases use the all-pair bound. These four models use unit ceilings. If \(d+D=E\), combine the exterior components in the residue \(J\) with frequency \(n-u\), so the four anchors are \(F_0,F_1,J_0,J_2\). Now \(J_0=1\) after gauging and \(J_2=\nu\) is a single common unit phase. Multiply the raw \(J_2\) weights by \(\nu\) while forming the tests. All norms and the real pairing are invariant under whole-profile conjugation with angular reversal, so use \(\nu=e^{i\pi t}\), \(0\le t\le1\). This phase is fixed before the choice of row. Finite reversal gives \(u=1/2\) and, if needed, \(a\le b\) for order two. For order three it gives \(u=1/3\). At orders at least four orient so \(u\le1/2\). The proved lower ratio and radial bounds, with outer Parseval \(y^2,\beta^2\le T_0\), then put the parameters in \[ a=y^{1/8},\ b=\beta^{1/8}\in[3/8,199/200],\quad u\in[1/46,1/2],\quad t\in[0,1]. \] Use \(\Psi^{(2)}\) with \((\eta,\zeta)=(0,0),(1/5,0),(1/4,0)\) in the three order ranges, respectively. \subsection{Box comparisons} The embedded table S (Table~\ref{tab:digital}) gives the raw rows and the closed box partitions for the nine models. For the first six models let the parameters be \(a,b\in[0,1]\), and set the scaled frequencies \(m=hn,hn-r,hn+h+s_1\). For orders two and three the box is \([0,1]^3\) in \(a,b,t\) and \(m=hn,h(n-u)\). For the last model use the compact box just displayed with \(h=8\) and the same formulas. There, on the \(J\) column the pivot is fixed at each \(n\) across the whole box because \(0T\), where in the coincident model \(r=hu,\ s_1=h(1-u)\). At the last model use the lower value \(a_{\rm lo}^{16u_{\rm hi}}+b_{\rm lo}^{16(1-u_{\rm lo})}>T\). The energy bound uses the actual exterior coefficients or their smaller artificial values. For a reflection box \(a_{\rm lo}>b_{\rm hi}\); in the symmetric models finite reversal permits \(a\le b\). Reversal preserves the strict norm cap by using small approximation slack, and the models then use the new actual phases. For the 180 radial boxes the ceilings on \(y,\beta\) satisfy either \(\max(y,\beta)\le.19\) or \(\min(y,\beta)\le.15,\ \max(y,\beta)\le.23\). Applying \(B(U,.91;T)\) at the indicated ceilings gives a lower two-point value exceeding \(C_1\) by \(0.001076624\) and \(0.000286547\), respectively. For the 32 ordinary two-point boxes use \(a\le.65,b\le.87,c=.935\) and \(T\); the two adverse bounds are at most \(3/200,143/1000\). For the four weighted boxes use \(a\le.72,b\le.87,u\le.08,c=.94\). The last bound on \(u\) gives \(D/g\ge6,E/g\ge7\), so the left bound \(B(.72^8,.94;T)\) is at most \(7/200\) and the right bound \(B_{\rm w}(.87^8,.94)\) at most \(16/125\). Both types give a lower value greater than \(C_0\). For each remaining box substitute its row into \eqref{eq:transport} and evaluate the directed corner bound and \eqref{eq:interp}, with the modes, grid size and norms shown below. For fractional powers use rational endpoint power enclosures (with a fixed-point enclosure at \(2^{42}\) in the indicated fixed-point rows) and charge their radii in the corners. Each comparison gives the positive lower margin shown in the last column, in units of \(10^{-9}\) or larger. \[ \begin{array}{c|c|c|rrr|r} \text{model}&(h,r,s_1)&\text{degree, grid}& \text{boxes}&\text{comparisons}&\text{other}&\text{floor}\\ \hline \text{Euclidean}&(9,8,8)&7,512&143&97&46&49174\\ \text{midpoint }F&(1,1,1)&7,512&144&83&61&738\\ \text{midpoint }H&(1,1,1)&6,256&83&77&6&207092\\ \text{all pairs}&(1,1,1)&6,256&81&48&33&1776421\\ \text{two edges at }F&(1,1,1)&6,256&81&48&33&4485532\\ \text{two edges at }G&(3,2,3)&6,256&81&79&2&971842\\ \text{coincident order }2&(2,1,1)&6,256&446&329&117&48684\\ \text{coincident order }3&(3,1,2)&6,256&336&312&24&201452\\ \text{coincident order }\ge4&(8,8u,8(1-u))&6,256&1628&1216&412&24754 \end{array} \] The volumes and disjoint interiors verified by the box endpoints give coverage, including \(t=0,1\). Every parameter point for an actual profile therefore belongs to a box with a contradiction to Parseval, the radial bounds, the orientation, or \eqref{eq:testmargin}. This proves Theorem~\ref{thm:peak}. \section{Exclusion of a decrease after two equal gaps}\label{sec:drop} \begin{theorem}\label{thm:drop} If all three knot norms are at most \(C_0\), the three consecutive gaps cannot be \(d,d,E\) with \(0C_0^2-2\). For the other prunes put \(X=a_{\rm lo}^h,\ Y=b_{\rm lo}^r\). At the middle knot use the known-coefficient test vector \(X+z+\xi z^2,\ Yz^3\). The real pairing is at least \(N=2+X^2+Y^2\). Its squared Euclidean norm has mean \(N\), upper bound \(M=(2+X)^2+Y^2\), and variance \(2(1+2X^2+2X\cos(\pi t))\), by the first two Fourier harmonics. Use the lower endpoint \(V\) of this variance obtained from the upper phase endpoint. The square-root tangent identity and the quadratic correction of \(\Psi^{(2)}\) bound the angular test mean by \[ \sqrt N-\frac{V}{2\sqrt N(\sqrt M+\sqrt N)^2} -\eta(2+X^2)Y^2/M^{3/2}. \] Thus the prune is valid if \(N\) exceeds \(C_0\) times the upper bound. The exact tests use upward roots for the denominators of the subtractions and a lower cosine in \(V\). In all other boxes the table gives either a fixed raw row or eight vertex rows. Evaluate with modes \(-9,\ldots,9\) and \(512\) grid points as in Section~\ref{subsec:quadrature}. Each used vertex lower bound is computed after the coefficient, phase and integration allowances. For a row interpolated from vertices, lower the smallest of its directed floors by the mixed interpolation allowance \(w^2MB_2/8+wDB_1/4\) for each raw entry and its factor, summed as specified there. The raw row for a fixed box uses \(D=0\) and checks all eight corners. The exact cover and margin counts in the supplementary table D (Table~\ref{tab:digital}) are \[ \begin{array}{c|r|r|r|rr|r} o&\text{boxes}&\text{interpolated}&\text{fixed}& \text{energy}&\text{known test}&10^9\,\text{minimum margin floor}\\ \hline 2&1377&1213&0&81&83&9863\\ 3\text{--}5&2084&1859&0&93&132&11901\\ \ge6&2345&2154&21&87&83&229 \end{array} \] There are 8,329 distinct used vertex rows. Each comparison box gives a strictly positive margin throughout its closure. The other boxes contradict the indicated energy or single-knot inequalities. This gives the contradiction in every box. \end{proof} \section{Localization on arithmetic windows}\label{sec:arith} \subsection{Bilateral profiles} Suppose arbitrarily late exposed runs with common gap \(d\) have lengths tending to infinity. The value of \(d\) may vary with the run. Choose a fixed cap \(C\) with \(Q(f)0\). If the two imaginary parts have opposite weak signs, the real part of the nearest contribution \(i(v-u)\) to \(O_0\) has absolute value \(s_1+s_2\). The next two sine magnitudes are \(1/2\), and the remaining sum is bounded by \(2R^{10}/(1-R^5)\). Thus \(|O_0|\ge(s_1+s_2-h)_+\). Using the mean of the two squared values and \((s-h)_+^2\ge s^2-2sh\), their greater squared value is at least \[ 3-\left[(3R/c-1)(s_1^2+s_2^2)-2s_1s_2+2h(s_1+s_2)+6e\right]. \] The bracket is separately convex on \(0\le s_i\le R(240/289)\). At its four corners substitution bounds it by \(3-C_0^2-.003\). This contradiction excludes opposite weak signs, including zero. Hence after one common conjugation we can use the positive sector with \(t\in[.6,1]\). For the next cover (A4), use the fixed two circles as in the third cover but apply the sector of product degrees \(T_{n-1}\) for \(n=2,3\), with inner radius zero. In other degrees use disks. This gives \(|w|>.10\). For A5 use \(k=0,1,2\), with only the corrected group at depth 2, inner radii zero for \(n=2,3\), disks otherwise, and \(\epsilon=.0075\). The coefficient ceiling can be factored as \(R^{(n-2)(n-3)/2}\) with multipliers \(R^{T_k-3+(2\pm k)n}\), as follows by moving this radial power from the pairing factor. Enclose each individual power times its moment before summing, and use the sharper \(K=2\) tail. The conclusion is \(|w|>.15,\ t<.76\). For A6--A8 use \(k=0,1,2\) and the corrected group of depth 2 in addition, with \(R=.32,m=.15\), phase window \([.6,.76]\), annular supports for \(n=2,3,4\) and disk supports otherwise. Use the same factored multipliers as in A5 and the sharper tail. For A7 apply \eqref{eq:arithslope}. A6, A7 and A8 successively give the lower radius \(.17\), lower radius \(.19\), and then \(.62R^2\), the stronger bound at \(L_2\) is \(C_0|\alpha_2|\) and that at \(P_2\) is \(C_0|\beta_2|\). Indeed the scaled displacement-one term has modulus \(Ry_{\rm hi}\) and is excluded by ordering \(x\le y\), which is allowed by reflecting the whole model. An entry \(-1\) has both curvatures in the disk of radius \(1/40\) about \[ c_*=(-191206+204897i)/10^6. \] For this check it suffices to compare the four endpoint distances for each curvature. The angle span is \(<\pi\), and the dot products of \(c_*\) with its endpoint rays are positive. Over the arc the dot product with \(c_*\) thus attains its minimum at an endpoint, and squared distance is convex in radius. At all other leaves entry \(j+4\) uses weight list \(j\) (numbering from zero). Each comparison exceeds \(C_0\) after deducting the displayed errors. There are respectively \(69,103,935\) leaves of the three types. In the four subtrees specified by the first two splits, the lower comparison excesses in units of \(10^{-9}\) are \(2272,2463,15313,9381\). The exact tree covers the ordered box. Its reflection gives the other order with the same curvature orientation, since the normalized reversed profiles exchange the two tail products. This proves \(|w_j-c_*|<1/40\) for all \(j\). \section{The sharp arithmetic comparison}\label{sec:spectral} \begin{theorem}\label{thm:arithmetic} If a transcendental entire function has arbitrarily late equal-gap runs of unbounded length, then \(Q(f)\ge A_*\). \end{theorem} \begin{proof} If \(Q(f)0>g_1\). We use weights \(\lambda,(1-\lambda)/2,(1-\lambda)/2\) at \(0,a,-a\), respectively. At any one of these angles \(x_i\), put \(F=|H_k(x_i)|^2,\ F'=(|H_k|^2)'(x_i)\). We use the test \(F+K_i(F-A_*^2)^2+c_i(F')^2\). All constants of the comparison are given below. \[ \begin{array}{c|rr} &i=0&i=1\\ \hline K_i&14/5&1/5\\ c_i&31/1000&17/500\\ d_i \text{ (window width)}&3/20&7/100\\ s_i \text{ (step)}&1/6&7/100\\ L_i \text{ (curvature bound)}&6&14\\ b_i \text{ (bound for }|H_*'(x_i)|\text{)}&0&6/5\\ p_i \text{ (value retention)}&5/8&7/10\\ q_i \text{ (derivative retention)}&1/4&3/4 \end{array} \] Put \(A=171/100\). Expanding the products along the segment from \(w\) to \(w_k\) bounds \(|H_k-H_*|,|H_k'-H_*'|,|H_k''-H_*''|\) by \(e_* A_1,e_* A_x,e_* A_{xx}\), respectively, where \(A_1=13/5,\ A_x=91/10,\ A_{xx}=34\). For the derivative of order \(j=0,1,2\), the sum for the constant is \(2\sum_{n\ge1} J_n^j T_n R_*^{T_n-1}\). All segment points have modulus at most \(R_*\). On \(|x-x_i|\le d_i\), the lower bound for the second derivative of \(|H_k|^2\) is \[ (|H_*|^2)''-4e_* A_x|H_*'|-2e_*(A A_{xx}+A_1|H_*''|+e_* A_1 A_{xx})>-L_i. \] The last inequalities and \(7/25<\lambda<3/10,\ |H_*'(a)|0\). Suppose all knot norms are at most \(A_*\). Then \(D=A_*^2-F\ge0\) and \((F')^2\le D/\gamma_i\). The deviation bound gives \(D\le D_{\max}=111/500\), using \(2A A_1e_*-(A_1e_*)^20\). In the interior angle range test \(-g_1>0,\ (1-7/25)(-g_1)-(7/25)g_0>0,\ (3/10)g_0-(1-3/10)(-g_1)>0\), and \(6/5-|\partial_xH_*|>0\). Midpoint bisection of the \(y\) range gives six leaves and lower margin exceeding \(0.001503\). For the angular second-derivative inequalities use \(y\in[391/400,1]\) about zero and \(y\in[563/1000,713/1000]\) about either interior angle. These contain the respective windows by \(|\sin x|\le|x|\) and \(|(\cos^2 x)'|\le1\). Substitute \((|H_*|^2)''=2\Re(\bar H_*H_*'')+2|H_*'|^2\) in the bound with \(-L_i\) and bisect in \(y\). There are six and four leaves with lower margins exceeding \(0.012456,0.238432\), respectively. For the Fourier test of the quadratic matrices, at zero angle use \(y=1\) and at the two interior angles take \(y\in[.633,.643]\). The interval matrix is the sum of \(h\mathcal E I\), the matrix for \(Z_\omega\), \(K_i p_i\) times the value row matrix and \(c_i q_i\) times the derivative row matrix. At each respective endpoint \(\lambda=7/25,3/10\) form the weighted sum, subtract \((4/3)I\), and test for positive lower endpoints of both diagonals and their product minus the off-diagonal square. Bisect any unaccepted rectangle in \(X\) if its \(X\)-width is at least three times its \(y\)-width, and in \(y\) otherwise. The resulting cover has 46 leaves, depth at most 10, and minimum lower endpoint among these three tests greater than \(0.357326\). Convexity of positive definiteness gives the bound for every intermediate weight. The finite test uses the polynomial frequency recurrences above throughout \([-1,1]\). By Fourier orthogonality, the angle-weighted mean of the test is at least \(A_*^2+(4/3-1)\Delta^2\). If \(\Delta>0\), it exceeds the pointwise cap \(A_*^2\). If \(\Delta=0\), all curvatures are \(w_*\) and the norm is \(A_*\). In both cases some knot norm is at least \(A_*\). The periodic closing argument now proves the theorem. \end{proof} \appendix \section{Tables for the exact comparisons} The numerical data of the proof are provided as JSON in the \texttt{anc/} folder of the source archive. Table~\ref{tab:digital} identifies the five tables and the corresponding formulas. Their entries are rational numbers, integer test weights and subdivision data. Table D is provided in five numbered JSON parts. Run \texttt{python anc/rebuild\_d.py} from the extracted source directory to recreate the original \texttt{anc/D-equal-gaps.json}. A rational endpoint is recorded as an integer or a fraction, and endpoints are in consecutive lower--upper pairs. Numbering of rows and group indices starts at zero unless a row identifier is given explicitly. \begin{table}[ht] \centering \begin{tabular}{cll} \hline Table & Data file & Comparison \\ \hline P & \textsf{P-plateau.json} & Radial Schur comparison in Section~\ref{sec:plateau} \\ S & \textsf{S-peaks.json} & Strict peaks in Section~\ref{sec:peak} \\ D & \textsf{D-equal-gaps-01.json} through \textsf{-05.json} & Equal gaps in Section~\ref{sec:drop} \\ A & \textsf{A-arithmetic-windows.json} & Covers in Section~\ref{subsec:arithcovers} \\ K & \textsf{K-adjacent-pairs.json} & Adjacent-pair disk in Section~\ref{sec:arith} \\ \hline \end{tabular} \caption{Exact comparison data.}\label{tab:digital} \end{table} \paragraph{Table P.} Each entry lists the endpoints of a rectangle in \(c,u\). Use the four rejection tests following the definition of \(P=(41/100)^2\). \paragraph{Tables S and D.} For each model, \textsf{rows} gives a row identifier and an array \textsf{weights}. Read the weights in order of knot, component and increasing integer mode. For distinct-residue models of S these are real integers. In the two-component models each entry consists of two consecutive integers giving real and imaginary parts. Divide all weight coordinates by \(2^{30}\). For vertex-dependent rows, \textsf{point} identifies the corner and \textsf{vertexFloor} is a lower bound for the corner margin in units of \(10^{-9}\). The box records have either a single row identifier, eight vertex row identifiers in lexicographic order of the endpoints, or a reason for exclusion. In the variable ratio model S the fixed-point power enclosures are indicated by the arithmetic field \textsf{fixed}. A comparison stores the lower margin floor after the interpolation and integration allowances, in the same units. \paragraph{Table A.} The groups A1 through A11 follow the comparison order in Section~\ref{subsec:arithcovers}. Each part lists the row lists \textsf{weights} and the cover records. In A1 a row lists its weight followed by the two angular parameters. Otherwise a row begins with its angular group, then the weight and the two angular parameters. In A1--A4 the group number is the virtual depth. In A5 the depth-2 rows are corrected. In A6--A8 the order is \(0,1,2,\text{corrected }2\). At greater depth the ordinary group is followed immediately by the corrected group at each depth \(k\ge2\). Ordinary cover records list four endpoints followed by a row-list index. A radial boundary part lists the lower and upper radial boundary in its root box, and each record lists its two phase endpoints and row-list index. Use \eqref{eq:arithslope} on the radial interval. \paragraph{Table K.} Use the nine fixed angular groups and the midpoint tree described in Section~\ref{sec:arith}. The angular row format is the same as in table A. A tree entry referring to a row list is offset by 4 to distinguish it from a coordinate split. \begin{thebibliography}{9} \bibitem{Bloom} T. F. Bloom, \emph{Erd\H{o}s Problem \#513}, \url{https://www.erdosproblems.com/513}. \bibitem{HT} Y. He and Q. Tang, \emph{Generalizing the Clunie--Hayman construction in an Erd\H{o}s maximum-term problem}, \href{https://arxiv.org/abs/2602.12217}{arXiv:2602.12217}, 2026. \bibitem{CH} J. Clunie and W. K. Hayman, \emph{The maximum term of a power series}, J. Analyse Math. \textbf{12} (1964), 143--186. \end{thebibliography} \end{document}