\documentclass[11pt]{amsart} \pdfsuppressptexinfo=-1 \usepackage[margin=1.1in]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools} \usepackage{booktabs,enumitem} \usepackage[hidelinks]{hyperref} \hypersetup{pdftitle={An explicit bound below 2/pi for the maximum term of an entire function}, pdfauthor={Tejasvi Singh Tomar}} \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} \setlength{\emergencystretch}{2em} \raggedbottom \newcommand{\R}{\mathbb{R}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\C}{\mathbb{C}} \newcommand{\T}{\mathbb{T}} \newcommand{\re}{\operatorname{Re}} \newcommand{\ip}[2]{\langle #1,#2\rangle} \newcommand{\norm}[1]{\lVert #1\rVert} % Values of the certificate (one place). \newcommand{\BETA}{0.63005} \newcommand{\BETAFR}{12601/20000} \newcommand{\LAMSTAR}{0.630041} \newcommand{\JSTAR}{0.6281} \newcommand{\NKLEAF}{176{,}240} \newcommand{\NKSKIP}{15{,}276} \newcommand{\NKTAIL}{18} \newcommand{\NGRID}{1421} \newcommand{\KMARGIN}{2.1\cdot10^{-7}} % Release slots. \newcommand{\ZENODODOI}{10.5281/zenodo.23716683} \newcommand{\REPOURL}{https://github.com/tejasvi/maximum-term-bound/tree/v4} \begin{document} \title[An explicit bound for the maximum term]{An explicit bound below $2/\pi$ for the maximum term of an entire function} \author{Tejasvi Singh Tomar} \renewcommand{\urladdrname}{{\itshape ORCID}} \urladdr{\href{https://orcid.org/0000-0003-4668-3639}{https://orcid.org/0000-0003-4668-3639}} \subjclass[2020]{Primary 30D20; Secondary 42A05, 65G20, 68V20} \keywords{Entire functions, maximum term, maximum modulus, central index, Wiman--Valiron theory, Newton polygon, computer-assisted proof, formal verification} \begin{abstract} For a transcendental entire function $f$ let $\mu(r)$ be its maximum term and $M(r)$ its maximum modulus on $|z|=r$. Erd\H{o}s asked for the largest possible value $B$ of $\liminf_{r\to\infty}\mu(r)/M(r)$. An argument of Clunie gives $B\le 2/\pi=0.636619\ldots$, and Clunie and Hayman proved $B\le 2/\pi-c$ for an absolute constant $c>0$ that they did not compute. We prove $B\le\BETA$. We make the stability argument of Clunie and Hayman quantitative by testing $f$ against trigonometric polynomials at one or both of two consecutive radii where the central index jumps. The proof is computer-assisted, and the theorem is also proved in the Lean proof assistant. \end{abstract} \maketitle \section{Introduction} Let $f(z)=\sum_{k\ge0}a_kz^k$ be a transcendental entire function, and for $r>0$ let \[ M(r)=\max_{|z|=r}|f(z)|,\qquad \mu(r)=\max_{k\ge0}|a_k|r^k \] be its maximum modulus and its maximum term. By Cauchy's inequality $\mu(r)\le M(r)$. The maximum term and the \emph{central index}, the largest $k$ with $|a_k|r^k=\mu(r)$, are basic quantities of the Wiman--Valiron theory of entire functions \cite{Hay74}. Erd\H{o}s \cite[p.~249]{Er61} asked how large the ratio $\mu(r)/M(r)$ can stay, that is, for the value of \[ B:=\sup_f\ \liminf_{r\to\infty}\frac{\mu(r)}{M(r)}, \] the supremum being over all transcendental entire functions (for polynomials the ratio tends to $1$). This is Problem~2.14(c) in the Hayman--Lingham collection \cite{HL19}.\footnote{It is also Problem~\#513 in Bloom's list of Erd\H{o}s problems \cite{EP513} and constant 51a in the repository of optimization constants \cite{TaoRepo}.} Clearly $B\le1$, and a suitable lacunary power series, all of whose terms except at most two are negligible compared with $\mu(r)$ at every large radius, shows that $B\ge1/2$. An unpublished example of K\H{o}v\'ari shows that $B>1/2$ \cite[p.~146]{CH64}. Gray and Shah \cite[Lemma~1]{GS63} published an argument of Clunie that gives $B\le2/\pi$. Clunie and Hayman showed that the same argument gives $\mu(r)/M(r)\le2/\pi$ at every radius where the central index jumps, and that this is sharp for the upper limit over these radii \cite[Theorem~1]{CH64}. For the lower limit they proved $4/70$ \cite[Theorem~2]{CH64}. Their constant comes from a stability argument, sketched below \cite[Lemmas~2.2--2.4]{CH64}. Its value is not computed there \cite[p.~156]{CH64}, and as far as we know no explicit value has been recorded since then (see \cite{EP513,HL19}). Recently He and Tang \cite{HT26} generalized the Clunie--Hayman construction and proved $B>1/1.70919>0.5850724$ \cite[Theorem~2.8 and Proposition~3.5]{HT26}, and a certified computation reported by Sothanaphan \cite{So26} gives $B>0.5850788$. \begin{theorem}\label{thm:main} For every transcendental entire function $f$, \[ \liminf_{r\to\infty}\frac{\mu(r)}{M(r)}\le\BETA . \] Consequently $B\le\BETA$. \end{theorem} The bound improves $2/\pi=0.636619\ldots$ by more than $0.0065$. \subsection*{Idea of the proof} Clunie's argument is short. Write $c_k=a_kr^k$ and $g(\theta)=f(re^{i\theta})=\sum_kc_ke^{ik\theta}$. At a radius where two indices $n0$ put $c_k(r)=a_kr^k$ for $k\ge0$ and $c_k(r)=0$ for $k<0$. \begin{lemma}[window bound]\label{lem:window} For every $r>0$ and every finite $S\subset\Z$, $M(r)\ge N_S\bigl((c_k(r))_{k\in S}\bigr)$. \end{lemma} \begin{proof} The function $g(\theta)=f(re^{i\theta})$ satisfies $\norm g_\infty=M(r)$ and $\widehat g(k)=c_k(r)$ for every $k\in\Z$. \end{proof} \section{Vertices, events and the two reductions}\label{sec:events} \subsection{The Newton polygon} The facts in this subsection are classical (see, e.g., \cite{Hay74}), but we include short proofs to keep the paper self-contained. Let $f$ be transcendental and $L(t)=\log\mu(e^t)=\max_k\ell_k(t)$, where $\ell_k(t)=\log|a_k|+kt$ for $a_k\ne0$. The maximum is attained because $|a_k|e^{kt}\to0$ as $k\to\infty$. The \emph{central index} $\nu(r)$ is the largest $k$ with $|a_k|r^k=\mu(r)$. It is nondecreasing in $r$. Call $k$ a \emph{vertex} if $\ell_k(t)=L(t)$ for some $t\in\R$, and let $v_00$, contradicting $\ell_k(t)=L(t)$. The indices attaining $L$ on a bounded interval $[a,b]$ are at most $\nu(e^b)$, so only finitely many $I_k$ meet $[a,b]$. There are infinitely many vertices, since otherwise $|a_k|r^k\le\max_{j\le N}|a_j|r^j$ for all $r$ and $k>N$ would force $a_k=0$ for $k>N$. As the $I_{v_j}$ cover $\R$, are ordered and are locally finite, consecutive ones meet: \[ t_j:=\max I_{v_{j-1}}=\min I_{v_j}\qquad(j\ge1), \] so that $t_1\le t_2\le\cdots$, $t_j\to\infty$, and at $r=e^{t_j}$ both $v_{j-1}$ and $v_j$ attain the maximum term. \begin{definition} The \emph{event} at $j\ge1$ consists of the vertices $v_{j-1}0$; \item $|c_{v_{j-1}}(r_0)|=|c_{v_j}(r_0)|=\mu(r_0)$ and $|c_{v_j}(r_1)|=|c_{v_{j+1}}(r_1)|=\mu(r_1)$; \item $c_k(r_1)=c_k(r_0)e^{ks}$ for all $k$. \end{enumerate} \subsection{Equal gaps} At an equal-gap event we use the five coefficients at $v_j+dl$, $|l|\le2$: the two pairs that tie at $r_0$ and at $r_1$, and one further coefficient on each side (wider windows are discussed in Remark~\ref{rem:seven}). Let $K=[0,\infty)\times[0,1]^2\times\R^3$, with points $\xi=(\sigma,x,y,\varphi_{-2},\varphi_{-1},\varphi_2)$, and \begin{align*} C_0(\xi)&=\bigl(xe^{i\varphi_{-2}},\ e^{i\varphi_{-1}},\ 1,\ e^{-\sigma},\ ye^{-2\sigma}e^{i\varphi_2}\bigr),\\ C_1(\xi)&=\bigl(xe^{-2\sigma}e^{i\varphi_{-2}},\ e^{-\sigma}e^{i\varphi_{-1}},\ 1,\ 1,\ ye^{i\varphi_2}\bigr),\qquad \Lambda(\xi)=\min_{i=0,1}\frac1{N_5(C_i(\xi))}. \end{align*} ($N_5(C_i(\xi))\ge1$ by Lemma~\ref{lem:dual} with $T_\tau=1$.) \begin{proposition}\label{prop:equal} At an equal-gap event there is $\xi\in K$ with $\min\{\mu(r_0)/M(r_0),\ \mu(r_1)/M(r_1)\}\le\Lambda(\xi)$. \end{proposition} \begin{proof} Let $n=v_j$, $\lambda=c_n(r_0)$ and $b_l=c_{n+dl}(r_0)/\lambda$ for $l=-2,\dots,2$; so $|\lambda|=\mu(r_0)$, $|b_{-1}|=|b_0|=1$ and $x:=|b_{-2}|\le1$. By (E2) and (E3), $|b_1|e^{ds}=1$, so $|b_1|=e^{-\sigma}$ with $\sigma=ds$, and (E1) at $r_1$ gives $y:=|b_2|e^{2\sigma}\le1$. Choose $\alpha$ with $d\alpha=-\arg b_1$ and replace $b_l$ by $b_le^{idl\alpha}$; then $b_0=1$, $b_1=e^{-\sigma}$, and with $\varphi_l=\arg b_l$ ($l=-2,-1,2$; arbitrary if $b_l=0$) we have $b=C_0(\xi)$. Lemma~\ref{lem:window} with $S=\{n+dl:|l|\le2\}$ and Lemma~\ref{lem:props}(a)--(c),(f) give $M(r_0)\ge\mu(r_0)N_5(C_0(\xi))$. At $r_1$ we have $\mu(r_1)=|c_n(r_1)|$ and $c_{n+dl}(r_1)/c_n(r_1)=b_le^{l\sigma}$, which after the same rotation is $C_1(\xi)$; so $M(r_1)\ge\mu(r_1)N_5(C_1(\xi))$. \end{proof} \subsection{Mixed gaps} When the two gaps differ, the three vertices are not equally spaced. We then work at a single radius: the coefficients on the progression of the larger gap contain the two that tie there, and the coefficient of the remaining vertex, which lies off that progression, carries the information about the other jump. At a mixed event let $g=\gcd(d,d')$, $a=d/g$, $b=d'/g$ (so $\gcd(a,b)=1$, $a\ne b$), and \[ p:=\max(a,b)\ge2,\qquad \sigma:=gs. \] For $p\ge2$, $W\ge0$ and $\varepsilon\ge0$ let \[ R_p(W,\varepsilon)=\max_{\varphi\in\R}\ \frac1p\sum_{j=0}^{p-1}\bigl|W+\varepsilon e^{i(\varphi+2\pi j/p)}\bigr|-W, \] and for $\tau\in\C^5$ (indices $-2,\dots,2$) and $\varepsilon\ge0$ let \[ U_p(\tau,\varepsilon)=\norm{T_\tau}_1+\frac1{2\pi}\int_0^{2\pi}R_p\bigl(|T_\tau(X)|,\varepsilon\bigr)\,dX . \] For $s\ge0$ and $\kappa\ge1$ let $Z(s,\kappa)$ be the set of vectors $z=(z_{-2},z_{-1},1,1,z_2)$ with $|z_{-2}|\le e^{-2\kappa s}$, $|z_{-1}|\le e^{-\kappa s}$ and $|z_2|\le1$. \begin{proposition}\label{prop:mixed} At a mixed event there are $s_1\ge0$ and $z\in Z\bigl(s_1,\tfrac p{p-1}\bigr)$ such that for every $\tau\in\C^5$ and $\varepsilon\ge0$ with $\re\ip z\tau+\varepsilon e^{-s_1}>0$, \[ \min\Bigl\{\frac{\mu(r_0)}{M(r_0)},\frac{\mu(r_1)}{M(r_1)}\Bigr\}\le\frac{U_p(\tau,\varepsilon)}{\re\ip z\tau+\varepsilon e^{-s_1}} . \] \end{proposition} \begin{proof} Work on the progression $v_j+g\Z$ and write $w_k=c_{v_j+gk}(r_0)/c_{v_j}(r_0)$ and $w_k(\sigma)=c_{v_j+gk}(r_1)/c_{v_j}(r_1)$; by (E3), $w_k(\sigma)=w_ke^{k\sigma}$. By (E1) and (E2), $\mu(r_0)=|c_{v_j}(r_0)|$, $\mu(r_1)=|c_{v_j}(r_1)|$, $w_0=w_0(\sigma)=1$, $|w_{-a}|=1$, $|w_b(\sigma)|=1$, and $|w_k|\le1$, $|w_k(\sigma)|\le1$ for all $k$. \emph{Case $ab$} ($p=a$). Now the larger gap is the one between the two vertices that tie at $r_0$, so we work at $r_0$ on the progression $v_j+ag\Z$, which contains them. The third vertex, $v_{j+1}=v_j+bg$, lies off this progression and supplies the off-lattice coefficient. Here $|w_{-2a}|\le1$, $|w_{la}|\le e^{-la\sigma}$ ($l=1,2$; by (E1) at $r_1$), and the off-lattice coefficient $w_b$ has modulus $e^{-b\sigma}$, with $b$ prime to $a$. Reversing the window (Lemma~\ref{lem:props}(g)) moves the coefficient at $k$ to $-k$. The rotation of Lemma~\ref{lem:props}(b) by an $\alpha$ with $e^{ia\alpha}w_{-a}=1$ then makes the coefficient at $a$ equal to $1$ and keeps the one at $0$, so that the coefficients at $la$, $l=-2,\dots,2$, form the vector $z=(w_{2a}e^{-2ia\alpha},\,w_ae^{-ia\alpha},\,1,\,1,\,w_{-2a}e^{2ia\alpha})$. Since $|w_{2a}|\le e^{-2a\sigma}$, $|w_a|\le e^{-a\sigma}$ and $|w_{-2a}|\le1$, this vector lies in $Z(s_1,a/b)\subset Z(s_1,p/(p-1))$ with $s_1=b\sigma$, and the off-lattice coefficient, now at $-b$, has modulus $e^{-b\sigma}=e^{-s_1}$. Lemma~\ref{lem:window} on the window $v_j+g(\{-2a,-a,0,a,2a\}\cup\{b\})$, Lemma~\ref{lem:props}(a)--(c),(f),(g), and Lemma~\ref{lem:B3} with $S_L=\{-2a,\dots,2a\}$ (step $a$) give $M(r_0)\ge\mu(r_0)\,(\re\ip z\tau+\varepsilon e^{-s_1})/U_a(\tau,\varepsilon)$. \end{proof} \section{Averaging over roots of unity}\label{sec:averaging} \begin{lemma}\label{lem:B3} Let $p\ge2$, let $e\in\Z$ with $\gcd(e,p)=1$, let $S_L\subset p\Z$ be finite, $S=S_L\cup\{e\}$, and let $c\in\C^S$ with $|c_e|=\eta$. For $\tau\in\C^{S_L}$ and $\varepsilon\ge0$ put $Q(X)=\sum_{k\in S_L}\tau_ke^{-i(k/p)X}$. Then \[ N_S(c)\ \ge\ \frac{\re\sum_{k\in S_L}c_k\tau_k+\varepsilon\eta}{\norm Q_1+\frac1{2\pi}\int_0^{2\pi}R_p\bigl(|Q(X)|,\varepsilon\bigr)\,dX} \] whenever the numerator is positive. \end{lemma} \begin{proof} Let $\chi=\arg c_e$ (any value if $\eta=0$) and $T(\theta)=Q(p\theta)+\varepsilon e^{-i(e\theta+\chi)}$. This is $T_{\tau'}$ for the vector $\tau'\in\C^S$ that extends $\tau$ by $\tau'_e=\varepsilon e^{-i\chi}$, and $\ip c{\tau'}=\sum_{S_L}c_k\tau_k+\varepsilon\eta$, so by Lemma~\ref{lem:dual} it suffices to bound $\norm T_1$. Split $[0,2\pi)$ into the arcs $[2\pi j/p,2\pi(j+1)/p)$ and substitute $\theta=(X+2\pi j)/p$ in the $j$-th arc. Since $k/p\in\Z$ for $k\in S_L$, $Q(p\theta)=Q(X)$, while $e\theta=eX/p+2\pi ej/p$. Hence \[ \norm T_1=\frac1{2\pi}\int_0^{2\pi}\frac1p\sum_{j=0}^{p-1}\bigl|Q(X)+\varepsilon e^{-i(eX/p+\chi)}\omega^j\bigr|\,dX,\qquad \omega=e^{-2\pi ie/p}. \] As $\gcd(e,p)=1$, $\omega^j$ runs over all $p$-th roots of unity. For fixed $X$ write $Q(X)=We^{i\alpha}$. The inner average equals $\frac1p\sum_j|W+\varepsilon e^{i(\phi+2\pi j/p)}|$ for some $\phi$, which is at most $W+R_p(W,\varepsilon)$. No cancellation between the lattice part and the off-lattice term is needed. \end{proof} For $\varepsilon>0$ we have $R_p(W,\varepsilon)=\varepsilon r_p(W/\varepsilon)$ with \[ r_p(u):=R_p(u,1)=\max_\varphi\frac1p\sum_{j=0}^{p-1}|u+e^{i(\varphi+2\pi j/p)}|-u, \] and $R_p(W,0)=0$. Let also $r_\infty(u)=\frac1{2\pi}\int_0^{2\pi}|u+e^{it}|\,dt-u=\frac2\pi(u+1)E\bigl(\frac{4u}{(u+1)^2}\bigr)-u$, where $E(m)=\int_0^{\pi/2}(1-m\sin^2x)^{1/2}dx$. The formula follows from $|u+e^{it}|^2=(u+1)^2-4u\sin^2(t/2)$. \begin{lemma}\label{lem:rp} Let $p\ge2$. \begin{enumerate}[label=\textup{(\roman*)}] \item $r_p$ is nonincreasing on $[0,\infty)$. \item $r_p(u)\le\sqrt{u^2+1}-u\le1/(2u)$. \item For $p_0\ge2$, $p\ge p_0$ and $u\ge0$, $r_p(u)\le r_\infty(u)+\operatorname{err}(p_0,u)$, where $\operatorname{err}(p_0,1)=2\pi/(3p_0^2)$ and for $u\ne1$ \[ \operatorname{err}(p_0,u)=\min\Bigl\{\frac{2\pi\min(u,1)}{3p_0^2},\ \frac{4\max(u,1)\,|\binom{1/2}{p_0}|\,q^{p_0}}{1-q^{p_0}}\Bigr\},\qquad q=u\ (u<1),\quad q=1/u\ (u>1). \] \end{enumerate} \end{lemma} \begin{proof} (i) Write $\psi_j=\varphi+2\pi j/p$. For fixed $\varphi$ the derivative in $u$ of $\frac1p\sum_j|u+e^{i\psi_j}|-u$ is $\frac1p\sum_j\re(u+e^{i\psi_j})/|u+e^{i\psi_j}|-1\le0$ wherever it exists, and the function is Lipschitz. A maximum of nonincreasing functions is nonincreasing. (ii) By the Cauchy--Schwarz inequality, $\frac1p\sum_j|u+e^{i\psi_j}|\le(\frac1p\sum_j|u+e^{i\psi_j}|^2)^{1/2}=(u^2+1)^{1/2}$, because $\sum_je^{i\psi_j}=0$ for $p\ge2$. (iii) Let $F(t)=|u+e^{it}|$. Its Fourier series converges absolutely, so by Poisson summation $\frac1p\sum_jF(\varphi+2\pi j/p)=\sum_{m\in\Z}\widehat F(mp)e^{imp\varphi}$ and $r_p(u)-r_\infty(u)\le\sum_{m\ne0}|\widehat F(mp)|$. We bound the right-hand side in two ways. \emph{First bound.} We have $F'(t)=-u\sin t/F(t)$. For $u\ne1$, the function $u\sin t/F(t)$ is unimodal on $[0,\pi]$; indeed its derivative has the sign of $\frac{B_u}2c^2+A_uc+\frac{B_u}2$ with $c=\cos t$, $A_u=u^2+1$, $B_u=2u$, which has exactly one root in $(-1,0)$. Its maximum, attained at $c=-u$ for $u<1$ and at $c=-1/u$ for $u>1$, equals $\min(u,1)$. So $F'$ has total variation $4\min(u,1)$ over a period. This also holds for $u=1$, where $F=2|\cos(t/2)|$. Two integrations by parts give $|\widehat F(n)|\le4\min(u,1)/(2\pi n^2)$, and $2\sum_{m\ge1}4\min(u,1)/(2\pi m^2p^2)=2\pi\min(u,1)/(3p^2)$. \emph{Second bound} ($u\ne1$). We have $F(t)=\max(u,1)\,|1+qe^{it}|$ (for $u<1$ use $|u+e^{it}|=|1+ue^{-it}|$ and the evenness of $F$). With $(1+w)^{1/2}=\sum_{j\ge0}\gamma_jw^j$, $\gamma_j=\binom{1/2}j$, we get $\widehat F(n)=\max(u,1)\sum_{j\ge0}\gamma_{j+n}\gamma_jq^{2j+n}$ for $n\ge0$. Since $|\gamma_j|$ decreases for $j\ge1$ and $\sum_{j\ge1}|\gamma_j|x^j=1-\sqrt{1-x}$, this gives $|\widehat F(n)|\le\max(u,1)|\gamma_n|q^n(2-\sqrt{1-q^2}) \le2\max(u,1)|\gamma_n|q^n$, and then $\sum_{m\ne0}|\widehat F(mp)|\le4\max(u,1)|\gamma_p|q^p/(1-q^p)$. Both bounds are nonincreasing in $p$. \end{proof} \section{The two finite-dimensional problems}\label{sec:finite} Throughout, $\beta_0=\BETAFR=\BETA$. \subsection{Equal gaps} \begin{lemma}[symmetries]\label{lem:sym} For $\xi=(\sigma,x,y,\varphi_{-2},\varphi_{-1},\varphi_2)\in K$ let $\rho(\xi)=(\sigma,y,x,\varphi_2+2\varphi_{-1},\varphi_{-1},\varphi_{-2}-2\varphi_{-1})$ and $\bar\xi=(\sigma,x,y,-\varphi_{-2},-\varphi_{-1},-\varphi_2)$. Then $N_5(C_0(\rho\xi))=N_5(C_1(\xi))$, $N_5(C_1(\rho\xi))=N_5(C_0(\xi))$ and $N_5(C_i(\bar\xi))=N_5(C_i(\xi))$; hence $\Lambda(\rho\xi)=\Lambda(\bar\xi)=\Lambda(\xi)$. \end{lemma} \begin{proof} Reversing $C_1(\xi)$ (Lemma~\ref{lem:props}(g)) gives $(ye^{i\varphi_2},1,1,e^{-\sigma}e^{i\varphi_{-1}},xe^{-2\sigma}e^{i\varphi_{-2}})$, and the rotation $c_k\mapsto c_ke^{-ik\varphi_{-1}}$ turns it into $C_0(\rho\xi)$. Reversing and rotating $C_0(\xi)$ in the same way gives $C_1(\rho\xi)$. Finally $C_i(\bar\xi)=\overline{C_i(\xi)}$. \end{proof} Given $\xi\in K$, reduce $\varphi_{-1}$ modulo $2\pi$ into $[-\pi,\pi]$, apply $\xi\mapsto\bar\xi$ if $\varphi_{-1}<0$, then apply $\rho$ if $x>y$ (it does not change $\varphi_{-1}$), and finally reduce $\varphi_{-2}$ and $\varphi_2$ modulo $2\pi$. By Lemma~\ref{lem:sym}, $\Lambda$ is unchanged, and the result lies in \[ D_K=[0,\infty)\times\{(x,y)\in[0,1]^2:x\le y\}\times[-\pi,\pi]\times[0,\pi]\times[-\pi,\pi]. \] \begin{lemma}[tail]\label{lem:Ktail} For all $x\in[0,1]$ and $\psi\in\R$ there is $\tau\in\C^5$ with $T_\tau\ne0$ and $\re\ip{(xe^{i\psi},1,1,0,0)}\tau\ge1.61\,\norm{T_\tau}_1$; in particular $N_5\bigl((xe^{i\psi},1,1,0,0)\bigr)\ge1.61$. Consequently $\Lambda(\xi)\le1/N_5(C_0(\xi))<0.6285$ whenever $\sigma\ge4$. \end{lemma} \begin{proof} The first statement is a certified computation (Section~\ref{sec:cert}); the second follows by Lemma~\ref{lem:dual}. For the third, the rotation $c_k\mapsto c_ke^{ik\varphi_{-1}}$ maps $(xe^{i\varphi_{-2}},e^{i\varphi_{-1}},1,0,0)$ to $(xe^{i\psi},1,1,0,0)$ with $\psi=\varphi_{-2}-2\varphi_{-1}$, and Lemma~\ref{lem:props}(e) gives $N_5(C_0(\xi))\ge1.61-e^{-\sigma}-e^{-2\sigma}\ge 1.61-e^{-4}-e^{-8}>1.5913$. \end{proof} \begin{theorem}[equal gaps]\label{thm:K} $\Lambda(\xi)\le\beta_0$ for every $\xi\in K$. \end{theorem} \begin{proof} By Lemma~\ref{lem:sym} it suffices to treat $\xi\in D_K$, and by Lemma~\ref{lem:Ktail} only $\sigma\le4$. The certificate of Section~\ref{sec:cert} splits $[0,4]\times[0,1]^2\times[-P,P]\times[0,P]\times[-P,P]$ into finitely many boxes. The boxes that lie entirely in $\{x>y\}$ do not meet $D_K$; every other box carries an index $i$ and a vector $\tau$ with $T_\tau\ne0$ and $\re\ip{C_i(\xi)}\tau\ge\norm{T_\tau}_1/\beta_0$ on the box, and Lemma~\ref{lem:dual} gives $N_5(C_i(\xi))\ge1/\beta_0$. \end{proof} \subsection{Mixed gaps} For $p\ge2$, $s_1\ge0$ and $z\in Z(s_1,p/(p-1))$ let \[ \Phi_p(z,s_1)=\inf\Bigl\{\frac{U_p(\tau,\varepsilon)}{\re\ip z\tau+\varepsilon e^{-s_1}}:\ \tau\in\C^5,\ \varepsilon\ge0,\ \re\ip z\tau+\varepsilon e^{-s_1}>0\Bigr\}. \] \begin{theorem}[mixed gaps]\label{thm:J} $\Phi_p(z,s_1)\le\beta_0$ for all $p\ge2$, $s_1\ge0$ and $z\in Z(s_1,p/(p-1))$. \end{theorem} \begin{proof} \emph{Tail $s_1\ge4$.} Take $\varepsilon=0$, so that $U_p(\tau,0)=\norm{T_\tau}_1$. Let $\tau$ be given by Lemma~\ref{lem:Ktail} for $xe^{i\psi}=z_2$, and let $\tau'_k=\tau_{-k}$, so that $T_{\tau'}(\theta)=T_\tau(-\theta)$ and $\norm{T_{\tau'}}_1=\norm{T_\tau}_1$. Then $\ip z{\tau'}=\ip{(z_2,1,1,0,0)}\tau+z_{-1}\tau_1+z_{-2}\tau_2$. Since $|\tau_k|\le\norm{T_\tau}_1$ ($\tau_k$ is a Fourier coefficient of $T_\tau$) and $|z_{-1}|\le e^{-s_1}$, $|z_{-2}|\le e^{-2s_1}$ (as $p/(p-1)\ge1$), we get $\re\ip z{\tau'}\ge(1.61-e^{-4}-e^{-8})\norm{T_{\tau'}}_1>1.5913\,\norm{T_{\tau'}}_1$. Hence $\Phi_p(z,s_1)<0.6285$. \emph{Compact part $s_1\le4$.} The periods are split into the groups $\{2\}$, $\{3\}$, $\{4\}$, $\{5,6\}$, $\{7,\dots,10\}$, $\{11,\dots,16\}$, $\{17,\dots,24\}$ and $\{p\ge25\}$. For a group $G$ let $\kappa_G=\inf_{p\in G}p/(p-1)$ (so $\kappa_G=1$ for the last group) and let $\bar r_G$ be the step function of Section~\ref{sec:cert}, which satisfies $\bar r_G\ge r_p$ for all $p\in G$. Let $U_G(\tau,\varepsilon)$ be $U_p(\tau,\varepsilon)$ with $R_p(W,\varepsilon)$ replaced by $\varepsilon\bar r_G(W/\varepsilon)$ (and by $0$ if $\varepsilon=0$). For $p\in G$ we have $Z(s_1,p/(p-1))\subset Z(s_1,\kappa_G)$ and $U_p\le U_G$. The certificate covers, for each group, all pairs $(s_1,z)$ with $s_1\in[0,4]$ and $z\in Z(s_1,\kappa_G)$, in the coordinates \begin{gather*} z=\bigl(m_{-2}e^{-2\kappa_Gs_1}e^{i\vartheta_{-2}},\ m_{-1}e^{-\kappa_Gs_1}e^{i\vartheta_{-1}},\ 1,\ 1,\ m_2e^{i\vartheta_2}\bigr),\\ (s_1,m_{-2},m_{-1},m_2,\vartheta_{-2},\vartheta_{-1},\vartheta_2)\in[0,4]\times[0,1]^3\times[-\pi,\pi]\times[0,\pi]\times[-\pi,\pi], \end{gather*} by boxes carrying $(\tau,\varepsilon)$ with $T_\tau\ne0$ and $\re\ip z\tau+\varepsilon e^{-s_1}\ge U_G(\tau,\varepsilon)/\beta_0$ on the box (so the left side is positive). The range $\vartheta_{-1}\in[-\pi,0]$ follows by conjugation: $(\bar\tau,\varepsilon)$ works for $\bar z$, because $|T_{\bar\tau}(X)|=|T_\tau(-X)|$. \end{proof} \subsection{Proof of Theorem~\ref{thm:main}} Let $f$ be transcendental. By Section~\ref{sec:events} it has events at every $j\ge1$, with $r_0=e^{t_j}\to\infty$. At an equal-gap event, Proposition~\ref{prop:equal} and Theorem~\ref{thm:K} give a radius $r\in\{r_0,r_1\}$ with $\mu(r)/M(r)\le\beta_0$. At a mixed event, Proposition~\ref{prop:mixed} and Theorem~\ref{thm:J} give the same. Hence $\mu(r)/M(r)\le\beta_0$ for arbitrarily large $r$. \qed \section{The certificates}\label{sec:cert} \subsection{Branch and bound} Each certificate is a list of boxes, products of closed intervals with rational endpoints, obtained by repeated bisection of a fixed initial grid. A box for Theorem~\ref{thm:K} carries an index $i\in\{0,1\}$ and a vector $\tau\in\mathbb Q(i)^5$; a box for Theorem~\ref{thm:J} carries $\tau\in\mathbb Q(i)^5$ and a rational $\varepsilon\ge0$; a box for Lemma~\ref{lem:Ktail} carries $\tau$. The phases range over $[-P,P]$ (or $[0,P]$) with the rational $P=3.14159265358980>\pi$. We obtained the vectors from a floating-point optimization of the dual problem at the centre of the box and rounded them to multiples of $2^{-24}$. A box that cannot be certified is bisected, and the vectors of its parent are tried first. A box of the equal-gap domain lying entirely in $\{x>y\}$ is recorded as covered by the symmetry $\rho$. A leaf is valid if a lower bound of the linear form $\re\ip{C_i(\xi)}\tau$ (respectively $\re\ip z\tau+\varepsilon e^{-s_1}$) over the box is positive and at least $1/\beta_0$ times an upper bound of $\norm{T_\tau}_1$ (respectively of $U_G(\tau,\varepsilon)$). A leaf for Lemma~\ref{lem:Ktail} is valid if a lower bound of $\re\ip{(xe^{i\psi},1,1,0,0)}\tau$ over the box is positive and at least $1.61$ times an upper bound of $\norm{T_\tau}_1$. Upper bounds for $\norm{T_\tau}_1$ come from an adaptive subdivision of $[0,2\pi]$. For the correction term in $U_G$, on each piece the integrand $\varepsilon\bar r_G(|T_\tau(X)|/\varepsilon)$ is at most $\varepsilon\bar r_G(w/\varepsilon)$ for a rigorous lower bound $w$ of $|T_\tau|$ on the piece, because $\bar r_G$ is nonincreasing. The first program takes $w$ from a ball enclosure of $T_\tau$. The second uses a chord lower bound of $|T_\tau|^2$ along the piece, integrates the step function exactly, and uses the last table value beyond the grid. \subsection{The tables} For $p=2,\dots,24$ there is a table of integers $n_j$ with $n_j2^{-40}\ge r_p(u_j)$ on a grid $0=u_04/7$ \cite[pp.~156--158]{CH64} ($q_n=0$ for $n\equiv0,1$ and $q_n=1$ for $n\equiv2,3\pmod4$), but the moduli differ from those of their series. The leaf of the equal-gap certificate with the smallest margin (Table~\ref{tab:certs}) lies near this configuration. It carries $i=1$ and, to three decimals, \[ \tau=(-0.146-0.003i,\ -0.006-0.026i,\ 0.830-0.020i,\ 0.741+0.023i,\ -0.027+0.020i), \] for which $\norm{T_\tau}_1\approx1$ and $\re\ip{C_1(\xi)}\tau\ge\norm{T_\tau}_1/\beta_0$ on the leaf. For the mixed problems, the largest value we found is about $\JSTAR$. \end{remark} \begin{remark}[wider windows]\label{rem:seven} Wider windows may lower the limit of the five-coefficient formulation. With seven coefficients centred at the middle of five consecutive vertices with equal gaps, tested at the four jump radii between them, the best configuration we know for the equal-gap problem has value about $0.6059$ (numerically). It is again real, with the Clunie--Hayman sign pattern. The exact value of that problem is not known: it is at least this number and, by Lemma~\ref{lem:props}(d) applied at the two middle radii, at most the value of the six-parameter problem. A proof along these lines would also need a different treatment of mixed events: numerically, the five-coefficient single-radius bound of Proposition~\ref{prop:mixed} tends to $(\sqrt5-1)/2\approx0.618$ as $s_1\to\infty$ (the off-lattice coefficient becomes negligible), so that argument alone cannot go below $0.618$. \end{remark} \section*{Use of AI tools} Generative AI tools were used to find and prove the results, to write the programs and the Lean formalization, and to draft the text. The author reviewed the arguments, the programs and the formal proofs, and takes responsibility for the content. \begin{thebibliography}{99} \bibitem{EP513} T. F. Bloom, \emph{Erd\H{o}s Problem \#513}, \url{https://www.erdosproblems.com/513}, accessed 20 October 2026. \bibitem{CH64} J. Clunie and W. K. Hayman, The maximum term of a power series, \emph{J. Analyse Math.} \textbf{12} (1964), 143--186. \bibitem{Er61} P. Erd\H{o}s, Some unsolved problems, \emph{Magyar Tud. Akad. Mat. Kutat\'o Int. K\"ozl.} \textbf{6} (1961), 221--254. \bibitem{GS63} A. Gray and S. M. Shah, A note on entire functions and a conjecture of Erd\"os, \emph{Bull. Amer. Math. Soc.} \textbf{69} (1963), 573--577. \bibitem{Hay74} W. K. Hayman, The local growth of power series: a survey of the Wiman--Valiron method, \emph{Canad. Math. Bull.} \textbf{17} (1974), 317--358. \bibitem{HL19} W. K. Hayman and E. F. Lingham, \emph{Research Problems in Function Theory}, Fiftieth anniversary edition, Problem Books in Mathematics, Springer, Cham, 2019; arXiv:1809.07200. \bibitem{HT26} Y. He and Q. Tang, Generalizing the Clunie--Hayman construction in an Erd\H{o}s maximum-term problem, arXiv:2602.12217, 2026. \bibitem{Joh17} F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, \emph{IEEE Trans. Comput.} \textbf{66} (2017), 1281--1292. \bibitem{Mathlib} The mathlib Community, The Lean mathematical library, in: \emph{Proc. 9th ACM SIGPLAN Int. Conf. Certified Programs and Proofs (CPP 2020)}, 2020, 367--381. \bibitem{Lean4} L. de Moura and S. Ullrich, The Lean 4 theorem prover and programming language, in: \emph{Automated Deduction -- CADE 28}, Lecture Notes in Comput. Sci. \textbf{12699}, Springer, 2021, 625--635. \bibitem{So26} N. Sothanaphan, A certified computation for an improved He--Tang parameter choice in Erd\H{o}s' maximum-term problem, note dated 27 February 2026, available through \url{https://www.erdosproblems.com/forum/thread/513}. \bibitem{TaoRepo} T. Tao et al., \emph{Optimization constants in mathematics}, constant 51a, repository \url{https://github.com/teorth/optimizationproblems}. \end{thebibliography} \end{document}