\documentclass[11pt]{amsart} \pdfsuppressptexinfo=-1 \usepackage[margin=1.1in]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools} \usepackage{booktabs,enumitem,array} \usepackage[hidelinks]{hyperref} \hypersetup{pdftitle={Upper and lower bounds for the Schur-Siegel-Smyth trace problem}, 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{\C}{\mathbb{C}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\Q}{\mathbb{Q}} \newcommand{\tr}{\operatorname{tr}} \newcommand{\Res}{\operatorname{Res}} \newcommand{\supp}{\operatorname{supp}} \newcommand{\file}[1]{\texttt{#1}} \newcommand{\LUNUM}{81383685642728309283453} \newcommand{\LUDEN}{45035996273705240000000} \newcommand{\LUDEC}{1.8070809} \newcommand{\LULONG}{1.8070808325} \newcommand{\EPSU}{1/100000} \newcommand{\NPOLYU}{21} \newcommand{\NNODESU}{2{,}325} \newcommand{\NHATSU}{2{,}323} \newcommand{\NHATSUNZ}{1{,}510} \newcommand{\BETAU}{0.88363} \newcommand{\MMINU}{4.05\cdot10^{-6}} \newcommand{\NWINUA}{102} \newcommand{\NCHKUA}{18{,}677} \newcommand{\NCHKUB}{14{,}091} \newcommand{\NWINUB}{99} \newcommand{\MINLBUB}{3.1\cdot10^{-9}} \newcommand{\MINMIDUB}{3.89\cdot10^{-6}} \newcommand{\DIPU}{0.676625} \newcommand{\SUPPULO}{0.054} \newcommand{\SUPPUHI}{5.227} \newcommand{\LPVALU}{1.8070689} \newcommand{\LUNUMA}{1660935360692531465493} \newcommand{\LUDENA}{919101964769480000000} \newcommand{\LUDECA}{1.8071286} \newcommand{\NCHKUAA}{15{,}509} \newcommand{\NCHKUBA}{10{,}051} \newcommand{\XLO}{1/100} \newcommand{\XHI}{30} \newcommand{\LLNUM}{901111277} \newcommand{\LLDEN}{500000000} \newcommand{\LLDEC}{1.8022225} \newcommand{\LLLONG}{1.802222554} \newcommand{\NNODESL}{10{,}386} \newcommand{\NHATSL}{10{,}384} \newcommand{\NHATSLNZ}{10{,}237} \newcommand{\SVAL}{0.535256} \newcommand{\IVAL}{-2.75\cdot10^{-6}} \newcommand{\NCHKLA}{142{,}137} \newcommand{\NCHKLB}{142{,}343} \newcommand{\SLACKL}{1.02\cdot10^{-6}} \newcommand{\BINDL}{1.413232} \newcommand{\LAMSOCP}{1.802223655} \newcommand{\LTAIL}{0.4867} \newcommand{\RTAIL}{18.03} \newcommand{\RALPHA}{2.7229} \newcommand{\SUPPLLO}{0.0488} \newcommand{\SUPPLHI}{5.3520} \newcommand{\NWINL}{68} \newcommand{\NNATL}{153} \newcommand{\NNATU}{78} \newcommand{\SUPPDOI}{10.5281/zenodo.23563971} \newcommand{\REPOURL}{https://github.com/tejasvi/schur-siegel-smyth-trace-bounds-v4} \begin{document} \title[The Schur--Siegel--Smyth trace problem]{Upper and lower bounds for the Schur--Siegel--Smyth trace problem} \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]{11R06, 11R80; 31A15, 65G20, 90C05} \keywords{Totally positive algebraic integers, Schur--Siegel--Smyth trace problem, distributions of conjugates, logarithmic potential, auxiliary functions, ball arithmetic, computer-assisted proof} \begin{abstract} The Schur--Siegel--Smyth trace constant $C_{86}$ is the lower limit of $\tr(\alpha)/\deg(\alpha)$ over totally positive algebraic integers $\alpha$. We prove $\LLDEC0$ only finitely many totally positive algebraic integers satisfy $\tr(\alpha)<(c-\varepsilon)\deg(\alpha)$. We follow the numbering of the repository \cite{Tao}; the constant is denoted $\lambda_{\mathrm{SSS}}$ in \cite{OSS25}. For a prime $p\ge3$, the number $4\cos^2(\pi/p)$ is a totally positive algebraic integer of degree $(p-1)/2$ and trace $p-2$, so $C_{86}\le2$. Schur proved $C_{86}\ge\sqrt e=1.6487\ldots$ \cite{Sch18}, and Siegel proved $C_{86}>1.7336$ \cite{Sie45}. Smyth \cite{Smy84} introduced the auxiliary function method: if $x-\sum_Q\lambda_Q\log|Q(x)|\ge\lambda$ for all $x>0$, where $Q$ runs through a finite set of nonzero integer polynomials and $\lambda_Q\ge0$, then $C_{86}\ge\lambda$. He proved $C_{86}\ge1.7719$, and better choices of polynomials have since raised the bound to $1.793145\ldots$ \cite{WWW21}. Serre showed that the method cannot prove a bound larger than a constant $c_S=1.898302\ldots$ \cite{AP08}, \cite[Section~1]{SO24}. Orloski, Talebizadeh Sardari and Smith \cite{OSS25} added the condition that limits of the uniform measures on the conjugates have nonnegative logarithmic energy, which comes from the integrality of discriminants, and proved $C_{86}\ge1.80203$. In their proof \cite[Corollary~1.9]{OSS25}, the reference measure is given by floating-point data and several of the required inequalities are checked numerically. Giraudin \cite{Gir26} has recently reported $C_{86}\ge1.80220$ in the same framework, with eighteen polynomials and a step reference density. His inequality is verified by two programs that bisect $[0,14]$ into about $1.7$ million intervals, with bounds evaluated in floating-point arithmetic; the energy of the reference measure is computed in double and extended precision and decreased by $10^{-12}$, and a further $10^{-9}$ is subtracted from every bound. Upper bounds come from measures. Smith \cite[Theorem~1.5]{Smi24} proved that a probability measure $\mu$ on an interval of length greater than $4$ is the weak limit of the uniform measures on the roots of a sequence of distinct monic irreducible integer polynomials whose roots lie in the interval, as soon as $\int\log|Q|\,d\mu\ge0$ for every nonzero $Q\in\Z[x]$. Applied to a measure of Serre, this gives $C_{86}<1.89831$ \cite[Theorem~1.1 and Example~5.16]{Smi24}, so $C_{86}<2$. Orloski and Talebizadeh Sardari \cite{SO24} constructed better measures, checking the infinitely many conditions with a finite set of polynomials, and proved $C_{86}\le1.8216$. Pinho \cite{Pin26} enlarged their set of four polynomials to eleven and obtained $C_{86}<1.8117251$, with a certificate checked in exact rational and ball arithmetic. Qu and Yang \cite{QY26} showed that $C_{86}$ is also the limit of $\tr(\alpha)/\deg(\alpha)$ along a sequence of totally positive units. \begin{theorem}\label{thm:main} \[ \frac{\LLNUM}{\LLDEN}\le C_{86}\le\frac{\LUNUM}{\LUDEN}. \] In particular $\LLDEC\LLDEC\deg(\alpha)$, and infinitely many satisfy $\tr(\alpha)<\LUDEC\deg(\alpha)$. \end{theorem} The two fractions are $\LLLONG$ and $\LULONG\ldots$, so Theorem~\ref{thm:main} places $C_{86}$ in an interval of length less than $0.005$. The upper bound is new. For it we use a measure $\mu$ on $[0,6]$ that is a mixture of a measure with a piecewise linear density and a small multiple of the equilibrium measure of $[0,6]$, together with $\NPOLYU$ integer polynomials: the seven of \cite{OSS25}, the four quintics added by Pinho, three polynomials of degree~$10$ found by McKee and Smyth \cite{MS04} and used by Giraudin, and seven further polynomials of degrees $2$ to~$8$ found by column generation (Table~\ref{tab:poly}). Lemma~\ref{lem:U1} reduces the conditions of Smith's theorem to one inequality on $[0,6]$ and one integral inequality for each of the $\NPOLYU$ polynomials, which two programs check in ball arithmetic \cite{Arb,FLINT}, in which every computed quantity is an interval guaranteed to contain the exact value. The lower bound uses the irreducible factors of the polynomials of \cite{Gir26} and is larger than the bound $1.80220$ reported there. It is larger than $1.8022211895\ldots$, the numerical infimum of the dual function of the certificate of \cite{Gir26} \cite[Remark~2]{Gir26}; for the constraints used there, \cite[Remark~3]{Gir26} reports a discrete ceiling of $1.8022246$. What differs is the verification. The certificate consists of integer polynomials and rational numbers: the multipliers, the nodes and weights of a piecewise linear reference density, and $\lambda$. Lemma~\ref{lem:T1} reduces the lower bound to an inequality $F\ge0$ on $(0,\infty)$, and two programs check this inequality in ball arithmetic. Since the reference density is piecewise linear, its potential and energy have closed forms (Lemma~\ref{lem:closed}), and no numerical integration is needed. We also give a self-contained proof of Lemma~\ref{lem:T1} that does not assume that limit measures have compact support. Both bounds of Theorem~\ref{thm:main} are also proved in Lean~4 \cite{Lean4} with Mathlib \cite{Mathlib} (Section~\ref{sec:lean}): the lower bound without hypotheses, and the upper bound under one named hypothesis, Smith's Theorem~1.5 for intervals.\ The finite checks on the certificates are evaluated by compiled code in one version of these proofs and by Lean's kernel in another. Section~\ref{sec:lower} proves the lower bound and Section~\ref{sec:upper} the upper bound. Section~\ref{sec:cert} describes the two certificates, how they were found, and the output of the programs. Section~\ref{sec:lean} describes the Lean proofs, and Section~\ref{sec:remarks} discusses how far the two methods can go. \section{The lower bound}\label{sec:lower} \subsection{Limit measures and the energy}\label{sec:lemma} For a finite Borel measure $\rho$ on $\R$ we write $U_\rho(x)=\int\log|x-y|\,d\rho(y)$ for its logarithmic potential, $I(\rho,\eta)=\int U_\eta\,d\rho$ and $I(\rho)=I(\rho,\rho)$ for its logarithmic energy, whenever the integrals exist. A signed measure $\sigma$ has finite energy if $\iint|\log|x-y||\,d|\sigma|(x)\,d|\sigma|(y)<\infty$. With this sign convention, $I(\sigma)\le0$ for every signed measure $\sigma$ of total mass zero with finite energy (the argument is recalled in the proof of Lemma~\ref{lem:T1}), so $I$ is concave on probability measures of finite energy. Let $\alpha$ be a totally positive algebraic integer of degree $d$ with conjugates $\alpha_1,\dots,\alpha_d$, and let $\mu_\alpha=\frac1d\sum_j\delta_{\alpha_j}$ be the uniform probability measure on them. Then $\int x\,d\mu_\alpha=\tr(\alpha)/d$. Recall that $C_{86}$ is the largest $c$ such that for each $\varepsilon>0$ only finitely many totally positive algebraic integers satisfy $\tr(\alpha)<(c-\varepsilon)\deg(\alpha)$. \begin{lemma}\label{lem:T1} Let $\mathcal A$ be a finite set of nonzero polynomials in $\Z[x]$, let $\lambda\in\R$, $\lambda_Q\ge0$ for $Q\in\mathcal A$ and $\lambda_0\ge0$, and let $\nu$ be a compactly supported probability measure on $\R$ with a bounded density. Suppose that \begin{equation}\label{eq:F} F(x)=x-\lambda-\sum_{Q\in\mathcal A}\lambda_Q\log|Q(x)|-\lambda_0\bigl(2U_\nu(x)-I(\nu)\bigr)\ge0 \end{equation} for every $x>0$ that is not a root of a polynomial in $\mathcal A$. Then $C_{86}\ge\lambda$. \end{lemma} This is the dual form of the method of Orloski, Talebizadeh Sardari and Smith \cite[Section~1.4]{OSS25} with a single reference measure $\nu$, and for $\lambda_0=0$ it is Smyth's auxiliary function method \cite{Smy84}. The key input is that limits $\mu$ of the measures $\mu_\alpha$ satisfy $I(\mu)\ge0$. For compactly supported $\mu$ this also follows from a result of Smith \cite[Proposition~5.7]{Smi24}: if $\int\log|Q|\,d\mu\ge0$ for every nonzero $Q\in\Z[x]$, then $I(\mu)\ge0$ in our sign convention. As said in the introduction, our proof of Lemma~\ref{lem:T1} is self-contained and does not assume that $\mu$ has compact support. It uses the discriminants of the minimal polynomials directly and truncates the logarithm near the diagonal. \begin{proof} Let $L=C_{86}$, so $L\le2$. By the maximality of $L$, for every $\delta>0$ infinitely many totally positive algebraic integers satisfy $\tr(\alpha)<(L+\delta)\deg(\alpha)$, and only finitely many satisfy $\tr(\alpha)<(L-\delta)\deg(\alpha)$. Hence we can choose distinct totally positive algebraic integers $\alpha_n$ of degrees $d_n$ with $\tr(\alpha_n)/d_n\to L$. For each $D$, only finitely many totally positive algebraic integers have degree at most $D$ and trace at most $3D$, because their conjugates lie in $(0,3D]$ and the coefficients of their minimal polynomials are bounded integers. Hence $d_n\to\infty$. We discard the finitely many $\alpha_n$ that are roots of polynomials in $\mathcal A$, and finitely many more so that $\int x\,d\mu_n\le L+1$ for all $n$, where $\mu_n=\mu_{\alpha_n}$. \emph{Arithmetic input.} Let $P_n$ be the minimal polynomial of $\alpha_n$. For $Q\in\mathcal A$, the product $\prod_jQ(\alpha_{n,j})=\pm\Res(P_n,Q)$ is a nonzero integer, so $\int\log|Q|\,d\mu_n\ge0$. The discriminant $\prod_{i\ne j}(\alpha_{n,i}-\alpha_{n,j})$ of $P_n$ is a nonzero integer up to sign, so \begin{equation}\label{eq:discr} \frac1{d_n^2}\sum_{i\ne j}\log|\alpha_{n,i}-\alpha_{n,j}|\ge0. \end{equation} \emph{Limit and integrability.} By Markov's inequality $\mu_n((R,\infty))\le(L+1)/R$, so the sequence is tight. After passing to a subsequence, $\mu_n$ converges weakly to a probability measure $\mu$ on $[0,\infty)$, and $\int x\,d\mu\le L$ because $x$ is continuous and nonnegative. For $R\ge e$ and $x\ge R$ we have $\log(1+x)\le2(\log R/R)\,x$, so \[ \sup_n\int_{x>R}\log(1+x)\,d\mu_n\le\frac{2(L+1)\log R}{R}\to0\qquad(R\to\infty). \] Therefore $\int\varphi\,d\mu_n\to\int\varphi\,d\mu$ for every continuous $\varphi$ on $[0,\infty)$ with $|\varphi(x)|\le c(1+\log(1+x))$, and the same holds for continuous $\psi$ on $[0,\infty)^2$ with $|\psi(x,y)|\le c(1+\log(1+x)+\log(1+y))$ and the product measures $\mu_n\times\mu_n\to\mu\times\mu$. \emph{The constraints pass to the limit.} Fix $Q\in\mathcal A$ and $M>0$, and let $\varphi_M=\max(\log|Q|,-M)$. This function is continuous, $\varphi_M\ge\log|Q|$, and $|\varphi_M(x)|\le M+c_Q(1+\log(1+x))$. Hence $\int\varphi_M\,d\mu=\lim_n\int\varphi_M\,d\mu_n\ge\limsup_n\int\log|Q|\,d\mu_n\ge0$. As $M\to\infty$, $\varphi_M$ decreases to $\log|Q|$, and $\varphi_0=\max(\log|Q|,0)\le c_Q(1+\log(1+x))$ is $\mu$-integrable, so monotone convergence gives $\int\log|Q|\,d\mu\ge0$. Similarly, let $\psi_M(x,y)=\max(\log|x-y|,-M)$, which is continuous with $|\psi_M|\le M+\log(1+x)+\log(1+y)$. The diagonal terms of $\sum_{i,j}\psi_M(\alpha_{n,i},\alpha_{n,j})$ equal $-M$, so \eqref{eq:discr} gives \[ 0\le\frac1{d_n^2}\sum_{i\ne j}\log|\alpha_{n,i}-\alpha_{n,j}|\le\iint\psi_M\,d\mu_n\,d\mu_n+\frac M{d_n}. \] Letting $n\to\infty$ gives $\iint\psi_M\,d\mu\,d\mu\ge0$. As $M\to\infty$, $\psi_M$ decreases to $\log|x-y|$, and $\psi_0\le\log(1+x)+\log(1+y)$ is $\mu\times\mu$-integrable, so monotone convergence gives $I(\mu)\ge0$. In particular $I(\mu)$ is finite, $\iint|\log|x-y||\,d\mu\,d\mu<\infty$, and $\mu$ has no atoms. \emph{Energy inequality.} We claim that $I(\mu)\le2I(\mu,\nu)-I(\nu)$. Let $\sigma=\mu-\nu$, a signed measure of total mass zero. For $r>0$ we have $-\log r=\frac12\int_0^\infty(e^{-sr^2}-e^{-s})\,\frac{ds}s$, and the integrand has constant sign, so $\int_0^\infty|e^{-sr^2}-e^{-s}|\,\frac{ds}s=2|\log r|$. The double integral $\iint|\log|x-y||\,d|\sigma|\,d|\sigma|$ is finite: for $\mu\times\mu$ by the previous step, and for $\nu\times\nu$ and $\mu\times\nu$ because $\nu$ has compact support and a bounded density $w$, so that $\int\log^-|x-y|\,d\nu(y)\le2\|w\|_\infty$, while $\log^+|x-y|\le\log(1+|x|)+\log(1+|y|)$ and $\int\log(1+x)\,d\mu\le\log(1+L)$ by Jensen's inequality. In particular the diagonal $\{x=y\}$ is $|\sigma|\times|\sigma|$-null. Fubini's theorem therefore gives \[ -I(\sigma)=\frac12\int_0^\infty\Bigl(\iint e^{-s|x-y|^2}\,d\sigma(x)\,d\sigma(y)\Bigr)\frac{ds}s\ge0, \] since $\sigma(\R)=0$ and the Gaussian kernel is positive definite. By bilinearity, $I(\sigma)=I(\mu)-2I(\mu,\nu)+I(\nu)$, which proves the claim. \emph{Conclusion.} Since $\mu$ has no atoms, the finitely many roots of the polynomials in $\mathcal A$ and the point $0$ are $\mu$-null, so $F\ge0$ holds $\mu$-almost everywhere. Each term of $F$ is $\mu$-integrable, so integrating gives \[ L\ge\int x\,d\mu\ge\lambda+\sum_Q\lambda_Q\int\log|Q|\,d\mu+\lambda_0\bigl(2I(\mu,\nu)-I(\nu)\bigr)\ge\lambda+\lambda_0I(\mu)\ge\lambda.\qedhere \] \end{proof} \begin{remark}\label{rem:T1} Only $Q(\alpha)\ne0$ is used, so the polynomials in $\mathcal A$ need not be irreducible or have real roots. The measure $\nu$ can be any compactly supported probability measure with a bounded density: in particular, $I(\nu)\ge0$ is not needed, and $\nu$ need not satisfy the constraints $\int\log|Q|\,d\nu\ge0$. \end{remark} \subsection{Piecewise linear reference densities}\label{sec:hat} We take for $\nu$ a continuous piecewise linear density. Let $e_00$, let $g=\sum_k\rho_kh_k$ and $\nu=g/s$. With $\lambda_0=s$, the function \eqref{eq:F} becomes \begin{equation}\label{eq:Fg} F(x)=x-\lambda-\sum_Q\lambda_Q\log|Q(x)|-2U_g(x)+\frac{I(g)}{s}. \end{equation} The second derivative of $g$ in the sense of distributions is $g''=\sum_iC_i\delta_{e_i}$, with rational coefficients $C_i$ determined by the slopes of $g$, and $\sum_iC_i=\sum_iC_ie_i=0$ because $g$ vanishes outside $[e_0,e_N]$. \begin{lemma}\label{lem:closed} Let $H(t)=\frac{t^2}2\log|t|-\frac34t^2$ and $J(t)=\frac{t^4}{24}\log|t|-\frac{25}{288}t^4$, with $H(0)=J(0)=0$. Then \[ U_g(x)=\sum_iC_iH(e_i-x),\qquad I(g)=\sum_{i,j}C_iC_jJ(e_i-e_j). \] \end{lemma} \begin{proof} A direct computation gives $H''(t)=\log|t|$ and $J''=H$, so $J''''(t)=\log|t|$. The functions $H$ and $J$ are even, $H$ is $C^1$ and $J$ is $C^3$. Since $g$ is Lipschitz with compact support, two integrations by parts give $\int\log|x-y|\,g(y)\,dy=\int H''(y-x)g(y)\,dy=\int H(y-x)\,dg''(y)$, with vanishing boundary terms. This is the first formula. For the second we integrate by parts twice in each variable. \end{proof} Lemma~\ref{lem:closed} reduces $F$ and $I(g)$ to finitely many elementary terms with exact rational arguments $e_i-x$ at rational points $x$. The same formulas give $F''(x)=\sum_Q\lambda_Q\frac{Q'(x)^2-Q(x)Q''(x)}{Q(x)^2}-2\sum_iC_i\log|e_i-x|$ away from the nodes and the roots. \subsection{Verification}\label{sec:verify} We verify \eqref{eq:Fg} for a certificate consisting of the polynomials $Q\in\mathcal A$, rational multipliers $\lambda_Q$, rational nodes and weights, a rational $\lambda$, and two rational cut points $00$ and $x_{\mathrm{lo}}$ is small. On the right, assume that the support of $g$ lies in $[0,x_{\mathrm{hi}}]$ and $x_{\mathrm{hi}}\ge1$. For $x\ge x_{\mathrm{hi}}$ we have $|Q(x)|\le\|Q\|_1x^{\deg Q}$, where $\|Q\|_1$ is the sum of the absolute values of the coefficients, and $U_g(x)\le s\log x$, because $00$ we choose a window: an interval $W$ with rational endpoints on which a single ball evaluation of $F$, with each $\log|Q|$ replaced by the logarithm of an upper bound for $|Q|$ on $W$, gives a positive lower bound. This bound is valid on all of $W$, roots included. If the ball contains a node $e_i$, the term $H(e_i-x)$ is enclosed by $|H(t)|\le r^2(\frac12\log\frac1r+\frac34)$ for $|t|\le r<0.6$, which holds because $t^2\log(1/t)$ is increasing on $[0,e^{-1/2}]$. The nodes and the window ends cut the rest of $[x_{\mathrm{lo}},x_{\mathrm{hi}}]$ into intervals on which $F$ is smooth, except for logarithmic singularities of $F''$ at nodes that are endpoints of the interval. On each such interval we use the bound below, and we bisect the interval when it does not give a positive lower bound. The correctness of the result does not depend on how the roots were located: on an interval that contains a root of some $Q$ with $\lambda_Q>0$, the ball enclosure of $F''$ described below is not finite, so the interval is never accepted. \begin{lemma}\label{lem:chord} Let $f$ be continuously differentiable on $[p,q]$ with $f'$ absolutely continuous, and let $\Delta=q-p$. Then for $p\le x\le q$, \[ f(x)\ge\min\bigl(f(p),f(q)\bigr)-\frac\Delta4\int_p^q\bigl(f''(t)\bigr)_+\,dt. \] \end{lemma} \begin{proof} Let $\ell$ be the linear function with $\ell(p)=f(p)$ and $\ell(q)=f(q)$. Then $f(x)=\ell(x)-\int_p^qG(x,t)f''(t)\,dt$, where $G(x,t)=(t-p)(q-x)/\Delta$ for $t\le x$ and $G(x,t)=(x-p)(q-t)/\Delta$ for $t\ge x$. Since $0\le G\le\Delta/4$, the integral is at most $\frac\Delta4\int_p^q(f'')_+$, and $\ell(x)\ge\min(f(p),f(q))$. \end{proof} We evaluate $F(p)$ and $F(q)$ with $Q(p)$, $Q(q)$ and the differences $e_i-p$, $e_i-q$ computed exactly. Write $F''=R+S$, where $S$ is the sum of the terms $-2C_i\log|e_i-x|$ over the nodes $e_i\in\{p,q\}$. Since $(R+S)_+\le R_++\sum(-2C_i\log|e_i-x|)_+$, we have \[ \int_p^q(F'')_+\le\Delta\,\bar R_++\sum_{e_i\in\{p,q\}}\int_0^\Delta\bigl(-2C_i\log t\bigr)_+\,dt, \] where $\bar R$ is an upper bound for $R$ on $[p,q]$ obtained by a single ball evaluation on a ball that contains $[p,q]$. With $m=\min(\Delta,1)$, the last integral equals $2C_im(1-\log m)$ if $C_i>0$, equals $2|C_i|(\Delta\log\Delta-\Delta+1)$ if $C_i<0$ and $\Delta>1$, and vanishes otherwise. Alternatively, a term with $C_i<0$ can be bounded on $[p,q]$ by $2|C_i|\log\Delta$ and added to $\bar R$. \subsubsection*{Implementation} Two programs carry out this procedure. They were written separately, share no code apart from the ball arithmetic library \cite{FLINT}, and rebuild every quantity from the certificate. Since they implement the same method, their agreement does not test the method itself, which is proved above. It tests the implementations, in particular of the potential, the energy and the root location, which the two programs compute in different ways. The first computes the coefficients $C_i$ and evaluates $U_g$ and $I(g)$ by Lemma~\ref{lem:closed}, with $200$-bit balls. It brackets the real roots by floating-point approximations confirmed by exact sign changes, and confirms their number with a Sturm sequence. On an interval it encloses each $Q$ and its first two derivatives through their exact Taylor coefficients at the rational midpoint, because Horner's scheme on balls overestimates polynomials of degree~$10$ with large coefficients too much. When a node with $C_i>0$ is an endpoint, it accepts the interval only if $\Delta\le1$, where the integral above equals $2C_i\Delta(1-\log\Delta)$; for the terms with $C_i<0$ at an endpoint it uses the alternative bound $2|C_i|\log\Delta$. The second uses the coefficients $C_i$ only for $F''$, and recomputes them from the values of $g$ at the nodes. It writes $g(y)=a+by$ on each piece $[e_k,e_{k+1}]$ and computes $U_g$ from the primitive \[ \int(a+by)\log|x-y|\,dy=(a+bx)\,G_1(y-x)+b\,G_2(y-x), \] where $G_1(t)=t\log|t|-t$ and $G_2(t)=\frac{t^2}2\log|t|-\frac{t^2}4$; on balls around $0$ it uses $|G_1(t)|\le r\log\frac1r+r$ for $|t|\le r<0.36$ and $|G_2(t)|\le\frac{r^2}2\log\frac1r+\frac{r^2}4$ for $|t|\le r<0.6$. It computes $I(g)$ from closed-form primitives of $t^n\log|t|$ over pairs of pieces, and it works with $256$-bit balls. It encloses all complex roots of each $Q$, and requires every root whose enclosure meets the real line to be simple and to have a bracket with an exact sign change, the brackets being disjoint; it then narrows each bracket by exact bisection to width at most $2^{-110}$. On an interval $I$ it bounds the part of $F''$ that comes from $Q$ through the roots, by $-(\log|Q|)''=\sum_\rho\operatorname{Re}(x-\rho)^{-2}\le\sum_\rho\operatorname{dist}(\rho,I)^{-2}$, and it bounds $|Q|$ on $I$ by $|a_Q|\prod_\rho\max_{x\in I}|x-\rho|$, so that no polynomial is evaluated on a wide ball. \section{The upper bound}\label{sec:upper} \subsection{A criterion}\label{sec:crit} As in Section~\ref{sec:lower}, $U_\mu(z)=\int\log|z-y|\,d\mu(y)$ denotes the logarithmic potential of a finite Borel measure $\mu$ on $\R$, now for $z\in\C$. \begin{lemma}\label{lem:U1} Let $X>4$, and let $\mu$ be a Borel probability measure on $[0,X]$ whose potential $U_\mu$ is finite and continuous on $\C$. Let $\mathcal B$ be a finite set of primitive polynomials in $\Z[x]$ of positive degree, irreducible over $\Q$, no two of which differ only in sign, and let $b_Q>0$ for $Q\in\mathcal B$, with $\beta=\sum_Qb_Q\deg Q<1$. Suppose that \begin{enumerate}[label=\textup{(U\arabic*)},ref=\textup{(U\arabic*)}] \item\label{u1} $h(x)=U_\mu(x)-\sum_{Q\in\mathcal B}b_Q\log|Q(x)|\ge0$ for every $x\in\supp\mu$ that is not a root of a polynomial in $\mathcal B$, and \item\label{u2} $\int\log|Q|\,d\mu\ge0$ for every $Q\in\mathcal B$. \end{enumerate} Then $\int\log|P|\,d\mu\ge0$ for every nonzero $P\in\Z[x]$, and $C_{86}\le\int x\,d\mu$. \end{lemma} \begin{proof} Since $U_\mu$ is finite, $\mu$ has no atoms, and for a polynomial $P$ with leading coefficient $a$ and roots $\rho_1,\dots,\rho_d$ the function $\log|P(x)|=\log|a|+\sum_k\log|x-\rho_k|$ is $\mu$-integrable, with \begin{equation}\label{eq:logP} \int\log|P|\,d\mu=\log|a|+\sum_kU_\mu(\rho_k). \end{equation} \emph{Step 1: $h(z)\ge0$ for every $z\in\C$ that is not a root of a polynomial in $\mathcal B$.} Define $h(z)$ for $z\in\C$ by the formula in \ref{u1}, with the value $+\infty$ at the roots. Since $U_\mu$ is continuous and $b_Q>0$, the function $h\colon\C\to(-\infty,\infty]$ is lower semicontinuous, and it is superharmonic on $D=\C\setminus\supp\mu$, where $U_\mu$ is harmonic. As $|z|\to\infty$ we have $h(z)=(1-\beta)\log|z|-\sum_Qb_Q\log|a_Q|+o(1)\to\infty$, where $a_Q$ is the leading coefficient of $Q$. Choose $R>X$ such that $h>0$ on $\{|z|\ge R\}$, and let $D_R=D\cap\{|z|0$. The maximum principle for the subharmonic function $-h$ on the bounded domain $D_R$ \cite[Theorem~2.3.1]{Ran95} therefore gives $h\ge0$ on $D_R$, and hence on $\C$. \emph{Step 2: $\int\log|P|\,d\mu\ge0$ for every nonzero $P\in\Z[x]$.} The integral is additive in $P$, and a nonzero integer $c$ contributes $\log|c|\ge0$, so it suffices to consider $P$ primitive, irreducible over $\Q$ and of degree $d\ge1$. If $P=\pm Q$ with $Q\in\mathcal B$, we use \ref{u2}. Otherwise $P$ has no common root with any $Q\in\mathcal B$, because both are primitive and irreducible and $P\ne\pm Q$. Let $a$ be the leading coefficient of $P$ and $\rho_1,\dots,\rho_d$ its roots. For $Q\in\mathcal B$, the resultant $\Res(P,Q)=a^{\deg Q}\prod_kQ(\rho_k)$ is a nonzero integer, so $\sum_k\log|Q(\rho_k)|\ge-\deg Q\log|a|$. By Step~1, $U_\mu(\rho_k)\ge\sum_Qb_Q\log|Q(\rho_k)|$, and summing over $k$ in \eqref{eq:logP} gives \[ \int\log|P|\,d\mu\ge\log|a|-\sum_{Q\in\mathcal B}b_Q\deg Q\,\log|a|=(1-\beta)\log|a|\ge0. \] \emph{Step 3: the sequence.} The interval $[0,X]$ has logarithmic capacity $X/4>1$ \cite[Example~I.3.5]{ST97}. By Step~2 and \cite[Theorem~1.5]{Smi24}, there are distinct monic irreducible polynomials $R_1,R_2,\dots\in\Z[x]$ with all roots in $[0,X]$ whose root distributions $\frac1{\deg R_k}\sum_{R_k(\rho)=0}\delta_\rho$ converge weakly to $\mu$. At most one $R_k$ equals $x$, and the roots of every other $R_k$ lie in $(0,X]$, so $R_k$ is the minimal polynomial of a totally positive algebraic integer $\alpha_k$, and these $\alpha_k$ are distinct. Since $x$ is continuous on $[0,X]$, $\tr(\alpha_k)/\deg(\alpha_k)=\int x\,d\mu_{R_k}\to\int x\,d\mu$, where $\mu_{R_k}$ is the root distribution of $R_k$. So for every $\varepsilon>0$ infinitely many totally positive algebraic integers satisfy $\tr(\alpha)<(\int x\,d\mu+\varepsilon)\deg(\alpha)$, and $C_{86}\le\int x\,d\mu$. \end{proof} Smith states the condition of his theorem as $\int\log|Q|\,d\mu\ge0$ for every nonzero $Q\in\Z[x]$, where $\log0=-\infty$, so that the condition fails when $\mu$ charges a root of $Q$; here $\mu$ has no atoms. Step~2 is the argument of \cite[Proposition~5.15]{Smi24}, whose hypothesis is that $h\ge0$ on all of $\C$ (Smith writes the potential with the opposite sign). Step~1 reduces this hypothesis to the support of $\mu$, so no estimates outside $[0,X]$ are needed; it uses $\beta<1$, which makes $h$ tend to $\infty$ at infinity. Bounds of this kind were first obtained in \cite[Example~5.16]{Smi24} and \cite{SO24}, and Pinho's certificate checks the same conditions \cite{Pin26}. \subsection{The measure}\label{sec:upmeasure} We take $X=6$ and \[ \mu=(1-\varepsilon)\mu_0+\varepsilon\omega,\qquad\varepsilon=\EPSU, \] where $\omega$ is the equilibrium measure of $[0,6]$ and $\mu_0=G(x)\,dx$ has a continuous piecewise linear density as in Section~\ref{sec:hat}: for rational nodes $0=e_00$. We apply Lemma~\ref{lem:U1} with the weights $(1-\varepsilon)b_Q$. The conditions become \begin{gather} \beta=(1-\varepsilon)\sum_Qb_Q\deg Q<1,\label{eq:S}\\ (1-\varepsilon)\Bigl(U_{\mu_0}(x)-\sum_Qb_Q\log|Q(x)|\Bigr)+\varepsilon\log\frac32\ge0\qquad(0\le x\le6),\label{eq:P}\\ \int\log|Q|\,d\mu=(1-\varepsilon)\sum_{Q(\rho)=0}U_{\mu_0}(\rho)+\varepsilon\sum_{Q(\rho)=0}U_\omega(\rho)\ge0\qquad(Q\in\mathcal B),\label{eq:M} \end{gather} and the bound is $\int x\,d\mu=(1-\varepsilon)\sum_kw_k\frac{e_{k-1}+e_k+e_{k+1}}3+3\varepsilon$. Since $U_\omega\ge\log\frac32$, condition \eqref{eq:M} follows from \[ (1-\varepsilon)\sum_{Q(\rho)=0}U_{\mu_0}(\rho)+\varepsilon\deg Q\log\frac32\ge0. \] So the admixture of $\omega$ adds $\varepsilon\log\frac32$ to the left side of \eqref{eq:P} and $\varepsilon\deg Q\log\frac32$ to the integrals in \eqref{eq:M}, and raises the bound by $\varepsilon(3-\int x\,d\mu_0)$. It absorbs the small negative values that the measure $\mu_0$ found by linear programming has after rounding (Section~\ref{sec:certup}). \subsection{Verification}\label{sec:upverify} Two programs check \eqref{eq:S}, \eqref{eq:P} and \eqref{eq:M}, and compute $\int x\,d\mu$ exactly. Both check in exact arithmetic that the weights are nonnegative and sum to $1$, that the nodes increase and lie in $[0,6]$, that $b_Q\ge0$, and that the polynomials are distinct, primitive and irreducible (by factorization in FLINT \cite{FLINT}) with only real roots; and both check \eqref{eq:S} exactly. All other quantities are enclosed in balls. The first program brackets the real roots of each $Q$ by rational intervals of width at most $2^{-110}$, confirmed by exact sign changes and counted with a Sturm sequence, and evaluates \eqref{eq:M} with $U_{\mu_0}$ on these brackets. For \eqref{eq:P} it proceeds as in Section~\ref{sec:verify}: around each root of a polynomial $Q\in\mathcal B$ it chooses a window on which one ball evaluation gives a positive lower bound, with $\log|Q|$ replaced by the logarithm of an upper bound for $|Q|$, and on the intervals between the nodes and the window ends it applies Lemma~\ref{lem:chord} to the left side of \eqref{eq:P}, bisecting where the bound is not positive. The upper bound for the second derivative on an interval $[p,q]$ adds the polynomial part $\sum_Qb_Q(Q'^2-QQ'')/Q^2$, evaluated with exact Taylor coefficients at the midpoint, to bounds for the terms $C_i\log|e_i-x|$ of $U_{\mu_0}''$: an integral bound for a node at an endpoint, the supremum for the other nodes near the interval, and a Taylor bound at the midpoint for distant nodes. The second program was written independently of the first, from Lemma~\ref{lem:U1} and the certificate format only, and shares no code with it apart from FLINT. It computes $U_{\mu_0}$ and $U_{\mu_0}'$ on each piece from the antiderivatives of $(a+b\,t)\log|t|$, encloses the roots with FLINT, and proves \eqref{eq:P} on each interval between consecutive breakpoints (the nodes, the rational roots, and windows of width at most $2^{-110}$ around the irrational roots) by a second-order Taylor bound at the midpoint, in which the regular part of the second derivative is bounded on the interval and the logarithmic singularity of a node at an endpoint is integrated; it bisects on failure. On each window it bounds the left side of \eqref{eq:P} from below by one ball evaluation, with $\log|x-\theta|$ for the root $\theta$ replaced by the logarithm of the width of the window. The two programs compute the potential, the root enclosures and the positivity bounds in different ways. \section{The certificates}\label{sec:cert} \begin{table} \caption{The polynomials of the two certificates with their weights, rounded to ten decimal places: the multipliers $\lambda_Q$ of the lower-bound certificate of Theorem~\ref{thm:main}, which are multiples of $2^{-48}$, and the weights $b_Q$ of the upper-bound certificate, which are multiples of $2^{-44}$. A dash means that $Q$ is not in the certificate. The three polynomials of degree~$10$ are from \cite{MS04}.}\label{tab:poly} \small \begin{tabular}{@{}>{\raggedright\arraybackslash}p{9.6cm}ll@{}} \toprule $Q$ & $\lambda_Q$ & $b_Q$\\ \midrule $x$ & $0.5034276784$& $0.0585032675$\\ $x-1$ & $0.4379921423$& $0.0491080836$\\ $x-2$ & $0.0423886965$& $0.0333960240$\\ $x^2-3x+1$ & $0.1563506267$& $0.0221368799$\\ $x^3-5x^2+6x-1$ & $0.0581548388$& $0.0226234061$\\ $x^4-7x^3+13x^2-7x+1$ & $0.0172965782$& $0.0114475294$\\ $x^4-7x^3+14x^2-8x+1$ & $0.0176047286$& $0.0125654815$\\ $x^5-9x^4+28x^3-35x^2+15x-1$ & $0.0023197148$& $0.0069739625$\\ $x^5-9x^4+26x^3-29x^2+11x-1$ & $0.0006606314$& $0.0088863765$\\ $x^5-9x^4+27x^3-32x^2+13x-1$ & $0.0003887454$& $0.0039245213$\\ $x^5-9x^4+27x^3-31x^2+12x-1$ & $0.0025850757$& $0.0080087411$\\ $x^3-6x^2+9x-1$ & $0.0000002526$& --\\ $x^{10}-18x^9+134x^8-537x^7+1265x^6-1798x^5+1526x^4-743x^3+194x^2-24x+1$ & $0.0003929732$& $0.0057035573$\\ $x^{10}-18x^9+135x^8-549x^7+1320x^6-1920x^5+1662x^4-813x^3+206x^2-24x+1$ & $0.0003957846$& $0.0055010705$\\ $x^{10}-18x^9+134x^8-538x^7+1273x^6-1822x^5+1560x^4-766x^3+200x^2-24x+1$ & $0.0004134755$& $0.0054838227$\\ $x^8-15x^7+90x^6-277x^5+467x^4-428x^3+200x^2-42x+3$ & -- & $0.0054037982$\\ $x^8-15x^7+89x^6-269x^5+445x^4-402x^3+187x^2-40x+3$ & -- & $0.0047710666$\\ $x^7-13x^6+63x^5-143x^4+157x^3-78x^2+16x-1$ & -- & $0.0049122996$\\ $x^2-4x+2$ & -- & $0.0129514327$\\ $x^2-4x+1$ & -- & $0.0131738152$\\ $x^6-11x^5+43x^4-72x^3+51x^2-14x+1$ & -- & $0.0067676993$\\ $x^3-6x^2+9x-3$ & -- & $0.0066472539$\\ \bottomrule \end{tabular} \end{table} Table~\ref{tab:poly} lists the polynomials of both certificates. The first seven are those of \cite[Corollary~1.9]{OSS25}. Pinho \cite{Pin26} added the two quartics and the next five polynomials to the four polynomials $x$, $x-1$, $x^2-3x+1$ and $x^3-5x^2+6x-1$ of \cite{SO24}; the polynomial $x-2$ is not in his set. All polynomials in the table are monic, irreducible and have only real roots, which lie in $[0,6)$. \subsection{The lower bound}\label{sec:certlow} The reference density has $\NNODESL$ nodes in $[\SUPPLLO,\SUPPLHI]$, chosen by the collocation method described below. The weights $\rho_k$ are multiples of $2^{-48}$, and $\NHATSLNZ$ of the $\NHATSL$ weights are nonzero. The support of $g$ lies in the same interval, $s=\SVAL$ and $I(\nu)=I(g)/s^2=\IVAL$; by Remark~\ref{rem:T1}, Lemma~\ref{lem:T1} does not require $I(\nu)\ge0$. Finally, $\lambda=\LLNUM/\LLDEN$, $x_{\mathrm{lo}}=\XLO$ and $x_{\mathrm{hi}}=\XHI$. All fifteen polynomials have positive multipliers, but $x^3-6x^2+9x-1$, whose multiplier is about $2.5\cdot10^{-7}$, is not needed: both programs also accept the certificate without it (Table~\ref{tab:tests}). Three earlier certificates are also in the accompanying repository: one found by the same method with $6{,}173$ nodes and $\lambda=63361515736283/35157522000000>1.8022179$; one with the same fifteen polynomials, $2{,}683$ nodes and $\lambda=901101631/500000000=1.802203262$, and one with the twelve polynomials of Table~\ref{tab:poly} other than those of degree~$10$, $1{,}753$ nodes and $\lambda=185185220/102755699>1.8021892$. Both programs accept them. \subsubsection*{How it was found} The search plays no role in the proof. The reference density was found by collocation. Its support is a union of $66$ intervals, obtained from the support of an approximate solution with $4{,}149$ hats and split at every root of a polynomial with positive multiplier. Each interval carries at least $60$ nodes, spaced like Chebyshev--Lobatto points, and the hats centred at the ends of the intervals carry no mass. For fixed intervals, the conditions $F=0$ at the hat centres, $\int\log|Q|\,d\nu=0$ for the polynomials with positive multiplier, and $I(\nu)=0$ form a linear system in the weights and the multipliers together with one quadratic equation, which we solved directly. The ends of the intervals were moved $26$ times, until every weight was nonnegative and no end moved. With $\nu$ fixed, a linear program on about $3.3\cdot10^5$ points re-optimised the multipliers and gave $\lambda=1.802219008$, with minimum $-4.6\cdot10^{-8}$ of $F$ on these points. Starting from this solution, we then gave every interval on which $F$ fell below $-10^{-6}$ between nodes $30\%$ more nodes and solved again. After seven such rounds the measure had $10{,}254$ hats and $F$ fell at most $1.3\cdot10^{-6}$ below zero between nodes; $17$ negative weights were set to $0$, so that $I(\nu)$ is no longer $0$. The linear program for the multipliers then gave $\lambda=\LAMSOCP$, with minimum $-6.4\cdot10^{-8}$ of $F$ on its points. We rounded the weights to multiples of $2^{-48}$ and took $\lambda$ smaller by $1.1\cdot10^{-6}$. \subsubsection*{Output of the two programs} Both accept the certificate. Their enclosures of $s^2I(\nu)$ agree to all $38$\ digits that the first program prints. The lower bound on the left tail is $\LTAIL$. On the right tail, $a=\RALPHA$ and $x_{\mathrm{hi}}-a\log x_{\mathrm{hi}}-b=\RTAIL$. The polynomials with $\lambda_Q>0$ have $\NWINL$ real roots in $[x_{\mathrm{lo}},x_{\mathrm{hi}}]$, each of which gets a window. On the rest of $[x_{\mathrm{lo}},x_{\mathrm{hi}}]$ the first program made $\NCHKLA$ interval checks and the second $\NCHKLB$. The smallest lower bound for $F$ at an endpoint of a checked interval, in the second program, was $\SLACKL$, at $x=\BINDL$. Both programs still accept the certificate with $\lambda$ increased by $10^{-6}$ and reject it with $\lambda$ increased by $1.1\cdot10^{-6}$. \subsection{The upper bound}\label{sec:certup} The density $G$ has $\NNODESU$ nodes in $[0,6]$, mostly with spacings $1/250$, $1/1000$ and $1/2000$ ($2{,}287$ of the $2{,}324$ gaps) and $37$ other gaps, of up to $0.763$, and $\NHATSUNZ$ of the $\NHATSU$ weights $w_k$ are nonzero. The support of $\mu_0$ lies in $[\SUPPULO,\SUPPUHI]$ and has many gaps. The weights are rational with $\sum_kw_k=1$, and $\beta=\BETAU\ldots$. The bound is $\int x\,d\mu=(1-\varepsilon)\int x\,d\mu_0+3\varepsilon=\LUNUM/\LUDEN$, where $\int x\,d\mu_0=1.80706890\ldots$. \subsubsection*{How it was found} With $\varepsilon=0$, minimizing $\int x\,d\mu_0$ over the weights $w_k\ge0$ with $\sum_kw_k=1$ and $b_Q\ge0$, subject to $\int\log|Q|\,d\mu_0\ge0$ for $Q\in\mathcal B$, a bound on $\sum_Qb_Q\deg Q$, and \eqref{eq:P} at the points of a grid, is a linear program, because \eqref{eq:P} and \eqref{eq:M} are linear in $(w,b)$. We first solved it for step densities and chose the polynomials by column generation: starting from the fifteen polynomials of the lower certificate, three of them of degree~$10$ from \cite{MS04}, we added polynomials whose reduced cost is negative from a list of $23{,}631$ irreducible totally positive polynomials of degree at most~$8$, until none was left. The final set has $22$ polynomials, and $x^3-6x^2+9x-1$ gets weight~$0$. For this set we solved the linear program with $\NHATSU$ hats, using HiGHS \cite{HiGHS} through SciPy, with nodes on a grid over the support of the step solution. After each solution we added to the grid the local minima at which the left side of \eqref{eq:P} is negative on a fine grid of $600{,}000$ points. After the eighth solution the minimum on the fine grid was $-1.7\cdot10^{-7}$, and the value was $\LPVALU$. We then rounded the weights $w_k$ to multiples of $2^{-50}$ and rescaled them to sum $1$, rounded the $b_Q$ down to multiples of $2^{-44}$, and chose $\varepsilon=\EPSU$. The term $\varepsilon\log\frac32=\MMINU$ covers the small negative values that remain: between the points of the fine grid, $U_{\mu_0}-\sum_Qb_Q\log|Q|$ dips to about $-4.5\cdot10^{-7}$ near $x=0.74436$, and $14$ of the $21$ integrals $\int\log|Q|\,d\mu_0$ are slightly negative, down to about $-3.8\cdot10^{-13}$. With $\varepsilon=10^{-6}$ both programs reject the certificate at that point (Table~\ref{tab:tests}). An earlier solution of the same linear program, on a coarser fine grid, gave a certificate with the same polynomials and nodes, other weights, and $\varepsilon=1/20000$; it proves $C_{86}\le\LUNUMA/\LUDENA<\LUDECA$. Both programs accept it as well, with $\NCHKUAA$ and $\NCHKUBA$ interval checks. \subsubsection*{Output of the two programs} Both accept the certificate and compute the bound exactly. The integrals in \eqref{eq:M} agree to all printed digits in the two programs; the smallest is $\MMINU$, for $Q=x$, and they are close to $\varepsilon\deg Q\log\frac32$ for every $Q$. For \eqref{eq:P}, the first program made $\NCHKUA$ interval checks and used $\NWINUA$ root windows, and the second made $\NCHKUB$ interval checks and used $\NWINUB$ windows, treating the three rational roots $0$, $1$ and $2$ as breakpoints instead. In the second program the smallest proven lower bound on an interval was $\MINLBUB$, and the smallest value at the midpoint of a checked interval was $\MINMIDUB$, at $x=\DIPU$, in the normalization $F=U_{\mu_0}-\sum_Qb_Q\log|Q|+\varepsilon\log\frac32/(1-\varepsilon)$. No midpoint fell at the dip near $x=0.74436$, where $F$ is about $3.6\cdot10^{-6}$. \subsubsection*{Tests of the programs} Both reject modified certificates (Table~\ref{tab:tests}). For the lower bound we varied $\lambda$, dropped polynomials, changed a coefficient of one of them, scaled multipliers and weights, and moved $x_{\mathrm{hi}}$; for the upper bound we decreased $\varepsilon$, scaled the weights $b_Q$ and dropped polynomials. Each program either finds a point where an inequality fails or cannot confirm a root window and stops. \begin{table} \caption{Modified certificates. Accepted: the program verifies the certificate. Rejected: a checked bound fails, or the program cannot confirm a root window and stops. The polynomials are numbered as in Table~\ref{tab:poly}.}\label{tab:tests} \small \begin{tabular}{@{}lll@{}} \toprule modification & first program & second program\\ \midrule \multicolumn{3}{@{}l}{\emph{lower bound}}\\ $\lambda$ increased by $5\cdot10^{-7}$ or $10^{-6}$ & accepted & accepted\\ $\lambda$ increased by $1.1\cdot10^{-6}$, $1.2\cdot10^{-6}$, $2\cdot10^{-6}$, $3\cdot10^{-6}$ or $10^{-5}$ & rejected & rejected\\ polynomial $2$, $4$, $8$ or $11$ dropped & rejected & rejected\\ one or all three polynomials of degree~$10$ dropped & rejected & rejected\\ polynomial $12$ dropped & accepted & accepted\\ coefficient of $x^5$ in polynomial $14$ increased by $1$ & rejected & rejected\\ $\lambda_Q$ of polynomial $15$ multiplied by $1.1$; $\lambda_x$ multiplied by $1.01$ & rejected & rejected\\ all $\rho_k$ multiplied by $1.001$; largest $\rho_k$ multiplied by $1.1$ & rejected & rejected\\ $x_{\mathrm{hi}}=6$ & rejected & rejected\\ \midrule \multicolumn{3}{@{}l}{\emph{upper bound}}\\ $\varepsilon=10^{-6}$, $10^{-7}$ or $10^{-14}$ & rejected & rejected\\ all $b_Q$ multiplied by $1.001$ & rejected & rejected\\ polynomial $2$ or $22$ dropped & rejected & rejected\\ \bottomrule \end{tabular} \end{table} \section{The Lean proofs}\label{sec:lean} Both bounds are proved in Lean~4 \cite{Lean4} with Mathlib \cite{Mathlib}. For $B\in\R$ let $\mathrm{TPTraceLe}(B)$ be the set of algebraic integers $\alpha\in\C$ whose minimal polynomial has only positive real roots and that satisfy $\tr(\alpha)\le B\deg(\alpha)$, where $\tr(\alpha)$ is the sum of these roots. The theorem \file{Trace86.trace\_lower\_N} states that $\mathrm{TPTraceLe}(\lambda_L-\varepsilon)$ is finite for every $\varepsilon>0$, where $\lambda_L=$ $\LLNUM/\LLDEN$, and it has no hypotheses. Its corollary \file{Trace86.trace\_lower\_dec\_N} states that $\mathrm{TPTraceLe}(\LLDEC)$ is finite. The theorem \file{Trace86.trace\_upper\_N} states that if the proposition \file{SmithThm15} holds, then $\mathrm{TPTraceLe}(\lambda_U+\varepsilon)$ is infinite for every $\varepsilon>0$, where $\lambda_U$ is the upper bound of Theorem~\ref{thm:main}. By the definition of $C_{86}$, the two theorems give $\lambda_L\le C_{86}$ and, assuming \file{SmithThm15}, $C_{86}\le\lambda_U$. The proposition \file{SmithThm15} is Smith's Theorem~1.5 \cite{Smi24} for an interval $[a,b]$ with $b-a>4$. Since Mathlib sets $\log0=0$, its premise states Smith's condition for each nonzero $Q\in\Z[x]$ as three conditions: $\mu$ does not charge the roots of $Q$, the function $\log|Q|$ is $\mu$-integrable, and $\int\log|Q|\,d\mu\ge0$. Together they are equivalent to Smith's condition with $\log0=-\infty$. All other mathematical steps are proved in Lean, including the energy inequality behind Lemma~\ref{lem:T1}, the closed forms of Lemma~\ref{lem:closed}, and the minimum principle and the resultant argument of Lemma~\ref{lem:U1}. The finite checks on the certificates are carried out by checkers written in Lean, whose soundness is proved in Lean, and they are evaluated with \texttt{native\_decide}: $\NNATL$ checks for the lower bound (pieces of $[1/100,30)$, the energy bound and the two tails) and $\NNATU$ for the upper bound ($73$ pieces of $[1/100,6]$ and five further checks, among them the interval $[0,1/100]$, the degree condition \eqref{eq:S}, the integrals \eqref{eq:M} and the value of the bound). Each check enters the proof as an auxiliary axiom tied to its statement. So, besides the three standard axioms of Lean, the trusted base includes Lean's compiler and the native evaluation of these checks. With the checks as named axioms, the exported proofs of both bounds are accepted by three independent kernel implementations: Lean's own kernel through \texttt{leanchecker}, nanoda \cite{nanoda}, with the checks permitted as axioms by name, and lean4lean \cite{Car24}. The proof of the lower bound was also replayed module by module in Lean's kernel. In a second version of both proofs, every check is decided by Lean's kernel, with \texttt{decide +kernel}, so that Lean's compiler is not part of the trusted base. The theorems \file{Trace86.trace\_lower\_K} and \file{Trace86.trace\_upper\_K} have the same statements as \file{Trace86.trace\_lower\_N} and \file{Trace86.trace\_upper\_N} and use the same certificates and checkers. The checks are split into modules: for the lower bound, $14{,}957$ modules for the end points of the pieces of the interval, one end point each, $599$ modules for the pieces, $25$ pieces each, $9{,}142$ modules for the pieces of the energy bound, and four modules for the data and the tails; for the upper bound, $7{,}169$ modules for the pieces of the interval, two pieces each, $21$ modules for the integrals \eqref{eq:M}, one for each polynomial, and two further modules. These theorems, and their forms \file{trace\_lower\_dec\_K} and \file{trace\_upper\_dec\_K} for the decimal bounds $\LLDEC$ and $1.8070809$, depend only on the three standard axioms \file{propext}, \file{Classical.choice} and \file{Quot.sound}. Building this version takes about $310$ CPU-hours for the lower bound and about $55$ for the upper bound, with up to about $7.6$~GB of memory for one module. All $24{,}865$ modules of the proof of the lower bound and all $7{,}313$ modules of the proof of the upper bound were replayed in Lean's kernel. With the generated checks stated as axioms, the complete proofs of both bounds were also accepted by \texttt{leanchecker}, nanoda and lean4lean, and so were the five modules with the literal data and the global checks and $59$ of the generated modules. \section{How far the two methods can go}\label{sec:remarks} \subsection*{The lower bound} The last two steps of the proof of Lemma~\ref{lem:T1} show that $\int x\,d\mu\ge\lambda$ for every probability measure $\mu$ on $[0,\infty)$ with $I(\mu)\ge0$ and $\int\log|Q|\,d\mu\ge0$ for $Q\in\mathcal A$ (if $\int x\,d\mu=\infty$ there is nothing to prove). So no certificate with the polynomial set $\mathcal A$ gives more than the infimum $\lambda_{\mathcal A}$ of $\int x\,d\mu$ over these measures. For the fifteen polynomials of the lower certificate, the program of Section~\ref{sec:certlow} gives $\LAMSOCP$, and a second-order cone program over step densities, a restriction of the problem that defines $\lambda_{\mathcal A}$, gives $1.8022490$ with $1{,}300$ steps and $1.8022344$ with $2{,}600$ steps, so numerically $\lambda_{\mathcal A}\le1.8022344$; without the three polynomials of degree~$10$ it gives $1.8022435$ with $1{,}300$ steps. These fifteen polynomials include the irreducible factors of the eighteen polynomials of \cite{Gir26}. In floating-point arithmetic we found no irreducible totally positive polynomial of degree at most~$6$ and trace at most $3d$, of degree~$7$ and trace at most $2.5d$, or of degree~$8$ and trace at most $2d$ ($503{,}124$ polynomials in all) whose constraint $\int\log|Q|\,d\mu\ge0$ is violated by the optimal measure with a piecewise linear density ($2{,}866$ hats) for these fifteen polynomials. So lower bounds much above $1.8022$ seem to need other constraints. \subsection*{The upper bound} Here every nonzero $P\in\Z[x]$ must satisfy $\int\log|P|\,d\mu\ge0$. The next proposition shows that a measure with a bounded density and positive energy satisfies these conditions for all squarefree, in particular all irreducible, polynomials of large degree. Since $\int\log|P|\,d\mu$ is additive in $P$, only irreducible polynomials of small and intermediate degree matter. \begin{proposition}\label{prop:highdeg} Let $\mu$ be a compactly supported probability measure on $\R$ with a density bounded by $M$ and energy $I(\mu)=\eta>0$, and let $P\in\Z[x]$ be squarefree of degree $d\ge1$ with leading coefficient $a$. Then for every $S>0$, \[ \int\log|P|\,d\mu\ge\frac{\log|a|}d+d\Bigl(\frac\eta2-M\sqrt{\frac\pi S}-\frac{E_1(S)}2\Bigr)-\frac{\gamma+\log S+E_1(S)}4, \] where $E_1(S)=\int_S^\infty e^{-u}\,du/u$ and $\gamma$ is Euler's constant. \end{proposition} \begin{proof} For $S>0$ and $x,y\in\C$ let $k_S(x,y)=\int_0^S(e^{-s|x-y|^2}-e^{-s})\,ds/s$. Then $k_S(x,x)=\gamma+\log S+E_1(S)$, and $k_S(x,y)=-2\log|x-y|+E_1(S)-E_1(S|x-y|^2)$ for $x\ne y$, so that $-2\log^+|x-y|\le k_S(x,y)\le k_S(x,x)$. Since the Gaussian kernel is positive definite, $\iint k_S\,d\sigma\,d\sigma=\int_0^S\bigl(\iint e^{-s|x-y|^2}\,d\sigma\,d\sigma\bigr)\,ds/s\ge0$ for every finite signed measure $\sigma$ on $\C$ with compact support and $\sigma(\C)=0$, by Fubini's theorem and these bounds. Let $z_1,\dots,z_d$ be the roots of $P$, let $\nu=\frac1d\sum_i\delta_{z_i}$, and apply this to $\sigma=\nu-\mu$. Since $P$ is squarefree, its discriminant $a^{2d-2}\prod_{i