\chapter{Exact rational bounds for the nonconvex example}\label{app:certificate} In this appendix, we prove the two finite statements used in the proof of \cref{prop:th-example}: an enclosure of the parameter $t$ of \eqref{eq:th-example-constants} between two rational numbers (\cref{lem:app-t}), and the sign pattern of the polynomial $P_t$ of \eqref{eq:th-example-U2} for all $t$ in this enclosure (\cref{lem:app-sign}). We also record the two rational comparisons used at the end of that proof. Every step is a finite computation with rational numbers. The sign pattern follows from the standard Bernstein-coefficient enclosure \cite{Farouki2012,Lorentz1986}, which we recall in \cref{lem:app-bernstein}. The computations can be repeated with the ancillary file \texttt{verify-thermal-example.py}, which is distributed with this paper. It uses only the Python standard library, with integer and rational arithmetic (\texttt{fractions.Fraction}); it uses no floating-point arithmetic, sampling, or root finding. It checks the inequalities of \cref{lem:app-t}, computes all $126$ Bernstein coefficients in \cref{tab:app-bernstein} at both endpoints of the enclosure, checks the four sign-and-margin bounds of \cref{lem:app-sign}, checks the endpoint identities \eqref{eq:app-endpoints} as a control of the conversion, and checks the comparisons in \eqref{eq:app-margins}. Throughout, we write \[ t_-=\frac{4539}{100000},\qquad t_+=\frac{4541}{100000},\qquad\mathcal L(u)=\frac{45+85u}{24+72u}. \] \begin{lemma}[Enclosure of the parameter]\label{lem:app-t} Let $t=(45-24\log6)/(72\log6-85)$. Then $t_-85/72$. The function $\ell\mapsto(45-24\ell)/(72\ell-85)$ is the inverse of $\mathcal L$ on $\ell>85/72$, and its derivative is $-1200/(72\ell-85)^2<0$. Its value at $\ell=\log6$ is $t$, and its values at $\mathcal L(t_\pm)$ are $t_\pm$. Then \eqref{eq:app-log6} implies that $t_-2/25&&\text{for }r\in[0,357/1000]\cup[892/1000,1],\\ P_t(r)&<-1/25&&\text{for }r\in[359/1000,890/1000],\\ P_t'(r)&<-86&&\text{for }r\in[357/1000,359/1000],\\ P_t'(r)&>73&&\text{for }r\in[890/1000,892/1000]. \end{aligned} \] In particular, $P_t$ has exactly two zeros $a_00$ on $[0,a_0)\cup(b_0,1]$ and $P_t<0$ on $(a_0,b_0)$. \end{lemma} \begin{table}[htb] \centering \small \begin{tabular}{@{}lllcc@{}} \toprule Polynomial & Interval & Sign $\varsigma$ & Lower bounds at $t_-$, $t_+$ & Coefficients\\ \midrule $P_t$ & $[0,\,357/1000]$ & $+1$ & $93$, $94$ & $26$\\ $P_t$ & $[359/1000,\,890/1000]$ & $-1$ & $56$, $57$ & $26$\\ $P_t$ & $[892/1000,\,1]$ & $+1$ & $92$, $92$ & $26$\\ $P_t'$ & $[357/1000,\,359/1000]$ & $-1$ & $87410$, $87410$ & $24$\\ $P_t'$ & $[890/1000,\,892/1000]$ & $+1$ & $74538$, $74529$ & $24$\\ \bottomrule \end{tabular} \caption{Bounds on the Bernstein coefficients of $P_t$ and $P_t'$. In each row, every Bernstein coefficient $B_k$ of the polynomial on the interval, computed by \eqref{eq:app-bernstein} with $n=12$ for $P_t$ and $n=11$ for $P_t'$, satisfies $1000\,\varsigma B_k>L$, where $L$ is the first listed integer at $t=t_-$ and the second at $t=t_+$. The last column counts the coefficients over both endpoints; the total is $126$.} \label{tab:app-bernstein} \end{table} \begin{proof} An exact computation of the coefficients \eqref{eq:app-bernstein} at $t=t_-$ and $t=t_+$ gives the bounds in \cref{tab:app-bernstein}. The lower bounds $93,94,92$ exceed $80=1000\cdot2/25$, the bounds $56,57$ exceed $40=1000\cdot1/25$, and the bounds $87410$ and $74538,74529$ exceed $86000$ and $73000$. The rest of the proof uses only these weaker margins, which are the ones checked by the ancillary script; the integer bounds in the table are stronger and record the size of the margins. By the remark before \cref{lem:app-sign}, the same strict bounds on the Bernstein coefficients hold for all $t\in[t_-,t_+]$, and \cref{lem:app-bernstein} transfers them to the polynomials on the intervals. This proves the four displayed bounds. The five intervals cover $[0,1]$. On the three long intervals, $P_t$ has no zero. On $[357/1000,359/1000]$, $P_t$ is strictly decreasing, positive at the left endpoint, and negative at the right endpoint, so it has exactly one zero $a_0$ there, which is simple because $P_t'(a_0)<0$. Similarly, $P_t$ has exactly one zero $b_0\in(890/1000,892/1000)$, and $P_t'(b_0)>0$. The sign pattern follows. \end{proof} Finally, the proof of \cref{prop:th-example} uses the following rational comparisons: \begin{equation}\label{eq:app-margins} \frac{12+132t_+}{24+72t_+}=\frac{149951}{227246}<\frac23,\qquad\frac14<11t_-\Big(\frac{25}{27}\Big)^8,\qquad11t_+\Big(\frac{25}{27}\Big)^8<\frac3{10}. \end{equation} The first is checked by cross-multiplication, since $3\cdot149951=449853<454492=2\cdot227246$. For the other two, $11t_\pm(25/27)^8$ are explicit rational numbers, approximately $0.26975$ and $0.26987$, and the two inequalities are again checked by cross-multiplication.