\documentclass[11pt]{amsart} \usepackage{amssymb,amsmath} \usepackage[margin=1in]{geometry} \usepackage{enumitem} \usepackage{booktabs} \usepackage{longtable} \usepackage{tikz} \usetikzlibrary{patterns} \usepackage{hyperref} \hypersetup{colorlinks=true, urlcolor=blue, linkcolor=black, citecolor=black, pdftitle={The Matsaev (1971) conjecture is false for every p other than 2, and other miscellanea}, pdfauthor={Dave Goel}} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{conjecture}[theorem]{Conjecture} \newtheorem{mainthm}{Theorem} \renewcommand{\themainthm}{\Alph{mainthm}} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newcommand{\sgn}{\operatorname{sgn}} \newcommand{\norm}[1]{\lVert #1\rVert} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\N}{\mathbb{N}} \newcommand{\Z}{\mathbb{Z}} \setlength{\emergencystretch}{3em} \title[The Matsaev conjecture is false for every $p\neq2$, and other miscellanea]{The Matsaev (1971) conjecture is false\\ for every $p\neq 2$, and other miscellanea} \author{Dave Goel} \email{david44hello@gmail.com} \subjclass[2020]{Primary 47A30; Secondary 47A60, 47B37, 46E30} \keywords{Matsaev's conjecture, von Neumann's inequality, $L^p$ spaces, contractions, shift operator, Bellman function} \date{First version: October 28, 2026. This version: November 16, 2026} \begin{document} \begin{abstract} The Matsaev (1971) conjecture asserts that for $11$, and $P_a(z)=1+az-Bz^2$. For every sufficiently small $a>0$ (how small depends only on $p$ and $B$) there is a real $2\times2$ matrix $T$ with $\norm{T}_p\le1$ on $\ell^p_2$ such that \[ \norm{P_a(T)}_p\;>\;\norm{P_a(S)}_p . \] For $p>2$ one can take $T=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2}$, for every sufficiently small $\varepsilon>0$ (depending on $p$, $B$ and $a$), where $D_*=\max_{0\le t\le1}(t-t^{p-1})/(1+t^p)>0$ depends only on $p$ (Lemma~\ref{lem:N1}). For $12$ in place of $p$; since the transpose of $R_\theta$ is $R_{-\theta}$, this is $T=e^{-\varepsilon D_*'/2}R_{-\pi/2-\varepsilon/2}$, where $D_*'$ is $D_*$ computed for $p'$: again a scaled rotation by slightly more than a right angle, now counterclockwise. \end{mainthm} \begin{mainthm}\label{thm:B} Let $P(z)=c_0+c_1z+c_2z^2$ with complex coefficients and $|c_0|=|c_2|$. Then $\norm{P(T)}_p\le\norm{P(S)}_p$ for every $1\le p\le\infty$ and every contraction $T$ on every $L^p$ space. \end{mainthm} \noindent\emph{Idea of the proof of Theorem~\ref{thm:A}.} Let $(x_n)_{n\in\Z}$ be a two-sided sequence with $x_{n+2}=-x_n$ for all $n$, and shift it by the same formula, $(Sx)_n=x_{n-1}$. Then $((Sx)_n,(Sx)_{n-1})=(x_{n-1},x_{n-2})=(x_{n-1},-x_n)=J(x_n,x_{n-1})$: on pairs of consecutive terms of such sequences, $S$ acts as $J$. Truncating long pieces of such sequences gives $\norm{P(J)}_p\le\norm{P(S)}_p$ for every $P$ (Lemma~\ref{lem:J}). For $P_0(z)=1-Bz^2$ we have $P_0(J)=I-BJ^2=(1+B)I$, so $1+B=\norm{P_0(J)}_p\le\norm{P_0(S)}_p\le1+B$ (Lemma~\ref{lem:J} and the triangle inequality); thus $\norm{P_0(S)}_p=1+B$, and $P_0(S)$ has many almost-maximising vectors: long truncations $x$ of every sequence with $x_{n+2}=-x_n$ give $\norm{P_0(S)x}_p/\norm{x}_p$ close to $1+B$. For $p\ne2$ and small $a>0$ we prove that $P_a(S)$ still has exactly the norm $C_a=\norm{P_a(J)}_p$ (Theorem~\ref{thm:SJ}). For $p>2$ the upper bound comes from a Bellman function, a function $G$ of two real variables with $|(P_a(S)x)_n|^p\le C_a^p|x_n|^p+G(x_{n-1},x_{n-2})-G(x_n,x_{n-1})$ for every real $x\in\ell^p(\N)$ (with $x_{-1}=x_{-2}=0$) and every $n\ge0$; in the sum over $n$ the $G$ terms cancel, so $\norm{P_a(S)}_p\le C_a$ (Lemma~\ref{lem:bellman}). For $12$ we compose $J$ with the rotation $R_{\varepsilon/2}$ and multiply by $e^{-\varepsilon D_*/2}$; this gives the matrix $T_\varepsilon=e^{-\varepsilon D_*/2}R_{\varepsilon/2}J=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2}$ of Theorem~\ref{thm:A}. The factor $e^{-\varepsilon D_*/2}$ compensates the largest possible rate of growth of the $\ell^p$ norm under the rotations $R_{\varepsilon/2}$ (at most $D_*/2$ per unit of $\varepsilon$), so that $T_\varepsilon$ is still a contraction (Figure~\ref{fig:contraction}). To see this, put $\phi(y)=\sgn y\,|y|^{p-1}$ (so that $\frac{d}{dy}|y|^p=p\,\phi(y)$) and, for real $\xi\ne0$, \[ Q(\xi)=\frac{\phi(\xi_1)\xi_2-\phi(\xi_2)\xi_1}{\norm{\xi}_p^p},\qquad\text{so that}\qquad Q(1,t)=\frac{t-t^{p-1}}{1+t^p}\quad(0\le t\le1). \] Since $\frac{d}{d\varepsilon}R_{\varepsilon/2}\xi\big|_{\varepsilon=0}=\frac12J\xi=\frac12(\xi_2,-\xi_1)$, the chain rule in each coordinate gives \[ \frac{d}{d\varepsilon}\norm{R_{\varepsilon/2}\xi}_p^p\Big|_{\varepsilon=0}=\frac p2\bigl(\phi(\xi_1)\xi_2-\phi(\xi_2)\xi_1\bigr) =\frac p2\,Q(\xi)\norm{\xi}_p^p, \] hence, dividing by $p\norm{\xi}_p^{p-1}$, $\frac{d}{d\varepsilon}\norm{R_{\varepsilon/2}\xi}_p\big|_{\varepsilon=0}=\frac12\,Q(\xi)\norm{\xi}_p$. Lemma~\ref{lem:N1} shows that the maximum of $Q$ over all $\xi\ne0$ is attained at $\xi=(1,t_*)$ for some $t_*\in(0,1)$; so $\max Q=D_*$, and the derivative above is at most $\frac{D_*}2\norm{\xi}_p$, with equality at $\xi=(1,t_*)$. As $R_{(\varepsilon+h)/2}=R_{h/2}R_{\varepsilon/2}$, the same bound holds at every $\varepsilon$, with $\xi$ replaced by $R_{\varepsilon/2}\xi$. Thus $\frac{d}{d\varepsilon}\log\norm{R_{\varepsilon/2}\xi}_p\le\frac{D_*}2$, and integration from $0$ to $\varepsilon$ gives $\norm{R_{\varepsilon/2}\xi}_p\le e^{\varepsilon D_*/2}\norm{\xi}_p$ for $\varepsilon\ge0$. As $J$ is an isometry of $\ell^p_2$ (it permutes the coordinates up to sign), $\norm{T_\varepsilon\xi}_p=e^{-\varepsilon D_*/2}\norm{R_{\varepsilon/2}J\xi}_p\le\norm{J\xi}_p=\norm{\xi}_p$; so $T_\varepsilon$ is a contraction on real vectors, and hence, by Lemma~\ref{lem:RC}, on $\ell^p_2$ (see also Lemma~\ref{lem:TK}). Let $\xi_a$ be a real vector at which $\norm{P_a(J)\xi}_p/\norm{\xi}_p$ is maximal, normalised by $(\xi_a)_1=1$ and close to $(1,t_*)$ (such a vector exists by Lemma~\ref{lem:SJmax}). The derivative in $\varepsilon$ at $\varepsilon=0$ of $\norm{P_a(T_\varepsilon)\xi_a}_p^p$ is \[ \tfrac12\,p\,(p-1)(1+B)^{p-2}E_*\,a\,(B-1)+o(a)\quad(a\to0),\qquad E_*=\frac{t_*^{p-2}(1+t_*^2)^2}{1+t_*^p}>0 \] (Lemma~\ref{lem:TK}). At $\varepsilon=0$ we have $T_0=J$, so $\norm{P_a(T_0)\xi_a}_p=\norm{P_a(J)}_p\norm{\xi_a}_p$ by the choice of $\xi_a$. Hence, for $B>1$ and every sufficiently small $a>0$, we get $\norm{P_a(T_\varepsilon)\xi_a}_p>\norm{P_a(J)}_p\norm{\xi_a}_p$, and so $\norm{P_a(T_\varepsilon)}_p>\norm{P_a(J)}_p$, for all sufficiently small $\varepsilon>0$; how small $\varepsilon$ must be depends on $a$. The gain $\norm{P_a(T_\varepsilon)}_p-\norm{P_a(J)}_p$ is positive, but it is small: it tends to $0$ as $\varepsilon\to0$, and the argument gives no explicit lower bound for it. Therefore an upper bound $\norm{P_a(S)}_p\le\norm{P_a(J)}_p+r$ with an error term $r>0$ would prove nothing unless $r$ were shown to be smaller than this gain. This is why we prove the exact equality $\norm{P_a(S)}_p=\norm{P_a(J)}_p$. For $B=1$ the gain vanishes to first order, and indeed for $B=1$ inequality \eqref{eq:matsaev} holds (Theorem~\ref{thm:B}). \medskip\noindent\emph{Further results.} The table lists the results of Part~II next to the best results known before. \begin{center}\footnotesize \renewcommand{\arraystretch}{1.1} \begin{tabular}{|p{0.45\textwidth}|p{0.45\textwidth}|} \hline \textbf{Best result before this paper} & \textbf{New result}\\ \hline No result near $p=2$. & The quadratic $1+2z-\frac65z^2$, with the scaled rotations $R_{\theta_0}/\norm{R_{\theta_0}}_p$, $\cos\theta_0=-\frac1{12}$, violates \eqref{eq:matsaev} for every $p\ne2$ close to $2$; proof by hand (Section~\ref{sec:near2})\\ \hline A computer verification over many patches of the unit sphere \cite{Dr11}. & A rational certificate of Drury's counterexample at $p=4$, verified in exact arithmetic (Section~\ref{sec:p4exact})\\ \hline $\norm{P(S)}_4<6.5278$, and the guess $\approx6.4548$ \cite{Dr11}. & Upper bounds, certified in exact rational arithmetic, for $\norm{P(S)}_p$, $P=1+2z-\frac{22}5z^2$, at $p=3,4,6,\dots,20$, within $2\cdot10^{-6}$ of $\norm{P(J)}_p$ for even $p$ (Section~\ref{sec:shiftnorms})\\ \hline Drury's explicit pair, for $p=4$ and $p=4/3$ only \cite{Dr11}. & Explicit counterexamples for every $p\in[40/39,40]$, $p\ne2$: four explicit real quadratics and explicit real $2\times2$ contractions, certified by computer (Section~\ref{sec:explicitfull})\\ \hline Known classes of contractions that satisfy \eqref{eq:matsaev} \cite{Pe81,Ar13}. & Elementary necessary conditions on a $2\times2$ counterexample (Section~\ref{sec:structure})\\ \hline \end{tabular} \end{center} Programs, certificates and logs for Part II are in the ancillary files of this paper: the archive \path{matsaev_1971_ancillary_files_v1.10.zip} (top folder \path{ancillary_files/}) in its Zenodo record, \url{https://doi.org/10.5281/zenodo.23691816} \section{Basic facts}\label{sec:basic} As in the Introduction, for a real number $y$ put $\phi(y)=\sgn y\,|y|^{p-1}$, so that $\frac{d}{dy}|y|^p=p\,\phi(y)$ for $p>1$; on real $\xi\in\ell^p_2$, $\phi$ and $|\cdot|^{p-2}$ act coordinatewise, and a product of two vectors, such as $|\xi|^{p-2}\psi$, is taken coordinatewise. With $\langle\xi,\psi\rangle=\xi_1\psi_1+\xi_2\psi_2$, $\tfrac{d}{dh}\norm{\xi+h\psi}_p^p\big|_{h=0}=p\langle\phi(\xi),\psi\rangle$ for real $\xi,\psi$ and $p>1$. \begin{lemma}[real and complex norms]\label{lem:RC} Let $1\le p<\infty$. A real $2\times2$ matrix has the same norm on the real and on the complex $\ell^p_2$, and for a real polynomial $P$ the operator $P(S)$ has the same norm on the real and on the complex $\ell^p$. \end{lemma} \begin{proof} This is the Marcinkiewicz--Zygmund theorem \cite{MZ39} (see also \cite{HK05}). For a complex number $z=y_1+iy_2$ we have $y_1\cos\theta+y_2\sin\theta=|z|\cos(\theta-\arg z)$, so $\int_0^{2\pi}|y_1\cos\theta+y_2\sin\theta|^p\,d\theta=\beta_p|z|^p$ with $\beta_p=\int_0^{2\pi}|\cos\theta|^p\,d\theta$. Let $A$ be a real matrix (a $2\times2$ matrix, or the infinite matrix of $P(S)$), and let $x=u+iv$ with real vectors $u,v$. Applying this identity to each coordinate of $Ax=Au+iAv$ gives $\norm{Ax}_p^p=\beta_p^{-1}\int_0^{2\pi}\norm{A(u\cos\theta+v\sin\theta)}_p^p\,d\theta$, and the same identity for $A=I$ gives $\norm{x}_p^p=\beta_p^{-1}\int_0^{2\pi}\norm{u\cos\theta+v\sin\theta}_p^p\,d\theta$. Write $\norm{A}_{\mathrm{real}}$ for the norm of $A$ on real vectors. Since $\norm{A(u\cos\theta+v\sin\theta)}_p\le\norm{A}_{\mathrm{real}}\norm{u\cos\theta+v\sin\theta}_p$ for every $\theta$, we get $\norm{Ax}_p\le\norm{A}_{\mathrm{real}}\norm{x}_p$: the complex norm of $A$ is at most its real norm. The reverse inequality is clear. \end{proof} \begin{lemma}[a lower bound for $\norm{P(S)}_p$]\label{lem:J} For every polynomial $P$ and $1\le p\le\infty$, $\norm{P(J)}_p\le\norm{P(S)}_p$. \end{lemma} \begin{proof} Let $\xi\in\C^2$ and let $x$ be the sequence with $(x_{n},x_{n-1})=J^{-n}\xi$ for $n\in\Z$. This is consistent: if $J^{-n-1}\xi=(b_1,b_2)$, then $J^{-n}\xi=J(b_1,b_2)=(b_2,-b_1)$, so both rules give $x_n=b_2$. Then $x_{n+2}=-x_n$ (as $J^{-2}=-I$), and with $X_n=(x_n,x_{n-1})$ we have $X_{n-k}=J^kX_n$, so $y_n:=\sum_kc_kx_{n-k}=(P(J)X_n)_1$. Since $X_{n+1}=J^{-1}X_n=-JX_n$ and $P(J)$ commutes with $J$, $(P(J)X_{n+1})_1=-(P(J)X_n)_2$. Hence, with $Y_n=P(J)X_n$ (so $Y_{n+1}=-JY_n$ and $Y_{n+2}=-Y_n$), the numbers $|y_n|,|y_{n+1}|,|y_{n+2}|,|y_{n+3}|$ are $|(Y_n)_1|,|(Y_n)_2|,|(Y_n)_1|,|(Y_n)_2|$; since $\norm{Y_n}_p=\norm{J^{-n}P(J)\xi}_p=\norm{P(J)\xi}_p$ ($J$ is an isometry of $\ell^p_2$), over any four consecutive indices $\sum|y_n|^p=2\norm{P(J)\xi}_p^p$, and likewise $\sum|x_n|^p=2\norm{\xi}_p^p$. For $j\ge1$ let $x^{(j)}\in\ell^p(\N)$ agree with $x$ for $0\le n<4j$ and vanish elsewhere. Then $\norm{x^{(j)}}_p^p=2j\norm{\xi}_p^p$, and $(P(S)x^{(j)})_n=y_n$ for $\deg P\le n<4j$, while $(P(S)x^{(j)})_n=0$ for $n\ge4j+\deg P$; so $\norm{P(S)x^{(j)}}_p^p=2j\norm{P(J)\xi}_p^p+O(1)$, and the ratio $\norm{P(S)x^{(j)}}_p^p/\norm{x^{(j)}}_p^p$ tends to $\norm{P(J)\xi}_p^p/\norm{\xi}_p^p$ as $j\to\infty$ (the sums $\sum_{n\ge0}|(P(S)x^{(j)})_n|^p$ and $\sum_{0\le n<4j}|y_n|^p=2j\norm{P(J)\xi}_p^p$ differ only in the at most $2\deg P$ terms with $n<\deg P$ or $4j\le n<4j+\deg P$, and in these terms $|(P(S)x^{(j)})_n|$ and $|y_n|$ are at most $\sum_k|c_k|\max_m|x_m|$). Hence $\norm{P(S)}_p\ge\norm{P(J)\xi}_p/\norm{\xi}_p$ for every $\xi\ne0$, which is the claim. For $p=\infty$ the inequality is trivial: $\norm{P(J)}_\infty\le\sum|c_k|=\norm{P(S)}_\infty$. \end{proof} For $12$, $B>0$, and $P_a(z)=1+az-Bz^2$; put $C_a:=\norm{P_a(J)}_p$ (by Lemma~\ref{lem:RC}, a maximum over real vectors). For a real sequence $x$, the \emph{window} at $n$ is $(u,v,w)=(x_n,x_{n-1},x_{n-2})$; then $(P_a(S)x)_n=u+av-Bw$. \begin{lemma}[a Bellman-function bound for $\norm{P_a(S)}_p$]\label{lem:bellman} Let $C>0$, and let $G:\R^2\to\R$ be continuous with $G(0,0)=0$. If \begin{equation}\label{eq:bellman} F(u,v,w):=C^p|u|^p+G(v,w)-G(u,v)-|u+av-Bw|^p\;\ge\;0\qquad\text{for all }(u,v,w)\in\R^3, \end{equation} then $\norm{P_a(S)}_p\le C$. We call such a $G$ a \emph{Bellman function} for $P_a$ and $C$. \end{lemma} \begin{proof} For a finitely supported real $x\in\ell^p(\N)$ (with $x_{-1}=x_{-2}=0$, and $x_n=0$ for $n>n_0$) sum \eqref{eq:bellman} over the windows at $0\le n\le n_0+2$: the $G$ terms telescope to $G(0,0)-G(0,0)=0$, so $\norm{P_a(S)x}_p^p\le C^p\norm{x}_p^p$. By density this holds for all real $x\in\ell^p(\N)$, and by Lemma~\ref{lem:RC} also for complex $x$. \end{proof} \begin{lemma}[the case $a=0$]\label{lem:B0} With $G_0(v,w)=B(1+B)^{p-1}(|v|^p+|w|^p)$ and $C=1+B$, \eqref{eq:bellman} holds for $a=0$: $F_0(u,v,w)=(1+B)^{p-1}\bigl(|u|^p+B|w|^p\bigr)-|u-Bw|^p\ge0$, with equality exactly on the plane $\Pi=\{w=-u\}$, the set of windows of the sequences with $x_{n+2}=-x_n$. \end{lemma} \begin{proof} The identities $G_0(v,w)-G_0(u,v)=B(1+B)^{p-1}(|w|^p-|u|^p)$ and $(1+B)^p-B(1+B)^{p-1}=(1+B)^{p-1}$ give the formula for $F_0$. Write $u-Bw=(1+B)\bigl(\tfrac1{1+B}u+\tfrac B{1+B}(-w)\bigr)$; strict convexity of $|\cdot|^p$ ($p>1$) gives $|u-Bw|^p\le(1+B)^{p-1}(|u|^p+B|w|^p)$, with equality if and only if $u=-w$. \end{proof} \begin{lemma}[growth of $F_0$ away from $\Pi$]\label{lem:B0q} With $\nu=u+w$, there is $\gamma_1=\gamma_1(p,B)>0$ with $F_0(u,v,w)\ge\gamma_1(|u|+|\nu|)^{p-2}\nu^2$ for all real $u,v,w$. \end{lemma} \begin{proof} The function $F_0$ does not depend on $v$, and both sides are homogeneous of degree $p$ in $(u,\nu)$, so it suffices to consider $\mathcal K=\{|u|+|\nu|=1\}$. On $\mathcal K$ with $\nu\ne0$, $F_0>0$ by Lemma~\ref{lem:B0}. Near $\nu=0$ on $\mathcal K$, $|u|$ is close to $1$, and the bases of the three $p$-th powers in $F_0$ (the numbers $y$ in its terms $|y|^p$), namely $u$, $w=\nu-u$ and $u-Bw=(1+B)u-B\nu$, are close to $\pm1,\mp1,\pm(1+B)$; so $F_0$ is real-analytic there, $F_0=0$ at $\nu=0$, $\partial_\nu F_0=pB(1+B)^{p-1}\phi(\nu-u)+pB\,\phi\bigl((1+B)u-B\nu\bigr)$, which at $\nu=0$ is $-pB(1+B)^{p-1}\phi(u)+pB(1+B)^{p-1}\phi(u)=0$, and $\partial_\nu^2F_0=p(p-1)B(1+B)^{p-2}|u|^{p-2}>0$ at $\nu=0$. Hence $F_0/\nu^2$ extends to a positive continuous function on $\mathcal K$, and $\gamma_1$ is its minimum. \end{proof} \emph{The function $Q$.} As in the Introduction, for $\xi\in\R^2\setminus\{0\}$ put $Q(\xi)=\langle\phi(\xi),J\xi\rangle/\norm{\xi}_p^p=\bigl(\phi(\xi_1)\xi_2-\phi(\xi_2)\xi_1\bigr)/\norm{\xi}_p^p$. Since $J^2=-I$, $P_a(J)=(1+B)I+aJ$, so $\norm{P_a(J)\xi}_p^p=(1+B)^p\norm{\xi+hJ\xi}_p^p$ with $h=a/(1+B)$. The derivative formula of Section~\ref{sec:basic} gives the term of first order in $h$, namely $hp\langle\phi(\xi),J\xi\rangle=hp\,Q(\xi)\norm{\xi}_p^p$; the second derivative in $h$, namely $p(p-1)\sum_{i=1}^2|\xi_i+h(J\xi)_i|^{p-2}(J\xi)_i^2$, is bounded for $\xi$ on the Euclidean unit circle and $|h|\le1$ because $p>2$. Taylor's formula in $h$, and homogeneity in $\xi$, therefore give \begin{equation}\label{eq:Qexp} \norm{P_a(J)\xi}_p^p=(1+B)^p\norm{\xi}_p^p\Bigl(1+\tfrac{pa}{1+B}\,Q(\xi)\Bigr)+O(a^2)\norm{\xi}_p^p, \end{equation} so, to first order in $a$, the vectors $\xi$ that maximise $\norm{P_a(J)\xi}_p/\norm{\xi}_p$ are the maximisers of $Q$. For real $\sigma$ put $f(\sigma)=Q(1,\sigma)=\sigma(1-|\sigma|^{p-2})/(1+|\sigma|^p)$. \begin{lemma}[the maximum of $Q$]\label{lem:N1} Let $p>2$. The function $f$ has exactly one critical point $t_*$ in $(0,1)$; it is a nondegenerate maximum, and $D_*:=\max_{\xi\ne0}Q(\xi)=f(t_*)>0$ is attained exactly on the nonzero multiples of $(1,t_*)$ and of $J(1,t_*)$. Moreover \begin{equation}\label{eq:ST} 1+t_*^{2p-2}=(p-1)\,t_*^{p-2}(1+t_*^2). \end{equation} \end{lemma} \begin{proof} For $00$ on $(0,1)$ and $f(0)=f(1)=0$, $t_*$ is the only maximiser of $f$ on $[0,1]$. Finally, $Q$ is invariant under $\xi\mapsto J\xi$, $\xi\mapsto-\xi$ and $\xi\mapsto c\xi$ ($c>0$). Replacing $\xi$ by $J\xi$ if necessary, we may assume $|\xi_2|\le|\xi_1|$; then $\xi$ is a multiple of $(1,\sigma)$ with $|\sigma|\le1$, and $Q(\xi)=f(\sigma)$. Since $f(\sigma)\le0$ for $\sigma\in[-1,0]$, the maximum of $Q$ is $f(t_*)$, attained exactly at the nonzero multiples of $(1,t_*)$ and of $J(1,t_*)$. \end{proof} For example, $t_*=2\sin(\pi/12)$, $D_*=2^{-3/2}$ for $p=4$, and $t_*=(\sqrt5-1)/2$, $D_*=\frac12$ for $p=6$ (Figure~\ref{fig:f}). \begin{figure}[ht] \centering \input{fig_matsaev_f} \caption{The function $f(t)=(t-t^{p-1})/(1+t^p)$ of Lemma~\ref{lem:N1} on $[0,1]$, for $p=3,4,6$. The dot marks its maximum $D_*=f(t_*)$ (for $p=4$ and $p=6$, the values given above). Plotted numerically; an illustration, not used in the proofs.} \label{fig:f} \end{figure} \begin{figure}[ht] \centering \input{fig_matsaev_contraction} \caption{The real unit circle of $\ell^4_2$ (black); its image under the rotation $R_{\pi/2+\varepsilon/2}$ (red, dashed), which leaves the unit ball; and its image under $T_\varepsilon=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2}$ (blue), which stays inside it (Lemma~\ref{lem:TK}). Here $\varepsilon=0.8$, larger than in Theorem~\ref{thm:A}, so that the curves are easy to tell apart. Computed numerically; an illustration, not used in the proofs.} \label{fig:contraction} \end{figure} \begin{lemma}[the maximising vectors of $P_a(J)$]\label{lem:SJmax} Let $p>2$ and $B>0$. We measure distances on the Euclidean unit circle of $\R^2$ by angle. Let $e_*=(1,t_*)/\norm{(1,t_*)}_2$, let $\mathcal O_0=\{\pm e_*,\pm Je_*\}$, and for $s\in\R^2$ put $\Sigma_a(s)=2\bigl(C_a^p\norm{s}_p^p-\norm{P_a(J)s}_p^p\bigr)$; thus $\Sigma_a\ge0$ by the definition of $C_a$. There are $\rho\in(0,\frac12\arctan t_*)$, $a_1>0$ and $\delta,\eta>0$ such that for $00$ such that $\partial_\theta^2Q\le-m$ on the four closed arcs of angular radius $2\rho$ centred at the points of $\mathcal O_0$ (the \emph{$2\rho$-arcs}); these arcs are disjoint and avoid the coordinate axes. Write \begin{gather*} \frac{\Sigma_a(s)}a=2\norm{s}_p^p\bigl(c(a)-H(a,s)\bigr),\qquad\text{where}\\ H(a,s)=\frac1a\Bigl(\frac{\norm{P_a(J)s}_p^p}{\norm{s}_p^p}-(1+B)^p\Bigr),\qquad c(a)=\max_{s\ne0}H(a,s)=\frac{C_a^p-(1+B)^p}a . \end{gather*} By \eqref{eq:Qexp}, $H(a,s)\to H(0,s):=p(1+B)^{p-1}Q(s)$ uniformly on the unit circle, hence $c(a)\to p(1+B)^{p-1}D_*$ and, uniformly on the unit circle, \[ \Sigma_a(s)/a\;\longrightarrow\;\Sigma'(s):=2p(1+B)^{p-1}\norm{s}_p^p\bigl(D_*-Q(s)\bigr)\qquad(a\to0^+). \] (As $\Sigma_0=0$, $\Sigma'(s)$ is the right derivative of $\Sigma_a(s)$ at $a=0$.) On the $2\rho$-arcs both coordinates of $s$ have absolute value at least $\sin(\arctan t_*-2\rho)$, so for $|a|<\frac12(1+B)\sin(\arctan t_*-2\rho)$ no coordinate of $P_a(J)s=(1+B)s+aJs$ vanishes there. For such $a$, $\norm{P_a(J)s}_p^p/\norm{s}_p^p$ is real-analytic in $(a,s)$ and equals $(1+B)^p$ at $a=0$, so $H$ (with the value $H(0,s)$ above) is real-analytic in $(a,s)$, also at $a=0$. In particular $H(a,\cdot)\to H(0,\cdot)$ in $C^2$ on these arcs. By Lemma~\ref{lem:N1}, $\Sigma'\ge0$ vanishes on the unit circle exactly on $\mathcal O_0$. Put \[ \delta=\tfrac12\min\{\Sigma'(s):\ \norm{s}_2=1,\ s\text{ at angular distance}\ge\rho/2\text{ from }\mathcal O_0\}>0,\qquad m_1=\tfrac12p(1+B)^{p-1}m . \] By the uniform convergence, and the $C^2$ convergence with $\partial_\theta^2H(0,\cdot)=p(1+B)^{p-1}\partial_\theta^2Q\le-2m_1$, there is $a_1>0$ such that for $02$ and $B>0$ there is $a_0=a_0(p,B)>0$ such that $\norm{P_a(S)}_p=\norm{P_a(J)}_p$ for $0\le a0$ we construct $G=G_a$ with $C=C_a$ in Lemma~\ref{lem:bellman}. \emph{Outline.} Write a window as $\hat s+\nu e_3$ with the \emph{state} $s=(u,v)$, $\hat s=(u,v,-u)\in\Pi$, $\nu=u+w$ and $e_3=(0,0,1)$. At $a=0$, $F_0\ge0$ vanishes exactly on $\Pi$ and grows like $(|u|+|\nu|)^{p-2}\nu^2$ off it (Lemmas~\ref{lem:B0} and~\ref{lem:B0q}). First, along a sequence with $x_{n+2}=-x_n$ the states run through the cycle $s,-Js,-s,Js$, and the sum of $F$ (with $C=C_a$) over the four windows does not depend on $G$: it is $\Sigma_a(s)$. Second, we take $G_a=G_0+\frac12\Delta_a+k_a$, with $\Delta_a$ and $k_a$ defined in Steps~1 and~2, and $F_a$ denotes $F$ with $C=C_a$ and $G=G_a$. Then $F_a$ takes the same value $\frac14\Sigma_a(s)$ at each of the four windows of the cycle; in particular $F_a(\hat s)=\frac14\Sigma_a(s)\ge0$, which is zero exactly when $s=0$ or $s$ lies on a ray through a point of $\mathcal O_a$, and grows quadratically off these rays (Lemma~\ref{lem:SJmax}). Third, since we need $F_a\ge0$, the gradient of $F_a$ must vanish wherever $F_a$ vanishes. Along $\Pi$ this holds automatically. In the direction $e_3$, across $\Pi$, it is achieved by the term $k_a$, which is of size $O(a)$ and does not change $F_a$ on $\Pi$. Fourth, $F_a-F_0$ and its gradient are of size $O(a)$ on bounded sets (Step~5). Away from $\mathcal O_a$, $\Sigma_a\ge a\delta$ absorbs this error for small $\nu$, and $F_0$ absorbs it for large $\nu$; near $\mathcal O_a$, $F_a$ is bounded below by a positive quadratic form in the distance to $\mathcal O_a$ and $\nu$. Throughout, $J(u,v)=(v,-u)$, and $\rho$, $a_1$, $\delta$, $\eta$, $\mathcal O_a$, $s_a$ are those of Lemma~\ref{lem:SJmax}; the constant $a_1$ will be decreased finitely many times. \emph{Step 1: the sum over a cycle of windows.} For a sequence with $x_{n+2}=-x_n$, if the window at $n$ is $\hat s$, then the window at $n+1$ is $\widehat{-Js}$, and $(v,w)=(v,-u)=Js$ on $\Pi$. Thus, with \[ \Delta_a(s)=C_a^p|u|^p-|(1+B)u+av|^p , \] $F(\hat s)=\Delta_a(s)+G(Js)-G(s)$ on $\Pi$, for $C=C_a$. Since $\Delta_a$ is even and $-Js,-s,Js$ complete the cycle of $s$, the $G$ terms telescope around the cycle and \begin{equation}\label{eq:Sigma} \sum_{s'\in\{s,-Js,-s,Js\}}F(\widehat{s'})=2\bigl(\Delta_a(s)+\Delta_a(Js)\bigr)=\Sigma_a(s), \end{equation} because $\Delta_a(s)+\Delta_a(Js)=C_a^p(|u|^p+|v|^p)-|(1+B)u+av|^p-|(1+B)v-au|^p$ and $P_a(J)=(1+B)I+aJ$. \emph{Step 2: the Bellman function.} By Step~1, $F(\hat s)=\frac14\Sigma_a(s)$ for every $s$ if $G(Js)-G(s)=\frac12\bigl(\Delta_a(Js)-\Delta_a(s)\bigr)$; the choice $G=\frac12\Delta_a$ satisfies this. Adding to $G$ a function invariant under $J$ does not change $F$ on $\Pi$. Accordingly, with $G_0$ as in Lemma~\ref{lem:B0}, we put \[ G_a=G_0+\tfrac12\Delta_a+k_a,\qquad k_a=\lambda_1(a)k^{(1)}+\lambda_2(a)k^{(2)},\quad k^{(1)}(s)=(u^2+v^2)^{p/2},\ k^{(2)}(s)=|u|^p+|v|^p, \] with $\lambda(a)=(\lambda_1(a),\lambda_2(a))$ to be fixed in Step~4. The functions $G_0$, $k^{(1)}$, $k^{(2)}$ are even, $J$-invariant and homogeneous of degree $p$; the functions $k^{(1)}$ and $k^{(2)}$ are $C^2$ (as $p>2$); and $G_a(0,0)=0$. For $a=0$, $F_a$ is $F_0$ of Lemma~\ref{lem:B0}, as $\Delta_0=0$ and $\lambda(0)=0$ (see Step~4). On $\Pi$ the $J$-invariant parts cancel, so \begin{equation}\label{eq:onPi} F_a(\hat s)=\Delta_a(s)+\tfrac12\bigl(\Delta_a(Js)-\Delta_a(s)\bigr)=\tfrac14\Sigma_a(s)\ \ge0 . \end{equation} \emph{Step 3: the bases of the $p$-th powers near the zeros of $F_a$.} The function $|y|^p$ is real-analytic in $y$ for $y\ne0$, and we use its analyticity only there. We therefore show that the bases of all the $p$-th powers in $F_a$ stay away from $0$ near the windows $\hat s$, $s\in\mathcal O_a$. For $a>0$ these bases are $u$, $v$, $w$ (from $|u|^p$, $G_0$, $k^{(2)}$), $(1+B)u+av$, $(1+B)v+aw$ (from $\Delta_a(u,v)$ and $\Delta_a(v,w)$) and $u+av-Bw$. The function $k^{(1)}$ is smooth where $u^2+v^2$ and $v^2+w^2$ are positive. For $\nu_0>0$ let \[ \mathcal U(\nu_0)=\{(s,\nu):\ \norm{s}_2=1,\ s\text{ within angular distance }2\rho\text{ of }\mathcal O_0,\ |\nu|\le\nu_0\}, \] with $s=(u,v)$ and $w=\nu-u$. At $a=0$ and $\nu=0$ the six bases are $u,v,-u,(1+B)u,(1+B)v,(1+B)u$, nonzero as $s$ avoids the axes. Hence we can fix $\nu_0\in(0,1)$ and $\kappa,q_*>0$ such that on $\mathcal U=\mathcal U(\nu_0)$ the six numbers $u$, $v$, $w$, $(1+B)u$, $(1+B)v$, $u-Bw$ have absolute value at least $\kappa$, and $\partial_\nu^2F_0\ge4q_*$ (at $\nu=0$, $\partial_\nu^2F_0=p(p-1)B(1+B)^{p-2}|u|^{p-2}$, by the proof of Lemma~\ref{lem:B0q}). For $a>0$ the bases $(1+B)u+av$, $(1+B)v+aw$ and $u+av-Bw$ differ from $(1+B)u$, $(1+B)v$ and $u-Bw$ by $av$, $aw$ and $av$; as $|v|\le1$ and $|w|\le2$ on $\mathcal U$, these differences are at most $2a$ in absolute value. Therefore all six bases have absolute value at least $\kappa/2$ on $\mathcal U$ once $a\le\kappa/4$, and we shrink $a_1$ to at most $\kappa/4$. The terms $k^{(1)}(u,v)$ and $k^{(1)}(v,w)$ are also real-analytic on $\mathcal U$, as $u^2+v^2=1$ and $v^2+w^2\ge w^2\ge\kappa^2$ there. \emph{Step 4: the derivative across $\Pi$.} At the windows $\hat s$, $s\in\mathcal O_a$, $F_a=\frac14\Sigma_a=0$ by \eqref{eq:onPi} and Lemma~\ref{lem:SJmax}(i); since we need $F_a\ge0$, its gradient must vanish there. Along $\Pi$ it does, by \eqref{eq:onPi}; across $\Pi$ we impose it by the choice of $\lambda(a)$. Let $N_a(s)=\partial_\nu F_a(\hat s+\nu e_3)|_{\nu=0}$. The term $k_a(v,w)-k_a(u,v)$ contributes $\partial_2k_a(Js)$ to $N_a(s)$, where $\partial_1,\partial_2$ are the partial derivatives in the first and second argument. For a $J$-invariant $C^1$ function $k$ (such as $k_a$), differentiating $k(J\xi)=k(\xi)$ gives $\nabla k(J\xi)=J\nabla k(\xi)$ (as $J^{\mathsf t}=J^{-1}$), so this contribution is $-\partial_1k_a(s)$, and its contribution to $N_a(Js)$ is $-\partial_1k_a(Js)=-\partial_2k_a(s)$. Write $N_a^0$ for $N_a$ with $k_a=0$. We choose $\lambda(a)$ so that \[ \nabla k_a(s_a)=\bigl(N_a^0(s_a),\,N_a^0(Js_a)\bigr),\quad\text{i.e. }N_a(s_a)=N_a(Js_a)=0 . \] This is a $2\times2$ linear system. Here $s_0=e_*$ is a positive multiple of $(1,t_*)$; at $(1,t_*)$ the gradients of $k^{(1)},k^{(2)}$ are $p(1+t_*^2)^{(p-2)/2}(1,t_*)$ and $p(1,t_*^{p-1})$, and at $s_0$ they are positive multiples of these, so they are independent because $02$. Moreover $N_0^0=\partial_\nu F_0=0$ on all of $\Pi$ ($F_0\ge0$ vanishes on $\Pi$), and $N_a^0$ and $s_a$ are real-analytic near $a=0$ (Lemma~\ref{lem:SJmax}(iv)); so, after strictly decreasing $a_1$, the system is uniquely solvable for $0\le a\le a_1$, $\lambda(a)$ is real-analytic there, $\lambda(0)=0$, and $|\lambda(a)|\le\Lambda a$ there, for some constant $\Lambda$. As $F_a$ is even, $N_a(-s)=-N_a(s)$, so $N_a=0$ on all of $\mathcal O_a$. \emph{Step 5: a uniform $C^1$ bound.} Put $\mathcal E_a=F_a-F_0$, i.e. \[ \mathcal E_a=(C_a^p-(1+B)^p)|u|^p+\tfrac12\bigl(\Delta_a(v,w)-\Delta_a(u,v)\bigr)+k_a(v,w)-k_a(u,v)-\bigl(|u+av-Bw|^p-|u-Bw|^p\bigr). \] With $\norm{\cdot}_\infty$ the maximum norm on $\R^3$ there is $L$, depending on $p$, $B$ and $\Lambda$, with \begin{equation}\label{eq:Ea} |\mathcal E_a(\mathbf x)|\le La\norm{\mathbf x}_\infty^p,\qquad|\nabla\mathcal E_a(\mathbf x)|\le La\norm{\mathbf x}_\infty^{p-1}\qquad(\text{windows }\mathbf x=(u,v,w)\in\R^3,\ 0\le a2$) $|\phi(y+h)-\phi(y)|\le(p-1)(|y|+|h|)^{p-2}|h|$; with $y=u-Bw$, $h=av$, $|y|\le(1+B)\norm{\mathbf x}_\infty$ and $|h|\le a\norm{\mathbf x}_\infty$, these bound $|u+av-Bw|^p-|u-Bw|^p$ and its gradient (whose remaining part, $ap\,\phi(u+av-Bw)$ in the $v$-direction, is at most $ap(1+B+a)^{p-1}\norm{\mathbf x}_\infty^{p-1}$); after we write $\Delta_a(s)=(C_a^p-(1+B)^p)|u|^p-\bigl(|(1+B)u+av|^p-|(1+B)u|^p\bigr)$, they also bound the terms with $\Delta_a$. The terms with $k_a$ are bounded because $|\lambda(a)|\le\Lambda a$. \emph{Step 6: $F_a\ge0$.} Since $F_a$ is homogeneous of degree $p$, it suffices to consider windows $\hat s+\nu e_3$ with $\norm{s}_2=1$ or $s=0$. If $s=0$, $F_a(0,0,w)=F_0(0,0,w)+\mathcal E_a(0,0,w)\ge \bigl(B(1+B)^{p-1}-B^p-La\bigr)|w|^p\ge0$ once $La\le B(1+B)^{p-1}-B^p$, a positive number as $1+B>B$. If $\norm{s}_2=1$ and $\mathbf x=\hat s+\nu e_3$, then $F_a(\mathbf x)=F_0(\mathbf x)+\mathcal E_a(\hat s)+\bigl(\mathcal E_a(\mathbf x)-\mathcal E_a(\hat s)\bigr)$ with $\mathcal E_a(\hat s)=F_a(\hat s)=\frac14\Sigma_a(s)$; along the segment from $\hat s$ to $\mathbf x$ the maximum norm is at most $1+|\nu|$, so \eqref{eq:Ea} and Lemma~\ref{lem:B0q} give \begin{equation}\label{eq:star} F_a(\mathbf x)\;\ge\;\tfrac14\Sigma_a(s)+\gamma_1(|u|+|\nu|)^{p-2}\nu^2-La(1+|\nu|)^{p-1}|\nu| . \end{equation} \emph{Case 1: $s$ at angular distance $\ge\rho$ from $\mathcal O_a$.} (This case contains $s=(0,\pm1)$ and $s=(\pm1,0)$, where $u=0$ or $v=0$ and $F_a$ need not be real-analytic; the argument uses only \eqref{eq:star}.) Let $\nu_1=\min\bigl(1,\delta/(2^{p+2}L)\bigr)$. For $|\nu|\le\nu_1$, \eqref{eq:star} and Lemma~\ref{lem:SJmax}(ii) give $F_a\ge a\delta/4-2^{p-1}La\nu_1\ge a\delta/8$. For $|\nu|\ge\nu_1$, $(|u|+|\nu|)^{p-2}\nu^2\ge|\nu|^p$ and $(1+|\nu|)^{p-1}\le(1+1/\nu_1)^{p-1}|\nu|^{p-1}$, so $F_a\ge|\nu|^p\bigl(\gamma_1-La(1+1/\nu_1)^{p-1}\bigr)\ge0$ for small $a$. \emph{Case 2: $s$ within angular distance $\rho$ of $\mathcal O_a$.} Write $r=r(s)$. As $\mathcal O_a$ lies within $\rho/2$ of $\mathcal O_0$, $s$ lies within $3\rho/2<2\rho$ of $\mathcal O_0$, so for $|\nu|\le\nu_0$ the point $(s,\nu)$ is in $\mathcal U$, where all bases of all $p$-th powers in $F_a$ are at least $\kappa/2$ in absolute value (Step 3). Then $F_a$ is real-analytic in $(a,s,\nu)$ there, and by Taylor's formula in $\nu$, \[ F_a(\hat s+\nu e_3)=\tfrac14\Sigma_a(s)+\nu N_a(s)+\nu^2q_a(s,\nu),\qquad q_a(s,\nu)=\int_0^1(1-\tau)\,\partial_\nu^2F_a(\hat s+\tau\nu e_3)\,d\tau . \] By Step 3, $\partial_\nu^2F_0\ge4q_*$ on $\mathcal U$; as $F_a$ is real-analytic in $(a,s,\nu)$ on the compact set $[0,a_1]\times\mathcal U$, there is $L_2$ with $|\partial_\nu^2(F_a-F_0)|\le L_2a$ there, so $q_a\ge q_*$ on $\mathcal U$ once $L_2a\le2q_*$. Since $N_0=N_0^0=0$ on all of $\Pi$ (Step~4, as $\lambda(0)=0$) and $N_a$ is real-analytic in $a$, the derivative of $N_a$ along the unit circle is $O(a)$ on the $2\rho$-arcs. As $N_a$ vanishes on $\mathcal O_a$, there is a constant $L_1$ with $|N_a(s)|\le L_1ar$. With Lemma~\ref{lem:SJmax}(iii), \[ F_a\ \ge\ \tfrac14a\eta r^2-L_1ar|\nu|+q_*\nu^2\ \ge\ 0\qquad\text{once }(L_1a)^2\le a\eta q_* , \] since then the discriminant $(L_1a)^2-4\cdot\frac14a\eta\cdot q_*$ of this quadratic form in $(r,|\nu|)$ is $\le0$. For $|\nu|\ge\nu_0$, \eqref{eq:star} with $\Sigma_a\ge0$, $(|u|+|\nu|)^{p-2}\nu^2\ge|\nu|^p$ and $(1+|\nu|)^{p-1}\le(1+1/\nu_0)^{p-1}|\nu|^{p-1}$ gives $F_a\ge|\nu|^p\bigl(\gamma_1-La(1+1/\nu_0)^{p-1}\bigr)\ge0$ for small $a$. \emph{Order of the constants.} The constants are chosen in this order: $\gamma_1$ (Lemma~\ref{lem:B0q}); $\rho$, $m$, $\delta$, $\eta$ and $a_1$ (Lemma~\ref{lem:SJmax}); $\nu_0$, $\kappa$, $q_*$ (Step 3); $\Lambda$ (Step 4, after the last shrinking of $a_1$); $L$ (Step 5); $L_1$, $L_2$, $\nu_1$ (Step 6); finally, $a_0$ is the smallest of $a_1$ and the finitely many upper bounds on $a$ imposed in Step~6. For $02$, $\norm{P_a(S)}_p=\norm{P_a(S)}_{p'}=\norm{P_a(J)}_{p'}=\norm{P_a(J^{\mathsf t})}_p$ for $0\le a\norm{P_a(J)}_p$]\label{lem:TK} Let $p>2$, $B>1$, $K_*=\tfrac12(J-D_*I)$ and $T_\varepsilon=\exp(\varepsilon K_*)J$. Then $\norm{T_\varepsilon}_p\le1$ for all $\varepsilon\ge0$, on both the real and the complex $\ell^p_2$, and there is $a_2>0$ such that for each $0\norm{P_a(J)}_p$ for all sufficiently small $\varepsilon>0$ (depending on $a$). \end{lemma} \begin{proof} \emph{The matrix $T_\varepsilon$ is a contraction.} For real $\xi$ and $\xi(\varepsilon)=\exp(\varepsilon K_*)\xi$, $\frac{d}{d\varepsilon}\norm{\xi(\varepsilon)}_p^p=p\langle\phi(\xi(\varepsilon)),K_*\xi(\varepsilon)\rangle=\tfrac p2\norm{\xi(\varepsilon)}_p^p\, (Q(\xi(\varepsilon))-D_*)\le0$. Hence $\exp(\varepsilon K_*)$, and so $T_\varepsilon=\exp(\varepsilon K_*)J$ ($J$ is an isometry), is a contraction of the real $\ell^p_2$, and by Lemma~\ref{lem:RC} also of the complex $\ell^p_2$. (As $\exp(\varepsilon K_*)=e^{-\varepsilon D_*/2}R_{\varepsilon/2}$, $T_\varepsilon=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2}$.) \emph{Choice of test vector.} By Lemma~\ref{lem:SJmax}(i), (iv), for $00$. Using $K_*J=-\frac12(I+D_*J)$, the remaining terms are $(p-1)(1+B)^{p-2}BE_*-\frac12(1+B)^{p-1}\bigl(\norm{\xi_0}_p^p+D_*\langle\phi(\xi_0),J\xi_0\rangle\bigr)$, that is \[ \Gamma'(0)=(1+B)^{p-2}\Bigl[-\tfrac12(1+B)(1+t^p)(1+D_*^2)+(p-1)BE_*\Bigr]. \] By \eqref{eq:ST}, $(1+t^p)^2(1+D_*^2)=(1+t^p)^2+(t-t^{p-1})^2=(1+t^2)(1+t^{2p-2})=(p-1)t^{p-2}(1+t^2)^2$, that is $(1+t^p)(1+D_*^2)=(p-1)E_*$. Therefore \[ \Gamma(a)=a\,(1+B)^{p-2}(p-1)E_*\,\frac{B-1}{2}+o(a), \] which is positive for $01$. For such $a$, $p\,\Gamma(a)$ is positive, so for all small $\varepsilon>0$, $\norm{P_a(T_\varepsilon)}_p\ge\norm{P_a(T_\varepsilon)\xi_a}_p/\norm{\xi_a}_p>\norm{P_a(J)\xi_a}_p/\norm{\xi_a}_p =\norm{P_a(J)}_p$. \end{proof} \section{Proofs of Theorems A and B}\label{sec:proofs} \begin{proof}[Proof of Theorem~\ref{thm:A}] Let $p>2$ and $B>1$. For $0\norm{P_a(J)}_p$, and Theorem~\ref{thm:SJ} gives $\norm{P_a(J)}_p\ge\norm{P_a(S)}_p$. Since $\ell^p_2$ is an $L^p$ space, this contradicts \eqref{eq:matsaev}. This holds for real and for complex scalars alike: $T_\varepsilon$ is a contraction of both spaces (Lemma~\ref{lem:TK}), and the real operators $P_a(T_\varepsilon)$, $P_a(J)$ and $P_a(S)$ have the same norms on the real and on the complex spaces (Lemma~\ref{lem:RC}). For $12$ and use Lemma~\ref{lem:dual}: $T_\varepsilon^{\mathsf t}$ is a contraction of $\ell^p_2$ and $\norm{P_a(T_\varepsilon^{\mathsf t})}_p=\norm{P_a(T_\varepsilon)}_{p'}>\norm{P_a(S)}_{p'}=\norm{P_a(S)}_p$. Finally, if $T$ is any of these contractions, $T\oplus0$ is a contraction of $\ell^p_2\oplus_p\ell^p=\ell^p$ with $P_a(T\oplus0)=P_a(T)\oplus I$, hence $\norm{P_a(T\oplus0)}_p=\max(\norm{P_a(T)}_p,1)\ge\norm{P_a(T)}_p$, so \eqref{eq:matsaev} also fails on $\ell^p$ itself. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:B}] If $c_0=c_2=0$, then $P(z)=c_1z$, and $\norm{P(T)}_p\le|c_1|=\norm{P(S)}_p$ because $S$ is an isometry. Otherwise choose $\zeta\in\C$ with $|\zeta|=1$ and $\zeta^2=-c_2/c_0$; then $P(z)=c_0\widetilde P(\zeta z)$ with $\widetilde P(z)=1+\alpha z-z^2$, $\alpha=c_1/(c_0\zeta)\in\C$. Since $\zeta T$ is a contraction and $\zeta S=USU^{-1}$ with the isometry $(Ux)_n=\zeta^nx_n$ (as in the Background), it suffices to treat $\widetilde P$: if $\norm{\widetilde P(T')}_p\le\norm{\widetilde P(S)}_p$ for every contraction $T'$, then $\norm{P(T)}_p=|c_0|\norm{\widetilde P(\zeta T)}_p\le|c_0|\norm{\widetilde P(S)}_p=|c_0|\norm{U\widetilde P(S)U^{-1}}_p =|c_0|\norm{\widetilde P(\zeta S)}_p=\norm{P(S)}_p$. For a contraction $T$ on $L^p=L^p(\Omega,\mathfrak m)$ and $x\in L^p$, \[ \widetilde P(T)x=\tfrac12\bigl[(2x+\alpha Tx)+T(\alpha x-2Tx)\bigr]. \] By convexity of $|\cdot|^p$ (pointwise) and $\norm{T}_p\le1$, \begin{align*} \norm{\widetilde P(T)x}_p^p&\le\tfrac12\norm{2x+\alpha Tx}_p^p+\tfrac12\norm{\alpha x-2Tx}_p^p =\tfrac12\int_\Omega\norm{M(x(\omega),Tx(\omega))}_p^p\,d\mathfrak m(\omega)\\ &\le\tfrac12\norm{M}_p^p\bigl(\norm{x}_p^p+\norm{Tx}_p^p\bigr)\le\norm{M}_p^p\norm{x}_p^p, \end{align*} where $M=\begin{pmatrix}2&\alpha\\ \alpha&-2\end{pmatrix}$ acts pointwise on $(x,Tx)$. As $M=V\widetilde P(J)$ with $V=\mathrm{diag}(1,-1)$, $\norm{M}_p=\norm{\widetilde P(J)}_p\le\norm{\widetilde P(S)}_p$ by Lemma~\ref{lem:J}. (For $p=\infty$, replace each average of $p$-th powers by a maximum and the integral by an essential supremum.) \end{proof} \begin{corollary}[exact norms for $|c_0|=|c_2|$]\label{cor:exactB} Let $P(z)=c_0+c_1z+c_2z^2$ with $|c_0|=|c_2|>0$, and let $\zeta\in\C$ satisfy $|\zeta|=1$ and $\zeta^2=-c_2/c_0$. For $1\le p\le\infty$, $\norm{P(S)}_p=\norm{P(\bar\zeta J)}_p=\max_{|\mu|=1}\norm{P(\mu J)}_p$. \end{corollary} \begin{proof} With $P(z)=c_0\widetilde P(\zeta z)$ as in the proof of Theorem~\ref{thm:B}, $\norm{P(S)}_p=|c_0|\norm{\widetilde P(S)}_p$. The proof of Theorem~\ref{thm:B} with $T=S$ gives $\norm{\widetilde P(S)}_p\le\norm{\widetilde P(J)}_p$, and Lemma~\ref{lem:J} gives equality; here $c_0\widetilde P(J)=P(\bar\zeta J)$. For every $|\mu|=1$, Lemma~\ref{lem:J} applied to $z\mapsto P(\mu z)$ gives $\norm{P(\mu J)}_p\le\norm{P(\mu S)}_p=\norm{P(S)}_p$. (The factor $\bar\zeta$ cannot be omitted: for $P(z)=1+z^2$, $P(J)=0$, while, with $\zeta=i$, $\norm{P(S)}_p=\norm{P(-iJ)}_p=\norm{2I}_p=2$.) \end{proof} \begin{remark}[Theorem~\ref{thm:B} is not the trivial bound]\label{rem:triv} Theorem~\ref{thm:B} does not reduce to the trivial bound $\norm{P(T)}_p\le\sum|c_k|$: for $\widetilde P(z)=1+\alpha z-z^2$ with real $\alpha\ne0$ ($\zeta=1$ in Corollary~\ref{cor:exactB}) and $12$ by \eqref{eq:Qexp}, as $D_*>0$. As $F_a(\hat s_a)=0$, a nonzero derivative across $\Pi$ there makes $F_a$ negative nearby. \end{remark} \begin{remark}\label{rem:SJ} As $p\to2^+$, $Q\to0$ and the curvature of $\Sigma'$ at $\mathcal O_0$ tends to $0$. For $B\ne1$, let $\bar a(p)=\sup\{b\ge0:\ \norm{P_a(S)}_p=\norm{P_a(J)}_p\text{ for all }a\in[0,b]\}$. Then $\bar a(p)\to0$ as $p\to2$. Indeed, for $p=2$, $B\ne1$ and every $a>0$, $\norm{P_a(S)}_2=\max_{|z|=1}|P_a(z)|>|P_a(i)|=\norm{P_a(J)}_2$, because $\frac{d}{d\theta}|P_a(e^{i\theta})|^2=2a(B-1)\ne0$ at $\theta=\pi/2$; and both norms depend continuously on $p$ (for $\norm{P_a(S)}_p$ by the Riesz--Thorin theorem \cite[Thm.~1.1.1]{BL76}). (The assumption $B\ne1$ is needed: for $B=1$, Corollary~\ref{cor:exactB} gives $\norm{P_a(S)}_p=\norm{P_a(J)}_p$ for every $a\ge0$ and every $p$.) \end{remark} By Remark~\ref{rem:SJ}, in the proof of Theorem~\ref{thm:A} the admissible coefficients $a$ shrink to $0$ as $p\to2$, so that proof gives no single polynomial for a whole punctured neighbourhood of $p=2$. Here an explicit quadratic and an explicit family of scaled rotations violate \eqref{eq:matsaev} for every $p\ne2$ in a neighbourhood of $p=2$, with a proof by hand. Throughout this section $M_P=\max_{|z|=1}|P(z)|$, and $\varepsilon$ denotes $p-2$ (it is not the rotation parameter $\varepsilon$ of Theorem~\ref{thm:A}). \begin{theorem}\label{thm:N2} Let $P(z)=1+2z-\tfrac65z^2$, let $\theta_0\in(\pi/2,\pi)$ with $\cos\theta_0=-\tfrac1{12}$, and put $T_p=R_{\theta_0}/\norm{R_{\theta_0}}_p$. There is $\delta>0$ such that for every $p$ with $0<|p-2|<\delta$, $T_p$ is a contraction on $\ell^p_2$ and $\norm{P(T_p)}_p>\norm{P(S)}_p$. \end{theorem} At $p=2$ both sides equal $M_P$. Moving $p$ away from $2$, the left side grows linearly in $|p-2|$ (Lemma~\ref{lem:near2A}), while the right side has slope $0$ (Lemma~\ref{lem:near2Z} and the paragraph after its proof), because $|P|$ attains its maximum on the circle at an angle that is an irrational multiple of $\pi$. Lemma~\ref{lem:near2A} treats $p<2$ and $p>2$ alike; Lemma~\ref{lem:near2Z} reduces $p<2$ to $p'>2$ by Lemma~\ref{lem:dual}. \subsection{The norm of a rotation near \texorpdfstring{$p=2$}{p=2}} For a probability vector $\mu$ let $H(\mu)=-\sum_i\mu_i\log\mu_i$, and put $\mathcal H(\omega)=H(\cos^2\omega,\sin^2\omega)$. For $v\ne0$ in $\R^n$, \begin{equation}\label{eq:entropy} \tfrac{d}{dp}\log\norm{v}_p\big|_{p=2}=-\tfrac14H(\mu),\qquad\mu_i=|v_i|^2/\norm{v}_2^2, \end{equation} as one sees by differentiating $\frac1p\log\sum_i|v_i|^p$. The function $\mathcal H$ is even and has period $\pi/2$; for $t=\tan\omega\in(0,1]$, $|\mathcal H'(\omega)|=|\sin2\omega\,\log\cot^2\omega|=4t|\log t|/(1+t^2)\le4/e$; by these symmetries, $|\mathcal H'|\le4/e$ everywhere. \begin{lemma}\label{lem:near2D} For every $\theta$, $\norm{R_\theta}_p=1+|p-2|\,\sigma(\theta)+o(p-2)$ as $p\to2$, where $\sigma(\theta)=\frac14\max_\omega\bigl(\mathcal H(\omega)-\mathcal H(\omega-\theta)\bigr)\ge0$. Moreover $\sigma$ depends only on the distance from $\theta$ to $\frac\pi2\Z$, and $\sigma(\theta)\le\operatorname{dist}(\theta,\frac\pi2\Z)/e$. \end{lemma} \begin{proof} By Lemma~\ref{lem:RC} we may work over $\R$. With $x_\omega=(\cos\omega,\sin\omega)$, $R_\theta x_\omega=x_{\omega-\theta}$, so $\norm{R_\theta}_p=\max_\omega\psi_\omega(p)$ with $\psi_\omega(p)=\norm{x_{\omega-\theta}}_p/\norm{x_\omega}_p$. Each $\psi_\omega(2)=1$, and $(\omega,p)\mapsto\psi_\omega(p)$ is $C^1$ (as $t\mapsto t^p\log t$ is continuous on $[0,1]$), with $\partial_p\psi_\omega(2)=\frac14(\mathcal H(\omega)-\mathcal H(\omega-\theta))$ by \eqref{eq:entropy}. By uniform continuity of $\partial_p\psi_\omega(p)$ on compact sets, $\psi_\omega(2+\varepsilon)=1+\varepsilon\,\partial_p\psi_\omega(2)+o(\varepsilon)$ uniformly in $\omega$, so $\norm{R_\theta}_{2+\varepsilon}=1+\varepsilon\max_\omega\partial_p\psi_\omega(2)+o(\varepsilon)$ for $\varepsilon>0$ and $1+\varepsilon\min_\omega\partial_p\psi_\omega(2)+o(\varepsilon)$ for $\varepsilon<0$ \cite{Da67}. Here $\max_\omega\partial_p\psi_\omega(2)=\sigma(\theta)$ by definition, and $\min_\omega\partial_p\psi_\omega(2)=-\sigma(\theta)$, since the substitution $\omega\mapsto\theta-\omega$ changes the sign of $\mathcal H(\omega)-\mathcal H(\omega-\theta)$ ($\mathcal H$ is even). The last two claims follow from the symmetries of $\mathcal H$ and the bound on $|\mathcal H'|$. \end{proof} \begin{lemma}\label{lem:near2A} Let $P$ be a real polynomial and $|P|$ maximal on the circle at $z_0=e^{i\theta_0}$, with $\sin\theta_0\neq0$. Put $\rho=z_0P'(z_0)/P(z_0)$, $\alpha=\arg P(z_0)$, $\tau_p=1/\norm{R_{\theta_0}}_p$, $T_p=\tau_pR_{\theta_0}$. Then $\rho\ge0$ is real, \begin{equation}\label{eq:near2exact} \norm{P(T_p)}_p=|P(\tau_pz_0)|\;\norm{R_{\alpha(\tau_p)}}_p,\qquad\alpha(\tau)=\arg P(\tau z_0), \end{equation} and $\norm{P(T_p)}_p\ge M_P\bigl(1+g\,|p-2|\bigr)+o(p-2)$ with $g=\sigma(\alpha)-\rho\,\sigma(\theta_0)$. \end{lemma} \begin{proof} The span of $I$ and $J$ is an algebra isomorphic to $\C$ ($J^2=-I$), with $R_\theta\leftrightarrow e^{i\theta}$; so $P(\tau R_{\theta_0})=\operatorname{Re}P(\tau z_0)\,I+\operatorname{Im}P(\tau z_0)\,J=|P(\tau z_0)|\,R_{\alpha(\tau)}$, which gives \eqref{eq:near2exact}. At a maximum of $\log|P(e^{it})|$ its $t$-derivative, $-\operatorname{Im}\rho$, vanishes, and $|P(rz_0)|\le M_P$ for $r<1$ gives $\operatorname{Re}\rho\ge0$. Hence $|P(\tau z_0)|=M_P\bigl(1-\rho(1-\tau)\bigr)+O((1-\tau)^2)$, and $1-\tau_p=|p-2|\,\sigma(\theta_0)+o(p-2)$ by Lemma~\ref{lem:near2D}. For the second factor a lower bound suffices. For $p>2$ fix $\omega_*$ with $\sigma(\alpha)=\frac14(\mathcal H(\omega_*)-\mathcal H(\omega_*-\alpha))$, and for $p<2$ fix $\omega_*$ with $-\sigma(\alpha)=\frac14(\mathcal H(\omega_*)-\mathcal H(\omega_*-\alpha))$ (Lemma~\ref{lem:near2D} and its proof); let $\Psi(p,\tau)=\norm{R_{\alpha(\tau)}x_{\omega_*}}_p/\norm{x_{\omega_*}}_p\le\norm{R_{\alpha(\tau)}}_p$. $\Psi$ is $C^1$ near $(2,1)$, $\Psi(2,\tau)=1$ for all $\tau$, and $\partial_p\Psi(2,1)=\operatorname{sgn}(p-2)\,\sigma(\alpha)$, so $\Psi(p,\tau_p)=1+|p-2|\,\sigma(\alpha)+o(p-2)$. Multiplying the two expansions gives the claim. \end{proof} \begin{remark}[the family $P_\eta$] For $P_\eta(z)=1+2z-(1+\eta)z^2$, $\eta>0$, one has $|P_\eta(e^{it})|^2=(2-\eta\cos t)^2+(2+\eta)^2\sin^2t$, maximal exactly at $\cos t=-\eta/(2(1+\eta))$; there $P_\eta(z_0)/z_0=2(2+\eta)\cos(\theta_0/2)e^{-i\theta_0/2}$ and $P_\eta'(z_0)=2(1+\eta)(1+\cos\theta_0-i\sin\theta_0)$ (recall $\rho=P_\eta'(z_0)/(P_\eta(z_0)/z_0)$), so $\alpha=\theta_0/2$, $M_{P_\eta}=2(2+\eta)\cos(\theta_0/2)$ and $\rho=2(1+\eta)/(2+\eta)$ for every $\eta$. The value $\eta=\frac7{10}$ is the polynomial of item~(a) of Theorem~\ref{thm:explicit}, and $\eta=\frac{17}5$ is Drury's polynomial; for the latter, numerically $g\approx-0.038<0$, so Lemma~\ref{lem:near2A} gives no counterexample with Drury's polynomial near $p=2$ (Drury's matrix, which is not a scaled rotation, gives a counterexample at $p=4$). \end{remark} \begin{lemma}\label{lem:near2g} For $P(z)=1+2z-\frac65z^2$, $|P|$ is maximal on the circle exactly at $e^{\pm i\theta_0}$ with $\cos\theta_0=-\frac1{12}$, and $M_P=\frac{11\sqrt{66}}{30}$, $\rho=\frac{12}{11}$, $\alpha=\theta_0/2$, and $g\ge0.1366>0$. \end{lemma} \begin{proof} With $u=\cos t$, $|P(e^{it})|^2=\frac{221}{25}-\frac45u-\frac{24}5u^2$, maximal at $u=-\frac1{12}$ with value $M_P^2=\frac{1331}{150}$. At $z_0=e^{i\theta_0}$, $P(z_0)/z_0=\frac{11}{60}(11-\sqrt{143}\,i)$ and $P'(z_0)=\frac15(11-\sqrt{143}\,i)$, so $\rho=\frac{12}{11}$; and $\tan(\theta_0/2)=\sin\theta_0/(1+\cos\theta_0)=\sqrt{143}/11$ gives $\arg(P(z_0)/z_0)=-\theta_0/2$, so $\alpha=\theta_0/2$. Let $\delta_0=\theta_0-\frac\pi2=\arcsin\frac1{12}<0.083431$. By Lemma~\ref{lem:near2D}, $\sigma(\theta_0)\le\delta_0/e<0.030693$. Taking $\omega=\pi/4$ in the definition of $\sigma$, $\sigma(\alpha)\ge\frac14\bigl(\mathcal H(\frac\pi4)-\mathcal H(\frac\pi4-\alpha)\bigr)=\frac14\bigl(\log2-\mathcal H(\frac{\delta_0}2)\bigr)$, since $\frac\pi4-\alpha=-\frac{\delta_0}2$ and $\mathcal H$ is even. Here $\sin^2\frac{\delta_0}2=\frac12(1-\frac{\sqrt{143}}{12})<0.0017392$, so $\mathcal H(\frac{\delta_0}2)<0.012792$ and $\sigma(\alpha)>0.170088$. Hence $g>0.170088-\frac{12}{11}\cdot0.030693>0.1366$. The decimal bounds are roundings in the safe direction, checked by hand from $\arcsin x\le x+\frac{x^3}{6(1-x^2)}$ for $0\le x<1$ (the Taylor coefficients of $\arcsin$ after the first are at most $\frac16$), $e>2.71828$, $\log2>0.693147$, and $\mathcal H(\omega)\le s\,(1+\log\frac1s)$ for $s=\sin^2\omega\in(0,1)$ (as $-\log(1-s)\le\frac s{1-s}$; the right side increases in $s$), with $\log\frac1{0.0017392}<6.3544$. \end{proof} \subsection{An upper bound for \texorpdfstring{$\norm{P(S)}_p$}{the norm of P(S)} near \texorpdfstring{$p=2$}{p=2}} \begin{lemma}\label{lem:near2Z} Let $P(z)=c_0+c_1z+c_2z^2$ be real with $c_0c_2<0$, and suppose $|P|$ attains its maximum $M_P$ on the circle exactly at $e^{\pm i\theta_0}$, with $\theta_0/\pi$ irrational. Then $\norm{P(S)}_p\le M_P+o(p-2)$ as $p\to2$. \end{lemma} \begin{proof} By Lemma~\ref{lem:dual} it suffices to take $p=2+\varepsilon$, $\varepsilon>0$. We use windows $(u,v,w)=(x_n,x_{n-1},x_{n-2})$ and $y=c_0u+c_1v+c_2w$, as in Lemma~\ref{lem:bellman} (whose proof applies verbatim to any real quadratic $P$). \emph{(Z1) $p=2$, exactly.} Put $q^2=-c_0c_2>0$. As a quadratic in $\cos t$, $|P(e^{it})|^2$ is concave ($4c_0c_2<0$) with maximiser $\cos\theta_0=-c_1(c_0+c_2)/(4c_0c_2)$, so $M_P^2-|P(e^{it})|^2=4q^2(\cos t-\cos\theta_0)^2$. Matching coefficients shows that the quadratic form $\mathcal G_0(s,t)=(M_P^2-c_0^2+c_0c_2)s^2+(4q^2\cos\theta_0-2c_0c_1)st+c_2(c_2-c_0)t^2$ satisfies \begin{equation}\label{eq:near2F0} F_0(u,v,w):=M_P^2u^2+\mathcal G_0(v,w)-\mathcal G_0(u,v)-y^2=q^2\,(u-2v\cos\theta_0+w)^2\ \ge0 . \end{equation} Its zero set is the plane $\Pi_{\theta_0}=\{u-2v\cos\theta_0+w=0\}$: the windows of the sequences $x_n=r\cos(\theta+n\theta_0)$. On such a window, with $\mathbf e(\theta)=(\cos\theta,\cos(\theta-\theta_0))$, the two states are $(u,v)=r\,\mathbf e(\theta)$ and $(v,w)=r\,\mathbf e(\theta-\theta_0)$, and $y=rM_P\cos(\theta+\gamma)$ for a fixed $\gamma$, since $|c_0+c_1e^{-i\theta_0}+c_2e^{-2i\theta_0}|=M_P$. The curve $E=\{\mathbf e(\theta)\}$ is an ellipse (as $\sin\theta_0\ne0$); let $N$ be its gauge ($N=1$ on $E$) and $\theta(s)$ the angle of $s/N(s)$ on $E$. \emph{(Z2) The ansatz.} Fix $\kappa>0$ and put $C_\varepsilon=M_P(1+\kappa\varepsilon)$, $\mathcal G_\varepsilon(s)=N(s)^\varepsilon \bigl(\mathcal G_0(s)+\varepsilon N(s)^2k(\theta(s))\bigr)$ with a continuous $k$ on $\R/2\pi\Z$ chosen below, and $F_\varepsilon=C_\varepsilon^p|u|^p+\mathcal G_\varepsilon(v,w)-\mathcal G_\varepsilon(u,v)-|y|^p$. Then $F_\varepsilon=F_0+\varepsilon F_1+R_\varepsilon$ with $F_1$ continuous and $|R_\varepsilon|\le K\varepsilon^2$ on the Euclidean unit sphere of $\R^3$ (the terms are $|t|^{2+\varepsilon}$, $N^\varepsilon\mathcal G_0$ and $\varepsilon N^{\varepsilon+2}k$, and $|t|^2\log^2|t|$ is bounded near $0$). On $\Pi_{\theta_0}$ with $r=1$ both states lie on $E$, so $F_1=\Omega(\theta)+k(\theta-\theta_0)-k(\theta)$ with \[ \Omega(\theta)=M_P^2\bigl[(\log M_P+2\kappa)\cos^2\theta+\cos^2\theta\log|\cos\theta|-\cos^2(\theta+\gamma)\bigl(\log M_P+\log|\cos(\theta+\gamma)|\bigr)\bigr], \] and $\int_0^{2\pi}\Omega=2\pi\kappa M_P^2$: the other terms cancel by translation invariance. \emph{(Z3) A coboundary.} As $\theta_0/2\pi$ is irrational, the rotation $\theta\mapsto\theta-\theta_0$ is uniquely ergodic, so the averages $A_L\Omega=\frac1L\sum_{j0$ on the sphere, by \eqref{eq:near2F0}. Hence $F_\varepsilon\ge0$ on the sphere, and so on $\R^3$, for all small $\varepsilon>0$. \emph{(Z5) Conclusion.} Summing over all windows of a finitely supported real sequence, the $\mathcal G_\varepsilon$ terms telescope ($\mathcal G_\varepsilon(0)=0$), so $\norm{P(S)x}_p\le C_\varepsilon\norm{x}_p$ for every finitely supported real $x$, hence for every real $x\in\ell^p(\N)$, and Lemma~\ref{lem:RC} extends this to complex $x$. Hence $\norm{P(S)}_{2+\varepsilon}\le M_P(1+\kappa\varepsilon)$ for all small $\varepsilon>0$; as $\kappa>0$ was arbitrary, the lemma follows. \end{proof} Conversely, $\norm{P(S)}_p\ge\norm{P(S)}_2=M_P$ for every $p\in(1,\infty)$: $\log\norm{P(S)}_{1/s}$ is convex in $s$ by the Riesz--Thorin theorem \cite[Thm.~1.1.1]{BL76}, it takes the same value at $s=1/p$ and $s=1/p'$ by Lemma~\ref{lem:dual}, and $s=\frac12$ is the midpoint. So $\norm{P(S)}_p=M_P+o(p-2)$, that is, the slope at $p=2$ is $0$. Only the upper bound is used. \begin{proof}[Proof of Theorem~\ref{thm:N2}] By Lemma~\ref{lem:near2g}, $|P|$ is maximal on the circle exactly at $e^{\pm i\theta_0}$ and $g>0$; $c_0c_2=-\frac65<0$; and $\theta_0/\pi$ is irrational, since otherwise $2\cos\theta_0=-\frac16$ would be an algebraic integer. By Lemmas~\ref{lem:near2A} and~\ref{lem:near2Z}, $\norm{P(T_p)}_p\ge M_P(1+g|p-2|)+o(p-2)>M_P(1+\frac g2|p-2|)\ge\norm{P(S)}_p$ for all $p\ne2$ close to $2$. \end{proof} \begin{remark} The constant $\delta$ of Theorem~\ref{thm:N2} is not explicit. Item~(a) of Theorem~\ref{thm:explicit} gives an explicit range, $p\in[21/11,2)\cup(2,2.1]$, for the polynomial with $\eta=\frac7{10}$, by certified computation. Numerically, for $P$ of Theorem~\ref{thm:N2}, $g=0.1425$ (the bound above is $0.1366$). \end{remark} \section{Drury's counterexample at \texorpdfstring{$p=4$}{p=4} with exact arithmetic}\label{sec:p4exact} Drury \cite{Dr11} disproved Matsaev's conjecture at $p=4$, and so at $p=4/3$, with the polynomial \begin{equation}\label{eq:druryP} P(z)=1+2z-\tfrac{22}{5}z^2 \end{equation} and a real $2\times2$ matrix $T_{\mathrm{Dr}}$ with six-decimal entries. His proof needs two upper bounds, $\norm{T_{\mathrm{Dr}}}_4\le1$ and $\norm{P(S)}_4<6.5278$. Both are certified by a general algorithm for the norm of a finite real matrix: a Ste\v{c}kin-type estimate on patches of the unit sphere, searched depth first in exact rational arithmetic in Maple, in under $24$ hours on a PC \cite[Sec.~3--8]{Dr11}. The only step that is not exact is the final comparison $6.52773\,\norm{\nu}_{4/3}\norm{\mu}_{4/3}^{-1}<6.5278$ \cite[p.~324 and footnote~3, p.~329]{Dr11}, where $\mu$ and $\nu$ are Drury's two weight vectors (Lemma~\ref{lem:W} and Remark~\ref{rem:W}(3) below). In this section we prove Drury's result again, with the same polynomial and with a two-decimal matrix close to Drury's, so that every step is a short finite computation in exact rational arithmetic. On the matrix side, the proof needs a $3\times3$ integer Gram matrix and an integer test vector (Proposition~\ref{prop:Tprime}); every number can be checked by hand. On the shift side, the proof uses the weighted window bound of Lemma~\ref{lem:W}, which is Drury's own argument, with a $15\times15$ positive definite Gram matrix (Proposition~\ref{prop:certS}). Remark~\ref{rem:druryT} treats Drury's own matrix in the same way. No new value of $p$ is obtained here (Theorem~\ref{thm:A} covers every $p\ne2$, without numerical computation). \subsection{A weighted window bound for the shift} Let $16.53254$}{T' with norm of P(T') above 6.53254}} Let \[ T'=\begin{pmatrix}-0.27&0.83\\-0.96&-0.11\end{pmatrix}. \] This is Drury's matrix $T_{\mathrm{Dr}}$ (Remark~\ref{rem:druryT}) rounded to two decimals, $\left(\begin{smallmatrix}-0.29&0.81\\-0.96&-0.11\end{smallmatrix}\right)$, with the entries $t_{11}$ and $t_{12}$ each moved by $+0.02$; it was found by a search over the $5^4$ two-decimal matrices whose entries differ from those of the rounded matrix by at most $0.02$. \begin{proposition}\label{prop:Tprime} $\norm{T'}_4<1$, on real and on complex $\ell^4_2$. For $P$ as in \eqref{eq:druryP} and $v=(50,-53)$, $\norm{P(T')v}_4/\norm{v}_4>6.53254$. \end{proposition} \begin{proof} \emph{The norm of $T'$.} For real $(x,y)$, $100\,T'(x,y)^{\mathsf t}=(-27x+83y,\,-96x-11y)^{\mathsf t}$. So $\norm{T'(x,y)}_4\le\norm{(x,y)}_4$ is the inequality $\Phi(x,y)\ge0$ for the integer binary quartic \begin{align*} \Phi(x,y)&:=10^8(x^4+y^4)-(27x-83y)^4-(96x+11y)^4\\ &\;=14533903\,x^4-32393628\,x^3y-36823302\,x^2y^2+61241892\,xy^3+52527038\,y^4 . \end{align*} (For example, $10^8-27^4-96^4=10^8-531441-84934656=14533903$.) With $Z=(x^2,xy,y^2)$ we have $\Phi=Z^{\mathsf t}HZ$ for the integer matrix \[ H=\begin{pmatrix}14533903&-16196814&-27500000\\-16196814&18176698&30620946\\-27500000&30620946&52527038\end{pmatrix}. \] Indeed $H_{11}$, $2H_{12}$, $2H_{13}+H_{22}$, $2H_{23}$, $H_{33}$ are the five coefficients of $\Phi$; the only free entry is $H_{13}=-27500000$, and then $H_{22}=-36823302+2\cdot27500000$. The leading principal minors of $H$ are \[ m_1=14533903,\qquad m_2=1841581841698,\qquad m_3=\det H=899424407723373376, \] all positive, so $H$ is positive definite. Explicitly, with $r=H_{11}H_{23}-H_{12}H_{13}=-370526067762$, \[ \Phi=\frac1{m_1}\bigl(m_1x^2-16196814\,xy-27500000\,y^2\bigr)^2+\frac1{m_1m_2}\bigl(m_2\,xy+r\,y^2\bigr)^2+\frac{m_3}{m_2}\,y^4 , \] as one checks by expanding. If $y\ne0$ the last term is positive, and if $y=0\ne x$ the first is $m_1x^4>0$. So $\Phi>0$ off the origin, and by homogeneity and compactness of the unit sphere $\norm{T'}_4<1$ on real vectors. As $T'$ is real, Lemma~\ref{lem:RC} gives the same on complex vectors. (Numerically, $1-\norm{T'}_4^4=5.0\times10^{-4}$.) \emph{The test vector.} $T'^2=\left(\begin{smallmatrix}-0.7239&-0.3154\\0.3648&-0.7847\end{smallmatrix}\right)$, so \[ P(T')=I+2T'-\tfrac{22}5T'^2=\begin{pmatrix}3.64516&3.04776\\-3.52512&4.23268\end{pmatrix},\qquad 50000\,P(T')=\begin{pmatrix}182258&152388\\-176256&211634\end{pmatrix}. \] For $v=(50,-53)$ this gives $50000\,P(T')v=(1036336,\,-20029402)$, and $\norm{v}_4^4=50^4+53^4=14140481$. Hence \[ \frac{\norm{P(T')v}_4^4}{\norm{v}_4^4}=\frac{1036336^4+20029402^4}{50000^4\cdot14140481}=1821.0876\ldots,\qquad \frac{\norm{P(T')v}_4}{\norm{v}_4}=6.5325495\ldots . \] As integers, $(1036336^4+20029402^4)\cdot10^{20}>653254^4\cdot14140481\cdot50000^4$, which is the claim. \end{proof} \begin{theorem}[Drury \cite{Dr11}; exact certificate]\label{thm:p4exact} For $P(z)=1+2z-\frac{22}5z^2$, the matrix $T'$ is a contraction on $\ell^4_2$ and \[ \norm{P(T')}_4\;>\;6.53254\;>\;6.529466499\;\ge\;\norm{P(S)}_4. \] So Matsaev's conjecture fails at $p=4$; with the transpose $T'^{\mathsf t}$ it also fails at $p=4/3$. Every step of the proof is a finite computation in exact rational arithmetic. \end{theorem} \begin{proof} Propositions~\ref{prop:Tprime} and~\ref{prop:certS}; $\ell^4_2$ is the $L^4$ space of two atoms. For $p=4/3$, by Lemma~\ref{lem:dual}, $\norm{T'^{\mathsf t}}_{4/3}=\norm{T'}_4<1$ and $\norm{P(T'^{\mathsf t})}_{4/3}=\norm{P(T')}_4>\norm{P(S)}_4=\norm{P(S)}_{4/3}$. \end{proof} The relative margin is small, about $0.047\%$. The test vector also beats Drury's own bound $\norm{P(S)}_4<6.5278$, so Proposition~\ref{prop:Tprime} together with \cite[Prop.~2]{Dr11} is a second proof. \begin{remark}[Drury's own matrix]\label{rem:druryT} Drury's matrix is $T_{\mathrm{Dr}}=\left(\begin{smallmatrix}-0.292187&0.812218\\-0.958489&-0.113428\end{smallmatrix}\right)$ \cite[Prop.~1]{Dr11}, with exactly these entries (they are also the entries checked in his verification worksheet\footnote{\url{https://www.math.mcgill.ca/drury/research/matsaev/matsaev.html}}); he shows $\norm{P(T_{\mathrm{Dr}})}_4>6.5381$ with the test vector $\xi=(0.821725,-0.858838)$. The same two exact checks work for $T_{\mathrm{Dr}}$. Write the unscaled quartic $x^4+y^4-(t_{11}x+t_{12}y)^4-(t_{21}x+t_{22}y)^4$, with the entries $t_{jk}$ of $T_{\mathrm{Dr}}$, as $Z^{\mathsf t}H_{\mathrm{Dr}}Z$ with the free entry $(H_{\mathrm{Dr}})_{13}=-7243/25000$ (so $(H_{\mathrm{Dr}})_{22}$ is the coefficient of $x^2y^2$ minus $2(H_{\mathrm{Dr}})_{13}$). The three leading principal minors of $H_{\mathrm{Dr}}$ are exact rationals (with denominators of $24$, $47$ and $69$ digits), and are positive: $1.486996\times10^{-1}$, $1.034277\times10^{-5}$ and $9.811364\times10^{-10}$. So $\norm{T_{\mathrm{Dr}}}_4<1$. Numerically $\norm{T_{\mathrm{Dr}}}_4=0.999990231$, so all six decimals are needed. Already $v=(1,-1)$ gives $\norm{P(T_{\mathrm{Dr}})v}_4/\norm{v}_4=6.533390>C_S$ (an exact comparison); numerically $\norm{P(T_{\mathrm{Dr}})}_4=6.538155009$, a margin of about $0.13\%$ over $C_S$. The program \path{e_p4_exact/programs/tside.py} repeats these exact checks for $T_{\mathrm{Dr}}$, and the Gram certificate of Proposition~\ref{prop:Tprime} for $T'$ (log \path{e_p4_exact/logs/tside_v020.log}). \end{remark} \begin{remark}[Drury's guess]\label{rem:druryguess} Drury wrote: ``We think that $\norm{P(S)}_{\ell^4\to\ell^4}\approx6.4548$ but this is a guess'' \cite[p.~324]{Dr11}. Theorem~\ref{thm:shiftnorms} bears this out: $\norm{P(S)}_4\le6.454826$, which is within $2\times10^{-6}$ of the numerically computed $\norm{P(J)}_4=6.454824957$, a lower bound for $\norm{P(S)}_4$ by Lemma~\ref{lem:J}. So the margin of Drury's matrix over the shift is about $1.3\%$, not $0.13\%$. The bound $6.529466499$ of Proposition~\ref{prop:certS} is about $1.2\%$ above $\norm{P(J)}_4$. \end{remark} \section{Shift norms of Drury's polynomial}\label{sec:shiftnorms} In this section $P(z)=1+2z-\frac{22}5z^2$ is Drury's polynomial \eqref{eq:druryP}. Since $J^2=-I$, $P(J)=\frac{27}5I+2J$, and Lemma~\ref{lem:J} gives the lower bound $\norm{P(S)}_p\ge\norm{P(J)}_p$. We give upper bounds at $p=3$ and at the even $p=4,6,\dots,20$, by window certificates in the sense of Lemma~\ref{lem:bellman}. Section~\ref{sec:explicitfull} uses these bounds as upper bounds for $\norm{P(S)}_p$ in item~(c) of Theorem~\ref{thm:explicit}. \begin{theorem}[computer-assisted]\label{thm:shiftnorms} Let $P(z)=1+2z-\frac{22}5z^2$. \begin{enumerate}[leftmargin=2em,label=(\alph*)] \item For $p\in\{4,6,\dots,20\}$, $\norm{P(S)}_p\le C_p$, with $C_p$ as in Table~\ref{tab:tisos}. \item $\norm{P(S)}_3\le C_3:=248143/40000=6.203575$. \end{enumerate} By Lemma~\ref{lem:dual} the same bounds hold on $\ell^{p'}$. \end{theorem} The proof is a computer check of explicit certificates, in exact rational arithmetic (for (b) also in ball arithmetic); the two subsections below describe the certificates and the programs. \begin{table}[ht] \centering \renewcommand{\arraystretch}{1.15} \begin{tabular}{|r|l|l|r|l|} \hline $p$ & $C_p$ & $C_p$ (decimal) & size of $\mathcal Q$ & $\norm{P(J)}_p$ (numerical)\\ \hline $4$ & $3227413/500000$ & $6.454826$ & $6$ & $6.454824957$\\ $6$ & $6737251/1000000$ & $6.737251$ & $10$ & $6.737249200$\\ $8$ & $13783/2000$ & $6.891500$ & $15$ & $6.891498511$\\ $10$ & $6988033/1000000$ & $6.988033$ & $21$ & $6.988031138$\\ $12$ & $1410781/200000$ & $7.053905$ & $28$ & $7.053903233$\\ $14$ & $3550829/500000$ & $7.101658$ & $36$ & $7.101656124$\\ $16$ & $3568923/500000$ & $7.137846$ & $45$ & $7.137844301$\\ $18$ & $7166211/1000000$ & $7.166211$ & $55$ & $7.166209723$\\ $20$ & $7189041/1000000$ & $7.189041$ & $66$ & $7.189039566$\\ \hline \end{tabular} \caption{The bounds $\norm{P(S)}_p\le C_p$ for $P(z)=1+2z-\frac{22}5z^2$, proved by the certificates of Definition~\ref{def:tisos}; the size of their Gram matrix $\mathcal Q$ (the number of monomials of degree $p/2$ in three variables), and the lower bound $\norm{P(J)}_p$ of Lemma~\ref{lem:J}, computed in floating point.}\label{tab:tisos} \end{table} In every row, $C_p$ is within $2\times10^{-6}$ of the numerically computed $\norm{P(J)}_p$, which is a lower bound for $\norm{P(S)}_p$ by Lemma~\ref{lem:J}; so numerically $\norm{P(S)}_p=\norm{P(J)}_p$ at these $p$. At $p=3$ the numerically computed value is $\norm{P(J)}_3=6.202999138$, and the bound of part (b) is within $0.0093\%$ of it. \begin{remark}\label{rem:p2} The equality $\norm{P(S)}_p=\norm{P(J)}_p$ cannot hold for every $p$. At $p=2$, $\norm{P(S)}_2=\max_{|z|=1}|P(z)|=\sqrt{19683/550}=5.98224646\ldots$, while $\norm{P(J)}_2=|\frac{27}5+2i|=5.758472\ldots$. \end{remark} \subsection{Sum-of-squares window certificates at even \texorpdfstring{$p$}{p}} Let $p=2m$ be even and let $P(z)=c_0+c_1z+c_2z^2$ be real. As in Lemma~\ref{lem:bellman} we use windows $(u,v,w)=(x_n,x_{n-1},x_{n-2})$ and $y=c_0u+c_1v+c_2w$. Let $Z(u,v,w)$ be the vector of the $\binom{m+2}2$ monomials of degree $m$ in $u,v,w$. \begin{definition}\label{def:tisos} A \emph{TISOS certificate} (translation-invariant sum of squares, window length $3$) for the bound $\norm{P(S)}_p\le C$ is a pair $(G,\mathcal Q)$, where $G(s,t)$ is a real homogeneous polynomial of degree $p$ in two variables and $\mathcal Q$ is a real symmetric positive semidefinite matrix indexed by the monomials in $Z$, such that, as polynomials in $(u,v,w)$, \begin{equation}\label{eq:tisos} C^pu^p-y^p+G(v,w)-G(u,v)=Z(u,v,w)^{\mathsf t}\,\mathcal Q\,Z(u,v,w). \end{equation} \end{definition} Write $\mathcal Q=\sum_i\lambda_iq_iq_i^{\mathsf t}$ with $\lambda_i\ge0$ (spectral theorem). Then the right side of \eqref{eq:tisos} is $\sum_i\lambda_i\,(q_i^{\mathsf t}Z)^2$, a sum of squares of polynomials. The name refers to the sum over all windows of a finitely supported real sequence $x$: the $G$ terms telescope, as in the proof of Lemma~\ref{lem:bellman}, and \[ C^p\norm{x}_p^p-\norm{P(S)x}_p^p=\sum_nZ(x_n,x_{n-1},x_{n-2})^{\mathsf t}\,\mathcal Q\,Z(x_n,x_{n-1},x_{n-2}), \] a single sum of squares whose terms are translates of each other. \begin{proposition}\label{prop:tisos} If a TISOS certificate for $\norm{P(S)}_p\le C$ exists, then $\norm{P(S)}_p\le C$ (on complex $\ell^p$). \end{proposition} \begin{proof} As $p$ is even, $u^p=|u|^p$ and $y^p=|y|^p$, so the left side of \eqref{eq:tisos} is the function $F$ of \eqref{eq:bellman} for this $P$ and this $G$. By \eqref{eq:tisos}, $F\ge0$ on $\R^3$, as $\mathcal Q$ is positive semidefinite. $G$ is continuous, and $G(0,0)=0$ since $G$ is homogeneous of degree $p>0$. Lemma~\ref{lem:bellman}, whose proof applies verbatim to every real quadratic $P$, gives $\norm{P(S)}_p\le C$. \end{proof} \emph{How the certificates are stored.} The certificate files give $C$ and $\mathcal Q$, but not $G$; they record \eqref{eq:tisos} in a folded form, which determines $G$. Two monomials $u^{\alpha_1}v^{\alpha_2}w^{\alpha_3}$ of the same degree are in the same \emph{translation class} if their exponent vectors $(\alpha_1,\alpha_2,\alpha_3)$ agree after the zero entries at both ends are deleted. For example, $u^3v$ and $v^3w$ are in the same class, but $u^3v$ and $uv^3$ are not. The substitution $(u,v)\mapsto(v,w)$, which moves the window by one index, maps $u^av^b$ to $v^aw^b$. So the classes are $\{u^p,v^p,w^p\}$; the pairs $\{u^av^b,\,v^aw^b\}$ with $a,b\ge1$; and the single monomials $u^av^bw^c$ with $a,c\ge1$. \begin{lemma}\label{lem:fold} Let $E(u,v,w)$ be a homogeneous polynomial of degree $p\ge1$. Suppose that for each translation class, the coefficients of $E$ on the monomials of that class add up to $0$. Then $E=G(v,w)-G(u,v)$ for a homogeneous polynomial $G$ of degree $p$. \end{lemma} \begin{proof} We treat the classes one at a time and add the results. A single monomial $u^av^bw^c$ with $a,c\ge1$ has coefficient $0$. On a pair $\{u^av^b,v^aw^b\}$ with $a,b\ge1$, $E$ has the form $e\,u^av^b-e\,v^aw^b$, which is $G(v,w)-G(u,v)$ for $G(s,t)=-e\,s^at^b$. On $\{u^p,v^p,w^p\}$, $E$ has the form $e_uu^p+e_vv^p+e_ww^p$ with $e_u+e_v+e_w=0$, which is $G(v,w)-G(u,v)$ for $G(s,t)=-e_us^p+e_wt^p$. \end{proof} Apply Lemma~\ref{lem:fold} to $E=Z^{\mathsf t}\mathcal QZ-(C^pu^p-y^p)$. If, for each translation class, the coefficients of $Z^{\mathsf t}\mathcal QZ$ and of $C^pu^p-y^p$ have the same sum over the class (the \emph{folded identity}), then \eqref{eq:tisos} holds for some $G$. (Conversely, \eqref{eq:tisos} implies the folded identity, since every $G(v,w)-G(u,v)$ has class sums $0$; this is not needed.) \emph{The computer check.} For $P(z)=1+2z-\frac{22}5z^2$ and $p=4,6,\dots,20$, the certificates are the ancillary files \texttt{cert\_p$p$\_W3\_C$C_p$.txt} (for example \texttt{cert\_p4\_W3\_C6.454826.txt}) in the folder \path{c_drury_28_20/certificates/}. Each gives $p$, the window length $3$, the coefficients $1,2,-\frac{22}5$, the number $C_p$, the list of the monomials of degree $m$ and the upper triangle of $\mathcal Q$. A monomial is listed by the multiset of its window positions $0,1,2$; position $i$ holds $x_{n+i}$ in the window $(x_n,x_{n+1},x_{n+2})$, so positions $0,1,2$ are $w,v,u$ in our notation, and $y=c_0u+c_1v+c_2w$. (The identity holds in this orientation, not in the mirrored one; the two orientations bound the same norm, by the reflection $n\mapsto-n$ in the proof of Lemma~\ref{lem:dual}.) Each certificate is checked in exact rational arithmetic by two programs written independently of each other, and once more by a third program; they check that \begin{enumerate}[leftmargin=2em,label=(\roman*)] \item the listed monomials are all the monomials of degree $m$ in three variables, each once; \item the folded identity holds, class by class; \item $\mathcal Q$ is positive definite (all leading principal minors, or all pivots of symmetric elimination, are positive). \end{enumerate} By Lemma~\ref{lem:fold} and Proposition~\ref{prop:tisos} this proves Theorem~\ref{thm:shiftnorms}(a). The programs and their logs, all in \path{c_drury_28_20/}, are: \begin{itemize}[leftmargin=2em]\raggedright \item \texttt{tisos\_verify.py}: leading principal minors by fraction-free (Bareiss) elimination; the coefficients of $C^pu^p-y^p$ by multiplying out dense polynomials. Log \texttt{logs/tisos\_bellman\_verify\_v010.log}. \item \texttt{verify\_tisos.py}: $LDL^{\mathsf t}$ elimination; it also checks the summed identity, with no folding, on $40$ random integer sequences. Logs \texttt{logs/verify\_tisos\_1939.log}, \texttt{logs/verify\_tisos\_high1\_2017.log} and \texttt{logs/verify\_tisos\_high2\_2017.log}. \item \texttt{tisos\_read.py}: its own parser and elimination; it tries both orientations of the window. Logs \texttt{logs\_rerun/tisos\_read\_cert\_p$p$\_W3\_C$C_p$\_run.log}. \end{itemize} As controls, lowering $C$ by $10^{-6}$, or changing a single entry of $\mathcal Q$ by $10^{-6}$, breaks (ii). The matrices $\mathcal Q$ were found by a semidefinite program in floating point and then made exact; how they were found plays no role in the check. \subsection{A Bellman certificate at \texorpdfstring{$p=3$}{p=3}} For odd $p$, $|t|^p$ is not a polynomial and \eqref{eq:tisos} is not available. Instead we take an explicit $G$ and check $F\ge0$ in Lemma~\ref{lem:bellman} by a certified branch-and-bound computation. We give a separate bound at $p=3$ because interpolation (Riesz--Thorin) between the bounds at $p=2$ and $p=4$ gives only $\norm{P(S)}_3\le6.2933$; Section~\ref{sec:explicitfull} uses it on $[2.8,3.6]$ (Table~\ref{tab:runsc}). \begin{proposition}\label{prop:bell3} Let $C_3=248143/40000$, let $h_3(\sigma_1,\sigma_2)=\sum_{j=1}^8\omega_j\,|\alpha_j\sigma_1+\beta_j\sigma_2|^3$ with the rational numbers of Table~\ref{tab:bell3}, and $y=u+2v-\frac{22}5w$. Then \begin{equation}\label{eq:bell3} F_3(u,v,w):=C_3^3\,|w|^3-|y|^3+h_3(w,v)-h_3(v,u)\;\ge\;0\qquad\text{for all }(u,v,w)\in\R^3, \end{equation} and hence $\norm{P(S)}_3\le C_3=6.203575$. \end{proposition} \begin{table}[ht] \centering \renewcommand{\arraystretch}{1.15} \begin{tabular}{|c|r|r|r|} \hline $j$ & $\alpha_j$ & $\beta_j$ & $\omega_j$\\ \hline $1$ & $1$ & $0$ & $-5479839/125000$\\ $2$ & $231/250$ & $383/1000$ & $-1674041/125000$\\ $3$ & $707/1000$ & $707/1000$ & $-10856541/100000$\\ $4$ & $383/1000$ & $231/250$ & $55056591/1000000$\\ $5$ & $0$ & $1$ & $19116003/500000$\\ $6$ & $-383/1000$ & $231/250$ & $-113888233/1000000$\\ $7$ & $-707/1000$ & $707/1000$ & $45200287/500000$\\ $8$ & $-231/250$ & $383/1000$ & $-372648/15625$\\ \hline \end{tabular} \caption{The data of $h_3(\sigma_1,\sigma_2)=\sum_j\omega_j|\alpha_j\sigma_1+\beta_j\sigma_2|^3$ in Proposition~\ref{prop:bell3} (the three columns of the file \texttt{bell\_p3\_M8\_m4.txt}).}\label{tab:bell3} \end{table} \begin{proof}[Proof of the implication] Put $G(s,t)=h_3(t,s)+C_3^3\bigl(|s|^3+|t|^3\bigr)$. It is continuous, $G(0,0)=0$, and $G(v,w)-G(u,v)=h_3(w,v)-h_3(v,u)+C_3^3\bigl(|w|^3-|u|^3\bigr)$. Hence \[ C_3^3|u|^3-|y|^3+G(v,w)-G(u,v)=F_3(u,v,w)\ge0, \] and Lemma~\ref{lem:bellman} (for $p=3$ and the real quadratic $P$) gives $\norm{P(S)}_3\le C_3$. The weights $\omega_j$ have both signs; no step needs $h_3\ge0$. \end{proof} \emph{The computer check of \eqref{eq:bell3}.} $F_3$ is even and $F_3(t\xi)=|t|^3F_3(\xi)$ for real $t$. Every $\xi\ne0$ in $\R^3$ is $\pm\max_i|\xi_i|$ times a point of one of the three faces $\{\xi_k=1,\ |\xi_i|\le1\ (i\ne k)\}$ of the cube $[-1,1]^3$. So it suffices that $F_3>0$ on these faces. Three programs, written independently of each other, check this. Each covers the faces by boxes and accepts a box only with a proved positive lower bound for $F_3$ on it. \begin{enumerate}[leftmargin=2em,label=(\roman*)]\raggedright \item \texttt{bellman\_verify.py}, in ball arithmetic (python-flint \texttt{arb}): adaptive bisection; a box is accepted when a lower bound for $F_3$ at its centre exceeds $\sum_i r_i\sup_{\text{box}}|\partial F_3/\partial\xi_i|$, with $r_i$ the half-widths (mean value theorem; $F_3$ is $C^1$). Log \texttt{logs/tisos\_bellman\_verify\_v010.log}. \item \texttt{verify\_bellman\_p3.py}, in exact rational arithmetic: the same kind of branch and bound, with $F_3$ at the centre computed exactly ($p=3$ is an integer) and exact interval bounds for the derivatives. Log \texttt{logs/verify\_bellman\_p3\_0031.log}. In its control run (\texttt{logs/verify\_bellman\_p3\_controlgrid\_0031.log}), $C_3$ is lowered to $6.2029<\norm{P(J)}_3$, and an exact search on a $200\times200$ grid of each face finds a point with $F_3<0$. \item \texttt{bellman\_bernstein.py}, in exact rational arithmetic, by a different method: on a box where a linear form $\alpha\sigma_1+\beta\sigma_2$ has constant sign (decided at the corners), the term $\omega|\alpha\sigma_1+\beta\sigma_2|^3$ is a polynomial; the sum of these terms is written in the tensor Bernstein basis of degree $(3,3)$ on the box, and its least coefficient is a lower bound; a term $\omega|\alpha\sigma_1+\beta\sigma_2|^3$ whose form changes sign on the box contributes the lower bound $\min(0,\omega)\max_{\text{corners}}|\alpha\sigma_1+\beta\sigma_2|^3$. The three faces need $28$, $70$ and $82$ boxes. Log \texttt{logs/bellman\_bernstein\_p3\_v020.log}. Its control runs (\texttt{logs/bellman\_bernstein\_p3\_controls\_v020.log}) fail: for example, with $C=6.2029<\norm{P(J)}_3$ it finds rational points where $F_3<0$. \end{enumerate} All files are in \path{c_drury_28_20/}. In the certificate file \path{certificates/bell_p3_M8_m4.txt}, the inequality is written for the window $(x_n,x_{n+1},x_{n+2})=(w,v,u)$, as $C^3|u_0|^3-|c_2u_0+c_1u_1+c_0u_2|^3+h(u_0,u_1)-h(u_1,u_2)$, which is \eqref{eq:bell3}. In floating point the minimum of $F_3$ on the faces is about $0.04$, that is about $1.7\times10^{-4}$ of $C_3^3=238.74$. \section{Explicit counterexamples for every \texorpdfstring{$p\in[40/39,40]$, $p\ne2$}{p in [40/39,40], p != 2}}\label{sec:explicitfull}\label{sec:explicit} Theorem~\ref{thm:A} is an existence statement. For each $p\ne2$ it needs a small coefficient $a$, and the admissible range $00$, the form used in Lemma~\ref{lem:bellman}. \begin{theorem}\label{thm:explicit} Let $p\in[40/39,40]$, $p\neq2$, and let $q=\max(p,p')$, so that $q\in(2,40]$. Let $(P,T)$ be given by the first item that applies: \begin{enumerate}[leftmargin=2em,label=(\alph*)] \item $q\in(2,2.1]$: $P=P_{7/10}$ and $T=T_q=R_\varphi/U(q)$, where $\varphi\in(\pi/2,\pi)$, $\cos\varphi=-\frac7{34}$, and $U(q)\ge\norm{R_\varphi}_q$ is the explicit bound of Definition~\ref{def:Ua}; \item $q\in[2.1,2.8]$: $P=P_A$, and $T=T_i/\varrho_i$, where $[p_{\mathrm{lo}},p_{\mathrm{hi}}]$ is a piece of item (b) (Definition~\ref{def:piece}; there are $15$) with $p_{\mathrm{lo}}\le q\le p_{\mathrm{hi}}$, and $T_i$, $\varrho_i$ are its matrix and its norm bound; \item $q\in[2.8,20]$: $P=P_{\mathrm{Dr}}$, and $T=T_i/\varrho_i$, where $[p_{\mathrm{lo}},p_{\mathrm{hi}}]$ is a piece of item (c) (there are $445$) with $p_{\mathrm{lo}}\le q\le p_{\mathrm{hi}}$, and $T_i$, $\varrho_i$ are its matrix and its norm bound; \item $q\in[20,40]$: $P=P_D$, and $T=T_i/\varrho_i$, where $[p_{\mathrm{lo}},p_{\mathrm{hi}}]$ is a piece of item (d) (there are $100$) with $p_{\mathrm{lo}}\le q\le p_{\mathrm{hi}}$, and $T_i$, $\varrho_i$ are its matrix and its norm bound. \end{enumerate} In (b)--(d), $T_i$ is a real matrix with six-decimal entries and $\varrho_i$ is a number with $8$ decimals; both are printed on the line of the piece in the log named in Section~\ref{ssec:item-b}, \ref{ssec:item-c} or~\ref{ssec:item-d}. Put $T_*=T$ if $p>2$ and $T_*=T^{\mathsf t}$ if $p<2$. Then \[ \norm{T_*}_p\le1\qquad\text{and}\qquad\norm{P(T_*)}_p\;>\;\norm{P(S)}_p, \] so Matsaev's inequality \eqref{eq:matsaev} fails at $p$, for a real quadratic and a real $2\times2$ contraction. \end{theorem} For every piece of (b)--(d), the programs \path{verify_R2.py} and \path{R2_read.py} read $p_{\mathrm{lo}}$, $p_{\mathrm{hi}}$, $T_i$ and $\varrho_i$ from the log and prove that $\varrho_i\ge\max(\norm{T_i}_{p_{\mathrm{lo}}},\norm{T_i}_{p_{\mathrm{hi}}})$ and that $L>\widehat C(p_{\mathrm{hi}})$ for $T_i/\varrho_i$ (Definition~\ref{def:piece}, Lemma~\ref{lem:piece}, Section~\ref{ssec:explicit-scheme}). For $p<2$ the four ranges are $[21/11,2)$, $[14/9,21/11]$, $[20/19,14/9]$ and $[40/39,20/19]$, the duals of $(2,2.1]$, $[2.1,2.8]$, $[2.8,20]$ and $[20,40]$. The relative margins (Definition~\ref{def:piece}) are between $0.2900\%$ and $1.6958\%$ in (b), between $0.0139\%$ and $1.9101\%$ in (c), and between $0.0228\%$ and $0.0994\%$ in (d); in (a) the relative margin exceeds $0.0329\%$ for $p\in[400/199,2.1]$ and tends to $0$ as $p\to2$. The rest of the section is organised as follows. Section~\ref{ssec:explicit-scheme} defines anchors, window certificates, the interpolation between anchors and the pieces, and proves the criterion that every computer check verifies. Section~\ref{ssec:worked} works out one piece of item (c) in full. Section~\ref{ssec:anchors} lists all anchors. Sections~\ref{ssec:item-a}--\ref{ssec:item-d} treat items (a)--(d). Section~\ref{ssec:twice} lists what was checked twice and by which programs, and Section~\ref{ssec:explicit-proof} assembles the proof of Theorem~\ref{thm:explicit}. \subsection{Anchors, window certificates and pieces}\label{ssec:explicit-scheme} \subsubsection*{Anchors} \begin{definition}[anchor]\label{def:anchor} Let $P$ be a real polynomial. An \emph{anchor} for $P$ is a value $p_j\in[2,\infty]$ together with a number $C(p_j)$ and a proof that $\norm{P(S)}_{p_j}\le C(p_j)$. At $p_j=2$ we take $C(p_j)=M_P:=\max_{|z|=1}|P(z)|$, at $p_j=\infty$ we take $C(p_j)=\sum|c_k|$ (Lemma~\ref{lem:exactanchors}), and at $21$, $C>0$ and finitely many rational triples $(\alpha_j,\beta_j,\omega_j)$. Put \begin{equation}\label{eq:bellcertF} \begin{gathered} h(\sigma_1,\sigma_2)=\sum_j\omega_j\,|\alpha_j\sigma_1+\beta_j\sigma_2|^p,\\ F(u_0,u_1,u_2)=C^p|u_0|^p-|c_2u_0+c_1u_1+c_0u_2|^p+h(u_0,u_1)-h(u_1,u_2). \end{gathered} \end{equation} The certificate is \emph{valid} if $F\ge0$ on $\R^3$. \end{definition} The certificate files list $p$, the coefficients, $C$, the number of terms and a line $\alpha_j\ \beta_j\ \omega_j$ for each term; the anchors of this section use $8$ or $12$ terms (Table~\ref{tab:anchors}). Here $(u_0,u_1,u_2)=(x_{n-2},x_{n-1},x_n)$, so $c_2u_0+c_1u_1+c_0u_2=(P(S)x)_n$. \begin{lemma}\label{lem:bellcert} If a Bellman certificate for $P=1+az-Bz^2$ is valid, then $\norm{P(S)}_p\le C$. \end{lemma} \begin{proof} This is the proof of the implication in Proposition~\ref{prop:bell3}, with $3$, $C_3$ and $h_3$ replaced by $p$, $C$ and $h$. In the window $(u,v,w)=(x_n,x_{n-1},x_{n-2})=(u_2,u_1,u_0)$ of Lemma~\ref{lem:bellman}, the continuous function $G(\sigma_1,\sigma_2)=h(\sigma_2,\sigma_1)+C^p(|\sigma_1|^p+|\sigma_2|^p)$, which has $G(0,0)=0$, gives $C^p|u|^p+G(v,w)-G(u,v)-|u+av-Bw|^p=F(w,v,u)\ge0$. This is \eqref{eq:bellman} with $a=c_1$ and $B=-c_2$, and Lemma~\ref{lem:bellman} gives $\norm{P(S)}_p\le C$ (its proof uses only $G(0,0)=0$; all anchors of this section have $p>2$, as in that lemma). \end{proof} \begin{lemma}[reduction to three faces]\label{lem:faces} The function $F$ of \eqref{eq:bellcertF} satisfies $F(-u)=F(u)$ and $F(tu)=t^pF(u)$ for $t>0$. Hence $F\ge0$ on $\R^3$ if and only if $F\ge0$ on the three faces $\mathcal C_k=\{u:\ u_k=1,\ |u_i|\le1\ (i\ne k)\}$, $k=0,1,2$, of the cube $[-1,1]^3$. \end{lemma} \begin{proof} Every term of $F$ is a multiple of $|\ell(u)|^p$ with $\ell$ linear, which gives both identities. If $u\ne0$, let $k$ be an index with $|u_k|=\max_i|u_i|$; then $u=\pm|u_k|\,u'$ with $u'\in\mathcal C_k$, and $F(u)=|u_k|^pF(u')$. \end{proof} \emph{How a Bellman certificate is checked.} Each face is cut into boxes, and on each box a rigorous lower bound for $F$ is computed; a box whose bound is not positive is cut into smaller boxes, and the check fails if the subdivision does not terminate. All data are rational, the box corners are exact rationals, and the evaluation uses ball arithmetic (the Arb library \cite{Jo17} through python-flint) or exact rational arithmetic. Three kinds of lower bound are used by the different programs: \begin{enumerate}[leftmargin=2em,label=(\roman*)] \item first order (mean value theorem; $F$ is $C^1$ for $p>1$): $F\ge F(u^\circ)-\sum_i\sup_{\mathrm{box}}|\partial_iF|\,r_i$, with $u^\circ$ the centre and $r_i$ the half-widths; the derivative of $|\ell|^p$ in $\ell$, $p\,\sgn(\ell)|\ell|^{p-1}$, is increasing, so it is bounded by its values at the extreme values of $\ell$, which are taken at box corners (\path{bellman_check.py}, \path{bellman_verify.py}; for integer $p$ in exact rationals, \path{verify_bellman_p3.py}, \path{verify_bellman_evenp.py}); \item second order (for $p\ge2$): each term $\omega|\ell|^p$ is replaced by a quadratic lower bound on the box. If $\ell$ keeps a strict sign on the box and $\min|\ell|\ge2\max|\delta|$, where $\delta=|\ell|-|\ell(u^\circ)|$, $u^\circ$ is the centre and $\min$, $\max$ are taken over the box, the bound is the Taylor polynomial of degree $2$ of $|\ell|^p$ at $u^\circ$ with the remainder bound $p(p-1)(p-2)\max\bigl(\min|\ell|^{p-3},\max|\ell|^{p-3}\bigr)\max|\delta|^3/6$. Otherwise it uses convexity: $|\ell|^p$ is at least its tangent at $\ell(u^\circ)$ (used if $\omega>0$) and at most $M^{p-2}\ell^2$ with $M=\max|\ell|$ (used if $\omega<0$). The resulting quadratic is bounded below through its coefficients in the Bernstein basis (\path{bellman_taylor.py}); \item for integer $p$: on a box where $\ell$ has constant sign, $|\ell|^p=\pm\ell^p$ is a polynomial, expanded exactly in the Bernstein basis of the box (\path{bellman_bernstein.py}). \end{enumerate} \subsubsection*{Translation-invariant sum-of-squares certificates} The anchors at $p=4,6,\dots,20$ of item (c) and at $p=20,22,\dots,32$ of item (d) are TISOS certificates in the sense of Definition~\ref{def:tisos}: for $p=2m$, a positive semidefinite matrix $\mathcal Q$, indexed by the $(m+1)(m+2)/2$ monomials of degree $m$ in three variables, for which \eqref{eq:tisos} holds. By Proposition~\ref{prop:tisos} such a certificate proves $\norm{P(S)}_p\le C$. The files store \eqref{eq:tisos} in the folded form of Lemma~\ref{lem:fold}, and in their names \path{W3} records the window length $3$. The files of item (d) also have a line \path{tau}, which gives a rational number $\gamma$ with $0<\gamma\le C^p$; in these files $\gamma$ takes the place of $C^p$ in \eqref{eq:tisos}, so each is a TISOS certificate for the bound $\gamma^{1/p}$, and $\gamma^{1/p}\le C$. \emph{How a TISOS certificate is checked} (all steps in exact rational arithmetic; for item (c) the programs are listed in Section~\ref{sec:shiftnorms}). (V1) The listed monomials are exactly all monomials of degree $m$ in three variables, each once. (V2) The folded identity holds, class by class (Lemma~\ref{lem:fold}). (V3) $\mathcal Q$ is positive definite: all leading principal minors are positive (Sylvester's criterion), computed by fraction-free elimination, or an exact $LDL^{\mathsf t}$ factorisation has positive pivots. (V4) For the files with a line \path{tau}, $0<\gamma\le C^p$. The program \path{verify_tisos.py} also verifies the summed identity directly on random integer sequences, with no folding. Two orientations of the convolution are possible in \eqref{eq:tisos} ($P(S)$ or its reflection $n\mapsto-n$); they have the same norm (proof of Lemma~\ref{lem:dual}), and the program \path{tisos_read.py} tries both. \subsubsection*{Interpolation between anchors} \begin{definition}[interpolated shift bound]\label{def:Chat} Let $2=p_0\widehat C(p_{\mathrm{hi}})$, then for every $p\in[p_{\mathrm{lo}},p_{\mathrm{hi}}]$ \[ \norm{T/\varrho}_p\le1\qquad\text{and}\qquad\norm{P(T/\varrho)}_p\;\ge\;L\;>\;\widehat C(p_{\mathrm{hi}})\;\ge\;\norm{P(S)}_p. \] In particular $\norm{P(T/\varrho)}_p/\norm{P(S)}_p-1$ is at least the relative margin, for every $p$ in the piece. \end{lemma} \begin{proof} (1) By Lemma~\ref{lem:interp}(iii), $\norm{T}_p\le\varrho$, so $T/\varrho$ is a contraction. (2) For a fixed vector $y$, $p\mapsto\norm{y}_p$ is non-increasing; hence $\norm{P(T/\varrho)v}_p\ge\norm{P(T/\varrho)v}_{p_{\mathrm{hi}}}$ and $\norm{v}_p\le\norm{v}_{p_{\mathrm{lo}}}$, so $\norm{P(T/\varrho)}_p\ge\norm{P(T/\varrho)v}_p/\norm{v}_p\ge L$. (3) By Lemma~\ref{lem:interp}(i)--(ii), $\norm{P(S)}_p\le\widehat C(p)\le\widehat C(p_{\mathrm{hi}})$. \end{proof} \begin{lemma}[$p<2$]\label{lem:dualpiece} Let $T$ be a real $2\times2$ matrix and $P$ a real polynomial. If $\norm{T}_p\le1$ and $\norm{P(T)}_p>\norm{P(S)}_p$, then $\norm{T^{\mathsf t}}_{p'}\le1$ and $\norm{P(T^{\mathsf t})}_{p'}>\norm{P(S)}_{p'}$. \end{lemma} \begin{proof} Lemma~\ref{lem:dual}: $\norm{T^{\mathsf t}}_{p'}=\norm{T}_p$, $\norm{P(T^{\mathsf t})}_{p'}=\norm{P(T)}_p$ and $\norm{P(S)}_{p'}=\norm{P(S)}_p$. \end{proof} \emph{The certifier.} The program \path{interval_certify.py} (the same file in parts b, c and d) takes a range of $p$, a step, a minimal step, the coefficients of $P$, an anchor table and a table of matrices. It walks through the range in pieces of the given step; for each piece it takes the row of the matrix table whose $p$ is nearest to the midpoint of the piece, proves a norm bound at both ends, chooses a test vector, and checks $L>\widehat C(p_{\mathrm{hi}})$. If a piece fails, it halves the step (not below the minimal step). It works in interval arithmetic (mpmath \cite{mpmath}, 30 digits); the piece ends are exact rationals. The norm bound: every real direction is $\pm(\cos\vartheta,\sin\vartheta)$ with $\vartheta\in[0,\pi]$; a floating point search gives a target value (the largest ratio found, times $1+10^{-6}$), and the interval $[0,\pi]$ is cut into cells until, on every cell, an interval enclosure of $\norm{T(\cos\vartheta,\sin\vartheta)}_p/\norm{(\cos\vartheta,\sin\vartheta)}_p$ is below the target; then $\varrho$ is the larger of the two targets. The test vector is found by a floating point search (it maximises $\norm{P(T/\varrho)x}_p/\norm{x}_p$ at the midpoint); only the final inequality is certified, so how $v$ was found does not matter. For each piece the log prints a line with the piece, $T$, $\varrho$ (as ``R$=$''), $L$ (as ``T-side$>=$''), $\widehat C(p_{\mathrm{hi}})$ (as ``shift$<=$'') and the relative margin; at the end it prints that the pieces tile the range with no gap. The printed values of $\varrho$, $L$ and $\widehat C$ are rounded to $8$, $6$ and $6$ decimals; the certifier compares before rounding, and two of the re-checking programs below prove the criterion again with the printed $\varrho$. \emph{The re-checking programs.} Every piece line of every certifier log is re-proved by two further programs: \path{read_R2_pieces.py} and \path{verify_R2.py} in item (b), \path{verify_R2.py} and \path{R2_read.py} in item (c), \path{R2_read.py} and \path{read_pieces.py} in item (d). Each reads $p_{\mathrm{lo}}$, $p_{\mathrm{hi}}$ and $T$ from the log and proves the piece criterion again, with its own certified norm bound, its own test vector and its own interpolation, in ball arithmetic \cite{Jo17} with exact rational inputs. Their norm bounds use other methods than the certifier: \path{verify_R2.py}, \path{R2_read.py} and \path{read_R2_pieces.py} write every direction as a multiple of $(1,t)$ or $(t,1)$ with $|t|\le1$, and \path{read_pieces.py} uses angle cells with a mean-value bound and bounds the matrix side on $8$ sub-pieces of each piece. The anchor values are read from the certificate files, except in \path{read_R2_pieces.py}, which reads the anchor table and takes the maximum of $\widehat C$ over the whole piece. So in this mode the contraction proved by a re-checker is $T/\varrho'$ with its own $\varrho'$. In a second mode (\texttt{-{}-printed-R} and \texttt{-{}-verify-printed-rho}), \path{verify_R2.py} and \path{R2_read.py} also read the printed $\varrho$ and prove $\varrho\ge\max(\norm{T}_{p_{\mathrm{lo}}},\norm{T}_{p_{\mathrm{hi}}})$ and the piece criterion for $T/\varrho$ itself, on every piece of (b), (c) and (d); this is the claim of Theorem~\ref{thm:explicit}. In their control runs every $\varrho$ is multiplied by $99999/100000$. Every re-check proves the piece criterion (Lemma~\ref{lem:piece}) on every piece. \subsection{A worked example}\label{ssec:worked} We go through one piece of item (c) step by step, using only the definitions above and a few lines of arithmetic. The polynomial is Drury's $P_{\mathrm{Dr}}(z)=1+2z-\frac{22}5z^2$, and the piece is $[3.59,3.6]=[359/100,18/5]$ of the run on $[3.55,4]$ (line 15 of \path{c_drury_28_20/logs/cert_left_tisos_v070.log}, broken here into three lines; $\varrho$ is printed as \texttt{R=}): \begin{verbatim} [359/100, 18/5] = [3.5900000, 3.6000000] T=-0.301279 0.818028 -0.955951 -0.131758 R=1.00000101 T-side>=6.461403 shift<=6.400525 margin=+0.9511% \end{verbatim} \noindent\emph{Step 1: the matrix and its norm bound.} The matrix is the row $p=3.60$ of \path{c_drury_28_20/tables/T_table_1836.txt}: \[ T=\begin{pmatrix}-0.301279&0.818028\\-0.955951&-0.131758\end{pmatrix}. \] In floating point, $\norm{T}_{3.59}\approx0.9998831$ and $\norm{T}_{3.6}\approx1.0000000057$, so $\norm{T}_{3.6}$ slightly exceeds $1$. The certifier proves $\norm{T}_{3.59}\le0.99988412$ and $\norm{T}_{3.6}\le\varrho=1.00000101$. By Lemma~\ref{lem:interp}(iii), $T/\varrho$ is a contraction on $\ell^p_2$ for every $p\in[3.59,3.6]$. \noindent\emph{Step 2: the test vector and the matrix side.} Take $v=(1,-1.03683)$ (the certifier's vector, rescaled so that $v_1=1$ and rounded to $6$ significant digits). Then \[ P_{\mathrm{Dr}}(T/\varrho)=I+2\tfrac T\varrho-\tfrac{22}5\bigl(\tfrac T\varrho\bigr)^2\approx\begin{pmatrix}3.438829&3.194691\\-3.733330&4.100869 \end{pmatrix}, \] $y=P_{\mathrm{Dr}}(T/\varrho)v\approx(0.126478,\,-7.985234)$, $\norm{y}_{3.6}\approx7.985235$ and $\norm{v}_{3.59}\approx1.235836$, so \[ L=\norm{y}_{3.6}/\norm{v}_{3.59}=6.4614024\ldots \] (with the certifier's own unrounded $\varrho$ and $v$ the value is $6.4614025\ldots$; the log prints $6.461403$). By step~(2) of the proof of Lemma~\ref{lem:piece}, $\norm{P_{\mathrm{Dr}}(T/\varrho)}_p\ge6.461402$ for every $p\in[3.59,3.6]$. \noindent\emph{Step 3: the shift side.} The anchor table of this run (\path{c_drury_28_20/tables/anchors_tisos_4_10.txt}) has anchors at $p=2,4,6,8,10,\infty$. The two adjacent to $p\in[3.59,3.6]$ are $p=2$, with the exact value $M_{P_{\mathrm{Dr}}}=\sqrt{19683/550}\approx5.9822465$, and $p=4$, with $C_4=3227413/500000=6.454826$ from the TISOS certificate \path{c_drury_28_20/certificates/cert_p4_W3_C6.454826.txt} (Theorem~\ref{thm:shiftnorms}(a)). For $2\le p\le4$, $\lambda=(\frac12-\frac1p)/(\frac12-\frac14)=2-\frac4p$; at $p=3.6$, $\lambda=\frac89$ and \[ \widehat C(3.6)=M_{P_{\mathrm{Dr}}}^{1/9}\,C_4^{8/9}=6.4005254\ldots \] (the log prints $6.400525$). By Lemma~\ref{lem:interp}, $\norm{P_{\mathrm{Dr}}(S)}_p\le\widehat C(p)\le6.4005254$ for every $p\in[3.59,3.6]$. \noindent\emph{Step 4: conclusion.} $6.461402>6.4005254$; the relative margin is $L/\widehat C(3.6)-1=0.9511\%$. So for every $p\in[3.59,3.6]$ the real matrix $T/\varrho$ is a contraction on $\ell^p_2$ with $\norm{P_{\mathrm{Dr}}(T/\varrho)}_p>\norm{P_{\mathrm{Dr}}(S)}_p$, and by Lemma~\ref{lem:dualpiece}, $T^{\mathsf t}/\varrho$ does the same for every $p\in[18/13,359/259]$. The same piece is also a piece of the run on $[2.8,3.6]$, which has the anchor at $p=3$ as well ($C_3=248143/40000$); there the shift-side bound is $6.369965$ and the margin $1.4354\%$. More anchors give a smaller shift-side bound. \subsection{Table of all anchors}\label{ssec:anchors} Table~\ref{tab:anchors} lists every anchor used in Theorem~\ref{thm:explicit}. A certificate file \path{X} of item (a), (b), (c) or (d) is \path{a_window_2_21/certificates/X}, \path{b_PA_21_28/certificates/X}, \path{c_drury_28_20/certificates/X} or \path{d_PD_20_40/certificates/X}. ``Bellman ($k$)'' is a Bellman certificate with $k$ terms (Definition~\ref{def:bellcert}); ``TISOS ($k$)'' is a TISOS certificate with $k$ monomials (Definition~\ref{def:tisos}). {\footnotesize \renewcommand{\arraystretch}{1.15} \begin{longtable}{|c|c|l|l|l|} \caption{All anchors: proved upper bounds $C$ for $\norm{P(S)}_p$.}\label{tab:anchors}\\ \hline item, $P$ & $p$ & $C\ge\norm{P(S)}_p$ & kind & certificate file\\ \hline\endfirsthead \hline item, $P$ & $p$ & $C\ge\norm{P(S)}_p$ & kind & certificate file\\ \hline\endhead \hline\endfoot (a) $P_{7/10}$ & $2$ & $\sqrt{19683/1700}\approx3.4026806$ & exact & ---\\ & $2.1$ & $17074961/5000000=3.4149922$ & Bellman (12) & \path{bell_p21_eta07_M12.txt}\\ \hline (b) $P_A$ & $2$ & $M_{P_A}$ (Table~\ref{tab:exactanchors}) $\approx5.2291914$ & exact & ---\\ & $2.1$ & $52358261/10000000=5.2358261$ & Bellman (12) & \path{bell_PA_p21_M12.txt}\\ & $2.2$ & $26238519/5000000=5.2477038$ & Bellman (12) & \path{bell_PA_p22_M12.txt}\\ & $2.35$ & $52745691/10000000=5.2745691$ & Bellman (12) & \path{bell_PA_p235_M12.txt}\\ & $2.5$ & $26540459/5000000=5.3080918$ & Bellman (12) & \path{bell_PA_p25_M12.txt}\\ & $2.65$ & $26800967/5000000=5.3601934$ & Bellman (12) & \path{bell_PA_p265_M12.txt}\\ & $2.8$ & $13555947/2500000=5.4223788$ & Bellman (12) & \path{bell_PA_p28_m3_M12.txt}\\ & $\infty$ & $3332879/500000=6.665758$ & $\sum|c_k|$ & ---\\ \hline (c) $P_{\mathrm{Dr}}$ & $2$ & $\sqrt{19683/550}\approx5.9822465$ & exact & ---\\ & $3$ & $248143/40000=6.203575$ & Bellman (8) & \path{bell_p3_M8_m4.txt}\\ & $4$ & $3227413/500000=6.454826$ & TISOS (6) & \path{cert_p4_W3_C6.454826.txt}\\ & $6$ & $6737251/1000000=6.737251$ & TISOS (10) & \path{cert_p6_W3_C6.737251.txt}\\ & $8$ & $13783/2000=6.891500$ & TISOS (15) & \path{cert_p8_W3_C6.891500.txt}\\ & $10$ & $6988033/1000000=6.988033$ & TISOS (21) & \path{cert_p10_W3_C6.988033.txt}\\ & $12$ & $1410781/200000=7.053905$ & TISOS (28) & \path{cert_p12_W3_C7.053905.txt}\\ & $14$ & $3550829/500000=7.101658$ & TISOS (36) & \path{cert_p14_W3_C7.101658.txt}\\ & $16$ & $3568923/500000=7.137846$ & TISOS (45) & \path{cert_p16_W3_C7.137846.txt}\\ & $18$ & $7166211/1000000=7.166211$ & TISOS (55) & \path{cert_p18_W3_C7.166211.txt}\\ & $20$ & $7189041/1000000=7.189041$ & TISOS (66) & \path{cert_p20_W3_C7.189041.txt}\\ & $\infty$ & $37/5=7.4$ & $\sum|c_k|$ & ---\\ \hline (d) $P_D$ & $2$ & $\sqrt{2673/32}\approx9.1395432$ & exact & ---\\ & $20$ & $9844747/1000000=9.844747$ & TISOS (66) & \path{cert_PD_p20_W3_C9.844747.txt}\\ & $22$ & $1971627/200000=9.858135$ & TISOS (78) & \path{cert_PD_p22_W3_C9.858135.txt}\\ & $24$ & $986943/100000=9.869430$ & TISOS (91) & \path{cert_PD_p24_W3_C9.869430.txt}\\ & $26$ & $9879081/1000000=9.879081$ & TISOS (105) & \path{cert_PD_p26_W3_C9.879081.txt}\\ & $28$ & $9887417/1000000=9.887417$ & TISOS (120) & \path{cert_PD_p28_W3_C9.887417.txt}\\ & $30$ & $4947343/500000=9.894686$ & TISOS (136) & \path{cert_PD_p30_W3_C9.894686.txt}\\ & $32$ & $4950539/500000=9.901078$ & TISOS (153) & \path{cert_PD_p32_W3_C9.901078.txt}\\ & $34$ & $49534407/5000000=9.9068814$ & Bellman (8) & \path{bell_PD_M8_m2e5v2_p34.txt}\\ & $36$ & $99119789/10000000=9.9119789$ & Bellman (8) & \path{bell_PD_M8_m2e5v2_p36.txt}\\ & $38$ & $49582631/5000000=9.9165262$ & Bellman (8) & \path{bell_PD_M8_m2e5v2_p38.txt}\\ & $40$ & $99205169/10000000=9.9205169$ & Bellman (8) & \path{bell_PD_M8_m5_p40.txt}\\ & $\infty$ & $10$ & $\sum|c_k|$ & ---\\ \end{longtable}} \noindent The TISOS certificates of item (d) have a line \path{tau}: they prove $\norm{P_D(S)}_p\le\gamma^{1/p}$, and $\gamma\le C^p$ is checked exactly (Section~\ref{ssec:explicit-scheme}). The certificates were found by programs that are not included; the checks do not depend on how a certificate was found. \subsection{The polynomial \texorpdfstring{$P_{7/10}$: $p\in(2,2.1]$}{P\_{7/10}: p in (2,2.1]} and \texorpdfstring{$[21/11,2)$}{[21/11,2)}}\label{ssec:item-a} Here $P=P_{7/10}=1+2z-\frac{17}{10}z^2$; it is the member $\eta=\frac7{10}$ of the family $1+2z-(1+\eta)z^2$ of Section~\ref{sec:near2}. Its maximum on the circle is attained at $e^{\pm i\varphi}$ with $\cos\varphi=-\frac7{34}$ (Table~\ref{tab:exactanchors}); we take $\varphi\in(\pi/2,\pi)$, so $\sin\varphi=\sqrt{1107}/34$. Theorem~\ref{thm:N2} shows, with a proof by hand, that the member $\eta=\frac15$ of this family violates \eqref{eq:matsaev} on a punctured neighbourhood of $p=2$ that is not explicit; item (a) gives an explicit one, $[21/11,2)\cup(2,2.1]$, for the member $\eta=\frac7{10}$. On this range the margin tends to $0$ as $p\to2$, since at $p=2$ both sides equal $M_P$. So here we do not use pieces with a fixed decimal matrix: the matrix is a scaled rotation that depends continuously on $p$, and the interval next to $p=2$ is handled with a bound on a derivative. For $\vartheta\in\R$ put $N_p(\vartheta)=(|\cos\vartheta|^p+|\sin\vartheta|^p)^{1/p}=\norm{(\cos\vartheta,\sin\vartheta)}_p$; it is even and has period $\pi/2$. Since $R_\theta(\cos\vartheta,\sin\vartheta)=(\cos(\vartheta-\theta),\sin(\vartheta-\theta))$, \begin{equation}\label{eq:rotnorm} \norm{R_\varphi}_p=\max_\vartheta\frac{N_p(\vartheta-\varphi)}{N_p(\vartheta)}=\max_\vartheta\frac{N_p(\vartheta+\varphi)}{N_p(\vartheta)} \end{equation} (substitute $\vartheta\mapsto-\vartheta$ and use that $N_p$ is even). \begin{definition}[the rotation bound $U(p)$]\label{def:Ua} Let $\bar u=10064142767/10^{10}$. For $p\in[2,2.1]$ put \[ \lambda(p)=\frac{\frac12-\frac1p}{\frac12-\frac{10}{21}}=21-\frac{42}p\in[0,1],\qquad U(p)=\bar u^{\,\lambda(p)},\qquad T_p=R_\varphi/U(p). \] \end{definition} The programs \path{window_cert.py} and \path{window_check.py} both prove $\norm{R_\varphi}_{2.1}\le\bar u$ in interval arithmetic from \eqref{eq:rotnorm}: the interval $\vartheta\in[0,\pi/2]$ is cut into cells until an enclosure of the quotient is below $\bar u$ on every cell. (In fact $\norm{R_\varphi}_{2.1}=1.0064142762007\ldots$, about $5\cdot10^{-10}$ below $\bar u$.) \begin{lemma}\label{lem:windowa} (i) $\norm{T_p}_p\le1$ for $p\in[2,2.1]$. (ii) For $\tau>0$, put $\alpha(\tau)=\arg P(\tau e^{i\varphi})$. For every $\vartheta_*\in\R$ and $p\ge1$, \[ \norm{P(\tau R_\varphi)}_p\ \ge\ |P(\tau e^{i\varphi})|\ \frac{N_p(\vartheta_*+\alpha(\tau))}{N_p(\vartheta_*)} . \] (iii) For $p\in[2,2.1]$, $\norm{P(S)}_p\le M_P^{1-\lambda(p)}C_{2.1}^{\lambda(p)}$ with $C_{2.1}=17074961/5000000$. \end{lemma} \begin{proof} (i) $\norm{R_\varphi}_2=1$ ($R_\varphi$ is orthogonal) and $\norm{R_\varphi}_{2.1}\le\bar u$ (proved after Definition~\ref{def:Ua}). By the convexity in Lemma~\ref{lem:interp}(iii), $\log\norm{R_\varphi}_{1/s}$ lies below its chord between $s=\frac{10}{21}$ and $s=\frac12$, so $\norm{R_\varphi}_p\le1^{1-\lambda(p)}\bar u^{\,\lambda(p)}=U(p)$. (ii) The span of $I$ and $J$ is an algebra isomorphic to $\C$ ($J^2=-I$), with $R_\theta\leftrightarrow e^{i\theta}$; so $P(\tau R_\varphi)=|P(\tau e^{i\varphi})|\,R_{\alpha(\tau)}$, as in \eqref{eq:near2exact}. For the test vector $x=(\cos \vartheta_*,-\sin \vartheta_*)$, $R_\alpha x=(\cos(\vartheta_*+\alpha),-\sin(\vartheta_*+\alpha))$, and $\norm{x}_p=N_p(\vartheta_*)$, $\norm{R_\alpha x}_p=N_p(\vartheta_*+\alpha)$. (iii) Lemma~\ref{lem:interp}(i) with the anchors $p=2$ ($M_P$, Lemma~\ref{lem:exactanchors}) and $p=2.1$ (Table~\ref{tab:anchors}); between these two anchors the interpolation parameter is the same $\lambda(p)$. \end{proof} \emph{The function $\Theta$.} With $\vartheta_*=157/200$, $s=1/p$, $\lambda=21-42s$, $\tau=\bar u^{\,-\lambda}=1/U(p)$ and $\alpha=\alpha(\tau)$, put \[ \Theta(s)=\log|P(\tau e^{i\varphi})|+\log N_p(\vartheta_*+\alpha)-\log N_p(\vartheta_*)-\bigl[(1-\lambda)\log M_P+\lambda\log C_{2.1}\bigr]. \] By Lemma~\ref{lem:windowa}, $\Theta(s)>0$ implies that $T_p$ is a contraction with $\norm{P(T_p)}_p>\norm{P(S)}_p$; the relative margin at $p$ is at least $e^{\Theta(s)}-1>\Theta(s)$. At $s=\frac12$: $\lambda=0$, $\tau=1$, $|P(e^{i\varphi})|=M_P$ and $N_2\equiv1$, so $\Theta(\frac12)=0$ exactly. \emph{The checks.} (W1) The interval $s\in[10/21,199/400]$, that is $p\in[400/199,2.1]$, is cut into $400$ equal pieces, and on each piece an interval enclosure of $\Theta$ (in Arb ball arithmetic, with the whole piece as a single ball) has a positive lower end. The smallest lower end is $0.000330007$ (\path{window_cert.py}) and $0.000329987$ (\path{window_check.py}). (W2) On the last piece $s\in[199/400,1/2]$, that is $p\in[2,400/199]$, interval values of $\Theta$ cannot be positive, because $\Theta(\frac12)=0$. Instead the programs enclose the derivative $\Theta'(s)$ on the whole piece, by forward-mode differentiation (every quantity is carried together with its derivative in $s$, both as balls). The enclosure is negative: $\Theta'\in-0.145090\pm0.0151$ (\path{window_cert.py}, a single ball for the whole piece), and the upper ends on $8$ sub-pieces are between $-0.1434$ and $-0.1424$ (\path{window_check.py}). On this piece the bases $\cos \vartheta_*$, $\sin \vartheta_*$, $\cos(\vartheta_*+\alpha)$, $\sin(\vartheta_*+\alpha)$ of the $p$-th powers stay away from $0$ (the programs assert it), so $\Theta$ is smooth there and \[ \Theta(s)=\Theta(\tfrac12)-\int_s^{1/2}\Theta'(\sigma)\,d\sigma=-\int_s^{1/2}\Theta'(\sigma)\,d\sigma>0\qquad(199/400\le s<1/2). \] Together, $\Theta>0$ on $[10/21,1/2)$, that is, for every $p\in(2,2.1]$. For $p\in[21/11,2)$ the matrix is $T_{p'}^{\mathsf t}=R_{-\varphi}/U(p')$ (Lemma~\ref{lem:dualpiece}). \emph{Files} (in \path{a_window_2_21/}; commands are run from \path{programs/}). \begin{itemize}[leftmargin=1.5em]\raggedright \item Anchor \path{certificates/bell_p21_eta07_M12.txt}: checked by \path{bellman_check.py} (\path{logs/bellman_check_p21.log}) and by \path{bellman_taylor.py} (\path{logs/bellman_taylor_p21_v020.log}); the controls of the latter are in \path{logs/bellman_taylor_p21_controls_v020.log}. \item The range $(2,2.1]$: {\small \begin{verbatim} window_cert.py 7/10 21/10 17074961/5000000 785/1000 1/400 400 10064142767/10000000000 \end{verbatim}} \noindent(arguments: $\eta$, $p_1=2.1$, $C_{2.1}$, $\vartheta_*=785/1000=157/200$, the width $1/400$ of the last piece in $s$, the number of pieces, $\bar u$), log \path{logs/window_cert_p21_ubar_v040.log}. \item Independent re-implementation from the written argument: \texttt{window\_check.py -{}-run}, log \path{logs/window_check_v030.log}; its control run, with $C_{2.1}=3.43$ in place of $17074961/5000000$, fails as it must (\path{logs_rerun/window_check_test_run.log}). \end{itemize} \subsection{The polynomial \texorpdfstring{$P_A$: $p\in[2.1,2.8]$}{P\_A: p in [2.1,2.8]} and \texorpdfstring{$[14/9,21/11]$}{[14/9,21/11]}}\label{ssec:item-b} \begin{itemize}[leftmargin=1.5em]\raggedright \item \emph{Polynomial:} $P_A=1+\frac{13043}{6250}z-\frac{1789439}{500000}z^2$. \item \emph{Anchors:} $p=2$ (exact), $2.1$, $2.2$, $2.35$, $2.5$, $2.65$, $2.8$ (Bellman certificates) and $\infty$ (Table~\ref{tab:anchors}); anchor table \path{b_PA_21_28/tables/anchors_PA.txt}. \item \emph{Matrices:} the table \path{b_PA_21_28/tables/T_table_PA_5col_0030.txt} has $22$ matrices with six-decimal entries, for $p=2.05,2.10,\dots,3.10$; each piece uses the row nearest to its midpoint. \item \emph{Run} (one command, from \path{b_PA_21_28/programs/}): {\small \begin{verbatim} interval_certify.py 2.1 2.8 0.05 0.0001 1,13043/6250,-1789439/500000 ../tables/anchors_PA.txt ../tables/T_table_PA_5col_0030.txt \end{verbatim}} \noindent log \path{b_PA_21_28/logs/cert_PA_21_28_v080.log}: $15$ pieces tiling $[2.1,2.8]$, namely $[2.1,2.125]$, $[2.125,2.15]$ and $13$ pieces of width $0.05$. \item \emph{Relative margins:} from $0.2900\%$ (on $[2.15,2.2]$) to $1.6958\%$ (on $[2.55,2.6]$). \item \emph{Re-checks of the pieces} (logs in \path{b_PA_21_28/logs/}): \path{read_R2_pieces.py} (\path{read_R2_pieces_PA_21_28_v010.log}, with two controls: all anchors raised by $2\%$, which certifies no piece, and a norm bound below the true norm at $p=2.1$, which cannot be proved) and \path{verify_R2.py} (\path{verify_R2_PA_21_28_v040.log}; its control, with the anchors scaled by $51/50$, is \path{verify_R2_PA_21_28_control_v040.log}). With the printed $\varrho$: \path{verify_R2_rho_PA_21_28_v050.log} and \path{R2_read_rho_PA_21_28.log}, each with a control log whose name adds \path{_control}. \item \emph{Checks of the anchors:} \path{bellman_check.py} (\path{bellman_check_PA_p21.log}, \dots, \path{bellman_check_PA_p28.log}) and \path{bellman_taylor.py} (\path{bellman_taylor_PA_4_v020.log} for $p=2.1$ to $2.5$, \path{bellman_taylor_PA_p265_v020.log}, \path{bellman_taylor_PA_p28_p2725_v020.log}), with controls with lowered $C$. \end{itemize} \subsection{Drury's polynomial: \texorpdfstring{$p\in[2.8,20]$}{p in [2.8,20]} and \texorpdfstring{$[20/19,14/9]$}{[20/19,14/9]}}\label{ssec:item-c} \begin{itemize}[leftmargin=1.5em]\raggedright \item \emph{Polynomial:} Drury's $P_{\mathrm{Dr}}=1+2z-\frac{22}5z^2$ \cite{Dr11}. \item \emph{Anchors:} $p=2$ (exact, $\max_{|z|=1}|P_{\mathrm{Dr}}(z)|=\sqrt{19683/550}$), $p=3$ (Bellman, $\le248143/40000$), $p=4,6,\dots,20$ (TISOS, $\le6.454826$, $6.737251$, $6.891500$, $6.988033$, $7.053905$, $7.101658$, $7.137846$, $7.166211$, $7.189041$) and $p=\infty$ ($\sum|c_k|=7.4$); Table~\ref{tab:anchors}. The anchors at $p=3,4,\dots,20$ are those of Theorem~\ref{thm:shiftnorms}. \item \emph{Runs:} four runs of \path{interval_certify.py} with the coefficients \path{1,2,-22/5}, in Table~\ref{tab:runsc}; together $80+45+160+160=445$ pieces. The first two overlap on $[3.55,3.6]$; together the runs cover $[2.8,20]$. \end{itemize} \begin{table}[ht] \centering\small \renewcommand{\arraystretch}{1.2} \begin{tabular}{|c|c|c|c|c|} \hline run & step / minimal step & anchors used & pieces & relative margins\\ \hline $[2.8,3.6]$ & $0.01$ / $0.0003125$ & $2,3,4$ & $80$ & $1.4354\%$--$1.9101\%$\\ $[3.55,4]$ & $0.01$ / $0.0003125$ & $2,4$ & $45$ & $0.9148\%$--$1.2489\%$\\ $[4,12]$ & $0.05$ / $0.0003125$ & $4,6,8,10,\infty$ & $160$ & $0.0790\%$--$1.0464\%$\\ $[12,20]$ & $0.05$ / $0.0001$ & $12,14,16,18,20$ & $160$ & $0.0139\%$--$0.0942\%$\\ \hline \end{tabular} \caption{The four runs of item (c); ``anchors used'' are the anchors next to the range of the run.}\label{tab:runsc} \end{table} \noindent The files of the four runs (anchor table; matrix table; log), all in \path{c_drury_28_20/}: \begin{itemize}[leftmargin=2em] \item $[2.8,3.6]$: \path{tables/anchors_tisos_p3_preview.txt}; \path{tables/T_table_280_440.txt}; \path{logs/cert_down_p3preview_v070.log}; \item $[3.55,4]$: \path{tables/anchors_tisos_4_10.txt}; \path{tables/T_table_1836.txt}; \path{logs/cert_left_tisos_v070.log}; \item $[4,12]$: \path{tables/anchors_tisos_4_10.txt}; \path{tables/T_table_280_1200.txt}; \path{logs/cert_up_A7_4_12_v080.log}; \item $[12,20]$: \path{tables/anchors_tisos.txt}; \path{tables/T_table_10_20_5col.txt}; \path{logs/cert_up_A7_12_20_v080.log}. \end{itemize} \noindent All piece steps were the nominal ones (no piece needed halving). The matrix tables give a matrix with six-decimal entries for each value of $p$ on a grid of step $0.05$ (step $0.1$ in \path{T_table_10_20_5col.txt}); a matrix may serve several adjacent pieces. \emph{Relative margins:} from $0.0139\%$ (on $[19.9,19.95]$ and $[19.95,20]$, near $p=20$) to $1.9101\%$ (on $[2.99,3]$, near $p=3$). \emph{The anchors used on $[2.8,3]$ and on $[10,12]$.} On $[2.8,3]$ the shift-side bound interpolates between the anchors at $p=2$ and $p=3$. On $[10,12]$ the run on $[4,12]$ interpolates between the anchors at $p=10$ and $p=\infty$, since its anchor table \path{anchors_tisos_4_10.txt} has no anchor at $p=12$; at $p=12$ this gives $6.988033^{5/6}\,7.4^{1/6}\approx7.055066$, which is enough on every piece of $[10,12]$. The re-check \path{verify_R2.py} also uses the anchor at $p=12$ ($\norm{P_{\mathrm{Dr}}(S)}_{12}\le7.053905$); \path{R2_read.py} uses the same anchors as the run. \begin{itemize}[leftmargin=1.5em]\raggedright \item \emph{Re-checks of the pieces} (logs in \path{c_drury_28_20/logs/}): \path{verify_R2.py} (\path{verify_R2_280_360_v030.log}, \path{verify_R2_355_12_v030.log} for the runs on $[3.55,4]$ and $[4,12]$ together, \path{verify_R2_12_20_v030.log}) and \path{R2_read.py} (\path{R2_read_280_360.log}, \path{R2_read_355_4.log}, \path{R2_read_4_12.log}, \path{R2_read_12_20.log}). Each reports all pieces certified, with no failure and an exact tiling. With the printed $\varrho$: \path{verify_R2_rho_280_360_v050.log}, \path{verify_R2_rho_355_12_v050.log}, \path{verify_R2_rho_12_20_v050.log}, and \path{R2_read_rho_280_360.log}, \path{R2_read_rho_355_4.log}, \path{R2_read_rho_4_12.log}, \path{R2_read_rho_12_20.log}, each \path{R2_read} log with a control log whose name adds \path{_control}. \item \emph{Checks of the anchors:} each TISOS certificate by \path{tisos_verify.py} (\path{tisos_bellman_verify_v010.log}) and \path{verify_tisos.py} (\path{verify_tisos_1939.log}, \path{verify_tisos_high1_2017.log}, \path{verify_tisos_high2_2017.log}), and again by \path{tisos_read.py} (\path{logs_rerun/}); the Bellman certificate at $p=3$ by \path{bellman_verify.py} (ball arithmetic; \path{tisos_bellman_verify_v010.log}), \path{verify_bellman_p3.py} (exact rationals; \path{verify_bellman_p3_0031.log}) and \path{bellman_bernstein.py} (exact Bernstein bounds; \path{bellman_bernstein_p3_v020.log}). \end{itemize} \emph{Exact rational matrices at $p=14,16,18,20$.} At these four values of $p$ there are also matrices with rational entries, for which every step is exact rational arithmetic. For each such $p=2m$ the certificate \path{c_drury_28_20/certificates/TSIDE_cert_p

.txt} gives a rational $T$, a rational test vector $v$, and an exact Gram matrix $H$, positive definite, such that for $w=(x,y)$ \[ \Phi(x,y):=x^p+y^p-(Tw)_1^p-(Tw)_2^p=Z^{\mathsf t}HZ,\qquad Z=(x^m,x^{m-1}y,\dots,y^m); \] as $p$ is even, $\Phi(x,y)=\norm{w}_p^p-\norm{Tw}_p^p$, so $\norm{T}_p\le1$. A second proof of $\norm{T}_p\le1$: $\Phi$ is homogeneous of even degree, $\Phi(1,0)>0$, and $\Phi(t,1)$ has no real root (exact count of real roots), so $\Phi>0$ off the origin. Finally, $\norm{P_{\mathrm{Dr}}(T)v}_p^p/\norm{v}_p^p>C^p$ exactly, with $C$ the TISOS anchor at $p$: {\footnotesize \renewcommand{\arraystretch}{1.25} \begin{center} \begin{tabular}{|c|c|c|c|c|c|} \hline $p$ & $T$ & $v$ & $\norm{P_{\mathrm{Dr}}(T)}_p\ge$ & $C$ & margin\\ \hline $14$ & $\begin{pmatrix}-13/1000&991/1000\\-443/500&-129/1000\end{pmatrix}$ & $(1,\ 953413/1000000)$ & $7.103397338$ & $7.101658$ & $0.0245\%$\\ $16$ & $\begin{pmatrix}-113/1000&899/1000\\-993/1000&-1/100\end{pmatrix}$ & $(-239149/250000,\ 1)$ & $7.140643266$ & $7.137846$ & $0.0392\%$\\ $18$ & $\begin{pmatrix}-39/5000&1243/1250\\-4567/5000&-12/125\end{pmatrix}$ & $(1,\ 191789/200000)$ & $7.168119861$ & $7.166211$ & $0.0266\%$\\ $20$ & $\begin{pmatrix}-823/10000&37/40\\-9957/10000&-59/10000\end{pmatrix}$ & $(-24031/25000,\ 1)$ & $7.190460162$ & $7.189041$ & $0.0197\%$\\ \hline \end{tabular} \end{center}} \noindent The certificates are written by \path{evenp_final.py} (\path{logs/evenp_final_v030_12to18.log}, \path{logs/evenp_final_v030_20.log}) and checked independently by \path{read_evenp_Tside.py} (exact, sympy; it takes $T$, $v$, $H$ from the certificate and $C$ from the TISOS file; \path{logs/read_evenp_Tside_v060.log}, with two controls: $1.01\,T$ and $1.01\,C$) and by \path{verify_evenp.py} (\path{logs/verify_evenp_direct_2016.log}). With the TISOS anchors, which are also exact, these four values of $p$ (and their duals) are proved by exact rational arithmetic alone. \subsection{The polynomial \texorpdfstring{$P_D$: $p\in[20,40]$}{P\_D: p in [20,40]} and \texorpdfstring{$[40/39,20/19]$}{[40/39,20/19]}}\label{ssec:item-d} \begin{itemize}[leftmargin=1.5em]\raggedright \item \emph{Polynomial:} $P_D=1+z-8z^2$. \item \emph{Anchors:} $p=2$ (exact, $\sqrt{2673/32}$), $p=20,22,\dots,32$ (TISOS, with a line \path{tau}), $p=34,36,38,40$ (Bellman) and $p=\infty$ ($10$); Table~\ref{tab:anchors}; anchor table \path{d_PD_20_40/tables/anchors_PD_20_40.txt}. \item \emph{Matrices:} \path{d_PD_20_40/tables/T_table_PD_20_40_5col.txt}, $101$ matrices with six-decimal entries for $p=20.0,20.2,\dots,40.0$. \item \emph{Run} (one command, from \path{d_PD_20_40/programs/}): {\small \begin{verbatim} interval_certify.py 20.0 40.0 0.2 0.0005 1,1,-8 ../tables/anchors_PD_20_40.txt ../tables/T_table_PD_20_40_5col.txt \end{verbatim}} \noindent log \path{d_PD_20_40/logs/cert_PD_20_40_v080.log}: $100$ pieces of width $0.2$ tiling $[20,40]$. \item \emph{Relative margins:} from $0.0228\%$ (on $[39.8,40]$) to $0.0994\%$ (on $[20,20.2]$). \item \emph{Re-checks of the pieces} (logs in \path{d_PD_20_40/logs/}): \path{R2_read.py} (\path{R2_read_PD_20_40.log}; the anchors are read from the eleven certificate files) and \path{read_pieces.py} (\path{read_pieces_PD_20_40_0546.log}; its own smallest margin is $0.0229\%$; its control, with every anchor multiplied by $1.001$, has worst margin $-0.0770\%$). With the printed $\varrho$: \path{verify_R2_rho_PD_20_40_v050.log} and \path{R2_read_rho_PD_20_40.log}, each with a control log whose name adds \path{_control}. \item \emph{Checks of the anchors:} TISOS $p=20,\dots,32$ by \path{tisos_read.py} (\path{tisos_read_PD_0146.log}) and \path{verify_tisos.py} (\path{verify_tisos_PD_0154.log}); Bellman $p=34,36,38$ by \path{verify_bellman_evenp.py} (exact rationals; \path{verify_bellman_evenp_PD_0147.log}) and \path{bellman_check.py} (\path{bellman_check_PD_p34.log}, \path{bellman_check_PD_p36.log}, \path{bellman_check_PD_p38.log}); Bellman $p=40$ by \path{bellman_check.py} (\path{bellman_check_PD_p40.log}) and \path{bellman_taylor.py} (\path{bellman_taylor_PD_p40_v020.log}). Each has a control with lowered $C$. \end{itemize} \subsection{Independent checks}\label{ssec:twice} Table~\ref{tab:twice} lists, for each kind of claim, the programs that check it. Programs in the same row were written independently of each other; they use different methods or different arithmetic. {\footnotesize \renewcommand{\arraystretch}{1.2} \begin{longtable}{|p{0.22\textwidth}|p{0.33\textwidth}|p{0.37\textwidth}|} \caption{Independent checks. Logs are in the folder \texttt{logs/} of the item's part unless said otherwise.}\label{tab:twice}\\ \hline \raggedright claim & \raggedright first check & \raggedright second (and third) check\tabularnewline \hline\endfirsthead \hline \raggedright claim & \raggedright first check & \raggedright second (and third) check\tabularnewline \hline\endhead \hline\endfoot \raggedright (a) anchor $p=2.1$ & \raggedright \path{bellman_check.py}: first-order box bounds, balls & \raggedright \path{bellman_taylor.py}: second-order Taylor and Bernstein bounds, balls\tabularnewline \raggedright (a) the range $(2,2.1]$ & \raggedright \path{window_cert.py}: $400$ pieces; an enclosure of $\Theta'$ on the last piece & \raggedright \path{window_check.py}: re-implementation from the written argument; enclosures of $\Theta'$ on $8$ sub-pieces\tabularnewline \hline \raggedright (b) six Bellman anchors & \raggedright \path{bellman_check.py} & \raggedright \path{bellman_taylor.py}\tabularnewline \raggedright (b) $15$ pieces & \raggedright \path{interval_certify.py}: mpmath intervals & \raggedright \path{read_R2_pieces.py} and \path{verify_R2.py}: own norm bound, own test vector, balls and exact rationals\tabularnewline \hline \raggedright (c) TISOS anchors $p=4,\dots,20$ & \raggedright \path{tisos_verify.py}: exact, folded identity, Sylvester & \raggedright \path{verify_tisos.py}: exact, folded and summed identity, $LDL^{\mathsf t}$; \path{tisos_read.py}: exact\tabularnewline \raggedright (c) Bellman anchor $p=3$ & \raggedright \path{bellman_verify.py}: balls & \raggedright \path{verify_bellman_p3.py}: exact rationals; \path{bellman_bernstein.py}: exact Bernstein bounds\tabularnewline \raggedright (c) $445$ pieces & \raggedright \path{interval_certify.py} (four runs) & \raggedright \path{verify_R2.py} and \path{R2_read.py}\tabularnewline \raggedright (c) exact matrices $p=14,\dots,20$ & \raggedright \path{evenp_final.py}: exact Gram matrix and Sturm count & \raggedright \path{read_evenp_Tside.py}: exact (sympy), own Gram check and root count; \path{verify_evenp.py}\tabularnewline \hline \raggedright (d) TISOS anchors $p=20,\dots,32$ & \raggedright \path{tisos_read.py} & \raggedright \path{verify_tisos.py}\tabularnewline \raggedright (d) Bellman anchors $p=34,36,38$ & \raggedright \path{verify_bellman_evenp.py}: exact rationals & \raggedright \path{bellman_check.py}: balls\tabularnewline \raggedright (d) Bellman anchor $p=40$ & \raggedright \path{bellman_check.py} & \raggedright \path{bellman_taylor.py}\tabularnewline \raggedright (d) $100$ pieces & \raggedright \path{interval_certify.py} & \raggedright \path{R2_read.py} and \path{read_pieces.py}\tabularnewline \hline \raggedright (b)--(d) the printed $\varrho_i$ (Theorem~\ref{thm:explicit}) & \raggedright \path{verify_R2.py} \texttt{-{}-printed-R}: balls & \raggedright \path{R2_read.py} \texttt{-{}-verify-printed-rho}: balls\tabularnewline \end{longtable}} \emph{Controls.} Every control run named in this paper fails, as it must. They use a changed constant $C$, a scaled weight of a Bellman certificate, anchors raised by a small factor, a lowered norm bound, or a perturbed matrix. A log with ``control'' in its name is such a run. \emph{Re-runs.} The folder \path{logs/} of each part holds the logs of the runs this section relies on. The folder \path{logs_rerun/} holds the same checks run again with exactly the shipped files, one log per command; its first line names the program, its md5 and the full command, and its last line the exit code. The certifier re-runs reproduce every piece line of the logs. Every log in \path{logs/} prints the md5 of the program, anchor table and matrix table as they were run; \path{MANIFEST.tsv} relates these to the shipped files, which differ from the originals only in comments, printed texts and, for a few programs, the paths they read (\path{README.txt} lists every change). The programs need Python 3.12, python-flint 0.9.0, mpmath 1.2.1, numpy 2.5.4 and sympy 1.12, and run on one thread; each is run from the \path{programs/} folder of its part. \subsection{Proof of Theorem~\ref{thm:explicit}}\label{ssec:explicit-proof} Let $q=\max(p,p')\in(2,40]$. By Lemma~\ref{lem:dualpiece} it suffices to prove the claim at $q$ with $T_*=T$. If $q\in(2,2.1]$: the checks (W1)--(W2) of Section~\ref{ssec:item-a} give $\Theta(1/q)>0$, and Lemma~\ref{lem:windowa} gives $\norm{T_q}_q\le1$ and $\norm{P_{7/10}(T_q)}_q>\norm{P_{7/10}(S)}_q$. If $q\in[2.1,40]$: the pieces of the runs of items (b)--(d) tile $[2.1,2.8]$, $[2.8,3.6]\cup[3.55,4]\cup[4,12]\cup[12,20]$ and $[20,40]$ with no gap (each certifier log ends with this statement, and each re-checker checks the tiling again). So $q$ lies in some piece, and on that piece the anchors of Table~\ref{tab:anchors} (valid by Lemmas~\ref{lem:exactanchors} and~\ref{lem:bellcert}, Proposition~\ref{prop:tisos} and the checks of Section~\ref{ssec:twice}) and the two facts $\varrho_i\ge\max(\norm{T_i}_{p_{\mathrm{lo}}},\norm{T_i}_{p_{\mathrm{hi}}})$ and $L>\widehat C(p_{\mathrm{hi}})$ for $T_i/\varrho_i$, proved for the printed $\varrho_i$ by \path{verify_R2.py} and \path{R2_read.py} (Section~\ref{ssec:explicit-scheme}), give the claim by Lemma~\ref{lem:piece}. \qed \section{Necessary conditions for matrix counterexamples}\label{sec:structure} Theorem~\ref{thm:A} and Drury's theorem \cite{Dr11} give real $2\times2$ counterexamples to \eqref{eq:matsaev}. This section collects simple necessary conditions for matrix counterexamples, mostly of size $2\times2$, and it checks where the known counterexamples lie. All the conditions come from a single known class of contractions that satisfy \eqref{eq:matsaev}: those whose entrywise modulus is a contraction (Lemma~\ref{lem:modulus}). These conditions are elementary; they show where a search need not look. Throughout this section $10$. Conversely, if this ratio is positive, put $\alpha_1=0$, $\beta_1=\arg t_{11}$, $\beta_2=\arg t_{12}$ and $\alpha_2=\arg t_{21}-\arg t_{11}$. Then \eqref{eq:phases} holds for $(1,1)$, $(1,2)$ and $(2,1)$ by construction, and for $(2,2)$ it reads $\arg t_{21}-\arg t_{11}+\arg t_{12}\equiv\arg t_{22}$, which is the assumption. \emph{Some entry zero.} Consider the graph whose four vertices are the two rows and the two columns, with an edge from row $j$ to column $k$ whenever $t_{jk}\neq0$. The unknown $\alpha_j$ sits at row $j$, and $\beta_k$ at column $k$. The four possible edges form a cycle of length four; with at most three of them, the graph has no cycle. In each connected component, set the unknown at one vertex equal to $0$ and walk outward: each edge leads to a vertex not yet reached, and \eqref{eq:phases} for that edge fixes the unknown there. Unknowns at isolated vertices are arbitrary. (For example, if $t_{11}=0$: $\alpha_1=0$, $\beta_2=\arg t_{12}$, $\alpha_2=\arg t_{22}-\beta_2$, $\beta_1=\arg t_{21}-\alpha_2$.) Finally, $t_{11}t_{22}\,\overline{t_{12}t_{21}}=0$ if some entry is $0$, and $t_{11}t_{22}\,\overline{t_{12}t_{21}}=|t_{12}t_{21}|^2\,t_{11}t_{22}/(t_{12}t_{21})$ otherwise; this gives the equivalent form. \end{proof} \begin{corollary}[necessary conditions for a $2\times2$ counterexample]\label{cor:where} Let $T$ be a $2\times2$ counterexample at some $p\in(1,\infty)$. Then: \begin{enumerate}[leftmargin=2em,label=(\roman*)] \item $\norm{|T|}_p>1$, and $T$ is invertible; \item all four entries of $T$ are nonzero, and $t_{11}t_{22}/(t_{12}t_{21})\notin(0,\infty)$; \item if $T$ is real, then $t_{11}t_{12}t_{21}t_{22}<0$, so $T$ has one or three negative entries. \end{enumerate} \end{corollary} \begin{proof} (i) follows from Lemma~\ref{lem:modulus} and Corollary~\ref{cor:rankone}, since a $2\times2$ matrix of rank two is invertible. (ii) is Lemma~\ref{lem:phase}. (iii) For real nonzero entries, the ratio in (ii) is positive if and only if $t_{11}t_{22}$ and $t_{12}t_{21}$ have the same sign, that is, if and only if $t_{11}t_{12}t_{21}t_{22}>0$. This product is negative exactly when an odd number of the four entries is negative. \end{proof} \begin{remark}\label{rem:phase} (1) A $2\times2$ matrix of rank one with four nonzero entries has $t_{11}t_{22}=t_{12}t_{21}$, so the ratio in Lemma~\ref{lem:phase}(b) is $1$. Thus, for $n=2$, Lemma~\ref{lem:phase} contains Corollary~\ref{cor:rankone}. (2) Among real matrices with four nonzero entries, Corollary~\ref{cor:where}(iii) excludes half of the $16$ sign patterns. For complex matrices the excluded set is small: the ratio is a positive real number only on a set of measure zero. (3) For $n\times n$ matrices, $T$ has the form in Lemma~\ref{lem:phase}(a) if and only if the system \eqref{eq:phases} is solvable on the bipartite graph of the nonzero entries of $T$ (rows and columns as vertices). By the standard criterion for such systems on a graph (solve along a spanning tree of each component, as above; the remaining equations are conditions on cycles), this holds if and only if, around every cycle of that graph, the alternating sum of the arguments of the entries is $0$ modulo $2\pi$. \end{remark} \subsection{Real sign patterns, and the known counterexamples} \begin{remark}[two sign classes]\label{rem:signclasses} Three operations take a real $2\times2$ counterexample at $p$ to another real counterexample at the same $p$. \begin{enumerate}[leftmargin=2.5em,label=(m\arabic*)] \item $T\mapsto VTV$ with $V=\mathrm{diag}(1,-1)$. It changes the signs of $t_{12}$ and $t_{21}$. Here $V=V^{-1}$ is an isometry of $\ell^p_2$, and $P(VTV)=VP(T)V$. \item $T\mapsto XTX$ with the coordinate exchange $X=\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)$. It exchanges $t_{11}$ with $t_{22}$, and $t_{12}$ with $t_{21}$. Again $X=X^{-1}$ is an isometry, and $P(XTX)=XP(T)X$. \item $T\mapsto-T$, with $P$ replaced by $\widetilde P(z)=P(-z)$. Then $\widetilde P(-T)=P(T)$, and $\norm{\widetilde P(S)}_p=\norm{P(-S)}_p=\norm{P(S)}_p$, because $-S=USU^{-1}$ with the isometry $U=\mathrm{diag}((-1)^n)$ of $\ell^p(\N)$. \end{enumerate} In addition, transposition takes a counterexample at $p$ to one at $p'$, by Lemma~\ref{lem:dual}. All four operations keep the signs of $t_{11}t_{22}$ and of $t_{12}t_{21}$, and by Corollary~\ref{cor:where}(iii) exactly one of these two products is negative. Hence (m1) and (m3) carry every real $2\times2$ counterexample at $p$ to a counterexample in one of two classes: \begin{center} \emph{Class I:} $t_{11},t_{22},t_{12}>0>t_{21}$;\qquad\qquad \emph{Class II:} $t_{11}>0>t_{22}$ and $t_{12},t_{21}>0$. \end{center} Indeed, if $t_{12}t_{21}<0$, then $t_{11}t_{22}>0$, and (m3), if needed, makes $t_{11}>0$, hence $t_{22}>0$; then (m1), if needed, also makes $t_{12}>0>t_{21}$. If $t_{11}t_{22}<0$, then $t_{12}t_{21}>0$; (m3), if needed, makes $t_{11}>0>t_{22}$, and then (m1), if needed, makes $t_{12},t_{21}>0$. The sign of $t_{11}t_{22}$ separates the two classes; equivalently, $\det T>0$ in class I and $\det T<0$ in class II. We say that a real $2\times2$ matrix \emph{belongs to} class I (or II) if (m1) and (m3) carry it into that class. So a search among real $2\times2$ matrices may fix all signs and search two families of matrices with positive entries. \end{remark} \begin{remark}[the known counterexamples are in class I]\label{rem:classI} (1) \emph{The matrices of Theorem~\ref{thm:A}.} The matrix of Lemma~\ref{lem:TK} is $T_\varepsilon=\exp(\varepsilon K_*)J=e^{-\varepsilon D_*/2}R_{\varepsilon/2}J=e^{-\varepsilon D_*/2}R_{\pi/2+\varepsilon/2}$, that is, \[ T_\varepsilon=e^{-\varepsilon D_*/2}\begin{pmatrix}-\sin(\varepsilon/2)&\cos(\varepsilon/2)\\-\cos(\varepsilon/2)&-\sin(\varepsilon/2)\end{pmatrix}. \] For $0<\varepsilon<\pi$ all four entries are nonzero, three of them are negative, and $t_{11}t_{22}/(t_{12}t_{21})=-\tan^2(\varepsilon/2)<0$. So $T_\varepsilon$ passes the tests (ii) and (iii) of Corollary~\ref{cor:where}, and it belongs to class I ($-VT_\varepsilon V$, obtained by (m3) and (m1), satisfies the inequalities of class I). The same holds for its transpose, which Theorem~\ref{thm:A} uses for $p<2$. At $\varepsilon=0$ the matrix is $J$, whose diagonal is zero. So $J$ has the form in Lemma~\ref{lem:phase}(a) (indeed $|J|=X$), and the lemma gives $\norm{P(J)}_p\le\norm{P(S)}_p$, in agreement with Lemma~\ref{lem:J}. Thus the matrices of Lemma~\ref{lem:TK} leave the set of Lemma~\ref{lem:phase}(a), where no counterexample can be, as soon as $\varepsilon>0$. More generally, a scaled rotation $cR_\theta$ with $c>0$ and $\sin\theta\cos\theta\neq0$ has $t_{11}t_{22}=c^2\cos^2\theta>0>-c^2\sin^2\theta=t_{12}t_{21}$, so it belongs to class I. The scaled rotations of Theorem~\ref{thm:N2} and of Theorem~\ref{thm:explicit}(a) are of this kind, with $\cos\theta<0$. (2) \emph{Drury's matrix.} Drury's contraction at $p=4$ \cite{Dr11} is the matrix of Remark~\ref{rem:druryT}, \[ T_{\mathrm{Dr}}=\begin{pmatrix}-0.292187&0.812218\\-0.958489&-0.113428\end{pmatrix}. \] Its products are $t_{11}t_{22}=0.033142187036>0$ and $t_{12}t_{21}=-0.778502018602<0$ (exact products of the decimals), so the ratio is about $-0.0426$, three entries are negative, and $T_{\mathrm{Dr}}$ belongs to class I: $-VT_{\mathrm{Dr}}V$ satisfies the inequalities of class I. By Lemma~\ref{lem:phase} the ratio must be negative, since $T_{\mathrm{Dr}}$ is a counterexample. In the same way, Corollary~\ref{cor:where}(i) forces $\norm{|T_{\mathrm{Dr}}|}_4>1$. (3) \emph{The matrices of Theorem~\ref{thm:explicit}.} The matrices $T_i$ of all $560$ pieces of Theorem~\ref{thm:explicit}(b)--(d), as printed in the logs named in Sections~\ref{ssec:item-b}--\ref{ssec:item-d}, and the four rational matrices at $p=14,16,18,20$ in Section~\ref{ssec:item-c} have the sign pattern $\left(\begin{smallmatrix}-&+\\-&-\end{smallmatrix}\right)$; so does the scaled rotation $T_q$ of item~(a), as $\cos\varphi<0<\sin\varphi$. So $T_\varepsilon$, $T_{\mathrm{Dr}}$ and the matrices of Theorem~\ref{thm:explicit} have the same sign pattern, $\left(\begin{smallmatrix}-&+\\-&-\end{smallmatrix}\right)$, and their transposes, used for $p<2$, have the transposed pattern. All known counterexamples belong to class I; the results of this section do not exclude class II. \end{remark} \subsection*{Acknowledgements} The author thanks S.~K.~Ray for correspondence about this work.\footnote{When the author sent S.~K.~Ray a preprint of this paper on 28~October~2026, he replied that he had arrived at the result independently, by a somewhat different approach.} This work was done overwhelmingly by AI models and supervised by the author. The models were Anthropic's Claude and OpenAI's ChatGPT. The author takes full responsibility for the paper. \begin{thebibliography}{99} \bibitem{AS77} M. A. Akcoglu and L. Sucheston, Dilations of positive contractions on $L_p$ spaces, Canad. Math. Bull. 20 (1977), 285--292. \bibitem{Ar13} C. Arhancet, On Matsaev's conjecture for contractions on noncommutative $L^p$-spaces, J. Operator Theory 69 (2013), 387--421 (arXiv:1009.1292). \bibitem{BL76} J. Bergh and J. L\"ofstr\"om, Interpolation Spaces. An Introduction, Grundlehren der Mathematischen Wissenschaften 223, Springer, Berlin, 1976. \bibitem{CRW78} R. R. Coifman, R. Rochberg and G. Weiss, Applications of transference: the $L^p$ version of von Neumann's inequality and the Littlewood--Paley--Stein theory, in: Linear Spaces and Approximation (Oberwolfach 1977), P.~L. Butzer and B. Sz.-Nagy (eds.), Internat. Ser. Numer. Math. 40, Birkh\"auser, Basel, 1978, 53--67. \bibitem{Da67} J. M. Danskin, The Theory of Max-Min and its Application to Weapons Allocation Problems, Springer, 1967. \bibitem{Dr11} S. W. Drury, A counterexample to a conjecture of Matsaev, Linear Algebra Appl. 435 (2011), 323--329. \bibitem{Fa15} S. Fackler, Regularity properties of sectorial operators: counterexamples and open problems, in: Operator Semigroups Meet Complex Analysis, Harmonic Analysis and Mathematical Physics, W. Arendt, R. Chill and Y. Tomilov (eds.), Operator Theory: Advances and Applications 250, Birkh\"auser, Cham, 2015, 171--197 (arXiv:1407.1142). \bibitem{Gs} D. Goel, The Matsaev (1971) conjecture is false for every $p\neq2$, Zenodo (2026), \url{https://doi.org/10.5281/zenodo.23691821} \bibitem{HK05} O. Holtz and M. Karow, Real and complex operator norms, preprint (2005), arXiv:math/0512608. \bibitem{Jo17} F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, IEEE Trans. Comput. 66 (2017), 1281--1292. \bibitem{MZ39} J. Marcinkiewicz and A. Zygmund, Quelques in\'egalit\'es pour les op\'erations lin\'eaires, Fund. Math. 32 (1939), 115--121. \bibitem{mpmath} The mpmath development team, mpmath: a Python library for arbitrary-precision floating-point arithmetic, version 1.2.1 (2021), \url{https://mpmath.org} \bibitem{Ni74} N. K. Nikol'skii, Five problems about invariant subspaces, Zap. Nauchn. Sem. LOMI 23 (1971), 115--127 (Russian); English: Five problems on invariant subspaces, J. Soviet Math. 2 (1974), 441--450. \bibitem{Pe76} V. V. Peller, An analogue of J. von Neumann's inequality for the space $L_p$, Dokl. Akad. Nauk SSSR 231 (1976), 539--542 (Russian). \bibitem{Pe78} V. V. Peller, Approximations by isometries and V. I. Matsaev's hypothesis for absolute contractions of the space $L^p$, Funktsional. Anal. i Prilozhen. 12 (1978), 38--50 (Russian); English: Funct. Anal. Appl. 12 (1978), 29--38. \bibitem{Pe81} V. V. Peller, An analogue of an inequality of J. von Neumann, isometric dilation of contractions, and approximation by isometries in spaces of measurable functions, Trudy Mat. Inst. Steklov. 155 (1981), 103--150 (Russian); English: Proc. Steklov Inst. Math. (1983), no.~1, 101--145. \bibitem{Ra20} S. K. Ray, On multivariate Matsaev's conjecture, Complex Anal. Oper. Theory 14 (2020), Paper No. 42 (arXiv:1703.00733). \end{thebibliography} \end{document}