% ========================================================================= \section{A three-mean support of order nine}\label{sec:mean} The nested support reaches order nine with two radicals placed one inside the other, and its arguments cost divisions. A second support of the same order uses two radicals placed side by side, each of them produced by one arc. It is a combination of three classical means of $1$ and $t$. \subsection{Formula} The exact support is $\kappa_*=1+(p-1)q_*$ with \[ q_*(t)=\frac{t-t^{h}}{(p-1)(t^{h}-1)},\qquad q_*(1)=1,\qquad q_*(1/t)=q_*(t)/t , \] a homogeneous symmetric mean of $1$ and $t$; for integer $p$ it is the arithmetic mean of the $p-1$ mean proportionals $t^{k/p}$ between $1$ and $t$. For real $p>2$ define \begin{equation}\label{eq:mcoeff} \begin{gathered} m=\frac{(p-2)(47p^{2}+21p+4)}{60\,p\,(4p^{2}-1)},\qquad g=\frac{16\,(p+1)(4p^{2}-1)}{15\,p\,(5p^{2}+4)},\\[2pt] n=\frac{49\,p\,(p+1)(p^{2}-4)}{10\,(4p^{2}-1)(5p^{2}+4)},\qquad d=\frac{5p^{2}+4}{84\,p^{2}} , \end{gathered} \end{equation} all positive, with $2m+g-n=1$, and \begin{equation}\label{eq:qM} q_M(t)=m\,(1+t)+g\sqrt t-n\sqrt{t+d\,(t-1)^{2}},\qquad \Tm_{M}(s)=s\,\frac{1+(p-1)q_M(t)}{t+(p-1)q_M(t)}, \end{equation} that is, the support $\kappa=1+(p-1)q_M(t)$ in \eqref{eq:support}. With the arithmetic mean $\mathrm A=(1+t)/2$, the geometric mean $\mathrm G=\sqrt t$ and the \emph{conic mean} $\mathrm C_d=\sqrt{\mathrm G^{2}+4d(\mathrm A^{2}-\mathrm G^{2})}$, \[ q_M=2m\,\mathrm A+g\,\mathrm G-n\,\mathrm C_d , \] a combination with weights of sum one; one weight is negative. The radicand $t+d(t-1)^{2}$ is positive for every $t>0$, so $q_M$ is defined on the whole positive axis, and $q_M(1/t)=q_M(t)/t$: the log-error map is odd. In terms of $h=1/p$, $m=\frac{(1-2h)(4h^{2}+21h+47)}{60(4-h^{2})}$, $g=\frac{16(1+h)(4-h^{2})}{15(4h^{2}+5)}$, $n=\frac{49(1+h)(1-4h^{2})}{10(4-h^{2})(4h^{2}+5)}$ and $d=\frac{4h^{2}+5}{84}$. For cube roots \begin{equation}\label{eq:qMcubic} q_M(t)=\frac{7}{90}(1+t)+\frac{64}{63}\sqrt t-\frac{6}{35}\sqrt{t+\frac{7}{108}(t-1)^{2}} . \end{equation} The three terms belong to one family. The decentered arc is $q_5^{2}=t+a_9(t-1)^{2}$, with the $a_9$ of \eqref{eq:abc}; the members $a=\tfrac14$ and $a=0$ of the family $q^{2}=t+a(t-1)^{2}$ are $\mathrm A$ and $\mathrm G$, and $\mathrm C_d$ is the member $a=d$. Each member divided by $\sqrt t$ is $\sqrt{1+aw}$, a function of $w=(t-1)^{2}/t$ alone, as is $q_*/\sqrt t$. A combination of members therefore matches $q_*/\sqrt t$ coefficient by coefficient in $w$: the arc, with one free parameter, matches one coefficient, and \eqref{eq:qM}, with three free parameters (four coefficients, one normalization), matches three. The two members $\mathrm A$ and $\mathrm G$ are imposed; only the third is chosen by the matching. \subsection{Construction by three arcs}\label{sec:mean-construction} Let $\lambda=H/X$, so that $M=(0,H-\lambda)$ and $E=(W,H-\lambda t)$. Put \[ \gamma=1+2(p-1)m,\qquad \zeta=\frac1{2d}-1>1,\qquad K_g=\frac{(p-1)g}{\gamma}\,\lambda,\qquad K_n=\frac{(p-1)n\sqrt d}{\gamma}\,\lambda , \] and prepare four verticals, the horizontal through $M$ and three points: \[ \begin{gathered} v:\ x=W\Bigl(1-\frac1\gamma\Bigr),\qquad v_1:\ x=x_1=\frac{(p-1)m}{\gamma}W,\qquad v_2:\ x=x_2=\frac{K_g}{2\lambda}W,\qquad v_3:\ x=x_3=\frac{K_n}{\lambda}W,\\[2pt] F=(x_2,\,H-\lambda),\qquad G_2=(x_2,\,H-\lambda+K_g),\qquad G_3=(x_3,\,H-\lambda+K_n(1+\zeta)). \end{gathered} \] The vertical $v$ is at abscissa $2x_1$. The construction below requires that the distance from $v_1$ to $v$ equal $K_n\sqrt{\zeta^{2}-1}$, which fixes the shape of the rectangle: \begin{equation}\label{eq:aspect} \frac{WX}{H}=\frac nm\sqrt{\frac{1-4d}{4d}} . \end{equation} The right side is a constructible number depending on $p$ only; it equals $3.7256\ldots$ for $p=3$ and decreases to $2.238\ldots$ as $p\to\infty$. The line $ME$ is drawn for the transport. Its intersections $J_i=ME\cap v_i$ are free, and their heights are affine in $t$: $y_{J_i}=H-\lambda+(x_i/W)\lambda(1-t)$. Three arcs follow. \begin{enumerate}[label=\arabic*.] \item \emph{Conic radical.} With centre $J_1$ and compass opening $J_3G_3=K_n(t+\zeta)$, draw an arc and take its upper intersection $D_1$ with $v$. Its height above $J_1$ is $K_n\sqrt{(t+\zeta)^{2}-(\zeta^{2}-1)}=K_n\sqrt{(t+d(t-1)^{2})/d}$. This is the device of the decentered arc, with a moving centre. \item \emph{Mean proportional.} With centre $J_2$, draw the arc through $G_2$; its radius is $\tfrac12K_g(1+t)$ and $J_2$ is at distance $\tfrac12K_g|1-t|$ from the horizontal through $M$, so the arc meets that horizontal at a point $Z$ with $FZ=K_g\sqrt t$. \item \emph{Transfer.} With centre $D_1$ and opening $FZ$, draw an arc and take its lower intersection $D_M$ with $v$. \end{enumerate} Then \[ H-y_{D_M}=\frac\lambda\gamma\bigl[1+(p-1)m(1+t)\bigr]-K_n\sqrt{\frac{t+d(t-1)^{2}}{d}}+K_g\sqrt t =\frac\lambda\gamma\bigl[1+(p-1)q_M(t)\bigr], \] and since $W-x_{D_M}=W/\gamma$ the slope of $AD_M$ is $\tfrac{H}{WX}[1+(p-1)q_M(t)]$: the join $AD_M$ is the support. Both radicands are positive for $t>0$, so the three intersections exist at every state. The mobile work is three arcs and one join, against one arc and one join for the decentered arc. Without the condition \eqref{eq:aspect} the centre of the first arc is obtained by one additional join, projecting $E$ from a prepared point onto the vertical at distance $K_n\sqrt{\zeta^{2}-1}$ from $v$; we count the protocol with \eqref{eq:aspect}. \begin{figure}[h] \centering\figMean \caption{The three-mean support for $p=3$, $X=2$, $H=4$, $s=3/4$, with $W=7.451\ldots$ from \eqref{eq:aspect}, drawn from the exact coordinates at equal axis scales. Dashed: prepared verticals and the horizontal through $M$. The free points $J_1,J_2,J_3$ are on the transport line $ME$ (green). The compass openings $J_3G_3$ and $FZ$ are the thick orange segments; the three arcs are orange; the support $AD_M$ is blue. Grey: the triangle chain.}\label{fig:mean} \end{figure} \subsection{Global convergence and order nine} \begin{theorem}[Three-mean support; formalized in Lean]\label{thm:mean} For real $p>2$, $X>0$ and every $s_0>0$: \begin{enumerate}[label=\textup{(\roman*)}] \item $q_M(t)>0$ for every $t>0$; \item the iteration $s_{j+1}=\Tm_M(s_j)$ converges to $\alpha$ monotonically without crossing it; \item the local order is exactly nine, \begin{equation}\label{eq:cM} e_+=c_M(p)\,e^{9}+O(e^{11}),\qquad c_M(p)=\frac{(p^{2}-1)(p^{2}-4)(4p^{2}-1)(17p^{2}+8)}{2^{11}\,3^{4}\,5^{2}\,7^{2}}>0 , \end{equation} and $c_M(3)=23/20736$. \end{enumerate} \end{theorem} For comparison $c_9(3)=7/648=224/20736$: at $p=3$ the constant is about ten times smaller than that of the nested support, and $c_M(p)/c_9(p)\to1020/21311\approx0.048$ as $p\to\infty$. \begin{proof} Work in the variable $\tau=\sqrt t>0$, in which only one radical remains. Put $c=p-1$ and \[ S_M(\tau)=\tau^{2}+d\,(\tau^{2}-1)^{2},\qquad \rho_M=\sqrt{S_M},\qquad \ell_M(\tau)=m(1+\tau^{2})+g\,\tau,\qquad q(\tau)=\ell_M(\tau)-n\,\rho_M(\tau), \] so that $q_M(t)=q(\sqrt t)$ and $S_M\ge\tau^{2}>0$. \emph{Positivity.} Clearing denominators gives the identity \begin{equation}\label{eq:mean-pos} \ell_M^{2}-n^{2}S_M=\frac{Q_0\,(\tau-1)^{4}+Q_1\,\tau\,(\tau-1)^{2}+Q_2\,\tau^{2}}{900\,p^{2}(4p^{2}-1)(5p^{2}+4)}, \end{equation} with $Q_0=(p-2)^{2}(2p-1)(313p^{3}+272p^{2}+200p+16)$, $Q_1=60\,p\,(p-2)(2p-1)(71p^{3}+74p^{2}+40p-8)$ and $Q_2=180\,p^{2}(149p^{4}+49p^{3}-141p^{2}-196p-20)$. For $p>2$, $Q_0\ge0$, $Q_1>0$ and $Q_2>0$, the last because $149p^{4}+49p^{3}-141p^{2}-196p-20=149\varrho^{4}+1241\varrho^{3}+3729\varrho^{2}+4596\varrho+1800$ with $\varrho=p-2$. Hence $\ell_M^{2}>n^{2}S_M$, and as $\ell_M>0$ we get $\ell_M>n\rho_M$: this is (i). \emph{The slope.} Let \[ \Lambda_M(\tau)=\log\bigl(\tau^{2}+c\,q\bigr)-\log\bigl(1+c\,q\bigr)-\frac2p\log\tau , \qquad\text{so that}\qquad -\Lambda_M(\sqrt t)=\frac{\log t}{p}+\log\frac{1+c\,q_M}{t+c\,q_M}=e_+ . \] Both logarithms are defined by (i). Differentiating, with $q'=\ell_M'-nS_M'/(2\rho_M)$, \[ \tau\rho_M(\tau^{2}+cq)(1+cq)\,\Lambda_M' =2\tau^{2}(1+cq)\rho_M+c\,\tau(1-\tau^{2})\bigl(\ell_M'\rho_M-\tfrac12nS_M'\bigr) -\tfrac2p(\tau^{2}+cq)(1+cq)\rho_M . \] The right side is a cubic polynomial in $\rho_M$. Reducing it with $\rho_M^{2}=S_M$ and substituting \eqref{eq:mcoeff} gives \begin{equation}\label{eq:mean-slope} \Lambda_M'(\tau)=-\,\frac{(p^{2}-4)(p^{2}-1)\,\bigl(2\Psi_2\,\rho_M-7\Psi_1\bigr)} {1800\,p^{3}(4p^{2}-1)^{2}(5p^{2}+4)^{2}\;\tau\,\rho_M\,(\tau^{2}+cq)(1+cq)} , \end{equation} where $\Psi_1$ is a palindromic sextic and $\Psi_2$ a palindromic quartic in $\tau$ with coefficients in $\mathbb Z[p^{2}]$, listed in Table~\ref{tab:mean-cert}. Every factor of the denominator is positive. \emph{The sign.} The three coefficients of $\Psi_2$ in Table~\ref{tab:mean-cert} are positive for $p>2$, so $\Psi_2>0$ for $\tau>0$. Rationalizing, \begin{equation}\label{eq:mean-square} 4\Psi_2^{2}\,S_M-49\Psi_1^{2}=\frac{4\,(4p^{2}-1)^{3}(5p^{2}+4)^{2}}{21\,p^{2}}\;(\tau-1)^{8}\;\Psi_3 , \end{equation} a polynomial identity after clearing $84p^{2}$, where $\Psi_3$ is a third palindromic quartic whose three coefficients are again positive. Let $\tau>0$, $\tau\ne1$. If $\Psi_1\le0$ then $2\Psi_2\rho_M>0\ge7\Psi_1$. Otherwise both sides are positive and \eqref{eq:mean-square} gives $(2\Psi_2\rho_M)^{2}>(7\Psi_1)^{2}$. In both cases $2\Psi_2\rho_M>7\Psi_1$, and by \eqref{eq:mean-slope} $\Lambda_M'<0$ for $\tau\neq1$. The factor $(\tau-1)^{8}$ is the contact of order eight. \emph{Convergence.} $\Lambda_M$ is strictly decreasing and $\Lambda_M(1)=0$, because $q(1)=2m+g-n=1$. Hence $e_+=-\Lambda_M(\sqrt t)$ has the sign of $t-1$, that is the sign of $e$: the state does not cross $\alpha$. Meanwhile $\Tm_M(s)/s-1=(1-t)/(t+c\,q_M)$ moves the state toward $\alpha$. The orbit is monotone and bounded between $s_0$ and $\alpha$; its positive limit is a fixed point by continuity, and a fixed point has $t=1$. This is (ii). \emph{Local order.} Both $q_M/\sqrt t$ and $q_*/\sqrt t$ are even analytic functions of $\log t$, hence power series in $w$, and exact expansion gives \begin{equation}\label{eq:mean-D4} \frac{q_M-q_*}{\sqrt t}=D_4^{M}\,w^{4}+O(w^{5}),\qquad D_4^{M}=\frac{(2-h)(1+h)(2+h)(1-2h)(1+2h)(8h^{2}+17)}{203212800} . \end{equation} Since $w=\xi^{2}/t$, $\kappa-\kappa_*=(p-1)D_4^{M}\xi^{8}+O(\xi^{9})$, and Lemma~\ref{lem:contact} with $\omega=8$ gives $e_+=(p-1)p^{7}D_4^{M}e^{9}+O(e^{10})$, which is \eqref{eq:cM}; oddness removes the term in $e^{10}$. In Lean the expansion \eqref{eq:mean-D4} is obtained without Taylor series. Put $y=\tfrac12\log t$, so that $\sqrt t=e^{y}$ and the conic radical is $e^{y}\sqrt{1+dw}$. If $R_4$ is the quartic Taylor polynomial of $\sqrt{1+dw}$, then $1+dw-R_4^{2}$ is $w^{5}$ times a polynomial, so replacing the radical by $e^{y}R_4(w)$ costs $O(w^{5})$. After this replacement and after subtracting $D_4^{M}e^{y}w^{4}$, the difference with $q_*$, multiplied by $60(4-h^{2})(4h^{2}+5)(1-h)(e^{2hy}-1)e^{8y}$, is a sum of $22$ exponentials $c_j(h)e^{(k_j+2hl_j)y}$ whose moments $\sum_jc_j(k_j+2hl_j)^{n}$ vanish for $n<11$; it is therefore $O(y^{11})$, and $w/y^{2}\to4$ gives $(q_M-q_*)/y^{8}\to256\,D_4^{M}$. \end{proof} \begin{table}[h] \centering\small \caption{The certificate polynomials of Theorem~\ref{thm:mean}. $\Psi_1$ is given by its coefficients on $\tau^{6}+1$, $\tau^{5}+\tau$, $\tau^{4}+\tau^{2}$, $\tau^{3}$; $\Psi_2$ and $\Psi_3$ by their coefficients on $(\tau-1)^{4}$, $\tau(\tau-1)^{2}$, $\tau^{2}$. Each coefficient of $\Psi_2$ and $\Psi_3$ is positive for $p>2$: written in powers of $p-2$ it has nine positive coefficients.}\label{tab:mean-cert} \begin{tabular}{@{}lll@{}} \toprule $\Psi_1$ & $\tau^{6}+1$ & $1175p^{8}+2755p^{6}+2352p^{4}+880p^{2}+128$\\ & $\tau^{5}+\tau$ & $5120p^{8}-3584p^{6}-3264p^{4}+1984p^{2}-256$\\ & $\tau^{4}+\tau^{2}$ & $-6635p^{8}+13877p^{6}+17808p^{4}+1808p^{2}-128$\\ & $\tau^{3}$ & $75776p^{8}-121856p^{6}+54912p^{4}-9344p^{2}+512$\\ \midrule $\Psi_2$ & $(\tau-1)^{4}$ & $30185p^{8}+32563p^{6}+16572p^{4}+8032p^{2}+128$\\ & $\tau(\tau-1)^{2}$ & $54180p^{8}+100044p^{6}+85680p^{4}+32256p^{2}$\\ & $\tau^{2}$ & $262836p^{8}-335160p^{6}+310464p^{4}$\\ \midrule $\Psi_3$ & $(\tau-1)^{4}$ & $489845p^{8}+647956p^{6}+361344p^{4}+123904p^{2}-1024$\\ & $\tau(\tau-1)^{2}$ & $804380p^{8}+1583824p^{6}+2050176p^{4}+1033216p^{2}-4096$\\ & $\tau^{2}$ & $3829896p^{8}-3081456p^{6}+2225664p^{4}+2128896p^{2}$\\ \bottomrule \end{tabular} \end{table} The proof has the structure of the proof of Theorem~\ref{thm:arc}, one radical, one rationalization and one sign discussion, and not that of Theorem~\ref{thm:global}, whose two nested radicals required a two-region certificate with $345$ terms: in the variable $\sqrt t$ three quartics suffice. \begin{remark}[Other members of the family; computer algebra]\label{rem:family} The same matching with fewer members gives lower orders. With $\mathrm A$ and one free conic mean, \[ q_7(t)=m_7\,(1+t)+n_7\sqrt{t-\varepsilon_7\,(t-1)^{2}},\qquad m_7=\frac{(h+3)(1-2h)}{4(4-h^{2})},\quad n_7=\frac{5(1+h)}{2(4-h^{2})},\quad \varepsilon_7=\frac{1-4h^{2}}{60}, \] has contact six: a single radical gives order seven, with $e_+=-\tfrac{(p^{2}-1)(p^{2}-4)(4p^{2}-1)}{302400}e^{7}+O(e^{9})$; for cube roots $q_7=\tfrac1{14}(1+t)+\tfrac67\sqrt{t-(t-1)^{2}/108}$. But the conic parameter is negative: the radicand is positive only for $w<60/(1-4h^{2})$, and the negative constant means that the state crosses the root at each step. The same happens for the combination of two free conic means, which has contact eight with one negative parameter. Imposing both $\mathrm A$ and $\mathrm G$ is what makes the third parameter $d$ positive in \eqref{eq:qM}. These statements are exact series computations; no convergence theorem is claimed for $q_7$. \end{remark}