% ========================================================================= \section{Geometric work: inventory, comparison and efficiency}\label{sec:work} Throughout this section $p\ge3$ is an integer (or the numerator $\mu$ of a rational degree, Proposition~\ref{prop:rational}), the state $R(s)$ is supplied, and counts follow Section~\ref{sec:convention}. For every support the power stage uses $2p-1$ parallels and the join $L_1B$, and transport adds the join $ME$ and two parallels (the parallel through $P$ and the horizontal readout). The supports differ in what they add: nothing for MB and AK, whose support lines are prepared; one arc and one join for centered and decentered AD; two joins for projective AD; three arcs and one join for the three-mean support; and the arithmetic of Table~\ref{tab:inventory} for the nested support. \subsection{Complete inventory of the nested support} Fixed preparation includes $a_9,b_9,c_9$ and $\lambda_9=(p-1)/X$; every residual-dependent length is rebuilt at each step. Obtain $E=R(s^{p})$ by the triangle stage, then $R(t)$ by the two-join fixed scale $X$. Stop if $t=1$. Otherwise use \begin{equation}\label{eq:w} w=t+\frac1t-2, \end{equation} which needs one five-join division and two rail sums or differences instead of a separate squaring. Then construct successively \[ K=1+\frac{b_9w}{1+c_9w},\qquad k_9=\sqrt{1\cdot K},\qquad \rho_9=1+\frac{a_9w}{k_9},\qquad q_9=\sqrt{t\,\rho_9}. \] Each mean proportional uses the semicircle macro of Section~\ref{sec:macros}, at most $2\jn+\pl+3\ar$ including diameter preparation and transfer to the rail. Scale $q_9$ to $\lambda_9q_9$ in two joins; its rail length is $H\lambda_9q_9$. One arc centred at $M$ transfers this length downward on the left rail and gives $D_9=(0,\;H-\tfrac HX(1+(p-1)q_9))$. Join $AD_9$ and $ME$, draw the parallel to $ME$ through $P$, and return horizontally. \begin{table}[h] \centering\small \caption{Per-step upper inventory of the nested support, from $R(s)$ to $R(s_+)$, integer $p\ge3$. A sum or nonnegative difference is one rail arc, a general division five joins, a fixed scale two joins.}\label{tab:inventory} \begin{tabular}{@{}lccc@{}} \toprule Stage & $\jn$ & $\pl$ & $\ar$\\ \midrule Triangle power endpoint $E=R(s^{p})$ & 1 & $2p-1$ & 0\\ Residual $t=Xs^{p}$ & 2 & 0 & 0\\ Width $w=t+1/t-2$ & 5 & 0 & 2\\ Inner argument $K$ (two scales, one division, two sums) & 9 & 0 & 2\\ First mean $k_9=\sqrt{1\cdot K}$ & 2 & 1 & 3\\ Outer argument $\rho_9$ (scale, division, sum) & 7 & 0 & 1\\ Second mean $q_9=\sqrt{t\rho_9}$ & 2 & 1 & 3\\ Scale $\lambda_9q_9$ and transfer from $M$ to $D_9$ & 2 & 0 & 1\\ Support join $AD_9$ & 1 & 0 & 0\\ Join $ME$, transport parallel, horizontal readout & 1 & 2 & 0\\ \midrule Total & 32 & $2p+3$ & 12\\ \bottomrule \end{tabular} \end{table} \begin{proposition}[Nested-radical work bound]\label{prop:inventory} For integer $p\ge3$, $X>0$ and a supplied state $R(s)$, the protocol above realizes the nested-radical correction with at most \begin{equation}\label{eq:inventory} 32\jn+(2p+3)\pl+12\ar=(4p+38)\jn+(6p+21)\ar, \end{equation} that is $10p+59$ traces per nonterminal step; $89$ traces for cube roots. \end{proposition} \begin{proof} The rail identities \eqref{eq:incid} establish every division and fixed scale. Both mean-proportional inputs are positive, and the prescribed circle intersection with the transferred length gives the positive root. The point $D_9$ gives $\kappa=1+(p-1)q_9$ in \eqref{eq:support}. Adding the ten rows gives the left side of \eqref{eq:inventory}, and $\pl=2\jn+3\ar$ gives the right side. Unit shortcuts, reused objects and an exact-root stop can only reduce the count. \end{proof} If charged, the final inversion adds $5\jn$ once, and supplying $u_0$ instead of $s_0$ another $5\jn$. \subsection{Inventory of the three-mean support} Fixed preparation consists of the four verticals $v,v_1,v_2,v_3$, the horizontal through $M$ and the points $F,G_2,G_3$ of Section~\ref{sec:mean-construction}, in a rectangle of the shape \eqref{eq:aspect}. The residual is not transferred to a rail: the line $ME$, drawn for the transport, carries it to the three verticals at no cost. \begin{table}[h] \centering\small \caption{Per-step upper inventory of the three-mean support, from $R(s)$ to $R(s_+)$, integer $p\ge3$.}\label{tab:mean-inventory} \begin{tabular}{@{}lccc@{}} \toprule Stage & $\jn$ & $\pl$ & $\ar$\\ \midrule Triangle power endpoint $E=R(s^{p})$ & 1 & $2p-1$ & 0\\ Arc 1: centre $J_1$, opening $J_3G_3$, onto $v$ at $D_1$ & 0 & 0 & 1\\ Arc 2: centre $J_2$ through $G_2$, onto the horizontal of $M$ at $Z$ & 0 & 0 & 1\\ Arc 3: centre $D_1$, opening $FZ$, onto $v$ at $D_M$ & 0 & 0 & 1\\ Support join $AD_M$ & 1 & 0 & 0\\ Join $ME$, transport parallel, horizontal readout & 1 & 2 & 0\\ \midrule Total & 3 & $2p+1$ & 3\\ \bottomrule \end{tabular} \end{table} \begin{proposition}[Three-mean work bound]\label{prop:mean-inventory} For integer $p\ge3$, $X>0$, a rectangle satisfying \eqref{eq:aspect} and a supplied state $R(s)$, the protocol of Section~\ref{sec:mean-construction} realizes the three-mean correction with at most \begin{equation}\label{eq:mean-inventory} 3\jn+(2p+1)\pl+3\ar=(4p+5)\jn+(6p+6)\ar, \end{equation} that is $10p+11$ traces per step; $41$ traces for cube roots. In an arbitrary rectangle one more join suffices. \end{proposition} \begin{proof} Section~\ref{sec:mean-construction} shows that the three arcs meet their lines at every state and that $AD_M$ has the slope $\tfrac{H}{WX}[1+(p-1)q_M]$, so that \eqref{eq:support} gives $\Tm_M$. The points $J_1,J_2,J_3$ and the two compass openings are intersections and distances of drawn or prepared objects, which are not charged. Adding the six rows gives the left side of \eqref{eq:mean-inventory}, and $\pl=2\jn+3\ar$ gives the right side. For the last assertion replace $J_1$ by the intersection of $v_1'$, the vertical at distance $K_n\sqrt{\zeta^{2}-1}$ from $v$, with the join of $E$ to a prepared centre chosen so that this intersection has the height $H-\tfrac\lambda\gamma[1+(p-1)m(1+t)]$. \end{proof} The bound relies on the convention of Section~\ref{sec:convention} more than the other inventories do: it uses seven prepared objects and three free intersections where the decentered arc uses three prepared objects, and it prescribes the shape of the rectangle. The comparison with the nested inventory, which was written without such shortcuts, should be read with this in mind (Section~\ref{sec:limits}). \subsection{Seven supports at a common state} \begin{table}[h] \centering\small \caption{Mobile work per step with the triangle power stage, integer $p\ge3$.}\label{tab:mobile} \begin{tabular}{@{}lccccll@{}} \toprule Construction & Order & $\jn$ & $\pl$ & $\ar$ & Traces & Scalar scope\\ \midrule MB rectangle & 1 & 2 & $2p+1$ & 0 & $10p+7$ & $X>1$\\ AK rectangle & 2 & 2 & $2p+1$ & 0 & $10p+7$ & global\\ AD rectangle & 3 & 3 & $2p+1$ & 1 & $10p+9$ & global\\ Projective AD & 4 & 4 & $2p+1$ & 0 & $10p+9$ & global\\ Decentered AD & 5 & 3 & $2p+1$ & 1 & $10p+9$ & global, $p>2$\\ Nested radicals & 9 & 32 & $2p+3$ & 12 & $10p+59$ & global, $p>2$\\ Three means & 9 & 3 & $2p+1$ & 3 & $10p+11$ & global, $p>2$\\ \bottomrule \end{tabular} \end{table} For cube roots the per-step bounds are $37,37,39,39,39,89,41$. Table~\ref{tab:cubic} runs all seven from the same state $X=2$, $s_0=3/4$ (so $u_0=4/3$, not the $u_0=1$ of the illustration) with the same stopping test on the reciprocal relative error. \begin{table}[h] \centering\small \caption{Steps / summed trace bounds to reach $|s/\alpha-1|\le10^{-d}$ for $p=3$, $X=2$, $s_0=3/4$. Each total is the per-step bound times the number of steps. Counts agree at 250 and 500 working digits.}\label{tab:cubic} \begin{tabular}{@{}lccccc@{}} \toprule Construction & Order & $d=6$ & $d=12$ & $d=30$ & $d=60$\\ \midrule MB rectangle & 1 & 8 / 296 & 17 / 629 & 44 / 1628 & 90 / 3330\\ AK rectangle & 2 & 3 / 111 & 4 / 148 & 5 / 185 & 6 / 222\\ AD rectangle & 3 & 2 / 78 & 3 / 117 & 3 / 117 & 4 / 156\\ Projective AD & 4 & 2 / 78 & 2 / 78 & 3 / 117 & 3 / 117\\ Decentered AD & 5 & 1 / 39 & 2 / 78 & 2 / 78 & 3 / 117\\ Nested radicals & 9 & 1 / 89 & 1 / 89 & 2 / 178 & 2 / 178\\ Three means & 9 & 1 / 41 & 1 / 41 & 2 / 82 & 2 / 82\\ \bottomrule \end{tabular} \end{table} For cube roots the nested support needs fewer steps but more traces than the decentered arc at every target in the table: $89,89,178,178$ against $39,78,78,117$. Higher order alone does not give a drawing-work advantage; the next subsection shows that for the nested support this is a statement about small $p$. The three-mean support takes the same numbers of steps as the nested support at $41$ traces each: $41,41,82,82$. Against the decentered arc it wins at $d=12$ and $d=60$ and loses by a few traces at $d=6$ and $d=30$, where both need the same number of steps. \subsection{Efficiency index} For an iteration of order $\varrho$ costing $N$ units per step, the classical efficiency index is $\varrho^{1/N}$~\cite{ostrowski,traub64}: the number of correct digits is asymptotically multiplied by $\varrho^{1/N}$ per unit of work. We compare its logarithm $\iota=\log\varrho/N(p)$, with $N(p)$ the trace bound of Table~\ref{tab:mobile}. \begin{proposition}[Efficiency ordering]\label{prop:index} For every integer $p\ge3$, $\iota_{\mathrm{AK}}<\iota_{\mathrm{AD}}<\iota_{\mathrm{proj}}<\iota_{\mathrm{dec}}$. The nested support satisfies $\iota_{\mathrm{nest}}>\iota_{\mathrm{dec}}$ if and only if \[ p>\frac{59\log5-9\log9}{10\,\log(9/5)}=12.79\ldots, \] that is, for integer degrees, if and only if $p\ge13$. It exceeds $\iota_{\mathrm{AD}}$ from $p=5$ and $\iota_{\mathrm{proj}}$ from $p=8$. The three-mean support satisfies $\iota_{M}>\iota_{\mathrm{dec}}$ and $\iota_{M}>\iota_{\mathrm{nest}}$ for every $p>0$: it has the largest index of the seven supports at every integer degree $p\ge3$. \end{proposition} \begin{proof} The last three of the first group share $N=10p+9$ and have increasing orders, and $\log2/(10p+7)<\log3/(10p+9)$ because $10p\log(3/2)>9\log2-7\log3$, whose right side is negative. The inequality $\log9/(10p+59)>\log5/(10p+9)$ is linear in $p$ and rearranges to the stated threshold; the comparisons with orders three and four give $p>4.1$ and $p>7.6$ in the same way. Finally $\log9/(10p+11)>\log5/(10p+9)$ reads $10p\log(9/5)>11\log5-9\log9$, whose right side is negative because $5^{11}<9^{9}$, and $\iota_M>\iota_{\mathrm{nest}}$ because the order is the same and $10p+11<10p+59$. \end{proof} \begin{table}[h] \centering\small \caption{$10^{3}\iota=10^{3}\log\varrho/N(p)$. The largest value in each column is in bold.}\label{tab:index} \begin{tabular}{@{}lccccccc@{}} \toprule Construction & $p=3$ & $5$ & $8$ & $12$ & $13$ & $20$ & $50$\\ \midrule AK & 18.73 & 12.16 & 7.97 & 5.46 & 5.06 & 3.35 & 1.37\\ AD & 28.17 & 18.62 & 12.34 & 8.52 & 7.90 & 5.26 & 2.16\\ Projective AD & 35.55 & 23.50 & 15.58 & 10.75 & 9.97 & 6.63 & 2.72\\ Decentered AD & 41.27 & 27.28 & 18.08 & 12.48 & 11.58 & 7.70 & 3.16\\ Nested radicals & 24.69 & 20.16 & 15.81 & 12.27 & 11.63 & 8.48 & 3.93\\ Three means & \textbf{53.59} & \textbf{36.02} & \textbf{24.15} & \textbf{16.77} & \textbf{15.58} & \textbf{10.41} & \textbf{4.30}\\ \bottomrule \end{tabular} \end{table} The reason is visible in Table~\ref{tab:mobile}: the power stage costs about $10p$ traces for every support, so the $50$ additional traces of the nested support are a fixed overhead that large $p$ amortizes, and the ratio of indices tends to $\log9/\log5\approx1.37$. The three-mean support has an overhead of two traces only, so it leads at every degree: the ratio of its index to that of the decentered arc is $1.30$ at $p=3$ and tends to the same limit. The index is an asymptotic quantity built from upper inventories; at a finite accuracy the integer number of steps makes the comparison lumpy, as Table~\ref{tab:p20} shows. \begin{table}[h] \centering\small \caption{Steps / summed trace bounds for $p=20$, $X=2$, $s_0=3/4$ (per-step bounds $209$, $259$ and $211$).}\label{tab:p20} \begin{tabular}{@{}lccccc@{}} \toprule Construction & $d=30$ & $d=60$ & $d=250$ & $d=1000$ & $d=2500$\\ \midrule Decentered AD & 4 / 836 & 5 / 1045 & 5 / 1045 & 6 / 1254 & 7 / 1463\\ Nested radicals & 3 / 777 & 3 / 777 & 4 / 1036 & 5 / 1295 & 5 / 1295\\ Three means & 3 / 633 & 3 / 633 & 4 / 844 & 4 / 844 & 5 / 1055\\ \bottomrule \end{tabular} \end{table} At $p=20$ the nested support has a smaller total than the decentered arc at four of the five targets and loses at $d=1000$, where saving a single step (five instead of six) does not pay for the higher step cost. The three-mean support has the smallest total at every target of the table. % ========================================================================= \section{Direct root enclosures for real degrees}\label{sec:enclosure} The power stage can be avoided altogether by constructing the root directly. The input is a positive length $X$ and a real degree $p$, also given as a length; the output is an approximation of $r=X^{1/p}$ with prescribed relative error. A real input length is assumed to be available: this does not assert that every real number is constructible from a unit segment. This calculator is the only one in the paper whose every step is a finite construction for irrational $p$. We first treat $X>1$ and $p\ge1$; the cases $X=1$ and $p=1$ are immediate, and Proposition~\ref{prop:allp} removes both restrictions. \subsection{One mean proportional per subdivision} Set $\theta=1/p$, $a=1$, $b=X$ and $\lambda=\theta$. Here $a,b$ are moving rail values. The invariant is \begin{equation}\label{eq:inv} 0\tfrac12.\end{cases} \end{equation} Both branches preserve \eqref{eq:inv}. Thus $p$ determines the choices and the weight, never the length of a power chain. After $n$ subdivisions \begin{equation}\label{eq:gap} \frac ba=X^{2^{-n}},\qquad z:=\frac ba-1\ge0. \end{equation} No logarithm, exponential or power $s^{p}$ is evaluated. The branch test compares two constructed lengths, $\lambda$ and $\tfrac12$; for irrational $p$ equality never occurs, and for dyadic $1/p$ the process stops with an exact endpoint. \begin{figure}[h] \centering\figSubdivision \caption{A subdivision constructs one mean proportional. The semicircle has diameter $OS=a+b$, centre $C$ and junction $OT=a$; the orange perpendicular $TG$ is the mean $m$. The same device is rotated onto the vertical rail in the counted protocol. The figure uses $a=1.2$, $b=2.8$ only to show the incidence.}\label{fig:subdiv} \end{figure} \subsection{A fifth-order final enclosure} The harmonic--arithmetic readout of the companion gallery~\cite{gallery}, $L_2=ab/[(1-\lambda)b+\lambda a]$ and $U_2=(1-\lambda)a+\lambda b$, encloses $r$ by the weighted arithmetic--geometric mean inequality, with \[ \frac{U_2}{L_2}-1=\lambda(1-\lambda)\frac{z^{2}}{1+z}\le\sinh^{2}\!\Bigl(\frac{\log X}{2^{n+1}}\Bigr), \] a width quadratic in $z$. A single rational readout of degree two gives a fifth-order width. It is the diagonal $[2/2]$ Pad\'e approximant of $(1+z)^{\tau}$ at zero~\cite{dlmf,baker}. For $0\le\tau\le1$ put \begin{equation}\label{eq:pade} \Pi_\tau(z)=12+6(2+\tau)z+(\tau+1)(\tau+2)z^{2},\qquad \psi_\tau(z)=\frac{\Pi_\tau(z)}{\Pi_{-\tau}(z)}, \end{equation} whose denominators are positive for $z\ge0$, and \begin{equation}\label{eq:L5} L_5=a\,\psi_\lambda(z),\qquad U_5=\frac{b}{\psi_{1-\lambda}(z)}. \end{equation} \begin{proposition}[Enclosure and finite termination]\label{prop:enclosure} For $X\ge1$, real $p\ge1$ and every subdivision depth, \begin{equation}\label{eq:width} L_5\le X^{1/p}\le U_5,\qquad \frac{U_5}{L_5}-1=\frac{\lambda(1-\lambda)(2-\lambda)(1+\lambda)\,z^{5}}{\Pi_\lambda(z)\,\Pi_{1-\lambda}(z)} \le\frac{z^{5}}{256}. \end{equation} Given $\eps>0$, choose a positive rational $\beta$ with $\beta^{5}/256\le\eps$, stop when $b\le(1+\beta)a$ and return $L_5$. The procedure terminates and its relative error is at most $\eps$. A sufficient depth depends on $X$ and $\eps$ only, not on $p$. \end{proposition} \begin{proof} For $0\le\tau\le1$, direct differentiation gives \[ (1+z)\bigl(\Pi_\tau'\Pi_{-\tau}-\Pi_\tau\Pi_{-\tau}'\bigr)-\tau\,\Pi_\tau\Pi_{-\tau} =-\tau(1-\tau^{2})(4-\tau^{2})\,z^{4}\le0. \] Hence $\log\psi_\tau(z)-\tau\log(1+z)$ is nonincreasing from zero, so $\psi_\tau(z)\le(1+z)^{\tau}$. With $\tau=\lambda$ this is $L_5\le r$; with $\tau=1-\lambda$ and $(1+z)^{\lambda}(1+z)^{1-\lambda}=1+z$ it is $r\le U_5$. Multiplying out gives the equality in \eqref{eq:width}. With $\chi=\lambda(1-\lambda)\le\tfrac14$ the numerator coefficient is $\chi(2+\chi)\le\tfrac9{16}$, and $\Pi_\lambda,\Pi_{1-\lambda}\ge12$, so the width is at most $z^{5}/256$. At the stopping test $0\le z\le\beta$, so $r/L_5-1\le\eps$ and $(r-L_5)/r\le\eps/(1+\eps)$. By \eqref{eq:gap}, $z\to0$, so the test is eventually met; any $n$ with $2^{n}\log(1+\beta)\ge\log X$ suffices. Equivalently $X\le(1+\beta)^{2^{n}}$, which can be checked in exact rational arithmetic for rational input. \end{proof} It suffices to draw $L_5$ alone; the analytic upper bound certifies its accuracy without constructing a second Pad\'e quotient. For $00$, $X\neq1$. \begin{enumerate}[label=(\roman*)] \item If $X>1$ and $01$, $p\ge1$.}\label{tab:subdiv} \begin{tabular}{@{}lccc@{}} \toprule Operation & $\jn$ & $\pl$ & $\ar$\\ \midrule Initial weight $\lambda=1/p$ & 5 & 0 & 0\\ $n$ subdivisions and weight updates & $4n$ & $n$ & $4n$\\ At most $n+1$ stopping tests & $2n+2$ & 0 & 0\\ One final lower Pad\'e readout & 49 & 0 & 7\\ \midrule Total, parallels expanded & $8n+56$ & 0 & $7n+7$\\ \bottomrule \end{tabular} \end{table} Thus $15n+63$ traces suffice for $X>1$, $p\ge1$. The other cases of Proposition~\ref{prop:allp} add: $5k+1$ traces in case (i), for the $k$ products $X^{2},\dots,X^{k+1}$ and the difference $1/p-k$; two five-join inversions in case (ii), giving $15n+73$ for $0