\documentclass[10pt,a4paper]{article} \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage{lmodern} \usepackage[a4paper,left=23mm,right=23mm,top=24mm,bottom=24mm]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools} \usepackage{graphicx,booktabs,array,tabularx,enumitem,xcolor} \usepackage[font=small,labelfont=bf]{caption} \usepackage{fancyhdr} \usepackage{microtype} \usepackage{placeins} \usepackage[colorlinks=true,linkcolor=blue!45!black,citecolor=blue!45!black,urlcolor=blue!45!black]{hyperref} \hypersetup{pdfauthor={Ivan Besevic},pdftitle={Geometric Root Calculators}} \input{figs} \pagestyle{fancy}\fancyhf{} \fancyhead[L]{\small Geometric root calculators} \fancyhead[R]{\small V29 / Research preprint} \fancyfoot[C]{\small\thepage} \renewcommand{\headrulewidth}{0.3pt} \setlength{\parindent}{0pt}\setlength{\parskip}{5pt plus 1pt} \setlist{itemsep=2pt,topsep=3pt} \theoremstyle{plain} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} % ---- notation macros --------------------------------------------------- \newcommand{\jn}{\mathsf{j}} % one join \newcommand{\pl}{\mathsf{p}} % one constructed parallel (= 2j + 3c) \newcommand{\ar}{\mathsf{c}} % one compass arc \newcommand{\Om}{\Omega} % projection centres on the top line \newcommand{\cE}{\mathcal{E}}\newcommand{\cC}{\mathcal{C}}\newcommand{\cD}{\mathcal{D}} \newcommand{\cU}{\mathcal{U}}\newcommand{\cV}{\mathcal{V}}\newcommand{\cR}{\mathcal{R}} \newcommand{\cJ}{\mathcal{J}}\newcommand{\cA}{\mathcal{A}}\newcommand{\cB}{\mathcal{B}} \newcommand{\cK}{\mathcal{K}}\newcommand{\cM}{\mathcal{M}}\newcommand{\cP}{\mathcal{P}} \newcommand{\Tm}{\mathrm{T}} \newcommand{\eps}{\varepsilon} \DeclareMathOperator{\disc}{disc} \begin{document} {\small\scshape Constructive geometry \,/\, root iterations \,/\, global convergence} \vspace{6pt} {\LARGE\bfseries Geometric Root Calculators\par} \vspace{3pt} {\large Radical supports of orders five and nine and real-degree enclosures in the Pandrosion rectangle\par} \vspace{8pt} Ivan Besevic\footnote{Computational assistance: the AI systems ChatGPT (OpenAI) and Claude (Anthropic) were used for derivations, proof development, figures and manuscript preparation. The author is responsible for the claims and their attribution. Section~\ref{sec:evidence} states, claim by claim, which results are formal proofs, written proofs, computer-algebra identities or numerical checks.} \quad$\cdot$\quad Independent research\\ Version 29 (revision of V28) \quad$\cdot$\quad 13 November 2026\\ \url{https://github.com/ivan-fr/pandrosion} \vspace{6pt} \begin{quote}\small \textbf{Abstract.} We study ruler-and-compass iterations for the positive root $X^{1/p}$ that share one rectangle, one power stage and one readout, and differ only in a \emph{support line}. Every support approximates the same target function $\kappa_*(t)=(t-1)/(t^{1/p}-1)$ of the residual $t$, and the contact order of the approximation fixes the order of the iteration. Besides the classical supports of orders one to four we treat three radical supports. A decentered arc gives a globally monotone fifth-order iteration. Two supports have local order \emph{exactly nine for every real $p>2$} and converge monotonically from every positive start: a support built from two successive mean proportionals, with error constant $7/648$ for cube roots, and a combination of the arithmetic mean, the geometric mean and one conic mean of $1$ and $t$, with error constant $23/20736$ for cube roots, drawn with three arcs. For integer $p\ge 3$ one step costs at most $10p+9$ elementary traces for the decentered arc, $10p+11$ for the three-mean support and $10p+59$ for the nested support; measured by the efficiency index, and under the stated counting conventions, the three-mean support is preferable for every $p\ge3$. We then separate what is a finite construction from what is only a scalar statement: the fast iterations are finite constructions for rational degrees, whereas for irrational $p$ a single step is not. For arbitrary real $p\neq 0$ a power-free subdivision with one final Pad\'e readout encloses the root with relative width at most $z^5/256$ in at most $15n+63$ traces after $n$ subdivisions (for $X>1$, $p\ge 1$). Global convergence and the local orders with their constants, for the three radical supports and every real $p>2$, are checked in Lean, as are the degree, approximation and efficiency statements; the drawing inventories are written protocols. \end{quote} % ========================================================================= \section{Introduction} A geometric iteration must realize its arithmetic by intersections, parallels and length transfers. The Pandrosion rectangle does this for $p$th roots: a chain of similar triangles produces the power $s^p$ of the current estimate, a line through a prepared point transports the residual, and a horizontal reads the next estimate on the same rail. The only part that changes from one method to another is the \emph{support}, a line through the corner $A$ whose slope encodes a correction factor. \paragraph{Results.} \begin{enumerate}[label=\arabic*.] \item \textbf{One target function} (Section~\ref{sec:coords}). All supports approximate $\kappa_*(t)=(t-1)/(t^{1/p}-1)$; contact of order $\omega$ at $t=1$ gives an iteration of order $\omega+1$ with an explicit error constant (Lemma~\ref{lem:contact}). Newton, Halley and the $[1/1]$ Pad\'e support are the cases $\omega=1,2,3$. \item \textbf{Decentered arc} (Section~\ref{sec:arc}). One moving arc and one join give $\omega=4$: a fifth-order iteration, monotone and convergent from every positive start for real $p>2$ (Theorem~\ref{thm:arc}). \item \textbf{Nested radicals} (Section~\ref{sec:nested}). Two successive mean proportionals give $\omega=8$. The iteration has order exactly nine for every real $p>2$, with the closed-form constant of Theorem~\ref{thm:order9}; it is globally monotone (Theorem~\ref{thm:global}). \item \textbf{Three means} (Section~\ref{sec:mean}). A combination of the arithmetic mean, the geometric mean and one conic mean of $1$ and $t$ also gives $\omega=8$. It is drawn with three arcs and one join, has order exactly nine with a smaller constant, and is globally monotone (Theorem~\ref{thm:mean}). \item \textbf{Work} (Section~\ref{sec:work}). Per-step inventories of the nested and three-mean supports (Propositions~\ref{prop:inventory} and~\ref{prop:mean-inventory}), a comparison of seven supports at a common state, and an efficiency-index comparison (Proposition~\ref{prop:index}): the nested support overtakes the decentered arc at $p=13$, and the three-mean support has the largest index for every $p\ge3$. \item \textbf{Degrees} (Sections~\ref{sec:degrees} and~\ref{sec:enclosure}). A classification of what is constructible in finitely many traces per step, summarized in Table~\ref{tab:degrees}, and a direct enclosure for every real $p\neq 0$ (Propositions~\ref{prop:enclosure} and~\ref{prop:allp}). \end{enumerate} \begin{table}[h] \centering\small \caption{Which calculator applies to which degree. ``Finite step'' means that one iteration step is a finite ruler-and-compass construction from the current state.}\label{tab:degrees} \begin{tabularx}{\linewidth}{@{}lXX@{}} \toprule Degree $p$ & Fast rectangle iterations (orders 2--9) & Direct subdivision (Section~\ref{sec:enclosure})\\ \midrule Integer $p\ge 3$ & Finite step; $10p+9$, $10p+11$ or $10p+59$ traces & Yes\\ Rational $p=\mu/\nu$ & Finite step, as degree $\mu$ applied to $X^{\nu}$ (Proposition~\ref{prop:rational}) & Yes\\ Irrational $p$ & Step not finite: $u^{p}$ is itself a real power. Available only through a rational approximation of $p$ (Proposition~\ref{prop:irrational}) & Yes; this is the recommended route\\ \bottomrule \end{tabularx} \end{table} \paragraph{What ``construct'' means here.} As in Pandrosion's recursive geometry, nothing in this paper constructs $X^{1/p}$ exactly in finitely many operations; $\sqrt[3]{2}$ is not constructible. A calculator is a rule whose every step is a finite construction and whose successive outputs converge to the root, with a stated rate or a stated enclosure. The distinction that matters is therefore not exact versus approximate, but whether \emph{each step} is finite. Scalar convergence theorems are stated for real $p$; drawing inventories are stated only where a step is a finite construction. \paragraph{Relation to earlier versions.} The classical supports and the decentered arc were introduced in the earlier preprint~\cite{v20}; they are restated here with complete proofs so that the comparison is self-contained. New in the present line of versions are the nested-radical support and its inventory. V28 added the contact lemma, the general ninth-order constant, the efficiency-index comparison, and the degree classification of Table~\ref{tab:degrees} with the rational and irrational cases treated separately. New in this revision is the three-mean support of Section~\ref{sec:mean}, with its Lean proofs, its construction and its inventory; the comparisons of Section~\ref{sec:work} are updated accordingly. The linear baseline, called MA in earlier versions, is renamed MB after its support line $MB$ (Appendix~\ref{app:mb}). Limits of the results are collected once, in Section~\ref{sec:limits}. % ========================================================================= \section{Coordinates, arithmetic and the common support equation}\label{sec:coords} \subsection{States and notation} For $X>0$ and $p\neq 0$ let $r=X^{1/p}$ be the root and $\alpha=1/r$ its reciprocal. The rectangle carries a \emph{reciprocal state} $s>0$ approaching $\alpha$; the \emph{direct state} is $u=1/s$. The \emph{residual} is \[ t=Xs^{p}=X/u^{p}, \qquad t=1 \iff s=\alpha . \] The logarithmic error is $e=\log(s/\alpha)$, so that $t=e^{pe}$. An iteration is written $s_+=s\,\Phi(t)$. Table~\ref{tab:notation} fixes the main symbols. \begin{table}[h] \centering\small \caption{Notation.}\label{tab:notation} \begin{tabularx}{\linewidth}{@{}lX@{}} \toprule $X,\ p,\ r,\ \alpha$ & input, degree, root $X^{1/p}$, reciprocal root $X^{-1/p}$\\ $s,\ u,\ t,\ e$ & reciprocal state, direct state $1/s$, residual $Xs^p$, log error $\log(s/\alpha)$\\ $h,\ \xi,\ x$ & $h=1/p$;\ \ $\xi=t-1$;\ \ $x=\log t$\\ $W,H$; $A,B,C,O$ & rectangle width and height; its corners (Section~\ref{sec:rect})\\ $R(q),L(q),\Om_f$ & right-rail point, left-rail point, projection centre on the top line\\ $M,E,P,T$ & prepared point, power endpoint $R(s^{p})$, state point $R(s)$, intersection on the support\\ $\kappa$, $\kappa_*$ & normalized support slope; exact slope $(t-1)/(t^{h}-1)$\\ $\jn,\pl,\ar$ & one join, one constructed parallel ($=2\jn+3\ar$), one compass arc\\ $q_5,\ q_9,\ q_M$ & normalized slopes $(\kappa-1)/(p-1)$ of the decentered, nested and three-mean supports\\ $m,g,n,d$;\ $\tau,\lambda$ & (Section~\ref{sec:mean} and its inventory only) coefficients of $q_M$;\ \ $\tau=\sqrt t$,\ \ $\lambda=H/X$\\ $a,b$ & generic positive rail values; in Section~\ref{sec:enclosure}, the moving endpoints\\ $m,\lambda,z$ & (Section~\ref{sec:enclosure} only) mean proportional $\sqrt{ab}$, weight, gap $b/a-1$\\ $\omega$ & order of contact of a support with $\kappa_*$ at $t=1$\\ $\mu,\nu$ & numerator and denominator of a rational degree $p=\mu/\nu$\\ \bottomrule \end{tabularx} \end{table} \subsection{Rectangle and line arithmetic}\label{sec:rect} In the rectangle $C=(0,0)$, $B=(W,0)$, $O=(0,H)$, $A=(W,H)$, with $W,H>0$, define \begin{equation}\label{eq:rails} R(q)=(W,H(1-q)),\qquad L(q)=(0,H(1-q)),\qquad \Om_f=\Bigl(\frac{Wf}{f-1},\,H\Bigr)\quad(f\neq 1). \end{equation} Line incidences perform multiplication and division: \begin{equation}\label{eq:incid} \Om_fR(q)\cap OC=L(fq),\qquad \Om_fL(b)\cap AB=R(b/f). \end{equation} Substituting the coordinates into the line equations proves both identities. Unit factors and zero values reuse existing points instead of a centre at infinity. \paragraph{Triangle power stage.} Let $B_j=(Ws^{j},H(1-s^{j}))\in OB$ and $L_j=L(s^{j})$. Horizontals transfer between these points and the rails. The parallel to $L_1B$ through $B_j$ meets the left rail at $L_{j+1}$, since its ordinate there is $H(1-s^{j})+H(1-s)s^{j}=H(1-s^{j+1})$. For integer $p$, repetition gives $E=R(s^{p})$ with $2p-1$ parallels and the join $L_1B$. \paragraph{Products.} For positive rail values $a,b$ with $b\neq 1$, draw $V=CR(b)\cap OA=\Om_{1/b}$ and project $R(a)$ successively through $\Om_{1/2}$, $V$, $\Om_{1/2}$, $\Om_{1/4}$. By~\eqref{eq:incid} the successive values are $L(a/2)$, $R(ab/2)$, $L(ab/4)$, $R(ab)$: a general product costs five joins. Division is symmetric: project $R(a)$ through $V$ to $L(a/b)$ and return through $\Om_{1/2},\Om_{1/4},\Om_{1/2}$, with values $R(2a/b)$, $L(a/(2b))$, $R(a/b)$; again five joins. Multiplication by a \emph{prepared} constant $d>0$ uses the two centres $\Om_{d/[2(1+d)]}$ and $\Om_{1/[2(1+d)]}$ and costs two joins. \subsection{The common support equation and its target} Prepare $M=(0,H(1-X^{-1}))$. With $P=R(s)$ and $E=R(s^{p})$, the line $ME$ has slope $\sigma=H(1-t)/(WX)$. Let a \emph{support} be a line through $A$ of slope $H\kappa/(WX)$. The parallel to $ME$ through $P$ meets the support at $T$, and the horizontal through $T$ meets the right rail at $P_+=R(s_+)$ with \begin{equation}\label{eq:support} s_+=s\,\frac{\kappa}{\kappa-1+t}. \end{equation} Indeed the two lines are $y-H=\tfrac{H\kappa}{WX}(x-W)$ and $y-H(1-s)=\sigma(x-W)$; eliminating $x$ gives $H-y=Hs\kappa/(\kappa-1+t)$. The update is exact, $s_+=\alpha$, precisely when $\kappa/(\kappa-1+t)=t^{-h}$, that is when \begin{equation}\label{eq:kstar} \kappa=\kappa_*(t):=\frac{t-1}{t^{h}-1}\qquad(\kappa_*(1)=p). \end{equation} The function $\kappa_*$ is not constructible from $t$; every support below is a constructible approximation of it. \begin{lemma}[Contact lemma; limit form formalized in Lean]\label{lem:contact} Let $\kappa(t)$ be analytic and positive near $t=1$ with $\kappa(t)-\kappa_*(t)=c\,\xi^{\omega}+O(\xi^{\omega+1})$, $\xi=t-1$, $\omega\ge1$. Then the iteration \eqref{eq:support} satisfies \[ e_+=c\,p^{\,\omega-1}e^{\,\omega+1}+O(e^{\,\omega+2}). \] \end{lemma} \begin{proof} From \eqref{eq:support}, $e_+=e+\log\kappa-\log(\kappa-1+t)$, and this vanishes for $\kappa=\kappa_*$. Since $\partial_\kappa[\log\kappa-\log(\kappa-1+t)]=(t-1)/[\kappa(\kappa-1+t)]$, which equals $\xi/p^{2}+O(\xi^{2})$ near $\kappa=p$, we get $e_+=c\,\xi^{\omega+1}/p^{2}+O(\xi^{\omega+2})$. Finally $\xi=e^{pe}-1=pe+O(e^{2})$. \end{proof} \begin{table}[h] \centering\small \caption{The supports as approximations of $\kappa_*$. Leading constants are for the reciprocal-state log error $e$; for odd orders they are unchanged in the direct state.}\label{tab:supports} \begin{tabular}{@{}llccl@{}} \toprule Support & $\kappa(t)$ & contact $\omega$ & order & leading constant of $e_+$\\ \midrule AK (Newton) & $p$ & 1 & 2 & $-(p-1)/2$\\[2pt] AD (Halley) & $\tfrac12[p+1+(p-1)t]$ & 2 & 3 & $(p^{2}-1)/12$\\[2pt] Projective AD & $\dfrac{2p[(2p-1)t+p+1]}{(p+1)t+5p-1}$ & 3 & 4 & $-(p^{2}-1)(2p-1)/72$\\[6pt] Decentered AD & $1+(p-1)q_5(t)$ & 4 & 5 & $(p-2)(p^{2}-1)(3p+1)/1440$\\[2pt] Nested radicals & $1+(p-1)q_9(t)$ & 8 & 9 & $c_9(p)$, Theorem~\ref{thm:order9}\\[2pt] Three means & $1+(p-1)q_M(t)$ & 8 & 9 & $c_M(p)$, Theorem~\ref{thm:mean}\\ \bottomrule \end{tabular} \end{table} AK is the value of $\kappa_*$ at $t=1$, AD its first-order Taylor polynomial and projective AD its $[1/1]$ Pad\'e approximant; these three are classical (Newton under inversion, Halley, and a Pad\'e correction) and are documented in Appendix~\ref{app:classical}. The linear MB construction uses the support $MB$, which does not pass through $A$, and is kept as a baseline in Appendix~\ref{app:mb}. \subsection{Work convention}\label{sec:convention} A \emph{trace} is one join or one compass arc. A constructed parallel expands to $\pl=2\jn+3\ar$, five traces. Intersections of drawn objects are free. \emph{Fixed preparation}, meaning every object that depends only on $p,X,W,H$ and not on the current state, is not charged, nor are decisions (comparisons of two constructed lengths) or the single final inversion that converts $s$ into $u=1/s$. Section \ref{sec:enclosure} charges its input and output inversions explicitly. All counts are upper bounds for the specified protocols; a trace count is neither a time nor an energy. The name Pandrosion refers to the addressee of Book~III of Pappus' \emph{Collection}~\cite{pappus}. Our rectangle follows a modern reconstruction~\cite{launay}; the accelerated supports are modern and are not attributed to the ancient text. % ========================================================================= \section{A decentered arc with global fifth-order convergence}\label{sec:arc} \subsection{Formula and construction} For real $p>2$ put \begin{equation}\label{eq:arc} S(t)=(p-2)(t^{2}+1)+(10p+4)t,\qquad q_5(t)=\sqrt{\frac{S(t)}{12p}},\qquad \Tm_{\mathrm{arc}}(s)=s\,\frac{1+(p-1)q_5(t)}{t+(p-1)q_5(t)}, \end{equation} that is, the support $\kappa=1+(p-1)q_5(t)$ in~\eqref{eq:support}. The radicand is positive for $t>0$. For integer $p\ge3$ the radical is produced by one arc. Put \[ \eta=(p-1)\sqrt{\frac{p-2}{12p}},\qquad \zeta=\frac{5p+2}{p-2}>1,\qquad d_0=\frac{W}{\eta},\qquad \delta=\frac{H}{X}\sqrt{\zeta^{2}-1}, \] and prepare \[ G_0=\Bigl(W,\;H+\frac{H\zeta}{X}\Bigr),\qquad O_{\mathrm{arc}}=\Bigl(W-d_0+\delta,\;H-\frac{H}{X\eta}\Bigr),\qquad\text{the vertical } x=W-d_0 . \] Draw the arc centred at $O_{\mathrm{arc}}$ with radius $EG_0=(H/X)(t+\zeta)$ and take its lower intersection $D_{\mathrm{arc}}$ with the prepared vertical. Its distance below the centre is \[ \sqrt{EG_0^{2}-\delta^{2}}=\frac{H}{X}\sqrt{t^{2}+2\zeta t+1}, \] so the slope of $AD_{\mathrm{arc}}$ is $\tfrac{H}{WX}\bigl[1+\eta\sqrt{t^{2}+2\zeta t+1}\bigr] =\tfrac{H}{WX}\bigl[1+(p-1)q_5(t)\bigr]$, because $(p-2)\cdot2\zeta=10p+4$. All preparation parameters use rational operations and square roots of positive quantities. Centered AD and decentered AD use the same mobile operations, one arc and one join; only the fixed preparation differs. This is the precise sense in which order five is obtained at the mobile cost of order three. \begin{figure}[h] \centering\figDecentered \caption{The decentered-arc device for $p=7$, $X=W=2$, $H=4$, $s=3/4$, drawn from the exact coordinates. (a) The whole circle. The compass opening $EG_0$ (thick orange segment) is carried to the prepared centre $O_{\mathrm{arc}}$; the circle meets the prepared vertical at $D_{\mathrm{arc}}$, and the right triangle with horizontal leg $\delta$ shows why its vertical leg is a square root. The small grey box is the rectangle at the same scale. (b) The rectangle enlarged: triangle chain (grey), transport lines (green), support $AD_{\mathrm{arc}}$ (blue) and the readout $P_+$. Each panel has equal axis scales.}\label{fig:arc} \end{figure} \subsection{Global monotone convergence} \begin{theorem}[Formalized in Lean]\label{thm:arc} For real $p>2$, $X>0$ and every $s_0>0$, the iteration $s_{n+1}=\Tm_{\mathrm{arc}}(s_n)$ converges to $\alpha$ monotonically without crossing it, and \begin{equation}\label{eq:arc-order} e_+=\frac{(p-2)(p^{2}-1)(3p+1)}{1440}\,e^{5}+O(e^{7}). \end{equation} \end{theorem} \begin{proof} Write $g=(p-1)q_5$. Since $S(1/t)=S(t)/t^{2}$ we have $q_5(1/t)=q_5(t)/t$, hence $\Phi(1/t)=1/\Phi(t)$ for $\Phi=(1+g)/(t+g)$: the log-error map $e\mapsto e_+$ is odd. Put \[ \Lambda(t)=\frac{\log t}{p}+\log(1+g(t))-\log(t+g(t)), \] so that $e_+=\Lambda(t)$, and define the polynomials \begin{align*} N_1(t)&=(p-1)(t^{2}+1)+(10p+2)t,\\ N_2(t)&=(t+1)\bigl(6pt-(t-1)^{2}\bigr),\\ N_3(t)&=(p^{2}-5p-2)(t-1)^{2}+12p(3p+1)t . \end{align*} Differentiation and factorization give the two identities \begin{align} \frac{d\Lambda}{d\log t}&=\frac{(p-2)(p-1)^{2}\,\bigl(q_5N_1-N_2\bigr)}{12p^{2}\,g(1+g)(t+g)},\label{eq:arc-der}\\ q_5^{2}N_1^{2}-N_2^{2}&=\frac{p+1}{12p}\,(t-1)^{4}\,N_3 .\label{eq:arc-id} \end{align} Here $N_1>0$. If $N_2\le0$ then $q_5N_1-N_2>0$. Otherwise $(t-1)^{2}<6pt$, and $N_3>0$: directly when $p^{2}-5p-2\ge0$, and otherwise from the rearrangement \[ N_3=6p^{2}(p+1)t+(p^{2}-5p-2)\bigl((t-1)^{2}-6pt\bigr), \] whose two terms are positive. Then \eqref{eq:arc-id} gives $q_5N_1>N_2$ for $t\neq1$. Hence $\Lambda$ is strictly increasing with $\Lambda(1)=0$: the new log error has the sign of the old one. Meanwhile $\Phi(t)-1=(1-t)/(t+g)$ 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 $\Phi(t)=1$ forces $t=1$. Near $t=1$, rationalize $q_5N_1-N_2$ with \eqref{eq:arc-id}. At $t=1$ one has $q_5=1$, $g=p-1$, $N_1=N_2=12p$ and $N_3=12p(3p+1)$, so \[ \lim_{t\to1}\frac{d\Lambda/d\log t}{(t-1)^{4}}=\frac{(p-2)(p^{2}-1)(3p+1)}{288\,p^{5}} . \] Substituting $t=e^{pe}$ and integrating yields the coefficient in \eqref{eq:arc-order}; oddness removes the $e^{6}$ term. \end{proof} At $p=3$ the constant is $1/18$. At $p=2$ the scalar formula is an exact square-root correction, but the displayed preparation is singular and is replaced by a single mean proportional. \input{nested} \input{mean} \input{work} \input{back} \end{document}