% ========================================================================= \section{Attribution and algebraic approximation}\label{sec:attribution} High local order and geometric iteration are not new principles. Traub organizes classical iteration formulas and their error terms~\cite{traub61,traub64}. Gander discusses Halley's method and radical corrections~\cite{gander}; Scavo and Thoo analyze its geometry~\cite{scavo}. Melman connects geometry with global convergence, and Amat, Busquier and Guti\'errez develop geometric constructions for nonlinear iteration~\cite{melman,amat}. Here AK is Newton under inversion and centered AD is Halley, realized in the shared rectangle; their scalar formulas are classical. Reading all supports as approximations of one function $\kappa_*$ (Lemma~\ref{lem:contact}) is the usual principle that the order of an iteration is the order of contact of its iteration function with the exact one, specialized to the rectangle. The decentered and nested supports are algebraic corrections defined by explicit positive radicals. Algebraic approximants generalize rational approximation and form an established methodology~\cite{sergeev,shafer}; the two-radical coefficient matching of Section~\ref{sec:nested} belongs to it. Section~\ref{sec:algebraic} proves a degree-four distinction at the cubic specialization, and Appendix~\ref{app:equiv} gives a specified quadratic comparison for the decentered support. The subdivision mean proportional is Euclid's~\cite{euclid}; its final readout is an ordinary Pad\'e approximant~\cite{dlmf,baker}, with the enclosure and width proved in Section~\ref{sec:enclosure}. Counting particular constructions follows the concerns of g\'eom\'etrographie and nomography~\cite{lemoine,ocagne}, and the efficiency index is Ostrowski's~\cite{ostrowski}. The three-mean support is a linear combination of three members of the family $q^{2}=t+a(t-1)^{2}$, and its method is not new either. Since $q_*$ is a homogeneous mean of $1$ and $t$, the support belongs to a classical subject, the approximation and bounding of means such as the logarithmic and extended means by combinations of the arithmetic, geometric and quadratic means~\cite{carlson,stolarsky}. Its coefficient matching is of Pad\'e type in a wider sense: in the variable $w$ it is a quadrature rule for the kernel $\sqrt{1+aw}$ with two prescribed nodes, in the way ordinary Pad\'e approximants of Stieltjes series are Gaussian quadratures for the kernel $1/(1+aw)$~\cite{baker}; and in the variable $\sqrt t+1/\sqrt t$ it is a quadratic algebraic approximant subject to one constraint (Appendix~\ref{app:equiv}). Raising the order of a root iteration by an affine combination of known iterations is also an established device, used for $n$th roots with rational iterations in~\cite{dubeau}. What this paper contributes is the stated rectangle realizations, their domains, the scalar convergence proofs with explicit constants, the protocol bounds, and the classification of degrees. % ========================================================================= \section{Evidence and reproducibility}\label{sec:evidence} \begin{table}[h] \centering\small \caption{Status of each claim. ``Lean'': checked by the Lean kernel, in the environments described below. ``CAS'': exact identity between polynomials or rational functions, recomputed with a computer-algebra system. ``Numerical'': high-precision evaluation, not a proof.}\label{tab:evidence} \begin{tabularx}{\linewidth}{@{}l>{\raggedright\arraybackslash}p{0.19\linewidth}>{\raggedright\arraybackslash}X@{}} \toprule Claim & Domain & Status\\ \midrule Lemma~\ref{lem:contact}, Table~\ref{tab:supports} & analytic supports & Lean (limit form); constants of AK, AD and projective AD: CAS\\ Theorem~\ref{thm:arc} & real $p>2$ & Lean (global convergence and the fifth-order limit); $O(e^{7})$ remainder written\\ Theorem~\ref{thm:order9} & real $p>2$ & Lean (limit $\hat e_+/\hat e^{\,9}\to c_9(p)$, positivity, $c_9(3)$); $O(\hat e^{\,11})$ remainder written\\ Theorem~\ref{thm:cubic} & $p=3$, $X>0$ & Lean\\ Theorem~\ref{thm:global} & real $p>2$ & Lean; certificate structure also CAS\\ Theorem~\ref{thm:mean} & real $p>2$ & Lean (positivity, global convergence, the limit $e_+/e^{9}\to c_M(p)$, positivity of $c_M$, $c_M(3)$); identities also CAS; $O(e^{11})$ remainder written\\ Remark~\ref{rem:family} & real $p>2$ & CAS (exact series)\\ Proposition~\ref{prop:rational} & rational $p$ & Lean, for the three radical supports\\ Proposition~\ref{prop:scalar} & real $p\neq0$ (scalar) & Lean\\ Proposition~\ref{prop:irrational} & real $p>0$ & Lean\\ Proposition~\ref{prop:inventory} & integer $p\ge3$ & written inventory; totals Lean; incidence checks\\ Proposition~\ref{prop:mean-inventory} & integer $p\ge3$ & written inventory; total Lean; incidence simulation\\ Proposition~\ref{prop:index}, Table~\ref{tab:index} & real and integer $p\ge3$ & Lean; table values numerical\\ Tables~\ref{tab:cubic}, \ref{tab:p20} & stated examples & numerical\\ Proposition~\ref{prop:enclosure} & $X\ge1$, $p\ge1$ & written proof; scalar part in Lean (retained)\\ Proposition~\ref{prop:allp} & real $p\neq0$, $X>0$ & Lean (invariant, gap, inversion)\\ Degree four (Section~\ref{sec:algebraic}) & $p=3$, $t=2$ & exact rational computation\\ Appendix~\ref{app:equiv}, three-mean paragraph & $p=3$ & exact rational computation (series only)\\ \bottomrule \end{tabularx} \end{table} \paragraph{Lean.} The nested-radical results have two project Docker/Lean audits totalling 104 theorems: 38 cubic, 14 real-parameter certificates, 44 real-parameter analysis and 8 degree-extension theorems. They use only \texttt{propext}, \texttt{Classical.choice} and \texttt{Quot.sound}, with no admitted proof. The sources, pinned toolchain, hashes and audit records accompany the paper~\cite{lean4,mathlib,supplNR}. The four nested-radical modules are those of V27-NR; for V28 they were recompiled from source in a separate environment (Lean~4.33.1, Mathlib~v4.33.1, Mathlib itself built from source), the count of 104 theorems was confirmed, and the main theorems (global convergence, the extension to every nonzero degree and the cubic limit) were checked to depend on these three axioms only. V28 added six modules, 125 theorems in all, with the same axioms and no admitted proof. \begin{itemize} \item \texttt{DecenteredArcGlobal} (30): Theorem~\ref{thm:arc}, global monotone convergence of the decentered arc for every real $p>2$, by the sign argument of Section~\ref{sec:arc}. \item \texttt{ContactLimit} (12): Lemma~\ref{lem:contact} in limit form, and the analytic tool used for the local orders: a finite sum $\sum_jc_je^{\lambda_jx}$ whose first $N$ moments $\sum_jc_j\lambda_j^{n}$ vanish is $O(x^{N})$. \item \texttt{NestedOrderNine} (21): Theorem~\ref{thm:order9}, the limit $\hat e_+/\hat e^{\,9}\to c_9(p)$ for every real $p>2$. The expansion \eqref{eq:Yseries} of $Y$ through $w^{4}$ is an exponential sum of 27 terms with twelve vanishing moments; the nested radical is expanded algebraically and the result is identified with \eqref{eq:D4}. \item \texttt{ArcOrderFive} (10): the fifth-order limit \eqref{eq:arc-order} with its constant. \item \texttt{DegreesEfficiency} (47), depending on Mathlib only: the exact contact identity; the positivity of $N_9$ and $c_9$, $c_9(3)=7/648$ and $c_9(p)=\tfrac12(p-1)p^{7}D_4(1/p)$; Propositions~\ref{prop:irrational} and~\ref{prop:index}, including $12<\text{crossover}<13$ from $9^{129}<5^{179}$ and $5^{189}<9^{139}$; the trace totals of Section~\ref{sec:work}; the invariant of Proposition~\ref{prop:allp}. \item \texttt{DegreesNested} (5): Proposition~\ref{prop:rational} for the nested support and for the decentered arc, without hypothesis and with integer powers only. \end{itemize} This revision adds three modules for the three-mean support, 93 theorems in all, compiled in the project Docker environment (Lean~4.33.1, Mathlib~v4.33.1), with no admitted proof and no warning; their main theorems were checked to depend on the same three axioms only. \begin{itemize} \item \texttt{ThreeMeanGlobal} (57): parts (i) and (ii) of Theorem~\ref{thm:mean} for every real $p>2$, in the variable $\sqrt t$: the identities \eqref{eq:mean-pos}, \eqref{eq:mean-slope} and \eqref{eq:mean-square}, the positivity of the nine certificate polynomials, the enclosure of the correction between $1$ and $t^{1/p}$, global convergence, and the agreement of the Lean definition with \eqref{eq:qM}. \item \texttt{ThreeMeanOrderNine} (29): part (iii), the limit $\hat e_+/\hat e^{\,9}\to c_M(p)$ for every real $p>2$, with $c_M>0$ and $c_M(3)=23/20736$. The radical is replaced by its quartic Taylor polynomial at an exact algebraic cost, and the remaining exponential sum has $22$ terms and ten vanishing moments (the tool of \texttt{ContactLimit}). \item \texttt{ThreeMeanWork} (7): the total $10p+11$ of Table~\ref{tab:mean-inventory}; the inequalities $\iota_M>\iota_{\mathrm{dec}}$ and $\iota_M>\iota_{\mathrm{nest}}$ for every $p>0$, the first from $5^{11}<9^{9}$; and Proposition~\ref{prop:rational} for the three-mean support. \end{itemize} The local orders are stated as limits of $e_+/e^{\,\omega+1}$; the refinements $O(e^{7})$ and $O(e^{11})$ of the remainders, which follow from oddness, are written arguments. \begin{center}\small \begin{tabularx}{\linewidth}{@{}lX@{}} \toprule Lean module & Certified scope\\ \midrule \texttt{NestedRadicalCubic} & Cubic global convergence and the relative-error limit $7/648$, for every positive target.\\ \texttt{NestedRadicalRealCertificate} & Exact polynomial identities and positive-coefficient certificates.\\ \texttt{NestedRadicalReal} & Original formula for every real $p>2$; support equivalence and the small-degree counterexample.\\ \texttt{NestedRadicalAllDegrees} & Degree doubling, preservation of the reported start and convergence for every $p\neq0$.\\ \texttt{DecenteredArcGlobal} & New in V28: global convergence of the decentered arc, real $p>2$.\\ \texttt{ContactLimit} & New in V28: contact lemma in limit form; exponential sums with vanishing moments.\\ \texttt{NestedOrderNine} & New in V28: order nine with constant $c_9(p)$ for every real $p>2$.\\ \texttt{ArcOrderFive} & New in V28: order five of the decentered arc with its constant.\\ \texttt{DegreesEfficiency} & New in V28: contact identity, constants, irrational degrees, efficiency ordering and crossover, trace totals, subdivision start for every positive degree.\\ \texttt{DegreesNested} & New in V28: rational degrees for the decentered and nested iterations.\\ \texttt{ThreeMeanGlobal} & New in V29: positivity and global convergence of the three-mean support, real $p>2$.\\ \texttt{ThreeMeanOrderNine} & New in V29: order nine of the three-mean support with constant $c_M(p)$.\\ \texttt{ThreeMeanWork} & New in V29: trace total, efficiency inequalities and rational degrees for the three-mean support.\\ \texttt{GeometricSubdivision} & Scalar interval invariant, real-degree rational bracket and width; stored seven-declaration audit, not recompiled for this revision~\cite{supplS}.\\ \bottomrule \end{tabularx} \end{center} \paragraph{Computer algebra and numerics.} The accompanying script \texttt{verify\_v29.py} recomputes, with exact rational arithmetic: the identities \eqref{eq:arc-der} (at sample points), \eqref{eq:arc-id}, \eqref{eq:cubic-id}; the series \eqref{eq:Yseries}--\eqref{eq:D4} and the constants of Table~\ref{tab:supports}; the positivity of $N_9(2+\rho)$; the factorization defining $\cP$ in Theorem~\ref{thm:global}, its degrees, and the positivity and term counts $345$, $49$, $24$, $112$ of the four certificates; the Pad\'e identity and the width equality \eqref{eq:width}; and the appendix formulas. Numerically it checks \eqref{eq:Dprime} against a numerical derivative, the constant $c_9(p)$ against the iteration at $p=2.5,3,4,7,20$, the monotone enclosure $\min(u,r)2$, both globally monotone: the nested radicals, with constant $c_9(p)$, and the three means, with the smaller constant $c_M(p)$ and a mobile cost of three arcs and one join. In trace counts, under the stated conventions, the three-mean support has the largest efficiency index at every degree $p\ge3$; between the two others the decentered arc leads for $p\le12$ and the nested support for $p\ge13$. These iterations are finite constructions for rational degrees; for a general real degree the power-free subdivision with a Pad\'e readout gives a certified enclosure of relative width at most $z^{5}/256$ in $15n+63$ traces. Open questions: \begin{enumerate}[label=\arabic*.] \item Formalize a cost semantics for the drawing protocols, so that the inventories, and not only their totals, are machine-checked, and so that prepared objects can be weighed. \item Shorten the nested inventory. Its overhead over the decentered arc is $50$ traces, of which $19$ build the two arguments $K$ and $\rho_9$ and $20$ the two mean proportionals. The three-mean support shows that order nine does not require this overhead; whether the nested formula itself has a short protocol is open. \item The family of Section~\ref{sec:mean} gives order seven with a single radical (Remark~\ref{rem:family}), but on a bounded range of residuals and with a state that crosses the root. Is there a single-radical support of order greater than five that is defined and monotone for every $t>0$? Which order do three conic means, or four, reach with positive radicands? \item Analyse the two-level recursion in which the residual $u^{p}$ of an irrational degree is itself enclosed by subdivision: which inner accuracy preserves the outer order? \item Analyse conditioning and finite drawing error. \item Compare the exact formulas with earlier algebraic approximants and with the literature on means under specified transformations. \item Extend to the three-mean support the half-plane contraction proved for the decentered arc in the complex companion~\cite{companion}. Numerically the boundary criterion of that paper holds for $p$ from $2.0001$ to $10^{4}$, with $\sup_y\operatorname{Arg}\varphi(iy)\,p/\pi\approx0.503$ against the bound $1$; this is an observation, not a proof. \end{enumerate} % ========================================================================= \FloatBarrier \appendix \section{Classical rectangle supports}\label{app:classical} \subsection{AK and centered AD: orders two and three} \emph{AK: a fixed line.} Prepare $K=(W(1-X/p),0)$. The support $AK$ has slope $Hp/(WX)$, so $\kappa=p$ and $\Tm_{\mathrm{AK}}(s)=ps/(p-1+t)$, Newton's method under inversion. It has order two and converges from every positive start. \emph{AD: a centered length transfer.} Prepare \[ O_{\mathrm{AD}}=\Bigl(W-\frac{2W}{p-1},\;H-\frac{H(p+1)}{X(p-1)}\Bigr). \] Transfer $AE=Hs^{p}$ downward from this centre to $D_{\mathrm{AD}}$ and join $AD_{\mathrm{AD}}$. Then $\kappa=\tfrac12[p+1+(p-1)t]$, giving \begin{equation}\label{eq:AD} \Tm_{\mathrm{AD}}(s)=s\,\frac{p+1+(p-1)t}{p-1+(p+1)t}. \end{equation} This Halley map has cubic order and global monotone convergence. Earlier geometric interpretations appear in~\cite{gander,scavo}; here the rail encoding and readout are shared with AK. \begin{figure}[h] \centering\figAK \caption{Pandrosion AK, order two, at $p=3$, $X=2$, $s=3/4$ ($W=2$, $H=4$). The fixed support $AK$ (blue) replaces the MB support, while the triangle chain (grey) and the correction direction (green) are retained. The inset magnifies the dashed window around $P$, $T$ and $P_+$.}\label{fig:ak} \end{figure} \begin{figure}[h] \centering\figAD \caption{Pandrosion AD, order three, at the same state. The orange length $AE$ is transferred onto the prepared vertical through $O_{\mathrm{AD}}$; the lower arc point determines the support $AD_{\mathrm{AD}}$ (blue). Here $O_{\mathrm{AD}}=C$.}\label{fig:ad} \end{figure} Both drawings retain the triangle stage, the green transport and the horizontal readout; only the support changes. \subsection{Projective AD: order four} Two joins supply the support. Prepare the horizontal line and the point \[ \ell:\ y=H\Bigl(1+\frac{5p-1}{(p+1)X}\Bigr),\qquad Z_{\mathrm{proj}}=\Bigl(W\Bigl(1+\frac{5p-1}{2p(2p-1)}\Bigr),\;H\Bigl(1+\frac{p+1}{(2p-1)X}\Bigr)\Bigr). \] (The line $\ell$ of this appendix is unrelated to the length $\ell$ of Figure~\ref{fig:means}.) Set $D_{\mathrm{proj}}=Z_{\mathrm{proj}}E\cap\ell$ and join $AD_{\mathrm{proj}}$. Substitution in \eqref{eq:support} gives the $\kappa$ of Table~\ref{tab:supports} and \begin{equation}\label{eq:proj} \Tm_{\mathrm{proj}}(s)=s\,\frac{2p\bigl((2p-1)t+p+1\bigr)}{Q_{\mathrm{proj}}(t)},\qquad Q_{\mathrm{proj}}(t)=(p+1)t^{2}+2(2p-1)(p+1)t+(2p-1)(p-1). \end{equation} The direct-root correction is the $[2/1]$ Pad\'e approximant; the reciprocal-state factor displayed here is $[1/2]$. Direct differentiation gives \[ \Tm_{\mathrm{proj}}'(s)=-\frac{2p(p^{2}-1)(2p-1)(t-1)^{3}}{Q_{\mathrm{proj}}(t)^{2}}, \] so the map has its global maximum $\alpha$ at $s=\alpha$. Also $\Tm_{\mathrm{proj}}(s)/s-1=(1-t)\bigl((p+1)t+5p-1\bigr)/Q_{\mathrm{proj}}(t)$. Thus a first step lands at or below $\alpha$ and subsequent steps increase to $\alpha$; the order is exactly four. The earlier preprint gives further dynamics and the original triangle-chain realization~\cite{v20}. \begin{figure}[h] \centering\figProjective \caption{Projective AD, order four, at the same cubic state. The orange line $Z_{\mathrm{proj}}E$ meets the prepared horizontal $\ell$ at $D_{\mathrm{proj}}$; the blue support $AD_{\mathrm{proj}}$, extended below $A$, receives the green correction. The inset magnifies the dashed window and separates $P$, $T$ and $P_+$.}\label{fig:proj} \end{figure} \section{The linear MB reference}\label{app:mb} MB is a baseline, not an accelerated method. Its support is the line $MB$, from the prepared point $M$ to the lower-right vertex $B$; it does not pass through $A$. Earlier versions called this construction MA: the name referred to the same lower-right vertex, which is $B$ in the labelling of this paper. The parallel to $ME$ through $P$ meets $MB$, and a horizontal returns the next state. For $s\neq1$, \[ \Tm_{\mathrm{MB}}(s)=1-\frac{X-1}{X\,\Sigma_p(s)},\qquad \Sigma_p(s)=\sum_{j=0}^{p-1}s^{j}. \] For $X>1$ this map is increasing and $\Tm_{\mathrm{MB}}(s)-s=(1-Xs^{p})/(X\Sigma_p(s))$, so it converges monotonically to $\alpha$ from every positive start, with the removable extension at $s=1$. Its derivative at $\alpha$ is $1-p\alpha^{p-1}/\Sigma_p(\alpha)\in(0,1)$: convergence is linear. \begin{figure}[h] \centering\figMB \caption{Pandrosion MB, the linear reference construction, at $p=3$, $X=2$, $s=3/4$. Thin lines construct the power by similar triangles. The green parallel meets the blue support $MB$; the horizontal gives $P_+$. At $s=1$ the geometric intersection is coincident despite the removable scalar extension.}\label{fig:mb} \end{figure} \FloatBarrier \section{Equivalence questions for the radical corrections}\label{app:equiv} Whether the arc is a known formula ``up to conjugacy'' depends on the class of coordinate changes allowed. Let $\varphi$ denote the correction factor $\Phi_{\mathrm{arc}}(t)=(1+g)/(t+g)$ with $g=(p-1)q_5$. Eliminating the radical gives the quadratic relation \begin{equation}\label{eq:quad} P_p(t,\varphi):=(p-1)^{2}S(t)(\varphi-1)^{2}-12p\,(t\varphi-1)^{2}=0 . \end{equation} The explicit radical in \eqref{eq:arc} selects the branch; the polynomial alone loses sign information and has a double root at $t=1$. \paragraph{Homographic changes.} For each real $p>2$, $\Phi_{\mathrm{arc}}$ is not rational in $t$. Indeed $\disc_tS=96p(p+1)$ and $S(0)=p-2>0$, so $S$ has two simple nonzero roots and is not a square in $\mathbb{C}(t)$; if the correction were rational, so would be $\sqrt{S(t)/(12p)}=(1-t\Phi_{\mathrm{arc}})/[(p-1)(\Phi_{\mathrm{arc}}-1)]$. Invertible homographies of the input or output preserve rationality. For integer $p\ge3$, $S(y^{p})$ also has only simple roots, so the state map $y\,\Phi_{\mathrm{arc}}(y^{p})$ is not rational and cannot be M\"obius-conjugate to a rational state map, even as a germ at $y=1$. This state-map exclusion is asserted only for integer degrees. \paragraph{A specified quadratic normalization.} A diagonal type $(2,2,2)$ quadratic Hermite--Pad\'e (Shafer) form for $f(t)=t^{-1/p}$ requires polynomials of degree at most two with $A(t)f^{2}+B(t)f+C(t)=O((t-1)^{8})$~\cite[eqs.~(1.4)--(1.6)]{cabay}. Direct substitution into \eqref{eq:quad} gives instead, with $\xi=t-1$, \[ P_p\bigl(1+\xi,(1+\xi)^{-1/p}\bigr)=\frac{(p-2)(p-1)^{2}(p+1)(3p+1)}{60p^{4}}\,\xi^{6}+O(\xi^{7}), \] contact six, not eight. The coefficients of $P_p$ have no common factor and its discriminant in $\varphi$ is $48p(p-1)^{2}(t-1)^{2}S(t)$, so it is irreducible over $\mathbb{C}(t)$; any polynomial with the same degree bounds defining this branch is a constant multiple of $P_p$. This excludes that particular diagonal approximant in these coordinates, without excluding other algebraic families or transformed targets. \paragraph{Local analytic conjugacy.} Distinct formulas may share a local analytic normal form: B\"ottcher's theorem conjugates a superattracting germ $F(e)=ce^{\varrho}+O(e^{\varrho+1})$, $c\neq0$, to its monomial normal form~\cite{buff}. This supplies neither a finite incidence protocol nor a global domain or drawing inventory. The explicit construction and its proved dynamics are therefore distinct from a claim of a new local conjugacy class. \paragraph{The three-mean correction as a constrained quadratic approximant.} Put $\sigma=\sqrt t+1/\sqrt t$, so that $w=\sigma^{2}-4$. Then \[ \frac{q_5}{\sqrt t}=\sqrt{1+a_9(\sigma^{2}-4)},\qquad \frac{q_M}{\sqrt t}=m\,\sigma+g-n\sqrt{1+d\,(\sigma^{2}-4)} , \] and $y=q_M/\sqrt t$ satisfies $(y-m\sigma-g)^{2}=n^{2}\bigl(1+d(\sigma^{2}-4)\bigr)$, a quadratic relation whose coefficients have degrees $0,1,2$ in $\sigma$. In this variable the three-mean support is therefore a quadratic algebraic approximant of $q_*/\sqrt t$ at $\sigma=2$ in the sense of~\cite{shafer}, but not the Hermite--Pad\'e approximant of its type: its discriminant is constrained to have no term of degree one in $\sigma$. The constraint is what keeps the radicand a quadratic polynomial in $t$, that is, one arc. The unconstrained form has one more free coefficient. For $p=3$ it is \[ \frac{16}{33}-\frac{\sigma}{66}+\frac1{99}\sqrt{5\sigma^{2}+736\sigma+1424} =\frac{q_*}{\sqrt t}+\frac{(\sigma-2)^{5}}{19131876}+O\bigl((\sigma-2)^{6}\bigr), \] one order more than \eqref{eq:mean-D4}, where the error is $\tfrac{23}{354294}(\sigma-2)^{4}$; but multiplied by $t$ its radicand contains $\sqrt t$, so the corresponding support is a nested radical. This is an exact rational computation on the series at $p=3$; the corresponding iteration has not been analysed. % ========================================================================= \begin{thebibliography}{99}\small \bibitem{pappus} Pappus of Alexandria. \emph{Book III of the Mathematical Collection}. English translation by John B. Little, following Hultsch; College of the Holy Cross, Bookshelf 63, 2023. \url{https://crossworks.holycross.edu/hc_books/63/}. \bibitem{launay} Micka\"el Launay. Pandrosion et la duplication du cube. Micmaths, video presentation, 2024. \url{https://www.youtube.com/watch?v=7T4neM660SM}. Rectangle construction and correction, approximately 16:37--19:10. \bibitem{sergeev} Alexei V. Sergeev and David Z. Goodson. Summation of asymptotic expansions of multiple-valued functions using algebraic approximants: application to anharmonic oscillators. \emph{J. Phys. A: Math. Gen.} 31:4301--4317, 1998. \url{https://avsergeev1.narod.ru/papers_37_iop_paper.pdf}. \bibitem{euclid} Euclid. \emph{Elements}, Book VI, Proposition 13. Online edition by David E. Joyce, Clark University. \url{https://mathcs.clarku.edu/~djoyce/elements/bookVI/propVI13.html}. Accessed 17 October 2026. \bibitem{gallery} Ivan Besevic. Real exponents: geometric subdivision. Pandrosion Geometry Workshop, companion interactive construction, 2026. \url{https://ivan-fr.github.io/pandrosion/gallery/geometric-subdivision.html}. Local source snapshot included with the V26-S supplement. \bibitem{dlmf} NIST Digital Library of Mathematical Functions. Section 3.11(iv): Pad\'e approximations. \url{https://dlmf.nist.gov/3.11#iv}. Accessed 17 October 2026. \bibitem{baker} George A. Baker, Jr. and Peter Graves-Morris. \emph{Pad\'e Approximants}, 2nd ed. Encyclopedia of Mathematics and its Applications 59, Cambridge University Press, 1996. doi:10.1017/CBO9780511530074. In particular Section~1.5, Theorems 1.5.2--1.5.3, on argument and value homographic invariance. \bibitem{traub61} J. F. Traub. On a class of iteration formulas and some historical notes. \emph{Comm. ACM} 4(6):276--278, 1961. doi:10.1145/366573.366606. \bibitem{traub64} J. F. Traub. \emph{Iterative Methods for the Solution of Equations}. Prentice-Hall, 1964. \bibitem{ostrowski} A. M. Ostrowski. \emph{Solution of Equations and Systems of Equations}. Academic Press, New York, 1960. \bibitem{gander} Walter Gander. On Halley's iteration method. \emph{Amer. Math. Monthly} 92(2):131--134, 1985. doi:10.2307/2322644. \bibitem{scavo} T. R. Scavo and J. B. Thoo. On the geometry of Halley's method. \emph{Amer. Math. Monthly} 102(5):417--426, 1995. doi:10.2307/2975033. \bibitem{melman} A. Melman. Geometry and convergence of Euler's and Halley's methods. \emph{SIAM Review} 39(4):728--735, 1997. doi:10.1137/S0036144595301140. \bibitem{amat} S. Amat, S. Busquier and J. M. Guti\'errez. Geometric constructions of iterative functions to solve nonlinear equations. \emph{J. Comput. Appl. Math.} 157(1):197--205, 2003. doi:10.1016/S0377-0427(03)00420-5. \bibitem{shafer} R. E. Shafer. On quadratic approximation. \emph{SIAM J. Numer. Anal.} 11(2):447--460, 1974. doi:10.1137/0711037. \bibitem{lemoine} \'Emile Lemoine. \emph{G\'eom\'etrographie ou Art des constructions g\'eom\'etriques}. Scientia 18, C. Naud, Paris, 1902. \url{https://catalogue.bnf.fr/ark:/12148/cb36049032t}. \bibitem{ocagne} Maurice d'Ocagne. \emph{Esquisse d'ensemble de la nomographie}. M\'emorial des sciences math\'ematiques 4, Gauthier-Villars, 1925. \url{https://www.numdam.org/item/MSM_1925__4__1_0/}. \bibitem{lean4} Leonardo de Moura and Sebastian Ullrich. The Lean 4 theorem prover and programming language. In \emph{CADE 28}, LNCS 12699, pages 625--635. Springer, 2021. doi:10.1007/978-3-030-79876-5\_37. \bibitem{mathlib} The mathlib Community. The Lean mathematical library. In \emph{CPP 2020}, pages 367--381. ACM, 2020. doi:10.1145/3372885.3373824. \url{https://leanprover-community.github.io/papers/mathlib-paper.pdf}. \bibitem{supplNR} Ivan Besevic. Nested-radical Pandrosion support: derivation, geometric example and Lean convergence certificates. Source bundle accompanying \emph{Geometric Root Calculators}, V27-NR, 18 October 2026. Includes \texttt{NestedRadicalCubic}, \texttt{NestedRadicalRealCertificate}, \texttt{NestedRadicalReal}, \texttt{NestedRadicalAllDegrees}, audit scripts, source hashes and numerical checks. \bibitem{supplS} Ivan Besevic. Geometric subdivision and Pad\'e readout: supplementary source and verification records. Source bundle accompanying \emph{Geometric Root Calculators}, V26-S, 2026. Includes \texttt{GeometricSubdivision.lean}, its retained axiom audit and the independent subdivision verifier. \bibitem{v20} Ivan Besevic. Reciprocal root geometry, version 20. Research preprint, 2026. \url{https://github.com/ivan-fr/pandrosion/tree/main/paper/reciprocal_geometry_v20}. \bibitem{cabay} Stan Cabay and George Labahn. A fast, reliable algorithm for calculating Pad\'e--Hermite forms. Technical Report CS-89-09, University of Waterloo, 1989. \url{https://cs.uwaterloo.ca/research/tr/1989/CS-89-09.pdf}. Equations (1.4)--(1.6), printed page 1, PDF page 5. \bibitem{buff} Xavier Buff, Adam L. Epstein and Sarah Koch. B\"ottcher coordinates, 2011. \url{https://arxiv.org/abs/1104.2981}. Introduction: the classical one-variable theorem. \bibitem{carlson} B. C. Carlson. The logarithmic mean. \emph{Amer. Math. Monthly} 79:615--618, 1972. \bibitem{stolarsky} Kenneth B. Stolarsky. Generalizations of the logarithmic mean. \emph{Math. Mag.} 48(2):87--92, 1975. doi:10.2307/2689825. \bibitem{dubeau} Fran\c{c}ois Dubeau. Newton's method and high-order algorithms for the $n$th root computation. \emph{J. Comput. Appl. Math.} 224:66--76, 2009. doi:10.1016/j.cam.2008.04.014. \bibitem{companion} Ivan Besevic. Complex geometric root calculators, version C3. Research preprint, companion to this paper, 13 November 2026. \end{thebibliography}