\documentclass[11pt]{amsart} \usepackage{amsmath} \usepackage{amssymb} \usepackage{amsthm} \usepackage{booktabs} \usepackage{array} \usepackage{hyperref} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \setlength{\emergencystretch}{2em} \newcommand{\NN}{\mathbb{N}} \newcommand{\ZZ}{\mathbb{Z}} \newcommand{\QQ}{\mathbb{Q}} \newcommand{\RR}{\mathbb{R}} \newcommand{\CC}{\mathbb{C}} \newcommand{\cO}{\mathcal{O}} \newcommand{\cD}{\mathcal{D}} \newcommand{\cC}{\mathcal{C}} \newcommand{\cP}{\mathcal{P}} \newcommand{\Nm}{\operatorname{N}} \newcommand{\eu}{\phi_{\mathrm E}} \newcommand{\Res}{\operatorname{res}} \newcommand{\lean}[1]{{\def\_{\textunderscore\allowbreak}\texttt{#1}}} \begin{document} \hypersetup{pdftitle={Linear independence of reciprocal and reciprocal-cube sums of Lucas sequences}, pdfauthor={Shawn Krisman}} \title[Reciprocal and reciprocal-cube sums]{Linear independence of reciprocal and reciprocal-cube sums of Lucas sequences} \author{Shawn Krisman} \subjclass[2020]{Primary 11J72; Secondary 11B39, 33D15, 41A21, 68V20} \keywords{Lucas sequences, Fibonacci numbers, Pell numbers, linear independence, irrationality, $q$-series, Hankel determinants, Lean} \begin{abstract} Let $\eta>1$ be a unit of a real quadratic order, of norm $Q=\pm1$, and let $U_n=(\eta^n-\bar\eta^n)/(\eta-\bar\eta)$ be its Lucas sequence. We prove that for each sign $\delta=\pm1$ the three numbers $1$, $\sum_{n\ge1}\delta^n/U_n$ and $\sum_{n\ge1}(\delta Q)^n/U_n^3$ are linearly independent over $\QQ$. For the Fibonacci numbers this says that $1$, $\sum_n(-1)^n/F_n$, $\sum_nF_n^{-3}$ are linearly independent, and so are $1$, $\sum_nF_n^{-1}$, $\sum_n(-1)^n/F_n^3$; for $F_{2n}$ it gives $1$, $\sum_nF_{2n}^{-1}$, $\sum_nF_{2n}^{-3}$. In particular $\sum_n(-1)^n/U_n^3$ is irrational for every such sequence. The proof is the Hankel-determinant construction of our earlier papers on $\sum_nU_n^{-3}$, run with the operator $c_3\theta(\theta+1)+c_1$, where the integers $(c_3,c_1)$ come from a hypothetical linear relation, and with a sign attached to the weights. The whole argument is formalised in Lean~4 with Mathlib, using only the standard axioms. \end{abstract} \maketitle % =========================================================================== \section{Introduction} % =========================================================================== For integers $P,Q$ let $U_n=U_n(P,Q)$ be the Lucas sequence $U_0=0$, $U_1=1$, $U_{n+1}=PU_n-QU_{n-1}$. When $Q=\pm1$ and $D=P^2-4Q>0$ is not a square, $U_n=(\eta^n-\bar\eta^n)/\sqrt D$ with $\eta=(P+\sqrt D)/2$ a unit of the order $\ZZ[\eta]$ of norm $\eta\bar\eta=Q$. Every unit $\eta>1$ of a real quadratic order arises in this way, with $P=\eta+\bar\eta\ge1$ if $Q=-1$ and $P\ge3$ if $Q=1$; we call these pairs $(P,Q)$ \emph{admissible}. The Fibonacci numbers are $U_n(1,-1)$, the Pell numbers $\cP_n$ are $U_n(2,-1)$, and $F_{kn}/F_k=U_n(L_k,(-1)^k)$, where $L_k$ is the $k$-th Lucas number. \begin{theorem}\label{thm:main} Let $(P,Q)$ be admissible, $U_n=U_n(P,Q)$, and $\delta\in\{1,-1\}$. Then the numbers \[ 1,\qquad A_\delta=\sum_{n\ge1}\frac{\delta^n}{U_n},\qquad B_\delta=\sum_{n\ge1}\frac{(\delta Q)^n}{U_n^3} \] are linearly independent over $\QQ$. \end{theorem} The four cases are: \begin{center} \small \begin{tabular}{llll} \toprule $Q=\Nm(\eta)$ & $\delta$ & $A_\delta$ & $B_\delta$\\ \midrule $+1$ ($P\ge3$) & $+1$ & $\sum1/U_n$ & $\sum1/U_n^3$\\ $+1$ & $-1$ & $\sum(-1)^n/U_n$ & $\sum(-1)^n/U_n^3$\\ $-1$ ($P\ge1$) & $-1$ & $\sum(-1)^n/U_n$ & $\sum1/U_n^3$\\ $-1$ & $+1$ & $\sum1/U_n$ & $\sum(-1)^n/U_n^3$\\ \bottomrule \end{tabular} \end{center} So the signs of the two sums agree when $\Nm(\eta)=1$ and differ when $\Nm(\eta)=-1$. The method forces this pairing; the other pairings, for example $1$, $\sum_nF_n^{-1}$, $\sum_nF_n^{-3}$, are not covered. \begin{corollary}\label{cor:main} \textup{(i)} For every admissible $(P,Q)$, $\sum_{n\ge1}(-1)^n/U_n(P,Q)^3$ is irrational. \textup{(ii)} The numbers $1$, $\sum_n(-1)^n/F_n$, $\sum_nF_n^{-3}$ are linearly independent over $\QQ$, and so are $1$, $\sum_nF_n^{-1}$, $\sum_n(-1)^n/F_n^3$. \textup{(iii)} For $k\ge1$ and $\delta=\pm1$, the numbers $1$, $\sum_n\delta^n/F_{kn}$, $\sum_n(\delta(-1)^k)^n/F_{kn}^3$ are linearly independent over $\QQ$. In particular $1$, $\sum_nF_{2n}^{-1}$, $\sum_nF_{2n}^{-3}$ are linearly independent. \textup{(iv)} The numbers $1$, $\sum_n(-1)^n/\cP_n$, $\sum_n\cP_n^{-3}$ are linearly independent over $\QQ$, and so are $1$, $\sum_n\cP_n^{-1}$, $\sum_n(-1)^n/\cP_n^3$. \end{corollary} \begin{proof} (i) Take $\delta=-Q$, so that $B_\delta=\sum_n(-1)^n/U_n^3$; a relation $B_\delta=r$ with $r\in\QQ$ would be a nontrivial relation among $1,A_\delta,B_\delta$. (ii) and (iv) are the cases $(P,Q)=(1,-1)$ and $(2,-1)$ with $\delta=\mp1$. (iii) The pair $(L_k,(-1)^k)$ is admissible ($L_k\ge1$, and $L_k\ge3$ for $k\ge2$), and $U_n=F_{kn}/F_k$ by Binet's formula. So $A_\delta=F_k\sum_n\delta^n/F_{kn}$ and $B_\delta=F_k^3\sum_n(\delta(-1)^k)^n/F_{kn}^3$, and multiplying the members of a family by nonzero rationals does not affect linear independence over $\QQ$. \end{proof} \subsection*{Context} The weight-one sums are classical. Andr\'e-Jeannin proved that $\sum_nx^n/w_n$ is irrational for a class of binary recurrences $w_n$ and integers $x$ that includes every $U_n(P,\pm1)$ with $x=\pm1$, in particular $\sum_n\delta^n/U_n\notin\QQ$ \cite{AJ89}. Linear independence results for weight-one Lambert series at Pisot-type algebraic integers, with applications to reciprocal sums of Fibonacci and Lucas numbers, were proved by Bundschuh and V\"a\"an\"anen \cite{BV05}, and for reciprocal sums of Fibonacci and Lucas numbers along certain sequences of pairwise coprime indices, such as the squares of primes, by Duverney, Suzuki and Tachiya \cite{DST20}. For the even powers, Duverney, Ke.~Nishioka, Ku.~Nishioka and Shiokawa proved that $\sum_nU_n^{-2s}$ ($s\ge1$) is transcendental for both signs of $Q$ \cite{DNNS97}, and Elsner, Shimomura and Shiokawa determined the algebraic relations among such numbers \cite{ESS07,ESS08,ESS12}. These sums are values of quasi-modular forms at algebraic points, to which Nesterenko's theorem \cite{Nes96} applies (see \cite[\S3.4]{Wal10} for the Fibonacci zeta function). Wang \cite[Cor.~5.2]{Wan25} expresses the generalised sums $\sum_nn^pU_n^{-s}$ through derivatives of Eisenstein series when $p\equiv s\pmod2$ and $Q=1$, and when $p$ and $s$ are even and $Q=-1$ (in his notation the sign is $\alpha\beta=\pm1$), and derives algebraic independence results for them; his Cor.~5.2 concerns the non-alternating sums. The sums of Theorem~\ref{thm:main} have $p=0$ and $s\in\{1,3\}$, so the non-alternating ones are outside these parity classes. Wang also remarks that some alternating sums $\sum_n(-1)^{n-1}n^pU_n^{-s}$ admit such expressions under parity conditions that he does not state; we have not found such an expression for $p=0$, $s\in\{1,3\}$, and we know of no modular interpretation of the sums considered here. For weight three, $\sum_nF_n^{-3}\notin\QQ$ is the main result of \cite{Kri26}, and $\sum_nU_n^{-3}\notin\QQ$ for every admissible $(P,Q)$ is proved in \cite{Kri26b}; both papers have Lean formalisations. The alternating sums $\sum_n(-1)^{n-1}U_n^{-3}$, for the Lucas sequence of every quadratic unit $\eta>1$, (and the analogous sums over $U_{kn}$) were shown to be irrational by twinheterodox \cite{Twin26}, in a preprint that builds on \cite{Kri26b}, is described by its author as primarily AI-generated, and has no formal proof. Corollary~\ref{cor:main}(i) is that statement, and the case $\delta=Q$ of Theorem~\ref{thm:main} with no weight-one term contains the main result of \cite{Kri26b}. Mixed-weight independence results are known for Lambert-type series at integer bases: Bundschuh and V\"a\"an\"anen \cite{BV05b} prove that $1$, $\sum_n1/(q^n-1)$, $\sum_n(-1)^{n-1}/(q^n-1)^2$ are linearly independent over $\QQ$ for suitable rational $q$, and a preprint posted while this paper was being completed \cite{Twin26b} (described by its author as primarily AI-generated; no formal proof) claims the linear independence of $1$ and $\sum_n(b^n-1)^{-j}$, $j=1,2,3$, for integers $b\ge2$. Since $b^n-1=(b-1)U_n(b+1,b)$, these are reciprocal sums of Lucas sequences with $Q=b$, not $\pm1$, and they do not cover the sums considered here. For the Lucas sequences of quadratic units we are not aware of an earlier linear independence statement involving a weight-one and a weight-three reciprocal sum, such as $1$, $\sum_n(-1)^n/F_n$, $\sum_nF_n^{-3}$. \subsection*{The method} Suppose $q_0+q_1A_\delta+q_3B_\delta=0$ with integers $q_i$, not all zero. Then $(q_1,q_3)\ne(0,0)$, and we put $(c_3,c_1)=(Dq_3,2q_1)$. With $t=\eta^{-1}$, $\rho=\delta t$ and $w=Qt^2$, the number \begin{equation}\label{eq:Xi} \Xi=\sum_{N\ge1}\rho^N\Bigl(\frac{2c_3w^N}{(1-w^N)^3}+\frac{c_1}{1-w^N}\Bigr) =\frac{2c_3B_\delta+c_1DA_\delta}{D\sqrt D} \end{equation} is then $-2q_0/\sqrt D$, a rational multiple of $\sqrt D$ (Lemma~\ref{lem:reduction}). We show that this is impossible for every $(c_3,c_1)\in\ZZ^2\setminus\{0\}$, by the construction of \cite{Kri26,Kri26b}: a moving-denominator Hankel determinant in the line of Uvarov \cite{Uva69}, Krattenthaler \cite{Kra21}, Pr\'evost \cite{Pre96} and Zudilin \cite{Zud16,Zud17}, which \cite{Kri26} adapted from an unrefereed, AI-assisted research draft on a $q$-zeta value \cite{UPL,QZ3}. The linear functional is now \[ L(g)=\sum_{n\ge1}\rho^n\,[\Theta g](w^n),\qquad\Theta=c_3\theta(\theta+1)+c_1,\quad\theta=s\tfrac{d}{ds}, \] and its moments against $1/(1-w^as)$ are $\rho^{-a}(\Xi-S_a)$, with partial sums $S_a$ of \eqref{eq:Xi}. The Hankel determinants $\Delta_K(X)$ of $L$ against $D_K(s)=\prod_{a\le K}(1-w^as)$ satisfy four estimates: \begin{center} \small \begin{tabular}{lll} \toprule & size (per $K^3\log\eta$) & \\ \midrule real decay $|\Delta_K(\Xi)|$ & $-2/9$ & Proposition~\ref{prop:decay}\\ conjugate $|\sigma\Delta_K(-\Xi)|$ & $-1$ & Proposition~\ref{prop:conj}\\ denominators $\log|\Nm C_K(t)|$ & $\le c_*=1.1954\ldots<11/9$ & Propositions~\ref{prop:global}, \ref{prop:cost}\\ nonvanishing & infinitely many $K$ & Proposition~\ref{prop:nonvan}\\ \bottomrule \end{tabular} \end{center} Then $C_K(t)D^K\Delta_K(\Xi)$ is a nonzero element of $\ZZ[\eta]$ of norm $<1$ for large $K$, a contradiction. Compared with \cite{Kri26b}, which is the case $\delta=Q$, $(c_3,c_1)=(1,0)$, the changes are these. The weight-one part $c_1$ of the operator has no factor $\theta$; this costs a factor $|\rho|^m=t^m$ in the decay estimate instead of the factor $|w|^{m-R}$ that $\theta$ supplies (Section~\ref{sec:decay}). The moment $\nu_0=L(1)$ no longer vanishes, so the roots of $1-\alpha_0=1-\rho$ join the denominators (Section~\ref{sec:den}); the root types depend on $\delta$, but the elementary root count used in the proof does not. In the conjugate estimate the generating function $\Gamma$ of the tails acquires a weight-one part with $\Gamma(0)\ne0$ (Section~\ref{sec:conj}). The Cauchy transform of $L$ has simple poles when $c_3=0$, and is still not rational (Section~\ref{sec:nonvan}). Everything else is unchanged, and we cite \cite{Kri26b} for the steps that are literally the same. The constants in the estimates do not depend on $(P,Q,\delta)$; they depend on $(c_3,c_1)$ only through lower-order terms. Applied to the unit $\eta^k$, whose Lucas sequence is $U_{kn}/U_k$, Theorem~\ref{thm:main} gives the analogous statement for the sums over $U_{kn}$, with the sign $\delta Q^k$ in place of $\delta Q$. \subsection*{Formal verification} The theorem, its corollaries and every step of the proof are formalised in Lean~4 with Mathlib (\url{https://github.com/telaviv/lucas-reciprocal-independence}; Section~\ref{sec:lean}), using only the standard axioms. The analytic ingredients are formalised for every real $t$ with $0a}\rho^N\bigl(2c_3w^N/(1-w^N)^3+c_1/(1-w^N)\bigr)= \rho^{-a}\sum_{N>a}\psi_N$. \end{proof} Let $D_K(s)=\prod_{a=1}^K(1-w^as)$, with poles $x_a=w^{-a}$, $|x_a|=\eta^{2a}$. For a polynomial $f_0$ of degree $<2K-1$, \[ \frac{f_0}{D_K}=g_0+\sum_{a=1}^K\frac{\Res_a(f_0)}{1-w^as},\qquad \Res_a(f_0)=\frac{f_0(x_a)}{\prod_{b\ne a}(1-w^{b-a})}, \] with $g_0$ a polynomial of degree $\le\deg f_0-K$. Let $L_X$ be the linear map with $L_X(s^k)=\nu_k$ and $L_X(1/(1-w^as))=\lambda_a(X-S_a)$, $X$ an indeterminate, and \[ \Delta_K(X)=\det\bigl[L_X(s^{i+j}/D_K)\bigr]_{0\le i,jK$ and $\zeta$ is an $\alpha$-root, then $-v(\Delta_{K,\tau})\le\max(0,K-1-k_{\min}(d))$. \end{lemma} \begin{proof} Work in the monomial basis. Since $e>K$, no difference $|a-b|a}\psi_N$. Write $N=a+M$; then $w^N=r_aw^M$ and $r_a^{-1}\delta'^at^a\rho^N=r_a^{-1}(\delta'\delta)^at^{2a}\delta^Mt^M=r_a^{-1}w^a\delta^Mt^M=\delta^Mt^M$. Hence the pole moment is $\Gamma(r_a)$, where \[ \Gamma(r)=2c_3r\sum_{M\ge1}\frac{\delta'^Mt^{3M}}{(1-rw^M)^3}+c_1\sum_{M\ge1}\frac{\delta^Mt^M}{1-rw^M} =\sum_{n\ge0}\gamma_nr^n, \] using $\delta^Mt^Mw^M=\delta'^Mt^{3M}$. The first sum is the function $\Gamma$ of \cite[Proposition~5.1]{Kri26b} with the sign $\delta'^M$; the second is new. $\Gamma$ is analytic on $|r|<\eta^2$, and $\gamma_n=\gamma_n^{(3)}+\gamma_n^{(1)}$ with \[ \gamma_n^{(3)}=c_3n(n+1)\sum_{M\ge1}\delta'^Mt^{3M}w^{M(n-1)}\ (n\ge1),\qquad\gamma_0^{(3)}=0, \] \[ \gamma_n^{(1)}=c_1\sum_{M\ge1}\delta^Mt^Mw^{Mn}\ (n\ge0). \] So $|\gamma_n^{(3)}|\le|c_3|n(n+1)t^{2n+1}/(1-t^3)$ and $|\gamma_n^{(1)}|\le|c_1|t^{2n+1}/(1-t)$. Unlike in \cite{Kri26b}, $\Gamma(0)=\gamma_0=c_1\delta t/(1-\delta t)$ is nonzero when $c_1\ne0$. Partial fractions give \[ \frac{s^k}{F^\sigma(s)}=\sum_{j=0}^{k-K}h_{k-K-j}(r)\,s^j+\sum_{a=1}^K\frac{r_a^k}{(F^\sigma)'(r_a)(s-r_a)}, \] and with $\sum_ar_a^m/(F^\sigma)'(r_a)=h_{m-K+1}(r)$ and $|r_a|\le t^2<\eta^2$ we get \[ Y_k(-\Xi)=\sum_{j=0}^{k-K}h_{k-K-j}(r)\,\sigma\nu_j+\sum_{n\ge0}\gamma_nh_{k+n-K+1}(r). \] The term $n=0$ is $\gamma_0h_{k-K+1}(r)$, which vanishes for $kt^2), \] using $\theta(\theta+1)(z-s)^{-1}=2sz/(z-s)^3$. Since $|\rho^n|=t^n$, the series converges locally uniformly on $\CC\setminus(\{0\}\cup\{w^n\})$, so $\cC$ is meromorphic on $\CC\setminus\{0\}$; it is analytic at $\infty$ and $O(1/z)$. \emph{$\cC$ is not rational.} The nodes $w^n$ are pairwise distinct, as $0<|w|<1$, and near $w^n$ all terms but the $n$-th are analytic. If $c_3\ne0$, $\cC$ has a pole of order exactly $3$ at $w^n$, with leading coefficient $2c_3\rho^nw^{2n}\ne0$ (the weight-one term has order $\le1$). If $c_3=0$, then $c_1\ne0$ and $\cC$ has a simple pole at $w^n$ with residue $c_1\rho^n\ne0$. Either way $\cC$ has infinitely many poles. The rest is the argument of \cite[Proposition~6.1]{Kri26b}, which uses only that $\cC$ is not rational, that $L$ is a linear combination of $2$-jets at the nodes $w^n$ with weights $O(t^n)$, that $0<|w|<1$, and that the exterior nodes $x_a=w^{-a}$ are distinct with $|x_a|\to\infty$; we repeat it. For $\deg R\le n-1$ put $H_{n,R}(z)=L\bigl(R(s)/((z-s)D_n(s))\bigr)$, analytic for $|z|>t^2$ and in particular at every $x_a$, and \[ T_{n,R}(z)=L\Bigl(\frac{R(z)D_n(s)-D_n(z)R(s)}{(z-s)D_n(s)}\Bigr), \] a polynomial in $z$ of degree $\le n-1$. Then $\cC R-T_{n,R}=D_nH_{n,R}$. If $\Delta_n(\Xi)=0$, a kernel vector gives $R_n\ne0$ of degree $\le n-1$ with $L(s^jR_n/D_n)=0$ for $jt^2+\epsilon_0$ ($\epsilon_0>0$), gives $H_{n,R_n}(z)=O(z^{-n-1})$, so $E_n=O(z^{-1})$ and $E_n(x_a)=0$ for $a\le n$. If $\Delta_n(\Xi)=\Delta_{n+1}(\Xi)=0$, then $V_0=T_{n+1}R_n-T_nR_{n+1}=R_{n+1}E_n-R_nE_{n+1}$ is a polynomial, $O(z^{n-1})$, vanishing at $x_1,\dots,x_n$; so $V_0=0$ and $T_n/R_n=T_{n+1}/R_{n+1}$. If $\Delta_n(\Xi)=0$ for all $n\ge N$, all $T_n/R_n$ equal one reduced fraction $T/R_0$, so $R_n=U'_nR_0$, $T_n=U'_nT$, and $E:=\cC R_0-T\ne0$ because $\cC$ is not rational. Then $U'_nE=E_n=O(1/z)$ forces $E=c_0z^{-q}(1+o(1))$ with $c_0\ne0$, $q\ge1$, and $\deg U'_n\le q-1$; but $E(x_a)\ne0$ for all large $a$ while $E_n(x_a)=0$, so $U'_n(x_a)=0$ for all large $a\le n$, too many zeros once $n$ is large. \end{proof} % =========================================================================== \section{\texorpdfstring{Proof of Theorem~\ref{thm:main}}{Proof of the theorem}}\label{sec:proof} % =========================================================================== Suppose $q_0+q_1A_\delta+q_3B_\delta=0$ with integers $q_i$, not all zero (a rational relation becomes one after clearing denominators). Fix $(c_3,c_1)=(Dq_3,2q_1)\ne(0,0)$ as in Lemma~\ref{lem:reduction}, so that $D\Xi\in\cO$ and $\sigma\Xi=-\Xi$. For $K$ in the infinite set of Proposition~\ref{prop:nonvan} put \[ y_K=C_K(t)\,D^K\,\Delta_K(\Xi) \] (here $D^K$ is the $K$-th power of the discriminant, not the polynomial $D_K$). By Proposition~\ref{prop:global}, $C_K(t)\Delta_K(X)=\sum_{j\le K}b_jX^j$ with $b_j\in\cO$, and $D^K\Xi^j=D^{K-j}(D\Xi)^j\in\cO$; so $y_K\in\cO$. Moreover $y_K\ne0$, because $\Delta_K(\Xi)\ne0$ and $C_K(t)\ne0$ ($t$ is not a root of unity). Since $\Xi\in\QQ(\sqrt D)$, $\sigma(\Delta_K(\Xi))=(\sigma\Delta_K)(\sigma\Xi)=\sigma\Delta_K(-\Xi)$, so $\sigma y_K=C_K(\sigma t)D^K\sigma\Delta_K(-\Xi)$. The norm $y_K\,\sigma y_K$ of the nonzero algebraic integer $y_K$ is a nonzero rational integer, so $|y_K\,\sigma y_K|\ge1$, and by Propositions~\ref{prop:decay}, \ref{prop:conj} and \ref{prop:cost} \begin{multline*} 0\le\log|\Nm C_K(t)|+2K\log D+\log|\Delta_K(\Xi)|+\log|\sigma\Delta_K(-\Xi)|\\ \le\bigl(c_*-\tfrac29-1\bigr)K^3\log\eta+O(K^2\log K), \end{multline*} where the implied constant depends on $D$ and $(c_3,c_1)$, which are fixed. Since $c_*-\tfrac{11}9=\tfrac{11}8\beta_7-\tfrac89=-0.02684\ldots<0$, this fails for large $K$. \qed \begin{remark}\label{rem:numerics} Not used in the proof. With the constants of Remark~\ref{rem:sharp} the margin is at least $0.0914$ ($Q=-1$) and $0.1214$ ($Q=1$) per $K^3\log\eta$, for both $\delta$ and all $(c_3,c_1)$. We checked the construction numerically, as a cross-check only. (1) The exact multiplicities of the cyclotomic factors in the denominators of $\Delta_{K,\tau}(X)$, computed in $\QQ(\tau)[X]$, never exceed the bounds $b_{K,d}$: for all four $(Q,\delta)$, several pairs $(c_3,c_1)$ and $K\le9$ (1645 pairs $(K,d)$), and in an independent computation for $K\le10$ (2388 pairs); more than half of the bounds are attained. The terms $k=0$ are needed: for $Q=-1$ and $c_1\ne0$ the multiplicity of $\Phi_2$ ($\delta=-1$) or $\Phi_1$ ($\delta=1$) is $4$ at $K=2$ and $8$ at $K=3$, above the values $3$ and $7$ of the bound without these terms. (2) At finite $K$, with exact determinants over $\QQ(\sqrt D)$ and interval arithmetic, the left side above without the term $2K\log D$, and with the norm of the ideal generated by the coefficients of $\Delta_K$ in place of $C_K(t)$, divided by $K^3\log\eta$, is negative for $6\le K\le11$ in all cases tried; at $K=11$ it is about $-0.22$ for the Fibonacci numbers ($\delta=1$, $(c_3,c_1)=(1,2)$), $-0.21$ for the Pell numbers ($\delta=1$, $(1,1)$) and $-0.25$ for $(P,Q)=(3,1)$ ($\delta=-1$, $(3,-2)$). For $c_3=0$ the measured decay of $\Delta_K(\Xi)$ is much faster than $-2/9$, about $-0.70$ per $K^3\log\eta$. The scripts are available from the author. \end{remark} % =========================================================================== \section{Formal verification}\label{sec:lean} % =========================================================================== The Lean development is the repository \begin{center} \url{https://github.com/telaviv/lucas-reciprocal-independence} \end{center} (Lean code as of commit \texttt{73a88f1}); all file names below refer to it. It (library \lean{LucasInd}, Lean~4.35.0, Mathlib v4.35.0, about $10\,100$ lines) generalises the formalisation of \cite{Kri26b} and proves \begin{center} \lean{theorem LucasInd.linearIndependent\_sums (hA : Admissible a E) (h$\delta$ : $\delta$ = 1 $\vee$ $\delta$ = -1) :}\\ \lean{LinearIndependent $\mathbb Q$ ![(1 : $\mathbb R$), sumA a E $\delta$, sumB a E $\delta$]}, \end{center} where \lean{Admissible a E} means $E=1$, $a\ge1$ or $E=-1$, $a\ge3$; \lean{Useq a E} is the sequence with $U_{n+2}=aU_{n+1}+EU_n$, so $(a,E)=(P,-Q)$; and \lean{sumA a E $\delta$} and \lean{sumB a E $\delta$} are the series $\sum_{n\ge0}\delta^{n+1}/U_{n+1}$ and $\sum_{n\ge0}(\delta\cdot(-E))^{n+1}/U_{n+1}^3$, written with a \lean{tsum}. Thus $\varepsilon:=-E=Q$ and \lean{sumB} is $B_\delta$. \lean{LucasInd.rat\_relation\_trivial} restates the theorem as: a relation $q_0+q_1A_\delta+q_3B_\delta=0$ with $q_i\in\QQ$ forces $q_0=q_1=q_3=0$. The proof uses only the axioms \lean{propext}, \lean{Classical.choice} and \lean{Quot.sound}; there is no \lean{sorry}, no \lean{axiom} declaration, no \lean{native\_decide} and no \lean{implemented\_by}. \emph{What has to be trusted.} For the theorem itself, only the definitions of \lean{Admissible}, \lean{Useq}, \lean{sumA} and \lean{sumB} in \lean{LucasInd/Defs.lean} (and Mathlib's \lean{tsum} and \lean{LinearIndependent}) have to be read. A divergent \lean{tsum} has the value $0$, which would make the family dependent, so the statement cannot hold for a degenerate reason; moreover \lean{LucasInd.summable\_sums} shows that both series converge. The corollaries, in \lean{LucasInd/Final.lean}, are \lean{LucasInd.irrational\_alt} (Corollary~\ref{cor:main}(i)), \lean{LucasInd.fib\_twin\_one} and \lean{LucasInd.fib\_twin\_two} (part (ii), written with \lean{Nat.fib}), \lean{LucasInd.fib\_mul} and \lean{LucasInd.fib\_two\_mul} (part (iii)), and \lean{LucasInd.pell\_indep} (part (iv), with \lean{pell} $=$ \lean{Useq 2 1}). In them the term of index $m=n+1$ carries the sign $(-1)^m$, as in Corollary~\ref{cor:main}. The identity $F_{kn}=F_kU_n(L_k,(-1)^k)$ is \lean{LucasInd.fib\_mul\_eq}, with the Lucas numbers \lean{lucasNum}, in \lean{LucasInd/Corollaries.lean}. \subsection*{Definitions and the chain} \lean{LucasInd/Defs.lean} defines the construction of Section~\ref{sec:setup} over an arbitrary field with signs $\varepsilon=Q$ and $\delta$, integers $c_3,c_1$ and a parameter $\tau$ (\lean{wpar} $=\varepsilon\tau^2$, \lean{alpha}, \lean{nu}, \lean{psi}, \lean{S}, \lean{node}, \lean{F}, \lean{res}, \lean{wrho} $=\varepsilon\delta\tau=w/\rho$, \lean{mConst}, \lean{mLin}, \lean{nrm}, \lean{Delta}); the operator \lean{opP} $=\Theta$, the real functional \lean{Lfun} and \lean{XiR} as functions of $(\varepsilon,\delta,c_3,c_1,t)$; and the clearing polynomial \lean{CKpoly} $=\prod_{d\le4K}\Phi_d^{b_{K,d}}$ with exponents \lean{nExp} $=$ \lean{nW} $+$ \lean{nAlpha}, where \lean{nAlpha} uses the $\alpha$-roots with $0\le k\le K-2$ (\lean{IsAlphaRoot}). \lean{LucasInd/Statements.lean} states the ingredients as propositions; the proofs assume \lean{Good $\varepsilon$ $\delta$ t}, i.e.\ $\varepsilon,\delta\in\{\pm1\}$ and $0