\documentclass[11pt,letterpaper]{article} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[left=1in,right=1in,top=.88in,bottom=.89in]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools,microtype} \usepackage{xcolor} \usepackage[colorlinks=true,linkcolor=blue!40!black,citecolor=blue!40!black, urlcolor=blue!40!black]{hyperref} \hypersetup{pdftitle={Irrational values of Dirichlet series with periodic coefficients}, pdfauthor={twinheterodox},pdfsubject={Uniform irrationality counts and special values of Dirichlet series}, pdfkeywords={Dirichlet series, periodic coefficients, irrationality, Hankel determinants}} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \numberwithin{equation}{section} \newcommand{\Q}{\mathbb Q} \newcommand{\Z}{\mathbb Z} \newcommand{\R}{\mathbb R} \newcommand{\C}{\mathbb C} \newcommand{\Qb}{\overline{\mathbb Q}} \newcommand{\Lfun}{\mathcal L_n} \newcommand{\grids}{\mathcal Q_B} \newcommand{\denoms}{\mathcal B_B} \newcommand{\dd}{\,\mathrm d} \DeclareMathOperator{\lcm}{lcm} \title{\fontsize{18}{22}\selectfont Irrational values of Dirichlet series with periodic coefficients} \author{twinheterodox} \date{18 November 2026} \begin{document} \maketitle \begin{abstract} Let \(f\) be a nonzero rational-valued function of period \(D\), and write \(L_f(j)=\sum_{n\ge1} f(n)n^{-j}\). For either prescribed parity we give an explicit lower bound of order \(\sqrt{s\log\log s/\log s}\) for the number of exponents \(2\le j\le s\) of that parity for which \(L_f(j)\) is irrational. The bound is uniform in \(f\) for fixed \(D\). It applies, for example, to every nonzero rational projection of a quartic Dirichlet character modulo five. If \(f\) is even or odd, at most \(D-1\) exponents \(j\ge2\) of the matching parity give zero. The nonzero values at distinct exponents of that parity are linearly independent over the algebraic numbers. \end{abstract} \section{A uniform count in each parity} Fix a positive integer \(D\). Assign a rational weight to each residue class modulo \(D\), with at least one weight nonzero, and let \(f(n)\) be the weight of the class of \(n\). For real \(x>1\) define the Dirichlet series \(L_f(x)=\sum_{n\ge1}f(n)n^{-x}\). The constant choice \(f\equiv1\) gives the Riemann zeta function \(\zeta\). A rational-valued Dirichlet character \(f=\chi\) gives its \(L\)-function \(L(x,\chi)\). Weights of opposite signs allow cancellation between residue classes. Given an integer cutoff \(s\), how many even exponents \(2\le j\le s\) must give an irrational value \(L_f(j)\), and how many odd exponents must do so? We give a lower bound for each count that tends to infinity with \(s\), uniformly over the weights for fixed \(D\). For \(\delta=0\) (even) or \(\delta=1\) (odd), write \[ N_{f,\delta}(s)=\#\{2\le j\le s:j\in\Z,\ j\equiv\delta\pmod2,\ L_f(j)\notin\Q\},\qquad Y(s)=\sqrt{\frac{s\log\log s}{\log s}} . \] The constants are defined by the exact prime powers \(p^a\parallel D\). An empty product equals one. \[ \kappa=\frac{\zeta(2)\zeta(3)}{\zeta(6)},\qquad K_D=\kappa\prod_{p^a\parallel D}\frac{p^{2-a}}{p^2-p+1},\qquad R_D=\prod_{p\mid D}(1+p^{-1}). \] \begin{theorem}\label{thm:main} For every \(\epsilon>0\) there is an \(s_0(D,\epsilon)\) such that all integers \(s\ge s_0(D,\epsilon)\), both parities \(\delta\), and every such \(f\) satisfy \[ N_{f,\delta}(s)\ \ge\ \bigl(2\sqrt{K_D/R_D}-\epsilon\bigr)Y(s). \] If \(f\) vanishes on all residues not coprime to \(D\), the coefficient \(2\sqrt{K_D/R_D}\) can be replaced by \(2\sqrt{K_D}\). If \(f(-n)=(-1)^\delta f(n)\) for every \(n\), at most \(D-1\) integers \(j\ge2\) of parity \(\delta\) have \(L_f(j)=0\). The nonzero values at distinct such exponents are linearly independent over \(\Qb\). \end{theorem} One choice of weights applies throughout each count, though it may change with \(s\). In the \emph{matching parity}, defined by \(f(-n)=(-1)^\delta f(n)\) for all \(n\), a Fourier expansion gives \(L_f(j)=\alpha_j\pi^j\) with \(\alpha_j\) algebraic. The nonzero values are thus distinct and transcendental. When the reflection identity fails, a discrepancy between reflected weights makes a determinant nonzero. \newpage \subsection*{Examples and earlier results} Let \(\chi\) be the quartic character modulo five with \(\chi(2)=\sqrt{-1}\). For any \((a,b)\in\Q^2\setminus\{(0,0)\}\), the coefficients of \(a\Re L(j,\chi)+b\Im L(j,\chi)\) on residues \(1,2,3,4\) are \(a,b,-b,-a\). Theorem~\ref{thm:main} gives the coefficient \(2\sqrt{5\kappa/21}\) at even exponents, including the imaginary part alone, whose first coefficient is zero. The two odd quadratic characters \(\chi_4,\chi_{-8}\), viewed modulo eight, likewise give \(2\sqrt{\kappa/6}\) for every nonzero rational combination of \(\beta(j)=L(j,\chi_4)\) and \(L(j,\chi_{-8})\). Their values at \(1,3,5,7\) are respectively \((1,-1,1,-1)\) and \((1,1,-1,-1)\). Fischler proved a uniform result for complex periodic coefficients. His Theorem~2 gives at least \(2^{(1-\epsilon)\log s/\log\log s}\) values in each parity outside any prescribed finite-dimensional rational vector space \cite{Fischler}. For real characters of conductor at least two, Kumar and Tosi obtained an irrationality count of order \(\sqrt{s/\log s}\) in the opposite parity \cite[Theorem~1]{KT}. For individual characters of modulus \(D\ge3\) with \(4\nmid D\), Mistiaen obtained a dimension bound of that order over the cyclotomic field \(\Q(e^{2\pi\sqrt{-1}/D})\) \cite[Th\'eor\`eme~1.0.5]{Mistiaen}, again in the opposite parity. Our conclusion concerns a count uniform in rational coefficients. It attains the scale \(Y(s)\) proved for odd zeta values by Fischler, Lai, Sprang and Zudilin (FLSZ) \cite[Theorem~1.1]{FLSZ}, and for even beta values in \cite[Theorem~1.1]{Beta}. For \(D=1\) and for \(\chi_4\) with \(D=4\) we recover their respective coefficients \(2\sqrt\kappa\) and \(2\sqrt{\kappa/3}\). \subsection*{The two counts behind the proof} When \(f(n)=(-1)^\delta f(-n)\), a Fourier evaluation and a bound on real zeros give the last assertion of Theorem~\ref{thm:main} (Section~\ref{sec:matching}). For the other case, we use the elimination on rational grids of Fischler, Sprang and Zudilin \cite[\S5]{FSZ} and the totient selection of Lai and Yu \cite[\S2]{LY}. The Hankel determinant estimate and the criterion for a nonzero determinant come from FLSZ \cite[\S\S3--5]{FLSZ}, in the form developed for the character modulo four in \cite[\S\S3--6]{Beta}. We determine the grid cost for general \(D\) and show how a discrepancy between two reflected residue classes survives every nontrivial elimination. Each grid \(1/q,\ldots,q/q\) supplies one weight. If fewer exponents give irrational values than there are grids, we can choose a nonzero vector of weights that cancels every irrational term. After clearing denominators, the resulting determinants would be integers of absolute value below one throughout a range of sizes, yet at least one would be nonzero. Proving this last assertion requires some care because the weights can have either sign. Write \(\varphi\) for Euler's totient. For \(B>1\) take the grid denominators \(\grids=\{q\ge1:D\mid q,\ \varphi(q)\le B\}\). For general support the allowed reduced denominators are \(\denoms=\{b\ge2:\varphi(\lcm(b,D))\le B\}\). For support on residues coprime to \(D\) take \(\denoms=\{b\in\grids:b\ge2\}\). Put \[ \Theta=\{a/b:b\in\denoms,\ 1\le a1\), grouping the series by the residues of \(qv+j\) gives \begin{equation}\label{eq:hurwitz} \sum_{j=1}^q f(j)\zeta(i,j/q)=q^i L_f(i),\qquad \zeta(i,\theta)=\sum_{v\ge0}(v+\theta)^{-i}. \end{equation} This is where the restriction \(D\mid q\) is used. Write \((t)_r=t(t+1)\cdots(t+r-1)\) and \((t)_0=1\). For even \(n\ge2\) use the rational function \[ R_n(t)=A^n(n!)^{s-h} \frac{\prod_{\theta\in\Theta}(t+\theta)_n}{(t)_{n+1}^{\,s}} . \] The numerator vanishes at the reflected shifts needed in the finite Hurwitz tails. Since \(n\) is even, reflection gives \(R_n(-n-t)=eR_n(t)\). For a polynomial \(U\) of degree at most \((n+1)s-nh-2\), put \begin{equation}\label{eq:functional} \Lfun(U)=A_0\sum_{q\in\grids}w_q\sum_{j=1}^q f(j)\sum_{v\ge0} R_n(t)\,[U(t+n/2)+U(-t-n/2)]\big|_{t=v+j/q}. \end{equation} The degree condition makes the series absolutely convergent. The two polynomial evaluations cancel for every odd test monomial, whatever \(\delta\) is. For the Dirichlet exponent the reflection \(R_n(-n-t)=(-1)^s R_n(t)\) forces the sum of the coefficients at each pole order of parity opposite to \(s\) to vanish. Equation~\eqref{eq:hurwitz} turns the surviving Hurwitz terms into \(L_f\)-values. For \(2m\le(n+1)s-nh\) define the determinant in the monomial basis and the factors for a second basis by \[ \Delta_{n,m}=\det_{0\le i,jn/2,\\ eA_0W(-x)\widehat R(x), &x<-n/2,\\ 0,&|x|\le n/2 . \end{cases} \end{equation} Extend the masses by zero off \(X\). The first line follows by collecting equal positive evaluation points in \eqref{eq:functional}, and the second from \(\widehat R(-x)=e\widehat R(x)\). In particular the mirror points \(x\) and \(-x\) have equal masses. For \(\theta\in\Theta\cup\{1\}\), periodicity compares \(k+\theta\) with \(-k+\theta\). These points are on opposite tails in the \emph{same} class modulo one, at distance \(2k\). For sufficiently large positive integers \(k\), the kernel is regular and nonzero there. Dividing the masses in \eqref{eq:mass} by its value at each point gives \(A_0W(\theta)\) and \(eA_0W(-\theta)\). The determinant criterion below asks whether a difference equation has a rational solution. Such a solution would make these two numbers residues of a one-periodic meromorphic function, forcing them to agree. Return to \(f=\Im\chi\) modulo five, whose values at \(1,2,3,4\) are \(0,1,-1,0\). Take even exponents, so \(e=1\), and give the 5- and 10-grids weights \(4\) and \(-1\), with all other weights zero. Since \(4\cdot5^2-10^2=0\), the coefficient of \(L_f(2)\) cancels. Both grids hit \(1/5=2/10\). The 5-grid samples the zero at \(1\), whereas the 10-grid samples the first discrepancy between \(f(j)\) and \(f(-j)\), at \(j=2\). Thus \[ W(1/5)=4f(1)-f(2)=-1,\qquad W(-1/5)=4f(-1)-f(-2)=1 . \] A one-periodic function cannot have residues \(-A_0\) and \(A_0\) at \(k+1/5\) and \(-k+1/5\), respectively. In general put \(g(j)=f(j)-ef(-j)\). When \(g\) is nonzero, let \(j_0\) be the first positive index with \(g(j_0)\ne0\), and let \(q_*\) be the largest grid denominator with \(w_{q_*}\ne0\). Set \(\theta_*=j_0/q_*\), so \(0<\theta_*\le1\) since \(j_0\le D\le q_*\). A smaller grid hits \(\theta_*\) only if \(q\theta_*\) is a positive integer below \(j_0\), where the discrepancy is zero. Consequently \begin{equation}\label{eq:witness} W(\theta_*)-eW(-\theta_*)=w_{q_*}g(j_0)\ne0 . \end{equation} At least one of \(f(j_0)\) and \(f(-j_0)\) is nonzero, and the set of fractions is closed under reflection. Thus \(X\) contains this class even if only one tail carries mass there. The endpoint \(\theta_*=1\) is included in \(X\). \begin{proposition}\label{prop:nonzero} Suppose \(g\) is not identically zero. Put \(d=s+h+2\) and \(T=(n+1)s-nh-1\). If \(d\le T\) and \(M\ge1\) is an integer with \(2(M+d-1)-20\) (and is zero for \(x<1\)). Subtract the proposed constant after dividing by \(x\). For \(p\ge3\), the limsup of the absolute error is at most \(\sigma\) times itself, with \(\sigma<1\); the summable weights bound the tail uniformly. For \(p=2\), \(t_1=1\) gives \(2g(x)\) on the right, and the remaining weights sum to one, giving a factor \(1/2\). Restoring the terms \(k\ge a\) multiplies \(K_M\) by \(\sum_{k\ge a}t_k^{-1}/(1+\sigma)=p^{2-a}/(p^2-p+1)\). Induction gives \(|\grids|\sim K_D B\). For support on units, \(h=\sum_{q\in\grids,\,q\ge2}\varphi(q)\); partial summation gives \(h\sim K_DB^2/2\). For general support, include temporarily \(b=1\) and group by \(q=\lcm(b,D)\). One has \[ \sum_{\lcm(b,D)=q}\varphi(b)= \varphi(q)\prod_{\substack{p^a\parallel D\\v_p(q)=a}}\frac{p}{p-1}. \] At a prime with \(v_p(q)=a=v_p(D)\), the local sum is \(\sum_{j=0}^a\varphi(p^j)=p^a\). At the other primes the exponent of \(b\) is fixed. For any subset of the primes dividing \(D\), requiring \(v_p(q)=v_p(D)\) at each selected prime has relative density \(\prod_p(1-1/p)\) among \(\grids\), by inclusion--exclusion and \(K_{Dp}=K_D/p\) for \(p\mid D\). Partial summation gives the same proportion with weight \(\varphi(q)\). Expanding the displayed product as factors \(1+\mathbf1_{v_p(q)=a}/(p-1)\) therefore multiplies the leading coefficient by \(\prod_{p\mid D}(1+1/p)=R_D\). Finally every \(b\in\denoms\) has \(\varphi(b)\le B\). The full-grid estimates \cite[(2.2) and Lemma~2.4]{LY}, \cite[Lemmas~2.1--2.2]{FLSZ} give \(b\ll B\log\log B\), \(h+1\le B^2\) eventually, and a contribution \(O(B^2(\log\log B)^2)\) to \(\log A\) from the second product of \eqref{eq:Acost}. In the first product the terms with \(\varphi(b)\le B/\log B\) have total \(\varphi\)-weight \(O((B/\log B)^2)\) by the unrestricted count. On the remainder, \(\log B-\log\log B\le\log b\le\log B+O(\log\log B)\). These bounds give the stated asymptotic. \end{proof} \section{The size of the same integer}\label{sec:size} The analytic basis puts zeros at the earliest grid points, where \(R_n(t)\) is largest. Since this basis is monic, changing to it preserves the unscaled determinant. This is the estimate of FLSZ \cite[Proposition~4.1]{FLSZ} and of the even-beta paper \cite[Proposition~5.1]{Beta}. We give the bounds here to keep track of the dependence on the elimination weights and on \(f\). Let \(B_0=1+\max\grids\) and \(C_0=4A_0 B_0^s\max_{1\le j\le D}|f(j)|\sum_{q\in\grids}q|w_q|\). If \(2m\le(s-B^2-10)n\), then \begin{equation}\label{eq:size} |J_{n,m}|\le(C_0m)^m A^{mn}\exp\{m^2+(2s+10B^2)mn\} \left(\frac{10B^2n}{m}\right)^{(s-B^2)mn-2m^2}. \end{equation} To see which matrix sizes are useful, take \(B=\lambda Y(s)\) for fixed \(\lambda>0\) and consider \(m/n\to C_s\) at each cutoff \(s\). After taking logarithms in \eqref{eq:size} and dividing by \(mn\), the two terms that govern the leading rate are the arithmetic cost \(\log A\) and \((s-B^2-2C_s)\log(10B^2/C_s)\). As \(s\) grows we want \(C_s/B^2\) large so that the logarithm is negative, and \(C_s=o(s)\) so that its multiplier stays asymptotic to \(s\). Taking \[ C_s=\frac{s}{\log\log s}\qquad\text{gives}\qquad \frac{C_s}{B^2}=\frac{\log s}{\lambda^2(\log\log s)^2}. \] The logarithm of this ratio is asymptotic to \(\log\log s\), giving decay on the \(s\log\log s\) scale of the arithmetic cost. To prove \eqref{eq:size}, observe that for each \(v\), imposing zeros at both signs of all \(h+1\) candidate shifts costs \(2(h+1)\) degrees. A polynomial of degree \(i\) can therefore remove the first \(\nu_i=\lfloor i/(2h+2)\rfloor\) levels. Use the monic polynomial \[ V_i(z)=z^{\,i-(2h+2)\nu_i} \prod_{\theta\in\Theta\cup\{1\}}\prod_{v=0}^{\nu_i-1} (z^2-(v+\theta+n/2)^2). \] It vanishes at both signs of \(t+n/2\) for every active \(t=v+j/q\) with \(v<\nu_i\), including \(j/q=1\). For \(v\ge\nu_i\) each root has absolute value at most \(t+n/2\), giving \(|V_i(\pm(t+n/2))|\le(t+n)^i\). Let \(S_i\) be the sum for \(\Lfun(V_i^2)\) with absolute weights. Positivity of \(R_n(t)\) for \(t>0\), Cauchy--Schwarz and the determinant expansion give \[ |\Delta_{n,m}|\le m!\prod_{i10n\) and \(e_i-1\ge E>0\). Integral comparison of the tail \(\sum_{v\ge\nu_i}(1+(v+j/q)/n)^{-e_i}\) gives at most \(2(1+\nu_i/n)^{-(e_i-1)}\). Since \(h+1\le B^2\), \(1+\nu_i/n>(i+1)/(2B^2n)\). Consequently \[ S_i\le C_0 A^n \exp(B^2n)n^{2i} \left(\frac{2B^2n}{i+1}\right)^E . \] When the last base exceeds one we have simply bounded the preceding tail factor \((1+\nu_i/n)^{-(e_i-1)}\) by one. In the product of these estimates, \[ \prod_{i=0}^{m-1} n^{2i}\left(\frac{2B^2n}{i+1}\right)^E =n^{m(m-1)}\left(\frac{(2B^2n)^m}{m!}\right)^E . \] Use \(m!\ge(m/\mathrm e)^m\) in this denominator. The factor \(H_{n,m}\) supplies the reciprocal power of \(n\). Indeed the bounds \(d_n\le\exp(2n)\) and \(\log E_{n,i}\ge i\log n-i-n\) give \(H_{n,m}\le n^{-m(m-1)}\exp(m^2+2(s+1)mn)\). Using \(m!\le m^m\) in the determinant expansion now yields \eqref{eq:size}. \subsection*{Completing the count when the reflected profiles differ} Now fix \(0<\lambda<2/\sqrt G\). With \(B=\lambda Y(s)\) and \(C_s=s/\log\log s\) as above, for all sufficiently large \(s\), \[ B^21). \] The numerator is \(e^{-t}\) times a nonzero polynomial of degree at most \(D-1\) in \(e^{-t}\). Thus \(H\) changes sign at most \(D-1\) times. Its Mellin transform has at most as many distinct real zeros on \(x>1\) as \(H\) has sign changes. Indeed, a function of constant sign that is nonzero on an interval has a Mellin transform of strict sign. If \(t_0\) is a sign-change point, \(H(t)\log(t/t_0)\) has exactly one fewer sign change. Its Mellin transform, multiplied by \(t_0^{-x}\), is the derivative of \(t_0^{-x}\int_0^\infty t^{x-1}H(t)\dd t\). Each differentiation is justified locally for \(x>1\) by \(H(t)=O(1/t)\) at zero and \(H(t)=O(e^{-t})\) at infinity, even after finitely many logarithmic factors. Induction and Rolle's theorem give the claimed bound. Since \(\Gamma(x)>0\), there are at most \(D-1\) real zeros of \(L_f\) on \((1,\infty)\). A zero at a matching integer can occur. Take \(f(n)=-8\) on multiples of three and \(f(n)=1\) otherwise. This even period-three function satisfies \(L_f(x)=(1-9\cdot3^{-x})\zeta(x)\), so \(L_f(2)=0\). In general, at least the number of eligible exponents minus \(D-1\) give distinct transcendental values on the matching parity. This linear count completes Theorem~\ref{thm:main}, with a cutoff depending only on \(D,\epsilon\). \paragraph{Counting distinct values for an arbitrary profile.} The same sign-change argument bounds repeated values. Every real value \(c\) of \(L_f(x)\) has at most \(D\) preimages on \(x>1\). Indeed \(\Gamma(x)(L_f(x)-c)\) has kernel \(H(t)-ce^{-t}\). With \(z=e^{-t}\), the numerator divided by \(z\) is \(\sum_{r=1}^D f(r)z^{r-1}-c(1-z^D)\), a nonzero polynomial of degree at most \(D\). Hence there are at least \(N_{f,\delta}(s)/D\) distinct irrational numbers uniformly for varying \(f\). For each fixed \(f\) the full leading constants also count distinct irrational numbers. If \(v\ge2\) is the first nonzero coefficient index beyond 1, dominated convergence gives \[ v^j\bigl(L_f(j)-L_f(j+2)\bigr)\longrightarrow f(v)(1-v^{-2})\ne0. \] Such a \(v\) exists by periodicity. Each parity subsequence is eventually strictly monotone, so discarding finitely many initial terms suffices. \newpage \section{Prescribed support and simultaneous projections}\label{sec:projections} Restricting the support of \(f\) can reduce the number of fractions while leaving all the grid weights available. Consider an odd period-eight profile, which vanishes at 0 and 4. Its allowed numerator classes have \((j,8)\in\{1,2\}\). A reduced denominator \(b\) with \((b,8)=4\) requires an even numerator on the original grid. Its smallest possible grid denominator is \(2b\), so \(\varphi(2b)=2\varphi(b)\le B\). For \(8\mid b\), the unit class already permits the grid \(q=b\) and the full cutoff \(\varphi(b)\le B\). The fraction count grows quadratically with the cutoff. Halving it for the first class therefore quarters that class's contribution. More generally, let \(\mathcal G\) be any nonempty set of divisors of \(D\) containing \((j,D)\) whenever \(f(j)\ne0\). For each \(d\mid D\), \(c_d\) below is the density coefficient for denominators with \((b,D)=d\), and \(B/m_d\) is their totient cutoff. Put \[ c_d=\sum_{e\mid D/d}\mu(e)K_{de},\qquad m_d=\min_{\substack{g\in\mathcal G\\D/g\mid d}} g\prod_{\substack{p\mid g\\p\nmid d}}(1-1/p),\qquad G_{\mathcal G}=\sum_{d\mid D}\frac{c_d}{m_d^2}. \] Here \(\mu\) is the M\"obius function. An empty minimum is \(+\infty\) and contributes zero. \begin{corollary}\label{cor:support} For these profiles the coefficient in Theorem~\ref{thm:main} can be taken to be \(2K_D/\sqrt{G_{\mathcal G}}\). \end{corollary} \begin{proof} The denominators required for whole classes \(\mathcal G\) are exactly the \(b\ge2\) for which \(D/g\mid b\) and \(\varphi(gb)\le B\) for some \(g\in\mathcal G\). In fact \(j/q=(j/g)/(q/g)\) with \(g=(j,D)\). The integer \(j/g\) is coprime to \(D/g\). Thus the reduced denominator \(b\) is divisible by \(D/g\), and \(gb\mid q\). Conversely each reduced numerator \(a\) of such a \(b\) occurs as \(ga/(gb)\) with \((ga,D)=g\). In particular the reduced fractions remain closed under reflection.% \footnote{In \texttt{checks/} (GitHub) or \texttt{anc/checks/} (Hexagon), run \texttt{python3 grid\_checks.py}. It compares the denominator description with direct enumeration on 3,110 combinations of support and cutoff.} For \(d=(b,D)\) the ratio \(\varphi(gb)/\varphi(b)\) is \(g\prod_{p\mid g,\,p\nmid d}(1-1/p)\). The cutoff is therefore \(\varphi(b)\le B/m_d\). By M\"obius inversion from Lemma~\ref{lem:density}, the count of \(b\) with \((b,D)=d\) and \(\varphi(b)\le x\) is \(c_dx+o(x)\). Partial summation gives \(h=(G_{\mathcal G}/2+o(1))B^2\). This coefficient is positive since all multiples \(b\) of \(D\) with \(\varphi(b)\le B/g\) occur for any fixed \(g\in\mathcal G\). The proof of the bound for \(\log A\) and the two determinant estimates now apply with \(G=G_{\mathcal G}\). Only finitely many supports occur at fixed \(D\), so a common cutoff suffices. \end{proof} In the period-eight example \(c_4=K_4-K_8=\kappa/6\) and \(c_8=K_8=\kappa/6\). The two cutoffs give \(G_{\{1,2\}}=c_4/4+c_8=5\kappa/24\) and hence the coefficient \(2\sqrt{2\kappa/15}\). For unit support alone the coefficient is \(2\sqrt{\kappa/6}\). \paragraph{Simultaneous detection for all rational directions.} Let \(\mathcal V\) be an \(r\)-dimensional rational vector space of \(D\)-periodic functions, \(r\ge1\). Suppose the uniform bound for the chosen parity at a fixed cutoff guarantees at least \(M\ge1\) irrational indices for every nonzero \(f\in\mathcal V\), where \(M\) is an integer. There are \(\lceil M/r\rceil\) disjoint sets of at most \(r\) exponents each, such that \emph{on each set} every nonzero \(f\in\mathcal V\) has an irrational value. The sets are common to all \(f\), although the exponent witnessing irrationality within a set may depend on \(f\). Indeed, after deleting fewer than \(M\) indices, the remaining rational-linear maps \(f\mapsto L_f(j)+\Q\) into \(\R/\Q\) have trivial common kernel. Choose maps successively that lower the dimension of the intersection of kernels, until it reaches zero. This uses at most \(r\) indices. Remove those indices and repeat. For the quartic character modulo five this produces disjoint pairs or singletons that work simultaneously for all rational directions \((a,b)\). \paragraph{Verification.} An independent AI audit of the 18 November 2026 working proof verified the uniform constants for both parities, the signed nonvanishing criterion, the arithmetic normalization and the matching-parity conclusion. A separate exposition review examined the ten-page manuscript of the same date. \begin{thebibliography}{99} \bibitem{Dressler} R.~E.~Dressler, \href{https://doi.org/10.2140/pjm.1970.34.371}{A density which counts multiplicity}, \emph{Pacific J. Math.} \textbf{34} (1970), 371--378. \bibitem{Fauzan} A.~Fauzan, \href{https://doi.org/10.5281/zenodo.22826419}{\(\zeta(5)\) is irrational}, Zenodo, 17 September 2026. \bibitem{Fischler} S.~Fischler, \href{https://doi.org/10.1112/jlms.12290}{Irrationality of values of \(L\)-functions of Dirichlet characters}, \emph{J. London Math. Soc.} \textbf{101} (2020), 857--876. \bibitem{FLSZ} S.~Fischler, L.~Lai, J.~Sprang and W.~Zudilin, \href{https://arxiv.org/abs/2611.16156}{Many more odd zeta values are irrational}, arXiv:2611.16156 (2026). \bibitem{FSZ} S.~Fischler, J.~Sprang and W.~Zudilin, \href{https://doi.org/10.1112/S0010437X1900722X}{Many odd zeta values are irrational}, \emph{Compos. Math.} \textbf{155} (2019), 938--952. \bibitem{KT} V.~Kumar and R.~Tosi, \href{https://arxiv.org/abs/2610.41458}{Irrationality of values of \(L\)-functions for real Dirichlet characters}, arXiv:2610.41458 (2026). \bibitem{LY} L.~Lai and P.~Yu, \href{https://arxiv.org/abs/1911.08458}{A note on the number of irrational odd zeta values}, \emph{Compos. Math.} \textbf{156} (2020), 1699--1717. \bibitem{Mistiaen} L.~Mistiaen, \href{https://arxiv.org/abs/2512.02612}{Ind\'ependance lin\'eaire de valeurs de fonctions \(L\) de Dirichlet}, arXiv:2512.02612 (2025). \bibitem{DLMF} F.~W.~J.~Olver et al., eds., \href{https://dlmf.nist.gov/24.8}{NIST Digital Library of Mathematical Functions}, \S24.8(i), Fourier series for Bernoulli polynomials. \bibitem{Sprang} J.~Sprang, \href{https://arxiv.org/abs/1802.09410}{Infinitely many odd zeta values are irrational. By elementary means}, arXiv:1802.09410 (2018). \bibitem{Beta} twinheterodox, \href{https://hexagonmath.org/2611.00600v1}{An asymptotic lower bound for the number of irrational even beta values}, Hexagon 2611.00600v1 (2026). \bibitem{Zdet} W.~Zudilin, \href{https://doi.org/10.1007/s00365-016-9333-7}{A determinantal approach to irrationality}, \emph{Constr. Approx.} \textbf{45} (2017), 301--310. \end{thebibliography} \end{document}