\documentclass[11pt]{article} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[margin=1in]{geometry} \usepackage{amsmath,amssymb,amsthm,mathtools} \usepackage{microtype} \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref} \hypersetup{pdftitle={An asymptotic lower bound for the number of irrational even beta values},pdfauthor={twinheterodox},pdfsubject={Irrationality counts for even Dirichlet beta values}} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newcommand{\Z}{\mathbb Z} \newcommand{\Q}{\mathbb Q} \newcommand{\R}{\mathbb R} \newcommand{\C}{\mathbb C} \newcommand{\Lcal}{\mathcal L} \newcommand{\sgn}{\operatorname{sgn}} \newcommand{\Res}{\operatorname{Res}} \newcommand{\dd}{\,\mathrm d} \DeclareMathOperator{\lcm}{lcm} \DeclareMathOperator{\rank}{rank} \DeclareMathOperator{\Span}{Span} \numberwithin{equation}{section} \title{\Large An asymptotic lower bound for the number of irrational even beta values} \author{twinheterodox} \date{November 17, 2026} \begin{document} \maketitle \begin{abstract} For even $s$, let $I(s)$ count the irrational values among $\beta(2),\beta(4),\ldots,\beta(s)$. We prove \[ \liminf_{\substack{s\to\infty\\2\mid s}} \frac{I(s)}{\sqrt{s\log\log s/\log s}} \geq 2\sqrt{\frac{\zeta(2)\zeta(3)}{3\zeta(6)}} \approx 1.6098018250 . \] Here $\beta$ denotes the Dirichlet beta function and $\zeta$ the Riemann zeta function. The bound improves the growth rate in our earlier counting result by a factor $\sqrt{\log\log s}$. It has the same growth rate as the recent theorem of Fischler, Lai, Sprang and Zudilin for irrational odd zeta values, with $1/\sqrt3$ of their leading constant. The conclusion remains valid after any fixed initial set of even beta values is omitted. \end{abstract} \section{The counting problem and the character twist} The Dirichlet beta function is \[ \beta(u)=\sum_{v=0}^{\infty}\frac{(-1)^v}{(2v+1)^u} =L(u,\chi_4),\qquad \chi_4(j)= \begin{cases} 0,&2\mid j,\\ 1,&j\equiv1\pmod4,\\ -1,&j\equiv3\pmod4. \end{cases} \] Its positive odd integer values are nonzero rational multiples of powers of $\pi$. The even values have a different arithmetic character. Our aim is to count how many of them must be irrational, without specifying which ones. Write \[ I(s)=\#\{i:2\leq i\leq s,\ 2\mid i,\ \beta(i)\notin\Q\},\qquad \kappa=\frac{\zeta(2)\zeta(3)}{\zeta(6)},\qquad Y(s)=\sqrt{\frac{s\log\log s}{\log s}}. \] Here and throughout, $s$ denotes the largest allowed index, rather than the number of entries in the list. \begin{theorem}\label{thm:main} For every $\varepsilon>0$, all sufficiently large even integers $s$ satisfy \[ I(s)\geq\left(2\sqrt{\frac{\kappa}{3}}-\varepsilon\right)Y(s). \] \end{theorem} Deleting finitely many indices changes $I(s)$ by a bounded amount, whereas $Y(s)\to\infty$. Thus the same lower limit holds if, for example, the list starts with $\beta(4)$. The theorem gives neither a numerical threshold for $s$ nor a lower bound for the dimension of the span of these values. Our earlier explicit counting theorem \cite[Theorem~1.1]{THcount} gives \[ I(s)\geq\left\lfloor\pi\sqrt{\frac{2}{135}} \sqrt{\frac{s}{\log s}}\right\rfloor \qquad(s\geq160,\quad 2\mid s). \] Theorem~\ref{thm:main} improves the growth rate by a factor $\sqrt{\log\log s}$. Its starting endpoint depends on $\varepsilon$ and is not computed here, whereas the earlier bound has the stated explicit range. That proof used the rational grids of Lai and Yu \cite[Section~2]{LY} in the real-character setting of Kumar and Tosi \cite{KT}, together with an auxiliary-prime test for nonvanishing from Lai and Sprang \cite[Lemma~2.3]{LS}. The endpoint certificates are in the \texttt{even-beta-counts/checks} directory of the project repository~\cite{beta4repo}. For odd zeta values, Zudilin \cite[pp.~1--2]{Z18} combined two shifted sums to eliminate $\zeta(3)$. Fischler, Sprang and Zudilin \cite[Section~5]{FSZ} used families of sums to eliminate several coefficients at once. Lai and Yu \cite[Section~2]{LY} chose the shifts using Euler's totient function. Fischler, Lai, Sprang and Zudilin \cite[Theorem~1.1]{FLSZ} took Hankel determinants of the resulting moments and obtained the odd-zeta counting bound with leading constant $2\sqrt\kappa$. Kumar and Tosi \cite[Theorem~1]{KT} state a counting bound of order $\sqrt{s/\log s}$ for the noncritical values of a real Dirichlet character, including $\chi_4$. Their stated constant at conductor four is \[ c(4)=\left(\frac{\kappa}{2} \log\frac{\sqrt{4e^2+16}+4}{\sqrt{4e^2+16}-4}\right)^{1/2}. \] The proof below does not use that theorem or its denominator estimates. Our construction is a character twist of \cite[Sections~2--6]{FLSZ}. On a grid whose denominator is divisible by four, the signs $\chi_4(j)$ combine Hurwitz zeta values into beta values. Only odd $j$ contribute. We therefore put zeros in the numerator only at the fractions used by the character and their integer translates. This retains every elimination weight and reduces the leading logarithmic cost of the numerator by a factor of three. We can consequently use a grid cutoff larger by a factor of $\sqrt3$. Together with the density $1/3$ of usable denominators, this explains the ratio $1/\sqrt3$ between the two leading constants. An even pole order selects the even indices under reflection. The combined grid weights are odd, so we also check that the reflected measure has the opposite residues needed for nonvanishing. The arithmetic normalization, analytic test polynomials and signed-measure criterion follow \cite{FLSZ} with the restricted numerator. The argument assumes familiarity with elementary prime valuations, polynomial linear algebra and residues of meromorphic functions. We give the signed-measure criterion in full before applying it to the character weights. \section{The available grids}\label{sec:grids} The grid family is chosen to supply many elimination weights while keeping the rational function's numerator cost under control. For $B>1$, let \begin{align*} \Psi_B&=\{b\in\Z_{>0}:\varphi(b)\leq B\},\\ \mathcal Q_B&=\{q\in\Psi_B:4\mid q\},\\ \Theta_B&=\left\{\frac aq:q\in\mathcal Q_B,\ 1\leq a0$, let $F(x)$ count all positive integers $b$ with $\varphi(b)\leq x$, let $g(x)$ count the odd such integers, and let $T(x)$ count those divisible by four. Separating the power of two in $b$ gives \[ F(x)=2g(x)+T(x),\qquad T(x)=\sum_{j\geq1}g(x/2^j). \] The two copies of $g$ correspond to $b$ odd and $b$ twice an odd integer; in both cases the totient is unchanged. It follows that \[ 2T(x)=F(x/2)+T(x/2),\qquad T(x)=\sum_{j\geq1}2^{-j}F(x/2^j). \] These sums are finite for fixed $x$, since $F(y)=0$ for $y<1$. By \eqref{eq:totient-count}, $F(y)/y\to\kappa$ and $F(y)\leq Cy$ for $y\geq1$, with an absolute constant $C$. Dividing the last sum by $x$ and applying dominated convergence with majorant $C4^{-j}$ yields $T(x)/x\to\kappa\sum_{j\geq1}4^{-j}=\kappa/3$. \end{proof} The same count determines how many numerator factors we pay for. Each usable denominator contributes exactly $\varphi(q)$ reduced fractions. Writing $T(t)$ for the count in Lemma~\ref{lem:grid-count}, partial summation in the totient variable gives \begin{equation}\label{eq:grid-size} h_B=\sum_{q\in\mathcal Q_B}\varphi(q) =B T(B)-\int_0^B T(t)\dd t =\left(\frac{\kappa}{6}+o(1)\right)B^2 . \end{equation} In particular $h_B+1\leq B^2$ for all sufficiently large $B$, since $\kappa<6$. Every $q\in\Psi_B$ satisfies $q\ll B\log\log B$ \cite[(2.2)]{LY}. Hence $\log A_1(B)\leq h_B(1+o(1))\log B$. For the reverse bound discard the denominators with $\varphi(q)0$, has degree $nh-(n+1)s$, and satisfies \begin{equation}\label{eq:reflection} R_n(-n-t)=R_n(t). \end{equation} Indeed, reflection pairs $\theta$ with $1-\theta$ in the numerator; $n$ is even, and the sign from the denominator disappears because $s$ is even. Suppose, for a contradiction, that $I(s)<|\mathcal Q_B|$. There are then integers $w_q$, not all zero, such that \begin{equation}\label{eq:elimination} \sum_{q\in\mathcal Q_B}w_q q^i=0 \quad\text{whenever }2\leq i\leq s,\quad2\mid i,\quad\beta(i)\notin\Q. \end{equation} This homogeneous rational system has fewer equations than unknowns; clearing denominators gives integral weights. Choose a positive integer $A_0$ clearing the denominators of the remaining, rational, beta values in the same range. Both $A_0$ and the $w_q$ are fixed as $n$ varies. Writing $\zeta(i,\theta)=\sum_{v\geq0}(v+\theta)^{-i}$ for the Hurwitz zeta function, the character identity underlying the construction is, for $i>1$ and $4\mid q$, \begin{equation}\label{eq:character} \sum_{j=1}^q\chi_4(j)\zeta(i,j/q)=q^i\beta(i). \end{equation} For example, the grid $q=4$ gives $\zeta(i,1/4)-\zeta(i,3/4)=4^i\beta(i)$. In general, the integers in the Hurwitz sums are $qv+j$, and $\chi_4(qv+j)=\chi_4(j)$; this proves \eqref{eq:character} by absolutely convergent summation. For a polynomial $P$, set \begin{equation}\label{eq:functional} \Lcal_n(P)=A_0\sum_{q\in\mathcal Q_B}w_q\sum_{j=1}^q\chi_4(j) \sum_{v=0}^{\infty} R_n(v+j/q) \left[P(v+j/q+n/2)+P(-v-j/q-n/2)\right]. \end{equation} We call $P$ admissible if \begin{equation}\label{eq:admissible} \deg P\leq(n+1)s-nh-2. \end{equation} This condition makes the summand $O(v^{-2})$ and guarantees absolute convergence. The proof uses the same moment matrix in two bases. The first makes its scaled determinant an integer. The second bounds the absolute value of that integer by a number less than one. To obtain a contradiction, we allow the size $m$ to vary and prove that at least one permitted size has nonzero determinant. We write \begin{equation}\label{eq:det} \Delta_{n,m}=\det\bigl(\Lcal_n(t^{i+j})\bigr)_{0\leq i,j1$, at every $k\in\Z$; \item $n!(t+j)/(t)_{n+1}$, with $0\leq j\leq n$, at $k=-j$. \end{enumerate} \end{lemma} \begin{proof} The case $r=0$ and Taylor orders above the degree are immediate. For the first assertion with $r\geq1$, a Taylor coefficient of order $\ell\leq r$ is a sum of products obtained by deleting $\ell$ factors from $r$ consecutive integers, divided by $r!$. For any prime $p$ and integer $v\geq1$, the original progression has at least $\lfloor r/p^v\rfloor$ factors divisible by $p^v$. Deleting $\ell$ factors loses at most $\ell$ of them. Hence each product has valuation at least \[ v_p(r!)-\ell\lfloor\log_p r\rfloor \geq v_p(r!)-\ell v_p(d_n). \] This proves integrality after multiplication by $d_n^\ell$. For the second assertion, write its numerator as $\prod_{i=0}^{n-1}(bt+a+bi)$. If $p\nmid b$, this arithmetic progression has at least $\lfloor n/p^v\rfloor$ multiples of each $p^v$; the same deletion argument applies, and the extra factors $b^\ell$ are integral. If $p\mid b$, the prefactor $p^{\lfloor n/(p-1)\rfloor}$ alone clears $n!$, since $v_p(n!)\leq\lfloor n/(p-1)\rfloor$. All the numerator coefficients are integers, proving the assertion at these primes too. For the last assertion set $t=-j+u$. The constant coefficient is $(-1)^j\binom nj$, and the remaining factor is \[ \prod_{\substack{0\leq k\leq n\\k\ne j}} \left(1+\frac{u}{k-j}\right)^{-1}. \] Each Taylor coefficient of order $\ell$ is a sum of products of $\ell$ reciprocals of nonzero integers of absolute value at most $n$. Multiplication by $d_n^\ell$ clears them. \end{proof} \subsection{The coefficients and the finite tails} \begin{lemma}\label{lem:arithmetic} If an admissible $U\in\Q[t]$ satisfies \begin{equation}\label{eq:derivative-integrality} \frac{d_n^\ell}{\ell!}U^{(\ell)}(k)\in\Z \qquad(k\in\Z,\ \ell\geq0), \end{equation} then $d_n^s\Lcal_n(U)\in\Z$. \end{lemma} \begin{proof} Write \[ \widetilde R_n(t)=R_n(t)\left[U(t+n/2)+U(-t-n/2)\right] =\sum_{k=0}^n\sum_{i=1}^s\frac{a_{i,k}}{(t+k)^i}. \] The decomposition has no polynomial part by admissibility. The arithmetic building-block argument of \cite[(3.5)--(3.6)]{FLSZ} gives \begin{equation}\label{eq:coeff-int} d_n^{s-i}a_{i,k}\in\Z\qquad(1\leq i\leq s,\ 0\leq k\leq n). \end{equation} For clarity, $(t+k)^s\widetilde R_n(t)$ factors into $s$ copies of $n!(t+k)/(t)_{n+1}$, the blocks \[ b^n\prod_{p\mid b}p^{\lfloor n/(p-1)\rfloor} \frac{(t+a/b)_n}{n!} \quad(b\in\mathcal Q_B,\ 1\leq a0$, independent of $n$ and $m$, such that, whenever \begin{equation}\label{eq:size-range} 2m\leq(s-B^2-10)n, \end{equation} we have \begin{equation}\label{eq:size-bound} H_{n,m}|\Delta_{n,m}| \leq (C_0m)^m A^{mn}e^{m^2+(2s+10B^2)mn} \left(\frac{10B^2 n}{m}\right)^{(s-B^2)mn-2m^2}. \end{equation} The constant may depend on the fixed parameters $s,B,A_0$ and $w_q$. \end{proposition} \begin{proof} We give the estimate of \cite[Proposition~4.1]{FLSZ} with the character weights included. Put $\nu_i=\lfloor i/(2h+2)\rfloor$ and take the monic polynomials \[ V_i(z)=z^{i-(2h+2)\nu_i} \prod_{\theta\in\Theta\cup\{1\}}\prod_{v=0}^{\nu_i-1} \left(z^2-(v+\theta+n/2)^2\right). \] They have degree $i$, so $\Delta_{n,m}=\det(\Lcal_n(V_iV_j))$. At every contributing point $t=v+j/q$ we have $\chi_4(j)\ne0$ and $j/q\in\Theta$. Thus both $V_i(t+n/2)$ and $V_i(-t-n/2)$ vanish if $v<\nu_i$. For $v\geq\nu_i$, all their roots in absolute value are no larger than $t+n/2$, which gives \[ |V_i(t+n/2)|=|V_i(-t-n/2)|\leq(t+n)^i. \] Let $S_i$ be the sum defining $\Lcal_n(V_i^2)$ with every weight $w_q\chi_4(j)$ replaced by its absolute value. Cauchy--Schwarz and the determinant expansion give \begin{equation}\label{eq:det-majorant} |\Lcal_n(V_iV_j)|\leq\sqrt{S_iS_j},\qquad |\Delta_{n,m}|\leq m!\prod_{i10n$. An integral comparison therefore yields \[ \sum_{v\geq\nu_i}\left(\frac{n}{v+j/q+n}\right)^{e_i} \leq\frac{2}{(1+\nu_i/n)^{e_i-1}}. \] Since $h+1\leq B^2$ and $n\geq2$, we also have $1+\nu_i/n>(i+1)/(2B^2n)$. Set $E=(s-B^2)n-2m>0$. When $2B^2n/(i+1)<1$ the preceding inequalities allow the exponent to be reduced to $E$; when that ratio is at least one, the same bound follows by first using $(1+\nu_i/n)^{-(e_i-1)}\leq1$. Thus \[ S_i\leq C_0 A^n e^{B^2n}n^{2i} \left(\frac{2B^2n}{i+1}\right)^E, \qquad C_0=4A_0B_0^s\sum_{q\in\mathcal Q_B}q|w_q|. \] We used only $|\chi_4(j)|\leq1$. Combining with \eqref{eq:det-majorant} and $m!\geq(m/e)^m$ gives \begin{equation}\label{eq:raw-det-bound} |\Delta_{n,m}|\leq(C_0m)^m A^{mn}e^{B^2mn}n^{m(m-1)} \left(\frac{2eB^2n}{m}\right)^{Em}. \end{equation} Finally, the elementary bounds $d_n\leq e^{2n}$ and \[ \log E_{n,i}\geq i\log n-i-n \] give \[ H_{n,m}\leq n^{-m(m-1)}e^{m^2+2(s+1)mn}. \] Multiply this by \eqref{eq:raw-det-bound}, use $2e<10$ and $B^2+2\leq10B^2$, and obtain \eqref{eq:size-bound}. \end{proof} \section{Nonvanishing for the signed character measure}\label{sec:nonvanishing} Zudilin \cite[Proposition~2]{Zdet} used Hankel determinants of positive moments in an irrationality criterion. Fischler, Lai, Sprang and Zudilin \cite[Section~1]{FLSZ} credit Fauzan's positive-moment construction for $\zeta(5)$ \cite[Section~2.4]{Fauzan} as the prototype for their determinant argument. The weights in \eqref{eq:functional} can have either sign. We use the signed-measure criterion of Fischler, Lai, Sprang and Zudilin \cite[Proposition~5.4]{FLSZ} and give its proof to identify the hypothesis our character weights must satisfy. \subsection{A signed-measure criterion} Let $X\subset\R$ be closed and discrete, and let $\mu=\sum_{x\in X}a_x\delta_x$ be a real signed measure. Suppose, for some positive integer $D$, that \begin{equation}\label{eq:growth} \sum_{x\in X}|a_x|(1+|x|)^{D-1}<\infty. \end{equation} Extend $a_x$ by zero outside $X$. Its Cauchy transform and moments are \[ f(z)=\sum_{x\in X}\frac{a_x}{z-x},\qquad \mu_k=\sum_{x\in X}a_x x^k\quad(0\leq kn/2,\\ 0,&\text{otherwise}. \end{cases} \end{equation} On the positive tail this follows by putting $x=v+j/q+n/2$; $n/2$ is an integer. On the negative tail, the oddness of $w$ and the extra factor $\sgn(x)$ make the reflected mass equal to the positive one, just as the two evaluations of $P$ in \eqref{eq:functional} require. The sign factor cannot be dropped. The measure is even, but its individual masses can have either sign. The ratio of consecutive kernel values is \begin{equation}\label{eq:kernel-ratio} \frac{\widehat R_n(z+1)}{\widehat R_n(z)} =\frac{\mathcal B_n(z)}{\mathcal A_n(z)}, \end{equation} where \begin{align} \mathcal A_n(z)&=(z+1+n/2)^{s+1}(z-n/2) \prod_{\theta\in\Theta}(z+\theta-n/2),\label{eq:A-rec}\\ \mathcal B_n(z)&=(z-n/2)^{s+1}(z+1+n/2) \prod_{\theta\in\Theta}(z+\theta+n/2).\label{eq:B-rec} \end{align} These are the coefficients of \cite[(5.11)]{FLSZ} with the products taken over our $\Theta$. Both have degree \[ d=s+h+2. \] Away from the central gap, \eqref{eq:kernel-ratio} and periodicity give $\mathcal A_n(x)a_{x+1}=\mathcal B_n(x)a_x$. At a transition into the positive tail, write $x=n/2-1+\theta$, where $0<\theta<1$ is a grid fraction. The factor indexed by $1-\theta$ in $\mathcal A_n(x)$ vanishes. At a transition out of the negative tail, $x=-n/2-\theta$, and the factor indexed by $\theta$ in $\mathcal B_n(x)$ vanishes. Integral endpoints carry no mass by \eqref{eq:w-parity}. Points outside $X$ remain outside $X$ after an integer shift. Thus the recurrence holds for the extension by zero at every real $x$. Since $\deg R_n=nh-(n+1)s$, condition \eqref{eq:growth} holds with \begin{equation}\label{eq:moment-D} D=(n+1)s-nh-1. \end{equation} Indeed the summands at infinity are $O(|x|^{-2})$, and $X$ is a finite union of translates of $\Z$. \begin{lemma}\label{lem:no-rational} For this measure and the polynomials \eqref{eq:A-rec}--\eqref{eq:B-rec}, the rational equation in Proposition~\ref{prop:signed} has no solution. \end{lemma} \begin{proof} Suppose $r\in\C(z)$ solves it. The Cauchy transform $f$ solves the same equation. Subtracting and using \eqref{eq:kernel-ratio}, we find that \[ g(z)=\frac{f(z)-r(z)}{\widehat R_n(z)} \quad\text{satisfies}\quad g(z+1)=g(z) \] as a meromorphic identity. Choose the largest denominator with a nonzero elimination weight, $q_* =\max\{q:w_q\ne0\}$, and set $\theta_*=1/q_*$. At either point $x=\pm T+\theta_*$, with integral $T$, a grid denominator $q$ contributes only if $q_*\mid q$. Maximality leaves just $q=q_*$ among the nonzero weights. Moreover $\chi_4(q_*(\pm T)+1)=1$, so \[ w(T+\theta_*)=w(-T+\theta_*)=w_{q_*}\ne0. \] Reflection pairs $T+\theta_*$ with $-T-\theta_*$, where the masses agree. Periodicity pairs $T+\theta_*$ with $-T+\theta_*$, where the grid weight is the same and the sign of $x$ changes. Choose $T$ large enough that these latter points lie outside the central gap and avoid the finitely many poles of $r/\widehat R_n$ and all zeros or poles of $\widehat R_n$. From \eqref{eq:mass}, the residues of $g$ at these two points are respectively \[ A_0w_{q_*}\quad\text{and}\quad -A_0w_{q_*}. \] But the points differ by the integer $2T$, so a one-periodic meromorphic function must have equal residues there. This is a contradiction. \end{proof} \begin{corollary}\label{cor:nonzero} Let $M\in\Z_{\geq1}$. If $d\leq D$ and $2(M+d-1)-2