\documentclass[12pt]{amsart} \pagestyle{headings} \textwidth 6.5in \sloppy \flushbottom \parindent 1em \marginparwidth 48pt \marginparsep 10pt \columnsep 10mm \renewcommand{\baselinestretch}{1.09} \usepackage[margin=1in]{geometry} \usepackage{microtype} \usepackage{amssymb} \usepackage{array} \usepackage[linktocpage=true,colorlinks,citecolor=blue,linkcolor=blue,urlcolor=blue,pagebackref]{hyperref} \newtheorem{theorem}{Theorem} \newtheorem{lemma}{Lemma} \newtheorem{corollary}{Corollary} \theoremstyle{definition} \newtheorem{definition}{Definition} \newtheorem{conjecture}{Conjecture} \newtheorem{example}{Example} \newtheorem{xca}{Exercise} \theoremstyle{remark} \newtheorem{remark}{Remark} \providecommand*{\theoremautorefname}{Theorem} \providecommand*{\lemmaautorefname}{Lemma} \providecommand*{\corollaryautorefname}{Corollary} \providecommand*{\definitionautorefname}{Definition} \providecommand*{\conjectureautorefname}{Conjecture} \providecommand*{\remarkautorefname}{Remark} \numberwithin{equation}{section} \begin{document} \title[Pair correlation and Hooley's conjecture]{Pair correlation of \texorpdfstring{$L$}{L}-zeros and Hooley's conjecture} % Author and publication metadata are placeholders for the author to complete. \author{Travor Zihao Liu} \address{Department of Mathematics, Stanford University, 380 Jane Stanford Way, Stanford, CA 94305, USA} \curraddr{} \email{tzliu@stanford.edu} \subjclass[2020]{Primary 11M50, 11N13; Secondary 11M26, 11N05} %\date{} %\dedicatory{} %\commby{} \begin{abstract} In this article, we study the form factor for zeros of Dirichlet $L$-functions, pairing zeros of the same $L$-function and then averaging over primitive characters modulo a prime conductor. Unconditionally, we show for $|\alpha|<1$ that the form factor agrees with the random matrix prediction from the Circular Unitary Ensemble (CUE). Under GRH and the Hardy--Littlewood twin primes conjecture, we extend this agreement to $|\alpha|<2$. For $|\alpha|\ge1$, we prove that the CUE prediction is equivalent to Hooley's conjecture for the variance of primes in arithmetic progressions with prime moduli. We also explore applications of these form-factor results to the simplicity of zeros. \end{abstract} \maketitle \section{Introduction}\label{sec:introduction} In 1973, assuming the Riemann hypothesis, Montgomery~\cite{MontgomeryPairCorrelation1973} studied the pair correlation of $\zeta$-zeros. Writing the nontrivial zeros as $\frac12+i\gamma,\frac12+i\gamma'$ and $N(T)$ the number of zeros with $0\le\gamma\le T$, he conjectured that, for every interval $[\alpha,\beta]$, \begin{align*} \lim_{T\to\infty}&{\#\left\{\gamma,\gamma'\in[0,T]:{\gamma-\gamma'\over 2\pi}\log T\in[\alpha,\beta]\right\}\over N(T)} \\ &=\mathbf 1_{0\in[\alpha,\beta]}+\int_\alpha^\beta\left[1-\left({\sin\pi u\over \pi u}\right)^2\right]\mathrm d u. \end{align*} This is known as the \emph{pair correlation conjecture}. The integrand in this conjecture coincides with the limiting pair-correlation density for the eigenvalues in the Circular Unitary Ensemble (CUE) model, establishing a connection between the random matrix theory and $\zeta$-zeros. This connection was later extended to a broad class of random matrix models for statistics of both zeros and values of $L$-functions: Rudnick--Sarnak (1996)~\cite{RudnickSarnakZeros1996} generalized the CUE prediction to high zeros of automorphic $L$-functions. Katz--Sarnak (1999)~\cite{KatzSarnak1999,KatzSarnak1999Zeros} developed a philosophy that associates each reasonable family of $L$-functions with a random matrix group known as the \emph{symmetry type} based on the statistics of low-lying zeros. Keating--Snaith (2000)~\cite{KeatingSnaith2000Zeta,KeatingSnaith2000LFunctions} formulated moment conjectures for $\zeta(\frac12+it)$ and central values of $L$-functions based on the symmetry type. A detailed discussion of these topics can be found in Soundararajan's 2022 ICM plenary lecture \cite{SoundararajanDistributionValues2021}. Montgomery originally formulated his conjecture using the Fourier-side statistic \begin{equation*} F(\alpha,T)={1\over N(T)}\sum_{0<\gamma,\gamma'\le T}T^{i\alpha(\gamma-\gamma')}w(\gamma-\gamma'), \end{equation*} where $w(u)$ is a suitable smooth weight with $w(0)=1$. This statistic is known as the \emph{form factor}. Montgomery's \emph{strong pair correlation conjecture} states, in the sense of distributions, that \begin{equation}\label{eq:zeta-strong-pcc} F(\alpha,T)\longrightarrow\delta(\alpha)+\min(|\alpha|,1), \qquad(T\to\infty). \end{equation} The continuous term has a \emph{ramp} for $|\alpha|<1$ and a \emph{plateau} for $|\alpha|\ge1$. The delta mass at the origin comes from the diagonal contribution $\gamma=\gamma'$. Montgomery himself verified this limit for $|\alpha|<1$ under RH, allowing him to establish conditional results on the simplicity of and gaps between $\zeta$-zeros. Just as the abscissae of $\zeta$-zeros govern the exponent in the error term of the prime number theorem, the pair correlation of the ordinates of these zeros has a prime-side interpretation. In 1987, Goldston--Montgomery~\cite{GoldstonMontgomeryPairCorrelation1987} established under RH that the strong pair correlation conjecture\footnote{More precisely, Goldston--Montgomery used a suitable uniform version of strong pair correlation; the distributional formulation above does not by itself assert this uniformity.} is equivalent to \begin{equation*} J(X,H):=\int_1^X\left(\sum_{x0$ and ${x\over(\log x)^A}x$ that \begin{equation}\label{eq:friedlander-goldston} G(x,q)=x\log x-x-{x^2\over\varphi(q)}+O_A\left({x\over(\log x)^A}\right)+O((\log q)^3), \end{equation} and they conjectured that \eqref{eq:hooley-asymptotic} holds throughout $x^{\frac12+\varepsilon}0$. In 2023, Fiorilli--Martin~\cite{FiorilliMartinDisproof2023} showed that this range is essentially optimal in the sense that the bound $G(x,q)\ll x\log q$ can fail when $q\asymp\log\log x$. In 2026, Leung~\cite{LeungMomentsProgressions2026} computed higher moments and showed under the Hardy--Littlewood $k$-tuple conjectures that $\psi(x;q,a)$ behaves like a Gaussian with mean ${x\over\varphi(q)}$ and variance ${x\log q\over\varphi(q)}$ as the reduced residue class $a$ varies, $x\to+\infty$, and $x^{1-\varepsilon}\le q\le x/\log x$. To focus on the number-theoretic content and avoid technical complications arising from the sharp cutoff, we work with a smoothed version of $G(x,q)$. Given $W\in C_c^\infty(0,\infty)$, define \begin{equation}\label{eq:weighted-variance} G_W(x,q)=\sum_{\substack{1\le a\le q\\ (a,q)=1}}\left|\sum_{n\equiv a(q)}\Lambda(n)W\left(\frac nx\right)-{1\over\varphi(q)}\sum_{(n,q)=1}\Lambda(n)W\left(\frac nx\right)\right|^2. \end{equation} We formulate the corresponding smooth conjecture as follows: \begin{conjecture}[Smooth Hooley]\label{conj:smooth-hooley} For every $\varepsilon>0$ and every fixed $W\in C_c^\infty(0,\infty)$, \begin{equation*} G_W(x,q)\sim\left(\int_0^\infty| W(u)|^2\mathrm d u\right)x\log q, \end{equation*} as $x\to\infty$, uniformly for $x^\varepsilon\le q\le x$. \end{conjecture} In 2025, Hayani~\cite{HayaniWeightedHooley2025} established the analog of Fiorilli--Martin for $G_W(x,q)$, so we choose not to study the range $q0$, \begin{equation*} \widehat f\left(-{\alpha\log q\over2\pi}\right)=0, \end{equation*} for $q\gg_\sigma 1$. Thus, the second main term vanishes, and the error term in \autoref{thm:form-factor-ramp} simplifies to $O(\sqrt{\log q})$ in this range. \end{remark} \begin{corollary}[Restricted-support pair correlation]\label{cor:restricted-pair-correlation} For every band-limited $\phi\in L^1(\mathbb R)$ with $B_\phi<1$, \begin{equation*} R(f,\phi,q)\sim{\log q\over2\pi}\left(\int_{\mathbb R}| f(t)|^2\mathrm d t\right)\left(\widehat\phi(0)+\int_{\mathbb R}|\alpha|\widehat\phi(\alpha)\mathrm d\alpha\right), \end{equation*} as $q\to\infty$ through odd primes. \end{corollary} We record a conditional application of \autoref{cor:restricted-pair-correlation} to the simplicity of $L$-zeros: \begin{corollary}\label{cor:smooth-simplicity} Assume GRH. For every nonnegative band-limited Schwartz function $f$, \begin{equation}\label{eq:smooth-simplicity-bound} \sideset{}{^*}{\sum}_\chi\sum_{\substack{\gamma_\chi\\ \rho_\chi\ \mathrm{simple}}}| f(\gamma_\chi)|^2\ge\left(\frac23-o_f(1)\right)\sideset{}{^*}{\sum}_\chi\sum_{\gamma_\chi}| f(\gamma_\chi)|^2. \end{equation} \end{corollary} \begin{remark}\label{rem:simple-interval} The estimate above is the smooth analog of the assertion that, for every nondegenerate fixed interval $[a,b]$, \begin{equation*} \sideset{}{^*}{\sum}_\chi\sum_{\substack{\gamma_\chi\in[a,b]\\ \rho_\chi\ \mathrm{simple}}}1\ge\left(\frac23-o(1)\right)\sideset{}{^*}{\sum}_\chi\sum_{\gamma_\chi\in[a,b]}1. \end{equation*} Formally, this is obtained from \eqref{eq:smooth-simplicity-bound} by taking $|f|^2$ to be $\mathbf 1_{[a,b]}$. We do not pursue here the approximation needed to pass from band-limited Schwartz weights to a sharp cutoff. \end{remark} We can now state the affirmative answer announced above: for prime moduli, Hooley's conjecture for $G_W(x,q)$ is equivalent to \autoref{conj:form-factor-plateau}. \begin{theorem}\label{thm:equivalence} Fix $B\ge1$. Restricted to prime moduli, the following are equivalent: \begin{enumerate} \item For every $W\in C_c^\infty(0,\infty)$, as $x\to\infty$, \begin{equation*} G_W(x,q)\sim x\log q\int_0^\infty| W(u)|^2\mathrm d u, \end{equation*} uniformly for primes $q$ satisfying $x^{1/B}\le q\le x$. \item For every band-limited Schwartz function $f$, as $q\to\infty$ through primes, \begin{equation*} K(f,\alpha,q)\sim{\log q\over2\pi}\int_{\mathbb R}| f(t)|^2\mathrm d t, \end{equation*} uniformly for $1\le|\alpha|\le B$. \end{enumerate} \end{theorem} It should be noted that the Goldston--Montgomery correspondence mentioned earlier was established under RH, but our \autoref{thm:equivalence} does not rely on any unproved hypotheses. \begin{theorem}\label{thm:hardy-littlewood-variance} Assume GRH and the following smooth Hardy--Littlewood conjecture. For every $W\in C_c^\infty(0,\infty)$, extend $W$ by zero to $\mathbb R$, put \begin{equation}\label{eq:weight-autocorrelation} A_W(t)=\int_{\mathbb R}W(u)\overline{W(u+t)}\mathrm d u, \end{equation} and, for $k\ne0$, let \begin{equation*} \mathfrak S(k)= \begin{cases} 2\displaystyle\prod_{p>2}\left(1-{1\over(p-1)^2}\right)\displaystyle\prod_{\substack{p\mid k\\ p>2}}{p-1\over p-2},&k\text{ even},\\ 0,&k\text{ odd}. \end{cases} \end{equation*} The conjecture asserts that, for every $\eta>0$, as $x\to\infty$ and uniformly for nonzero integers $k$ with $|k|\ll x$, \begin{equation}\label{eq:smooth-hardy-littlewood} \sum_n\Lambda(n)\Lambda(n+k)W\left({n\over x}\right)\overline{W\left({n+k\over x}\right)}=\mathfrak S(k)xA_W\left({k\over x}\right)+O\left(x^{\frac12+\eta}\right). \end{equation} Then, for every $\varepsilon>0$, the asymptotic in \autoref{conj:smooth-hooley} holds uniformly for primes in the range $x^{\frac12+\varepsilon}\le q\le x$. More precisely, uniformly for primes $q$ satisfying $x^{\frac12+\varepsilon}\le q\ll x$, \begin{equation}\label{eq:hardy-littlewood-variance} G_W(x,q)=x\log q\int_0^\infty| W(u)|^2\mathrm d u+O(x). \end{equation} \end{theorem} By \autoref{thm:equivalence}, \autoref{thm:hardy-littlewood-variance} has the following immediate consequence. \begin{corollary}[Hardy--Littlewood extension]\label{cor:hardy-littlewood-form-factor} Under the hypotheses of \autoref{thm:hardy-littlewood-variance}, \autoref{conj:form-factor-plateau} holds in the range $1\le|\alpha|<2$, uniformly on compact subranges. \end{corollary} Combining \autoref{thm:form-factor-ramp} with \autoref{cor:hardy-littlewood-form-factor}, we obtain the following conditional extensions of \autoref{cor:restricted-pair-correlation} and \autoref{cor:smooth-simplicity}. \begin{corollary}\label{cor:conditional-pair-correlation} Under the hypotheses of \autoref{thm:hardy-littlewood-variance}, for every band-limited $\phi\in L^1(\mathbb R)$ with $B_\phi<2$, \begin{equation*} R(f,\phi,q)\sim{\log q\over2\pi}\left(\int_{\mathbb R}| f(t)|^2\mathrm d t\right)\left(\widehat\phi(0)+\int_{\mathbb R}\min(|\alpha|,1)\widehat\phi(\alpha)\mathrm d\alpha\right), \end{equation*} as $q\to\infty$ through odd primes. \end{corollary} \begin{corollary}\label{cor:conditional-simplicity} Under the hypotheses of \autoref{thm:hardy-littlewood-variance}, for every nonnegative band-limited Schwartz function $f$, \begin{equation}\label{eq:conditional-simplicity-bound} \sideset{}{^*}{\sum}_\chi\sum_{\substack{\gamma_\chi\\ \rho_\chi\ \mathrm{simple}}}| f(\gamma_\chi)|^2\ge\left({11\over12}-o_f(1)\right)\sideset{}{^*}{\sum}_\chi\sum_{\gamma_\chi}| f(\gamma_\chi)|^2. \end{equation} \end{corollary} The setup considered here is not the only $q$-analog of Montgomery's pair correlation for Dirichlet $L$-functions. Among previous works, Chandee--Lee--Liu--Radziwi{\l}{\l}~\cite{ChandeeLeeLiuRadziwill2014} is closest to our zero-side family: they studied correlations for primitive Dirichlet $L$-functions with conductors $q\asymp Q$. By contrast, we fix a single prime conductor $q$ and average only over the primitive characters modulo $q$. Earlier, {\"O}zl{\"u}k~\cite{OzlukThesis1982,OzlukPairCorrelation1996} considered a related modulus-averaged statistic over all characters modulo $q\le Q$. In both {\"O}zl{\"u}k's work and that of Chandee--Lee--Liu--Radziwi{\l}{\l}, the additional averaging over the modulus allows one to exploit Hooley's asymptotic evaluation of the Barban--Davenport--Halberstam variance $H(x,Q)$ in the range $x^{\frac12+\varepsilon}0$ is an absolute constant. Since $\widetilde V(1)=\int_0^\infty| W(u)|^2\mathrm d u$ and $\widetilde V'(1)=O(1)$, combining this asymptotic with the $O(x^{1/2}\log(2x))$ error from replacing $\Lambda(n)^2$ by $\Lambda(n)\log n$ evaluates the first term in \eqref{eq:variance-diagonal-expansion} up to $O(x)$. Together with the $O(x)$ bound for the second term, this proves the lemma. \end{proof} \begin{remark}\label{rem:arithmetic-transition} The key point is the implication $m\equiv n\pmod q\Rightarrow m=n$ for $m,n0$, suppose $0\le\alpha\le 1-\eta$, and put $x=q^\alpha$. Therefore, for $q\gg_\eta 1$, one has $q\gg x$. Substitution of \autoref{lem:large-modulus-variance} into \autoref{thm:form-factor-variance} gives \begin{align}\label{eq:form-factor-weighted-ramp} K(f,\alpha,q) &=\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\log x+{| W(1/x)|^2\over x}(\log q)^2 \\ &\quad+O\left(\left(1+{| W(1/x)|\log q\over\sqrt x}\right)\sqrt{\log(2x)}\right).\notag \end{align} Using $\log x=\alpha\log q$ and the two identities in \eqref{eq:fourier-mellin-identities} proves the asserted formula. For $-1+\eta\le\alpha<0$, define \begin{equation*} W_-(u)=u^{-1}W(1/u), \qquad \widetilde W_-(s)=\widetilde W(1-s), \qquad f_-(t)=f(-t). \end{equation*} By \eqref{eq:zero-symmetry}, replacing each zero ordinate of $L(s,\chi)$ by the negative of the corresponding ordinate of $L(s,\overline\chi)$ in \eqref{eq:form-factor-definition} gives \begin{align}\label{eq:form-factor-symmetry} K(f,\alpha,q) &={1\over q-2}\sideset{}{^*}{\sum}_\chi\left(\sum_{\gamma_{\overline\chi}}f(-\gamma_{\overline\chi})q^{-i\alpha\gamma_{\overline\chi}}\right)\left(\sum_{\gamma_{\overline\chi}'}f^\#(-\gamma_{\overline\chi}')q^{i\alpha\gamma_{\overline\chi}'}\right)\notag\\ &={1\over q-2}\sideset{}{^*}{\sum}_\chi\left(\sum_{\gamma_\chi}f_-(\gamma_\chi)q^{-i\alpha\gamma_\chi}\right)\left(\sum_{\gamma_\chi'}(f_-)^\#(\gamma_\chi')q^{i\alpha\gamma_\chi'}\right)\notag\\ &=K(f_-,-\alpha,q). \end{align} Moreover, a change of variables gives \begin{equation*} \int_{\mathbb R}| f_-(t)|^2\mathrm d t=\int_{\mathbb R}| f(t)|^2\mathrm d t, \qquad \widehat{f_-}(\xi)=\widehat f(-\xi). \end{equation*} Applying the positive-$\alpha$ result to $K(f_-,-\alpha,q)$ completes the proof. \end{proof} \section{The plateau and Hooley's conjecture}\label{sec:plateau-hooley} \begin{proof}[Proof of \autoref{thm:equivalence}] Fix $B\ge1$. Suppose first that the prime-side asymptotic holds. Fix a band-limited Schwartz function $f$, let $W$ be its unique partner, and put $x=q^\alpha$ with $1\le\alpha\le B$. Then $x^{1/B}\le q\le x$, so \autoref{thm:form-factor-variance} gives \begin{align}\label{eq:hooley-to-form-factor} K(f,\alpha,q) &=\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\log q+{q-1\over q-2}o_W(\log q)+{| W(1/x)|^2\over x}(\log q)^2 \\ &\quad+O\left(\left(1+{| W(1/x)|\log q\over\sqrt x}\right)\sqrt{\log q}\right),\notag \end{align} uniformly in this range. By \autoref{rem:fourier-term-away-from-zero} and \eqref{eq:fourier-mellin-identities}, the second main term vanishes and the error simplifies to $O(\sqrt{\log q})$. The first identity in \eqref{eq:fourier-mellin-identities} therefore turns \eqref{eq:hooley-to-form-factor} into the positive-$\alpha$ plateau. Finally, \eqref{eq:form-factor-symmetry} extends this conclusion to $1\le|\alpha|\le B$. Conversely, assume the form-factor asymptotic. Fix $W\in C_c^\infty(0,\infty)$, and let $f$ be its unique partner. Given $x^{1/B}\le q\le x$, put $\alpha=\log x/\log q$, so that $1\le\alpha\le B$. By \autoref{rem:fourier-term-away-from-zero} and \eqref{eq:fourier-mellin-identities}, \autoref{thm:form-factor-variance} reduces to \begin{equation}\label{eq:plateau-to-hooley} K(f,\alpha,q)={q-1\over q-2}{G_W(x,q)\over x}+O\left(1+\left({G_W(x,q)\over x}\right)^{1/2}\right). \end{equation} Put $Y=G_W(x,q)x^{-1}$. Then \eqref{eq:plateau-to-hooley} and the assumed form-factor asymptotic imply \begin{equation*} Y\ll \log q+Y^{\frac12}, \end{equation*} which implies $Y\ll \log q$, so the error in \eqref{eq:plateau-to-hooley} is $O(\sqrt{\log q})$, which is negligible compared with the main term. The first identity in \eqref{eq:fourier-mellin-identities} now recovers the prime-side asymptotic in \autoref{thm:equivalence}, uniformly for $x^{1/B}\le q\le x$. \end{proof} \section{The Hardy--Littlewood extension}\label{sec:hardy-littlewood} We now use the smooth Hardy--Littlewood conjecture stated in \autoref{thm:hardy-littlewood-variance} to establish an asymptotic for $G_W(x,q)$ in the range $x^{\frac12+\varepsilon}\le q\ll x$. The extension to $1\le|\alpha|<2$ then follows from \autoref{thm:equivalence}. The derivation in this section is a smooth adaptation of arguments in \S2 and \S4 of Friedlander--Goldston~\cite{FriedlanderGoldstonVariance1996}. \begin{lemma}\label{lem:singular-series} For $W\in C_c^\infty(0,\infty)$, put \begin{equation*} A_W(t)=\int_{\mathbb R}W(u)\overline{W(u+t)}\mathrm d u. \end{equation*} Uniformly for odd primes $q$ with $q\ll x$, \begin{equation}\label{eq:singular-series} \sum_{j\ne0}\mathfrak S(jq)A_W\left({jq\over x}\right)={x\over q-1}|\widetilde W(1)|^2-\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\log{x\over q}+O(1). \end{equation} \end{lemma} \begin{proof} Put \begin{equation*} h_W(t)=A_W(t)+A_W(-t)\qquad(t\ge0). \end{equation*} Since $\mathfrak S(k)=\mathfrak S(-k)$, Mellin inversion gives \begin{equation}\label{eq:singular-contour} \sum_{j\ne0}\mathfrak S(jq)A_W\left({jq\over x}\right)={1\over2\pi i}\int_{(2)}\left({x\over q}\right)^sD_q(s)\widetilde h_W(s)\mathrm d s, \end{equation} where \begin{align}\label{eq:dirichlet-series} D_q(s)&=\sum_{j\ge1}{\mathfrak S(jq)\over j^s}=\mathfrak S(2q)\zeta(s)\zeta(s+1)F_q(s),\notag\\ F_q(s)&=2^{-s}\left(1-2^{-s-1}\right)\left(1-q^{-s-1}\right)\prod_{\substack{p>2\\ p\ne q}}\left(1+{1\over(p-2)p^s}\right)\left(1-p^{-s-1}\right). \end{align} The product defining $F_q(s)$ converges absolutely for $\mathop{\rm Re}s>-\frac12$ and satisfies $F_q(s)\ll1$ uniformly over odd primes $q$ and $s$ with $-\frac14\le\mathop{\rm Re}s\le2$. Shifting the line of integration in \eqref{eq:singular-contour} to $\mathop{\rm Re}s=-\frac14$, we have \begin{align}\label{eq:singular-residues} \sum_{j\ne0}\mathfrak S(jq)A_W\left({jq\over x}\right) &=\mathop{\rm Res}_{s=1}\left(\left({x\over q}\right)^sD_q(s)\widetilde h_W(s)\right)\notag\\ &\quad+\mathop{\rm Res}_{s=0}\left(\left({x\over q}\right)^sD_q(s)\widetilde h_W(s)\right)+\mathcal E_W(x,q), \end{align} where \begin{equation}\label{eq:singular-remainder} \mathcal E_W(x,q)={1\over2\pi i}\int_{(-\frac14)}\left({x\over q}\right)^sD_q(s)\widetilde h_W(s)\mathrm d s. \end{equation} As $s\to1$, one has $\zeta(s)\sim(s-1)^{-1}$ and hence $(s-1)D_q(s)\to q/(q-1)$. Since \begin{equation*} \widetilde h_W(1)=\int_{\mathbb R}A_W(t)\mathrm d t=|\widetilde W(1)|^2, \end{equation*} the residue at $s=1$ is \begin{equation}\label{eq:one-residue} \mathop{\rm Res}_{s=1}\left(\left({x\over q}\right)^sD_q(s)\widetilde h_W(s)\right)={x\over q-1}|\widetilde W(1)|^2. \end{equation} For the residue at zero, logarithmic differentiation of the $F_q(s)$ in \eqref{eq:dirichlet-series} gives \begin{align*} {F_q'(s)\over F_q(s)} &=-\log2+{2^{-s-1}\log2\over1-2^{-s-1}}+{q^{-s-1}\log q\over1-q^{-s-1}}\\ &\quad+\sum_{\substack{p>2\\ p\ne q}}\left(-{((p-2)p^s)^{-1}\log p\over1+((p-2)p^s)^{-1}}+{p^{-s-1}\log p\over1-p^{-s-1}}\right),\\ {F_q'(0)\over F_q(0)} &=-\log2+\log2+{\log q\over q-1}+\sum_{\substack{p>2\\ p\ne q}}\left(-{\log p\over p-1}+{\log p\over p-1}\right)={\log q\over q-1}. \end{align*} Together with $\mathfrak S(2q)F_q(0)=1$, this yields \begin{equation*} \mathfrak S(2q)F_q(s)=1+{\log q\over q-1}s+O_q(s^2). \end{equation*} Since \begin{equation*} \zeta(s)=-\frac12-\frac12(\log2\pi)s+O(s^2), \qquad \zeta(s+1)={1\over s}+\gamma+O(s), \end{equation*} we obtain the explicit Laurent expansion \begin{equation}\label{eq:D-zero} D_q(s)=-{1\over2s}-\frac12\left(\gamma+\log2\pi+{\log q\over q-1}\right)+O_q(s). \end{equation} Since $h_W(0)=2\int_0^\infty| W(u)|^2\mathrm d u$, the meromorphic continuation of its Mellin transform satisfies \begin{equation}\label{eq:h-zero} \widetilde h_W(s)={2\over s}\int_0^\infty| W(u)|^2\mathrm d u+\left.{\mathrm d\over\mathrm d s}\left(s\widetilde h_W(s)\right)\right|_{s=0}+O(s). \end{equation} Combining \eqref{eq:D-zero} and \eqref{eq:h-zero} with $({x\over q})^s=1+s\log({x\over q})+O(s^2)$ shows that the residue at $s=0$ is \begin{align}\label{eq:zero-residue} &-\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\log{x\over q}-\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\left(\gamma+\log2\pi+{\log q\over q-1}\right)\notag\\ &\qquad-\frac12\left.{\mathrm d\over\mathrm d s}\left(s\widetilde h_W(s)\right)\right|_{s=0}. \end{align} The last two terms in \eqref{eq:zero-residue} are $O(1)$ uniformly in $q$. Finally, the standard polynomial bounds for $\zeta(s)$, the uniform bound for $F_q(s)$ above, and the rapid vertical decay of $\widetilde h_W(s)$ give \begin{equation*} \mathcal E_W(x,q)\ll \left({x\over q}\right)^{-\frac14}=\left({q\over x}\right)^{\frac14}\ll 1. \end{equation*} Inserting this estimate and the two residue evaluations into \eqref{eq:singular-residues} proves \eqref{eq:singular-series}. \end{proof} \begin{proof}[Proof of \autoref{thm:hardy-littlewood-variance}] We may suppose that $x$ is sufficiently large, so that $q$ is odd. Since the inner $n$-sum is nonzero only when $|j|\ll x/q$, summing \eqref{eq:smooth-hardy-littlewood} in these ranges and applying \autoref{lem:singular-series} gives \begin{align}\label{eq:hardy-littlewood-off-diagonal} &\sum_{j\ne0}\sum_n\Lambda(n)\Lambda(n+jq)W\left({n\over x}\right)\overline{W\left({n+jq\over x}\right)}\notag\\ &\qquad=x\sum_{j\ne0}\mathfrak S(jq)A_W\left({jq\over x}\right)+O\left(x^{\frac32+\eta}q^{-1}\right)\notag\\ &\qquad={x^2\over q-1}|\widetilde W(1)|^2-x\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\log{x\over q}+O\left(x+x^{\frac32+\eta}q^{-1}\right). \end{align} Opening \eqref{eq:weighted-variance} and using orthogonality of residue classes gives \begin{align}\label{eq:hardy-littlewood-variance-expansion} G_W(x,q) &=\sum_{(n,q)=1}\Lambda(n)^2\left|W\left({n\over x}\right)\right|^2\notag\\ &\quad+\sum_{j\ne0}\sum_n\Lambda(n)\Lambda(n+jq)W\left({n\over x}\right)\overline{W\left({n+jq\over x}\right)}\notag\\ &\quad-{1\over q-1}\left|\sum_{(n,q)=1}\Lambda(n)W\left({n\over x}\right)\right|^2+O\left((\log(2x))^2\right), \end{align} where the error comes from dropping the coprimality condition in the sum over $j\ne0$. The diagonal calculation in the proof of \autoref{lem:large-modulus-variance} gives \begin{equation}\label{eq:hardy-littlewood-diagonal} \sum_{(n,q)=1}\Lambda(n)^2\left|W\left({n\over x}\right)\right|^2=x\log x\int_0^\infty| W(u)|^2\mathrm d u+O(x). \end{equation} Contour integration and GRH also give \begin{equation*} \sum_{(n,q)=1}\Lambda(n)W\left({n\over x}\right)=\sum_n\Lambda(n)W\left({n\over x}\right)+O(\log(2x))=x\widetilde W(1)+O\left(x^{\frac12+\eta}\right). \end{equation*} After squaring and dividing by $q-1$, this becomes \begin{equation}\label{eq:hardy-littlewood-mean-square} {1\over q-1}\left|\sum_{(n,q)=1}\Lambda(n)W\left({n\over x}\right)\right|^2={x^2\over q-1}|\widetilde W(1)|^2+O\left(x^{\frac32+\eta}q^{-1}\right). \end{equation} Substituting \eqref{eq:hardy-littlewood-off-diagonal}, \eqref{eq:hardy-littlewood-diagonal}, and \eqref{eq:hardy-littlewood-mean-square} into \eqref{eq:hardy-littlewood-variance-expansion}, we obtain \begin{align*} G_W(x,q) &=x\left(\int_0^\infty| W(u)|^2\mathrm d u\right)\left(\log x-\log{x\over q}\right)+O\left(x+x^{\frac32+\eta}q^{-1}\right)\\ &=x\log q\int_0^\infty| W(u)|^2\mathrm d u+O\left(x+x^{\frac32+\eta}q^{-1}\right). \end{align*} Taking $\eta=\varepsilon$ and using $q\ge x^{\frac12+\varepsilon}$ proves \eqref{eq:hardy-littlewood-variance}. \end{proof} \begin{remark}\label{rem:unified-conditional-variance} Combining \autoref{thm:hardy-littlewood-variance} with \autoref{lem:large-modulus-variance} gives, for every $\varepsilon>0$, \begin{equation}\label{eq:unified-conditional-variance} G_W(x,q)=x\min(\log x,\log q)\int_0^\infty| W(u)|^2\mathrm d u+O(x), \end{equation} as $x\to\infty$, uniformly for primes $q$ satisfying $q\ge x^{\frac12+\varepsilon}$. \end{remark} \section{Pair correlation}\label{sec:pair-correlation} \begin{proof}[Proof of \autoref{cor:restricted-pair-correlation}] Start with the unconditional identity \eqref{eq:fourier-identity}. Choose $\sigma$ with $B_\phi<\sigma<1$; then \autoref{thm:form-factor-ramp} is uniform on the support of $\widehat\phi$. The first ramp term contributes \begin{equation*} {\log q\over2\pi}\left(\int_{\mathbb R}| f(t)|^2\mathrm d t\right)\int_{\mathbb R}|\alpha|\widehat\phi(\alpha)\mathrm d\alpha. \end{equation*} Changing variables by $u=-\alpha\log q/(2\pi)$ and applying dominated convergence and Plancherel, we find that the second term of \autoref{thm:form-factor-ramp} contributes \begin{align*} &{(\log q)^2\over4\pi^2}\int_{\mathbb R}\widehat\phi(\alpha)\left|\widehat f\left(-{\alpha\log q\over2\pi}\right)\right|^2\mathrm d\alpha \\ &\qquad={\log q\over2\pi}\int_{\mathbb R}\widehat\phi\left(-{2\pi u\over\log q}\right)|\widehat f(u)|^2\mathrm d u \\ &\qquad\sim{\log q\over2\pi}\widehat\phi(0)\int_{\mathbb R}| f(t)|^2\mathrm d t. \end{align*} The contributions coming from the $O$-terms are \begin{equation*} \int_{\mathbb R}|\widehat\phi(\alpha)|\sqrt{\log(2q^{|\alpha|})}\mathrm d\alpha\ll_\phi\sqrt{\log q}, \end{equation*} \begin{align*} &\log q\int_{\mathbb R}|\widehat\phi(\alpha)|\left|\widehat f\left(-{\alpha\log q\over2\pi}\right)\right|\sqrt{\log(2q^{|\alpha|})}\mathrm d\alpha \\ &\qquad=2\pi\int_{\mathbb R}\left|\widehat\phi\left(-{2\pi u\over\log q}\right)\right||\widehat f(u)|\sqrt{\log2+2\pi| u|}\mathrm d u\ll_\phi 1. \end{align*} Both of these are $o(\log q)$, proving the corollary. \end{proof} \begin{proof}[Proof of \autoref{cor:conditional-pair-correlation}] Combining \eqref{eq:unified-conditional-variance} with \autoref{thm:form-factor-variance}, \eqref{eq:form-factor-symmetry}, and \eqref{eq:fourier-mellin-identities}, and using the observation in \autoref{rem:fourier-term-away-from-zero} when $\alpha$ is away from zero, we obtain, uniformly on compact subsets of $0<|\alpha|<2$, \begin{equation*} K(f,\alpha,q)\sim{\log q\over2\pi}\min(|\alpha|,1)\int_{\mathbb R}| f(t)|^2\mathrm d t. \end{equation*} Now choose $B$ with $B_\phi0$, \begin{equation*} \operatorname{supp}\widehat{\phi_1}\subset[-B,-\sigma]\cup[\sigma,B]. \end{equation*} The contribution of $\phi_0$ is given by \autoref{cor:restricted-pair-correlation}, while substitution of the uniform asymptotic into \eqref{eq:fourier-identity} gives the asserted contribution of $\phi_1$. Recombining the two contributions proves the asserted asymptotic. The identity \begin{equation*} \widehat\phi(0)+\int_{\mathbb R}\min(|\alpha|,1)\widehat\phi(\alpha)\mathrm d\alpha=\phi(0)+\int_{\mathbb R}\left(1-\left({\sin\pi u\over\pi u}\right)^2\right)\phi(u)\mathrm d u \end{equation*} shows that it is equivalent to \eqref{eq:full-pair-correlation-real}. \end{proof} \section{Simplicity of \texorpdfstring{$L$}{L}-zeros}\label{sec:simple-zeros} \begin{proof}[Proofs of \autoref{cor:smooth-simplicity} and \autoref{cor:conditional-simplicity}] Assume GRH and fix a nonnegative band-limited Schwartz function $f$. If $\rho_\chi$ is a zero, let $m_{\rho_\chi}$ denote its multiplicity, and put \begin{align*} M_f(q)&={1\over q-2}\sideset{}{^*}{\sum}_\chi\sum_{\gamma_\chi}| f(\gamma_\chi)|^2,\\ S_f(q)&={1\over q-2}\sideset{}{^*}{\sum}_\chi\sum_{\substack{\gamma_\chi\\ \rho_\chi\ \mathrm{simple}}}| f(\gamma_\chi)|^2. \end{align*} For $0<\beta<2$, let \begin{equation}\label{eq:fejer-kernel} \Phi_\beta(u)=\left({\sin\pi\beta u\over\pi\beta u}\right)^2, \qquad \widehat{\Phi_\beta}(\alpha)=\max\left(0,{\beta-|\alpha|\over\beta^2}\right). \end{equation} Applying \autoref{lem:explicit-formula} to $|f|^2$ gives \begin{equation}\label{eq:one-level-mass} M_f(q)={\log q\over2\pi}\int_{\mathbb R}| f(t)|^2\mathrm d t+O(1). \end{equation} Since $\Phi_\beta$ is nonnegative and $\Phi_\beta(0)=1$, positivity under GRH implies \begin{equation*} {1\over q-2}\sideset{}{^*}{\sum}_\chi\sum_{\gamma_\chi}| f(\gamma_\chi)|^2m_{\rho_\chi}\le R(f,\Phi_\beta,q). \end{equation*} The inequality $\mathbf 1_{m=1}\ge2-m$ for every positive integer $m$ therefore gives \begin{equation}\label{eq:fejer-simplicity-bound} S_f(q)\ge2M_f(q)-R(f,\Phi_\beta,q). \end{equation} For $0<\beta<1$, \autoref{cor:restricted-pair-correlation} and \eqref{eq:one-level-mass} give \begin{equation*} R(f,\Phi_\beta,q)\sim\left(\frac1\beta+\frac\beta3\right)M_f(q), \end{equation*} because \begin{equation*} \widehat{\Phi_\beta}(0)+\int_{\mathbb R}|\alpha|\widehat{\Phi_\beta}(\alpha)\mathrm d\alpha=\frac1\beta+\frac\beta3. \end{equation*} It follows from \eqref{eq:fejer-simplicity-bound} that $S_f(q)\ge(2-\frac1\beta-\frac\beta3-o_{f,\beta}(1))M_f(q)$. Letting $\beta\to1^-$ proves \autoref{cor:smooth-simplicity}. Under the additional smooth Hardy--Littlewood conjecture in \autoref{thm:hardy-littlewood-variance}, \autoref{cor:conditional-pair-correlation} applies to $\Phi_\beta$ when $B_{\Phi_\beta}=\beta\in(1,2)$. In this case, \begin{equation*} \widehat{\Phi_\beta}(0)+\int_{\mathbb R}\min(|\alpha|,1)\widehat{\Phi_\beta}(\alpha)\mathrm d\alpha=1+{1\over3\beta^2}, \end{equation*} so \eqref{eq:fejer-simplicity-bound} gives $S_f(q)\ge(1-{1\over3\beta^2}-o_{f,\beta}(1))M_f(q)$. Letting $\beta\to2^-$ proves \autoref{cor:conditional-simplicity}. \end{proof} \section{Comparison with related work}\label{sec:comparisons} Our study may be seen as a $q$-analog of Montgomery's~\cite{MontgomeryPairCorrelation1973} 1973 investigation of the pair correlation of $\zeta$-zeros. He considered the pair correlation of a single $\zeta(s)$. We compute the pair correlation of zeros of a single $L(s,\chi)$ and then average over primitive $\chi$ modulo $q$. His results concerned zeros in $[0,T]$ as $T\to\infty$, while we keep the height level fixed and let $q\to\infty$ through primes. Montgomery's pair correlation statistics is connected to primes in short intervals via the Goldston--Montgomery correspondence~\cite{GoldstonMontgomeryPairCorrelation1987}, while ours is connected to primes in arithmetic progressions via \autoref{thm:form-factor-variance}. Montgomery, Yıldırım~\cite{YildirimPairCorrelation1991}, and Kandhil--Languasco--Moree~\cite{KLMPairCorrelation2026} studied form factors for zeros whose ordinates lie in an enlarging height window ($[0,T]$ or $[-T,T]$ with $T\to\infty$). Montgomery considered pairs of zeros of a single $\zeta(s)$, while Yıldırım and Kandhil--Languasco--Moree paired a zero of $L(s,\chi_1)$ with a zero of $L(s,\chi_2)$ and then summed over all $\chi_1,\chi_2\bmod q$, so the two zeros in a pair need not belong to the same $L$-function. {\"O}zl{\"u}k~\cite{OzlukThesis1982}, Chandee--Lee--Liu--Radziwi{\l}{\l}~\cite{ChandeeLeeLiuRadziwill2014}, and our work studied form factors of zeros of Dirichlet $L$-functions with ordinates in a fixed height window. These works constructed form factors by first pairing zeros of the same $L(s,\chi)$ and then average over $\chi$, but they averaged $\chi$ differently. {\"O}zl{\"u}k averaged over all $\chi$ modulo $q\le Q$, so the same primitive $L$-function could reappear through induced characters, whereas Chandee--Lee--Liu--Radziwi{\l}{\l} averaged directly over primitive characters with conductors $q\asymp Q$. We only average over primitive characters with a fixed conductor $q$, requiring $q$ to be prime. The family we consider has also been studied through other statistics: Hughes--Rudnick~\cite{HughesRudnickLinearStatistics2003} studied its one-level density, with a scaling that restricted attention to low-lying zeros whose ordinates are $\ll 1/\log q$. We study pair correlation over this same family in a fixed height window as the prime conductor tends to infinity. In 2014, Keating--Rudnick~\cite{KeatingRudnickVariance2014} studied the function-field analog of the variance of primes in arithmetic progressions. In their model, the variance is related to traces of powers of Frobenius matrices, the function field analog of the form factor. To describe their setting, let $\mathbb F_{\mathfrak q}$ be a finite field, let $Q\in\mathbb F_{\mathfrak q}[T]$ be squarefree of degree $d\ge2$, and let $\mathcal M_n$ denote the set of monic polynomials of degree $n$. Set \begin{equation*} G(n;Q)=\sum_{\substack{A\bmod Q\\ (A,Q)=1}}\left|\sum_{\substack{N\in\mathcal M_n\\ N\equiv A\pmod Q}}\Lambda_{\mathfrak q}(N)-{\mathfrak q^n\over\Phi(Q)}\right|^2, \end{equation*} where $\Phi(Q)$ is the number of invertible residue classes modulo $Q$ and $\Lambda_{\mathfrak q}$ is the analog of the von Mangoldt function. Keating--Rudnick first treated the range $1\le n1$ in relevant calculations. We therefore restrict $q$ to prime moduli. \enlargethispage{2\baselineskip} \bibliographystyle{amsplain} \bibliography{refs} \end{document}