\documentclass[11pt]{amsart} \usepackage{shortcuts} \usepackage{amssymb} \title{Smooth Numbers in Arithmetic Progressions with Uniformity in the Residue Class} \author{Sarvagya Jain} \address{Department of Mathematics and Statistics, University of Turku, Turku, Finland} \email{sarvagyajain.math@gmail.com} \begin{document} \begin{abstract} We improve the level of distribution of smooth numbers in arithmetic progressions to $x^{5/8-\epsilon}$, with uniformity in the residue class. In particular, for every $\epsilon,A>0$ and integer $k\geq1$, there exist $\delta,C>0$ such that \[ \sum_{\substack{q\leq x^{5/8-\epsilon}\\(q,a_1a_2)=1}}\tau_k(q)\sup_{0\leq z\leq x} \Biggl|\Psi(z,y;a_1\overline{a_2},q) -\frac{1}{\phi(q)}\Psi_q(z;y)\Biggr| \ll_{\epsilon,A,k}\frac{\Psi(x;y)}{(\log x)^A}, \] uniformly for $x\geq2$, $(\log x)^C\leq y\leq x^{1/C}$, and $a_1,a_2\in\Z\setminus\{0\}$ satisfying \[ |a_1|\leq x^{1+\delta},\qquad |a_2|\leq x^\delta,\qquad (a_1,a_2)=1. \] Here $\Psi(z;y)$ counts the $y$-smooth integers up to $z$, $\Psi_q(z;y)$ counts those coprime to $q$, and $\Psi(z,y;a,q)$ counts those congruent to $a$ modulo $q$; $\tau_k$ is the $k$-fold divisor function. The proof combines dispersion with second and fourth spectral moments. \end{abstract} \maketitle \section{Introduction} Let $P^+(n)$ denote the largest prime factor of a positive integer $n$, with $P^+(1)=1$. We call $n$ $y$-smooth if $P^+(n)\leq y$. For $x\geq0$, $y\geq2$, $q\in\Z_{>0}$ and $a\in\Z$, write \begin{align*} \Psi(x;y)&:=\abs{\{n\in\Z_{>0}:n\leq x,\ P^+(n)\leq y\}},\\ \Psi_q(x;y)&:=\abs{\{n\in\Z_{>0}:n\leq x,\ P^+(n)\leq y,\ (n,q)=1\}},\\ \Psi(x,y;a,q)&:=\abs{\{n\in\Z_{>0}:n\leq x,\ P^+(n)\leq y,\ n\equiv a\pmod q\}}. \end{align*} For $(a,q)=1$, the expected count in the class $a\pmod q$ is $\Psi_q(x;y)/\phi(q)$, where $\phi$ is Euler's totient function. We study the discrepancy from this count, averaged over the moduli, while allowing the residue class and the summation endpoint to vary. Our main result gives exponent of distribution $5/8$, uniformly in residue classes of the form $a_1\overline{a_2}\pmod q$. Here $\overline{a_2}$ denotes the inverse of $a_2$ modulo $q$, and $\tau_k(q)$ is the number of ordered factorisations of $q$ into $k$ positive integers. \begin{theorem}\label{thm: distribution} For any $\epsilon,A>0$ and integer $k\geq1$, there exist $\delta,C>0$ such that the following holds. Let $x\geq2$ and $a_1,a_2\in\Z\setminus\{0\}$ satisfy \[ \abs{a_1}\leq x^{1+\delta},\qquad \abs{a_2}\leq x^\delta,\qquad (a_1,a_2)=1. \] Uniformly for $(\log x)^C\leq y\leq x^{1/C}$, we have \begin{align*} &\sum_{\substack{q\leq x^{5/8-\epsilon}\\(q,a_1a_2)=1}}\tau_k(q)\sup_{0\leq z\leq x} \Biggl|\Psi(z,y;a_1\overline{a_2},q) -\frac1{\phi(q)}\Psi_q(z;y)\Biggr| \ll_{\epsilon,A,k}\frac{\Psi(x;y)}{(\log x)^A}. \end{align*} \end{theorem} Section~\ref{sec: outline} outlines the argument. The convolution estimate used in the proof is stated in Section~\ref{subsec: convolution estimates}. We deduce Theorem~\ref{thm: distribution} in Section~\ref{sec: proof of distribution} and prove the convolution estimate in Section~\ref{sec: proof of triple conv}, using spectral moments and Kloosterman sums. \subsection{AI Usage} This manuscript was largely generated using ChatGPT/Codex, including much of the mathematical argumentation, exposition, and \LaTeX{} code. It builds on an earlier draft by the author concerning Balog's additive-representation problem, which used the exponent of distribution $153/256$. The spectral-moment strategy was suggested by ChatGPT in response to the author's prompts. The author takes responsibility for the final manuscript. \subsection{Acknowledgements} The author thanks Kaisa Matom\"aki for her encouragement throughout this project and for providing a ChatGPT subscription. The author was supported by the Research Council of Finland (grant numbers 346307, 364214, and 370133) and by the University of Turku Graduate School's Exactus fellowship while working on this project. \subsection{Notation} The letter $p$, with or without subscripts, is reserved for primes; sums and products indexed by $p$ are therefore over primes. All logarithms are natural. We write $\Z_{>0}$ for the positive integers, $(a,b)$ for the greatest common divisor, and $[a,b]$ for the least common multiple of $a$ and $b$; the same conventions apply to more than two integers. Divisors in sums and products are positive unless otherwise specified. For a summation variable, $n\sim N$ means $N0$, with $G\geq0$. Subscripts indicate permitted dependence of the implied constant, as in $O_\epsilon$ or $\ll_\epsilon$, and constants may change from line to line. For positive quantities, $F\asymp G$ means $F\ll G\ll F$. In a specified limit, $F=o(G)$ means $F/G\to0$. In particular, $F(x)\ll x^{o(1)}$ means $F(x)\ll_\epsilon x^\epsilon$ for every $\epsilon>0$. We write $\mu$ for the M\"obius function and $\tau=\tau_2$ for the divisor function. The radical is $\operatorname{rad}(n)=\prod_{p\mid n}p$, and $d\mid n^\infty$ means that every prime factor of $d$ divides $n$. A star on a residue sum restricts it to reduced residues; a star on a character sum restricts it to primitive characters. Overbars on complex coefficients denote complex conjugation; overbars on integers in modular expressions denote multiplicative inverses modulo the indicated modulus. For a bounded function $F$, write $\|F\|_\infty=\sup_u|F(u)|$. We use $J_\nu$ for the Bessel function of the first kind and $K_\nu$ for the modified Bessel function of the second kind. The spectral notation is introduced in Section~\ref{sec: proof of triple conv} and Appendix~\ref{sec: regular spectrum} as needed. \section{Outline}\label{sec: outline} We explain the exponent $5/8$ and the uniformity in the residue class, following the order of the overview in~\cite[Section~2]{PascadiLvlDist}. Throughout this sketch, estimates are understood up to factors $x^{o(1)}$, and we suppress smooth weights and divisor sums. To keep the notation simple, take $a_2=1$ and $a_1=a$ with $1\leq a\leq x$. Coprimality conditions needed for multiplicative inverses are understood. The full proof treats both signs and restores the small powers of $x$ needed for the stated residue range. \subsection{Reduction to exponential sums} The factorisation of smooth numbers allows us to choose three factor lengths $M,N,L$ with $MNL=x$. After treating characters of small conductor separately, we have to bound \[ \sum_{\substack{r\sim R\\(r,a)=1}} \left|\sum_{m\sim M}\sum_{n\sim N}\sum_{\ell\sim L} \alpha_m\beta_n\gamma_\ell\,\mcE_D(mn\ell\overline a;r)\right|, \] where $\mcE_D$ is the character remainder defined in Section~\ref{subsec: convolution estimates}. We seek a saving over its natural size $x$, with $R$ as large as possible. The limiting choice of lengths is \begin{equation}\label{eqn: outline factor lengths} M=L=\frac xR,\qquad N=\frac{R^2}{x},\qquad K=NL=R. \end{equation} These lengths indicate the exponent; they will be adjusted slightly at the end of the outline. Linnik's dispersion method begins with Cauchy--Schwarz in $r,m$. Opening the square introduces two copies of $n\ell$. Poisson summation in $m$ produces a frequency of length \[ H\approx\frac RM=N. \] The zero frequency combines with the other dispersion main terms. In the nonzero frequencies, write one of the products as $k\asymp K$ and pass from $r$ to the complementary divisor \[ t=\frac{n\ell-k}{r},\qquad |t|\ll\frac KR\approx1. \] Reciprocity and a further application of Cauchy--Schwarz lead to sums of the form \begin{equation}\label{eqn: outline Kloosterman fractions} \mathcal K= \sum_{n,n'\sim N}\alpha_{n,n'} \sum_{\substack{h,h'\sim H\\n'h\neq nh'}} \sum_{\substack{k\asymp K\\\ell\asymp L}} e\left(\frac{a(n'h-nh')\overline{nn'\ell}}k\right), \qquad e(u)=e^{2\pi iu}. \end{equation} Here the bounded complementary divisor is suppressed. The coefficients $\alpha_{n,n'}$ are bounded, and the $h,h',k,\ell$ sums have smooth weights. The diagonal $n'h=nh'$ is estimated separately. If $\mathcal R$ denotes the resulting off-diagonal dispersion error, the second Cauchy--Schwarz step gives schematically \[ |\mathcal R|\ll\frac MR(KL)^{1/2}|\mathcal K|^{1/2}. \] The dispersion main term has size $MK^2/R=xK/R$, so it remains to save a power of $x$ in \begin{equation}\label{eqn: outline normalised error} \left(\frac{|\mathcal R|}{MK^2/R}\right)^4 \ll\frac{(KL)^2|\mathcal K|^2}{K^8} =\frac{|\mathcal K|^2}{N^6L^4}. \end{equation} \subsection{The spectral estimates} For the spectral calculation, first take $n,n'$ squarefree with $(n,n')=1$ and $|n'h-nh'|\asymp N^2$. We display only the positive frequencies. Completing the $\ell$-sum in~\eqref{eqn: outline Kloosterman fractions} introduces a frequency $m$ of length $K/L\approx N$ and classical Kloosterman sums \[ S(m\overline q,av;k),\qquad q=nn',\quad v=n'h-nh'. \] Kuznetsov's formula then introduces forms of level $q$. Thus the relevant sizes are \[ q\asymp Q=N^2,\qquad m\asymp N,\qquad v\asymp Q. \] The spectral level $Q$ is different from the original modulus size $R$. Common factors of $n,n'$ and smaller values of $|v|$ are handled in the full proof. Write $\rho_{f,\mathfrak a}(n)$ for the Fourier coefficients in the normalisation of Section~\ref{sec: proof of triple conv}, and $\lambda_f(n)$ for the Hecke eigenvalues. We first consider the exceptional Maass contribution $\mathcal K_{\mathrm{exc}}$. Put $\theta_f=\sqrt{1-4\lambda_f}$ and $\kappa_f=i\theta_f/2$, where $\lambda_f$ without an argument is the Laplace eigenvalue. The Kuznetsov transform contributes the factor $\max(1,X)^{\theta_f}$, where \[ X^2\approx\frac{K^2Q}{aNQ}=\frac xa. \] The inverse Hecke relation separates the residue $a$ from the Fourier index $av$: \[ \rho_{f,0}(av)=\sum_{e\mid(a,v)}\mu(e)\lambda_f(a/e)\rho_{f,0}(v/e) \qquad((a,q)=1). \] Ignoring these divisor sums for the sketch, the two Fourier polynomials have the shapes \[ P_f=\sum_{m\asymp N}\psi(m/N)\rho_{f,\infty}(m),\qquad L_f=\sum_{v\asymp Q}c_v \rho_{f,0}(v), \] where $\psi$ is a fixed smooth cutoff and, for each pair $n,n'$, the coefficients $c_v$ come from the smooth sum over $h,h'$ subject to $n'h-nh'=v$. The coefficients of $P_f$ are common to all levels; those of $L_f$ retain this additive structure. The residue now occurs through $\lambda_f(a)$. A pointwise bound for this eigenvalue would lose a power of $a$. Instead, we apply H\"older with exponents $2,4,4$, keeping it in a second moment. Put \[ W=N,\qquad V=\max\left\{1,\frac{x}{a\sqrt N}\right\}. \] Then $V^{\theta_f/2}W^{\theta_f/4}\geq\max(1,X)^{\theta_f}$ at the limiting sizes, up to a constant. Consequently, \begin{align*} |\mathcal K_{\mathrm{exc}}|^2\ll{}&L^2Q \left(\sum_{q\asymp Q}\sum_f V^{\theta_f}|\lambda_f(a)|^2\right)\\ &\quad\times\left(\sum_{q\asymp Q}\sum_f\frac{W^{\theta_f}|P_f|^4}{\cosh^2(\pi\kappa_f)}\right)^{1/2} \left(\sum_{q\asymp Q}\sum_f\frac{|L_f|^4}{\cosh^2(\pi\kappa_f)}\right)^{1/2}. \end{align*} Here $f$ runs over the exceptional Maass forms at level $q$, and the levels satisfy $(q,a)=1$. The fourth moment of $P_f$, obtained from Watt's estimate and the Hecke relation, admits the weight $W=Q^2/N^3=N$. To bound the fourth moment of $L_f$, we use the additive structure of $c_v$, Poisson summation and Voronoi summation. At these lengths, Lemmas~\ref{lem: improved short fourth moment} and~\ref{lem: improved long fourth moment} give \[ \left(\sum_{q\asymp Q}\sum_f\frac{W^{\theta_f}|P_f|^4}{\cosh^2(\pi\kappa_f)}\right)^{1/2}\ll N, \qquad \left(\sum_{q\asymp Q}\sum_f\frac{|L_f|^4}{\cosh^2(\pi\kappa_f)}\right)^{1/2}\ll Q. \] For the Hecke factor, with $\theta=7/32$, the fixed-index second moment in Lemma~\ref{lem: improved Hecke second moment} gives \[ \sum_{q\asymp Q}\sum_f V^{\theta_f}|\lambda_f(a)|^2 \ll Q(Q+\sqrt a) \max\left\{1,\frac{x}{\sqrt N\max\{a,Q^2\}}\right\}^{\theta}. \] The factor after $Q$ is nondecreasing in $a$, up to an absolute constant, since $\theta<1/2$. For $a\leq x$ and $Q\leq\sqrt x$, it is therefore $O(\sqrt x)$. Combining the three estimates yields \begin{equation}\label{eqn: outline spectral bound} |\mathcal K_{\mathrm{exc}}|^2\ll L^2Q\cdot Q\sqrt x\cdot N\cdot Q =L^2N^7\sqrt x. \end{equation} This is the step that allows the residue to grow with $x$. The exceptional weights are retained in the moments, and the residue stays fixed while its Hecke eigenvalue is averaged over forms and levels. Proposition~\ref{propn: improved squarefree Kloosterman} proves the same bound for $\mathcal K$, including the real Maass, holomorphic and Eisenstein contributions and the suppressed common factors. \subsection{The exponent \texorpdfstring{$5/8$}{5/8}} Substituting~\eqref{eqn: outline spectral bound} into~\eqref{eqn: outline normalised error} gives \[ \left(\frac{|\mathcal R|}{MK^2/R}\right)^4 \ll\frac{N\sqrt x}{L^2} =\frac{R^4}{x^{5/2}}. \] Thus a power saving is available when $R\leq x^{5/8-\epsilon}$. The condition $Q=N^2\leq\sqrt x$ used above gives the same endpoint. At the exact lengths~\eqref{eqn: outline factor lengths}, however, the diagonal ratio $N^2L^3/(xR)$ equals $1$. To obtain a saving there as well, choose a small fixed $\lambda>0$ and take \[ M=\frac{x^{1-\lambda}}R,\qquad L=\frac{x^{1-3\lambda}}R,\qquad N=\frac{R^2}{x^{1-4\lambda}}. \] Then $MNL=x$, $L/M=x^{-2\lambda}$ and $NL/R=x^\lambda$. Including the complementary-divisor cost from Proposition~\ref{propn: improved triple convolution}, the two relevant ratios become \[ \frac{N^2L^3}{xR}=x^{-\lambda},\qquad \left(\frac{NL}{R}\right)^{C_*}\frac{N\sqrt x}{L^2} =R^4x^{-5/2+(C_*+10)\lambda}. \] Both save a fixed power for $R\leq x^{5/8-\epsilon}$ when $\lambda$ is sufficiently small in terms of $\epsilon$. Section~\ref{sec: proof of distribution} checks the remaining length conditions. The auxiliary squarefree and coprimality restrictions, the range of $a_2$, and the small extension $|a_1|\leq x^{1+\delta}$ are accounted for in the convolution proof without changing this exponent. \section{Convolution estimate}\label{subsec: convolution estimates} We introduce the character sums used to separate small and large conductors in the proof of Theorem~\ref{thm: distribution}. Let \begin{align*} \omega_D(k;r):=\sum_{\substack{\chi\in \mcX_D\\\operatorname{cond}(\chi)\mid r}}\chi(k), \quad\text{where}\quad \mcX_D := \set{\chi \text{ primitive}: \operatorname{cond}(\chi)\leq D}. \end{align*} Note that \begin{align*} \abs{\omega_D(k;r)}\leq \abs{\mcX_D}\leq D^2. \end{align*} Let \begin{align*} \mcE_D(k; r) := \mathbf{1}_{k\equiv 1 \pmod r} - \frac{\mathbf{1}_{(k, r)=1}}{\phi(r)}\omega_D(k;r)=\frac{\mathbf{1}_{(k, r)=1}}{\phi(r)}\sum_{\substack{\chi\pmod r\\ \operatorname{cond}(\chi)> D}}\chi(k). \end{align*} The following convolution estimate controls the contribution of characters with large conductors in Theorem~\ref{thm: distribution}. \begin{proposition}\label{propn: improved triple convolution} There is an absolute constant $C_*>1$ with the following property. For every $\kappa>0$, there is a $\delta>0$ such that the following holds. Let $M,N,L,R\gg1$, $x=MNL$, and let $a_1,a_2\in\Z\setminus\{0\}$ satisfy \[ |a_1|\leq x^{1+\delta},\qquad |a_2|\leq x^\delta,\qquad (a_1,a_2)=1. \] Let $(\alpha_m)_{m\sim M}$, $(\beta_n)_{n\sim N}$ and $(\gamma_\ell)_{\ell\sim L}$ be $1$-bounded complex sequences. Suppose \begin{equation}\label{eqn: improved convolution hypotheses} \begin{gathered} N\geq x^\kappa,\qquad NL\leq x^{2/3-5\kappa},\qquad L\leq x^{-\kappa}M,\qquad M\leq R\leq x^{-\kappa}NL,\\ N^2\leq x^{1/2-\kappa},\qquad \frac{N^2L^3}{xR}\leq x^{-\kappa},\qquad \left(\frac{NL}{R}\right)^{C_*}\frac{N\sqrt x}{L^2}\leq x^{-\kappa}. \end{gathered} \end{equation} Then, for every $1\leq D\leq x^\kappa$, \[ \sum_{\substack{r\sim R\\(r,a_1a_2)=1}} \left|\sum_{m\sim M}\sum_{n\sim N}\sum_{\ell\sim L} \alpha_m\beta_n\gamma_\ell\, \mcE_D(mn\ell\overline{a_1}a_2;r)\right| \ll_\kappa\frac{x(\log x)^4}{\sqrt{\min\{x^\delta,D\}}}. \] \end{proposition} The proof is given in Section~\ref{sec: proof of triple conv}. We follow the dispersion arguments of Drappeau~\cite[Théorème~3]{Drappeau} and Pascadi~\cite[Proposition~6.3]{PascadiLvlDist}. The dependence on the residue is controlled by a second moment for its Hecke eigenvalue and fourth moments for the two remaining spectral polynomials. We first assume that $\beta_n$ is supported on squarefree integers and that the auxiliary coprimality condition~\eqref{eqn: coprime support for triple convolution} below holds. These restrictions are removed at the end of the proof by separating powerful parts and common factors, and applying the multiplicative large sieve. \section{Proof of Theorem~\ref{thm: distribution}}\label{sec: proof of distribution} We use the standard smooth-number estimate \begin{equation}\label{eqn: coarse smooth count} \Psi(x;(\log x)^C)=x^{1-1/C+o(1)}\qquad(x\to\infty) \end{equation} for each fixed $C>1$, which follows from~\cite[Theorems~1 and~2(i)]{Hildebrand-Tenenbaum}. By monotonicity in $y$, for every fixed $c>0$ we therefore have $\Psi(x;y)\geq x^{1-c}$ when $y\geq(\log x)^C$, provided $C$ and then $x$ are sufficiently large. \begin{proof} We use Proposition~\ref{propn: improved triple convolution} and give the factor lengths explicitly. Write $\Delta(z;q):=\Psi(z,y;a_1\overline{a_2},q)-\Psi_q(z;y)/\phi(q)$. We first prove the estimate without the supremum over $z$, that is, for any fixed $B>0$, \begin{equation}\label{eqn: weighted endpoint distribution} \sum_{\substack{q\leq x^{5/8-\epsilon}\\(q,a_1a_2)=1}} \tau_k(q)\abs{\Delta(x;q)} \ll_{\epsilon,B,k}\frac{\Psi(x;y)}{(\log x)^B}. \end{equation} Following the proof of~\cite[Theorem~7.1]{PascadiLvlDist}, with $q_0=1$, $y_1=y$, $y_2=x$, and the given $k$, split the characters at conductor $D=x^\eta$, where $\eta>0$ is sufficiently small. The argument in~\cite[p.~40, paragraph beginning ``For $S_{\mathrm{small}}$'']{PascadiLvlDist}, with $A=B+1$ and summation over dyadic ranges of moduli, gives, for sufficiently large $C$, \[ \sum_{q\leq x^{5/8-\epsilon}}\frac{\tau_k(q)}{\phi(q)} \sum_{\substack{\chi\pmod q\\1<\operatorname{cond}(\chi)\leq D}} \Biggl|\sum_{\substack{1\leq n\leq x\\P^+(n)\leq y}}\chi(n)\Biggr| \ll_{\epsilon,B,k}\frac{\Psi(x;y)}{(\log x)^B}. \] This bound is uniform in the residue: after the triangle inequality, the factors $\overline{\chi(a_1)}\chi(a_2)$ disappear as they have modulus one for $(q,a_1a_2)=1$. For $D\leq R\leq x^{1/2}$, Harper's estimate~\cite[Proposition~2]{HarperBV} gives \begin{align*} &\sum_{q\leq R}\frac1{\phi(q)} \sum_{\substack{\chi\pmod q\\\operatorname{cond}(\chi)>D}} \Biggl|\sum_{\substack{1\leq n\leq x\\P^+(n)\leq y}}\chi(n)\Biggr|\\ &\qquad\ll (\log x)^{7/2}\sqrt{\Psi(x;y)} \left(R+x^{1/2-\eta}(\log x)^2\right), \end{align*} provided $\eta$ is sufficiently small and $C$ is sufficiently large. Taking $R=x^{1/2-c}$, with $c>0$ sufficiently small and fixed, and using $\Psi(x;y)\leq x$, gives $O(x^{1-c_1})$ for some $c_1>0$. This bound is independent of the residue class. For the remaining moduli we use Proposition~\ref{propn: improved triple convolution}. We give the length calculation, since it determines the improved exponent. We may assume $0<\epsilon<1/100$. Choose \[ 0<\lambda<\min\left\{\frac1{1000}, \frac{\epsilon}{100(C_*+100)}\right\},\qquad \kappa=\frac\lambda4, \] and take the constant $c$ in the preceding paragraph smaller than $\lambda/1000$. Let $\delta_*>0$ be supplied by Proposition~\ref{propn: improved triple convolution} with parameter $\kappa$. Choose $\zeta,\delta>0$ sufficiently small in terms of $\epsilon,\lambda,\delta_*$, and then choose the conductor exponent $\eta>0$ sufficiently small in terms of $\zeta,\kappa,\delta_*$. The constant $C$ is increased last. First discard the integers $n\leq x^{1-\zeta}$. The kernel $\mcE_D$ defined in Section~\ref{subsec: convolution estimates} satisfies \[ \sum_{\substack{q\leq x^{5/8-\epsilon}\\(q,a_1a_2)=1}} \left|\sum_{\substack{n\leq x^{1-\zeta}\\P^+(n)\leq y}} \mcE_D(na_2\overline{a_1};q)\right| \ll x^{1-\zeta+2\eta+o(1)}+x^{5/8-\epsilon+o(1)}. \] Indeed, the progression term is bounded by summing the divisors of $a_2n-a_1$; the possible equality $a_2n=a_1$ contributes at most one term for each modulus. The other term is bounded using $|\omega_D|\leq D^2$ and $\sum_{q\leq x}1/\phi(q)\ll\log x$. Since $2\eta<\zeta$, this is a power saving. Split the remaining integers into intervals $(X,2X]$ with $X\gg x^{1-\zeta}$, and split the moduli into dyadic intervals $q\sim R$. By the choices above, the retained blocks satisfy \[ X^{1/2-\lambda/100}\leq R\leq X^{5/8-\epsilon/2},\qquad |a_1|\leq X^{1+\delta_*},\quad |a_2|\leq X^{\delta_*},\quad D\leq X^{\min\{\kappa,\delta_*\}}. \] For each such block put \[ M_0=\frac{X^{1-\lambda}}R,\qquad L_0=\frac{X^{1-3\lambda}}R,\qquad N_0=\frac{R^2}{X^{1-4\lambda}},\qquad K_0=N_0L_0. \] Then \begin{align*} M_0N_0L_0&=X,& \frac{K_0}R&=X^\lambda,& \frac{L_0}{M_0}&=X^{-2\lambda},& \frac{N_0^2L_0^3}{XR}&=X^{-\lambda}. \end{align*} Writing $R=X^\rho$, the two critical inequalities are \begin{align*} \log(N_0^2)&=(4\rho-2+8\lambda)\log X\\ &\leq\left(\frac12-2\epsilon+8\lambda\right)\log X <\left(\frac12-\kappa\right)\log X,\\ \log\left\{\left(\frac{K_0}R\right)^{C_*} \frac{N_0\sqrt X}{L_0^2}\right\} &=\left(4\rho-\frac52+(C_*+10)\lambda\right)\log X\\ &\leq\bigl(-2\epsilon+(C_*+10)\lambda\bigr)\log X <-\epsilon\log X. \end{align*} The remaining conditions in~\eqref{eqn: improved convolution hypotheses} follow from \begin{gather*} \log N_0=(2\rho-1+4\lambda)\log X>3\lambda\log X,\\ M_0\leq R,\qquad K_0=RX^\lambda\leq X^{2/3-5\kappa}. \end{gather*} All these inequalities retain a fixed positive margin under sufficiently small fixed-power changes in the lengths. Now use the smooth-number decomposition in the proof of~\cite[Theorem~6.4]{PascadiLvlDist}, with these target lengths. Its separated, $1$-bounded coefficient sequences have dyadic lengths differing from $M_0,N_0,L_0$ by powers of $y$ of bounded degree, and their product is comparable to $X$. The separation integrals and dyadic sums have polylogarithmic total mass, as in~\cite[Section~3.7]{Drappeau}. Taking $C$ sufficiently large makes these length changes smaller than the margins above. The residue and cutoff inequalities likewise hold at the actual product of the lengths. Proposition~\ref{propn: improved triple convolution} therefore gives $O(x^{1-\kappa_1})$ for some $\kappa_1>0$, after summing the blocks and the separation integrals. Combining the two ranges gives, for some $c_2>0$, \[ \sum_{\substack{q\leq x^{5/8-\epsilon}\\(q,a_1a_2)=1}}\frac1{\phi(q)} \Biggl|\sum_{\substack{\chi\pmod q\\\operatorname{cond}(\chi)>D}} \overline{\chi(a_1)}\chi(a_2) \sum_{\substack{1\leq n\leq x\\P^+(n)\leq y}}\chi(n)\Biggr| =O(x^{1-c_2}). \] For fixed $k$, multiplication by $\max_{q\leq x}\tau_k(q)=x^{o(1)}$ leaves $O(x^{1-c_2/2})$, which is $O(\Psi(x;y)(\log x)^{-B})$ by~\eqref{eqn: coarse smooth count} once $C$ is sufficiently large. Thus, with suitable constants depending on $\epsilon,B,k$, we have established~\eqref{eqn: weighted endpoint distribution}. We now introduce the supremum, following the argument in the proof of~\cite[Corollaire~1]{Drappeau}. Put $Q=x^{5/8-\epsilon}$ and $t_0=x^{1-\gamma}$, where $\gamma>0$ will be chosen sufficiently small in terms of $\epsilon,A,k$. We first consider $0\leq z0}$, let $\mathcal B_q$ be a Petersson-orthonormal basis of Maass cusp forms for $\Gamma_0(q)$ with trivial nebentypus. For a primitive form $g$ of level $s\mid q$, its oldspace at level $q$ is the span of $g(\delta z)$ for $\delta\mid q/s$. Choose simultaneous eigenforms for the Laplacian, the unramified Hecke operators and reflection, separately in each such oldspace. Thus \[ f(-\overline z)=\varepsilon_f f(z),\qquad \varepsilon_f\in\{-1,1\}. \] Write the Laplace eigenvalue as $\lambda_f=1/4+\kappa_f^2$ and let $\mathcal B_q^{\mathrm{exc}}$ denote the exceptional forms, for which $\kappa_f=i\theta_f/2$. All spectral sums in this subsection are over exceptional forms. Kim--Sarnak~\cite[Appendix~2]{Kim-Sarnak} gives \begin{equation}\label{eqn: defn of theta} \theta:=\frac7{32},\qquad \theta_f\leq\theta,\qquad |\lambda_f(m)|\leq\tau(m)m^{\theta/2}\quad((m,q)=1), \end{equation} where $\lambda_f(m)$ denotes the normalised Hecke eigenvalue. At the cusps $\infty$ and $0$ take \[ \sigma_\infty=I,\qquad \sigma_0=\begin{pmatrix}0&-q^{-1/2}\\q^{1/2}&0\end{pmatrix}. \] Write $\rho_{f,\mathfrak a}(n)$ for the Fourier coefficients in the normalisation of~\cite[Section~3]{PascadiLargeSieve}. The spectral weights below are positive: on the exceptional spectrum, $\cosh(\pi\kappa_f)=\cos(\pi\theta_f/2)\geq\cos(\pi\theta/2)>0$. The matrix $\sigma_0$ acts by $z\mapsto-1/(qz)$. The corresponding operator $f(z)\mapsto f(-1/(qz))$ is the Fricke involution. It preserves each primitive oldspace and commutes with the operators above, so we choose the forms in $\mathcal B_q$ to be its eigenfunctions as well. Its eigenvalues are $\pm1$, and hence \[ \rho_{f,0}(n)=\pm\rho_{f,\infty}(n), \] with a sign depending only on $f$. It therefore suffices to prove the Fourier-moment bounds at $\infty$; the bounds at $0$ are identical. The auxiliary coprimality condition~\eqref{eqn: coprime support for triple convolution} allows us to use the unramified Fourier-coefficient factorisation of~\cite{Assing-Blomer-Li,Lichtman}. For $b,n\in\Z_{>0}$, $(b,q)=1$, and $\mathfrak a\in\{\infty,0\}$, the inverse Hecke relation is \begin{equation}\label{eqn: inverse Hecke relation} \rho_{f,\mathfrak a}(bn)=\sum_{d\mid(b,n)} \mu(d)\lambda_f(b/d)\rho_{f,\mathfrak a}(n/d). \end{equation} At infinity this is M\"obius inversion of the usual Hecke relation, and the relation $\rho_{f,0}(n)=\pm\rho_{f,\infty}(n)$ gives the same identity at $0$. Reflection gives $\rho_{f,\mathfrak a}(-n)=\varepsilon_f \rho_{f,\mathfrak a}(n)$. The Hecke relation holds for oldforms in the chosen basis; no nonvanishing assumption on $\rho_{f,\mathfrak a}(1)$ is needed, since we do not divide by that coefficient. Put $e(t)=e^{2\pi it}$ and \[ \widehat F(t)=\int_\R F(u)e(-tu)\,du. \] Coefficient sequences are extended by zero outside their indicated support. A smooth function has complexity at most $Z$ if its $j$th derivative is $O_j(Z^j)$ on its fixed compact support. Constants $C_0,C_1,\ldots$ in this section are absolute after fixing the supports and the finitely many smoothness orders used. \begin{lemma}[Exceptional spectral large-sieve bounds]\label{lem: improved moment inputs} Let $q\in\Z_{>0}$, $\eta>0$, and $\mathfrak a\in\{\infty,0\}$. \begin{enumerate} \item[(i)] For $J\geq1/2$ and any complex sequence $(b_n)$ supported on $n\asymp J$, \begin{equation}\label{eqn: improved ordinary second moment} \sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{1}{\cosh(\pi\kappa_f)} \left|\sum_n b_n\rho_{f,\mathfrak a}(n)\right|^2 \ll_\eta(qJ)^\eta\left(1+\frac Jq\right)\sum_n|b_n|^2. \end{equation} \item[(ii)] For $b\in\Z_{>0}$ with $(b,q)=1$ and $00$ and let $g$ be a Petersson-normalised primitive exceptional Maass form of level $s$. For any fixed even nebentypus, \begin{equation}\label{eqn: improved first coefficient} s^{-1-\eta} \ll_\eta\frac{|\rho_{g,\infty}(1)|^2}{\cosh(\pi\kappa_g)} \ll_\eta s^{-1+\eta}. \end{equation} If the nebentypus is trivial, $s\mid q$, and $f\in\mathcal B_q$ belongs to the oldspace generated by $g$, then \begin{equation}\label{eqn: improved oldform expansion} \frac{\rho_{f,\infty}(n)}{\sqrt{\cosh(\pi\kappa_f)}} =\sum_{\delta\mid q/s}c_f(\delta)\lambda_g(n/\delta), \qquad |c_f(\delta)|\ll_\eta q^{-1/2+\eta}\sqrt\delta. \end{equation} Here $\lambda_g(u)=0$ when $u$ is not a positive integer. \end{lemma} \begin{proof} Equation~\eqref{eqn: improved first coefficient} follows from~\cite[(0.9)--(0.10), Corollary~0.3]{HoffsteinLockhart}: their Petersson norm is divided by the volume, whereas ours is not, and $\operatorname{vol}(\Gamma_0(s)\backslash\mathbb H)=s^{1+o(1)}$. The lower bound is the one recorded immediately before their~(0.10). For the standard orthonormal oldform basis,~\cite[(3.8)--(3.11)]{BlomerKhan} gives~\eqref{eqn: improved oldform expansion} with the stated coefficient bound, since $\theta<1$. A unitary change to our chosen basis costs at most $\sqrt{\tau(q/s)}$ by Cauchy--Schwarz. This is absorbed into $q^\eta$. \end{proof} \begin{lemma}[Watt's mean-value estimate]\label{lem: improved Watt estimate} Let $\eta>0$, $Q,Z\geq1$, $J_1,J_2\geq1/2$, and \[ 00}$ with $(b,q)=1$ and $V\geq1$, \[ \sum_{f\in\mathcal B_q^{\mathrm{exc}}} V^{\theta_f}|\lambda_f(b)|^2 \ll_\eta(qbV)^\eta(q+\sqrt b) \max\left\{1,\frac{V}{\max\{1,q^2/b\}}\right\}^{\theta}. \] \end{lemma} \begin{proof} For a primitive form $g$ of level $s\mid q$, let $\mathcal B_q(g)$ be the basis of its oldspace at level $q$. The normalised lift $g/[\Gamma_0(s):\Gamma_0(q)]^{1/2}$ and Cauchy--Schwarz give \begin{equation}\label{eqn: oldspace first coefficient mass} \sum_{f\in\mathcal B_q(g)} \frac{|\rho_{f,\infty}(1)|^2}{\cosh(\pi\kappa_f)} \geq\frac{|\rho_{g,\infty}(1)|^2} {[\Gamma_0(s):\Gamma_0(q)]\cosh(\pi\kappa_g)} \gg_\eta q^{-1-\eta}, \end{equation} by~\eqref{eqn: improved first coefficient} and $[\Gamma_0(s):\Gamma_0(q)]\ll_\eta(q/s)q^\eta$, with sufficiently small exponents in these inputs. Since $(b,q)=1$, we have $\rho_{f,\infty}(b)=\lambda_g(b)\rho_{f,\infty}(1)$ throughout this oldspace. Its dimension is $\tau(q/s)$, and $\theta_f=\theta_g$. Summing over the primitive forms therefore gives \[ \sum_{f\in\mathcal B_q^{\mathrm{exc}}}V^{\theta_f}|\lambda_f(b)|^2 \ll_\eta q^{1+\eta} \sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{V^{\theta_f}|\rho_{f,\infty}(b)|^2}{\cosh(\pi\kappa_f)}. \] Apply~\eqref{eqn: improved fixed index} at level $q$, first with $X=\min\{V,\max(1,q^2/b)\}$ and then using $\theta_f\leq\theta$ to increase the weight to $V$. Choosing the intermediate small exponents below $\eta$ proves the claim. \end{proof} \begin{lemma}[Fourth moments]\label{lem: improved ordinary fourth moment} For $|\beta_n|\leq1$ supported on $n\asymp J$, independently of $f$, \begin{equation}\label{eqn: improved ordinary fourth moment} \sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{1}{\cosh^2(\pi\kappa_f)} \left|\sum_n\beta_n\rho_{f,\mathfrak a}(n)\right|^4 \ll_\eta(qJ)^\eta \left(\frac{J^2}{q}+\frac{J^4}{q^2}\right), \qquad \mathfrak a=\infty,0. \end{equation} For primitive exceptional forms of fixed level $s$ and nebentypus $\omega$ of conductor $m_\omega$, one also has \begin{equation}\label{eqn: improved primitive fourth moment} \left(\sum_{\substack{g\ \mathrm{primitive}\ (s,\omega)\\g\ \mathrm{exceptional}}} \left|\sum_n\beta_n\lambda_g(n)\right|^4\right)^{1/4} \ll_\eta(sJ)^\eta m_\omega^{C_0} \left(s^{1/4}\sqrt J+J\right). \end{equation} \end{lemma} \begin{proof} The Hecke relation, with $\omega$ extended by zero at primes dividing $s$, gives \[ \left(\sum_{n\asymp J}\beta_n\lambda_g(n)\right)^2 =\sum_{k\ll J^2}c_k\lambda_g(k),\qquad c_k=\sum_{uv=k}\sum_{\substack{du,dv\asymp J}}\omega(d)\beta_{du}\beta_{dv}. \] Thus $|c_k|\ll J\tau(k)/\sqrt k$ and $\sum_k|c_k|^2\ll_\eta J^{2+\eta}$. The nebentypus large sieve~\cite[Proposition~4.7]{DrappeauDispersion} has factor $1+m_\omega^{1/2}J/s$ at length $J$. Apply it at length $O(J^2)$ and use~\eqref{eqn: improved first coefficient}; the fourth moment of the Hecke polynomial is \[ \ll_\eta(sJ)^\eta \left(sJ^2+m_\omega^{1/2}J^4\right). \] This proves~\eqref{eqn: improved primitive fourth moment}. For trivial nebentypus, insert~\eqref{eqn: improved oldform expansion} and apply H\"older to its divisor sum. For each $s\mid q$ and $\delta\mid q/s$, the resulting contribution is bounded, apart from the displayed small powers, by \[ \frac{\delta^2}{q^2} \left(s+\frac{J^2}{\delta^2}\right)\frac{J^2}{\delta^2} =\frac{sJ^2}{q^2}+\frac{J^4}{\delta^2q^2} \leq\frac{J^2}{q}+\frac{J^4}{q^2}. \] Summing over $s,\delta$ and the $\tau(q/s)$ oldforms costs only a divisor factor, which is absorbed into the small power. \end{proof} We record the transform bounds used in the two fourth-moment arguments. For a primitive exceptional form $g$ of level $s$ and nebentypus conductor $m_\omega$, suppose $(a,c)=1$, $(c,m_\omega)=1$, $h=(c,s)$ is an exact divisor of $s$, meaning $(h,s/h)=1$, and $(c/h,s)=1$. Exact-divisor Voronoi summation~\cite[Proposition~5.4, Corollary~6.3]{AssingCorbett} expresses \begin{equation}\label{eqn: improved Voronoi} \sum_n\lambda_g(n)e(an/c)F(n/J) =\frac{J}{\sqrt{\mathfrak C}}\sum_{\pm}\epsilon_{g,\pm} \sum_{n\geq1}\lambda_{g^*}(n)e(a_\pm n/c) [\mathcal H_{\kappa_g}^{\pm}F](nJ/\mathfrak C), \qquad \mathfrak C=\frac{sc^2}{h}. \end{equation} Here $|\epsilon_{g,\pm}|\leq1$; $g^*$ is the corresponding Atkin--Lehner partner, and the dual residues are fixed on each family of fixed level and nebentypus. The map on primitive forms is bijective. For $c=1$ and trivial nebentypus, we may take $g^*=g$, absorbing its eigenvalue under $g(z)\mapsto g(-1/(sz))$ into $\epsilon_{g,\pm}$. The transforms, with the parity sign included in $\epsilon_{g,-}$, are \begin{align*} \mathcal H_t^+F(v)&=\frac{\pi i}{\sinh(\pi t)} \int_0^\infty\bigl(J_{2it}(4\pi\sqrt{vu})-J_{-2it}(4\pi\sqrt{vu})\bigr)F(u)\,du,\\ \mathcal H_t^-F(v)&=4\cosh(\pi t)\int_0^\infty K_{2it}(4\pi\sqrt{vu})F(u)\,du. \end{align*} At $t=0$ the removable limits are understood. The conductor is $[c^2,m_\omega c,s]=sc^2/h$. To see the normalisation directly, put $s_1=s/h$ and choose $d_0,b_0\in\Z$ with $as_1d_0-b_0c=1$. Since $(s_1,c)=1$, the matrix \[ W=\frac1{\sqrt{s_1}} \begin{pmatrix}as_1&b_0\\cs_1&s_1d_0\end{pmatrix} \] is an Atkin--Lehner matrix for the exact divisor $s_1$ of $s$, and \[ W^{-1}\left(\frac ac+iy\right) =-\frac{d_0}c+\frac{i}{c^2s_1y}. \] Thus the dual conductor is $c^2s_1=\mathfrak C$ and the prefactor is $1/(c\sqrt{s_1})=\mathfrak C^{-1/2}$. The condition $(c,m_\omega)=1$ places the nebentypus conductor in $s_1$, so the exact-divisor cusp relation gives the stated partner family and a scalar of modulus one. \begin{lemma}\label{lem: improved Hankel} Let $F$ have fixed compact support in $(0,\infty)$ and complexity $Z$. For a fixed dyadic cutoff $\psi$ put $F_{t,\rho}(u)=\psi(u)[\mathcal H_t^\pm F](\rho u)$. For $t\in i[-\theta/2,\theta/2]$, every fixed integer $j\geq0$ and real $B\geq0$ satisfy \[ \|F_{t,\rho}^{(j)}\|_\infty\ll_{j,B}(1+Z)^{C_{j,B}} \begin{cases} \rho^{-\theta/2}(1+|\log\rho|),&0<\rho\leq1,\\ \rho^{-B},&\rho\geq1. \end{cases} \] The same bounds hold for $\int_\R(1+|\zeta|)^j\sup_t|\widehat F_{t,\rho}(\zeta)|\,d\zeta$, after increasing $C_{j,B}$. \end{lemma} \begin{proof} For small arguments, the Bessel expansions give powers $v^{\pm it}$ and a logarithm at $t=0$. The same bounds hold after applying $(v\,d/dv)^j$. For $v\geq1$, the change of variables $z=4\pi\sqrt{vu}$ identifies the transform with the Bessel transforms in~\cite[Lemma~3.4]{PascadiLargeSieve}, applied to \[ \phi_v(z)=\frac{z^2}{8\pi^2v}F\left(\frac{z^2}{16\pi^2v}\right). \] Their bounds give an initial polynomial estimate. Repeatedly move the Bessel differential operator onto $F$; each integration by parts contributes $v^{-1}$, with costs controlled by the derivatives of $F$. This proves arbitrary decay, also after applying $(v\,d/dv)^j$. Integration by parts in the Fourier transform gives \[ |\widehat F_{t,\rho}(\zeta)|\ll_j(1+|\zeta|)^{-j-2} \sum_{r=0}^{j+2}\|F_{t,\rho}^{(r)}\|_\infty. \] Multiplying by $(1+|\zeta|)^j$ and integrating proves the second assertion, after increasing $C_{j,B}$. Fourier inversion writes $F_{t,\rho}(u)$ as an integral of $e(\zeta u)$ with coefficient $\widehat F_{t,\rho}(\zeta)$. Minkowski's inequality and the uniform bound on this coefficient then allow us to apply the spectral moment to sums whose coefficients are independent of the form. \end{proof} \begin{lemma}\label{lem: improved short fourth moment} Let $Q,U\geq1$, $\zeta\in\R$, and let $\psi$ be a fixed smooth function supported in a compact subinterval of $(0,\infty)$. Put \[ P_f(U,\zeta):=\sum_n\psi(n/U)e(\zeta n/U)\rho_{f,\infty}(n), \qquad W:=\max\left\{1,\frac{Q^2}{U^3}\right\}. \] Then \begin{equation}\label{eqn: improved short fourth moment} \sum_{q\asymp Q}\sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{W^{\theta_f}|P_f(U,\zeta)|^4}{\cosh^2(\pi\kappa_f)} \ll_{\eta,\psi}(QU)^\eta(1+|\zeta|)^{C_0}U^2. \end{equation} The cutoff and the phase in $P_f$ are the same for every level. \end{lemma} \begin{proof} First take primitive exceptional forms at levels $s\asymp S$ and write \[ H_g(J,\zeta)=\sum_n\psi(n/J)e(\zeta n/J)\lambda_g(n). \] Hecke multiplication, Minkowski,~\eqref{eqn: improved first coefficient}, and~\eqref{eqn: improved Watt input} give \begin{align*} \left(\sum_{s\asymp S}\sum_g X^{\theta_g}|H_g(J,\zeta)|^4\right)^{1/2} &\ll_\eta S^{1/2+\eta}(1+|\zeta|)^{C_0} \sum_{d\ll J}\left(\left(S+\frac{J^2}{d^2}\right)\frac{J^2}{d^2}\right)^{1/2}\\ &\ll_\eta(SJ)^\eta(1+|\zeta|)^{C_0} \left(SJ+\sqrt S\,J^2\right) \end{align*} for $X\ll\max(1,S^2/J^3)$. Indeed, $H_g^2=\sum_{(d,s)=1}\sum_{u,v}\psi(du/J)\psi(dv/J)e(\zeta d(u+v)/J)\lambda_g(uv)$. For each fixed $d$, the restriction $(d,s)=1$ is dropped only after forming a nonnegative moment; Watt's estimate applies for $X\ll S^2d^3/J^3$. If $S^2d^3/J^3<1$, Lemma~\ref{lem: improved ordinary fourth moment} gives the bound at weight $1$. For $J\leq\sqrt S$ the squared bound is $O((SJ)^\eta(1+|\zeta|)^{C_0}S^2J^2)$. For $J\geq\sqrt S$, use~\eqref{eqn: improved Voronoi} with $c=1$. The dual indices $n\asymp\rho S/J$ have length $J_\rho=\rho S/J$ and prefactor $J/\sqrt s$. Lemma~\ref{lem: improved Hankel} separates both the spectral parameter and $s/S$ before the moment is applied. The preceding estimate applies to these indices for \[ X\ll\max\left\{1,\frac{S^2}{J_\rho^3}\right\} =\max\left\{1,\frac{J^3}{S}\rho^{-3}\right\}. \] At weight $J^3/S$, the fourth moment of this dual sum, before including the transform bound from Lemma~\ref{lem: improved Hankel}, is therefore at most \[ \frac{J^4}{S^2} \left(S^2J_\rho^2+SJ_\rho^4\right)\max(1,\rho^3)^\theta \leq S^2J^2(\rho^2+\rho^4)\max(1,\rho^3)^\theta. \] For $\rho\leq1$, multiplying the fourth root of this bound by the factor $\rho^{-\theta/2}(1+|\log\rho|)$ from Lemma~\ref{lem: improved Hankel} gives $O((S^2J^2)^{1/4}\rho^{(1-\theta)/2}(1+|\log\rho|))$; for $\rho\geq1$, that lemma gives the factor $O_B(\rho^{-B})$ for any fixed $B>0$. Minkowski thus gives, at every $J\geq1/2$, \begin{equation}\label{eqn: improved primitive smooth fourth} \sum_{s\asymp S}\sum_g X^{\theta_g}|H_g(J,\zeta)|^4 \ll_\eta(SJ)^\eta(1+|\zeta|)^{C_0}S^2J^2 \max\left\{1,\frac{X}{\max(1,S^2/J^3,J^3/S)}\right\}^{\theta}. \end{equation} To include oldforms, put $q=sm$, fix $\delta\mid m$, and apply~\eqref{eqn: improved oldform expansion} and H\"older. With $S=Q/m$ and $J=U/\delta$, the size in~\eqref{eqn: improved primitive smooth fourth} becomes \[ \frac{\delta^2}{Q^2}S^2J^2=\frac{U^2}{m^2}. \] The denominator $\max(1,S^2/J^3,J^3/S)$ in~\eqref{eqn: improved primitive smooth fourth} is at least $\max(1,Q^2\delta^3/(m^2U^3))$, so taking $X=W$ costs at most $m^{2\theta}$. The divisor sums and oldspace multiplicities give \[ \ll_\eta(QU)^\eta(1+|\zeta|)^{C_0}U^2 \sum_{m\geq1}\frac{\tau(m)^{C_1}}{m^{2-2\theta}} \ll_\eta(QU)^\eta(1+|\zeta|)^{C_0}U^2. \] The series converges since $2-2\theta>1$; intermediate divisor exponents are chosen smaller than the final $\eta$. If $J=U/\delta<1$, nonempty support forces $J\gg1$, so replacing $J$ by $1$ changes the bound only by a constant; empty supports contribute nothing. \end{proof} \begin{lemma}\label{lem: improved rational twist} Fix an integer $B\geq1$. Suppose $q=Dq_0$, where $D,d,q_0\in\Z_{>0}$, $q_0$ is squarefree, $(q_0,d)=1$, $D\mid d^\infty$, and $D\leq d^B$. Let $c_0\mid q_0$, $c_1\mid d^B$, $a_0,a_1\in\Z$, and $\alpha=a_0/c_0+a_1/c_1$. For a fixed smooth compactly supported $\psi$ and $V\geq1/2$, \begin{equation}\label{eqn: improved rational twist} \begin{aligned} &\left(\sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{1}{\cosh^2(\pi\kappa_f)} \left|\sum_n\psi(n/V)e(\alpha n)\rho_{f,\mathfrak a}(n)\right|^4\right)^{1/4}\\ &\qquad\ll_{\eta,B,\psi}(qV)^\eta d^{C_B} \left(\frac{\sqrt V}{q^{1/4}}+\sqrt{c_0}\right), \quad\mathfrak a=\infty,0. \end{aligned} \end{equation} Replacing $\psi(u)$ by $\psi(u)e(\zeta u)$ multiplies the right-hand side by at most $(1+|\zeta|)^{C_B}$, after increasing $C_B$. \end{lemma} \begin{proof} Reduce the two rational fractions first; replacing $c_0,c_1$ by divisors only improves the bound. Expand an oldvector generated by a primitive form $g$ of level $s\mid q$ using~\eqref{eqn: improved oldform expansion}. After dividing by $\sqrt{\cosh(\pi\kappa_g)}$, a term indexed by $\delta\mid q/s$ has coefficient $O_\eta(q^{-1/2+\eta}\sqrt\delta)$ and length $J=V/\delta$. Let $s_0,\delta_0$ denote the parts of $s,\delta$ coprime to $d$. Then $s_0\delta_0\mid q_0$ and $(s_0,\delta_0)=1$. Let $c'$ be the reduced additive denominator away from $d$. Then \[ c'\mid\frac{c_0}{(c_0,\delta_0)}. \] The remaining denominator, say $b$, divides a fixed power of $d$. For clarity, remove this latter denominator exactly. In a primitive polynomial $\sum_n F(n/J)e(An/c')e(un/b)\lambda_g(n)$, write $h=(n,b)$, $n=hv$, and $m=b/h$. The inverse Hecke relation gives \[ \lambda_g(hv)=\sum_{\substack{j\mid(h,v)\\(j,s)=1}} \mu(j)\lambda_g(h/j)\lambda_g(v/j). \] Put $v=jw$. Nonempty terms have $(j,m)=(w,m)=1$, and \begin{align*} \mathbf1_{(w,m)=1}e(ujw/m) &=\sum_{\chi\pmod m}\gamma_\chi(uj)\chi(w),\\ \gamma_\chi(v)&=\frac1{\phi(m)} \sum_{z\bmod m}^{*}e(vz/m)\overline{\chi(z)}, \qquad \sum_\chi|\gamma_\chi(v)|\leq\sqrt{\phi(m)}. \end{align*} Let $g_\chi$ be the primitive twist by the primitive character inducing $\chi$. Its level $s'$ divides $sm^2$, its spectral parameter is unchanged, and its level away from $d$ is $s_0$. The conductor bound follows from the principal-series conductor formula and~\cite[Proposition~4]{RamakrishnanYang} for the essentially discrete-series cases. Remove the remaining unit restriction by \[ \mathbf1_{(w,m)=1}\lambda_{g_\chi}(w) =\sum_{\ell\mid\operatorname{rad}(m)^2} \mu_{g_\chi}(\ell)\lambda_{g_\chi}(w/\ell), \] where $\mu_f$ is multiplicative, $\mu_f(p)=-\lambda_f(p)$, and $\mu_f(p^2)=\omega_f(p)$ for $p\nmid\operatorname{level}(f)$ and $0$ otherwise; here $\omega_f$ is the nebentypus of $f$. These coefficients come from the inverse Euler factors of $f$. The three identities express the original polynomial as a linear combination of \[ \sum_zF(w_0z/J)e(Aw_0z/c')\lambda_{g_\chi}(z), \qquad w_0=hj\ell\leq d^{O_B(1)}. \] The coefficients have total absolute value at most $d^{C_B}$: use the displayed bound for $\sum_\chi|\gamma_\chi(uj)|$ and the elementary estimate $|\lambda_f(n)|\leq\tau(n)n^{1/2}$, also at ramified indices, since all remaining indices are bounded by $d^{O_B(1)}$. Twisting is injective for each fixed character. Partitioning by $s'$ and its nebentypus creates only $d^{O_B(1)}$ families. For each fixed $h,j,\chi,\ell$, the polynomial coefficients $F(w_0z/J)e(Aw_0z/c')$ are common within each family. The scalar factors $\lambda_g(h/j)\mu_{g_\chi}(\ell)$ may depend on the form; we bound them uniformly by $d^{O_B(1)}$ before applying the moment estimate. For each such family the length is $J'=V/(\delta w_0)$. The nebentypus conductor is supported on $d$, whereas $(c',d)=1$ and $(c',s')$ is an exact divisor of $s'$. Apply~\eqref{eqn: improved Voronoi} with $\mathfrak C=s'c'^2/(c',s')$. The essential arithmetic bound is \begin{equation}\label{eqn: improved rational conductor} \frac{\delta\mathfrak C}{q}\ll_B d^{C_B}c_0. \end{equation} Indeed, write $q_0=s_0\delta_0d_0$ with pairwise coprime factors and put $v_0=c'/(c',s_0)$. Then $v_0\mid d_0$ and \[ \frac{\delta_0s_0c'^2}{q_0(c',s_0)} =c'\frac{v_0}{d_0}\leq c'\leq c_0. \] The prime powers supported on $d$ have polynomial size, proving~\eqref{eqn: improved rational conductor}. For dual indices $n\asymp\mathfrak C/J'$,~\eqref{eqn: improved primitive fourth moment}, including the oldform coefficient and Voronoi prefactor, gives \begin{align*} q^{-1/2}\sqrt\delta\,\frac{J'}{\sqrt{\mathfrak C}} \left(s'^{1/4}\sqrt{\frac{\mathfrak C}{J'}} +\frac{\mathfrak C}{J'}\right) &=q^{-1/2}\sqrt\delta \left(s'^{1/4}\sqrt{J'}+\sqrt{\mathfrak C}\right)\\ &\ll_B d^{C_B}\left(\frac{\sqrt V}{q^{1/4}}+\sqrt{c_0}\right). \end{align*} Here $s'\ll qd^{O_B(1)}$ and $\delta J'\leq V$. For dual indices $n\asymp\rho\mathfrak C/J'$, the two terms acquire factors $\rho^{1/2}$ and $\rho$, respectively. The transform bound in Lemma~\ref{lem: improved Hankel} makes the resulting bounds summable over dyadic $\rho$. Summing the finite decompositions proves the estimate. \end{proof} \begin{lemma}\label{lem: improved long fourth moment} Let $d,e,b,m\in\Z_{>0}$ and let $n,n'\asymp N$ be squarefree. Put \[ g=(n,n'),\qquad n=gs,\quad n'=gs',\qquad S=N/g,\qquad q=dgs s'. \] Thus $g,s,s'$ are squarefree and pairwise coprime. Suppose $1/2\leq H\leq N$, $Y\geq1/2$, $Y\ll HS$, and $(e,q)=1$. For fixed smooth functions $\Phi,\psi_1,\psi_2$ supported away from zero and phases $\omega_1,\omega_2\in\R/\Z$, define \[ c_v=\sum_{s'h-sh'=v}\psi_1(h/H)\psi_2(h'/H)e(h\omega_1+h'\omega_2). \] Suppose $\omega_1=A/m+\xi/H$, with $A\in\Z$ and $\xi\in\R$. Let $\mathfrak d$ be a positive integer divisible by $d,b,m$, and assume $d,b,m,1+|\xi|\leq\mathfrak d^{B_0}$ for a fixed $B_0$. Then \begin{equation}\label{eqn: improved long fourth moment} \sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{1}{\cosh^2(\pi\kappa_f)} \left|\sum_v\Phi(ev/Y)c_{ev}\rho_{f,0}(bv)\right|^4 \ll_\eta\mathfrak d^{C_0}(qYe)^\eta g^4 \left(\frac{Y^2}{q}+\min\left\{\frac{Y^4}{q^2},q\right\}\right). \end{equation} There is no fixed positive power of $e$ in this bound. Replacing $\Phi(u)$ by $\Phi(u)e(\zeta u)$ multiplies the right-hand side by at most $(1+|\zeta|)^{C_0}$, after increasing $C_0$. \end{lemma} \begin{proof} The solutions of $s'h-sh'=ev$ form one arithmetic progression, so $|c_{ev}|\ll1+H/S\ll g$. Insert zeros at the other indices $bv$ and apply~\eqref{eqn: improved ordinary fourth moment}; this gives the bound with $Y^2/q+Y^4/q^2$. For the other bound, solving for $h$ and applying Poisson summation to $h'\equiv-e\overline s\,v\pmod{s'}$ gives the exact identity \begin{align} c_{ev}={}&\frac H{s'}e(ev\omega_1/s') \sum_{k\in\Z}e(-ke\overline s\,v/s')\notag\\ &\quad\times\int_\R \psi_1\left(\frac{ev+sHu}{s'H}\right)\psi_2(u) e\left(H\left(\omega_2+\frac{s}{s'}\omega_1-\frac{k}{s'}\right)u\right)\,du. \label{eqn: improved additive Poisson} \end{align} After multiplication by $\Phi(ev/Y)$, normalised derivatives in $ev/Y$ cost only fixed powers of $1+|\xi|$, since $Y/(Hs')\ll1$. Integration by parts in $u$ shows that the sum over $k$ of the supremum norms of any fixed number of derivatives with respect to $ev/Y$ is $O((1+|\xi|)^{C_0}(1+S/H))$, uniformly in the centre $s\omega_1+s'\omega_2$ of the $k$-range. Put $w=bv$ and insert $\mathbf1_{b\mid w}=b^{-1}\sum_{\nu\bmod b}e(\nu w/b)$. The rational part of each phase has denominator dividing $bms'$. Its part coprime to $\mathfrak d$ divides the corresponding part of $s'$ and is at most $O(S)$. The remaining smooth phase is $e(\xi ew/(bHs'))$, whose normalised frequency is $\xi Y/(Hs')$; in particular it has no loss in $e$. The part of $q$ supported on $\mathfrak d$ is at most $d\operatorname{rad}(\mathfrak d)$, because $gss'$ is squarefree. Lemma~\ref{lem: improved rational twist} therefore applies, with polynomial cost in $\mathfrak d$. Minkowski and~\eqref{eqn: improved additive Poisson} give \[ \begin{aligned} &\left(\sum_f\frac{1}{\cosh^2(\pi\kappa_f)}\left|\sum_v\Phi(ev/Y)c_{ev}\rho_{f,0}(bv)\right|^4\right)^{1/4}\\ &\qquad\ll_\eta\mathfrak d^{C_0}(qYe)^\eta \frac HS\left(1+\frac SH\right) \left(\frac{\sqrt Y}{q^{1/4}}+\sqrt S\right). \end{aligned} \] Now $1+H/S\ll g$, $Y\ll HS\leq gS^2\ll q$, and $\sqrt S\ll q^{1/4}$, so the last display is $O_\eta(\mathfrak d^{C_0}(qYe)^\eta gq^{1/4})$. Combining its fourth power with the first bound proves~\eqref{eqn: improved long fourth moment}. \end{proof} \subsection{Kloosterman sums and the exponential convolution} For integers $m,n$ and $c\geq1$, put \[ S(m,n;c):=\sum_{\substack{z\bmod c\\(z,c)=1}} e\left(\frac{mz+n\overline z}{c}\right). \] For $a\in\Z\setminus\{0\}$, $K,L>0$, and finitely supported coefficients $(u_{q,v})$ with $q,v\in\Z_{>0}$ and $(q,a)=1$, put \[ \mathcal T=\sum_{q,v}u_{q,v} \sum_{\substack{k,\ell\geq1\\(k,q\ell)=1}} W_1(\ell/L)W_2(k/K)e\left(\frac{av\overline{q\ell}}k\right), \] where $W_1,W_2$ are smooth functions supported in fixed compact subintervals of $(0,\infty)$. Poisson summation in $\ell$ gives the identity \begin{equation}\label{eqn: improved Poisson completion} \mathcal T=L\sum_{q,v}u_{q,v} \sum_{(k,q)=1}\frac{W_2(k/K)}k \sum_{m\in\Z}\widehat W_1(mL/k)S(m\overline q,av;k). \end{equation} Fix a smooth function $\psi$ with compact support in $(0,\infty)$. For $U>0$ and $\sigma\in\{-1,1\}$, define the smooth nonzero-frequency block \[ \mathcal T_\sigma(U):=L\sum_{q,v}u_{q,v} \sum_{(k,q)=1}\frac{W_2(k/K)}k \sum_{m\geq1}\psi(m/U)\widehat W_1(\sigma mL/k) S(\sigma m\overline q,av;k). \] Let $\mathcal T_\sigma^{\mathrm{exc}}(U)$ denote its exceptional Maass contribution in Kuznetsov's formula at the cusps $\infty$ and $0$. Write \[ P_f(U,\zeta):=\sum_{m\geq1}\psi(m/U)e(\zeta m/U)\rho_{f,\infty}(m). \] The coefficients of this polynomial are independent of the level $q$ and the form $f$. \begin{lemma}[Spectral estimate for a completed block]\label{lemma: improved Poisson Kuznetsov} With the notation above, let $x\geq2$, $0<\eta<1$, and $K,L,Q,Y,U\geq1/2$. Suppose $K,L,Q,Y,|a|$ are bounded by fixed powers of $x$, that $(u_{q,v})$ is supported on $q\asymp Q$, $v\asymp Y$, and that $W_1,W_2$ have complexity at most $x^\eta$. Assume \[ U\ll\frac KLx^{2\eta},\qquad X:=\frac{K\sqrt Q}{\sqrt{|a|UY}}\geq x^{-C\eta}, \] where $C>0$ is fixed. For every fixed $B>0$ and sufficiently large fixed $J$, there is a constant $C_1>0$ such that \begin{align}\label{eqn: improved Kuznetsov majorant} |\mathcal T_\sigma^{\mathrm{exc}}(U)| \ll{}&Lx^{C_1\eta}\int_{\R^2}\frac{d\zeta\,d\xi}{(1+|\zeta|+|\xi|)^J} \sum_{q\asymp Q}\sqrt q\sum_{f\in\mathcal B_q^{\mathrm{exc}}} \frac{\max(1,X)^{\theta_f}}{\cosh(\pi\kappa_f)}|P_f(U,\zeta)|\notag\\ &\qquad\times\left|\sum_vu_{q,v}e(\xi v/Y)\rho_{f,0}(|a|v)\right| +x^{-B}\sum_{q,v}|u_{q,v}|. \end{align} The constants may depend on $B,J,C$, the fixed polynomial bounds, supports and smoothness orders; the implicit constant may also depend on $\eta$. \end{lemma} \begin{proof} With the scaling matrices $\sigma_\infty,\sigma_0$ chosen above, the Kloosterman sum $S_{\infty,0}$ between these cusps satisfies \[ S(\sigma m\overline q,av;k)=S_{\infty,0}(\sigma m,av;k\sqrt q) \qquad((k,q)=1). \] Apply the geometric-test-function Kuznetsov formula, as in~\cite[Proposition~3.5 and the proof of Corollary~5.11]{PascadiLargeSieve}. For $a>0$ and $\sigma=1$, keep $\psi(m/U)$ outside the transform and use the test function \[ \phi_{m,v,q}(z)=W_2\left(\frac{4\pi\sqrt{amv}}{K\sqrt q\,z}\right) \widehat W_1\left(\frac{Lm\sqrt q\,z}{4\pi\sqrt{amv}}\right), \qquad z\asymp X^{-1}. \] The factor $L/k=L\sqrt q/(k\sqrt q)$ accounts for $L\sqrt q$ in~\eqref{eqn: improved Kuznetsov majorant}. The Bessel-transform bounds in~\cite[Lemma~3.4]{PascadiLargeSieve} give $x^{O(\eta)}\max(1,X)^{\theta_f}$ on the exceptional spectrum. Extend the remaining dependence on $m/U$ smoothly from the support of $\psi$, and Fourier-separate it from $v/Y$ before taking absolute values. The resulting kernel is bounded by $x^{O(\eta)}(1+|\zeta|+|\xi|)^{-J}$, uniformly in $q$ and $f$. Keeping $\psi(m/U)$ outside this separation gives exactly $P_f(U,\zeta)$, with coefficients common to all levels. The other signs of $\sigma$ and $a$ follow from the opposite-sign Kuznetsov formula and reflection. There is no diagonal term in this geometric-test-function formula. \end{proof} \begin{lemma}\label{lemma: improved spectral envelope} For $Q\geq1$, $1\leq Y\leq Q$ and $W\geq1$, put \[ B(Y;Q)=\frac YQ+\min\left\{\frac{Y^2}{Q^{3/2}},1\right\}. \] Then \begin{equation}\label{eqn: improved spectral envelope} B(Y;Q)\max\left\{1,\frac Q{Y\sqrt W}\right\}^{\theta} \ll\max\{1,Q^{\theta/4}W^{-\theta/2}\}. \end{equation} \end{lemma} \begin{proof} The contribution of $Y/Q$ is at most $1$, since, when $Q>Y\sqrt W$, \[ \frac YQ\left(\frac Q{Y\sqrt W}\right)^\theta =\left(\frac YQ\right)^{1-\theta}W^{-\theta/2}\leq1. \] The other contribution increases up to $Y=Q^{3/4}$ and decreases thereafter. Its value at that point is $\max\{1,Q^{\theta/4}W^{-\theta/2}\}$. \end{proof} \begin{proposition}\label{propn: improved squarefree Kloosterman} Let $x\geq2$, $\eta,\delta>0$, and $K,L,N,H\gg1$ be of at most fixed polynomial size in $x$. Put \begin{equation}\label{eqn: improved Kloosterman distortions} \Lambda=\max\left\{1,\frac K{NL},\frac HN, \frac{N^2}{\sqrt x},\frac{K^2}{xN},\frac{xH}{K^2}\right\}. \end{equation} Let $a,t\in\Z\setminus\{0\}$, $d\in\Z_{>0}$, $(a,d)=1$, and $|a|\leq x^{1+\delta}$. Suppose \[ \omega_1=A/m+\xi/H,\qquad A\in\Z,\quad m\in\Z_{>0}, \] while $\omega_2\in\R/\Z$ is arbitrary. Let $\mathfrak d$ be a positive integer divisible by $d,|t|,m$, with $1+|\xi|\leq\mathfrak d^{B_0}$. Let $|\alpha_{n,n'}|\leq1$ be supported on squarefree $n,n'\asymp N$ satisfying $(nn',at)=1$. For fixed smooth weights of complexity at most $x^\eta$, define \begin{align*} \mathcal K={}&\sum_{n,n'}\alpha_{n,n'} \sum_{\substack{h,h'\in\Z\\n'h\neq nh'}} \psi_1(h/H)\psi_2(h'/H)e(h\omega_1+h'\omega_2)\\ &\qquad\times\sum_{(k,dnn'\ell)=1} W_1(\ell/L)W_2(k/K) e\left(\frac{at(n'h-nh')\overline{dnn'\ell}}k\right). \end{align*} The $h,h'$ supports are bounded away from zero. There is an absolute constant $C_0$, depending only on $B_0$, the fixed supports and the required smoothness orders, such that \begin{equation}\label{eqn: improved squarefree Kloosterman} |\mathcal K|^2\ll(\mathfrak d\Lambda)^{C_0} x^{C_0(\eta+\delta)}L^2N^7\sqrt x. \end{equation} Additional conditions on $n,n'$ may be included in $\alpha_{n,n'}$. \end{proposition} \begin{proof} Write \[ g=(n,n'),\qquad n=gs,\quad n'=gs',\qquad (g,ss')=(s,s')=1. \] Squarefreeness gives the coprimalities in this display. Cancelling $g$ in the reciprocal phase gives level $q=dgss'$ and additive index $at(s'h-sh')$. This preserves the condition on $k$, since $nn'$ and $gss'$ have the same prime divisors. First suppose $\Lambda=1$, and temporarily suppress factors bounded by fixed powers of $\mathfrak d$. Group $g\asymp G$ and $|s'h-sh'|\asymp Y$, and put \[ Q=N^2/G,\qquad 1\leq Y\ll HN/G\leq Q. \] For a fixed $q$, there are at most $\tau_3(q/d)$ triples $(g,s,s')$. H\"older over these triples costs $x^{O(\eta)}$, and the coefficients $\alpha_{n,n'}$ are included in $u_{q,v}$ and hence do not alter $P_f(U,\zeta)$. Rapid decay of $\widehat W_1$ restricts the nonzero completion frequencies, with negligible error, to \[ |m|\ll\frac KLx^{2\eta}\ll Nx^{2\eta}. \] Split the signs of $s'h-sh'$ and the nonzero completion frequency, absorbing the sign of $t(s'h-sh')$ into $a$. Apply Lemma~\ref{lemma: improved Poisson Kuznetsov}, with $|t|Y$ in place of $Y$, to each block $\mathcal T_\sigma(U)$ with $U\ll Nx^{2\eta}$. We first estimate the exceptional contribution. In~\eqref{eqn: inverse Hecke relation}, a divisor $e\mid|a|$ gives \[ e_0=(e,t),\qquad e_1=e/e_0,\qquad c_{e_1v}\rho_{f,0}((t/e_0)v),\qquad (e_1,q)=1, \] where $c_v$ is the coefficient defined in Lemma~\ref{lem: improved long fourth moment}. Write the resulting polynomial as \[ L_{f,e}(Y):=\sum_v\Phi(e_1v/Y)c_{e_1v} \rho_{f,0}((|t|/e_0)v), \] with its smooth cutoff and normalised Fourier twist included in $\Phi$. Lemma~\ref{lem: improved long fourth moment} applies with $e\leftarrow e_1$ and $b\leftarrow|t|/e_0$. All divisor sums and Fourier integrations below cost $x^{O(\eta)}\mathfrak d^{O(1)}$. We write $P_f(U)$ for $P_f(U,\zeta)$ and suppress the dependence on the Fourier variables in $L_{f,e}(Y)$. Put \[ W=\max\{1,Q^2/U^3\},\qquad X^2=\frac{K^2Q}{|a t|UY},\qquad V=\max\{1,X^2/\sqrt W\}. \] Let $\mathcal K_G(U,Y)$ denote the exceptional contribution with $g\asymp G$, $|m|\asymp U$ and $|s'h-sh'|\asymp Y$. Then $V^{\theta_f/2}W^{\theta_f/4}\geq\max(1,X)^{\theta_f}$. H\"older with exponents $2,4,4$, applied over both levels and forms in~\eqref{eqn: improved Kuznetsov majorant}, gives \begin{align*} |\mathcal K_G(U,Y)|^2 \ll{}&x^{O(\eta)}L^2Q \left(\sum_{q\asymp Q}\sum_fV^{\theta_f}|\lambda_f(|a|/e)|^2\right)\\ &\quad\times\left(\sum_{q\asymp Q}\sum_f\frac{W^{\theta_f}|P_f(U)|^4}{\cosh^2(\pi\kappa_f)}\right)^{1/2} \left(\sum_{q\asymp Q}\sum_f\frac{|L_{f,e}(Y)|^4}{\cosh^2(\pi\kappa_f)}\right)^{1/2}. \end{align*} Lemmas~\ref{lem: improved Hecke second moment}, \ref{lem: improved short fourth moment}, and~\ref{lem: improved long fourth moment} bound the three factors, respectively, by \[ Q(Q+\sqrt{|a|})\mathcal F^\theta,\qquad U,\qquad G^2Q B(Y;Q), \] up to $x^{O(\eta)}$, where \[ \mathcal F=\max\left\{1, \frac{K^2Q}{UY\max\{|a|,Q^2\}\sqrt W}\right\}. \] Indeed, the exact denominator for $|a|/e$ is $\max\{|a|,eQ^2\}$, and $|t|\geq1$ can only decrease this factor. Consequently, \begin{equation}\label{eqn: improved Holder Kloosterman} |\mathcal K_G(U,Y)|^2 \ll x^{O(\eta)}L^2G^2UQ^3(Q+\sqrt{|a|})B(Y;Q)\mathcal F^\theta. \end{equation} The residue-dependent factor in~\eqref{eqn: improved Holder Kloosterman} is nondecreasing in $|a|$, up to an absolute constant: below $Q^2$ its denominator is constant; above $Q^2$, use $Q+\sqrt{|a|}\asymp\sqrt{|a|}$ and $\theta<1/2$. We may therefore replace $|a|$ by $x$, at cost $x^{O(\delta)}$. Since $Q\leq\sqrt x$ and $K^2\leq xN$, the $U$-dependent part is at most \[ U\max\left\{1,\frac{NQ}{UY\sqrt{\max\{1,Q^2/U^3\}}}\right\}^{\theta} =\begin{cases} U\max\{1,N\sqrt U/Y\}^{\theta},&U\leq Q^{2/3},\\ U\max\{1,NQ/(UY)\}^{\theta},&U\geq Q^{2/3}. \end{cases} \] Both expressions are nondecreasing. Substituting $U=N$, and absorbing the Fourier truncation into $x^{O(\eta)}$, gives, with $W_G=\max\{1,N/G^2\}$, \[ |\mathcal K_G(Y)|^2 \ll x^{O(\eta+\delta)}\frac{L^2N^7\sqrt x}{G} B(Y;Q)\max\left\{1,\frac Q{Y\sqrt{W_G}}\right\}^{\theta}. \] Lemma~\ref{lemma: improved spectral envelope} applies, and \[ Q^{\theta/4}W_G^{-\theta/2} =G^{3\theta/4}\quad(G\leq\sqrt N),\qquad Q^{\theta/4}\leq G^{3\theta/4}\quad(G\geq\sqrt N). \] Thus the exceptional contribution is \[ |\mathcal K_{G,\rm exc}|^2 \ll x^{O(\eta+\delta)}L^2N^7\sqrt x\,G^{-1+3\theta/4}. \] The dyadic $G$-sum is convergent, even after taking square roots, since $-1+3\theta/4=-107/128<0$. The regular contributions satisfy the required bound by Proposition~\ref{propn: regular Kloosterman} in Appendix~\ref{sec: regular spectrum}. The zero Poisson frequency, still with nonzero additive index, is bounded using \[ |S(0,atv;k)|\leq(atv,k),\qquad \sum_{k\asymp K}\frac{(atv,k)}k\ll\tau(|atv|). \] Its size is $O(x^{O(\eta)}LN^2H^2)$, whose square is at most $x^{O(\eta)}L^2N^8$, also admissible. Finally restore $\mathfrak d$ and $\Lambda$. The levels satisfy $q\asymp dQ$; replacing $dQ$ by $Q$ in the displayed powers costs $\mathfrak d^{O(1)}$, while $B(Y;dQ)\leq B(Y;Q)$. The weight available for the fourth moment of $P_f$ satisfies $\max\{1,(dQ)^2/U^3\}\geq\max\{1,Q^2/U^3\}$. The parameter $X$ in Lemma~\ref{lemma: improved Poisson Kuznetsov} satisfies \[ X^2\gg\frac{K^2N}{|a t|UH} \gg(\mathfrak d\Lambda)^{-O(1)}x^{-O(\eta+\delta)}, \] since $U\ll\Lambda Nx^{2\eta}$ and $xH/K^2\leq\Lambda$. Thus the Fourier integrals in~\eqref{eqn: improved Kuznetsov majorant} contribute at most $(\mathfrak d\Lambda)^{C_0}x^{C_0(\eta+\delta)}$, after increasing $C_0$. Similarly, $Q\leq\Lambda\sqrt x$, $K^2\leq\Lambda xN$, and $H\leq\Lambda N$ introduce fixed powers of $\Lambda$ in the preceding estimates. For $Q0$ and $\omega\in\R/\Z$, put \[ T_H(\omega):=\min_{m\in\Z_{>0}}\{m+2H\|m\omega\|\}, \qquad \|u\|:=\min_{j\in\Z}|u-j|. \] The following estimate transfers the Kloosterman bound to the exponential sum arising from dispersion. \begin{lemma}\label{lemma: improved exponential convolution} Let $x=MNL\geq2$, $0<\eta<1$, $a\in\Z\setminus\{0\}$, and $v,d_1,d_2\in\Z_{>0}$, with $(a,vd_1d_2)=1$. Suppose $1\ll M,K,N,L,H,R\ll x$, $K\geq1$, and \[ |a|\leq x^{1+\eta},\qquad vd_1d_2\leq x^\eta, \] \[ x^{-\eta}NL\leq K\leq x^\eta NL,\quad L\leq x^\eta M,\quad R\leq x^\eta K,\quad H\leq x^\eta R/M,\quad N^2\leq x^{1/2+\eta}. \] Let $(u_k)_{K1\implies u_k=0,\qquad (n\ell,avd_1)>1\implies\beta_n\lambda_\ell=0. \] Suppose also that $\beta_n=0$ unless $n$ is squarefree. Let $\Phi,\Psi$ be fixed smooth functions supported in compact subintervals of $(0,\infty)$, and let $\omega\in\R/\Z$ satisfy $T_H(\omega)\leq x^\eta$. Put \begin{align*} \mcR:={}&\sum_{\substack{r\sim R\\(r,avd_1d_2)=1}} \frac Mr\Psi(r/R) \sum_{\substack{k,n,\ell\\d_1k\equiv d_2n\ell\pmod r\\(d_1k,d_2n\ell)=1}} u_k\beta_n\lambda_\ell\\ &\qquad\times\sum_{h\in\Z}e(h\omega)\Phi(h/H) e\left(-\frac{ha\overline{vd_1d_2k}}r\right). \end{align*} There is an absolute constant $C_0>0$ such that \begin{equation}\label{eqn: improved exponential estimate} \frac{|\mcR|}{MK^2/R} \ll x^{C_0\eta} \left\{\left(\frac{N^2L^3}{xR}\right)^{1/2} +\max\left(1,\frac KR\right)^{C_0} \left(\frac{N\sqrt x}{L^2}\right)^{1/4}\right\}. \end{equation} The constant $C_0$ is independent of $\eta$; the implicit constant may depend on $\eta$ and the fixed smooth weights. \end{lemma} \begin{proof} We follow the complementary-divisor argument in~\cite[Proposition~1]{Drappeau}, with the smooth $h$-sum of~\cite[Lemma~6.2]{PascadiLvlDist}. Put \[ \nu=vd_1d_2,\qquad T=\max\left\{1,4\frac{\max(d_1K,d_2NL)}R\right\} \ll x^{2\eta}\max(1,K/R). \] Write $rt=d_2n\ell-d_1k$. The case $t=0$ would force $d_1k=d_2n\ell=1$, which is excluded by the supports. Reciprocity gives \[ -\frac{ah\overline{\nu k}}r \equiv\frac{aht\overline{\nu d_2n\ell}}k +\frac{ah\overline{rk}}\nu-\frac{ah}{\nu kr} \pmod1. \] We retain the last exponential. Put \[ D=1+\frac{|a|H}{\nu KR},\qquad \nu D=\nu+\frac{|a|H}{KR}\ll x^{3\eta}. \] For fixed $k,\ell,t$, write $r(u)=(d_2u\ell-d_1k)/t$ and $b(u)=aH/(\nu kr(u))$. Choose a fixed nonnegative smooth cutoff $\Phi_0$ equal to $1$ on $[1,4]$ and on the support of $\Phi$, and supported away from zero. Fourier inversion gives \[ \Phi(h/H)e\left(-\frac{ah}{\nu kr(n)}\right) =\Phi_0(h/H)\int_\R\widehat\Phi(s+b(n))e(sh/H)\,ds. \] On the interval $[u_0,u_1]$ defined by $N\leq u\leq2N$ and $R\leq r(u)\leq2R$, set \[ W_s(u)=\frac R{r(u)}\Psi(r(u)/R)\widehat\Phi(s+b(u)). \] Since $b(u)$ is monotone and $|b(u)|\ll D$, differentiation and rapid decay give \[ \int_\R\sup_{k,\ell,t} \left(|W_s(u_1)|+\int_{u_0}^{u_1}|W_s'(u)|\,du\right)ds\ll D. \] The supremum is over contributing triples. Partial summation in $n$, followed by the interval separation, character decomposition and Cauchy--Schwarz argument in~\cite[Section~3.5, (3.22)]{Drappeau}, now gives \begin{equation}\label{eqn: improved exponential reduction} |\mcR|\ll x^{C_1\eta}\frac MR\sqrt{KLT} \sup_{\substack{|s|\leq x^\eta D\\\rho\in\R/\Z,\ \nu\rho=0}} \mcB(s,\rho)^{1/2}+O(x^{-100}). \end{equation} Here, after applying the triangle inequality to the $t$-sum, we may take \begin{align*} \mcB(s,\rho):={}&\sum_{1\leq|t|\leq T} \sum_{\substack{n,n'\sim N\ \mathrm{squarefree}\\(nn',at)=1\\t\mid n-n'}} \Biggl|\sum_{h,h'\in\Z}\Phi_0(h/H)\Phi_0(h'/H) e(h\alpha_1+h'\alpha_2)\\ &\qquad\times\sum_{\substack{k,\ell\\(k,dnn'\ell)=1}} \Phi_0(k/K)\Phi_0(\ell/L) e\left(\frac{at(n'h-nh')\overline{dnn'\ell}}k\right)\Biggr|, \end{align*} where \[ d=\nu d_2,\qquad \alpha_1=-\alpha_2=\omega+\rho+s/H. \] The squarefree support is retained when the square is opened. Further restrictions on $n,n',t$ can be dropped after the displayed absolute value. The divisor and separation losses have total cost $x^{O(\eta)}\nu D\ll x^{C_1\eta}$. The Fourier tails are negligible. Moreover, \[ T_H(\alpha_1)\leq\nu(1+2|s|)T_H(\omega)\ll x^{5\eta}, \] so we can write $\alpha_1=A/m+\xi/H$ with $m,1+|\xi|\ll x^{5\eta}$ by choosing $m$ attaining the minimum in $T_H(\alpha_1)$ and $A$ nearest to $m\alpha_1$. The part of $\mcB$ with $n'h=nh'$ is at most \[ \mcB_0\ll x^\eta KLTNH. \] Indeed, if $g=(n,n')$, its nonzero solutions are $h=(n/g)j$, $h'=(n'/g)j$, numbering $O(Hg/N)$. There are no solutions when this quantity is too small, and $\sum_{n,n'\sim N}(n,n')\ll N^2\log(2N)$. For the remaining terms fix $t$, linearise the outer absolute values, and apply Proposition~\ref{propn: improved squarefree Kloosterman}. The support gives $(a,dnn')=1$, and the condition $t\mid n-n'$ is included in its coefficients. The quantity $\Lambda$ defined in~\eqref{eqn: improved Kloosterman distortions} satisfies $\Lambda\ll x^{C_2\eta}$, since \[ \frac K{NL}\leq x^\eta,\quad \frac HN\leq x^{4\eta},\quad \frac{N^2}{\sqrt x}\leq x^\eta,\quad \frac{K^2}{xN}\leq x^{3\eta},\quad \frac{xH}{K^2}\leq x^{3\eta}. \] The auxiliary integer $d|t|m\lceil1+|\xi|\rceil$ is at most $x^{C_3\eta}\max(1,K/R)$. Summing the resulting estimate over $t$ gives \[ \mcB_1\ll x^{C_4\eta}\max(1,K/R)^{C_4} TLN^{7/2}x^{1/4}. \] The fourth power of its contribution in~\eqref{eqn: improved exponential reduction}, divided by $(MK^2/R)^4$, is at most \[ x^{C_5\eta}\max(1,K/R)^{C_5} \frac{(KLT)^2T^2L^2N^7\sqrt x}{K^8} \ll x^{C_6\eta}\max(1,K/R)^{C_6} \frac{N\sqrt x}{L^2}. \] For the diagonal, the corresponding squared ratio is \[ x^{C_7\eta}\frac{KLT\cdot KLTNH}{K^4} \ll x^{C_8\eta}\frac{N^2L^3}{xR}. \] Here $K\asymp NL$ up to $x^\eta$, $T\ll x^{2\eta}\max(1,K/R)$ and $H\leq x^\eta R/M$; when $K1\implies\beta_n\gamma_\ell=0. \end{equation} Both assumptions will be removed below. We follow the dispersion argument in~\cite[Sections~3.1--3.6]{Drappeau} and~\cite[Proposition~6.3]{PascadiLvlDist}. Put \[ K=NL,\qquad u_k=\sum_{n\ell=k}\beta_n\gamma_\ell, \qquad f(m)=\Phi_0(m/M), \] where $\Phi_0\geq0$ is smooth, supported on $[1/2,3]$, and equal to $1$ on $[1,2]$. Thus $|u_k|\leq\tau(k)$ and $u_k=0$ when $(k,a_1)>1$. Define \begin{align*} \mcS_1&:=\sum_{\substack{r\sim R\\(r,a_1a_2)=1}} \sum_{(m,r)=1}f(m) \left|\sum_{k\equiv a_1\overline{a_2m}\pmod r}u_k\right|^2,\\ \mcS_2&:=\sum_{\substack{r\sim R\\(r,a_1a_2)=1}}\frac1{\phi(r)} \sum_{(m,r)=1}f(m) \sum_{\substack{(k_1,r)=1\\k_2\equiv a_1\overline{a_2m}\pmod r}} u_{k_1}\overline{u_{k_2}}\omega_D(mk_1\overline{a_1}a_2;r),\\ \mcS_3&:=\sum_{\substack{r\sim R\\(r,a_1a_2)=1}}\frac1{\phi(r)^2} \sum_{(m,r)=1}f(m) \left|\sum_{(k,r)=1}u_k\omega_D(mk\overline{a_1}a_2;r)\right|^2. \end{align*} Cauchy--Schwarz gives \begin{equation}\label{eqn: dispersion cauchy schwarz} \mcD(D)^2\ll MR(\mcS_1-2\Re\mcS_2+\mcS_3). \end{equation} The corresponding main terms are \begin{align*} X_1&:=\sum_{\substack{r\sim R\\(r,a_1a_2)=1}}\frac1r \sum_{\substack{00$, provided $|a_1|\leq x^{1+\delta_1}$, $|a_2|\leq x^{\delta_1}$ and $D\leq x^{s_1}$, with $\delta_1,s_1>0$ sufficiently small in terms of $\kappa$. These arguments are uniform in $a_1$: the argument for $\mcS_3$ is independent of its size, and the one for $\mcS_2$ allows $|a_1|\leq x^{1+\delta_1}$. The inequalities $K\leq x^{2/3-5\kappa}$ and $M=x/K$ leave a fixed power-saving margin in both estimates. For $\mcS_1$, retain coprimality to $a_1$ throughout Drappeau's decomposition~\cite[Section~3.4]{Drappeau}. In its notation, \[ v=(k_1,k_2),\quad d_1\mid v^\infty,\quad e_1\mid a_2^\infty,\quad e_2\mid a_2,\quad d_2=(k_2/(ve_2),v), \] Thus $(a_1,vd_1e_1e_2d_2)=1$. Under the substitution \[ (v,d_1,d_2)\longleftarrow \left(\frac{|a_2|v}{e_2d_2},e_1d_1,e_2d_2\right), \qquad (vd_1d_2)_{\rm new}=|a_2|vd_1e_1, \] the new parameters remain coprime to $a_1$; the sign of $a_2$ is absorbed into the residue parameter. Dividing the coefficient indices by these factors preserves~\eqref{eqn: coprime support for triple convolution} and squarefreeness of the $n$-variable. Use cutoffs $x^\eta$ in this decomposition, where $\eta>0$ is sufficiently small in terms of $\kappa$, and apply Lemma~\ref{lemma: improved exponential convolution} to the retained ranges, with a sufficiently large constant multiple of $\eta+\delta_1$ as its small parameter, as in~\cite[proof of Proposition~6.3]{PascadiLvlDist}. The new lengths differ from $K,N,L$ by factors between $x^{-C\eta-C\delta_1}$ and $x^{C\eta+C\delta_1}$. The conditions $N^2\leq x^{1/2-\kappa}$, $L/M\leq x^{-\kappa}$ and $R/K\leq x^{-\kappa}$ therefore imply the size hypotheses of that lemma. Its diagonal and nonzero additive frequencies give \[ \left(\frac{|\mcS_1-\widehat f(0)X_1|}{xK/R}\right)^4 \ll x^{C_1(\eta+\delta_1)} \left\{\left(\frac{N^2L^3}{xR}\right)^2 +\left(\frac KR\right)^{C_*}\frac{N\sqrt x}{L^2}\right\} +x^{-c_2\eta}. \] Here $C_*$ is fixed large enough to include the polynomial costs of the complementary divisors and the other auxiliary integers. The last term, with $c_2>0$, bounds the discarded gcd ranges, using $R/K\leq x^{-\kappa}$ and choosing the divisor losses sufficiently small after $\eta$. By~\eqref{eqn: improved convolution hypotheses}, choosing $\eta$ and then $\delta_1$ sufficiently small gives a power saving. Consequently \[ \mcS_i=\widehat f(0)X_i+O_\kappa(x^{1-c}K/R) \qquad(i=1,2,3) \] for some $c>0$. It remains to estimate the main terms. Parseval on the unit group gives \begin{align*} X_1-2\Re X_2+X_3 &\leq\frac1R\sum_{\substack{r\sim R\\(r,a_1a_2)=1}}\frac1{\phi(r)} \sum_{\substack{\chi\ \mathrm{primitive}\\\operatorname{cond}(\chi)>D\\\operatorname{cond}(\chi)\mid r}} \left|\sum_{(k,r)=1}u_k\chi(k)\right|^2\\ &\leq\frac1R\sum_{DD\geq1$, one has $s\geq2$, and therefore $d\mid r/s$ implies $d\leq R$. Partial summation and the multiplicative large sieve~\cite[Lemme~6]{Drappeau}, with $|u_k|\leq\tau(k)$, yield \begin{align*} X_1-2\Re X_2+X_3 &\ll\frac{(\log R)^2(\log K)^3}{R} \sum_{d\leq R}\frac{\tau(d)^3}{\phi(d)} \left(\frac K{dD}+R\right)\frac Kd\\ &\ll(\log x)^5\left(\frac{K^2}{RD}+K\right). \end{align*} Decrease $s_1$ so that $s_1\leq\kappa$. Then $RD\leq K$, and substitution in~\eqref{eqn: dispersion cauchy schwarz}, using $\widehat f(0)\ll M$ and $MK=x$, proves \begin{equation}\label{eqn: squarefree coprime improved convolution} \mcD(D)\ll_\kappa\frac{x(\log x)^3}{\sqrt D}+x^{1-c_1} \qquad(1\leq D\leq x^{s_1}) \end{equation} for some $01$ is $(v,k)=1$. Since $\phi(dk)\geq\phi(d)\phi(k)$, \[ T_2\ll x^{o(1)}\sum_{k\leq2R}\frac1k \sum_{d\leq D}\frac1d\sum_{\chi\pmod d}^{*} \left|\sum_{(v,k)=1}b_v\chi(v)\right|. \] For each fixed $k$, the coefficients are independent of $d$. Cauchy--Schwarz and the multiplicative large sieve give \begin{align*} &\sum_{EB\\b\text{ powerful}}}\frac1b\ll B^{-1/2},\qquad \sum_{b\text{ powerful}}b^{-1+c_1}<\infty. \end{equation} Take $\delta_1,s_1,c_1$ from~\eqref{eqn: squarefree coprime improved convolution} with $\kappa/2$ in place of $\kappa$, choose $0<\eta_0\ll\min\{\kappa,\delta_1,s_1\}$, and then choose $\delta_0,s_0>0$ sufficiently small in terms of $\eta_0$, with $s_0\leq\eta_0/8$. For the terms whose powerful part exceeds $x^{\eta_0}$, set \[ b_v:=\sum_{\substack{mn\ell=v\\m\sim M,\ n\sim N,\ \ell\sim L\\ \text{powerful part of }n>x^{\eta_0}}}\alpha_m\beta_n\gamma_\ell. \] Then $|b_v|\leq\tau_3(v)$ and \[ B=\sum_v|b_v|\ll x^{1-\eta_0/2+o(1)}. \] Equation~\eqref{eqn: sparse convolution contribution} bounds this contribution by \[ x^{1-\eta_0/2+o(1)}+Rx^{o(1)} +(D+\sqrt x)x^{1/2-\eta_0/4+o(1)}, \] which saves a power for $D\leq x^{s_0}$. For a retained powerful $b\leq x^{\eta_0}$, put \[ \beta^{(b)}_{n'}=\beta_{bn'}\mathbf1_{n'\text{ squarefree}}\mathbf1_{(n',b)=1}. \] The coprimality assumption gives $(b,a_1)=1$. Moduli sharing a prime factor with $b$ contribute zero to both terms of the kernel. On the other moduli the block is the squarefree problem with \[ x'=x/b,\qquad N'=N/b,\qquad a_1'=a_1,\qquad a_2'=a_2b, \] and $M,L,R$ unchanged. For sufficiently small $\eta_0,\delta_0,s_0$, the residue and cutoff conditions hold at $x'$, and the length conditions hold with $\kappa/2$. In particular, \[ \frac{(N')^2L^3}{x'R}=\frac1b\frac{N^2L^3}{xR},\qquad \left(\frac{N'L}{R}\right)^{C_*}\frac{N'\sqrt{x'}}{L^2} =b^{-C_*-3/2}\left(\frac{NL}{R}\right)^{C_*}\frac{N\sqrt x}{L^2}. \] Sum~\eqref{eqn: squarefree coprime improved convolution} over these powerful $b$. The convergent sums in~\eqref{eqn: powerful reciprocal bounds} give, for some $c_0>0$, \begin{equation}\label{eqn: coprime improved convolution} \mcD(D)\ll_\kappa\frac{x(\log x)^3}{\sqrt D}+x^{1-c_0} \qquad(1\leq D\leq x^{s_0}), \end{equation} uniformly for $|a_1|\leq x^{1+\delta_0}$, $|a_2|\leq x^{\delta_0}$, with no squarefree restriction. We finally remove~\eqref{eqn: coprime support for triple convolution} and restore the full cutoff range. Take $\delta_0,s_0,c_0$ from~\eqref{eqn: coprime improved convolution} with $\kappa/2$ in place of $\kappa$, and choose \[ 0<\delta\ll\eta\ll\min\{\kappa,\delta_0,s_0,c_0\}. \] If $D>x^{\eta/4}$, expanding the difference of the character sums gives \[ |\mcD(D)-\mcD(x^{\eta/4})| \leq\sum_{r\sim R}\frac1{\phi(r)} \sum_{\substack{\chi\pmod r\\x^{\eta/4}<\operatorname{cond}(\chi)\leq D}} \left|\sum_{m\sim M}\alpha_m\chi(m) \sum_{n\sim N}\beta_n\chi(n) \sum_{\ell\sim L}\gamma_\ell\chi(\ell)\right|. \] Bound the $\ell$-sum by $O(L)$ and write $r=dk$, where $d$ is the conductor. The inducing primitive character $\chi^*\pmod d$ satisfies \[ \chi(m)=\chi^*(m)\mathbf1_{(m,k)=1}. \] For each fixed $k$, the restricted coefficients are $1$-bounded and independent of $d$. Cauchy--Schwarz and the multiplicative large sieve bound a block $Ex^\eta$, the coefficients \[ b_v:=\sum_{\substack{mn\ell=v\\m\sim M,\ n\sim N,\ \ell\sim L\\(n\ell,a_1)>x^\eta}} \alpha_m\beta_n\gamma_\ell \] satisfy $|b_v|\leq\tau_3(v)$ and \[ B\ll x^{1+o(1)}\sum_{\substack{g\mid|a_1|\\g>x^\eta}}\frac1g \ll x^{1-\eta+o(1)}. \] By~\eqref{eqn: sparse convolution contribution}, their contribution is $O(x^{1-\eta/2+o(1)}+Rx^{o(1)})$, which saves a power. For the remaining terms put \[ d=(n,a_1),\qquad e=(\ell,a_1/d),\qquad g=de=(n\ell,a_1)\leq x^\eta, \] and write $n=dn'$, $\ell=e\ell'$. The coefficients \[ \beta^{(d)}_{n'}=\beta_{dn'}\mathbf1_{(n',a_1/d)=1},\qquad \gamma^{(d,e)}_{\ell'}=\gamma_{e\ell'}\mathbf1_{(\ell',a_1/g)=1} \] satisfy~\eqref{eqn: coprime support for triple convolution} with $a_1/g$. Division by $g$ modulo $r$ gives the parameters \[ x'=x/g,\qquad N'=N/d,\qquad L'=L/e,\qquad a_1'=a_1/g, \] with $M,R,a_2$ unchanged. The original moduli form a subset of those coprime to $a_1'a_2$. The length conditions hold with $\kappa/2$; the critical ratios are \[ \frac{(N')^2(L')^3}{x'R}=\frac1{de^2}\frac{N^2L^3}{xR},\qquad \left(\frac{N'L'}R\right)^{C_*}\frac{N'\sqrt{x'}}{(L')^2} =d^{-C_*-3/2}e^{-C_*+3/2} \left(\frac{NL}R\right)^{C_*}\frac{N\sqrt x}{L^2}. \] Take $C_*\geq2$, so both improve. All other length, residue and cutoff conditions retain their margins because $g\leq x^\eta$ and $\eta,\delta$ are sufficiently small. Apply~\eqref{eqn: coprime improved convolution} and sum the blocks, using \[ \sum_{de\mid|a_1|}\frac1{de} =\sum_{g\mid|a_1|}\frac{\tau(g)}g \leq\left(\frac{|a_1|}{\phi(|a_1|)}\right)^2 \ll(\log\log x)^2. \] The resulting estimate is $x(\log x)^4/\sqrt D+O(x^{1-c})$ for some $c>0$, with the number of blocks absorbed in the power-saving error. Decrease $c$ to include the saving in~\eqref{eqn: change of conductor cutoff}. Restoring the original cutoff and decreasing $\delta$ so that $\delta\leq\eta/4$ and $\delta<2c$ gives \[ \mcD(D)\ll_\kappa \frac{x(\log x)^4}{\sqrt{\min\{x^\delta,D\}}} \qquad(1\leq D\leq x^\kappa). \] \end{proof} \appendix \section{The regular spectrum}\label{sec: regular spectrum} \setcounter{theorem}{0} \renewcommand{\thetheorem}{\thesection.\arabic{theorem}} \renewcommand{\theHtheorem}{\thesection.\arabic{theorem}} By the regular spectrum we mean the real Maass, holomorphic and continuous spectra. For real Maass forms we take $\kappa_f\geq0$. We record the estimates needed to complete the proof of Proposition~\ref{propn: improved squarefree Kloosterman}. The argument follows the exceptional case, but there is no factor $X^{\theta_f}$. For the holomorphic and continuous spectra we use the Petersson and Plancherel weights, respectively, with the Fourier normalisations of~\cite[Section~3]{PascadiLargeSieve}. \begin{lemma}[Ordinary spectral large sieve]\label{lem: regular large sieve} Let $q\in\Z_{>0}$, $\eta>0$, $T\geq1$, and $\mathfrak a\in\{\infty,0\}$. For a complex sequence $(b_n)$ supported on $n\asymp J$, $J\geq1/2$, \begin{equation}\label{eqn: regular second moment} \sum_{\substack{f\in\mathcal B_q\\0\leq\kappa_f\leq T}} \frac1{\cosh(\pi\kappa_f)} \left|\sum_n b_n\rho_{f,\mathfrak a}(n)\right|^2 \ll_\eta(qJT)^\eta\left(T^2+\frac Jq\right)\sum_n|b_n|^2. \end{equation} The same bound holds for the holomorphic and continuous spectra, replacing the Maass sum and its weight by the corresponding Petersson sum or Plancherel integral, with spectral parameters truncated at $T$. For $b\in\Z_{>0}$ with $(b,q)=1$, one also has \begin{equation}\label{eqn: regular fixed index} \sum_{\substack{f\in\mathcal B_q\\0\leq\kappa_f\leq T}} \frac{|\rho_{f,\mathfrak a}(b)|^2}{\cosh(\pi\kappa_f)} \ll_\eta(qbT)^\eta\left(T^2+\frac{\sqrt b}{q}\right). \end{equation} \end{lemma} \begin{proof} These are~\cite[Lemmas~3.8--3.9]{PascadiLargeSieve}. As in Lemma~\ref{lem: improved moment inputs}, fixed compact supports and bounded lengths are covered by a fixed partition of unity. \end{proof} \begin{lemma}[Real Maass moments]\label{lem: regular moments} Let $\eta>0$ and $T\geq1$. For $b,q\in\Z_{>0}$ with $(b,q)=1$, \begin{equation}\label{eqn: regular Hecke moment} \sum_{\substack{f\in\mathcal B_q\\0\leq\kappa_f\leq T}}|\lambda_f(b)|^2 \ll_\eta(qbT)^\eta(1+T)^{C_0}(q+\sqrt b). \end{equation} For $J\geq1/2$ and $|\beta_n|\leq1$ supported on $n\asymp J$, independently of $f$, \begin{equation}\label{eqn: regular fourth moment} \sum_{\substack{f\in\mathcal B_q\\0\leq\kappa_f\leq T}} \frac1{\cosh^2(\pi\kappa_f)} \left|\sum_n\beta_n\rho_{f,\mathfrak a}(n)\right|^4 \ll_\eta(qJT)^\eta(1+T)^{C_0} \left(\frac{J^2}{q}+\frac{J^4}{q^2}\right), \quad\mathfrak a=\infty,0. \end{equation} Under the hypotheses and notation of Lemma~\ref{lem: improved long fourth moment}, \begin{equation}\label{eqn: regular long fourth moment} \begin{aligned} &\sum_{\substack{f\in\mathcal B_q\\0\leq\kappa_f\leq T}} \frac1{\cosh^2(\pi\kappa_f)} \left|\sum_v\Phi(ev/Y)c_{ev}\rho_{f,0}(bv)\right|^4\\ &\qquad\ll_\eta\mathfrak d^{C_0}(qYeT)^\eta(1+T)^{C_0}g^4 \left(\frac{Y^2}{q}+\min\left\{\frac{Y^4}{q^2},q\right\}\right). \end{aligned} \end{equation} Here $C_0$ may be increased between the three estimates. \end{lemma} \begin{proof} For a primitive real Maass form $g$ of level $s$ and fixed even nebentypus, the first-coefficient estimate used in Lemma~\ref{lem: improved Fourier coefficients} takes the form \[ s^{-1-\eta}(1+T)^{-C_0} \ll_\eta\frac{|\rho_{g,\infty}(1)|^2}{\cosh(\pi\kappa_g)} \ll_\eta s^{-1+\eta}(1+T)^{C_0} \qquad(0\leq\kappa_g\leq T). \] The oldspace mass bound~\eqref{eqn: oldspace first coefficient mass} now acquires a factor $(1+T)^{-C_0}$. The argument of Lemma~\ref{lem: improved Hecke second moment}, applied directly at level $q$ with~\eqref{eqn: regular fixed index}, proves~\eqref{eqn: regular Hecke moment}. The expansion~\eqref{eqn: improved oldform expansion} also holds, with an additional factor $(1+T)^{C_0}$ in the bound for $c_f(\delta)$. The Hecke multiplication argument in Lemma~\ref{lem: improved ordinary fourth moment} and the nebentypus large sieve~\cite[Proposition~4.7]{DrappeauDispersion} give, for primitive forms of level $s$ and even nebentypus $\omega$ of conductor $m_\omega$, \[ \left(\sum_{\substack{g\ \mathrm{primitive}\ (s,\omega)\\0\leq\kappa_g\leq T}} \left|\sum_n\beta_n\lambda_g(n)\right|^4\right)^{1/4} \ll_\eta(sJT)^\eta(1+T)^{C_0}m_\omega^{C_0} \left(s^{1/4}\sqrt J+J\right). \] Inserting~\eqref{eqn: improved oldform expansion} and summing over the primitive levels and divisors as in Lemma~\ref{lem: improved ordinary fourth moment} proves~\eqref{eqn: regular fourth moment}. For~\eqref{eqn: regular long fourth moment}, the rational-twist argument of Lemma~\ref{lem: improved rational twist} also applies to real Maass forms. The Voronoi identity~\eqref{eqn: improved Voronoi} has the same normalisation for real spectral parameters. The twists and the Voronoi partner preserve the spectral parameter, and the preceding primitive fourth moment applies at the resulting levels and nebentypus. The only additional dependence comes from the transforms. With the notation of Lemma~\ref{lem: improved Hankel}, \[ \begin{aligned} &\sup_{|t|\leq T}\|F_{t,\rho}^{(j)}\|_\infty +\int_\R(1+|\zeta|)^j\sup_{|t|\leq T}|\widehat F_{t,\rho}(\zeta)|\,d\zeta\\ &\qquad\ll_{j,B}(1+T+Z)^{C_{j,B}} \begin{cases} 1+|\log\rho|,&0<\rho\leq1,\\ \rho^{-B},&\rho\geq1. \end{cases} \end{aligned} \] Indeed, the small-argument powers $v^{\pm it}$ have modulus one, with a logarithm at $t=0$; the hyperbolic and gamma factors cancel their exponential dependence on $t$. The differential-operator and Fourier-inversion arguments are otherwise identical. Thus Lemma~\ref{lem: improved rational twist} holds for the real spectrum truncated at $T$, with an additional factor $(1+T)^{C_0}$. Applying it to~\eqref{eqn: improved additive Poisson}, and using~\eqref{eqn: regular fourth moment} for the other bound, gives~\eqref{eqn: regular long fourth moment}. The divisor parameter $e$ still has no fixed positive power. \end{proof} Let $\mathcal T_\sigma^{\mathrm{real}}(U)$, $\mathcal T_\sigma^{\mathrm{hol}}(U)$ and $\mathcal T_\sigma^{\mathrm{Eis}}(U)$ denote the real Maass, holomorphic and Eisenstein contributions to the block $\mathcal T_\sigma(U)$ defined before Lemma~\ref{lemma: improved Poisson Kuznetsov}. \begin{lemma}[Completed blocks on the regular spectrum]\label{lem: regular completed block} Under the hypotheses of Lemma~\ref{lemma: improved Poisson Kuznetsov}, the estimate~\eqref{eqn: improved Kuznetsov majorant} holds for $\mathcal T_\sigma^{\mathrm{real}}(U)$ upon replacing the exceptional sum by \[ \sum_{\substack{f\in\mathcal B_q\\0\leq\kappa_f\leq x^{C_1\eta}}} \frac1{\cosh(\pi\kappa_f)}|P_f(U,\zeta)| \left|\sum_vu_{q,v}e(\xi v/Y)\rho_{f,0}(|a|v)\right|. \] For $\mathcal T_\sigma^{\mathrm{hol}}(U)$ and $\mathcal T_\sigma^{\mathrm{Eis}}(U)$, use the corresponding Petersson sum over weights at most $x^{C_1\eta}$ or Plancherel integral over $|t|\leq x^{C_1\eta}$, with no exceptional factor. The remainder and the dependence of constants are as in that lemma. \end{lemma} \begin{proof} Use the same Kuznetsov test function and Fourier separation. Since $X^{-1}\leq x^{C\eta}$, the Bessel-transform estimates in~\cite[Lemma~3.4]{PascadiLargeSieve} allow the spectral parameters to be truncated at $x^{C_1\eta}$. Integration by parts, followed by Lemma~\ref{lem: regular large sieve}, bounds the tails by $O_{B,\eta}(x^{-B}\sum_{q,v}|u_{q,v}|)$. On the remaining spectra the transform costs $x^{O(\eta)}$, with no factor $X^{\theta_f}$. Keeping the cutoff $\psi$ outside the Fourier separation again makes the coefficients of $P_f$ common to all levels. \end{proof} \begin{proposition}\label{propn: regular Kloosterman} Under the hypotheses of Proposition~\ref{propn: improved squarefree Kloosterman}, let $\mathcal K_{\mathrm{reg}}$ be the sum of the real Maass, holomorphic and Eisenstein contributions obtained by completing the $\ell$-sum and applying Kuznetsov to the nonzero completion frequencies. Then \[ |\mathcal K_{\mathrm{reg}}|^2 \ll(\mathfrak d\Lambda)^{C_0}x^{C_0(\eta+\delta)}L^2N^7\sqrt x. \] \end{proposition} \begin{proof} Use the reductions in the proof of Proposition~\ref{propn: improved squarefree Kloosterman}: first take $\Lambda=1$, suppress fixed powers of $\mathfrak d$, and group \[ g\asymp G,\qquad Q=N^2/G,\qquad U\ll Nx^{2\eta},\qquad Y\ll HN/G\leq Q. \] Apply Lemma~\ref{lem: regular completed block}. The factors $(1+T)^{C_0}$ at the cutoff $T=x^{C_1\eta}$ are absorbed into $x^{O(\eta)}$. For real Maass forms, apply the inverse Hecke relation and H\"older with exponents $2,4,4$ as before. Lemma~\ref{lem: regular moments} gives the three factors \[ Q(Q+\sqrt{|a|}),\qquad U\left(1+\frac{U^2}{Q}\right)^{1/2},\qquad G^2Q B(Y;Q), \] up to $x^{O(\eta)}$, with $B(Y;Q)$ as in Lemma~\ref{lemma: improved spectral envelope}. Since $Q\leq\sqrt x$, $|a|\leq x^{1+\delta}$ and $B(Y;Q)\ll1$, we obtain \begin{align*} |\mathcal K_{G,\mathrm{real}}(U,Y)|^2 &\ll x^{O(\eta+\delta)}L^2Q\cdot Q\sqrt x\cdot U\left(1+\frac{U^2}{Q}\right)^{1/2}\cdot G^2Q\\ &\ll x^{O(\eta+\delta)}L^2N^7\sqrt x\, G^{-1}\sqrt{1+G} \ll x^{O(\eta+\delta)}L^2N^7\sqrt x\,G^{-1/2}. \end{align*} Here the harmless enlargement of $U\leq N$ is included in $x^{O(\eta)}$. For holomorphic forms, Deligne's bound~\cite{Deligne} gives $|\lambda_f(b)|\leq\tau(b)$ at unramified indices. For Eisenstein series, the same bound follows from \[ \lambda_{\chi,t}(b)=\sum_{uv=b}\chi(u)\overline{\chi(v)}(u/v)^{it}. \] Fix a divisor $e\mid|a|$ in the inverse Hecke relation and put $e_1=e/(e,t)$, as in the main proof. Restriction to multiples of $e_1$ does not increase the coefficient energy: \[ \sum_v|c_{e_1v}|^2\leq\sum_w|c_w|^2 \ll H^2(1+HG/N). \] The last bound follows by counting equal differences $s'h-sh'$. The divisor sums and the Hecke eigenvalues cost $x^{O(\eta)}$, after allowing fixed powers of $\mathfrak d$. There are $O(N^2/G)$ pairs $(n,n')$ with $(n,n')\asymp G$. Cauchy--Schwarz and the two ordinary large sieves from Lemma~\ref{lem: regular large sieve} therefore give, using $U\ll Nx^{2\eta}$, $N\ll Q$ and $Y\ll Q$, \begin{align*} |\mathcal K_{G,\mathrm{hol}}(U,Y)|^2+ |\mathcal K_{G,\mathrm{Eis}}(U,Y)|^2 &\ll x^{O(\eta)}L^2Q\cdot QU\cdot \frac{N^2H^2}{G}(1+HG/N)\\ &\ll x^{O(\eta)}L^2N^9G^{-2} \leq x^{O(\eta)}L^2N^7\sqrt x\,G^{-2}. \end{align*} The last inequality uses $N^2\leq\sqrt x$. The fixed powers of $\mathfrak d$ also absorb the small multipliers in the Fourier indices. The dyadic $G$-sums converge after taking square roots, and the $U,Y$ sums cost only $x^{O(\eta)}$. Finally, restore $\mathfrak d$ and $\Lambda$. As in the length comparisons in the main proof, the level becomes $dQ$, while the inequalities $H\leq N$, $U\leq N$, $Y\leq Q$ and $Q\leq\sqrt x$ are enlarged by fixed powers of $\mathfrak d\Lambda$ and $x^{O(\eta)}$. The spectral cutoff has the same type of enlargement. All preceding bounds therefore lose at most $(\mathfrak d\Lambda)^{C_0}x^{C_0(\eta+\delta)}$, proving the result. \end{proof} \setlength{\parskip}{0pt} \begin{thebibliography}{99} \bibitem{Assing-Blomer-Li} E. Assing, V. Blomer, and J. Li. Uniform Titchmarsh divisor problems. \textit{Adv. Math.} \textbf{393} (2021), 108076. \href{https://doi.org/10.1016/j.aim.2021.108076}{doi:10.1016/j.aim.2021.108076}. \bibitem{AssingCorbett} E. Assing and A. Corbett. Vorono\"i summation via switching cusps. Preprint (2019), arXiv:1904.02025v1 [math.NT]. \url{https://arxiv.org/abs/1904.02025v1}. \bibitem{BlomerKhan} V. Blomer and R. Khan. Twisted moments of $L$-functions and spectral reciprocity. Preprint (2018), arXiv:1706.01245v2 [math.NT]. \url{https://arxiv.org/abs/1706.01245v2}. \bibitem{Deligne} P. Deligne. La conjecture de Weil. I. \textit{Publ. Math. Inst. Hautes \'{E}tudes Sci.} \textbf{43} (1974), 273--307. \bibitem{Drappeau} S. Drappeau. Théorèmes de type Fouvry-Iwaniec pour les entiers friables. \textit{Compos. Math.} \textbf{151} (2015), no.~5, 828--862. \bibitem{DrappeauDispersion} S. Drappeau. Sums of Kloosterman sums in arithmetic progressions, and the error term in the dispersion method. Preprint (2016), arXiv:1504.05549v4 [math.NT]. \url{https://arxiv.org/abs/1504.05549v4}. \bibitem{HarperBV} A. J. Harper. Bombieri--Vinogradov and Barban--Davenport--Halberstam type theorems for smooth numbers. Preprint (2012), arXiv:1208.5992v1 [math.NT]. \url{https://arxiv.org/abs/1208.5992v1}. \bibitem{Hildebrand-Tenenbaum} A. Hildebrand and G. Tenenbaum. On integers free of large prime factors. \textit{Trans. Amer. Math. Soc.} \textbf{296} (1986), no.~1, 265--290. \bibitem{HoffsteinLockhart} J. Hoffstein and P. Lockhart. Coefficients of Maass forms and the Siegel zero. \textit{Ann. of Math.} \textbf{140} (1994), no.~1, 161--176. \bibitem{Kim-Sarnak} H. H. Kim. Functoriality for the exterior square of $\text{GL}_4$ and the symmetric fourth of $\text{GL}_2$. \textit{J. Amer. Math. Soc.} \textbf{16} (2003), 139--183. With Appendix 1 by D. Ramakrishnan and Appendix 2 by H. H. Kim and P. Sarnak. \bibitem{Lichtman} J. D. Lichtman. Primes in arithmetic progressions to large moduli, and Goldbach beyond the square-root barrier. Preprint (2023), arXiv:2309.08522v1 [math.NT]. \url{https://arxiv.org/abs/2309.08522v1}. \bibitem{PascadiLargeSieve} A. Pascadi. Large sieve inequalities for exceptional Maass forms and the greatest prime factor of $n^2+1$. \textit{Forum Math. Pi} \textbf{14} (2026), e8. \href{https://doi.org/10.1017/fmp.2026.10025}{doi:10.1017/fmp.2026.10025}. \bibitem{PascadiLvlDist} A. Pascadi. On the exponents of distribution of primes and smooth numbers. Preprint (2025), arXiv:2505.00653v2 [math.NT]. \url{https://arxiv.org/abs/2505.00653v2}. \bibitem{RamakrishnanYang} D. Ramakrishnan and L. Yang. A constraint for twist equivalence of cusp forms on $\mathrm{GL}(n)$. Preprint (2020), arXiv:1906.01047v3 [math.NT]. \url{https://arxiv.org/abs/1906.01047v3}. \end{thebibliography} \end{document}