\section{An independent classical bound for quadratic cores}\label{sec:older-route} The quadratic divisor pairing has two factors, so completed Kloosterman sums and positive restoration give the required classical order bounds. Fix an even modulus \(L_0\), a nonprincipal quadratic literal character \(\chi\bmod L_0\), and put \[ \begin{gathered} H=\ker\chi,\qquad \mathcal C=\langle p:p\nmid L_0,\ \chi(p)=1\rangle,\\ V(n)=\tau(n)\1_{\mathcal C}(n),\qquad J(N)=\prod_{\substack{p\mid N\\p\nmid L_0,\ \chi(p)=-1}} \frac{p-1}{p-2}. \end{gathered} \] All arithmetic data in the following lemma are fixed. \begin{lemma}[Classical quadratic core bound]\label{lem:older-quadratic-filter} Let \(A,B\) be coprime positive \(L_0\)-supported integers and \(U,V_0\subset H\) fixed residue subsets. For any positive \(A',B'\), define \[ \Gamma_{A',B';U,V_0}(N)=\frac{L_0}{|H|^2} \#\{(a,b)\in U\times V_0:A'a+B'b=N\bmod L_0\}. \] Uniformly over all sufficiently large targets with \(\Gamma_{A,B;U,V_0}(N)>0\), \[ \#\{Ax+By=N:x,y\in\mathcal C,\ x\bmod L_0\in U,\ y\bmod L_0\in V_0\} \asymp\frac{NJ(N)}{\log N}. \] Zero finite density gives zero count identically. In particular the lemma applies to unions of unit-square residue states. Its proof uses only fixed-character means, fixed-progression prime Mertens, the classical Weil--Estermann bound, and finite sieve inequalities. \end{lemma} \paragraph{Proof plan and parameter order.} Completed two-factor boxes give a shifted weighted family, uniform under the divisibility predicates needed by a finite lower sieve. The rough mass has lower-bound scale \(NJ(N)/(\log N)^2\). Restoration supplies the missing logarithm through the positive product in \eqref{eq:older-restoredmass}; the valuation patterns \((1,1),(1,0),(0,1)\) recover its labels \(d,u,v\), and the residual rough divisor weights are bounded. This makes restoration injective at the numerical-value level. Choose the box exponent and core family first, then the saving \(\eta\) and external exponent \(\theta\) from \eqref{eq:older-uniform-reserve}. Next choose \(\alpha,\xi\) by \eqref{eq:older-smallbudgets}, the Bonferroni depth, the roughness exponent, and finally the restoration range, in that order. All remain fixed as \(N\to\infty\). \subsection{Two-factor boxes with a fixed period} For bounded functions \(\omega_1,\omega_2\) of fixed period \(c\), intervals \(I_1,I_2\), and \((a,m)=(c,m)=1\), let \[ S(a)=\sum_{\substack{r\in I_1,\ s\in I_2\\rs\equiv a\ (m)}} \omega_1(r)\omega_2(s). \] Then, for every \(\epsilon>0\), \begin{equation}\label{eq:older-box} S(a)-\frac1{\varphi(m)}\sum_{b\bmod m}^{*}S(b) \ll_{c,\epsilon}m^{1/2+\epsilon}. \end{equation} The estimate is uniform in the intervals, including intervals longer than \(cm\), and permits subintervals of arbitrary lengths. Here is a direct completion proof. Put \(Q=cm\), \[ W_i(\nu)=\sum_{n\in I_i}\ee{-\nu n/Q},\qquad G_i(t)=\sum_{r\bmod c}\omega_i(r)\ee{tr/c}. \] CRT and finite Fourier inversion express \(S(a)\) as \[ Q^{-2}\sum_{\nu,k\bmod Q} W_1(\nu)W_2(k)G_1(m^{-1}\nu)G_2(m^{-1}k) K_1(c^{-1}\nu,ac^{-1}k;m), \] where \(K_1(u,v;m)=\sum_{z\bmod m}^{*}\ee{(uz+v\bar z)/m}\); inverses in \(G_i\) are taken modulo \(c\). The terms \(\nu=0\bmod m\) or \(k=0\bmod m\) are independent of the unit \(a\), since the resulting Ramanujan sum is unchanged by unit multiplication. For all other frequencies use \[ |K_1(u,v;m)|\le\tau(m)m^{1/2}(u,v,m)^{1/2}. \] This is the ordinary Weil--Estermann bound stated and applied in \cite[printed p.~44]{heathbrown}. The primary source's Lemma~6 is its singleton-padded divisor-box consequence. In centered frequency ranges, \(|W_i(\nu)|\ll\min(|I_i|,Q/|\nu|)\) and \(|G_i|\le c\). The remaining double sum is bounded using \[ \sum_{\substack{0<|\nu|,|k|\le Q/2\\m\nmid \nu,\ m\nmid k}} \frac{(\nu,k,m)^{1/2}}{|\nu k|} \le \sum_{d\mid m}d^{1/2} \left(\sum_{\substack{0<|\nu|\le Q/2\\d\mid \nu}}\frac1{|\nu|}\right)^2 \ll_c \log^2(2m)\sum_{d\mid m}d^{-3/2}. \] The \(Q^2\) from the two geometric sums cancels the completion normalization. This proves a bound \(O_c(m^{1/2}\tau(m)\log^2(2m))\) around an \(a\)-independent quantity. Averaging that quantity over unit \(a\) changes the error by at most a factor two and proves \eqref{eq:older-box}. The case \(m=1\) has discrepancy zero. Tensor partial summation also allows bounded-variation interval weights. \subsection{Shifted kernels, means, and fixed filters} For \(t\in\mathcal C\) put \(F_t(n)=V(tn)/V(t)\). At a split prime with \(b=v_p(t)\), \[ F_t(p^e)=1+\frac{e}{b+1}; \] at a nonsplit prime it is zero for \(e>0\). With literal characters, its Dirichlet series is \[ \sum_{n\ge1}\frac{F_t(n)}{n^s} =L(s,\chi_0^{L_0})L(s,\chi)\sum_{j\ge1}\frac{e_t(j)}{j^s}. \] The correction factors are \(1-p^{-2s}\) at inert primes and \(1-\frac{b}{b+1}p^{-s}\) at split primes dividing \(t\). Deleting all Euler factors at primes dividing another integer \(D\) replaces its split local correction by \((1-p^{-s})^2\). Thus all these kernels satisfy, uniformly in valuations, \begin{equation}\label{eq:older-kernel} |e(j)|\le\tau(j)^C,\qquad \sum_j|e(j)|j^{-3/5}\ll_{\epsilon}(tD)^\epsilon. \end{equation} Every kernel index is an \(L_0\)-unit with \(\chi(j)=1\). We record explicitly the elementary mean used below. If \(\xi\) is a character of a fixed modulus \(c\) with the same prime support as \(L_0\), then the twisted series has character factors \(\xi,\xi\chi\). It has a principal factor only when the restriction of \(\xi\) is \(1\) or \(\chi\); in those cases \(F_t\xi=F_t\). Otherwise both character sums have bounded partial sums, and Dirichlet hyperbola at \(\sqrt X\) gives \(O_c(\sqrt X)\) for their convolution. In the principal cases, split the principal factor from the nonprincipal one by the same hyperbola argument to obtain its residue times \(X\), with error \(O_c(\sqrt X)\). Summing kernels by \eqref{eq:older-kernel}, including their main tails, therefore gives \begin{equation}\label{eq:older-mean} \sum_{n\le X}F_t(n)\1_{(n,D)=1}\xi(n) =\delta_{\xi\in\{1,\chi\}}\rho_{t,D}X +O_\epsilon((tDX)^\epsilon X^{3/5}). \end{equation} Here the two displayed characters mean their inflations to \(c\). The error and main remain valid with any fixed periodic unit filter, after finite character expansion. For later normalization write \[ M_{t,p}=(1-p^{-1})\sum_{e\ge0}F_t(p^e)p^{-e},\qquad \rho_t=\frac{\varphi(L_0)}{L_0}\prod_{p\nmid L_0}M_{t,p}. \] The product is taken in natural prime order; its convergence and positivity follow equally from the character factorization and \(L(1,\chi)>0\). A deleted split prime divides this mean by its original local series. The residue measure on the fixed unit group is uniform on \(H\). \begin{proposition}[Uniform shifted weighted family]\label{prop:older-weighted} There exist \(\eta,\theta>0\), with \(\theta\) chosen sufficiently small after \(\eta\), such that the following holds. Let \(t_1,t_2,u,v\in\mathcal C\), each at most \(N^\theta\), with no restriction on their gcds. For fixed residue subsets \(U,V_0\subset H\), \begin{align*} &\sum_{\substack{Au x+Bv y=N\\x,y\ge1,\ x\bmod L_0\in U,\ y\bmod L_0\in V_0}} F_{t_1}(x)F_{t_2}(y)\\ &\quad=\frac{N}{ABuv}\, \Gamma_{Au,Bv;U,V_0}(N)\rho_{t_1}\rho_{t_2} \prod_{p\nmid L_0}\frac{D_p}{M_{t_1,p}M_{t_2,p}} +O(N^{1-\eta}), \end{align*} uniformly in every target valuation. Here \[ D_p=\lim_{m_p\to\infty}p^{-m_p} \sum_{\substack{a,b\bmod p^{m_p}\\Au a+Bv b=N\ (p^{m_p})}} F_{t_1}(p^{\min(v_p(a),m_p)}) F_{t_2}(p^{\min(v_p(b),m_p)}). \] All factors, including zero factors, are retained. The same formula holds after imposing any fixed finite combination of small divisibility and exact-valuation tests whose resulting shifts and coefficients stay in this family. \end{proposition} \begin{proof} First suppose \((u,v)=(uv,N)=1\). For \(g=(x,N)\), write \(x=gx'\), \(y=gy'\), \(M=N/g\); a nonzero term has \(g\in\mathcal C\). Its exact prefactor is \(F_{t_1}(g)F_{t_2}(g)\), its new shifts are \(t_1g,t_2g\), and \((x',M)=1\). Transport the residue states to \(U/g,V_0/g\). Every such multiplication is a unit bijection, so the corresponding finite root count, with its transported target, is unchanged. Choose a small fixed \(\lambda>0\), put \(K=N^\lambda\), and truncate \(g\) and the two kernels at \(K\). Large gcds cost \(O(N^{1-\lambda+\epsilon})\). For either kernel, counting multiples and using \eqref{eq:older-kernel} gives \[ N^{1+\epsilon}\sum_{j>K}\frac{|e(j)|}{j} +N^\epsilon\sum_{j\le N}|e(j)| \ll N^{1-2\lambda/5+\epsilon}+N^{3/5+\epsilon}. \] Discard also \(Au x\le N^{1-\lambda}\), at \(O(N^{1-\lambda+\epsilon})\). All implicit valuation and shift factors here are divisor-bounded. We suppress primes on \(x',y'\) in the remainder of this core argument. The fixed filter on the original first unit variable is \[ w(x)=\1_{x\bmod L_0\in U/g}\, \1_{B\mid M-Au x}\, \1_{(z,L_0)=1,\ z\bmod L_0\in v(V_0/g)},\qquad z=(M-Au x)/B. \] It has fixed period \(c=BL_0\). Expand it on unit classes by characters modulo \(c\), retaining every scalar phase of every kernel index. The filter ensures that \(t=(M-Au x)/(Bvk)\) is an \(L_0\)-unit with \(\chi(t)=1\). Consequently the exact symmetric divisor pairing is \[ (\chi_0^{L_0}*\chi)(t) =2\sum_{\substack{a\mid t\\a<\sqrt t}}\chi(a) +\1_{t=\square}\chi(\sqrt t). \] The square endpoint costs \(O(N^{1/2+O(\lambda+\theta)+\epsilon})\). For a second kernel \(k\le K\), the remaining congruence is \[ Au x\equiv M\pmod{vka},\qquad Au x+Bvka^20\). At a prime outside \(uMvkL_0\), its correction is explicitly \(1+(W_p-1)\chi(p)p^{-s}\); at primes of \(uM\) it has the finite deletion \(1-\chi(p)p^{-s}\), and at primes of \(vk\) no new correction is required. Since the partial sums of the fixed nonprincipal \(\chi\) are bounded, convolution with this correction gives \[ \sum_{a\le Y}b_k(a)\ll_\epsilon (t_1g\,uMvkY)^\epsilon . \] In particular \(\sum b_k(a)/a\) converges, its tail after \(Y\) is \(O(N^\epsilon/Y)\), and \(\sum_{a\le Y}a\,b_k(a)=O(N^\epsilon Y)\). Consequently \[ \begin{split} &\sum_{aK\) tail is therefore \(O(N^{1-2\lambda/5+\epsilon})\). Both errors are included in \eqref{eq:older-uniform-reserve}. Here is the Euler calculation of the completed sum, without an implicit normalization. Write \(z=p^{-1}\), \(q_p=1-z\), \[ A_i=\mathcal A(p^i),\qquad B_i=\mathcal B(p^i),\qquad P_{\mathcal B,p}(z)=\sum_{i\ge0}e_{\mathcal B}(p^i)z^i. \] The kernel identity says \[ \frac{P_{\mathcal B,p}(z)}{1-\chi(p)z} =(1-z)\sum_{i\ge0}B_i z^i =M_{\mathcal B,p}. \] The double sum over \(k,a\), after extracting its real factor \(M/(uv)\), has the following local factors: \begin{align*} \mathcal L_p &=1+W_p\left(\frac{P_{\mathcal B,p}(z)}{1-\chi(p)z}-1\right) &&(p\nmid uvM)\\ &=\frac{M_{\mathcal A,p}+M_{\mathcal B,p}-1} {M_{\mathcal A,p}} &&(p\nmid uvM),\\ \mathcal L_p&=1 &&(p\mid uM),\\ \mathcal L_p &=\frac{W_pP_{\mathcal B,p}(z)}{1-\chi(p)z} =\frac{M_{\mathcal B,p}}{M_{\mathcal A,p}} &&(p\mid v). \end{align*} For \(p\mid uM\), both local indices \(k,a\) are excluded. For \(p\mid v\), the factor \(W_p\) is already present even at \(v_p(k)=v_p(a)=0\); this explains that row of the table. The first mean is \[ \rho_{\mathcal A,M} =\alpha_0\prod_{p\nmid L_0}M_{\mathcal A,p} \prod_{\substack{p\mid M\\p\nmid L_0}}H_{\mathcal A,p}^{-1}. \] At \(p\mid M\) its combined factor is therefore \(M_{\mathcal A,p}/H_{\mathcal A,p}=q_p\). Multiplying the other rows by this first mean gives the completed fixed-\(g\) expression \begin{equation}\label{eq:older-completed-g} \begin{aligned} &C_g\,\frac{M}{Auv}\,2\alpha_0\mu \prod_{\substack{p\mid M\\p\nmid L_0}}q_p \prod_{p\mid u}M_{\mathcal A,p}\\ &\qquad{}\times \prod_{p\mid v}M_{\mathcal B,p} \prod_{\substack{p\nmid L_0uvM}} (M_{\mathcal A,p}+M_{\mathcal B,p}-1). \end{aligned} \end{equation} Intermediate products are taken in natural prime order. Their factorization into the fixed \(L(1,\chi)\) and absolutely convergent corrections justifies every completion; the final normalized product in the proposition is absolutely convergent. We may extend the divisor sum from \(g\le K\) to all \(g\mid N\). The means and local factors in \eqref{eq:older-completed-g}, with their normalized shifts, are \(N^{O(\epsilon)}\) uniformly, whereas \(C_gM\ll_\epsilon N^{1+\epsilon}/g\). The added main is \(O(N^{1-\lambda+O(\epsilon)})\). At \(p^n\Vert N\), use now the original local coefficients \(A_i=F_{t_1}(p^i)\), \(B_i=F_{t_2}(p^i)\). The choices \(v_p(g)=in}A_i z^i +A_n\sum_{j>n}B_j z^j\right). \end{split} \] At \(p\mid u\) or \(p\mid v\), the target is a unit and the respective factors remain \(M_{A,p}\) or \(M_{B,p}\), exactly the original local line integral. Finally \eqref{eq:older-filter-normalization} implies \[ \frac{2\alpha_0\mu}{Auv} =\frac{\Gamma_{Au,Bv;U,V_0}(N)\alpha_0^2}{ABuv}, \qquad \rho_{t_1}\rho_{t_2}=\alpha_0^2 \prod_{p\nmid L_0}M_{t_1,p}M_{t_2,p}. \] Together with \eqref{eq:older-completed-g}, these identities give precisely the proposition's main term in the unit-coefficient target core, including the bad coefficient \(B\). For completeness, the general coefficient reductions also have an explicit local check. At a good prime write \(a=v_p(u),b=v_p(v),n=v_p(N)\), and \(t=\min(a,b)\). If \(n0,b=0\), solving for the unit-coefficient variable gives \[ D_p^{\rm red}= \begin{cases} B_nM_{A,p},&n0,a=0\). A zero normalizing coefficient makes the relevant density zero. These formulas retain arbitrarily large coefficient and target valuations. Globally remove \(r=(u,v)\), requiring \(r\mid N\), and divide the coefficients and target by \(r\). For the coprime remaining coefficients put \(d=(u,N)\), \(e=(v,N)\). Then \(e\mid x,d\mid y\); the exact substitutions are \[ x=ex',\quad y=dy',\quad (t_1,t_2)\mapsto(t_1e,t_2d),\quad (u,v,N)\mapsto(u/d,v/e,N/(de)), \] with prefactor \(F_{t_1}(e)F_{t_2}(d)\) and states \(U/e,V_0/d\). The new coefficients are prime to the new target. All removed factors belong to \(\mathcal C\). Their unit multipliers biject the finite root sets, with the transported target and states. The displayed local formulas show that the radial integrals transform with exactly the same prefactors and common-coefficient gains. Thus reexpressing the transformed main in the original variables gives the stated \(D_p\)-product, for every gcd profile. All estimates in \eqref{eq:older-uniform-reserve} are uniform in these profiles and in the shift valuations. They depend on divisor bounds and \eqref{eq:older-kernel}, not on an unbounded number of local cases. When the external family exponent is decreased as below, transformed targets are at least \(N^{1-3\theta}\), transformed parameters lie in the proved core range, and the divisor-bounded prefactors preserve a smaller positive saving. Choose \(\lambda\) small, then \(\theta\) much smaller. Equations \eqref{eq:older-box-budget} and \eqref{eq:older-boundary}, the kernel saving \(2\lambda/5\), and the gcd and small-summand saving \(\lambda\), leave a fixed positive core saving. Choose \(\eta\) below a hundredth of the boundary saving, and decrease the external family exponent again. The transformed targets are at least \(N^{1-3\theta}\) and all transformed parameters stay inside the core family. Divisor prefactors preserve a smaller saving. Finally \(b\mid x\) is the exact substitution \(x=bx'\), which changes the shift to \(t_1b\), coefficient to \(ub\), state to \(U/b\), and introduces \(F_{t_1}(b)\). The analogous substitution on the second side, and finite differences for exact valuations, prove the predicate assertion. \end{proof} \subsection{A finite lower sieve and positive restoration} We now prove Lemma~\ref{lem:older-quadratic-filter}. Fix an admissible target. Its finite density \(\Gamma(N)\) belongs to a fixed finite positive set, and hence has a positive lower bound depending only on the fixed data. At good primes write \(x=p^{-1}\), \(q_p=1-x\), and use the unshifted coefficients \(A_i=B_i=V(p^i)\). Let \(I_{p,e}\) be the radial integral in the preceding proof, where \(e=v_p(N)\). The weighted mass with both variables units is \[ c_e(p)= \begin{cases}1-2/p,&e=0,\\1-1/p,&e\ge1.\end{cases} \] At an inert prime \(I_{p,e}=c_e(p)\). At a split prime define the bad-event probability \(g_e(p)=1-c_e(p)/I_{p,e}\). The explicit geometric series gives, uniformly in \(e\), \begin{equation}\label{eq:older-roughprob} 0\le g_e(p)<1,\qquad g_e(p)=4/p+O(p^{-2}). \end{equation} For example, at \(e=0\), \(I_{p,0}=2(1-p^{-1})^{-1}-1\). For \(e\ge1\), the \(i=0,1\) terms give \(I_{p,e}=1+3/p+O(p^{-2})\); the remaining polynomial-geometric tail is uniformly \(O(p^{-2})\). The formula at \(e=1\) has the same expansion directly. At every finite collection of primes \(I_{p,e}\) is uniformly bounded, whereas \(c_e(p)\ge1/3\), so the good-event probabilities have a uniform positive lower bound. Put \(M_p=(1-p^{-1})\sum_iV(p^i)p^{-i}\) and \[ \mathcal V(z)=\prod_{\substack{p\le z\\p\nmid L_0}} \frac{1-2/p}{M_p^2},\qquad E_p=\frac{p-1}{p-2},\qquad E_{p,e}=\begin{cases}1,&e=0,\\E_p,&e\ge1,\end{cases} \qquad a_p=\frac{p-1}{p(p-2)}. \] At a split prime \(M_p=(1-p^{-1})^{-1}\), so the factor in \(\mathcal V\) is \(1-4/p+O(p^{-2})\). At an inert prime it is \(1+O(p^{-2})\). Fixed-character prime Mertens, and the split-prime density \(1/2\), therefore give \begin{equation}\label{eq:older-roughdimension} \mathcal V(z)\asymp(\log z)^{-2}. \end{equation} This is the dimension-two rough weighted count which must recover one power of \(\log N\). Choose small fixed \(\alpha,\xi>0\) after the weighted family. We use squarefree pairwise coprime triples \(d,u,v\) with \[ d,u,v\in\mathcal C,\quad d\mid N,\quad (uv,N)=1,\quad duv\le N^\alpha, \] all primes at most \(w=N^\omega\). At target \(M=N/d\), prescribe \(v_p(x)=1\) for \(p\mid u\), \(v_p(y)=1\) for \(p\mid v\), and require that no other prime at most \(z=N^\gamma\) divide \(xy\). Count with weight \(V(x)V(y)/(2^{\omega(u)}2^{\omega(v)})\) and with residue states \(x\in U/d,\ y\in V_0/d\). Thus the lifted values \(dx,dy\) have the original states. All these predicates are covered by Proposition~\ref{prop:older-weighted}. Indeed exact valuation one is the difference of divisibility by \(p\) and by \(p^2\), and a roughness predicate expands using \[ \1_{p\mid xy}=\1_{p\mid x}+\1_{p\mid y} -\1_{p\mid x,\ p\mid y}. \] For a tested squarefree divisor \(\ell\le N^\xi\), each resulting shift and coefficient is at most \(N^{2\alpha+\xi}\). The expansion has \(N^{o(1)}\) terms. It suffices to choose \begin{equation}\label{eq:older-smallbudgets} 2\alpha+\xi<(1-\alpha)\theta,\qquad \alpha+\xi<\eta/4 . \end{equation} Forbidden shifts give zero terms; common-divisibility terms use the noncoprime part of the weighted family. The finite density remains exactly \(\Gamma(N)\) in every lower term. After writing \(x=ux'\), \(y=vy'\), its states are \(U/(du),V_0/(dv)\) and its equation is \(Au x'+Bv y'=N/d\). The map \((x',y')\mapsto(du x',dv y')\) bijects its finite roots with the original roots modulo \(L_0\). This is why split primes with nontrivial residue states may be restored. Restricting restoration to identity-state primes would lose the exponent needed below. Here is the exact positive local model after imposing all roughness conditions. Its main mass for a triple is \begin{equation}\label{eq:older-triplemass} \mathcal M_{d,u,v} =C_N(z)\frac{N}{ABd}\mathcal V(z) \prod_{\substack{p\mid N/d\\p\le z,\ p\nmid L_0}}E_p \prod_{p\mid uv}a_p . \end{equation} The positive factor \(C_N(z)\) includes \(\Gamma(N)\rho_1^2\) and the Euler tail above \(z\). It is bounded above and below by fixed positive constants, uniformly in the target and the triples. To verify this, outside the target the tail factors are \(1+O(p^{-2})\); at target primes they are \(1+O(1/p)\) uniformly in valuation. At most \(1/\gamma\) target primes exceed \(z\). All cofactor primes are below \(w