\section{Fixed shifts and wide major arcs}\label{sec:fixed-shift-comparison} For fixed \(k=h\ne0\), the shared-label comparison \eqref{eq:gr-comparison}--\eqref{eq:gr-comparison-term} remains valid. The frequency estimate has a different scale. After localizing \(D_i\asymp U_i\), put \[ U_*=U_1U_2,\qquad Y=xU_*,\qquad W=U_*L^{-B_0}. \] The multiplier is now \(m(h\theta)\). The integer bilinear bound \eqref{eq:gr-multiplier} confines its significant part to \(O_h(L^{2B_0})\) arcs about rationals \(a/q\), with \[ q\le |h|L^{B_0},\qquad |\theta-a/q|\le C_h/W. \] In particular, their physical width is \(L^{B_0}/U_*\), rather than \(L^{B_0}/Y\). The target-grid proposition does not estimate these arcs: on them \(|\beta|Y\) can be of order \(xL^{B_0}\). We extract the already established Dirichlet-frequency argument of \cite[Sections 6.2--6.4]{walsh}, with the arithmetic hypotheses \textup{(C1)--(C3)} in place of its particular divisor model. \begin{proposition}[Fixed-shift comparison energy] \label{prop:fixed-comparison-energy} Let \(F\) have the form \eqref{eq:gr-energy-F}, with a selected small bank whose label does not enter \(b\), and with its exact prime count equal to one. Suppose \[ U_i\ge\exp(L^{1/3}),\qquad U_i\le\exp(O((\log L)L^{2/5})),\qquad Y=xU_1U_2. \] For every fixed \(B_0,D>0\), conditions \textup{(C1)--(C3)} and the zero-free input imply, at sufficiently deep fixed model depth, \begin{equation}\label{eq:fixed-wide-energy} \int_{|\beta|\le C_h/W} |\widehat F(a/q+\beta)|^2\,d\beta \ll_{D,h}YL^{-D} \end{equation} uniformly for \(q\le |h|L^{B_0}\), where \(\widehat F(\theta)=\sum_nF(n)\ee{\theta n}\). Consequently every fixed-shift comparison in \eqref{eq:gr-comparison-term} has an arbitrary prescribed logarithmic saving. No binary cancellation estimate is assumed. \end{proposition} \subsection{Conversion to Dirichlet frequencies} We use the following existing two mean-square inequalities, in their normalizations from \cite[Lemma 6.3 and (6.14)]{walsh}. If \(a_n\) is supported on \(n\asymp Y\), then \begin{equation}\label{eq:fixed-additive-mellin} \int_{|\beta|\le C/W}\left|\sum_na_n\ee{\beta n}\right|^2d\beta \ll_C\frac1Y\int_{\mathbb R}(1+W|t|/Y)^{-2} \left|\sum_na_nn^{it}\right|^2dt. \end{equation} Here is the averaging step in that inequality, following \cite[Lemmas 2.1--2.2]{soundararajan}. Put \(\eta=W/(4C)\), \(A(v)=\sum_{n\le v}a_n\), \(P(t)=\sum_na_nn^{it}\), and \(J_\eta=\int_0^\infty|A(v+\eta)-A(v)|^2dv\). The interval kernel has Fourier transform of modulus at least \(c\eta\) on \(|\beta|\le C/W\), so Plancherel bounds the left side by \(O(\eta^{-2}J_\eta)\). For \(2\eta\le\nu\le3\eta\), split the increment at \(v+\nu\), square, and average in \(\nu\). Changing \(v\) to \(v+\eta\) in the second increment gives, if the coefficients are supported on \([aY,bY]\), \[ J_\eta\ll\eta^{-1}\int_\eta^{3\eta} \int_{aY/2}^{2bY}|A(v+\nu)-A(v)|^2dv\,d\nu. \] Substitute \(\nu=\Delta v\). Since \(v\asymp Y\), this is at most \[ \frac{CY^2}{\eta} \int_{c_1\eta/Y}^{c_2\eta/Y}\int_0^\infty |A(v(1+\Delta))-A(v)|^2\frac{dv}{v}\,d\Delta. \] Logarithmic Plancherel evaluates the inner integral exactly as \[ \frac1{2\pi}\int_{\mathbb R}|P(t)|^2 \left|\frac{(1+\Delta)^{it}-1}{t}\right|^2dt. \] The kernel is \(O(\min\{\eta^2/Y^2,|t|^{-2}\})\), and the \(\Delta\)-interval has length \(O(\eta/Y)\). Thus \(J_\eta\ll Y\int\min\{\eta^2/Y^2,|t|^{-2}\}|P(t)|^2dt\). Multiplying by \(\eta^{-2}\) proves \eqref{eq:fixed-additive-mellin}; no identification of fixed additive windows with fixed multiplicative windows is used. The other inequality, the ordinary Dirichlet-polynomial mean-value theorem, is \begin{equation}\label{eq:fixed-dirichlet-mean} \int_I\left|\sum_{n\le CR}a_nn^{it}\right|^2dt \ll_C (|I|+R)\sum_n|a_n|^2, \end{equation} uniformly in the location of the interval \(I\). Neither inequality is a new distributional input. Apply \eqref{eq:fixed-additive-mellin} to \(a_n=F(n)\ee{an/q}\) and set \[ \mathcal T(t)=Y^{-1}\sum_nF(n)\ee{an/q}n^{it},\qquad T_0=(Y/W)L^{B_1}=xL^{B_0+B_1}. \] The pointwise list bound and \textup{(C1)} give \(\sum_n|F(n)|^2\ll YL^C\). Thus on a dyadic interval \(T\le |t|\le2T\), \eqref{eq:fixed-dirichlet-mean} gives \[ \int_{T\le|t|\le2T}|\mathcal T(t)|^2dt \ll L^C(1+T/Y). \] Summing this with the weight in \eqref{eq:fixed-additive-mellin}, for \(|t|>T_0\), gives \begin{equation}\label{eq:fixed-frequency-tail} \int_{|t|>T_0}(1+W|t|/Y)^{-2}|\mathcal T(t)|^2dt \ll L^C\bigl(L^{-2B_1}+W^{-1}L^{-B_1}\bigr). \end{equation} The exponent \(B_1\) is chosen after the required saving, so this tail is arbitrarily small. It remains to prove \begin{equation}\label{eq:fixed-dirichlet-energy} \int_{|t|\le T_0}|\mathcal T(t)|^2dt\ll_D L^{-D}. \end{equation} \subsection{Low frequencies and recordwise centering} Choose a fixed exponent \(B_2\), to be specified after the smooth prime estimates, and put \(T_c=L^{B_2}\). In the low range write \(n=rv\), where \(r=n_{\mathcal P_y}\) is the complete smooth record and \(v\) is rough. Every list weight is a function of \(r\) alone and is bounded by \(L^C\). The terms with \(r>E\) have population \(O(YL^C\exp(-cL^{1/10}))\), by Rankin's inequality with parameter \(c/\log y\), as in \eqref{eq:record-tail}; their contribution to \(\mathcal T(t)\) is negligible uniformly in \(t\). For each remaining record, \(\ee{arv/q}\) is a bounded periodic test of period at most \(q\). The phase \((rv)^{i(t+\sigma)}\psi(rv/Y)\) has fixed logarithmic derivatives bounded by fixed powers of \(L\), throughout \(|t|\le T_c\). Condition \textup{(C3)} therefore gives any prescribed saving for that record. Summing uses only \[ \sum_{r\ \mathcal P_y\text{-smooth}}\frac1r =\prod_{p\le y}(1-1/p)^{-1}\ll L^C. \] Choose the precision in \textup{(C3)} after \(B_2\) and all list costs. This gives \begin{equation}\label{eq:fixed-low} \int_{|t|\le T_c}|\mathcal T(t)|^2dt\ll_D L^{-D}. \end{equation} The arithmetic progression test is on \(v\), so this argument does not require \(r\) or the endpoint to be a unit at \(L_0\). \subsection{High frequencies and a long cofactor} For this part remove selected-bank prime squares, whose population is \(O(Y/P_{\min})\). Applied to their bounded coefficients, \eqref{eq:fixed-dirichlet-mean} bounds their normalized integrated energy on \([-T_0,T_0]\) by \(L^C(1+T_0/Y)/P_{\min}\), hence by less than every logarithmic power. Do not impose the cutoff \(n_{\mathcal P_y}\le E\) here: it would couple the selected prime to its cofactor. The remaining terms have a unique selected-bank prime and can be written \(n=pm\). Exact invariance \textup{(C2)} gives \(\xi(pm)=\xi(m)\). The distinguished list coefficient ignores this prime, so after removing its factor \(V_g^{-1}\) the remaining coefficient \(\gamma(m)\) is independent of \(p\) and bounded by \(L^C\). It contains the condition that \(m\) has no selected-bank prime, together with the other list weights. Partition the selected bank into smooth dyadic pieces \(p\asymp P\), including its two sharp outer endpoints. Replace each sharp endpoint by a smooth cutoff across a strip of length \(\varepsilon P\), where \(\varepsilon=L^{-A_{\rm sm}}\). The normalized sum of prime labels in such a strip is at most \(O(\varepsilon+P^{-1})\), simply by counting integers. The Dirichlet mean-value bound for the cofactor below and Minkowski show that this replacement changes the normalized integrated energy by at most \[ L^C(\varepsilon+P_{\min}^{-1})^2. \] There are only polynomially many dyadic pieces in \(L\). Choose \(A_{\rm sm}\) large after the desired saving and these costs. The resulting smooth prime cutoffs have derivative bounds which are fixed powers of \(L\). Put \(R=Y/P\), and insert a smooth cofactor cutoff \(g(m/R)\) equal to one wherever the product cutoff can contribute. Separate the additive phase by the residue classes of \(p,m\) modulo \(q\), and Mellin-invert the product cutoff. Apart from the errors just bounded, \(\mathcal T(t)\) is a sum of at most polynomially many terms of the form \begin{equation}\label{eq:fixed-prime-cofactor} \frac{1}{2\pi V_g}\sum_{r,r'\bmod q}\ee{arr'/q} \int_{\mathbb R}\widetilde\psi(\nu)Y^{-i\nu} P_r(t+\sigma+\nu)R_{r'}(t+\sigma+\nu)\,d\nu, \end{equation} where harmless Mellin sign choices are fixed by this formula and \[ \begin{split} P_r(s)&=P^{-1}\sum_{\substack{p\equiv r\ (q)\\p\equiv c\ (L_0)}} w(p/P)p^{is},\\ R_{r'}(s)&=R^{-1}\sum_{m\equiv r'\ (q)} g(m/R)\gamma(m)m^{is}. \end{split} \] There is no unit restriction on \(r'\). For any interval of length at most \(4T_0\), \eqref{eq:fixed-dirichlet-mean} gives, uniformly in its location, \begin{equation}\label{eq:fixed-long-cofactor} \int_I|R_{r'}(t)|^2dt\ll L^C(1+T_0/R)\ll L^C. \end{equation} The final inequality uses the decisive scale separation \[ \frac{T_0}{R}\ll\frac{P L^{B_0+B_1}}{U_1U_2} \le\exp\bigl(CL^{1/4}-2L^{1/3}\bigr)L^{B_0+B_1}=o(1). \] It comes from the two independent big-prime products already present in the graph comparison, rather than a new energy bound. The Mellin transform in \eqref{eq:fixed-prime-cofactor} has Schwartz decay with polynomial-logarithmic seminorms. Choose a fixed exponent \(C_\nu\) large enough that its tail \(|\nu|>L^{C_\nu}\), in weighted \(L^1\), is smaller than the required logarithmic power. Uniformity in the location of \(I\) in \eqref{eq:fixed-long-cofactor}, and the trivial bound \(|P_r|\ll1\), justify removing this tail. Now choose \(B_2\) larger than \(C_\nu\), the bound for \(|\sigma|\), and the smooth-cutoff derivative costs. Throughout \(T_c<|t|\le T_0\), the argument \(s=t+\sigma+\nu\) then has \(|s|\gg L^{B_2}\). In \eqref{eq:fixed-high-prime} fix, for example, the integration-by-parts order \(j=2\) first. After choosing the requested exponent \(A\), enlarge \(B_2\) beyond the resulting fixed derivative costs. This gives \[ \sup_{T_c<|t|\le T_0,\ |\nu|\le L^{C_\nu}} |P_r(t+\sigma+\nu)|\ll_A L^{-A} \] for any requested \(A\). The smooth prime estimate applies because \(T_0=xL^{B_0+B_1}\le x^2\) for large \(x\). Choose this \(A\) after the fixed partition and Mellin costs. Combine the estimate with \eqref{eq:fixed-long-cofactor}, then apply Minkowski to the Mellin integral and sum the dyadic and residue pieces. This proves \[ \int_{T_c<|t|\le T_0}|\mathcal T(t)|^2dt\ll_D L^{-D}. \] Together with \eqref{eq:fixed-low} this is \eqref{eq:fixed-dirichlet-energy}; together with \eqref{eq:fixed-frequency-tail} and \eqref{eq:fixed-additive-mellin} it proves \eqref{eq:fixed-wide-energy}. The parameter order is explicit. First fix the graph comparison, its list bounds, and its external cutoff costs. Then fix the major arcs and the required energy saving, the tail cutoff \(B_1\), the bank smoothing precision, and the Mellin truncation. Choose the high-frequency boundary \(B_2\) after those choices. Finally choose the model depth through \textup{(C3)} at the required low-frequency precision. No exponent in the graph or the endpoint moment bound depends on that final depth. On the arcs, Cauchy--Schwarz, \eqref{eq:fixed-wide-energy}, and Parseval for the other endpoint bound the normalized Fourier comparison by every requested inverse power of \(L\), after paying the polynomially many arcs. Off the arcs, \eqref{eq:gr-multiplier} and the two endpoint norms do so. The normalization is \(1/Y\), exactly as in \eqref{eq:gr-fouriercomparison}, with \(k=h\). The harmonic conversion \(\Psi(t)=t\Phi(t)\) therefore gives the same centered banked replacement as for reflected sums. This completes the proof of Proposition~\ref{prop:fixed-comparison-energy} and supplies the fixed-shift branch used in Section~\ref{sec:assembly}. \subsection{Classical estimates and the height range}\label{sec:fixed-classical-boundary} The polynomial-logarithmic modulus here is an advantage over the reflected target grid, but the height is a separate constraint. Write \(Z=\exp(L^{1/16})\). The classical Dirichlet zero-free region, together with Siegel's theorem in its ineffective all-character form, gives \eqref{eq:fixed-smooth-primes} with error \[ O\!\left(P L^{C_w}e^{-cL^{1/16}}\right) \qquad(|t|\le 2Z, q\le L^B). \] For exact classical statements see \cite[Proposition 51, Theorem 62(ii) and Exercise 45(i)--(ii)]{taonotes}. Indeed, a rectangle of height \(3Z^2\) has a zero-free strip of width \(2\delta\), with \(\delta=cL^{-1/16}\). For a possible exceptional character, choose \(\varepsilon<1/(32(B+1))\) in Siegel's theorem; its zero then has distance \(\gg L^{-1/32}\) from one, larger than the strip width for sufficiently large \(x\). On the line \(\Re s=1-\delta\), the regularized logarithmic derivative has polynomial-logarithmic size. The same translated smooth Mellin contour as above, truncated at imaginary part \(Z^2\), therefore saves \(P^{-\delta}\le e^{-cL^{1/16}}\); fixed sufficiently high derivatives control the horizontal and truncation terms. Prime powers are smaller. Constants and the threshold in this classical estimate may be ineffective. The factor two absorbs the retained polynomial-logarithmic Mellin translations when the outer frequency has absolute value at most \(Z\). This covers the high-prime argument only up to \(Z\). Its required upper height is \(T_0=xL^{B_0+B_1}\). At that endpoint the classical strip has width of order \(L^{-1}\), whereas \(\log P\le CL^{1/4}\); the contour saving \(\exp(-c\log P/L)\) does not tend to zero. Thus a direct Siegel--Walfisz substitution does not prove the remaining weighted prime--cofactor energy on \(Z<|t|\le T_0\). Walsh's classical-recovery remark \cite[p.~2]{walsh} motivates a different high-frequency argument, but the proof here retains the zero-free input on that range. Such a replacement for fixed shifts would not remove the target-sized conductor issue for reflected sums.