\section{From a certified continuum function to sequences}\label{sec:lift} This section proves that the continuum statement of hypothesis (2) below, together with the non-vanishing of the diagonals and of $T_0(g)$, gives the theorem. The argument is FM's construction procedure, the proof of \cite[Theorem~6.3]{FM}, with their mask, extended by a re-solve of the forced functions after masking and by a reduction of the twisted sums to the rows of \S\ref{sec:construction}. \begin{theorem}\label{thm:lift} Let $0 < \eta < \alpha' < 1/3$ with $\eta \le 1/4$, let $m^* = \max\{j : j\eta < \alpha'\}$, and let free data of the form \eqref{eq:free} be given, with the $h_j$ forced as in \S\ref{sec:construction} and the constants $c^* = c + t^* g$ solving $T_0 = 0$. Suppose: \begin{enumerate} \item $|B_j(\sigma)| \ge \beta_0 > 0$ for $1 \le j \le m^*$ and $\sigma \in [j\eta, \alpha']$, and $|T_0(g)| \ge \tau_0 > 0$; \item $f(1) < 0$, and for some $m > 0$ the inequality $f(u) \ge (1 - m) f(1)$ of \eqref{eq:target} holds at every non-prime $\eta$-rough vector $u$ with no subsum equal to $1/2$ (such $u$ differ from $E$); and $|f| \le C$ on all vectors. \end{enumerate} Then for every $0 < \nu < \eta$ and every $0 \le \alpha < \alpha'$ there are $\eps > 0$ and a bounded function $F_\eps$ on $\eta$-rough vectors, both independent of $B$ and $x$, with $F_\eps(1) = -1$ and $F_\eps \ge -1$, such that the following holds. For $x \ge 2$ let $a_n = a^{(x)}_n := 1 + F_\eps(v(n))$ for squarefree $n \in (x/2, x]$ whose prime factors are all at least $n^\eta$, $a_n := 1$ for the other $n \in (x/2, x]$ and $a_n := 0$ for $n \notin (x/2, x]$, and let $b_n := \mathbf 1_{(x/2, x]}(n)$. Then $0 \le a_n \le 1 + \sup|F_\eps|$ for every $x$, and $a_p = 0$ for every prime $p \in (x/2, x]$. For every $B > 0$ there is $x_0$, depending on $B$ and on the data above, such that for every $x \ge x_0$ the sequence $w_n = a_n - b_n$ satisfies (I) at $\gamma = 1/2$, (II) on $1 < \ell \le x^\nu$, (w) and $(\Hoo_\alpha)$. In particular $C^-(1/2, 0, \nu \mid \Hoo_\alpha) = 0$. \end{theorem} \subsection{Setup and the mask} Fix $\alpha < \alpha'$ and $0 < \nu < \eta$. Fix $\eps > 0$ with $\nu + \eps < \eta$ and $\eps < 1/2 - \alpha'$, and so small that the $O(\eps)$ bounds of Steps~1--3 give $|B_{j,\eps}| \ge \beta_0/2$ at every level, $|T_{0,\eps}(g)| \ge \tau_0/2$, $f_\eps(1) < 0$ and the margin inequality of Step~3. These thresholds depend only on $\eta$, $\alpha'$, the free data, $g$, $C$, $\beta_0$, $\tau_0$, $m$ and $f(1)$, not on $x$; $\eps$ is fixed before $x \to \infty$. Put $I_\eps = [\nu, \nu + \eps] \cup [1/2 - \eps, 1/2 + \eps]$ and $\chi_\eps(u) = \mathbf 1[u$ has no subsum in $I_\eps]$. In Lemma~\ref{lem:L1} implied constants depend only on $\eta$ and $\sup|g_0|$ \cite[(6.6)]{FM}; in Steps~1--3 they depend on $\eta$, $\alpha'$, $C$, the free data ($c^*$, the weights $w_k$, the $h_j$), $g$, $\beta_0$ and $\tau_0$, but not on $\eps$, $x$ or the tuple; in Step~4 and Lemma~\ref{lem:L3} they also depend on $\eps$, through the bound and the Lipschitz data of $F_\eps$. \begin{lemma}[Masking costs $O(\eps)$, uniformly]\label{lem:L1} Let $g_0$ be bounded free data, piecewise Lipschitz on finitely many polytopes, let $g$ be its extension by (6.3), and let $g_\eps$ be the extension of $g_0 \chi_\eps$. Then: \begin{enumerate} \item[(a)] $g_\eps$ satisfies (6.2), the continuum Type I relation at level $1/2$ \cite[Theorem~6.4]{FM}; \item[(b)] $g_\eps(u) = 0$ whenever $u$ has a subsum in $I_\eps$, and $|g(u) - g_\eps(u)| \ll \eps$ otherwise \cite[(6.6)]{FM}; \item[(c)] for every row functional $R$ (the functional $T_0$, and the twisted and untwisted functionals at every tuple $s$ of $t \le m^*$ small components with $|s| \le \alpha'$), $|R(g) - R(g_\eps)| \ll \eps$, uniformly in the tuple. \end{enumerate} \end{lemma} \begin{proof} (a) is \cite[Theorem~6.4]{FM}; (b) is \cite[(6.6)]{FM} and the remark preceding it in \cite[\S6.2]{FM}. For (c), each row is a finite sum over $\eta$-rough ordered completions $z$ of the tuple $s$ to total $1$, with uniformly bounded densities and total masses. The half $[\nu, \nu + \eps]$ of $I_\eps$ is vacuous, since the subsums of an $\eta$-rough vector are $0$ or at least $\eta > \nu + \eps$. A subsum that contains no completion coordinate is at most $|s| \le \alpha'$, and one that contains all of them is at least $1 - |s| \ge 1 - \alpha'$. Since $\alpha' < 1/2 - \eps$, neither meets $[1/2 - \eps, 1/2 + \eps]$. Every other subsum contains some, but not all, completion coordinates. Holding the other coordinates fixed, vary one included and one excluded completion coordinate with their sum fixed: the subsum moves with slope $1$, so the set of such variations on which it lies in $I_\eps$ has length $O(\eps)$. Summing over the finitely many index sets, the set $U_\eps$ of completions with a subsum in $I_\eps$ has measure $O(\eps)$, uniformly in the tuple. For completions with one component, $U_\eps$ is empty; this includes the prime vector in $T_0$. On $U_\eps$ the masked extension vanishes and the unmasked one is bounded, and off $U_\eps$ part (b) applies. This proves (c). \end{proof} \subsection{Step 1: the masked forced functions} Unmasked, the rows see the small part only through $(j, \tau)$. The mask sees subsums such as $s_i + r$ separately, so the masked rows depend on the tuple. FM allow free data that are bounded and Lipschitz on finitely many polytopes (Lemma~\ref{lem:L2}), so we let the forced functions depend on the tuple: \[ f_{0,\eps}(u) := c^*[j, k, b]\,\chi_\eps(u) + \mathbf 1\bigl[1 \le j \le m^*,\ s_1 + \dots + s_j \le \alpha'\bigr]\, w_k\, h_{j,\eps}(s_1, \dots, s_j)\,\chi_\eps(u), \] where $s_1, \dots, s_j$ are the small components of $u$. The masked coefficients are the actual tuple rows of the (6.3) extensions of the masked basis data, with the mask kept inside every fragmentation integral. They are invariant under permutations of the tuple, so $h_{j,\eps}$ is symmetric and the free data are well defined on multisets. Solve for $j = m^*, m^* - 1, \dots, 1$ in turn. In the masked row $T_{j,\eps}(s_1, \dots, s_j)$, the function $h_{j,\eps}$ enters only through free data whose small part is exactly the tuple, and whose rough part, counted after fully fragmenting any large component, has $k$ parts with $w_k \ne 0$; its coefficient is the masked diagonal $B_{j,\eps}(s) := \sum_k w_k B^{(k)}_{j,\eps}(s)$, where $B^{(k)}_{j,\eps}(s)$ is the contribution of the completions whose rough part has $k$ parts. The functions $h_{J,\eps}$ with $J > j$ enter through completions adding exactly $J - j$ small components, with the total of all $J$ small components at most $\alpha'$ (as free components, inside the remaining free part, or as fragments of the large part), and they are already defined. The functions with $J < j$ do not enter. A vector with $m^* + 1$ or more small components carries no $h$. So \[ T_{j,\eps} = A_{j,\eps}\cdot c^* + B_{j,\eps}\, h_{j,\eps} + W^{(j)}_\eps[h_{j+1,\eps}, \dots, h_{m^*,\eps}]. \] By Lemma~\ref{lem:L1}(c) and induction on the solved levels, $|B_{j,\eps}(s) - B_j(|s|)| \ll \eps$, $|A_{j,\eps} - A_j| \ll \eps$ and $|W^{(j)}_\eps[h_{j+1,\eps}, \dots, h_{m^*,\eps}](s) - W^{(j)}[h_{j+1}, \dots, h_{m^*}](|s|)| \ll \eps$, uniformly in the tuple. Since $|B_j| \ge \beta_0$, for small $\eps$ \[ h_{j,\eps} := -\bigl(A_{j,\eps}\cdot c^* + W^{(j)}_\eps\bigr)/B_{j,\eps} \] is well defined, $|h_{j,\eps}(s) - h_j(|s|)| \ll \eps$ uniformly, and $T_{j,\eps} \equiv 0$ at every $j$-tuple with total $\le \alpha'$. Because the masked data are defined with the same weights $w_k$, each masked row has a single unknown, whose coefficient is a fixed combination of masked diagonals; no consistency condition between the parts with different $k$ arises. \subsection{Step 2: parity} The direction $g$ has no small component, so $A_j \cdot g = 0$ for $j \ge 1$, masked or not, and moving along $c^* + t g$ leaves every $T_{j,\eps}$ unchanged. Now $T_{0,\eps}(c^*) = O(\eps)$ and $T_{0,\eps}(g) = T_0(g) + O(\eps)$ with $|T_0(g)| \ge \tau_0$. So $t_\eps := -T_{0,\eps}(c^*)/T_{0,\eps}(g) = O(\eps)$, and $c_\eps := c^* + t_\eps g$ gives $T_{0,\eps} = 0$ and $T_{j,\eps} \equiv 0$ for every $1 \le j \le m^*$. Let $f_\eps$ be the resulting masked extension. \subsection{Step 3: positivity and normalisation} On vectors with a subsum in $I_\eps$ the masked extension is zero. This includes $E$, since every fragmentation of $E$ has a subsum equal to $1/2$, and $E$ is the only vector with two components $\ge 1/2$. On every other vector, the boundedness of (6.3), Lemma~\ref{lem:L1}(b) and the $O(\eps)$ bounds above give $f_\eps = f + O(\eps)$ uniformly. The prime vector $(1)$ is unmasked, so $f_\eps(1) = f(1) + O(\eps) < 0$, and every unmasked non-prime vector has no subsum in $[1/2 - \eps, 1/2 + \eps]$, in particular none equal to $1/2$, so it differs from $E$ and has the margin of hypothesis (2). We may assume $m \le 1$, since \eqref{eq:target} for $m$ implies it for every smaller $m > 0$. So for small $\eps$, $f_\eps(u) \ge (1 - m/2)f_\eps(1)$ on every non-prime vector: on the unmasked ones by that margin, and on the masked ones because there $f_\eps = 0$ while $(1 - m/2)f_\eps(1) < 0$. Put $F_\eps := f_\eps/|f_\eps(1)|$. Then $F_\eps(1) = -1$, $F_\eps \ge -1 + m/2$ on every other vector, $|F_\eps| \le C'$, $T_0(F_\eps) = 0$, and $T_t(F_\eps)(s) = 0$ at every tuple of $t \le m^*$ small components with $|s| \le \alpha'$; and $F_\eps$ satisfies (6.2) by Lemma~\ref{lem:L1}(a), whose appeal to \cite[Theorem~6.4]{FM} uses axiom (b) of \cite[Definition~6.2]{FM}, supplied by Lemma~\ref{lem:L2}. \begin{lemma}[Regularity]\label{lem:L2} $F_\eps$ is symmetric and satisfies FM's Definition~6.2(a, b): in each dimension it is a finite sum of functions supported and Lipschitz on convex polytopes. \end{lemma} \begin{proof} Dimensions are bounded by $1/\eta$. The class, bin and profile conditions (the number of rough parts, the bin of the largest one), the ordering conditions, the mask and the partition inequalities defining $L_{1/2}$ are finitely many affine conditions. For each level in turn, expand the finite fragmentation and Lemma~\ref{lem:K} subset sums that define $A_{j,\eps}$, $B_{j,\eps}$ and $W^{(j)}_\eps$, keeping $\chi_\eps$. The domains are polytopes with fixed normals and affine dependence on the tuple, every coordinate that appears in a denominator is at least $\eta$, and the integrals of bounded integrands over polytopes with fixed normals moving affinely with the tuple are Lipschitz, with chambers created only by point evaluations and by the affine walls of the integrand. $B_{j,\eps}$ is bounded away from zero, so $h_{j,\eps}$ is piecewise Lipschitz; refine the chambers before the next level. The extension (6.3) and the normalisation preserve this. \end{proof} \subsection{Step 4: the sequences} Put $w_n := F_\eps(v(n))$ for squarefree $n \in (x/2, x]$ whose prime factors are all at least $n^\eta$ and $w_n := 0$ for the other $n \in (x/2, x]$, and let $a_n := 1 + w_n \ge 0$ and $b_n := 1$ on $(x/2, x]$, with $a_n = b_n = 0$ elsewhere, as in Theorem~\ref{thm:lift}. Then $a_p = 1 + F_\eps(1) = 0$ for every prime $p \in (x/2, x]$. (I) at $\gamma = 1/2$ follows from the argument for \cite[(6.7)--(6.9)]{FM} in the proof of \cite[Theorem~6.3]{FM}, applied to the already masked $F_\eps$ (Lemma~\ref{lem:L2}); the strip $[1/2 - \eps, 1/2 + \eps]$ supplies their case $m > x^{\gamma - \eps}$, and no second mask is applied. (II) on $1 < \ell \le x^\nu$ holds because every perturbation term vanishes: such an $\ell$ has a prime factor at most $x^\nu$, and $x^\nu < (x/2)^\eta$ for large $x$, so $\ell$ divides no element of the support of $w$. (w) holds since $a_n$ is bounded uniformly in $x$. For $(\Hoo_\alpha)$, which is imposed on $w = a - b$, we bound $\sum_{d \mid n} \mu(n) w_n$ for $d \le x^\alpha$ (by \S\ref{sec:setting}, Hooley's form for $a_n$ then follows as well). A non-squarefree $d$ contributes $0$. If $d$ has a prime factor below $(x/2)^\eta$, the sum is $0$. If $d$ has $m^* + 1$ or more prime factors $\ge (x/2)^\eta$, then $d \ge (x/2)^{(m^*+1)\eta} > x^\alpha$ for large $x$, since $(m^* + 1)\eta \ge \alpha' > \alpha$. For $d = 1$, the proof of Lemma~\ref{lem:L3} with $t = 0$ ($\lambda = 0$, $q = w$ and $\int_{1 - \log 2/\log x}^{1} x^w\,dw/w = \Li(x) - \Li(x/2)$) gives $\sum \mu(n) w_n = (\Li(x) - \Li(x/2))\, T_0(F_\eps) + O(x e^{-(\log x)^{1/5}}) = O(x e^{-(\log x)^{1/5}})$. For $d = p_1 \cdots p_t$ with $1 \le t \le m^*$, every $p_i \ge (x/2)^\eta$ and $d \le x^\alpha$, Lemma~\ref{lem:L3} gives $\ll (x/d)e^{-(\log x)^{1/5}}$. Summing, \[ \sum_{d \le x^\alpha} \Bigl|\sum_{d \mid n} \mu(n) w_n\Bigr| \ll x e^{-(\log x)^{1/5}} \sum_{t=0}^{m^*} \Bigl(\sum_{p \le x} \frac1p\Bigr)^t \ll x e^{-(\log x)^{1/5}} (\log\log x)^{m^*} \ll_B x (\log x)^{-B}. \] \begin{lemma}[Twisted sums at products of $t$ primes]\label{lem:L3} Let $1 \le t \le m^*$. Uniformly over primes $p_1 < \dots < p_t$ with $p_1 \ge (x/2)^\eta$ and $d = p_1 \cdots p_t \le x^\alpha$, \[ S^\mu_d := \sum_{\substack{x/2 < n \le x\\ d \mid n}} \mu(n) w_n \ll_\eps \frac{x}{d}\, e^{-(\log x)^{1/5}} . \] \end{lemma} \begin{proof} We have $S^\mu_d = \mu(d) \sum_r \mu(r) w_{dr} = (-1)^t \sum_r \mu(r) w_{dr}$, over squarefree $r$ coprime to $d$ with $x/(2d) < r \le x/d$. This is FM's sum $S_m(c, d')$ of \cite[(6.8)]{FM} with their $m$ equal to our $d$, $\lambda_i = \log p_i/\log x$, $\lambda = \sum \lambda_i \le \alpha < 1/2 - \eps$ and $(c, d') = (1/2, 1)$, except for the factor $(-1)^t\mu(r)$; FM's (6.8) holds for every $m \le x^{1/2 - \eps}$, whatever its number of prime factors. FM's pieces fix $k = \omega(dr)$, the polytope piece of $F_\eps$ and the meshing; in each piece $\omega(r) = k - t$, so $\mu(r) = (-1)^{k-t}$. FM's Lemma~5.11, applied to each piece (normalised by a constant $K_\eps \ge 1$ controlling the sup norm and the Lipschitz constant, then multiplied back; its hypotheses hold, since $\eta \le 1/4$ and $d$ is squarefree with $d \le x^\alpha < x^{1-\eta}$), gives \[ S^\mu_d = \sum_{k \ge t + 1} (-1)^k\, \mathrm{main}_k + O_\eps\bigl((x/d)\, e^{-(\log x)^{1/5}}\bigr), \] where $\mathrm{main}_k$ is FM's untwisted main term in total dimension $k$ and the sign includes $\mu(d) = (-1)^t$. Their sum is the integral \[ \int_{1 - \lambda + \log(1/2)/\log x}^{1 - \lambda} x^w \sum_{k \ge t+1} (-1)^k \int_{\substack{u_{t+1} + \dots + u_k = w\\ u_{t+1} < \dots < u_k}} \frac{F_\eps\bigl(\lambda_1/q, \dots, \lambda_t/q, u_{t+1}/q, \dots, u_k/q\bigr)}{u_{t+1} \cdots u_k}\, d\boldsymbol u\, dw, \] with $q = \lambda + w$. Rescale $u_i = q z_i$: the projection measure contributes $q^{k-t-1}$ and the denominator $q^{-(k-t)}$, so the inner sum is exactly $q^{-1}\, T_t(F_\eps)(\lambda_1/q, \dots, \lambda_t/q)$, with no factorial or logarithmic factor. If some $\lambda_i/q < \eta$, the integrand vanishes, since $F_\eps$ is supported on $\eta$-rough vectors. Otherwise $\sum_i \lambda_i/q = \lambda/(\lambda + w) \le \alpha(1 + o(1)) < \alpha'$ for large $x$, and $T_t(F_\eps) = 0$ at that tuple by Step~3. Uniformity in $(p_1, \dots, p_t)$ is that of Lemma~5.11. \end{proof} This proves Theorem~\ref{thm:lift}. \qed