\subsection{Self-sums with prime support in a fixed coset}\label{sec:cosets} The comparison theorem also covers a prime coset which contains no identity class. The additional arithmetic issue is that its character twists can have negative or genuinely complex singular exponents. Those branches enter the model used for centering, even though they are smaller than its positive leading part. Fix an even integer \(m\), put \(G=(\Z/m\Z)^\times\), and let \(K\le G\), \(a\in G\). Write \[ \mathscr A=aK,\qquad H=\langle K,a\rangle,\qquad r=[G:K],\qquad d=\operatorname{ord}_{G/K}(aK),\qquad \rho=1/r. \] Define \[ \mathcal P_{\mathscr A}=\{p:p\nmid m,\ p\bmod m\in\mathscr A\},\qquad C_{\mathscr A}=\langle\mathcal P_{\mathscr A}\rangle,\qquad f=\1_{C_{\mathscr A}}. \] Thus \(1\in C_{\mathscr A}\). The generated residue group is \(H\), whereas the prime density is \(\rho\); these are different data. Put \begin{align} c_{\mathscr A}&=\frac1{\Gamma(\rho)}\lim_{s\downarrow1}(s-1)^\rho \prod_{p\in\mathcal P_{\mathscr A}}(1-p^{-s})^{-1}, \label{eq:coset-marginal}\\ \Gamma_{\mathscr A}(N)&=\frac{m}{|H|^2} \#\{(u,v)\in H^2:u+v\equiv N\pmod m\},\notag\\ \mathcal E_{\mathscr A}&=\{p:p\nmid m,\ p\bmod m\notin\mathscr A\},\notag\\ C_{\mathcal E_{\mathscr A}}&=\prod_{p\in\mathcal E_{\mathscr A}} \left(1-\frac1{(p-1)^2}\right),\qquad J_{\mathscr A}(N)=\prod_{\substack{p\mid N\\p\in\mathcal E_{\mathscr A}}} \frac{p-1}{p-2}.\notag \end{align} The constant \(c_{\mathscr A}\) is positive; an explicit expression is given below. For \(d>1\), set \[ g_{\mathscr A}=\rho\bigl(1-\cos(2\pi/d)\bigr),\qquad \delta_{\mathscr A}=\min\left\{\frac1{8192\varphi(m)},\frac{3g_{\mathscr A}}{10}\right\}. \] For \(d=1\), use \(g_{\mathscr A}=+\infty\) and \(\delta_{\mathscr A}=1/(8192\varphi(m))\). \begin{theorem}\label{thm:coset} Under the analytic inputs of Section~\ref{sec:inputs}, uniformly in all sufficiently large targets with \(\Gamma_{\mathscr A}(N)>0\), \[ \#\{(x,y)\in C_{\mathscr A}^2:x+y=N\} =c_{\mathscr A}^2\Gamma_{\mathscr A}(N)C_{\mathcal E_{\mathscr A}}J_{\mathscr A}(N) \frac{N}{(\log N)^{2-2\rho}} \left(1+O((\log N)^{-\delta_{\mathscr A}})\right). \] If \(\Gamma_{\mathscr A}(N)=0\), the count is identically zero. The condition \(\Gamma_{\mathscr A}(N)>0\) is equivalent to solvability in every finite quotient. Consequently it characterizes eventual representability. The modulus and coset remain fixed, and the error is uniform in all target prime valuations. \end{theorem} The exponent incorporates both discarded banks and the gap between the leading and secondary character branches. In particular a fixed positive exponent works for every fixed coset. For \(d\ge2\), \(1-\cos(2\pi/d)=2\sin^2(\pi/d)\ge8/d^2\), so one may always replace \(\delta_{\mathscr A}\) by the simpler, weaker \(1/(8192\varphi(m)^3)\). There is no uniformity in growing \(m\). The new assertion concerns unfiltered self-sums. A filter on the whole numerical value need not be preserved by multiplication by a prime in a nonidentity coset. \subsubsection{Every fixed-character branch} For a character \(\psi\), define \(F_\psi(s)=\sum_n\psi(n)f(n)n^{-s}\), with all characters inflated to a common modulus containing \(m\). Fourier expansion of the prime coset gives \begin{equation}\label{eq:coset-character-factorization} F_\psi(s)=G_\psi(s) \prod_{\chi\in K^\perp}L(s,\psi\chi)^{\rho\overline{\chi(a)}}. \end{equation} The correction is uniformly absolutely convergent for \(\Re s>1/2\). Indeed the coefficient of \(p^{-s}\) in the logarithm is \[ \rho\psi(p)\sum_{\chi\in K^\perp}\overline{\chi(a)}\chi(p) =\psi(p)\1_{p\bmod m\in\mathscr A}; \] the remaining logarithmic terms are bounded by a fixed multiple of \(\sum_p p^{-2\Re s}\). Literal Euler factors include all conductor deletions. Analytic logarithms continued from the Euler half-plane specify all complex powers. A zeta factor occurs precisely for a fixed character \(\psi\in K^\perp\); its exponent is \begin{equation}\label{eq:coset-exponents} \alpha_\psi=\rho\psi(a). \end{equation} All other twists have no singular branch at one. The exponents \(\rho\) occur exactly for \(\psi\in H^\perp\). The remaining exponents are \(\rho\) times nontrivial \(d\)-th roots of unity, each with the same multiplicity. In particular, \(\rho-\Re\alpha_\psi\ge g_{\mathscr A}\) outside \(H^\perp\). For the untwisted series, an explicit positive correction is \[ \log G_{1,p}(s)=\sum_{j\ge2} \frac{\1_{p\bmod m\in\mathscr A}-\1_{(p\bmod m)^j\in\mathscr A}}{j p^{js}} \qquad(p\nmid m). \] Thus \begin{equation}\label{eq:coset-constant} c_{\mathscr A}=\frac{(\varphi(m)/m)^\rho G_1(1)}{\Gamma(\rho)} \prod_{\substack{\chi\in K^\perp\\\chi\ne1}} L(1,\chi)^{\rho\overline{\chi(a)}}. \end{equation} Here \(G_1(1)\) is an absolutely convergent positive Euler product. Conjugate characters are paired. Continuation along \(s>1\) makes the expression positive and unambiguous. The ordinary one-variable asymptotic, and its complex twisted versions, also follow from \cite[Theorem 1.2 and (1.9)--(1.10)]{delabretechetenenbaum}. That fixed-function theorem is not used in place of the moving-cutoff estimate which follows. \subsubsection{The arithmetic model and its one-variable means} If \(K=G\), then \(f\) is the indicator of integers coprime to \(m\), and we use the exact model \(\mathcal D=f\). Counting the finite residue classes on the line proves Theorem~\ref{thm:coset} with \(O_m(1)\) error. In the model construction below assume \(K\ne G\), so \(\rho<1\) and the model upper bound tends to zero. Keep an outer scale \(X\), with \(L=\log X\), and let \[ y=e^{L^{2/5}},\quad \mathcal P_y=\{p:p\le y\},\quad V_y=\prod_{p\le y}(1-1/p),\quad E=e^{\sqrt L}. \] For \(n=A v\), where \(A=n_{\mathcal P_y}\), the integer \(v\) is \(\mathcal P_y\)-rough. The outer scale and cutoff are held fixed under multiplication by bank primes. For \(\psi\in K^\perp\), write \[ F_{\psi,R}(s)=F_\psi(s) \prod_{\substack{p\le y\\p\in\mathcal P_{\mathscr A}}} (1-\psi(p)p^{-s}) =(s-1)^{-\alpha_\psi}H_{\psi,R}(s). \] Since \(\psi(p)=\psi(a)\) on the allowed prime set, fixed-character Mertens and a disk of radius \(c/\log y\) give \begin{equation}\label{eq:coset-rough-regular} |H_{\psi,R}(1)|\asymp(\log y)^{-\Re\alpha_\psi},\qquad \frac{|H_{\psi,R}^{(j)}(1)|}{j!} \ll_j|H_{\psi,R}(1)|(C\log y)^j. \end{equation} For the derivative bound, the logarithm of each finite Euler product relative to its value at one is \(O(|s-1|\sum_{p\le y}(\log p)/p)=O(1)\) on that disk; the fixed regular factor is analytic and nonzero there. Choose a sufficiently large fixed depth \(J\), and put \begin{equation}\label{eq:coset-hankel} M_\psi(t)=\sum_{j=0}^J\frac{H_{\psi,R}^{(j)}(1)}{j!} \frac{(\log t)^{\alpha_\psi-j-1}} {\Gamma(\alpha_\psi-j)}. \end{equation} Write \(M_0=M_1\) for the positive principal branch. On any fixed power range, \(M_0(t)>0\), and \begin{equation}\label{eq:coset-secondary-gap} |M_\psi(t)|\ll L^{-3g_{\mathscr A}/5}M_0(t) \qquad(\psi\notin H^\perp). \end{equation} In fact its individual exponent is \(3(\rho-\Re\alpha_\psi)/5\): the factors \(L\) and \(\log y\) in \eqref{eq:coset-rough-regular} and \eqref{eq:coset-hankel} give exactly this ratio. Conjugate branches are conjugate functions. Every \(\eta\in H^\perp\) is one on the support of \(f\), so \(F_{\psi\eta,R}=F_{\psi,R}\) and \(M_{\psi\eta}=M_\psi\) exactly. Define \begin{equation}\label{eq:coset-model} \mathcal D_J(n)=\frac{f(A)}{V_y} \sum_{\psi\in K^\perp}\overline{\psi(v)}M_\psi(v), \qquad \xi_J(n)=f(n)-\mathcal D_J(n). \end{equation} Grouping characters modulo \(H^\perp\) factors the bracket by \(|G/H|\1_{v\bmod m\in H}\). Its leading term is positive and all other terms satisfy \eqref{eq:coset-secondary-gap}. Therefore \(\mathcal D_J\) is real and nonnegative for large \(X\), and \begin{equation}\label{eq:coset-model-domination} \mathcal D_{0,J}(n)=|G/H|\1_{v\bmod m\in H}\frac{f(A)M_0(v)}{V_y}, \qquad |\mathcal D_J-\mathcal D_{0,J}|\ll L^{-3g_{\mathscr A}/5}\mathcal D_{0,J}. \end{equation} For \(d=1\), the difference is zero. For a precise extension, use the cutoff from Section~\ref{sec:model-preparation}, immediately after \eqref{eq:hankel-model}, with the present outer scale \(X\): set every \(M_\psi(t)=0\) for \(t\le X^{1/4}\), and multiply \eqref{eq:coset-hankel} by \(\eta_0(\log t/L)\) above that point. It equals the displayed formula for \(t\ge X^{1/3}\). The common real cutoff preserves conjugacy, \(M_{\psi\eta}=M_\psi\), positivity of the combined model, and \eqref{eq:coset-secondary-gap} on the transition range. Condition (C3) only uses \(t\gg X^{1/2}/E>X^{1/3}\). Consequently, on \(1\le n\le X^3\), \[ 0\le \mathcal D_J(n)\ll L^{-3(1-\rho)/5},\qquad |\xi_J(n)|\ll1. \] These exponents are independent of the fixed depth. For an allowed \(p\le y\), multiplication changes \(A\) to \(pA\) and leaves \(v\) unchanged. Hence, exactly, \begin{equation}\label{eq:coset-invariance} f(pn)=f(n),\qquad \mathcal D_J(pn)=\mathcal D_J(n),\qquad \xi_J(pn)=\xi_J(n). \end{equation} The characters in \eqref{eq:coset-model} are evaluated at the rough part \(v\), rather than at the whole integer \(n\). \begin{lemma}\label{lem:coset-centering} For every fixed \(B,A_0>0\), a sufficiently large fixed depth \(J\) gives the recordwise estimate \eqref{eq:comparison-rough-means} for \(\xi_J\). It holds uniformly for every \(\mathcal P_y\)-supported record \(r_0\le E\), every \(Y\in[X^{1/2},X^2]\), bounded periodic tests of period at most \(L^B\), and the smooth weights permitted there. In particular it includes every fixed-character contribution. \end{lemma} \begin{proof} First consider an \(\mathcal P_y\)-rough variable at scale \(T\in[X^{2/5},X^2]\). Expand a periodic test in characters after refining its period to \(Q=\operatorname{lcm}(q,m)\). All primes dividing \(Q\) lie below \(y\). For a character \(\vartheta\bmod Q\), \eqref{eq:coset-character-factorization} holds literally, with every small allowed Euler factor removed. A principal constituent occurs only if \(\vartheta\) is inflated from \(K^\perp\). In that case the rough series is precisely \(F_{\vartheta,R}\) defined above; it does not acquire a new conductor-dependent regular part. Complex powers require a bound on the analytic logarithm, rather than merely on the modulus of an \(L\)-function. For a principal character put \(L^\#(s)=(s-1)L(s)/(s+1)\), and for a nonprincipal character put \(L^\#=L\). The zero-free input makes \(L^\#\) analytic and nonzero on \(\Re s>11/12\). Its logarithm is fixed by the Euler logarithm on \(\Re s=2\), together with the displayed pole-removing factor. On a fixed disk centered at \(2+it\), contained in that half-plane, convexity gives \(\Re\log L^\#\ll\log(Q(2+|t|))\), while the center logarithm is bounded by the same quantity. Borel--Carath\'eodory on a smaller fixed disk therefore gives \begin{equation}\label{eq:coset-analytic-log} |\log L^\#(\sigma+it)|\ll\log(Q(2+|t|)) \qquad(23/24\le\sigma\le2). \end{equation} The disks can have fixed radii because their left boundary remains strictly to the right of \(11/12\). The consistent center logarithm is uniformly controlled by its Euler series. Thus every fixed complex power in \eqref{eq:coset-character-factorization} has polynomial growth in conductor and height; its imaginary exponent causes no uncontrolled argument factor. Use smooth Mellin inversion to height \(\exp(L^{1/10})\), and move the line to \(1-c/\log y\). This lies inside the preceding strip for large \(X\). The finite Euler factors cost only \(L^{O(1)}\), and \eqref{eq:coset-analytic-log} costs \(\exp(O(L^{1/10}))\). The factor \(T^{s-1}\) supplies \(\exp(-cL^{3/5})\), which dominates both. Sufficiently many fixed derivatives of the smooth weight dispose of the Mellin tails. A twist without a principal constituent therefore has mean \(O_A(TL^{-A})\) for every fixed \(A\). For \(\vartheta\in K^\perp\), Hankel integration around one gives exactly the density \(M_\vartheta\) in \eqref{eq:coset-hankel}. The derivative bound \eqref{eq:coset-rough-regular} gives relative Taylor remainder \(O_J(L^{-3(J+1)/5})\) for that branch. All branch magnitudes are at most a fixed multiple of the principal magnitude. This argument applies without change to negative and complex \(\alpha_\vartheta\); the reciprocal Gamma factor is the Hankel integral of \((s-1)^{-\alpha_\vartheta+j}\). On the model side the unrestricted rough population has series \(\zeta(s)\prod_{p\le y}(1-p^{-s})\), principal density \(V_y\), and no other principal branch. Its twist by \(\vartheta\overline\psi\) is principal exactly when \(\vartheta=\psi\) on the units modulo \(Q\). Hence \(V_y^{-1}\sum_{\psi\in K^\perp}\overline\psi(v)M_\psi(v)\) matches each of the original character means with its exact coefficient. This proves arbitrary logarithmic centering after choosing \(J\) beyond the periodic-character and smooth-derivative costs. A factor \(v^{iu}\), \(|u|\le L^B\), only enlarges those fixed costs. Finally, for a fixed record \(r_0\le E\), multiplicativity gives \(f(r_0v)=f(r_0)f(v)\), and \eqref{eq:coset-model} has the same factor \(f(r_0)\). The remaining scale is \(T=Y/r_0\), within the preceding power range for large \(X\). If the record is forbidden both terms vanish. Otherwise the just-proved estimate is \(O((Y/r_0)L^{-A_0})\), uniformly in the record and test. This is the required recordwise assertion, with the depth chosen after the requested accuracy. \end{proof} \subsubsection{Common banks and the independent sieve check} Choose one residue \(a_0\in aK\) and take all small and big banks from primes \(p\equiv a_0\pmod m\), deleting \(p\mid N\). Their fixed density is \(1/\varphi(m)\), irrespective of \(\rho\). Equation \eqref{eq:coset-invariance} verifies condition (C2) of Theorem~\ref{thm:model-comparison}; the bounds above verify (C1), and Lemma~\ref{lem:coset-centering} verifies (C3). Apply Lemma~\ref{lem:gr-ideal} with \(L_0=m\) and \(c=a_0\). Its quotient character frame already retains the common unit phase on products with the same slot type. Thus these banks require no new graph construction or mark-length loss. We verify the remaining envelope assumption independently of the comparison. Put \(\kappa_{\mathscr A}=2-2\rho\) and \(\Sigma_N=NJ_{\mathscr A}(N)L^{-\kappa_{\mathscr A}}\). This satisfies \eqref{eq:comparison-scale}, for example with \(\kappa_0=3\), since \(0\le\kappa_{\mathscr A}<2\), \(J_{\mathscr A}(N)\ge1\), and the prime-product bound \(J_{\mathscr A}(N)\ll(\log\log(3N))^2\) is uniform in \(N\). At a bank prime \(p\nmid N\), the exact two-coordinate valuation law is \eqref{eq:bank-law}. For coprime bank records \(r_1,r_2\le E\), their divisibilities select one interval progression of length \(|\mathcal I|/(r_1r_2)\) on \(x+y=N\). Exactness excludes two roots at every unused bank prime, and one remaining root at a prime dividing \(r_1r_2\). Together with \(1/(r_1r_2)\), these factors are exactly its probability under the independent bank law. Shared bank-prime divisibility is impossible. Use the upper Selberg sieve at \(z=N^{1/10}\), with primes up to \(w=z^{1/16}\). This contains all banks. Every good forbidden prime has two roots, merging precisely when it divides \(N\). Fixed-coset prime Mertens gives the following lower bounds for the nonbank sieve denominator, up to fixed positive constants: \[ \begin{array}{c|c} \text{weights}&\text{sieve denominator}\\ \hline f,f&(\log w)^{2(1-\rho)}/J_{\mathscr A}(N),\\ \mathcal D_J,f&(\log y)^{1-\rho}(\log w)^{1-\rho}/J_{\mathscr A}(N),\\ \mathcal D_J,\mathcal D_J&(\log y)^{2(1-\rho)}/J_{\mathscr A}(N). \end{array} \] For a model factor use its nonnegative domination by \(2D_{0,J}\), its bound \(L^{-3(1-\rho)/5}\), and its exclusion of every forbidden prime at most \(y\). These bounds turn each row into exactly \(L^{-\kappa_{\mathscr A}}J_{\mathscr A}(N)\). Fixed residue conditions can be discarded for this upper estimate. Its CRT endpoint error per retained record pair is \(O(z^{4+\epsilon})\); summing at most \(E^2\) pairs gives \(N^{0.45+o(1)}\). The exponent is independent of the model depth. Rankin's bound with parameter \(c/\log y\) gives \[ \#\{n\le N:n_{\mathcal P_y}>E\} \ll NL^{O(1)}e^{-cL^{1/10}}. \] It applies to every actual/model pair because the weights are bounded. Multiplying the record-stratum estimate by an arbitrary nonnegative record function bounded by \(L^B\), and summing, proves \eqref{eq:comparison-envelope}, including its interval-length factor and its errors. This is condition (C4). Its fixed logarithmic error exponents do not depend on the eventual density depth. All four hypotheses of Theorem~\ref{thm:model-comparison} have now been checked by invariance, one-variable estimates, or a positive upper sieve. No additive cancellation or residue-energy conclusion has been included as an arithmetic assumption. \subsection{The main term and the exact congruence criterion} \begin{lemma}\label{lem:coset-finite-images} A residue \(t\bmod M\) is represented by \(C_{\mathscr A}\) if and only if its reduction modulo \(\gcd(M,m)\) is that of some \(h\in H\), and \(p\nmid t\) at each forbidden prime \(p\mid M\), \(p\nmid m\). \end{lemma} \begin{proof} Necessity follows from prime support. For sufficiency extract an allowed-prime prefix \(D\) with valuations \(\min(v_p(t),v_p(M))\) at allowed primes dividing \(M\). It belongs to \(C_{\mathscr A}\), and its residue belongs to \(H\). The desired unit tail has state \(hD^{-1}\in H\). The positive monoid generated by \(\mathscr A\) in the finite group is \(H\), so choose a nonempty word \(s_1\cdots s_k=hD^{-1}\), with \(s_i\in\mathscr A\). An identity state also has a nonempty word, by taking a positive power of any element of \(\mathscr A\). Realize all but the last letter by primes in their reduced classes modulo \(m\), avoiding \(M\). For the last letter, impose its class modulo \(m\), and the remaining unit-tail congruences modulo the prime powers of \(M\). At an allowed prime with prefix valuation \(e0\). Conversely fix such a root. At primes dividing \(m\), lift the equation while preserving its unit states. At every other prime, both coefficients are units and \(p\ge3\). The equation modulo \(p\) has \(p\) pairs; requiring a unit on each forbidden coordinate removes at most two. A surviving pair lifts, and CRT combines the choices. Lemma~\ref{lem:coset-finite-images} realizes each coordinate by an actual element of \(C_{\mathscr A}\). This proves solvability in every finite quotient. Finally the positive asymptotic proves eventual global representability. \end{proof} \subsubsection{Two cosets which test the branches} For \(m=4\), \(K=\{1\}\), \(a=3\), the generated group is \(H=G\), while \(\rho=1/2\). The exponents are \(1/2,-1/2\). Let \(C^-_4\) contain the positive integers supported on primes \(3\pmod4\). With \(P_3=\prod_{p\equiv3(4)}(1-p^{-2})^{-1}\), its marginal is \(c_{\mathscr A}^2=2P_3/\pi^2\). Indeed \[ F(s)^2=(1-2^{-s})\frac{\zeta(s)}{L(s,\chi_{-4})}F(2s). \] Here \(\Gamma_{\mathscr A}(N)=2\) on every even target. Since \(\prod_{p\text{ odd}}(1-p^{-2})^{-1}=\pi^2/8\), \[ 2c_{\mathscr A}^2C_{\mathcal E_{\mathscr A}} =\frac12\prod_{p\equiv1(4)} \left(1-\frac2{p(p-1)}\right)=:A_-. \] Thus Theorem~\ref{thm:coset} gives \[ \#\{x+y=N:x,y\in C^-_4\} = A_-\prod_{\substack{p\mid N\\p\equiv1(4)}}\frac{p-1}{p-2} \frac N{\log N} \left(1+O((\log N)^{-1/16384})\right) \] for every sufficiently large even \(N\). Odd targets have zero count. The negative branch is needed for centering, although it does not change this leading constant. For a genuinely complex example, take \(m=10\), \(K=\{1\}\), \(a=3\), and let \(C_{3,10}\) be supported on primes \(3\pmod{10}\). Choose the literal quartic character \(\chi\bmod10\) with \(\chi(3)=i\). Then \(H=G\), \(\rho=1/4\), \(d=4\), and the exponents of \(1,\chi,\chi^2,\overline\chi\) are \[ \frac14,\quad\frac i4,\quad-\frac14, \quad-\frac i4. \] The rough model is explicitly real: \[ \mathcal D_J(n)=\frac{f(A)}{V_y} \left\{M_1(v)+2\Re\bigl(\overline{\chi(v)}M_\chi(v)\bigr) +\chi(v)^2 M_{\chi^2}(v)\right\}. \] The complex terms have relative size \(O(L^{-3/20})\), and the negative branch has relative size \(O(L^{-3/10})\). They cannot be omitted from an arbitrarily accurate centered model. The leading constant is given by \eqref{eq:coset-marginal}; equivalently, \[ c_{3,10}=\frac{(2/5)^{1/4}G_1(1)}{\Gamma(1/4)} L(1,\chi)^{-i/4}L(1,\chi^2)^{-1/4} L(1,\overline\chi)^{i/4}, \] with the positive determination and convergent correction specified in \eqref{eq:coset-constant}. The forbidden good residues are \(1,7,9\pmod{10}\). For even \(N\), \[ \Gamma_{\mathscr A}(N)= \begin{cases}5/2,&5\mid N,\\15/8,&5\nmid N.\end{cases} \] This follows by counting pairs of nonzero residues modulo five: there are four at a zero target and three otherwise. Consequently every even target is congruence-admissible, and \[ \#\{x+y=N:x,y\in C_{3,10}\} =c_{3,10}^2\Gamma_{\mathscr A}(N)C_{\mathcal E_{\mathscr A}}J_{\mathscr A}(N) \frac N{(\log N)^{3/2}} \left(1+O((\log N)^{-1/32768})\right). \] This includes all sufficiently large even targets, uniformly in their prime valuations. The example tests complex character contributions without changing the comparison theorem or its inputs. The common-bank condition is substantive. Two different prime prescriptions can be disjoint and supply no common allowed prime class. Theorem~\ref{thm:coset} makes no assertion about arbitrary pairs of prime cosets.