\section{The inverse on the whole annulus} \label{sec:radial-reference-inverse} Spectral gaps of the individual pieces do not automatically survive gluing. A homogeneous solution can leave one piece, return through its neighbour, and satisfy a new matching condition. We must therefore solve the matching problem for the whole annulus, not just list the inverses of the pieces. This section does so and completes the proof of Proposition~\ref{prop:rad-fixed-theorem}. \subsection{Overview} Everything is done one frequency at a time. Fix an even frequency $j$ and let $n$ be its band index, $j\in\Band_n$ (Section~\ref{subsec:bands}). At each end the annulus is divided into three regions (Figure~\ref{fig:frequency-regions}). \begin{itemize} \item The \emph{inner region}: the central model, the cells $mn$, governed by the recurrence of Lemma~\ref{lem:radial-deep-inverse}. \end{itemize} The inverse is assembled in four steps: an inner inverse $G_h$ with zero boundary values at the two entrances of the outer cores (Section~\ref{subsec:rad-localized-hilbert-assembly}); an outer inverse $R_o$ at each end, with the natural condition at the endpoint (Section~\ref{subsec:outer-inverse}); a $2\times2$ linear system for the two boundary values that remain, one for each end (Section~\ref{subsec:terminal-matching}); and weighted estimates for the resulting formula (Section~\ref{subsec:global-estimate}). Solving on the pieces first and then solving a finite system for the values on the cuts is again an instance of the Schur complement formula \cite{SjostrandZworski2007}; the matrices $\mS$ and $D_b$ below are Schur complements in this sense. One principle organizes the estimates. A correction created at the transition cell is large: it costs $\delta_n^{-1}$, the transfer factor $e^{C_o\tau_{n,j}}$ of the core, and coefficient powers. If its output stays in cell $n$ or beyond, this cost is absorbed by the coefficient weight allowed in the output norm there. If its output returns to the interior, it must first cross the positive ramp preceding the transition join, which attenuates it by $e^{-c\tau_{n-1,j}}$; and $\tau_{n-1,j}/\tau_{n,j}\asymp e^{n}$, so the attenuation wins. \begin{figure}[htbp] \centering \begin{tikzpicture}[x=1cm,y=1cm,>=stealth] \draw[->] (0,0)--(12,0); \draw[fill=orange!18] (.1,.3) rectangle (4.65,.95); \draw[fill=green!18] (4.65,.3) rectangle (5.75,.95); \draw[fill=red!12] (5.75,.3) rectangle (8.2,.95); \draw[fill=blue!13] (8.2,.3) rectangle (11.7,.95); \node at (2.38,.62) {inner region: cells $mn$}; \node[align=center,font=\small] at (2.3,1.55) {$\tau_{m,j}\ge C_*q_m$:\\local spectral inverses}; \node[align=center,font=\small] at (7,1.55) {selected matching angle\\$\ge\delta_n$}; \node[align=center,font=\small] at (10,1.55) {contracting recurrence,\\natural condition at the endpoint}; \draw[->,thick,red!70!black] (7.8,-.35)--(5.9,-.35); \draw[->,thick,green!50!black] (5.65,-.35)--(4.1,-.35); \node[align=center,font=\small] at (6,-1.05) {a cost $e^{C_o\tau_{n,j}}$ created in the core is attenuated by $e^{-c\tau_{n-1,j}}$\\before it enters the interior, and $\tau_{n-1,j}/\tau_{n,j}\asymp e^n$}; \end{tikzpicture} \caption{The decomposition of one end of the annulus at a frequency $j\in\Band_n$. The other end is treated at the same time. The position of the division changes with the frequency; the profile is chosen once.} \label{fig:frequency-regions} \end{figure} \paragraph{Notation of this section.} We write $w$ for the profile. Estimates are made in the Hilbert space $L^2(d\mu)$, $d\mu=\varpi\,dr$, with the conventions \begin{equation}\label{eq:rad-eq-hn1} B_j=-\varpi^{-1}L_j, \qquad F=-\varpi^{-1}f,\qquad \langle u,F\rangle_{L^2(d\mu)}=-\int\overline u\,f\,dr , \end{equation} so that $L_ju=f$ is the same as $B_ju=F$; we also write $\mathscr H_j=-L_j$. For a border column $k=-\varpi^{-1}g_0$ the Hilbert adjoint is $k^*u=-\ell_0u$; so after a change of sign of the scalar equation the bordered operator is symmetric. This convention is used in all energy identities below. Finite matrices that act on boundary values are written in sans-serif type when their letter is also used for a scale ($\mA,\mB,\mK,\mS$). The letters $\lambda$, $\Lambda$, $\kappa$, $T$, $D$, $E$ have the local meanings stated in each subsection. \subsection{Completing the profile while keeping the seed} \label{subsec:rad-cone-bump-target-match} We first finish the construction of the profile, in such a way that the solution of $L_2\phi=0$ is preserved exactly. The device is to prescribe the solution and to solve for the profile. \begin{lemma}[Connectors]\label{lem:connector} On an interval put $y=\log r$ and $w=re^f$, and let $\phi$ be a nonvanishing function with logarithmic slope $\mathsf k=\phi_y/\phi$. Then $L_2\phi=0$ is equivalent to \begin{equation}\label{eq:rad-eq-sp1} f_y=\frac{\mathsf k^2+\mathsf k_y-4}{4-\mathsf k}, \qquad\text{that is,}\qquad 1+f_y=\frac{rw'}{w}=\frac{\mathsf k^2-\mathsf k+\mathsf k_y}{4-\mathsf k}. \end{equation} \end{lemma} \begin{proof} In the variable $y$ the equation $r^2L_2\phi=0$ reads $w\phi_{yy}+(w_y-w)\phi_y-4w_y\phi=0$. Divide by $w\phi$ and use $\phi_{yy}/\phi=\mathsf k_y+\mathsf k^2$ and $w_y/w=1+f_y$. \end{proof} Thus any function $\mathsf k(y)$ with $\mathsf k<4$ and $\mathsf k^2-\mathsf k+\mathsf k_y>0$ produces, by integrating first $f_y$ and then $\phi_y=\mathsf k\phi$, a positive interval together with a solution of $L_2\phi=0$ on it. For constant $\mathsf k=\pm2$ the right side of \eqref{eq:rad-eq-sp1} is $1$: the interval is a cone and $\phi=r^{\pm2}$. This is more reliable than interpolating $w$ and hoping to keep the kernel. \paragraph{The central side.} By \eqref{eq:rad-eq-sc2} the logarithmic slope of the central solution $\phi_0$ tends to $2$ on the left and to $-2$ on the right, with all derivatives. Cut the central model far out on both sides and interpolate $\mathsf k$ to the constant $\pm2$ over a unit interval of $y$, with cutoffs that match all derivatives. Since the right side of \eqref{eq:rad-eq-sp1} is $1$ at the constant, closeness in $C^1$ gives $1+f_y\ge\tfrac34$. This \emph{connector} ends in an exact cone on which $\phi=r^{2}$ (left) or $\phi=r^{-2}$ (right). \paragraph{The tail side.} On the auxiliary cone of a tail with selected amplitudes, the natural solution at frequency two is $C_+r^2+C_-r^{-2}$. At the right tail the incoming line is spanned by $r^2$, so the nonzero angle of Lemma~\ref{lem:radial-product-selection} gives $C_-\ne0$; far along the cone, towards $r=0$, the slope of the natural solution tends to $-2$. At the left tail the incoming line is $r^{-2}$, hence $C_+\ne0$ and the slope tends to $2$ towards $r=\infty$. In both cases the natural solution does not vanish far along the cone, and the same interpolation of $\mathsf k$ connects it to an exact cone on which the solution is $r^{-2}$ (right tail) or $r^{2}$ (left tail), keeping $1+f_y\ge\tfrac34$. \paragraph{Assembly.} The exact cones on the central side and on the tail side carry the same solution, $r^{-2}$ on the right and $r^{2}$ on the left. A change of radius \begin{equation}\label{eq:rad-eq-ha1} \widetilde r=Rr,\qquad \widetilde w(Rr)=R\,w(r) \end{equation} preserves $w/r$, the operators $B_j$ and the measure $\varpi\,dr$; it places each tail at the required radius. A rescaling of the amplitude of the tail then matches the additive constant of $f$, without changing homogeneous solutions or travel angles, and a constant multiple of the tail solution matches its value. Cone segments of arbitrary length may be left between the connectors. The result is one smooth profile $w_*$, flat at both endpoints by Lemma~\ref{lem:tail-flat}, and one function $\phi_*$ with $L_{2,w_*}\phi_*=0$ which is flat at both endpoints by Lemma~\ref{lem:radial-deep-inverse}(iii). Only positive intervals were added, so the negative components and their travel angles are those of the models; part~(ii) of Proposition~\ref{prop:rad-fixed-theorem} follows from Section~\ref{subsec:rad-central-singlepoint} and Lemma~\ref{lem:morse-arithmetic}, with $\gamma_m^{-1}=3\sqrt{d_m}$. From now on $w=w_*$. It remains to prove that there is no kernel besides $\phi_*$ and that the inverse satisfies the stated estimates. \subsection{A tool: cutting a model inside a positive buffer} \label{subsec:buffer} The spectral information of Section~\ref{sec:rosen-reference-packet} concerns models on the whole line, whereas the profile contains only finite pieces of them. The following lemma says that cutting a model inside a region where its potential is large and positive changes its inverse, and its boundary data, by an exponentially small amount. It is formulated for the Schr\"odinger operator in the model coordinate. \begin{lemma}[Cutting inside a positive buffer]\label{lem:fence-compression} Let $H_0=-\partial_X^2+V$ be self-adjoint on a full model, possibly with a compact symmetric scalar border supported away from the cuts, and let $\mathcal H$ denote the resulting operator. Assume $\mathcal H$ is invertible with $\|\mathcal H^{-1}\|\le B_0$. Let there be one or two cuts $c_i$, and at each cut numbers $m_i>0$ and $\Sigma_i>0$ such that \begin{itemize} \item $V\ge m_i^2$ on the exterior of the cut and on the inward interval of length $(\Sigma_i+6)/m_i$ adjacent to it (the \emph{buffer}); \item $V\le Cm_i^2$ on two layers of width $4/m_i$: one at the cut and one at the inner end of the buffer; \item an exterior wall, if there is one, is at distance at least $4/m_i$ from the cut. \end{itemize} Let $\alpha_i^{\rm in}$ be the least energy $\int(|v'|^2+V|v|^2)\,dX$ of a function on the buffer with $v=1$ at the cut and $v=0$ at the inner end of the buffer, and $\alpha_i^{\rm out}$ the least energy of a function on the exterior with $v=1$ at the cut. Then $\alpha_i^{\rm in},\alpha_i^{\rm out}\asymp m_i$. Put $D=\diag(\alpha_i^{\rm in}+\alpha_i^{\rm out})$. Let $\mathcal H_D$ be the operator on the retained interval with the Dirichlet condition at the cuts; let $T$ be its Poisson map, which sends boundary values at the cuts to the homogeneous solution with zero scalar equation; and let $\Lambda$ be its matrix of outward derivatives at the cuts. If $B_0m_i^2e^{-\Sigma_i}$ is sufficiently small at every cut, then \begin{equation}\label{eq:buffer-estimates-new} \begin{gathered} \|\mathcal H_D^{-1}\|\le B_0+C\sum_i m_i^{-2},\qquad \|TD^{-1/2}\|\le C\Bigl(\sum_i m_i^{-2}\Bigr)^{1/2},\\ D^{-1/2}\bigl(\Lambda-\diag\alpha_i^{\rm in}\bigr)D^{-1/2} =O(E),\qquad |E_{ik}|\le CB_0\,m_im_k\,e^{-\Sigma_i-\Sigma_k}. \end{gathered} \end{equation} In particular $\Lambda$ is positive definite. \end{lemma} In words: the cut model has essentially the inverse of the full model; a boundary value at a cut penetrates only a distance $1/m_i$; and the response of the cut model to a boundary value is, up to a doubly exponentially attenuated error, that of the buffer alone. \begin{proof} \step{1: the energies} Expanding $\int|v'\pm m_iv|^2\ge0$ shows that the energy of a function with $v=1$ at the cut is at least $m_i$. A test function supported within distance $2/m_i$ of the cut gives the matching upper bound. \step{2: an exact Green column at each cut} Join at the cut the minimizers of the two energies. After division by $\alpha_i^{\rm in}+\alpha_i^{\rm out}$ this is the Green function, with pole at the cut, of the operator with a Dirichlet wall at the inner end of the buffer. Multiply it by a cutoff which vanishes near that wall, and call the result $K_i$. Comparison with the solutions of $-v''+m_i^2v=0$ gives \[ \mathcal H(K_i,0)=(\delta_{c_i}+f_i,\,0),\qquad \|K_i\|_2\le Cm_i^{-3/2},\qquad \|f_i\|_2\le Cm_i^{1/2}e^{-\Sigma_i} . \] The commutator term $f_i$ is an ordinary $L^2$ function supported in the inner layer, and the support of the cutoff does not meet the border. Therefore \[ \boldsymbol g_i=(K_i,0)-\mathcal H^{-1}(f_i,0) \] satisfies $\mathcal H\boldsymbol g_i=(\delta_{c_i},0)$ exactly. \step{3: the matrix of cut values} Let $\Gamma_{ik}$ be the value at $c_i$ of the field component of $\boldsymbol g_k$. The correction $\mathcal H^{-1}(f_k,0)$ solves the homogeneous equation in the buffer at $c_i$ and satisfies the exterior condition, so on its way from the inner layer to the cut it decays by a further factor $e^{-\Sigma_i}$. Passing from an $L^2$ bound to a pointwise bound costs $Cm_i^{1/2}$: where $V\ge m_i^2$ the modulus of this solution is monotone, so its value at a point is at most $m_i^{1/2}$ times its $L^2$ norm over the preceding interval of length $1/m_i$. With $\|\mathcal H^{-1}(f_k,0)\|\le B_0\,Cm_k^{1/2}e^{-\Sigma_k}$ and $D_{ii}\asymp m_i$ this gives \[ D^{1/2}\,\Gamma\,D^{1/2}=I+E,\qquad |E_{ik}|\le CB_0\,m_im_k\,e^{-\Sigma_i-\Sigma_k} , \] which also follows from Green's identity applied to two of the commutator terms. Under the smallness assumption this matrix is invertible. \step{4: the cut operator} Let $\iota$ extend a source on the retained interval by zero, let $\mathcal R$ restrict a field to the retained interval, both keeping the scalar, and let $\boldsymbol g_I$ be the restricted Green columns. Subtracting the cut values gives \[ \mathcal H_D^{-1}=\mathcal R\,\mathcal H^{-1}\iota -\boldsymbol g_I\,\Gamma^{-1}\boldsymbol g_I^*, \qquad T=\boldsymbol g_I\,\Gamma^{-1}. \] The kernel is trivial: a homogeneous solution with zero values at the cuts, extended by zero, solves the full equation with delta sources at the cuts, and the invertibility of $\Gamma$ forces their coefficients to vanish. Moreover \[ \|\boldsymbol g_i\|\,D_{ii}^{1/2} \le C\bigl(m_i^{-1}+B_0\,m_i\,e^{-\Sigma_i}\bigr)\le Cm_i^{-1}, \] which gives the first two bounds in \eqref{eq:buffer-estimates-new}. Finally, the interior and exterior derivatives of a Green column add up to the coefficient of its delta function, so $\Lambda=\Gamma^{-1}-\diag\alpha_i^{\rm out}$; this proves the third bound and the positivity. The symmetric border cancels in all the Green identities used. \end{proof} \paragraph{Translation to radial quantities.} Lemma~\ref{lem:fence-compression} is applied to the models through the dictionary of Lemma~\ref{lem:liouville}. For $u=\mathcal P^{-1/2}\psi$, \begin{equation}\label{eq:rad-eq-hn2} \|u\|_{L^2(d\mu)}^2=12\lambda\int\upsilon^2|\psi|^2\,dX, \qquad wu_r=\lambda\sqrt{\mathcal P} \Bigl(\psi_X-\frac{\mathcal P_X}{2\mathcal P}\,\psi\Bigr), \end{equation} where $\upsilon=r_X/r$; working with $\sqrt\lambda\,\psi$ removes the amplitude $\lambda$ from all formulas. The quantity $m_i$ of the lemma is a lower bound for $\sqrt V=j\sqrt{|\rho|}$ on the buffer. The buffer is a fixed fraction of the ramp at a distance from the turning point, where $\rho$ vanishes. For the two ramps of a Morse cell $m$, with $\Delta X$ the length of the buffer, \[ (m_i,\ \Delta X,\ \upsilon)\asymp \bigl(j\sqrt{A_m},\ 1,\ A_m\bigr)\ \text{(short ramp)},\qquad \bigl(jA_m^{9/2},\ A_m^{-4},\ A_m^5\bigr)\ \text{(long ramp)}. \] In both cases the buffer has action $\Sigma_i=c\,\tau_{m,j}$, with $\tau_{m,j}=j\sqrt{A_m}$ as in \eqref{eq:cap-action}, and $\upsilon/m_i\le C/j$. The inverse of the full Morse model costs a coefficient power, by \eqref{eq:rad-eq-rm1}. For the central model, a cut at $|X|=X_0$ has $m_i\asymp jX_0$ and action $c\,jX_0^2$, and the weighted inverse of the full model is bounded by Lemma~\ref{lem:central-weighted}. By the second formula in \eqref{eq:rad-eq-hn2}, the Liouville factor contributes to the radial flux. At the cuts of the central model, with $\epsilon_i=\pm1$ the outward sign and $\mathcal P_i$ the value of $\mathcal P$ at the cut, the matrix of outward radial fluxes is \begin{equation}\label{eq:rad-eq-hn3} \Lambda_r=\lambda\,\diag(\sqrt{\mathcal P_i})\, \Lambda\,\diag(\sqrt{\mathcal P_i}) +\lambda\,a_c\,\diag(\epsilon_i\mathcal P_i). \end{equation} Its leading diagonal part is therefore \begin{equation}\label{eq:kappa-def} \kappa_i=\lambda\,\mathcal P_i\,(\alpha_i^{\rm in}+\epsilon_ia_c), \end{equation} which is positive and comparable to $\lambda\mathcal P_im_i$ at far cuts. For a Morse cell $\mathcal P=1$ and $\kappa_i=\lambda\alpha_i^{\rm in}$. The term with $a_c$ must be kept for the fluxes to match exactly. \subsection{The inner region} \label{subsec:rad-localized-hilbert-assembly} Fix $j\in\Band_n$. The inner domain consists of the central model, the cells $m0$; here $m^2=o(q_m)$ was absorbed in $c_1$, and $c_2\log(1+t)\le c_3t/4+C$. So after $C_*$ is fixed, increasing $N$ makes these errors smaller than any prescribed tolerance, simultaneously for all $m0$ with $e^{-Cn^2}\le\lambda/\lambda_n\le e^{Cn^2}$. We describe solutions on the core by the state $(u,z)$ with $z=wu_{r_o}/\lambda$. \begin{lemma}[Inputs from the core and the deep tail] \label{lem:transition-inputs} Let $\mathsf T_{\rm core}$ be the transfer matrix of $L_ju=0$ across the core in the state $(u,z)$, and let $P_d$ be the natural solution on the deep tail with value one at its entrance. Then \begin{equation}\label{eq:transition-inputs-new} \begin{gathered} \|\mathsf T_{\rm core}\|+\|\mathsf T_{\rm core}^{-1}\| \le\Pi_n:=e^{Cn^2+C_o\tau_{n,j}}, \qquad \int_{\rm core}\varpi\,dr_o\le\lambda\,e^{Cn^2},\\ \|P_d\|_{L^2(d\mu)}\le\sqrt\lambda\,e^{Cn^2}, \qquad \|B_{d,D}^{-1}\|\le Cj^{-4}, \end{gathered} \end{equation} where $B_{d,D}$ is $B_j$ on the deep tail with the Dirichlet condition at its entrance. \end{lemma} \begin{proof} The last bound is \eqref{eq:rad-eq-rm20}. The mass of the core is $12\lambda_n\int(r_X/r)^2\,dX$ plus the negligible mass of the join, and is estimated by \eqref{eq:rad-eq-rm6}. The contraction \eqref{eq:rad-eq-dg4} bounds $\|P_d\|^2$ by a constant times the first mass of the deep tail; if an initial resistance edge comes first, the value at the next mass entrance is at most $C/(\Lambda\mu R_0)$ and the mass of the edge is small, with the same conclusion. For the transfer matrix, consider first a positive piece. In the state $(u,z)$ the equation is $u_{r_o}=(\lambda/w)\,z$, $z_{r_o}=(j^2w_r/(\lambda r))\,u$. Replace $z$ by $\varsigma z$ with a constant $\varsigma>0$ and apply Gronwall's inequality; the best choice of $\varsigma$ balances the integrals of the two coefficients, and $\varsigma^{\pm1}\le e^{Cn^2}$ on the pieces of the core. This bounds the transfer in either direction by \[ e^{Cn^2}\exp\biggl(2j\sqrt{\int\frac{dr_o}{w}\, \int\frac{w_r}{r}\,dr_o}\,\biggr). \] The product under the square root is $O(A_n)$ on each ramp, by Lemma~\ref{lem:cell}, and smaller on the join. On the negative component use $u_{XX}+j^2\rho\,u=0$ and the energy $\mathsf e=|u_X|^2+(1+j^2\rho)|u|^2$, which satisfies \[ |\mathsf e_X|\le\Bigl[1+\frac{j^2|\rho_X|}{1+j^2\rho}\Bigr]\mathsf e . \] The component has length $O(n^2)$ in $X$ and $\rho$ has a single maximum, so the integral of the second term is at most $2\log(1+j^2\max\rho)$. Converting between $(u,u_X)$ and $(u,z)$ costs $e^{Cn^2}$. \end{proof} \paragraph{A Robin inverse.} Let $V$ be the natural solution on the outer region. Its value at the entrance $r_{o,-}$ of the core may vanish, so we do not divide by it. Instead normalize its full state there, and let $U$ be the solution with the perpendicular state: \begin{equation}\label{eq:rad-eq-tm1} v=\Bigl(V,\ \frac{f_V}{\lambda}\Bigr)(r_{o,-}),\quad |v|=1, \qquad f_u:=wu_{r_o}, \qquad \Bigl(U,\ \frac{f_U}{\lambda}\Bigr)(r_{o,-})=v^\perp:=(-v_2,v_1). \end{equation} Then $w\,(UV_{r_o}-U_{r_o}V)=-\lambda$, and the kernel \[ K_o(r_o,t_o)=\lambda^{-1}\,U(\min(r_o,t_o))\,V(\max(r_o,t_o)) \] defines operators \begin{equation}\label{eq:outer-robin-new} R_oF=\int K_o(r_o,t_o)\,F(t_o)\,d\mu(t_o),\qquad G_of=-\int K_o(r_o,t_o)\,f(t_o)\,dt_o , \end{equation} which invert $B_j$ in $L^2(d\mu)$ and $L_j$ for the pairing $dr_o$, respectively, with the natural condition at the endpoint and the Robin condition $v\cdot(u,f_u/\lambda)=0$ at the entrance. For a solution $u=R_oF$, Green's identity with $V$ gives the complementary boundary quantity, \begin{equation}\label{eq:rad-eq-tm2} v^\perp\cdot\Bigl(u,\ \frac{f_u}{\lambda}\Bigr)(r_{o,-}) =\frac1\lambda\int V\,\varpi\,F\,dr_o, \qquad \|V\|_{L^2(d\mu)}\le\sqrt\lambda\,\mathcal E_n^C\,\Pi_n ; \end{equation} the boundary term at the endpoint vanishes by Section~\ref{subsec:deep-green}. \begin{lemma}[The outer inverse]\label{lem:outer-inverse} The operator $R_o$ satisfies \begin{equation}\label{eq:rad-eq-tm3} \|R_o\|\le\mathcal E_n^C\,\Pi_n^2 , \end{equation} and, for the localizations $\vartheta_k,\vartheta_m$ to cells $k,m\ge n-O(1)$, \begin{equation}\label{eq:rad-eq-tm4} \|\vartheta_k\,R_o\,\vartheta_m\| \le\mathcal E_k^C\,\mathcal E_m^C\,\Pi_n^2\,e^{-|S_k-S_m|}. \end{equation} The localized bounds \eqref{eq:rad-eq-tm4} hold also for $G_o$ between the spaces $L^2(dr_o)$. \end{lemma} The global bound \eqref{eq:rad-eq-tm3} has no counterpart for $G_o$: no inverse of $L_j$ is bounded on $L^2(dr_o)$ over the whole outer region. Indeed, let $u_\epsilon$ be a fixed smooth bump rescaled to the points at distance between $\epsilon$ and $2\epsilon$ from the flat endpoint. Since $w$ is flat there and $L_ju=(wu')'-j^2w'u/r$, \[ \|L_ju_\epsilon\|_{L^2(dr_o)} \le C_N\,\epsilon^N\,\|u_\epsilon\|_{L^2(dr_o)} \qquad\text{for every }N . \] Only the weighted bound and the localized bounds are used below. \begin{proof} The factor $\lambda^{-1}$ in $K_o$ causes no loss. On the core it is cancelled by the mass bound in \eqref{eq:transition-inputs-new}. On the deep tail, restricting the kernel gives exactly \[ (R_o)_{d,d}=B_{d,D}^{-1}+P_d\,K_o(r_{o,+},r_{o,+})\,P_d^*, \qquad |K_o(r_{o,+},r_{o,+})| \le\lambda^{-1}\,\mathcal E_n^C\,\Pi_n^2 , \] where $r_{o,+}$ is the entrance of the deep tail, and the two factors $P_d$ cancel $\lambda^{-1}$. The blocks between core and deep tail contain one factor $P_d$ each. This proves \eqref{eq:rad-eq-tm3}. The localized bound follows from \eqref{eq:rad-eq-dj5} for the first term and from \eqref{eq:rad-eq-rm17} for the factors $P_d$: a product of two such decay factors, both measured from the entrance, is bounded by $e^{-|S_k-S_m|}$. Passing to $L^2(dr_o)$ in a localized bound multiplies source and output by $\varpi^{\mp1/2}$ on their cells $m$ and $k$, which are powers of $\mathcal E_m$ and $\mathcal E_k$. These powers are not bounded uniformly in the cell, which is why the global bound is not transferred. \end{proof} \subsection{Matching the two remaining boundary values} \label{subsec:terminal-matching} Let $i$ label the two truncation points, where the inner domain meets the outer core at each end. Put \[ \lambda_i=\frac{j^2\,W_{{\rm in},i}}{r_{{\rm in},i}}, \] with $W_{{\rm in},i}$ and $r_{{\rm in},i}$ the profile and the radius at that point; its ratio to $\lambda_n$ lies between $e^{-Cn^2}$ and $e^{Cn^2}$, so Section~\ref{subsec:outer-inverse} applies with $\lambda=\lambda_i$. We use the normalized boundary coordinates \[ x_i=\sqrt{\lambda_i}\,u_i,\qquad z_i=\frac{f_{u,i}}{\sqrt{\lambda_i}} . \] In these coordinates the stand-in solution of Section~\ref{subsec:height-selection}, which vanishes at a cut in the preceding positive ramp, has the ratio of outward flux to value \[ d_i=\frac{f_u}{\lambda_iu}=\frac{r\,u_{r_o}}{j^2u} \in\bigl(0,\tfrac12\bigr], \qquad D_0:=\diag(d_i), \] by the estimate proving \eqref{eq:rad-eq-js6}; in the first band $d_i=1/j$. \paragraph{The true inner boundary matrix.} Let $B_h$ be the inner operator in $L^2(d\mu)$, including the symmetric border, and $\mathsf q_h$ its quadratic form. We compare the true matrix of normalized outward fluxes of the inner problem with $D_0$. Take the stand-in solution with boundary value $\lambda_i^{-1/2}$, multiply it by a cutoff near the inner end of its ramp, and extend it by zero to a field $\mathsf L_i$ on the inner domain; let $\mathsf F_i=B_h\mathsf L_i$, with zero scalar component. Separate fixed fractions of the ramp are reserved for this cutoff and for the buffer of the neighbouring model. Let $\mathsf L$ and $\mathsf F$ have the columns $\mathsf L_i$ and $\mathsf F_i$, and put $P_h=\mathsf L-G_h\mathsf F$. \begin{lemma}[The inner boundary matrix] \label{lem:inner-boundary-matrix} The field $P_hx$ is the solution of the homogeneous inner problem with normalized boundary values $x$. For both signs, and with $\omega_i$ the constant value of $W$ on the gap which carries the cutoff, \begin{equation}\label{eq:rad-eq-tm6} \begin{gathered} \|W^{\pm1}\mathsf L_i\|\,\omega_i^{\mp1} \le\frac{C\sqrt{d_i}}{j},\qquad \|W^{\pm1}\mathsf F_i\|\,\omega_i^{\mp1} \le\sqrt{d_i}\;\mathcal E_n^C\,e^{-\Sigma_{b,i}},\\ \Sigma_{b,i}=c\,\tau_{n-1,j}\quad\text{(in the first band: } \Sigma_{b,i}=c\,jL). \end{gathered} \end{equation} The matrix $D_b$ of normalized outward fluxes of $P_h$, defined by $x^*D_bx=\mathsf q_h(P_hx,P_hx)$, is \begin{equation}\label{eq:rad-eq-tm7} D_b=\mathsf q_h(\mathsf L,\mathsf L)-\mathsf F^*G_h\mathsf F =D_0+E_b, \qquad |(E_b)_{ik}|\le \mathcal E_n^C\,\sqrt{d_id_k}\;e^{-\Sigma_{b,i}-\Sigma_{b,k}} . \end{equation} \end{lemma} \begin{proof} The field $\mathsf Lx$ has the boundary values $x$ and $B_h\mathsf Lx=\mathsf Fx$, while $G_h\mathsf Fx$ vanishes at the boundary and $B_hG_h\mathsf Fx=\mathsf Fx$; so $P_hx$ is the homogeneous solution with boundary values $x$. The bounds \eqref{eq:rad-eq-tm6} follow from positivity and the commutator formula \[ \mathscr H_j(\chi u)-\chi\,\mathscr H_ju =-2w\chi'u'-(w\chi')'u . \] The first bound uses the energy $d_i$ of the stand-in solution and \eqref{eq:rad-eq-hn5}; the second uses the attenuation up to the cutoff layer. The two fields $\mathsf L_i$ have disjoint supports, and their energy matrix is diagonal with entries $d_i\,(1+O(\mathcal E_n^Ce^{-2\Sigma_{b,i}}))$. For \eqref{eq:rad-eq-tm7}, complete the square in the quadratic form. Since $B_hG_h\mathsf F=\mathsf F$ and $G_h\mathsf Fx$ vanishes at the boundary, integration by parts gives \[ \mathsf q_h(\mathsf Lx,G_h\mathsf Fx) =\langle\mathsf Fx,G_h\mathsf Fx\rangle =\mathsf q_h(G_h\mathsf Fx,G_h\mathsf Fx) , \] so in the expansion of $\mathsf q_h(\mathsf Lx-G_h\mathsf Fx, \mathsf Lx-G_h\mathsf Fx)$ the two cross terms cancel one of the two correction terms. The error $E_b$ consists of $-\mathsf F^*G_h\mathsf F$ and of the deviation of the energy matrix of the fields $\mathsf L_i$ from $D_0$; both are bounded by \eqref{eq:rad-eq-tm6} and the second estimate in \eqref{eq:inner-inverse-estimates-new}. This is why the error contains \emph{two} attenuation factors. The border is symmetric and the fields $\mathsf L_i$ have zero scalar component, so no scalar term is missing. \end{proof} Any dependence of the interior on the selected tails enters only through $E_b$. \paragraph{The matching system.} Let $v_i$ be the normalized entrance state \eqref{eq:rad-eq-tm1} of the natural solution $V_i$ at the end $i$, and put \[ \mA=\diag(v_{1,i}),\qquad \mB=\diag(v_{2,i}),\qquad \mathcal V=\bigl(V_i/\sqrt{\lambda_i}\bigr)_i , \] so that $\mA^2+\mB^2=I$. \begin{lemma}[The matching system]\label{lem:matching-system} Let $F_h$ be a source in the inner domain and $F_o$ a source in the outer regions. Consider the field which is $G_hF_h+P_hx$ in the inner domain and, in the outer region at the end $i$, is $R_oF_o$ plus a multiple of $V_i$. Its values and fluxes are continuous at the truncation points if and only if the normalized boundary values $x$ and fluxes $z$ satisfy \begin{equation}\label{eq:rad-eq-tm8} \begin{gathered} z=D_bx-P_h^*F_h,\qquad -\mB x+\mA z=\mathcal V^*F_o,\\ \text{that is,}\qquad \mK x=\mathcal V^*F_o+\mA P_h^*F_h,\qquad \mK=\mA(D_0+E_b)-\mB . \end{gathered} \end{equation} The coefficients of the natural solutions $V_i/\sqrt{\lambda_i}$ in the outer regions are then the entries of $\mA x+\mB z$. \end{lemma} \begin{proof} The outward flux of $P_hx$ is $D_bx$, and by Green's identity that of $G_hF_h$, which vanishes at the boundary, is $-P_h^*F_h$. The outer solution at the end $i$ is $R_oF_o$ plus a multiple of $V_i$; by \eqref{eq:rad-eq-tm1} and \eqref{eq:rad-eq-tm2} its normalized state $(x_i,z_i)$ at the truncation point is $(\mathcal V^*F_o)_i\,v_i^\perp$ plus a multiple of $v_i$, the multiple being the coefficient of $V_i/\sqrt{\lambda_i}$ in the outer solution. Continuity of values and fluxes at the truncation points means that the two normalized states agree. Taking the component along $v_i^\perp$ gives the second equation of \eqref{eq:rad-eq-tm8}, and the component along $v_i$ shows that the coefficients of the natural solutions are $\mA x+\mB z$. Eliminating $z$ gives the last line. \end{proof} \begin{lemma}[The matching matrix]\label{lem:matching} After $C_*$, $N$ and the lengths of the cone heads have been chosen suitably, the matrix $\mK$ is invertible and $\|\mK^{-1}\|\le C\delta_n^{-1}$, uniformly in $n$ and in $j\in\Band_n$. \end{lemma} \begin{proof} The diagonal part $\mA D_0-\mB$ of $\mK$ is controlled by the selected angle. The state used in Lemma~\ref{lem:radial-product-selection} is $(u,-u\pm f_u/\lambda_i)$ at the truncation point, which differs from $(u,f_u/\lambda_i)$ by a fixed linear map, and transport across the join is close to the identity by \eqref{eq:rad-eq-js14}. Hence \begin{equation}\label{eq:rad-eq-tm9} |v_{1,i}d_i-v_{2,i}| =\sqrt{1+d_i^2}\;|\sin\angle(v_i,(1,d_i))| \ge c\,\delta_n . \end{equation} This holds also when $v_{1,i}=0$. The direction $(1,d_i)$ was fixed before the amplitude ratio was varied. It therefore suffices to have \begin{equation}\label{eq:rad-eq-tm10} \delta_n^{-1}\|E_b\|<\frac c4, \qquad\text{and we also arrange}\qquad \mathcal E_n^{C_{\rm b}}\,\delta_n^{-1}\, e^{C_o\tau_{n,j}-c\tau_{n-1,j}}\le e^{-c\tau_{n-1,j}/2} \end{equation} for use in Section~\ref{subsec:global-estimate}. Here $C_{\rm b}$ is the largest coefficient exponent which occurs in \eqref{eq:rad-eq-tm6}, \eqref{eq:rad-eq-tm7} and in the bounds \eqref{eq:boundary-column-weights} below. Like every coefficient exponent it depends only on the exponent $\ea$ of the weights, and it is fixed before $C_*$ and $N$. Both inequalities hold because \[ \frac{\tau_{n-1,j}}{\tau_{n,j}}\asymp e^n,\qquad \tau_{n-1,j}\ge C_*q_{n-1},\qquad \frac{q_n}{q_{n-1}}\to1,\qquad \frac{n^2}{q_{n-1}}\to0 . \] The first relation absorbs $C_o\tau_{n,j}$ into a quarter of $c\tau_{n-1,j}$, and the others absorb $Cq_n+n^2+|\log c_0|$ into another quarter, provided $C_*$ is chosen after the coefficient exponent $C_{\rm b}$ and $N$ is large; what remains is the factor $e^{-c\tau_{n-1,j}/2}$. The first inequality in \eqref{eq:rad-eq-tm10} even has two attenuation factors, by \eqref{eq:rad-eq-tm7}. In the first band the role of $\tau_{n-1,j}$ is played by $jL$, and the length $L$ of the cone heads is taken large. \end{proof} \subsection{The matched inverse and its weighted bound} \label{subsec:global-estimate} Put $\mK_1=\mA+\mB D_b$. Solving \eqref{eq:rad-eq-tm8} and reconstructing the coefficient of the natural solution in the outer regions, which is $\mK_1x-\mB P_h^*F_h$, gives the inverse for the frequency $j$ on the whole annulus, \begin{equation}\label{eq:rad-eq-tm11} G=\begin{pmatrix}G_h&0\\0&R_o\end{pmatrix} +\begin{pmatrix}P_h&0\\0&\mathcal V\end{pmatrix} \begin{pmatrix} \mK^{-1}\mA&\mK^{-1}\\\mK^{-T}&\mK_1\mK^{-1} \end{pmatrix} \begin{pmatrix}P_h^*&0\\0&\mathcal V^*\end{pmatrix}, \end{equation} in the decomposition into inner and outer sources and fields. Since $D_b$ is symmetric and $\mA,\mB$ are commuting diagonal matrices with $\mA^2+\mB^2=I$, direct multiplication gives \[ \mK\mA=\mA\mK^T,\qquad \mK^T\mK_1=\mK_1^T\mK,\qquad \mK_1\mK^{-1}\mA-\mB=\mK^{-T}. \] These identities prove the formula and show that the middle matrix is symmetric; its norm is $O(\delta_n^{-1})$ by Lemma~\ref{lem:matching}. The boundary weights at the two ends are taken equal, which is possible because the band index is the same at both ends. We now prove that $G$ is bounded between the weighted spaces. The direct terms $G_h$ and $R_o$ are covered by Lemmas~\ref{lem:inner-inverse} and~\ref{lem:outer-inverse}. The correction terms are of two kinds, according to where their output lies, and each kind is paid for in its own way. \begin{proposition}[Weighted bound at one frequency] \label{prop:global-reference-bound} For every compact interval $[\gamma_-,\gamma_+]\subset(0,1)$ there are exponents $\ea$, $\el>\ea$ and choices of $C_*$, $N$, $X_0$, $L$ such that, with the weights $W$ of \eqref{eq:rad-eq-hn4} and $W_1=W\,(\mathcal E/\mathcal E_N)^{-\el}$, \begin{equation}\label{eq:global-reference-bound-new} \|W_1\,G\,W^{-1}\|\le C \end{equation} for all even $j\ge2$ and all $\gamma\in[\gamma_-,\gamma_+]$, where at $j=2$ the temporary border is included. The same bound holds between the spaces of source densities for $dr$ and fields in $L^2(dr)$, after an adjustment of the exponents. \end{proposition} \begin{proof} \step{1: terms whose source and output are both outer} These are the terms of \eqref{eq:rad-eq-tm11} built from $R_o$, $\mathcal V$ and the leading parts $\mathsf L$ of $P_h$; their sources and outputs lie in cells of index at least $n-O(1)$. By \eqref{eq:rad-eq-tm4} and $\|\mK^{-1}\|\le C\delta_n^{-1}$, with fixed constants $\alpha_1,\alpha_2$, \[ \|\vartheta_k\,G_{\rm out}\,\vartheta_m\| \le C\exp\{\alpha_1\,(q_k+q_m)+\alpha_2\,q_n-|S_k-S_m|\} ; \] here $\alpha_2$ accounts for the cost of the core, since $\delta_n^{-1}e^{C_o\tau_{n,j}}\le c_0^{-1}\exp(n^2+C_oC_*q_n)$. The output is outer, so $q_n\le c_1q_k+C_1$; put $\alpha_3=\alpha_2c_1$. Conjugating by the weights gives, with $\Delta S=S_k-S_m$, \begin{equation}\label{eq:outer-weighted-block} \begin{aligned} \log\|\vartheta_k\,W_1G_{\rm out}W^{-1}\vartheta_m\| &\le C-(\el-2\alpha_1-\alpha_3)\,q_k\\ &\quad +(\alpha_1-\ea)\,(q_m-q_k)-|\Delta S|+\gamma\,\Delta S . \end{aligned} \end{equation} Choose $\el>\max(\ea,2\alpha_1+\alpha_3)$; this is done after $C_*$. Since $q_n$ varies slowly compared with $S_n$, \[ |q_m-q_k|\le\varepsilon\,|S_m-S_k|+C_\varepsilon, \qquad \varepsilon=\frac{1-\gamma_+}{2\,(1+|\alpha_1-\ea|)}, \] and \eqref{eq:outer-weighted-block} is at most $C_\gamma-\tfrac12(1-\gamma_+)\,|S_k-S_m|$. With $s_0=\inf_m(S_{m+1}-S_m)>0$ and $\eta=\tfrac12(1-\gamma_+)$, \[ \sup_k\sum_m e^{-\eta|S_k-S_m|}\le1+\frac2{e^{\eta s_0}-1}, \] and the same holds with $k$ and $m$ exchanged. Schur's test for block kernels, with the bounded overlap of the localizers and the two ends, bounds this part of $W_1GW^{-1}$. Note that raising the output exponent $\el$ helps precisely because the output index $k$ is at least about $n$; it would not help for an output at a fixed place in the interior. \step{2: terms which contain a column $G_h\mathsf F$} Insert $P_h=\mathsf L-G_h\mathsf F$ in \eqref{eq:rad-eq-tm11}. The terms not covered by Step~1 are those in which the column $G_h\mathsf F$ occurs at least once. With the symmetric matrix $\mathsf X=\mK^{-1}\mA=\mA\mK^{-T}$ there are five of them: \begin{center} \renewcommand{\arraystretch}{1.15} \begin{tabular}{@{}lll@{}} \toprule & term & from source to output\\ \midrule (a) & $-(G_h\mathsf F)\,\mK^{-1}\,\mathcal V^*$ & outer to inner\\ (a$'$) & $-\mathcal V\,\mK^{-T}\,(G_h\mathsf F)^*$ & inner to outer\\ (b) & $-(G_h\mathsf F)\,\mathsf X\,\mathsf L^*$ & inner to inner\\ (b$'$) & $-\mathsf L\,\mathsf X\,(G_h\mathsf F)^*$ & inner to inner\\ (c) & $(G_h\mathsf F)\,\mathsf X\,(G_h\mathsf F)^*$ & inner to inner\\ \bottomrule \end{tabular} \end{center} Each is a product of a column, a matrix indexed by the truncation points, and a row. Let $\Omega$ be the common weight at the two truncation points, and $W_h,W_o$ the restrictions of $W$ to the inner and outer regions. Since $\Omega$ is a multiple of the identity, it commutes with $\mK^{-1}$ and with $\mathsf X$. For both signs, \begin{equation}\label{eq:boundary-column-weights} \begin{gathered} \|W_h^{\pm1}\,G_h\mathsf F\,\Omega^{\mp1}\| \le\mathcal E_n^{C_1}\,e^{-c\tau_{n-1,j}},\qquad \|W_h^{\pm1}\,\mathsf L\,\Omega^{\mp1}\|\le\mathcal E_n^{C_1},\\ \|\mK^{-1}\|+\|\mathsf X\|\le C\delta_n^{-1},\\ \|W_o^{\pm1}\,\mathcal V\,\Omega^{\mp1}\| \le C_\gamma\,\mathcal E_n^{C_2}\,e^{C_o\tau_{n,j}} . \end{gathered} \end{equation} The first bound is the second bound of \eqref{eq:inner-inverse-estimates-new} applied to \eqref{eq:rad-eq-tm6}, with $d_i\le\tfrac12$; the second is the first bound of \eqref{eq:rad-eq-tm6}. In both, the weight $\omega_i$ of \eqref{eq:rad-eq-tm6} has been replaced by $\Omega$: the two belong to the same cell or to neighbouring cells, so by \eqref{eq:adjacent-weight-ratio-new} their ratio is a coefficient power. The third bound is Lemma~\ref{lem:matching}. For the fourth, the part of the column $\mathcal V$ in cell $k$ is bounded by $\mathcal E_n^C\mathcal E_k^C\,e^{C_o\tau_{n,j}}\,e^{-(S_k-S_n)}$; multiplying by either power of the weight changes the exponent by $(C\pm\ea)(q_k-q_n)-(1\mp\gamma)(S_k-S_n)$, and the slow variation of $q$ used in Step~1 makes the sum of the squares finite, leaving a coefficient power at $n$ and a constant $C_\gamma$. The norm of a row is the norm of the adjoint column, so $\|\Omega^{\pm1}\,\mathcal V^*\,W_o^{\mp1}\| =\|W_o^{\mp1}\,\mathcal V\,\Omega^{\pm1}\|$, and likewise for $(G_h\mathsf F)^*$ and $\mathsf L^*$. Insert $\Omega^{-1}\Omega$ on both sides of the middle matrix and multiply the bounds \eqref{eq:boundary-column-weights}, taking the column with the upper signs and the row, through its adjoint, with the lower signs: \begin{equation}\label{eq:one-boundary-cross-block} \begin{aligned} \text{(a), (a$'$):}\quad& C_\gamma\,\mathcal E_n^{C_1+C_2}\,\delta_n^{-1}\, e^{C_o\tau_{n,j}-c\tau_{n-1,j}},\\ \text{(b), (b$'$):}\quad& C\,\mathcal E_n^{2C_1}\,\delta_n^{-1}\,e^{-c\tau_{n-1,j}},\\ \text{(c):}\quad& C\,\mathcal E_n^{2C_1}\,\delta_n^{-1}\,e^{-2c\tau_{n-1,j}} . \end{aligned} \end{equation} These are bounds for the norms of $W\,(\cdot)\,W^{-1}$ of the five terms. By the second inequality in \eqref{eq:rad-eq-tm10}, in which $C_{\rm b}\ge\max(C_1+C_2,\,2C_1)$, each of them is at most $C_\gamma\,e^{-c\tau_{n-1,j}/2}\le C_\gamma$; in the first band $c\,\tau_{n-1,j}$ is replaced by $c\,jL$. Finally $\mathcal E/\mathcal E_N\ge1$ on the tails and is bounded below on the middle part, so $W_1\le C\,W$, and the bounds hold with the output weight $W_1$. \step{3: conclusion} All terms of \eqref{eq:rad-eq-tm11} have been covered. The passage between $L^2(d\mu)$ and the spaces for $dr$ costs coefficient powers on each cell, which are absorbed by enlarging $\ea$ and $\el$. \end{proof} \begin{lemma}[One geometry for all exponents] \label{lem:radial-uniform-exponents} The profile, the amplitude ratios, the central cuts and the lengths of the cone heads can be chosen once, so that \eqref{eq:global-reference-bound-new} holds for every compact interval $[\gamma_-,\gamma_+]\subset(0,1)$. Only the constants and the exponents $\ea,\el$ depend on the interval. \end{lemma} \begin{proof} All the finite matching problems, that is the inversion of $\mS$ and of $\mK$, involve the weights only through the ratios of neighbouring weights, and \eqref{eq:adjacent-weight-ratio-new} is uniform for $|\gamma|\le1$. Fix one exponent $\ea$ large enough for the conversion of sources, and choose $C_*$, $N$, the central cuts and the cone heads using that uniform bound. The invertibility of $\mS$ and $\mK$ then does not depend on $\gamma$. With the geometry fixed, the only place where $\gamma$ matters is the summation over the cells of the deep tail. Conjugation by $H^{-\gamma}$ changes the exponential in \eqref{eq:rad-eq-tm4} to \[ e^{-|S_k-S_m|+\gamma(S_k-S_m)}\le e^{-(1-|\gamma|)\,|S_k-S_m|}, \] whose row and column sums are finite for $|\gamma|\le\gamma_+<1$. The coefficient factors are absorbed by $|q_k-q_m|\le\varepsilon|S_k-S_m|+C_\varepsilon$, which holds because $q'(x)/S'(x)\to0$; here $\varepsilon$ is chosen after $\gamma_+$, which changes the summation constant and the output exponent but not the geometry. The terms which enter the interior already carry the attenuation \eqref{eq:rad-eq-tm10}. So the dependence on the interval is confined to constants which multiply an existing bound. \end{proof} \paragraph{One radial derivative.} Derivatives of the solution are estimated from the equation, not by differentiating the construction. On a slightly enlarged cell $m$, \begin{equation}\label{eq:rad-eq-rm39} \begin{aligned} \|u_r\|_2&\le\mathcal E_m^C\bigl(\|f\|_2+(1+j)\,\|u\|_2\bigr),\\ \|u_{rr}\|_2&\le\mathcal E_m^C\bigl(\|f\|_2+(1+j^2)\,\|u\|_2\bigr). \end{aligned} \end{equation} For the first, use the interpolation inequality $\|u_r\|_2^2\le C\|u\|_2(\|u_{rr}\|_2+\|u\|_2)$ on an interval, substitute $wu_{rr}=f-w'u_r+j^2w'u/r$, and absorb the resulting term $C\|u\|_2\|u_r\|_2$ by Young's inequality; then substitute back for the second. Rescaling the interval and dividing by $w$ cost coefficient powers, and a cutoff on the enlarged cell gives the estimate without any boundary condition. At $j=2$ the source $f$ is replaced by $f-\nu g_0$, whose second term is compactly supported and controlled by $|\nu|$. Summing the first inequality over the modes and the cells, with the weights of the output norm, gives \eqref{eq:reference-radial-derivative}: the factor $1+j$ is one angular derivative of $u$, and the inverse commutes with $\partial_\theta$. In an outer core, any fixed power of $j$ is itself a coefficient power, because $j<\jn_n$. \subsection{The zero mode, the domain, and the final border} \label{subsec:final-border} It remains to treat the mode $j=0$, to say on which domain the operator is inverted, and to replace the temporary border. \paragraph{The zero mode.} For a source $f_0$ with $\int f_0\,dr=0$ put \begin{equation}\label{eq:rad-eq-rm49} p_0(r)=\int_{r_-}^r f_0(t)\,dt=-\int_r^{r_+}f_0(t)\,dt,\qquad u_0(r)=-\int_r^{r_+}\frac{p_0(t)}{w(t)}\,dt . \end{equation} Then $(wu_0')'=f_0$, and $u_0$ vanishes at the right endpoint. Summing over the cells beyond $n$ gives $|p_0|\le\mathcal E_n^CH_n^\gamma\|f_0\|_{Y_{\gamma,0}}$ on cell $n$ of either tail. Division by $w$ and one more integration cost coefficient powers, so $u_0$ converges to a constant $a_L$ at the left endpoint, $a_L$ is a bounded functional of the source, and $u_0-a_L\chi_L$ has the required decay at both ends. The solution is unique: a homogeneous solution with nonzero flux would be a multiple of $\int dr/w$, which is excluded by its growth, and a constant is excluded by the condition at the right endpoint. \begin{definition}[The weighted graph domain] \label{def:radial-graph-domain} Let $\mathcal Y_\gamma^{\rm rad}$ be the space of sources $(f,c)$ with finite norm \eqref{eq:rad-eq-source-norm} at $s=0$ and with $\int\av f\,dr=0$. Let $\mathcal X_\gamma^{\rm rad}$ be the space of pairs $(u,\nu)$, $u=a_L\chi_L+v$, with finite output norm \[ |a_L|^2+|\nu|^2+ \sum_m\mathcal E_m^{-2\eb}H_m^{-2\gamma} \|\vartheta_mv\|_{L^2(dr\,d\theta)}^2 , \] the angular measure being normalized. The \emph{graph domain} is \[ \mathcal D_\gamma^{\rm rad} =\bigl\{U=(u,\nu)\in\mathcal X_\gamma^{\rm rad}:\ \mathbb L_wU\in\mathcal Y_\gamma^{\rm rad}\bigr\}, \qquad \|U\|_{\mathcal D_\gamma^{\rm rad}} =\|U\|_{\mathcal X_\gamma^{\rm rad}} +\|\mathbb L_wU\|_{\mathcal Y_\gamma^{\rm rad}} , \] where the differential expression is understood in the sense of distributions. The definition is used first with the temporary border and then with the final one. \end{definition} The kernel statement of Proposition~\ref{prop:rad-fixed-theorem}(i) will be proved for the kernel on this domain. This suffices, because the domain contains every solution of $L_ju=0$, $j\ge2$, which is bounded near both endpoints: the total mass $\int\mathsf p_j\,dx$ of a tail is finite, so a bounded solution is square integrable for $\mathsf p_j\,dx$, lies on the natural line by Lemma~\ref{lem:radial-deep-inverse}(ii), and therefore decays like $H_m$ by \eqref{eq:rad-eq-rm17}. \begin{lemma}[Two-sided inverse and closure] \label{lem:rad-green-completion} Let $\mathbb G_w$ be defined on $\mathcal Y_\gamma^{\rm rad}$ by the matched inverses \eqref{eq:rad-eq-tm11} on the modes $j\ge2$ and by \eqref{eq:rad-eq-rm49} on the mode $j=0$. Then $\mathbb G_w:\mathcal Y_\gamma^{\rm rad}\to\mathcal D_\gamma^{\rm rad}$ is bounded, with $\eb=\el-\ea$ adjusted as in Proposition~\ref{prop:global-reference-bound}, and \begin{equation}\label{eq:rad-reference-piecewise-left-identity} \mathbb L_w\mathbb G_w=I,\qquad \mathbb G_w\mathbb L_w=I\ \text{ on }\mathcal D_\gamma^{\rm rad}. \end{equation} The graph domain is the closure, in the graph norm, of the smooth fields obtained from smooth compactly supported sources with finitely many modes. Green's identity holds on it with the pairing $dr$, including the border. \end{lemma} \begin{proof} \step{1: right inverse} On smooth compactly supported sources with finitely many modes, every block of the construction is an inverse of the corresponding piece, and values and fluxes match exactly at all interfaces. So $\mathbb L_w\mathbb G_w=I$ on these sources, and by the bound \eqref{eq:global-reference-bound-new} and density it holds on $\mathcal Y_\gamma^{\rm rad}$. \step{2: uniqueness} Let $U\in\mathcal D_\gamma^{\rm rad}$ with $\mathbb L_wU=0$; we show $U=0$. At each frequency $j\ge2$ the component of $U$ is a solution of the homogeneous equation which is locally $H^2$, because $w>0$ in the interior; so its values and fluxes agree at every interface. The weights place it in $L^2(\mathsf p_j\,dx)$ near each endpoint, since $\gamma S_m-Cq_m\to+\infty$ for every fixed $C$. By Lemma~\ref{lem:radial-deep-inverse}(ii) it lies on the natural line at both ends. Its two boundary values at the truncation points then satisfy the homogeneous system \eqref{eq:rad-eq-tm8} and vanish because $\mK$ is invertible, and the inner and outer inverses show that it is zero. The zero mode is unique by the computation above. \step{3: closure and reciprocity} If a sequence converges in the graph norm, its limit and the limit of its images are related by $\mathbb L_w$ in the sense of distributions, so the graph domain is complete. Smooth compactly supported sources with finitely many modes are dense in $\mathcal Y_\gamma^{\rm rad}$: truncate and mollify, and restore the integral condition with a fixed bump of integral one. Their images are smooth, and flat at the endpoints in all modes $j\ge2$. By Steps~1 and~2 every element of the graph domain is a limit of such images. Green's identity follows from the local identities and the symmetry of the middle matrix in \eqref{eq:rad-eq-tm11}, with the conversion \eqref{eq:rad-eq-hn1}. \end{proof} \paragraph{The final border.} The function $\phi_*$ of Section~\ref{subsec:rad-cone-bump-target-match} spans the kernel of the unbordered operator at $j=2$ in the graph domain; a larger kernel would contradict the injectivity of the bordered operator. For the same reason $\ell_0\phi_*\ne0$. Now choose a function $\chi\ge0$, not identically zero, supported in a positive connector on the right of the central model where $\phi_*'/\phi_*<0$. Integrating by parts, \begin{equation}\label{eq:ct-positive-shape} a_*:=\int\phi_*\,L'_{2,w}[w\chi]\,\phi_*\,dr =\int w\chi\Bigl(-\phi_*'^2+\frac{8\phi_*\phi_*'}r -\frac{4\phi_*^2}{r^2}\Bigr)dr<0 . \end{equation} Put \begin{equation}\label{eq:rad-eq-ct9} \chi_*=\frac{\chi}{a_*},\qquad g_*=L'_{2,w}[w\chi_*]\,\phi_*,\qquad \ell_*u=\int g_*u_2\,dr , \end{equation} so that $\int g_*\phi_*\,dr=1$, that is, $\ell_*(\phi_*e_2)=1$. The bordered equation with the final border is solved as follows: for a source $f_2$ in the second mode, the scalar unknown is $\nu=\int\phi_*f_2\,dr$; the source $f_2-\nu g_*$ is orthogonal to $\phi_*$ and is solved with the temporary border; and a multiple of $\phi_*$ is added to satisfy the row $\ell_*$. This change of finite rank is bounded in all the weighted spaces and preserves the inverse identities and the closure of the graph. From now on $\mathbb L_{w_*}$ and $\mathbb G_{w_*}$ refer to the final border. \begin{proof}[Proof of Proposition~\ref{prop:rad-fixed-theorem}] The profile and part~(ii) were obtained in Section~\ref{subsec:rad-cone-bump-target-match}; the simple zeros of $w_*'$ are the zeros of $\rho$ in the models. Parts~(i) and~(iv) are Lemma~\ref{lem:rad-green-completion} with the final border, together with Proposition~\ref{prop:global-reference-bound} and \eqref{eq:rad-eq-rm39}; part~(iii) is \eqref{eq:rad-eq-ct9}; and part~(v) is Lemma~\ref{lem:radial-uniform-exponents}. \end{proof} \subsection{What Part~\ref{part:nonlinear} uses} \label{subsec:interface} The nonlinear argument uses the reference profile only through Proposition~\ref{prop:rad-fixed-theorem}, the coefficient bounds of Lemma~\ref{lem:tail-flat}, and the localized form of the inverse estimate: the block of $\mathbb G_{w_*}$ from cell $m$ to cell $k$ is bounded by a coefficient power times $e^{-\gamma(S_k-S_m)}$, for every $\gamma$ in a compact subinterval of $(0,1)$, whatever the sign of $S_k-S_m$; this is \eqref{eq:global-reference-bound-new} read block by block. What has been proved is an estimate at angular order zero, without loss of angular derivatives, with its first radial derivative, on the maximal graph domain. Higher radial derivatives will be obtained from the equation, which replaces two radial derivatives by at most two angular ones and lower-order terms. At each fixed total order $s$ the \emph{complete} product of coefficients produced by this recursion is bounded by $C_s\mathcal E_m^{C_s}$. To absorb such products for all $s$ at once we shall use the slightly larger weight \begin{equation}\label{eq:M-weight} M_m=\exp\bigl\{1+m^{5/2}\log^2(e+m)\bigr\}, \end{equation} which satisfies, for every fixed $s$ and every $\varepsilon>0$, \[ \mathcal E_m^{C_s}\le C_sM_m,\qquad |\log M_k-\log M_m|\le\varepsilon\,|S_k-S_m|+C_\varepsilon . \] The first inequality holds because $\log(e+m)\to\infty$, the second because the ratio of the derivatives of $\log M$ and of $S$ tends to zero. So complete products of coefficients cost one power of $M$, whatever the order.