\section{Inverting the corrected linearization near the reference} \label{sec:stable-inverse} \label{sec:grn-fixed-phase-green} The goal of this section is Proposition~\ref{prop:rewrite-volume-inverse}: for every state $\bx$ in the neighbourhood of Proposition~\ref{prop:uniform-physical-geometry}, the operator $\widehat L_{\bx}$, bordered by the derivative of the equation with respect to the metric parameter $\nu$, has an inverse which satisfies tame estimates. The argument has to respect one constraint. The reference inverse loses coefficient weight, and a tame estimate with a loss is not a bounded inverse on a single Banach space; it cannot by itself justify a Neumann series. We therefore proceed in three steps. \begin{enumerate}[label=\textup{\arabic*.}] \item We put the problem in canonical form and identify a single Hilbert space of sources on which the \emph{relative} operator $\mathscr R_{\bx}=(\mathbb P_{\bx}-\mathbb P_*)\,\mathbb G_*$ is bounded, with small norm. The loss of the reference inverse is compensated by the factor $H^\beta$ carried by every difference of coefficients (Sections~\ref{subsec:canonical-problem}--\ref{subsec:relative}). \item A Neumann series on that one space gives the solution. \item All higher regularity of this same solution is then recovered from the equation: angular derivatives by commutators, radial derivatives by solving the equation for its highest radial derivative (Section~\ref{subsec:higher}). \end{enumerate} In this order of argument, one neighbourhood of the reference state serves for all orders of differentiation. The result of the section is the following proposition. Its proof occupies Sections~\ref{subsec:canonical-problem}--\ref{subsec:higher}; the last lines are at the end of Section~\ref{subsec:higher}. \begin{proposition}[The inverse at a nearby state] \label{prop:rewrite-volume-inverse} Let $k_{\bx}=D_\nu\mathcal F_q(w,u)$ be the derivative of the equation with respect to the scalar $\nu$. For every state in $\mathfrak N$ and every $f\in\mathcal Y_\beta$ there is a unique pair $G_{\bx}f=(\eta,\mu)$ with \[ \widehat L_{\bx}\eta+\mu\,k_{\bx}=f,\qquad \eta\in\mathcal X_\beta\cap\ker\ell_*,\qquad \mu\in\R , \] where $\eta$ tends to zero at the right endpoint and to a constant at the left endpoint. For every $s$, with the base order $b$ of Lemma~\ref{lem:two-weights-closed-domain} and the number $d_*$ fixed below, \begin{equation}\label{eq:readable-physical-inverse-tame} \|G_{\bx}f\|_s \le C_s\bigl[\|f\|_{s+d_*} +(1+\mathfrak n_{s+d_*})\,\|f\|_{b+d_*}\bigr] . \end{equation} The map $(\bx,q,f)\mapsto G_{\bx}f$ is smooth tame. \end{proposition} Throughout this section and the next, $d_*$ denotes one fixed integer, the largest of the following losses of derivatives: \begin{itemize} \item $d_{\rm geo}+2$, for the coefficients of the relative operator and their commutators with $\Lambda$ (Lemma~\ref{lem:relative}); here $d_{\rm geo}=14$ is the number of Proposition~\ref{prop:uniform-physical-geometry}; \item $d_{\rm tot}$, the loss in \eqref{eq:rewrite-global-mixed-bound}; \item the sum of $d_{\rm tot}$, $L_0+1$ and the losses of the transformations between canonical and original variables (Proposition~\ref{prop:uniform-physical-geometry}), for \eqref{eq:readable-physical-inverse-tame}. \end{itemize} These numbers are determined by the reference profile and by the base order $b$; they do not depend on the orders $s$ and $t$, nor on $d_*$. Every estimate in which $d_*$ occurs remains true when $d_*$ is replaced by a larger number, because the seminorms $\mathfrak n_s$ and $\mathfrak g_{s,m}$ are nondecreasing in $s$. So the maximum can be taken at the outset. \subsection{The linear problem in canonical form} \label{subsec:canonical-problem} \paragraph{Transformations.} Use the cylinders of Section~\ref{sec:fixed-geometry}. Let $\Psi_{\bx}$ denote the coordinate map on each piece, $J_{\bx}$ its Jacobian determinant and $h_{\bx}$ the conjugating factor in the original variables. Define \[ \mathscr C_{\bx}u=\frac{u}{h_{\bx}}\circ\Psi_{\bx}^{-1}, \qquad \mathscr S_{\bx}f=\frac{h_{\bx}f}{|J_{\bx}|}\circ\Psi_{\bx}^{-1}, \qquad \mathbb L_{\bx}= \mathscr S_{\bx}\,\widehat L_{\bx}\,\mathscr C_{\bx}^{-1} . \] Thus $\mathscr C_{\bx}$ maps a height to its canonical unknown, $\mathscr S_{\bx}$ maps a source to its canonical source, and $\mathbb L_{\bx}$ is the operator in canonical form; on a positive cylinder $\mathbb L_{\bx}=-P_{\bx}$. The pairing with $dr\,d\theta$ is preserved. A \emph{piece} is one of the cylinders: a half of a negative component, or a positive component. A function on the collection of all pieces is written $F=(F_m)$. \paragraph{Source norms.} For real $t$ and integers $s\ge0$ let $\mathbb V_{t,s}$ be the space of $F=(F_m)$ with \begin{equation}\label{eq:vol-source-spaces} \|F\|_{t,s}^2 =\sum_m M_m^{2t}\,H_m^{-2\beta}\, \|F_m\|_{L^2_{\rm rad}H^s_\varphi}^2<\infty , \end{equation} where $m$ runs over all pieces, the norm is $L^2$ in the radial variable of the cylinder with $s$ angular derivatives, and $M_m,H_m$ are the weights of the cell containing the piece. No matching condition and no integral condition is imposed on sources. The index $t$ counts powers of the weight $M$, the index $s$ angular derivatives. \paragraph{Scalars and the augmented operator.} The linear problem involves three scalars besides the canonical unknown $W$. \begin{itemize} \item \emph{The endpoint constant $a_L$.} A height is represented as \[ \eta=\mathscr C_{\bx}^{-1}W+a_L\chi_L,\qquad \ell_*^cW=0 , \] so that $\eta$ tends to $a_L$ at the left endpoint and to $0$ at the right one. The coordinates agree with the reference ones on the interval $J_*$ of Section~\ref{subsec:positive-crossing}, which contains the support of $g_*$; hence the normalization $\ell_*\eta=0$ takes a form $\ell_*^cW=0$ which does not depend on the state. \item \emph{The metric increment $\mu$.} It multiplies the column \[ k_{\bx}=D_\nu\mathcal F_q(w,u),\qquad k_{\bx}^c=\mathscr S_{\bx}k_{\bx}, \] where $D_\nu$ is the derivative of the residual with respect to $\nu$ for fixed $u$ and $q$; it acts on the metric and on the radial height $h$. At the reference state $k_*=g_*e_2$, the border column of Proposition~\ref{prop:rad-fixed-theorem}. This scalar removes the obstruction in the second mode. \item \emph{An auxiliary scalar $\lambda$.} Choose a function $\psi$ of $r$, compactly supported in $J_*$, with $\Qint\psi=1$, and put $c_*=\mathscr S_*\psi=\mathscr S_{\bx}\psi$. The scalar $\lambda$ multiplies $c_*$ and removes the integral obstruction, so that arbitrary sources can be treated. For compatible sources it will turn out to be zero. \end{itemize} With $s_{\bx}=\mathscr S_{\bx}\widehat L_{\bx}\chi_L$, the \emph{augmented operator} is \begin{equation}\label{eq:vol-abstract-complete-operator} \mathbb P_{\bx}\bW =\mathbb L_{\bx}W+a_L\,s_{\bx}+\mu\,k_{\bx}^c+\lambda\,c_*, \qquad \bW=(W,a_L,\mu,\lambda) . \end{equation} The norm of an output $\bW$ is the norm \eqref{eq:vol-source-spaces} of $W$ plus the absolute values of the three scalars, which are assigned to the finite middle part; we write $\|\bW\|^{\rm vol}_{t,s}$. \paragraph{The reference inverse in this form.} At the reference state let $\mathfrak g_*=\mathscr C_*1$ and $\Qint_*F=\langle\mathfrak g_*,F\rangle$; this functional is continuous by the coefficient bounds, it is the integral $\Qint$ in canonical form, and $\Qint_*c_*=1$. Given $F$, put $\lambda=\Qint_*F$ and apply the inverse of Proposition~\ref{prop:rad-fixed-theorem} to the compatible source $F-\lambda c_*$. This defines a two-sided inverse $\mathbb G_*$ of $\mathbb P_*$. In the zero mode it is explicit: with $f_*=\mathscr S_*^{-1}F$, \[ p_0(r)=\int_{r_-}^r\bigl(\av{f_*}-\lambda\psi\bigr)(t)\,dt, \qquad u_0(r)=-\int_r^{r_+}\frac{p_0(t)}{w_*(t)}\,dt , \] as in \eqref{eq:rad-eq-rm49}; near either endpoint $p_0$ is the primitive starting at that endpoint, division by $w_*$ costs a coefficient power and not an endpoint exponent, and subtracting the left endpoint value times $\chi_L$ gives the field and the scalar $a_L$. \subsection{Two endpoint exponents give decay off the diagonal} \label{subsec:two-exponents} By Proposition~\ref{prop:rad-fixed-theorem}(iv),(v) the reference inverse is bounded with the endpoint weight $H^{-\gamma}$ for every $\gamma$ in a compact subinterval of $(0,1)$. Fix \[ 0<\epsilon<\min(\beta,1-\beta). \] The inverses at the exponents $\beta-\epsilon$ and $\beta+\epsilon$ coincide on smooth compactly supported sources: their difference is a homogeneous solution in the graph domain, hence zero by Lemma~\ref{lem:rad-green-completion}. So both estimates apply to the same operator $\mathbb G_*$. Using each of them on one side of the diagonal converts the two bounds into decay in the distance between source and output (Figure~\ref{fig:two-weights}). Obtaining decay of an inverse away from the diagonal from its boundedness in a family of exponentially weighted spaces is the idea of Combes and Thomas \cite{CombesThomas1973}. \begin{lemma}[A fixed loss of coefficient weight] \label{lem:two-weights-closed-domain} There is a fixed $L_0$ such that for every real $t$ and every $s\ge0$ \begin{equation}\label{eq:reference-negative-material-extension} \|\mathbb G_*F\|^{\rm vol}_{t-L_0,s}\le C_{t,s}\,\|F\|_{t,s} . \end{equation} The first radial derivative of $\mathbb G_*F$ satisfies the same bound with at most one additional angular derivative of the source. For a fixed integer $b\ge2$, \begin{equation}\label{eq:fixed-base-graph-domain} \mathbb D_*:=\mathbb G_*(\mathbb V_{0,b}),\qquad \|\bW\|_{\mathbb D_*}=\|\mathbb P_*\bW\|_{0,b} , \end{equation} is a Hilbert space on which $\mathbb P_*$ is closed. It contains every smooth $\bW$ in the transmission domain of Section~\ref{subsec:transmission-domain} which satisfies the conditions at the endpoints and the normalization, and for which $\mathbb P_*\bW\in\mathbb V_{0,b}$. \end{lemma} \begin{proof} \step{1: blocks} Let $\bar S_m=S_m-S_N$ on the tails and $\bar S_m=0$ on the finite middle. Let $\pi_k$ be the restriction to the piece $k$ and $\iota_m$ the extension by zero of a source on the piece $m$. By Section~\ref{subsec:interface}, the estimate at the exponent $\gamma$ gives \[ \|\pi_k\,\mathbb G_*\,\iota_m\|_{H^s_\varphi\to H^s_\varphi} \le C_s\,\mathcal E_k^{C_s}\,\mathcal E_m^{C_s}\, e^{-\gamma(\bar S_k-\bar S_m)} . \] There is no loss of angular derivatives, because the reference inverse acts on each angular mode with bounds uniform in the mode. Multiply by the weights $H_k^{-\beta}H_m^{\beta}$ of \eqref{eq:vol-source-spaces}, and use $\gamma=\beta+\epsilon$ when $\bar S_k\ge\bar S_m$ and $\gamma=\beta-\epsilon$ otherwise. The resulting block $K_{km}$ satisfies \begin{equation}\label{eq:global-two-weight-block-bound} \|K_{km}\|_{H^s_\varphi\to H^s_\varphi} \le C_s\,\mathcal E_k^{C_s}\,\mathcal E_m^{C_s}\, e^{-\epsilon|\bar S_k-\bar S_m|} . \end{equation} The estimate for the first radial derivative in Proposition~\ref{prop:rad-fixed-theorem}(iv) gives the same block bound with one more angular derivative. \step{2: summation} For every $\delta>0$, $|\log M_k-\log M_m|\le\delta\,|\bar S_k-\bar S_m|+C_\delta$, because the derivative of $n^{5/2}\log^2(e+n)$ with respect to $2n^3/3$ tends to zero. With $\mathcal E^{C_s}\le C_sM$ we get, for $L_0\ge2$, \[ M_k^{t-L_0}\,\|K_{km}\|\,M_m^{-t} \le C_s\,M_k^{2-L_0}\Bigl(\frac{M_k}{M_m}\Bigr)^{t-1} e^{-\epsilon|\bar S_k-\bar S_m|} \le C_{t,s}\,e^{-(\epsilon/2)|\bar S_k-\bar S_m|} , \] where $\delta$ was chosen with $|t-1|\,\delta\le\epsilon/2$. This matrix has bounded row and column sums, since the numbers $S_n$ are uniformly separated and each index carries a bounded number of pieces. Schur's test proves \eqref{eq:reference-negative-material-extension}. \step{3: the domain} By definition $\mathbb D_*$ is isometric to $\mathbb V_{0,b}$, hence complete. If $\bW_i\to\bW$ in the space of outputs and $\mathbb P_*\bW_i\to F$ in $\mathbb V_{0,b}$, then $\bW_i=\mathbb G_*\mathbb P_*\bW_i\to\mathbb G_*F$, so $\bW=\mathbb G_*F$: the operator is closed. Approximating a source by smooth compactly supported ones, the estimate gives convergence of the solutions on every piece, the equation passes to the limit in the sense of distributions, and the bound for the first radial derivative identifies the traces, so the transmission conditions hold in the weak sense. Finally, a smooth $\bW$ as in the last statement differs from $\mathbb G_*\mathbb P_*\bW$ by a normalized homogeneous solution with the natural behaviour at the endpoints, which vanishes by Lemma~\ref{lem:rad-green-completion}. \end{proof} \begin{figure}[htbp] \centering \begin{tikzpicture}[x=.92cm,y=.9cm,font=\small,>=Stealth] \draw[->] (0,0)--(6.2,0) node[right] {depth $\bar S_m$ of the source}; \draw[->] (0,0)--(0,4.9) node[above] {depth $\bar S_k$ of the output}; \fill[blue!7] (0,0)--(4.5,4.5)--(0,4.5)--cycle; \fill[orange!9] (0,0)--(5.7,0)--(5.7,4.5)--(4.5,4.5)--cycle; \draw[thick] (0,0)--(4.5,4.5); \node at (1.45,3.4) {use $\gamma=\beta+\epsilon$}; \node at (4,1.2) {use $\gamma=\beta-\epsilon$}; \node[align=left,anchor=west] at (6.55,2.6) {after conjugation by $H^{-\beta}$:\\[2pt] $e^{-\epsilon|\bar S_k-\bar S_m|}$}; \end{tikzpicture} \caption{The same reference inverse is estimated with two endpoint weights, each used on one side of the diagonal. Together they give decay in the distance between the depths of the source and of the output.} \label{fig:two-weights} \end{figure} \subsection{Positive cylinders: an estimate without boundary conditions} \label{subsec:cubic-estimate} On a positive cylinder the perturbation \eqref{eq:readable-positive-difference} is of second order, the same order as the operator. It can nevertheless be controlled by the operator, because its coefficients vanish to third order at the two ends. The following estimate is the reason for that requirement. In this subsection $\langle j\rangle=(1+j^2)^{1/2}$. \begin{lemma}[Weighted graph estimate]\label{lem:cubic} On $00$ we have $M_i^C\le C_\delta\,e^{\delta\bar S_i}$, and $\bar S_m\le\bar S_k+|\bar S_k-\bar S_m|$. Choosing $\delta<\min(\beta/4,\epsilon/2)$ bounds \eqref{eq:rewrite-relative-block} by \[ C'\,\mathfrak n_{b_0}\, e^{-(\beta/2)\bar S_k-(\epsilon/2)|\bar S_k-\bar S_m|} , \] which has bounded row and column sums. Increasing $b_0$ to cover the finitely many low derivatives used, and then decreasing $\epsilon_0$, gives \eqref{eq:rewrite-base-relative}. \step{2: higher orders} In the product estimates keep one coefficient factor at the high order. The same summation proves the bound for $\mathscr R_{\bx}$ in \eqref{eq:compact-volume-relative-bound}; the output may carry any power $M^t$ because of the factor $H_k^\beta$. For the commutator, assign the frequency zero to the scalars $a_L,\lambda$ and the frequency two to $\mu$. For multiplication by a coefficient, the matrix of the commutator in Fourier modes gains the factor $|k|-|j|$, which is bounded by $|k-j|$; so one more derivative of the coefficient gives the same bound. The commutators with the scalar columns are the terms $\Lambda(s_{\bx}-s_*)\,a_L$ and $(\Lambda-2)(k_{\bx}^c-k_*^c)\,\mu$, which are of the same type. \step{3: the factorization} By Step~1 the operator extends to $\mathbb V_{0,b}$. Approximating by smooth compactly supported sources identifies the extension with the difference of the differential operators, in the sense of distributions, on $\mathbb D_*$; here it is used that the transmission conditions are the same for all states. This is \eqref{eq:current-fixed-domain-factorization}. \end{proof} \subsection{Solving, and recovering all derivatives} \label{subsec:higher} By \eqref{eq:rewrite-base-relative} we may define, for $F\in\mathbb V_{0,b}$, \[ \widetilde F=(I+\mathscr R_{\bx})^{-1}F, \qquad \bW=\mathbb G_*\widetilde F , \] where the inverse is a Neumann series on $\mathbb V_{0,b}$. By \eqref{eq:current-fixed-domain-factorization}, $\mathbb P_{\bx}\bW=F$. We now show that this solution is as regular as the source allows, with tame estimates. Three steps recover, in turn, the angular derivatives of $\widetilde F$, its coefficient weight, and the radial derivatives of $W$. \paragraph{Step A: angular derivatives of $\widetilde F$.} This step rests on a general lemma. Let $\mathcal H$ be a Hilbert space with a self-adjoint operator $\Lambda$ whose spectrum lies in $\{0,1,2,\ldots\}$, let $\|v\|_p=\|\langle\Lambda\rangle^pv\|$, and let $\mathcal H^p$ be the space of vectors for which this norm is finite. \begin{lemma}[Regularity of an inverse from a commutator] \label{lem:ang-base-inverse-regularity} Let $R$ be a bounded operator on $\mathcal H$ which maps every $\mathcal H^p$ into itself, and suppose that $I+R$ has a bounded inverse on $\mathcal H$. Suppose the commutator $\mathscr K=[\Lambda,R]$ satisfies, for all integers $p\ge0$, \begin{equation}\label{eq:commutator-hypothesis} \|\mathscr Kv\|_p\le a_p\,\|v\|_p+c_p\,\|v\|_0 , \end{equation} with nondecreasing constants $a_p$ and $c_p$. Then $(I+R)^{-1}$ maps every $\mathcal H^p$ into itself, and for $p\ge1$ \begin{equation}\label{eq:commutator-conclusion} \|(I+R)^{-1}f\|_p\le C_p\bigl(\|f\|_p+(1+c_{p-1})\,\|f\|_0\bigr), \end{equation} where $C_p$ depends only on $p$, on $a_0,\ldots,a_{p-1}$, on $c_0$ and on the norm of $(I+R)^{-1}$ on $\mathcal H$. \end{lemma} The constants $c_p$ with $p\ge1$ may be arbitrarily large: they enter linearly, and they do not enter the constant $C_p$. \begin{proof} \step{1: the solution is regular} Let $U_\delta=e^{i\delta\Lambda}$. For $v\in\mathcal H^1$ the function $t\mapsto U_{\delta-t}RU_tv$ is differentiable with derivative $-i\,U_{\delta-t}\mathscr KU_tv$, so \[ [U_\delta,R]=i\int_0^\delta U_{\delta-t}\,\mathscr K\,U_t\,dt , \] first on $\mathcal H^1$ and then on $\mathcal H$, because $\mathscr K$ is bounded on $\mathcal H$ by \eqref{eq:commutator-hypothesis} with $p=0$. Let $u=(I+R)^{-1}f$. Then \[ (I+R)\,\frac{U_\delta u-u}{\delta} =\frac{U_\delta f-f}{\delta}-\frac{[U_\delta,R]\,u}{\delta}. \] If $f\in\mathcal H^1$ the right side is bounded in $\mathcal H$ uniformly in $\delta$, hence so are the difference quotients of $u$; by Fatou's lemma for the spectral decomposition of $\Lambda$, $u\in\mathcal H^1$, and letting $\delta\to0$ gives $(I+R)\Lambda u=\Lambda f-\mathscr Ku$. Now let $p\ge2$, $f\in\mathcal H^p$ and suppose $u\in\mathcal H^{p-1}$. Then \[ (I+R)\,\Lambda^{p-1}u=\Lambda^{p-1}f -\sum_{k=0}^{p-2}\Lambda^k\,\mathscr K\,\Lambda^{p-2-k}u , \] and the right side lies in $\mathcal H^1$, because $\Lambda^{p-2-k}u\in\mathcal H^{k+1}$ and $\mathscr K$ maps $\mathcal H^{k+1}$ into itself. By the case just treated $\Lambda^{p-1}u\in\mathcal H^1$, that is, $u\in\mathcal H^p$. By induction, for $f\in\mathcal H^p$, \begin{equation}\label{eq:commuted-equation} (I+R)\,\Lambda^pu=\Lambda^pf -\sum_{k=0}^{p-1}\Lambda^k\,\mathscr K\,\Lambda^{p-1-k}u . \end{equation} \step{2: the estimate} For $p=1$, \eqref{eq:commuted-equation} and the bound $a_0+c_0$ for $\mathscr K$ on $\mathcal H$ give the claim. Let $p\ge2$ and $A=1+c_{p-1}$. By \eqref{eq:commutator-hypothesis}, $\mathscr K$ is bounded on $\mathcal H$ with norm at most $a_0+c_0$, and from $\mathcal H^{p-1}$ with the norm $(\|v\|_{p-1}^2+A^2\|v\|_0^2)^{1/2}$ to $\mathcal H^{p-1}$ with norm at most $a_{p-1}+1$. Both pairs of norms are weighted $\ell^2$ norms over the spectrum of $\Lambda$. Interpolating between them, by the three-lines lemma applied to $z\mapsto\langle\Lambda\rangle^{(p-1)z}\,\mathscr K\, (\langle\Lambda\rangle^{2(p-1)}+A^2)^{-z/2}$ on vectors with finitely many nonzero spectral components, for which this family is analytic and bounded on the closed strip, and using $(\langle j\rangle^{2(p-1)}+A^2)^{\vartheta} \le\langle j\rangle^{2(p-1)\vartheta}+A^{2\vartheta}$ for $0\le\vartheta\le1$, we get \[ \|\mathscr Kv\|_k\le C\bigl(\|v\|_k+A^{k/(p-1)}\,\|v\|_0\bigr), \qquad 0\le k\le p-1 . \] Apply this to $v=\Lambda^{p-1-k}u$. Since $\|u\|_{p-1-k}\le\|u\|_{p-1}^{1-k/(p-1)}\|u\|_0^{k/(p-1)}$, Young's inequality gives \[ A^{k/(p-1)}\,\|u\|_{p-1-k} \le(A\|u\|_0)^{k/(p-1)}\,\|u\|_{p-1}^{1-k/(p-1)} \le\|u\|_{p-1}+A\,\|u\|_0 . \] So every term of the sum in \eqref{eq:commuted-equation} is bounded in $\mathcal H$ by $C(\|u\|_{p-1}+A\|u\|_0)$. Applying $(I+R)^{-1}$ to \eqref{eq:commuted-equation}, \[ \|u\|_p\le C\bigl(\|f\|_p+\|u\|_{p-1}+A\,\|u\|_0\bigr). \] Finally $\|u\|_{p-1}\le\varepsilon\|u\|_p+C_\varepsilon\|u\|_0$, and $\|u\|_0\le C\|f\|_0$; absorbing the term $\varepsilon\|u\|_p$ proves \eqref{eq:commutator-conclusion}. \end{proof} We apply the lemma with $\mathcal H=\mathbb V_{0,b}$, $\mathcal H^p=\mathbb V_{0,b+p}$, $\Lambda=|D_\varphi|$ and $R=\mathscr R_{\bx}$. By \eqref{eq:compact-volume-relative-bound} with $t=0$, the operator $\mathscr R_{\bx}$ maps every $\mathbb V_{0,s}$ into itself, and its commutator with $\Lambda$ satisfies \eqref{eq:commutator-hypothesis} with fixed constants $a_p$ and with $c_p=C_p\,\mathfrak n_{b+p+d_*}$; in particular $c_0$ is bounded on the neighbourhood. Hence $\widetilde F=(I+\mathscr R_{\bx})^{-1}F$ has all the angular regularity of $F$, and \begin{equation}\label{eq:angular-resolvent-bound} \|\widetilde F\|_{0,s} \le C_s\bigl[\|F\|_{0,s}+(1+\mathfrak n_{s+d_*})\,\|F\|_{0,b}\bigr], \qquad s\ge b . \end{equation} \paragraph{Step B: coefficient weight of $\widetilde F$.} Use the identity $\widetilde F=F-\mathscr R_{\bx}\widetilde F$ for the function already constructed. By \eqref{eq:compact-volume-relative-bound} the last term has every coefficient weight once the angular regularity of $\widetilde F$ is known. Hence \begin{equation}\label{eq:full-source-tame-resolvent} \|\widetilde F\|_{t,s} \le C_{t,s}\bigl[\|F\|_{t,s} +(1+\mathfrak n_{t+s+d_*})\,\|F\|_{0,b}\bigr] . \end{equation} \paragraph{Step C: radial derivatives of $W$.} The corrected source $\widetilde F$ is controlled only in its angular derivatives. For the radial derivatives of $W$ we therefore use the equation $\mathbb P_{\bx}\bW=F$ with the \emph{original} source. Subtract the three scalar columns, \[ F^\circ=F-a_L\,s_{\bx}-\mu\,k_{\bx}^c-\lambda\,c_* , \] and rescale every half of a negative component to unit radial length with a variable $y$; the positive pieces are already normalized, and the rescaling factors are coefficient powers. Solving the equation for its highest radial derivative gives on every piece \begin{equation}\label{eq:rewrite-normal-recursion-base} \partial_y^2W=\sum_{(i,j)\in I}A_{ij}\,\partial_y^i \partial_\varphi^jW+B\,F^\circ,\qquad I=\{(1,1),(0,2),(1,0),(0,1),(0,0)\} . \end{equation} On a negative half this is the equation $W_{\xi\xi}=\xi W_{\varphi\varphi}-VW+F^\circ$. On a positive piece the coefficient of $\partial_y^2$ is bounded below by $\mathcal E_m^{-C}$ on the whole neighbourhood. The coefficients $A_{ij}$ and $B$ obey the calculus of Proposition~\ref{prop:uniform-physical-geometry}, with a fixed loss $d_{\rm geo}$ of derivatives of the state. Differentiating, for $a\ge2$, \begin{equation}\label{eq:rewrite-mixed-normal-recursion} \begin{split} \partial_y^a\partial_\varphi^b W ={}&\sum_{(i,j)\in I}\sum_{p=0}^{a-2}\sum_{q=0}^{b} \binom{a-2}{p}\binom bq\, (\partial_y^p\partial_\varphi^q A_{ij})\; \partial_y^{a-2-p+i}\partial_\varphi^{b-q+j}W\\ &+\partial_y^{a-2}\partial_\varphi^b(B\,F^\circ) . \end{split} \end{equation} Every term on the right contains at most $a-1$ radial derivatives of $W$ and has total order at most $a+b$, and a derivative that falls on a coefficient lowers the order of the remaining factor by the same amount. Iterating, the expansion terminates in angular derivatives of $W$, angular derivatives of $\partial_yW$, and derivatives of $F^\circ$; at total order $s$ their orders are at most $s$, $s-1$ and $s-2$. Starting from the angular regularity and the first radial derivative, which are known, this identity proves the existence of all mixed derivatives in turn. To make this quantitative, fix a base order $b_1\ge4$ and for $s\ge b_1+2$ put \[ \mathfrak a_{s,m} =\|W_m\|_{L_y^2H_\varphi^s} +\|\partial_yW_m\|_{L_y^2H_\varphi^{s-1}} +\|F_m^\circ\|_{H^{s-2}_{y,\varphi}} . \] The finite expansion of \eqref{eq:rewrite-mixed-normal-recursion} gives \begin{equation}\label{eq:rewrite-local-mixed-bound} \|W_m\|_{H^s_{y,\varphi}} \le C_s\,\mathcal E_m^{C_s}\, \bigl[\mathfrak a_{s,m} +(1+\mathfrak n_{s+d_{\rm geo},m})\,\mathfrak a_{b_1,m}\bigr] . \end{equation} Indeed, in each term of the expansion the orders of the derivatives of the coefficients and of the remaining factor add up to at most $s$. This is the situation of Lemma~\ref{lem:one-high-factor}(ii), applied after extension to the full cylinder: one factor is taken in the norm of order $s$, or $s+d_{\rm geo}$ for a coefficient, and all the others in norms of the fixed order $b_1$. Where a factor is known only in the mixed norms of $\mathfrak a_{s,m}$, interpolate its intermediate norms between the corresponding entries of $\mathfrak a_{b_1,m}$ and $\mathfrak a_{s,m}$ by Lemma~\ref{lem:one-high-factor}(i). The Sobolev embedding is used at the fixed order $b_1$ only. The reciprocal of the leading coefficient $A$ is $A_*^{-1}(1+(A-A_*)/A_*)^{-1}$, where $|A-A_*|\le\tfrac12A_*$; so Lemma~\ref{lem:one-high-factor}(iii) applies to it with $\Phi(x)=(1+x)^{-1}$ on $|x|\le\tfrac12$. The exponent $C_s$ of the reference coefficients in \eqref{eq:rewrite-local-mixed-bound} may grow with $s$. It is important to absorb the \emph{complete} product only after the expansion: $\mathcal E_m^{C_s}\le C_sM_m$ costs one power of the weight, whereas paying for each step of the recursion separately would not give a fixed loss. Define the mixed norms \[ \bigl(\|Z\|_{t,s}^{\rm mix}\bigr)^2 =\sum_m M_m^{2t}\,H_m^{-2\beta} \sum_{a+b\le s}\|\partial_y^a\partial_\varphi^b Z_m\|_2^2 . \] The first entry of $\mathfrak a_{s,m}$ is controlled by \eqref{eq:reference-negative-material-extension}. For the second, use the bound for the first radial derivative in Lemma~\ref{lem:two-weights-closed-domain} at the angular order $s-1$; its one additional angular derivative requires exactly the order $s$ of $\widetilde F$. Summing \eqref{eq:rewrite-local-mixed-bound} over the pieces yields \begin{equation}\label{eq:rewrite-global-mixed-bound} \|W\|_{t,s}^{\rm mix} \le C_{t,s}\bigl[ \|\widetilde F\|_{t+L_0+1,s} +\|F\|_{t+1,s-2}^{\rm mix} +(1+\mathfrak n_{t+s+d_{\rm tot}})\,\|F\|_{0,b_1}^{\rm mix}\bigr] , \end{equation} where the fixed number $d_{\rm tot}$ covers $L_0+1$, $d_{\rm geo}$ and the base and embedding orders. The scalar columns are compactly supported and satisfy the same estimates, and the scalars themselves are bounded by the base inverse. In the term with the high norm of the state, the weight $M^{t+L_0+1}$ is placed on the localized norm of the state, and the low norm of the field is taken in the output space of the reference inverse, with weight $-L_0$. The order $b_0$ controls the finitely many base norms that occur, which is part of requirement (N5); the finitely many orders below $b_1+2$ follow from the same computation. \paragraph{Conclusion.} Insert \eqref{eq:full-source-tame-resolvent} into the first term of \eqref{eq:rewrite-global-mixed-bound}. The indices of weight and of derivatives in that term add up to $t+s+L_0+1$, and in the source term to $t+s-1$. So an estimate for all $t+s\le r$ loses a fixed number of orders of the grading, independent of $r$. The transformations between canonical and original variables have fixed tame losses by Proposition~\ref{prop:uniform-physical-geometry}, and the jets match across the roots as explained in Section~\ref{subsec:transmission-domain}. All these losses are included in $d_*$, and we obtain \eqref{eq:readable-physical-inverse-tame}. Undoing the transformations, $\bW$ corresponds to a height $\eta$ and scalars $\mu,\lambda$ with \[ \widehat L_{\bx}\eta+\mu\,k_{\bx}+\lambda\,\psi=f, \qquad \ell_*\eta=0 . \] Integrate this equation over the annulus. The fluxes of $\eta$ vanish at the endpoints, so $\Qint\widehat L_{\bx}\eta=0$; and $\Qint k_{\bx}=0$, by differentiating the conservation law of Lemma~\ref{lem:scalar-integral} with respect to $\nu$. Hence $\lambda=\Qint f$, which is zero for $f\in\mathcal Y_\beta$. \begin{proof}[End of the proof of Proposition~\ref{prop:rewrite-volume-inverse}] Existence and the estimate \eqref{eq:readable-physical-inverse-tame} have now been proved. The factorization \eqref{eq:current-fixed-domain-factorization} on $\mathbb D_*$ shows that $G_{\bx}$ is a two-sided inverse, and uniqueness among smooth solutions follows because $\mathbb D_*$ contains all of them (Lemma~\ref{lem:two-weights-closed-domain}). Continuity in the state follows from the resolvent identity, first in the base space and then, by the higher estimates, in every order. The coefficients form a smooth tame family. A smooth tame family of linear maps which has unique solutions and a continuous tame inverse has a smooth tame inverse (Theorem~\ref{thm:hamilton-inverse} below); its hypotheses, uniqueness and continuity, are exactly what has just been verified. \end{proof}