\section{From the linear seed to a nonlinear solution} \label{sec:solving-the-equation} \label{sec:exact-return-height} We can now explain why the unknowns were chosen as in Table~\ref{tab:unknowns}, and solve the system. The height solves the curvature equation. The compact metric parameter $\nu$ removes its obstruction in the second mode. One control $\kappa_m$ per negative component adjusts the mean travel angle, and the odd charts adjust the oscillating part of the return maps; these two tasks are separated by taking the angular mean. The return residual depends on the height only through $qu$, so the feedback from the return equations to the height equation carries the small factor $q$, and the coupled linear system is solved by a perturbation of a triangular one. Objects of the full system are written in bold type: the state $\bx$, the residual $\mathbf R$, the approximate inverse $\mathbf V$ and the error $\mathbf Q$. \subsection{The residual and its derivative} For a state $\bx=(u,\nu,\kappa,h_L)$ as in \eqref{eq:rewrite-state} and a parameter $q\in\R$ let \begin{equation}\label{eq:readable-joint-residual} \mathbf R(\bx,q)=(\mathcal A,\mathfrak e), \qquad \mathcal A=\mathcal F_q(w,u), \qquad \mathfrak e=\mathfrak e(qu,\kappa,h_L), \end{equation} where $w$ is the profile \eqref{eq:full-height-metric-controls} and $\mathfrak e=(\mathfrak e_m)_m$ is the return residual \eqref{eq:full-height-one-chart} on every negative component. The return residual depends on the height through $Z=qu$, on $\kappa$ and on $h_L$; it does not depend on $\nu$ or on the endpoint constant of $u$, because $\chi_*$ and the derivative of $\chi_L$ are supported in positive components. The spaces of increments and of residuals are \[ \mathfrak X=(\mathcal X_\beta\cap\ker\ell_*)\times\R \times\mathcal K_\beta\times\mathcal Z_{\beta,\mathrm{odd}}, \qquad \mathfrak Y=\mathcal Y_\beta\times\mathcal Z_{\beta,\mathrm{even}} . \] The second factor of $\mathfrak Y$ includes the angular means; the first incorporates the conservation law of Lemma~\ref{lem:scalar-integral}. By \eqref{eq:normalized-integral} and the calculus of Section~\ref{sec:fixed-geometry}, $\mathbf R$ is a smooth tame map into $\mathfrak Y$ on the open set $\mathfrak N$ of \eqref{eq:the-neighbourhood}. It vanishes at \[ q=0,\qquad \bx_0=(\Phi_*,0,0,\id), \] since $\mathscr L_{w_*}\Phi_*=0$ and the return maps of the reference profile are the rotations $R_{\pm T_m}$. \paragraph{The derivative of the return residual.} Write \begin{equation}\label{eq:joint-circle-linear-rows} C_{\bx}^{\circ}=D_Z\mathfrak e,\qquad R_{\bx}^{\circ}=\av{C_{\bx}^{\circ}},\qquad \mathcal D_{\bx}=D_\kappa\mathfrak e,\qquad \mathsf J_{\bx}=\av{\mathcal D_{\bx}} . \end{equation} The residual on the component $m$ depends only on the height near that component, on its chart and on $\kappa_m$. So $\mathsf J_{\bx}$ is a diagonal matrix. At a state with radial height and trivial charts, $\mathfrak e_m=T_m(w)-T_m(w_*)$, and by the normalization of the controls in Lemma~\ref{lem:quantitative-optical-controls}, $\mathsf J=I$ at the reference. On a neighbourhood of low order the entries of $\mathsf J_{\bx}$ stay uniformly away from zero, and $\mathsf J_{\bx}^{-1}$ is smooth tame. At the reference state, the mean $R_0^\circ$ of the derivative with respect to the height annihilates all nonzero angular frequencies, by invariance under rotation, and on a radial increment $Z$ it is \begin{equation}\label{eq:rewrite-reference-circle-row} (R_0^\circ Z)_m =-\int v_0\,\frac{1+w_*^2}{w_*}\,Z_r\,dr , \end{equation} the integral being over the $m$th negative component. Indeed, for a radial height $h+Z$ the characteristic speed \eqref{eq:completed-characteristics} is $v_0\,w_*\sqrt D/(w_*+Z_r)$ with $D=1-2w_*Z_r-Z_r^2$, and its logarithmic derivative at $Z=0$ is $-(1+w_*^2)Z_r/w_*$. In particular $R_0^\circ\Phi_*=0$. \paragraph{The corrected linear system.} Let $K_{\bx}=D_\kappa\mathcal A$ and, as before, $k_{\bx}=D_\nu\mathcal A$. For an increment $(\eta,\mu,\varkappa,\psi\circ h_L)$ of the state $(u,\nu,\kappa,h_L)$ and a right side $(f,g)\in\mathfrak Y$ we consider \begin{equation}\label{eq:corrected-coupled-linear-system} \begin{split} \widehat L_{\bx}\eta+k_{\bx}\,\mu+K_{\bx}\,\varkappa&=f, \qquad \ell_*\eta=0,\\ q\,C_{\bx}^{\circ}\eta+\mathcal D_{\bx}\,\varkappa +\mathcal M_T\psi&=g . \end{split} \end{equation} This is the derivative of $\mathbf R$ with respect to the state, except for three terms that have been dropped: the two residual corrections in \eqref{eq:readable-completed-height-derivative} and the term $\mathcal Q_h$ in \eqref{eq:full-height-chart-derivative}. Each of them is proportional to $\mathcal A$ or to $\mathfrak e$. The structure of the system is shown in Figure~\ref{fig:coupled-columns}. \begin{figure}[htbp] \centering \begin{tikzpicture}[x=1.8cm,y=.85cm,font=\small] \fill[blue!7] (0,1) rectangle (4,2); \fill[orange!10] (0,0) rectangle (4,1); \draw (0,0) rectangle (4,2); \draw (0,1)--(4,1); \foreach \x in {1,2,3} {\draw (\x,0)--(\x,2);} \node at (.5,2.4) {$\eta$}; \node at (1.5,2.4) {$\mu$}; \node at (2.5,2.4) {$\varkappa$}; \node at (3.5,2.4) {$\psi$}; \node at (.5,1.5) {$\widehat L$}; \node at (1.5,1.5) {$k$}; \node at (2.5,1.5) {$K$}; \node at (3.5,1.5) {$0$}; \node at (.5,.5) {$qC^\circ$}; \node at (1.5,.5) {$0$}; \node at (2.5,.5) {$\mathcal D$}; \node at (3.5,.5) {$\mathcal M_T$}; \node[anchor=east] at (-.1,1.5) {curvature equation}; \node[anchor=east] at (-.1,.5) {return equations}; \node[anchor=west] at (4.12,.5) {$\av{\mathcal D}=\mathsf J,\quad\av{\mathcal M_T\psi}=0$}; \end{tikzpicture} \caption{The corrected linear system and its four columns: the increments $\eta,\mu,\varkappa,\psi$ of the height, of $\nu$, of the controls $\kappa$ and of the charts. Taking the mean of the second row removes the chart column. The feedback from the second row to the height is multiplied by $q$.} \label{fig:coupled-columns} \end{figure} \subsection{The control column gains all coefficient weights} Let $G_{\bx}$ be the inverse of Proposition~\ref{prop:rewrite-volume-inverse} and put $U_{\bx}=G_{\bx}K_{\bx}$: the response $(\eta,\mu)$ of the first equation to an increment of the controls. The natural space for increments of the controls is \[ B_\beta=\Bigl\{\varkappa=(\varkappa_m):\ \|\varkappa\|_{B_\beta}^2=\sum_mH_m^{-2\beta}|\varkappa_m|^2<\infty\Bigr\}, \] which has no weight $M$ at all. \begin{lemma}[Smoothing by the control column]\label{lem:control-smoothing} For every $s$, with a fixed loss $d_*$, \begin{equation}\label{eq:weak-control-smoothing} \|K_{\bx}\varkappa\|_{\mathcal Y_\beta,s} +\|U_{\bx}\varkappa\|_{\mathcal X_\beta,s}+|(U_{\bx}\varkappa)_{\mu}| \le C_s\,(1+\mathfrak n_{s+d_*})\,\|\varkappa\|_{B_\beta} , \end{equation} where the middle term is the norm of the height component of $U_{\bx}\varkappa$ and the last term is its scalar component. \end{lemma} \begin{proof} \step{1: two decaying factors} Differentiating the formula \eqref{eq:normalized-integral} for $\mathcal F_q$ with respect to the controls gives \[ K_{\bx}\varkappa=\int_0^1D_\kappa\bigl[\,r\,D_z\mathcal N_{h}(h+tqu)\bigr] [\varkappa]\;u\,dt , \] where $h'=w$ and the derivative $D_\kappa$ acts on $w$ and on $h$. Every term of the result contains the increment of the control and a derivative of positive order of the height $u$: the curvature residual does not change when a constant is added to the height. This is true also at $q=0$, where the derivative is that of $\mathscr L_wu$. On the component $m$ write $\varkappa_m=H_m^\beta b_m$. The derivative of the height contributes another factor $H_m^\beta$; the endpoint constant of $u$ does not contribute, because the supports of the controls lie in negative components. After division by the weight $H_m^\beta$ of the source space a factor $H_m^\beta$ remains. \step{2: summation} At the order $s$, all coordinate, cutoff and reciprocal factors and the derivatives of the control $\chi_m$ together cost a coefficient power $\mathcal E_m^{C_s}$ and a fixed power of $M_m$; the remaining factor $H_m^\beta$ bounds their product, by \eqref{eq:full-material-envelope}. The product estimate \eqref{eq:weighted-product} puts one factor of the state at the high order. The supports of the controls are disjoint, so summing over $m$ gives the bound for $K_{\bx}$ with $(\sum_m|b_m|^2)^{1/2} =\|\varkappa\|_{B_\beta}$. Composition with $G_{\bx}$ adds its fixed loss. \end{proof} \subsection{Solving the corrected system} \begin{proposition}[The approximate right inverse]\label{prop:approx-inverse} On $\mathfrak N$ the system \eqref{eq:corrected-coupled-linear-system} has for every $(f,g)\in\mathfrak Y$ the solution \begin{equation}\label{eq:joint-height-circle-inverse} \begin{split} \varkappa&=(I-q\,\mathsf J^{-1}R^\circ U)^{-1}\, \mathsf J^{-1}\,\bigl(\av g-q\,R^\circ Gf\bigr),\\ (\eta,\mu)&=Gf-U\varkappa,\\ \psi&=\mathcal M_T^{-1}\bigl(g-q\,C^\circ\eta-\mathcal D\varkappa\bigr), \end{split} \end{equation} where the subscript $\bx$ is omitted and $R^\circ$, $C^\circ$ act on the height component of $G$ and $U$. The map \[ \mathbf V(\bx,q)\,(f,g)=(\eta,\mu,\varkappa,\psi\circ h_L) \] is smooth tame, and linear in $(f,g)$. \end{proposition} \begin{proof} \step{1: elimination} The first row gives $(\eta,\mu)=Gf-U\varkappa$. Taking the angular mean of the second row removes $\mathcal M_T\psi$ and gives \[ (I-q\,\mathsf J^{-1}R^\circ U)\,\varkappa =\mathsf J^{-1}\,\bigl(\av g-q\,R^\circ Gf\bigr). \] By Lemma~\ref{lem:control-smoothing}, $R^\circ U$ is a bounded operator from $B_\beta$ to the \emph{same} space $B_\beta$, with norm controlled by finitely many norms of the state. Requirement (N6) is that $b_0$ controls these norms and that $\epsilon_0$, which bounds $|q|$ because $|q|\le\mathfrak n_{b_0}$, is so small that $\|q\,\mathsf J^{-1}R^\circ U\|_{B_\beta\to B_\beta}<\tfrac12$. This gives the first line of \eqref{eq:joint-height-circle-inverse}, and the second follows. By the equation for $\varkappa$, the argument of $\mathcal M_T^{-1}$ in the third line has mean zero on every component, so \eqref{eq:rewrite-circle-inverse} applies. \step{2: regularity} Higher norms of $\varkappa$ are read off from the identity \[ \varkappa=\mathsf J^{-1}\bigl(\av g-q\,R^\circ Gf\bigr) +q\,\mathsf J^{-1}R^\circ U\varkappa , \] satisfied by the solution already found. By Lemma~\ref{lem:control-smoothing} the last term has all higher norms, bounded by $\|\varkappa\|_{B_\beta}$ and one high norm of the state; the first term has the norms of the sources with a fixed loss. So $\varkappa$ lies in all the spaces $\mathcal K_{\beta,s}$, with tame bounds, and substitution gives the estimates for $\eta$, $\mu$ and $\psi$. \step{3: smoothness} The resolvent identity gives continuity in the base space, and the identity of Step~2 with its estimates gives continuity in every order. By Theorem~\ref{thm:hamilton-inverse} below, applied to the inverse in the first line, all three formulas are smooth tame. \end{proof} \subsection{The quadratic error and the Nash--Moser theorem} We use two theorems of Hamilton. Tame spaces and smooth tame maps were defined in Section~\ref{subsec:spaces}. Implicit function theorems which compensate a loss of derivatives by smoothing go back to Nash \cite{Nash1956} and Moser \cite{Moser1966}. H\"ormander gave versions in \cite{Hormander1976,Hormander1985}; the second is the one cited in the Nash--Moser--H\"ormander iteration for local isometric embedding in \cite[Section~6]{HanHongLin2003}, see also \cite{HanHong2006}. We use Hamilton's formulation \cite{Hamilton1982} because it allows an inverse of the linearization up to a quadratic error. \begin{theorem}[Hamilton: families of inverses] \label{thm:hamilton-inverse} Let $U$ be an open subset of a tame space, let $\mathfrak H,\mathfrak K$ be tame spaces, and let $L:U\times\mathfrak H\to\mathfrak K$ be a smooth tame map which is linear in its second argument. Suppose that $L(x)$ is bijective for every $x\in U$, and that the family of inverses $(x,y)\mapsto L(x)^{-1}y$ is continuous and tame. Then this family is smooth tame. \end{theorem} This is \cite[Part~II, Theorem~3.1.1]{Hamilton1982}. Existence, uniqueness, continuity and tame bounds for the inverses are hypotheses; the theorem supplies their derivatives. It was used in this way in Propositions~\ref{prop:rewrite-volume-inverse} and~\ref{prop:approx-inverse}. \begin{theorem}[Hamilton: implicit function theorem with quadratic error] \label{thm:hamilton-quadratic} Let $\mathfrak X,\mathfrak P,\mathfrak Y$ be tame spaces and let $\mathbf R:U\subset\mathfrak X\times\mathfrak P\to\mathfrak Y$ be a smooth tame map on an open set. Suppose there are smooth tame maps $\mathbf V(\bx,q)\,\mathbf y$ with values in $\mathfrak X$, linear in $\mathbf y\in\mathfrak Y$, and $\mathbf Q(\bx,q)\,[\mathbf r,\mathbf y]$ with values in $\mathfrak Y$, bilinear in $(\mathbf r,\mathbf y)\in\mathfrak Y\times\mathfrak Y$, such that \begin{equation}\label{eq:rewrite-hamilton-hypothesis} D_{\bx}\mathbf R(\bx,q)\,\mathbf V(\bx,q)\,\mathbf y =\mathbf y+\mathbf Q(\bx,q)\,[\mathbf R(\bx,q),\mathbf y] \end{equation} for all $(\bx,q)\in U$ and all $\mathbf y$. Then near every zero $(\bx_0,q_0)$ of $\mathbf R$ there is a smooth tame map $q\mapsto\bx(q)$ with $\bx(q_0)=\bx_0$ and $\mathbf R(\bx(q),q)=0$. \end{theorem} This is \cite[Part~III, Theorem~3.3.1]{Hamilton1982}. The point of the hypothesis is that $\mathbf V$ need only be an inverse \emph{up to an error proportional to the residual}; but the identity must hold exactly, with the two arguments of $\mathbf Q$ independent. The theorem gives a solution, not its uniqueness, which we do not need. \begin{proposition}[The quadratic error identity]\label{prop:quadratic-error} For $\mathbf r=(\mathbf r_1,\mathbf r_2)\in\mathfrak Y$ and $\mathbf y=(f,g)\in\mathfrak Y$, let $(\eta,\mu,\varkappa,\psi)$ be given by \eqref{eq:joint-height-circle-inverse} and define \[ \mathbf Q(\bx,q)\,[\mathbf r,\mathbf y] =\Bigl(q\,\partial_\theta\bigl(B\,\mathbf r_1\,\eta_\theta\bigr) +\mathscr E_{\bx}(\mathbf r_2)\,\eta,\ \ \mathcal Q_h(\mathbf r_2,\psi)\Bigr), \] where in $\mathcal Q_h$ the points of evaluation in \eqref{eq:full-height-chart-derivative} are those of the actual state and only the explicit factors $\mathfrak e$, $\mathfrak e'$ are replaced by $\mathbf r_2$, $\mathbf r_2'$, as was done for $\mathscr E_{\bx}$ in Proposition~\ref{prop:optical-residual-completion}. Then $\mathbf Q$ is a smooth tame family of bilinear maps $\mathfrak Y\times\mathfrak Y\to\mathfrak Y$, and \eqref{eq:rewrite-hamilton-hypothesis} holds. \end{proposition} \begin{proof} For the increments \eqref{eq:joint-height-circle-inverse}, the identities \eqref{eq:readable-completed-height-derivative} and \eqref{eq:full-height-chart-derivative} give the true derivative \begin{equation}\label{eq:joint-height-circle-quadratic-error} D_{\bx}\mathbf R\,(\eta,\mu,\varkappa,\psi\circ h_L) =\Bigl(f+q\,\partial_\theta\bigl(B\,\mathcal A\,\eta_\theta\bigr) +\mathscr E_{\bx}(\mathfrak e)\,\eta,\ \ g+\mathcal Q_h(\mathfrak e,\psi)\Bigr), \end{equation} which is \eqref{eq:rewrite-hamilton-hypothesis} with $\mathbf r=\mathbf R=(\mathcal A,\mathfrak e)$. The increments are linear in $\mathbf y$, so $\mathbf Q$ is bilinear. It takes values in $\mathfrak Y$. Its first component has integral zero: one term is an angular derivative, the other is in divergence form with compact support inside the negative components. Its second component is even. Finally, every term is a product of two factors each of which decays like $H^\beta$, so by the calculus of Section~\ref{sec:fixed-geometry} and the estimates for flows, $\mathbf Q$ is smooth tame with values in the spaces of weight $\beta$. \end{proof} \begin{theorem}[The solution branch]\label{thm:weak-branch} There are $q_1>0$ and a smooth tame map \begin{equation}\label{eq:rewrite-weak-branch} q\mapsto\bx(q)=(u_q,\nu_q,\kappa_q,h_{L,q}),\qquad |q|