\section{The endpoint gain and the closed surface} \label{sec:final-proof} The branch of Theorem~\ref{thm:weak-branch} was constructed with the weight $H^\beta$, because the reference inverse is available only for exponents below one. This already suffices for the geometry: every element of $\mathcal X_\beta$ is flat at the endpoints. In this section we first show that the corrections to the metric are in fact of size $q^2H^{2\beta}$, which quantifies how the final metric differs from the reference one, and then we state the annular theorem in its precise form and prove Theorems~\ref{thm:annulus-intro} and~\ref{thm:main}. The gain comes from two identities for angular means. The first recovers the mean of the height from an expression which is quadratic in the oscillating part. The second recovers the controls $\kappa$ from that mean and from another quadratic expression. A product of two factors of size $H^\beta$ has size $H^{2\beta}$, and $2\beta>1$. In this section a quantity of size $H^\beta$ is called \emph{weak} and one of size $H^{2\beta}$ \emph{strong}; both are flat at the endpoints. \subsection{Two identities for means} \paragraph{The mean of the height.} Let $\mathfrak F_w(U)$ be the connection period \eqref{eq:cart-mean-path-flux}. By \eqref{eq:mean-period-derivative} its derivative at $U=0$ is $w\av{U_r}$. Let \begin{equation}\label{eq:posterior-period-bilinear} \mathcal B_{w,U}(v_1,v_2) =\int_0^1(1-t)\,D_U^2\mathfrak F_w(tU)[v_1,v_2]\,dt \end{equation} be its Taylor remainder of second order. For a solution the period vanishes identically, because its radial derivative is the mean of the residual and it is zero at the endpoints. Hence, with $u^{\rm av}=\av u$, \begin{equation}\label{eq:mean-recovered-derivative} 0=\mathfrak F_w(qu)=q\,w\,(u^{\rm av})' +q^2\,\mathcal B_{w,qu}(u,u), \qquad\text{that is,}\qquad (u^{\rm av})'=-\frac qw\,\mathcal B_{w,qu}(u,u) . \end{equation} \paragraph{The mean of the return residual.} Use the additive coordinates $\mathbf v=(Z,\kappa,a)$ with $Z=qu$ and $h_L=\id+a$, and put \[ \mathcal C(\mathbf v)=\av{\mathfrak e(\mathbf v)}, \qquad \mathcal B^\circ_{\mathbf v}(v_1,v_2) =\int_0^1(1-t)\,D^2\mathcal C(t\mathbf v)[v_1,v_2]\,dt . \] By Section~\ref{sec:solving-the-equation} the derivative of $\mathcal C$ at $\mathbf v=0$ is the identity on $\kappa$, zero on the odd increments of the charts, and $R_0^\circ$ on the mean of the height. At a solution therefore \begin{equation}\label{eq:weak-posterior-control} 0=\kappa+q\,R_0^\circ u^{\rm av} +\mathcal B^\circ_{\mathbf v}(\mathbf v,\mathbf v) . \end{equation} Formulas \eqref{eq:mean-recovered-derivative} and \eqref{eq:weak-posterior-control} already explain the exponent $2\beta$. What remains is to verify the corresponding estimates in the stronger topology, with all derivatives in the parameter, so that the conclusion holds for the whole branch. \subsection{Two weak factors give one strong remainder} Fix a convex neighbourhood of $(U,\nu,\kappa,a)=(0,0,0,0)$ in the spaces of weight $\beta$, small in one norm of finite order, on which $D$ and $Q/w$ are positive and the charts are uniformly increasing. Both Taylor remainders above are integrals over segments in this neighbourhood; the map $\mathcal C$ does not depend on $\nu$. \begin{lemma}[Quadratic remainders in the doubled weight] \label{lem:posterior-two-weight} The maps $(w,U,v_1,v_2)\mapsto w^{-1}\mathcal B_{w,U}(v_1,v_2)$ and $(\mathbf v,v_1,v_2)\mapsto\mathcal B^\circ_{\mathbf v}(v_1,v_2)$ are smooth tame with values in the space $\mathcal X^0_{2\beta}$ of radial functions and in $\mathcal K_{2\beta}$, respectively. For every number $k$ of derivatives with respect to the state there is a loss $d_k$, independent of the order $s$, such that \begin{equation}\label{eq:posterior-multilinear-bound} \begin{split} &\bigl\|D^k\mathcal B[\eta_1,\ldots,\eta_k]\, (\eta_{k+1},\eta_{k+2})\bigr\|_{2\beta,s}\\ &\quad\le C_{s,k}\Bigl[\, \sum_{i=1}^{k+2}\|\eta_i\|_{s+d_k} \prod_{j\ne i}\|\eta_j\|_{b+d_k} +(1+\|\mathbf y\|_{s+d_k}) \prod_j\|\eta_j\|_{b+d_k}\Bigr], \end{split} \end{equation} where $\mathcal B$ is either of the two maps, $\mathbf y$ is the state, the norms of the arguments are those of weight $\beta$, and compact scalars are measured by their absolute values. The same bound holds for all derivatives of $\mathcal C$ of order at least two. \end{lemma} \begin{proof} \step{1: counting weak factors} On the component $m$, every increment of a control or of a chart, and every derivative of positive order of a height, carries a factor $H_m^\beta$. The endpoint constants of heights are harmless: only derivatives of positive order enter, and those of $\chi_L$ are supported in a compact positive region and vanish on the supports of all controls on negative components. Differentiation does not destroy a weak factor: for $u=H^\beta\widetilde u$, \[ H^{-\beta}\,\Dr u=\Dr\widetilde u+\beta\,(\Dr\log H)\,\widetilde u , \] and $\Dr\log H$ is bounded. Products and restrictions to the pieces cost fixed numbers of derivatives and reference factors $\mathcal E_m^{C_{s,k}}$. So it suffices to show that every term of the expressions in question contains two weak factors, one from each of the two distinguished arguments: these give the weight $H_m^{2\beta}$ of the output; the highest derivative is placed on one argument or on one coefficient of the state; the complete product of coefficients is absorbed into one power of $M$; and finite overlap together with $\|ab\|_{\ell^2}\le\|a\|_{\ell^2}\|b\|_{\ell^\infty}$ sums over the components. \step{2: the period} With $k=w+U_r$, $p=U_\theta/r$ and $E=1+w^2$, the integrand of $\mathfrak F_w(U)$ is \[ \frac{1+kp_\theta/(1-p^2)}{\sqrt{E(1-p^2)-k^2}}-1 , \] a smooth function of $w$ and of the derivatives $(U_r,U_\theta/r,U_{\theta\theta}/r)$ on the neighbourhood. Each second derivative with respect to $U$ contains one such derivative from each argument, and further differentiation with respect to the state does not remove them. The factor $w^{-1}=w_*^{-1}\exp(-\nu\chi_*-\sum_m\kappa_m\chi_m)$ is part of the product of coefficients; it costs a coefficient power, not an endpoint exponent. \step{3: flows} For the map $\mathcal C$ consider a normalized characteristic flow $\dot\Theta=G(t,\Theta;\mathbf y)$, in the form \eqref{eq:fixed-strip-root-flow} near a root. Its first and second variations with respect to the state satisfy \[ \dot\Theta_i=G_\theta\,\Theta_i+G_i, \qquad \dot\Theta_{12}=G_\theta\,\Theta_{12} +G_{\theta\theta}\,\Theta_1\Theta_2 +(G_1)_\theta\,\Theta_2+(G_2)_\theta\,\Theta_1+G_{12} . \] The integrating factor of the homogeneous equation is bounded above and below, by Lemma~\ref{lem:uniform-characteristic-flow-constants}. The source $G_i$ contains one weak factor and $G_{12}$ two, so by variation of constants $\Theta_i$ contains one and $\Theta_{12}$ two. Composition of flows does not change the weight of the component. \step{4: inverses} If $\iota=h^{-1}$, then \begin{equation}\label{eq:posterior-inverse-variation} \iota_i=-\frac{h_i\circ\iota}{h'\circ\iota},\qquad \iota_{12}=-\frac{h_{12}\circ\iota +(h_1'\circ\iota)\,\iota_2+(h_2'\circ\iota)\,\iota_1 +(h''\circ\iota)\,\iota_1\iota_2}{h'\circ\iota} . \end{equation} The denominator is bounded below, and every term of the second numerator contains both variations. Apply this to $f^{-1}$ and $h_R^{-1}$ in \eqref{eq:full-height-one-chart}; taking even or odd parts and means preserves the weights. \step{5: higher variations} Label the directions of differentiation by a finite set $I$. In the chain rule for $G(t,\Theta;\mathbf y)$, a subset $J\subset I$ acts directly on the state and the blocks $B$ of a partition of $I\setminus J$ act on the flow, giving terms \[ \bigl(D_{\mathbf y}^{|J|}\partial_\theta^{\ell}G\, [\eta_j:j\in J]\bigr)\circ\Theta\; \prod_{B}\Theta_B , \] where $\ell$ is the number of blocks. Keep the term $G_\theta\,\Theta_I$ on the left side; all other terms involve only variations of lower order and contain every label exactly once. By induction both distinguished factors are retained, and the identity for inverses has the same structure. The additional weak factors are bounded in low norms. The tame estimates for products and compositions put only one factor at the high order, and the formulas for the coefficients, the regularization at the roots and the restrictions have fixed order; $k$ variations add a cost depending on $k$ but not on $s$. This proves \eqref{eq:posterior-multilinear-bound}. \step{6: smoothness} These are estimates for the actual derivatives. Taylor's formula with integral remainder and one more variation gives continuity and differentiability in the norms of weight $2\beta$, and by iteration smoothness in the stronger Fr\'echet topology. \end{proof} \subsection{Recovering the mean and the controls} \begin{lemma}[A primitive keeps the endpoint exponent] \label{lem:weighted-primitive} Let $\gamma>0$. For a radial function $f\in\mathcal X^0_\gamma$ let $\mathscr Pf$ be its primitive which vanishes at the right endpoint. Then $\mathscr Pf=v+a_L\chi_L$ with $v\in\mathcal X^0_\gamma$ and $a_L\in\R$, and $\mathscr P:\mathcal X^0_\gamma\to\mathcal X_\gamma$ is a tame linear map on radial functions. \end{lemma} This is a Hardy inequality with weights \cite{Muckenhoupt1972}. For the weights of this paper, which are exponential in the variable $\varrho$, the proof is short. \begin{proof} Near an endpoint, orient $\varrho$ towards it. Then $(\log H)'\to-1$ in the variable $\varrho$, so far enough out \[ \frac{d}{d\varrho}\log H^\gamma\le-\frac\gamma2 : \] the ratio of $H^\gamma$ at $\sigma\ge\varrho$ to its value at $\varrho$ is at most $e^{-\gamma(\sigma-\varrho)/2}$. Near the right endpoint $\mathscr Pf(\varrho)=-\int_\varrho^\infty f\,\ell\,d\sigma$, where the factor $\ell\le d_\partial$ comes from $dr=\ell\,d\varrho$ and is bounded. So the kernel of $\mathscr P$, conjugated by the weight $H^{-\gamma}$, is bounded by $Ce^{-\gamma(\sigma-\varrho)/2}$ for $\sigma\ge\varrho$ and vanishes otherwise. Its integrals in either variable are bounded, and Schur's test gives the bound in $L^2(d\varrho)$. One derivative of the primitive is $\ell f$; further derivatives with respect to $\varrho$ fall on $f$, on the weight and on $\ell$, whose logarithmic derivatives are bounded. A factor $M^c$ with $c\ge0$ causes no loss, since $M(\varrho)\le M(\sigma)$ for $\sigma\ge\varrho$. Near the left endpoint the same argument applies to $\mathscr Pf-a_L$, where $a_L=-\int_{r_-}^{r_+}f\,dr$ is a continuous functional of $f$, because this difference is the primitive which vanishes at the left endpoint. On the compact middle part integration is bounded in every Sobolev norm, and the derivative of $\chi_L$ is supported there. \end{proof} \paragraph{The mean.} By \eqref{eq:mean-recovered-derivative} the mean of the height on the solution branch is given by the explicit map \begin{equation}\label{eq:rewrite-mean-recovery} \mathcal U(\bx,q) =-q\,\mathscr P\bigl[w^{-1}\mathcal B_{w,qu}(u,u)\bigr], \qquad u_q^{\rm av}=\mathcal U(\bx(q),q) . \end{equation} By Lemma~\ref{lem:posterior-two-weight}, $\mathcal U$ is smooth with values in $\mathcal X_{2\beta}$, and it is $O(q)$ in every norm; in particular the difference of the endpoint values of $u_q$ is $O(q)$. \paragraph{The controls.} The operator $R_0^\circ$ of \eqref{eq:rewrite-reference-circle-row} involves only the derivative of its argument. It maps the part of $\mathcal U$ in $\mathcal X^0_{2\beta}$ to $\mathcal K_{2\beta}$ with a fixed loss, and annihilates the multiple of $\chi_L$, whose derivative is supported in a positive component. By \eqref{eq:weak-posterior-control}, \begin{equation}\label{eq:rewrite-control-recovery} \kappa_q=-q\,R_0^\circ\,\mathcal U(\bx(q),q) -\mathcal B^\circ_{\mathbf v(q)}(\mathbf v(q),\mathbf v(q)) . \end{equation} Since $\mathbf v(q)=O(q)$ on the branch, both terms on the right are $O(q^2)$ in every norm of $\mathcal K_{2\beta}$. The identity also shows that $q\mapsto\kappa_q$ is smooth with values in $\mathcal K_{2\beta}$, and that $\dot\kappa_0=0$, where the dot denotes the derivative with respect to $q$. \paragraph{The compact parameter.} Differentiating the equation $\mathcal F_q(w_q,u_q)=0$ at $q=0$, \[ \mathscr L_{w_*}\dot u_0+\dot\nu_0\,k_*+K_*\,\dot\kappa_0 +\partial_q\mathcal F_q(w_*,\Phi_*)\big|_{q=0}=0, \qquad \ell_*\dot u_0=0 . \] The last term is quadratic in the derivatives of $\Phi_*$, which contain only the angular frequency two; products of $\cos2\theta$ and $\sin2\theta$ contain only the frequencies zero and four. So this term has no component in the second mode. Since also $\dot\kappa_0=0$, the second mode of the equation is the homogeneous bordered equation of Proposition~\ref{prop:rad-fixed-theorem}, whose only solution is zero. Hence $\dot\nu_0=0$ and $\nu_q=O(q^2)$. \begin{figure}[htbp] \centering \begin{tikzpicture}[x=6.6cm,y=1cm,font=\small,>=Stealth] \draw[->] (0,0)--(1.75,0) node[right] {endpoint exponent}; \fill[blue!8] (.5,-.12) rectangle (1,.12); \foreach \x/\lab in {0/0,.5/{1/2},.75/{\beta=3/4},1/1,1.5/{2\beta=3/2}} {\draw (\x,-.12)--(\x,.12); \node[below] at (\x,-.16) {$\lab$};} \draw[->,thick,orange!85!black] (.75,.4) to[bend left=30] (1.5,.4); \node[above] at (1.12,.95) {two weak factors in each remainder of a mean}; \node[align=center] at (.75,-1.05) {admissible range\\$1/2<\beta<1$}; \end{tikzpicture} \caption{The two roles of the exponent $\beta=\tfrac34$. Exponents near $\beta$ and below one are needed for the reference inverse. The equations for the means multiply two factors of weight $\beta$ and give an exponent above one, which yields the bound for the metric in Theorem~\ref{thm:new-annular-height}. Flatness of the solution needs only a positive exponent.} \label{fig:exponent-gain} \end{figure} \subsection{The annular theorem} For a profile $w$ let $h_w$ denote the radial function with $h_w'=w$, with a fixed normalization of the additive constant. \begin{theorem}[A nonradial height with flat ends] \label{thm:new-annular-height} Let $w_*$, $\phi_*$, $\chi_*$ be as in Proposition~\ref{prop:rad-fixed-theorem} and let $\chi_m$ be the controls of Lemma~\ref{lem:quantitative-optical-controls}. There are $q_0>0$ and a smooth family \[ (u_q,\nu_q,\kappa_q,h_{L,q}),\qquad |q|0$. \item \textup{(Profile.)} $w_q>0$ on $(r_-,r_+)$, and $w_q$ extends smoothly by zero across both endpoints. The zeros of $w_q'$ are those of $w_*'$, and $w_q=w_*$ near each of them. \item \textup{(Height.)} $z_q$ is even and $\pi$-periodic in $\theta$, and has a constant extension at each endpoint; the two constants may differ. Moreover $\ell_*u_q=1$, so $z_q$ is not radial for $q\ne0$. \item \textup{(Returns.)} On every negative component $m$, the two characteristic return maps $f_{q,m,\pm}$ of the linearized equation at $z_q$ satisfy \begin{equation}\label{eq:annular-return-condition} h_{R,q,m}\circ f_{q,m,\pm}\circ h_{L,q,m}^{-1}=R_{\pm T_m} \end{equation} for odd increasing charts commuting with the translation by $\pi$, where $T_m$ is the travel angle of $w_*$. \item \textup{(Estimates.)} For every $s$, \begin{equation}\label{eq:annular-branch-estimates} \|u_q-\Phi_*\|_{\mathcal X_\beta,s}\le C_s\,|q|,\qquad |\nu_q|+\|\kappa_q\|_{\mathcal K_{2\beta},s}\le C_s\,q^2 , \end{equation} and the family is smooth in $q$ for these topologies. In particular, for all $a\ge0$ and $\epsilon>0$, \[ |\partial_r^a(w_q-w_*)|\le C_{a,\epsilon}\,q^2\,H^{1-\epsilon} \] near the endpoints, and $w_q-w_*=O(q^2)$ in every $C^a$ norm. \end{enumerate} \end{theorem} \begin{proof} The family is the branch of Theorem~\ref{thm:weak-branch}. Choose $q_0$ so that $[-q_0,q_0]$ lies in its domain; every seminorm of the branch is bounded on this compact interval, with constants depending on the order but with $q_0$ independent of the order. \emph{Estimates.} The first bound in \eqref{eq:annular-branch-estimates} is contained in Theorem~\ref{thm:weak-branch}; the second and the smoothness in the stronger topology follow from \eqref{eq:rewrite-mean-recovery}, \eqref{eq:rewrite-control-recovery} and the bound for $\nu_q$. On the tails $\chi_*$ vanishes, and differentiating the formula for $w_q$ gives $|\partial_r^a(w_q-w_*)|\le C_a\,q^2\,H^{2\beta}E_0^{C_a}$; since $2\beta>1$, \eqref{eq:full-material-envelope} gives the stated bound. On the compact middle part use the smoothness of the family and $\nu_q=O(q^2)$. \emph{Profile.} The controls vanish near every turning circle, so the germ of the metric there is unchanged, and by Lemma~\ref{lem:common-geometric-neighborhood} the sign of $w_q'$ is that of $w_*'$ elsewhere. Flatness of $w_q$ follows from that of $w_*$ and of the controls. \emph{Equation and returns.} At a zero of $\mathbf R$ both residual corrections vanish, so $\widehat L$ is the true linearization and its exact returns are \eqref{eq:annular-return-condition}. The inequality $D>0$ of Lemma~\ref{lem:common-geometric-neighborhood} is equivalent to $|dz_q|_{g_q}<1$, and $\mathcal A=0$ means that the curvature residual vanishes, which by Proposition~\ref{prop:scalar-core} is the flatness of the remainder metric. \emph{Height.} The symmetry is that of the function spaces. Since $\ell_*u_q=1$, the second angular mode of $u_q$ is not zero. By Lemma~\ref{prop:full-physical-tame} every element of $\mathcal X_\beta$ has constant extensions at the endpoints; the same holds for all derivatives in $q$. Finally $h_{w_q}$ has constant extensions because $w_q$ is flat. \end{proof} \subsection{Proof of Theorems~\ref{thm:annulus-intro} and~\ref{thm:main}} The construction involves two parameters with different roles, and it is worth separating them once more. The small parameter $q$ was used to construct one nonradial isometric realization of one rotationally symmetric metric. The phase $\tau$ moves that realization. \begin{proof}[Proof of Theorem~\ref{thm:annulus-intro}] Fix $q$ with $0<|q|