\section{External results and the order of choices} \label{app:dependencies} The proof is self-contained except for the results listed in Table~\ref{tab:external}. The table records each of them in the form in which it is used, so that the reader can see exactly what is taken from the literature. The rigidity and flexibility theorems quoted in the introduction give context only; none of their conclusions is used. \renewcommand{\arraystretch}{1.12}\setlength{\tabcolsep}{4pt} \begin{longtable}{@{}p{.21\linewidth}p{.47\linewidth}p{.25\linewidth}@{}} \caption{External results in the form used.}\label{tab:external}\\ \toprule Result & Statement used & Where\\ \midrule \endfirsthead \toprule Result & Statement used & Where\\ \midrule \endhead Gauss equation and structure equations & For a surface with metric $g$ and second fundamental form $\mathrm{II}$ in a three-manifold, $K_g=\overline K+\det_g\mathrm{II}$, with $\overline K$ the sectional curvature of the tangent plane. For an oriented orthonormal coframe with $d\eta_1=\alpha\wedge\eta_2$, $d\eta_2=-\alpha\wedge\eta_1$: $d\alpha=-K\,\eta_1\wedge\eta_2$. & Lemma~\ref{lem:darboux-invariant}, Proposition~\ref{prop:scalar-core}\\ Primitives and coverings & A closed one-form on a simply connected domain is exact. A proper local diffeomorphism of the plane which is the identity outside a compact set is a diffeomorphism. & Proposition~\ref{lem:disk-development}\\ Ordinary differential equations & Existence, uniqueness and smooth dependence on data and parameters; Gronwall's inequality. & radial models, flows, Proposition~\ref{prop:obs-monotone-tail}\\ Quadratic forms \cite[Theorems~2.12, 2.13]{Teschl2009} & A densely defined closed quadratic form which is bounded below is the form of a unique self-adjoint operator. & the models of Section~\ref{sec:radial-models}; Lemma~\ref{lem:radial-deep-inverse}\\ Spectral theorem \cite[Chapter~3]{Teschl2009} & For a self-adjoint operator $A$ and $z$ outside its spectrum, $\|(A-z)^{-1}\|=\dist(z,\sigma(A))^{-1}$. & Lemmas~\ref{lem:central-gap}, \ref{lem:morse-arithmetic}\\ Weyl's theorem \cite[Theorem~6.19]{Teschl2009} & If the difference of the resolvents of two self-adjoint operators is compact, their essential spectra coincide. & Section~\ref{subsec:morse}\\ Harmonic oscillator \cite[Section~8.3]{Teschl2009} & The operator $-\partial_x^2+\omega^2x^2$ on the line has the simple eigenvalues $(2n+1)\omega$, $n\ge0$, and no other spectrum. & Section~\ref{sec:worked-example}, Lemma~\ref{lem:central-gap}\\ Countable products & A sequence in a countable product of compact intervals has a subsequence which converges in every coordinate. & Lemma~\ref{lem:radial-product-selection}\\ Three-lines lemma & Let $T_z$ be a family of bounded operators such that, for every vector $v$, the function $z\mapsto T_zv$ is analytic for $0<\operatorname{Re}z<1$, and continuous and bounded for $0\le\operatorname{Re}z\le1$. If $\|T_z\|\le M_0$ for $\operatorname{Re}z=0$ and $\|T_z\|\le M_1$ for $\operatorname{Re}z=1$, then $\|T_z\|\le M_0^{1-\vartheta}M_1^{\vartheta}$ for $\operatorname{Re}z=\vartheta$. Boundedness on the closed strip cannot be omitted. & Lemma~\ref{lem:ang-base-inverse-regularity}\\ Extension operator \cite{Seeley1964} & On a half-line or an interval there is a linear extension operator to the line which is bounded in every Sobolev norm $H^s$, $s\ge0$. & after Lemma~\ref{lem:one-high-factor}\\ Hamilton's theorem on families of inverses & Theorem~\ref{thm:hamilton-inverse}. & Propositions~\ref{prop:rewrite-volume-inverse}, \ref{prop:approx-inverse}\\ Hamilton's theorem with quadratic error & Theorem~\ref{thm:hamilton-quadratic}. & Theorem~\ref{thm:weak-branch}\\ \bottomrule \end{longtable} The Airy asymptotics \cite{DLMFAiry} quoted in Section~\ref{sec:worked-example} serve as a heuristic only. \paragraph{Two elementary facts used repeatedly.} \emph{Schur's test.} If the blocks $K_{nm}$ of an operator matrix satisfy $\sup_n\sum_m\|K_{nm}\|\le A$ and $\sup_m\sum_n\|K_{nm}\|\le B$, then the operator is bounded on the $\ell^2$ sum with norm at most $\sqrt{AB}$: apply the Cauchy--Schwarz inequality with the weights $\|K_{nm}\|$ in each row, and then the bound for the columns. \emph{Neumann series.} If $\|R\|<1$ on a Banach space, then $\sum_{k\ge0}(-R)^k$ converges in norm to a two-sided inverse of $I+R$. \paragraph{The order of the choices.} The constants of the construction are chosen in the following order; no later choice changes an earlier one. \begin{enumerate}[label=\textup{\arabic*.}] \item \emph{The models.} The parameters of the central oscillator and of the Morse models, the exponent $12$ of the joins and the constant $c_0$ of the matching angle (Sections~\ref{sec:radial-models}, \ref{sec:radial-tails}). \item \emph{The profile.} The exponent $\ea$ of the source weight, then $C_*$, then the initial index $N$, then the cuts of the central model and the lengths of the cone heads; then the amplitude ratios $\sigma_n$, by Lemma~\ref{lem:radial-product-selection}; then the connectors. These choices use only $|\gamma|\le1$ and are therefore the same for all endpoint exponents (Lemma~\ref{lem:radial-uniform-exponents}). \item \emph{The reference inverse.} A compact interval of endpoint exponents $\gamma\in(0,1)$ changes the constants of the estimates, not the profile. The final border $(g_*,\ell_*)$ is fixed after the geometry (Section~\ref{subsec:final-border}). \item \emph{The spaces.} The exponent $\beta=\tfrac34$, the root zones of width $\delta_n^{\rm root}$, the controls $\chi_n$ and the cutoffs (Section~\ref{subsec:root-geometry}). \item \emph{The neighbourhood.} The order $b_0$ and the number $\epsilon_0$ of \eqref{eq:the-neighbourhood}, by the requirements (N1)--(N6) of Section~\ref{subsec:neighbourhood}, in this order. The last of them also restricts the parameter $q$. \item \emph{The losses.} The number $d_*$, as the largest of the finitely many losses listed at the beginning of Section~\ref{sec:stable-inverse}. \end{enumerate} Estimates of higher order change constants and use a fixed shift of the order; they never require a new choice of the profile, of the neighbourhood or of the range of $q$. The two inequalities $\tfrac12<\beta<1$ serve different purposes: $\beta<1$ leaves room for two neighbouring exponents at which the reference inverse is available, and $\beta>\tfrac12$ makes the product of two decaying factors decay faster than $H$.