\section{Reaching a flat endpoint with an infinite chain of cells} \label{sec:radial-tails} \label{sec:radial-selection} By Proposition~\ref{prop:obs-monotone-tail} the profile must have infinitely many turning circles near each endpoint. We obtain them by repeating one Morse cell infinitely often, with rapidly shrinking lengths and amplitudes. This section constructs such a \emph{tail} and analyses the equation $L_ju=0$ on it. Its results are the following. \begin{itemize} \item The tail is a smooth profile which extends by zero across the endpoint, with explicit bounds on all coefficients (Lemmas~\ref{lem:cell} and~\ref{lem:tail-flat}). \item On the part of the tail which is deep enough for a given frequency $j$, the equation reduces \emph{exactly} to a three-term recurrence with a dominant diagonal. Hence there is a unique solution decaying towards the endpoint, it loses the factor $e^{-2n^2+o(n^2)}$ across cell $n$, and the operator has a bounded inverse (Lemma~\ref{lem:radial-deep-inverse}). \item The accumulated loss $e^{-2n^3/3}$ beats every fixed number of derivatives, so the decaying solution is flat at the endpoint. \item The amplitudes of the cells can be chosen so that, at every frequency, the decaying solution is quantitatively transverse to the solution arriving from the interior (Lemma~\ref{lem:radial-product-selection}). \end{itemize} \subsection{The budget of lengths, amplitudes and errors} For $n\ge N$ let $d_n$ be the least integer which is not a square and is at least $e^{n^2}$; thus $d_n=e^{n^2}+O(1)$. Let $A_n$ and $B_n$ be the Morse parameters \eqref{eq:rad-eq-rm5} for $d=d_n$, so that \[ A_n\asymp e^{-n^2},\qquad B_n=d_n^{-9}\asymp A_n^9 . \] The amplitude of cell $n$ is \begin{equation}\label{eq:rad-eq-js2} \lambda_n=e^{-n^{5/2}}\prod_{k=N}^n\sigma_k, \qquad 1\le\sigma_k\le2 . \end{equation} The \emph{amplitude ratios} $\sigma_k$ are free parameters until Section~\ref{subsec:height-selection}; they change $\log\lambda_n$ by at most $O(n)$, which does not affect any of the comparisons of scales below. The initial index $N$ will be taken large. Table~\ref{tab:endpoint-budget} lists the scales and the reason for each. The choices can be understood before any computation. A derivative of fixed order on a cell of length $e^{-n^2}$ costs $e^{C_sn^2}$. For the profile to be flat its amplitude must decay faster than that, and the exponent $n^{5/2}$ achieves it. The solution must decay faster still, in order to pay for reciprocals of the coefficients; its cubic exponent is not chosen but comes out of the recurrence. These requirements leave the whole interval $2<5/2<3$ for the middle exponent; no optimal value is singled out. \begin{table}[htbp] \centering \begin{tabularx}{\linewidth}{@{}l l Y@{}} \toprule Quantity & Size & Purpose\\ \midrule Length of cell $n$ & $A_n\asymp e^{-n^2}$ & The lengths are summable, so the cells accumulate at a finite endpoint.\\ Amplitude of cell $n$ & $\lambda_n=e^{-n^{5/2}+O(n)}$ & Beats the cost $e^{C_sn^2}$ of $s$ derivatives: the profile is flat.\\ Length of a join & $\asymp A_n^{12}$ & A large change of amplitude is crossed with negligible transfer error.\\ Matching angle & $\delta_n=c_0e^{-n^2}$ & The excluded sets of amplitude ratios are summable over all frequency bands.\\ Decay of the solution & $H_n=e^{-S_n}$, $S_n=\tfrac23n^3$ & Produced by the recurrence; gives integrability and flatness of the selected solution.\\ Coefficient allowance & $\mathcal E_n=e^{n^{5/2}\log(e+n)}$ & Absorbs fixed powers of coefficients and of changes of coordinates, while staying below $e^{\varepsilon S_n}$.\\ \bottomrule \end{tabularx} \caption{The scales of the tail. Their compatibility is expressed by \eqref{eq:rad-scale-separation} and verified in this section.} \label{tab:endpoint-budget} \end{table} All comparisons of scales concern a fixed number of derivatives; the constants may grow with that number. \subsection{One cell} \label{subsec:one-cell} Let $A=A_n$, $B=B_n$ and $\rho=A(e^{-X}-e^{-2X})-B$. Let $X_L0$, and we retain the interval \[ X_c\le X\le X_R+c_RA^{-4},\qquad X_c=-\log\tfrac32, \] for a fixed sufficiently small $c_R>0$. It consists of three parts (Figure~\ref{fig:cell}): \begin{itemize} \item the \emph{short positive ramp} $[X_c,X_L]$, where $\rho<0$ and $w$ rises; \item the \emph{negative component} $(X_L,X_R)$, where $\rho>0$ and $w$ falls; its length in $X$ is about $\log(A/B)\asymp n^2$; \item the \emph{long positive ramp} $[X_R,X_R+c_RA^{-4}]$, where $-B\le\rho<0$ and $w$ rises slowly. \end{itemize} \begin{figure}[htbp] \centering \begin{tikzpicture}[x=1cm,y=1cm,>=Stealth,font=\small] \fill[teal!9] (.5,0) rectangle (2.3,2.9); \fill[orange!12] (2.3,0) rectangle (6.4,2.9); \fill[teal!9] (6.4,0) rectangle (10.4,2.9); \fill[teal!22] (10.4,0) rectangle (11.3,2.9); \draw[->] (0,0)--(12.4,0) node[right] {$r$}; \draw[thick,blue!70!black] (.5,1.55) .. controls (1.2,1.8) and (1.8,2.15) .. (2.3,2.2) .. controls (3.4,2.3) and (4.9,.5) .. (6.4,.42) .. controls (7.8,.36) and (9.4,.62) .. (10.4,.72) .. controls (10.75,.76) and (11.0,2.0) .. (11.3,2.5); \draw[thick,blue!70!black,dashed] (11.3,2.5)--(11.9,2.75); \draw[thick,blue!70!black,dashed] (.05,1.4)--(.5,1.55); \foreach \x/\lab in {.5/X_c,2.3/X_L,6.4/X_R}{ \draw (\x,-.08)--(\x,.08); \node[below] at (\x,-.08) {$\lab$};} \draw (10.4,-.08)--(10.4,.08); \node[below] at (10.2,-.08) {$X_R+c_RA^{-4}$}; \node[align=center] at (1.4,3.35) {short\\positive ramp}; \node[align=center] at (4.35,3.35) {negative\\component}; \node[align=center] at (8.4,3.35) {long\\positive ramp}; \node[align=center] at (10.85,3.35) {join}; \node at (1.4,.9) {$\rho<0$}; \node at (4.0,.9) {$\rho>0$}; \node at (8.4,1.4) {$-B\le\rho<0$}; \node[blue!70!black] at (3.3,2.55) {$w$}; \draw[decorate,decoration={brace,mirror,amplitude=5pt}] (.5,-.75)--(6.4,-.75) node[midway,below=6pt] {$I_n$: mass interval}; \draw[decorate,decoration={brace,mirror,amplitude=5pt}] (6.4,-.75)--(11.3,-.75) node[midway,below=6pt] {$J_n$: resistance interval}; \end{tikzpicture} \caption{One cell and the join following it, in the direction of increasing $r$ (not to scale). The tick marks give the values of the model coordinate $X$. Each of the three parts of the cell has radial length comparable to $A_n$; the join has length comparable to $A_n^{12}$ or $A_{n\pm1}^{12}$. Along the join the profile rises by a factor $e^{4n^2+o(n^2)}$. The intervals $I_n$ and $J_n$ are used in Section~\ref{subsec:mass-resistance}.} \label{fig:cell} \end{figure} \begin{lemma}[The cell is a radial profile]\label{lem:cell} On the retained interval $r_X>0$. More precisely, with constants independent of $n$, \begin{equation}\label{eq:rad-eq-rm6a} r_X(X_c)=\tfrac38A+O(A^2),\qquad r(X_c)=1-\tfrac{15}{16}A+O(A^2), \end{equation} and \begin{equation}\label{eq:rad-eq-rm6} \begin{gathered} r=1+O(A),\qquad cA^5\le r_X\le CA,\qquad r_X\asymp A^5\ \text{on the long positive ramp},\\ \Delta r\asymp A,\qquad \int_{X_c}^{X_R}r_X^2\,dX\asymp A^2,\qquad \int_{\text{each positive ramp}}\sqrt{-\rho}\,dX\asymp\sqrt A . \end{gathered} \end{equation} Here $\Delta r$ is the radial length of each of the three parts. \end{lemma} \begin{proof} On $[X_c,X_R]$ the integral $\int(Y-X_c)|\rho(Y)|\,dY$ is $O(A)$, so the Volterra equation \[ r(X)=1-r_X(X_R)\,(X_R-X) -\int_X^{X_R}(Y-X)\,\rho(Y)\,r(Y)\,dY \] gives $r=1+O(A)$ there. For \eqref{eq:rad-eq-rm6a} evaluate the leading terms with $t=e^{-Y}$: \[ \int_0^{3/2}(1-t)\,dt=\frac38,\qquad \int_0^{3/2}\log\frac{3/2}{t}\,(1-t)\,dt=\frac{15}{16}; \] the constant $B$ contributes only $O(B\log^2(A/B))=o(A^2)$. On the negative component $r_X(X)=r_X(X_R)+\int_X^{X_R}\rho\,r>0$. On the short positive ramp $r_{XX}=-\rho r>0$, so the minimum of $r_X$ is its value at $X_c$, which is positive by \eqref{eq:rad-eq-rm6a}. On the long positive ramp $00$ without changing $\rho$ or the operators $H_{j,M}$. Between a cell and its neighbour in the direction of increasing $r$ we insert a \emph{join}: an interval of length \begin{equation}\label{eq:rad-eq-rm7} \ell_{n,m}=c_C\,r_o\,\min(A_n,A_m)^{12},\qquad m=n\pm1, \end{equation} where $n,m$ are the indices of the two cells, $r_o$ is the radius at which the join starts, and $c_C>0$ is a fixed constant. The radius factor of the neighbour is chosen so that it starts at $r_o+\ell_{n,m}$. By Lemma~\ref{lem:cell} the ratio of consecutive radius factors is $1+O(A_n+A_m)$, so the recursion converges at either end; on the left the limiting radius may be prescribed. Hence all radii stay in a compact subset of $(0,\infty)$, the $\mathfrak r_n$ are bounded above and below, and the cells accumulate at finite endpoints. \paragraph{The profile on a join.} Let $W_L$ and $W_R$ be the values of $w$ at the two ends of a join. At the right endpoint the join leads from the end of the long ramp of cell $n$, where $w\asymp\lambda_nA_n^5$, to the beginning of cell $n+1$, where $w\asymp\lambda_{n+1}A_{n+1}$; at the left endpoint it leads from cell $n+1$ to cell $n$. In both cases the profile rises, and \begin{equation}\label{eq:height-contrast-new} \log\frac{W_R}{W_L}=4n^2+O(n^{3/2}+n). \end{equation} To match the two neighbouring profiles to infinite order, let $f_L,f_R$ be the logarithms of the two neighbouring model profiles, choose cutoffs $\chi_L,\chi_R$ with disjoint supports, equal to one near the left and right end of the join respectively, and define $f=\log w$ on the join by \begin{equation}\label{eq:rad-eq-rm8} f'=\chi_Lf_L'+\chi_Rf_R'+C_1\,(1-\chi_L-\chi_R), \qquad C_1=\frac{\log(W_R/W_L)-\int\chi_Lf_L'-\int\chi_Rf_R'} {\int(1-\chi_L-\chi_R)}, \end{equation} with $f=\log W_L$ at the left end. Then $f=\log W_R$ at the right end, and $w=e^f$ agrees with the two models to infinite order. The two integrals in the numerator are $O(\ell_{n,m}/\min(A_n,A_m))$, while the required change of amplitude is of order $n^2$. Hence $C_1>0$, and after $N$ is increased \begin{equation}\label{eq:join-slope} \frac{w'}{w}\ge\frac2r\qquad\text{on every join.} \end{equation} \paragraph{Why the exponent $12$.} The exponent in \eqref{eq:rad-eq-rm7} looks extravagant; it is chosen so that a join is invisible to every frequency that has to cross it carefully. The contrast of amplitudes across the join costs a factor $e^{4n^2+o(n^2)}$. The largest frequency assigned to cell $n$ (Section~\ref{subsec:bands}) satisfies $j^2\le e^{n^2+o(n^2)}$. The length of the join contributes $e^{-12n^2+o(n^2)}$. The product is $e^{-7n^2+o(n^2)}$, far below the matching angle $e^{-n^2}$ that we shall allow. The same choice makes the mass and the resistance of a join negligible in Section~\ref{subsec:mass-resistance}. No minimal exponent is needed. \begin{lemma}[Flatness and coefficient bounds]\label{lem:tail-flat} Let $d_\partial$ be the distance to the endpoint. On cell $n$ and on the joins adjacent to it, \begin{equation}\label{eq:geometry-endpoint-distance} cA_{n+1}\le d_\partial\le CA_n, \qquad \log\frac{d_0}{d_\partial}=n^2+O(n) \end{equation} for a fixed $d_0>0$, and for every $s\ge0$ \begin{equation}\label{eq:geometry-profile-jets} \begin{gathered} e^{-n^{5/2}-Cn^2}\le w\le e^{-n^{5/2}+Cn^2},\\ |\partial_r^sw|\le C_s\,e^{-n^{5/2}+C_sn^2},\qquad |\partial_r^s(w^{-1})|\le C_s\,e^{n^{5/2}+C_sn^2}. \end{gathered} \end{equation} The constants do not depend on the amplitude ratios $\sigma_k\in[1,2]$. Consequently $w$ extends smoothly by zero across the endpoint, and all fixed-order derivatives of $w$ and of $w^{-1}$ on cell $n$ are bounded by coefficient powers $C_s\mathcal E_n^{C_s}$. \end{lemma} \begin{proof} The length of cell $n$ is comparable to $A_n$ by Lemma~\ref{lem:cell}, and $A_{n+1}/A_n=e^{-2n-1}(1+o(1))$; summing the lengths of the cells beyond $n$ gives \eqref{eq:geometry-endpoint-distance}. All lengths of intervals and their reciprocals are bounded by $e^{Cn^2}$. On a model $\partial_r=r_X^{-1}\partial_X$ and $r_X\ge c\,\mathfrak r_nA_n^5$, and the differential equation \eqref{eq:rad-eq-rm4} bounds every fixed derivative of $\log r_X$ by $e^{C_sn^2}$. On a join \eqref{eq:rad-eq-rm8} gives the same bound for $\log w$. A derivative of $e^{\pm\log w}$ is $e^{\pm\log w}$ times a polynomial in the derivatives of $\log w$, which proves \eqref{eq:geometry-profile-jets}. Since $e^{-n^{5/2}+C_sn^2}d_\partial^{-k}\to0$ for all $s,k$ by \eqref{eq:geometry-endpoint-distance}, the extension by zero is smooth. The last statement follows from $n^2=o(q_n)$. \end{proof} \subsection{Why a mass--resistance chain appears} \label{subsec:mass-resistance} We now study $L_ju=0$ on a tail. The equation has the unpleasant feature that its zeroth-order coefficient $j^2w'/r$ is enormous on the joins. A change of the unknown removes it. \begin{lemma}[Gauge]\label{lem:gauge} Let $y=\log r$ and $u=r^{j^2}V$. Then \begin{equation}\label{eq:rad-eq-rm13} r^2L_ju=r^{j^2}\,\frac{w}{\mathsf p_j}\, \bigl[(\mathsf p_jV_y)_y+\Lambda\,\mathsf p_jV\bigr], \qquad \mathsf p_j=w\,r^{2j^2-1},\qquad \Lambda=j^2(j^2-1). \end{equation} \end{lemma} \begin{proof} In the variable $y$ one has $r^2L_ju=wu_{yy}+(w_y-w)u_y-j^2w_yu$. Substituting $u=e^{j^2y}V$, the terms containing $w_yV$ cancel, and what remains is $e^{j^2y}\,[wV_{yy}+((2j^2-1)w+w_y)V_y+(j^4-j^2)wV]$. \end{proof} So $L_ju=0$ is equivalent to the Sturm--Liouville equation $-(\mathsf p_jV_y)_y=\Lambda\,\mathsf p_jV$, in which the profile enters only through the density $\mathsf p_j$. For $j=2$ this density is $\mathsf p=wr^7$ and $\Lambda=12$. \paragraph{Mass and resistance.} Divide each cell into two intervals (Figure~\ref{fig:cell}): the \emph{mass interval} $I_n$, consisting of the short positive ramp and the negative component, and the \emph{resistance interval} $J_n$, consisting of the long positive ramp and the join that follows it in the direction of increasing $r$. For a density $\mathsf p$ define \begin{equation}\label{eq:mass-resistance-def-new} \mu_n=\int_{I_n}\mathsf p\,dy,\quad R_n^I=\int_{I_n}\frac{dy}{\mathsf p},\quad \nu_n=\int_{J_n}\mathsf p\,dy,\quad R_n=\int_{J_n}\frac{dy}{\mathsf p}. \end{equation} We call $\mu_n,\nu_n$ \emph{masses} and $R_n^I,R_n$ \emph{resistances}. The names come from the following elementary observation about an equation $-(\mathsf pV_y)_y=\Lambda\mathsf pV$. \begin{itemize} \item On an interval of small resistance the value $V$ is nearly constant, and the flux $\mathsf pV_y$ changes by $-\Lambda$ times the mass times that value. \item On an interval of small mass the flux is nearly constant, and the value changes by the flux times the resistance. \end{itemize} In our tails $I_n$ carries almost all the mass and almost no resistance, and $J_n$ the opposite. Let $a_n$ be the value of $V$ on $I_n$ and $\varphi_n$ the flux on $J_n$. The two rules give \[ \varphi_n=\varphi_{n-1}-\Lambda\mu_na_n,\qquad a_{n+1}=a_n+R_n\varphi_n . \] Eliminating the fluxes yields a three-term recurrence for the $a_n$ whose diagonal coefficient $\Lambda\mu_n$ is very large compared with the off-diagonal coefficients $1/R_n$. For a solution that decays, the dominant balance is $a_{n+1}\approx-a_n/(\Lambda\mu_{n+1}R_n)$: the values alternate in sign and decrease rapidly (Figure~\ref{fig:mass-resistance-chain}). Section~\ref{subsec:exact-reduction} turns this picture into an exact statement. The picture is that of a vibrating string. In the variable $x=\int dy/\mathsf p$ the equation reads $-V_{xx}=\Lambda\,\mathsf p^2\,V$: the resistance of an interval is its length in $x$, and its mass is $\int\mathsf p^2\,dx=\int\mathsf p\,dy$. A string whose mass sits at isolated points satisfies a three-term recurrence, as in Stieltjes's work on continued fractions; for the theory of strings with a general distribution of mass see \cite{KacKrein1974}. We use only the picture, not that theory. \begin{figure}[htbp] \centering \begin{tikzpicture}[x=1cm,y=1cm,>=stealth] \draw[->] (0,0)--(12.1,0); \node[align=center,font=\small] at (10.75,1.45) {towards the\\flat endpoint}; \draw[fill=orange!20] (.2,.25) rectangle (2.5,.85); \draw[fill=blue!13] (2.5,.25) rectangle (5.25,.85); \draw[fill=orange!20] (5.25,.25) rectangle (7.25,.85); \draw[fill=blue!13] (7.25,.25) rectangle (9.7,.85); \draw[dashed] (9.7,.55)--(11.5,.55); \node at (1.35,.55) {$I_n$: mass}; \node at (3.87,.55) {$J_n$: resistance}; \node at (6.25,.55) {$I_{n+1}$}; \node at (8.48,.55) {$J_{n+1}$}; \foreach \x/\lab in {.2/a_n,5.25/a_{n+1},9.7/a_{n+2}} {\draw (\x,-.1)--(\x,.15);\node[below] at (\x,-.1) {$\lab$};} \node[align=center,font=\small] at (1.35,1.6) {value nearly\\constant,\\flux changes}; \node[align=center,font=\small] at (3.87,1.6) {flux nearly\\constant,\\value changes}; \node[align=center] at (6,-1.2) {$b_n y_n=c_{n-1}y_{n-1}+c_ny_{n+1}$,\qquad $a_n=(-1)^ny_n$\\[2pt] $b_n\approx\Lambda\mu_n$, \qquad $c_n\approx R_n^{-1}$}; \end{tikzpicture} \caption{The intervals of a tail and the node equation satisfied by the values $a_n$ of a solution at the entrances of the mass intervals. Large products $\Lambda\mu_nR_n$ make successive node values of the decaying solution small.} \label{fig:mass-resistance-chain} \end{figure} \begin{lemma}[Masses and resistances of the tail]\label{lem:mass-res} For the density $\mathsf p=wr^7$ on the tail constructed above, \begin{equation}\label{eq:rad-eq-rm10} \mu_n\asymp\lambda_nA_n^2,\qquad R_n^I\le\frac{C}{\lambda_n}\log\frac{A_n}{B_n},\qquad \nu_n\le C\lambda_nA_n^6,\qquad R_n\asymp\frac1{\lambda_nA_n^{4}}, \end{equation} hence \begin{equation}\label{eq:rad-eq-rm11} \mu_nR_n\asymp A_n^{-2},\qquad \mu_nR_n^I\le CA_n^2\log\frac{A_n}{B_n},\qquad \nu_nR_n\le CA_n^2, \end{equation} and \begin{equation}\label{eq:rad-eq-rm12} \log(\mu_{n+1}R_n)=2n^2+O(n^{3/2}+n). \end{equation} \end{lemma} \begin{proof} Since $dy=dr/r$ and $dr=r_X\,dX$, the measures in \eqref{eq:mass-resistance-def-new} are $wr^6\,dr$ and $dr/(wr^8)$, and on a model $w\,dr=\lambda r_X^2\,dX$ and $dr/w=dX/\lambda$. The radius is bounded above and below, so \eqref{eq:rad-eq-rm6} gives \eqref{eq:rad-eq-rm10} for the model parts. A join of length $\ell$ adds at most $W_R\ell$ to the mass and $\ell/W_L$ to the resistance of $J_n$; relative to $\nu_n$ and $R_n$ these are $e^{-7n^2+O(n^{3/2}+n)}$ and $e^{-11n^2+O(n)}$. The products \eqref{eq:rad-eq-rm11} follow, and so does \eqref{eq:rad-eq-rm12} from $\log(\lambda_{n+1}/\lambda_n)=O(n^{3/2})$ and $\log(A_{n+1}^2/A_n^4)=2n^2+O(n)$. \end{proof} \paragraph{Other frequencies.} For a general frequency the density is $\mathsf p_j$. After division by the constant $r_\partial^{2(j^2-4)}$, where $r_\partial$ is the radius of the endpoint, \[ \frac{\mathsf p_j}{\mathsf p} =\Bigl(\frac r{r_\partial}\Bigr)^{2(j^2-4)}, \] and on the part of the tail beyond cell $N'$ the logarithm of this ratio is bounded by $Cj^2A_{N'}$. So Lemma~\ref{lem:mass-res} holds for $\mathsf p_j$, with uniform constants, on every part of the tail where $j^2A_{N'}$ is bounded. We shall use the slightly stronger condition \begin{equation}\label{eq:deep-frequency-condition-new} j^2A_{N'}\sqrt{\log(A_{N'}/B_{N'})}\le c \end{equation} for a small constant $c$; by \eqref{eq:rad-eq-rm11} it makes $\Lambda\mu_nR_n^I$ and $\Lambda\nu_nR_n$ small for all $n\ge N'$. \subsection{The exact reduction to a recurrence} \label{subsec:exact-reduction} Use the coordinate $x=\log r$ at the right endpoint and $x=-\log r$ at the left endpoint, so that $x$ always increases outward. Definition~\ref{def:deep-tail} below is formulated for an abstract sequence of intervals $I_n,J_n$ in outward order, in which a resistance interval $J_n$ follows each mass interval $I_n$. At the right endpoint these are the intervals of the cells. At the left endpoint the outward order of the intervals of the cells is $J_N,I_N,J_{N+1},I_{N+1},\ldots$; there the definition is applied with the interval $J_{n+1}$ of the cells in the place of its $J_n$, and with $J_N$ as a separate \emph{initial resistance edge} $J_0$, whose resistance and mass are denoted $R_0$ and $\nu_0$. The shift of index affects only the lower-order terms of \eqref{eq:rad-eq-rm12}. The following definition isolates what the reduction needs. \begin{definition}[Deep tail]\label{def:deep-tail} Let $\mathsf p>0$ be a density on a half-line in $x$ which is divided into consecutive intervals $I_n,J_n$ ($n\ge N'$), possibly preceded by an initial resistance edge $J_0$, and let $\Lambda>0$. With the masses and resistances \eqref{eq:mass-resistance-def-new} formed with $dx$, the data form a \emph{deep tail} with constants $\eta\ll1\ll K$ if for all $n$ \begin{equation}\label{eq:rad-eq-dg1} \begin{gathered} \Lambda\mu_nR_n^I+\Lambda\nu_nR_n\le\eta,\qquad \Lambda\mu_nR_n\ge K,\quad \Lambda\mu_{n+1}R_n\ge K,\\ \nu_n\le\eta\,\min(\mu_n,\mu_{n+1}), \end{gathered} \end{equation} and if, when there is an initial edge, $\Lambda\mu_{N'}R_0\ge K$, $\Lambda\nu_0R_0\le\eta$ and $\nu_0\le\eta\mu_{N'}$. \end{definition} By Lemma~\ref{lem:mass-res} and the remark on other frequencies, the part of a constructed tail beyond cell $N'$ is a deep tail for $\mathsf p_j$ and $\Lambda=j^2(j^2-1)$ whenever \eqref{eq:deep-frequency-condition-new} holds and $N'$ is large enough; $\eta$ and $K^{-1}$ can be made as small as we wish. \begin{lemma}[Deep inverse and the natural line] \label{lem:radial-deep-inverse} \label{lem:radial-flat-line} On a deep tail, with $\eta$ and $K^{-1}$ sufficiently small, let \[ A_{\mathsf p}=-\mathsf p^{-1}\partial_x\,\mathsf p\,\partial_x \] act in $L^2(\mathsf p\,dx)$ with the Dirichlet condition at the entrance and the Friedrichs condition at the outer endpoint. \begin{enumerate} \item $\|(A_{\mathsf p}-\Lambda)^{-1}\|\le C/\Lambda$. \item Up to multiples, the equation $(A_{\mathsf p}-\Lambda)V=0$ has exactly one solution which is square integrable for $\mathsf p\,dx$ near the outer endpoint. Its values $a_n$ at the entrances of the mass intervals $I_n$ are nonzero and alternate in sign: $a_n=(-1)^ny_n$ with $y_n>0$ and \begin{equation}\label{eq:rad-eq-dg4} \frac{y_{n+1}}{y_n} =\frac{1+O(\eta)+O(K^{-1})}{\Lambda\mu_{n+1}R_n}, \qquad \frac{\sqrt{\mu_{n+1}}\,y_{n+1}}{\sqrt{\mu_n}\,y_n}\le\frac CK<\frac12 . \end{equation} \item For the constructed tails and $\mathsf p=\mathsf p_j$, the corresponding solution $u=r^{j^2}V$ of $L_ju=0$ is smooth and flat at the endpoint. \end{enumerate} \end{lemma} The line spanned by the solution in (ii) is the \emph{natural line} of Section~\ref{subsec:layout}. \begin{proof} The reduction is carried out on the differential equation itself, so that the operator norm and the endpoint condition come out of one argument. An \emph{edge} is $\mathsf E_n=I_n\cup J_n$; its two endpoints are \emph{nodes}, and node $n$ is the entrance of $I_n$. \step{1: one edge in the resistance coordinate} On $\mathsf E_n$ introduce $ds=dx/\mathsf p$. The edge has length $L_n=R_n^I+R_n$ in $s$; with $\omega=\mathsf p^2$ we have $\mathsf p\,dx=\omega\,ds$ and \[ (A_{\mathsf p}-\Lambda)u=f \quad\Longleftrightarrow\quad -u_{ss}-\Lambda\,\omega\,u=\omega f . \] The Dirichlet Green function of $-\partial_s^2$ on the edge is $G_0(s,t)=\min(s,t)\,[L_n-\max(s,t)]/L_n$. The small quantities are \[ \delta_n^{\rm edge}=\Lambda\int\omega(s)\,G_0(s,s)\,ds\le3\eta, \qquad \frac{R_n^I}{L_n}\le\frac\eta K , \] by the first line of \eqref{eq:rad-eq-dg1}. The inequality $G_0(s,z)G_0(z,t)\le G_0(s,t)G_0(z,z)$ bounds the Neumann series of the Dirichlet Green function $G_\Lambda$ of $-\partial_s^2-\Lambda\omega$ term by term. With $l_0=1-s/L_n$, $l_1=s/L_n$, and $v_0,v_1$ the solutions of the homogeneous equation on the edge with boundary values $(1,0)$ and $(0,1)$, the same argument applies with one factor replaced by $l_i$, and we obtain \begin{equation}\label{eq:rad-eq-dg2} G_0\le G_\Lambda\le(1-\delta_n^{\rm edge})^{-1}G_0, \qquad l_i\le v_i\le(1-\delta_n^{\rm edge})^{-1}\,l_i . \end{equation} For $u$ vanishing at both ends of the edge, $\Lambda\int\omega|u|^2\le\delta_n^{\rm edge}\int|u_s|^2$. Hence the direct sum $G_D$ of the Dirichlet inverses of all edges has norm at most $C\eta/\Lambda$ on $L^2(\mathsf p\,dx)$. \step{2: the boundary matrix of an edge} For boundary values $(a,a')$ at the two ends of $\mathsf E_n$ the homogeneous solution is $av_0+a'v_1$, and its outward derivatives at the two ends, in the variable $s$, are given by the matrix \[ \begin{pmatrix}d_{n,0}&-c_n\\-c_n&d_{n,1}\end{pmatrix},\qquad c_n=\frac1{L_n}+\Lambda\int\omega\,l_0v_1,\qquad d_{n,i}=\frac1{L_n}-\Lambda\int\omega\,l_iv_i ; \] the formulas follow from Green's identity with the functions $l_i$. Since $l_0l_1=G_0(s,s)/L_n$, \eqref{eq:rad-eq-dg2} gives \begin{equation}\label{eq:rad-eq-dg3} c_n=\frac{1+O(\eta)}{R_n},\qquad b_n:=-d_{n-1,1}-d_{n,0} =\Lambda\mu_n\,\bigl(1+O(\eta)+O(K^{-1})\bigr). \end{equation} For the second formula: on $I_n$ we have $l_0=1+O(R_n^I/L_n)$, so that interval contributes $\Lambda\mu_n(1+O(\eta))$ to $\Lambda\int\omega\,l_0v_0$, and $J_n$ contributes at most $C\Lambda\nu_n\le C\eta\Lambda\mu_n$. The preceding edge contributes to $\Lambda\int\omega\,l_1v_1$ at most $C\Lambda$ times $\nu_{n-1}+\mu_{n-1}(R_{n-1}^I/L_{n-1})^2$, and the second term is at most $\eta^2K^{-2}\mu_n$ by \eqref{eq:rad-eq-dg1}. The two terms $1/L_n$ and $1/L_{n-1}$ are at most $\Lambda\mu_n/K$ each. The same computation covers an initial resistance edge. \step{3: interpolation of node values} For a sequence $a=(a_n)$ of node values let $Ta$ be the function which on each edge is the homogeneous solution with these boundary values, $(Ta)|_{\mathsf E_n}=a_nv_0+a_{n+1}v_1$. Let $\mathfrak h=\ell^2(\mu_n)$. Then \begin{equation}\label{eq:rad-eq-dj1} c\,\|a\|_{\mathfrak h}^2\le\|Ta\|_{L^2(\mathsf p\,dx)}^2 \le C\,\|a\|_{\mathfrak h}^2 . \end{equation} Indeed, by \eqref{eq:rad-eq-dg2} the function $Ta$ is close to $a_n$ on $I_n$, which carries the mass $\mu_n$. On $I_n$ the neighbouring value enters with the factor $l_1\le R_n^I/L_n$, and the sum of these contributions is $O(\eta^2/K^2)\,\|a\|_{\mathfrak h}^2$ after a shift of index. The contribution of $J_n$ is controlled by the last inequality of \eqref{eq:rad-eq-dg1}. \step{4: the node operator} The quadratic form of $A_{\mathsf p}-\Lambda$ on interpolants is, by Step~2, \[ (\mathcal Da,a)_{\mathfrak h} =-\sum_n b_n|a_n|^2-2\operatorname{Re}\sum_n c_n\,a_n\overline{a_{n+1}} , \] which defines a symmetric operator $\mathcal D$ on $\mathfrak h$. In the coordinates $\sqrt{\mu_n}\,a_n$ its diagonal entries have size $\Lambda$ by \eqref{eq:rad-eq-dg3}, while each off-diagonal entry is bounded by $C/(R_n\sqrt{\mu_n\mu_{n+1}})\le C\Lambda/K$. By diagonal dominance \[ \|\mathcal D^{-1}\|\le\frac C\Lambda , \] independently of the number of edges. \step{5: the inverse} These algebraic facts give an inverse of the operator because they respect its form domain. On the one hand, the energy of an interpolant is bounded by $C\Lambda\|a\|_{\mathfrak h}^2$, by the formula for $\mathcal D$ and \eqref{eq:rad-eq-dj1}. On the other hand, averaging $u(x_n)-u(x)$ over $I_n$, where $x_n$ is node $n$, gives \[ \mu_n\,|u(x_n)|^2\le2\,\|u\|_{L^2(I_n,\mathsf p\,dx)}^2 +2\mu_nR_n^I\int_{I_n}|u_s|^2\,ds , \] so every function in the form domain has node values in $\mathfrak h$ and splits as $u=Ta+u^\circ$ with $u^\circ$ vanishing at all nodes. The two parts are orthogonal for the form of $A_{\mathsf p}-\Lambda$, because $Ta$ solves the homogeneous equation on each edge. Therefore \begin{equation}\label{eq:rad-eq-dj2} (A_{\mathsf p}-\Lambda)^{-1}=G_D+T\,\mathcal D^{-1}T^*, \qquad (T^*f)_n=\frac1{\mu_n}\int\mathsf p\,(T\mathbf e_n)\,f\,dx , \end{equation} where $\mathbf e_n$ is the $n$th unit sequence. Truncations of the tail by a Dirichlet wall at a far node obey the same formula and converge strongly, since the diagonally dominant series for $\mathcal D^{-1}$ converges uniformly on finite compressions; their limit is the Friedrichs realization identified by the decomposition of the form domain. Steps~1, 3 and~4 now give~(i). \step{6: the decaying solution} A homogeneous solution is the interpolant of its node values, and continuity of the flux at node $n$ reads $-b_na_n=c_{n-1}a_{n-1}+c_na_{n+1}$. With $a_n=(-1)^ny_n$ this is \[ b_n\,y_n=c_{n-1}\,y_{n-1}+c_n\,y_{n+1}. \] Put $\epsilon_n=c_n^2/(b_nb_{n+1})\le C/K^2$ and $t_n=b_{n+1}y_{n+1}/(c_ny_n)$. Eliminating backwards from the outer end, \[ t_n=\frac1{1-\epsilon_{n+1}t_{n+1}} . \] For large $K$ these maps preserve the interval $[1,2]$ and contract it by a factor at most $\tfrac12$. Hence the solutions with a Dirichlet wall at a far node converge, as the wall recedes, to a unique sequence with $t_n\in[1,2]$ for all $n$. By \eqref{eq:rad-eq-dg3} it satisfies \eqref{eq:rad-eq-dg4}, and by \eqref{eq:rad-eq-dj1} its interpolant is square integrable. Any solution independent of this one has node values of the form \[ z_n=y_n\Bigl(C+\mathcal W\sum_{kN), \end{equation} and let the first band $\Band_N$ consist of all even $j$ with $2\le j<\jn_N$. The \emph{bands} $\Band_n$, $n\ge N$, partition the even frequencies. For $j\in\Band_n$: \begin{itemize} \item every earlier cell $mn$ form a deep tail, because \begin{equation}\label{eq:rad-eq-rm22} j^2A_{n+1}\sqrt{\log(A_{n+1}/B_{n+1})} \le C\,C_*^2\,q_n^2\,\frac{A_{n+1}}{A_n}\,(n+1)\longrightarrow0 . \end{equation} \end{itemize} So there is no range of frequencies left unexamined between the local spectral inverses and the deep inverse. The decomposition depends on the frequency; the profile does not. This structure is the key to the assembly of the global inverse in Section~\ref{sec:radial-reference-inverse}. \paragraph{An auxiliary cone at the entrance of a tail.} To define an incoming line for the first band, attach at the entrance of each tail an auxiliary cone $w=cr$. For the right tail the cone is continued inward to $r=0$, for the left tail outward to $r=\infty$. The transition between cone and first cell is a short positive interval with $rw'/w\ge1$, constructed by applying \eqref{eq:rad-eq-rm8} to $\log(w/r)$. On the cone, in $y=\log r$, \[ B_j=\tfrac1{12}\,(-\partial_y^2+j^2),\qquad \varpi\,dr=12c\,dy ; \] the solutions regular at the far end of the cone are $r^{j}$ for the right tail and $r^{-j}$ for the left tail, and the Dirichlet inverse of $B_j$ has norm at most $12/j^2$. The cone serves only to specify an incoming line. The final profile retains a finite piece of it, the \emph{cone head}. \subsection{Choosing the amplitudes so that the lines do not meet} \label{subsec:height-selection} Fix an endpoint, a band index $n$ and a frequency $j\in\Band_n$. Consider the join which precedes cell $n$ in the outward order (for $n=N$: the transition from the auxiliary cone). Let $W_{\rm in}$ be the value of the profile at its inner end and $\sigma_nW^0_{\rm out}$ its value at the outer end. By \eqref{eq:rad-eq-js2}, $W_{\rm in}$ does not depend on $\sigma_n$, and $W^0_{\rm out}$ is the outer value for $\sigma_n=1$. Put \begin{equation}\label{eq:rad-eq-js4} \Xi_n=\frac{W^0_{\rm out}}{W_{\rm in}},\qquad \Upsilon_n=\max(\Xi_n,\Xi_n^{-1}),\qquad \log\Upsilon_n=4n^2+O(n^{3/2}+n). \end{equation} At the first join the amplitude $c$ of the auxiliary cone is chosen, independently of $\sigma_N$, so that the contrast is the same. The radii and all cut points are independent of the amplitude ratios. The role of $\sigma_n$ is the following. Everything beyond the join is multiplied by $\sigma_n$; this does not change the homogeneous equation there, hence not the natural line. The incoming line is determined by what lies before the join, and does not depend on $\sigma_n$ either. What $\sigma_n$ changes is the scale of the flux $wu'$ by which the two lines are compared across the join. We show that this is enough to move them apart. \paragraph{A state variable adapted to the join.} On a join $w'/w$ is huge, and the pair $(u,wu')$ is not a convenient state. Define instead \begin{equation}\label{eq:rad-eq-js5} \zeta_j=\frac{ru_r}{j^2}-u . \end{equation} If $L_ju=0$, then $(w\zeta_j)'=-(j^2-1)\,wu_r/j^2$, in which $w'$ no longer appears. For the state $Y=(u,\ w\zeta_j/W_{\rm in})$ this gives \begin{equation}\label{eq:rad-eq-js7} Y_r=\frac1r \begin{pmatrix} j^2&j^2\,W_{\rm in}/w\\ -(j^2-1)\,w/W_{\rm in}&-(j^2-1) \end{pmatrix}Y . \end{equation} The coefficients involve only the sizes of $w$ and $w^{-1}$ on the join, which lie between those at its ends. In this subsection the \emph{incoming line} at the join means the line of the solution which vanishes at a cut inside the positive ramp preceding the join. The true incoming line differs from it by an exponentially small amount, which is accounted for in Section~\ref{subsec:terminal-matching}. Let \[ \delta_n=c_0\,e^{-n^2}. \] \begin{lemma}[The excluded amplitude ratios for one frequency] \label{lem:excluded-interval} Let $C_*$ be large. For every $n\ge N$, every $j\in\Band_n$ and every choice of the ratios $\sigma_k$, $k\ne n$, the set of $\sigma_n\in[1,2]$ for which the incoming and the natural line at the join preceding cell $n$ make an angle of sine less than $\delta_n$, in the state \eqref{eq:rad-eq-js7} at the outer end of the join, is contained in an interval of length at most $C\delta_n$. The constant $C$ does not depend on $n$, on $j$, on the other ratios or on the contrast $\Xi_n$. \end{lemma} \begin{proof} \step{1: the join is almost transparent} By the Volterra series, the transfer matrix of \eqref{eq:rad-eq-js7} across the join differs from the identity by at most $e^{\eta_{n,j}}-1$, where \begin{equation}\label{eq:rad-eq-js14} \eta_{n,j}=Cj^2\,\frac{\ell_n}{r_{\min}}\,\Upsilon_n \le C\,C_*^2\,q_n^2\,e^{-7n^2+O(n^{3/2}+n)}=o(e^{-n^2}), \end{equation} $\ell_n$ being the length of the join. This is the computation behind the exponent $12$ in \eqref{eq:rad-eq-rm7}: the steepness of the profile on the join causes no difficulty. \step{2: the incoming line has a definite slope} Normalize the solution which defines the incoming line to $u=1$ at the inner end of the join. Then \begin{equation}\label{eq:rad-eq-js6} a:=\frac{\zeta_j}{u}\in\bigl[-\tfrac32,-\tfrac12\bigr]. \end{equation} Indeed, on the ramp $u_{XX}=j^2|\rho|\,u$, and comparison with the Riccati equation for a constant potential gives $|u_X/u|\le Cj\sqrt b\,\coth(cj\sqrt A)$, where $b$ is the size of $|\rho|$ on the ramp and $A=A_{n-1}$. The ramp is either the short one, with $|\rho|\asymp A$, length $\asymp1$ in $X$ and $r_X\asymp A$, or the long one, with $|\rho|\asymp A^9$, length $\asymp A^{-4}$ and $r_X\asymp A^5$. In both cases $r_X\asymp\sqrt{Ab}$, so \[ \Bigl|\frac{ru_r}{j^2u}\Bigr|\le\frac{C}{j\sqrt{A_{n-1}}}<\frac12 \] once $C_*$ is large, because $j\sqrt{A_{n-1}}\ge C_*q_{n-1}$. At the left endpoint only the sign of $ru_r$ changes. In the first band the cone gives $ru_r/(j^2u)=\pm1/j$, which is at most $\tfrac12$ in absolute value even for $j=2$. \step{3: the excluded set} Let $(\beta_1,\beta_2)$ span the natural line in the frame $(u,\zeta_j)$ at the outer end of the join, computed for $\sigma_n=1$. In the common frame $Y$ the natural line is then spanned by \[ \mathbf b(\sigma)=(\beta_1,\ \sigma\,\Xi_n\,\beta_2), \qquad \sigma=\sigma_n , \] while the incoming line at the inner end is spanned by $(1,a)$; see Figure~\ref{fig:height-selection}. If $\beta_1=0$ the angle between the two lines is bounded below. Otherwise put $b_\sigma=\sigma\Xi_n\beta_2/\beta_1$; the sine of the angle is \[ \frac{|a-b_\sigma|}{\sqrt{1+a^2}\,\sqrt{1+b_\sigma^2}} . \] If this is small, $b_\sigma$ lies in a fixed compact interval of negative numbers, on which $|db_\sigma/d\sigma|=|b_\sigma|/\sigma$ is bounded below. Therefore the set of $\sigma\in[1,2]$ for which the sine is less than $2\delta_n$ is contained in an interval of length at most $C\delta_n$. By Step~1 the state of the incoming solution at the outer end of the join is $(1,a)+O(\eta_{n,j})$, and normalization of vectors is Lipschitz near $(1,a)$. Since $\eta_{n,j}=o(\delta_n)$ by \eqref{eq:rad-eq-js14}, an angle of sine less than $\delta_n$ at the outer end forces a sine less than $2\delta_n$ for the line $(1,a)$, and the claim follows. None of the bounds used depends on $j$, on the other ratios or on the size of $\Xi_n$. \end{proof} \begin{figure}[htbp] \centering \begin{tikzpicture}[x=2.1cm,y=1.5cm,>=stealth] \draw[->] (-.12,0)--(2,0) node[right] {$u$}; \draw[->] (0,.22)--(0,-1.65) node[below] {$w\zeta_j/W_{\rm in}$}; \draw[thick,blue!75!black,->] (0,0)--(1.65,-.98); \node[anchor=west,blue!75!black] at (1.25,-1.12) {incoming line $(1,a)$}; \draw[teal!65!black,->] (0,0)--(1.1,-.4); \draw[teal!65!black,->] (0,0)--(1.1,-.8); \draw[teal!65!black,->] (0,0)--(1.1,-1.2); \draw[->,teal!65!black] (1.16,-.4)--(1.16,-1.17); \node[anchor=west,align=left] at (1.25,-.35) {natural line $\mathbf b(\sigma)=(\beta_1,\sigma\Xi_n\beta_2)$:\\ $\sigma$ rescales the flux coordinate}; \end{tikzpicture} \caption{Selection of an amplitude ratio in the plane of boundary states. The slope of the incoming line lies in a fixed negative interval. As $\sigma$ varies, the natural line sweeps past it with speed bounded below, so the two lines are close only for a short interval of $\sigma$.} \label{fig:height-selection} \end{figure} \begin{lemma}[Simultaneous selection of the amplitudes] \label{lem:radial-product-selection} If $N$ is large enough, then at each endpoint the amplitude ratios $\sigma_n\in[1,2]$, $n\ge N$, can be chosen so that for every $n$ and every $j\in\Band_n$ the incoming and natural lines at the join preceding cell $n$ make an angle of sine at least $\delta_n$, in the sense of Lemma~\ref{lem:excluded-interval}. \end{lemma} \begin{proof} The band $\Band_n$ contains at most $\jn_n=C_*q_ne^{n^2/2}(1+o(1))$ frequencies, so the total length of the excluded intervals satisfies \begin{equation}\label{eq:rad-eq-js13} \sum_{n\ge N}C\,\#\Band_n\,\delta_n \le C\,C_*c_0\sum_{n\ge N}q_n\,e^{-n^2/2}<\frac12 \end{equation} for large $N$. The excluded interval of Lemma~\ref{lem:excluded-interval} for a pair $(n,j)$ depends on the ratios $\sigma_k$ with $k\ne n$, but the bound for its length does not, and this is all we need. Consider first finitely many bands $N\le n\le N_1$, with $\sigma_k=1$ for $k>N_1$, and Lebesgue measure on the cube $[1,2]^{N_1-N+1}$. By Fubini's theorem and the uniform one-dimensional bound, the set of ratios violating the condition for a given pair $(n,j)$ has measure at most $C\delta_n$, and by \eqref{eq:rad-eq-js13} the union of these sets has measure less than one. So a good choice exists for every $N_1$. Extend each of these choices by ones and pass to a subsequence converging in every coordinate. For a fixed frequency, the transfer across finitely many cells is continuous in the ratios, and the natural line is continuous in the product topology by Remark~\ref{rem:line-continuity}. Each condition on the angle is closed, hence holds for the limit. The two endpoints are treated independently. \end{proof} \subsection{Summary} We have constructed, at each endpoint, a smooth flat tail with the following properties, for every even frequency $j$ with band index $n$: beyond cell $n$ the equation $L_ju=0$ has a unique decaying solution, which is flat at the endpoint, and $L_j$ has a Dirichlet inverse with the localized bound \eqref{eq:rad-eq-dj5}; at the entrance of cell $n$ the natural line makes an angle at least $\delta_n$ with the line arriving through the preceding positive ramp. We have \emph{not} yet shown that gluing the tails to the central model preserves the spectral gaps of the pieces. That global question is the subject of the next section. Its answer uses both halves of the present construction: the selected angle at the transition cell, and the strong attenuation through the positive ramp that precedes it.