\documentclass[12pt]{article} \usepackage{amsmath, amssymb, amsthm, mathrsfs} \usepackage{geometry} \usepackage[hidelinks]{hyperref} \usepackage{enumitem} \usepackage{booktabs} \usepackage{graphicx} \geometry{margin=2.5cm} \allowdisplaybreaks \theoremstyle{plain} \newtheorem{theorem}{Theorem} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{lemma}[theorem]{Lemma} \theoremstyle{definition} \newtheorem{remark}{Remark}[section] \newtheorem{proposition}{Proposition}[section] \newtheorem{example}{Example} \newtheorem{definition}[theorem]{Definition} \newcommand{\BT}{\operatorname{BT}} \newcommand{\dev}{\operatorname{dev}} \newcommand{\HH}{\mathbb{H}} \newcommand{\RR}{\mathbb{R}} \hypersetup{ colorlinks=true, linkcolor=blue, citecolor=blue, urlcolor=blue } \title{An unconditional estimate and the structure of the B\"acklund orbit} \author{Anton Yu. Kalmykov \thanks{\small{Independent researches, Corresponding author e-mail: anton@kalmykov.net}}\\\small{Russia, Moscow}\\[0.5ex] Alexander Timmermans\\\small{Holland, Utrecht} } \date{\today} \begin{document} \maketitle \begin{abstract} Let $f$ be an arbitrary $C^\infty$ solution of the sine-Gordon equation $f_{xy}=\sin f$ in the square $W=[-L,L]^2$, and let $\omega_n$ be obtained by $n$ successive B\"acklund transformations with arbitrary real parameters $a_1,\dots,a_n>0$ and arbitrary normalization constants $c_j$. For a solution $\omega$ let $r(\omega)$ be the associated smooth map of a surface of curvature $K=-1$ (Chebyshev net, common initial frame at the origin). We prove the unconditional estimate $\|r(\omega_n)-r((-1)^n f)\|_{L^\infty(W)}\le K_n\sum_{j=1}^n 1/a_j$, $K_n=3L^2+(1+L)\pi(2T_*+\pi)3^{n-1}$, $T_*=\sup_{|x|\le L}\int_{-L}^{L}|f_y|\,dy$, with no restrictions on the $a_j$, their ratios, the $c_j$, or $L$. The proof reduces the problem to a cascade of forced pendulum equations, controlled by an integral Lyapunov estimate and by cumulative estimates for the orthogonal Gauss-Weingarten propagator. For the Amsler seed $f=Y(xy)$ we show $T_*\le 2L^2$, which gives an explicit constant; the order $1/a$ is sharp. We also study the dependence on the data: smoothness along finite-dimensional families, continuity on the Fr\'echet data space, convergence of the immersions in $C^0$ and $W^{1,p}$ although the solutions $\omega_n$ themselves need not converge, partial limits along arbitrary index sets for $a_j\to\infty$ and for $a_j\to0$, and a two-chamber structure of the orbit under B\"acklund chains. \end{abstract} \noindent\textbf{Keywords:} sine-Gordon equation; B\"acklund transformation; surfaces of constant negative curvature; Chebyshev nets; Amsler surfaces; Bianchi permutability theorem. \noindent\textbf{2020 MSC:} 35Q51; 37K35; 53A05; 53A35. \newpage \tableofcontents \newpage \section{Introduction} \subsection{Historical context} The connection between surfaces of constant negative curvature in \(\mathbb R^3\) and the sine-Gordon equation is one of the oldest topics in the geometric theory of integrable systems. If on a surface with \(K=-1\) one chooses asymptotic coordinates in which both families of asymptotic lines are parametrized by arclength (a Chebyshev net), then the angle \(\omega\) between them satisfies the equation \(\omega_{xy}=\sin\omega\); conversely, every solution defines a smooth map whose image has \(K=-1\) at the points where \(\sin\omega\ne0\) (see Remark~\ref{rem:regularity}). The classical corpus on pseudospherical coordinates and surfaces of constant curvature goes back to Darboux \cite{Darboux} and Eisenhart \cite{Eisenhart}; a modern exposition within soliton theory is in the monograph of Rogers and Schief \cite{RogersSchief}. The B\"acklund transformation---the machinery for generating new solutions of sine-Gordon (and, geometrically, new surfaces with \(K=-1\)) from a given one---was introduced by B\"acklund in his Lund paper of 1883 \cite{Baecklund1883} as a geometric transformation of pseudospherical surfaces; the triply orthogonal variant was systematically studied by Bianchi \cite{Bianchi1892}. The permutability property (the composition of two B\"acklund transformations with parameters \(a\) and \(b\) is independent of the order) is the Bianchi theorem; a modern exposition is in \cite[Chs.~1--2]{RogersSchief}. The role of the parameter \(a\) as the ``width of the transition layer'' of the kink solution and the relation to multisoliton solutions are reflected in the classical works on sine-Gordon physics: the survey of Barone--Esposito--Magee--Scott \cite{Barone}, and Hirota's exact multisoliton solution \cite{Hirota}. The general program of ``soliton surfaces''---immersions associated with spectral problems---was developed by Sym \cite{Sym}. Separate mention should be made of surfaces of constant negative curvature with specific structure: the Kuen surface \cite{Kuen} and the family of Amsler surfaces \cite{Amsler}. The Amsler seed \(f(x,y)=Y(xy)\), where \(Y\) solves \((uY')'=\sin Y\), occupies the position of a canonical example in the present work: its special structure (dependence only on \(xy\)) allows one to compute the constant \(T_*\) explicitly (Corollary~\ref{cor:amsler} below). \subsection{Statement and result} Fix \(n\ge1\) real parameters \(a_j>0\) and constants \(c_j\in\mathbb R\). Starting from the seed \(\omega_0=f\), we construct the chain \(\omega_j=\mathrm{BT}_{a_j}(\omega_{j-1})\) (system \eqref{eq:BT} below), normalized by the condition \(\omega_j(0,0)=c_j\). Each step changes the solution; the first question of the present text is: \emph{how far can the immersion \(r(\omega_n)\) move away from the immersion \(r((-1)^n f)\) of the mirror seed if the parameters \(a_j\) are large?} The answer (Theorem~\ref{thm:main}) requires no assumptions except smoothness of the seed: the norm of the difference of immersions is majorized by a quantity linear in \(\sum 1/a_j\), with an explicit constant independent of the normalization constants \(c_j\) and of the mutual ratios of the parameters, and with no restriction on the size of the window. It is essential that the constant is written in terms of the single integral characteristic of the seed \(T_*=\sup_x\int|f_y|\,dy\)---the total variation of its \(y\)-derivative along columns. The proof splits into three independent blocks. 1. \textbf{Pendulum cascade.} Along each column \(x=\mathrm{const}\), the difference of neighboring iterates is reduced to the chain of equations \(\theta_j'=-a_j\sin\theta_j-F_j\) with known forcing \(F_j\) (Lemmas~\ref{lem:7}, \ref{lem:9}). No information about the location of solutions relative to the equilibria \(\pi\mathbb Z\) is used: the only input is the integral Lyapunov estimate of Lemma~\ref{lem:8}, which asserts that for the equation \(\theta'=-a\sin\theta-F\) the quantity \(a\cdot\int|\sin\theta|\) is controlled by the constant \(\pi\) and the total integral of \(|F|\). 2. \textbf{Orthogonal propagator.} Both immersions are constructed with the Gauss--Weingarten frame with a common initial frame; the difference of the frames satisfies a linear inhomogeneous equation with a skew-symmetric matrix, for which a cumulative (not Gronwall) norm estimate holds (Lemmas~\ref{lem:4a}--\ref{lem:4b}). Independence of the difference of immersions from the integration path (Lemma~\ref{lem:4c}) makes the choice of path a free parameter of the proof. 3. \textbf{Assembly.} The error on the axis (Lemmas~\ref{lem:5}--\ref{lem:6}, contribution \(O(S\cdot L^2)\)) and the error in the columns (Lemmas~\ref{lem:10}--\ref{lem:11}, contribution \(\int d(\omega_n-\mathrm{ref}_n)\,dy\) with weight \((1+L)\)) add up to the final constant \(K_n\). The method is notable in that the estimate does not track the dynamics of the pendulums in phase (no count of transitions between the branches \(2\pi\mathbb Z\)): all phase data are reduced to the integral of \(d(\omega_n-\mathrm{ref}_n)\), which is then majorized from above. This is precisely why the result is unconditional. The second, complementary half of the paper (\S\ref{sec:smooth}) studies the dependence of the immersions on the data, for which the Gronwall-type constants are allowed to degenerate at large \(a\). \emph{Smoothness.} The dependence on the data is \(C^\infty\): the Goursat solution map \(\Psi\colon(\alpha,\beta)\mapsto f\) of the seed is a \(C^\infty\) bijection of the data space (Lemma~\ref{lem:sg-goursat-smooth}), each B\"acklund step depends \(C^\infty\) on \((a,c)\) and on the seed (Lemma~\ref{lem:bt-smooth}), the immersion depends \(C^\infty\) on \(\omega\) (Lemma~\ref{lem:immersion-smooth}), and their composition \((a,c,\alpha,\beta)\mapsto r(\omega_n)\) is \(C^\infty\) jointly in \((x,y,p)\) into every \(C^k\) along every finite-dimensional smooth family of data (Theorem~\ref{thm:smooth-finite}); along curves of data continuity follows from Corollary~\ref{cor:paths}. Beyond the finite-dimensional statement, \(\Phi_n\) is continuous---in fact locally Lipschitz in the \(C^k\)-seminorms, with a loss of \(n+1\) derivatives---on the whole product \((0,\infty)^n\times\mathbb R^n\times\mathcal D\) with its Fr\'echet topology (Theorem~\ref{thm:continuity-D}). The total limit \(a_j\to\infty\) of the immersions exists by Theorem~\ref{thm:main}---although the solutions themselves need not have a \(C^0\) limit (Proposition~\ref{prop:no-C0})---with convergence in \(C^0\) and every \(W^{1,p}\) at sharp rates (Theorem~\ref{thm:S-sharp}), and, for the model kink, an explicit first-order limit off a boundary layer of width \(O(1/a)\) (Proposition~\ref{prop:infinitesimal}). In the coordinates \(1/a_j\) the immersion extends to the closed orthant: partial limits \(a_j\to\infty\) along arbitrary index sets, uniformly in the constants (Theorem~\ref{thm:S-partial}), with a stratified Lipschitz structure at the faces (Corollary~\ref{cor:S-lipschitz}); the opposite limits \(a_j\to0\) are treated in the transposed chart (immersion \(r^\sigma\)), where no reflection occurs (Theorem~\ref{thm:S-zero}, Corollary~\ref{cor:S-zero-partial}). The resulting global picture is a reflection chamber: the immersions of the chains of a fixed seed are covered by two alternating, mirror-symmetric, connected unions glued along the trailing faces \(a_n=\infty\) (\S\ref{sec:chamber}, Theorem~\ref{thm:chamber}; their disjointness is not decided), the ``mirror'' being a proper rotation (Corollary~\ref{cor:S-mirror}). A companion interactive demonstration (numerical verification of the estimate for \(n=1,2\) on the Amsler seed and on \(\mathrm{BT}^2\)-chains) has been published in \cite{KalmykovTimmermans}; the numerical materials of the present article---in \S\ref{sec:numerics}---are illustrations and are not used in the proof. \section{Statement and main result} \textbf{Data.} \(L>0\), \(n\ge1\); \(f\in C^\infty(W)\) (here and below this means that all partial derivatives exist in the interior of \(W\) and extend continuously to the closed square; \(C^k(W)\) is understood in the same way) is an \emph{arbitrary} solution of \(f_{xy}=\sin f\) in the window \(W=[-L,L]^2\); the parameters \(a_1,\dots,a_n>0\) and the constants \(c_1,\dots,c_n\in\mathbb R\) are \emph{arbitrary}. Put \[ T_*:=\sup_{|x|\le L}\int_{-L}^{L}|f_y(x,y)|\,dy \] (finite: \(f\) is smooth on the compact set \(W\)). \textbf{Chain.} \(\omega_0:=f\); \(\omega_j\) is the solution of system \eqref{eq:BT} with parameter \(a_j\) and seed \(\omega_{j-1}\), normalized by the condition \(\omega_j(0,0)=c_j\) (existence and uniqueness: Lemma~\ref{lem:bt}): \begin{equation}\label{eq:BT}\tag{BT} (q+g)_x=-\frac2a\sin\frac{q-g}{2},\qquad (q-g)_y=-2a\sin\frac{q+g}{2}. \end{equation} \textbf{Reference.} \(\mathrm{ref}_n:=(-1)^n f\). Denote \(S:=\sum_{j=1}^{n}1/a_j\). The map \(r(\omega)\) is from \S\ref{sec:immersion} (common initial frame at the origin). \begin{theorem}\label{thm:main} \[ \|r(\omega_n)-r(\mathrm{ref}_n)\|_{L^\infty(W)} \le K_n\cdot S, \qquad K_n=3L^2+(1+L)\cdot\pi(2T_*+\pi)\cdot 3^{n-1}. \] \end{theorem} \begin{corollary}[Amsler seed]\label{cor:amsler} Let \(f(x,y)=Y(xy)\), \(Y\in C^\infty(\mathbb R)\), \((uY')'=\sin Y\), \(Y(0)=\pi s\), \(s\in(0,1)\). Then \(T_*\le 2L^2\) and \[ \|r(\omega_n)-r(\mathrm{ref}_n)\|_{L^\infty(W)} \le \bigl[3L^2+(1+L)\pi(4L^2+\pi)3^{n-1}\bigr]\cdot S. \] \end{corollary} There are no conditions on the size of \(a_j\), on their order and ratios, on the constants \(c_j\), or on the size of the window. The constant depends only on \((L,n,T_*)\); in Corollary~\ref{cor:amsler} it depends only on \((L,n)\). Plan of the paper: \S\ref{sec:prelim}---existence and uniqueness of the BT image; \S\ref{sec:immersion}---construction of the immersion and its correctness; \S\ref{sec:axis}---estimate along the axis \(y=0\); \S\S\ref{sec:column}--\ref{sec:cascade}---pendulum cascade and integral estimates in the columns; \S\ref{sec:assembly}---assembly; \S\ref{sec:proofs}---completion of the proof of Theorem~\ref{thm:main}; \S\ref{sec:remarks}---limitations of the result; \S\ref{sec:smooth}---dependence on the seed and the parameters: smooth dependence, continuity on the data space, the total and partial limits, the infinitesimal regime; \S\ref{sec:chamber}---the B\"acklund reflection chamber; \S\ref{sec:numerics}---numerical illustrations (not part of the proof). \section{Preliminary facts}\label{sec:prelim} \subsection{Elementary inequalities} Denote \(d(t):=\operatorname{dist}(t,2\pi\mathbb Z)\), \(d_\pi(t):=\operatorname{dist}(t,\pi\mathbb Z)\). \begin{equation}\label{eq:E1}\tag{E1} d_\pi(t_1+t_2)\le d_\pi(t_1)+d_\pi(t_2). \end{equation} \emph{Proof.} Choose \(k_i\in\mathbb Z\) with \(|t_i-k_i\pi|=d_\pi(t_i)\); then \((k_1+k_2)\pi\in\pi\mathbb Z\) and \(|t_1+t_2-(k_1+k_2)\pi|\le d_\pi(t_1)+d_\pi(t_2)\). \(\square\) \begin{equation}\label{eq:E2}\tag{E2} d_\pi(t)\le(\pi/2)\cdot|\sin t|. \end{equation} \emph{Proof.} Both sides are \(\pi\)-periodic and even, so it suffices to take \(t\in[0,\pi/2]\), where \(\sin t\ge 2t/\pi\) (concavity of \(\sin\)). \(\square\) \begin{equation}\label{eq:E3}\tag{E3} |e^{i\alpha}-e^{i\beta}|\le d(\alpha-\beta). \end{equation} \emph{Proof.} \(|e^{i\alpha}-e^{i\beta}|=2|\sin((\alpha-\beta)/2)|\) is even and \(2\pi\)-periodic in \(\alpha-\beta\), and for \(\tau\in[0,\pi]\) we have \(2\sin(\tau/2)\le\tau\). \(\square\) \begin{equation}\label{eq:E4}\tag{E4} d(2t)=2\cdot d_\pi(t). \end{equation} \emph{Proof.} \(d(2t)=\min_k|2t-2\pi k|=2\cdot\min_k|t-\pi k|\). \(\square\) \subsection{Existence and uniqueness of the BT image} \begin{lemma}\label{lem:bt} Let \(g\in C^\infty(W)\) solve \(g_{xy}=\sin g\), \(a>0\), \(c\in\mathbb R\). Then on \(W\) there exists a unique \(q\in C^\infty(W)\) satisfying \eqref{eq:BT} and \(q(0,0)=c\). It solves \(q_{xy}=\sin q\). \end{lemma} \begin{proof} Put \(u:=(q+g)/2\), \(w:=(q-g)/2\). Then \eqref{eq:BT} is equivalent to the system for a \emph{single} function \(u\): \[ u_x=A(x,y,u):=-\frac1a\sin(u-g),\qquad u_y=B(x,y,u):=g_y-a\sin u, \] with \(q=2u-g\), \(w=u-g\) (indeed, \(w_y=u_y-g_y=-a\sin u\iff(q-g)_y=-2a\sin u\)). \emph{Compatibility.} The partial derivatives for fixed \(u\) are: \[ A_y=\frac1a\cos(u-g)\,g_y,\quad A_u=-\frac1a\cos(u-g),\quad B_x=g_{xy}=\sin g,\quad B_u=-a\cos u. \] Then \[ A_y+A_uB =\frac1a\cos(u-g)g_y-\frac1a\cos(u-g)(g_y-a\sin u) =\cos(u-g)\sin u, \] \[ B_x+B_uA=\sin g+\cos u\sin(u-g). \] The difference equals \[ \cos(u-g)\sin u-\cos u\sin(u-g)-\sin g =\sin(u-(u-g))-\sin g=0. \] Thus \begin{equation}\label{eq:K}\tag{K} A_y+A_uB=B_x+B_uA \end{equation} identically. \emph{Construction.} \(A\) and \(B\) are smooth. On the compact set \(W\) we have the explicit bounds \(|A|\le1/a\) and \(|A_u|\le1/a\), so \(A\) is bounded and Lipschitz in \(u\) uniformly in \((x,y)\); by the standard global existence-uniqueness theorem for ODEs, the equation \(u_x=A(x,0,u)\) with initial value \(u(0,0)=(c+g(0,0))/2\) has a unique solution on the whole interval \(x\in[-L,L]\). Then for each \(x\) solve \(u_y=B(x,y,u)\) in \(y\in[-L,L]\) with initial value \(u(x,0)\); here \(|B_u|=a|\cos u|\le a\) and \(g_y\) is bounded on \(W\), so \(B\) is globally Lipschitz in \(u\) and the solution exists on the whole interval. The solution depends smoothly on the parameter \(x\), hence \(u\in C^\infty(W)\). \emph{Equality \(u_x=A\) everywhere.} Let \(\Phi:=u_x-A(x,y,u)\); \(\Phi(x,0)=0\) by construction. Differentiating \(u_y=B(x,y,u)\) with respect to \(x\), we obtain \(u_{xy}=B_x+B_u u_x\), while \(\partial_y[A(x,y,u)]=A_y+A_uB\). Hence \[ \Phi_y=B_x+B_u u_x-A_y-A_uB =B_u\Phi+(B_x+B_uA-A_y-A_uB)=B_u\Phi \] by \eqref{eq:K}. The linear homogeneous ODE with \(\Phi(x,0)=0\) gives \(\Phi\equiv0\). \emph{Uniqueness.} Any solution of \eqref{eq:BT} gives a solution of the system for \(u\); by the uniqueness theorem for ODEs it coincides with the constructed one along the axis, and then along the columns. \emph{\(q\) solves SG.} From the system: \(u_x=-(1/a)\sin w\), \(w_y=-a\sin u\). Then \[ u_{xy}=-\frac1a\cos w\,w_y=\cos w\sin u,\qquad w_{yx}=-a\cos u\,u_x=\cos u\sin w, \] whence \(q_{xy}=u_{xy}+w_{xy}=\sin(u+w)=\sin q\). \end{proof} By induction, all \(\omega_j\in C^\infty(W)\) are uniquely defined and solve SG. \section{The immersion map \(r(\omega)\)}\label{sec:immersion} For \(\omega\in C^\infty(W)\), define the matrices (the rows of \(F\) are the vectors \(e_1,e_2,n\)): \[ P(p)=P_0+pJ,\qquad P_0=\begin{pmatrix}0&0&0\\0&0&1\\0&-1&0\end{pmatrix}, \qquad J=\begin{pmatrix}0&-1&0\\1&0&0\\0&0&0\end{pmatrix}, \] \[ Q(\omega)= \begin{pmatrix} 0&0&\sin\omega\\ 0&0&-\cos\omega\\ -\sin\omega&\cos\omega&0 \end{pmatrix}. \] \textbf{Definition of \(r(\omega)\).} 1. \emph{Axis.} \(F(\cdot,0)\) solves \(F_x=P(\omega_x(x,0))F\), \(F(0,0)=I\); \(r(x,0):=\int_0^x e_1(\xi,0)\,d\xi\). 2. \emph{Columns.} For each \(x\), the matrix \(F(x,\cdot)\) solves \(F_y=Q(\omega(x,y))F\) with initial value \(F(x,0)\) from step 1; \(r(x,y):=r(x,0)+\int_0^y(\cos\omega\,e_1+\sin\omega\,e_2)\,dt\). The frame at the origin is common for all \(\omega\): \(F(0,0)=I\), \(r(0,0)=0\). \textbf{On the axis \(\omega\) enters only through \(\omega_x\); in the columns, only through \(\sin\omega\) and \(\cos\omega\).} \begingroup \renewcommand{\thetheorem}{4a} \begin{lemma}[norms]\label{lem:4a} (a) \(P(\cdot)\) and \(Q(\cdot)\) are skew-symmetric, hence \(F(x,y)\) is orthogonal and \(\|F\|_{\mathrm{op}}=1\). (b) The operator norm of a skew-symmetric \(3\times3\) matrix equals the Euclidean length of its axis vector. Hence \[ \|P(p)\|=\sqrt{1+p^2},\qquad \|Q(\omega)\|=1, \] \[ \|P(p)-P(p')\|=|p-p'|,\qquad \|Q(\omega)-Q(\omega')\| =|e^{i\omega}-e^{i\omega'}| \le d(\omega-\omega') \] by \eqref{eq:E3}. \end{lemma} \endgroup \begin{proof} (a) Since \(P(p)\) and \(Q(\omega)\) are skew-symmetric, the solution of \(F'=MF\) with skew-symmetric \(M\) satisfies \[ (F^\top F)'=F^\top(M^\top+M)F=0. \] With \(F(0)=I\) this gives \(F^\top F=I\), so \(F\) is orthogonal and \(\|F\|_{\mathrm{op}}=1\). (b) For a skew-symmetric \(3\times3\) matrix \(M=(m_{ij})\), define its axis vector \(\alpha:=(m_{32},m_{13},m_{21})\). A direct entry-by-entry verification gives \((Mv)_i=(\alpha\times v)_i\) for all \(v\in\mathbb R^3\), hence \(|Mv|\le|\alpha|\,|v|\) with equality when \(v\perp\alpha\); therefore \(\|M\|_{\mathrm{op}}=|\alpha|\). For \(P_0=\begin{pmatrix}0&0&0\\0&0&1\\0&-1&0\end{pmatrix}\) this gives \(\alpha_{P_0}=(-1,0,0)\); for \(J\) it gives \(\alpha_J=(0,0,1)\); for \(Q(\omega)\) it gives \(\alpha_{Q(\omega)}=(\cos\omega,\sin\omega,0)\). Since \(P(p)=P_0+pJ\) has axis \(\alpha_{P_0}+p\alpha_J=(-1,0,p)\), we obtain \[ \|P(p)\|=\sqrt{1+p^2},\qquad \|Q(\omega)\|=1. \] The difference formula \(\|P(p)-P(p')\|=|p-p'|\) follows from linearity in \(p\), and \(\|Q(\omega)-Q(\omega')\|=|e^{i\omega}-e^{i\omega'}|\le d(\omega-\omega')\) follows from \eqref{eq:E3}. \end{proof} \begingroup \renewcommand{\thetheorem}{4b} \begin{lemma}[orthogonal propagator]\label{lem:4b} If \(M(t)\) is skew-symmetric and \(D'=M(t)D+R(t)\), then \[ \|D(t)\|\le \|D(0)\|+\left|\int_0^t\|R(\tau)\|\,d\tau\right|. \] \end{lemma} \endgroup \begin{proof} For a fixed vector \(v\), put \(u(t)=D(t)v\). Then \[ \frac12\frac{d}{dt}\|u\|^2 =\langle u,Mu\rangle+\langle u,Rv\rangle =\langle u,Rv\rangle, \qquad \Bigl|\frac12\frac{d}{dt}\|u\|^2\Bigr|\le \|u\|\,\|R\|\,\|v\|, \] because \(\langle u,Mu\rangle=0\) (this gives a two-sided bound, valid for both signs of \(t\)). Hence, for $\epsilon>0$ and the smooth positive function $\nu_\epsilon:=(\|u\|^2+\epsilon^2)^{1/2}$ (so that no differentiability issue arises at zeros of $u$), \[ |\nu_\epsilon'|=\left|\frac{\langle u,u'\rangle}{\nu_\epsilon}\right|\le\|R\|\,\|v\|. \] Integrating gives $\nu_\epsilon(t)\le\nu_\epsilon(0)+\|v\|\bigl|\int_0^t\|R(\tau)\|\,d\tau\bigr|$, and letting $\epsilon\to0$, \[ \|D(t)v\|\le \left(\|D(0)\|+\left|\int_0^t\|R(\tau)\|\,d\tau\right|\right)\|v\|. \] Taking the supremum over \(\|v\|=1\) yields the claim. \end{proof} \begingroup \renewcommand{\thetheorem}{4c} \begin{lemma}[path independence and geometric meaning]\label{lem:4c} Let \(\omega\in C^\infty(W)\) solve \(\omega_{xy}=\sin\omega\). Then throughout \(W\) we have \(F_x=P(\omega_x)F\) and \(r_x=e_1\). Consequently, the pair \((F,r)\) coincides with the solution of the Gauss--Weingarten system of the Chebyshev net \[ r_x=e_1,\qquad r_y=\cos\omega\,e_1+\sin\omega\,e_2, \] \[ e_{1x}=-\omega_x e_2,\qquad e_{2x}=\omega_x e_1+n,\qquad n_x=-e_2, \] \[ e_{1y}=\sin\omega\,n,\qquad e_{2y}=-\cos\omega\,n,\qquad n_y=-\sin\omega\,e_1+\cos\omega\,e_2, \] and \textbf{does not depend on the integration path}. \end{lemma} \endgroup \begin{proof} \emph{Zero-curvature condition.} Since \(P=P_0+\omega_xJ\), we have \(P_y=\omega_{xy}J\). Since \(Q=Q(\omega)\) depends on \((x,y)\) only through \(\omega\), we have \(Q_x=\omega_x\,Q'(\omega)\), where \[ Q'(\omega):=\partial_\omega Q(\omega)= \begin{pmatrix} 0&0&\cos\omega\\ 0&0&\sin\omega\\ -\cos\omega&-\sin\omega&0 \end{pmatrix}. \] A direct computation of commutators gives \[ [J,Q(\omega)]=Q'(\omega),\qquad [P_0,Q(\omega)]= \begin{pmatrix} 0&\sin\omega&0\\ -\sin\omega&0&0\\ 0&0&0 \end{pmatrix}. \] Hence \[ [P,Q]=[P_0,Q]+\omega_x[J,Q]=[P_0,Q]+\omega_x Q'(\omega), \] so the terms containing \(\omega_x\) in \(-Q_x+[P,Q]\) cancel: \[ -Q_x+[P,Q]=[P_0,Q]. \] Therefore \[ P_y-Q_x+[P,Q] =\omega_{xy}J+[P_0,Q] =(\sin\omega-\omega_{xy}) \begin{pmatrix} 0&1&0\\ -1&0&0\\ 0&0&0 \end{pmatrix}=0. \] \emph{\(F_x=P F\).} Let \(\Psi:=F_x-P(\omega_x)F\). Then \(\Psi(x,0)=0\) by definition. Differentiating \(F_y=QF\) with respect to \(x\): \[ \Psi_y=Q\Psi+(Q_x+QP-P_y-PQ)F=Q\Psi \] by the zero-curvature condition. The linear homogeneous ODE with \(\Psi(x,0)=0\) gives \(\Psi\equiv0\). \emph{\(r_x=e_1\).} Let \(\Xi:=r_x-e_1\), \(\Xi(x,0)=0\). Then \[ \Xi_y=\partial_x(\cos\omega\,e_1+\sin\omega\,e_2)-e_{1y}. \] Substituting \(e_{1x}=-\omega_x e_2\), \(e_{2x}=\omega_x e_1+n\), we obtain \[ \partial_x(\cos\omega\,e_1+\sin\omega\,e_2)=\sin\omega\,n=e_{1y}. \] Thus \(\Xi_y=0\) and \(\Xi\equiv0\). Now \(F\) and \(r\) satisfy both systems \(F_x=PF\), \(F_y=QF\), \(r_x=e_1\), \(r_y=\cos\omega\,e_1+\sin\omega\,e_2\) throughout \(W\); the solution of such a compatible system with prescribed value at a point is unique, and one may integrate along any path. \end{proof} Lemma~\ref{lem:4c} applies both to \(\omega_n\) and to \(\mathrm{ref}_n\), because both solve SG (\(-f\) solves SG since \(\sin\) is odd). Therefore the scheme ``axis, then columns'' is a device of the proof, not a choice of path in the definition of the surface. \begin{remark}[Regularity]\label{rem:regularity} \(r_x\times r_y=\sin\omega\,n\), therefore \(r(\omega)\) is an immersion exactly where \(\sin\omega\ne0\). For arbitrary \(c_j\) this is not guaranteed: \(r(\omega)\) is a smooth map, possibly with singular points. The estimates of the theorem do not depend on this. \end{remark} \section{The axis \(y=0\)}\label{sec:axis} \setcounter{theorem}{4} \begin{lemma}[derivative of the difference on the axis]\label{lem:5} Let \(\delta_j:=\omega_j-(-1)^j f\). Then \textbf{throughout the window \(W\)} and for all \(j\ge1\), \[ |\partial_x\delta_j|\le 2\sum_{k\le j}\frac1{a_k}. \] In particular, \(|\partial_x(\omega_n-\mathrm{ref}_n)|\le2S\) in \(W\). \end{lemma} \begin{proof} The \(x\)-row of \eqref{eq:BT} is \[ \partial_x(\omega_j+\omega_{j-1})=-\frac2{a_j}\sin\frac{\omega_j-\omega_{j-1}}2, \] that is, \[ \partial_x\omega_j=-\partial_x\omega_{j-1}-\frac2{a_j}\sin(\cdot). \] Subtracting \((-1)^j\partial_x f\), we obtain \[ \partial_x\delta_j=-\partial_x\delta_{j-1}-\frac2{a_j}\sin(\cdot). \] Hence \(|\partial_x\delta_j|\le|\partial_x\delta_{j-1}|+2/a_j\), while \(\delta_0=\omega_0-f=0\). \end{proof} \begin{lemma}[estimate of the frame and position on the axis]\label{lem:6} Let \(F_n,F_{\mathrm{ref}}\) be the matrices \(F\) for \(\omega_n\) and \(\mathrm{ref}_n\), and \(D:=F_n-F_{\mathrm{ref}}\). Then for all \(x\in[-L,L]\), with \(S:=\sum1/a_j\), \[ \|D(x,0)\|\le2S|x|,\qquad |r(\omega_n)(x,0)-r(\mathrm{ref}_n)(x,0)|\le S x^2. \] \end{lemma} \begin{proof} On the axis, \[ D_x=P_nD+(P_n-P_{\mathrm{ref}})F_{\mathrm{ref}},\qquad D(0,0)=0, \] where \(P_n=P(\partial_x\omega_n(x,0))\) is skew-symmetric. By Lemma~\ref{lem:4b}, Lemma~\ref{lem:4a}(b), and Lemma~\ref{lem:5}, \[ \|D(x,0)\|\le\int_0^{|x|}|\partial_x(\omega_n-\mathrm{ref}_n)(\xi,0)|\,d\xi \le2S|x|. \] Furthermore, \(\partial_x\Delta r\) is the first row of \(D\), and its length is at most \(\|D\|\); hence \[ |\Delta r(x,0)|\le\int_0^{|x|}2S\xi\,d\xi=S x^2. \] \end{proof} The discrepancy of the \textbf{values} \(\omega_n(x,0)-\mathrm{ref}_n(x,0)\) themselves (for instance, nonzero modulo \(2\pi\) for arbitrary \(c_j\) along the whole axis) does not appear here: on the axis the map sees only \(\omega_x\). \section{Column: pendulum cascade and Lyapunov function}\label{sec:column} Fix \(x\in[-L,L]\); all functions are functions of \(y\in[-L,L]\), and the prime denotes \(\partial_y\). Let \(\varphi(y):=f(x,y)\) and \(h_j(y):=\omega_j(x,y)\), \(h_0=\varphi\). The \(y\)-row of \eqref{eq:BT} is \[ h_j'=h_{j-1}'-2a_j\sin\frac{h_j+h_{j-1}}2. \] Introduce \(\hat h_j:=(-1)^j h_j\) (\(\hat h_0=\varphi\)) and \[ z_j:=\frac{\hat h_j-\varphi}{2},\qquad z_0=0,\qquad \theta_j:=z_j-z_{j-1}. \] Then \(z_n=\sum_{k\le n}\theta_k\), and, since \(\mathrm{ref}_n=(-1)^n f\), \begin{equation}\label{eq:6.1}\tag{6.1} \omega_n-\mathrm{ref}_n=(-1)^n(\hat h_n-\varphi)=(-1)^n2z_n. \end{equation} \begin{lemma}[cascade]\label{lem:7} For \(j=1,\dots,n\), \[ (z_j+z_{j-1})'=-\varphi'-a_j\sin\theta_j, \] and therefore \[ \theta_j'=-a_j\sin\theta_j-F_j,\qquad F_j:=\varphi'+2z_{j-1}'. \] \end{lemma} \begin{proof} From \(h_j=(-1)^j\hat h_j\) and \(h_{j-1}=-(-1)^j\hat h_{j-1}\), \[ h_j+h_{j-1}=(-1)^j(\hat h_j-\hat h_{j-1}),\qquad h_j-h_{j-1}=(-1)^j(\hat h_j+\hat h_{j-1}). \] Substituting into the \(y\)-row: \[ (-1)^j(\hat h_j+\hat h_{j-1})' =-2a_j\sin\left(\frac{(-1)^j(\hat h_j-\hat h_{j-1})}{2}\right) =-2a_j(-1)^j\sin\left(\frac{\hat h_j-\hat h_{j-1}}2\right), \] whence \[ (\hat h_j+\hat h_{j-1})'=-2a_j\sin\left(\frac{\hat h_j-\hat h_{j-1}}2\right). \] Since \(\hat h_j=\varphi+2z_j\), we obtain \[ 2\varphi'+2(z_j+z_{j-1})'=-2a_j\sin(z_j-z_{j-1}), \] which is the first equality. The second follows from \(\theta_j'=(z_j+z_{j-1})'-2z_{j-1}'\). \end{proof} \begin{lemma}[integral Lyapunov estimate]\label{lem:8} Let \(J\) be a compact interval, \(\theta\in C^1(J)\), \(F\in C^0(J)\), \(a>0\), and \(\theta'=-a\sin\theta-F\) on \(J\). Then \[ a\cdot\int_J|\sin\theta|\,dy\le\pi+\int_J|F|\,dy. \] No assumptions are made on the size of \(F\) relative to \(a\), on the initial angle, or on the orientation of \(J\). \end{lemma} \begin{proof} Let \(G(t):=d(t)=\operatorname{dist}(t,2\pi\mathbb Z)\): it is a \(1\)-Lipschitz function with values in \([0,\pi]\); outside the points \(\pi\mathbb Z\) it is differentiable. The function \(\Psi:=G\circ\theta\) is Lipschitz, hence absolutely continuous and differentiable almost everywhere. Consider the set \(E:=\{y\in J:\theta(y)\in\pi\mathbb Z\}\). \begin{itemize} \item If \(y\notin E\), then \[ \Psi'(y)=\operatorname{sgn}(\sin\theta)\cdot\theta' =-a|\sin\theta|-\operatorname{sgn}(\sin\theta)\cdot F. \] \item If \(y\in E\) is a limit point of \(E\), then by continuity and discreteness of \(\pi\mathbb Z\) there exist \(y_k\to y\) with \(\theta(y_k)=\theta(y)\); hence \(\theta'(y)=0\), and from the equation \(F(y)=-\theta'(y)-a\sin\theta(y)=0\). Then \(\Psi\) is differentiable at \(y\) with \(\Psi'(y)=0\), because \[ |\Psi(y+h)-\Psi(y)|\le|\theta(y+h)-\theta(y)|=o(h). \] The expression \(-a|\sin\theta|-\operatorname{sgn}(\sin\theta)F\) there is also zero (with \(\operatorname{sgn}0:=0\)). \item Isolated points of \(E\) form a countable set of measure zero. \end{itemize} Thus almost everywhere \[ \Psi'=-a|\sin\theta|-\sigma F,\qquad \sigma:=\operatorname{sgn}(\sin\theta)\in\{-1,0,1\}. \] Integrating over \(J=[\alpha,\beta]\) (absolute continuity): \[ \Psi(\beta)-\Psi(\alpha)=-a\int|\sin\theta|-\int\sigma F, \] whence \[ a\int|\sin\theta|=\Psi(\alpha)-\Psi(\beta)-\int\sigma F \le\pi+\int|F|, \] since \(0\le\Psi\le\pi\). \end{proof} \emph{Meaning.} \(d(\theta)\) is a Lyapunov function decreasing at rate \(a|\sin\theta|\) in the absence of forcing; being bounded by the number \(\pi\), it bounds the total ``time away from equilibria'' in both directions in \(y\) at once. \section{Cascade estimates}\label{sec:cascade} For fixed \(x\) (integrals over \(y\in[-L,L]\)): \[ T:=\int|\varphi'|\,dy\le T_*,\qquad V_k:=\int|z_k'|\,dy,\qquad I_k:=\int|\sin\theta_k|\,dy,\qquad V_0=0. \] \begin{lemma}[cascade inequalities]\label{lem:9} For \(k=1,\dots,n\): (I) \(a_kI_k\le\pi+T+2V_{k-1}\); (II) \(V_k\le(3^k-1)(T+\pi/2)\); (III) \(a_kI_k\le3^{k-1}(2T+\pi)\). \end{lemma} \begin{proof} (I) Lemma~\ref{lem:8} for \(\theta_k\), \(a=a_k\), and \(F=F_k\) from Lemma~\ref{lem:7}; here \[ \int|F_k|\le\int|\varphi'|+2\int|z_{k-1}'|=T+2V_{k-1}. \] (II) From Lemma~\ref{lem:7}: \[ z_k'=-z_{k-1}'-\varphi'-a_k\sin\theta_k, \] therefore \[ V_k\le V_{k-1}+T+a_kI_k. \] By Part(I), \[ V_k\le3V_{k-1}+2T+\pi. \] Induction: \(V_0=0\); if \(V_{k-1}\le(3^{k-1}-1)(T+\pi/2)\), then \[ V_k\le3(3^{k-1}-1)(T+\pi/2)+2(T+\pi/2) =(3^k-1)(T+\pi/2). \] (III) By Parts (I) and (II): \[ a_kI_k\le\pi+T+(3^{k-1}-1)(2T+\pi) =3^{k-1}(2T+\pi)-T \le3^{k-1}(2T+\pi). \] \end{proof} The estimate depends neither on \(a_j\), nor on the initial values \(z_k(0)\), nor on \(x\). \begin{lemma}[integral estimate on a column]\label{lem:10} For each \(x\in[-L,L]\), \[ \Lambda(x):=\int_{-L}^{L}d(\omega_n-\mathrm{ref}_n)(x,y)\,dy \le\pi(2T+\pi)\sum_{k\le n}\frac{3^{k-1}}{a_k} \le\pi(2T_*+\pi)3^{n-1}S. \] \end{lemma} \begin{proof} By \eqref{eq:6.1}, the evenness of \(d\), and \eqref{eq:E4}, \[ d(\omega_n-\mathrm{ref}_n)=d(2z_n)=2d_\pi(z_n). \] Since \(z_n=\sum\theta_k\), by \eqref{eq:E1} and \eqref{eq:E2}, \[ d_\pi(z_n)\le\sum_k d_\pi(\theta_k) \le\frac{\pi}{2}\sum_k|\sin\theta_k|. \] Therefore \[ \Lambda\le\pi\sum_k I_k \le\pi(2T+\pi)\sum_k\frac{3^{k-1}}{a_k} \] by Lemma~\ref{lem:9}(III); it remains to use \(T\le T_*\) and \(3^{k-1}\le3^{n-1}\). \end{proof} \section{Assembly: error of the map}\label{sec:assembly} \begin{lemma}\label{lem:11} Let \(\Lambda_y(x):=\int d(\omega_n-\mathrm{ref}_n)(x,t)\,dt\) over \(t\) between \(0\) and \(y\) (so that \(\Lambda_y(x)\le\Lambda(x)\)). Then for all \((x,y)\in W\), \[ |r(\omega_n)(x,y)-r(\mathrm{ref}_n)(x,y)| \le S x^2+2S|x||y|+(1+|y|)\Lambda_y(x). \] \end{lemma} \begin{proof} Consider \(y\ge0\) (for \(y<0\) the argument is mirrored). In the column \(x\), denote \(D(y):=F_n(x,y)-F_{\mathrm{ref}}(x,y)\) and \(d(y):=d(\omega_n-\mathrm{ref}_n)(x,y)\). \emph{Frame.} We have \[ D'=Q(\omega_n)D+(Q(\omega_n)-Q(\mathrm{ref}_n))F_{\mathrm{ref}}, \] the matrix \(Q(\omega_n)\) is skew-symmetric, and by Lemma~\ref{lem:4a}(b) and the orthogonality of \(F_{\mathrm{ref}}\), \[ \|(Q(\omega_n)-Q(\mathrm{ref}_n))F_{\mathrm{ref}}\|\le d(y). \] Lemma~\ref{lem:4b} and Lemma~\ref{lem:6} give \[ \|D(y)\|\le\|D(0)\|+\int_0^y d(t)\,dt \le2S|x|+\Lambda_y(x). \] \emph{Position.} Write \(\Delta r_y=r(\omega_n)_y-r(\mathrm{ref}_n)_y\) and split \begin{align*} \Delta r_y &=\bigl(\cos\omega_n\,e_{1,n}+\sin\omega_n\,e_{2,n}\bigr) -\bigl(\cos\omega_{\mathrm{ref}}\,e_{1,\mathrm{ref}}+\sin\omega_{\mathrm{ref}}\,e_{2,\mathrm{ref}}\bigr)\\ &=(\cos\omega_n-\cos\omega_{\mathrm{ref}})\,e_{1,n} +(\sin\omega_n-\sin\omega_{\mathrm{ref}})\,e_{2,n}\\ &\quad+\cos\omega_{\mathrm{ref}}\,(e_{1,n}-e_{1,\mathrm{ref}}) +\sin\omega_{\mathrm{ref}}\,(e_{2,n}-e_{2,\mathrm{ref}}), \end{align*} where \(e_{i,n}:=e_i(\omega_n)\), \(e_{i,\mathrm{ref}}:=e_i(\mathrm{ref}_n)\). The first two terms, expanded in the orthonormal pair \((e_{1,n},e_{2,n})\), have norm \[ \bigl((\cos\omega_n-\cos\omega_{\mathrm{ref}})^2+(\sin\omega_n-\sin\omega_{\mathrm{ref}})^2\bigr)^{1/2} =|e^{i\omega_n}-e^{i\omega_{\mathrm{ref}}}|\le d(y) \] by \eqref{eq:E3}. The last two terms equal the row \((\cos\omega_{\mathrm{ref}},\sin\omega_{\mathrm{ref}},0)\) acting on the matrix \(D(y)=F_n(x,y)-F_{\mathrm{ref}}(x,y)\), whose norm is at most \(|(\cos\omega_{\mathrm{ref}},\sin\omega_{\mathrm{ref}},0)|\cdot\|D(y)\|=\|D(y)\|\). Hence \(|\Delta r_y|\le d(y)+\|D(y)\|\). \emph{Conclusion.} Integrating over \(t\in[0,y]\) and using Lemma~\ref{lem:6} for \(|\Delta r(x,0)|\): \[ |\Delta r(x,y)| \le|\Delta r(x,0)|+\int_0^y(d+\|D\|)\,dt \le Sx^2+\Lambda_y+2S|x|y+\int_0^y\Lambda_t\,dt \le Sx^2+2S|x||y|+(1+|y|)\Lambda_y(x), \] since \(\Lambda_t\) is nondecreasing in \(t\), and therefore \(\int_0^y\Lambda_t\,dt\le y\Lambda_y\). \end{proof} \section{Proofs of Theorem~\ref{thm:main} and Corollary~\ref{cor:amsler}}\label{sec:proofs} \textbf{Theorem~\ref{thm:main}.} For \(|x|,|y|\le L\), we have \[ Sx^2+2S|x||y|\le3L^2S, \] and, by Lemma~\ref{lem:10}, \[ (1+|y|)\Lambda_y(x)\le(1+L)\Lambda(x) \le(1+L)\pi(2T_*+\pi)3^{n-1}S. \] Substituting into Lemma~\ref{lem:11} and taking the supremum over \(W\): \[ \|r(\omega_n)-r(\mathrm{ref}_n)\|_{L^\infty(W)} \le[3L^2+(1+L)\pi(2T_*+\pi)3^{n-1}]S=K_nS. \] \(\square\) There is no case distinction according to the size of \(S\): Lemmas~\ref{lem:6}, \ref{lem:10}, and \ref{lem:11} hold for all data, and no Gronwall estimates are used in them. \textbf{Corollary~\ref{cor:amsler} (Amsler).} Let \(Y\in C^\infty(\mathbb R)\), \((uY')'=\sin Y\), \(Y(0)=\pi s\). \begin{itemize} \item \emph{\(|Y'|\le1\).} Since \(Y\in C^\infty(\mathbb R)\), \(Y'\) is bounded in a neighborhood of \(0\), so \(uY'(u)\to0\) as \(u\to0\). Integrating \((uY')'=\sin Y\) from \(0\) to \(u\), we obtain \[ uY'(u)=\int_0^u\sin Y, \] whence \(|uY'(u)|\le|u|\), and \(|Y'|\le1\). \item \emph{\(f(x,y):=Y(xy)\) solves SG.} \(\partial_x f=yY'(xy)\), and \(\partial_y\) of this expression equals \[ Y'(u)+uY''(u)=(uY')'=\sin Y(u),\qquad u=xy. \] \item \emph{\(T_*\le2L^2\).} \(|\partial_y f|=|x|\cdot|Y'(xy)|\le L\), and the length of the integration interval is \(2L\). \end{itemize} Applying Theorem~\ref{thm:main} with this \(f\): \[ \|\Delta r\|\le[3L^2+(1+L)\pi(4L^2+\pi)3^{n-1}]S. \] \(\square\) \section{Remarks}\label{sec:remarks} \begin{remark}[Sharpness in the sup-norm and integral form]\label{rem:sharp} \(\sup d(\omega_n-\mathrm{ref}_n)\) can reach \(\pi\) in thin layers if the initial data of the columns are far from the reference; the theorem needs only the integral \(\Lambda(x)=\int d\,dy\), and it is small regardless of the number and location of such layers. The passage from the pointwise norm to the integral is the essence of the method. \end{remark} \begin{remark}[Axial constants]\label{rem:axial} The difference of the values \(\omega_n(x,0)-\mathrm{ref}_n(x,0)\), generated by the constants \(c_j\), is not seen by the map on the axis (Lemma~\ref{lem:6}: only \(\omega_x\) occurs there) and enters the columns as the initial data of Lemma~\ref{lem:8}, whose contribution to the right-hand side does not exceed \(\pi\), that is, \(O(1/a)\) in the final normalization. Therefore the constants \(c_j\) are absent from the estimate. \end{remark} \begin{remark}[Choice of the integration path as optimal control]\label{rem:path} The difference \(\Delta r\) does not depend on the path (Lemma~\ref{lem:4c})---the path is a free parameter \emph{of the proof}: it determines only the tightness of the upper bound. For a monotone path \(\gamma\), the bound functional decomposes as \[ F[\gamma]=S x^2\ \text{(independent of the path)} +\int2S\,x(t)\dot y\,dt +\int D_d(t)(|\dot x|+|\dot y|)\,dt +\int d\,\dot y\,dt, \] where \(D_d(t)=\int d\,ds\) is the accumulated \(d\)-contribution. For seeds for which \(T(x)=\int|f_y(x,\cdot)|\) is minimal on the vertical (for example, for the Amsler seed \(T(0)=0\), since \(f_y(0,y)=xY'(xy)|_{x=0}=0\)), the optimal path is \((0,0)\to(0,y)\to(x,y)\); it gives \[ |\Delta r(x,y)|\le S x^2+(1+|x|+|y|)\Lambda(0) \] and the estimate \(K'_n=L^2+(1+2L)\pi^2 3^{n-1}\) (for \(\Lambda(0)\le\pi^2 3^{n-1}S\)). \emph{Proof.} By Lemma~\ref{lem:4c}, $F_x=P(\omega_x)F$ holds throughout $W$ for both $\omega_n$ and $\mathrm{ref}_n$. On $x=0$ we have $D(0,0)=0$, so Lemma~\ref{lem:4b} gives $\|D(0,t)\|\le\Lambda(0)$ and, as in the proof of Lemma~\ref{lem:11}, $|\Delta r(0,y)|\le\int_0^{|y|}(d+\|D\|)\,dt\le(1+|y|)\Lambda(0)$. Along the horizontal segment at height $y$, $D_x=P_nD+(P_n-P_{\mathrm{ref}})F_{\mathrm{ref}}$ and Lemma~\ref{lem:5} (valid throughout $W$) give $\|D(x,y)\|\le\|D(0,y)\|+2S|x|\le\Lambda(0)+2S|x|$; since $\partial_x\Delta r$ is the first row of $D$, $|\Delta r(x,y)|\le|\Delta r(0,y)|+|x|\Lambda(0)+Sx^2$. Finally $\Lambda(0)\le\pi^2 3^{n-1}S$ by Lemma~\ref{lem:10} with $T(0)=0$. The base \(3^{n-1}\) does not depend on the choice of path---this is a structural limitation of the cascade technique (Lemma~\ref{lem:9}(II)). For a general seed, minimization of \(F[\gamma]\) is a substantial open problem (a possible formulation is in terms of Hamilton--Jacobi--Bellman for a functional with memory). \end{remark} \begin{remark}[Limitations of the result] \label{rem:limits} (a) The constant \(3^{n-1}\) has not been optimized; uniformity in \(n\) is not claimed. However, the order \(1/a\) is sharp for every fixed \(n\): Corollary~\ref{cor:sharp-all-n} exhibits a one-parameter family ($f\equiv0$, $c_j=\pi$, $a_j=s^{-j}$) along which the immersion error is bounded below by $c_n\sum_j 1/a_j$ with $c_n=c_0/4^{\,n-1}>0$. The constant $c_n$ is exponentially small in $n$, so this does not give uniformity in \(n\).\\ (b) The immersions are normalized by a common frame at the origin; the estimate compares the maps under \emph{the same} frame at the origin.\\ (c) \(r(\omega)\) is an immersion only when \(\sin\omega\ne0\) (see Remark~\ref{rem:regularity}); the estimate remains valid through singular points.\\ (d) The window \(W\) is compact; the dependence on \(L\) is polynomial.\\ (e) The statement is real: for complex parameters \(a_j\) the geometry has a different character (poles of complex solutions appear), and the theorem is not formulated outside the real case. \end{remark} \section{Dependence on the seed and on the B\"acklund parameters: smoothness, limits, and the infinitesimal regime} \label{sec:smooth} Throughout this section $n\ge1$ is fixed, $W=[-L,L]^2$, and \[ \mathcal S:=\{f\in C^\infty(W):\ f_{xy}=\sin f\ \text{in}\ W\}. \] Theorem~\ref{thm:main} describes the limit of $r(\omega_n)$ as \emph{all} parameters $a_j$ tend to infinity, uniformly in the normalization constants $c_j$. Here we study the dependence of the immersions on the data systematically: smooth dependence on the seed and on $(a_j,c_j)$ at finite parameters (\S\ref{sec:smooth-finite}); the total limit $a_j\to\infty$ and its first-order (infinitesimal) structure (\S\ref{sec:smooth-infty}); one-step asymptotics outside the boundary layer (\S\ref{sec:smooth-layer}); the orientation-reversal symmetry and the optimal topology of the total limit (\S\ref{sec:S-limit}); the remaining limits---in the constants $c_j$ (\S\ref{sec:smooth-other}) and in $a_j\to0$ (\S\ref{sec:S-zero}); partial limits along arbitrary index sets and the closed orthant (\S\ref{sec:S-partial}), including the partial limits $a_j\to0$ (Corollary~\ref{cor:S-zero-partial}) and the stratified Lipschitz structure at the faces (Corollary~\ref{cor:S-lipschitz}); and unbounded seed data (\S\ref{sec:smooth-other}). As everywhere in the paper, all statements are analytic; the numerical illustrations of \S\ref{sec:numerics} are not used. \medskip\noindent\textbf{Convention.} Maps defined on the Fr\'echet spaces $C^\infty(W)$, $\mathcal D$, $\mathcal S$ are called $C^\infty$ if their compositions with all $C^\infty$ finite-dimensional families are $C^\infty$ (jointly in $(x,y,p)$); continuity is understood in the Fr\'echet topology. Banach-space statements (implicit function theorem, Gronwall) are applied to the $C^k$-norms, $k<\infty$. \subsection{The Goursat parametrization of the seed} \label{sec:smooth-goursat} Every $f\in\mathcal S$ is determined by its \emph{characteristic data} \[ \alpha(x):=f(x,0),\qquad \beta(y):=f(0,y),\qquad \alpha(0)=\beta(0), \] and conversely, arbitrary $\alpha,\beta\in C^\infty([-L,L])$ with $\alpha(0)=\beta(0)$ determine a unique element of $\mathcal S$ (this is the Goursat problem for the sine-Gordon equation; Lemma~\ref{lem:sg-goursat-smooth} below). We therefore introduce the data space \[ \mathcal D:=\{(\alpha,\beta)\in C^\infty([-L,L])^2:\ \alpha(0)=\beta(0)\} \] (a closed affine subspace of the Fr\'echet space $C^\infty([-L,L])^2$) and the \emph{solution map} \[ \Psi\colon\mathcal D\to\mathcal S,\qquad \Psi(\alpha,\beta):=f . \] By ``a $C^\infty$ \emph{family of seeds}'' $p\mapsto g_p$ ($p$ in an open set $P\subset\mathbb R^m$) we mean a map such that $g_p\in\mathcal S$ for all $p$ and all parameter derivatives exist and are continuous as maps into $C^k(W)$ for every $k$; this is equivalent to joint smoothness of $(x,y,p)\mapsto g_p(x,y)$. \begin{lemma}[smooth dependence for first-order systems] \label{lem:goursat-smooth} Let $A,B\colon W\times\mathbb R\times P\to\mathbb R$ be $C^\infty$ jointly ($P\subset\mathbb R^m$ open), let $u_0\colon P\to\mathbb R$ be $C^\infty$, and assume that for each $p$ in a neighbourhood of $p_0$ the system \[ u_x=A(x,y,u,p),\qquad u_y=B(x,y,u,p),\qquad u(0,0)=u_0(p) \] has a solution $u(\cdot,\cdot\,;p)\in C^\infty(W)$. (This holds, for instance, if $A,B$ satisfy the compatibility identity $A_y+A_uB=B_x+B_uA$ and $|A|$, $|A_u|$, $|B_u|$ are bounded on $W\times\mathbb R\times\{|p-p_0|\le1\}$: then the axis ODE in $x$ and the column ODEs in $y$ have global solutions, and the argument of Lemma~\ref{lem:bt} ($\Phi:=u_x-A$ solves $\Phi_y=B_u\Phi$, $\Phi(x,0)=0$) gives $u_x=A$ on all of $W$. For the systems of Lemma~\ref{lem:bt-smooth} these conditions hold.) Then $(x,y,p)\mapsto u(x,y;p)$ is $C^\infty$ jointly; in particular $p\mapsto u(\cdot,\cdot\,;p)$ is a $C^\infty$ family in the sense above. For any direction $\dot p$, the first variation $U:=\partial_p u\cdot\dot p$ solves the linear system \[ U_x=A_u\,U+A_p\dot p,\qquad U_y=B_u\,U+B_p\dot p,\qquad U(0,0)=u_0'(p)\dot p, \] and higher variations are obtained by differentiating this linear system. \end{lemma} \begin{proof} Along the axis $y=0$ the function $u(\cdot,0;p)$ solves the ODE $u_x=A(x,0,u,p)$ with initial value $u_0(p)$; the right-hand side is $C^\infty$ in $(x,u,p)$, so the classical theorem on smooth dependence of ODE solutions on parameters and data gives $u(x,0;p)$ jointly $C^\infty$ in $(x,p)$. For fixed $x$, the column function $u(x,\cdot\,;p)$ solves $u_y=B(x,y,u,p)$ with initial value $u(x,0;p)$, now with the pair $(x,p)$ as parameter; the same classical theorem gives $u(x,y;p)$ jointly $C^\infty$ in $(x,y,p)$. Differentiating the two ODEs in $p$ yields the stated linear system for $U$ (on the axis first, then in the columns), and iteration yields all higher variations. \end{proof} \begin{lemma}[smooth dependence of the seed on the characteristic data] \label{lem:sg-goursat-smooth} The map $\Psi\colon\mathcal D\to\mathcal S$ is a well-defined bijection and is $C^\infty$ in the $C^\infty$-topologies. The first variation $h:=\delta f$ in the direction $(\delta\alpha,\delta\beta)$ solves the linear Goursat problem \begin{equation}\label{eq:smooth-var} h_{xy}=\cos f\cdot h,\qquad h(x,0)=\delta\alpha(x),\qquad h(0,y)=\delta\beta(y), \end{equation} and higher variations are obtained by differentiating \eqref{eq:smooth-var}. \end{lemma} \begin{proof} \emph{Step 1: a Volterra estimate (no smallness of $L$ is needed).} For $\phi,h\in C^0(W)$ with $|\phi|\le1$ put $(\mathcal V_\phi h)(x,y):=\int_0^x\!\!\int_0^y\phi\,h\,dt\,ds$ (oriented integrals). By induction on $m$, $|(\mathcal V_\phi^mh)(x,y)|\le\frac{|xy|^m}{(m!)^2}\|h\|_{C^0(W)}$; the induction step is $\int_0^{|x|}\!\int_0^{|y|}\frac{(st)^{m-1}}{((m-1)!)^2}\,dt\,ds=\frac{|xy|^m}{(m!)^2}$. Hence $\|\mathcal V_\phi^m\|\le L^{2m}/(m!)^2$, the Neumann series converges, and $\mathrm{Id}-\mathcal V_\phi$ is invertible on $C^0(W)$ with $\|(\mathrm{Id}-\mathcal V_\phi)^{-1}\|\le\sum_mL^{2m}/(m!)^2\le e^{2L}$, for every $L>0$. \emph{Step 2: existence, uniqueness and regularity on all of $W$.} Put $h_0(x,y):=\alpha(x)+\beta(y)-\alpha(0)$ and $T(f):=h_0+\int_0^x\!\!\int_0^y\sin f\,dt\,ds$ on $C^0(W)$. Since $|\sin a-\sin b|\le|a-b|$, the same induction gives $|(T^mf_1-T^mf_2)(x,y)|\le\frac{|xy|^m}{(m!)^2}\|f_1-f_2\|_{C^0(W)}$, so $T^m$ is a contraction for $m$ large and $T$ has exactly one fixed point $f\in C^0(W)$; it satisfies $f(x,0)=\alpha(x)$, $f(0,y)=\beta(y)$. \emph{Regularity.} The operator $\iint\colon C^k(W)\to C^{k+1}(W)$ is bounded: for $i,j\ge1$, $\partial_x^i\partial_y^j\iint g=\partial_x^{i-1}\partial_y^{j-1}g$, and $\partial_x^i\iint g=\int_0^y\partial_x^{i-1}g\,dt$ (similarly in $y$) exist and are continuous for $i\le k+1$. (In general $\iint g$ is \emph{not} $C^{k+2}$.) From $f=h_0+\iint\sin f$: $f\in C^k\Rightarrow f\in C^{k+1}$, hence $f\in C^\infty(W)$ and $f_{xy}=\sin f$, i.e.\ $f\in\mathcal S$. Conversely, if $f\in\mathcal S$ has characteristic data $(\alpha,\beta)$, integrating $f_{xy}=\sin f$ over the rectangle with corners $(0,0)$, $(x,y)$ gives $f=h_0+\iint\sin f$, so $f$ is that fixed point. Thus $\Psi\colon\mathcal D\to\mathcal S$ is a well-defined bijection. \emph{Step 3: smooth dependence in $C^0$.} Let $\Phi\colon C^0(W)\times C^0(W)\to C^0(W)$, $\Phi(f,h_0):=f-h_0-\iint\sin f$. It is $C^\infty$ (the Nemytskii operator $\sin$ is $C^\infty$ on the Banach space $C^0$, and $\iint$ is bounded linear), and $D_f\Phi(f,h_0)=\mathrm{Id}-\mathcal V_{\cos f}$ is invertible by Step~1. By the implicit function theorem in Banach spaces and the global uniqueness of Step~2, $h_0\mapsto f$ is a $C^\infty$ map $C^0(W)\to C^0(W)$. The map $(\alpha,\beta)\mapsto h_0$ is linear and continuous from $\mathcal D$ into $C^k(W)$ for every $k$. (Smoothness on the Fr\'echet space $\mathcal D$ is understood, as throughout this section, along $C^\infty$ finite-dimensional families $p\mapsto(\alpha_p,\beta_p)$; all maps below are compositions of continuous linear maps with $C^\infty$ maps between Banach spaces, so this is automatic.) \emph{Step 4: bootstrap in $C^k$.} We claim: for every $k$, the map $(\alpha,\beta)\mapsto f(\alpha,\beta)$ is $C^\infty$ into $C^k(W)$. For $k=0$ this is Step~3. If it holds for $k$, then $(\alpha,\beta)\mapsto\sin f$ is $C^\infty$ into $C^k(W)$ ($\sin\colon C^k\to C^k$ is $C^\infty$, $C^k(W)$ being a Banach algebra), $\iint\colon C^k\to C^{k+1}$ is bounded linear and $h_0$ is linear and continuous into $C^{k+1}(W)$; the identity $f=h_0+\iint\sin f$ gives the claim for $k+1$. \emph{Step 5: variational equation.} Differentiating $f=h_0+\iint\sin f$ in the direction $(\delta\alpha,\delta\beta)$ gives \[ h=\delta\alpha(x)+\delta\beta(y)-\delta\alpha(0) +\int_0^x\!\!\int_0^y\cos f\cdot h, \] whence $h_{xy}=\cos f\cdot h$, $h(x,0)=\delta\alpha(x)$, $h(0,y)=\delta\beta(y)$. The linear problem has a unique solution by Step~1 ($\mathrm{Id}-\mathcal V_{\cos f}$ is invertible), and higher variations follow by iteration. \end{proof} \subsection{Smooth dependence at finite parameters} \label{sec:smooth-finite} \begin{lemma}[smooth dependence of one B\"acklund step] \label{lem:bt-smooth} Let $P\subset\mathbb R^m$ be open, let $a\colon P\to(0,\infty)$ and $c\colon P\to\mathbb R$ be $C^\infty$, and let $p\mapsto g_p$ be a $C^\infty$ family of seeds in $\mathcal S$. Let $q_p\in C^\infty(W)$ solve \eqref{eq:BT} with parameter $a(p)$ and seed $g_p$, normalized by $q_p(0,0)=c(p)$ (existence and uniqueness: Lemma~\ref{lem:bt}; note that the seed must solve the sine-Gordon equation, otherwise the construction of Lemma~\ref{lem:bt} is not available). Then $p\mapsto q_p$ is a $C^\infty$ family of seeds. Writing $q$, $g$, $a$, $c$ for the values at a fixed $p$ and considering a direction $\dot p$ with $\dot a$, $\dot c$, $h:=\delta g$ (so that $h_{xy}=\cos g\cdot h$ by Lemma~\ref{lem:sg-goursat-smooth}), the first variation $V:=\delta q$ solves: \begin{itemize} \item for the variation in $a$ (with $\delta g=0$, $\delta c=0$): \[ V_x=\frac{2}{a^2}\sin\frac{q-g}{2} -\frac1a\cos\frac{q-g}{2}\,V, \qquad V_y=-2\sin\frac{q+g}{2}-a\cos\frac{q+g}{2}\,V, \qquad V(0,0)=0; \] \item for the variation in $c$ (with $\delta g=0$, $\delta a=0$): \[ V_x=-\frac1a\cos\frac{q-g}{2}\,V, \qquad V_y=-a\cos\frac{q+g}{2}\,V, \qquad V(0,0)=\delta c; \] \item for the variation in the seed: \[ V_x=-h_x-\frac1a\cos\frac{q-g}{2}\,(V-h), \qquad V_y=h_y-a\cos\frac{q+g}{2}\,(V+h), \qquad V(0,0)=0. \] \end{itemize} Higher variations are obtained by differentiating these linear systems. \end{lemma} \begin{proof} By Lemma~\ref{lem:bt}, $q_p=2u_p-g_p$, where $u_p$ solves \[ u_x=-\frac1{a}\sin(u-g),\qquad u_y=g_y-a\sin u,\qquad u(0,0)=\frac{c+g(0,0)}2 . \] The right-hand sides are $C^\infty$ in $(x,y,u,p)$ jointly: in $u$ and $p$ through $a,c$, and in $(x,y,p)$ through $g_p$ and $g_{p,y}$, which are jointly $C^\infty$ by the definition of a smooth family (and, for the seed itself, by Lemma~\ref{lem:sg-goursat-smooth}). Lemma~\ref{lem:goursat-smooth} then gives that $u_p$, hence $q_p=2u_p-g_p$, is a $C^\infty$ family. For the variation systems, let $U:=\delta u$; then $V=2U-h$ and, from $u_x=-(1/a)\sin(u-g)$, $u_y=g_y-a\sin u$: \emph{variation in $a$} ($h=0$, $U=V/2$): \[ U_x=\frac1{a^2}\sin(u-g)-\frac1a\cos(u-g)\,U,\qquad U_y=-\sin u-a\cos u\,U,\qquad U(0,0)=0, \] and multiplying by $2$ gives the first displayed system (use $u-g=(q-g)/2$, $u=(q+g)/2$). \emph{variation in $c$}: $U(0,0)=\delta c/2$ and $U_x=-(1/a)\cos(u-g)U$, $U_y=-a\cos u\,U$; since $V=2U$, the second displayed system follows. \emph{variation in the seed}: from $\delta(u_x)=-(1/a)\cos(u-g)(U-h)$ and $\delta(u_y)=h_y-a\cos u\,U$ we get \[ U_x=-\frac1a\cos\frac{q-g}{2}\,(U-h),\qquad U_y=h_y-a\cos\frac{q+g}{2}\,U, \] and substituting $U=(V+h)/2$ gives $V_x=-h_x-\frac1a\cos\frac{q-g}{2}(V-h)$ and $V_y=h_y-a\cos\frac{q+g}{2}(V+h)$. \end{proof} \begin{lemma}[smooth dependence of the immersion] \label{lem:immersion-smooth} For $\omega\in C^\infty(W)$ let $(F(\omega),r(\omega))$ be constructed as in \S\ref{sec:immersion} (axis, then columns). Then $\omega\mapsto(F(\omega),r(\omega))$ is $C^\infty$ from $C^\infty(W)$ to $C^\infty(W,\mathrm{SO}(3))\times C^\infty(W,\mathbb R^3)$ in the sense that it is $C^\infty$ into every $C^k$. For a $C^\infty$ family $t\mapsto\omega(t)$ with $\eta=\partial_t\omega$, the variations $D=\partial_tF$, $V=\partial_t r$ satisfy, on the axis, \[ D_x=P(\omega_x)D+\eta_x\,J\,F,\qquad V_x=\text{first row of }D, \] and, in the columns, \[ D_y=Q(\omega)D+Q'(\omega)\eta F, \qquad V_y=-\eta\sin\omega\,e_1+\eta\cos\omega\,e_2 +\cos\omega\,D_{1\cdot}+\sin\omega\,D_{2\cdot}, \] where $D_{i\cdot}$ is the $i$-th row of $D$. If $\omega\in\mathcal S$, then by Lemma~\ref{lem:4c} also $F_x=P(\omega_x)F$ holds throughout $W$, and the full Gauss--Weingarten system is available. \end{lemma} \begin{proof} On the axis, $F(\cdot,0)$ solves $F_x=P(\omega_x(x,0))F$ with $F(0,0)=I$; the coefficient depends continuously (in every $C^k([-L,L])$) on $\omega_x(\cdot,0)$, which depends continuously on $\omega\in C^{k+1}(W)$. Smooth dependence of ODE solutions on the coefficient function (a $C^\infty$ map into the space of continuous matrix functions) gives that $\omega\mapsto F(\cdot,0)$ is $C^\infty$ into every $C^k$. In the columns, $F(x,\cdot)$ solves $F_y=Q(\omega(x,y))F$ with initial value $F(x,0)$; treating $x$ as a parameter and using the smooth dependence already established gives $F$ jointly, and $r$ follows by integrating the $C^\infty$-in-$(\omega,F)$ integrands $e_1$ and $\cos\omega\,e_1+\sin\omega\,e_2$. Differentiating in $t$ yields the stated systems ($P$ is affine in $p$, $Q'$ is as in the proof of Lemma~\ref{lem:4c}, and $\partial_t r_y$ uses $\partial_t\cos\omega=-\eta\sin\omega$, $\partial_t\sin\omega=\eta\cos\omega$ together with $\partial_t e_i=D_{i\cdot}$). Iteration gives all higher variations. \end{proof} The next lemma quantifies the continuity of the immersion in $\omega$ by the same cumulative (Gronwall-free) technique as in Lemmas~\ref{lem:6}--\ref{lem:11}. \begin{lemma}[quantitative continuity of the immersion] \label{lem:immersion-continuity} Let $\omega\in\mathcal S$ and $\omega+\eta\in\mathcal S$, $\eta\in C^\infty(W)$. Put \[ A(x):=\int_0^{|x|}|\eta_\xi(\xi,0)|\,d\xi, \qquad B(x,y):=\int_{\min(0,y)}^{\max(0,y)}|\eta(x,t)|\,dt . \] Then for all $(x,y)\in W$, \[ |r(\omega+\eta)(x,y)-r(\omega)(x,y)| \;\le\; \int_0^{|x|}\!\!A(\xi)\,d\xi \;+\;B(x,y)\;+\;|y|\,\bigl(A(x)+B(x,y)\bigr). \] \end{lemma} \begin{proof} Let $F,\tilde F$ be the frames of $\omega$ and $\omega+\eta$ and $D:=\tilde F-F$, $\Delta r:=r(\omega+\eta)-r(\omega)$. \emph{Axis.} On $y=0$, $D_x=P(\omega_x)D+\bigl(P(\tilde\omega_x)-P(\omega_x)\bigr)\tilde F$; the matrix $P(\omega_x)$ is skew-symmetric and, by Lemma~\ref{lem:4a}(b), $\|P(\tilde\omega_x)-P(\omega_x)\|=|\eta_x|$. Lemma~\ref{lem:4b} gives $\|D(x,0)\|\le A(x)$, and since $\partial_x\Delta r$ is the first row of $D$, $|\Delta r(x,0)|\le\int_0^{|x|}\|D(\xi,0)\|\,d\xi\le\int_0^{|x|}A(\xi)d\xi$. \emph{Columns.} For fixed $x$, \[ D_y=Q(\tilde\omega)D+\bigl(Q(\tilde\omega)-Q(\omega)\bigr)F , \] $Q(\tilde\omega)$ is skew-symmetric, and by Lemma~\ref{lem:4a}(b) and \eqref{eq:E3}, $\|Q(\tilde\omega)-Q(\omega)\|\le d(\eta)\le|\eta|$. Lemma~\ref{lem:4b} gives $\|D(x,y)\|\le A(x)+B(x,y)$. For the position, decompose as in the proof of Lemma~\ref{lem:11}: \[ \Delta r_y =(\cos\tilde\omega-\cos\omega)\,\tilde e_1 +(\sin\tilde\omega-\sin\omega)\,\tilde e_2 +\cos\omega\,D_{1\cdot}+\sin\omega\,D_{2\cdot}, \] where the first two terms have norm $|e^{i\tilde\omega}-e^{i\omega}|\le|\eta|$ and the last two are bounded by $\|D\|$. Hence $|\Delta r_y|\le|\eta|+\|D\|$, and integrating over $t\in[0,y]$ (the case $y<0$ is mirrored) gives \[ |\Delta r(x,y)| \le|\Delta r(x,0)|+\int_{\min(0,y)}^{\max(0,y)}\bigl(|\eta|+\|D\|\bigr)dt \le\int_0^{|x|}\!\!A\,d\xi+B(x,y)+|y|\bigl(A(x)+B(x,y)\bigr), \] because $B(x,t)$ is nondecreasing as $t$ runs from $0$ to $y$. \end{proof} \begin{theorem}[smooth dependence for finite parameters] \label{thm:smooth-finite} Fix $n$. Let $P\subset\mathbb R^m$ be open, let $a\colon P\to(0,\infty)^n$ and $c\colon P\to\mathbb R^n$ be $C^\infty$, and let $\alpha_p,\beta_p$ be such that $(x,p)\mapsto\alpha_p(x)$ and $(y,p)\mapsto\beta_p(y)$ are jointly $C^\infty$ with $\alpha_p(0)=\beta_p(0)$ (equivalently, $p\mapsto g_p:=\Psi(\alpha_p,\beta_p)$ is a $C^\infty$ family of seeds). Then $\omega_j(p)$ ($j\le n$), $F(\omega_n(p))$ and $r(\omega_n(p))$ are jointly $C^\infty$ in $(x,y,p)$; equivalently, $p\mapsto r(\omega_n(p))$ is $C^\infty$ into every $C^k(W,\mathbb R^3)$ and $p\mapsto F(\omega_n(p))$ into every $C^k(W,\mathrm{SO}(3))$. In particular the maps $\widetilde\Phi_n(a,c,\alpha,\beta):=r(\omega_n)$ on $(0,\infty)^n\times\mathbb R^n\times\mathcal D$ and $\Phi_n(a,c,f):=r(\omega_n)$ on $(0,\infty)^n\times\mathbb R^n\times\mathcal S$ are $C^\infty$ along every such finite-dimensional smooth family; this is the sense of smoothness used in this paper. Continuity on the whole product, including the Fr\'echet factor $\mathcal D$, is Theorem~\ref{thm:continuity-D} below; Fr\'echet differentiability on $\mathcal D$ is not claimed. \end{theorem} \begin{proof} Induction over $j$. The seed $\omega_0=g_p$ is a $C^\infty$ family by Lemma~\ref{lem:sg-goursat-smooth}, and Lemma~\ref{lem:bt-smooth} propagates the property along the chain (each $\omega_j(p)$ is again a seed by Lemma~\ref{lem:bt}). Finally, Lemma~\ref{lem:immersion-smooth} composes with $\omega\mapsto(F(\omega),r(\omega))$. If the seed is given as a smooth family $f_p\in\mathcal S$, then $\alpha_p=f_p(\cdot,0)$, $\beta_p=f_p(0,\cdot)$ and $f_p=\Psi(\alpha_p,\beta_p)$, so $\Phi_n$ is covered as well. \end{proof} \begin{corollary}[continuity along paths of data] \label{cor:paths} In the setting of Theorem~\ref{thm:smooth-finite}, let $t\mapsto p(t)\in P$ be a curve. If it is continuous, then $t\mapsto r(\omega_n(p(t)))$ is continuous in every $C^k(W,\mathbb R^3)$; if it is $C^\infty$, then so is $t\mapsto r(\omega_n(p(t)))$ in every $C^k$. In particular this applies to deformations of the B\"acklund parameters and constants at a fixed seed ($P=(0,\infty)^n\times\mathbb R^n$) and to affine deformations $(\alpha_t,\beta_t)=(1-t)(\alpha_0,\beta_0)+t(\alpha_1,\beta_1)$ of the characteristic data. Moreover, for \emph{every} continuous curve $t\mapsto(a(t),c(t),\alpha_t,\beta_t)$ in $(0,\infty)^n\times\mathbb R^n\times\mathcal D$ the curve $t\mapsto r(\omega_n(t))$ is continuous in every $C^k$ (Theorem~\ref{thm:continuity-D}). A quantitative bound for the deformation is obtained by combining the variation systems of Lemmas~\ref{lem:bt-smooth}--\ref{lem:immersion-smooth} with Lemma~\ref{lem:immersion-continuity} (applied with $\eta=\delta\omega_n$). \end{corollary} \begin{proof} Theorem~\ref{thm:smooth-finite} composed with the curve ($C^\infty$ maps are continuous); the last assertion is Theorem~\ref{thm:continuity-D}. \end{proof} \subsection{Continuity on the data space: Banach estimates} \label{sec:smooth-banach} Theorem~\ref{thm:smooth-finite} is smoothness along finite-dimensional families. The following estimates give, by purely Banach-space arguments (no Fr\'echet calculus), continuity---in fact local Lipschitz continuity with respect to $C^m$-seminorms---of the whole chain on $(0,\infty)^n\times\mathbb R^n\times\mathcal D$. All constants are allowed to be large (they contain Gronwall factors, cf.\ Remark~\ref{rem:degenerate}); only finiteness is used. For $g\in C^k(W)$ put $[g]_k:=\max_{i+j\le k}\|\partial_x^i\partial_y^jg\|_{L^\infty(W)}$ (for matrix- or vector-valued $g$: the maximum over entries; products of $3\times3$ matrices cost a harmless factor $3$). Then $[g]_{k+1}=\max\bigl([g]_0,[g_x]_k,[g_y]_k\bigr)$. \begin{lemma}[Banach-algebra estimates]\label{lem:banach} Let $g,h\in C^k(W)$ and $\rho\ge2^k\max([g]_k,[h]_k)$. \begin{enumerate}[label=\textup{(\roman*)}] \item $[gh]_k\le2^k[g]_k[h]_k$. \item $[\sin g]_k\le\sinh\rho$, $[\cos g]_k\le\cosh\rho$, $[\sin g-\sin h]_k\le\cosh\rho\,[g-h]_k$, $[\cos g-\cos h]_k\le\sinh\rho\,[g-h]_k$. \item $\bigl[\int_0^x\!\!\int_0^yg\bigr]_{k+1}\le c_L[g]_k$, $c_L:=\max(1,L^2)$. \end{enumerate} \end{lemma} \begin{proof} (i) is the Leibniz rule: $\partial_x^i\partial_y^j(gh)$ is a sum of products of derivatives with binomial weights summing to $2^{i+j}\le2^k$. (ii) By (i), $N:=2^k[\cdot]_k$ is submultiplicative, so for $m\ge1$ $N(g^m-h^m)\le m\rho^{m-1}N(g-h)$ (write $g^m-h^m=(g-h)\sum g^ih^{m-1-i}$) and $N(g^m)\le\rho^m$; hence $[g^m-h^m]_k\le m\rho^{m-1}[g-h]_k$ and $[g^m]_k\le\rho^m$. The series of $\sin$ and $\cos$ converge absolutely in $C^k(W)$ and their sums are $\sin\circ g$, $\cos\circ g$; summing, $\sum_m m|c_m|\rho^{m-1}$ equals $\cosh\rho$ for $\sin$ and $\sinh\rho$ for $\cos$, and $\sum|c_m|\rho^m$ equals $\sinh\rho$, $\cosh\rho$ (for $\cos$ the term $m=0$ contributes $[1]_k=1$). (iii) Let $G=\iint g$. Then $|G|\le L^2[g]_0$; for $i,j\ge1$, $\partial_x^i\partial_y^jG=\partial_x^{i-1}\partial_y^{j-1}g$; and for $1\le i\le k+1$, $\partial_x^iG=\int_0^y\partial_x^{i-1}g\,dt$, similarly in $y$, bounded by $L[g]_k$. \end{proof} \begin{lemma}[Lipschitz dependence of the seed on the data]\label{lem:goursat-lip} Let $(\alpha_i,\beta_i)\in\mathcal D$, $f_i=\Psi(\alpha_i,\beta_i)$ ($i=1,2$), and $[\alpha_i]_m,[\beta_i]_m\le R$. Then $[f_i]_m\le C$ and \[ [f_1-f_2]_m\le C\bigl([\alpha_1-\alpha_2]_m+[\beta_1-\beta_2]_m\bigr), \] with $C=C(m,L,R)$. \end{lemma} \begin{proof} Write $f_i=h_0^i+\iint\sin f_i$ (Step~2 of Lemma~\ref{lem:sg-goursat-smooth}), $\Delta f=f_1-f_2$. \emph{$m=0$:} $|f_i|\le[h_0^i]_0+L^2$, and $\Delta f=\Delta h_0+\mathcal V_\varphi\Delta f$ with $\varphi:=\int_0^1\cos(f_2+\tau\Delta f)\,d\tau$, $|\varphi|\le1$, so Step~1 of that proof gives $[\Delta f]_0\le e^{2L}[\Delta h_0]_0$. \emph{$m\to m+1$:} by Lemma~\ref{lem:banach}, $[f_i]_{m+1}\le[h_0^i]_{m+1}+c_L\sinh\rho_m$ and $[\Delta f]_{m+1}\le[\Delta h_0]_{m+1}+c_L\cosh\rho_m[\Delta f]_m$, with $\rho_m=2^m\max[f_i]_m$. Finally $[h_0^1-h_0^2]_k\le2[\alpha_1-\alpha_2]_k+[\beta_1-\beta_2]_k$. \end{proof} \begin{lemma}[Lipschitz dependence of one B\"acklund step]\label{lem:bt-lip} Let $g_1,g_2\in\mathcal S$, $a_i\in[R^{-1},R]$, $|c_i|\le R$, $[g_i]_{k+1}\le R$, and let $q_i$ be the solution of \eqref{eq:BT} with seed $g_i$, parameter $a_i$, $q_i(0,0)=c_i$. Then $[q_i]_k\le C$ and \[ [q_1-q_2]_k\le C\bigl(|a_1-a_2|+|c_1-c_2|+[g_1-g_2]_{k+1}\bigr),\qquad C=C(k,L,R). \] \end{lemma} \begin{proof} Let $u_i=(q_i+g_i)/2$, so $u_{i,x}=-a_i^{-1}\sin(u_i-g_i)$, $u_{i,y}=g_{i,y}-a_i\sin u_i$, $q_i=2u_i-g_i$ (Lemma~\ref{lem:bt}); write $\Delta u=u_1-u_2$, etc. \emph{Bounds.} $|u_i(0,0)|\le R$, $|u_i(x,0)-u_i(0,0)|\le LR$ (as $|u_x|\le1/a$), and $|u_i(x,y)-u_i(x,0)|\le L(R+R)$; so $[u_i]_0\le C$. If $[u_i]_j\le C$, then $[u_i]_{j+1}=\max([u_i]_0,[u_{i,x}]_j,[u_{i,y}]_j)\le C'$ by Lemma~\ref{lem:banach}(ii), because $u_{i,x}$, $u_{i,y}$ are given by the formulas above; this gives $[u_i]_{k+1}\le C$, hence $[q_i]_k\le C$. \emph{Level $0$.} On the axis let $v(x)=\Delta u(x,0)$. Then $|v'|\le|a_1^{-1}-a_2^{-1}|+a_2^{-1}(|v|+[\Delta g]_0)$, so by Gronwall $|v(x)|\le e^{LR}\bigl(|v(0)|+L R^2|\Delta a|+LR[\Delta g]_0\bigr)$, with $|v(0)|\le\frac12(|\Delta c|+[\Delta g]_0)$. In a column, $\Delta u_y=\Delta g_y-\Delta a\,\sin u_1-a_2(\sin u_1-\sin u_2)$, so $|\Delta u_y|\le[\Delta g]_1+|\Delta a|+R|\Delta u|$ and $|\Delta u(x,y)|\le e^{RL}\bigl(|v(x)|+L([\Delta g]_1+|\Delta a|)\bigr)$. Thus $[\Delta u]_0\le C(|\Delta a|+|\Delta c|+[\Delta g]_1)$. \emph{Induction.} Suppose $[\Delta u]_j\le C(|\Delta a|+|\Delta c|+[\Delta g]_{j+1})$. From $\Delta u_x=-(a_1^{-1}-a_2^{-1})\sin(u_1-g_1)-a_2^{-1}\bigl(\sin(u_1-g_1)-\sin(u_2-g_2)\bigr)$ and the formula for $\Delta u_y$ above, Lemma~\ref{lem:banach}(ii) gives \[ [\Delta u_x]_j+[\Delta u_y]_j\le C\bigl(|\Delta a|+[\Delta u]_j+[\Delta g]_{j+1}\bigr), \] hence $[\Delta u]_{j+1}\le C(|\Delta a|+|\Delta c|+[\Delta g]_{j+1})$, which is stronger than required. Finally $\Delta q=2\Delta u-\Delta g$. \end{proof} \begin{lemma}[Lipschitz dependence of the immersion on $\omega\in\mathcal S$]\label{lem:imm-lip} Let $\omega_1,\omega_2\in\mathcal S$ with $[\omega_i]_{k+1}\le R$. Then $[F(\omega_i)]_k+[r(\omega_i)]_k\le C$ and \[ [F(\omega_1)-F(\omega_2)]_k+[r(\omega_1)-r(\omega_2)]_k\le C\,[\omega_1-\omega_2]_{k+1},\qquad C=C(k,L,R). \] \end{lemma} \begin{proof} Put $\eta=\omega_1-\omega_2$, $D=F(\omega_1)-F(\omega_2)$, $\Delta r=r(\omega_1)-r(\omega_2)$. \emph{Level $0$:} by Lemma~\ref{lem:immersion-continuity} (its proof gives $\|D\|\le A+B$), $[D]_0\le L([\eta_x]_0+[\eta]_0)$ and $[\Delta r]_0\le(3L^2+L)[\eta]_1$. \emph{Higher levels:} since $\omega_i\in\mathcal S$, Lemma~\ref{lem:4c} gives $F_x=P(\omega_x)F$, $F_y=Q(\omega)F$, $r_x=e_1$, $r_y=\cos\omega\,e_1+\sin\omega\,e_2$ throughout $W$. Hence \[ D_x=P(\omega_{1,x})D+\eta_xJF(\omega_2),\qquad D_y=Q(\omega_1)D+\bigl(Q(\omega_1)-Q(\omega_2)\bigr)F(\omega_2), \] and $[F(\omega_i)]_{j+1}=\max([F]_0,[F_x]_j,[F_y]_j)$ is bounded inductively by Lemma~\ref{lem:banach}. If $[D]_j\le C[\eta]_{j+1}$, then $[D_x]_j+[D_y]_j\le C([D]_j+[\eta]_{j+1})$ (Lemma~\ref{lem:banach}(ii) for $Q(\omega_1)-Q(\omega_2)$), so $[D]_{j+1}\le C[\eta]_{j+2}$. For $r$: $\Delta r_x$ is the first row of $D$, and $\Delta r_y=(\cos\omega_1-\cos\omega_2)e_1(\omega_1)+(\sin\omega_1-\sin\omega_2)e_2(\omega_1)+\cos\omega_2\,\Delta e_1+\sin\omega_2\,\Delta e_2$, so $[\Delta r]_{j+1}=\max([\Delta r]_0,[\Delta r_x]_j,[\Delta r_y]_j)\le C[\eta]_{j+2}$. \end{proof} \begin{theorem}[continuity on the data space]\label{thm:continuity-D} Fix $n$ and $k\ge0$, put $K:=k+n+1$, and let $B_R:=\{(a,c,\alpha,\beta):\ a_j\in[R^{-1},R],\ |c_j|\le R,\ [\alpha]_K,[\beta]_K\le R\}$. There is $C=C(n,k,L,R)$ such that for $p_1,p_2\in B_R$ \[ [F(\omega_n^{(1)})-F(\omega_n^{(2)})]_k+[r(\omega_n^{(1)})-r(\omega_n^{(2)})]_k \le C\Bigl(|a^{(1)}-a^{(2)}|+|c^{(1)}-c^{(2)}|+[\alpha^{(1)}-\alpha^{(2)}]_K+[\beta^{(1)}-\beta^{(2)}]_K\Bigr). \] Consequently $\widetilde\Phi_n$ is continuous (locally Lipschitz in the seminorms $[\cdot]_K$) on $(0,\infty)^n\times\mathbb R^n\times\mathcal D$ into every $C^k(W,\mathbb R^3)$, and $\Phi_n$ is continuous on $(0,\infty)^n\times\mathbb R^n\times\mathcal S$ ($\mathcal S\subset C^\infty(W)$ with its $C^\infty$-topology). In particular $t\mapsto r(\omega_n(t))$ is continuous for every continuous curve of data. \end{theorem} \begin{proof} By Lemma~\ref{lem:goursat-lip} with $m=K$, the seeds satisfy $[\omega_0^{(i)}]_K\le C$ and $[\omega_0^{(1)}-\omega_0^{(2)}]_K\le C([\Delta\alpha]_K+[\Delta\beta]_K)$. Apply Lemma~\ref{lem:bt-lip} successively with $k$ replaced by $K-j$ ($j=1,\dots,n$): it gives $[\omega_j^{(i)}]_{K-j}\le C$ and $[\Delta\omega_j]_{K-j}\le C(|\Delta a_j|+|\Delta c_j|+[\Delta\omega_{j-1}]_{K-j+1})$; after $n$ steps, $[\Delta\omega_n]_{k+1}\le C(|\Delta a|+|\Delta c|+[\Delta\omega_0]_K)$ and $[\omega_n^{(i)}]_{k+1}\le C$. Lemma~\ref{lem:imm-lip} (both $\omega_n^{(i)}\in\mathcal S$) gives the estimate. Since $\mathcal D$ is metrizable and every convergent sequence in it is bounded in each seminorm $[\cdot]_K$, continuity follows; $\Phi_n=\widetilde\Phi_n\circ(f\mapsto(f(\cdot,0),f(0,\cdot)))$ with a continuous linear restriction map. \end{proof} \begin{remark} Only continuity (with a loss of $n+1$ derivatives) is claimed; the constants contain $e^{O(RL)}$ factors and are useless as $a\to\infty$, where the cumulative estimates of Theorem~\ref{thm:main} take over (Remark~\ref{rem:degenerate}). Fr\'echet differentiability on $\mathcal D$ is not claimed. \end{remark} \begin{remark}[the continuity constants degenerate at large $a$] \label{rem:degenerate} The smoothness of Theorem~\ref{thm:smooth-finite} is a statement at \emph{fixed} parameters; its implicit constants degenerate as the parameters grow. For example, the variation in $c$ of one step (Lemma~\ref{lem:bt-smooth}) satisfies, on the axis, $|V(x,0)|\le|\delta c|\,e^{L/a}$ (because $|V_x|\le|V|/a$), and along a column $|V(x,y)|\le|\delta c|\,e^{L/a}e^{aL}$ (because $|V_y|\le a|V|$). Thus the Gronwall bound for the Lipschitz constant of $c\mapsto r(\omega_1)$ grows like $e^{aL}$; we do not claim that this growth is attained (away from the layer the column flow contracts the variation); similar (in general worse, iterated) bounds hold for variations in $a$ and along the chain. Uniform-in-$a$ control is provided instead by the \emph{cumulative} estimates of Theorem~\ref{thm:main}, which are independent of $c$ and non-Gronwall. The two mechanisms complement each other: Theorem~\ref{thm:smooth-finite} for finite perturbations at fixed $a$, Theorem~\ref{thm:main} for the regime $a\to\infty$. \end{remark} \begin{remark}[what the flow deforms: first variation of the metric] \label{rem:bending} Corollary~\ref{cor:paths} asserts continuity of the flow; the computation behind it also identifies, to first order, \emph{what} is deformed. Let $t\mapsto\omega(t)$ be a $C^1$ family of solutions with $\eta:=\partial_t\omega|_{t=0}$ (so $\eta_{xy}=\cos\omega\,\eta$ by Lemma~\ref{lem:sg-goursat-smooth}), and let $D=\partial_tF$, $\delta r=\partial_tr$ be the variations. Since $F(t)$ is orthogonal for all $t$ (Lemma~\ref{lem:4a}(a)), the matrix $M:=(\langle e_i,\delta e_j\rangle)_{ij}=FD^{T}$ is skew-symmetric. Writing $r_x=e_1$, $r_y=\cos\omega\,e_1+\sin\omega\,e_2$ and $\delta r_y=-\eta\sin\omega\,e_1+\eta\cos\omega\,e_2+\cos\omega\,\delta e_1 +\sin\omega\,\delta e_2$, a direct computation gives \[ \langle r_x,\delta r_x\rangle=0,\qquad \langle r_y,\delta r_y\rangle=0,\qquad \langle r_x,\delta r_y\rangle+\langle r_y,\delta r_x\rangle =-\sin\omega\;\eta , \] that is, \[ \mathrm{sym}\,\langle dr,d\,\delta r\rangle =\tfrac12\,\delta I =-\sin\omega\;\eta\;dx\odot dy . \] Thus the flow preserves the Chebyshev normalization ($E=G=1$) and the curvature $K\equiv-1$ \emph{exactly} along the whole family, and touches the metric only through the angle function; the linearized equation $\eta_{xy}=\cos\omega\,\eta$ is precisely the first-order condition of staying on the constant-curvature locus. Consequently the infinitesimal-bending identity $\mathrm{sym}\langle dr,d\,\delta r\rangle=0$ (Darboux) holds along the flow if and only if $\sin\omega\cdot\eta\equiv0$; if the nondegenerate locus $\{\sin\omega\neq0\}$ is dense in $W$ (e.g.\ for real-analytic solutions not identically $\equiv k\pi$), this forces $\eta\equiv0$, and then the homogeneous linearized systems of Lemmas~\ref{lem:immersion-smooth}--\ref{lem:bt-smooth} with the normalization $F(0,0)=I$, $r(0,0)=0$ give $D\equiv0$ and $\delta r\equiv0$. In other words: within the ASG normalization, the flow is metric-rigid---its only isometric directions are the infinitesimal rigid motions, which the normalization quotients out. Classical bendings of a $K=-1$ surface do exist, but they do not preserve the asymptotic gauge ($e=g=0$ in $II$) and are therefore not directions of this flow. At a degenerate base the identity is vacuous: for the model kink family of Proposition~\ref{prop:infinitesimal} the base $\omega\equiv0$ has $\sin\omega\equiv0$, the metric changes only in the $O(s)$-wide layer (at order $O(s^2)$ pointwise outside it), and the formal variation $q_s/s\to2\pi\,\delta(y)$ is \emph{not} a solution of the linearized equation---the family is a transition between branches ($k_- = k_+\pm1$, Proposition~\ref{prop:layer-asymptotics}), not a $C^1$-path in the solution space, which is exactly why the infinitesimal limit $\Delta_*$ of Proposition~\ref{prop:infinitesimal} is a singular layer field. \end{remark} \subsection{The total limit $a_j\to\infty$} \label{sec:smooth-infty} \begin{theorem}[continuous extension to the B\"acklund chart at infinity] \label{thm:smooth-infinity} Fix $n$ and $f\in\mathcal S$. As $\min_j a_j\to\infty$ (with arbitrary relative rates and arbitrary $c_j$), \[ \|r(\omega_n)-r((-1)^nf)\|_{L^\infty(W)} \le K_n(f)\sum_{j=1}^n\frac1{a_j}\;\longrightarrow\;0, \qquad K_n(f)=3L^2+(1+L)\pi(2T_*(f)+\pi)3^{n-1}. \] Consequently, in the coordinates $s_j=1/a_j\in(0,\infty)$, with $f\in\mathcal S$ carrying the $C^1(W)$-topology, the map $(s,c,f)\mapsto r(\omega_n)$ has at every boundary point $(0,c^0,f)$ the limit $r((-1)^nf)$, independent of $c^0$: $r(\omega_n(s,c,f'))\to r((-1)^nf)$ as $(s,c,f')\to(0,c^0,f)$. \end{theorem} \begin{proof} For a fixed seed the estimate is Theorem~\ref{thm:main} (unconditional in $c_j$ and in the ratios of the $a_j$). Let now $f_m\to f$ in $C^1(W)$. Then $T_*(f_m)\le2L\|\partial_yf_m\|_\infty$ is bounded, hence so is $K_n(f_m)\le K$. Put $\eta_m:=(-1)^n(f_m-f)$; both $(-1)^nf$ and $(-1)^nf_m$ lie in $\mathcal S$, so Lemma~\ref{lem:immersion-continuity} gives (with $A\le L\|\eta_x\|_\infty$, $B\le L\|\eta\|_\infty$) \[ \|r((-1)^nf_m)-r((-1)^nf)\|_\infty\le2L^2\|\eta_{m,x}\|_\infty+(L+L^2)\|\eta_m\|_\infty \le(2L^2+L)\|f_m-f\|_{C^1}, \] with $\|g\|_{C^1}:=\|g\|_\infty+\|g_x\|_\infty+\|g_y\|_\infty$. Therefore $\|r(\omega_n(s,c,f_m))-r((-1)^nf)\|_\infty\le K\sum_js_j+(2L^2+L)\|f_m-f\|_{C^1}\to0$, uniformly in $c$. \end{proof} Theorem~\ref{thm:smooth-infinity} concerns the \emph{immersions}. The solutions $\omega_n$ themselves need not converge: \begin{proposition}[no $C^0$-limit for $\omega_n$] \label{prop:no-C0} Let $n=1$, $f\equiv0$, $c=\pi$. Then \[ q_a(x,y)=4\arctan\bigl(e^{-x/a-ay}\bigr) \] solves \eqref{eq:BT} with seed $0$ and $q_a(0,0)=\pi$, but \[ q_a\to \begin{cases}0,&y>0\\ \pi,&y=0\\ 2\pi,&y<0\end{cases} \qquad(a\to\infty) \] pointwise on $W$, so the family $\{q_a\}$ has no limit in $C^0(W)$. Meanwhile $\|r(q_a)-r(0)\|_{L^\infty(W)}\le K_1/a\to0$ by Theorem~\ref{thm:main} (here $T_*(0)=0$, $S=1/a$): the immersion converges although the solution does not. \end{proposition} \begin{proof} With $\theta=-x/a-ay$ and $q_a=4\arctan e^{\theta}$ one has $\sin(q_a/2)=2e^{\theta}/(1+e^{2\theta})$ and \[ \partial_xq_a=-\frac4a\frac{e^{\theta}}{1+e^{2\theta}} =-\frac2a\sin\frac{q_a}{2}, \qquad \partial_yq_a=-4a\frac{e^{\theta}}{1+e^{2\theta}} =-2a\sin\frac{q_a}{2}, \] which is \eqref{eq:BT} with $g\equiv0$; and $q_a(0,0)=4\arctan1=\pi$. The pointwise limit is as stated (for $y=0$, $e^{-x/a}\to1$ uniformly on $[-L,L]$). If $q_a$ converged in $C^0(W)$, the limit would be a uniform limit of continuous functions, hence continuous, and it would coincide with the displayed step function; but the step function is discontinuous at $y=0$ while $W$ contains points with $y=\pm\varepsilon$ on both sides. The last claim is Theorem~\ref{thm:main}. \end{proof} \begin{proposition}[no $C^1$-extension of the immersion] \label{prop:no-C1} In the situation of Proposition~\ref{prop:no-C0}, put $q_s:=q_{1/s}$, $s>0$. Then $\|r(q_s)-r(0)\|_{L^\infty(W)}=O(s)$, but the difference quotient $(r(q_s)-r(0))/s$ is unbounded in $C^1(W,\mathbb R^3)$; in particular $s\mapsto r(q_s)$ (extended by $r(0)$ at $s=0$) is not a $C^1$-curve in $C^1(W,\mathbb R^3)$. \end{proposition} \begin{proof} The bound $\|r(q_s)-r(0)\|_{L^\infty}\le K_1\,s$ is Theorem~\ref{thm:main} with $n=1$, $a=1/s$, $T_*(0)=0$. On the axis, $(q_s)_x=-2s\sin(q_s/2)$, so $|(q_s)_x|\le2s$ and, by Lemma~\ref{lem:4b} with $\|P(p)-P(0)\|=|p|$ (Lemma~\ref{lem:4a}(b)), \[ \|F(q_s)(x,0)-F(0)(x,0)\|\le 2s|x| , \] hence $e_i(q_s)(x,0)\to e_i(0)(x,0)$ uniformly on $[-L,L]$. Also $q_s(x,0)=4\arctan e^{-sx}\to\pi$ uniformly. Therefore, at $y=0$, \[ r_y(q_s)(x,0) =\cos(q_s)e_1(q_s)+\sin(q_s)e_2(q_s) \;\longrightarrow\;-e_1(0)(x,0), \] while $r_y(0)(x,0)=e_1(0)(x,0)$. Hence $\sup_W|\partial_y(r(q_s)-r(0))|\ge 2-o(1)$, and \[ \Bigl\|\frac{r(q_s)-r(0)}{s}\Bigr\|_{C^1(W)} \;\ge\;\frac{2-o(1)}{s}\;\longrightarrow\;\infty . \qedhere \] \end{proof} Propositions~\ref{prop:no-C0} and \ref{prop:no-C1} show that the convergence in Theorem~\ref{thm:smooth-infinity} is genuinely $C^0$: the solution has a jump layer of size $2\pi$ and width $O(1/a)$ collapsing onto the axis $y=0$. The natural question is whether a \emph{first-order} (infinitesimal) limit exists. The answer is yes, in $C^0$ off the axis, and the limit is completely explicit; it is \emph{not} obtained by differentiating at any finite parameter, and it retains a jump across the axis---the imprint of the boundary layer. \begin{proposition}[infinitesimal limit for the model kink] \label{prop:infinitesimal} In the situation of Proposition~\ref{prop:no-C0} ($n=1$, $f\equiv0$, $c=\pi$, $q_s=q_{1/s}$), the difference quotient $(r(q_s)-r(0))/s$ converges as $s\to0^+$, uniformly on $\{|y|\ge\varepsilon'\}$ for every $\varepsilon'>0$ and pointwise on all of $W$, to the map $\Delta_*$ defined by the following linear construction. Let $F_0:=F(0)$ be the frame of the identically-zero solution with rows $e_1^{(0)},e_2^{(0)},n^{(0)}$, let $P_0,J,Q(\omega)$ be as in \S\ref{sec:immersion}, and let $\Phi_0(y):=\exp(y\,Q(0))$. Define \begin{itemize} \item[(a)] $\bar E$ as the solution of $\bar E_x=P_0\bar E-2\,J\,F_0(x,0)$, $\bar E(0,0)=0$; \item[(b)] $\Delta_*(x,0):=\int_0^x\bar E_{1\cdot}(\xi,0)\,d\xi$; \item[(c)] for $y>0$: \[ \Delta_*(x,y):=\Delta_*(x,0)+\bigl(-2e_1^{(0)}(x,0)+2e_2^{(0)}(x,0)\bigr) +\int_0^y E^{*}_{1\cdot}(x,t)\,dt, \] where $E^{*}(x,y):=\Phi_0(y)\bigl(\bar E(x)+M_+F_0(x,0)\bigr)$; \item[(d)] for $y<0$: \[ \Delta_*(x,y):=\Delta_*(x,0)+\bigl(2e_1^{(0)}(x,0)+2e_2^{(0)}(x,0)\bigr) +\int_0^y E^{*}_{1\cdot}(x,t)\,dt, \] where $E^{*}(x,y):=\Phi_0(y)\bigl(\bar E(x)+M_-F_0(x,0)\bigr)$; \end{itemize} with the constant matrices \[ M_+=\begin{pmatrix}0&0&2\\0&0&2\\-2&-2&0\end{pmatrix}, \qquad M_-=\begin{pmatrix}0&0&2\\0&0&-2\\-2&2&0\end{pmatrix}. \] The limit $\Delta_*$ has a jump across the axis: $\Delta_*(0,0^+)-\Delta_*(0,0)=(-2,2,0)\neq0$; in particular the convergence is not uniform near $y=0$ (and not in $C^1$, Proposition~\ref{prop:no-C1}). Moreover there is an absolute $c_0>0$ with \[ \|r(q_s)-r(0)\|_{L^\infty(W)}\;\ge\;c_0\,s \qquad\text{for all } s\le s_0 . \] \end{proposition} \begin{proof} \emph{Preliminaries.} Write $F_s:=F(q_s)$, $F_0:=F(0)$, $D:=F_s-F_0$, $E:=D/s$, $\Delta r:=r(q_s)-r(0)$, and \[ \hat q_s(z;x):=q_s(x,sz)=4\arctan e^{-sx-z},\qquad \hat q(z):=4\arctan e^{-z}. \] (The identity $q_s(x,sz)=4\arctan(e^{-sx-z})$ is exact; throughout the proof we suppress the parameter $x$ in $\hat q_s$, since in every estimate below the base point $x$ is held fixed.) We use repeatedly: $00$, \[ E(x,y)=\Phi_s(y)E(x,0) +\int_0^{y/s}\Phi_s(y)\Phi_s(sz)^{-1}\bigl(Q(\hat q_s(z))-Q(0)\bigr)F_0(x,sz)\,dz . \] Here $\|\Phi_s(y)-\Phi_0(y)\|\le\int_0^{|y|}\|Q(q_s)-Q(0)\|\,dt=O(s)$, and the integrand is dominated by $2\cdot4e^{sL}e^{-z}\|F_0\|_{L^\infty}\in L^1(0,\infty)$, uniformly in $x$. Splitting the integral at $z=Z$, letting $s\to0$ (on $[0,Z]$ use $\hat q_s\to\hat q$ uniformly, $\Phi_s(y)\Phi_s(sz)^{-1}\to\Phi_0(y)$ uniformly, $F_0(x,sz)\to F_0(x,0)$, and $E(x,0)\to\bar E(x)$) and then $Z\to\infty$ (dominated convergence) yields \[ E(x,y)\;\longrightarrow\;\Phi_0(y)\Bigl(\bar E(x) +\Bigl[\int_0^\infty\bigl(Q(\hat q(z))-Q(0)\bigr)dz\Bigr]F_0(x,0)\Bigr), \] uniformly in $(x,y)$ with $y\ge\varepsilon'$; by \eqref{eq:layer} the bracketed matrix is $M_+$ (entrywise: the $(1,3)$- and $(2,3)$-entries are $\int\sin\hat q=2$ and $\int(1-\cos\hat q)=2$, the others follow by skew-symmetry). For $y<0$ the same formula holds with the \emph{oriented} integral $\int_0^{y/s}=-\int_{y/s}^{0}$. With $z\in(y/s,0)$ and the domination $2\pi-\hat q_s(z)\le4e^{sL}e^{z}$ the limit is $\Phi_0(y)\bigl(\bar E(x)+M_-F_0(x,0)\bigr)$ with $M_-=-\int_{-\infty}^{0}\bigl(Q(\hat q(z))-Q(0)\bigr)dz$; by \eqref{eq:layer} its $(1,3)$-, $(2,3)$-entries are $-\int_{-\infty}^0\sin\hat q=2$ and $-\int_{-\infty}^0(1-\cos\hat q)=-2$, the others follow by skew-symmetry. This is $E^*$ as in (d), uniformly on $\{|y|\ge\varepsilon'\}$. \emph{4. Position limit.} From $r_y(q_s)=\cos q_s\,e_1(q_s)+\sin q_s\,e_2(q_s)$, $r_y(0)=e_1^{(0)}$ and $e_1(q_s)=e_1^{(0)}+D_{1\cdot}$, \[ \Delta r(x,y)=\Delta r(x,0) +\int_0^y\Bigl[(\cos q_s-1)\,e_1(q_s)+\sin q_s\,e_2(q_s) +D_{1\cdot}\Bigr](x,t)\,dt . \] Divide by $s$, substitute $t=sz$, and use dominated convergence with the dominators $|\cos\hat q_s-1|\le\hat q_s^2/2\le8e^{2sL}e^{-2z}$ and $|\sin\hat q_s|\le4e^{sL}e^{-z}$ (resp.\ their mirror for $z<0$): \[ \frac1s\int_0^y(\cos q_s-1)e_1(q_s)\,dt \longrightarrow e_1^{(0)}(x,0)\int_0^{\infty}(\cos\hat q-1)\,dz =-2e_1^{(0)}(x,0), \] \[ \frac1s\int_0^y\sin q_s\,e_2(q_s)\,dt \longrightarrow e_2^{(0)}(x,0)\int_0^\infty\sin\hat q\,dz =2e_2^{(0)}(x,0), \] and $\frac1s\int_0^yD_{1\cdot}\,dt=\int_0^yE_{1\cdot}(x,t)\,dt \to\int_0^yE^{*}_{1\cdot}(x,t)\,dt$ (uniformly on $y\ge\varepsilon'$). This is (c); the $y<0$ computation is identical with the interval $(y/s,0)$, the constants $-2$ and $-2$ of \eqref{eq:layer}, and an overall change of sign from reversing the integration, giving (d). \emph{5. Jump and sharpness.} From (b),(c), $\Delta_*(x,0^+)-\Delta_*(x,0)=(-2e_1^{(0)}+2e_2^{(0)})(x,0)$; at $x=0$ this is $(-2,2,0)$. A uniform limit of continuous functions on $W$ would be continuous, while $\Delta_*$ is discontinuous across $y=0$; hence the convergence cannot be uniform near the axis (it is uniform exactly on $\{|y|\ge\varepsilon'\}$). Finally, for fixed small $y_0>0$, $\Delta_*(0,y_0)=(-2,2,0)+O(y_0)$, so $|\Delta_*(0,y_0)|\ge\sqrt2$ for $y_0\le y_0'$, and therefore $\|r(q_s)-r(0)\|_{L^\infty}\ge s\,|\Delta_*(0,y_0)|/2\ge c_0s$ for all small $s$. \end{proof} \begin{corollary}[sharpness of the order in Theorem~\ref{thm:main}] \label{cor:sharp} In the family of Proposition~\ref{prop:no-C0} ($n=1$, $f\equiv0$, $c=\pi$, $a=1/s$) the immersion error is exactly of order $1/a$: the exponent $1$ in $\sum 1/a_j$ cannot be improved. \end{corollary} \begin{proof} Immediate from Proposition~\ref{prop:infinitesimal}: the first-order term $\Delta_*$ is nonzero, and the error is $s\,\Delta_*+o(s)$ off the axis. \end{proof} \begin{corollary}[order-sharpness for every fixed $n$] \label{cor:sharp-all-n} Fix $n\ge1$ and $L>0$. There exist constants $c_n=c_n(L)>0$ and $s_0=s_0(n,L)>0$ such that for all $s\in(0,s_0)$, with \[ f\equiv0,\qquad c_j=\pi\ \ (j=1,\dots,n),\qquad a_j:=s^{-j}, \] one has \[ \|r(\omega_n)-r((-1)^n f)\|_{L^\infty(W)} =\|r(\omega_n)-r(0)\|_{L^\infty(W)} \;\ge\; c_n\sum_{j=1}^n\frac1{a_j}. \] The constant may be taken $c_n=c_0/4^{\,n-1}$, where $c_0=c_0(L)>0$ is the constant of Corollary~\ref{cor:sharp}. \end{corollary} \begin{proof} Induction on $n$. \emph{Base $n=1$:} $a_1=1/s$ (chain constant $\pi$), $f\equiv0$, so $\omega_1=q_s=4\arctan(e^{-sx-y/s})$, and Corollary~\ref{cor:sharp} gives $\|r(q_s)-r(0)\|_\infty\ge c_0 s=c_0/a_1$ for all $s\le s_0(1)$, where $s_0(1)>0$ and $c_0>0$ are the constants of that corollary; here $\sum_{j=1}^1 1/a_j=1/a_1$. This proves the claim for $n=1$ with $c_1:=c_0$ and $s_0(1)$ as above. \emph{Step.} Assume the claim for $n-1$ with constant $c_{n-1}$, and append $a_n=s^{-n}$ (with chain constant $\pi$, not to be confused with the estimate constant $c_n$). Write $g:=\omega_{n-1}$. By Theorem~\ref{thm:main} applied with $n=1$ to the seed $g$ (which solves SG by Lemma~\ref{lem:bt}), \[ \|r(\omega_n)-r(-g)\|_{L^\infty}\le\frac{K_1(g)}{a_n}=K_1(g)\,s^n, \qquad K_1(g)=3L^2+(1+L)\pi\bigl(2T_*(g)+\pi\bigr). \] By Lemma~\ref{lem:S-orient}, $r(-g)=r(g)R$ with $R=\operatorname{diag}(1,-1,-1)$ orthogonal, and $r(0)R=r(0)$ (the immersion of the identically zero solution, $r(0)(x,y)=(x+y)\varepsilon_1$, is invariant under $R$); hence $\|r(-g)-r(0)\|_\infty=\|r(g)-r(0)\|_\infty$. The triangle inequality gives \[ \|r(\omega_n)-r(0)\|_\infty \ge\|r(g)-r(0)\|_\infty-K_1(g)\,s^n \ge c_{n-1}\sum_{j=1}^{n-1}\frac1{a_j}-K_1(g)\,s^n . \] By Lemma~\ref{lem:S-T} with $f\equiv0$ (so $T_*(f)=0$), $T_*(g)\le\pi(3^{n-1}-1)$ uniformly in $s$, hence \[ K_1(g)\le C_n:=3L^2+(1+L)\pi\bigl(2\pi(3^{n-1}-1)+\pi\bigr) \] independently of $s$. Also $\sum_{j=1}^{n-1}1/a_j=s+O(s^2)\ge s$ for $s\le\tfrac12$. Choose \[ s_0(n):=\min\Bigl(s_0(n-1),\,\tfrac12,\,\bigl(c_{n-1}/(2C_n)\bigr)^{1/(n-1)}\Bigr)>0 ; \] then for $s\in(0,s_0(n))$ the induction hypothesis applies (as $s0$, $c_j\in\mathbb R$: \begin{enumerate}[label=\textup{(\roman*)}] \item $\|r(\omega_n)-r(\mathrm{ref}_n)\|_{L^\infty(W)}\le K_nS$; \item $\|\partial_x(r(\omega_n)-r(\mathrm{ref}_n))\|_{L^\infty(W)}\le K_n'S$ and $\|F_n-F_{\mathrm{ref}}\|_{L^\infty(W)}\le K_n'S$; \item for every $1\le p<\infty$, \[ \|\partial_y(r(\omega_n)-r(\mathrm{ref}_n))\|_{L^p(W)} \le\pi^{1-\frac1p}\bigl(2L\pi(2T_*+\pi)3^{n-1}S\bigr)^{\frac1p}+(4L^2)^{\frac1p}K_n'S; \] \item hence $r(\omega_n)\to r(\mathrm{ref}_n)$ in $W^{1,p}(W)$ for all $p<\infty$, with rate $S^{1/p}$ for the derivative $\partial_yr$, and the family $\{r(\omega_n)\}$ is uniformly Lipschitz ($|r_x|=|r_y|=1$). \end{enumerate} The rates in \textup{(i)} and \textup{(iii)} cannot be improved, and the convergence does not hold in $W^{1,\infty}$ (Proposition~\ref{prop:S-nodiff}). \end{theorem} \begin{proof} (i) is Theorem~\ref{thm:main}. Apply Lemma~\ref{lem:S-comp} to $\omega_n,\mathrm{ref}_n$: here $\delta\le2S$ (Lemma~\ref{lem:5}) and $\Lambda_y(x)\le\Lambda(x)\le\pi(2T_*+\pi)3^{n-1}S$ (Lemma~\ref{lem:10}), which gives $\|D\|\le(2L+\pi(2T_*+\pi)3^{n-1})S=K_n'S$ and (ii). For (iii): $|\partial_y\Delta r|\le d_n+K_n'S$, $d_n:=d(\omega_n-\mathrm{ref}_n)\le\pi$, so $\iint d_n^p\le\pi^{p-1}\iint d_n\le\pi^{p-1}\cdot2L\cdot\pi(2T_*+\pi)3^{n-1}S$, and Minkowski gives (iii). (iv) follows from (i)--(iii). \end{proof} \begin{proposition}[sharpness; no $C^1$-convergence and no differentiability at infinity]\label{prop:S-nodiff} Let $n=1$, $f\equiv0$, $c_1=\pi$, $a=1/s$. Then $\omega_1=q_s:=4\arctan e^{-sx-y/s}$, $\mathrm{ref}_1\equiv0$, $r(0)(x,y)=(x+y)\varepsilon_1$ with $\varepsilon_i$ the standard basis, and for $00$ on $(0,\infty)$; it is therefore positive for small $s$. \end{proposition} \begin{proof} That $q_s$ solves \eqref{eq:BT} with $g\equiv0$ and $q_s(0,0)=\pi$ is a direct check ($\theta=-sx-y/s$, $(q_s)_x=-2s\sin\frac{q_s}2$, $(q_s)_y=-\frac2s\sin\frac{q_s}2$; this is the family of Proposition~\ref{prop:no-C0} with $a=1/s$). For $\omega\equiv0$: $P(0)=P_0$ and $Q(0)$ leave $e_1$ unchanged, so $e_1\equiv\varepsilon_1$ and $r(0)=(x+y)\varepsilon_1$. (a) At the origin the frame is $I$, so $r_y(q_s)(0,0)=\cos\pi\,\varepsilon_1=-\varepsilon_1$, whereas $r_y(0)(0,0)=\varepsilon_1$. (b) $\partial_y\Delta r=(\cos q_s,\sin q_s,0)F(q_s)-(1,0,0)F(0)=(\cos q_s-1,\sin q_s,0)F(0)+(\cos q_s,\sin q_s,0)(F(q_s)-F(0))$, so $|\partial_y\Delta r|\ge|e^{iq_s}-1|-\|F(q_s)-F(0)\|$. Here $|e^{iq_s}-1|=2|\sin(q_s/2)|=2\operatorname{sech}(sx+y/s)\ge2\operatorname{sech}2>\tfrac12$ on $\{|y|\le s\}$ if $sL\le1$, and by Lemma~\ref{lem:S-comp}, $\|F(q_s)-F(0)\|\le\delta L+\Lambda\le(2L+\pi^2)s$ (using $\delta\le2s$, $\Lambda\le\pi^2s$ from Lemmas~\ref{lem:5}, \ref{lem:10}, with $T_*(0)=0$). For $s$ small this is $\le\tfrac14$ on a set of measure $4Ls$. (c),(d) Compute at $x=0$. There $q_s(0,y)=\hat\Theta(y/s)$ and $\|F(q_s)(0,y)-I\|\le|y|$ (Lemma~\ref{lem:4b}, $\|Q\|=1$), so $|e_i(0,t)-\varepsilon_i|\le|t|$. Hence, for $T\in\mathbb R$, with $g(\tau)=(\cos\hat\Theta(\tau)-1)\varepsilon_1+\sin\hat\Theta(\tau)\,\varepsilon_2$, \[ \Delta_s(0,sT)=\frac1s\int_0^{sT}\bigl[(\cos q_s-1)e_1+\sin q_s\,e_2+(e_1-\varepsilon_1)\bigr]dt =\int_0^Tg(\tau)\,d\tau+E,\qquad|E|\le2sT^2 . \] (The integrand of the error is bounded by $(2+1+1)|t|$.) Now $\hat\Theta(-\tau)=2\pi-\hat\Theta(\tau)$, so $\sin\hat\Theta$ is odd and $\cos\hat\Theta$ is even, and $\sin(\hat\Theta/2)=\operatorname{sech}\tau$; hence $\int_{-1}^1g=-2\int_{-1}^1\operatorname{sech}^2\tau\,d\tau\,\varepsilon_1=-4\tanh1\,\varepsilon_1$, which gives (d). From (d), $\max(|\Delta_s(0,s)|,|\Delta_s(0,-s)|)\ge2\tanh1-2s\ge1$ for $s\le\tfrac14$, which is (c). The $\varepsilon_2$-component is $\int_0^1\sin\hat\Theta\,d\tau+O(s)$, and $\sin\hat\Theta>0$ on $(0,\infty)$. A uniformly convergent sequence of continuous functions on a compact set is equicontinuous, whence the non-differentiability. \end{proof} \subsection{One step outside the boundary layer} \label{sec:smooth-layer} Propositions~\ref{prop:no-C0}--\ref{prop:infinitesimal} exhibit a boundary layer of width $O(1/a)$ at the axis, across which $\omega$ jumps by $2\pi$. Outside the layer the solution is close to an explicit quasi-equilibrium: \begin{proposition}[one-step asymptotics off the boundary layer] \label{prop:layer-asymptotics} Let $g\in\mathcal S$, $a>0$, $c\in\mathbb R$, and let $q_a$ be the solution of \eqref{eq:BT} with seed $g$ and $q_a(0,0)=c$ (Lemma~\ref{lem:bt}); put $u:=(q_a+g)/2$ and $u_0:=(c+g(0,0))/2=u(0,0)$. Assume \[ d_0:=\operatorname{dist}(u_0,\pi\mathbb Z)>0 . \] Let $k_+$ be the even and $k_-$ the odd integer nearest to $u_0/\pi$ (both well defined because $u_0/\pi\notin\mathbb Z$, i.e.\ precisely because $d_0>0$: the nearest \emph{even} integer is ambiguous only when $u_0/\pi$ is an odd integer, the nearest \emph{odd} one only when $u_0/\pi$ is an even integer). Then for every $\varepsilon>0$ there are $a_0=a_0(d_0,\varepsilon,L,\|g\|_{C^2(W)})$ and $C=C(d_0,\varepsilon,L,\|g\|_{C^2(W)})$ such that for all $a\ge a_0$, uniformly in $x\in[-L,L]$, \begin{align*} \sup_{y\in[\varepsilon,L]}\Bigl| q_a(x,y)+g(x,y)-2\pi k_+-\frac{2}{a}g_y(x,y) \Bigr|&\le\frac{C}{a^2},\\ \sup_{y\in[-L,-\varepsilon]}\Bigl| q_a(x,y)+g(x,y)-2\pi k_-+\frac{2}{a}g_y(x,y) \Bigr|&\le\frac{C}{a^2}. \end{align*} \end{proposition} \begin{proof} Write $G:=\|g_y\|_{C^0}$, $G_2:=\|g_{yy}\|_{C^0}$ and let $a$ be large (several lower bounds below). The potential $u$ solves (Lemma~\ref{lem:bt}) \begin{equation}\label{eq:u-pend} u_x=-\frac1a\sin(u-g),\qquad u_y=g_y-a\sin u . \end{equation} \emph{Step 0 (axis drift).} From \eqref{eq:u-pend}, $|u_x|\le1/a$, so $|u(x,0)-u_0|\le L/a\le d_0/2$ for $a\ge2L/d_0$. Hence $\operatorname{dist}(u(x,0),\pi\mathbb Z)\ge d_0/2$ and, since $|u_0/\pi-k_+|\le1-\delta'$ with $\delta':=d_0/\pi$ and $L/(a\pi)\le\delta'/2$, \[ u(x,0)\in\bigl((k_+-1)\pi+\tfrac{d_0}2,\ (k_++1)\pi-\tfrac{d_0}2\bigr) \quad\text{for all }x . \] \emph{Step 1 (basin, forward).} Fix $x$ and integrate the second equation of \eqref{eq:u-pend} in $y$ from the initial value $u(x,0)$. Set $\rho_a:=K/\sqrt a$ with $K:=\pi(G+1)$, and assume $a\ge a_1$ so that $a\sin(d_0/4)\ge2G$ and $(2K/\pi)\sqrt a\ge2G$. \emph{Invariance:} the interval $I_a:=[k_+\pi-\rho_a,\,k_+\pi+\rho_a]$ is invariant: at $u=k_+\pi+\rho_a$ (recall $k_+$ is even, so $\sin u=\sin\rho_a>0$) \[ u_y\le G-a\sin\rho_a\le G-a\cdot\tfrac2\pi\rho_a =G-\tfrac{2K}{\pi}\sqrt a<0, \] and symmetrically at $u=k_+\pi-\rho_a$. Moreover $u$ cannot leave $((k_+-1)\pi,\ (k_++1)\pi)$: at distance $\rho_a$ from an endpoint the same estimate shows $u_y$ points inwards. \emph{Attraction:} while $\rho_a\le|u-k_+\pi|\le\pi-d_0/4<\pi$, the sign of $\sin(u-k_+\pi)$ is that of $u-k_+\pi$, so $u_y=g_y-a\sin(u-k_+\pi)$ points towards $k_+\pi$ whenever $a|\sin(u-k_+\pi)|>G$. We bound the speed in two sub-regions. For $\rho_a\le|u-k_+\pi|\le\pi/2$ we have $|\sin(u-k_+\pi)|\ge\frac2\pi|u-k_+\pi|\ge\frac2\pi\rho_a$, so \[ |u_y|\ge a\cdot\tfrac2\pi\rho_a-G =\tfrac{2K}{\pi}\sqrt a-G>0 . \] For $\pi/2<|u-k_+\pi|\le\pi-d_0/4$ we have $|\sin(u-k_+\pi)|=\sin\bigl(\pi-|u-k_+\pi|\bigr)\ge\sin(d_0/4)$, so \[ |u_y|\ge a\sin(d_0/4)-G\;\ge\;\tfrac a2\sin(d_0/4)>0 \] for $a\ge a_1$ (by the definition of $a_1$). In both sub-regions the velocity is directed towards $k_+\pi$ and bounded below by $\min\bigl(\tfrac{2K}{\pi}\sqrt a-G,\ \tfrac a2\sin(d_0/4)\bigr)$, so $u$ enters $I_a$ within time \[ \tau_1\le\frac{\pi/2}{\tfrac{2K}{\pi}\sqrt a-G} +\frac{\pi}{\tfrac a2\sin(d_0/4)} \le\frac{C_1}{\sqrt a}, \] uniformly in $x$ (enlarging $C_1$). \emph{Step 2 (tracking, forward).} For $y\ge\tau_1$ write $u=k_+\pi+e_a+v$ with $e_a(y):=g_y(x,y)/a$ (the exact equilibrium branch would be $e_a$), so \[ v'=u_y-e_a' =g_y-a\sin(e_a+v)-\frac{g_{yy}}a =-a\,v-\frac{g_{yy}}a+a\,\theta, \qquad |\theta|\le\frac{(|e_a|+|v|)^3}{6}, \] using $\sin(e_a+v)=e_a+v+O\bigl((|e_a|+|v|)^3\bigr)$ and $g_y=ae_a$. At $y=\tau_1$, $|v|\le\rho_a+G/a\le K'/\sqrt a$. Choose $a\ge a_2$ so that $5\ln a/a\le\varepsilon/3$ and $C_1/\sqrt a\le\varepsilon/3$. \emph{Round 1:} while $|e_a|+|v|\le K''/\sqrt a$ the forcing is bounded by $G_2/a+C''/\sqrt a\le C'''/\sqrt a$, so \[ |v(y)|\le e^{-a(y-\tau_1)}\frac{K'}{\sqrt a} +\int_{\tau_1}^{y}e^{-a(y-t)}\frac{C'''}{\sqrt a}\,dt \le e^{-a(y-\tau_1)}\frac{K'}{\sqrt a}+\frac{C'''}{a\sqrt a}; \] at $y\ge\tau_1+5\ln a/a$ the exponential term is $O(a^{-11/2})$, so $|v|\le C_4/a^{3/2}$, and $|e_a|+|v|\le C_5/a$. \emph{Round 2:} now the forcing is bounded by $G_2/a+C_6/a^2\le C_7/a$, and repeating the estimate gives, for $y\ge\tau_1+10\ln a/a$, \[ |v(y)|\le e^{-a(y-\tau_1-5\ln a/a)}\frac{C_4}{a^{3/2}} +\frac{C_7}{a^2}\le\frac{C_8}{a^2}. \] Since $\tau_1+10\ln a/a\le\varepsilon$ for $a\ge a_0$, this holds on $[\varepsilon,L]$, uniformly in $x$. Finally \[ q_a=2u-g=2\pi k_+-g+\tfrac2a g_y+2v \qquad\Rightarrow\qquad \Bigl|q_a+g-2\pi k_+-\tfrac2ag_y\Bigr|\le\frac{2C_8}{a^2}, \] which is the first estimate ($k_+$ even, so $(-1)^{k_+}=+1$). \emph{Step 3 (backward).} For $y<0$ substitute $\tau=-y$ and $\tilde u(\tau):=u(x,-\tau)$, so \[ \tilde u_\tau=-g_y(x,-\tau)+a\sin\tilde u . \] Now the stable equilibria (in the $\tau$-direction) are the odd multiples: near $\tilde u=k_-\pi+w$, $\tilde u_\tau=-g_y-a\sin w$ with equilibrium $w=-g_y(x,-\tau)/a$; the linearized rate is $+a>0$ towards it. The initial value $\tilde u(0)=u(x,0)$ lies in $((k_--1)\pi+d_0/2,\ (k_-+1)\pi-d_0/2)$ by Step 0 (with $k_-=k_+\pm1$), and the same basin argument (Step 1, with $I_a$ centred at $k_-\pi$ and the inward checks using $\cos(k_-\pi)=-1$, i.e.\ the sign of $\sin\tilde u$ reversed) gives $|\tilde u-k_-\pi|\le\rho_a$ for $\tau\ge\tau_1$, and the same two-round tracking with $e_a^-:=-g_y(x,-\tau)/a$ gives $|\tilde u-k_-\pi-e_a^-|\le C_8/a^2$ for $\tau\in[\varepsilon,L]$. Returning to $y=-\tau$: \[ u(x,y)=k_-\pi-\frac{g_y(x,y)}a+O(a^{-2}), \qquad q_a=2u-g=-g+2\pi k_--\frac{2}{a}g_y+O(a^{-2}), \] uniformly on $y\in[-L,-\varepsilon]$. This is the second estimate ($k_-$ odd, so $(-1)^{k_-}=-1$). \end{proof} \begin{remark}[structure of the layer] \label{rem:layer-structure} \emph{(a)} In general $k_-=k_+\pm1$: the two half-windows are governed by different constants, and $q_a$ jumps by $2\pi$ across the layer of width $O(\ln a/a)$ around $y=0$. This is consistent with Proposition~\ref{prop:no-C0} (there $u_0=\pi/2$, $k_+=0$, $k_-=1$, and the half-window limits are $0$ and $2\pi$); in particular a \emph{single} window-wide constant $k$ (as one might naively guess from $k_+=\operatorname{round}\bigl((c+g(0,0))/(2\pi)\bigr)$ at the origin) is correct only on the half-window $\{y>0\}$, and only when that round happens to be even. \emph{(b)} The hypothesis $d_0>0$ excludes the ``stall'' values $u_0\in\pi\mathbb Z$---and this is all that is needed: under $d_0>0$ the nearest even and the nearest odd integers $k_\pm$ are automatically unique (they are ambiguous only when $u_0/\pi$ is an integer, which $d_0>0$ excludes). By Step 0 of the proof the hypothesis can be checked directly on the data: $d_0>0$ is equivalent to $(c+g(0,0))/(2\pi)\notin\mathbb Z$. \emph{(c)} Inside the layer no uniform $O(1/a^2)$ expansion of $q_a$ exists; the contribution of the layer to the immersion is $O(1/a)$ (Theorem~\ref{thm:main}), and Proposition~\ref{prop:infinitesimal} shows this contribution is genuinely of first order---the layer, not the exterior, carries the sharp constant. \emph{(d)} For $n>1$ the exterior expansion would have to be propagated through subsequent B\"acklund steps, and the layer of each step interacts with the next one; we do not pursue this here (the resulting partial limits are nevertheless controlled by Theorem~\ref{thm:S-partial} below). \end{remark} \subsection{Other limits and cases} \label{sec:smooth-other} \subsubsection{The normalization constants $c_j$} \begin{lemma}[$4\pi$-periodicity in the constants] \label{lem:c-period} Fix a seed $g\in\mathcal S$ and $a>0$. If $q$ solves \eqref{eq:BT} with seed $g$, then $q+4\pi$ solves \eqref{eq:BT} with the same seed $g$ and the same $a$. Consequently $q_{c+4\pi}=q_c+4\pi$ (Lemma~\ref{lem:bt}), the whole chain satisfies $\omega_n(c_j+4\pi)=\omega_n(c_j)+4\pi$ in the shifted component, and the immersion does not change: $r(\omega+4\pi k)=r(\omega)$ for every $k\in\mathbb Z$, because on the axis the map sees only $\omega_x$ and in the columns only $\sin\omega,\cos\omega$ (\S\ref{sec:immersion}). \end{lemma} \begin{proof} Both equations of \eqref{eq:BT} are invariant under $q\mapsto q+4\pi$: $\sin\bigl(\frac{(q+4\pi)\pm g}{2}\bigr) =\sin\bigl(\frac{q\pm g}{2}+2\pi\bigr) =\sin\frac{q\pm g}{2}$, while the left-hand sides are unchanged. Uniqueness (Lemma~\ref{lem:bt}) identifies the shifted solution with the BT image of the shifted constant; induction along the chain gives the statement for $\omega_n$, and the last claim follows from the definition of $r$. \end{proof} The period is $4\pi$, \emph{not} $2\pi$: \begin{example}[different branches for $c$ and $c+2\pi$] \label{ex:branches} Let $n=1$, $g\equiv0$. Then \[ q_\pi=4\arctan\bigl(e^{-x/a-ay}\bigr), \qquad q_{3\pi}=4\pi-4\arctan\bigl(e^{-x/a-ay}\bigr) \] are the solutions of \eqref{eq:BT} with $q(0,0)=\pi$ and $q(0,0)=3\pi$ respectively (direct verification as in Proposition~\ref{prop:no-C0}; note $q_{3\pi}\ne q_\pi+2\pi$, so $c\mapsto q$ is not $2\pi$-periodic). The immersions differ: on the axis $\partial_xe_1(0,0)$ is the first row of $P(\omega_x(0,0))$, namely $(0,-\omega_x(0,0),0)$, and \[ \omega_x(0,0)=\mp\frac2a \quad\text{for }q_{\pi},\,q_{3\pi}, \] so $\partial_xe_1(0,0)=(0,\pm2/a,0)$; equal immersions would have equal frames along the axis ($e_1=\partial_x r(\cdot,0)$), a contradiction for $a>0$. \end{example} \begin{remark}[the limit $c_j\to\infty$] \label{rem:c-to-infinity} By Lemma~\ref{lem:c-period} the maps $(a_1,\dots,a_n,c_1,\dots,c_n,f)\mapsto r(\omega_n)$ factor through $(\mathbb R/4\pi\mathbb Z)^n$. Hence: (i) if $c_j\to\infty$ and each $c_j$ converges in $\mathbb R/4\pi\mathbb Z$, then $r(\omega_n)$ converges (Theorem~\ref{thm:smooth-finite}); (ii) convergence may fail without such a limit: by Example~\ref{ex:branches}, along a sequence with $c$ alternating between $\pi$ and $3\pi$ the immersions have two distinct limit points; (iii) the estimate of Theorem~\ref{thm:main} is uniform in the $c_j$ regardless---the non-convergence reflects the normalization, not the geometry. \end{remark} \subsubsection{The limit $a_j\to0$}\label{sec:S-zero} Let $(\sigma u)(x,y):=u(y,x)$. Then $\sigma\mathcal S=\mathcal S$ and $\sigma$ preserves $W$: if $u$ solves $u_{xy}=\sin u$, then $\partial_{xy}(\sigma u)=\partial_{yx}u\circ\sigma=\sin u\circ\sigma=\sin(\sigma u)$. \begin{lemma}[exact symmetries of the B\"acklund system]\label{lem:S-sym} For $g\in\mathcal S$, $a>0$, $c\in\mathbb R$: \[ \BT_a(g;c)=\sigma\,\BT_{1/a}(-\sigma g;c),\qquad\BT_a(-g;-c)=-\BT_a(g;c). \] \end{lemma} \begin{proof} Let $q=\BT_a(g;c)$, $Q=\sigma q$, $G=-\sigma g$. Then $(Q+G)_x=\sigma\bigl((q-g)_y\bigr)=-2a\sin\frac{\sigma q+\sigma g}2=-\frac2A\sin\frac{Q-G}2$ with $A=1/a$, and $(Q-G)_y=\sigma\bigl((q+g)_x\bigr)=-\frac2a\sin\frac{\sigma q-\sigma g}2=-2A\sin\frac{Q+G}2$; $Q(0,0)=q(0,0)=c$. By uniqueness (Lemma~\ref{lem:bt}), $Q=\BT_{A}(G;c)$, which is the first identity. The second: $(q,g)\mapsto(-q,-g)$ leaves \eqref{eq:BT} unchanged and maps $c\mapsto-c$. \end{proof} \begin{theorem}[the limit $a_j\to0$]\label{thm:S-zero} Define $r^\sigma(\omega)(x,y):=r(\sigma\omega)(y,x)$ \textup{(immersion of the same Chebyshev net, normalized at the origin by the frame $I$ adapted to the $y$-lines: $r^\sigma_y(0,0)=\varepsilon_1$)} and $T_*^\sigma(f):=\sup_{|y|\le L}\int_{-L}^L|f_x(x,y)|\,dx$. Then for all $a_j>0$, $c_j\in\mathbb R$ \[ \|r^\sigma(\omega_n)-r^\sigma(f)\|_{L^\infty(W)} \le\bigl[3L^2+(1+L)\pi(2T_*^\sigma(f)+\pi)3^{n-1}\bigr]\sum_{j=1}^na_j . \] For the Amsler seed $T_*^\sigma\le2L^2$. Thus, in contrast to $a_j\to\infty$ (the reflection $\omega\mapsto-\omega$ at each step), $a_j\to0$ acts as the identity: no reflection occurs. \end{theorem} \begin{proof} Put $\zeta_j:=\sigma\omega_j$. By Lemma~\ref{lem:S-sym} (first identity applied to the pair $(\omega_{j-1},c_j)$, then the second identity), $\zeta_j=\BT_{A_j}(-\zeta_{j-1};c_j)=-\BT_{A_j}(\zeta_{j-1};-c_j)$ with $A_j=1/a_j$. Let $\eta_j:=(-1)^j\zeta_j$. Using that identity repeatedly, $\eta_j=\BT_{A_j}(\eta_{j-1};(-1)^jc_j)$ with $\eta_0=\sigma f\in\mathcal S$ (induction: if $\zeta_{j-1}=(-1)^{j-1}\eta_{j-1}$, then $\zeta_j=-\BT_{A_j}((-1)^{j-1}\eta_{j-1};-c_j)=(-1)^j\BT_{A_j}(\eta_{j-1};(-1)^jc_j)$). So $\eta$ is a standard chain with parameters $A_j$ and seed $\sigma f$. Theorem~\ref{thm:main} gives $\|r(\eta_n)-r((-1)^n\sigma f)\|\le K_n(\sigma f)\sum1/A_j=K_n(\sigma f)\sum a_j$, where $T_*(\sigma f)=T_*^\sigma(f)$ (relabeling $x\leftrightarrow y$). Now $\sigma\omega_n=(-1)^n\eta_n$ and $\sigma f=(-1)^n\bigl((-1)^n\sigma f\bigr)$; by Lemma~\ref{lem:S-orient}, $r(\sigma\omega_n)-r(\sigma f)=(r(\eta_n)-r((-1)^n\sigma f))R^{\,n}$ and $R$ is orthogonal. Finally $r^\sigma(\omega)=r(\sigma\omega)\circ\sigma$ is a reparametrization. For $f=Y(xy)$, $|f_x|=|y|\,|Y'(xy)|\le L$. \end{proof} \begin{remark}[open]\label{rem:S-mixed} Combined regimes (some $a_j\to0$, others $\to\infty$) follow from Lemma~\ref{lem:S-sym} only for a chain in which \emph{all} the parameters tend to $0$ or all to $\infty$; the partial limits for $a_j\to0$ along an arbitrary index set are Corollary~\ref{cor:S-zero-partial} below. A genuinely mixed theorem (some $a_j\to0$ and some $a_k\to\infty$ simultaneously) requires a comparison lemma for the two normalizations simultaneously and is not proved here. \end{remark} \subsubsection{Unbounded seed data} \begin{proposition}[growth of $T_*$ on a sequence of seeds] \label{prop:Tstar-growth} Let $\beta_\lambda(y)=\lambda\phi(y)$ with $\phi\in C^\infty([-L,L])$, $\phi(0)=0$, $\phi'\not\equiv0$, and $\alpha_\lambda\equiv0$. Then $f_\lambda=\Psi(\alpha_\lambda,\beta_\lambda)$ satisfies \[ T_*(f_\lambda)\;\ge\;\lambda\int_{-L}^{L}|\phi'(y)|\,dy . \] Hence $T_*(f_\lambda)\to\infty$ linearly in $\lambda$, and the constant $K_n(f_\lambda)$ of Theorem~\ref{thm:main} grows at least linearly. \end{proposition} \begin{proof} On the characteristic line $x=0$, $f_\lambda(0,y)=\beta_\lambda(y)=\lambda\phi(y)$, so $\partial_yf_\lambda(0,y)=\lambda\phi'(y)$ and \[ T_*(f_\lambda)=\sup_{|x|\le L}\int_{-L}^L|\partial_yf_\lambda(x,y)|\,dy \ge\int_{-L}^L\lambda|\phi'(y)|\,dy . \qedhere \] \end{proof} \begin{remark}[no compactification of the seed data] \label{rem:seed-data-infinity} Smoothness in Theorem~\ref{thm:smooth-finite} is local in the data. If $(\alpha_\lambda,\beta_\lambda)$ leaves every bounded subset of $\mathcal D$, each $f_\lambda=\Psi(\alpha_\lambda,\beta_\lambda)$ is still a perfectly regular seed, and $\lambda\mapsto r(\omega_n(a,c,f_\lambda))$ is $C^\infty$ on compact $\lambda$-intervals (Theorem~\ref{thm:smooth-finite}); but by Proposition~\ref{prop:Tstar-growth} the constant $K_n(f_\lambda)$ may grow without bound, so the rate in Theorem~\ref{thm:smooth-infinity} is guaranteed uniformly in $\lambda$ only while $T_*(f_\lambda)$ stays bounded. Whether the actual error grows with $\lambda$ is not investigated (Proposition~\ref{prop:Tstar-growth} concerns the constant of the \emph{upper bound} only). No continuous extension of the flow to a ``boundary of the data space'' is claimed. \end{remark} \begin{remark}[periodicity of the seed itself] \label{rem:seed-periodic} For a \emph{single} solution, $r(f+2\pi k)=r(f)$: the axis sees only $f_x$ and the columns only $\sin f,\cos f$ (\S\ref{sec:immersion}). This does not descend to the chain: the B\"acklund system \eqref{eq:BT} depends on the seed itself through $(q\pm g)/2$, and $g\mapsto g+2\pi$ flips the sign of $\sin\bigl((q-g)/2\bigr)$ (it corresponds to $a\mapsto-a$), so $\mathrm{BT}_a(g+2\pi)$ is not $\mathrm{BT}_a(g)$ shifted. The correct period of the whole construction is $4\pi$ (Lemma~\ref{lem:c-period}, Example~\ref{ex:branches}). \end{remark} \subsection{Partial limits}\label{sec:S-partial} Let $J\subset\{1,\dots,n\}$ be the set of indices whose parameters $a_j$ will tend to infinity, $K=\{1,\dots,n\}\setminus J$. Put $c_0:=f(0,0)$; then $w_j(0,0)=(c_j-c_{j-1})/2$ is the value at the origin of $w=(q-g)/2$ in the $j$-th step. For $g\in\mathcal S$ and $w_0\in\mathbb R$ denote by $\BT_a^{\flat}(g;w_0)$ the B\"acklund image $q$ with $(q-g)(0,0)/2=w_0$, i.e.\ $\BT_a(g;\,g(0,0)+2w_0)$. \begin{definition}[reduced chain]\label{def:S-reduced} $\hat g_0:=f$; $\hat g_j:=-\hat g_{j-1}$ for $j\in J$; $\hat g_j:=\BT^{\flat}_{a_j}\bigl(\hat g_{j-1};\,\tfrac12(c_j-c_{j-1})\bigr)$ for $j\in K$. \end{definition} For $J=\{1,\dots,n\}$, $\hat g_n=(-1)^nf=\mathrm{ref}_n$. For $J\ne\{1,\dots,n\}$ the reduced chain depends on $a_k$, $k\in K$, and on $c$. Note that $\hat g_j$ need not satisfy $\hat g_j(0,0)=c_j$ across the steps $j\in J$; the definition fixes the normalization $w_0$ of each B\"acklund step, which is all the comparison below uses. \begin{lemma}[bound on the variation in $y$]\label{lem:S-T} For all $a_1,\dots,a_k>0$ and all $c_j$: $\sup_{|x|\le L}\int_{-L}^L|\partial_y\omega_k|\,dy\le\mathcal T_k:=T_*+2(3^k-1)(T_*+\pi/2)$. \end{lemma} \begin{proof} In the notation of \S\ref{sec:column}, $\omega_k=h_k=\varphi+2z_k$, so $\int|h_k'|\,dy\le T+2V_k$, and $V_k\le(3^k-1)(T+\pi/2)$ by Lemma~\ref{lem:9}(II), which holds for all $a_j>0$; take the supremum in $x$ and use $T\le T_*$. \end{proof} \begin{lemma}[a large step]\label{lem:S-large} Let $g\in\mathcal S$, $q=\BT_a(g;c)$. Then $\delta(q,-g)\le2/a$ and $\Lambda(q,-g)\le\pi(\pi+T(g))/a$, where $T(g):=\sup_{|x|\le L}\int_{-L}^L|g_y|\,dy$. \end{lemma} \begin{proof} $\partial_x(q+g)=-\frac2a\sin\frac{q-g}2$ throughout $W$ (the first equation of \eqref{eq:BT}), which gives the first claim. For the second apply the case $n=1$ of Lemmas~\ref{lem:8}--\ref{lem:10} with seed $g$: $d(q+g)=2d_\pi\bigl((q+g)/2\bigr)\le\pi\bigl|\sin\frac{q+g}2\bigr|$ by \eqref{eq:E4} and \eqref{eq:E2}, and $aI_1\le\pi+T(g)$ by Lemma~\ref{lem:9}(I). \end{proof} \begin{lemma}[a finite step is stable]\label{lem:S-small} Let $g,\hat g\in\mathcal S$, $a>0$, $q=\BT_a^{\flat}(g;w_0)$, $\hat q=\BT_a^{\flat}(\hat g;w_0)$ \textup{(the same $w_0$)}. Then, with $\delta=\delta(g,\hat g)$ and $\Lambda=\Lambda(g,\hat g)$, \[ \delta(q,\hat q)\le\kappa\,\delta,\qquad \Lambda(q,\hat q)\le\mu\,\Lambda+\nu\,\delta, \] \[ \kappa=1+\tfrac{2L}ae^{L/a},\qquad\mu=1+4aLe^{aL},\qquad\nu=4L^2e^{aL+L/a}. \] \end{lemma} \begin{proof} Let $w=(q-g)/2$, $\hat w=(\hat q-\hat g)/2$. On the axis, $w_x=u_x-g_x=-\frac1a\sin w-g_x$ (with $u=w+g$), so $|\partial_x(w-\hat w)|\le\delta+\frac1a|w-\hat w|$ and $w(0,0)=\hat w(0,0)=w_0$; Gronwall gives $|w-\hat w\,|(x,0)\le a\delta(e^{|x|/a}-1)\le\delta|x|e^{L/a}$. Since $q_x=-g_x-\frac2a\sin w$, $|\partial_x(q-\hat q)(x,0)|\le\delta+\frac2a|w-\hat w|\le\delta\kappa$. In the columns $w_y=u_y-g_y=-a\sin(w+g)$, and $|\sin\alpha-\sin\beta|\le|e^{i\alpha}-e^{i\beta}|\le d(\alpha-\beta)\le d(\alpha_1-\beta_1)+d(\alpha_2-\beta_2)$ for $\alpha=\alpha_1+\alpha_2$, $\beta=\beta_1+\beta_2$; hence $|(w-\hat w)_y|\le a|w-\hat w|+a\,d(g-\hat g)$, and Gronwall gives $|w-\hat w\,|(x,y)\le e^{a|y|}\bigl(|w-\hat w\,|(x,0)+a\Lambda\bigr)$. Finally $q-\hat q=2(w-\hat w)+(g-\hat g)$, so $d(q-\hat q)\le2|w-\hat w|+d(g-\hat g)$, and, integrating over an interval of length $2L$, $\Lambda(q,\hat q)\le4L\sup|w-\hat w|+\Lambda\le4Le^{aL}(L\delta e^{L/a}+a\Lambda)+\Lambda$. \end{proof} \begin{theorem}[partial limits]\label{thm:S-partial} For all $a_j>0$, $c_j\in\mathbb R$ and all $J\subset\{1,\dots,n\}$, \[ \|r(\omega_n)-r(\hat g_n)\|_{L^\infty(W)} \le\Bigl(\tfrac32L^2+1+L\Bigr)\sum_{j\in J}\frac{\pi(\pi+\mathcal T_{j-1})}{a_j}\prod_{\substack{k\in K\\ k>j}}\rho_k, \qquad\rho_k:=\kappa_k+\mu_k+\nu_k, \] where $\kappa_k,\mu_k,\nu_k$ are the constants of Lemma~\ref{lem:S-small} for $a=a_k$, and $\hat g_n$ is the reduced chain of Definition~\ref{def:S-reduced}. In particular, if $a_j\to\infty$ for $j\in J$ while $a_k$, $k\in K$, stay in a compact subset of $(0,\infty)$, then $r(\omega_n)\to r(\hat g_n)$ uniformly with rate $O(\sum_{j\in J}1/a_j)$, uniformly in the constants $c_j$. The steps $j\in J$ may occur anywhere in the chain, including non-trailing positions. \end{theorem} \begin{proof} Let $E_j:=\max\{\delta(\omega_j,\hat g_j),\,\Lambda(\omega_j,\hat g_j)\}$, $E_0=0$. If $j\in J$: $\omega_j-\hat g_j=(\omega_j+\omega_{j-1})-(\omega_{j-1}-\hat g_{j-1})$, so by Lemma~\ref{lem:S-large} (with seed $\omega_{j-1}$ and $T(\omega_{j-1})\le\mathcal T_{j-1}$ by Lemma~\ref{lem:S-T}), the subadditivity of $d$, and the same splitting for $\delta$, \[ \delta(\omega_j,\hat g_j)\le\delta(\omega_{j-1},\hat g_{j-1})+\frac2{a_j},\qquad \Lambda(\omega_j,\hat g_j)\le\Lambda(\omega_{j-1},\hat g_{j-1})+\frac{\pi(\pi+\mathcal T_{j-1})}{a_j}; \] since $\pi(\pi+\mathcal T_{j-1})\ge2$, $E_j\le E_{j-1}+\sigma_j/a_j$ with $\sigma_j=\pi(\pi+\mathcal T_{j-1})$. If $j\in K$: both $\omega_j$ and $\hat g_j$ are $\BT^{\flat}_{a_j}$-images with the same $w_0=\frac12(c_j-c_{j-1})$ (for $\omega_j$ this holds because $\omega_{j-1}(0,0)=c_{j-1}$; for $\hat g_j$ it is the definition), so Lemma~\ref{lem:S-small} gives $E_j\le\rho_jE_{j-1}$. By induction, $E_n\le\sum_{j\in J}\frac{\sigma_j}{a_j}\prod_{k\in K,\,k>j}\rho_k$. Lemma~\ref{lem:S-comp} gives $\|r(\omega_n)-r(\hat g_n)\|\le\frac32L^2\delta_n+(1+L)\Lambda_n\le(\frac32L^2+1+L)E_n$. \end{proof} \begin{corollary}[uniformity in the other parameters; the closed orthant]\label{cor:S-orthant} \textup{(i)} For $J=\{n\}$: $\|r(\omega_n)-r(-\omega_{n-1})\|_\infty\le(\frac32L^2+1+L)\,\pi(\pi+\mathcal T_{n-1})/a_n$, uniformly in $a_1,\dots,a_{n-1}>0$ and in all $c_j$; this strengthens Proposition~\ref{prop:trailing} below. \textup{(ii)} Put $\varepsilon_j:=1/a_j$. The map $(\varepsilon,c)\mapsto r(\omega_n)\in C^0(W)$, defined for $\varepsilon\in(0,\infty)^n$, extends continuously to $[0,\infty)^n\times\mathbb R^n$: for every $(\varepsilon^0,c^0)\in[0,\infty)^n\times\mathbb R^n$ the limit of $r(\omega_n(\varepsilon,c))$ as $(\varepsilon,c)\to(\varepsilon^0,c^0)$ through the interior exists and equals $r(\hat g_n)$, computed with $J=J(\varepsilon^0):=\{j:\varepsilon^0_j=0\}$ and the parameters $(1/\varepsilon^0_k)_{k\in K}$, $c^0$. The convergence is Lipschitz in $\varepsilon_J$, uniformly in $c$, and the boundary value is a smooth function of $(\varepsilon_K,c)$. \end{corollary} \begin{proof} (i) is Theorem~\ref{thm:S-partial} with $J=\{n\}$: here $\hat g_j=\omega_j$ for $jj}}\rho_k , \] where $\mathcal T^{\eta}_{j-1}:=T_*^\sigma(f)+2(3^{j-1}-1)\bigl(T_*^\sigma(f)+\pi/2\bigr)$ and $\rho_k$ are the constants of Lemma~\ref{lem:S-small} for the parameters $A_k=1/a_k$. In particular, $r^\sigma(\omega_n)\to r((-1)^n\hat g^{\eta}_n)\circ\sigma$ uniformly with rate $O(\sum_{j\in J}a_j)$. For $J=\{1,\dots,n\}$ this recovers Theorem~\ref{thm:S-zero} (there with the constant of Theorem~\ref{thm:main}, which differs from the one above). \end{corollary} \begin{proof} Apply Theorem~\ref{thm:S-partial} to the $\eta$-chain: its large parameters are $A_j\to\infty$ exactly for $j\in J$, its $K$-step parameters $A_k=1/a_k$ stay in a compact subset of $(0,\infty)$, and its constants $\tilde c_j=(-1)^jc_j$ range over all of $\mathbb R^n$ as $c$ does, so the uniformity in $\tilde c$ is the uniformity in $c$. This gives \[ \|r(\eta_n)-r(\hat g^{\eta}_n)\| \le\Bigl(\tfrac32L^2+1+L\Bigr)\sum_{j\in J}\frac{\pi(\pi+\mathcal T^{\eta}_{j-1})}{A_j}\prod_{\substack{k\in K\\ k>j}}\rho_k =\Bigl(\tfrac32L^2+1+L\Bigr)\sum_{j\in J}a_j\,\pi\bigl(\pi+\mathcal T^{\eta}_{j-1}\bigr)\prod_{\substack{k\in K\\ k>j}}\rho_k , \] where $\mathcal T^{\eta}_{j-1}$ is the quantity of Lemma~\ref{lem:S-T} for the $\eta$-chain, whose seed $\sigma f$ has $T_*(\sigma f)=T_*^\sigma(f)$ (relabeling $x\leftrightarrow y$). On the other hand, by the construction of the proof of Theorem~\ref{thm:S-zero}, $\sigma\omega_n=(-1)^n\eta_n$ and $r^\sigma(\omega_n)=r(\sigma\omega_n)\circ\sigma$; hence, by Lemma~\ref{lem:S-orient} ($r((-1)^nu)=r(u)R^{\,n}$ with orthogonal $R$) and the compatibility of $\sigma$ with the norms, \[ \|r^\sigma(\omega_n)-r\bigl((-1)^n\hat g^{\eta}_n\bigr)\circ\sigma\|_{L^\infty(W)} =\|r\bigl((-1)^n\eta_n\bigr)-r\bigl((-1)^n\hat g^{\eta}_n\bigr)\|_{L^\infty(W)} =\|r(\eta_n)-r(\hat g^{\eta}_n)\|_{L^\infty(W)} , \] which is the claim. For $J=\{1,\dots,n\}$ the reduced chain is $\hat g^{\eta}_n=(-1)^n\sigma f$, so the left-hand side is $\|r^\sigma(\omega_n)-r^\sigma(f)\|$ and the statement reduces to Theorem~\ref{thm:S-zero} (whose constant, coming from Theorem~\ref{thm:main}, differs from the one above). \end{proof} \begin{corollary}[stratified Lipschitz at the faces of the orthant]\label{cor:S-lipschitz} Fix $(\varepsilon^0,c^0)\in[0,\infty)^n\times\mathbb R^n$, put $J:=\{j:\ \varepsilon^0_j=0\}$ and $K$ its complement, and fix $\delta>0$ with $\varepsilon^0_k\ge2\delta$ for $k\in K$. Then for all $\varepsilon,\bar\varepsilon$ in the stratum \[ S_\delta:=\bigl\{\varepsilon:\ |\varepsilon_k-\varepsilon^0_k|\le\delta\ (k\in K),\ 0<\varepsilon_j\le\delta\ (j\in J)\bigr\} \] and all $|c-c^0|\le\delta$, $|\bar c-c^0|\le\delta$, \[ \|r(\omega_n(\varepsilon,c))-r(\omega_n(\bar\varepsilon,\bar c))\|_{L^\infty(W)} \le C_*\Bigl(\sum_{k\in K}|\varepsilon_k-\bar\varepsilon_k|+\sum_{j\in J}(\varepsilon_j+\bar\varepsilon_j)+|c-\bar c|\Bigr), \] with $C_*=C_*(n,L,T_*,\delta,\varepsilon^0)$; in particular $C_1=\bigl(\tfrac32L^2+1+L\bigr)\max_{j\in J}\pi(\pi+\mathcal T_{j-1})\prod_{k\in K}\bar\rho_k$, where $\bar\rho_k:=\max\{\kappa+\mu+\nu\ \text{(Lemma~\ref{lem:S-small} with parameter }a):\ a\in[1/(\varepsilon^0_k+\delta),\,1/(\varepsilon^0_k-\delta)]\}$. \end{corollary} \begin{proof} By Theorem~\ref{thm:S-partial} (as in the proof of Corollary~\ref{cor:S-orthant}(ii)), uniformly in $c$, \[ \|r(\omega_n(\varepsilon,c))-r(\hat g_n(\varepsilon_K,c))\| \le\Bigl(\tfrac32L^2+1+L\Bigr)\sum_{j\in J}\varepsilon_j\prod_{\substack{k\in K\\ k>j}}\rho_k(\varepsilon_k) \le C_1\sum_{j\in J}\varepsilon_j, \] since on the stratum $a_k=1/\varepsilon_k$ stays in the compact interval $[1/(\varepsilon^0_k+\delta),1/(\varepsilon^0_k-\delta)]\subset(0,\infty)$, so that $\rho_k\le\bar\rho_k$ (Lemma~\ref{lem:S-small}), and $\rho_k\ge1$. The same holds for $(\bar\varepsilon,\bar c)$. Finally, $(\varepsilon_K,c)\mapsto\hat g_n(\varepsilon_K,c)$ is $C^\infty$ on the compact stratum (Corollary~\ref{cor:S-orthant}(ii) and Theorem~\ref{thm:smooth-finite}), and $g\mapsto r(g)$ is locally Lipschitz from $C^1$ to $C^0$ (Lemma~\ref{lem:imm-lip}), so \[ \|r(\hat g_n(\varepsilon_K,c))-r(\hat g_n(\bar\varepsilon_K,\bar c))\| \le C_2\Bigl(\sum_{k\in K}|\varepsilon_k-\bar\varepsilon_k|+|c-\bar c|\Bigr). \] The triangle inequality gives the claim with $C_*=\max(C_1,C_2)$. \end{proof} \begin{remark}[what is not claimed]\label{rem:S-open} The constant of Theorem~\ref{thm:S-partial} contains $e^{a_kL}$ for the finite steps $k\in K$ \emph{after} a large step; the estimate is therefore useful only for fixed (or slowly growing) finite parameters, and it is not asserted that this exponential dependence is necessary. Corollary~\ref{cor:S-lipschitz} gives Lipschitz control only in the \emph{tangential} variables $(\varepsilon_K,c)$; in the normal directions $\varepsilon_j$, $j\in J$, the bound has the two-sided form $\varepsilon_j+\bar\varepsilon_j$ (through the base point), and a bound of the form $|\varepsilon_j-\bar\varepsilon_j|$ is not claimed. Proposition~\ref{prop:S-nodiff}(d) shows, for the model kink ($n=1$), that $\varepsilon\mapsto r(\omega_1)$ is not differentiable at $\varepsilon=0$ in $C^0$ (although it is Lipschitz at that point, Theorem~\ref{thm:S-sharp}(i)); it does not by itself exclude a Lipschitz bound between interior points. Finally, an asymptotic expansion of $r(\omega_n)$ in powers of $1/a_j$ is in general not available in $C^0$ (for the model kink, Proposition~\ref{prop:S-nodiff}); whether one exists in $L^p$, $p<\infty$, is not investigated here. \end{remark} \subsection{The B\"acklund reflection chamber} \label{sec:chamber} Theorem~\ref{thm:smooth-infinity} describes the \emph{total} limit (all $a_j\to\infty$ simultaneously). In this subsection we show that the \emph{one-step} (trailing) limit is already available from it, that the B\"acklund step commutes with the sign change $g\mapsto-g$ of the seed, and that these two facts combine into a precise version of the ``graph of levels connected through infinity'' picture: the orbit of a fixed seed splits into two alternating, mirror-symmetric, connected unions, glued along the faces $a_n=\infty$. Throughout, we fix $f\in\mathcal S$ and write, for $\varepsilon\in\{+1,-1\}$ and $n\ge0$, \[ M_n^{\varepsilon}:=\Bigl\{r\bigl(\omega_n(a,c,\varepsilon f)\bigr):\ a\in(0,\infty)^n,\ c\in\mathbb R^n\Bigr\} \ \subset\ C^0(W,\mathbb R^3), \qquad M_0^{\varepsilon}=\{r(\varepsilon f)\}, \] for the \emph{levels} of the orbit (no manifold structure is claimed or needed). \begin{lemma}[levels are path-connected] \label{lem:level-conn} Each $M_n^{\varepsilon}$ is path-connected in $C^0(W,\mathbb R^3)$. \end{lemma} \begin{proof} The parameter set $(0,\infty)^n\times(\mathbb R/4\pi\mathbb Z)^n$ is path-connected (Lemma~\ref{lem:c-period} allows the reduction modulo $4\pi$), and $\Phi_n$ is continuous into every $C^k(W,\mathbb R^3)$ by Theorem~\ref{thm:smooth-finite}. The image of a parameter path under $\Phi_n$ is a path in $M_n^{\varepsilon}$ joining the two given points (Corollary~\ref{cor:paths}). \end{proof} \begin{lemma}[sign equivariance of one B\"acklund step] \label{lem:bt-sign} If $q$ solves \eqref{eq:BT} with seed $g\in\mathcal S$, parameter $a$ and normalization $q(0,0)=c$, then $-q$ solves \eqref{eq:BT} with seed $-g$, the same $a$, and normalization $-q(0,0)=-c$; by uniqueness (Lemma~\ref{lem:bt}) it is \emph{the} normalized transform of $-g$. Consequently, along chains, \begin{equation}\label{eq:bt-sign} -\,\omega_n(a,c,f)\;\equiv\;\omega_n(a,-c,-f)\quad\text{on }W, \qquad\text{hence}\qquad M_n^{-\varepsilon}=\bigl\{r(-\omega):\ r(\omega)\in M_n^{\varepsilon}\bigr\}. \end{equation} \end{lemma} \begin{proof} Write $q=2u-g$ with $u$ solving $u_x=-(1/a)\sin(u-g)$, $u_y=g_y-a\sin u$, $u(0,0)=(c+g(0,0))/2$ (Lemma~\ref{lem:bt}), and put $g':=-g$, $c':=-c$, $u':=-u$, $q':=2u'-g'$. Then $u'_x=-u_x=(1/a)\sin(u-g)=-(1/a)\sin(u'-g')$, $u'_y=-u_y=-g_y+a\sin u=g'_y-a\sin u'$, and $u'(0,0)=-u(0,0)=(c'+g'(0,0))/2$, so $q'$ solves \eqref{eq:BT} with seed $g'$ and $q'(0,0)=c'$; uniqueness gives $q'=-q$. Iterating along the chain (each $\omega_j$ is again a seed, Lemma~\ref{lem:bt}) gives \eqref{eq:bt-sign}; the displayed equality of level sets follows by re-parametrizing $c\mapsto-c$. \end{proof} \begin{proposition}[one-step mirror at the trailing end] \label{prop:trailing} Fix $n\ge1$, the previous parameters $(a_1,\dots,a_{n-1})$, $(c_1,\dots,c_{n-1})$, and the seed $f$. Then, as $a_n\to\infty$, uniformly in $c_n$, \[ \bigl\|r(\omega_n)-r(-\omega_{n-1})\bigr\|_{L^\infty(W)} \ \le\ \frac{K_1(\omega_{n-1})}{a_n}\ \longrightarrow\ 0, \qquad K_1(g)=3L^2+(1+L)\pi\bigl(2T_*(g)+\pi\bigr)<\infty . \] \end{proposition} \begin{proof} $\omega_{n-1}\in\mathcal S$ by Lemma~\ref{lem:bt}, and $(\omega_{n-1},\omega_n)$ \emph{is} the one-step chain from the seed $g:=\omega_{n-1}$ with parameter $a_n$ and constant $c_n$. Apply Theorem~\ref{thm:smooth-infinity} with $n=1$ and seed $g$: it gives $\|r(\omega_1^{(g)})-r(-g)\|_{L^\infty}\le K_1(g)/a_1$, unconditionally in $c_1$. \end{proof} \begin{remark}[the branch ambiguity is immaterial for immersions] \label{rem:branch-immaterial} Proposition~\ref{prop:trailing} lands at $r(-\omega_{n-1})$ \emph{without} any $2\pi$-ambiguity, because Theorem~\ref{thm:main} is unconditional in $c_n$. The branch shifts $2\pi k$ that a naive mirroring prescription would suggest are in any case invisible at the level of immersions, since $r(\omega+2\pi k)=r(\omega)$ for a single solution (Remark~\ref{rem:seed-periodic}). \end{remark} \begin{theorem}[the two mirror unions are connected] \label{thm:chamber} Fix $f\in\mathcal S$ and define the two alternating unions \[ \mathcal C_+:=\bigcup_{n\ge0}M_n^{(-1)^n} =M_0^{+}\cup M_1^{-}\cup M_2^{+}\cup M_3^{-}\cup\cdots, \qquad \mathcal C_-:=\bigcup_{n\ge0}M_n^{(-1)^{n+1}} . \] Then: \emph{(i)} for every $n\ge1$ and $\varepsilon\in\{\pm1\}$, $M_{n-1}^{-\varepsilon}\subset\overline{M_n^{\varepsilon}}$ (closure in $C^0(W,\mathbb R^3)$); \emph{(ii)} $\mathcal C_+$ and $\mathcal C_-$ are connected; \emph{(iii)} the mirror symmetry holds as an equality of sets, $\{r(-\omega):\ r(\omega)\in\mathcal C_+\}=\mathcal C_-$ and vice versa; \emph{(iv)} the total-limit points lie in the correct union: $r\bigl((-1)^n\varepsilon f\bigr)\in\overline{M_n^{\varepsilon}}$ (Theorem~\ref{thm:smooth-infinity}), consistently with \emph{(i)}. \end{theorem} \begin{proof} \emph{(i)} Let $p=r(\omega_{n-1}(a,c,-\varepsilon f))\in M_{n-1}^{-\varepsilon}$. By \eqref{eq:bt-sign}, $p=r\bigl(-\omega_{n-1}(a,-c,\varepsilon f)\bigr)$, and Proposition~\ref{prop:trailing} applied to the prefix chain $\omega_{n-1}(a,-c,\varepsilon f)$ gives, as $a_n\to\infty$ (any $c_n$), \[ r\bigl(\omega_n((a,a_n),(-c,c_n),\varepsilon f)\bigr)\ \longrightarrow\ p, \] with all terms in $M_n^{\varepsilon}$; the prefix parameters range over all of $(0,\infty)^{n-1}\times\mathbb R^{n-1}$, so \emph{every} point of $M_{n-1}^{-\varepsilon}$ is obtained. \emph{(ii)} Let $D_N:=\bigcup_{0\le n\le N}M_n^{(-1)^n}$. Each $D_N$ is connected: $D_0=M_0^{+}$ is a point; and if $D_{N-1}$ is connected, then $M_N^{(-1)^N}\cup M_{N-1}^{(-1)^{N-1}}$ is connected by \emph{(i)} (a connected set together with a subset of its closure is connected), and it meets $D_{N-1}$ in $M_{N-1}^{(-1)^{N-1}}\neq\varnothing$; hence $D_N$ is connected. $\mathcal C_+=\bigcup_N D_N$ is an increasing union of connected sets, hence connected; $\mathcal C_-$ follows by symmetry (replacing $f$ with $-f$ exchanges the two). \emph{(iii)} By \eqref{eq:bt-sign}, $\{r(-\omega):r(\omega)\in M_n^{\varepsilon}\}=M_n^{-\varepsilon}$, and negation reverses the parity index: $M_n^{-( (-1)^n)}=M_n^{(-1)^{n+1}}$. \emph{(iv)} is Theorem~\ref{thm:smooth-infinity}. \end{proof} \begin{corollary}[iterated trailing limits] \label{cor:iterated} Fix $f$ and $n$. Sending $a_n\to\infty$, then $a_{n-1}\to\infty$, \dots, then $a_1\to\infty$ (each limit taken in the reduced chain, whose parameters are then finite), one obtains, after $d$ steps, \[ r\bigl(\omega_n(a,c,f)\bigr)\ \longrightarrow\ r\Bigl(\omega_{n-d}\bigl(a_1,\dots,a_{n-d},\,(-1)^d c,\,(-1)^d f\bigr)\Bigr), \] and, for $d=n$, the value $r\bigl((-1)^n f\bigr)$ of Theorem~\ref{thm:smooth-infinity}. In particular the iterated trailing limit and the total limit agree. \end{corollary} \begin{proof} Induction over $d$ with Proposition~\ref{prop:trailing} and \eqref{eq:bt-sign}: after one step, $r(\omega_n)\to r(-\omega_{n-1}) =r\bigl(\omega_{n-1}(a_1,\dots,a_{n-1},-c,-f)\bigr)$, which is a finite-parameter chain to which the next step applies verbatim; the sign and the constants flip once per step. \end{proof} \begin{remark}[what remains open: exactly two components?] \label{rem:chamber-open} Theorem~\ref{thm:chamber} gives two \emph{connected} unions; it does not claim that $\mathcal C_+\cap\mathcal C_-=\varnothing$, nor that these are the connected components of $\mathcal C_+\cup\mathcal C_-$. Both questions amount to possible \emph{cross-parity coincidences} $r(\omega)=r(\omega')$ between immersions of chains with opposite index $\varepsilon(-1)^n$; nothing in this article excludes or establishes them. By Lemma~\ref{lem:S-orient} and Theorem~\ref{thm:chamber}(iii), $\mathcal C_-=\{rR:\ r\in\mathcal C_+\}$ with $R=\operatorname{diag}(1,-1,-1)$, so $\mathcal C_+\cap\mathcal C_-=\varnothing$ is equivalent to $\mathcal C_+\cap\mathcal C_+R=\varnothing$. Note the robustness of the classification: any gluing produced by the mechanism of Theorem~\ref{thm:S-partial} (replacing each of $m$ infinite steps, wherever they sit in the chain, by the reflection $\omega\mapsto-\omega+2\pi k$) sends $M_n^{\varepsilon}$ into $M_{n-m}^{\varepsilon(-1)^m}$ up to $2\pi$ shifts, and the index $\varepsilon(-1)^{n}$ is preserved because $(-1)^m\cdot(-1)^{n-m}=(-1)^n$; such gluings therefore organize $\mathcal C_\pm$ further but cannot merge them. Finally, if the seed itself is allowed to vary, the picture becomes trivially connected: $\mathcal D$ is affine, so $(\alpha,\beta)$ and $(-\alpha,-\beta)$ are joined by a straight segment, and Corollary~\ref{cor:paths} yields a continuous path of immersions from $r(f)$ to $r(-f)$---a path of a different nature than B\"acklund iteration. \end{remark} \begin{remark}[reflection chamber: informal picture] \label{rem:chamber-weyl} The structure exhibited above is a chamber picture for the group $\mathbb Z/2$: the open chamber is $(0,\infty)^n$ (the levels), the walls are $a_j=\infty$, the wall reflection is $g\mapsto-g$ with $R^2=\mathrm{id}$ (Lemma~\ref{lem:bt-sign}), and crossing the \emph{trailing} wall $a_n=\infty$ freezes the immersion on the mirror of the previous level, while the solution itself in general has no $C^0$ limit (Proposition~\ref{prop:no-C0}; degenerate ``stall'' data, for which the layer disappears and $\omega_n$ does converge, are not excluded)---the ``stop of the seed change'' is the generic behaviour. Two honest caveats. First, the walls $a_j=\infty$, $j