\section{Introduction} \label{sec:introduction} \subsection{The result} A \emph{bending} of a closed surface is a continuous family of isometric embeddings of one fixed Riemannian metric into $\R^3$. It is nontrivial when the shape of the surface changes, that is, when the images are not all congruent. Connelly's flexible sphere \cite{Connelly1977} shows that nontrivial bendings exist for embedded polyhedra. Problem~1.1 of Ghomi's list \cite{Ghomi2019}, which is attributed there to Euler and Maxwell, asks whether a closed $C^2$ surface in $\R^3$ can be flexible; for the history of the question see also \cite{Alexandrov1947,Sabitov2023}. We construct a $C^\infty$ example. \begin{maintheorem}\label{thm:main} There are a smooth surface $\Sigma$ diffeomorphic to $S^2$, a smooth Riemannian metric $g_\Sigma$ on it, and a jointly smooth map $\mathcal F:\Sigma\times\R\to\R^3$ such that the maps $F_\tau=\mathcal F(\,\cdot\,,\tau)$ satisfy \[ F_\tau\text{ is an embedding},\qquad F_\tau^*g_{\mathrm E}=g_\Sigma,\qquad F_{\tau+1}=F_\tau \qquad\text{for all }\tau\in\R, \] where $g_{\mathrm E}$ is the Euclidean metric. At least two of the images $F_\tau(\Sigma)$ are not congruent under the Euclidean isometry group. An inner disk and a complementary cap of $\Sigma$ are fixed pointwise throughout the family. The area, the enclosed volume and the total oriented mean curvature of $F_\tau(\Sigma)$ do not depend on $\tau$. \end{maintheorem} Three conventions are implicit in this statement. A \emph{closed} surface is compact and without boundary. \emph{Nontrivial} always refers to noncongruence of the images: relabelling the points of a fixed surface by a family of intrinsic isometries moves material points but does not change the shape, and is not a bending in our sense. Families of this kind are the \emph{slide bendings} studied by Sabitov \cite{Sabitov1995,Sabitov2021,Sabitov2023}; they need not be induced by rigid motions, although the surface does not change as a set. The mean curvature is $(k_1+k_2)/2$ for the outward normal. \subsection{The mechanism} \label{subsec:mechanism} The family in Theorem~\ref{thm:main} is completely explicit once one particular disk is available. Let $D$ be a disk with a rotationally symmetric metric $g$, and write $\varrho_\vartheta$ for its intrinsic rotation by the angle $\vartheta$. Suppose $F:D\to\R^3$ is an isometric embedding. If $R_\vartheta$ denotes the rotation of $\R^3$ about the vertical axis, then \begin{equation}\label{eq:phase-family} F_\tau=R_{-2\pi\tau}\circ F\circ\varrho_{2\pi\tau} \end{equation} is again an isometric embedding of $(D,g)$ for every $\tau$, and its image is the rotated copy $R_{-2\pi\tau}F(D)$ of the original image. Two cases are uninteresting. If $F(D)$ is a surface of revolution, the image does not move at all. If $F(D)$ is not a surface of revolution, the image does move, but in general so does its boundary curve, and nothing can be attached to it. There is a third possibility. Suppose that $F(D)$ is not a surface of revolution, but that near the boundary of $D$ the embedding is the inclusion of a flat horizontal annulus, $F(x)=(x,c)$. There the two rotations in \eqref{eq:phase-family} cancel, so $F_\tau=F$ near the boundary for every $\tau$. All the maps $F_\tau$ can then be glued to one and the same complementary cap. The result is a closed surface in which a wavy annulus turns relative to a stationary shell (Figures~\ref{fig:surface-motion} and~\ref{fig:surface-schematic}). If the shell carries a marker which breaks its rotational symmetry, no rigid motion can account for the relative rotation, and the shape of the closed surface changes. In the terminology above, $F\circ\varrho_{2\pi\tau}$ is a slide bending of the disk, and the rotation $R_{-2\pi\tau}$ in \eqref{eq:phase-family} is chosen so that the collar stays fixed pointwise instead of the image. \begin{figure}[htbp] \centering \begin{tikzpicture}[scale=.91,>=Stealth,font=\small] \foreach \shift/\phase/\word in {0/0/initial,6.6/55/later}{ \begin{scope}[xshift=\shift cm] \fill[gray!15] (0,0) circle (2.05); \fill[blue!6] (0,0) circle (1.65); \draw[gray!65,thick] (0,0) circle (2.05); \draw[dashed] (0,0) circle (1.65); \foreach \radius in {.85,1.12,1.38}{ \draw[blue!70!black,thick,rotate=\phase] plot[domain=0:360,samples=65,variable=\t] ({(\radius+.10*cos(2*\t))*cos(\t)}, {(\radius+.10*cos(2*\t))*sin(\t)});} \filldraw[fill=green!15,draw=green!45!black] (0,0) circle (.52); \fill[red!75!black] (1.30,1.33) circle (2.4pt); \node at (0,-2.36) {\word\ position}; \end{scope}} \draw[->,thick] (2.65,0)--(3.95,0); \node[above] at (3.3,0) {phase $\tau$}; \node[below,align=center] at (3.3,-.1) {same\\metric}; \node[above] at (0,2.15) {fixed inner disk}; \draw[->,green!45!black] (-.5,2.0)--(-.18,.45); \node[above] at (6.6,2.15) {fixed marker on the cap}; \draw[->,red!75!black] (7.4,2.0)--(7.90,1.36); \end{tikzpicture} \caption{The relative rotation, seen from above. The blue curves are schematic height contours of the moving annulus; they are not numerical data for the constructed surface. The dashed circle bounds the planar collar, which stays fixed pointwise, as does the green inner disk. The red marker on the cap also stays fixed while the nonradial part turns, so no ambient rotation accounts for the whole motion.} \label{fig:surface-motion} \end{figure} \begin{figure}[htbp] \centering \includegraphics[width=\linewidth]{figures/surface-schematic.png} \caption{A schematic view of the closed surface in $\R^3$ at two values of $\tau$. \emph{This is an illustration and not a computation of the surface of Theorem~\ref{thm:main}:} the nonradial part is magnified by many orders of magnitude, and the infinitely many small oscillations near the two edges of the moving ring are not drawn. The colours show the sign of the nonradial part of the height, which has angular period $\pi$. Between the two pictures this pattern has turned about the axis; the inner disk, the planar collar outside the dashed circle and the cap with its marker have not moved.} \label{fig:surface-schematic} \end{figure} The motion itself therefore requires no analysis; in particular, no infinitesimal bending has to be integrated in the variable $\tau$. Everything rests on the existence of the disk, and the difficulty lies in the requirement that it be \emph{exactly} planar on a collar of its boundary while being nonradial inside. \subsection{The analytic core} Such a disk is produced as follows. Fill the middle of the disk by a flat piece, so that only an annulus $(r_-,r_+)\times\T$ remains, where $\T=\R/2\pi\Z$. On the annulus take the metric of a radial graph with slope $w$, \begin{equation}\label{eq:intro-metric-form} g=(1+w(r)^2)\,dr^2+r^2\,d\theta^2, \end{equation} and look for an isometric embedding of the form $(P,z)$, where $z$ is a height function and $P$ takes values in the horizontal plane. Such an embedding exists locally exactly when the \emph{remainder metric} $g-dz\otimes dz$ is positive definite and flat, and flatness is a single scalar equation for $z$, the Darboux equation. The disk of Section~\ref{subsec:mechanism} is then obtained from the following statement, which is the main analytic result of the paper. \begin{maintheorem}\label{thm:annulus-intro} There exist radii $00$ on $(r_-,r_+)$ and a smooth function $z$ on $(r_-,r_+)\times\T$ with the following properties, where $g$ is the metric \eqref{eq:intro-metric-form}. \begin{enumerate} \item $|dz|_g<1$, and the metric $g-dz\otimes dz$ is flat. \item $z(r,\theta+\pi)=z(r,\theta)$, and $z$ is not independent of $\theta$. \item $w$ extends smoothly by zero across $r_-$ and $r_+$, and $z$ extends smoothly by a constant across each of them. \end{enumerate} \end{maintheorem} Condition~(iii) says that $w$, and $z$ minus its boundary value, vanish to infinite order at both boundary circles. After a slight enlargement of the annulus this gives open collars on which $w=0$ and $z$ is constant, which is the exact planarity needed for the mechanism. A small error at the boundary would be useless; this is why flatness to infinite order is the right target. The precise form of Theorem~\ref{thm:annulus-intro}, with quantitative information on $w$ and $z$, is Theorem~\ref{thm:new-annular-height}. The passage from Theorem~\ref{thm:annulus-intro} to Theorem~\ref{thm:main} is carried out in Section~\ref{sec:geometric-closure} and uses only classical differential geometry. The rest of the paper proves Theorem~\ref{thm:annulus-intro}. \subsection{Three difficulties, and how they are met} \label{subsec:difficulties} We look for the height in the form $z=h+qu$, where $h'=w$ is the radial height whose graph has the metric $g$, and $q$ is a small parameter. To first order in $q$ the Darboux equation is \begin{equation}\label{eq:intro-linear} \mathscr L_wu=(wu_r)_r+\frac{w'}{r}\,u_{\theta\theta}=0 . \end{equation} A nonzero solution of this equation which is constant to infinite order at both ends is an infinitesimal version of what we want, and the plan is to choose $w$ so that one exists and then to correct it to a solution of the nonlinear equation. Three features of the problem shape everything that follows. \paragraph{Flat ends force infinitely many changes of type.} For $w>0$ the operator \eqref{eq:intro-linear} is elliptic where $w'>0$ and hyperbolic where $w'<0$. At a flat end $w$ tends to zero. If $w$ were monotone near that end, no nonzero angular mode of \eqref{eq:intro-linear} could vanish there together with its flux (Proposition~\ref{prop:obs-monotone-tail}). So $w$ must oscillate: near each end there are infinitely many circles on which the equation changes type, and they accumulate at the boundary. We build such a profile from infinitely many copies of one explicitly solvable cell, with lengths and amplitudes decaying fast enough for smoothness (Sections~\ref{sec:rosen-reference-packet} and~\ref{sec:radial-tails}). \paragraph{Two unrelated resonance conditions.} On each hyperbolic annulus, characteristics turn by a fixed \emph{travel angle} $T$, and solving along them involves dividing by $\sin(jT)$ at the angular frequency $j$. Independently, the solutions on neighbouring annuli must be matched, and the matching fails at eigenvalue-like resonances. A finite model in which both sets of resonances can be computed exactly (Section~\ref{sec:worked-example}) shows that neither condition implies the other. In the actual profile the travel angles are quadratic irrationals fixed by the choice of the cells, while matching resonances are avoided by adjusting the relative amplitudes of consecutive cells. The outcome is a two-sided inverse of $\mathscr L_w$ on the whole annulus, in spaces weighted towards the flat ends (Section~\ref{sec:radial-reference-inverse}). \paragraph{The inverse has to survive perturbation.} The Nash--Moser method needs an inverse of the linearization at every state near the reference one, not only at the reference. For a nearby state both the type-change circles and the characteristic return maps move, so its linearization does not even act on the same space. We remove this problem by subtracting from the linearization two terms, each proportional to an equation that is not yet solved: one fixes the type-change circles, the other fixes the return maps. Both terms vanish at a solution, and they produce exactly the quadratic error allowed by Hamilton's form of the implicit function theorem \cite{Hamilton1982} (Sections~\ref{sec:fixed-geometry}--\ref{sec:solving-the-equation}). \subsection{Why the problem was open, and what is new} \label{subsec:novelty} \paragraph{What the existing theory provides.} A closed surface which can be bent is not convex: two isometric closed convex surfaces are congruent \cite{CohnVossen1927,Herglotz1943,Pogorelov1973}. Since a closed surface of nonnegative curvature is convex, the curvature of a flexible closed surface takes both signs, and the equation for its isometric embeddings is of mixed type. For such equations the existence theorems known to us are local or semi-global. The local ones construct an embedding of a neighbourhood of a point at which the curvature changes sign, under a hypothesis on the way it vanishes: cleanly across a curve \cite{Lin1986,Han2005}, to finite order across a curve \cite{Khuri2007}, or to finite order on two curves which cross \cite{HanKhuri2010}. The semi-global ones construct an embedding of a narrow band along one closed curve on which the curvature vanishes: on both sides of the curve when the curvature changes sign cleanly \cite{Dong1993} or to finite order \cite{Cao2019}, and on one side of it in \cite{Li2012}. The angular variable is periodic there, as it is here, but a single curve of type change is involved, and in \cite{Li2012,Cao2019} the embedding has finite regularity. A theorem of either kind produces one surface, and a bending is a path of surfaces. In the other direction, an infinitesimal bending of a surface with a planar patch can be written down at once (Section~\ref{subsec:context}), but nothing guarantees that it is the derivative of a bending. \paragraph{What was missing.} The mechanism of Section~\ref{subsec:mechanism} is elementary, and it turns one surface into a path. What was not available is a surface to which it applies: a nonradial isometric embedding of a rotationally symmetric disk which is exactly planar near the boundary. The existence of such a disk is a problem of a different kind from the local and the semi-global ones. The equation has to be solved in $C^\infty$ on a whole annulus, with conditions at both boundary circles. It changes type on infinitely many circles which accumulate at the boundary (Proposition~\ref{prop:obs-monotone-tail}), so no hypothesis of finite order holds there. And its linearization has the two families of resonances of Section~\ref{subsec:difficulties}. One circumstance works in our favour. In an embedding problem the metric is given. Here it is not: Ghomi's question asks for some surface, so the profile $w$ is ours to choose, and we choose it so that the linearized equation can be solved explicitly. \paragraph{The new steps.} \begin{enumerate}[label=\textup{\arabic*.}] \item \emph{The profile.} A profile $w_*$ with infinitely many changes of type at each end, for which the linearized equation has a solution $\phi_*(r)\cos2\theta$ flat at both ends, has no such solution at the higher even frequencies, and has, after bordering by one scalar, an inverse bounded in weighted spaces at all frequencies (Proposition~\ref{prop:rad-fixed-theorem}). Its ingredients are classical: two solvable models, a chain recurrence, Schur complements; so is the design of a meridian on which only finitely many modes are admissible (Section~\ref{subsec:context}). What is new is their assembly into one inverse on the whole annulus, and in particular the choice of the amplitudes of the cells which avoids the matching resonances at all frequencies simultaneously (Lemma~\ref{lem:radial-product-selection}). \item \emph{The second residual correction.} Subtracting from the linearization a multiple of the residual of the equation is a device of Han and Khuri \cite[pp.~650--653]{HanKhuri2010}; here it fixes the circles on which the type changes (Section~\ref{subsec:first-correction}). On an annulus of hyperbolic type which is traversed completely this is not enough, because the return map of the characteristics also depends on the state. We add an equation which says that the return maps are rotations in suitable charts, with the charts as new unknowns, and subtract from the linearization a second term, proportional to the residual of this equation (Section~\ref{subsec:return-correction}). \item \emph{The inverse at nearby states.} After the two corrections all linearizations act on one domain, and they are inverted by a Neumann series on a single Hilbert space, although the reference inverse loses weight (Section~\ref{sec:stable-inverse}). \end{enumerate} The Nash--Moser theorem itself is used as a black box, in Hamilton's form. \subsection{Organization of the proof} \label{subsec:organization} Figure~\ref{fig:proof-map} shows the logical structure, and Table~\ref{tab:sections} lists what each section contributes. The paper has three parts. \begin{figure}[htbp] \centering \begin{tikzpicture}[font=\small,>=Stealth, box/.style={draw,rounded corners=2pt,align=center,inner sep=5pt, minimum height=9mm}] \node[box,fill=blue!7] (A) at (0,0) {Theorem~\ref{thm:main}: flexible sphere}; \node[box] (C) at (0,-1.45) {Theorem~\ref{thm:geometric-closure}: geometric closure\quad (\S\ref{sec:geometric-closure})}; \node[box,fill=blue!7] (B) at (0,-2.9) {Theorem~\ref{thm:annulus-intro} $=$ Theorem~\ref{thm:new-annular-height}: annular height with flat ends}; \node[box] (N) at (0,-4.35) {Nash--Moser branch $q\mapsto\bx(q)$\quad (\S\ref{sec:solving-the-equation})\quad and endpoint gain\quad (\S\ref{sec:final-proof})}; \node[box] (I) at (0,-5.8) {Proposition~\ref{prop:rewrite-volume-inverse}: inverse of the corrected linearization near the reference\quad (\S\ref{sec:stable-inverse})}; \node[box] (R) at (-4.1,-7.5) {Proposition~\ref{prop:rad-fixed-theorem}:\\ reference profile $w_*$, seed $\phi_*$,\\ inverse of $\mathscr L_{w_*}$\quad (\S\S\ref{sec:rosen-reference-packet}--\ref{sec:radial-reference-inverse})}; \node[box] (F) at (4.1,-7.5) {Corrected linearization\\ on a state-independent domain\\ (\S\ref{sec:scalar-core}, \S\ref{sec:fixed-geometry})}; \draw[->] (C)--(A); \draw[->] (B)--(C); \draw[->] (N)--(B); \draw[->] (I)--(N); \draw[->] (R)--(I); \draw[->] (F)--(I); \end{tikzpicture} \caption{The structure of the proof. An arrow means ``is used to prove''.} \label{fig:proof-map} \end{figure} \begin{table}[htbp] \centering \begin{tabularx}{\linewidth}{@{}l Y@{}} \toprule Section & What it provides\\ \midrule \S\ref{sec:geometric-closure} & Theorem~\ref{thm:annulus-intro} implies Theorem~\ref{thm:main}: planar development, the marked cap, noncongruence, invariants.\\ \S\ref{sec:scalar-core} & The Darboux equation as the curvature of a connection form; its linearization and conservation law; why the ends must oscillate; the two resonances in a finite model; the first residual correction.\\ \S\ref{sec:rosen-reference-packet} & Statement of what the reference profile must provide (Proposition~\ref{prop:rad-fixed-theorem}); the two solvable models from which it is built.\\ \S\ref{sec:radial-tails} & The infinite chain of cells at each end; the mass--resistance recurrence and the decaying solution; choice of the cell amplitudes.\\ \S\ref{sec:radial-reference-inverse} & The inverse of $\mathscr L_{w_*}$ on the whole annulus; proof of Proposition~\ref{prop:rad-fixed-theorem}.\\ \S\ref{sec:fixed-geometry} & Weighted spaces; the unknowns of the nonlinear problem; the second residual correction; coordinates in which every nearby linearization has the same transmission conditions.\\ \S\ref{sec:stable-inverse} & The inverse of the corrected linearization at nearby states, with tame estimates (Proposition~\ref{prop:rewrite-volume-inverse}).\\ \S\ref{sec:solving-the-equation} & The coupled linear system and Hamilton's theorem: a smooth branch of solutions.\\ \S\ref{sec:final-proof} & Improved decay of the metric corrections; the precise annular theorem (Theorem~\ref{thm:new-annular-height}); proof of Theorem~\ref{thm:main}.\\ Appendix~\ref{app:app-obstructions} & A side remark: no finite angular ansatz can solve the nonlinear equation.\\ Appendix~\ref{app:dependencies} & The external results in the form used, and the order in which the constants are chosen.\\ Appendix~\ref{app:notation} & Index of notation.\\ \bottomrule \end{tabularx} \caption{Contents by section.} \label{tab:sections} \end{table} \emph{Part~\ref{part:geometry}} (Sections~\ref{sec:geometric-closure} and~\ref{sec:scalar-core}) contains the geometry and the derivation of the scalar equation. It is self-contained and elementary. \emph{Part~\ref{part:reference}} (Sections~\ref{sec:rosen-reference-packet}--\ref{sec:radial-reference-inverse}) is one-dimensional. After separation of angular modes, the linear equation \eqref{eq:intro-linear} becomes a family of ordinary differential equations, and the whole part is devoted to the construction of one profile $w_*$ for which these equations have the right kernel and a controlled inverse. Its only output is Proposition~\ref{prop:rad-fixed-theorem}; a reader willing to accept that proposition can go directly from Section~\ref{sec:radial-what} to Part~\ref{part:nonlinear}. \emph{Part~\ref{part:nonlinear}} (Sections~\ref{sec:fixed-geometry}--\ref{sec:final-proof}) solves the nonlinear equation near the reference and proves the two theorems. The estimates are arranged so that precision is spent where the decisions above are implemented: the matching of infinitely many intervals, the common domain of the perturbed operators, and the decay at the endpoints are proved in detail, while the spectra of the models and the elementary operator theory are quoted. The specific numbers that appear in the construction, such as the exponents $2n^3/3$ and $n^{5/2}$, the power $12$ and the weight exponent $3/4$, are each introduced together with the inequalities they have to satisfy; none of them is optimal, and Table~\ref{tab:endpoint-budget} collects them. \subsection{Context} \label{subsec:context} \paragraph{Continuous families versus pairs.} Ghomi's problem \cite[Problem~1.1]{Ghomi2019} asks for a continuous family. In his survey of the work of Cohn-Vossen, Alexandrov puts the question in the words ``Do there exist closed surfaces that admit a continuous finite bending?'' \cite{Alexandrov1947}. Pairs are easier to obtain than paths. Replacing a dimple by its reflection across a flat rim gives two different isometric realizations of one metric, and an infinitesimal bending $z$ of a closed surface $x$ gives the isometric pair $x\pm tz$ \cite{CohnVossen1930,Alexandrov1947}; neither gives a path between the two surfaces. Sa Earp and Toubiana \cite{SaEarpToubiana2002} construct smooth closed surfaces, some of genus zero, which have isometric and noncongruent smooth realizations arbitrarily close to them in every $C^k$ topology, as well as families of isometric surfaces parametrized by a nowhere dense compact subset of $[0,1]$. Ghomi, Qiu, Raffaelli and Zhang \cite{GhomiEtAl2026} recently used the second observation to construct, for every genus $g\ge1$, a metric on the closed surface of genus $g$ with uncountably many smooth tight isometric embeddings no two of which are congruent. They state that it is not known whether these surfaces are flexible, and that no smooth counterpart of Connelly's flexible polyhedra has been found. Theorem~\ref{thm:main} provides a path in the smooth category. \paragraph{Infinitesimal versus finite bendings.} A normal vector field supported inside a planar patch has zero first variation of the metric. This observation already appears in Cohn-Vossen's discussion of closed infinitesimally nonrigid surfaces \cite[footnote~2]{CohnVossen1930}; it does not solve the nonlinear isometry equations. Efimov's survey \cite{Efimov1948Survey} and his work on deformations at points of zero curvature \cite{Efimov1948Zero} belong to the historical setting of this distinction, and the later survey \cite[Sections~1.1--1.4]{IvanovaSabitov1991} treats finite bendings, their infinitesimal equations and higher-order extension. Our starting point is also an infinitesimal bending, the solution of \eqref{eq:intro-linear}; the analysis of Parts~\ref{part:reference} and~\ref{part:nonlinear} is what turns it into an actual height. \paragraph{Surfaces of revolution and their harmonics.} The main subject of Cohn-Vossen's paper is close to Part~\ref{part:reference}. He expands an infinitesimal bending of a closed surface of revolution in Fourier modes in the angle, obtains for each mode an ordinary differential equation along the meridian, and constructs meridians for which a mode of order at least two has a solution regular at both poles; he also shows that his surfaces admit infinitesimal bendings of second order \cite{CohnVossen1930,Alexandrov1947}. In \S6 of that paper he considers the extension of an infinitesimal bending to higher order for a fixed metric. With a single admissible mode $m$, the second order involves the modes $0$ and $2m$; at the third order the mode $m$ reappears, with a compatibility condition which he leaves open. Later work sharpens the control of the admissible modes. Reshetnyak \cite[Theorem~1*]{Reshetnyak1962} and Trotsenko \cite[Theorem~1*]{Trotsenko1980} show that for every $m\ge1$ there are distinct integers $k_1,\ldots,k_m\ge2$ and a closed surface of revolution, smooth in the first paper and analytic in the second, for which the equation of the mode $k$ has a nonzero solution regular at both poles exactly when $k\in\{k_1,\ldots,k_m\}$. The integers are not prescribed in advance. Milka \cite[\S2]{Milka1965} exchanges the roles of the unknowns: he prescribes a component of the bending field in a given mode and recovers the meridian by integration. For surfaces of revolution which are flattened at the poles, Sabitov shows that in the non-analytic class there are nonrigid ones with arbitrary prescribed orders of flattening \cite{Sabitov2013}, and gives conditions under which a given harmonic of an infinitesimal bending extends to second order \cite{Sabitov2014}. The design problem of Section~\ref{sec:radial-models} is of the same kind. What the nonlinear problem adds are three requirements: the ends are not poles on the axis but circles of positive radius at which the profile is flat to infinite order; the single admissible mode is prescribed; and the inverse on all other modes is bounded uniformly in the frequency. \paragraph{Rigidity.} On the rigid side, hypotheses on curvature and on boundary data are decisive. Besides the theorem on closed convex surfaces quoted in Section~\ref{subsec:novelty}, Ghomi and Spruck \cite[Theorem~1.1]{GhomiSpruck2020} prove that two properly oriented $C^{2,1}$ positively curved isometric immersions which agree on a curve segment agree everywhere on their connected domain. The annulus of Theorem~\ref{thm:annulus-intro} has curvature of both signs, with infinitely many transitions at each flat end, so neither strict positivity nor a finite collection of nondegenerate interfaces describes its geometry. \paragraph{Other categories.} Connelly's flexible embedded polyhedral sphere \cite{Connelly1977} establishes flexibility for polyhedra. Rounding its edges would change the metric, so it does not furnish a smooth bending. The $C^1$ theorems of Nash and Kuiper \cite{Nash1954,Kuiper1955} concern another regularity regime; in the smooth problem the second fundamental form and the type of the Darboux equation must be controlled at every scale. \paragraph{Smooth isometric embedding.} Our analytic method is closest to the theory of smooth isometric embedding. Smoothing already compensates for the loss of derivatives in Nash's argument \cite{Nash1956}. Han, Hong and Lin \cite[Theorem~1.1]{HanHongLin2003} use a Nash--Moser--H\"ormander scheme for nonpositive curvature under a specified degeneracy condition, and Han and Khuri treat local problems of mixed type with finite-order changes of type \cite{HanKhuri2010,HanKhuri2013}; the monograph \cite{HanHong2006} is a general reference. The modification of the linearization by a multiple of the residual in \cite[pp.~650--653]{HanKhuri2010} is the precedent for our residual corrections. What these works do not have to face is the simultaneous control of infinitely many interfaces and of both flat ends on a single closed domain. \paragraph{The invariants.} The three invariants in Theorem~\ref{thm:main} follow directly from the form \eqref{eq:phase-family} of the motion. They should be compared with general results of different scope: Sabitov's volume theorem \cite{Sabitov1996} is polyhedral, and Almgren and Rivin \cite{AlmgrenRivin1998} prove the invariance of the mean-curvature integral under smooth bendings. Motions in which curvature precesses around a shell also arise in the mechanical model of Hamouche, Maurini, Vidoli and Vincenti \cite{Hamouche2017}; that comparison supplies neither the flat collars nor the closed embedded surface. \subsection{Conventions} \label{subsec:conventions} Throughout, $\T=\R/2\pi\Z$, the angular average of a function is \[ \av f=\frac1{2\pi}\int_0^{2\pi}f\,d\theta , \] and $R_\vartheta$ denotes a rotation by the angle $\vartheta$, of the plane, of $\R^3$ about the vertical axis, or of the circle ($R_\vartheta\theta=\theta+\vartheta$), as the context requires. A prime denotes a derivative with respect to the radius $r$ unless another variable is indicated by a subscript. The letters $C$ and $c$ denote positive constants, large and small respectively, whose values may change from line to line; a subscript indicates dependence, as in $C_s$. We write $a\asymp b$ when $cb\le a\le Cb$. Constants that are fixed once and referred to later carry a distinguishing mark, as in $C_*$, $c_0$ or $N$. The paper uses a fairly large number of symbols. Each of Parts~\ref{part:reference} and~\ref{part:nonlinear} begins by fixing its vocabulary, and Appendix~\ref{app:notation} is an index of notation. Some letters have a local meaning inside a single proof; this is always said where it happens.