\documentclass[11pt]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage[margin=1.25in]{geometry} \usepackage{amssymb,mathtools} \usepackage{microtype} \usepackage[colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue]{hyperref} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\Q}{\mathbb{Q}} \newcommand{\F}{\mathbb{F}} \newcommand{\HH}{\mathbb{H}} \newcommand{\ad}{\operatorname{ad}} \newcommand{\ind}{\operatorname{ind}} \newcommand{\cs}{\operatorname{cs}} \newcommand{\coker}{\operatorname{coker}} \newcommand{\im}{\operatorname{im}} \newcommand{\tr}{\operatorname{tr}} \newcommand{\rk}{\operatorname{rk}} \newcommand{\B}{\mathcal{B}} \newcommand{\G}{\mathcal{G}} \newcommand{\spinc}{\mathfrak{s}} \begin{document} \title[Spherical space forms in K3]{Spherical space forms in K3:\\ an instanton parity argument} \date{15 November 2026} \author{P. B. Kronheimer} \author{T. S. Mrowka} \begin{abstract} Let $R$ be a connected subdiagram of the Dynkin diagram $E_8$, and let $Y_R$ be the spherical three-manifold which bounds the corresponding plumbing of $(-2)$-spheres. We show that if a negative-definite four-manifold $N$ with $H_1(N;\Z)=0$ and oriented boundary $Y_R$ is embedded in the K3 surface, then the intersection form of $N$ is the negative-definite root lattice $-R$. As a consequence, the traces of $-2$ and $-4$ surgery on the left-handed trefoil do not embed in K3. Heegaard Floer correction terms and the theory of lattices reduce the theorem to the exclusion of one smaller lattice for each of $R=D_5$, $D_6$, $D_7$ and $E_7$. In these cases, stretching the K3 surface along $Y_R$ exhibits a zero-dimensional moduli space of instantons as the zero set of a section of a real two-plane bundle over a compact surface. The quaternionic nature of the linearized anti-self-duality equations on K3 shows that the Euler number of this bundle is even, while the corresponding Donaldson invariant is odd. \end{abstract} \maketitle \section*{Foreword}\label{foreword:section} With the exception of this foreword, the present paper is entirely AI-generated, in response to prompts provided by the human authors. The mathematics was produced by GPT-6 Astra, and the paper was redrafted by Claude Fable 5.2. The main result is Theorem~\ref{thm:main}. A particular case states that the 4-manifold $N$ obtained by adding a 2-handle to the 4-ball along a left-handed trefoil with framing $-2$ cannot be smoothly embedded in a $K3$ surface. (See Corollary~\ref{cor:trace}.) These results arose from prompts motivated by a question which the authors learned of from Anubhav Mukherjee. Let $S$ be the standard 3-sphere in the neck of the connected sum $K3 \# K3$. Does the connected sum contain any sphere different from $S$? Here one can interpret different as up to isotopy; or one can ask whether $\pm[S]$ is the only non-zero class in $\pi_{3}$ that is represented by an embedding. In the latter version, the question is closely related to the question of whether a diffeomorphism of $K3 \# K3$ can ``mix'' the homology of the two summands. The clearest strategy to construct such a mixing diffeomorphism starts with a hypothetical embedded $2$-sphere $\Sigma$ in $K3\# K3$ with $\Sigma \cdot \Sigma = -2$ such that the class $[\Sigma] \in H_{2}(K3)\oplus H_{2}(K3)$ has non-zero projection onto both sides. Then reflection in $[\Sigma]$ arises from a diffeomorphism of $K3\# K3$ which does not preserve the direct sum decomposition. (Indeed, if $[\Sigma]$ is the sum of two such \emph{primitive} classes, then it turns out that this would imply that the diffeomorphism group of $K3 \# K3$ acts by the full isometry group of the homology latttice.) To try and construct such a 2-sphere, one might start with a cuspidal elliptic fiber on one side, and seek on the other side an embedded disk with relative self-intersection number $-2$ and boundary a trefoil in the 3-sphere $S$. Tubing the trefoil in $S$ to the cusp gives the sought-after 2-sphere. But finding the disk in punctured $K3$ is equivalent to embedding the 4-dimensional handlebody $N$ mentioned above, which Corollary~\ref{cor:trace} says cannot be done. Of course, only the simplest construction of such a sphere is ruled out by this argument, because the hypothetical sphere $\Sigma$ might meet $S$ in an arbitrary knot or link instead of the trefoil. ChatGPT proved the corollary first, for the case of the $-2$ framed left-handed trefoil. It began by giving an argument conditional on the superconformal simple type conjecture \cite{MMP, FL}. If $N$ embeds in $K3$, then the manifold obtained by replacing this piece with an $E_{7}$ plumbing is a manifold whose Seiberg-Witten invariants can be computed sufficiently to obtain a contradiction to that conjecture. The AI then found an unconditional proof using Donaldson invariants, which is the one presented here. In outline, here is the proof by contradiction as it applies to a hypothetical embedding of $N = X_{-2}(T_{-})$ in $K3$. With a suitable $\mathrm{SO}(3)$ bundle on $K3$, one obtains a zero-dimensional ASD moduli space, and the Donaldson invariant is known to be $\pm 1$. On the other hand, stretching the metric along $\partial N$ leads to a gluing description which presents the moduli space as the zero set of a section of an obstruction bundle $L\to M$, where $L$ has real rank $2$ and $M$ is a compact, oriented $2$-manifold. This describes the Donaldson invariant as the Euler class $\langle e(L) , [M] \rangle$. However, $[L] - [TM]$ has an interpretation as the element of $KO(M)$ arising from the index of the ASD deformation complex. The quaternionic structure of the tangent bundle of $K3$ implies that this element is even, and it follows that the Donaldson invariant of $K3$, computed as $\langle e(L) , [M] \rangle$, is even, which contradicts the known result. The justification of the required gluing result in section 5 of the paper is verbose. The AI initially found no suitable models for the required analytic steps in the literature. The human authors pointed to the cited papers of Taubes, which provided both the overall strategy for the gluing theorem and the analytic details. With this background, GPT-6 wrote a gluing argument which was later condensed by Fable. After completing the proof of the corollary, the generalization given in Theorem~\ref{thm:main} was suggested by the authors, as something that might follow from similar arguments. As noted in the paper, the instanton parity argument is needed only in the cases $D_{5}$, $D_{6}$, $D_{7}$ and $E_{7}$, to rule out the lattices $\langle -4 \rangle$, $-(A_{1} \oplus A_{1})$, $-A_{3}$ and $-A_{1}$ respectively. The above corollary and the parity argument in its proof are both new to the authors. Distributing the preprint in this form, the authors would be glad to hear from others if these results are known. \subsection*{Acknowledgements.} The work of the first author was supported by NSF grant DMS-2304877. Both authors were supported by Simons Foundation Awards \#994330 and \#14760-09, \emph{Simons Collaboration on New Structures in Low-Dimensional Topology}. \clearpage \section{Introduction}\label{sec:intro} Let $R$ be a Dynkin diagram of type $A$, $D$ or $E$. We use the same letter for the corresponding root lattice, which we take to be positive definite. Let $P_R$ be the four-manifold obtained by plumbing disk bundles of Euler number $-2$ over the two-sphere according to the diagram $R$, so that the intersection form of $P_R$ is $-R$, and let $Y_R=\partial P_R$ be its oriented boundary. Thus $Y_R$ is the quotient of $S^3$ by a finite subgroup of $SU(2)$. This paper is concerned with embeddings of the manifolds $Y_R$ in the K3 surface. Our main result is the following. \begin{theorem}\label{thm:main} Let $R$ be a connected subdiagram of the Dynkin diagram $E_8$, and let $N$ be a compact, connected, codimension-zero submanifold of the K3 surface whose oriented boundary is $Y_R$. Suppose that $N$ is negative definite and that $H_1(N;\Z)=0$. Then the intersection form of $N$ is isomorphic to $-R$. \end{theorem} The diagrams $R$ to which the theorem applies are $A_n$ for $1\le n\le 7$, $D_n$ for $4\le n\le 7$, and $E_6$, $E_7$ and $E_8$. With the hypotheses of the theorem, the group $H_2(N;\Z)$ is free and $H_3(N;\Z)$ is zero (Proposition~\ref{prop:reduction}), so $N$ has the same integral homology as the plumbing $P_R$. In three cases, the conditions which the theorem imposes on $N$ do not need to be assumed. If $R$ is $E_7$, $D_5$ or $D_7$, and $Y_R$ is embedded in the K3 surface in any manner, then it divides the K3 surface into two pieces, and one of the two is negative definite, has oriented boundary $Y_R$, and has vanishing first homology. Theorem~\ref{thm:main} applies to that piece, whose intersection form is therefore $-R$. See Theorems~\ref{thm:octahedral} and~\ref{thm:dihedral}. For a knot $K\subset S^3$ and an integer $m$, let $X_m(K)$ denote the trace of $m$-surgery on $K$: the four-manifold obtained by attaching a two-handle to the four-ball along $K$ with framing $m$. \begin{corollary}\label{cor:trace} Let $K$ be the left-handed trefoil. Then the trace $X_{-2}(K)$ does not admit an orientation-preserving smooth embedding in the K3 surface. The same is true of $X_{-4}(K)$. \end{corollary} \begin{proof} As oriented manifolds, the boundary of $X_{-2}(K)$ is $Y_{E_7}$ \cite[Lemma~1.2]{DLME}, and the boundary of $X_{-4}(K)$ is $Y_{D_5}$ \cite{Moser}. The traces are simply connected, and their intersection forms are $\langle-2\rangle$ and $\langle-4\rangle$ respectively, rather than $-E_7$ and $-D_5$. \end{proof} \subsection*{Outline of the proof} The first part of the proof of Theorem~\ref{thm:main} uses Heegaard Floer correction terms and the theory of lattices. The manifold $N$ is spin, so the inequality of Ozsv\'ath and Szab\'o \cite{OS} bounds $b_2(N)$ in terms of the correction terms of $Y_R$, and also determines $b_2(N)$ modulo~$8$. The classification of even definite lattices of small rank and determinant then shows that the intersection form of $N$ is $-R$ when $R$ is $D_4$, $E_6$ or $E_8$. The same holds for $R=A_n$, by an argument which also uses Donaldson's diagonalization theorem. In the four remaining cases, there is one possibility for the intersection form other than $-R$: \begin{equation}\label{eq:four-cases} \begin{tabular}{c|cccc} $R$ & $E_7$ & $D_5$ & $D_6$ & $D_7$\\ \hline \rule{0pt}{2.6ex}other possibility for $Q_N$ & $\langle-2\rangle$ & $\langle-4\rangle$ & $-(A_1\oplus A_1)$ & $-A_3$ \end{tabular} \end{equation} To exclude these four possibilities we use instantons. Consider the case $R=E_7$, and suppose that a K3 surface $X$ is decomposed as $N\cup_Y C$, where $Y=Y_{E_7}$ and $N$ has intersection form $\langle-2\rangle$. Let $P\to X$ be an $SO(3)$ bundle with $p_1(P)=-6$ whose second Stiefel--Whitney class is non-zero on the generator of $H_2(N)$. The moduli space of anti-self-dual connections in $P$ has formal dimension zero, and the corresponding Donaldson invariant, which counts the points of this moduli space with signs, is $\pm1$. We examine the moduli space as the metric on $X$ is stretched along $Y$. On the manifold $N$ with a cylindrical end attached, the bundle $P$ carries a reducible anti-self-dual connection $A_0$ of action $1/8$, asymptotic to a reducible flat connection $\alpha$ on $Y$. A dimension count shows that, when the neck is long, every anti-self-dual connection in $P$ is close to $A_0$ on $N$, and on $C$ is close to a point of a moduli space $M$ of instantons asymptotic to $\alpha$, where $M$ is a compact two-manifold. The connection $A_0$, however, belongs to a moduli space of formal dimension $-3$. Its stabilizer is a circle and, however the equations are perturbed, the cohomology group $H^2_{A_0}$ of its deformation complex is not zero: it is a copy of $\C$ on which the stabilizer acts with weight one. In this situation, gluing theory describes the moduli space on $X$ not as a copy of $M$ but as the zero set of a section of a real two-plane bundle $E\to M$, the bundle of obstruction spaces. The Donaldson invariant is therefore equal, modulo~$2$, to the Euler number of $E$. The proof is completed by showing that this Euler number is even. The gluing construction provides a family of connections on $X$ parametrized by $M$, and the index of the family of linearized operators $d_A^*\oplus d_A^+$ is the element $[TM]-[E]$ of $KO(M)$. For a hyperk\"ahler metric on $X$, the operator $d_A^*\oplus d_A^+$ can be identified with a Dirac operator coupled to the real vector bundle $\ad P$, and so commutes with a quaternionic structure. The index of any family of such operators therefore lies in the image of $KSp\to KO$, and in particular its second Stiefel--Whitney class is zero. It follows that $w_2(E)=w_2(TM)$, so the Euler number of $E$ has the same parity as the Euler number of the surface $M$, which is even. The cases $R=D_5$, $D_6$ and $D_7$ are dealt with in the same way, and differ only in the numerical data. The method of gluing in the presence of an obstruction goes back to Taubes \cite{TauIndef} and Donaldson \cite{Don86}, and is a standard part of the theory \cite{Don,DK}. Obstructed gluing at a reducible connection, for cobordisms between rational homology spheres, is studied in \cite[Section~5]{DME}. We have collected the analysis in Section~\ref{sec:analytic}, and the reader who is willing to take Proposition~\ref{prop:gluing} on trust can omit that section. \subsection*{Organization} Sections~\ref{sec:lattice}--\ref{sec:parity} deal with the case $R=E_7$. Section~\ref{sec:lattice} contains the lattice-theoretic part of the argument, Section~\ref{sec:moduli} describes the moduli spaces on the two pieces, and Section~\ref{sec:parity} contains the parity argument. The transversality and gluing results on which these depend are proved in Section~\ref{sec:analytic}. The cases $R=D_5$, $D_6$ and $D_7$ are treated in Section~\ref{sec:dihedral}, and the remaining cases in Section~\ref{sec:ade}, where the proof of Theorem~\ref{thm:main} is completed. \section{The case \texorpdfstring{$R=E_7$}{R=E7}: constraints on the lattice}\label{sec:lattice} In this section and the next two, $Y$ denotes the manifold $Y_{E_7}$. It is the quotient of $S^3$ by the binary octahedral group $O^*$, a group of order $48$ whose abelianization is $\Z/2$. For this manifold we prove a statement which is a little stronger than Theorem~\ref{thm:main}, in that the hypotheses on $N$ become conclusions. \begin{theorem}\label{thm:octahedral} Suppose that $Y=Y_{E_7}$ is smoothly embedded in a K3 surface $X$. Then $Y$ separates $X$ into two pieces, \[ X=W\cup_Y C, \] where $W$ has oriented boundary $Y$ and is negative definite. The groups $H_1(W;\Z)$ and $H_1(C;\Z)$ are zero, and the intersection form of $W$ is isomorphic to $-E_7$. In particular, $b^+(C)=3$ and $b^-(C)=12$. \end{theorem} The proof occupies this section and the next three. Here we prove the following. \begin{proposition}\label{prop:alternatives} In the situation of Theorem~\ref{thm:octahedral}, one of the two pieces, $W$, is negative definite and has oriented boundary $Y$. The groups $H_1(W;\Z)$ and $H_1(C;\Z)$ are zero, and the intersection form $Q_W$ is isomorphic either to $-E_7$ or to $\langle-2\rangle$. \end{proposition} \begin{proof} Since $H_3(X;\Z)=0$, the submanifold $Y$ separates $X$ into two pieces. Let $V$ denote either of them. In the Mayer--Vietoris sequence, the groups $H_2(Y;\Z)$ and $H_1(X;\Z)$ are zero, so $H_2(V;\Z)$ injects into $H_2(X;\Z)$ and is therefore free, while $H_1(V;\Z)$ is a quotient of $H_1(Y;\Z)$ and is therefore finite. The exact sequence of the pair $(V,Y)$ and Poincar\'e--Lefschetz duality then give \begin{equation}\label{eq:det} |H_1(Y;\Z)|=|\det Q_V|\,|H_1(V;\Z)|^2. \end{equation} The left-hand side is $2$, so $H_1(V;\Z)=0$ and $|\det Q_V|=2$. The form $Q_V$ is even, because $X$ is spin. The manifold $Y$ is an $L$-space and $X$ is symplectic, so one of the two pieces has $b^+=0$ \cite[Theorem~1.2]{Muk}. (If $b^+$ were positive on both sides, then the Ozsv\'ath--Szab\'o invariants of $X$ would factor through $HF_{\mathrm{red}}(Y)=0$, and they would vanish, which they do not for a symplectic manifold.) We call this piece $W$. It is negative definite, because $Q_W$ is non-degenerate. To see which way the boundary of $W$ is oriented, we use the correction terms of $Y$. There are two spin$^c$ structures on $Y$, both of them spin, and their correction terms are \begin{equation}\label{eq:dterms} d(Y,\spinc)=7/4\quad\text{and}\quad 1/4. \end{equation} See \cite[Lemma~1.2]{DLME}, where the values are given for $-Y$. If $V$ is a negative-definite spin manifold whose boundary is a rational homology sphere, and $\spinc$ is the spin structure which the boundary inherits, then \[ b_2(V)\le 4\,d(\partial V,\spinc) \] by \cite[Theorem~9.6]{OS}. Since both correction terms of $-Y$ are negative, the oriented boundary of $W$ is $Y$ and not $-Y$, and the rank $r=b_2(W)$ is at most $7$. It is at least $1$, because $|\det Q_W|=2$. Let $L$ be the lattice $-Q_W$, which is even and positive definite, of determinant $2$. Its discriminant group has one non-zero element, on which the discriminant quadratic form takes a value $q$ which is either $1/2$ or $3/2$ modulo~$2$. The Milgram formula \cite[Appendix~4]{MH} gives \[ e^{2\pi i r/8}=\frac{1+e^{\pi i q}}{\sqrt2}, \] so $r\equiv1\pmod 8$ if $q=1/2$, and $r\equiv 7\pmod 8$ if $q=3/2$. So either $r=1$, in which case $L=\langle2\rangle$, or $r=7$ and $q=3/2$. In the second case, the lattice $L\oplus A_1$ has an even unimodular overlattice of index $2$. This overlattice is positive definite of rank $8$, and is therefore isomorphic to $E_8$ \cite[Chapter~V]{Serre}. In it, $L$ is the orthogonal complement of a root, so $L$ is isomorphic to $E_7$. \end{proof} To prove Theorem~\ref{thm:octahedral}, it remains to show that $Q_W$ cannot be $\langle-2\rangle$. \section{The moduli spaces when \texorpdfstring{$N$}{N} has rank one}\label{sec:moduli} We suppose from now on that the second alternative in Proposition~\ref{prop:alternatives} occurs, and we write $N$ for $W$. So $X$ is a K3 surface, decomposed as \[ X=N\cup_Y C,\qquad Q_N\cong\langle-2\rangle, \] and our aim is to derive a contradiction. By Proposition~\ref{prop:alternatives}, the first homology groups of $N$ and $C$ are zero. The group $H_2(N;\Z)$ is generated by a class $x$ with $x\cdot x=-2$, and we have \begin{equation}\label{eq:chisig} \chi(N)+\sigma(N)=1,\qquad b^+(C)=3,\quad b^-(C)=18,\qquad \chi(C)+\sigma(C)=7. \end{equation} The class $x$ is primitive in $H_2(X;\Z)$, because the intersection form of $X$ is even. \subsection{The bundle} Since $x$ is primitive and has square $-2$, there is an automorphism of the lattice $H_2(X;\Z)$ which carries $x$ to $e-f$, where $e$ and $f$ are the standard basis of a hyperbolic summand \cite[Proposition~3.3(i)]{GHS}. Let $e'$ and $f'$ be the standard basis of a second hyperbolic summand, orthogonal to the first, and let $w\in H^2(X;\Z)$ be the class dual to $f+e'-3f'$. Then \begin{equation}\label{eq:w} w[x]=1,\qquad w^2=-6. \end{equation} Let $\mathcal{E}\to X$ be the $U(2)$ bundle with $c_1(\mathcal{E})=w$ and $c_2(\mathcal{E})=0$, and let $P=\ad(\mathcal{E})$ be the associated $SO(3)$ bundle, so that \[ w_2(P)=w \bmod 2,\qquad p_1(P)[X]=-6. \] We adopt the usual conventions for such bundles, as in \cite{KM}: we consider connections $A$ in $P$, and the gauge group $\G$ consists of the automorphisms of $\mathcal{E}$ of determinant one, so that an irreducible connection has stabilizer $\{\pm1\}$ in $\G$. We write $\kappa$ for the action of a connection, normalized so that $\kappa=-p_1(P)/4$ on a closed manifold. For the bundle $P$ we have $\kappa=3/2$, and the moduli space $M(X,P)$ of anti-self-dual connections has formal dimension \[ 8\kappa-3\bigl(1+b^+(X)\bigr)=0. \] The corresponding Donaldson invariant, which counts the points of a regular moduli space $M(X,P)$ with signs, is \begin{equation}\label{eq:odd} D_X^w(1)=\pm1. \end{equation} This is the standard calculation for the K3 surface \cite[Proposition~7.13]{KM}. We shall use only the fact that the number of points is odd. \subsection{Flat connections on \texorpdfstring{$Y$}{Y}} The restriction of $w$ to $N$ is the generator $z$ of $H^2(N;\Z)\cong\Z$ with $z[x]=1$, and its restriction to $Y$ is the non-zero element of $H^2(Y;\Z)=\Z/2$. The flat connections in $P|_Y$, modulo gauge transformations of determinant one, therefore correspond to the conjugacy classes of homomorphisms $\pi_1(Y)=O^*\to U(2)$ whose determinant is the non-trivial character $\epsilon\colon O^*\to\{\pm1\}$. \begin{lemma}\label{lem:flat} \textup{(a)} There are two such conjugacy classes. They are represented by $\alpha=1\oplus\epsilon$ and by $\beta$, the pull-back of the irreducible two-dimensional representation of the symmetric group $S_3$ by the quotient maps $O^*\to S_4\to S_3$. \textup{(b)} The rho-invariants of the corresponding flat connections in $P|_Y$ are \[ \rho_Y(\alpha)=-4,\qquad\rho_Y(\beta)=-\frac{17}{3}. \] \textup{(c)} Their Chern--Simons invariants satisfy \[ \cs(\beta)-\cs(\alpha)=-\frac13\pmod\Z. \] \end{lemma} Here $\rho_Y(\gamma)$ is the Atiyah--Patodi--Singer rho-invariant of the flat bundle $\ad\gamma$, and $\cs$ is the Chern--Simons invariant with values in $\R/\Z$. Our sign conventions for both are fixed by the way they enter the formulae \eqref{eq:aps} and \eqref{eq:residues} below. The proof of the lemma is a calculation with the character table of $O^*$, and is given in Section~\ref{subsec:boundary}. The flat connection $\alpha$ is reducible: because the characters $1$ and $\epsilon$ are distinct, its stabilizer in $\G/\{\pm1\}$ is a circle. The flat connection $\beta$ is irreducible. Both are non-degenerate, in that \[ H^1(Y;\ad\alpha)=H^1(Y;\ad\beta)=0, \] because $\pi_1(Y)$ is finite. We write $h_\gamma$ for the dimension of the stabilizer of $\gamma$, so $h_\alpha=1$ and $h_\beta=0$. \begin{remark}\label{rem:O2} The holonomy group of the $SO(3)$ connection $\ad\alpha$ has order two, so its centralizer in $SO(3)$ is a copy of $O(2)$. The elements in the non-identity component of $O(2)$ are automorphisms of $P|_Y$ which do not lift to automorphisms of $\mathcal{E}|_Y$ of determinant one, and they play no role in what follows. \end{remark} \subsection{Moduli spaces on \texorpdfstring{$N$}{N} and \texorpdfstring{$C$}{C}}\label{subsec:moduli} Let $N^+$ and $C^+$ be the manifolds obtained from $N$ and $C$ by attaching cylindrical ends $[0,\infty)\times Y$ and $(-\infty,0]\times Y$, with metrics which are products on the ends. For $T>0$, let $X_T$ be the Riemannian manifold \[ X_T=N\cup\bigl([-T,T]\times Y\bigr)\cup C, \] which is a copy of $X$ containing a neck of length $2T$. Let $V$ be either $N$ or $C$, and let $\gamma$ be either $\alpha$ or $\beta$. We write $M_\kappa(V,\gamma)$ for the moduli space of anti-self-dual connections in $P$ on $V^+$ which are asymptotic to $\gamma$ on the end and have action $\kappa$. By the Atiyah--Patodi--Singer index theorem, the formal dimension of this moduli space is \begin{equation}\label{eq:aps} d=8\kappa-\frac32\bigl(\chi(V)+\sigma(V)\bigr)-\frac12h_\gamma +\frac12\rho_{\partial V}(\gamma), \end{equation} where $\rho_{\partial N}=\rho_Y$ and $\rho_{\partial C}=-\rho_Y$. On $M_\kappa(V,\gamma)$, the action $\kappa$ is determined modulo $\Z$ by $\gamma$. To find these residues, we consider reducible connections on $N$. A reduction of $P|_N$ to the circle has the form $P=\R\oplus L$, where $L$ is a complex line bundle with $c_1(L)\equiv w\pmod2$, so that $c_1(L)=kz$ with $k$ odd. Because $b^+(N)=0$ and $H_1(N;\Z)=0$, each such line bundle carries an anti-self-dual connection on $N^+$ with curvature in $L^2$, unique up to gauge. Since $z^2=-1/2$, the corresponding connection in $P$ has action \[ \kappa=-\tfrac14c_1(L)^2=k^2/8. \] The reductions for $k$ and $-k$ give the same connection in $P$. So the reducible anti-self-dual connections on $N^+$ have action $1/8$, $9/8$, $25/8$ and so on, and there is one of each. We write $A_0$ for the reducible connection with action $1/8$. The limit of $A_0$ on the end is $\alpha$, because $\alpha$ is the only reducible flat connection. So $\kappa\equiv1/8$ on $M_\kappa(N,\alpha)$, and from Lemma~\ref{lem:flat}(c) we obtain \begin{equation}\label{eq:residues} \kappa\equiv\frac18\ \text{ on } M_\kappa(N,\alpha), \qquad \kappa\equiv\frac18-\frac13=\frac{19}{24}\ \text{ on } M_\kappa(N,\beta) \pmod\Z. \end{equation} The residues on $C$ are obtained by subtracting these from $3/2$. In the following table we list the pairs of non-negative values $\kappa_N$ and $\kappa_C$ allowed by \eqref{eq:residues} with $\kappa_N+\kappa_C=3/2$, together with the formal dimensions $d_N$ and $d_C$ of $M_{\kappa_N}(N,\gamma)$ and $M_{\kappa_C}(C,\gamma)$ given by \eqref{eq:aps}, \eqref{eq:chisig} and Lemma~\ref{lem:flat}(b). \begin{equation}\label{eq:table} \begin{array}{c|rr|rr} \gamma&\kappa_N&d_N&\kappa_C&d_C\\\hline \rule{0pt}{2.6ex}\alpha&1/8&-3&11/8&2\\ \alpha&9/8&5&3/8&-6\\ \beta&19/24&2&17/24&-2 \end{array} \end{equation} In each row, $d_N+d_C+h_\gamma$ is zero, as it must be because $M(X,P)$ has formal dimension zero. The connection $A_0$ belongs to a moduli space of formal dimension $-3$. Its deformation complex decomposes according to the decomposition $\ad P=\R\oplus L$. The summand with coefficients in $\R$ has $H^0=\R$, the Lie algebra of the stabilizer of $A_0$, and its other cohomology groups are zero because $b_1(N)$ and $b^+(N)$ are zero. The summand with coefficients in $L$ is a complex of complex vector spaces, with $H^0=0$, and its index is $-1$. So $H^2_{A_0}$ has complex dimension at least one, with equality if and only if $H^1_{A_0}=0$. In the next proposition, the moduli spaces are defined using perturbations of the anti-self-duality equations on $N^+$ and $C^+$, which are of the type described in Section~\ref{subsec:perturb} and are supported in compact subsets of the interiors of $N$ and $C$. The same perturbations are used on $X_T$ to define $M(X_T,P)$. Perturbing the equations does not affect the formal dimensions or the residues of $\kappa$. If the perturbations are small, then there is still exactly one reducible solution on $N^+$ with action $1/8$, and we continue to call it $A_0$. \begin{proposition}\label{prop:moduli} The perturbations can be chosen so that the following hold. \begin{enumerate} \item[\textup{(a)}] The moduli space $M_{1/8}(N,\alpha)$ consists of the single point $[A_0]$. The cohomology group $H^1_{A_0}$ is zero, and $H^2_{A_0}$ is isomorphic to $\C$, with the stabilizer of $A_0$ acting with weight one. \item[\textup{(b)}] The moduli spaces $M_{3/8}(C,\alpha)$ and $M_{17/24}(C,\beta)$ are empty. The moduli space \[ M=M_{11/8}(C,\alpha) \] is a compact smooth two-manifold. It consists of irreducible connections $B$ with $H^2_B=0$. \item[\textup{(c)}] Let $T_i$ be a sequence tending to infinity, and let $[A_i]$ belong to $M(X_{T_i},P)$. Then, after passing to a subsequence, there is a point $[B]$ in $M$ such that the connections $A_i$ converge to $A_0$ on compact subsets of $N^+$ and to $B$ on compact subsets of $C^+$. No action is lost in the limit: $\kappa(A_0)+\kappa(B)=3/2$. \end{enumerate} \end{proposition} \begin{proof} In Section~\ref{subsec:perturb} we show that the perturbations can be chosen so that the following hold. First, the connection $A_0$ has $H^1_{A_0}=0$ and $H^2_{A_0}\cong\C$, and the irreducible parts of the moduli spaces on $N^+$ are regular (Lemma~\ref{lem:normal}). Second, the moduli spaces on $C^+$ with $\kappa\le 3/2$ contain no reducibles and are regular (Lemma~\ref{lem:complement}). Third, the action of every solution on $N^+$ or $C^+$ is greater than $-5/24$; since the residues of $\kappa$ are $1/8$ and $19/24$ on $N$, and $3/8$ and $17/24$ on $C$, the action is then non-negative. A regular moduli space of negative formal dimension is empty. So the irreducible part of $M_{1/8}(N,\alpha)$ is empty, which proves \textup{(a)}, and the two moduli spaces on $C^+$ in the second and third rows of \eqref{eq:table} are empty also. The moduli space $M$ is a smooth two-manifold. To see that it is compact, recall that a sequence in $M$ has a subsequence which converges weakly to a point of some $M_{\kappa'}(C,\gamma)$ with $\kappa'\le11/8$. If $\kappa'$ is less than $11/8$, then the difference is accounted for by bubbles, each of which carries action at least $1$, and by anti-self-dual connections on the cylinder $\R\times Y$. The only possibilities allowed by the residues are $M_{3/8}(C,\alpha)$ and $M_{17/24}(C,\beta)$, and these are empty. The proof of \textup{(c)} is similar. A sequence $[A_i]$ as in \textup{(c)} has a subsequence which converges weakly to a pair of solutions on $N^+$ and $C^+$, belonging to moduli spaces $M_{\kappa_N}(N,\gamma)$ and $M_{\kappa_C}(C,\gamma')$, where $\kappa_N$ and $\kappa_C$ are non-negative and $\kappa_N+\kappa_C\le3/2$. Equality holds only if there are no bubbles and no action is lost in the neck, in which case $\gamma=\gamma'$. Among the moduli spaces on $C^+$ with $\kappa_C\le 3/2$, only $M$ can be non-empty. So $\kappa_C=11/8$ and $\kappa_N\le1/8$. Since $\kappa_N$ is non-negative and congruent to $1/8$ or $19/24$ modulo $\Z$, it is equal to $1/8$, the limiting flat connection is $\alpha$, and no action is lost. By \textup{(a)}, the limit on $N^+$ is $A_0$. \end{proof} \subsection{Proof of Lemma~\ref{lem:flat}}\label{subsec:boundary} We refer to \cite[Appendix~A.4]{Helle} for the character table of $O^*$. The group has three irreducible two-dimensional representations: the defining representation $Q$ of $O^*\subset SU(2)$, the tensor product $Q\otimes\epsilon$, and $\beta$. The first two have determinant $1$, and $\beta$ has determinant $\epsilon$. A reducible representation with determinant $\epsilon$ is the sum of two characters whose product is $\epsilon$, and so is $1\oplus\epsilon$. This proves \textup{(a)}. For \textup{(b)}, let $\pm\theta_g$ be the arguments of the eigenvalues of $g\in O^*$ in the representation $Q$. With $Y$ oriented as the boundary of the negative-definite plumbing, the rho-invariant is given by \[ \rho_Y(\gamma)=\frac1{48}\sum_{g\ne1} \cot^2\Bigl(\frac{\theta_g}{2}\Bigr) \bigl(\tr\ad\gamma(g)-3\bigr). \] The representations $\ad\alpha$ and $\ad\beta$ factor through the quotient $S_4$ of $O^*$, and each element of $S_4$ is the image of two elements $\pm g$ of $O^*$. In the following table, the third column is the sum of $\cot^2(\theta_g/2)$ over the two elements of $O^*$ lying over an element of the given conjugacy class in $S_4$. \[ \begin{array}{c|c|c|cc} \text{class in }S_4&\text{size}& \sum\cot^2(\theta_g/2)& \tr\ad\alpha-3&\tr\ad\beta-3\\\hline \rule{0pt}{2.6ex}(12)&6&2&-4&-4\\ (123)&8&10/3&0&-3\\ (1234)&6&6&-4&-4 \end{array} \] The remaining conjugacy classes in $S_4$ make no contribution, because $\ad\alpha$ and $\ad\beta$ are trivial there. So \[ \rho_Y(\alpha)=\frac{-48-144}{48}=-4, \qquad \rho_Y(\beta)=\frac{-48-80-144}{48}=-\frac{17}{3}. \] For \textup{(c)}, recall that $H^4(BO^*;\Z)$ is cyclic of order $48$, generated by $e=c_2(Q)$. We have $c_2(\alpha)=0$. To find $c_2(\beta)$ we restrict to subgroups. On a Sylow $2$-subgroup of $O^*$, which is a generalized quaternion group of order $16$, the representation $\beta$ is the sum of the trivial character and a non-trivial one, so $c_2(\beta)$ is zero modulo $16$. On a cyclic subgroup of order $3$, the representation $\beta$ coincides with $Q$, so $c_2(\beta)$ is equal to $e$ modulo $3$. Hence $c_2(\beta)=16e$. With the conventions of \cite[Appendix~B]{Helle}, the Chern--Simons invariant of $Q$ is $-1/48$. Since $\alpha$ and $\beta$ have the same determinant, it follows that \[ \cs(\beta)-\cs(\alpha)=-\frac{16}{48}=-\frac13\pmod\Z. \] In \eqref{eq:residues}, we have used the fact that if two connections on $N^+$ are asymptotic to $\gamma$ and $\gamma'$, then the difference of their actions is congruent to $\cs(\gamma)-\cs(\gamma')$ modulo~$\Z$. For this, the Chern--Simons invariant is defined using the form $\zeta$ with $d\zeta=(8\pi^2)^{-1}\tr(F\wedge F)$, which is the normalization for which an anti-self-dual connection on a closed manifold has action $\kappa=-p_1/4$. \section{The parity argument}\label{sec:parity} We continue with the notation of Section~\ref{sec:moduli}. The gluing theorem of Section~\ref{sec:analytic} leads to the following description of the moduli space on $X_T$ when $T$ is large. We write $\B^*(X_T,P)$ for the space of irreducible connections in $P$ modulo the gauge group $\G$, and $\tau$ for the perturbation of the equations on $X_T$, so that $M(X_T,P)$ is the set of solutions of $F_A^++\tau(A)=0$ in $\B^*(X_T,P)$. \begin{proposition}\label{prop:gluing} If $T$ is sufficiently large then, after the addition of a further small perturbation to $\tau$, the following hold. \begin{enumerate} \item[\textup{(a)}] There are an open subset $U$ of $\B^*(X_T,P)$, an oriented real two-plane bundle $E\to U$, and a smooth bundle map $\iota$ from $E$ to the bundle over $U$ whose fibre at $[A]$ is $\Omega^+(X_T;\ad P)$, such that the space \[ \Sigma=\bigl\{\,([A],v) : [A]\in U,\ v\in E_{[A]},\ F_A^++\tau(A)+\iota_A(v)=0\,\bigr\} \] is cut out transversely and is a compact surface, diffeomorphic to $M$. \item[\textup{(b)}] The moduli space $M(X_T,P)$ is regular and is contained in $U$. So $M(X_T,P)$ is the zero set of the section $s$ of $E|_\Sigma$ given by $s([A],v)=v$, and $s$ vanishes transversely. The number of its zeros is odd. \end{enumerate} \end{proposition} The surface $\Sigma$ is a thickened moduli space, of the kind which arises whenever a moduli space is described by a finite-dimensional obstruction bundle. To say that it is cut out transversely means the following. Write $\sigma(A)=F_A^++\tau(A)$. At a point $([A],v)$ of $\Sigma$, the linearization of the equation $\sigma(A)+\iota_A(v)=0$, together with the Coulomb gauge condition, is the operator \begin{equation}\label{eq:augmented} \begin{gathered} \mathcal{D}_{(A,v)}\colon\Omega^1(X_T;\ad P)\oplus E_{[A]} \longrightarrow\Omega^0(X_T;\ad P)\oplus\Omega^+(X_T;\ad P),\\ (a,u)\longmapsto\bigl(d_A^*a,\ (d\sigma)_A(a)+(\partial_a\iota)(v) +\iota_A(u)\bigr), \end{gathered} \end{equation} and we require this operator to be surjective. Its kernel is then the tangent space to $\Sigma$ at $([A],v)$. The last statement in part \textup{(b)} holds because the number of points in the moduli space, counted with signs, is the Donaldson invariant \eqref{eq:odd}. On the other hand, the number of zeros of a transverse section of $E|_\Sigma$ is given, modulo~$2$, by the second Stiefel--Whitney class. So Proposition~\ref{prop:gluing} has the following consequence, in which $[\Sigma]$ is the fundamental class with coefficients $\F_2$: \begin{equation}\label{eq:count} \bigl\langle w_2(E),[\Sigma]\bigr\rangle=1. \end{equation} Because $\G/\{\pm1\}$ acts freely on the irreducible connections, there is a universal $SO(3)$ bundle over $\B^*(X_T,P)\times X_T$. Its restriction to $\Sigma\times X_T$ carries a family of connections parametrized by $\Sigma$. So the operators \[ D_A=d_A^*\oplus d_A^+\colon\Omega^1(X_T;\ad P)\longrightarrow \Omega^0(X_T;\ad P)\oplus\Omega^+(X_T;\ad P), \] for $[A]$ in $\Sigma$, form a family of elliptic operators over $\Sigma$, which has an index $\ind D$ in $KO(\Sigma)$. \begin{lemma}\label{lem:index} In $KO(\Sigma)$, we have \begin{equation}\label{eq:ko} [T\Sigma]-[E|_\Sigma]=\ind D. \end{equation} \end{lemma} \begin{proof} The operators $\mathcal{D}_{(A,v)}$ in \eqref{eq:augmented} form a family of surjective Fredholm operators over $\Sigma$, and the kernels form the bundle $T\Sigma$. So the index of this family is $[T\Sigma]$. On the other hand, $\mathcal{D}_{(A,v)}$ differs from the operator $(a,u)\mapsto D_Aa$ by terms which are compact: the derivative of the perturbation $\tau$ is compact, and the terms involving $\iota$ have finite rank. The index of the family $(a,u)\mapsto D_Aa$ is $\ind D+[E|_\Sigma]$. \end{proof} \begin{proposition}\label{prop:ksp} Let $X$ be a K3 surface with any Riemannian metric, let $B$ be a compact space, and let $\mathbf{P}\to B\times X$ be an $SO(3)$ bundle carrying a family of connections on $X$ parametrized by $B$. Then the index of the family of operators $D_A=d_A^*\oplus d_A^+$ lies in the image of the map $KSp(B)\to KO(B)$. In particular, its first and second Stiefel--Whitney classes are zero. \end{proposition} \begin{proof} The index of the family does not depend on the Riemannian metric, so we may suppose that the metric is hyperk\"ahler. The spin bundles $S^+$ and $S^-$ of $X$ are quaternionic line bundles: we regard their fibres as right $\HH$-modules, and Clifford multiplication by one-forms and the Dirac operator are then $\HH$-linear. For the hyperk\"ahler metric, the bundle $S^+$ is flat and trivial. Let $\phi$ be a parallel section of $S^+$ of unit length. Clifford multiplication on $\phi$ gives isomorphisms of real vector bundles \[ \Lambda^1\longrightarrow S^-,\qquad \Lambda^0\oplus\Lambda^+\longrightarrow S^+, \] and because $\phi$ is parallel, these isomorphisms carry the operator $d^*\oplus d^+$ to the Dirac operator $\Gamma(S^-)\to\Gamma(S^+)$, up to constant factors. In the same way, the operator $D_A$ becomes the Dirac operator coupled to the real vector bundle $\ad P$ with the connection $A$. This coupled Dirac operator is again $\HH$-linear. The quaternionic structure is the same for every member of the family, so the index of the family is the class of a virtual quaternionic vector bundle on $B$. Finally, the real vector bundle which underlies a quaternionic vector bundle has $w_1=0$ and $w_2=0$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:octahedral}] By Proposition~\ref{prop:alternatives}, we have only to exclude the case considered in Section~\ref{sec:moduli}, in which $Q_N$ is $\langle-2\rangle$. In that case, let $\Sigma$ and $E$ be as in Proposition~\ref{prop:gluing}. By Proposition~\ref{prop:ksp}, the class $\ind D$ in \eqref{eq:ko} has $w_1=0$ and $w_2=0$, and so the Stiefel--Whitney classes of $T\Sigma$ and $E|_\Sigma$ are equal. Because $E$ is oriented, it follows that $\Sigma$ is orientable and that \[ \bigl\langle w_2(E),[\Sigma]\bigr\rangle =\bigl\langle w_2(T\Sigma),[\Sigma]\bigr\rangle. \] The right-hand side is the Euler number of $\Sigma$ modulo~$2$, which is zero because $\Sigma$ is a closed orientable surface. This contradicts \eqref{eq:count}. (If $M$ is empty, then $\Sigma$ is empty and \eqref{eq:count} fails for that reason.) So $Q_W$ is isomorphic to $-E_7$, and the values of $b^+(C)$ and $b^-(C)$ follow. \end{proof} \section{Regularity and gluing}\label{sec:analytic} In this section we prove the transversality statements which were used in the proof of Proposition~\ref{prop:moduli}, and we prove Proposition~\ref{prop:gluing}. The methods are not new. The technique for gluing in the presence of an obstruction, in which the equations are first solved modulo a finite-dimensional space associated with the small eigenvalues of the linearized operator, goes back to Taubes \cite{TauIndef}; see also \cite{Don86} and \cite[Section~7.2]{DK}. For manifolds with cylindrical ends, we follow \cite[Chapter~4]{Don}. Some constants in our argument grow polynomially with the length of the neck, while the error terms decay exponentially, and to deal with this one needs a slice theorem in which the constants are controlled by the lowest eigenvalue of the Laplacian $d_A^*d_A$. Results of that kind are also due to Taubes: see \cite[Section~6]{TauFramework} and \cite[Lemma~A.1]{TauPath}, and the account in \cite[Theorems~1.1 and~1.2]{Fee}. The small eigenvalues of Dirac-type operators on manifolds with long necks, and the obstruction bundles to which they give rise in families, are discussed in \cite{Nic}. The perturbations of the equations which we use are the holonomy perturbations of \cite{Kro}. What is particular to the present situation is that the ends are rational homology spheres, on which the flat connections may be reducible. Moduli spaces of this sort are studied in detail in \cite{ME} and \cite{DME}, and we refer to those papers at several points. \subsection{Perturbations and regularity}\label{subsec:perturb} Let $V$ be an oriented Riemannian four-manifold with a cylindrical end $[0,\infty)\times Y$, where $Y$ is a rational homology sphere, and let $P\to V$ be an $SO(3)$ bundle, with the conventions of Section~\ref{sec:moduli}. For a non-degenerate flat connection $\gamma$ on the end, the moduli spaces $M_\kappa(V,\gamma)$ are defined in the standard way \cite[Chapter~4]{Don}, using Sobolev spaces $L^2_k$ with $k\ge4$ and a small exponential weight on the end. For a solution $A$, we write $H^0_A$, $H^1_A$ and $H^2_A$ for the cohomology groups of its deformation complex, so that $H^0_A$ is the Lie algebra of the stabilizer of $A$ and $H^2_A$ is the cokernel of the linearized equations. The solution is \emph{regular} if $H^2_A=0$. The elements of $H^2_A$ are represented by self-dual forms which decay exponentially on the end. The equations are perturbed using holonomy perturbations, as in \cite[Section~3]{Kro}. In place of $F_A^+=0$ we consider \[ \sigma(A)=F_A^++\tau(A)=0, \] where $\tau$ is a sum of terms of the following kind. One takes a ball $B$ in the interior of $V$, a map $q\colon S^1\times B\to V$ with $q(1,x)=x$, and a self-dual form $\omega$ supported in $B$. The corresponding term is obtained from the holonomy of $A$ around the loops $q(\cdot,x)$, for $x$ in $B$, by multiplying by $\omega$ and projecting to $\Omega^+(V;\ad P)$. The perturbations form a Banach space, they are supported in a compact subset of the interior of $V$, and they are gauge-equivariant: if $A$ is compatible with a reduction $P=\R\oplus L$, then $\tau(A)$ takes values in the summand $\R$. Their use on manifolds with cylindrical ends, in the presence of reducible connections, is described in \cite[Section~4]{ME} and \cite[Section~3.2]{DME}. In \cite{Kro}, the map $q$ is required to be a submersion. We require more, namely that each map $q(z,\cdot)$ is an embedding of $B$. The derivative of $\tau$ is then built from pull-backs by diffeomorphisms, so that it and its formal adjoint are bounded on Sobolev spaces of every order, negative orders included, which we need in the gluing argument. The usual transversality arguments apply without change to perturbations of this restricted kind. A solution of the perturbed equations with action $\kappa$ satisfies \begin{equation}\label{eq:energy} \kappa=c_0\bigl(\|F_A^-\|^2_{L^2}-\|F_A^+\|^2_{L^2}\bigr) \ge-c_0\|\tau(A)\|^2_{L^2}, \end{equation} where $c_0$ is a positive constant which depends only on the normalization of $\kappa$. The $L^2$ norm of $\tau(A)$ is bounded independently of $A$, by a constant which is small when $\tau$ is small. So for any $\varepsilon>0$, all solutions have $\kappa>-\varepsilon$ when $\tau$ is sufficiently small. If a solution $A$ is compatible with a reduction $P=\R\oplus L$, then the linearization of the equations at $A$ in Coulomb gauge is the sum of a \emph{tangential} operator, which acts on ordinary one-forms, and a \emph{normal} operator $D^\nu_A$, which acts on one-forms with values in $L$ and is complex-linear. The circle in the stabilizer of $A$ acts on $L$ with weight one. \begin{lemma}\label{lem:normal} Suppose that $H_1(V;\Z)=0$ and $b^+(V)=0$. Let $\alpha$ be a flat connection on the end with $H^1(Y;\ad\alpha)=0$ whose stabilizer is a circle. Let $P=\R\oplus L$ be a reduction of $P$ which is compatible with $\alpha$ on the end, for which $c_1(L)$ is non-zero in $H^2(V;\Q)$ and for which the normal operators have complex index $-1$. Then for all sufficiently small perturbations $\tau$, there is exactly one solution $A_0$ compatible with this reduction, up to gauge. For a residual set of small perturbations, the normal operator $D^\nu_{A_0}$ is injective, so that \[ H^1_{A_0}=0,\qquad H^2_{A_0}\cong\C, \] and the irreducible solutions in the moduli spaces $M_\kappa(V,\gamma)$ are regular, for all $\gamma$ and for all $\kappa$ below any given bound. \end{lemma} \begin{proof} When $\tau=0$, the solutions compatible with the reduction are the anti-self-dual connections in $L$ with curvature in $L^2$, and there is exactly one of these up to gauge, because $b^+(V)=0$ and $H_1(V;\Z)=0$. The tangential operator at this solution has kernel $H^1(V;\R)=0$, and its cokernel consists only of the constants. So the solution persists when $\tau$ is small, by the implicit function theorem. No others appear, because solutions in a line bundle cannot bubble and cannot lose action on the end; compare \cite[Section~4.4]{ME}. The normal operators $D^\nu_\tau$ at these solutions are complex Fredholm operators of index $-1$. Among such operators, those whose kernel has complex dimension $j$ form a submanifold of real codimension $2j(j+1)$, whose normal space at an operator with kernel $K$ and cokernel $H$ is $\operatorname{Hom}_\C(K,H)$. Let $K$ and $H$ be the kernel and cokernel of $D^\nu_0$. It is enough to show that the derivative of $\tau\mapsto D^\nu_\tau$ at $\tau=0$ maps onto $\operatorname{Hom}_\C(K,H)$. For surjectivity onto a finite-dimensional space is an open condition, so the map $\tau\mapsto D^\nu_\tau$ is then transverse to all these submanifolds for $\tau$ near zero, and by the Sard--Smale theorem $D^\nu_\tau$ is injective for a residual set of small $\tau$. The surjectivity is proved by the argument which is used in \cite[Section~4.6]{ME} and \cite[Section~3.5]{DME} to obtain transversality in the directions normal to the reducible connections. A non-zero element of $K$ is detected by the derivative of the holonomy around some loop in a compact subset of $V$, because the flat line bundle $L|_Y$ has no parallel sections. A non-zero element of $H$ cannot vanish on an open set. And the only constraint on the derivative of a holonomy perturbation in the directions normal to the reducible connections is equivariance under the circle, which for a real-linear map from $K$ to $H$ is the same as complex-linearity. The perturbations can be taken to vanish to first order at $A_0$ along the reducible locus, so that $A_0$ does not move. The regularity of the irreducible solutions for generic $\tau$ is proved as in \cite[Sections~3 and~5]{Kro}; see also \cite[Section~4.6]{ME}. \end{proof} \begin{lemma}\label{lem:complement} Suppose that $H_1(V;\Z)=0$ and $b^+(V)>0$, and that $w_2(P)$ is non-zero on the end. Then for a residual set of small perturbations $\tau$, the moduli spaces $M_\kappa(V,\gamma)$ contain no reducible solutions and are regular, for all $\gamma$ and for all $\kappa$ below any given bound. \end{lemma} \begin{proof} Since $H^1(V;\F_2)=0$, every reducible connection is compatible with a reduction $P=\R\oplus L$. The class $c_1(L)$ reduces modulo~$2$ to $w_2(P)$, so it is non-zero on the end; and it is non-zero in $H^2(V;\Q)$, because $H^2(V;\Z)$ has no torsion. The solutions compatible with such a reduction are the zeros of a Fredholm section of index $b_1(V)-b^+(V)$, which is negative, so for generic $\tau$ there are none; compare \cite[Section~4.4]{ME}. The regularity of the irreducible solutions is proved as before. \end{proof} These lemmas supply what was needed in the proof of Proposition~\ref{prop:moduli}. Lemma~\ref{lem:normal} applies to the manifold $N^+$ of Section~\ref{sec:moduli}, with the reduction for which $c_1(L)=z$: the formal dimension $-3$ of $M_{1/8}(N,\alpha)$ is the sum of $-1$, from the stabilizer, and twice the complex index of the normal operator. Lemma~\ref{lem:complement} applies to $C^+$, where $b_1-b^+=-3$. The lower bound for the action comes from \eqref{eq:energy}, with $\varepsilon=5/24$. \subsection{The gluing theorem}\label{subsec:gluing} Let $X$ be a closed, oriented four-manifold, decomposed as $N\cup_Y C$ along a rational homology sphere $Y$, and let $P\to X$ be an $SO(3)$ bundle, with the same conventions as before. We form $N^+$, $C^+$ and $X_T$ as in Section~\ref{subsec:moduli}. We fix holonomy perturbations on $N^+$ and $C^+$, and we use their sum on $X_T$. Let $\alpha$ be a flat connection on $P|_Y$ with $H^1(Y;\ad\alpha)=0$ whose stabilizer $\Gamma_\alpha$ is a circle. We make the following hypotheses. \begin{enumerate} \item[(i)] $A_0$ is a reducible solution on $N^+$, asymptotic to $\alpha$, with $H^1_{A_0}=0$. The space $H=H^2_{A_0}$ is isomorphic to $\C$, and the stabilizer $\Gamma_{A_0}$, which is a circle, acts on it with weight one. \item[(ii)] $M$ is a compact two-manifold which is a union of components of the moduli space of solutions on $C^+$ asymptotic to $\alpha$. It consists of irreducible, regular solutions. \item[(iii)] If $T_i$ is a sequence tending to infinity, and $[A_i]$ belongs to $M(X_{T_i},P)$, then after passing to a subsequence there is a point $[B]$ in $M$ such that the connections $A_i$ converge to $A_0$ on compact subsets of $N^+$ and to $B$ on compact subsets of $C^+$, without loss of action. \end{enumerate} Restriction to the end identifies $\Gamma_{A_0}$ with $\Gamma_\alpha$. Let $\widetilde M\to M$ be the moduli space of solutions framed at infinity, which is a principal $\Gamma_\alpha$-bundle over $M$, and let \[ E_0=\widetilde M\times_{\Gamma_\alpha}H \] be the associated complex line bundle. \begin{theorem}\label{thm:gluing} Suppose that hypotheses \textup{(i)--(iii)} hold. Then there are constants $T_0$ and $K$ such that the following is true for all $T\ge T_0$. Let $\nu$ be a holonomy perturbation on $X_T$, supported in the complement of the neck, whose norm is at most $e^{-KT}$. Then assertion~\textup{(a)} of Proposition~\ref{prop:gluing} holds for the equations on $X_T$ defined by $\tau+\nu$, with a diffeomorphism $M\to\Sigma$ which is covered by an isomorphism of oriented bundles $E_0\to E|_\Sigma$; and the moduli space $M(X_T,P)$ is contained in $U$. For a residual set of such $\nu$, the moduli space $M(X_T,P)$ is regular. \end{theorem} \begin{remark}\label{rem:parameter} The gluing parameter for the pair $(A_0,B)$ is an element of $\Gamma_\alpha$. Because $\Gamma_{A_0}$ maps isomorphically to $\Gamma_\alpha$, the gluing parameter is absorbed by the action of $\Gamma_{A_0}$, which is why $\Sigma$ is a copy of $M$ rather than of $\widetilde M$. In the situation of Remark~\ref{rem:O2}, there is a second way to identify the two copies of $\ad\alpha$ over the neck, using the non-identity component of $O(2)$. The bundle which results is not $P$: its second Stiefel--Whitney class differs from $w_2(P)$ by the image of the non-zero element of $H^1(Y;\F_2)$ under the coboundary map to $H^2(X;\F_2)$, and this image is not zero because $H^1(N;\F_2)$ and $H^1(C;\F_2)$ are zero. \end{remark} \subsection{Proof of Theorem~\ref{thm:gluing}} We first take $\nu=0$. \subsubsection*{Approximate solutions} Because $\alpha$ is non-degenerate, the connection $A_0$ approaches $\alpha$ exponentially on the end of $N^+$, and so do the solutions $B$ representing the points $m$ of $M$ on the end of $C^+$, uniformly in $m$. We use cut-off functions to make $A_0$ and $B$ equal to $\alpha$ at distance greater than $T$ along the ends, and we join the results across the neck of $X_T$. This requires an identification of the two copies of $\alpha$. The possible identifications form a copy of $\Gamma_\alpha$, and because $\Gamma_{A_0}$ maps onto $\Gamma_\alpha$, they all give gauge-equivalent connections. So we obtain connections $A_T(m)$, whose gauge-equivalence classes depend only on $m$, with \begin{equation}\label{eq:preglue} \|\sigma(A_T(m))\|_{L^2_{k-1}}\le Ce^{-\lambda T}, \end{equation} and with similar bounds for the derivatives with respect to $m$. Here and below, $C$, $\lambda$ and $q$ denote positive constants which are independent of $T$ and $m$. The manifolds $X_T$ have uniformly bounded geometry, so the constants in the Sobolev and elliptic estimates do not depend on $T$. In the same way, the self-dual forms on $N^+$ which represent the elements of $H$ can be cut off and transferred to $X_T$. The result depends on the identification, through the action of $\Gamma_{A_0}$ on $H$, and we obtain linear maps \[ \iota_m\colon(E_0)_m\longrightarrow\Omega^+(X_T;\ad P). \] \subsubsection*{The linearized equations} Let $\delta_m\colon\Omega^1(X_T;\ad P)\to\Omega^+(X_T;\ad P)$ be the derivative of $\sigma$ at $A_T(m)$. \begin{lemma}\label{lem:spectral} There is a constant $c>0$ such that, for all $m$ and all sufficiently large $T$, the operator $\delta_m\delta_m^*$ has exactly two eigenvalues in the interval $[0,Ce^{-\lambda T}]$, counted with multiplicity, and no other spectrum in $[0,c]$. If $\Pi_m$ is the projection onto the span of the two eigenvectors, then $\Pi_m\circ\iota_m$ is an isomorphism from $(E_0)_m$ to the image of $\Pi_m$. \end{lemma} \begin{proof} The forms $\iota_m(v)$ are annihilated by $\delta_m^*$ up to an error of size $Ce^{-\lambda T}$, which accounts for two small eigenvalues and for the last statement. If the lemma were false, there would be self-dual forms $u_i$ on $X_{T_i}$, with $T_i\to\infty$, which have $L^2$ norm $1$, are orthogonal to the image of $\iota_{m_i}$, and have $\delta_{m_i}^*u_i\to0$ in $L^2$. On the neck, where $A_T(m)$ is the pull-back of $\alpha$, we identify self-dual forms with time-dependent one-forms on $Y$ by $\eta\mapsto dt\wedge\eta+*_Y\eta$, and $\delta\delta^*$ becomes the operator \begin{equation}\label{eq:neckoperator} \tfrac12\bigl(-\partial_t^2+\Delta_\alpha\bigr), \end{equation} where $\Delta_\alpha$ is the Hodge Laplacian on $\Omega^1(Y;\ad\alpha)$. This Laplacian is bounded below by a positive constant, because $H^1(Y;\ad\alpha)=0$, so the $L^2$ norm of $u_i$ cannot escape into the neck. A subsequence therefore converges on compact subsets of $N^+$ and $C^+$ to $L^2$ solutions of the adjoint equations, not both zero. But the limit on $C^+$ is zero, because the points of $M$ are regular, and the limit on $N^+$ is an element of $H$ which is orthogonal to $H$. \end{proof} The space $H^0(Y;\ad\alpha)$, which is not zero, does not enter in \eqref{eq:neckoperator}. It does affect the Laplacian $d_A^*d_A$ on zero-forms, and so the Coulomb gauge condition. For $A=A_T(m)$, and for connections $A$ whose $L^2_k$ distance from $A_T(m)$ is at most $T^{-q}$, we have \begin{equation}\label{eq:scalar} \|\xi\|_{L^2}\le CT\,\|d_A\xi\|_{L^2},\qquad \xi\in\Omega^0(X_T;\ad P). \end{equation} This is obtained from a Poincar\'e inequality on a compact subset of $C$, where the connections representing the points of $M$ are irreducible, by integrating along the neck. So $d_A^*d_A$ has an inverse $G_A$, whose norms as an operator between Sobolev spaces are bounded by powers of $T$, and $P_A=1-d_AG_Ad_A^*$ is the projection onto the kernel of $d_A^*$. The powers of $T$ do no harm, because they multiply terms which are exponentially small. For example, if $A=A_T(m)$ and $R_m$ denotes the operator $\xi\mapsto[\sigma(A),\xi]$, then the gauge-invariance of $\sigma$ gives $\delta_md_A=R_m$, and so \begin{equation}\label{eq:coulomb} \delta_mP_A\delta_m^*=\delta_m\delta_m^*-R_mG_AR_m^*, \end{equation} where the last term has norm at most $CT^qe^{-2\lambda T}$, by \eqref{eq:preglue}. With Lemma~\ref{lem:spectral}, this shows that $(1-\Pi_m)\delta_m$ has a right inverse $S_m$ whose image lies in the kernel of $d_A^*$ and whose norm, as an operator from $L^2_{k-1}$ to $L^2_k$, is at most $CT^q$. \subsubsection*{The thickened moduli space} For $m$ in $M$, consider the equation \begin{equation}\label{eq:stabilized} \sigma\bigl(A_T(m)+a\bigr)+\iota_m(v)=0 \end{equation} for a one-form $a$ in the image of $S_m$ and an element $v$ of $(E_0)_m$. Its linearization at $(a,v)=(0,0)$ is the map $(a,v)\mapsto\delta_ma+\iota_m(v)$, which is invertible, with an inverse of norm at most $CT^q$: this follows from the properties of $S_m$ and from Lemma~\ref{lem:spectral}. The left-hand side of \eqref{eq:stabilized} is of size $Ce^{-\lambda T}$ at $(0,0)$, and the non-linear terms are quadratic, with constants bounded by powers of $T$. So by the contraction mapping theorem, there is a $Q$ such that, for large $T$, the equation has a unique solution $(a_m,v_m)$ with \begin{equation}\label{eq:ball} \|a\|_{L^2_k}+|v|\le T^{-Q}. \end{equation} The solution depends smoothly on $m$. The map $m\mapsto[A_T(m)]$ embeds $M$ in $\B^*(X_T,P)$, and up to an exponentially small error, the image of $S_m$ is a complement in the Coulomb slice at $A_T(m)$ to the tangent space of the embedded surface. So the connections $A_T(m)+a$, with $a$ in the image of $S_m$ and $\|a\|_{L^2_k}1$. So $d=1$, and there is a vector $v$ in $\Lambda$ with $u\cdot v=1$. After adding a multiple of $u$ to $v$, we may suppose that $v^2=0$, so that $u$ and $v$ span a hyperbolic plane in $\Lambda$. Now let $w_0$ be any class in $H^2(X;\Z)$ which restricts to $w_N$. If we identify $H^2(X;\Z)$ with $H_2(X;\Z)$, then the classes which restrict to zero on $N$ are the elements of $\Lambda$. So we may subtract a combination of $u$ and $v$ from $w_0$ to make it orthogonal to both, without changing its restriction to $N$. The square of $w_0$ is even, and the class \[ w=w_0+u-\tfrac12(6+w_0^2)\,v \] has the required properties. \end{proof} Let $P$ be the adjoint bundle of the $U(2)$ bundle $\mathcal{E}$ with $c_1(\mathcal{E})=w$ and $c_2(\mathcal{E})=0$. As in Section~\ref{sec:moduli}, we have $\kappa=3/2$, the moduli space $M(X,P)$ has formal dimension zero, and the Donaldson invariant is $\pm 1$. \subsection{Flat connections} The group $H^2(Y_n;\Z)$ is the group of characters of $\Gamma_m$, which has four elements. Let $\delta$ be the character with \[ \delta(a)=1,\qquad\delta(b)=-1, \] and let $\chi$ be the character which corresponds to the restriction of $w$ to $Y_n$. When $n$ is odd, the class $w_N$ generates $L_n^*/L_n\cong\Z/4$, so $\chi$ has order four and $\chi^2=\delta$. When $n=6$, the character $\chi$ is not trivial and is not $\delta$, because its self-linking number is non-zero, whereas $\delta$ corresponds to the non-zero element of $D_6^*/D_6$ which has self-linking number zero. In all three cases, therefore, \[ \chi(a)=-1, \] and the four characters of $\Gamma_m$ are $1$, $\delta$, $\chi$ and $\delta\chi$. \begin{lemma}\label{lem:dihedral-flat} \textup{(a)} Up to conjugacy, there are two homomorphisms $\Gamma_m\to U(2)$ with determinant $\chi$. They are \[ \alpha_0=1\oplus\chi,\qquad\alpha_1=\delta\oplus\delta\chi. \] \textup{(b)} The rho-invariants of the corresponding flat connections are \[ \rho_{Y_n}(\alpha_0)=\rho_{Y_n}(\alpha_1)=-m. \] \end{lemma} \begin{proof} The irreducible two-dimensional representations of $\Gamma_m$ are the representations $\rho_j$, for $1\le j\le m-1$, given by \[ \rho_j(a)=\begin{pmatrix}e^{\pi ij/m}&0\\0&e^{-\pi ij/m}\end{pmatrix}, \qquad \rho_j(b)=\begin{pmatrix}0&(-1)^j\\1&0\end{pmatrix}. \] See \cite[Appendix~A.2]{Helle}. The determinant of $\rho_j$ is $1$ or $\delta$, according as $j$ is odd or even, and is never $\chi$. A reducible representation with determinant $\chi$ is the sum of two characters whose product is $\chi$, and there are two such pairs. This proves \textup{(a)}. For \textup{(b)}, we use the formula for the rho-invariant from Section~\ref{subsec:boundary}, with $4m$ in place of $48$. For both $\alpha_0$ and $\alpha_1$, the character of the adjoint representation is $1+\chi+\overline\chi$. So the contribution of $a^j$ is zero when $j$ is even, and when $j$ is odd it is $-4\cot^2(\pi j/2m)$. Each of the $2m$ elements $ba^j$ has $\theta=\pi/2$, and the sum of $\tr\ad\alpha_i-3$ over these elements is $-4m$. Since \[ \sum_{\substack{0-\kappa_0$, which by \eqref{eq:dihedral-residues} implies that $\kappa\ge0$. The three parts of Proposition~\ref{prop:moduli} then hold, with the following changes. In \textup{(a)}, the moduli space is $M_{\kappa_0}(N,\alpha_0)$. In \textup{(b)}, the moduli spaces on $C^+$ in the second and third rows of \eqref{eq:dihedral-table} are empty, and the compact two-manifold is \[ M=M_{3/2-\kappa_0}(C,\alpha_0). \] The proofs are the same. For the compactness of $M$ and for part \textup{(c)}, we note that an anti-self-dual connection on $\R\times Y_n$ which runs between $\alpha_0$ and $\alpha_1$ has action at least $1/2$, while a bubble, or a non-trivial anti-self-dual connection on the cylinder with the same limit at both ends, has action at least $1$. So the only moduli spaces on $C^+$ which could arise from weak limits are the ones in the table. The hypotheses of Theorem~\ref{thm:gluing} are therefore satisfied, and Proposition~\ref{prop:gluing} holds in the present situation. The argument of Section~\ref{sec:parity} then gives a contradiction, as before, and Proposition~\ref{prop:dihedral} is proved. \section{The remaining cases}\label{sec:ade} In this section we complete the proof of Theorem~\ref{thm:main}. Throughout, $R$ is a connected subdiagram of the Dynkin diagram $E_8$, and $N$ is a submanifold of a K3 surface $X$ which satisfies the hypotheses of the theorem: it is negative definite, its oriented boundary is $Y_R$, and $H_1(N;\Z)=0$. We write $C$ for the closure of the complement of $N$ in $X$. \begin{proposition}\label{prop:reduction} The group $H_2(N;\Z)$ is free, and $H_3(N;\Z)$ is zero. The intersection form $Q_N$ either is isomorphic to $-R$, or is one of the four forms listed in \eqref{eq:four-cases}. The group $H_1(C;\Z)$ is zero, except possibly when $R=A_7$, in which case it is either zero or $\Z/2$. \end{proposition} \begin{proof} The complement $C$ is connected, and both $N$ and $C$ are spin. As in the proof of Proposition~\ref{prop:alternatives}, for $V=N$ or $C$ the group $H_2(V;\Z)$ is free, the group $H_1(V;\Z)$ is finite, and \begin{equation}\label{eq:ade-det} |H_1(Y_R;\Z)|=|\det Q_V|\,|H_1(V;\Z)|^2. \end{equation} The boundary map $H_4(X;\Z)\to H_3(Y_R;\Z)$ in the Mayer--Vietoris sequence is an isomorphism, so $H_3(N;\Z)$ and $H_3(C;\Z)$ are zero. From \eqref{eq:ade-det} we see that $L=-Q_N$ is an even, positive-definite lattice whose determinant is $|H_1(Y_R;\Z)|=\det R$. Let $r$ be its rank. \subsubsection*{Correction terms} The spin$^c$ structures on $Y_R$ are indexed by the elements $\lambda$ of $R^*/R$, in such a way that the spin structures correspond to the elements with $2\lambda=0$, and the formula of \cite[Corollary~1.5]{OSPlumbing} for the boundary of a negative-definite plumbing gives \begin{equation}\label{eq:ade-d} d(Y_R,\spinc_\lambda)=\frac{\rk R}{4}-\min_{v\in R}|\lambda+v|^2. \end{equation} From the minimal norms of the cosets of the root lattices in their duals, we obtain the correction terms of the spin structures: \begin{equation}\label{eq:ade-spin-d} \begin{array}{c|l} R&d(Y_R,\spinc),\ \spinc\text{ spin}\\\hline \rule{0pt}{2.6ex}A_n,\ n\text{ even}&n/4\\ A_n,\ n\text{ odd}&n/4,\ -1/4\\ D_n,\ n\text{ odd}&n/4,\ n/4-1\\ D_n,\ n\text{ even}&n/4,\ n/4-1,\ 0,\ 0\\ E_6&3/2\\ E_7&7/4,\ 1/4\\ E_8&2 \end{array} \end{equation} (For $D_n$ the four cosets have minimal norms $0$, $1$, $n/4$ and $n/4$, and the last two have order two only when $n$ is even. For $A_n$ the coset corresponding to $j\in\Z/(n+1)$ has minimal norm $j(n+1-j)/(n+1)$, and it has order two when $j=(n+1)/2$.) Let $\spinc$ be the spin structure which $Y_R$ inherits from $N$. Then \begin{equation}\label{eq:ade-rank} r\le4\,d(Y_R,\spinc),\qquad r\equiv4\,d(Y_R,\spinc)\pmod 8. \end{equation} The inequality is \cite[Theorem~9.6]{OS}, and the congruence follows from the grading shift of the map which the punctured manifold $N$ induces on $HF^\infty$. If $r=0$, then $N$ is a rational homology ball and $d(Y_R,\spinc)=0$. \subsubsection*{The diagrams $E_6$, $E_7$ and $E_8$} From \eqref{eq:ade-spin-d} and \eqref{eq:ade-rank}, the rank $r$ is $6$ for $E_6$, it is $1$ or $7$ for $E_7$, and it is $0$ or $8$ for $E_8$. In the last case, $r=0$ is excluded because the correction term is not zero, and $L$ is an even unimodular lattice of rank $8$, so it is isomorphic to $E_8$. For $E_6$, the discriminant form of $L$ is the negative of the discriminant form of $A_2$. So $L\oplus A_2$ has an even unimodular overlattice, which is isomorphic to $E_8$, and $L$ is the orthogonal complement of a sublattice $A_2$ in $E_8$, which is $E_6$. The case of $E_7$ was treated in Proposition~\ref{prop:alternatives}. \subsubsection*{The diagrams $D_n$} The determinant of $L$ is $4$, by \eqref{eq:ade-det} and the hypothesis that $H_1(N;\Z)$ is zero, so $r$ is not zero. (For $D_4$ and $D_6$, the correction terms alone do not exclude $r=0$.) From \eqref{eq:ade-spin-d} and \eqref{eq:ade-rank}, we then have $r=4$ for $D_4$, and \[ r\in\{1,5\},\qquad r\in\{2,6\},\qquad r\in\{3,7\} \] for $D_5$, $D_6$ and $D_7$ respectively. An even positive-definite lattice of determinant $4$ is $\langle4\rangle$ if its rank is $1$, and is $A_1\oplus A_1$ if its rank is $2$. When $r\ge3$, the discriminant form of $L$ is determined by the linking form of $Y_R$ and the value of $r$ modulo~$8$, through Milgram's formula, and it is the negative of the discriminant form of the lattice $L'$ in the following table. \[ \begin{array}{c|ccccc} r&3&4&5&6&7\\\hline \rule{0pt}{2.6ex}L'&D_5&D_4&A_3&A_1\oplus A_1&\langle4\rangle\\ L&A_3&D_4&D_5&D_6&D_7 \end{array} \] So $L\oplus L'$ has an even unimodular overlattice, isomorphic to $E_8$, in which $L$ is the orthogonal complement of $L'$. The Weyl group of $E_8$ acts transitively on the sublattices of $E_8$ of each of the types $A_3$, $D_4$ and $D_5$, on the pairs of orthogonal roots, and on the vectors of norm $4$. So the complement can be computed in the standard model, in which $E_8$ consists of the vectors $x$ in $\Z^8\cup(\Z+\tfrac12)^8$ with $\sum x_i$ even. A sublattice $D_k$ in the first $k$ coordinates, where $D_3=A_3$, has complement $D_{8-k}$ in the remaining coordinates. The orthogonal roots $e_1-e_2$ and $e_1+e_2$ have complement $D_6$, and the vector $2e_1$ has complement $D_7$. This gives the last row of the table. \subsubsection*{The diagrams $A_n$} When $n$ is odd, the spin structure with correction term $-1/4$ cannot be the one inherited from $N$. So $d(Y_R,\spinc)=n/4$, and from \eqref{eq:ade-rank} and the fact that $n\le7$ we obtain $r=n$. To identify the lattice, let $V$ be the disk bundle over the two-sphere with Euler number $-(n+1)$. Its boundary is $-Y_{A_n}$, so we can form the closed manifold $Z=N\cup V$. It is negative definite and has $H_1(Z;\Z)=0$, so by Donaldson's theorem its intersection form is diagonalizable. Changing the sign, we identify $H_2(Z;\Z)$ with the standard lattice $\Z^{n+1}$. Let $v\in\Z^{n+1}$ be the class of the zero section of $V$, so that $|v|^2=n+1$. The vector $v$ is primitive, because the quotient of $H_2(Z;\Z)$ by $H_2(V;\Z)$ is a subgroup of $H_2(N,Y_{A_n};\Z)$, which is free. The lattice $L$ is contained in the orthogonal complement of $v$, and it is equal to it because the two have the same rank and determinant. Because $v$ is primitive, the reduction of $v^\perp$ modulo~$2$ is the kernel of the linear form $u\mapsto v\cdot u$ on $(\Z/2)^{n+1}$. Because $v^\perp$ is even, the linear form $u\mapsto|u|^2$, which is not zero, vanishes on this kernel. So the two linear forms are equal, which means that every coordinate of $v$ is odd. Since $|v|^2=n+1$, every coordinate is $\pm 1$, and $v^\perp$ is isomorphic to $A_n$. \subsubsection*{The complement} By \eqref{eq:ade-det}, the group $H_1(C;\Z)$ is zero when $|H_1(Y_R; \Z)|$ is square-free. When $|H_1(Y_R;\Z)|=4$, which is the case for $A_3$ and for $D_n$, the group $H_1(C;\Z)$ could only have order $2$, and $Q_C$ would then be even and unimodular. But the signature of $Q_C$ is $r-16$, with $1\le r\le7$, and this is not divisible by $8$. In the remaining case, $R=A_7$, the relation \eqref{eq:ade-det} allows $H_1(C;\Z)$ to have order $1$ or $2$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] By Proposition~\ref{prop:reduction}, either $Q_N$ is isomorphic to $-R$ or we are in one of the four cases of \eqref{eq:four-cases}. The first of these is excluded by Theorem~\ref{thm:octahedral}, in which $N$ is the piece called $W$, and the other three are excluded by Proposition~\ref{prop:dihedral}. \end{proof} For $R=D_5$ and $R=D_7$, as for $R=E_7$, the hypotheses of Theorem~\ref{thm:main} can be removed. \begin{theorem}\label{thm:dihedral} Let $n$ be $5$ or $7$, and suppose that $Y_n=Y_{D_n}$ is smoothly embedded in a K3 surface $X$. Then $Y_n$ separates $X$ into two pieces, $X=W\cup_{Y_n}C$, where $W$ has oriented boundary $Y_n$ and is negative definite. The groups $H_1(W;\Z)$ and $H_1(C;\Z)$ are zero, and the intersection form of $W$ is isomorphic to $-D_n$. \end{theorem} \begin{proof} As in the proof of Proposition~\ref{prop:alternatives}, the manifold $Y_n$ is an $L$-space, and one of the two pieces, $W$, is negative definite. By \eqref{eq:ade-spin-d}, the correction terms of the two spin structures on $Y_n$ are $n/4$ and $n/4-1$. Both are positive, so the inequality of \cite[Theorem~9.6]{OS} shows that the oriented boundary of $W$ is $Y_n$ and that $b_2(W)\le n$. By \eqref{eq:ade-det}, the group $H_1(W;\Z)$ has order $1$ or $2$. If the order were $2$, then $Q_W$ would be even and unimodular, and its rank would be divisible by $8$. The rank is at most $7$, so $W$ would be a rational homology ball. The correction term of the spin structure inherited from $W$ would then be zero, which it is not. So $H_1(W;\Z)=0$, and Theorem~\ref{thm:main} shows that $Q_W$ is isomorphic to $-D_n$. The group $H_1(C;\Z)$ is zero by Proposition~\ref{prop:reduction}. \end{proof} \begin{thebibliography}{99} \bibitem{DLME} A.~Daemi, T.~Lidman and M.~Miller Eismeier, \emph{3-manifolds without any embedding in symplectic 4-manifolds}, Geom. Topol. \textbf{28} (2024), 3357--3372. \bibitem{DME} A.~Daemi and M.~Miller Eismeier, \emph{Instantons and rational homology spheres}, preprint, \href{https://arxiv.org/abs/2210.14071}{arXiv:2210.14071}. (Section numbers refer to version~2.) \bibitem{Don86} S.~K. Donaldson, \emph{Connections, cohomology and the intersection forms of $4$-manifolds}, J. Differential Geom. \textbf{24} (1986), 275--341. \bibitem{Don} S.~K. Donaldson, \emph{Floer homology groups in Yang--Mills theory}, Cambridge Tracts in Mathematics, vol.~147, Cambridge University Press, 2002. \bibitem{DK} S.~K. Donaldson and P.~B. Kronheimer, \emph{The geometry of four-manifolds}, Oxford Mathematical Monographs, Oxford University Press, 1990. \bibitem{Fee} P.~M.~N. Feehan, \emph{Critical-exponent Sobolev norms and the slice theorem for the quotient space of connections}, Pacific J. Math. \textbf{200} (2001), 71--118. \bibitem{FL} P.~M.~N.~Feehan and T.~G.~Leness, \emph{The \(SO(3)\) monopole cobordism and superconformal simple type}, Advances in Mathematics (2019), article 106817; arXiv:1408.5307, version~5. \href{https://arxiv.org/abs/1408.5307}{arxiv.org/abs/1408.5307}. See Equation~(1.2), Theorem~1.1, Hypothesis~2.5, and Remark~2.8. \bibitem{GHS} V.~Gritsenko, K.~Hulek and G.~K. Sankaran, \emph{Abelianisation of orthogonal groups and the fundamental group of modular varieties}, J. Algebra \textbf{322} (2009), 463--478. \bibitem{Helle} G.~O. Helle, \emph{Equivariant instanton Floer homology and calculations for the binary polyhedral spaces}, preprint, \href{https://arxiv.org/abs/2203.09471}{arXiv:2203.09471}. \bibitem{Kro} P.~B. Kronheimer, \emph{Four-manifold invariants from higher-rank bundles}, J. Differential Geom. \textbf{70} (2005), 59--112. \bibitem{KM} P.~B. Kronheimer and T.~S. Mrowka, \emph{Embedded surfaces and the structure of Donaldson's polynomial invariants}, J. Differential Geom. \textbf{41} (1995), 573--734. \bibitem{MMP} M.~Mari\~no, G.~Moore, and G.~Peradze, \emph{Four-manifold geography and superconformal symmetry}, Mathematical Research Letters \textbf{6} (1999), 429--437; arXiv:math/9812042. \href{https://arxiv.org/abs/math/9812042}{arxiv.org/abs/math/9812042}. \bibitem{ME} M.~Miller Eismeier, \emph{Equivariant instanton homology}, preprint, \href{https://arxiv.org/abs/1907.01091}{arXiv:1907.01091}. (Section numbers refer to version~3.) \bibitem{MH} J.~Milnor and D.~Husemoller, \emph{Symmetric bilinear forms}, Ergebnisse der Mathematik und ihrer Grenzgebiete, vol.~73, Springer, 1973. \bibitem{Moser} L.~Moser, \emph{Elementary surgery along a torus knot}, Pacific J. Math. \textbf{38} (1971), 737--745. \bibitem{Muk} A.~Mukherjee, \emph{A note on embeddings of 3-manifolds in symplectic 4-manifolds}, preprint, \href{https://arxiv.org/abs/2010.03681}{arXiv:2010.03681}. \bibitem{Nic} L.~I. Nicolaescu, \emph{On the Cappell--Lee--Miller gluing theorem}, Pacific J. Math. \textbf{206} (2002), 159--185. \bibitem{OS} P.~Ozsv\'ath and Z.~Szab\'o, \emph{Absolutely graded Floer homologies and intersection forms for four-manifolds with boundary}, Adv. Math. \textbf{173} (2003), 179--261. \bibitem{OSPlumbing} P.~Ozsv\'ath and Z.~Szab\'o, \emph{On the Floer homology of plumbed three-manifolds}, Geom. Topol. \textbf{7} (2003), 185--224. \bibitem{Serre} J.-P. Serre, \emph{A course in arithmetic}, Graduate Texts in Mathematics, vol.~7, Springer, 1973. \bibitem{TauPath} C.~H. Taubes, \emph{Path-connected Yang--Mills moduli spaces}, J. Differential Geom. \textbf{19} (1984), 337--392. \bibitem{TauIndef} C.~H. Taubes, \emph{Self-dual connections on $4$-manifolds with indefinite intersection matrix}, J. Differential Geom. \textbf{19} (1984), 517--560. \bibitem{TauFramework} C.~H. Taubes, \emph{A framework for Morse theory for the Yang--Mills functional}, Invent. Math. \textbf{94} (1988), 327--402. \end{thebibliography} \end{document}