The curve cone of almost complex 4-manifolds
Extracted source text ยท arXiv 1501.06744
\documentclass[11pt]{amsart}
\usepackage{amssymb,latexsym,graphicx, amscd}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{prop}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{question}[theorem]{Question}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{cor}[theorem]{Corollary}
\newtheorem{example}[theorem]{Example}
\newtheorem{conj}[theorem]{Conjecture}
\DeclareMathOperator{\ep}{\epsilon}
\DeclareMathOperator{\E}{\mathcal E}
\newcommand\C{{\mathbb C}}
\newcommand\f{\mathcal{F}}
\newcommand\h{{harm}}
\renewcommand\O{{\omega}}
\renewcommand\t{{\theta}}
\renewcommand\r{{\Re\mathfrak{e}\,}}
\newcommand\im{{\Im\mathfrak{m}\,}}
%\newcommand\E{{\mathbb E}}
\newcommand\R{{\mathbb R}}
\newcommand\Z{{\mathbb Z}}
\newcommand\T{{\mathbb T}}
\newcommand{\Cf}{$\mathcal{C}^\infty$-full}
\newcommand{\Cp}{$\mathcal{C}^\infty$-pure}
\newcommand{\de}[2]{\frac{\partial #1}{\partial #2}}
\newcommand\Span{{\hbox{\em Span}}}
\begin{document}
\title{The curve cone of almost complex $4$-manifolds}
\author{Weiyi Zhang}
\address{Mathematics Institute\\ University of Warwick\\ Coventry, CV4 7AL, England}
\email{weiyi.zhang@warwick.ac.uk}
\begin{abstract}
In this paper, we study the curve cone of an almost complex $4$-manifold which is tamed by a symplectic form. In particular, we prove the cone theorem as in Mori theory for all such manifolds using the Seiberg-Witten theory. For small rational surfaces and minimal ruled surfaces, we study the configuration of negative curves. As an application, we prove the Nakai-Moishezon type duality for all almost K\"ahler structures on $\mathbb CP^2\#k\overline{\mathbb CP^2}$ with $k\le 9$ and minimal ruled surfaces with a negative curve. This is proved using a version of Gram-Schmidt orthogonalization process for the $J$-tamed symplectic inflation.
\end{abstract}
\maketitle
\section{Introduction}
The study of the curve cone is of much importance to the birational geometry of an algebraic variety. The cone theorem for smooth varieties proved in \cite{Mor82}, which describes the structure of the curve cone by extremal rays, was the first major step of Mori's program. It was later generalized to a larger class of varieties by Koll\'ar, Reid, Shokurov and others. The proof for a general variety relies on the bend-and-break technique, where a characteristic $0$ proof is still lacking. However, there is an elementary proof for algebraic surfaces, see {\it e.g.} \cite{Reidsur}. Early applications of the notion of the curve cone include Nakai-Moishezon's and Kleiman's ampleness criteria.
We could also similarly define the curve cone $A_J(M)$ for a tamed almost complex manifold $(M, J)$:
$$A_J(M)=\{\sum a_i[C_i]| a_i> 0\}$$where $C_i$ are irreducible $J$-holomorphic subvarieties on $M$. Here an irreducible $J$-holomorphic subvariety is the image of a $J$-holomorphic map $\phi: \Sigma\rightarrow M$ from a complex connected curve $\Sigma$, where $\phi$ is an embedding off a finite set. More generally, a $J$-holomorphic subvariety is a finite set of pairs $\{(C_i, m_i), 1\le i\le n\}$, where each $C_i$ is irreducible $J$-holomorphic subvariety and each $m_i$ is a non-negative integer. Later, we sometimes say $J$-holomorphic curves (or simply, curves) instead of $J$-holomorphic subvarieties.
We will focus on dimension $4$. Hence by taking Poincar\'e duality (we will identify the curve classes with their Poincar\'e dual cohomology classes by abusing the notation), we have $A_J(M)$ sitting as a cone in vector space $ H_J^+(M)\subset H^2(M; \mathbb R)$. Here $H_J^+(M)$ is called the $J$-invariant cohomology which is introduced in \cite{LZ, DLZ} along with the $J$-anti-invariant $H_J^-(M)$. The almost complex
structure $J$ acts on the bundle of real 2-forms $\Lambda^2$ as an
involution, by $\alpha(\cdot, \cdot) \rightarrow \alpha(J\cdot,
J\cdot)$. This involution induces the splitting into $J$-invariant, respectively,
$J$-anti-invariant 2-forms
$\Lambda^2=\Lambda_J^+\oplus \Lambda_J^-$. Then we define
\begin{equation*}
H_J^{\pm}(M)=\{ \mathfrak{a} \in H^2(M;\mathbb R) | \exists \;
\alpha\in \Lambda_J^{\pm}, \, d\alpha=0 \mbox{ such that } [\alpha] =
\mathfrak{a} \}.
\end{equation*}
Usually, to prove a geometric result for a general (non-generic) tamed almost complex structure is far more delicate than the generic case, since we have to deal with very general subvarieties. The techniques developed in \cite{LZ-generic, LZrc} enable us to have a fairly clear structural picture for subvarieties in a $J$-nef class. In this paper, we would develop more techniques to work with a general tamed almost complex structure. In particular, we find that the curve cone and sometimes curve configurations behave in many aspects similar to that of algebraic varieties.
Throughout the paper, all the numerical calculations have the following geometric picture in mind: we think an element in $A_J(M)$ (or $A^{\mathbb Q}_J(M)=\overline{A}_J(M)\cap H^2(M, \mathbb Q)$) as the homology class of an $\mathbb R$ (or $\mathbb Q$) $J$-holomorphic subvarieties. Here an $\mathbb R$ (or $\mathbb Q$) subvariety allows $m_i$ in the definition of subvariety to be positive real numbers (or rational numbers). With this freedom in hand, we can choose some or all ``irreducible components" $C_i$ to be the extremal rays of $A_J(M)$.
Here a subcone $N \subset A_J(M)$ is called {\it extremal} if $u,v \in A_J(M), u+v\in N$ imply that $u, v\in N$. A $1$-dimensional extremal subcone is called a {\it extremal ray}. Hence, the numerical information of the given class in $A_J(M)$ would give information for these extremal rays, and vice versa. The geometry is thus hidden in the numerical information. A typical example is Corollary \ref{-1} which guarantees the existence of $-1$ rational curves by the numerical property of a class known to be in $A_J(M)$.
From another viewpoint, considering $\mathbb Q$-subvarieties of an integral class $e\in A_J(M)$ is the same as considering subvarieties in all the integral classes of the ray $\mathbb R^+ \cdot e\subset A_J(M)$. The complexity arises, thus one has to study all integral classes in the ray rather than just a single class, because the subvarieties in a class and its multiples do not change linearly, see {\it e.g.} Examples \ref{-K8blowup}, \ref{ruledex}.
For this sake, we will first understand the extremal rays. Our first result, the cone theorem for tamed almost complex $4$-manifolds, is a general structural result on the extremal rays of the ``negative" part (this is not to be confused with the notion of ``negative curves" used later, which means curves with negative self-intersection). There are several different proofs of the cone theorem in algebraic geometry (at least for algebraic surfaces). Mori's original proof uses the bend-and-break and the Kleiman's ampleness criterion. The argument in \cite{Reidsur} is bend-and-break free, but it proves rationality theorem first. Our proof proceeds in a totally different logical order. We do not need all these technical results before the cone theorem. In our paper, the rationality theorem, Proposition \ref{ratthm}, is proved as a corollary of the cone theorem. Moreover, we are only able to prove Nakai-Moishezon and Kleiman type results for certain rational surfaces.
However, the soul of the algebraic geometric proofs and our proof is the same: vanishing $\Rightarrow$ non-vanishing $\Rightarrow$ cone theorem.
%Recall that the logical order of the proofs of the cone theorem in algebraic geometry is the following: vanishing theorems $\Rightarrow$ base point free theorem $\Rightarrow$ rationality theorem $\Rightarrow$ cone theorem. Our proof proceeds in a totally different logical order, where the rationality theorem (Proposition \ref{ratthm}) is proved after the cone theorem and there is no corresponding version of base point free theorem.
We start by proving certain Seiberg-Witten invariants vanish. Using the Seiberg-Witten wall crossing formula \cite{LLwall}, we have non-vanishing results. Then Taubes' SW=Gr \cite{T, T3} would imply a class does not span an extremal ray if it could be written as the sum of two classes with non-trivial Seiberg-Witten invariants. This would eventually guarantee the extremal rays of the ``negative" part are generated by rational curves.
%The main tool of our proof is the Seiberg-Witten theory, particularly the Seiberg-Witten wall crossing formula \cite{LLwall} and Taubes' SW=Gr \cite{T, T3}. Furthermore, the adjunction formula guarantees the extremal rays of the ``negative" part are generated by rational curves.
The precise statement of the cone theorem is the following, which is a combination of Theorem \ref{conethm} and Proposition \ref{exray}.
\begin{theorem}\label{coneintro}
Let $(M, J)$ be a tamed almost complex $4$-manifold. Then $$\overline{A}_J(M)=\overline{A}_{K_J}(M)+\sum \mathbb R^+[L_i]$$ where $L_i\subset M$ are countably many smooth irreducible rational curves such that $-3\le K_J\cdot [L_i]<0$ which span the extremal rays $\mathbb R^+[L_i]$ of $\overline{A}_J(M)$.
Moreover, for any $J$-almost K\"ahler symplectic form $\omega$ and any given $\epsilon>0$, there are only finitely many extremal rays with $(K_J+\epsilon[\omega])\cdot [L_i]\le 0$.
In addition, an irreducible curve $C$ is an extremal rational curve if and only if
\begin{enumerate}
\item $C$ is a $-1$ rational curve;
\item $M$ is a minimal ruled surface or $\mathbb CP^2\# \overline{\mathbb CP^2}$, and $C$ is a fiber;
\item $M=\mathbb CP^2$ and $C$ is a projective line.
\end{enumerate}
\end{theorem}
In fact, we prove a slightly stronger version of cone theorem:
$$\overline{A}^{\mathbb Q}_J(M)=\overline{A}^{\mathbb Q}_{K_J}(M)+\sum \mathbb Q^+[L_i],$$where $\overline{A}^{\mathbb Q}_J(M)=\overline{A}_J(M)\cap H^2(M, \mathbb Q)$ and $\overline{A}^{\mathbb Q}_{K_J}(M)=\overline{A}_{K_J}(M)\cap H^2(M, \mathbb Q)$.
Here $K_J$ is the canonical class of $(M, J)$ and $A_{K_J}(M)=\{C\in A_J(M)| K_J\cdot C\ge 0\}$ is the ``positive" part of the curve cone. In general, this part is not generated by countably many extremal rays since it may have round boundary. This phenomenon happens in particular when we do not have sufficient curves as for a generic almost complex structure on manifolds with $b^+>1$, or even in the rational surfaces $\mathbb CP^2\#k\overline{ \mathbb CP^2}$ with $k> 9$.
On the other hand, there are indeed many cases whose curve cones are polytopes. The most well known examples are rational surfaces $\mathbb CP^2\#k\overline{ \mathbb CP^2}$ when $k\le 9$ and $S^2\times S^2$. The cases of $S^2$ bundles over $S^2$ are treated in \cite{LZ-generic} for example.
We give a careful analysis of the curve cone for all possible tamed almost complex structures on $\mathbb CP^2\#2\overline{ \mathbb CP^2}$ in Section $3$. In particular, we show that there are at least two embedded $-1$ rational curves for any tamed almost complex structures (Theorem \ref{cp2+2}). This should be compared to other results on the existence of embedded $-1$ rational curves. As a corollary of Theorem \ref{coneintro}, we give a proof of the fact that any tamed $J$ on a non-minimal symplectic $4$-manifold which is not diffeomorphic to a $S^2$ bundle contains at least one smooth $-1$ rational curve. In fact, it appears as a component of a $\mathbb Q$-variety representing an exceptional sphere class (Corollary \ref{-1}). On the other hand, based on an argument communicated to the author by Dusa McDuff, we show that if $M$ is not diffeomorphic to one point blow up of $S^2$ bundles, then there exists a tamed $J$ such that all $-1$ rational curves are disjoint and any possible numbers of such $-1$ rational curves could be realized (Theorem \ref{disjoint-1}).
Let $h_J^+(M)=\dim H_J^+(M)$. Then by the light cone lemma, when $h_J^+(M)=b^-(M)+1$ and $b^-(M)>1$, the boundary hyperplanes of the dual of the curve cone are determined by $J$-holomorphic curves with negative self-intersection, {\it i.e.} negative curves (see Lemma \ref{negbdy}). The equality $h_J^+(M)=b^-(M)+1$ holds for any complex structure $J$ or manifolds with $b^+(M)=1$. Hence to determine various cones, {\it e.g.} the curve cone and the almost K\"ahler cone, for $\mathbb CP^2\#k\overline{ \mathbb CP^2}$ when $k\le 9$, we are reduced to determine all possible negative curves. The classification is done in Section $4$. Especially, all negative curves (resp. non-positive curves) on $\mathbb CP^2\#k\overline{ \mathbb CP^2}$ with $k\le 9$ (resp. $k<9$) are rational curves. A complete list of the classes of all such rational curves are obtained in Propositions \ref{a<0}, \ref{a>0}, \ref{a>0'}. Moreover, we have the following more precise statement about negative curve configurations.
\begin{theorem}\label{main1}
For rational $4$-manifolds $\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k<8$, the set of all the possible configurations of negative self-intersection curves for tamed almost complex structures are the same as the set for complex structures.
\end{theorem}
Namely, given a tamed almost complex structures, we could find a complex structure (on the same manifold) such that their configurations of negative curves are the same. Here by a configuration of negative curves, we mean the set of homology classes and multiplicities of each intersection point.
This is also true for minimal ruled surfaces. However, as observed in \cite{CP}, this is not true for a non-minimal ruled surface, {\it e.g.} $(T^2\times S^2)\# \overline{\mathbb CP^2}$. A generic almost complex structure will have only two negative curves $E$ and $F-E$ where $F$ is class of $S^2$ and $E$ is the class of exceptional curve, while there is no complex structure on it having exactly those two negative curves. The paper \cite{CP} also gives such an example which is a minimal surface of general type. Currently, no such simply connected examples are known. On $\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k\ge 9$, it is related to the Nagata conjecture.
\begin{question}(Nagata)
For every $k> 9$, do we have complex structure on $\mathbb CP^2\# k\overline{\mathbb CP^2}$, every irreducible curve $C$, such that
$$d\ge \frac{\sum_{q=1}^km_q}{\sqrt{k}},$$
where $[C]=dH-\sum_{q=1}^km_qE_q$?
\end{question}
This is equivalent to say that $H-\frac{1}{\sqrt{k}}\sum E_i$ is on the closure of the K\"ahler cone by Nakai-Moishezon criterion. It is easy to see that a generic tamed almost complex structure satisfies the inequality.
The above discussion gives us a clearer picture on the polytopic boundary of the curve cone $A_J(M)$. We can apply it to understand the Nakai-Moishezon or the Kleiman type duality between the curve cone and the almost K\"ahler cone $\mathcal K_J^c=\{[\omega]\in H^2(M; \mathbb R)| \omega \mbox{ is compatible with }J\}$. When $b^+(M)=1$, it is shown in \cite{LZ} that $\mathcal K_J^c$ is equal to the tame cone $\mathcal K_J^t=\{[\omega]\in H^2(M; \mathbb R)| \omega \mbox{ tames }J\}$.
Let us introduce a couple more cones. First is the positive cone $\mathcal P=\{e\in H^2(M;\mathbb R)|e\cdot e >0\}$. The second cone $A_J^{\vee, >0}(M)$ (resp. $\overline{A}_J^{\vee, >0}(M)$) is the positive dual of $A_J(M)$ (resp. $\overline{A}_J(M)$) where the duality is taken within $H_J^+(M)$. Let $\mathcal P_J=A_J^{\vee, >0}(M) \cap \mathcal P$. Clearly, $\mathcal K_J^c\subset \mathcal P_J$. Then we can ask the following
\begin{question} \label{dualC}
For an almost K\"ahler structure $J$ on a closed, oriented $4$-manifold $M$ with $b^+(M)=1$, is
\begin{equation*}\label{dualak}
\mathcal K_J^c=\mathcal P_J=\overline{A}_J^{\vee, >0}(M) ?
\end{equation*}
When $b^+(M)>1$, is $\mathcal K_J^c$ a connected component of $\mathcal P_J$?
\end{question}
When $J$ is an integrable complex structure, $\mathcal K_J^c$ is the K\"ahler cone. A K\"ahlerian Nakai-Moishezon theorem is given by Buchdahl and Lamari in
dimension $4$ \cite{Bu, Lamari}, and Demailly-Paun \cite{DP} in arbitrary dimension to determine the K\"ahler cone completely.
It is worth noting that in algebraic geometry, all the proofs of the cone theorem relies on the Nakai-Moishezon or Kleiman ampleness criterion. However, our almost complex cone theorem (Theorem \ref{coneintro}) does not need the corresponding version, {\it i.e.} Question \ref{dualC}.
The key of Question \ref{dualC} is to construct almost K\"ahler forms. There are currently two methods of construction. The first is using Taubes' subvarieties-current-form technique \cite{T2009}. However, to argue it for an arbitrary $J$, we have to use spherical subvarieties as in \cite{LZ-generic}. It was successfully used to affirm Question \ref{dualC} for $S^2$ bundles over $S^2$. The method is applied to prove it for $\mathbb CP^2\#2\overline{\mathbb CP^2}$ in this paper. However, it cannot go further along this line. The limit of this method is discussed in Section \ref{cp2+3}.
The second method is reducing the construction of the almost K\"ahler cone to determine the more flexible tame cone $\mathcal K_J^t$ by virtue of the identity $\mathcal K_J^c=\mathcal K_J^t\cap H_J^+(M)$ established in \cite{LZ}. In this situation, we could apply the $J$-tamed symplectic inflation developed by McDuff \cite{Mc2} and Buse \cite{Bus}. More precisely, we apply Buse's inflation for negative $J$-holomorphic curves to our extremal ray of curve cone. The combinatorial version of it, the formal $J$-inflation, is introduced. A version of Gram-Schmidt orthogonalization process is applied in our situation to significantly simplify the otherwise too complicated calculation. We are able to prove the following general result when the curve cone has no round boundary.
\begin{theorem}\label{noround}
On a $4$-dimensional almost complex manifold $(M,J)$ with $\mathcal K_J^c\ne \emptyset$, and $h_J^+=b^-+1$, if $\mathcal P_J$ has no round boundary and each of the boundary hyperplane is determined by a smooth negative curve, then $\mathcal K_J^c$ is a connected component of $\mathcal P_J$.
\end{theorem}
Applying the discussion of the curve cone for rational and ruled surfaces in Sections 2-4, we have the both Nakai-Moishezon and Kleiman dualities.
\begin{theorem}\label{rr}
If $J$ is almost K\"ahler on $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k\le 9$ or a minimal irrational ruled surface with a negative curve, then $$\mathcal K_J^c=\mathcal K_J^t= \mathcal P_J=\overline{A}_J^{\vee, >0}(M).$$
\end{theorem}
We are also able to show that Question \ref{dualC} is true for a generic tamed $J$ on a manifold with $b^+(M)=1$.
Here a generic tamed almost complex structure means that it is chosen from a residual subset of all tamed almost complex structures.
\begin{theorem}\label{generic}
On a symplectic $4$-manifold with $b^+=1$, $\mathcal K_J^t=\mathcal K_J^c= \mathcal P_J$ for a generic tamed $J$.
\end{theorem}
The author is grateful to Tian-Jun Li and Clifford Taubes for very helpful suggestions on Theorem \ref{coneintro}. He is also in debt to Tian-Jun Li for pointing out reference \cite{Bus} and Dusa McDuff for many useful comments to greatly improve the presentation of the paper. The author is partially supported by AMS-Simons travel grant and EPSRC grant EP/N002601/1.
\section{The Cone Theorem}
In this section, we will give a proof of the cone theorem for tamed almost complex $4$-manifolds. Recall an almost complex structure $J$ is said to be tamed if there is a symplectic form $\omega$ such that the bilinear form $\omega(\cdot, J(\cdot))$ is positive definite. A taming form of $J$ is said to be compatible with $J$ if the bilinear form $\omega(\cdot, J(\cdot))$ is symmetric. If there is a symplectic form compatible with the almost complex structure $J$, then $J$ is called almost K\"ahler. Hence, a $J$-compatible symplectic form is also called a $J$-almost K\"ahler form.
Given a class $e\in H_2(M; \mathbb Z)$, introduce the $J$-genus of $e$, $$g_J(e)=\frac{1}{2}(e\cdot e+K\cdot e)+1,$$where $K=K_J$ is the canonical class of $J$.
Moreover, when $C$ is an irreducible subvariety, $g_J([C])$ is non-negative. In fact, if $\Sigma$
is the model curve of $C$, by the adjunction inequality,
\begin{equation}\label{adj inequality}
g_J(e_C)\geq g(\Sigma),
\end{equation}
with equality if and only if $C$ is smooth.
In particular, if $C$ is the class of an irreducible curve having $C^2<0$ and $K\cdot C<0$ then the curve must be a $-1$ rational curve.
In the following, a rational curve means an irreducible $J$-holomorphic subvariety of $J$-genus $0$. By \eqref{adj inequality}, such a curve has to be smooth.
\subsection{The Seiberg-Witten invariant}
In this subsection, we will give a very brief introduction to the Seiberg-Witten invariant, which will be the main tool to establish Theorem \ref{coneintro}. For a detailed introduction, see for example \cite{LLwall, LLb+=1} and references therein.
Let $M$ be an oriented $4$-manifold with a given Riemannian metric $g$ and a spin$^{c}$ structure $\mathcal L$ on $M$. Hence there are a pair of rank $2$ complex vector bundles $S^{\pm}$ with isomorphisms $\det(S^+)=\det(S^-)=\mathcal L$. The Seiberg-Witten equations are for a pair $(A, \phi)$ where $A$ is a connection of $\mathcal L$ and $\phi$ is a section of $S^+$. These equations are
$$D_A\phi=0$$
$$F_A^+=iq(\phi)+i\eta$$ where $q$ is a canonical map $q: \Gamma(S^+)\rightarrow \Omega^2_+(M)$ and $\eta$ is a self-dual $2$-form on $M$.
The group $C^{\infty}(M; S^1)$ naturally acts on the space of solutions and acts freely at irreducible solutions. Recall a reducible solution has $\phi=0$, and hence $F_A^+=i\eta$. The quotient is the moduli space and is denoted by $\mathcal M_M(\mathcal L, g, \eta)$. For generic pairs $(g, \eta)$, the Seiberg-Witten moduli space $\mathcal M_M(\mathcal L, g, \eta)$ is a compact manifold of dimension $$2d(\mathcal L)=\frac{1}{4}(c_1(\mathcal L)^2-(3\sigma(M)+2\chi(M)))$$ where $\sigma(M)$ is the signature and $\chi(M)$ is the Euler number. Furthermore, an orientation is given to $\mathcal M_M(\mathcal L, g, \eta)$ by fixing a homology orientation for $M$, {\it i.e.} an orientation of $H^1(M)\oplus H^2_+(M)$. The space of $g$-self-dual space $\mathcal H^+_g(M)$ is spanned by a single harmonic $2$-form $\omega_g$ of norm $1$ agreeing with the homology orientation.
Now, for the convenience of notation, we fix a (homotopy class of) almost complex structures. In particular, it determines a canonical class $K$ which is the first Chern class of the cotangent bundle. We denote $e:=\frac{c_1(\mathcal L)+K}{2}$. In the following, notice the first Chern class $c_1(\mathcal L)$ determines $\mathcal L$ and {\it vice versa}.
For a generic choice of $(g, \eta)$, the Seiberg-Witten invariant $SW_{M, g, \eta}(e)$ is defined as follows. If $d(\mathcal L)<0$, then the SW invariant is zero. If $d(\mathcal L)=0$, then the moduli space is a finite union of signed points and the SW invariant is the sum of the corresponding signs. If $d(\mathcal L)>0$, then the SW invariant is obtained by pairing the fundamental class of $\mathcal M_M(\mathcal L, g, \eta)$ with the maximal cup product of the Euler class of the $S^1$-bundle $\mathcal M^0_M(\mathcal L, g, \eta)$ over $\mathcal M_M(\mathcal L, g, \eta)$. Here $\mathcal M^0_M(\mathcal L, g, \eta)$ is called the based moduli space which is the quotient of the space of solution by $C_0^{\infty}(M, S^1)$ (the elements in $C^{\infty}(M, S^1)$ which map a base point in $M$ to $1$ in $S^1$).
If $b^+>1$, a generic path of $(g, \eta)$ contains no reducible solutions. Hence, the Seiberg-Witten invariant is an oriented diffeomorphism invariant in this case. Hence we can use the notation $SW(e)$ for the Seiberg-Witten invariant. We will also write $$\dim_{SW}(e)=2d(\mathcal L)=\frac{1}{4}((2e-K)^2-K^2)=e^2-K\cdot e$$ for the Seiberg-Witten dimension. In the case $b^+=1$, there might be reducible solutions on a $1$-dimensional family. Recall that the curvature $F_A$ represents the cohomology class $-2\pi ic_1(\mathcal L)$. Hence $F_A^+=i\eta$ holds only if $-2\pi c_1(\mathcal L)^+=\eta$. This happens if and only if the discriminant $\Delta_{\mathcal L}(g, \eta):=\int (2\pi c_1(\mathcal L)+\eta)\omega_g=0$. With this in mind, the set of pairs $(g, \eta)$ with positive (resp. negative) discriminant is called the positive (resp. negative) $\mathcal L$ chamber. We use the notation $SW^{\pm}(e)$ for the Seiberg-Witten invariants in these two chambers. Moreover, in the this paper, we will use $SW(e)$ instead of $SW^-(e)$ when $b^+=1$.
As one can easily see from the discussion above, $SW$ (or $SW^{\pm}$) could be defined more generally as map from $H^2(M, \mathbb Z)$ to $\Lambda^*H^1(M, \mathbb Z)$, although we do not need this generality for this paper.
We now assume $(M, \omega)$ is a symplectic $4$-manifold, and $J$ is a $\omega$-tamed almost complex structure. Then the theorems in \cite{T3} and \cite{LLb+=1} equate Seiberg-Witten invariants with Gromov-Taubes invariants that are defined by making a suitably counting of $J$-holomorphic subvarieties (In fact, the $SW(e)$ we defined here is essentially the Gromov-Taubes invariant in the literature). Especially, when $SW(e)\ne 0$, there is a $J$-holomorphic subvariety in class $e$ passing through $\dim_{SW}(e)$ given points. This is the key result from the Seiberg-Witten theory we will use in this section.
\subsection{The cone theorem}
We first have the following
\begin{lemma}\label{Kod>0}
Let $J$ be a tamed almost complex structure on a symplectic $4$-manifold which is not rational or ruled. Let $C$ be an irreducible $J$-holomorphic curve with $K\cdot [C]<0$. Then $C$ has to be a $-1$ rational curve.
\end{lemma}
\begin{proof}
If $M$ is a minimal symplectic $4$-manifold which is not rational or ruled, then by Taubes' theorem \cite{T} we always have a $J$-holomorphic subvariety in class $K$ when $b^+>1$ and by Li-Liu's theorem in class $2K$ when $b^+=1$. If $2K=\sum a_i[C_i]$, then $K\cdot C<0$ would imply $C$ is one of $C_i$. Let it be $C_1$. Then $[C]\cdot (2K-a_1[C])\ge 0$ implies $C^2<0$. Altogether, we have $C$ is a $-1$ rational curve. This is a contradiction.
If $M$ is not minimal, then $K$ could be written as $K_0+E_1+\cdots +E_k$, where $K_0\cdot E_i=0$ and $E_i$ are the exceptional classes. We also have $2K_0=\sum a_i[C_i]$. Thus $2K$ is also a linear combination of curve classes. The same argument as in the minimal case implies $C$ is a $-1$ rational curve.
\end{proof}
Next we prove a very useful lemma. In many applications later, our $C$ also has $SW([C])\ne 0$.
\begin{lemma}\label{int>0}
If $C$ is an irreducible $J$-holomorphic curve with $C^2\ge 0$ and $SW(e)\ne 0$, then $e \cdot [C]\ge 0$.
\end{lemma}
\begin{proof}
Since $SW(e)\ne 0$, we can represent $e$ by a possible reducible $J$-holomorphic subvariety. Since each irreducible curve $C'$ has $[C']\cdot [C]\ge 0$, we have $e\cdot [C]\ge 0$.
\end{proof}
\begin{prop}\label{SW}
Let $(M, J)$ be a tamed almost complex $4$-manifold. Let $C$ be an irreducible curve such that $K\cdot [C]<0$. Then $SW([C])\ne 0$. Moreover, there is a curve in class $[C]$ passing through any given point when $C$ is not a $-1$ rational curve.
\end{prop}
\begin{proof}
First we could assume $M$ is rational or ruled. Otherwise, it is proved in Lemma \ref{Kod>0}.
Notice $C^2+K\cdot [C]=2g-2$, $$[C]\cdot (K-[C])=-\dim_{SW} [C]=-\dim_{SW}(K-[C])=2K\cdot [C]+2-2g\le 0.$$ The equality holds if and only if $g=0, K\cdot [C]=-1$, and $C^2=-1$, {\it i.e.} $C$ is a $-1$ rational curve. Otherwise $C^2\ge 2g-2+2\ge 0$.
Hence we assume $[C]\cdot (K-[C])<0$. If $SW(K-[C])\neq 0$, by SW=Gr we have a (possibly reducible) $J$-holomorphic curve in class $K-[C]$. Hence the irreducible curve $C$ must be a component of this curve and $C^2<0$. This contradicts the adjunction and dimension formula. Thus $SW(K-[C])=0$.
Since $\dim_{SW} [C]\ge 0$, we have wall-crossing formula
$$|SW(K-[C])-SW([C])|=\begin{cases} 1 & \hbox{if $(M, \omega)$ rational,}\cr
|1+[C]\cdot T|^h &\hbox{if
$(M, \omega)$ irrationally ruled,}\cr\end{cases}$$
where $T$ is the unique positive fiber class and $h$ is the genus of base surface of irrationally ruled manifolds (see \cite{LLwall}).
In summary, $SW([C])\neq 0$. Hence there is a curve in class $[C]$ passing through any given point since $\dim_{SW} [C]>0$.
\end{proof}
Next we will prove the first two statements of Theorem \ref{coneintro}, the cone theorem.
\begin{theorem}\label{conethm}
Let $(M, J)$ be a tamed almost complex $4$-manifold. Then $$\overline{A}_J(M)=\overline{A}_{K_J}(M)+\sum \mathbb R^+[L_i]$$ where $L_i\subset M$ are countably many smooth irreducible rational curves such that $-3\le K\cdot [L_i]<0$ which span the extremal rays $\mathbb R^+[L_i]$ of $\overline{A}_J(M)$.
Moreover, for any $J$-almost K\"ahler symplectic form $\omega$ and any given $\epsilon>0$, there are only finitely many extremal rays with $(K+\epsilon[\omega])\cdot [L_i]\le 0$.
\end{theorem}
\begin{proof}
We first prove our statement of extremal rays $\mathbb R^+[L_i]$.
When $M$ is not rational or ruled, then Lemma \ref{Kod>0} verifies our claim. Especially, $L_i$ are finitely many rational curves with $K\cdot [L_i]=-1$.
In general, let $C$ be an irreducible curve with $K\cdot [C]<0$. By Proposition \ref{SW}, we have $SW([C])\neq 0$. Hence for any tamed almost complex structure $J$, there is a (possibly reducible) subvariety in class $[C]$ by Taubes' SW=Gr and Gromov compactness. Especially, it is true for a projective variety.
Now, assume $C$ is an irreducible curve with $g_J([C])>0$ and $K\cdot [C]<0$, or $g_J([C])=0$ and $K\cdot [C]< -3$ on a rational or ruled surface. We want to show that $[C]$ cannot span an extremal ray of the curve cone (we will say $[C]$ is not extremal for simplicity). We divide our discussion into the following cases.
\bigskip
\noindent {\bf Case 1: Irrational ruled surfaces} \\
$\bullet$ $M=\Sigma_h\times S^2$, $h> 1$, and its blowups \\
In this case, let $U$ be the class of the base $\Sigma_h$ and $T$ be the class of the fiber $S^2$. Then the canonical class $K=-2U+(2h-2)T+\sum_i E_i$. Let $[C]=aU+bT-\sum_i c_iE_i$.
Since both $[C]$ and $T$ pair negatively with $K$, by Lemma \ref{SW}, both classes have non-trivial Seiberg-Witten invariant. Applying Lemma \ref{int>0} to the pair $[C]$ and $T$, we have $a\ge 0$. We could assume $[C]$ is not one of the classes of exceptional curves $E_i$. Then applying Lemma \ref{int>0} to $[C]$ and $E_i$ gives $c_i\ge 0$.
The assumption $K\cdot [C]<0$ reads as $$a(2h-2)-2b+\sum c_i<0.$$Especially, we have $b>0$.
When $a=0$, we have $$-\sum c_i^2=C^2=2g(C)-2-K\cdot [C]>2g(C)-2\ge -2.$$ It works only when $g(C)=0$ and there is a unique nonzero $c_i$, say $c_1$, which equals to $1$. Moreover, $K\cdot [C]=-1$ would imply $[C]=T-E_1$.
When $a>0$, we take the smooth projection $f: C\rightarrow \Sigma_h$ to the base. It has degree $a=[C]\cdot T$.
Since $\Sigma_h$ has genus greater than one, and by Kneser's theorem, we have
$$2g(C)-2\ge a(2h-2)\ge 2a.$$
Now, we are planning to show that the class $[C]-T$ has non trivial Seiberg-Witten invariant. First we show $SW(K-([C]-T))=0$. If not, notice $SW(T)\ne 0$, we should have $(K-([C]-T))\cdot T\ge 0$ by Lemma \ref{int>0}. However, it contradicts to the calculation $(K-([C]-T))\cdot T=-2-[C]\cdot T<0$. Next we show the Seiberg-Witten dimension of it is nonnegative.
\begin{equation*} \label{SWdim}
\begin{array}{lll}
\dim_{SW}([C]-T)&=&([C]-T)^2-K\cdot ([C]-T)\\
&&\\
&=&C^2-K\cdot [C] -2-2[C]\cdot T\\
&&\\
&\ge&2g(C)-2-2a\\
&&\\
&\ge&0.\\
\end{array}
\end{equation*}
Finally the wall crossing formula implies $$|SW([C]-T)|=|SW(K-([C]-T))-SW([C]-T)|=|1+a|^h\ne 0.$$
Hence, we complete our argument that $[C]=([C]-T)+T$ is not extremal in this case.
\\
$\bullet$ $M=T^2\times S^2$ and its blowups \\
We use the same setting as the above case. That is, we assume $[C]=aU+bT-\sum_i c_iE_i$ and the canonical class $K=-2U+\sum_i E_i$.
As showed in the above case, we only need to show that when $a>0$, $[C]$ is not extremal. Without loss, we assume $c_1\ge c_2 \ge \cdots$.
Notice $$C^2=2ab-\sum c_i^2, \quad -K\cdot [C]=2b-\sum c_i>0.$$
We will first show that $[C]$ must not be extremal in the case that $c_1\le a$ and make a corresponding change in the $c_1>a$ case later.
%If $c_1\le a$, we have contradiction
$$C^2=2ab-\sum c_i^2\ge 2ab-a\sum c_i=a(2b-\sum c_i)=a(-K\cdot [C]).$$
Hence look at the classes $l[C]-T$, we have
\begin{equation*} \label{SWdim4}
\begin{array}{lll}
\dim_{SW}(l[C]-T)&=&(l[C]-T)^2-K\cdot (l[C]-T)\\
&&\\
&=&(l^2C^2-lK\cdot [C]) -2-2l [C]\cdot T\\
&&\\
&\ge&(l^2a+l)(-K\cdot [C])-2-2la\\
&&\\
&\ge& l^2a+l-2la-2\\
\end{array}
\end{equation*}
It is greater than $0$ if $l$ is large enough ({\it e.g.} $l>2a$).
Since $(K-(l[C]-T))\cdot T=-2-la<0$, Lemma \ref{int>0} implies $SW(K-(l[C]-T))= 0$. Apply the wall crossing formula $$|SW(l[C]-T)|=|SW(K-(l[C]-T))-SW(l[C]-T)|=|1+la|\ne 0.$$
Hence $[C]=\frac{1}{l}((l[C]-T)+T)$ where $l[C]-T$ is not proportional to $T$ since $a>0$. It is a decomposition of two non-proportional classes with non-trivial Seiberg-Witten invariant which is not extremal.
If $c_1>a$.
similar to the above, we first calculate the Seiberg-Witten dimension of the class $[C]-T+E_1$.
\begin{equation*} \label{SWdim2}
\begin{array}{lll}
\dim_{SW}([C]-T+E_1)&=&([C]-T+E_1)^2-K\cdot ([C]-T+E_1)\\
&&\\
&=&(C^2-K\cdot [C]) -2-2[C]\cdot T+2c_1\\
&&\\
&\ge&C^2-K\cdot [C]\\
&&\\
&>&0.\\
\end{array}
\end{equation*}
Again $(K-([C]-T+E_1))\cdot T=-2-a<0$ implies $SW(K-([C]-T+E_1))=0$. Then $$|SW([C]-T+E_1)|=|SW(K-([C]-T+E_1))-SW(l[C]-T+E_1)|=|1+a|\ne 0.$$ And $[C]=([C]-T+E_1)+(T-E_1)$ is not extremal.
\bigskip
$\bullet$ $M$ is a non-trivial $S^2$ bundle over $\Sigma_h$, $h\ge 1$ \\
Since the blow-ups of it are diffeomorphic to those of trivial bundle, we are only left with the non-trivial bundles.
Let $U$ be the class of a section with $U^2=1$ and $T$ be the class of the fiber. Then $K=-2U+(2h-1)T$. Let $[C]=aU+bT$. Again, we have $a\ge 0$. The condition $K\cdot [C]<0$ reads as $$2b>a(2h-3).$$
If $a=0$, $$0=C^2=2g(C)-2-K\cdot [C]>2g(C)-2.$$Thus $g(C)=0$ and $K\cdot C=-2$.
If $a>0$ and $h>1$, then we have $b>0$. Hence $C^2=a^2+2ab\ge 1+2a=1+2[C]\cdot T$. Hence $\dim_{SW}([C]-T)=C^2-K\cdot [C]-2-2[C]\cdot T\ge 0$. Moreover $SW(K-([C]-T))= 0$, since $(K-([C]-T))\cdot T=-2-a<0$. Then the wall crossing formula implies $$|SW([C]-T)|=|SW(K-([C]-T))-SW([C]-T)|=|1+a|^h\ne 0.$$
And $[C]=([C]-T)+T$ is not an extremal ray.
If $a>0$ and $h=1$, then we have $$C^2=a^2+2ab>0, \quad -K\cdot [C]=2b+a>0.$$ Look at the classes $l[C]-T$, we have
\begin{equation*} \label{SWdim3}
\begin{array}{lll}
\dim_{SW}(l[C]-T)&=&(l[C]-T)^2-K\cdot (l[C]-T)\\
&&\\
&=&(l^2C^2-lK\cdot [C]) -2-2l [C]\cdot T\\
&&\\
&=&(l^2a+l)(a+2b)-2-2la\\
&&\\
&\ge& l^2a+l-2la-2\\
\end{array}
\end{equation*}
It is greater than $0$ if $l>2a$ is large enough. Moreover $SW(K-(l[C]-T))= 0$, since $(K-(l[C]-T))\cdot T=-2-la<0$. Then the wall crossing formula shows
$$|SW(l[C]-T)|=|SW(K-(l[C]-T))-SW(l[C]-T)|=|1+la|^h\ne 0.$$
Hence $[C]=\frac{1}{l}((l[C]-T)+T)$ is not extremal.
\smallskip
Note our argument shows more specifically that any extremal ray paring negatively with the canonical class $K$ must be generated by either $E_i$, $T-E_i$ or in the minimal case $T$.
\bigskip
\noindent {\bf Case 2: Rational surfaces:}\\
Recall $C$ is an irreducible curve with $K\cdot [C]<0$, which would imply $SW([C])\ne 0$ by Lemma \ref{SW}. Suppose $C$ is not a rational curve with self-intersection $0$ or $-1$. Since $SW([C])\ne 0$ and $K\cdot [C]<0$, by Lemma \ref{int>0}, we know $[C]$ is in the closure of the cone $$P_{K_J}:=\{e\in H^2(M; \mathbb R)|e^2>0, e\cdot E>0 \hbox{ for any } E\in \mathcal E_{K_J}, e\cdot (-K_J)> 0\}$$where $\mathcal E_{K_J}$ is the set of $-1$ symplectic rational curves. Let $S$ be the set of homology classes which are represented by smoothly embedded spheres. We define $$S_{K_J}^+=\{e\in S|g_J(e)=0, e^2>0\}.$$
Using this notation, $$\mathcal E_{K_J}=\{e\in S| g_J(e)=0, e^2=-1\}.$$
By Proposition 5.20 in \cite{LZ-generic}, $P_{K_J}=\mathcal S_{K_J}^+$ where the latter is the open cone spanned by $S_{K_J}^+$ (and furthermore equals to the almost K\"ahler cone when $J$ is good generic). Hence $[C]=\sum a_i [C_i]$ where $a_i>0$ and $[C_i]\in S_{K_J}^+$. Notice we can choose $a_i$ to be rational numbers since all the classes here are rational (in fact, integral). Furthermore, as noted in \cite{LBL},
any class is in $S_{K_J}^{+}$ is Cremona equivalent to one of the following classes
\begin{enumerate}
\item $H$, $2H$,
\item $(n+1)H-nE_1, n\ge 1$,
\item $(n+1)H-nE_1-E_2, n\ge 1$.
\end{enumerate}
Here Cremona equivalence refers to the equivalence under the group of diffeomorphisms preserving the canonical class $K_J$.
It is easy to check that each of them could be written as sum of classes of rational curves with square $0$ or $1$. This implies extremal rays (with $K\cdot [C]<0$) have to be spanned by rational curves with $-3\le K\cdot [C]<0$. There are countably many such classes, since there are countable many $-1$ curve classes. This finishes the proof of the first statement of our cone theorem on extremal rays.
For the finiteness statement, it makes non-trivial sense only when $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k\ge 9$. This is because there are finitely many irreducible curves with $K\cdot [C]<0$. If $M$ is not rational or ruled, they are $-1$ rational curves $E_i$ where $M$. If $M$ is ruled they are $E_i$, $T-E_i$ or $T$ if it is minimal. If $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k<9$, they are $-1$ rational curves (there are finitely many when $k<9$), $H-E$ (when $k=1$) and $H$ (when $k=0$).
Now we are amount to show that when $k\ge 9$ there are only finitely many $-1$ rational curve classes with bounded symplectic energy $[\omega]\cdot E<\frac{1}{\epsilon}$. Since being symplectic is an open condition, $[\omega]-\delta H$ is still a class of symplectic form when $\delta>0$ is small. Moreover $E$ is always represented by a (possibly reducible) symplectic surface since $SW(E)\ne 0$. Hence $[\omega]\cdot E<\frac{1}{\epsilon}$ would imply $$H\cdot E\le\frac{1}{\delta}([\omega]-\delta H)\cdot E+ H\cdot E=\frac{1}{\delta}[\omega]\cdot E<\frac{1}{\epsilon\delta}.$$ On the other hand there are only finitely many classes $E=aH-\sum c_iE_i$ with $E^2=-1$ and $a>0$ is bounded from above. Especially, the finiteness statement implies the extremal rays $\mathbb R^+[L_i]$ are discrete.
To conclude our proof, we first see that $\overline{A}_{K_J}(M)+\sum \mathbb R^+[L_i]\subset \overline{A}_J(M)$ is a closed convex cone. This is because the above discussion implies $\mathbb R^+[L_i]$ can only have accumulate points in $\overline{A}_{K_J}(M)$. On the other hand, to show $\overline{A}_J(M)\subset \overline{A}_{K_J}(M)+\sum \mathbb R^+[L_i]$, we only need to show the inclusion for all classes $e\in \overline{A}_J(M)$ with $K\cdot e<0$. This is exactly what we have proved.
\end{proof}
\begin{remark}
What we have proved is actually a slightly stronger version of cone theorem:
$$\overline{A}^{\mathbb Q}_J(M)=\overline{A}^{\mathbb Q}_{K_J}(M)+\sum \mathbb Q^+[L_i],$$where $\overline{A}^{\mathbb Q}_J(M)=\overline{A}_J(M)\cap H^2(M, \mathbb Q)$ and $\overline{A}^{\mathbb Q}_{K_J}(M)=\overline{A}_{K_J}(M)\cap H^2(M, \mathbb Q)$.
\end{remark}
The next example shows that it is not true that any Seiberg-Witten nontrivial class of nonzero $J$-genus is an integral combination of curve classes.
\begin{example}\label{-K8blowup}
Let $M=\mathbb CP^2\#8\overline{\mathbb CP^2}$. Then $SW(-K)=1$ because of $SW(2K)=0$ and the wall crossing formula. However, $-K$ cannot be written as $m_1[C_1]+m_2[C_2]$ with $m_1m_2\ne 0$ and $m_i\in \mathbb Z$ such that $SW([C_i])\neq 0$ for $i=1,2$. This is because the Seiberg-Witten is a deformation invariant. Especially, $[C_1]$ and $[C_2]$ are classes of (possibly reducible) symplectic surfaces in a Del Pezzo surface, {\it i.e.} a $4$-manifold where $-K$ is the class of the symplectic form. Hence $(-K)\cdot [C_i]\ge 1$ since all the classes are integral. It contradicts to $(-K)\cdot (m_1[C_1]+m_2[C_2])=(-K)^2=1$.
On the other hand, it has the following decomposition with rational coefficients: $$-K=\frac{1}{2}(6H-3E_1-2E_2-\cdots-2E_8)+\frac{1}{2}E_1.$$
\end{example}
We have the following more specific description of the extremal rays.
\begin{prop}\label{exray}
Let $(M, J)$ be a tamed almost complex $4$-manifold. An irreducible curve $C$ is an extremal rational curve as in Theorem \ref{conethm} if and only if
\begin{enumerate}
\item $C$ is a $-1$ rational curve;
\item $M$ is a minimal ruled surface or $\mathbb CP^2\# \overline{\mathbb CP^2}$, and $C$ is a fiber;
\item $M=\mathbb CP^2$ and $C$ is a projective line.
\end{enumerate}
\end{prop}
\begin{proof}
In the above proof of Proposition \ref{conethm}, we see that when $M$ is not rational or ruled, all the extremal rays for non-minimal manifolds are $-1$ classes. For irrational ruled surfaces, our proof shows that any extremal curve class $C$ has $-2\le K\cdot C<0$. The $K\cdot C=-2$ case is when $C$ is the fiber class $T$. However when $M$ is non-minimal, $T=E+(T-E)$ is not extremal. %But there is at least one extremal class as the class of an irreducible component of the subvariety in class $E$ since $K\cdot E=-1<0$.
Similarly for rational surfaces, we know that any square $1$ (resp. $0$) sphere class is Cremona equivalent to $H$ (resp. $H-E$). When $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$, $k\ge 2$, we know both classes could be decomposed into two classes with non-trivial Seiberg-Witten invariant: $H=E+(H-E)$, $H-E=E'+(H-E-E')$. Hence they are not extremal. %Since there is at least one extremal class, it could only be the class of a $-1$ rational curve.
When $M=\mathbb CP^2\# \overline{\mathbb CP^2}$, $H=E+(H-E)$ is the sum of two classes with non-trivial Seiberg-Witten invariant. But it is possible that $H-E$ is the only extremal ray. In this case, the effective class $E$ degenerates as $n(H-E)+((n+1)E-nH)$. Then $(n+1)E-nH$ is the class of a $-(2n+1)$ section for the ruled surface. It corresponds to Hirzebruch surfaces $\mathbb F_{2n+1}$ when $J$ is complex.
\end{proof}
The following lemma gives our information on how a general extremal ray of $A_J(M)$ could be.
\begin{lemma}\label{-extremal}
If $C$ is an irreducible curve with $C^2<0$, then $[C]$ spanned an extremal ray of the curve cone $A_J(M)$.
\end{lemma}
\begin{proof}
If $\mathbb R^+[C]$ is not extremal, then $[C]=\sum a_i [C_i]$ where $a_i>0$ and $C_i$ are irreducible curves whose homology classes are not on $\mathbb R^+[C]$. For those $C_i$, $[C]\cdot [C_i]\ge 0$. Hence we have the following contradiction
$$0\le \sum a_i[C]\cdot [C_i]=[C]\cdot [C]<0.$$
\end{proof}
The following rationality theorem is originally used in algebraic geometry to prove the cone theorem. It is well known that the statement of cone theorem implies the statement of the rationality theorem.
\begin{prop}\label{ratthm}
Let $(M, J)$ be a tamed almost complex $4$-manifold for which $K_J$ is not $J$-nef ({\it i.e.} it pairs non-negatively with the curve cone $A_{J}(M)$). Let $\omega$ be an almost K\"ahler form on $(M, J)$ with $[\omega]\in H^2(M, \mathbb Q)$. Define the nef threshold of $[\omega]$ by $$t_0=t([\omega])=\sup\{t\in \mathbb R: tK_J+[\omega] \hbox{ is $J$-nef}\}.$$Then the nef threshold is a rational number.
\end{prop}
\begin{proof}
It is easy to see that $$t_0=\sup \frac{L_i\cdot [\omega]}{-K_J\cdot L_i}.$$ If there are finitely many extremal curves $L_i$, then $t_0$ is a rational number by this formula. In general, since $K_J$ is not $J$-nef, there exists a small number $\epsilon$ such that $\frac{1}{\epsilon}K+[\omega]$ is not $J$-nef which pairs negatively with only finitely many extremal curves. Hence $t_0$ is the supreme of $ \frac{L_i\cdot [\omega]}{-K_J\cdot L_i}$ for these finitely many $L_i$, which has to be a rational number. In fact, in this case, $t_0$ is an integer since the situation of the second statement of Theorem \ref{conethm} happens only when all $L_i$ are $-1$ rational curves.
Moreover, when $[\omega]\in H^2(M, \mathbb Z)$, then the denominator of $t_0$ is no greater than $3$. It is not an integer only when the last two cases of Proposition \ref{exray} happen.
\end{proof}
\begin{cor}\label{-1}
Let $M=N\#\overline{\mathbb CP^2}$ be a non-minimal symplectic $4$-manifold which is not diffeomorphic to $\mathbb CP^2\# \overline{\mathbb CP^2}$. Then for any tamed $J$ on $M$, there exists at least one smooth $J$-holomorphic $-1$ rational curve.
More precisely, for any exceptional class $E$, we have a decomposition $E=\sum a_i[C_i]$ with $0<a_i\in \mathbb Q$ and $C_i$ irreducible curves, such that there is at least one $-1$ rational curves $C_i$. %Furthermore, all $a_i$ could be chosen as rational numbers.
%More precisely, for any exceptional class $E$, there will be at least one $-1$ rational curves $C_i$ in any decomposition $E=\sum a_i[C_i]$ with $a_i>0$ and $C_i$ irreducible curves.
%as irreducible components of every subvariety representing class $E$.
\end{cor}
\begin{proof}
By the assumption, there is at least one $-1$ rational curve class $E\in H_2(M, \mathbb Z)$. If there is an irreducible $J$-holomorphic subvariety in class $E$, then we are done since it will be smooth by adjunction inequality. If not, since the curve cone is a convex cone, the class is written as $E=\sum a_i[C_i]$ where all $C_i$ are irreducible $J$-holomorphic subvarieties and all $[C_i]$ are extremal curve classes. Furthermore, $a_i\in \mathbb Q$ since all the classes $E$ and $[C_i]$ are in $H_2(M, \mathbb Z)$.
Since $K\cdot E=-1<0$, we know there is at least one $C_i$ (say $C_1$) such that $K\cdot [C_1]<0$. By the cone theorem, this irreducible $C_1$ has to be a rational curve with $-3\le K\cdot [C_1]<0$. %We claim that in our situation,
Moreover, by comparing to the list in Proposition \ref{exray}, $C_1$ has to be the class of a $-1$ rational curve.
\end{proof}
%In the above proof, we see that when $M$ is not rational or ruled, all the extremal rays for non-minimal manifolds are $-1$ classes. For ruled surfaces, our proof shows that any extremal curve class $C$ has $-2\le K\cdot C<0$. The $K\cdot C=-2$ case is when $C$ is the fiber class $T$. However when $M$ is non-minimal, $T=E+(T-E)$ is not extremal. But there is at least one extremal class as the class of an irreducible component of the subvariety in class $E$ since $K\cdot E=-1<0$.
%Similarly for rational surfaces, we know that any square $1$ (resp. $0$) sphere class is Cremona equivalent to $H$ (resp. $H-E$). When $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$, $k\ge 2$, we know both classes could be decomposed into two classes with non-trivial Seiberg-Witten invariant: $H=E+(H-E)$, $H-E=E'+(H-E-E')$. Hence they are not extremal. Since there is at least one extremal class, it could only be the class of a $-1$ rational curve.
%When $M=\mathbb CP^2\# \overline{\mathbb CP^2}$, it is possible that $H-E$ is the only extremal ray. In this case, the effective class $E$ degenerates as $n(H-E)+((n+1)E-nH)$. Then $(n+1)E-nH$ is the class of a $-(2n+1)$ section for the ruled surface. It corresponds to Hirzebruch surfaces $\mathbb F_{2n+1}$ when $J$ is complex.
Notice that the statement that $\mathbb CP^2\# k\overline{\mathbb CP^2}$ has at least one smooth $-1$ rational curve is first proved in \cite{pinn}. Actually, it shows that a class with minimal symplectic energy is such a class. However, our proof gives more precise result.
The second statement of the above corollary is crucial for our later applications.
It is interesting to compare our picture here for a general tamed almost complex structure to the bend-and-break in algebraic geometry. The bend-and-break technique in algebraic geometry starts with an irreducible curve $C'$ with $K\cdot [C']<0$. One chooses a normalization $f: C\rightarrow M$ of $C'$. Then it contains two parts. The first, the ``bend" part, is to compose the normalization with automorphisms of $C$, possibly in characteristic $p$ when $g(C)>1$, such that $-K\cdot f'(C)-g(C)\dim_{\mathbb C} M>0$. This would guarantee one could deform curves in a class which is a multiple of $[C']$. The second, the ``break" part, shows that this family must degenerate to $f''(C)+$(sum of rational curves).
Our argument is sort of a reverse process. We show that all the extremal rays with negative $K$ pairing are spanned by rational curves. And thus a higher multiple of the curve class $C$ with $K\cdot C<0$ will degenerate to a reducible curve with at least one extremal ray as one of its irreducible components.
In general, it is not true that we always have a reducible curve in class $[C]$ if $C$ is an irreducible $J$-holomorphic curve of positive genus such that $K\cdot [C]<0$ as we have seen in Example \ref{-K8blowup}. Here is an example for ruled surfaces.
\begin{example}\label{ruledex}
Let $M$ be the non-trivial $S^2$ bundle over $\Sigma_h$. Let $U$ be the class of a section with $U^2=1$. For $a=\lceil \frac{h-2}{2}\rceil$, we have $SW(U+aT)\ne 0$ and $SW(U+(a-1)T)=0$. Thus for a generic tamed almost complex structure, we do not have reducible curves in class $U+aT$. However, as shown in the proof of Theorem \ref{conethm}, we do have curves in class $lU-T$ when $l>2$. Hence there is always a reducible subvariety in class $l(U+aT)$.
\end{example}
On the other hand, as shown in our proof, the class $[C]$ itself contains reducible curves in many circumstances. We also have the following slight variant of the well known fact that rational or ruled $4$-manifolds are symplectically uniruled.
\begin{prop}\label{ample}
Let $(M, J)$ be a tamed almost complex $4$-manifold with a $J$-ample anti-canonical class $-K$ ({\it i.e.} $-K$ pairs positively with all curves). Then $M$ contains a rational curve. In fact, through every point of $M$ there is a rational curve $C$ such that $$0<-K\cdot [C]\le 3.$$
\end{prop}
\begin{proof}
By Taubes' theorem \cite{T}, when $-K$ is ample, $M$ has to be rational or ruled. Furthermore, it cannot be irrational ruled, otherwise $(-K)\cdot (U+aT)\le 0$ and $SW(U+aT)\ne 0$ where $a=\lceil \frac{h-2}{2}\rceil$ when $M$ is the non-trivial $S^2$ bundle over $\Sigma_h$ and $a=\lceil \frac{h-1}{2}\rceil$ when $M=S^2\times \Sigma_h\# k\overline{\mathbb CP^2}$. Hence, $M$ is rational and we choose homology basis such that $K=-3H+E_1+\cdots +E_k$.
Since $K\cdot [C]<0$ for all irreducible curves, then either the curve is a $-1$ rational curve or it has $(K-[C]) \cdot [D]<0$ for any irreducible curve (and then any subvariety) $D$. In both cases, $\dim_{SW}([C])=C^2-K\cdot [C]=2g-2-2K\cdot [C]\ge 0$. Hence $SW([C])\ne 0$ by Lemma \ref{int>0} and the wall-crossing formula.
Since $SW([C])\ne 0$ and $H$ is represented by an irreducible $J'$-holomorphic curve with positive self-intersection for a generic $J'$, by Lemma \ref{int>0}, $H\cdot [C]\ge 0$ for all curves $C$.
In other words, $H$ is $J$-nef when $-K$ is $J$-ample. Then by Theorem 1.5 of \cite{LZrc}, we know any irreducible component $C_i$ of a subvariety in class $H$ is a rational curve with $0<-K\cdot [C_i]<-K\cdot H=3$. Then the conclusion follows since there is a subvariety in class $H$ passing through any given point.
\end{proof}
\section{$\mathbb CP^2\#2\overline{\mathbb CP^2}$}\label{k=2}
This section serves as a link between Sections 2 and 4. We first recall some general results for the homology classes of irreducible subvarieties on $\mathbb CP^2\#k\overline{\mathbb CP^2}$. Then we apply them to give an explicit description of the negative curves on $\mathbb CP^2\#2\overline{\mathbb CP^2}$. In particular, we show there are at least two smooth $-1$ rational curves for any tamed almost complex structure on it in Theorem \ref{cp2+2}. A full description of the curve cone is given in Theorem \ref{cp2+2cone}.
The information on the curve cone helps us to obtain a Nakai-Moishezon type duality, Theorem \ref{NM2}. %Finally, we will discuss the configuration of negative curves for complex structures.
\subsection{The curve cone and the $K-$symplectic cone}
The $K-$symplectic cone for a class $K\in H^2(M;\mathbb Z)$ introduced in \cite{LL}:
\begin{equation}\label{Kcone}
\mathcal C_{M, K}=\{e\in H^2(M;\mathbb R)|e=[\omega] \hbox{ for some $\omega$ with $K_{\omega}=K$}\}.
\end{equation}
Here $K_{\omega}$ is the symplectic canonical class of $\omega$. Suppose $\mathcal C_{M,K}$ is non-trivial, $[\omega]\in \mathcal C_{M,K}$ and $b^+(M)=1$, by Theorem 3 in \cite{LL},
\begin{equation} \label{cmk}\mathcal C_{M,K}=\{e\in \mathcal{FP}(K)|e\cdot E>0, E \in \mathcal E_{K}
\}.\end{equation}
Here $\mathcal{FP}(K)$ is the connected component of $\mathcal{P}=\{e\in H^2(M, \mathbb Z)|
e^2>0 \}$ containing $[\omega]$. Notice by the light cone lemma, both $\mathcal{FP}(K)$ and $\mathcal
C_{M,K}$ are convex cones. Recall the following statement which is called the light cone lemma in the literature, which is in the guise of Cauchy-Schwartz inequality. The cone of elements with positive squares have two components. The forward cone means one of the two connected components containing a given element with positive square. In our applications, it is usually a class of symplectic form.
\begin{lemma}[light cone lemma]\label{light}
For the light cone of signature $(1, n)$ ($n\ne 0$), any two elements in the forward cone have non-negative dot product. Especially, if the dot product is zero then the two elements are proportional to each other.
\end{lemma}
We will state a structural description of the $K-$symplectic of rational surfaces, which might be known for experts. Without loss, we suppose $K=-3H+\sum_iE_i$.
\begin{prop}\label{corner01}
Let $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$.
\begin{enumerate}
\item When $k<9$, the $K-$symplectic cone is a cone over a polytope whose corners are the classes of the symplectic spheres with canonical class $K$ and self-intersection $0$ or $1$.
\item When $k\ge 9$, all the extremal rays $\mathbb R^+e$ of the $K-$symplectic cone having $K\cdot e<0$ are generated by the classes of the symplectic spheres with canonical class $K$ and self-intersection $0$ or $1$.
\end{enumerate}
\end{prop}
\begin{proof}
We first assume $k<9$. The $K-$symplectic cone is a polytope follows from Proposition 2.7(1) in \cite{FM}. In fact, the $K-$symplectic cone is a $P$-cell when $k<9$, and a $P$-cell has no round boundary. One could also see Section 5.4 of \cite{LZ-generic} for an overview of the results in \cite{FM} using notations similar to this paper's. For a more direct argument, see \cite{McSc}.
We now show that the corners of the polytope, which correspond to the extremal rays of the $K-$symplectic cone, are those classes which can be represented by the symplectic spheres with canonical class $K$ and self-intersection $0$ or $1$. This statement follows from the light cone lemma. The corners of the polytope are the intersection of several hyperplanes determined in our situation by classes in $\mathcal E_K$. Recall the $K-$symplectic cone is a cone over a polytope implies that the corners of the polytope are in $\overline{\mathcal P}$.
There are two types of corners. If the classes determine the hyperplanes around the corner ray $\mathbb R^+e$ are orthogonal to each others, without loss we can assume them to be $E_1, \cdots, E_l$. We know $l=k$ otherwise the intersection would include any classes $aH-bE_k$ with $a\ge b>0$, hence is not a ray. But when $l=k$, the primitive class corresponding to the corner is a class Cremona equivalent to $H$, {\it i.e.} the class of a symplectic sphere with self-intersection $1$.
If two of the classes determine the hyperplanes around a corner are not orthogonal, say $E$ and $E'$ (both in $\mathcal E_k$), then the corner class $e$ is also orthogonal to $E+E'$ which is of non-negative self-intersection. By the light cone lemma, $e$ is proportional to $E+E'$ and $(E+E')^2=0$. The equality happens only if $E\cdot E'=1$ and $(E+E')^2=0$. The equality happens only if $E\cdot E'=1$ and $E+E'$ is the class of a symplectic sphere of self-intersection $0$ (since $K\cdot (E+E')=-2$). If we choose $e$ to be primitive, then $e=E+E'$.
When $k\ge 9$, by Proposition 2.7 of \cite{FM}, the intersection $\mathcal C_{M, K}\cap \{e\in H^2(M; \mathbb R)| K\cdot e\le 0\}$ is a $P$-cell which has no round boundary. Hence, the extremal rays with $K\cdot e< 0$ corresponding to intersections of the ordinary walls ({\it i.e.} the hyperplanes determined by classes in $\mathcal E_K$ rather than the canonical class $K$). Then we have exactly the same two types of corners, and the same argument applies.
\end{proof}
\begin{remark}\label{k>9round}
The $K-$symplectic cone has round boundary when $k>9$.
For examples, one can take the anti-canonical class $e=-K$. The class $e\notin \mathcal C_{M, K}$ when $k>9$ since $e^2<0$ but $e\cdot E=1>0$ for all $E\in \mathcal E_K$.
Moreover, one can indeed get an open set of round boundary. Since $\mathcal C_{M, K}\ne \emptyset$, we choose an open set $B$ of it.
A class $-K+a_{\omega}[\omega]$, where $a_{\omega}>0$ and $[\omega]\in B$, is on the round boundary of $\mathcal C_{M, K}$ if $a_{\omega}$ solves $(-K+a_{\omega}[\omega])^2=0$ (since $(-K+a_{\omega}[\omega])\cdot E>0$ for all $E\in \mathcal E_K$). This is a quadratic equation and such $a_{\omega}$ exists since $K^2\cdot [\omega]^2<0$.
For a detailed discussion on the round boundary, see the new edition of \cite{McS}.
\end{remark}
The $K-$symplectic cone can be used to restrict the classes of negative square in the curve cone. The following lemma is simple but also very useful.
\begin{lemma} \label{curvecone}
Suppose $b^+(M)=1$. A cohomology class
is in the curve cone only if it is positive on some extremal ray of the $K-$symplectic cone.
\end{lemma}
\begin{proof}
Assume the class $e$ is in the curve cone $A_J(M)$ for some tamed almost complex structure $J$ with $K_J=K$. Let $\omega$ be a symplectic form taming $J$. Then $e\cdot [\omega]>0$. Since $\mathcal C_{M, K}$ is convex, there is an extremal ray pairing positively with $e$.
\end{proof}
%Remember, without loss, we always suppose $K=-3H+\sum_iE_i$. When $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$ with $k<9$, the $K-$symplectic cone is a cone over a polytope whose corners are primitive classes. This follows from Proposition 2.7(1) in \cite{FM}. In fact, the $K-$symplectic cone is a $P$-cell when $k<9$, and a $P$-cell has no round boundary. The corners of the polytope, which correspond to the extremal rays of the $K-$symplectic cone, are those classes which can be represented by the symplectic spheres with canonical class $K$ and self-intersection $0$ or $1$. This statement follows from the light cone lemma. The corners of the polytope are the intersection of several hyperplanes determined in our situation by classes in $\mathcal E_K$. Recall the $K-$symplectic cone is a cone over a polytope implies that the corners of the polytope are in $\overline{\mathcal P}$. There are two types of corners. If the classes determine the hyperplanes around the corner ray $\mathbb R^+e$ are orthogonal to each others, without loss we can assume them to be $E_1, \cdots, E_l$. We know $l=k$ otherwise the intersection would include any classes $aH-bE_k$ with $a\ge b>0$, hence is not a ray. But when $l=k$, the corner is a class Cremona equivalent to $H$, {\it i.e.} the class of a symplectic sphere with self-intersection $1$. If two of the classes determine the hyperplanes around a corner are not orthogonal, say $E$ and $E'$, then the corner class $e$ is also orthogonal to $E+E'$ which is of non-negative self-intersection. By the light cone lemma, $e$ is proportional to $E+E'$ and $(E+E')^2=0$. The equality happens only if $E\cdot E'=1$ and $E+E'$ is the class of a symplectic sphere of self-intersection $0$.
The next lemma is on the constraints of the curve classes
provided by the adjunction inequality.
\begin{lemma}\label{B}
Suppose a class $B=\alpha H+\sum \beta_i E_i$ has
an irreducible curve representative.
If $\alpha>0$ then $ |\beta_i|\leq |\alpha|$ for each $i$, and $|\alpha|=|\beta_i|$ only when $\alpha=-\beta_i=1$.
If $\alpha=0$, then $B=E_i-\sum_jE_{k_j}$.
If $\alpha< 0$, then $ |\beta_i|\leq |\alpha|+1$, and $|\beta_i|=|\alpha|+1$ only if $\beta_i=-\alpha+1$.
\end{lemma}
\begin{proof}
By adjunction formula, and the fact that $\gamma^2+\gamma\geq 0$ for any integer $\gamma$, we have
$$(\alpha-1)(\alpha-2)\geq \beta_i(\beta_i+1).$$
Then all the conclusions are clear when $\alpha\neq 0$.
For $\alpha=0$, we first get that $\beta_i=0$, $-2$ or $\pm 1$. However when $\alpha=0$, some $\beta_i$ is non-negative. Otherwise, $J$ will not be tamed by lemma \ref{curvecone}. Then $B=E_i-\sum_jE_{k_j}$ holds since $\sum \beta_i(\beta_i+1)\le 2$ by adjunction formula.
\end{proof}
\subsection{Curves on $\mathbb CP^2\#2\overline{\mathbb CP^2}$}
We have shown that there is at least one smooth $J$-holomorphic $-1$ rational curve for any tamed $J$ on non-minimal symplectic manifold except for $\mathbb CP^2\#\overline{\mathbb CP^2}$ in Corollary \ref{-1}. Now, we will show that when $M=\mathbb CP^2\#2\overline{\mathbb CP^2}$, we actually have at least two $-1$ sphere classes. Notice it is not true when $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$ for $k>2$.
\begin{theorem}\label{cp2+2}
There are at least two smooth $-1$ $J$-holomorphic rational curves for any tamed $J$ on $\mathbb CP^2\#2\overline{\mathbb CP^2}$.
\end{theorem}
\begin{proof}
First, there is at least one $-1$ smooth rational curve by Corollary \ref{-1}. We first assume the class $E_2$ has such a smooth representative.
\begin{lemma}\label{A} Suppose $M=\mathbb CP^2\#2\overline{\mathbb CP^2}$ and the class $E_2$ is the class of a smooth $-1$ rational curve.
If a class $A=aH+bE_1+cE_2\ne E_2$ with %$A\cdot A<0$ and
$a\le 0$ has an irreducible curve representative, then
(i) $b>0, a=1-b$ and $c=0$ or $-1$.
(ii) $A$ is a sphere class.
(iii) $A\cdot A<0$.
(iv) $A$ is the only such class.
\end{lemma}
\begin{proof}
First $c\leq 0$ since $A\cdot E_2=-c\geq 0$.
First, the set $\mathcal E_K=\{E_1, E_2, H-E_1-E_2\}$. It could be checked directly by adjunction formula, or from section 4 (in particular, Lemma \ref{a<0} and Proposition \ref{a>0}). Then by \eqref{cmk}, the extremal rays of the $K-$symplectic cone are spanned by
$$H, \quad H-E_1, \quad H-E_2.$$
As $a\le 0$ and $c\le 0$, we have
$A\cdot H\le 0$ and $A\cdot (H-E_2)\le 0$. Therefore by Lemma \ref{curvecone} $A\cdot (H-E_1)$ is strictly positive. This means that
$a+b>0$, i.e. $b>-a\ge 0$.
By the adjunction formula
\begin{equation}
(a-1)(a-2)\geq b(b+1)+c(c+1).
\end{equation}
The only possibility is as claimed in $(i)$, $A= (1-b)H+ bE_1$ or $A=(1-b)H+bE_1-E_2$ with $b\ge 1$. Items $(ii)$ and $(iii)$ are then direct to check.
For $(iv)$, suppose $A'=(1-b')H+b'E_1 $ or $(1-b')H+b'E_1-E_2$ is another such class. Then $A\cdot A'=1-(b+b')$ or $-(b+b')$ is negative.
\end{proof}
\begin{lemma}\label{cca<0}
Suppose $M=\mathbb CP^2\#2\overline{\mathbb CP^2}$ and the class $E_2$ is the class of a smooth $-1$ rational curve. If $A=(1-s)H+sE_1$ or $(1-s)H+sE_1-E_2$ with $s\ge 1$ is in the curve cone, then a class $D=H+vE_1+wE_2$ is in the curve
cone only if $v> -1$ or $v=-1, w\ge -1$.
\end{lemma}
\begin{proof}
Since $s\ge 1$, $D$ must be of the form $pA+\sum B_i$, where $B_i=\alpha_iH+\beta_iE_1+\gamma_iE_2$ and $p\ge 0$. By Lemma \ref{A} $(iv)$, we have $\alpha_i>0$ or $B_i=E_2$. Then pairing with $H$, we have
$$1=p(1-s)+\sum \alpha_i.$$ Now pairing with $E_1$, we have
$$v=ps+\sum \beta _i=p-1+ \sum (\alpha _i+\beta_i)\geq -1$$ by Lemmas \ref{B} and \ref{A}. Moreover, $v=-1$ only if $p=0$, and by Lemma \ref{B}, we have $$w=\sum \gamma_i \ge -\sum \alpha_i=-1.$$
\end{proof}
\begin{cor}\label{1to2}
Suppose $E_2$ is the class of a smooth $-1$ rational curve. Then so does either the class $H-E_1-E_2$ or the class $E_1$. And in the latter case,
both $E_1$ and $H-E_1-E_2$ are the classes of smooth $-1$ rational curves.
\end{cor}
\begin{proof}
Suppose $E_2$ has an embedded representative and $E_1$ does not.
Note the class $H-E_1-E_2$ is in the curve cone since $SW(H-E_1-E_2)\ne 0$.
By Corollary \ref{-1}, there is a $-1$ rational curve $C_i$ in a decomposition $H-E_1-E_2=\sum_ia_i[C_i]$ (in other words, an irreducible component of a $\mathbb R$-variety in class $H-E_1-E_2$).
By our assumption, this class cannot be $E_1$.
If this class is $H-E_1-E_2$, we are done.
If this class is $E_2$, then $H-E_1-lE_2$, with $l>1$, is in the curve cone. If we have an irreducible curve in class $(1-s)H+sE_1$ or $(1-s)H+sE_1-E_2$ with $s\ge 1$, it will contradict to the Lemma \ref{cca<0}. So all irreducible curves $C$ have $a=C\cdot H> 0$. By Lemma \ref{B}, all such irreducible curves $aH-b_1E_1-b_2E_2$ have $b_2\le a$. Hence $l\le 1$.
If both $E_1$ and $E_2$ have embedded representatives, same argument shows that neither can appear in the decomposition of $H-E_1-E_2$.
\end{proof}
Now, to finish the proof of Theorem \ref{cp2+2}, we are left with case that the class $H-E_1-E_2$ has an embedded representative.
By Corollary \ref{-1}, there will be a $-1$ rational curve as an irreducible component of the $\mathbb R$-subvariety representing class $E_1$ or $E_2$. Suppose that there is no irreducible curve with non-positive $H$ coefficient. Then
$H-E_1-E_2$ cannot appear as the class of such $-1$ rational curve. Thus
the $-1$ class component in $E_1$ is either $E_1$ or $E_2$. Hence in this situation,
there are at least two $-1$ rational curves.
Thus we assume that there is an irreducible curve in class
$A=aH+bE_1+cE_2$ with $a\le 0$.
Since $a\le 0$, either $a+b>0$ or $a+c>0$ by Lemma \ref{curvecone}.
Without loss, we assume that $a+b>0$. We will show that in this case $E_1$ or $E_2$ must have an embedded representative.
First by adjunction formula $$(a-1)(a-2)\ge b(b+1)+(c^2+c).$$
On the other hand $b\ge -a+1>0$. Hence the only possibility for the adjunction holds would be $a=-b+1$, $c=0$ or $-1$ and $g=0$. Then $A= (1-b)H+ bE_1$ or $A=(1-b)H+bE_1-E_2$ with $b\ge 1$.
If $E_1$ and $E_2$ do not have irreducible representative, then $H-E_1-E_2$ is the only class of extremal irreducible curve with $K\cdot C<0$ as shown in Corollary \ref{-1}. We now look at irreducible curves with $K\cdot C\ge 0$. First let $[C]=aH-b_1E_1-b_2E_2, a>0$. Then $K\cdot C\ge 0$ implies $$0<3a\le b_1+b_2.$$But by local positivity of intersections, $C\cdot (H-E_1-E_2)\ge 0$, which is $a\ge b_1+b_2$. It is a contradiction. Hence $a\le 0$ and the curves classes are calculated as above. Hence $E_i$ will be a linear combination of these classes and $H-E_1-E_2$. If we write a class as $aH+b_1E_1+b_2E_2$, then all the above classes will contribute non-positively to $2a+b_1+b_2$. However $E_i$ has positive $2a+b_1+b_2$. This is a contradiction. Hence there is an irreducible curve in class $E_1$ or $E_2$.
\end{proof}
The above discussion actually gives the following description of the curve cone. By Theorem \ref{cp2+2}, there is always an irreducible $J$-holomorphic curve in class $E_1$ or $E_2$. Hence, without loss, we could assume $E_2$ has an irreducible representative.
\begin{theorem}\label{cp2+2cone}
Let $J$ be a tamed almost complex structure on $\mathbb CP^2 \# 2\overline{\mathbb CP^2}$ such that there is a smooth $J$-holomorphic curve in the class $E_2$. Then the curve cone $A_J(M)$ is generated by $3$ classes. They are either
$$\alpha=(1-s)H+sE_1, \quad \beta=E_2, \quad \gamma=H-E_1-E_2,$$
or
$$ \alpha=(1-s)H+sE_1-E_2, \quad \beta=E_2, \quad \gamma=H-E_1-E_2,$$ where $s\ge 1$.
\end{theorem}
\begin{proof}
First by Lemma \ref{1to2}, there is always an irreducible curve in class $H-E_1-E_2$. By the argument in the last paragraph of the proof of Theorem \ref{cp2+2}, curve classes $C$ with $C\cdot H>0$ is always spanned by $H-E_1-E_2$, one of $E_i$ say $E_2$ and another irreducible curve with non-positive pairing with $H$. When $E_2$ has irreducible representative, the last curve class is $(1-s)H+sE_1$ or $(1-s)H+sE_1-E_2$ with $s\ge 1$. By Lemma \ref{A} (iv), such a curve is unique. This completes our proof.
\end{proof}
Now we can study the $\geq 0-$dual of the curve cone $A_J(M)$.
If $A_J$ is generated by $E_1, E_2, H-E_1-E_2$, then its $\geq 0-$dual is generated by
$H, H-E_1, H-E_2$.
Let us assume that $E_2$ is irreducible. We discuss the two cases in Theorem \ref{cp2+2cone}.
In the first case, the $\geq 0-$dual of $A_J$ is generated by $$A=sH-(s-1)E_1, \quad B=H-E_1, \quad C=sH-(s-1)E_1-E_2.$$
In the second case, the $\geq 0-$dual of $A_J$ is generated by $$A=sH-(s-1)E_1, \quad B=H-E_1, \quad C=(s+1)H-sE_1-E_2.$$
In both cases, $A$ and $C$ are in $S_{K_J}^+$ and $B$ is approximated by the sequence $pH-(p-1)E_1-E_2, p\to \infty$ in $S_{K_J}^+$.
All the above actually shows the following
\begin{prop}\label{s=p}
For any tamed $J$ on $\mathbb CP^2 \# 2\overline{\mathbb CP^2}$, we have $$\mathcal S_J=\mathcal P_J.$$
\end{prop}
Recall that $\mathcal P_J=A_J^{\vee, >0}(M) \cap \mathcal P$. The spherical cone $\mathcal S_J$ here is defined to be the interior of the convex cone generated by big $J$-nef classes ({\it i.e.} $J$-nef classes with positive square) in $S_{K_J}$ if it is of dimension $3$.
\subsection{Nakai-Moishezon type theorem for almost K\"ahler structure on $\mathbb CP^2 \# 2\overline{\mathbb CP^2}$}
With Proposition \ref{s=p} in hand, we can establish the Nakai-Moishezon and Kleiman type theorems for almost K\"ahler $J$ on $\mathbb CP^2 \# 2\overline{\mathbb CP^2}$.
\begin{theorem}\label{NM2}
Let $M=\mathbb CP^2\# 2\overline{\mathbb CP^2}$. For any almost K\"ahler $J$,
the $J$-compatible cone $\mathcal K_J^c(M)$
is dual to the $J$-curve cone $A_J(M)$, i.e.
$$\mathcal K_J^c= \mathcal P_J=A_J^{\vee, >0}(M) .$$
\end{theorem}
\begin{proof}
It is clear that $\mathcal K_J^c\subset \mathcal P_J.$ If $J$ is almost K\"ahler, we have $\mathcal S_J\subset \mathcal K_J^c$ (Lemma 5.18 in \cite{LZ-generic}). By Proposition \ref{s=p}, we have $$\mathcal K_J^c= \mathcal P_J=\mathcal S_J.$$
The second equality $\mathcal P_J=A_J^{\vee, >0}(M)$ holds because the classes $A, B, C$ and thus their positive combinations all have non-negative squares.
\end{proof}
We remark that the techniques in \cite{LZ-generic} to construct almost K\"ahler form for a tamed $J$ fail in this situation.
First we need to construct a Taubes current. Here a current is a differential form with distribution coefficients. Hence it represents a real cohomology class when pairing with smooth closed forms in the weak sense. A Taubes current is a closed, positive $J$-invariant current $\Phi$, which satisfies $$k^{-1}t^4\le \Phi(if_{B_t(x)}\sigma\wedge \bar{\sigma})<kt^4.$$ Here $\sigma$ denotes a point-wise unit length section of $T^{1,0}M|_{B_t(x)}$.
The usual technique for the construction is to integrate certain part of the moduli space of the subvarieties in a $J$-ample class $e$, {\it i.e.} a cohomology class pairing positively with any curve classes, with $g_J(e)=0$.
However in general, as in current situation, we do not have any such classes. We may only have big $J$-nef spherical classes. In this situation, we are still able to produce a weak version of Taubes current in class $e$: a closed, non-negative $J$-invariant current $\Phi_e$, satisfying $$0\le \Phi_e(if_B\sigma\wedge \bar{\sigma})<kt^4.$$ It will vanish along the vanishing locus $Z(e)$, {\it i.e.} the union of irreducible subvarieties $D_i$ such that $e\cdot D_i=0$. But over any $4$-dimensional compact submanifold with boundary $K$ of the complement $M(e)=M\backslash Z(e)$, it is a Taubes current with the constant $k>1$ depending only on $K$.
If we have sufficiently many big $J$-nef classes, we could produce genuine Taubes currents by the following Proposition 5.7 in \cite{LZ-generic}.
\begin{prop}\label{form}
Let $e_i$ be big $J$-nef classes in $S_{K_J}$ and the zero locus of $e_i$ is denoted by $Z_i$.
If $\cap Z_i =\emptyset$, then there is a Taubes current in the class $e=\sum_i a_ie_i$, with $a_i>0$.
\end{prop}
Finally, we apply the following regularization result of \cite{T2009} (see also \cite{Zmann}) to obtain an almost K\"ahler form in the class $e$.
\begin{theorem}
In a $4$-manifold $M$ with $b^+(M)=1$, if we have a Taubes current $T$,
then there is an almost K\"ahler form $\omega$, s.t. $[\omega]=[T]$.
\end{theorem}
Hence to construct an almost K\"ahler form by the subvariety-current-form method, we are reduced to prove that there exist big $J$-nef classes $e_i$ in $S_{K_J}$, such that the intersection of the zero locus $\cap Z(e_i) =\emptyset$. We claim it is impossible if our $J$ is assumed only to be tamed. In the below, $\alpha, \beta, \gamma$ are those classes in Theorem \ref{cp2+2cone}. The main point is there is no class $e\in S_J^+$ such that $Z(e)=\alpha$ by simple homological calculation. Since $\beta\cdot \gamma=1$, we have $\beta\cap \gamma\ne \emptyset$. Because of the above observation, any class $e\in S_J^+$ will have $\beta\cap \gamma \subset Z(e)$. Hence $\beta\cap \gamma \subset \cap Z(e_i)$, which is then not empty.
\section{Configurations of negative curves on rational and ruled surfaces}
Almost complex structures are different from complex structures at blowing up and down. More precisely, when we have an irreducible holomorphic $-1$ sphere, we can always blow it down by Castelnuovo's criterion. However, generally we cannot blow down a smooth $J$-holomorphic $-1$ rational curve for a tamed almost complex structure $J$.
In this section, we study the negative curves in a tamed almost complex rational or ruled $4$-manifolds, which might not be mentioned explicitly in each statement.
\subsection{Negative curves on $\mathbb CP^2\#k\overline{\mathbb CP^2}$ with $k\le 9$}
In this subsection, we will study the negative curves on rational surfaces. We will show that they have to be spheres for $\mathbb CP^2\#k\overline{\mathbb CP^2}$ with $k\le 9$. This will enable us to determine the curve cone and show that the configurations of negative curves for tamed almost complex structures are all realized by complex structures (Theorem \ref{main1}). Note that the case of $S^2\times S^2$ and $\mathbb CP^2\# \overline{\mathbb CP^2}$ have been done in \cite{LZ-generic}, $\mathbb CP^2\#2\overline{\mathbb CP^2}$ is done in section \ref{k=2}.
We first take a look at the irreducible curve classes $C=aH+\sum b_iE_i$ with $C^2<0$ and $a<0$.
\begin{lemma}\label{a<0}Let $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$.
If $C=aH+\sum b_iE_i$ with %$C^2<0$ and
$a<0$ is represented by an irreducible curve, then
\begin{itemize}
\item $C=-nH+(n+1)E_1-\sum_{k_j\neq 1}E_{k_j}$ up to diffeomorphism.
\item Or $C=f^*C'$, where $f$ is a diffeomorphism and $C'$ is a class with $a'>0$.
\end{itemize}
\end{lemma}
\begin{proof}
The proof is eventually similar to that of Lemma \ref{A}. As we suppose our canonical class $K=-3H+\sum E_i$, there are two types of the generators of extremal rays of the $K-$symplectic cone. The first type is the classes $F$ with $F^2=1$ which can be represented as a sphere. By \cite{LBL}, those are Cremona equivalent to $H$.
The second type of classes are those Cremona equivalent to $H-E_1$.
Let us first suppose that $C$ pairs non-positively with all the classes of the first type. By Lemma \ref{curvecone}, at least one of these classes equivalent to $H-E_1$ pairs positively with $C$. If we suppose it is $H-E_1$, then we know that $b_1>-a>0$. Then by the adjunction formula $(a-1)(a-2)\ge \sum b_i(b_i+1)$. Thus, the only possibility is as claimed, $C=-nH+(n+1)E_1-\sum_jE_{k_j}$ which has $g_J(C)=0$.
If $C\cdot F>0$ for some $F$ of the first type, then we can first change the class $F$ to $H$ by a diffeomorphism $f$ preserving the canonical class. The class $C$ changes to $C'$ at the same time and thus $C'\cdot H>0$. Thus $C'$ is a class with $a'>0$
and $C$ is pull-back of it by a diffeomorphism.
\end{proof}
The latter case could happen. For example when $C=-H+E_1+E_2+E_3$, $F=2H-E_1-E_2-E_3$. Then $C$ is equivalent to $H-E_1-E_2-E_3$ after a Dehn twist along $H-E_1-E_2-E_3$.
However, all $-1$ rational curve classes $C=aH+\sum b_iE_i$ have $a\ge 0$. This can be seen by applying Lemma \ref{int>0} to the Seiberg-Witten nontrivial classes $H$ and $C$.
The case when $a=0$ is investigated in Lemma \ref{B}. The only possible curves are $E_i-\sum_{k_j\neq i}E_{k_j}$.
Now, let us take a look at the case of $a>0$.
\begin{prop} \label{extreme}
Suppose $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$, $k\le 9$.
\begin{enumerate}
\item Then any irreducible curves $C$ with $C^2<0$ are smooth spheres.
\item If $k\le 8$, any irreducible curves with $C^2\le 0$ are smooth spheres.
\end{enumerate}
\end{prop}
\begin{proof}
Let $C=aH-\sum b_iE_i$. The case when $a\le 0$ is discussed above. The only undetermined case, the second bullet of Lemma \ref{a<0} is reduced to the case of $a>0$. Hence, we suppose $a \ge1$ below.
Because $C^2<0$, we can suppose
\begin{equation}\label{squ-}
c^2+a^2=b_1^2+\cdots b_k^2, \quad c\in \mathbb R\backslash \{0\}.
\end{equation}
In other words, $c^2=-C^2$.
If $C$ is an irreducible curve and is not a sphere, then by adjunction formula,
\begin{equation}\label{adj-}
c^2+3a\le b_1+\cdots+b_k.
\end{equation}
Hence, by Cauchy-Schwartz inequality, we have $(c^2+3a)^2\le k(c^2+a^2)\le 9(c^2+a^2)$, {\it i.e.} $$6ac^2+c^4-9c^2\le 0.$$ This inequality is possible only when $a=1$. Then \eqref{adj-} becomes $$b_1+\cdots +b_k\ge 3+c^2.$$ It contradicts to \eqref{squ-}, which reads as $$3+c^2\le b_1+\cdots +b_k\le b_1^2+\cdots +b_k^2= 1+c^2.$$
This contradiction shows that $C$ should be a sphere.
We also notice that when $k\le 8$, any irreducible curves with $C^2\le 0$ are spheres. This is because if $C^2=0$, formulae \eqref{squ-}, \eqref{adj-} lead to the contradiction $$ka^2\ge (3a)^2.$$
\end{proof}
Notice that the statement is sharp, in the sense that it is no longer true for $k\ge 10$ (resp. $k\ge 9$), since the anti-canonical class $K$ could be the class of an elliptic curve with $K^2<0$ (resp. $K^2\le 0$).
It is true that these negative curves form the extremal rays of the curve cone.
\begin{prop}\label{dA}
The curve cone $A_J(M)$ of $(M=\mathbb CP^2\#k\overline{\mathbb CP^2}, J)$, $k< 9$, is a polytopic cone generated by the classes of spheres with non-positive self-intersections.
\end{prop}
\begin{proof}
By Proposition \ref{extreme}, we just need to prove that any irreducible curve $C$ with $C^2>0$ cannot span an extremal ray of the curve cone $A_J(M)$.
We look at the class $n[C]$. We know that when $k\le 8$, there are only finitely many elements in $\mathcal E_K$. This is because first there are finitely many classes in $\mathcal E_K$ such that its pairing with $H$ is non-negative ({\it e.g.} Proposition \ref{a>0}). Second, since there are finitely many classes $F$ such that $F^2=1$ and $K\cdot F=-3$ when $k\ge 8$, the action of the Cremona group which gives the classes in the second bullet of Lemma \ref{a<0} have all orbits finite (see \cite{LBL}).
Now we assume the maximal possible pairing of $C$ with elements in $\mathcal E_K$ is $l$. We choose $n$ large enough such that
\begin{equation} \label{n2}
n^2C^2-n(K\cdot [C]+2l)-2>0, \quad n^2C^2-2n(K\cdot [C]+l)+6-k>0.
\end{equation}
This is possible, since the coefficient of the quadratic term is positive.
Now we claim $[C]=\frac{1}{n}((n[C]-E)+E)$, $E\in \mathcal E_K$, gives a decomposition with both classes having nontrivial Seiberg-Witten. Hence $[C]$ is not extremal.
First, we check the SW dimension
\begin{equation*}
\begin{array}{lll}
\dim_{SW}(n[C]-E)&=&(n[C]-E)^2-K\cdot (n[C]-E)\\
&&\\
&=&n^2C^2-nK\cdot [C] -2n [C]\cdot E-2\\
&&\\
&>&0\\
\end{array}
\end{equation*}
Take a symplectic form $\omega$ taming $J$. If $SW(K-(n[C]-E))\ne 0$, then $[\omega]\cdot (K-(n[C]-E))>0$. We also have $(K-(n[C]-E))^2>0$ by the second inequality of \eqref{n2}. On the other hand, by our assumptions, $[\omega]\cdot [C]>0$ and $C^2>0$. Hence by the light cone lemma, we have $(K-(n[C]-E))\cdot [C]\ge 0$, which contradicts to the first inequality of \eqref{n2}. Hence we have $SW(K-(n[C]-E))=0$. By wall crossing, $$|SW(n[C]-E)|=|SW(n[C]-E)-SW(K-(n[C]-E))|=1.$$ Apparently $SW(E)=1$. Hence $[C]=\frac{1}{n}((n[C]-E)+E)$ is not extremal.
Finally, since we have classified all the classes of irreducible curves with non-positive self-intersections when $k<9$ in Proposition \ref{a<0} and \ref{a>0}. Especially, there are finitely many such classes. Hence our conclusion follows. In particular, our curve cone has no round boundary.
\end{proof}
This should be compared with Lemma \ref{-extremal}.
%On the other hand, we have the following general fact.
Next, let us classify the negative irreducible curves with $a>0$ on $\mathbb CP^2\#k\overline{\mathbb CP^2}$ with $k<9$.
\begin{prop}\label{a>0}
Let $J$ be a tamed almost complex structure on $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$, $k< 9$, and $C=aH-\sum b_iE_i$ be an irreducible curve with $C^2< 0$, $a>0$. Then $[C]$ is one of the following:
\begin{enumerate}
\item $H-\sum E_{k_j}$;
\item $2H-\sum E_{k_j}$;
\item $3H-2E_m-\sum_{k_j\neq m} E_{k_j}$;
\item $4H-2E_{m_1}-2E_{m_2}-2E_{m_3}-\sum_{k_j\neq m_i} E_{k_j}$;
\item $5H-E_{m_1}-E_{m_2}-\sum_{k_j\neq m_i} 2E_{k_j}$;
\item $6H-3E_{m_1}-\sum_{k_j\neq m_1} 2E_{k_j}$.
\end{enumerate}
\end{prop}
\begin{proof}
Similar to Proposition \ref{extreme}, we have
$$c^2+a^2=b_1^2+\cdots b_k^2,$$
$$-2+c^2+3a\le b_1+\cdots+b_k.$$
Now, $(c^2+3a-2)^2\le k(c^2+a^2)\le 8(c^2+a^2)$ holds by Cauchy-Schwartz inequality. This can be written as
\begin{equation}\label{ac}
a^2-3a+(3c^2-9)a+(3a-12)c^2+c^4+4\le 0.
\end{equation}
First let us assume $C^2<0$.
The cases when $c^2<3$ ({\it i.e.} $-3<C^2<0$) actually follow from the classification of possible $-1$ and $-2$ sphere classes, see for example \cite{Man}. More precisely, when $c^2=1$, the classification is obtained in Proposition 26.1, diagram (IV.8) of \cite{Man}. Especially, it contains our classes (3)-(6). When $c^2=2$, the classification is in diagram (IV.2) there. Let us reproduce the proof for readers' convenience. When $c^2=1$, we have $$3a-\sum_{i=1}^r b_i=1, \quad a^2-\sum_{i=1}^r b_i^2=-1, \quad r<9.$$ By possibly adding some $b_i=0$ and $b_9=1$, we are reduced to solve $$3a-\sum_{i=1}^9 b_i=0, \quad a^2-\sum_{i=1}^9 b_i^2=-2, \quad b_9=1.$$ Rewrite the second equation, we have $$3a-\sum_{i=1}^9 b_i=0, \quad \sum_{i=1}^9 (a-3b_i)^2=18, \quad b_9=1.$$ In total there are three essentially different representations of $18$ as a sum of $9$ squares which are in the same residue class mod $3$: $$18=3^2+3^2+0^2+\cdots +0^2=(\pm 2)^2+(\pm 2)^2+(\pm 2)^2+(\mp 1)^2+\cdots (\mp 1)^2.$$ Up to the order of $b_i$, the solutions $(a;b_1, \cdots, b_9)$ are $$(3b; b+1, b-1, b, \cdots, b), \quad (3b\pm 2; b, b, b, b\pm 1, \cdots, b\pm 1).$$ Notice one of $b_i$ has to be $1$, this gives the list in our statement.
Similarly, when $c^2=2$, our equations are reduced to $$3a-\sum_{i=1}^9 b_i=0, \quad \sum_{i=1}^9 (a-3b_i)^2=18,\quad b_9=0.$$ Up to the order of $b_i$, we have the same general solutions, but now with one of $b_i$ being $0$. These will also lie in the list.
Now let us assume $C^2\le -3$, {\it i.e.} $c^2\ge 3$. Then we have $a\le 3$ by \eqref{ac}.
When $a=1$ or $2$, then by adjunction, $b_i=1$ or $0$. This corresponds to our classes (1) and (2).
When $a=3$, we have $$\sum (b_i-1)b_i=2.$$ Hence only one $b_i$ could be $2$ or $-1$, others are $1$ or $0$. However, if $a=3$ and $b_i$ are $\pm1$ or $0$, and $k<9$, then $C^2>0$, a contradiction. Hence, it lies in class (3).
\end{proof}
Notice all $-1$ curve classes could be realized by complex structures. Hence Proposition \ref{a>0} says that all the negative curves with $a>0$ are obtained from the $-1$ curves by further blow ups and one can obtain other negative curves by blowing up up to $6$ more times. No other classes occur even for a tamed almost complex structure.
The method used in the above proof could be extended to non-negative curves as well. We will prove the case for $C^2=0$ in the following.
\begin{prop}\label{a>0'}
Let $J$ be a tamed almost complex structure on $M=\mathbb CP^2\#k\overline{\mathbb CP^2}$, $k< 9$, and $C=aH-\sum b_iE_i$ be an irreducible curve with $C^2=0$. Then $[C]$ is one of the following:
\begin{enumerate}
\item $H-E_{m_1}$;
\item $2H- E_{m_1}-\cdots -E_{m_4}$;
\item $3H-2E_{m_1}-E_{m_2}-\cdots -E_{m_6}$;
\item $4H-2E_{m_1}-2E_{m_2}-2E_{m_3}-E_{m_4}-\cdots E_{m_7}$;
\item $5H-E_{m_1}-2E_{m_2}-\cdots 2E_{m_7}$;
\item $6H-3E_{m_1}-3E_{m_2}-2E_{m_3}-\cdots -2E_{m_6}-E_{m_7}-E_{m_8}$;
\item $7H-3E_{m_1}-\cdots -3E_{m_4}-2E_{m_5}-\cdots- 2E_{m_7}-E_{m_8}$;
\item $5H-3E_{m_1}-2E_{m_2}-\cdots -2E_{m_4}-E_{m_5}-\cdots -E_{m_8}$;
\item $8H-3E_{m_1}-\cdots -3E_{m_7}-E_{m_8}$;
\item $4H-3E_{m_1}-E_{m_2}-\cdots -E_{m_8}$;
\item $8H-4E_{m_1}-3E_{m_2}-\cdots -3E_{m_5}-2E_{m_6}-\cdots-2E_{m_8}$;
\item $7H-4E_{m_1}-3E_{m_2}-2E_{m_3}-\cdots -2E_{m_8}$;
\item $9H-4E_{m_1}-4E_{m_2}-3E_{m_3}-\cdots -3E_{m_7}-2E_{m_8}$;
\item $11H-4E_{m_1}-\cdots-4E_{m_7}-3E_{m_8}$;
\item $10H-4E_{m_1}-\cdots-4E_{m_4}-3E_{m_5}-\cdots -3E_{m_8}$.
\end{enumerate}
\end{prop}
\begin{proof}
First, by Proposition \ref{extreme}, all these $C$ are spheres.
We can still use \eqref{ac}. If $C^2=0$, then $c=0$. Hence we have $0<a\le 11$.
By Lemma \ref{int>0}, we have $b_i\ge 0$.
Now we have $$3a-\sum_{i=1}^r b_i=2, \quad a^2-\sum_{i=1}^r b_i^2=0, \quad r<9.$$ If $r<7$, then by adding $b_8=b_9=1$, we have $$3a-\sum_{i=1}^9 b_i=0, \quad a^2-\sum_{i=1}^9 b_i^2=-2, \quad b_8=b_9=1.$$ This is exactly the same case as we have done for $C^2<0$. These are the cases (1)-(5) in our list.
Now we assume $r=8$ in the following. Without loss, we could assume $b_1\ge \cdots \ge b_8\ge 1$. We have a quick estimate $8b_8^2\le a^2\le 121$, thus $b_8\le 3$. If $b_8=1$, we change it to $b_8=2$ and add $b_9=1$. Hence we have
$$3a-\sum_{i=1}^9 b_i=0, \quad a^2-\sum_{i=1}^9 b_i^2=-4, \quad b_8=2, \quad b_9=1.$$ Rewrite it, we have
$$3a-\sum_{i=1}^9 b_i=0, \quad \sum_{i=1}^9 (a-3b_i)^2=36,\quad b_8=2, \quad b_9=1.$$
In total there are seven essentially different representations of $36$ as a sum of $9$ squares which are in the same residue class mod $3$ and sum to $0$:
\begin{equation*}
\begin{array}{lll}
36&=&3^2+3^2+(-3)^2+(-3)^2+0^2+ \cdots +0^2\\
&&\\
&=&(\pm 4)^2+(\pm 1)^2+(\pm 1)^2+(\pm 1)^2+(\pm 1)^2+(\mp 2)^2+(\mp 2)^2+(\mp 2)^2+(\mp 2)^2\\
&&\\
&=&(\pm 5)^2+(\pm 2)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2\\
&&\\
&=&(\pm 6)^2+0^2+ \cdots +0^2.\\
\end{array}
\end{equation*}
Up to the order of $b_i$, the solutions $(a;b_1, \cdots, b_9)$ are $$(3b; b+1, b+1, b-1, b-1, b, \cdots, b);$$ $$(3b\pm 1; b\mp 1, b, b, b, b, b\pm 1, b\pm 1, b\pm 1, b\pm 1);$$
$$(3b\pm 2; b\mp 1, b, b\pm 1, \cdots, b\pm 1);$$ $$(3b; b\pm 2, b, \cdots, b).$$
Note $b_8=2, b_9=1$, and $b_8$ has to change back to $1$, this gives the list (6)-(10) in our statement. In particular, we remark that the last case in the above is impossible.
If $b_8=2$, we change it to $b_8=3$ and add $b_9=1$. Hence we have
$$3a-\sum_{i=1}^9 b_i=0, \quad a^2-\sum_{i=1}^9 b_i^2=-6, \quad b_8=3, \quad b_9=1.$$ Rewrite it, we have
$$3a-\sum_{i=1}^9 b_i=0, \quad \sum_{i=1}^9 (a-3b_i)^2=54,\quad b_8=3, \quad b_9=1.$$
In total there are seven essentially different representations of $36$ as a sum of $9$ squares which are in the same residue class mod $3$ and sum to $0$:
\begin{equation*}
\begin{array}{lll}
54&=&3^2+3^2+3^2+(-3)^2+(-3)^2+(-3)^2+ 0^2+ 0^2 +0^2\\
&&\\
&=&(\pm 5)^2+(\pm 2)^2+(\pm 2)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 4)^2\\
&&\\
&=&(\pm 4)^2+(\pm 4)^2+(\pm 1)^2+(\pm 1)^2+(\mp 2)^2+\cdots+(\mp 2)^2\\
&&\\
&=&(\pm 6)^2+(\mp 3)^2+(\mp 3)^2+0^2 \cdots +0^2.\\
\end{array}
\end{equation*}
Up to the order of $b_i$, the solutions $(a;b_1, \cdots, b_9)$ are $$(3b; b+1, b+1,b+1, b-1, b-1, b-1,b, b, b);$$ $$(3b\pm 2; b\mp 1, b, b, b\pm 1, \cdots , b\pm 1, b\pm 2);$$
$$(3b\pm 1; b\mp 1, b\mp 1, b, b, b\pm 1, \cdots, b\pm 1);$$ $$(3b; b\pm 2, b\mp 1, b\mp 1, b, \cdots, b).$$
Note $b_8=3, b_9=1$, especially there are exactly one $1$ among all $b_i$, we know the first and the third in the above are impossible. After changing $b_8$ back to $2$, we have the list (11)-(13) in our statement.
Finally, if $b_8=3$, we change it to $b_8=4$ and add $b_9=1$. Hence we have
$$3a-\sum_{i=1}^9 b_i=0, \quad a^2-\sum_{i=1}^9 b_i^2=-8, \quad b_8=4, \quad b_9=1.$$ Rewrite it, we have
$$3a-\sum_{i=1}^9 b_i=0, \quad \sum_{i=1}^9 (a-3b_i)^2=72,\quad b_8=4, \quad b_9=1.$$
In total there are seven essentially different representations of $36$ as a sum of $9$ squares which are in the same residue class mod $3$ and sum to $0$:
\begin{equation*}
\begin{array}{lll}
72&=&3^2+3^2+3^2+3^2+(-3)^2+(-3)^2+(-3)^2+ (-3)^2 +0^2\\
&&\\
&=&(\pm 8)^2+(\mp 1)^2+\cdots +(\mp 1)^2\\
&&\\
&=&(\pm 7)^2+(\pm 1)^2+(\pm 1)^2+(\pm 1)^2+(\mp 2)^2+\cdots+(\mp 2)^2\\
&&\\
&=&(\pm 6)^2+(\pm 3)^2+(\mp 3)^2+(\mp 3)^2+(\mp 3)^2+0^2 \cdots +0^2\\
&&\\
&=&6^2+(-6)^2+0^2 \cdots +0^2\\
&&\\
&=&(\pm 5)^2+(\pm 2)^2+(\pm 2)^2+(\pm 2)^2+(\mp 1)^2+(\mp 1)^2+(\mp 1)^2+(\mp 4)^2+(\mp 4)^2\\
&&\\
&=&(\pm 4)^2+(\pm 4)^2+(\pm 4)^2+(\mp 2)^2+\cdots+(\mp 2)^2.\\
\end{array}
\end{equation*}
Up to the order of $b_i$, the solutions $(a;b_1, \cdots, b_9)$ are $$(3b; b+1, b+1,b+1, b+1, b-1, b-1,b-1, b-1, b), \quad (3b\pm 2; b\mp 2, b\pm 1, \cdots , b\pm 1);$$
$$(3b\pm 1; b\mp 2, b, b, b, b\pm 1, \cdots, b\pm 1), \quad (3b; b\pm 2, b\pm 1, b\mp 1, b\mp 1, b\mp 1, b, b, b, b);$$
$$
(3b; b+2, b-2, b, \cdots, b), \quad (3b\pm 2; b\mp 1, b, b, b, b\pm 1, b\pm 1, b\pm 1, b\pm 2, b\pm 2);$$
$$(3b\pm 1; b\mp 1, b\mp 1, b\mp 1, b\pm 1, \cdots, b\pm 1).$$
Note $b_8=4, b_9=1$, especially there are exactly one $1$ and no $2$ among all $b_i$, we know that only the second and the third in the above are possible. After switching $b_8$ back to $3$, we have the list (14)-(15) in our statement.
\end{proof}
After completing this paper, the author was kindly informed by Dusa McDuff and Felix Schlenk that they \cite{McSc} have also obtained precisely this list of classes in Proposition \ref{a>0'} as corresponding to all full symplectic packings of a $4$-ball by no more than $8$ balls. In fact, our Propositions \ref{a>0'} and \ref{corner01} altogether give an alternative argument of their full packing result. Precisely, the list of full packing corresponds to classes in $\partial\overline{\mathcal P}\cap \mathcal C_{M, K}$, which are the classes of $K$-symplectic spheres with self-intersection $0$ (the list of Proposition \ref{a>0'}) by Proposition \ref{corner01}.
The next lemma basically shows that the phenomenon discovered in \cite{CP} for (elliptic) ruled surfaces cannot happen for small rational surfaces.
\begin{lemma}\label{dim0good}
Let $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$. Then
\begin{enumerate}
\item there is no irreducible curve $C$ such that $C^2, K\cdot [C]\ge 0$ when $k<9$.
\item For $k=9$, the only such curve classes are $-mK=3mH-mE_1-\cdots-mE_9$, $m>0$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $[C]=aH-b_1E_1-\cdots-b_kE_k$ with $k\le 9$. The conditions then read as $$a^2\ge b_1^2+\cdots +b_k^2, \quad b_1+\cdots +b_k\ge 3a.$$
Since it is a $J$-holomorphic curve class for a tamed $J$, with $C^2\ge 0$, we have $a>0$ by the Lemma \ref{int>0}. Hence $$ka^2\ge k(b_1^2+\cdots +b_k^2)\ge (b_1+\cdots +b_k)^2\ge 9a^2.$$
It is a contradiction if $k<9$. When $k=9$ the equality holds if and only if $a=3m>0$ and $b_i=m$.
\end{proof}
Especially, it shows that there are no curve classes such that $\dim_{SW}([C])=0$ and $g_J([C])>0$ when $k<9$. There is a technical lemma for higher genus curve classes, which we will need later.
\begin{lemma}\label{small+}
Let $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$, $k<9$. Let $C=aH-b_1E_1-\cdots-b_kE_k$ be a curve class.
\begin{enumerate}
\item If $C^2\ge 0$, then only classes with $a\le 2$ have $g_J(C)=0$.
\item If $g_J(C)=1$, then $C^2\ge 9-k$. The equality holds if and only if $C=3H-E_1-\cdots -E_k$.
\end{enumerate}
\end{lemma}
\begin{proof}
First $a>0$ as in last lemma. Let $C^2=c$, $g=g_J(C)$. Then by adjunction formula, the inequality $(b_1+\cdots +b_k)^2\le k(b_1^2+\cdots b_k^2)$ reads as $$(3a+2g-2-c)^2\le k(a^2-c).$$
When $a\le 2$, the inequality holds only if $g=0$ or $a=c=2, g=1, k=8$. But the latter case would imply $b_i=\frac{1}{2}$ which is a contradiction.
By Proposition \ref{extreme}, $C^2\ge 0$ if $g\ge 1$. Now we assume $g=1$ and $a \ge 3$. Then the inequality reads as $$(c-9)(c-(9-k))+(3-a)(6c-(3+a)(9-k))\le 0.$$ Hence if $c\le 9-k$, both terms are non-negative, with equality holds if and only if $a=3$ and $c=9-k$.
This in turn implies $C^2=9-k$ if and only if $C=3H-E_1-\cdots -E_k$.
\end{proof}
Finally we are ready to prove Theorem \ref{main1}, which is reformulated in the following.
\begin{theorem}\label{conf}
For rational $4$-manifolds $\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k<8$, the set of all the possible configurations of negative self-intersection curves for tamed almost complex structures are the same as the set for complex structures.
\end{theorem}
\begin{proof}
We first show it for $k\le 6$. Notice in this situation, by the above classification of irreducible negative curves, we know all the negative curves are rational curves and the intersections of them with a (different) $-1$ rational curve are either $0$ or $1$. To see this, we notice that any $-1$ rational curve class $E$ has non-trivial Seiberg-Witten invariant, which allows us to apply Lemma \ref{int>0} to conclude that $E\cdot H\ge 0$. Hence $-1$ curve classes are either $E_i$ or (1) and (2) in the list of Proposition \ref{a>0} when $k\le 6$. Then Lemma \ref{a<0} and Proposition \ref{a>0} gives the result.
For our main claim, we use induction. The claim for $k=1$ (and $S^2\times S^2$) is proved in \cite{LZ-generic} and $k=2$ is proved in Section 3. We look at configurations of negative curves satisfying the following three properties:
\begin{itemize}
\item all negative curves classes are chosen from the above classification and any two distinct ones intersect non-negatively;
\item there is a rational cohomology class $a\in H^2(M, \mathbb Q)$ such that $a\cdot a> 0$ and it pairs positively with these negative curve classes and classes with non-trivial Seiberg-Witten;
\item any $-1$ rational curve classes are positive linear combinations of these negative curve classes and classes with non-trivial Seiberg-Witten.
\end{itemize}
In particular, the negative curve configuration for a tamed almost complex structure on $\mathbb{C}P^2\#k\overline{\mathbb{C}P^2}$, $2\le k<9$, satisfies these properties as we have shown before, in particular, Proposition \ref{dA}. We will show that all the curve configurations satisfying above three properties could be realized as the negative curve configuration of a complex structure, at least when $k<8$. We use induction and start it with $k=2$ which is known to be true in section $3$.
The idea of our proof is to proceed by starting with a negative curve configuration on $\mathbb{C}P^2\#k\overline{\mathbb{C}P^2}$ satisfying above three properties, combinatorially blowing it down to get a negative curve configuration on $\mathbb{C}P^2\#(k-1)\overline{\mathbb{C}P^2}$, proving it also satisfies the three properties and using the inductive hypothesis to realize this blown down configuration as a complex curve configuration, and then doing a complex blowup of this new configuration to realize the original curve configuration.
For this goal, our proof contains two steps. The first step is to prove that combinatorially blowing down a $-1$ curve from a negative curve configuration with the three properties will give a configuration of negative curves with the same three properties. Then by our induction assumption, this is a complex curve configuration. The second step is to show that we can apply a complex blowup of this new configuration to realize the original curve configuration.
We start now with the first step. Notice our curve configuration is not related to an almost complex structure at this moment, but we fix a canonical class $K=-3H+E_1+\cdots+E_k$.
For any $3\le k<9$, any configuration of negative curves with the two properties includes at least a class of $-1$ rational curve by the same argument as Corollary \ref{-1} since it only involves adjunction formula. Let it be $E_k$. We combinatorially blow down this $-1$ curve, which means that remove $E_k$, and change any other curve classes to $C'=C+(C\cdot E_k)E_k$. Then remove all the non-negative curves from the configuration and keep the rest in the configuration if they still have negative square. We also define the new canonical divisor be $K'=K-E_k$. We check that $$C'^2=C^2+(C\cdot E_k)^2,\quad K'\cdot C'=K\cdot C-C\cdot E_k.$$ Hence if we let $g_K(e)=\frac{1}{2}(e\cdot e+K\cdot e)+1$, we have $$g_{K'}(C')\ge g_K(C), \quad C'^2-K'\cdot C'=C^2-K\cdot C+(C\cdot E_k)^2+C\cdot E_k\ge C^2-K\cdot C.$$ Especially, all the curve classes $C'$ in the new configuration still have $g_{K'}(C')\ge 0$. The equality holds if and only if $C\cdot E_k=0$ or $1$. Especially, when $k\le 6$, we have $g_{K'}(C')=0$. Thus all the new negative classes $C'$ are those classes in the previous classification since we only used adjunction formula and $g_{K'}(C')\ge 0$.
We claim that the new configuration still has a $-1$ rational curve class with respect to $K'$: $$C'^2=-1, \quad K'\cdot C'=K\cdot C'=-1.$$ To prove the claim, we notice that every class $E$ in $\mathcal E_{E_k^{\perp}}\subset \mathcal E_K$, {\it i.e.} those classes in $\mathcal E_K$ whose pairing with $E_k$ is zero, is $J$-effective, {\it i.e.} $E=\sum a_iC_i$. Since $E\cdot E_k=0$, we have $\sum a_i(C_i\cdot E_k)=0$. Hence $$E=\sum a_iC_i=\sum a_iC_i'=E'.$$ Since $K'\cdot E'=-1<0$, there is a $C_i'$ with $K'\cdot C_i'<0$. If $C_i'^2<0$, then we are done. If not, understand $C_i'$ as a class in original $\mathbb CP^2\# k\overline{\mathbb CP^2}$, which is orthogonal to $E_k$.
We could continue this process for $C_i'$ since it is not extremal and $K\cdot C_i'=-1$. This process will stop in finite steps because there is a positive lower bound of the pairing of $a$ in our second properties with curve classes since it is a rational cohomology class. Finally we will get a $-1$ rational curve class. This curve class is orthogonal to $E_k$ since $C'\cdot E_k=0$.
This is a part of our induction assumption. It is also direct to check that different new negative curve classes intersect non-negatively. On the other hand, to show all classes $C'$ with $SW(C')\ne 0$ and $E_k\cdot C'=0$ are positive linear combinations of these new negative curves and other Seiberg-Witten nontrivial classes, we only need to show it for $C'\in \mathcal E_{E_k^{\perp}}$ as argued in the cone theorem. We can always write $E\in \mathcal E_{E_k^{\perp}}$ as linear combination of $C_i$'s in the original configuration and those with non-trivial Seiberg-Witten invariant. Since $C'\cdot E_k=0$, $C'$ is the linear combination of $C_i'$ with the same coefficients where $C_i'=C_i+(C\cdot E_k)E_k$. It is direct to see that if $SW(C_i)\ne 0$ then $SW(C_i')\ne 0$ as well. Now we assume $C_i$ is originally a negative curve. If $C_i'$ is a negative curve, we are fine. If $C_i'$ is no longer a negative curve, from the classification of negative curves earlier in this section, we see that $SW(C_i')\ne 0$.
This finishes the first step.
Then we prove the second step: to show that we can reverse the process using complex blowups, at least when $k<8$. We first discuss the case when $k\le 6$. By induction, the above new configuration is realized by a complex structure on $\mathbb CP^2\# (k-1)\overline{\mathbb CP^2}$. Then we do complex blow up as following. When we remove $E_k$, if self-intersections and mutual intersections are unchanged for new negative curves configuration comparing to the the corresponding part of the original one, then we blow up a generic point of $\mathbb CP^2\# (k-1)\overline{\mathbb CP^2}$. Remember all the negative curves are rational curves and the intersections of them with a (different) $-1$ rational curve are either $0$ or $1$, so we only have the following three cases. If it changes self-intersection of certain curve but not the mutual intersections, we blow up a generic point of this curve. If new intersection is introduced, we blow up this intersection. The three different situations correspond to involving zero, one or more than one negative curves respectively when we do the combinatorial blow down. We want to make sure such ``generic" choices exist. The first case happens if some $-1$ curves become square $0$ classes when blowing down $E_k$. We check that these new square $0$ classes are nef with respect to any possible tamed almost complex structures. This is apparent since the new class is $C=C_1+C_2$ where $C_i$ are $-1$ curves (one of them is $E_k$) and $C_1\cdot C_2=1$. Hence by Proposition 4.5 in \cite{LZ-generic}, for any given point of $\mathbb CP^2\# (k-1)\overline{\mathbb CP^2}$, we have a possibly reducible rational curve in this class. Moreover, there is a unique such curve passing through any given point. By Theorem 1.4 of \cite{LZrc}, reducible curves happen only when all components are in the new negative curve configuration. Hence, for any points on $\mathbb CP^2\# (k-1)\overline{\mathbb CP^2}$ outside the negative curve locus, we have a smooth curve in the class $C$. Hence a blow up will send this curve to a negative curve. Also, by the classification above, no other types of negative curves are possible to be produced. The second and third cases are similar to argue. In these cases, if the component(s) we want to blow up are part of the reducible curve in a square $0$ rational curve class $C'$. Then the $-1$ class $C'-E$ in the blow up is represented by a reducible curve. When the point blown up is a generic point on a negative curve, this is the second case. If it is an intersection point, this is the third case above. Otherwise, we will have a smooth rational curve as above. We thus finish the proof for $k\le 6$.
The case of $k=7$ introduces a new type of $-1$ curves, {\it i.e.} class (3) in Proposition \ref{a>0}. Thus there are possibly a negative curve and a $-1$ rational curve having intersection number $2$. All the argument for $k\le 6$ went through, except now we will have a square $3$ class $C'=3H-E_1-\cdots -E_6$ (the corresponding $C=3H-E_1-\cdots -E_6-2E_7$) to deal with. This is a class of $J$-genus $1$. As above, by induction, we have a complex structure on $M'=\mathbb CP^2\#6\overline{\mathbb CP^2}$ such that the combinatorial blown down of the negative curve configuration is realized. We look at the the class $C'$. When we blow up at $M'$ at one point, we will have an embedded rational curve in class $E_7$, and a possibly reducible curve in class $C$. Hence for any given point on $M'$ there is a non-embedded (possibly reducible) subvariety in class $C'$ whose one singular (nodal, cusp or multiple) point is that point. These three types of singularities correspond to $E_7$ is secant to, tangent to, or an irreducible component of the subvariety representing $C$.
Since the class $C'$ is $J$-nef, we know there are two possible types of reducible curves: there is an elliptic curve class as a component, or all the curve classes are rational.
If the first case happens, then $[C']=m_0[C_0]+\sum m_i[C_i]$ where $g_J(C_0)=1$ and $g_J(C_i)=0$ for other $i$. Since $C'$ is $J$-nef, $[C']\cdot [C_i]\ge 0$. By Lemma \ref{small+}, we know $[C']^2\le [C_0]^2=-K_{M'}\cdot [C_0]=[C']\cdot [C_i]$. The inequality holds if and only if $C'=C_0$.
Hence in this case the curve in class $C'$ is irreducible.
For the second case, let $C'=\sum m_ie_i$. Remember the class $C'$ is nef. Use $1, \cdots, l$ to label the curves whose class has negative self-intersection. Then $$C'\cdot C'=\sum_{j=l+1}^n(m_j^2e_j\cdot e_j+m_je_j\cdot (C'-m_je_j))+\sum_{i=1}^lm_ie_i\cdot C'.$$ All terms are non-negative. Since the corresponding subvariety is connected by Proposition 4.25 of \cite{LZrc}, the second term is positive.
We want to show that for reducible curves, the double points are on some negative component. Hence we could assume $n-l>1$ otherwise we are done since double points are on some negative curves. Without loss, we assume $m_i=1$ for $i>l$ as in \cite{LZ-generic}.
First it is impossible to have $l=0$ and $n\ge 2$. If so there will be a cycle in the graph of the corresponding subvariety. Hence the second term will contribute at least $4$. If $l>0$ and $n-l>2$, then it is possible only when $n-l=3$ and $e_i\cdot e_j=0$ for $i, j>l$. Hence, for this case, all double points are on some negative curves.
Now we are left with the case of $n-l=2$. Then the intersection of the two curves cannot be greater than 1. Otherwise the second term contributes at least $4$. It also cannot be $1$ since there are at least two more intersections with negative curve components. To see this, let $e_i$ be a class with $i>l$. If it does not intersect the negative curves, then $e_i\cdot (C'-e_i)=1$. Hence it is straightforward to calculate that $g_J(C'-e_i)=1$ and $(C'-e_i)^2=2-C'\cdot e_i=1-e_i^2\le 1$. This contradicts to Lemma \ref{small+} (2). Since $n-l=2$, we know the second term is at least $4$. If the intersection is $0$, then by the light lemma, they are cohomologous.
This implies that except possibly one or two cohomologous square $0$ class, all the other connected components of the reducible curve are negative curves. Hence the non-generic blow-ups happen when the point is on the negative curve locus of the reducible curve. Thus the conclusion follows from the same argument for the case of $k\le 6$.
\end{proof}
When $k=8$, we have three more classes: (4)-(6) in Proposition \ref{a>0}. Hence there are negative curves with mutual intersection $3$. A similar argument should be enough to give a proof. The only difference is that a ``generic" blowup is no longer blowing up outside the negative locus. For example, $$6H-3E_1-2E_2-\cdots -2E_8=(3H-2E_1-E_2-\cdots -E_8)+(3H-E_1-\cdots-E_8).$$
Before blowing up $E_1$, both curves have positive squares (one nodal curve and one smooth curve intersect at the node). However, it is a non-generic phenomenon since it is a reducible curve. Another interesting new feature of this reducible curve is one of its component is of genus one although the original class is of genus $0$. Recall this cannot happen if the original class is $J$-nef by \cite{LZrc}. However, the following question still makes sense.
\begin{question}
Suppose $M$ is not diffeomorphic to $\mathbb CP^2\#k\overline{\mathbb CP^2}$. Let $E\in \mathcal E_{K_J}$. Is it true that for any subvarieties $\Theta=\{(C_i, m_i)\}$ in class $E$, {\it i.e.} $E=\sum m_i[C_i]$, we have $g_J([C_i])=0$?
\end{question}
For interested readers, the above decomposition of an exceptional class could also be seen from blowing down certain elliptic fibration of $E(1)\cong \mathbb CP^2\# 9\overline{\mathbb CP^2}$. More precisely, fix an elliptic fibration of $E(1)$ with an $I_2$ fiber and a $-1$ section $E$ (the existence of such fibration should be well known, see {\it e.g.} \cite{AZ}). Then the $I_2$ is constituted by two $-2$ rational curves $C_1$ and $C_2$. Choose a generic fiber $F$. We know $[C_1]+[C_2]=[F]=-K$. Let the section $E$ intersection with $C_2$ (and $F$). After blowing it down, $C_1$ is unchanged and $F$ becomes a smooth elliptic curve $F'$ with self-intersection $1$. We see $C_1$ and $F'$ are disjoint, and $[C_1]+[F']\in \mathcal E_K$ by the adjunction formula.
%\begin{remark}
Theorem 1.1 in \cite{AGN} states that the inclusion of the space of compatible integrable complex structures into the space of all compatible almost complex structures is a weak homotopy equivalence for a rational ruled surface. Our Theorem \ref{conf} indicates that it may hold for $\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k<9$.
%\end{remark}
\subsection{Complex Configurations for small rational surfaces}\label{cp2+3}
From Theorem \ref{conf}, to know all the possible curve configurations for tamed almost complex structures, we only need to know that for complex structures when our underlying manifold is a small rational surface. This subsection summarizes all such possibilities for rational surfaces of Euler number no greater than $6$. Based on the all possible curve configurations for $\mathbb CP^2\#3\overline{\mathbb CP^2}$, we will discuss the limitation of the construction of almost K\"ahler forms using spherical classes as in \cite{LZ-generic}.
%Let us digress on the configurations of negative curves on complex rational surfaces.
For $S^2\times S^2$, the possible types are Hirzebruch surfaces $\mathbb F_{2n}$. So the only irreducible negative curve is a $-2n$ curve which is in class $A-nB$ (or $B-nA$) where $A=[\{pt\}\times S^2], B=[S^2\times \{pt\}]$.
For $\mathbb CP^2\#\overline{\mathbb CP^2}$, the possible types are $\mathbb F_{2n+1}$. So the only negative curve is in class $(n+1)E-nH$.
For $\mathbb CP^2\#2\overline{\mathbb CP^2}$, we view this as blow up of $\mathbb F_{2n+1}$. We can either blow up at a point on the unique negative curve of $\mathbb F_{2n+1}$, or blow up at a point not on it. For the first case, our configuration is $E_2$, $H-E_1-E_2$ and $(n+1)E_1-nH-E_2$. For the latter case, our negative curves are $E_2$, $H-E_1-E_2$ and $(n+1)E_1-nH$.
In all the above cases, the dual of the curve cone is the $J$-spherical cones $\mathcal S_J$ which is the K\"ahler cone. As we see in Theorem \ref{cp2+2cone} and in \cite{LZ-generic}, these configurations realize all the possible configurations of negative curves for any almost complex structures on $S^2\times S^2$, $\mathbb CP^2\#\overline{\mathbb CP^2}$ and $\mathbb CP^2\#2\overline{\mathbb CP^2}$.
For $\mathbb CP^2\#3\overline{\mathbb CP^2}$, it is a further blowup at certain points on some complex structure of $\mathbb CP^2\#2\overline{\mathbb CP^2}$. we can blow up
\begin{itemize}
\item at a point not on negative curves (a generic point), then the negative curves are $E_3$, $E_2$, $H-E_1-E_2$, $H-E_1-E_3$ and $(n+1)E_1-nH-E_2$, or $E_3$, $E_2$, $H-E_1-E_2$, $H-E_1-E_3$, $H-E_2-E_3$ (if $n=0$) and $(n+1)E_1-nH$;
\item at a generic point of $E_2$, then the curves are $E_3$, $E_2-E_3$, $H-E_1-E_2$ and $(n+1)E_1-nH-E_2$, or $E_3$, $E_2-E_3$, $H-E_1-E_2$ and $(n+1)E_1-nH$;
\item at a generic point of $H-E_1-E_2$, then the curves are $E_3$, $E_2$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH-E_2$, or $E_3$, $E_2$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH$;
\item at a generic point of $(n+1)E_1-nH-E_2$ or $(n+1)E_1-nH$, we get $E_3$, $E_2$, $H-E_1-E_2$, $H-E_1-E_3$ and $(n+1)E_1-nH-E_2-E_3$, or $E_3$, $E_2$, $H-E_1-E_2$, $H-E_1-E_3$ and $(n+1)E_1-nH-E_3$;
\item at the intersection point of $H-E_1-E_2$ and $(n+1)E_1-nH$, then the curves are $E_3$, $E_2$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH-E_3$;
\item at the intersection point of $(n+1)E_1-nH-E_2$ and $E_2$, then the curves are $E_3$, $E_2-E_3$, $H-E_1-E_2$ and $(n+1)E_1-nH-E_2-E_3$;
\item at the intersection point of $E_2$ and $H-E_1-E_2$, then the curves are only $E_3$, $E_2-E_3$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH-E_2$, or $E_3$, $E_2-E_3$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH$;
\end{itemize}
Notice in the last case, we only have one irreducible $-1$ rational curve, which is in class $E_3$. For all the others, we have at least two smooth $-1$ rational curves.
Then we can show that if for an almost K\"ahler structure the configuration of the negative curves is like the first six cases, the Nakai-Moishezon type theorem as Theorem \ref{NM2} holds since $J$-spherical cones are equal to the K\"ahler cones.
\begin{prop}
If the configuration of negative curves for an almost K\"ahler structure is one of the first six bullets listed above, we have
$$\mathcal K_J^c=\mathcal S_J= \mathcal P_J=A_J^{\vee, >0}(M).$$
\end{prop}
\begin{proof}
First notice $ \mathcal P_J=A_J^{\vee, >0}(M)$. This is because, for any $J$, $A_J^{\vee, >0}(M)$ is contained in polytope with vertices $H$, $H-E_1$, $H-E_2$, $H-E_3$ and $2H-E_1-E_2-E_3$.
Then notice $\mathcal S_J\subset \mathcal K^c_J$ and $\mathcal K^c_J\subset \mathcal P_J$. Thus we could reduce the rest to show that $\mathcal S_J= \mathcal P_J$. To prove $\mathcal S_J= \mathcal P_J=A_J^{\vee, >0}(M)$, we notice for the first case $A_J^{\vee, >0}(M)$ is another triangular bipyramid with all vertices are spherical classes. These vertices could be represented or approximated by classes in $\mathcal S_J$. For the rest, they are all tetrahedra with at least two faces (thus span the tetrahedron) determined by $-1$ classes which can be generated by spherical classes.
\end{proof}
If the negative curves are $E_3$, $E_2-E_3$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH-E_2$, as in the last case, then the corners of $A_J^{\vee, >0}(M)$ are $(2n+3)H-(2n+1)E_1-E_2-E_3$, $(n+2)H-(n+1)E_1-E_2$, $(n+1)H-nE_1$ and $H-E_1$. The first one cannot be represented or approximated by a sphere, while the other three classes are all spheres. %It is clearer if we draw a picture.
All the spherical classes are on the boundary (or more precisely, on the edges) of the triangular bipyramid with vertices $H$, $H-E_1$, $H-E_2$, $H-E_3$ and $2H-E_1-E_2-E_3$. While the class $(2n+3)H-(2n+1)E_1-E_2-E_3$ is in the interior of the hexahedron.
If the curves are $E_3$, $E_2-E_3$, $H-E_1-E_2-E_3$ and $(n+1)E_1-nH$, then the corners of $A_J^{\vee, >0}(M)$ are $(2n+2)H-2nE_1-E_2$ and other spherical classes.
In other words, in both subcases of Case $7$, our spherical classes only span a face of the dual of curve cone. This is point the techniques in \cite{T2009, LZ-generic} does not work. However, this case will be covered by the inflation method in section \ref{infmethod}, see Theorem \ref{rr}.
\subsection{Configurations of smooth $-1$ rational curves}
In addition to study the possible configurations of all smooth negative curves, we could also look at the configurations of smooth $-1$ rational curves.
In this section, we assume $M_k=\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k\ge 1$. Dusa McDuff asks a couple of questions on possible numbers of $-1$ rational curves.
\begin{question}[McDuff]\label{mcduff}
\begin{enumerate}
\item What are the possible maximal numbers $l$ of disjoint embedded $-1$ rational curves for tamed almost complex structures on $M_k$? %How many ways, up to diffeomorphism of negative curve configurations, are there of achieving a certain $l$?
\item What are the possible numbers of embedded $-1$ rational curves for tamed almost complex structures on $M_k$, especially when $k\ge 9$?
\end{enumerate}
\end{question}
We first give a few remarks on Question \ref{mcduff} (2). It is very direct to work out the possible numbers of embedded $-1$ rational curves for a complex structure on $M_k$ with $k<9$. For examples, from section \ref{cp2+3}, we know there could be $0$ or $1$ embedded $-1$ rational curves on $M_1$, $2$ or $3$ embedded $-1$ rational curves on $M_2$, $1$, $2$, $3$, $4$, or $6$ embedded $-1$ rational curves on $M_3$. Theorem \ref{main1} implies that at least when $k<8$, working with a tamed almost complex structure would not produce more possibilities. %When $k=9$, it would be interesting to know that whether there is a number $L$ such that if one has more than $L$ embedded $-1$ rational curves, one has infinitely many.
Now, let us work with the first part of Question \ref{mcduff} (1). Apparently, $l\le b^-(M_k)=k$. By Corollary \ref{-1}, we have $l\ge 1$ when $k\ge 2$. Our Theorem \ref{cp2+2} says that when $k=2$, there are at least two $-1$ rational curves. However, from the discussion in section \ref{cp2+3}, the possible value of $l$ could be $1$ or $2$. The following result implies for a general $k\ge 1$, any integer $l$ with $1\le l\le k$ could be realized. This is based on an argument delivered to the author by McDuff.
\begin{prop}\label{rat-1}
Let $M_k=\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k\ge 3$ and $l$ an integer with $1\le l\le k$. There is an integrable (K\"ahler) $J$ on $M_k$ with exactly $l$ embedded $-1$ rational curves. Moreover, these rational curves are disjoint.
\end{prop}
\begin{proof}
We start with a line on $\mathbb CP^2$. We blow up at $l$ distinct points on this line, call these exceptional curves $E_1, \cdots, E_l$. We then blow up at the intersection of $E_l$ with the line. Call the new exceptional curve $E_{l+1}$, and blow up its intersection with the line. Continue this process to do $k-l$ blow ups. Now we have a complex structure on $M_k$ which contains negative curves in classes $$H-E_1-\cdots-E_k, E_1, \cdots, E_{l-1}, E_l-E_{l+1}, \cdots, E_{k-1}-E_k, E_k.$$
If there is another embedded $-1$ rational curve, say $E$, in class $dH-\sum m_iE_i$, then we have $$d\ge m_1+\cdots +m_{k}, \hbox{ and } m_i\ge 0,\quad \forall i.$$ In particular, it implies $d\ge 0$. Moreover $d\ne 0$ otherwise all $m_i$ have to be $0$ and $E=0$.
Recall that a class $x_0H-\sum x_iE_i$ is called {\it ordered} if $x_1\ge \cdots \ge x_k$. An ordered vector is {\it reduced} if $x_0\ge x_1+x_2+x_3$ and $x_i\ge 0$ for all $i$. Hence the class of $E$, when ordered, is a reduced class. However, by Lemma 3.4 of \cite{LBL}, there is no reduced class in $\mathcal E_K$.
Hence $E_1, \cdots, E_{l-1}, E_k$ are the only exceptional curves, which are disjoint to each others.
\end{proof}
We notice this result also gives a partial answer for Question \ref{mcduff} (2): when $k\ge 3$, there could be $1$ to $k$ embedded $-1$ rational curves.
There is a different construction for $l=1$ which is due to McDuff. %Start with integrable $J_1$ on $M_1$ with negative curves $C$ in class $nE_1-(n-1)H$. Blow up a point on $C$ to get $
By section 3.3, we have an integrable complex structure $J_2$ on $M_2$ with negative curves $$(n+1)E_1-nH-E_2, \quad H-E_1-E_2, \quad E_2.$$ Then blow up the intersection point of $H-E_1-E_2$ and $E_2$ to get $J_3$ on $M_3$. Then we blow up inductively the intersection point of $E_{i-1}-E_i$ and $E_i$ for $3\le i\le k$ to get $J_k$ on $X_k$ with negative curves $$(n+1)E_1-nH-E_2, \quad H-E_1-E_2-E_3, \quad E_i-E_{i+1}, i=2, \cdots, k-1, \quad E_k.$$
Use the similar argument as Proposition \ref{rat-1}, we can show that the only $-1$ rational curve lies in the class $E_k$. We leave the full detail to the interested readers.
Since we can choose any $n\ge 0$, there are infinitely many possible configurations for $l=1$ in Proposition \ref{rat-1}.
Proposition \ref{rat-1} could be extended to a general symplectic $4$-manifold.
\begin{theorem}\label{disjoint-1}
Let $M$ be diffeomorphic to $N\#k\overline{\mathbb CP^2}$ with $k\ge 1$ and $N$ a minimal symplectic $4$-manifold. We assume $M$ is not diffeomorphic to one point blow up of an $S^2$ bundle over surface. Given $1\le l\le k$, there is a tamed $J$ on $M$ such that there are exactly $l$ embedded $-1$ rational curves. Moreover, these rational curves are disjoint.
\end{theorem}
\begin{proof}
When $M$ is a rational surface, it follows from Proposition \ref{rat-1}.
When $M$ is irrationally ruled, we could also choose such a K\"ahler structure. We start with $N=S^2\times \Sigma_h$ endowed with product complex structure. Denote the fiber class by $T$. Blow up at one point, we have two negative curves $E_1, T-E_1$. Further blow up $l-1$ times at distinct points on $T-E_1$ other its intersection with $E_1$. Then we further blow up at the intersection of $E_l$ with the line. Call the new exceptional curve $E_{l+1}$, and blow up its intersection with the line. Continue this process to do $k-l$ blow ups. In total, we have done $k$ blow ups, and $M=N\#k\overline{\mathbb CP^2}$. Now we have negative curves in classes $$T-E_1-\cdots -E_k, E_1, \cdots, E_{l-1}, E_l-E_{l+1}, \cdots, E_{k-1}-E_k, E_k.$$ For irrational ruled surfaces, $-1$ rational curves could only appear in classes $E_i$ and $T-E_i$ for $1\le i\le k$. However, $(T-E_i)\cdot (T-E_1-\cdots -E_k)=-1<0$ for any $i\le k$. Hence $E_1, \cdots, E_{l-1}, E_k$ are the only exceptional curves, which are disjoint to each others.
For a non-rational and non-ruled symplectic manifold, we could choose a tamed almost complex structure on $N$ such that in a small ball, it is integrable. Then we first blow up $l$ distinct points in this ball and then blow up consecutively on $E_l$. Again $E_1, \cdots, E_{l-1}, E_k$ are the only exceptional curves, which are disjoint to each others.
\end{proof}
%First, we assume $1<l<k$. As discussed in section 3.3, we can choose an integrable $J_2$ on $M_2$ with negative curves in classes $$2E_1-H-E_2, \quad H-E_1-E_2, \quad E_2.$$ Now blow up at $l-1$ distinct points of $H-E_1-E_2$ other than the intersection of it with $E_2$. Then blow up $k-l-1$ times at $E_{l+1}$ consecutively at generic points on the newly obtained exceptional curve. Hence, we have a complex structure $J_k$ on $M_k$. It contains negative curves in classes $$2E_1-H-E_2, H-E_1- \cdots -E_{l+1}, E_2, \cdots, E_{l}, E_{l+1}-E_{l+2}, \cdots, E_{k-1}-E_{k}, E_k.$$
%If there is another embedded $-1$ rational curve, say $E$, in class $dH-\sum m_iE_i$, then we have $$d\ge m_1+\cdots +m_{l+1}, \quad m_{l+1}\ge\cdots \ge m_k,\quad m_i\ge 0,\quad \forall i.$$
%Hence $E_2, \cdots, E_l, E_k$ are the only exceptional curves, which are disjoint to each others.
%We have d\ge m_1+\cdots +m_{l+1}, and m_{l+1}\ge \cdots \ge m_k and all m_i\ge 0. Same argument as yours, we know there are no exceptional curve with d\ne 0. Hence E_2, \cdots, E_l, E_k are only l disjoint exceptional curve.
\subsection{Irrational ruled surfaces}
In this section, we discuss the cases of irrational ruled surfaces and prove Theorem \ref{ruled}.
In general, the complex structures of non-rational ruled surfaces are much more complicated than that of rational ones. Any such minimal surface $M$ could be viewed as the projectivization $\mathbb P(E)$ of a vector bundle of dimension two over $\Sigma_g$. The curve cone behaves quite different when $E$ is unstable from it is semi-stable. When $E$ is unstable, {\it e.g.} $E=L\oplus \mathcal O$, the corresponding ruled surface $\mathbb P(E)$ has a negative curve. This is because by definition, we have a line bundle quotient $A$ of negative degree $a$. Then $C=\mathbb P(A)$ is an effective curve in the class $aT+U$ with $C^2=2a, 2a+1<0$. Recall that $T$ is the class of the fiber $S^2$ and $U$ is the class of a section with $U^2=0$ or $1$. In this case, the curve cone $A(M)$ is always closed.
In contrast, when $E$ is semi-stable, the curve cone has different features. For convenience, we assume $E$ has even degree, and after twisting a line bundle we can then suppose $\deg E=0$. First this is always true that the nef cone is the same as the closure of the curve cone which is the first quadrant of the $U$-$T$ plane. This is because if there is an irreducible curve $C$ in the class $aT+bU$, then $C\in H^0(\mathbb P(E), \mathcal O_{\mathbb P(E)}(m)\otimes \pi^*A)=\Gamma(S^mE\otimes A)$ for some integer $m\ge 0$ and some line bundle $A$. It would imply $a\ge 0$ by semi-stability. On the other hand, $b\ge 0$ since there is always an irreducible curve in class $T$ and thus $[C]\cdot T\ge 0$. There is a famous example of Mumford showing that the curve cone might not be closed by the existence of the bundle $E$ over $\Sigma_g$ with $g>1$ such that $\Gamma(S^mE\otimes A)=0$ for all $m\ge 1$ whenever $\deg A\ge 0$.
The above discussion suggests that bizarre things may happen for non-negative curves. See the discussion in the end of this section. However, the configuration of negative curves is always very simple.
\begin{theorem}\label{ruled}
For minimal irrational ruled surfaces, i.e. $S^2$ bundles over $\Sigma_{h\ge 1}$, the set of all the possible configurations of negative self-intersection curves for tamed almost complex structures are the same as the set for complex structures.
\end{theorem}
\begin{proof}
We divide our discussion in two cases.\\
$\bullet$ $\Sigma_h\times S^2$, $h\ge 1$ \\
In this case, let $U$ be the class of the base $\Sigma_h$ and $T$ be the class of the fiber $S^2$. Then the canonical class $K=-2U+(2h-2)T$. We suppose $F$ is an irreducible $J$-holomorphic curve with negative square, and $[F]=aU+bT$ for some integers $a$ and $b$. Then $a\cdot b<0$.
The adjunction formula tells us that
$$-2b+(2h-2)a+2ab=2g(F)-2.$$
If we project $F$ to the base $\Sigma_h$, the degree of the map is $a$. Since $\Sigma_h$ has genus at least one, we have
$$2g(F)-2\ge a(2h-2).$$
Hence we have
$$-2b+(2h-2)a+2ab\ge a(2h-2),$$
and in turn,
$$2b(a-1)\ge 0.$$
Since $a\cdot b<0$, it implies $a=1$ and $b<0$. For the configuration, we know that at most one class of the type $U-kT$ with $k\ge 0$ could appear because the negative intersection of each other.
On the other hand, we could also show that $U-kT$ is the class of some complex curve for a complex structure on $\Sigma_h\times S^2$.
Suppose $L$ is a holomorphic line bundle with degree $2k\ge 0$. Then projectivization $\mathbb P(L\oplus \mathcal O)$ is topologically $\Sigma_h\times S^2$. Moreover, the section $S_{-k}=\mathbb P(L\oplus 0)$ of the $\mathbb P^1$ bundle has self-intersection $-2k$, which is in the class $U-kT$.\\
$\bullet$ Non-trivial $S^2$ bundles over $\Sigma_h$, $h\ge 1$ \\
Let $U$ be the class of a section with square $1$ and $T$ be the class of the fiber. Then the canonical class $K=-2U+(2h-1)T$. We suppose $F$ is an irreducible $J$-holomorphic curve with negative square, and $[F]=aU+bT$ for some integers $a$ and $b$. Then $a\cdot (a+2b)<0$.
The adjunction formula tells us that
$$-2b+(2h-1)a-2a+a^2+2ab=2g(F)-2 \ge a(2h-2),$$
which is equivalent to say that
$$(a+2b)(a-1)\ge0.$$
This again implies $a=1$ and $b<0$, which shows the negative curves are in classes $U-kT$.
The rest of the argument is exactly the same as the case of $\Sigma_h\times S^2$. Suppose $L$ is We only a holomorphic line bundle $L$ of degree $2k-1$. The section $S_{-k}=\mathbb P(L\oplus 0)$ is in the class $U-kT$.
\end{proof}
\begin{remark}
Notice that the Seiberg-Witten invariant calculation shows that there is a curve in class $aU+bT$, $a,b>0$ (let us focus on the trivial bundle case here, similar for the nontrivial bundle case), if and only if $ab+b+a-ah\ge 0$. This implies the closure of curve cone is still the $U-T$ plane. However, it is intriguing to see whether there is a generic complex structure in the sense that only curve classes are the Seiberg-Witten non-trivial classes.
\end{remark}
We now give an interpretation of the example in \cite{CP}. Consider the nontrivial $S^2$ bundle over $T^2$. The classes $U$ and $T$ have the same meaning as above. Then the canonical class $K=-2U+T$. Consider the class $-2K$, it is the class of a square zero torus and its Seiberg-Witten dimension is $0$. Hence, generically we only have a unique $J$-holomorphic curve in this class. The key observation of \cite{CP} is this is not true for complex structures: for any complex structures, there is always a $J$-holomorphic tori in class $-2K$ passing through any given point. Hence, after one blow-up at any point, we have a $-1$ $J$-holomorphic torus (possibly reducible) in class $4U-2T-E$. Notice its Seiberg-Witten dimension is negative, so generically there is no curve in this class.
It is interesting to see that if we blow down along the other $-1$ curve, the one in class $T-E$, we will have $S^2\times T^2$. The curve class is $4U+T$ in it (now our $U^2=0$), which is a genus $4$ class. Since the previous class in one point blow-up is represented as $4U+T-3E$ in our new basis, it implies every point is a triple point of a holomorphic curve in class $4U+T$ for any complex structures, which is of course not a generic phenomenon.
\section{Nakai-Moishezon duality for manifolds with abundant negative self-intersection curves}\label{infmethod}
In this section, we assume the tamed almost complex $4$-manifold $(M, J)$ has sufficiently many negative curves, such that $\mathcal P_J$ has no round boundary. We say there is no round boundary if the boundary is a cone over a polytope. Thus any class $e$ with $e^2=0$ should have $e\cdot C=0$ for some $C\in A_J(M)$.
As mentioned in the introduction, besides the subvarieties-current-form strategy, there is another way to attack Question \ref{dualC}. This is our main focus in this section. Alongside the main theorem in \cite{LZ}, we will need to construct $J$-tamed symplectic forms from an existing one. We use three operations in this section. The first one is the $J$-tamed inflation along curves with negative self-intersection (and sometimes along curves of square $0$), as described in Theorem \ref{buse} (Theorem \ref{mc} respectively). The second one is the summing of two $J$-tamed symplectic forms. The third one is rescaling, i.e. multiplying any $J$-tamed symplectic form with a positive number. The latter two make sense since $\mathcal K_J^t$ is a convex cone.
Let us begin with several lemmas. First, let us determine the polytopic boundary of $\mathcal P_J$.
\begin{lemma}\label{negbdy}
Let $h_J^+(M)=b^-(M)+1$ and $\mathcal P_J\ne \emptyset$.
\begin{enumerate}
\item If $b^-(M)>1$, all the boundary hyperplanes are determined by negative curves.
\item If $b^-(M)=1$, all the boundary hyperplanes are determined by non-positive curves.
\item If $b^-(M)=0$, then $\mathcal P_J$ is a single ray.
\end{enumerate}
\end{lemma}
\begin{proof}
Recall $\mathcal P_J\subset H_J^+(M)$, The third item is self-evident. Now we can assume $b^-(M)>0$.
By the light cone lemma, if $A$, $B$ are classes in $H_J^+$ with $A^2>0$, $B^2\ge 0$, then we have $A\cdot B>0$. This implies any positive curves cannot determine a polytopic boundary of $\mathcal P_J$.
If $A$, $B$ are classes in $H_J^+$ with $A^2, B^2\ge 0$, then we have $A\cdot B\ge 0$. And the equality holds if and only if $A$ is proportional to $B$. This implies any square $0$ curve classes will contribute a ray in the polygonal boundary of $\mathcal P_J$. It is a boundary hyperplane only when the cone $\mathcal P_J$ has dimension $2$, {\it i.e.} $b^-(M)=1$.
These give the proof of the first two facts.
\end{proof}
The next one is on the geometric property of a general $\mathcal P_J$.
\begin{lemma}\label{closed}
Let $C_i$'s be the irreducible $J$-holomorphic curves in $A_J(M)$. If $C_i^2<0$, for any class $A\in \mathcal P_J$, and any $0< \epsilon < \dfrac{A\cdot [C_i]}{-C_i^2}$, the class $(A+\epsilon [C_i])\in \mathcal P_J$. If $C_i^2\ge 0$, $A+\epsilon [C_i]\in \mathcal P_J$ for any $\epsilon>0$.
\end{lemma}
\begin{proof}
First, $(A+\epsilon [C_i]) \cdot [C_i]= A \cdot [C_i](1+\epsilon \dfrac{C_i^2}{A\cdot [C_i]})>0$.
When $i\neq j$, $(A+\epsilon [C_i]) \cdot [C_j]= A\cdot [C_j]+\epsilon [C_i]\cdot [C_j]>0$ because both terms are positive.
Finally, $(A+\epsilon [C_i])^2= A^2+ \epsilon A\cdot [C_i] +(A+\epsilon [C_i]) \cdot [C_i]>0$.
\end{proof}
Among the $3$ operations mentioned above for constructing $J$-tamed symplectic form, the $J$-tamed inflation is the most important one. One of the most effective tools to determine the symplectic cone of a $4$-manifold is the (positive) symplectic inflation process introduced by Lalonde and McDuff in \cite{LaMc} along symplectic curves with non-negative self-intersection. In \cite{LU}, this construction is extended to the case of negative self-intersection curves.
There is also a corresponding $J$-tamed version of it. McDuff, in \cite{Mc2}, proved the following result regarding the existence of (embedded) $J$-holomorphic curves with non-negative self-intersection.
\begin{theorem}[McDuff]\label{mc}
Let $J$ be a $\tau_0$-tame almost complex structure on a symplectic $4$-manifold $(M,\tau_0)$ that admits an embedded $J$-holomorphic curve $Z$ with $Z\cdot Z\ge 0$. Then there is a family $\tau_{\lambda}$, $\lambda\ge 0$, of symplectic forms that all tame $J$ and have cohomology class $[\tau_{\lambda}]=[\tau_0]+\lambda [Z]$. \end{theorem}
More recently, Buse (in \cite{Bus}) provided the corresponding version when $J$-holomorphic curves with negative self-intersection are in presence.
\begin{theorem}[Buse]\label{buse}
Fix a symplectic $4$-manifold $(M^4, J, \tau_0)$ such that $J$ is any $\tau_0$-tame almost complex structure. Assume that $M$ admits an embedded $J$-holomorphic curve $u:(\Sigma,j)\rightarrow (M^4,J)$ in a homology class $Z$ with $Z^2=-m$. For all $\epsilon>0$ there exist a family of symplectic forms $\tau_{\mu}$ all tame $J$ which satisfy
$$[\tau_{\mu}]=[\tau_0]+\mu Z$$
for all $0\le \mu \le \frac{\tau_0(Z)}{m}-\epsilon$.
\end{theorem}
For the convenience of discussion, let us introduce the notion of the {\em formal $J$-inflation}.
\begin{definition}
An operation on a class $A$ is called a formal $J$-inflation along the cohomology class $C \in H_J^+(M)$ with $A\cdot C \ge 0$ and $C^2<0$,
if $A$ is transformed to $A+\epsilon C$ with $0< \epsilon \le\dfrac{A\cdot C}{-C^2}$. When $\epsilon=\dfrac{A\cdot C}{-C^2}$, we call it a maximal formal $J$-inflation.
\end{definition}
A self-evident fact for this definition is that a class obtained from formal $J$-inflation could be approximated by genuine $J$-tamed symplectic inflations if the class $A\in \mathcal K_J^t$ and $C$ is the class of an embedded $J$-holomorphic curve with $C^2<0$. This fact will be used frequently in this section along with the main result in \cite{LZ} that $\mathcal K_J^t\cap H_J^+(M)=\mathcal K_J^c$.
Lemma \ref{closed} demonstrates that the closure of the dual cone $ \mathcal P_J$ is closed under the operation of formal $J$-inflation. Because $\mathcal P_J$ is a convex cone, it is also closed under summing and rescaling. Thus $\overline{\mathcal P_J}$ is closed under all the three operations. Moreover, after the three operations, the class will still stay in the same connected component of $\mathcal K_J^t$ as beginning.
\begin{lemma}\label{lightcone} Suppose $h_J^+(M)=b^-(M)+1$.
Let $C_1$ and $C_2$ be two smooth $J$-holomorphic curves with negative intersection, which provide two hyperplane pieces $\mathcal C_1$ and $\mathcal C_2$ of the boundary respectively. If the intersection $\mathcal C_1\cap \mathcal C_2\cap \overline{\mathcal P_J}\ne \emptyset$, then $([C_1]\cdot [C_2])^2\le C_1^2\cdot C_2^2$. Moreover, the equality holds if and only if there is a unique ray in the above intersection which is spanned by $[C_1]-\frac{[C_1]\cdot [C_2]}{C_2^2}[C_2]$.
\end{lemma}
\begin{proof}
Let us assume $[C_1]\cdot [C_2]\neq 0$ from now on.
Suppose $([C_1]\cdot [C_2])^2> C_1^2\cdot C_2^2$ and there is a class $A\in \mathcal C_1\cap \mathcal C_2\cap \overline{\mathcal P_J}$.
Hence we can construct a class $[C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1]$. Notice this class pairs non-negatively with all the curve classes. Moreover,
$$([C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1])^2=C_2^2-\frac{C_1\cdot C_2}{C_1^2}>0.$$
Since $A\cdot ([C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1])=0$ and $\mathcal P_J\subset H_J^+(M)$, we have $A^2<0$ by applying the light cone lemma to the $(1, b^-)$ space $H_J^+(M)$. This is a contradiction.
The equality case goes similarly, except for the last step we have $A$ is proportional to $[C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1]$. Notice $[C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1]$ and $[C_1]-\frac{C_1\cdot C_2}{C_2^2}[C_2]$ span the same ray when the equality holds.
\end{proof}
\begin{example}
Take $[C_1]=E_1$ and $[C_2]=H-E_1-E_2$, then the intersection $\mathcal C_1 \cap \mathcal C_2$ is the ray of $H-E_2$.
On the other hand, when $[C_1]=E_8$ and $[C_2]=6H-3E_1-2E_2-\cdots -2E_8$, then there is no class $A$ in $\mathcal C_1 \cap \mathcal C_2\cap \mathcal P$.
\end{example}
Next lemma describes what could be obtained if we only use the $J$-inflation along two negative curves alternatively. In the following, we say a class in $\overline{\mathcal P_J}$ is {\it achieved} by formal inflations, summing and rescaling if this class could be arbitrarily approximated by these three operations starting from a class in $\overline{\mathcal P_J}$.
\begin{lemma}\label{facetoedge}Suppose $h_J^+(M)=b^-(M)+1$.
Let $C_1$ and $C_2$ be two smooth $J$-holomorphic curves with negative intersection, and denote the boundary of $\mathcal P_J$ determined by them as $\mathcal C_1$ and $\mathcal C_2$ respectively. If $\mathcal C_1\cap \mathcal C_2\cap \overline{\mathcal P_J}\ne \emptyset$, then starting from any class $A\in \overline{\mathcal P_J}$, we will achieve a class in $\mathcal C_1\cap \mathcal C_2\cap \overline{\mathcal P_J}$ by taking formal inflations along $C_1$ and $C_2$ alternatively.
\end{lemma}
\begin{proof}
We may also assume the given class $A \in \mathcal C_1$, otherwise taking a maximal formal $J$-inflation along $C_1$.
We take maximal formal $J$-inflations along $C_1$ and $C_2$ alternatively. Namely, we suppose $A_0=A$ and when $k\ge 0$, $$A_{2k+1}=A_{2k}+l_{2k+1}[C_2], \quad l_{2k+1}=\dfrac{A_{2k}\cdot [C_2]}{-C_2^2};$$
$$A_{2k+2}=A_{2k+1}+l_{2k+2}[C_1], \quad l_{2k+2}=\dfrac{A_{2k+1}\cdot [C_1]}{-C_1^2}.$$
By calculating the coefficients $l_k$ inductively,
$$l_1=\dfrac{A\cdot [C_2]}{-C_2^2};$$
$$l_{2k+1}=l_1\cdot (\dfrac{([C_1]\cdot [C_2])^2}{C_1^2\cdot C_2^2})^k;$$
$$l_{2k}=l_1 \cdot \dfrac{[C_1]\cdot [C_2]}{-C_2^2} \cdot (\dfrac{([C_1]\cdot [C_2])^2}{C_1^2\cdot C_2^2})^{k-1}.$$
By Lemma \ref{lightcone}, we have $([C_1]\cdot [C_2])^2\le C_1^2\cdot C_2^2$.
First let us assume $([C_1]\cdot [C_2])^2=C_1^2\cdot C_2^2$. To consider the convergence of the classes $A_k$ is indeed to consider the convergence of the corresponding rays of $A_k$. Simple calculation shows that $A_k$ approaches the ray of $[C_2]-\frac{[C_1]\cdot [C_2]}{C_1^2}[C_1]$, which is the (unique) intersection of $\mathcal C_1\cap \mathcal C_2$ in $\mathcal P$.
If we have $([C_1]\cdot [C_2])^2< C_1^2\cdot C_2^2$, we have the limit of $A_k$, whose value is $$\lim_{k\rightarrow \infty}A_k=A+\frac{l_1}{1-x}([C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1]),$$ where $x=\frac{(C_1\cdot C_2)^2}{C_1^2\cdot C_2^2}<1$. When we vary the class $A$, we get different limiting classes. It is straightforward to check that the pairing with $[C_2]$ is $$A\cdot [C_2]+\frac{l_1}{1-x}(C_2^2-\frac{([C_1]\cdot [C_2])^2}{C_1^2})=A\cdot [C_2]-\frac{A\cdot [C_2]}{1-x}(1-x)=0.$$ Similarly the paring with $C_1$ is zero as well. Since the formal inflation keeps our class in $\overline{\mathcal P_J}$, our conclusion follows.
\end{proof}
There is a better viewpoint to see the above calculations: we are actually doing formal inflation along the ray determined by $[C_2']=[C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1]$. Notice $C_2'^2<0$ and $[C_2']\cdot [C_1]=0$. Hence when we do inflation along the class $[C_2']$, the new class will keep orthogonality with $C_1$. And the coefficient $\frac{l_1}{1-x}$ is nothing but the maximal inflation coefficient $\frac{A\cdot [C_2']}{-C_2'^2}$ by simple calculation.
\begin{lemma}\label{vertex}
Suppose $h_J^+(M)=b^-(M)+1$. Let $C_1$, $C_2$, $\cdots$, $C_n$ be smooth $J$-holomorphic curves with negative intersection, and denote the boundary of $\mathcal P_J$ determined by them as $\mathcal C_i$. Moreover, we assume $\cap_i\mathcal C_i\cap \overline{\mathcal P_J}$ is a ray spanned by the class $B$. Then given any class $A\in \overline{\mathcal P_J}$, one could achieve the class $B$ by taking formal $J$-inflations along $C_i$ (as well as summing and rescaling).
\end{lemma}
\begin{proof}
If there are two $C_i$'s, say $C_1$ and $C_2$, satisfy $([C_1]\cdot [C_2])^2=C_1^2\cdot C_2^2$, then by Lemma \ref{lightcone}, $\mathbb R^+B$ is the intersection $\mathcal C_1\cap \mathcal C_2\cap \mathcal P_J$. Thus, maximal formal $J$-inflations along $C_1$ and $C_2$ would approach the ray $B$ as the limit as argued in Lemma \ref{facetoedge}. Hence we assume $([C_i]\cdot [C_j])^2<C_i^2\cdot C_j^2$ for $i\neq j$. We do induction for the $n$. When $n=2$, it is Lemma \ref{facetoedge}.
Let us now show it for $n=3$, whose argument suggests the general induction step. We use the viewpoint after Lemma \ref{facetoedge}. We first find a class in $\mathcal C_1\cap \mathcal C_2$ by formally doing inflations for an arbitrary class $A\in \mathcal C_1\cap \mathcal P_J$ along orthogonal classes $[C_1]$ and $[C_2']=[C_2]-\frac{C_1\cdot C_2}{C_1^2}[C_1]$. For the new class $A_1$, we apply formal inflations along orthogonal classes $[C_1]$ and $[C_3']=[C_3]-\frac{C_1\cdot C_3}{C_1^2}[C_1]$. We thus obtain a class $A_2\in \mathcal C_1\cap \mathcal C_3$. Then we repeat this period-$2$ process. Notice $A_k\cdot C_1=0$ for all $k$. Hence, we are doing formal inflations along $[C_2']$ and $[C_3']$ alternatively. By the calculation as in Lemma \ref{facetoedge}, $A_k$ converges to $$A_1+\frac{l_1'}{1-x'}([C_3']-\frac{C_2'\cdot C_3'}{C_2'^2}[C_2'])$$ where $l'=\frac{A_1\cdot C_3'}{-C_3'^2}$ and $x'=\frac{(C_2'\cdot C_3')^2}{C_2'^2\cdot C_3'^2}$. By the viewpoint after Lemma \ref{facetoedge}, we are doing formal inflations along orthogonal classes $$[C_1],\quad [C_2'], \quad [C_3'']=[C_3']-\frac{C_2'\cdot C_3'}{C_2'^2}[C_2'].$$
The general induction step is a similar process of choosing orthogonal basis, an adaption of Gram-Schmidt process. Suppose in the cases of $n\le k$, we can find orthogonal classes as positive linear combinations of original $C_i$ and we get a class in $\cap \mathcal C_i$ by taking formal inflations along these orthogonal classes. Now we want to argue it for $n=k+1$. We start with $A\in \mathcal C_1\cap \mathcal P_J$ and obtain $A_i$ for $i\le k$ in turns by taking formal inflations along orthogonal classes $[C_1]$ and $[C_{i+1}']=[C_{i+1}]-\frac{C_1\cdot C_{i+1}}{C_1^2}[C_1]$. Repeat this period-$k$ process. Notice all $A_k\cdot C_1=0$. Hence, we are doing (period-$k$) formal inflations along $[C_i']$, $2\le i \le k+1$. Then the induction process implies we are actually doing formal inflations along a basis of orthogonal classes $[C_i'']$, $2\le i \le k+1$.
Notice any linear combinations of $\sum a_i[C_i]$ with $a_i>0$, especially $[C_i'']$, have non-positive square. Otherwise the there is a contradiction by light cone lemma as in Lemma \ref{lightcone}.
We thus finish our proof by finding the unique ray in the intersection $\cap \mathcal C_i$ spanned by $$A_1+\frac{A_1\cdot C_2''}{-C_2''^2}[C_2'']+\cdots +\frac{A_1\cdot C_{k+1}''}{-C_{k+1}''^2}[C_{k+1}''].$$ Notice that whatever the class $A$ we start with, we will arrive at the same ray.
\end{proof}
\begin{example}
Suppose $M=\mathbb CP^2\#3\overline{\mathbb CP^2}$. Let the negative curves be $E_3$, $E_1-E_2$, $H-E_1-E_2-E_3$ and $E_2-E_3$. Then the first three classes will determine a intersection $2H-E_1-E_2$. Actually, we have $[C_1]=E_3$, $[C_2']=E_1-E_2$, $[C_3']=[C_3'']=H-E_1-E_2$. They are orthogonal to each other. Then our intersection ray is spanned by $$A+(A\cdot E_3) E_3+ \frac{A\cdot (E_1-E_2)}{2} (E_1-E_2)+A\cdot (H-E_1-E_2) (H-E_1-E_2).$$
We start with different classes in the closure of $\mathcal P_J$, we will get the same ray spanned by $2H-E_1-E_2$.
For example, if we start with $H$, we get $2H-E_1-E_2$. If we start with $H-E_1$, we get $$H-E_1+\frac{1}{2}(E_1-E_2)=\frac{1}{2}(2H-E_1-E_2).$$
\end{example}
Now we are ready to prove Theorems \ref{noround} and \ref{rr}.
\begin{proof}(of Theorem \ref{noround})
As there is no round boundary, each connected component is a polytope (with perhaps infinitely many faces). We denote the known almost K\"ahler form by $\omega$. We want to prove that $\mathcal K_J^c$ is the connected component of $\mathcal P_J$ containing $[\omega]$. Since $\mathcal K_J^c\subset \mathcal P_J$ and $\mathcal K_J^t\cap H_J^+(M)=\mathcal K_J^c$, we only need to prove that $\mathcal K_J^t$ contains the connected component of $\mathcal P_J$ containing $[\omega]$. Since $\mathcal K_J^t$ is convex, we only need to prove that each extremal ray could be achieved through the formal $J$-inflation, summing and rescaling.
For each extremal ray, we could find boundary hyperplanes $\mathcal C_i$'s such that each $\mathcal C_i$ is determined by a smooth negative curve $C_i$. Thanks to Lemma \ref{vertex}, we could achieve this extremal ray by formal $J$-inflations along $C_i$'s, starting from the class $[\omega]$. After achieving each extremal ray, we could use summing and scaling to achieve the closure of the connected component of $\mathcal P_J$ containing $[\omega]$ since $\mathcal K_J^t$ is a convex cone.
Finally, as maximal formal $J$-inflations could be infinitesimally approximated by $J$-tamed symplectic inflation processes. We could achieve any class of the connected component of $\mathcal P_J$ containing $[\omega]$ by the three operations, thus all represented by $J$-tamed symplectic forms. Hence $\mathcal K_J^c$ is the connected component of $\mathcal P_J$ containing $[\omega]$ since $\mathcal P_J\subset H_J^+(M)$ and $\mathcal K_J^t\cap H_J^+(M)=\mathcal K_J^c$.
\end{proof}
There are many tamed almost complex manifolds whose $\mathcal P_J$ has a round boundary. For example, a generic almost complex structure on manifolds with $b^+>1$ will do since there are not enough curves.
It is also true that a generic tamed almost complex structure on $\mathbb CP^2\#k\overline{\mathbb CP^2}$ for $k>9$ also has a round boundary $\mathcal P_J$. First, by Lemma \ref{negbdy}, the boundary hyperplanes of $\mathcal P_J$ are determined by negative curves. On the other hand, only negative curves with non-trivial Seiberg-Witten are the classes of $-1$ rational curves since $2e^2=\dim_{SW}(e)+2g_J(e)-2\ge -2$.
Hence, generically the only negative curves are $-1$ rational curves. However, the condition $e\cdot E>0$ for all $E\in \mathcal E_K$ does not guarantee that $e^2\ge 0$ for $k>9$, by taking $e=-K$ for instance. More precisely, there is an open subset of $\partial\mathcal P$ whose all elements pairing positively with $\mathcal E_K$ as in Remark \ref{k>9round}. See the new edition of \cite{McS} for more discussions.
%And one can indeed get an open set of round boundary by varying the $J$-tamed symplectic form $\omega$ in the expression $-K+a_{\omega}[\omega]$ where $a_{\omega}>0$ is determined such that $(-K+a_{\omega}[\omega])^2=0$. Such $a_{\omega}$ exists since $K^2\cdot [\omega]^2<0$.
On the other hand, Theorem \ref{rr} shows in some interesting cases, the assumptions of Theorem \ref{noround} are satisfied.
When $b^+=1$, there is yet another cone which is relevant to Question \ref{dualC}, the $K-$symplectic cone $\mathcal C_{M, K}$ introduced as \eqref{Kcone}.
\begin{proof}(of Theorem \ref{rr})
First a rational or ruled surface has $b^+=1$, hence $h_J^+=b^-+1$ holds.
It is known that for rational surfaces $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$ with $k<9$ or for minimal ruled surfaces, $\mathcal C_{M, K}$ has no round boundary. Actually, it is a polytope whose extremal rays are Cremona equivalent to $H$ or $H-E_1$ \cite{LBL}.
Thus, $\mathcal P_J$ has no round boundary, since it is included in and cut along by more hyperplanes from the polytope $\mathcal C_{M, K}$. Moreover, it is connected. When $M=\mathbb CP^2\# k\overline{\mathbb CP^2}$ and $1<k<9$, the boundary of $\mathcal P_J$ is constituted of hyperplanes determined by curves with negative intersection because of Lemma \ref{negbdy}. By Proposition \ref{extreme}, all these curves are smooth rational curves. By applying Theorem \ref{noround}, we have $\mathcal K_J^c=\mathcal P_J$. Since $\mathcal C_{M,K}\subset \mathcal P$ in this case, we know $A_J(M)$ is a closed cone and $\mathcal P_J=A_J^{\vee, >0}(M)$.
For minimal ruled surfaces (and $\mathbb CP^2\# \overline{\mathbb CP^2}$), the boundary of (the closure of) the curve cone is constituted of two rays. One is the fiber class $T$. By a result of \cite{Mc2}, all the $J$-holomorphic curves in the fiber class $T$ are embedded. If there is a negative curve by assumption, then it is unique by Theorem \ref{ruled}. Hence the other ray is spanned by an irreducible curve class $C$ with self-intersection $-n<0$. Moreover, by Theorem \ref{ruled}, the class $C\cdot T=1$. Hence for the boundary of $\mathcal P_J$, one of the ray is the fiber class $T$ and the other is $A$ with $A\cdot C=0$ and $A^2=n$.
As observed in \cite{Bus}, the class $C$ is also represented by an embedded curve.
We do positive inflation along $T$ and negative inflation along $C$ (which determines the boundary A), $B+kT+lC$ with any $k>0$ and $0<l<\frac{B\cdot C}{n}$ spans $\mathcal P_J$.
Hence we have constructed $J$-tamed symplectic form in any any class of $\mathcal P_J$. Since $b^+=1$, we have $\mathcal K_J^c=\mathcal P_J$. By the above discussion, the boundary of $\overline{A}_J^{\vee, >0}$ contains two rays, one is the fiber class and the other is a class of non-negative square. Hence $\overline{A}_J^{\vee, >0}\subset \mathcal P$ and thus $\mathcal P_J=\overline{A}_J^{\vee, >0}$.
For the case of $M=\mathbb CP^2\# 9\overline{\mathbb CP^2}$, we know $\mathcal P_J$ is almost polytopic in the sense that only the class $-K$ is possibly on the round boundary. In other words, any class in the interior of $\mathcal P_J$ could be expressed as positive linear combinations of extremal rays. Moreover, $-K$ is the only possible curve class with non-positive self intersection which is not a rational curve. However, it does not contribute the vertices to $\mathcal P_J$. Then by Lemma \ref{vertex}, we can achieve these extremal rays by formal $J$-inflations along smooth rational curves. Hence we could obtain realize all the classes in $\mathcal P_J$ by $J$-tamed symplectic form, and thus $\mathcal K_J^c=\mathcal P_J$, as argued in Theorem \ref{noround}. Since $\mathcal C_{M,K}\subset \overline{\mathcal P}$ in this case, we know $\mathcal P_J=\overline{A}_J^{\vee, >0}(M)$.
Finally, the statement $\mathcal K_J^c=\mathcal K_J^t$ follows from \cite{LZ}.
\end{proof}
In the case of minimal ruled surfaces, the curve cone does not necessarily be closed if the other extremal ray has square $0$. However, we have irreducible positive curve classes $A_n$ arbitrarily close to the boundary ray $\mathbb R^+ C$. The author does not know how to show there is always an embedded one.
If there is such one in each $A_n$,
by Theorem \ref{mc}, given any class $B$ in $\mathcal P_J$, we do (positive) inflations along $A_n$ and $F$ to obtain $B+k_1F+k_2A_n$ which spans all $\mathcal P_J$.
We remark that actually Theorem \ref{noround} has more applications. One important case is
when we have a smooth representation of the anti-canonical class. In this case, all the negative curves are smooth rational curves with self-intersection $-1$ or $-2$. In fact, $\mathcal P_J$ is a polytope bounded by the hyperplanes determined by those rational curves and the curve in $-K$.
On a general four dimensional symplectic manifold, we do not usually have enough embedded $J$-holomorphic curves, although a generic almost complex structure on manifolds with $b^+=1$ does have so. Thus we have Theorem \ref{generic}.
\begin{proof}(of Theorem \ref{generic}) Now let us prove $\mathcal K_J^t=\mathcal C_{M,K}$ when $J$ is in a residual set of tamed almost complex structures.
We first pick up any symplectic form $\omega$ with integral cohomology class. Let $e$ be an integral class in $\mathcal C_{M,K}$. There is an integer $L$ such that for all the integrals $l>L$, $le-[\omega]$ and $le-[\omega]-K$ are both in $\mathcal C_{M,K}$. By Lemma 3.4 of \cite{LL}, $SW(le-[\omega])$ is nontrivial. Then $le-[\omega]$ could be represented by an embedded $J$-holomorphic curve for a generic $J$ tamed by $\omega$. We take the union of these residual subsets of $\omega$-tame almost complex structures and denote it by $\mathcal J_{e,\omega}$. By Theorem \ref{mc}, we know that $le=[\omega]+(le-[\omega])$ is represented by a $J$ tamed symplectic form for $J\in \mathcal J_{e,\omega}$. Take intersection of $\mathcal J_{e,\omega}$ for all integral class $e$, we get another generic subset $\mathcal J_{\omega}$ in all $\omega$-taming almost complex structures, since there are only countably many integral cohomology classes. Hence we have shown that for $J\in \mathcal J_{\omega}$, $\mathcal K_J^t=\mathcal C_{M,K}$.
Because the set of all integral symplectic forms is dense in the space of symplectic forms, any tamed almost complex structure is tamed by a symplectic form with integral cohomology class. Taking union of $\mathcal J_{\omega}$ for all symplectic forms with integral cohomology class, we achieve our final residual subset $\mathcal J$ in all tamed almost complex structures. Hence $\mathcal K_J^t=\mathcal C_{M, K}=\mathcal P_J$ for a generic tamed $J$.
Finally, by \cite{T2009} we have $\mathcal K_J^c\ne \emptyset$ for generic tamed $J$. Hence for all such $J$, $\mathcal K_J^t=\mathcal K_J^c$ by \cite{LZ}. Thus the proof of Theorem \ref{generic} completes.
\end{proof}
Notice the above proof is a renaissance of the argument in \cite{LL}. There is an alternative way to construct the residual set $\mathcal J$ using the strategy in \cite{T2009}.
We endeavour to prove Question \ref{dualC} for all tamed $J$ rather than a residual subset. However, we may not have enough embedded $J$-holomorphic curves to apply the $J$-inflation even if we always have sufficient irreducible curves to play with the formal $J$-inflation.
\begin{thebibliography}{99}
\bibitem{AGN} M. Abreu, G. Granja, N. Kitchloo, \textit{Compatible complex structures on symplectic rational ruled surfaces}, Duke Math. J. 148 (2009), no. 3, 539--600.
\bibitem{AZ} A. Akhmedov, W. Zhang, \textit{The fundamental group of symplectic $4$-manifolds with $b^+=1$}, arXiv:1506.08367.
\bibitem{Bu} N. Buchdahl, \textit{On compact K\"ahler surfaces}, Ann. Inst. Fourier (Grenoble) 49 (1999), no. 1, vii, xi, 287--302.
\bibitem{Bus} O. Buse, \textit{Negative inflation and stability in symplectomorphism groups of ruled surfaces}, Journal of Symplectic Geometry, Vol. 9, No. 2, 147--160, 2011.
\bibitem{CP} P. Cascini, D. Panov, \textit{Symplectic generic complex structures on four-manifolds with $b^+=1$}, J. Symplectic Geom. 10 (2012), no. 4, 493--502.
\bibitem{DP} J. Demailly, M. Paun, \textit{Numerical characterization of the K\"ahler cone
of a compact K\"ahler manifold}, Ann. of Math. (2) 159 (2004), no.3,
1247--1274.
\bibitem{D} S. K. Donaldson, \textit{Two-forms on
four-manifolds and elliptic equations}, Inspired by S. S. Chern,
153--172, Nankai Tracts Math., 11, World Sci. Publ., Hackensack, NJ,
2006.
\bibitem{DLZ} T. Draghici, T.J. Li, W. Zhang, \textit{Symplectic forms and cohomology decomposition of almost complex $4$-manifolds}, Int. Math. Res. Not. IMRN 2010, no. 1, 1--17.
\bibitem{FM} R. Friedman, J. Morgan, \textit{On the diffeomorphism types of certain algebraic surfaces. I},
J. Differential Geom. 27 (1988), no. 2, 297--369.
\bibitem{Gr} M. Gromov, \textit{Pseudoholomorphic curves in symplectic manifolds}, Invent. Math. 82 (1985), no. 2, 307--347.
\bibitem{LaMc} F. Lalonde, D. McDuff, \textit{The classification of ruled symplectic $4$-manifolds}, Math. Res. Lett. 3 (1996), no. 6, 769--778.
\bibitem{Lamari} A. Lamari, Le c\^{o}ne k\"ahl\'{e}rien d'une surface,
{\em J. Math. Pures Appl.} (9) 78 (1999), 249--263.
\bibitem{LBL} B.-H. Li, T.J. Li, \textit{Symplectic genus, minimal genus and diffeomorphisms}, Asian J. Math. 6 (2002), no. 1, 123--144.
\bibitem{LLwall} T.J. Li, A.-K. Liu, \textit{General wall crossing formula}, Math. Res. Lett. 2 (1995), no. 6, 797--810.
\bibitem{LLb+=1} T.J. Li, A.-K. Liu, \textit{The equivalence between SW and Gr in the case where $b^+=1$}, Internat. Math. Res. Notices 1999, no. 7, 335--345.
\bibitem{LL} T.J. Li, A.-K. Liu, \textit{Uniqueness of symplectic canonical class,
surface
cone and symplectic cone of $4$-manifolds with $b\sp +=1$}, J.
Differential Geom. 58 (2001), no. 2, 331--370.
\bibitem{LU} T.J. Li, M. Usher, \textit{Symplectic forms and surfaces of negative
square}, J. Symplectic Geom. 4 (2006), no. 1, 71--91.
\bibitem{LZ} T.J. Li, W. Zhang, \textit{Comparing tamed and compatible
symplectic cones and cohomological properties of almost complex manifolds},
Comm. Anal. Geom. 17 (2009), no. 4, 651--683.
\bibitem{LZ-generic} T.J. Li, W. Zhang, \textit{Almost K\"ahler forms on rational $4$-manifolds}, arXiv:1210.2377, to appear in Amer. J. Math.
\bibitem{LZrc} T.J. Li, W. Zhang, \textit{$J-$holomorphic curves in a nef class}, arXiv:1210.3337, to appear in Int. Math. Res. Not.
\bibitem{Man} Yu. I. Manin, \textit{Cubic forms. Algebra, geometry, arithmetic}, Translated from the Russian by M. Hazewinkel. Second edition. North-Holland Mathematical Library, 4. North-Holland Publishing Co., Amsterdam, 1986. x+326 pp.
% \bibitem{Mc1} D. McDuff, \textit{Notes on ruled symplectic $4$-manifolds}, Trans. Amer. Math. Soc. 345 (1994), no. 2, 623--639.
\bibitem{Mc2} D. McDuff, \textit{Symplectomorphism groups and almost complex structures}, Essays on geometry and related topics, Vol. 1, 2, 527-556, Monogr. Enseign. Math., 38, Enseignement Math., Geneva, 2001.
\bibitem{McS} D. McDuff, D. Salamon, \textit{Introduction to symplectic topology}, new edition. %Second edition. Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, 1998. x+486 pp. ISBN: 0-19-850451-9.
\bibitem{McSc} D. McDuff, F. Schlenk, \textit{Symplectic embedding problems}, work in preparation.
\bibitem{Mor82} S. Mori, \textit{Threefolds whose canonical bundles are not numerically effective}, Ann. of Math. (2) 116 (1982), no. 1, 133--176.
\bibitem{pinn} M. Pinsonnault, \textit {Maximal compact tori in the Hamiltonian group of $4$-dimensional symplectic manifolds}, J. Mod. Dyn. 2 (2008), no. 3, 431--455.
\bibitem{Reidsur} M. Reid, \textit{Chapters on algebraic surfaces}, Complex algebraic geometry (Park City, UT, 1993), 3--159, IAS/Park City Math. Ser., 3, Amer. Math. Soc., Providence, RI, 1997.
\bibitem{T} C. Taubes, \textit{$SW\Rightarrow Gr$: from the Seiberg-Witten equations to pseudo-holomorphic curves}, J. Amer. Math. Soc. 9 (1996), no. 3, 845--918.
\bibitem{T3} C. Taubes, \textit{$Gr=SW$: counting curves and connections},
J. Differential Geom. 52 (1999), no. 3, 453--609.
\bibitem{T2009} C. Taubes, \textit{Tamed to compatible: symplectic forms via moduli space integration}, J. Symplectic Geom. 9 (2011), no. 2, 161--250.
\bibitem{Zmann} W. Zhang, \textit{From Taubes currents to almost K\"ahler forms}, Math. Ann. 356 (2013), no. 3, 969--978.
\end{thebibliography}
\end{document}