Albanese map for Kähler manifolds with nef anticanonical bundle

Let \alpha: X \to A(X) be the Albanese morphism which is smooth by [Cao13], since the base is of dimension one. In particular, by [DPS94, Proposition 3.12], \tilde{q}(F)=0 where F is any fiber. Note that if K_X is \mathbb{Q}-effective, c_1(X)=0 and X is the product of a simply connected Calabi-Yau threefold with an elliptic curve up to a finite étale cover. In the following, we always assume that K_X is not \mathbb{Q}-effective. Up to a finite étale cover, we can assume that \pi_1(X) is free abelian, see [Pau97].

By [DHP23, Theorem 1.2], we can consider the relative K_X-MMP with respect to \alpha,

where the last step of \beta is a Mori fiber space. Notice that the restriction of the relative K_X-MMP with respect to \alpha to any fiber F is the K_F-MMP since K_X|_F=(K_X+F)|_F=K_F by the adjunction formula. But since any fiber F of \alpha is a compact Kähler threefold with nef anticanonical divisor such that \tilde{q}(F)=0 whose MMP consists of at most one step of a Mori fiber space,

\beta is a morphism and \beta consists of at most one step of a Mori fiber space. In fact, since \tilde{q}(F)=0, \pi_1(F) is finite by the structural theorem for compact Kähler threefolds with nef anticanonical divisor proven in [MW23]. Since \pi_2(A(X))=0, the homotopy exact sequence implies that \pi_1(F) is a subgroup of \pi_1(X) which is free abelian. Thus F is simply connected. Therefore, F is either a Calabi-Yau threefold, a product of \mathbb{P}^1 with a K3 surface or a rational connected threefold. In the third case, H^{2,0}(F)=H^{1,0}(F)=0 and thus H^{2,0}(X)=0, contradicting the assumption that X is non-projective. Thus the third possibility can be ruled out. Note that the dimension of Y is 4 or 3.

If \mathrm{dim}(Y)=4, any fiber F of \alpha is a simply connected Calabi-Yau threefold and \beta is an isomorphism. By Proposition 3.1 we have c_1(X)=0, contradicting the assumption that K_X is not \mathbb{Q}-effective.

If \mathrm{dim}(Y)=3, the map \gamma is a K3-fibration between compact Kähler manifolds. Because \mathrm{dim}(Y)=3, F is a product of \mathbb{P}^1 with a K3 surface and the restriction of \beta to F is the projection onto the K3 surface. In particular, the set-theoretic fiber of \gamma is alway a K3 surface. Since \beta is a \mathbb{Q}-conic bundle (see [MW23]) and X is smooth, \beta is a conic bundle outside a codimension at least 2 closed analytic subset of Y. The proof of [HP15b, Lemma 4.1] shows that Y is klt and so in particular Cohen-Macaulay. Since the fibers of \gamma are K3 sufaces, \gamma is flat by miracle flatness and hence smooth. It follows that Y is smooth.

Again by Proposition 3.1, up to some finite étale cover, Y is a product of A(X) with a K3 surface. Alternatively, we may argue as follows: Since -K_X is nef and \beta is a conic bundle (i.e. -K_X is \beta-ample) by (the proof of) Theorem 8.13 in [DHP23], -K_Y is psef by [MW23, Proposition 2.5]. On the other hand, Y is not uniruled which implies that K_Y is psef by [Bru06]. More precisely, the rational curves of Y are contained in the fibers of \gamma since A(X) contains no rational curves. Thus, if Y were uniruled the fibers of \gamma would be uniruled, too, contradicting the fact that the fibers of \gamma are K3 surfaces. Thus c_1(Y)=\frac{1}{4} \beta_*(c_1(-K_X)^2)=0 by conic bundle formula [MW23, Proposition 3.5].

By the Fischer-Grauert theorem [FG65], \beta is locally trivial since \mathbb{P}^1 is rigid. In particular, there exists a line bundle L on X such that -K_X=L^{\otimes 2} (see e.g. [ARM14, Lemma 6]). Since L is \beta-very ample, X=\mathbb{P}(\beta_* L). By [MW23, Claim 2.8], \beta_* L is weakly positively curved. On the other hand,

By [numer_flat, Main theorem], \beta_* L is numerically flat. Since a K3 surface is simply connected, \beta_* L is the pull back of some numerically flat vector bundle over the elliptic curve by the next Lemma.

By [DPS94, Theorem 1.18], E is a successive extension of hermitian flat vector bundles on X \times Y. Since any hermitian flat vector bundle corresponds to a representation of \pi_1(X \times Y) \simeq \pi_1(X), these hermitian flat vector bundles are pullbacks of some hermitian flat vector bundles on X.

Now consider an extension of pullbacks of hermitian flat vector bundles E_1, E_2 on X which corresponds to an element of H^1(X \times Y, p_1^* \mathrm{Hom}(E_2, E_1)), and thus an element of H^1(X, \mathrm{Hom}(E_2, E_1)) \otimes H^0(Y, \mathcal{O}_Y) \oplus H^0(X, \mathrm{Hom}(E_2, E_1)) \otimes H^1(Y, \mathcal{O}_Y) by the Künneth formula which gives H^1(X, \mathrm{Hom}(E_2, E_1)). In particular, the extension is the pullback of some extension of the hermitian flat vector bundles on X. Induction on the rank of E finishes the proof.

We start by recalling the following result of [TZ14, Remark 3.2]:

If g : X \to Y is a holomorphic submersion between compact Kähler manifolds with fibers X_y having trivial canonical bundle and base Y which admits a Kähler metric with non-negative Ricci curvature, then g is a holomorphic fiber bundle.

This result says that f is locally trivial. In other words, we have no deformation of complex structure among the fibers of f which in turn implies that the Weil-Petersson metric on T vanishes identically. To conclude, we use the fact that for a smooth fibration of Calabi-Yau manifolds, there is a close relation between the variation of complex structure and the positivity of the relative canonical bundle. To this end, we consider the fiberwise unique Ricci-flat Kähler-Einstein metrics \omega_s in the corresponding Kähler classes, that are given by the polarization. The relative volume form \omega_{X/T}^{n}=gdV, where \mathrm{dim} X_s=n, induces a hermitian metric g^{-1} on the relative canonical bundle K_{X/T}=K_X. Its curvature is given, up to a numerical constant, by the pullback of the Weil-Petersson form downstairs, i.e. the L^2 inner product of harmonic representatives of the Kodaira-Spencer classes, see e.g. [BCS20, Eqn. 1.1]:

But by the isotriviality, the Weil-Petersson form \omega^{WP} vanishes identically, hence
c_1(K_X, g^{-1})=0.

Let \alpha: X \to A(X) be the Albanese morphism which is smooth in codimension 1 by [Cao13]. If \mathrm{dim} (A(X))=n, by [Cao13], X is a modification of A(X). In particular, we have equality of Kodaira dimensions \kappa(X)=\kappa(A(X))=0 and thus K_X is psef. Since -K_X is nef, c_1(X)=0. Beauville-Bogomolov decomposition implies that in fact X \simeq A(X). Therefore, we only consider the case where \mathrm{dim}(A(X))<n. Without loss of generality, we may assume also that q(X) \geq 1.

Let Z be a codimension at least two closed analytic subset of A(X) such that \alpha: \alpha^{-1}(A(X) \setminus Z) \to A(X) \setminus Z is smooth.

Consider the universal cover \pi: \mathbb{C}^{2\mathrm{dim}(A(X))} \to A(X). By the codimension condition of Z, \pi: \pi^{-1}(A(X) \setminus Z) \to A(X) \setminus Z is the universal cover of A(X) \setminus Z.

Endow the local system R^k\alpha_*\mathbb{R} associated to \alpha: \alpha^{-1}(A(X) \setminus Z) \to A(X)\setminus Z with the geometric real variation of polarized Hodge structures, where k=n-\mathrm{dim}(A(X)), cf. [TZ14]. The construction of period domains and period mappings for real polarized Hodge structures by Griffiths gives a well-defined holomorphic period map p: \pi^{-1}(A(X) \setminus Z) \to \mathcal{D} where \mathcal{D} is the associated period domain. For example, the period domain is the upper half plane \mathbb{H} if the fiber F is an elliptic curve. In the case \mathrm{dim}(A(X))=n-2, as observed on page 305 of [Kob90], the period domain of real polarized K3 surfaces is the non-compact Hermitian symmetric domain O(2,19)/SO(2) \times O(1,19). By Harish-Chandra’s embedding theorem, the period domain can be realized as an open subset of a complex vector space. By Hartogs theorem, the period map p extends to a holomorphic function on \mathbb{C}^{2\mathrm{dim}(A(X))}. In the general case, the extension follows from [GS69, Corollary 9.5] based on the geodesic completeness of the period domain and the distance decreasing property of morphisms between manifolds whose curvature have strict negative upper bounds (which implies the continuous hence holomorphic extension).

The period map is thus constant by Liouville’s theorem in the case \mathrm{dim}(A(X))=n-1. In the general case, it follows from Schwarz lemma [TZ14, Lemma 3.3] and the existence of a hermitian metric on \mathcal{D} whose holomorphic sectional curvature is bounded form above by a strictly negative constant along the horizontal subbundle of T \mathcal{D} defined by Griffiths [Gri70].

But now the differential of the period map p at z is a composition of the Kodaira-Spencer map \rho_z: T_z (\pi^{-1}(A(X) \setminus Z)) \to H^1(\alpha^{-1}(\pi(z)), T_{\alpha^{-1}(\pi(z))}), and the cup product map H^1(\alpha^{-1}(\pi(z)), T_{\alpha^{-1}(\pi(z))}) \to T_{p(z)}\mathcal{D}. Since the second map is injective by infinitesimal Torelli, the Kodaira-Spencer map of the pullback family over \pi^{-1}(A(X) \setminus Z) is identically zero.

By [Kod05, Theorem 4.6], since h^1(F, T_F)=h^{1,n-1}(F) which is constant for the Kähler fibration, \alpha is locally trivial over A(X) \setminus Z. No deformation of complex structure among the fibers of \alpha implies that the Weil-Petersson metric on A(X) \setminus Z is identically zero. Again by the result of [BCS20], the pullback of the Weil-Petersson metric represents the first Chern class of relative canonical divisor K_{(X \setminus \alpha^{-1}(Z))/(A(X) \setminus Z)} modulo a constant factor, which implies that the first Chern class of X \setminus \alpha^{-1}(Z) is trivial.

The preimage of the non-flat locus W of \alpha is codimension at least 2 by Lemma 3.4. Since the dimension of the fibers is the expected one for a point on the base where the map is flat, the preimage of the non-smooth locus of \alpha is also codimension at least 2. Thus, by Lemma 3.3, H^2(X, \mathbb{C}) \simeq H^2(X \setminus \alpha^{-1}(Z), \mathbb{C} ) and hence c_1(X)=0.

By Poincaré duality, it is equivalent to show that for i \leq 2(k-1),

which follows from long exact sequence

Here, the vanishing of cohomologies follows from dimension reasons.

We follow the arguments of [CH19, Lemma 3.2]. Let \omega_X be a Kähler form on X. By Hironaka’s flattening theorem [Hir75], we have a diagram

such that \pi which is the composition \tilde{X} \to X_1 \to X, X_1 \to X, \tau are modifications, \tilde{X}, \tilde{Y} are smooth and \tilde{\alpha} is flat. Denote

where E,F effective divisors. Let \omega_{\tilde{X}} be a Kähler form on \tilde{X}. For any \epsilon >0, \pi^*(-K_X)+\epsilon \omega_{\tilde{X}} is Kähler. Since \beta(E) \neq \tilde{Y}, K_{\tilde{X}}+\pi^*(-K_X)+\epsilon \omega_{\tilde{X}} is Kähler when restricting to a general fiber of \beta. By [Pau12, Theorem 1.1], K_{\tilde{X}/\tilde{Y}}+\pi^*(-K_X)+\epsilon \omega_{\tilde{X}} is psef. Thus after passing to a limit, E-\beta^* F is psef and we get

Therefore, \beta^*(F) must be \pi-exceptional and, in particular, \alpha^{-1}(W) is of codimension at least 2.

Denote X=A\setminus Z where A is a torus and Z is of codimension at least 2. The obstruction of a \mathbb{P}^{r-1}-bundle over X being isomorphic to \mathbb{P}(E) for some holomorphic vector bundle E over X with \mathrm{det}(E)=\mathcal{O}_X lies in H^2(A \setminus Z, \mathbb{Z}_r) by the theory of Brauer groups.

Here, we follow the proof of [CP91, Lemma 7.4]. From the exact sequence

we see that the obstruction lies in

by the following Lemma 4.2. Thus the obstruction vanishes after passing to some finite étale cover. Notice that any finite étale cover of A \setminus Z extends uniquely to a finite étale cover of A. This finishes the proof of the lemma. We refer to [El82] for more information about Brauer groups.

By stratification of closed analytic subset (cf. e.g. [agbook, Chap II, Proposition 5.6]), without loss of generality, we may assume that X is the complement of a smooth complex submanifold of codimension at least k in a complex manifold Y. Consider long exact sequence

It is enough to show that for i \leq 2k-1, by considering a tubular neighborhood of Y \setminus X and the excision theorem,

where \mathbb{D} is the disk of dimension of the real codimension of Y \setminus X and \mathbb{D}^* is the punctured disk. Since H^\bullet(\mathbb{D}, \mathbb{D}^*,\mathbb{Z}) is finite generated free group, Künneth theorem implies that

by the codimension condition.

By Lemma 3.4 and [ARM14], since the Albanese map is smooth in codimension one with \mathbb{P}^{r-1} as the general fiber, there exists a closed analytic subset Z of codimension at least 3 in A(X) (non-flat locus of \alpha) such that X is a \mathbb{P}^{r-1}-bundle over A(X) \setminus Z. By Lemma 4.1, there exists a holomorphic vector bundle E over A(X) \setminus Z of rank r with \mathrm{det}(E)=\mathcal{O}_X such that

up to passing to some finite étale cover of X.

In particular, up to some finite étale cover, c_1(E)=0 and -K_X=\mathcal{O}_{X}(r) on A(X) \setminus Z. By the Leray-Hirsch theorem, we have

Let \omega be a normalised Kähler form on X (i.e. such that the restriction of \omega+c_1(K_X) to the general fiber of \alpha is trivial, hence for all fibers over A(X) \setminus Z by the Leray-Hirsch theorem). We claim that there exists \beta \in H^{1,1}(A(X), \mathbb{R}) such that we have

in H^{1,1}(X, \mathbb{R}). By Lemma 3.3, it is enough to show the existence of \beta \in H^{1,1}(A(X), \mathbb{R}) satisfying the equality on \alpha^{-1}(A(X) \setminus Z). Define

The claim follows by construction. (The same equality holds in more finer cohomology as integral Bott-Chern cohomology as shown in [Wu19a].)

In particular, -K_X is \alpha-ample and \alpha is projective. We claim that E can be extended over A(X) as a reflexive sheaf. The line bundle \mathcal{O}_{\mathbb{P}(E)}(1) can also be endowed with a smooth metric whose curvature has finite L^2 norm and whose local potentials are the local potentials of -K_{X} divided by r. By the removable singularity theorem (see [BS94, Lemma 1]), the line bundle \mathcal{O}_{\mathbb{P}(E)}(1) has a reflexive extension. Its direct image, denoted by \mathcal{F}, is a reflexive extension of E by [Har80, Cor.1.7]. (Note that the proof of [Har80] goes through as long as the preimage of the non-flat locus is of codimension at least two.) Note that the extended line bundle of \mathcal{O}_{\mathbb{P}(E)}(1) is nef and \alpha-ample as -K_X. By the positivity of direct images [MW23, Proposition 2.6], \mathcal{F} is weakly positively curved. Since c_1(E)=0, c_1(\mathcal{F})=0 on A(X).

Next, we show that \mathcal{F} is locally free by [BS94]. By nefness of the extended line bundle of \mathcal{O}_{\mathbb{P}(E)}(1) and the fact that Z is of codimension at least 3 (such that H^4(A(X) \setminus Z, \mathbb{R}) \simeq H^4(A(X), \mathbb{R})), c_1^2(\mathcal{F})-c_2(\mathcal{F})=-c_2(\mathcal{F}) contains a closed positive current whose restriction on A(X) \setminus Z represents \alpha_*(c_1(\mathcal{O}_{\mathbb{P}(E)}(1))^{r+1}). We refer to [Wu23, Remark 1] for the construction of Chern classes of a coherent sheaf over A(X) \setminus Z in the de Rham cohomologies. If \mathcal{F} is stable, by [BS94, Corollary 3], the equality case of the Bogomolov-Gieseker inequality implies that \mathcal{F} is a projectively flat vector bundle. Since c_1(\mathcal{F})=0, \mathcal{F} is hermitian flat.

We show in the following that \mathcal{F} is a numerically flat vector bundle following the arguments of [CCM21, Proposition 2.7], if \mathcal{F} is not stable. If \mathcal{F} is not stable, consider the Harder-Narasimhan filtration of \mathcal{F} respect to \omega, say

where \mathcal{F}_i /\mathcal{F}_{i-1} is \omega-stable for every i and \mu_1 \geq \mu_2 \geq \cdots \geq \mu_m, and where \mu_j=\mu_\omega(\mathcal{F}_j / \mathcal{F}_{j-1}) is the slope of \mathcal{F}_j / \mathcal{F}_{j-1} with respect to \omega. Now, consider the coherent subsheaf \mathcal{S}=\mathcal{F}_{m-1}. Without loss of generality, we can assume that \mathcal{S} is reflexive by taking the double dual if necessary, as this preserves the rank, first Chern class and slope. Then we get a short exact sequence

As a quotient of \mathcal{F}, \mathcal{Q} moduo its torsion part is also weakly positively curved. In particular, c_1(\mathcal{Q}) is pseudoeffective. Together with the stability condition of the filtration and the fact that

we can easily check that c_1(\mathcal{S})=c_1(\mathcal{Q})=0.

By [DPS94, Corollary 1.19], the inclusion \mathcal{S} \to \mathcal{F} is a bundle morphism on the locally free locus of \mathcal{F} (i.e. A(X) \setminus Z). In particular, \mathcal{Q} is locally free on A(X) \setminus Z. Compared to the case of an arbitrary weakly positive reflexive sheaf with vanishing first Chern class over a projective manifold as studied in [CCM21], there exist smooth metrics with lower bounds on the curvature of \mathcal{O}_{\mathbb{P}(E)}(1) (induced from -K_X and the nefness condition of -K_X). Now we deviate from [CCM21] and give a more simplified argument based on these metrics. Restriction of these metrics on \mathcal{O}_{\mathbb{P}(\mathcal{Q})}(1) to A(X) \setminus Z shows that over A(X) \setminus Z, the cohomology class

contains a closed positive (2,2)-current. Since Z is of dimension at least 3, this current extends across Z. In particular, -c_2(\mathcal{Q}^{**}) contains a closed postive current. On the other hand, \mathcal{Q}^{**} is stable. The equality case of the Bogomolov-Gieseker inequality is attained and thus \mathcal{Q}^{**} is also a hermitian flat vector bundle by [BS94].

(If the rank r=2, \mathcal{Q} can only be of rank 1. We have that c_1(\mathcal{Q}^{**})=c_1(\mathcal{Q})=0 and \mathcal{Q}^{**} is a hermitian flat line bundle.)

The natural morphism on the locally free locus of \mathcal{F}

implies that \mathcal{S} is weakly positively curved after extension to A(X) and it admits a sequence of smooth metrics with increasing lower bound on A(X) \setminus Z. Since

if \mathcal{S} is not stable, we can continue to think about its Harder-Narasimhan filtration and by induction on the rank conclude that \mathcal{S} is a numerically flat vector bundle.

By [numer_flat, Lemma 4], the restriction map

is surjective. Hence the extension class in H^1(A(X) \setminus Z, \mathrm{Hom}(\mathcal{Q}^{**}, \mathcal{S})) obtained from the Harder-Narasimhan filtration on A(X) \setminus Z can be extended to an extension class in H^1(A(X) \setminus Z, \mathrm{Hom}(\mathcal{Q}^{**}, \mathcal{S})). The extended class determines a vector bundle whose restriction to A(X) \setminus Z is isomorphic to \mathcal{F} by construction. It follows that this vector bundle coincides with \mathcal{F} on A(X) since \mathcal{F} is reflexive. Note that \mathcal{F} is a numerically flat vector bundle in this case.

In conclusion, we have shown that X is bimeromorphic to \mathbb{P}(\mathcal{F}) over A(X) which is isomorphic in codimension one. Following [KM98, Lemma 6.39], we show that it is in fact an isomorphism. Let p: \mathbb{P}(\mathcal{F}) \to A(X) be the natural projection. It is enough to show that for any \alpha-ample Cartier divisor D on X, the induced Cartier divisor D' on \mathbb{P}(\mathcal{F}) is p-nef. Since p is a projective morphism, it is enough to show for any (compact) curve C contained in a fiber of p, the intersection number (c_1(D') \cdot C) \geq 0. On the other hand, since p is locally trivial, C deforms to a curve over A(X) \setminus Z which is thus isomorphic to a curve in X. Thus (c_1(D') \cdot C) \geq 0 since D is \alpha-ample.

For convenience of the reader, we reproduce the arguments from [PS98, Proposition 0.4]. First, note that all non-minimal surfaces F with -K_F nef are isomorphic to the plane \mathbb{P}_2 blown up in at most 9 points in sufficiently general position.

We fix the Kähler class \omega on X. Consider the fiber F of \pi in the assumption and take a (-1)-curve E \subset F such that \omega \cdot E is minimal under all (-1)-curves in F. By assumption,

The normal bundle of E satisifes

Since H^1(E, \mathrm{Hom}(N_{F/X}|_E, N_{E/F}))=H^1(\mathbb{P}^1, \mathcal{O}(-1))=0, the short exact sequence splits and the normal bundle is of the form

In particular, E corresponds to a regular point of the Douady space. Note that the irreducible component of the Douady space containing E is of dimension of Y which is moreover compact. We will show in the following that it is in fact smooth.

To fix notations, there exists a compact complex space Y' and an irreducible effective divisor M \subset X \times Y', flat over Y' such that M \cap (X \times 0) = E, identifying X \times 0 with X. We let

and shall identify X with X \times t.

Claim 1 Every E_t is supported in some fiber F' of \pi.

Otherwise, \pi(E_t) will be a non-trivial effective curve. This contradicts the fact that the E_t are in the same cohomology class and so are the \pi(E_t). Note that [\pi(E)]=0 \in H^{2 \mathrm{dim} (Y)-2}(Y, \mathbb{Z}). In particular, there exists commuting diagram

The holomorphic map \varphi: Y' \to Y is surjective. More precisely, if it is not surjective, the image of Y' is a proper analytic subset of Y. In particular, by semicontinuity of the dimension of the fiber, the preimage of \pi(E) has strictly positive dimension. In other words, F contains a family of rational curves containing E of strictly positive dimension. This contradicts the fact that E is rigid in F.

To show that the morphism Y' \to Y is in fact locally isomorphic near the point coresponding to E, we need a local description of the Douady space using Kodaira’s theory, cf. [Kod62]. Locally, one can cover X by coordinate charts (z_i^\lambda,w_j^\lambda ), i=1,2; 1 \leq j \leq \mathrm{dim}(Y) where \lambda is the index of the coordinate chart such that \pi(z_i^\lambda,w_j^\lambda)=w_j^\lambda and E is locally given by \{z_2^\lambda= w_j^\lambda=0, \forall j \}. By Kodaira’s theory, the deformation of E is locally given by (z_1^\lambda, z_2^\lambda=\psi^\lambda(z_1^\lambda, t), w_j^\lambda=\varphi_j^\lambda(z_1^\lambda, t) ) for some local holomorphic functions \psi^\lambda(z_1^\lambda, t), \varphi_j^\lambda(z_1^\lambda, t) where t is the coordinate of the parameter space. The differential of the map t \mapsto (\psi^\lambda(z_1^\lambda, t), \varphi_j^\lambda(z_1^\lambda, t)) gives an identification of the tangent space of the parameter space at 0 with H^0(E, N_{E/X}). In particular, the differential of the map t \mapsto ( \varphi_j^\lambda(z_1^\lambda, t)) is isomorphic. (And \psi^\lambda(z_1^\lambda, t) is independent of t.) On the other hand, in local coordinates, \varphi(t)=(\varphi_j^\lambda(z_1^\lambda, t)) which implies that \varphi is locally isomorphic.

Note that the above calculation is valid for any smooth (-1)-rational curve.

Observe next that

Consider some E_{t_0}.

Write

with irreducible curves C_i (as a cycle, neglecting the possible embedded point). Since -K_{F'} is nef, we conclude (after renumbering possibly) that

and that

Since E_{t_0} is a deformation of E and the intersection number is invariant under deformation, we have that

Claim 2 C_0 is a (-1)-curve in F' and the C_i, i \geq 1, are (-2)-curves.

Define

Consider the map

by Serre duality. The last equality follows from the inequality (K_{F'}+C) \cdot \omega <0.

Since F' is rational, H^1(F', \mathcal{O}_{F'} / \mathcal{I})=H^1(F', \mathcal{O}_{F'} )=0.

Since H^1(E_{t_0}, \mathcal{O}_{E_{t_0}})=0, any irreducible component is a smooth rational curve. Adjunction formula shows that C_0 is a (-1)-curve in F' and the C_i, i \geq 1, are (-2)-curves.

Claim 3 E_{t_0} has only one irreducible component.

Now we look at the deformations of C_0 and obtain a family (C_s)_{s \in A.} For a small neighborhood \Delta' \subset Y' of t_0 the curve C_t is in \pi^{-1}(\varphi(t)) (strictly speaking there is a canonical map f: A \to Y, and f\vert_\Delta is biholomorphic on a small neighborhood \Delta \subset A of 0, so that we can identify t and f(t) for small t).

Therefore we can consider the (non-effective) family of cycles (E_t - C_t)_{t \in \Delta'} so that

By the choice of E_t, \omega \cdot E_t is minimal for general t, therefore \omega \cdot E_t \leq \omega \cdot C_t and we conclude

and therefore a_i = 0 for i \geq 1 so that E_{t_0} is irreducible and reduced.

Claim 4 Y' is smooth at t_0.

Note that E_{t_0} has no embedded point. In fact, an embedded point would increase the arithmetic genus, which would then contradict the fact that the arithmetic genus stays 0 in the flat family and that the reduced underlying space is a rational curve. Note that the reduced underlying space of E_{t_0} is a smooth (-1)-curve. Thus we have

and Y' is smooth at t_0.

In conclusion, Y' is smooth and M \to Y' is a \mathbb{P}^1-bundle. The

fact that any fiber of Y' represents some (-1)-curve in some fiber of \pi implies that Y' \to Y is étale. So set D = Y' and define S to be the irreducible component of \mathrm{pr}_1(M) \times_Y Y' mapping onto \mathrm{pr}_1(M).

First note that all non-minimal surfaces F with -K_F nef are isomorphic to the plane \mathbb{P}^2 blown up in at most 9 points in sufficiently general position (by the Kodaira-Enriques classification). All minimal rational ones are either \mathbb{P}^2 or \mathbb{P}^1 \times \mathbb{P}^1. The case of \mathbb{P}^2 is treated as follows: By [CP91, Lemma 7.4], up to some finite étale cover, we have X \simeq \mathbb{P}(E) for some vector bundles E over A(X) with trivial determinant. Since -K_X is nef, E is nef, hence numerically flat.

The case of \mathbb{P}^1 \times \mathbb{P}^1 can be treated similarly: Notice that the proof of [CP91, Lemma 9.4] works also in the compact Kähler setting. Up to some finite étale cover, we have X \simeq \mathbb{P}(E_1) \times_{A(X)} \mathbb{P}(E_2) for some vector bundles E_1,E_2 over A(X) with trivial determinant. By positivity of direct images applied to the submersions X \to \mathbb{P}(E_i) (i=1,2) with the condition that -K_X is nef, \mathbb{P}(E_i) (i=1,2) has nef anticanonical line bundle which implies that E_1 and E_2 are numerically flat.

The remaining case can be treated as follows: By the preceding remark 5.2, the factorisation as stated in case (3) exists. Thus, X_k is a compact manifold of class \mathcal{C} whose fibers of the Albanese morphism are either \mathbb{P}^2 or \mathbb{P}^1 \times \mathbb{P}^1. The conclusion follows from [CP91, Lemma 9.4] and [El82]. Note that proof of [CP91, Lemma 9.4] also works for compact manifolds of class \mathcal{C} by replacing the Chow variety by the corresponding Barlet space whose irreducible components are known to be compact. In particular, all X_i’s are in fact Kähler. Note that the Albanese morphism is always projective in this case. (In fact in this case, the Albanese morphism can be chosen to be a MRC quotient map.)

We claim that the Albanese morphism is locally constant (cf. [MW, Definition 2.3]) by applying [MW, Proposition 2.5]. Since all spaces are smooth, it is enough to show the existence of a relatively ample line bundle whose direct image is numerically flat. By [MW23, Theorem 2.6], it is enough to show the existence of a relatively ample and pseudoeffective line bundle. Denote by \pi: X \to X_k the composition with exceptional divisor E. Since -K_{X_k} is relatively ample with respect to the natural projection X_k \to A(X), for m \gg 0, m \pi^* (-K_{X_k})-E is \alpha-ample. On the other hand, -K_X=\pi^*(-K_{X_k})-E which implies that m \pi^* (-K_{X_k})-E=-mK_X+(m-1)E is pseudoeffective.

By [Cam81], to show that X is projective, it is enough to show that any two general points of X can be connected by a chain of (compact) curves. Without loss of generality, we can assume that the rational quotient is a holomorphic map since as a compact Kähler manifold, the projectivity of any modification of X would imply that X is Moishezon (and hence projective).

Take a very ample line bundle A over R(X) which exists by the projectivity assumption. By Bertini’s theorem, any two general points of R(X) can be connected by a smooth curve C defined as a complete intersection of sections in the linear system of A such that the general fiber over C is smooth and rationally connected. By Sard’s theorem, we can furthermore assume that the base change X_C is smooth. It is enough to show that X_C is projective which implies that any two points of X_C can be connected by a chain of (compact) curves.

By Kodaira’s criterion, as X_C is a compact Kähler manifold, it is enough to show that H^0(X_C, \Omega_{X_C}^2)=0. Let C_0 \subset C be the smooth locus such that each fiber is a rationally connected manifold. Any section vanishing over a non-empty Zariski open set of X_C has to vanish identically. Therefore, it is enough to show that H^0(X_{C_0}, \Omega_{X_{C_0}}^2)=0. We denote by f: X_{C_0} \to C_0 the restricted map which gives a short exact sequence of vector bundles

that induces a filtration on \Omega^2_{X_{C_0}}:

The graded pieces are

It is enough to show that the graded pieces have no global sections. R_0=0 since C is of dimension 1. H^0(X_{C_0}, \Omega^2_{X_{C_0}/C_0})=H^0(C_0, f_* \Omega^2_{X_{C_0}/C_0})=0 since by Grauert’s theorem, for any z \in C_0, H^0(X_z, \Omega^2_{X_{C_0}/C_0}|_{X_z})=0. Note that a rationally connected manifold has no non-trivial holomorphic forms. H^0(X_{C_0}, \Omega^1_{X_{C_0}/C_0} \otimes f^* \Omega^1_{{C_0}})=H^0(C_0, f_*(\Omega^1_{X_{C_0}/C_0} \otimes f^* \Omega^1_{{C_0}}))=H^0(C_0, f_*(\Omega^1_{X_{C_0}/C_0}) \otimes \Omega^1_{{C_0}}))=0 where the last equality follows from Grauert’s theorem. This finishes the proof of the lemma.

Consider the fibration \varphi: X \to Y predicted in Conjecture 1. By the well-known fact that a Calabi-Yau manifold can be algebraically approximated, there exists a deformation \pi: \mathfrak{Y} \to \Delta of Y over a small disk \Delta such that for Euclidean dense t \in \Delta, Y_t := \pi^{-1}(t) is a projective Calabi-Yau manifold. Let p: \tilde{\mathfrak{Y}} \to \mathfrak{Y} be the (holomorphic) universal cover of \mathfrak{Y}.

We claim that \tilde{Y}_t := (p \circ \pi)^{-1}(t) is the universal cover of Y_t for any t. Since a base change of étale morphism is still étale, the projection \tilde{Y}_t \to Y_t is étale (i.e. a unbranched covering map). Since \Delta is contractible, by Ehrensmann’s theorem, there exist 2 pointwise linear independent smooth vector fields over \mathfrak{Y}. These vector fields define a diffeomorphism between \mathfrak{Y} and Y_0 \times \Delta. Since p is étale, the pullback of these vector fields are well defined on \tilde{\mathfrak{Y}}. Up to a possible shrinking of \Delta, the pullback fields define a diffeomorphism between \tilde{\mathfrak{Y}} and \tilde{Y}_0 \times \Delta. In particular, for any t, \pi_1(\tilde{Y}_t)=\pi_1(\tilde{\mathfrak{Y}})=0 since \Delta is contractible. This finishes the proof of the claim.

Consider \mathfrak{X} := (\tilde{\mathfrak{Y}} \times F)/\pi_1(Y) \xrightarrow{\tilde{\varphi}} \mathfrak{Y} \to \Delta which defines a deformation of X. Since the group action is free, \mathfrak{X} is smooth. For any t \in \Delta, the morphism X_t:= (\pi \circ \tilde{\varphi})^{-1}(t) \to Y_t has rational connected fiber whose base is non-uniruled. Thus Y_t is a MRC quotient of X_t. For dense t, Lemma 6.1 implies that X_t is projective which finishes the proof. Note that for such dense t, -K_{X_t} is nef. Indeed, since -K_X is nef, by definition, for any \epsilon >0, there exists a smooth metric h_\epsilon such that its Chern curvature satisfies c_1(-K_X, h_\epsilon) \geq -\epsilon \omega. We pull back h_\epsilon to the universal cover q: \tilde{Y} \times F \to X of X whose restriction to \{ \tilde{y} \} \times F for any \tilde{y} \in \tilde{Y} defines a \pi_1(Y)-invariant metric on -K_F. The product metric defined by the restricted h_\epsilon and the Ricci flat metric on \tilde{Y}_t is \pi_1(Y)-invariant and thus descends to X_t which implies that -K_{X_t} is nef.