Failure of the Invariant Cycle Theorem over ℤ \mathbb{Z} Donu Arapura, François Greer, Yilong Zhang Abstract . We construct a counterexample to both the local and global invariant cycle theorems with integral coefficients. The example is a semistable one-parameter family of elliptic surfaces with p g = q = 1 p_{g}=q=1 and constant period map. The smooth fibers have the smallest possible discriminant, and are associated with Vinberg’s most algebraic K3 surface. Our construction generalizes the Shioda–Inose construction for rational double covers of K3 surfaces. 2020 Mathematics Subject Classification 14J27, 14D06 primary, 14J28, 14C30, 32G20 secondary 1. Introduction The classical local invariant cycle theorem, stated below, was conjectured by Griffiths [ Gri70 ] in 1970 and proved by Clemens and Schmid [ Cle77 ] [ GS75 , pp 121-122] shortly thereafter. Professor Griffiths kindly informed us that many of these ideas were worked out in a seminar at Berkeley in the 1960’s. We note that a different proof was found by Deligne [ Del80 , théorème 3.6] . Both proofs are nontrivial; the first relies on the existence of limit mixed Hodge structures and the second on the Weil conjectures. Theorem 1.1 (LICT Q ) . Let f\mathrel{\mathop{\mathchar 58\relax}}\mathcal{X}\to\Delta be a flat family of complex projective varieties over a disk. Suppose \mathcal{X} is smooth and f is smooth over the punctured disk \Delta^{*} . Let X_{t} be a general smooth fiber, then for each n , the restriction map H^{n}(\mathcal{X},\mathbb{Q})\twoheadrightarrow H^{n}(X_{t},\mathbb{Q})^{\pi_{1}(\Delta^{*})} surjects onto the invariant part. For semistable families (i.e., the central fiber \mathcal{X}_{0} is a reduced normal crossing divisor and the total space is smooth), the theorem follows from the exactness of the Clemens–Schmid sequence [ Cle77 , Mor84b ] H^{n}(X_{0},\mathbb{Q})\to H^{n}(X_{t},\mathbb{Q})\xrightarrow{N}H^{n}(X_{t},\mathbb{Q}), where N=\log(T) is the logarithm of the monodromy operator T , and \ker(N)=\ker(T-I) is precisely the invariant part H^{n}(X_{t},\mathbb{Q})^{\textup{inv}} (see Section 8 ). More generally, a much stronger form of Theorem 1.1 follows from the decomposition theorem [ BBD82 , Theorem 6.2.5] . However, the previous arguments only work rationally. A natural question to ask is Question 1.2 (LICT Z ) . Does the theorem still hold with integral coefficients? In other words, is the restriction map r\mathrel{\mathop{\mathchar 58\relax}}H^{n}(\mathcal{X},\mathbb{Z})\to H^{n}(X_{t},\mathbb{Z})^{\pi_{1}(\Delta^{*})} surjective? The answer to this question is no. A straightforward counterexample can be constructed as follows. Let \mathcal{X}\to\Delta be a local elliptic surface with a multiple fiber of multiplicity m>1 over 0 . Any class in the image of H^{2}(\mathcal{X},\mathbb{Z})\to H^{2}(X_{t},\mathbb{Z}) must be divisible by m . In particular, the generator of H^{2}(X_{t},\mathbb{Z}) , which is invariant, cannot lift to \mathcal{X} . One can preclude this sort of counterexample by working with semistable families. However, even with this assumption, we will see that the answer is no. This type of question was first studied by Friedman and Scattone [ Fri84 , FS86 ] , where they showed that LICT Z holds for a semistable family of K3 surfaces. Recently, Friedman and Griffiths showed it also holds for I-surfaces [ FG24 ] . In general, we show that LICT Z for H^{2} holds for any semistable family of surfaces with q=0 (see Proposition 8.2 ). 1.1. Failure of local ICT Z In this paper, we provide a negative example in which the family is semistable, and q=1 for the general fiber: Theorem 1.3 . The local invariant cycle “theorem” ( \textup{LICT}_{\mathbb{Z}} ) fails over \mathbb{Z} for a semistable family. More precisely, there exists a semistable degeneration of smooth elliptic surfaces with p_{g}=q=1 and torsion-free cohomology \mathcal{X}\to\Delta , such that the restriction map H^{2}(\mathcal{X},\mathbb{Z})\to H^{2}(X_{t},\mathbb{Z})^{\pi_{1}(\Delta^{*})} is not surjective. The counterexample comes from a one-parameter family of p_{g}=q=1 elliptic surfaces. We refer to such surfaces as elliptic-elliptic surfaces. They have K3-type Hodge structures on their second cohomology H^{2} , and nonzero first cohomology H^{1} . We construct a family of such surfaces over a complete curve with finite monodromy on H^{2} , and infinite monodromy on H^{1} . The family has several degenerations. Restricting to a neighborhood of a specific singular fiber, and then passing to a semistable model yields a family \mathcal{X}\to\Delta with a constant VHS on H^{2} , such that the image of H^{2}(\mathcal{X},\mathbb{Z}) is a proper subgroup of H^{2}(X_{t},\mathbb{Z}) . To prove this, we show that specialization map induces an injection on transcendental parts T_{X_{0}}\hookrightarrow T_{X_{t}}, which preserves the cup product pairings. We find that the central fiber X_{0} has a component birational to a specific elliptic K3 surface with discriminant 48 , and so does T_{X_{0}} . However, we show that the discriminant of T_{X_{t}} is 3 . Hence, the image of the restriction map is only an index 4 subgroup on transcendental lattices, so LICT Z fails. On the other hand, for divisor classes, LICT Z holds for this family (cf. Corollary 8.5 ). In the previous example, the maps in the Clemens–Schmid sequence are defined integrally. The following is a direct consequence of Theorem 1.3 . Corollary 1.4 . The Clemens–Schmid sequence fails to be exact over \mathbb{Z} for the semistable family in Theorem 1.3 . 1.2. Failure of global ICT Z For a proper surjective family \mathcal{X}\to B over a projective base B , Deligne proved the global invariant cycle theorem over \mathbb{Q} [ Del71 ] . See also [ Voi03 ] for a clear exposition. Theorem 1.5 (GICT Q [ Del71 ] ) . Let f\mathrel{\mathop{\mathchar 58\relax}}\mathcal{X}\to B be a proper surjective morphism between smooth projective varieties. Let U\subseteq B be the largest Zariski open subset over which f is smooth. Let X_{t} be a general fiber. Then the image of the restriction map H^{n}(\mathcal{X},\mathbb{Q})\to H^{n}(X_{t},\mathbb{Q}) is the subspace of invariants H^{n}(X_{t},\mathbb{Q})^{\pi_{1}(U,t)} . One can also ask if the global invariant cycle theorem holds integrally. It is a classical result that it holds for a Lefschetz pencil arising from hyperplane sections of a smooth projective variety [ Lam81 ] . We show that it can still be false in general: Theorem 1.6 . The global invariant cycle "theorem" ( GICT Z ) fails over \mathbb{Z} for a one-dimensional semistable family. More precisely, There is a smooth projective 3-fold \mathcal{X} , a smooth projective curve C , and a proper surjective morphism \mathcal{X}\to C with semistable singular fibers, such that the restriction map H^{2}(\mathcal{X},\mathbb{Z})\to H^{2}(X_{t},\mathbb{Z})^{\pi_{1}(U,t)} is not surjective. Note that in our example, the Leray spectral sequence still degenerates over \mathbb{Z} for the smooth fibration \mathcal{X}_{U}\to U because U is an open Riemann surface homotopy equivalent to a 1-dimensional CW complex, so the invariant cycles lift to H^{2}(\mathcal{X}_{U},\mathbb{Z}) . However, it fails to lift to H^{2}(\mathcal{X},\mathbb{Z}) due to Theorem 1.3 . 1.3. The construction The family \mathcal{X}\to\Delta used in the proof of Theorem 1.3 has constant period map. Each smooth fiber is Hodge-theoretically associated with a K3 surface Y , which is one of Vinberg’s “most algebraic” K3 surfaces [ Vin83 ] with discriminant d=3 , in the sense that there is an Hodge isometry T_{X_{t}}\cong T_{Y} . They are related by the following diagram, formally analogous to a Shioda–Inose structure [ SI77 , Mor84a ] : {\lx@inpgf@ignorespaces X_{t}} {\lx@inpgf@ignorespaces Y} {\lx@inpgf@ignorespaces Z} \scriptstyle{\lx@inpgf@ignorespaces\pi_{1}} \scriptstyle{\lx@inpgf@ignorespaces\pi_{2}} Shioda and Inose [ SI77 ] showed that Y admits a rational map of degree two onto the Kummer surface Z=\textup{Kum}(E_{\omega}\times E_{\omega}) associated to the self-product of the elliptic curve E_{\omega}=\mathbb{C}/(\mathbb{Z}+\mathbb{Z}\omega) with j=0 . The rational map \pi_{2} arises from a quadratic base change of an elliptic fibration on Z branched at two special fibers ( IV^{*},I_{0}^{*} ). We show that the quadratic base change branched at three special fibers ( II^{*},IV^{*},I_{0}^{*} ), plus an additional smooth fiber lying over t produces an elliptic-elliptic surface X_{t} . As t varies, the Hodge structure H^{2}(X_{t},\mathbb{Z}) remains constant. That is, the period map for this family sends all values of t to a point. The three singular members of the family occur when t collides with each of the special values lying under singular fibers of Z . Each of these singular surfaces contains a reduced component birational to a K3 surface; one of these K3 surfaces is Y . This forces \mathrm{disc}(T_{X_{t}})=3 . On the other hand, when t collides with the special value lying under the I_{0}^{*} fiber, the limiting K3 is different; it has discriminant d=48 . This allows us to deduce Theorem 1.3 . A second counterexample can be constructed similarly, beginning instead with Vinberg’s second most algebraic K3, which has discriminant d=4 (cf. Theorem 7.1 ). Outline In Section 2 , we recall basics on transcendental Hodge structures, elliptic surfaces, and Shioda–Inose structures. In Section 3 , we study quadratic base change of elliptic surfaces and construct a smooth family of elliptic-elliptic surfaces with constant VHS on H^{2} . In Section 4 , we extend the smooth family to a proper family with controlled singular fibers. In Section 5 , we prove Theorem 1.3 . In Section 6 , we prove Theorem 1.6 . In Section 7 , we construct a second counterexample using the same techniques. In Section 8 , we discuss some positive results for ICT Z . In Section 9 , we study the correspondence problem between the elliptic-elliptic surfaces in question and their associated K3 surfaces, and we pose an open question. 2. Preliminaries 2.1. Transcendental Hodge structures In this paper, the word surface , without any further qualification, will mean a smooth projective surface over \mathbb{C} . If S is a surface, the Néron–Severi group NS(S) is the subgroup of divisor classes in H^{2}(S,\mathbb{Z}) . The intersection pairing on NS(S)/\textup{torsion} is nondegenerate, so it carries a lattice structure. Its orthogonal complement T_{S} in H^{2}(S,\mathbb{Z}) is called the transcendental lattice. T_{S}\mathrel{\mathop{\mathchar 58\relax}}=NS(S)^{\perp}\subset H^{2}(S,\mathbb{Z}). By Wu’s formula, T_{S} is always an even lattice, meaning that for every x\in T_{S} , x\cdot x\in 2\mathbb{Z} . One has |\det(NS(S))|=|\det(T_{S})| , and this positive integer is called the discriminant of S and is denoted by \mathrm{disc}(S) . The complexification T_{S}\otimes\mathbb{C} contains H^{2,0}(S,\mathbb{C}) , and it carries a Hodge filtration induced from H^{2}(S,\mathbb{C}) . Hence, T_{S} (resp. T_{S,\mathbb{Q}} ) carries a weight two \mathbb{Z} -Hodge structure (resp. \mathbb{Q} -Hodge structure) induced from H^{2}(S,\mathbb{Z}) . We call \phi\mathrel{\mathop{\mathchar 58\relax}}T_{S}\to T_{S^{\prime}} a Hodge isometry if \phi is an isometry as a lattice and preserves the Hodge filtration. Lemma 2.1 . The transcendental Hodge structure T_{S,\mathbb{Q}}\subset H^{2}(S,\mathbb{Q}) is the smallest \mathbb{Q} sub-Hodge structure whose complexification contains H^{2,0}(S) . Furthermore T_{S}=T_{S,\mathbb{Q}}\cap H^{2}(S,\mathbb{Z})/\textup{torsion}. Proof. By the Lefschetz (1,1) theorem, NS(S)_{\mathbb{Q}} is the maximal Hodge structure contained in H^{1,1}(S) . hence the first statement follows from the fact that H^{2,0}(S) is orthogonal to H^{1,1}(S) . The second statement follows from the fact that T_{S} is primitive in H^{2}(S,\mathbb{Z})/\text{torsion} . ∎ Given a dominant rational map \pi\mathrel{\mathop{\mathchar 58\relax}}S\dashrightarrow S^{\prime} between surfaces. There is a pullback \pi^{*} and a pushforward \pi_{*} (of free abelian groups) \pi^{*}\mathrel{\mathop{\mathchar 58\relax}}T_{S^{\prime}}\to T_{S},\ \pi_{*}\mathrel{\mathop{\mathchar 58\relax}}T_{S}\to T_{S^{\prime}}. To see this, one can resolve the indeterminacy of \pi : S\xleftarrow{\sigma}\tilde{S}\xrightarrow{p}S^{\prime}, where \sigma is birational. We can define \pi^{*}\mathrel{\mathop{\mathchar 58\relax}}=\sigma_{*}\circ p^{*} and \pi_{*}\mathrel{\mathop{\mathchar 58\relax}}=p_{*}\sigma^{*} on transcendental lattices. It is well-defined since the transcendental lattice is birationally invariant (both \sigma_{*} and \sigma^{*} induce isomorphisms T_{S}\cong T_{\tilde{S}} ), and the definition is independent of the resolution \sigma . We remark that \pi^{*} and \pi_{*} do not preserve the intersection pairings in general. When p_{g}(S)=p_{g}(S^{\prime}) , then both \pi^{*} and \pi_{*} are injective and the image is a finite index sublattice. Thus there is an isomorphism of (unpolarized) \mathbb{Q} -Hodge structures: T_{S}\otimes\mathbb{Q}\cong T_{S^{\prime}}\otimes\mathbb{Q}. 2.2. Elliptic surfaces An elliptic surface is a smooth projective surface S together with map f\mathrel{\mathop{\mathchar 58\relax}}S\to C to a smooth projective curve C , whose general fiber is an elliptic curve. We call an elliptic surface relatively minimal if there is no ( -1 )-curve contained in a fiber of f . If f has a section, then the set of sections of f form an abelian group MW(S/C) , called the Mordell–Weil group, which is finitely generated as long as f has a singular fiber. The rank of this group is called the Mordell–Weil rank. There is a \mathbb{Q} -lattice structure on the torsion-free part of the Mordell–Weil group, called the Mordell–Weil lattice. One refers to [ SS10 , SS19 ] for details. Without further qualification, we will assume that all elliptic surfaces are relatively minimal and admit a section. The possible singular fibers in an elliptic surface were classified by Kodaira, Néron, and Tate [ SS19 , Section 5.4] into the following types: II,II^{*},III,III^{*},IV,IV^{*},I_{n},I_{n}^{*}\ \textup{with}\ n\geq 0. These singular fibers occur as the Kodaira–Néron models for any one-parameter family of elliptic curves over \Delta^{*} . Among all the singular fiber types, we are particularly interested in the following subset featuring the superscript ∗ : Definition 2.2 . The singular fiber types II^{*},III^{*},IV^{*},I_{n}^{*} , n\geq 0 are called star fibers . They are all reducible and non-reduced; the dual graph encoding the incidence of all components not meeting the zero section is a Dynkin diagram of type \mathrm{E}_{8},\mathrm{E}_{7},\mathrm{E}_{6} , and \mathrm{D}_{4+n} , respectively. These fiber types play an important role in the study of quadratic base change of elliptic surfaces. We record the following useful observation for later use: Lemma 2.3 . In each star fiber F , there are exactly four components with odd multiplicity, and they are disjoint. Proof. This can be directly verified from [ SS19 , Theorem 5.12] . ∎ The canonical bundle of an elliptic surface f\mathrel{\mathop{\mathchar 58\relax}}S\to C is given by \omega_{S}\mathrel{\mathop{\mathchar 58\relax}}=f^{*}(\omega_{C}\otimes\mathbb{L}_{S}), where \mathbb{L}_{S}=f_{*}(\omega_{S/C}) is the fundamental line bundle of degree \chi(S) on C . Lemma 2.4 . The degree of fundamental line bundle of an elliptic surface satisfies \deg\mathbb{L}_{S}=\frac{1}{12}e(S)=\frac{1}{12}\sum_{c\in\Delta}e(f^{-1}(c)), where e(\cdot) denotes the topological Euler characteristic, and \Delta is the subset of C that corresponds to the singular fibers. In particular, \deg\mathbb{L}_{S}>0 if and only if there is at least one singular fiber. Lemma 2.5 . The geometric genus p_{g}(S) of an elliptic surface satisfies p_{g}(S)=\deg\mathbb{L}_{S}+g(C)-1. Remark 2.6 . There are exactly two types of elliptic surfaces S\to C with p_{g}=1 that have positive \deg\mathbb{L}_{S} : one has g(C)=0 and \deg\mathbb{L}_{S}=2 , and is an elliptic K3 surface; the other has g(C)=1 and \deg\mathbb{L}_{S}=1 , and is called an elliptic-elliptic surface. An elliptic-elliptic surface S\to C is an elliptic surface with p_{g}=q=1 together with a section [ EGW24 , GZ24 ] . Equivalently, the base curve has genus one, and the degree of the fundamental line bundle is one also. The cohomology group H^{2}(S,\mathbb{Z}) carries a weight two Hodge structure with Hodge numbers h^{2,0}=h^{0,2}=1,h^{1,1}=12 . Therefore, S has a unique nonzero holomorphic 2-form \omega up to a scaling. It vanishes along a unique fiber, which is called the canonical fiber. Lemma 2.7 . [ GS85 , Lemma D] If an elliptic surface S has a section and at least one singular fiber, then its second cohomology H^{2}(S,\mathbb{Z}) is torsion-free. 2.3. Shioda–Inose structure Shioda and Inose [ SI77 ] showed that every K3 surface with maximal Picard rank has an associated pair of isogenous CM elliptic curves. More precisely, for a Picard maximal K3 surface Y , the transcendental lattice is even integral positive definite binary quadratic form T_{Y}\cong\begin{bmatrix}2a&b\\ b&2c\end{bmatrix};\ a,b,c\in\mathbb{Z} with discriminant d=4ac-b^{2} . Theorem 2.8 . A K3 surface with maximal Picard number is uniquely determined by its transcendental lattice. There is a pair of isogenous elliptic curves E and E^{\prime} associated with periods \tau=\frac{-b\pm\sqrt{-d}}{2a} . Shioda and Inose showed that there is a Hodge isometry \Phi\mathrel{\mathop{\mathchar 58\relax}}T_{Y}\cong T_{E\times E^{\prime}} which is geometrically realized by the diagram below. (1) {\lx@inpgf@ignorespaces Y} {\lx@inpgf@ignorespaces E\times E^{\prime}} {\lx@inpgf@ignorespaces\textup{Kum}(E\times E^{\prime}).} \scriptstyle{\lx@inpgf@ignorespaces\pi_{1}} \scriptstyle{\lx@inpgf@ignorespaces\pi_{2}} Here, X=\textup{Kum}(E\times E^{\prime}) is the Kummer surface associated with E\times E^{\prime} , and \pi_{i} is a rational double cover, such that (\pi_{1})_{*}\mathrel{\mathop{\mathchar 58\relax}}T_{Y}(2)\cong T_{X} . Consequently, they show that the Hodge isometry \Phi is induced by an algebraic cycle. To construct \pi_{1} , Shioda–Inose find a special elliptic fibration on the Kummer surface and realize Y via a quadratic base change: Lemma 2.9 . ( [ SI77 ] ) The Kummer surface X=\textup{Kum}(E_{1}\times E_{2}) associated with product of two elliptic curves admits an elliptic fibration X\to\mathbb{P}^{1} with a section and singular fiber types (i) II^{*},IV^{*},I_{0}^{*} , or (ii) II^{*},I_{b_{1}}^{*},I_{b_{2}}^{*} , b_{1}+b_{2}\leq 2 . Moreover, if \pi\mathrel{\mathop{\mathchar 58\relax}}Y\dashrightarrow X is the rational double cover arising from quadratic base change branched at the two star fibers different from II^{*} , then Y is a K3 surface and \pi_{*} induces a Hodge isometry of transcendental lattices \pi_{*}\mathrel{\mathop{\mathchar 58\relax}}T_{Y}(2)\cong T_{X}. In particular, the K3 surface Y admits a Nikulin involution with 8 fixed points. In general, Morrison [ Mor84a ] showed that if a K3 surface Y admits a transcendental Hodge isometry T_{Y}\cong T_{A} to the transcendental lattice of an abelian surface A , then there is a diagram similar to ( 1 ), but with E\times E^{\prime} replaced by A ; such a diagram is called a Shioda–Inose structure for the K3 surface Y . In this paper, we explore an analogous diagram and transcendental lattice comparison between K3 surfaces and elliptic-elliptic surfaces. 3. Quadratic base change and variations of Hodge structure In this section, we investigate certain special base changes of elliptic surfaces, in which the degree of the fundamental line bundle is smaller than expected, and study their Hodge structures. Our main application is to construct elliptic-elliptic surfaces as rational double covers of elliptic K3 surfaces. 3.1. Euler defect of base change Let X^{*}\to\Delta^{*} be a family of smooth elliptic curves parameterized by a punctured disk \Delta^{*} , and X\to\Delta be the Kodaira–Néron model with singular fiber F over 0 . Let \tilde{\Delta}\to\Delta be the 2 -to-1 cover totally branched at 0 , then the base change X\times_{\Delta}\tilde{\Delta} is birational to a new family of elliptic curves \tilde{X}\to\tilde{\Delta} , with Kodaira fiber type F^{\prime} . Using Kodaira’s notation, the base change fiber type table is listed below [ SS19 , p.105] for star fibers: F F^{\prime} II^{*} IV^{*} III^{*} I_{0}^{*} IV^{*} IV I_{n}^{*} I_{2n} Table 1 . Double cover and reduction of Kodaira’s star fibers Definition 3.1 . For each Kodaira singular fiber of type F , we define (2) \delta(F)\mathrel{\mathop{\mathchar 58\relax}}=\frac{1}{12}(2\cdot e(F)-e(F^{\prime})), with respect to a quadratic base change, where e(\cdot) is the Euler characteristic. Proposition 3.2 . For every Kodaira fiber of type F , \delta(F)\in\{0,1\} . We have \delta(F)=1 if and only if F is a star fiber. Proof. It is a direct consequence of the fact that e(II)=2 , e(III)=3 , e(IV)=4 , e(II^{*})=10 , e(III^{*})=9 , e(IV^{*})=8 , e(I_{n}^{*})=6+n , e(I_{n})=n . ∎ Let f\mathrel{\mathop{\mathchar 58\relax}}X\to C be an elliptic surface and \pi\mathrel{\mathop{\mathchar 58\relax}}C^{\prime}\to C a branched double cover. Let f^{\prime}\mathrel{\mathop{\mathchar 58\relax}}X^{\prime}\to C^{\prime} be the relatively minimal regular model of the fibered product X\times_{C}C^{\prime} . We want to investigate the relationship between the degree of fundamental line bundles d=\deg(\mathbb{L}_{X}) and d^{\prime}=\deg(\mathbb{L}_{X^{\prime}}) . Lemma 3.3 . There is an inequality d^{\prime}\leq 2d. Moreover, the difference (3) \delta\mathrel{\mathop{\mathchar 58\relax}}=2d-d^{\prime} equals the number of points p\in C such that • the fiber f^{-1}(p) is a star fiber, and • p is a branch point of the double cover C^{\prime}\to C . Proof. By Lemma 2.4 , d=e(X)/12 , and d^{\prime}=e(X^{\prime})/12 , and both e(X) and e(X^{\prime}) are determined by the Euler characteristic of the singular fibers. Hence, it suffices to compare the singular fibers on the two surfaces. When p\in C correspond a singular fiber F of f , then by Proposition 3.2 , there are three possibilities (i) p is unbranched for \pi , then it corresponds to two singular fibers on X^{\prime} of type F_{p} ; (ii) p is a branched point, and F_{p} is not of star type, and \delta(F_{p})=0 ; (iii) p is a branched point, and F_{p} is of star type, and \delta(F_{p})=1 . Therefore, only the last case contribute to the difference 2e(X)-e(X^{\prime}) . Using Lemma 2.4 again, we find 2e(X)-e(X^{\prime})=\sum_{p\in\Delta}\big(2e(F_{p})-e(F_{p}^{\prime})\big), where \Delta\subset C is the collection of points satisfy (iii), and F_{p}^{\prime} is the type of the singular fiber upstairs according to Table 1 . Then divide the equality by 12 and the result follows from the definition ( 2 ) of \delta . ∎ Definition 3.4 . We call \delta (cf. ( 3 )) the Euler defect of the quadratic base change of elliptic surface X\to C with respect to a double cover C^{\prime}\to C . We say the base change defective if \delta>0 . Remark 3.5 . The Euler defect can be defined in general for higher degree base change. This will be studied more systematically in a subsequent paper. In this paper, we focus on the case of quadratic base change case, when g(C^{\prime})=1 . 3.2. Defective double cover of elliptic K3 Now we start with an elliptic K3 surface f\mathrel{\mathop{\mathchar 58\relax}}X\to\mathbb{P}^{1} with (a certain number of) star fibers. We denote by f^{\prime}\mathrel{\mathop{\mathchar 58\relax}}X^{\prime}\to C^{\prime} the relatively minimal regular model of the elliptic surface that comes from the quadratic base change C^{\prime}\to\mathbb{P}^{1} . The geometry of the surface X^{\prime} is listed below for g(C^{\prime})=0,1 and \delta=2,3,4 below. Corollary 3.6 . (1) ( \delta=2 , K3 surface) If \mathbb{P}^{1}=C^{\prime}\to\mathbb{P}^{1} is the double cover branched at points lying under two star fibers, then f^{\prime}\mathrel{\mathop{\mathchar 58\relax}}X^{\prime}\to\mathbb{P}^{1} is an elliptic K3 surface. (2) ( \delta=3 , elliptic-elliptic surface) If E=C^{\prime}\to\mathbb{P}^{1} is the double cover branched at four points with three points lying under star fibers, then f^{\prime}\mathrel{\mathop{\mathchar 58\relax}}X^{\prime}\to E is an elliptic-elliptic surface. (3) ( \delta=4 , abelian surface) If E=C^{\prime}\to\mathbb{P}^{1} is the double cover branched at four points that all lie under star fibers. Then f^{\prime}\mathrel{\mathop{\mathchar 58\relax}}X^{\prime}\to E is trivial family, and X^{\prime} is an abelian surface isogenous to a product of elliptic curves. Proof. By Lemma 3.3 , the Euler defect of the three quadratic base change are \delta=2,3,4 , respectively. The elliptic K3 surface f\mathrel{\mathop{\mathchar 58\relax}}X\to\mathbb{P}^{1} has degree of fundamental line bundle d=2 . Therefore, by Lemma 3.3 , d^{\prime}=2\cdot 2-\delta , which equals to 2,1 , and 0 , respectively. In case (1), the base curve is \mathbb{P}^{1} and d^{\prime}=2 , so X^{\prime} is a K3 surface. In case (2), the base curve has genus one and d^{\prime}=1 , so X^{\prime} is elliptic-elliptic surface. In case (3), d^{\prime}=0 implies X^{\prime}\to C^{\prime} is a trivial elliptic family and isogeny to a product E\times E^{\prime} . ∎ Note that case (1) includes the Shioda–Inose double cover (Lemma 2.9 ); Case (3) is the standard rational double cover E\times E^{\prime}\dashrightarrow\textup{Kum}(E\times E^{\prime}) of the Kummer surface associated with product of two elliptic curves. Remark 3.7 . The construction in case (1) and (3) are rigid with respect to X\to\mathbb{P}^{1} ; While in the case (2), the construction of elliptic-elliptic surfaces has a free parameter, since one can vary the last branch fiber freely. The Hodge structure of the one-parameter family will be studied in Section 3.4 . Remark 3.8 . The assumptions in Corollary 3.6 put a lower bound on the Picard number \rho=\textup{rank}(NS(X)) of the K3 surface X . Indeed, case (1), (2), and (3) assume X\to\mathbb{P}^{1} has at least 2, 3, and 4 star fibers. A star fiber has at least 5 components (realized by I_{0}^{*} ), so by Shioda–Tate formula [ SS19 , Corollary 6.7] , \rho is at least 10,14 , and 18 , respectively. This also gives a lower bound of Picard number of X^{\prime} as well (see Lemma 3.12 ). 3.3. Birational geometry of defective double cover In this section, we analyze defective quadratic double covers from the point of view of birational geometry. Lemma 3.9 . Suppose that the special fiber F of a Kodaira–Néron model X\to\Delta is a star fiber over 0 . Let \tilde{\Delta}\to\Delta be a double cover branched at 0 . Then the normalization \tilde{X}^{\prime} of X\times_{\Delta}\tilde{\Delta} is smooth, and contains four disjoint (-1) curves. After contracting the four curves, one obtains the relatively minimal model X^{\prime}\to\tilde{\Delta} with singular fiber of type F^{\prime} listed in Table 1 . Proof. According to [ HM98 , p.125] , the composition p\mathrel{\mathop{\mathchar 58\relax}}\tilde{X}^{\prime}\to X\times_{\Delta}\tilde{\Delta}\to X is the double cover branched along the components with odd multiplicity on the fiber F . Then the result follows from the Lemma 2.3 . ∎ Of course, one can formulate a version of Lemma 3.9 for any compact elliptic surfaces. But from now on, we only focus on elliptic-elliptic surfaces. Proposition 3.10 . Let S\to E be an elliptic-elliptic surface arising from quadratic base change of an elliptic K3 surface X\to\mathbb{P}^{1} branched at three star fibers and one smooth fiber. Then the rational double cover S\dashrightarrow X factors through the following diagram {\lx@inpgf@ignorespaces S^{\prime}} {\lx@inpgf@ignorespaces X} {\lx@inpgf@ignorespaces S,} \scriptstyle{\lx@inpgf@ignorespaces p} \scriptstyle{\lx@inpgf@ignorespaces\sigma} \scriptstyle{\lx@inpgf@ignorespaces\pi} where \sigma is a blow-up at 12 points, and p is branched double cover ramified along the exceptional divisors. Proof. By assumption, the branch loci of E\to\mathbb{P}^{1} underlies three star fibers. Therefore, by Lemma 3.9 , the normalization S^{\prime} of E\times_{\mathbb{P}^{1}}X has 4\times 3=12 curves with self-intersection (-1) . Contract these curves, we obtain the relatively minimal model S\to E , which is an elliptic-elliptic surface by the fact that \deg(\mathbb{L}_{S})=1 (cf. Corollary 3.6 , (2)). ∎ In particular, there is an involution \iota^{\prime} on S^{\prime} fixing the 12 components and S^{\prime}/\iota^{\prime}\cong X . By descending the involution to S and commuting the \mathbb{Z}_{2} -quotient with the blow-up, we have Corollary 3.11 . There is a symplectic involution \iota\mathrel{\mathop{\mathchar 58\relax}}S\to S with 12 fixed points. The minimal resolution of the quotient S/\iota is the K3 surface X . Note that this is analogous to the Shioda–Inose structure for K3 surfaces [ vGS07 ] . The K3 surface Y that arises from rational double cover of an elliptic Kummer K3 surface is branched along two star fibers, and each contains 4 components with odd multiplicity (cf. Lemma 2.3 ). Therefore, they correspond to 8 fixed points of the Nikulin involution on Y . 3.4. Hodge theory of defective double cover We analyze the transcendental Hodge structure of the defective double cover of elliptic surfaces, and investigate a one-parameter family of elliptic-elliptic surface in Corollary 3.6 . Lemma 3.12 . In the setup of Proposition 3.10 , the maps \pi^{*}\mathrel{\mathop{\mathchar 58\relax}}T_{X}\hookrightarrow T_{S} and \pi_{*}\mathrel{\mathop{\mathchar 58\relax}}T_{S}\hookrightarrow T_{X} are injective morphisms of \mathbb{Z} -Hodge structures. In particular, there is an isomorphism of (unpolarized) \mathbb{Q} -Hodge structures T_{X}\otimes\mathbb{Q}\cong T_{S}\otimes\mathbb{Q}, and \pi^{*}(T_{X}) is a finite index sublattice of T_{S} . Proof. As explained in Section 2.1 , both \pi_{*} and \pi^{*} are defined on transcendental lattice for any rational map. Secondly, \pi has degree 2 , so \pi_{*}\pi^{*}\mathrel{\mathop{\mathchar 58\relax}}T_{X}\to T_{X} is 2\cdot Id . Hence, \pi^{*} is injective. Moreover, by Hartogs theorem, the pullback of holomorphic 2-form extends, hence \pi^{*}\otimes\mathbb{C}\mathrel{\mathop{\mathchar 58\relax}}H^{2,0}(X)\to H^{2,0}(S) preserves the Hodge filtration and \pi^{*} is a morphism of \mathbb{Z} -Hodge structures. and so is \pi_{*} . Since \dim H^{2,0}(S)=1 , T_{S}\otimes\mathbb{Q} is an irreducible \mathbb{Q} -Hodge structure, and there is an isomorphism \pi^{*}T_{X}\otimes\mathbb{Q}\cong T_{S}\otimes\mathbb{Q} of \mathbb{Q} -Hodge structures. In particular, \pi^{*}T_{X} has finite index in T_{S} . ∎ Remark 3.13 . By [ Mor84a ] , the morphisms \pi_{*}\mathrel{\mathop{\mathchar 58\relax}}T_{S}(2)\to T_{X} and \pi^{*}\mathrel{\mathop{\mathchar 58\relax}}T_{X}(2)\to T_{S} preserve the lattice pairings. Next, we allow the construction of Proposition 3.10 to vary in a family, and study the corresponding VHS. Lemma 3.14 . There is a non-isotrivial family of elliptic-elliptic surfaces f_{U}\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}_{U}\to U over a quasi-projective curve U , that arises from defective quadratic base changes of a fixed elliptic K3 surface X\to\mathbb{P}^{1} (cf. Corollary 3.6 ) at three star fibers. More precisely, there is a non-isotrivial family g_{U}\mathrel{\mathop{\mathchar 58\relax}}\mathcal{E}_{U}\to U of elliptic curves and map h_{U}\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}_{U}\to\mathcal{E}_{U} such that f_{U}=g_{U}\circ h_{U} , and for every t\in U , the induced morphism on fibers h_{t}\mathrel{\mathop{\mathchar 58\relax}}S_{t}\to E_{t} is an elliptic-elliptic surface that is the Kodaira–Néron model of the base change E_{t}\times_{\mathbb{P}^{1}}X . Proof. For convenience, assume that the star fibers of X\to\mathbb{P}^{1} lie over the values \{0,1,2\}\subset\mathbb{P}^{1}. Let C be a projective curve with a non-constant morphism \lambda\mathrel{\mathop{\mathchar 58\relax}}C\to\mathbb{P}^{1} of even degree, and let \mathcal{E} be a cyclic double cover of C\times\mathbb{P}^{1} branched at the 2-divisible divisor D=\Gamma_{\lambda}+\sum_{j=0}^{2}C\times\{j\}, where \Gamma_{\lambda} is the graph of \lambda . Let U=C\smallsetminus\textup{pr}_{1}(D_{sing}) be the open subset of C over which D is étale. Observe that g_{U}\mathrel{\mathop{\mathchar 58\relax}}\mathcal{E}_{U}\to U is an elliptic fibration with four distinguished sections, and it is non-isotrivial because \lambda was non-constant. Now shrink U further by removing \lambda^{-1}(c) for any other critical values of X\to\mathbb{P}^{1} besides \{0,1,2\} . There is a morphism \mathcal{E}_{U}\to\mathbb{P}^{1} given by the composition of the double cover above with the second projection map C\times\mathbb{P}^{1}\to\mathbb{P}^{1} . The fibered product \mathcal{E}_{U}\times_{\mathbb{P}^{1}}X has normalization \mathcal{S}^{\prime}_{U} , which is birational to the desired \mathcal{S}_{U} . We will construct a commutative diagram generalizing Proposition 3.10 to the case of a family over U . {\lx@inpgf@ignorespaces\mathcal{S}^{\prime}_{U}/U} {\lx@inpgf@ignorespaces X\times U} {\lx@inpgf@ignorespaces\mathcal{S}_{U}/U.} \scriptstyle{\lx@inpgf@ignorespaces P} \scriptstyle{\lx@inpgf@ignorespaces\Sigma} \scriptstyle{\lx@inpgf@ignorespaces\Pi} Here, P is a branched double cover, and \Sigma will be a birational contraction. Note that each fiber of \mathcal{S}^{\prime}_{U} contains 12 disjoint (-1) -curves. The monodromy action of \pi_{1}(U) on this configuration is trivial, since they come from 12 curves in X , so each such exceptional curve traces out a divisor R_{i}\subset\mathcal{S}_{U} . The P -pullback of an ample divisor on X\times U gives an ample divisor on \mathcal{S}^{\prime}_{U} which is also monodromy-invariant. Hence, by the Relative Cone Theorem [ KM98 , Theorem 3.25] , there is a sequence of contractions over U of the ruled divisors R_{i} , whose composite \Sigma\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}_{U}^{\prime}\to\mathcal{S}_{U} results in the desired family f_{U}\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}_{U}\to\mathcal{E}_{U}\to U . ∎ Proposition 3.15 . The family constructed in Lemma 3.14 above satisfies the following. • The VHS on H^{2} is locally constant. More precisely, the polarized variation of Hodge structure R^{2}f|{{}_{U}*}\mathbb{Z} admits a decomposition after tensoring with \mathbb{Q} R^{2}f|{{}_{U}*}\mathbb{Q}=\mathcal{T}_{\mathbb{Q}}\oplus\mathcal{NS}_{\mathbb{Q}}, where \mathcal{T} is the sublocal system defined by transcendental lattices T_{S_{t}} and \mathcal{NS}=\mathcal{T}^{\perp} is the sublocal system defined by Néron–Severi groups NS(S_{t}) . The factor \mathcal{T}_{\mathbb{Q}} has trivial monodromy, and \mathcal{NS}_{\mathbb{Q}} has finite monodromy. • The VHS on H^{1} is non-constant. Proof. The rational map \Pi\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}_{U}\dashrightarrow X\times U gives a morphism of \mathbb{Q} -VHS \Pi^{*}\mathrel{\mathop{\mathchar 58\relax}}H^{2}(X,\mathbb{Q})\otimes\underline{\mathbb{Q}}_{U}\to R^{2}f|{{}_{U}*}\mathbb{Q}, whose stalks are \pi_{t}^{*}\mathrel{\mathop{\mathchar 58\relax}}H^{2}(X,\mathbb{Q})\to H^{2}(S_{t},\mathbb{Q}) . By Lemma 3.12 , the image local system \mathcal{T}_{\mathbb{Q}}=\textup{Im}(\Pi^{*}|_{T_{X}}) is a constant sub-VHS isomorphic to T_{X}\otimes\underline{\mathbb{Q}}_{U} . Since NS(S_{t}) contains a monodromy-invariant ample class, the polarized Néron–Severi is an integral PVHS of type (1,1) , so its monodromy is both compact and discrete, hence finite. In particular, \mathcal{NS} is locally constant, and so is R^{2}f|_{U*}\mathbb{Z} . For the second statement, there is an isomorphism of Hodge structures H^{1}(S_{t},\mathbb{Z})\cong H^{1}(E_{t},\mathbb{Z}) . Since E_{t} varies non-isotrivially, the last statement follows from Lemma 3.14 . ∎ Corollary 3.16 . The weight 2 period map for the family \mathcal{S}_{U}\to U is constant. Remark 3.17 . Note that by Corollary 3.16 , infinitesimal Torelli fails for each surface S_{t} in the family. This is consistent with a criterion in [ GZ24 ] for infinitesimal Torelli in terms of the j -map at the canonical fiber associated with S_{t}\to E_{t} . 3.5. Examples of families with constant VHS In this section, we revisit some examples of one-dimensional families of elliptic-elliptic surfaces with constant period map [ Ike19 , EGW24 ] . We construct them in a uniform way using Corollary 3.6 . Example 3.18 . [ Ike19 ] Ikeda found a 4-dimensional family \mathcal{I}_{1} of elliptic-elliptic surfaces arising from Prym varieties of bielliptic curves of genus 3. The period image of \mathcal{I}_{1} is 3-dimensional, and the generic surface in \mathcal{I}_{1} has six I_{2} fibers. These surfaces also arise from quadratic base change of an elliptic K3 surface X\to\mathbb{P}^{1} with singular fibers of type I_{2}\times 3,I_{0}^{*}\times 3, branched at the three I_{0}^{*} fibers and an additional smooth fiber X_{t} . Note that a generic elliptic K3 surface with these singular fibers has \rho=17 , and varies with 3 moduli. The moving branch point provides an additional parameter, so the moduli for the corresponding elliptic-elliptic surfaces with six I_{2} fibers has dimension 4, and it dominates \mathcal{I}_{1} . Example 3.19 . [ EGW24 ] Engel–Greer–Ward found a 6-dimensional family \mathcal{I}_{2} of isotrivial elliptic-elliptic surfaces with six cuspidal fibers. The period image of \mathcal{I}_{2} is 5-dimensional. A subfamily of codimension two can also be constructed from quadratic base change of an elliptic K3 surface X\to\mathbb{P}^{1} with singular fibers of type II\times 3,I_{0}^{*}\times 3. Again, such elliptic K3 surfaces vary with 3 moduli, and the moving branch point provides an additional parameter. The family of elliptic-elliptic surfaces that arise from base change dominates a codimension two locus inside \mathcal{I}_{2} . 4. Surface degenerations Our construction begins with a specific instance of Corollary 3.6 . We start from the Kummer surface X=\textup{Kum}(E_{\omega}\times E_{\omega}) associated to the self-product of the elliptic curve E_{\omega}=\mathbb{C}/(\mathbb{Z}\oplus\omega\mathbb{Z}) , where \omega is a third root of unity. The K3 surface X has Picard number 20 , and admits an elliptic fibration [ SI77 ] (4) X\to\mathbb{P}^{1} with three singular fibers of types II^{*},IV^{*},I_{0}^{*} . By choosing a suitable coordinate on \mathbb{P}^{1} , we may assume that the three singular fibers lie over 0,1,2 respectively. The transcendental lattice of X is given by T_{X}\cong\mathrm{A}_{2}(2)=\begin{bmatrix}4&2\\ 2&4\end{bmatrix}. Let \lambda\mathrel{\mathop{\mathchar 58\relax}}\mathbb{P}^{1}\to\mathbb{P}^{1} be a nonconstant morphism, and let E_{t}\to\mathbb{P}^{1} be the elliptic curve branched at 0,1,2 and an additional point \lambda(t) . Let S_{t}\to E_{t} be the relatively minimal model of X\times_{\mathbb{P}^{1}}E_{t} . Then by Lemma 3.14 , we have: Lemma 4.1 . There exists a smooth family of elliptic-elliptic surfaces \{S_{t}\to E_{t}\}_{t\in U} parameterized by an open subset U\subset\mathbb{P}^{1} , with (i) constant VHS on the transcendental Hodge structure T_{S_{t}} ; (ii) non-constant VHS on H^{1} ; (iii) maximal Picard number \rho=h^{1,1}=12 ; (iv) the singular fibers of S_{t}\to E_{t} are of type IV^{*} , IV ; (v) the Mordell–Weil rank of S_{t}/E_{t} is 2. Proof. (i) and (ii) follow from Proposition 3.15 . (iii) follows from Lemma 3.12 ; since T_{S_{t}} has rank 2, \rho=h^{1,1}=12 . (iv) follows from Table 1 showing that branched double covers of II^{*},IV^{*},I_{0}^{*} become IV^{*},IV,I_{0} , respectively. (v) follows from the Shioda–Tate formula [ SS19 , Corollary 6.7] ; \rho=2+r+6+2 , so the Mordell–Weil rank r=2 . ∎ 4.1. A degeneration of elliptic-elliptic surfaces In this subsection, we compactify the smooth family from Lemma 4.1 . Our goal is to show the following Proposition 4.2 . There exists a smooth projective 3-fold \mathcal{S} and a proper flat family (5) g\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}\to\mathbb{P}^{1} which satisfies the following conditions: • When t\notin\{t_{0},\ldots,t_{5}\} , S_{t}\mathrel{\mathop{\mathchar 58\relax}}=g^{-1}(t) recovers the family from Lemma 4.1 . • When t=t_{i} , the fiber g^{-1}(t) is simple normal crossings (SNC), with a component that is reduced and birational to a K3 surface. We begin by compactifying the base family of elliptic curves to an elliptic surface \mathcal{E} . There are many ways of doing this, using the general setup of Lemma 3.14 , but for the sake of concreteness we give the simplest example here. Our elliptic surface \mathcal{E} will be the minimal resolution of a certain double cover of \mathbb{P}^{1}\times\mathbb{P}^{1} . In order to distinguish the two factors, we will denote the first by \mathbb{P}^{1}_{t} and the second by \mathbb{P}^{1}_{x} . Lemma 4.3 . There is a rational elliptic surface f\mathrel{\mathop{\mathchar 58\relax}}\mathcal{E}\to\mathbb{P}^{1}_{t} with six I_{2} fibers that admits a map h\mathrel{\mathop{\mathchar 58\relax}}\mathcal{E}\to\mathbb{P}_{x}^{1} , such that the restriction to the general fiber of f is a double cover h|_{E_{t}}\mathrel{\mathop{\mathchar 58\relax}}E_{t}\to\mathbb{P}^{1}_{x} branched at four points, including three fixed points x=0,1,2 . Proof. The elliptic surface \mathcal{E} is the minimal resolution of the double cover of \mathbb{P}_{t}^{1}\times\mathbb{P}^{1}_{x} branched along a divisor D of bidegree (4,2) , pictured below. The divisor D is the sum of 3 horizontal lines with the graph of a general double cover \lambda\mathrel{\mathop{\mathchar 58\relax}}\mathbb{P}^{1}_{t}\to\mathbb{P}^{1}_{x} . The first projection f\mathrel{\mathop{\mathchar 58\relax}}\mathcal{E}\to\mathbb{P}^{1}_{t} defines an elliptic fibration surface with fiber E_{t}=f^{-1}(t) . The second projection is a conic bundle. When t is general, under the projection \mathcal{E}\to\mathbb{P}^{1}_{x} , E_{t} can be realized as a double cover of \mathbb{P}^{1}_{x} branched at \{0,1,2,\lambda(t)\} . The Kummer K3 surface X has a special elliptic fibration X\to\mathbb{P}^{1}_{x} with three star fibers situated over \{0,1,2\}\in\mathbb{P}^{1}_{x} , as indicated in the figure. Figure 1 . Rational elliptic surface with its two projections By choosing \lambda general, the six singular points of D are nodes. Accordingly, the branched double cover will have an A_{1} -singularity over each singular point of D , so its desingularization \mathcal{E} gives an elliptic surface \mathcal{E}\to\mathbb{P}^{1}_{t} with six I_{2} fibers. Each I_{2} fiber consists of a main component, the proper transform of the nodal double cover of \mathbb{P}^{1}_{x} , together with an exceptional component from the blow up of the A_{1} -singularity. On the main component, h gives a double cover \mathbb{P}^{1}\to\mathbb{P}^{1}_{x} branched at the two values in \{0,1,2\} not equal to \lambda(t_{i}) . ∎ These singular fibers of f\mathrel{\mathop{\mathchar 58\relax}}\mathcal{E}\to\mathbb{P}^{1}_{t} come in pairs according to the value of \lambda(t)\in\mathbb{P}^{1}_{x} . We label these values of t as follows: \displaystyle\lambda^{-1}(0) \displaystyle=\{t_{0},t_{3}\}, \displaystyle\lambda^{-1}(1) \displaystyle=\{t_{1},t_{4}\}, \displaystyle\lambda^{-1}(2) \displaystyle=\{t_{2},t_{5}\}. The fibered product \mathcal{E}\times_{\mathbb{P}^{1}_{x}}X has normalization \mathcal{S}^{\prime} admitting a proper morphism, \mathcal{S}^{\prime}\to\mathbb{P}^{1}_{t}, This is a family of elliptic-elliptic surfaces which is smooth over U=\mathbb{P}^{1}\smallsetminus\{t_{0},t_{1},t_{2},t_{3},t_{4},t_{5}\} . We showed in Lemma 3.14 that \mathcal{S}^{\prime}_{U} admits a contraction \mathcal{S}^{\prime}_{U}\to\mathcal{S}_{U} which restricts to the minimal model contraction on each smooth surface fiber. Let \overline{\mathcal{S}}\to\mathbb{P}^{1} denote any flat projective family of surfaces extending \mathcal{S}_{U} , and let \mathcal{S}\to\overline{\mathcal{S}} be a resolution of singularities of the total space such that the fibers have simple normal crossings. We can also arrange that the fibers over each pair t_{0},t_{3} etc. are isomorphic. Lemma 4.4 . Each singular fiber of \mathcal{S}^{\prime}\to\mathbb{P}^{1} and of g\mathrel{\mathop{\mathchar 58\relax}}\mathcal{S}\to\mathbb{P}^{1} contains a unique reduced component birational to a K3 surface. Proof. The singular fibers of \mathcal{S}^{\prime}\to\mathbb{P}^{1} over \{t_{i}\}_{i=0,\dots,5} are readily described. First, observe that the fibers of \mathcal{E}\to\mathbb{P}^{1} are I_{2} configurations, featuring two reduced components isomorphic to \mathbb{P}^{1} meeting at two nodes. The fiber of \mathcal{E}\times_{\mathbb{P}^{1}_{x}}X\to\mathbb{P}^{1}_{t} over t_{i} is the union of Z_{1,i} and Z_{2,i} . The (main) component Z_{1,i} is reduced and birational to the base change of X\to\mathbb{P}^{1}_{x} branched at the two (other) star fibers. On the other hand, Z_{2,i} is described as follows: Z_{2,0}\cong Z_{2,3}\cong II^{*}\times\mathbb{P}^{1},\ Z_{2,1}\cong Z_{2,4}\cong IV^{*}\times\mathbb{P}^{1},\ Z_{2,2}\cong Z_{2,5}\cong I_{0}^{*}\times\mathbb{P}^{1}. Each Z_{2,i} contains nonreduced components due to the nonreduced curves in the star fibers. After normalization, the main component Z_{1,i}^{\nu} is a K3 surface blown up at 8 points, which proves the first statement. For the second statement, recall from Lemma 3.14 that we have a birational map \mathcal{S}^{\prime}\dasharrow\mathcal{S} with indeterminacy locus supported in the central fiber. Resolving this indeterminacy to produce a regular map, \textstyle{\widehat{\mathcal{S}^{\prime}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces} \scriptstyle{\eta} \textstyle{\mathcal{S}} \textstyle{\mathcal{S}^{\prime}\ignorespaces\ignorespaces\ignorespaces\ignorespaces} the blow up \widehat{\mathcal{S}^{\prime}} still contains a component birational to a K3 in each singular fiber. These components are non-uniruled divisors, and \mathcal{S} is smooth by assumption, so no such component is contracted by \eta , using [ KM98 , Prop. 1.3] . ∎ 4.2. The K3 components In this subsection, we will describe the geometry of the three K3 surfaces whose birational models occur in Proposition 4.2 . Notation 4.5 . Let Y_{0} (respectively, Y_{1} , and Y_{2} ) be the K3 surfaces from Proposition 4.2 . They are the minimal models of the main components of the fiber over t_{0},t_{1},t_{2} of \mathcal{S}\to\mathbb{P}^{1} . (Note the singular fibers over t_{3},t_{4},t_{5} are identical to t_{0},t_{1},t_{2} , respectively.) Each Y_{i} is obtained by taking the relatively minimal regular model of the normalized base change \mathbb{P}^{1}\times_{\mathbb{P}^{1}_{x}}X , where \mathbb{P}^{1}\to\mathbb{P}^{1}_{x} is the double cover branched at x=1,2 (respectively, x=0,2 , and x=0,1 ). 4.2.1. Analysis of Y_{0} : By construction, the elliptic K3 Y_{0}\to\mathbb{P}^{1} arises from a quadratic base change of ( 4 ) branched at IV^{*} and I_{0}^{*} . Hence by Table 2, it has singular fibers of types II^{*},II^{*},IV . Lemma 2.9 implies that T_{Y_{0}} has intersection pairing T_{X}=A_{2} , which has discriminant 3 . Theorem 2.8 guarantees that Y_{0} is the K3 surface arising from the minimal desingularization of (E_{\omega}\times E_{\omega})/\mu_{3} , acting anti-diagonally. This is one of Vinberg’s “most algebraic” K3 surfaces [ Vin83 ] . 4.2.2. Analysis of Y_{1} : The elliptic K3 Y_{1}\to\mathbb{P}^{1} arises from a quadratic base change of ( 4 ) branched at II^{*},I_{0}^{*} fibers, hence it has three singular fibers of type IV^{*} . The Shioda–Tate formula implies the elliptic surface is extremal. By the classification theorem [ SZ01 , Table 2, No. 219] , the Mordell–Weil group is MW(Y_{1}/\mathbb{P}^{1})\cong\mathbb{Z}/3\mathbb{Z} . Moreover, the transcendental lattice of Y_{1} is isometric to T_{Y_{0}} . In particular Proposition 4.6 . There is an isomorphism Y_{0}\cong Y_{1} , and \mathrm{disc}(Y_{0})=\mathrm{disc}(Y_{1})=3 . 4.2.3. Analysis of Y_{2} : The elliptic K3 (6) Y_{2}\to\mathbb{P}^{1} arises from base change of ( 4 ) branched at II^{*} and IV^{*} fibers. Hence it has singular fibers I_{0}^{*},I_{0}^{*},IV,IV^{*} . Our claim is that Proposition 4.7 . Y_{2}\ncong Y_{1}\cong Y_{0} . Proof. This follows from the fact that Vinberg’s most algebraic K3 surface Y_{0} does not have an elliptic fibration structure with singular fiber types I_{0}^{*},I_{0}^{*},IV,IV^{*} (cf. [ Vin83 , Nis96 , BZ25 ] ). ∎ We have a more precise description of the K3 surface Y_{2} : Proposition 4.8 . The transcendental lattice T_{Y_{2}} has discriminant \mathrm{disc}(Y_{2})=48. Moreover, the elliptic fibration ( 6 ) has Mordell–Weil group \mathbb{Z}^{2} and the Mordell–Weil lattice has discriminant \mathrm{disc}(\textup{MWL})=\frac{1}{3}. Proof. Let \pi\mathrel{\mathop{\mathchar 58\relax}}Y_{2}\dashrightarrow X be the rational double cover, then • \pi_{*}\pi^{*}=[2]|_{T_{X}} , • \pi^{*}\pi_{*}=[2]|_{T_{Y_{2}}} , where [2] denotes multiplication by 2 . Hence both \pi_{*} and \pi^{*} are injective. Since the image of [2]|_{T_{X}} has index 2^{rk(T_{X})}=2^{2} , \pi_{*}T_{Y_{2}} has index 2^{\alpha} in T_{X} , with the possible values \alpha=0,1,2 . Recall that T_{X}=\mathrm{A}_{2}(2) has discriminant 3\cdot 2^{2} , (2^{\alpha})^{2}=\frac{\mathrm{disc}(\pi_{*}T_{Y_{2}})}{\mathrm{disc}(T_{X})}=\frac{\mathrm{disc}(\pi_{*}T_{Y_{2}})}{3\cdot 2^{2}}. We have \mathrm{disc}(\pi_{*}T_{Y_{2}})=3\cdot 2^{2+2\alpha} . On the other hand, the pushforward \pi_{*} scales the intersection pairing on transcendental lattices by 2 . Hence, \mathrm{disc}(\pi_{*}T_{Y_{2}})=2^{2}\cdot\mathrm{disc}(T_{Y_{2}}). Combining these identities, we obtain \mathrm{disc}(T_{Y_{2}})=3\cdot 2^{2\alpha},\ \alpha=0,1,2. However, if \alpha=0 , then \mathrm{disc}(T_{Y_{2}})=3 , which contradicts Proposition 4.7 ; If \alpha=1 , then \mathrm{disc}(T_{Y_{2}})=12 , and there are two possibilities up to isometry: (1)\ T_{Y_{2}}=\begin{bmatrix}4&2\\ 2&4\end{bmatrix},\ \textup{or}\ (2)\ T_{Y_{2}}=\begin{bmatrix}2&0\\ 0&6\end{bmatrix}. Note that the first one is the Kummer K3 surface X=\textup{Kum}(E_{\omega}\times E_{\omega}) that we started with. However, there is a complete classification of elliptic fibrations for K3 surfaces of types (1) and (2) (see [ Nis96 ] and [ BGH+15 ] ), but none of them have the fibrations with singular fibers I_{0}^{*}\times 2 , IV^{*} , IV . Therefore \alpha\neq 1 , and the only possibility is \alpha=2 , hence \mathrm{disc}(T_{Y_{2}})=48 . We already know Y_{2} has maximal Picard number \rho=20 . By Shioda–Tate formula, Y_{2}\to\mathbb{P}^{1} has Mordell–Weil rank 2 . The configurations of the singular fibers implies that MW_{tor}=0 [ SS19 , Proposition 6.33 (iv)] . Therefore, the Mordell–Weil group is \mathbb{Z}^{2} . The discriminant formula [ SS19 , Corollary 6.39] implies \mathrm{disc}(NS(Y_{2}))=\mathrm{disc}(T_{Y_{2}})=3\times 3\times 4^{2}\times\mathrm{disc}(\textup{MWL}). Hence \mathrm{disc}(\textup{MWL})=\frac{1}{3} . ∎ Remark 4.9 (Mordell–Weil lattice) . A second proof that Y_{2} is not Vinberg’s most algebraic K3 surface (Proposition 4.7 ) independent from the classification of elliptic fibration structures goes as follows. Assuming the contrary, \mathrm{disc}(T_{Y_{2}})=3 . The discriminant formula for Mordell-Weil lattices implies that \det(\textup{MWL})=\frac{1}{48} . However, from the formula of the height pairing [ SS19 , Theorem 6.24] and knowledge of the singular fibers (I_{0}^{*},I_{0}^{*},IV,IV^{*}) , the lcm of the denominator of each entry in the Mordell–Weil lattice is 6 , which implies that the denominator of \mathrm{disc}(\textup{MWL}) divides 6\times 6=36 . This is a contradiction. 5. Failure of the local invariant cycle theorem over \mathbb{Z} The goal of this section is to prove Theorem 1.3 . It will follow from the following theorem. Theorem 5.1 . For any flat projective completion of the family \mathcal{S}_{U} with smooth total space, the LICT Z fails near t_{2} . It fails also for any semistable reduction near t_{2} . 5.1. The transcendental lattice of a singular surface. We now define the transcendental lattice of a reducible surface, extending the discussion in Section 2.1 . Given a singular (possibly reducible) projective surface Z , recall that there is a mixed Hodge structure on H^{2}(Z,\mathbb{Q}) with an increasing weight filtration W_{k}H^{2}(Z,\mathbb{Q}) and decreasing Hodge filtration F^{p}H^{2}(Z,\mathbb{Q}) (cf. [ Del71 ] ). In particular, Gr_{2}^{W}H^{2}(Z,\mathbb{Q}) is a pure Hodge structure of weight two. Working integrally, we have Gr^{W}_{2}H^{2}(Z,\mathbb{Z})\mathrel{\mathop{\mathchar 58\relax}}=\frac{H^{2}(Z,\mathbb{Z})}{W_{1}H^{2}(Z,\mathbb{Q})\cap H^{2}(Z,\mathbb{Z})}. The following definition generalizes Lemma 2.1 . Definition 5.2 . Define T_{Z,\mathbb{Q}}\subset Gr_{2}^{W}H^{2}(Z,\mathbb{Q}) to be the smallest \mathbb{Q} sub-Hodge structure whose complexification contains the (2,0) part on the right. Define T_{Z}\mathrel{\mathop{\mathchar 58\relax}}=T_{Z,\mathbb{Q}}\cap Gr^{W}_{2}H^{2}(Z,\mathbb{Z})/\textup{torsion}. Lemma 5.3 . Suppose that Z=\bigcup_{i=1}^{N}Z_{i} is an SNC surface. Then there is an isomorphism of \mathbb{Z} -Hodge structures (7) T_{Z}\cong\oplus_{i}T_{Z_{i}}. Proof. We can compute the cohomology of Z by a Mayer–Vietoris spectral sequence. With rational coefficients, it degenerates at the second page E_{2} , and computes the weight graded subquotients of the mixed Hodge structure. In particular, we obtain an exact sequence of Hodge structures (8) 0\to Gr^{W}_{2}H^{2}(Z,\mathbb{Z})/\text{torsion}\to\oplus_{i}H^{2}(Z_{i},\mathbb{Z})/\text{torsion}\xrightarrow{\rho}\oplus_{i