Professional Documents
Culture Documents
1. Introduction
One of the main results of M. Saito’s theory of Hodge modules is a Hodge-
theoretic proof of the decomposition theorem of Beilinson-Bernstein-Deligne-Gabber
[2], at least for those semisimple perverse sheaves underlying polarizable pure Hodge
modules. However, while parts of the theory works just as well on arbitrary complex
manifolds as on projective varieties, Saito’s proof of the decomposition theorem
[25] relies heavily on the assumption that the morphism in question be projective.
In particular, the argument uses dimensional induction to prove a relative hard
Lefschetz theorem by using a version of the Lefschetz hyperplane theorem for
holonomic D-modules.
Since the usual hard Lefschetz theorem for singular cohomology still holds in
the compact Kähler setting, it’s natural to ask whether the decomposition theorem
holds along proper relative Kähler morphisms, suitably defined. We use geometric
methods, inspired by Saito’s original argument, to prove this for polarizable complex
Hodge modules which admit an integral structure, in particular for Hodge modules
of geometric origin (see section 2.5).
Theorem 1.1. Let f : X → Y be a proper morphism of complex manifolds, and
suppose l ∈ H 2 (X, R(1)) is a relative Kähler class. Then for any M ∈ HM(Z) (X, n)p
a polarizable pure C-Hodge module of weight n which admits an integral structure,
we have the following:
(1) T f∗ M is strict.
(k)
(2) T f∗ M ∈ HM(Z) (Y, n + k)p
(−k) (k)
(3) The relative hard Lefschetz theorem holds, i.e. lk : T f∗ M → T f∗ M (k)
is an isomorphism for every k ≥ 0.
(−k)
(4) If M is polarized by the sesquilinear pairing S : M ⊗M → CX , then T f∗ M
is polarized by
(k,−k)
(α, β) 7→ T f∗ S(lk α, β).
2020 Mathematics Subject Classification. 14D07 (Primary) 32D20, 32C38, 14F10, 14F43
(Secondary).
1
2 MADS BACH VILLADSEN
(k)
Here T f∗ denotes the kth derived direct image functor of triples of D-modules
e
2
(see section 2.4) and aX : X → {pt} the constant map. A class l ∈ H (X, R(1)) is
said to be relatively Kähler along f if, locally over Y , there exists a closed 2-form el
∗
representing l and a Kähler form η on the base such that l + f η is a Kähler form
e
(see also section 2.2). An integral structure on a variation of Hodge structures is
just a local system with integer coefficients that recovers the complex local system
of the VHS; in general, an integral structure is a constructible complex with integer
coefficients.
Note that if KR is a real perverse sheaf which provides a real structure for M , then
the hard Lefschetz isomorphism on M is induced by a corresponding isomorphism
lk : p R−k f∗ KR → p Rk f∗ KR .
If f : X → Y is a projective morphism, our methods in fact apply to arbitrary
polarizable Hodge modules, not just those which admit an integral structure (see
Remark 3.2). This strictly generalizes [25, Théorème 5.3.1], since the Kähler cone
may be bigger than the ample cone on projective manifolds.
Corollary 1.2. With notation and assumptions as in Theorem 1.1, suppose F is a
subfield of C and that KF is a perverse sheaf over F underlying M , i.e. that there is
given an isomorphism between KF ⊗F C and the de Rham complex of the D-module
underlying M . Then there is a decomposition Rf∗ KF ≃ k p H k Rf∗ KF [−k] in
L
the derived category.
The decomposition over F is obtained by taking a decomposition over either R
or C depending on whether F is a real field, and then deforming this decomposition
using density of F in the larger field. The perverse sheaf KF is not required to be
compatible with the Hodge filtration on M in any way, hence the decomposition
of KF will not be compatible with the Hodge module structure. Even if M is
an F -Hodge module, however, the decomposition of KF will still not generally
be a decomposition of F -Hodge modules, since the deformation procedure is not
compatible with the Hodge filtrations. It is possible that there is some other
decomposition of KF which arises from a decomposition of F -Hodge modules, but
there seems to be no a priori reason why such a decomposition should exist unless
the map f is projective.
Kollár in [19] stated a series of conjectures about the existence of certain coherent
sheaves associated to generically defined VHS, satisfying certain properties enjoyed
by the canonical bundle and its higher direct images. In the projective case these
conjectures were proven by Saito [26]. As a consequence of Theorem 1.1, we can prove
parts of those conjectures for Kähler morphisms and integral VHS, see Theorem 4.1.
1.1. Outline of the proof. The central idea in this paper is to use period maps to
replace the Lefschetz pencils in Saito’s proof in the projective case. If one tries to
imitate Saito’s argument in the Kähler setting, at some point one wants to factor the
constant map aX : X → {pt} through some intermediate map f : X → Y to allow
for induction on the fibre dimension of the morphism. In the projective setting there
are plenty of rational maps available for this purpose, but a general compact Kähler
manifold will not admit any, even meromorphic, maps to lower dimensional varieties.
However, in Saito’s argument, X comes equipped with a VHS defined on a Zariski
open subset U of X. Since we assume an integral structure on our Hodge modules,
this VHS will also carry an integral structure, giving rise to a period map defined
on U . The case where the period map is trivial can be handled directly, and if the
HODGE MODULES AND KÄHLER MORPHISMS 3
1.2. Related work. The results of this paper are special cases of Mochizuki’s paper
[20] which proves a more general version of the decomposition theorem for Kähler
morphisms for tame twistor D-modules [20], hence in particular covering the case
of complex Hodge modules. The proof given here is significantly simpler, however,
and should be easier to follow for anyone familiar with Saito’s original argument in
the projective case. Further, this paper applies to all Hodge modules of geometric
origin, hence to a large fraction of the Hodge modules encountered in practice.
It is a theorem of Cattani-Kaplan-Schmid and Kashiwara-Kawai [4, 18] that
the intersection cohomology of a variation of Hodge structures defined on the
complement of a normal crossings divisor can be computed using L2 -cohomology
with respect to an appropriately chosen metric. Since one can prove the Kähler
identities for L2 -forms, this implies that the intersection cohomology satisfies hard
Lefschetz, and endows it with a polarizable Hodge structure. Mochizuki’s main new
result is that this L2 -Hodge filtration coincides with the proposed Hodge filtration
defined in terms of filtered D-modules by Saito (rather, Mochizuki proves a twistor
version of this result, but we will stick to Hodge modules in this paper. Note also
that the VHS version of this theorem was the never-published result announced in
[17]).
We note that the reduction lemma of [20, Section 6.3] that we use in the present
paper does not depend on any of the new analytic results that Mochizuki proves.
Separate from the results on L2 -cohomology, Saito [24] claimed to prove parts
of Theorem 1.1 (though for arbitrary real Hodge modules) conditional on results
announced by Kashiwara and Kawai [17]; however, these announced results were
4 MADS BACH VILLADSEN
never published (see also [23, Remarks, end of Section 2.5]). Saito later [27] gave a
corrected proof for the decomposition theorem for the constant sheaf.
Acknowledgements. I thank Christian Schnell for bringing my attention to this
problem and many useful discussions; Claude Sabbah for answering my questions
regarding the MHM project and polarizations, and detailed feedback on a draft of
this paper; Johan Commelin for his interest in the issue of signs in polarizations of
Hodge modules in the sense of Saito; and Thomas Krämer for reading and giving
feedback on a draft.
2. Preliminaries
2.1. Hodge structures, filtrations and polarizations. In this section, we spell
out some terminology and conventions regarding complex Hodge structures and
polarizations, following [MHM]. We will use Deligne’s sign convention [9] (used also
by both the MHM project and Saito [25]); note that this differs from the papers of
Cattani-Kaplan-Schmid [4, 5] by a factor (−1)n for pure weight n Hodge structures
[MHM, Remark 2.5.15]. This difference will be relevant in proving Lemma 2.8.
Throughout this section, let H be a finite-dimensional complex vector space.
A pure C-Hodge structure of weight w on H is the data H = L (H, F ′• H, F ′′• H),
where F and F are w-opposite decreasing filtrations, i.e. H ∼
′ ′′
= p Hp,w−p where
Hp,w−p = F ′p H ∩ F ′′w−p H. If R ⊆ R is a subring, then an R-Hodge structure on a
free R-module H is a C-Hodge structure on HC where additionally F ′′w−p = F ′p .
A mixed Hodge structure is the data H = (H, F ′• H, F ′′• H, W• H), where W• is
now an increasing exhaustive filtration such that the filtrations induced by F ′ and
F ′′ on grWj H define a pure Hodge structure of weight j for each j ∈ Z. For a ring
R ⊆ R, a mixed R-Hodge module is a free R-module H consists of a mixed Hodge
structure on HC such that the weight filtration W• is defined over R, and such that
the Hodge structures on the associated graded pieces of W• are R-Hodge structure.
Let Z(w) be the unique weight −2w integral Hodge structure on (2πi)w Z (by
which we mean an R-Hodge structure on (2πi)w Z ⊗Z R); let R(w) and C(w) be
the associated real and complex Hodge structures. For any trifiltered vector space
H = (H, F ′• , F ′′• , W• ), let H(w) = H ⊗Z Z(w) with the induced filtrations. Here
we consider any pure Hodge structure K of weight w, and in particular Z(w), as a
mixed Hodge structure with grW w K = K.
Given a Hodge structure, let CD : H → H be the linear automorphism given by
multiplication by (−1)q on Hp,q ; this is the Deligne-Weil operator associated to the
Hodge structure H.
Definition 2.1. Let H be a pure Hodge structure of weight w as above. A
polarization of H is a morphism of Hodge structures S : H ⊗ H → C(−w) such that
(1) S is hermitian, i.e. S(x, y) = S(y, x) for all x, y ∈ H, and
(2) the pairing h(x, y) = S(CD x, y) = S(x, CD y) on H is hermitian positive
definite.
While one can define rational and real polarizations [MHM, Section 2.5.c], we
will not consider these in this paper.
Given an sl2 -representation on H, by the associated sl2 -triple we will mean the
images (X, H, Y) of the standard generators of sl2 , where X increases weights, Y
lowers weights, and H is semisimple.
HODGE MODULES AND KÄHLER MORPHISMS 5
with respect to which N is self-adjoint, and such that the induced pairing on grW (N) H
polarizes the induced sl2 -Hodge structure.
Definition 2.5. Let (H, F ′ , F ′′ , W ) be a mixed Hodge structure of central weight
w.
Given commuting nilpotent operators N1 , . . . , Nk : H → H(−1), let
X
(8) C = C(N1 , . . . , Nk ) = { λl Nl | λl > 0}
l
Recall from [4, (2.11)] that (H, F ′ , F ′′ , W ) is a mixed Hodge structure polarized
by (S, C(N1 , . . . , Nk )) if and only if W = W (N) for every N ∈ C(N1 , . . . , Nk ) and
the map
(9) θ : Ck → D b
X
(z1 , . . . , zk ) 7→ exp zi Ni · F ′
b is the flag manifold of filtrations on H of type
is a nilpotent orbit (see [5]) where D
′ ′p ′p−1
F ; i.e. Ni F ⊂ F and there exists some α ∈ R such that θ is the period map
associated to a pure polarized variation of Hodge structures on
(10) {(z1 , . . . , zk ) ∈ Ck | Re zi < α}
Note that as we are now dealing with R-Hodge structures, we only need to
describe the single filtration F ′ .
Note that [4, 5] use nilpotent real endomorphisms, corresponding to 2πiNj for Nj
as above, hence the condition Re zi < α here rather than the condition Im zi > α.
Proof. Given (H, F ′ , F ′′ , W ; S, C) as in the lemma, let θ be the associated nilpotent
orbit as above.
Similarly, we can define a map
X
(11) θe: Ck−1 → D b : (z2 , . . . , zk ) 7→ exp f′
fi · F
zi N
e
proves the SL(2)-orbit theorem starting from the existence of the limiting mixed
Hodge structures, and [4, (2.11)] for C-Hodge structures follows (see [5, (4.66)]).
2.2. Kähler forms and Chern classes. As in [9, 25], we will define Chern classes
of line bundles on X using the exponential exact sequence
× exp
(12) 0 → Z(1) → OX −−→ OX → 1,
×
letting c1 : H 1 (X, OX ) → H 2 (X, Z(1)) be the connecting homomorphism; hence
Chern classes of line bundles will take values in Z(1) = 2πiZ.
Similarly, by a Kähler class we will mean a cohomology class ω ∈ H 2 (X, R(1))
such that the real cohomology class (2πi)−1 ω can be represented by a positive closed
real (1, 1)-form on X.
D M∗
f = Rf∗ (Mf⊗L D
eX→Y ) = Rf∗ (M
X→Y (DX ));
(15) f⊗ e Sp
DX
e
D
e X
(j)
we denote the cohomology modules by D f∗ M f.
Let DX,X = DX ⊗C DX , and let CX be the sheaf of currents of degree 0 on X, as
a right DX,X -module.
HODGE MODULES AND KÄHLER MORPHISMS 9
(16) s : M ⊗C M ′′ → CX .
We will see later that it is more natural to add a sign to this definition.
(k,−k) (k,−k)
Definition 2.11. The signed pushforward of s is T f∗ s = ϵ(k)D,D f∗ s,
where as usual ϵ(k) = (−1)k(k−1)/2 .
From now on, we specifically work in the filtered setting and let D
eX = RF DX .
Hodge modules will be certain D-triples.
e
Definition 2.12. A triple on X is the data T f = (Mf′ , M f′′ , s), where Mf′ , Mf′′
are DX -modules and s is a sesuilinear pairing between M , M , where in general
e ′ ′′
Ee(0,j) is the sheaf of C ∞ differential forms of type (0, j). The differentials are given
by
′∞
δ
e X,i ⊗ e Ee(0,j) −
e
e X,i−1 ⊗ e Ee(0,j)
Θ OX −→ Θ OX
e i ) ⊗ ϕ + ξ⌟de′ ϕ
ξ ⊗ ϕ 7→ δ(ξ
and de′′ , where de′ , de′′ are the usual holomorphic and antiholomorphic exterior deriva-
tives on differential forms.
The relative C ∞ Spencer resolution of M fcan now be written as
∞,•
(21) Sp∞
X→Y (M ) = M ⊗O
f f e Sp
X
fX ⊗ e D
OX
eX→Y ,
∞,•,•
where we take the simple complex associated to the double complex Spf
X defined
above.
e • = f∗ Sp∞
Thus we will let K X→Y (M ), and define T f∗ T as the complex of triples
f f
with terms
Z
′k e ′′−k
(22) K ,K
e , ϵ(k) f∗ s ,
f
∞
with the pairing s implicitly extended to C -coefficients.
The following is [MHM, Corollary 12.7.37]
f g
Lemma 2.13. Consider proper maps X − →Y − → Z and a triple T f on X. Then the
complexes T (g ◦ f )∗ T and T g∗T f∗ T (more precisely, the simple complex associated
f f
to the double complex that naturally computes the latter) are quasi-isomorphic, and
in particular the functors T f∗ admit Leray spectral sequences in the category of
triples.
We refer to [MHM] for the details of the proof, but since this commutation,
which essentially boils down to a sign comparison, plays a crucial role in this paper
(specifically in Lemma 3.3), let us explain the role of ϵ(k) in this computation.
Write h = g ◦ f . To compute the pushforward T h∗ T f, one considers, for Mf= M f′
or M , the pushforwards
f′′
(23) e • = h∗ M
K f⊗ e ⊗ Sp∞ (D eX ) ,
DX X→Z
together with the pairing between their underlying D-modules defined from s, taking
values in
(24) h∗ CX ⊗De Sp∞X,X→Y,Y
( D
e
X,X ) .
X,X
To
R get the final pushforward pairing, compose this with the integration morphism
h
.
To compute the double pushforward T g∗T f∗ T
f, however, one must consider the
double complexes
h i
(25) Ke •,• = g∗ f∗ Mf⊗ e Sp∞ (D eX ) ⊗ e Sp∞ ,
DX X→Y DY Y →Z
for Mf= Mf′ , Mf′′ , and the pairing between them induced by s taking values in
h i
(26) g∗ f∗ CX ⊗De Sp∞X,X→Y,Y
(De
X,X ) ⊗De Sp∞
Y,Y →Z,Z
.
X,X Y,Y
HODGE MODULES AND KÄHLER MORPHISMS 11
with s and h∗ . For a single triple Tf, there is no sign in this tensor swap morphism.
But in the computation of T g∗T f∗ T f, the pushforward T g∗ is applied to the complex
of triples T f∗ T , and the tensor swap operation thus introduces a sign to the effect
f
of (−1)ij on the pairing sij : K ′i,j ⊗ K ′′−i,−j → CZ . The introduction of the sign
ϵ, which satisfies ϵ(i + j) = ϵ(i)ϵ(j)(−1)ij , ensures consistency between T h∗ Tf and
T g∗T f∗ T .
f
adjoint pairing defined above. The Tate twist by l ∈ Z is T f(l) = (M f′ (l), M f′′ (−l), s),
where in general M f(l) = z l Mf (recall that M f is a module over D e = RF DX =
p Fp DX ).
L
A polarized Hodge module of weight w is the data of a holonomic object of
D−Triples(X)
e and a morphism S : T f→ T f∗ (−w), the polarization, satisfying
certain inductive conditions in terms of nearby and vanishing cycles (see[MHM,
Definition 14.2.2] for the full definition). In particular, T f is required to be strict,
hence M , M correspond to filtered DX -modules, and S is an isomorphism of
f ′ f′′
triples.
We will use the notation (M, S) for a polarized Hodge module and, by strictness,
consider M to be a triple ((M ′ , F • M ′ ), (M ′′ , F • M ′′ ), S). Note that S defines an
isomorphism between (M ′ , F • M ′ ) and (M ′′ , F • M ′′ )(w). Let (M , F ) = (M ′ , F );
through this isomorphism, (M, S) is isomorphic to a polarized Hodge module of
the form ((M , F ), (M , F )(w), S) with polarization (id, id), where S is defined by
composing s and S appropriately. In this way, we will occasionally refer to S as a
polarization.
For a proper holomorphic map f : X → Y , let T f∗ (M, S) = (T f∗ M, T f∗ S) in
the category of triples. The following now follows directly from the results of the
previous section.
f g
Lemma 2.14. Let X − → Y −→ Z be proper holomorphic maps, and (M, S) a
polarized Hodge module of weight w on X. Then the complexes T (g ◦ f )∗ (M, S) and
T g∗T f∗ (M, S) are quasi-isomorphic.
(g ◦ f )∗ Gn A ⊗ (g ◦ f )∗ Gn B (g ◦ f )∗ Gm+n (A ⊗ B)
M M M
g∗ Gi f∗ Gj A ⊗ g∗ Gk f∗ Gl B g∗ Gi+k f∗ Gj+l (A ⊗ B)
i+j=n k+l=m i+j=m
k+l=n
M
g∗ Gi+k (f∗ Gj A ⊗ f∗ Gl B)
i+j=m
k+l=n
Here the two vertical maps are the natural quasi-isomorphisms. The top horizontal
map represents (32), while the composition across the bottom three terms represents
(33).
Now in general, one can check that for complexes K, L of constructible sheaves,
the natural map ϕ : G• K ⊗ G• L → G• (K ⊗ L) looks like ϕ(α ⊗ β) = (−1)an α ⊗ β for
α ∈ Ga K m and β ∈ Gb Ln , where on both sides we consider α and β as discontinuous
sections of the constructible sheaves making up K and L respectively. On a local
section
(34) α ⊗ β ∈ g∗ Gi f∗ Gj A ⊗ g∗ Gk f∗ Gl B,
of the direct sum in the middle left of the diagram, a computation now shows that
this diagram commutes up to a sign of (−1)il .
Comparing the signs in the pairings H g∗ H f∗ S and H (g ◦ f )∗ S on R(g ◦ f )∗ K
f g
for a composition X − →Y − → Z and perverse sheaf K underlying a Hodge module,
one now sees that there are certain signs involved that depend on the constructible
degrees of the objects involved. These signs do not seem to be cancelled out by
anything else, and so it is not the case that H g∗ H f∗ S = H (g ◦ f )∗ S in general,
even on terms of the Leray spectral sequence, and there does not seem to be any
sign correction that one can insert on the level of Hodge modules to fix this, since
such a correction would be in terms of the perverse t-structure, not the constructible
one. Even if these “constructible signs” were cancelled out, one still needs to explain
the discrepancy factor (−1)ij between the sign factors ϵ(i + j) and ϵ(i)ϵ(j) that show
up in the natural polarizations on p Ri+j (g ◦ f )∗ K and p Ri g∗ p Rj f∗ K respectively.
This sign, defined in terms of the perverse t-structure, does not seem to result from
the tensor product maps Rf∗ K ⊗ Rf∗ K → Rf∗ (K ⊗ K) like their analogues in the
MHM project approach to polarizations do.
Hodge modules with rational structures on X of weight n. Let HM(Z) (X, n)p be the
full subcategory of HMQ (X, n)p consisting of those Hodge modules M such that
KQ is of the form KZ ⊗Z Q for some integral constructible complex KZ (note that
KZ is not part of the data!). We say that such an M admits an integral structure.
Note that Z-structures are preserved under the usual operations in the theory of
Hodge modules: The functor (−) ⊗Z Q : Dcb (X, Z) → Dcb (X, Q) commutes with the
six-functor formalism for constructible complexes, and similarly for extending to
complex coefficients. However, direct sum decompositions of rational structures will
not necessarily be compatible with underlying integral structures on the nose, but
only up to finite index.
For nearby and vanishing cycles, suppose f : X → ∆ is a holomorphic function,
submersive over the punctured disk ∆∗ . Let exp : H → ∆ be the composition of
the universal covering of ∆∗ by the upper half plane with the inclusion ∆∗ → ∆,
k: Xe → X the pullback of exp along f , and i : X0 = f −1 (0) → X. Recall that for a
constructible complex C, the nearby cycles along f are defined as
(35) ψf C = i−1 Rk∗ (k −1 C)
and the vanishing cycles ϕf C as the cone of the canonical map i−1 C → ψf C. This
construction goes through whether C has coefficients in Z or Q, and the construction
commutes with (−) ⊗Z Q. One can check that if C is a Q-complex which admits an
integral structure, then the unipotent nearby and vanishing cycles ψf,1 C and ϕf,1 C
also admit integral structures. More precisely, the usual eigenspace decomposition
over Q induces a decomposition over Z up to finite index (needed to clear the
denominators of the rational numbers involved in the maps in the decomposition).
There’s a notion of a perverse t-structure on Dcb (X, Z) which maps to the usual
one on Dcb (X, Q) under tensor product [2, Section 3.3], so we can take perverse
cohomology objects in Dcb (X, Z). In particular, if C ∈ Dcb (X, Q) admits an integral
structure and f : X → Y is a proper morphism, then each p Ri f∗ C admits an integral
structure. Similarly, we can take intermediate extensions of local systems with
integral coefficients, compatible with the usual intermediate extension functor for
rational local systems.
We recall the structure theorem for Hodge modules, see [MHM, Theorem 16.2.1
(Structure theorem)] and the S-decomposability property in [MHM, Theorem 14.2.17
(Main properties of polarizable Hodge modules)]. Note that while the proof of the
structure theorem uses the direct image theorem, it only does so in the specific
context of blowups, which are projective morphisms. Hence we are free to use the
structure theorem here.
Theorem 2.15. Any L polarizable pure Hodge module M admits a direct sum de-
composition M = Z MZ where each MZ is a polarizable pure Hodge module of
the same weight as M which has strict support on an irreducible closed subvariety
Z ⊂ X. Note that the sum is taken over distinct subvarieties.
For each MZ , there exists a Zariski open subset U ⊂ Z such that MZ |U corre-
sponds to a polarizable C-VHS on U , and MZ is the intermediate extension of this
VHS as a pure Hodge module.
Finally, M admits an integral structure if and only if each MZ does, and a given
MZ admits an integral structure if and only if the corresponding VHS MZ |U does.
Proof. The statements about Hodge modules are due to [MHM], as stated above.
The statement comparing integral structures on MZ and MZ |U is [28, Lemma 2.4].
HODGE MODULES AND KÄHLER MORPHISMS 15
If each MZ admits an integral structure, then obviously so does M . For the other
direction, suppose M admits an L integral structure, and fix one. We will first show
that the decomposition M = Z MZ is defined over Q. Arguing inductively over
the supports of M ordered by inclusion, it suffices to show that the decomposition
M = MZ ⊕ N is defined over Q in the case where Z is a maximal support, i.e.
not strictly contained in any other support. Let K, KZ be the complex perverse
sheaves associated to M and MZ . There is a dense open subset U ⊆ Z such
that H − dim Z K|U (the constructible cohomology sheaf in degree − dim Z) and
KZ |U [− dim Z] are both local systems, and by maximality of Z among the supports,
they are equal. It follows that KZ |U and its intermediate extension KZ naturally
have a rational structure induced from that of K. Note that the splitting maps
MZ → M → MZ are defined over Q, since they are entirely determined by their
restrictions to U , so N also has an induced rational structure.
To compare integral structures on M and the summands MZ , first observe that if
ρ : G → GL(A) is a representation of a group G on a finitely generated abelian group
A, and if ρQ : G → GL(AQ ) preserves a rational subspace V ⊂ AQ , then ρ preserves
the subgroup A ∩ V of A (interpreted as the preimage of V under A → AQ ). That
is, any rational subrepresentation of ρQ admits an integral structure.
Now let KQ be the rational structure on M . Given Z among the supports of
M , let KZ,Q be the rational structure on MZ , and take a non-empty Zariski open
subset U ⊂ Z such that KZ,Q |U [− dim Z] and H − dim Z KQ |U are local systems,
where H − dim Z is taken using the standard t-structure on constructible sheaves.
Then KZ,Q |U [− dim Z] is a local subsystem of H − dim Z KQ |U , and the latter admits
an integral structure by assumption. Thus KZ,Q admits an integral structure as
desired. □
2.6. A compactification theorem by Sommese.
Definition 2.16. A proper modification of a complex analytic space X is a proper
morphism f : X ′ → X such that there exists a dense, Zariski open subset U ⊂ X
such that f −1 (U ) is dense and f |f −1 (U ) : f −1 (U ) → U is an isomorphism.
Note that a proper modification is also a bimeromorphism, that is, the inverse
(f |f −1 (U ) )−1 : U → f −1 (U ) is also meromorphic [21, Remark 1.8].
Definition 2.17. A compact complex space X is in Fujiki’s class C if there exists
a surjective holomorphic map f : Y → X with Y a compact Kähler manifold.
Fujiki introduced and studied this class of spaces in the papers [11, 10, 12]. We
recall from Fujiki’s results that the class C is closed under taking subspaces and
meromorphic images, and that any irreducible component of the Douady space
of compact subvarieties of a space in class C is again a space in class C . By [33,
Théorème 3] and resolution of singularities, if X is in class C then there exists a
proper modification f : X ′ → X where X ′ is a compact Kähler manifold. In fact by
[34, Theorem 5] we can even take f to be projective.
Definition 2.18. Let X be a complex analytic space. By a compactification of X,
we mean an open embedding X ⊂ X of X in a compact complex analytic spaceX
such that X contains a non-empty Zariski open subset of X.
The following theorem goes back to Sommese [32, Proposition III, Remark III-C]
and is also stated in a similar form in [1, Theorem 7.7]. However, parts of Sommese’s
16 MADS BACH VILLADSEN
X X′ X′
π π′
Y Y′ Y′
where the vertical maps are surjective proper maps with connected fibres, the left
side horizontal maps are proper modifications, the right side horizontal maps are
compactifications where X ′ and Y ′ are compact Kähler manifolds, and X ′ and X
are bimeromorphic.
Proof. By replacing Y and X with proper modifications, we can arrange that π is
flat by Hironaka’s flattening theorem [16, Corollary 1], while still having connected
fibres. Note that Fujiki’s class C is closed under proper modifications. Abusing
notation, we will thus assume from the start that π is flat.
Now consider the Douady space of X. By functoriality and flatness of π, Y
maps to the Douady space; let C be an irreducible component containing the
image of Y . By [12], C is compact and in class C . Since the fibres of π are
connected, our assumptions on X and Y imply that the general fibre is irreducible,
so by Theorem 2.20 below applied to π, the functorial map j : Y → C is an open
embedding over a non-empty Zariski open subset of Y .
Let d : Z → C be the family of subvarieties of X parametrized by C, and
ϕ : Z → X the projection to X. Let U ⊂ Y be the Zariski open subset over which
the fibres of π are irreducible, and let ZU = ϕ−1 (π −1 (U )).
If for c ∈ C the fibre δ −1 (c) intersects ZU , then ϕ(δ −1 (c)) intersects an irreducible
fibre of π in a Zariski open subset, hence contains that entire fibre. Since the fibre and
ϕ(δ −1 (c)) are flat deformations of each other, they must coincide, so ZU = d−1 (j(U )).
In particular, applying Remmert’s proper mapping theorem to the complement of
ZU in Z, j(U ) is Zariski open in C. This implies that ϕ : Z → X is an isomorphism
when restricted to ZU , hence is a proper modification. This also shows that C is a
compactification of Y .
Finally, Z and C are compact in Fujiki class C . Hence there is a proper modifica-
tion π ′ : X ′ → Y ′ of d : Z → C such that X ′ and Y ′ are compact Kähler manifolds
as desired. In particular, we have also constructed bimeromorphic morphisms
X ′ → Z → X. □
is a polarized Hodge-Lefschetz structure for any 0 < c ≤Lc0 (keeping in mind the
abuse of terminology described in Remark 2.6), hence H ′j is a mixed Hodge
′ −1
structure polarized by S , C(ω, c0 ω + η).
By Lemma 2.8, passing to the associated graded of the weight filtration W (ω)
yields the polarized bi-sl2 -Hodge structure
M W (ω)
(37) Gri H ′j , ω, GrW (ω) η, S ′ ,
i,j
(j)
Further, GrW (ω) η is just the operator induced by the morphisms η : T π∗ M ′ →
(j+2)
T π∗ M ′ (1) by Theorem 1.1 applied to π. Looking at supports, we see that
this morphism vanishes on M , hence H j is in the GrW (ω) η-primitive part of
W (ω)
H ′j .
L
i,j Gri
Now T π∗ S ′ and S agree on the open locus U . Since M has strict support, a
pairing on M is uniquely determined by its restriction to U , so we get the desired
polarization on T aX∗ M . □
3.2. Normal crossings case. Let us first prove the decomposition theorem in the
case where Z is smooth and D is a simple normal crossings divisor. We can then
assume that Z = X for notational simplicity.
In general, T aX∗ M computes the intersection cohomology of the VHS V . Since
D has only simple normal crossings, this can be computed using L2 -cohomology [4,
18], and since the Kähler identities (and hence the whole Kähler package) hold in
that setting, T aX∗ M satisfies hard Lefschetz [18, Theorem 6.4.2].
By Lemma 3.1, it suffices to prove the decomposition theorem for aX ′ + M ′ where
π : X ′ → X is a composition of a finite morphism and a sequence of blowups in
smooth loci, and M ′ is the generic pullback of M . By Lemma 2.21 and since V has
quasi-unipotent local monodromy, we can thus from the start reduce to the case
20 MADS BACH VILLADSEN
(k)
As each summand on the right is a polarizable Hodge structure, so is T aX∗ M .
HODGE MODULES AND KÄHLER MORPHISMS 21
It remains to check the polarization formula for T aX∗ M with respect to a Kähler
class η on X, that is, to check Theorem 1.1(4). Consider first that η induces a
map η : T f∗ M → T f∗ M (1)[2] in the derived category of triples. Given a choice of
decomposition of T f∗ M , η decomposes as a sum of morphisms
(i) (i−l+2)
(41) T f∗ M → T f∗ M (1)[2 − l]
(i) (i−l+2)
in the derived category for various i and l, i.e. classes in Extl (T f∗ M, T f∗ M (1)).
We can collect these classes according to how much they shift the cohomological
degree to get elements
(i) (i−l+2)
M
(42) η−l ∈ Extl (T f∗ M, T f∗ M (1)).
i
(k) (j) (i)
For notational convenience, write H k = T aX∗ M and Hi,j = T aP1 ∗ T f∗ M ; then the
classes η−l induce operators
(43) η−l : Hi,j → Hi−l+2,j+l (1).
But as we are on P1 , the operators η−l vanish for l ≥ 3. Further, there’s a unique
choice of decomposition of T f∗ M such that η−1 = 0 and η0 and η−2 commute [7,
Proposition 3.5]. We will work with this decomposition throughout.
Let L be the pullback along f of a Kähler class on P1 . This acts on the
cohomology groups by L : Hi,j → Hi,j+2 (1). The two non-zero components of η are
η0 : Hi,j → Hi+2,j (1) and η−2 : Hi,j → Hi,j+2 (1). Since L and η commute, we see
using the bigrading that L commutes with η0 and η−2 as well.
Now choose a polarization S of M . Note that η0 acting on Hi,j is induced by the
(i) (i+2)
relative Lefschetz operator T f∗ M → T f∗ M (1) inducedby η. Hence relative
L
hard Lefschetz for η and hard Lefschetz for L imply that i,j Hi,j , η0 , L is a
bi-sl2 -Hodge structure (with the Lefschetz operators being the weight-increasing
operators), and Theorem 1.1 for f and for the constant map on P1 imply that this
structure is polarized by
M
(44) S ′ := T aP1 ∗T f∗ S|Hi,j ⊗H−i,−j .
i,j≥0
Concretely, this means that on the (η0 , L)-bi-primitive part of H−i,−j , the pairing
(45) (x, y) 7→ S ′ (η0i Lj x, CD y)
is a positive-definite hermitian form. We need to show that
M
(46) S ′′ = T aX∗ S|H k ⊗H −k
k≥0
L k
polarizes the sl2 -Hodge structure k H , η0 + η−2 . By Lemma 2.7, it suffices to
show that S ′′ polarizes the bi-sl2 -Hodge structure
M
(47) Hi,j , η0 , η−2 .
i,j
′ ′′
By Lemma 2.14, S and S agree on each of the spaces Hi,j ⊗ H−i,−j , so it suffices
to show that S ′ polarizes the structure (47).
Note that Hi,j can only be non-zero for j = −1, 0, 1, so η−2 and L can only act
non-trivially on the spaces Hi,−1 . On such a space, either operator is an isomorphism
22 MADS BACH VILLADSEN
∼
Hi,−1 − → Hi,1 (1). We will show that the automorphism L−1 η−2 of Hi,−1 only has
positive real eigenvalues.
To do this, first recall that η−2 and L commute with η0 , hence in particular
preserve the η0 -primitive decomposition of H−i,−1 , so it suffices to look at the
restricted isomorphisms
η−2 ,L
(48) Pη0 H−i,−1 −−−−→ Pη0 H−i,1 (1)
and the eigenvalues of L−1 η−2 on Pη0 H−i,−1 .
Suppose λ is an eigenvalue and s ∈ Pη0 (H−i,−1 )C a corresponding eigenvector,
such that η−2 s = λLs. Note that s is automatically η−2 and L-primitive. Then
(49) S ′ (η0i η−2 s, CD s) = λS ′ (η0i Ls, CD s).
The former is the self-pairing of s under a hermitian form, hence a real number,
and the latter is a (positive real) norm times λ, hence λ is real. Suppose λ is
negative. Then replacing η with η − λL, which is still a Kähler form on X, we see
that (η − λL)i+1 s = η0i+1 s = 0 since s is primitive, contradicting hard Lefschetz for
η − λL.
It follows that for every a ≥ 0, the map η−2 + aL : H−i,−1 → H−i,1 (1) is an
isomorphism. In particular, the bi-primitive spaces of the bi-sl2 -Hodge structures
M
(50) Hi,j , η0 , L
i,j
M
(51) Hi,j , η0 , η−2 + aL
i,j
Besides the hard Lefschetz theorem, establishing the polarization formula is the
main difficulty in our proof of Theorem 1.1.
3.3. Reduction to normal crossings case. By Lemma 3.1 and section 3.2,
applied to a log resolution of (Z, D), it only remains to prove Theorem 1.1(3) (hard
Lefschetz) for T aX∗ M in the general case.
We can apply the reduction argument of [20, Section 6.3] as follows. Take an
embedded log resolution π : X ′ → X of (Z, D). This is a sequence of blowups in
smooth (proper) subvarieties of Z such that the strict transform Z ′ of Z in X ′
is smooth, and is D′ = Z ′ ∩ π −1 (D) is a simple normal crossings divisor in Z ′ .
Let M ′ be the generic pullback of M to Z ′ , considered as a Hodge module on X ′ .
By the previous section, aX ′ + M ′ satisfies the conclusions of the decomposition
theorem, hence [20, Section 6.3] applies to the associated Twistor D-modules, so
hard Lefschetz holds for T aX∗ M . Note that since we only cite Mochizuki’s work for
the hard Lefschetz theorem, which is a statement on the level of perverse sheaves,
we do not need to check any compatibility between Hodge modules and Twistor
D-modules.
One could alternatively use the results of [31] to get the hard Lefschetz theorem.
3.4. Rational decomposition. Let us finally show how Corollary 1.2 follows from
Theorem 1.1. If F is a real field, let E = R, else let E = C; note that F is dense in
E in the Euclidean topology.
Let KE = KF ⊗F E, and note that Rf∗ KE admits a decomposition in the derived
category by the hard Lefschetz theorem in Theorem 1.1(3) and Deligne’s theorem
[8]. The following lemma implies the desired decomposition of Rf∗ KF .
Lemma 3.5. L Suppose C is a rational constructible complex such that CE is iso-
morphic to k
p
H k
CE [−k] in the derived category. Then C is isomorphic to
k H C[−k].
L p k
4. Kollár’s conjecture
Kollár [19] conjectured that to any VHS V defined on a smooth Zariski open
subset of a projective variety X, there should exist a coherent sheaf SX (V ) on X
24 MADS BACH VILLADSEN
References
[1] B. Bakker, Y. Brunebarbe, and J. Tsimerman. “o-minimal GAGA and
a conjecture of Griffiths”. Inventiones mathematicae 232.1 (Apr. 2023),
pp. 163–228.
[2] A. A. Beilinson, J. Bernstein, and P. Deligne. “Faisceaux pervers”. In:
Analysis and topology on singular spaces, I (Luminy, 1981). Vol. 100.
Astérisque. Soc. Math. France, Paris, 1982, pp. 5–171.
[3] Y. Brunebarbe and B. Cadorel. “Hyperbolicity of varieties supporting
a variation of Hodge structure”. Int. Math. Res. Not. IMRN 6 (2020),
pp. 1601–1609.
[4] E. Cattani, A. Kaplan, and W. Schmid. “L2 and intersection cohomologies
for a polarizable Variation of Hodge structure.” Inventiones mathematicae
87 (1987), pp. 217–252.
[5] E. Cattani, A. Kaplan, and W. Schmid. “Degeneration of Hodge structures”.
Ann. of Math. (2) 123.3 (1986), pp. 457–535.
REFERENCES 25
[6] C. W. Curtis and I. Reiner. Representation theory of finite groups and asso-
ciative algebras. Vol. Vol. XI. Pure and Applied Mathematics. Interscience
Publishers (a division of John Wiley & Sons, Inc.), New York-London,
1962, pp. xiv+685.
[7] P. Deligne. “Décompositions dans la catégorie dérivée”. In: Motives (Seat-
tle, WA, 1991). Vol. 55. Proc. Sympos. Pure Math. Amer. Math. Soc.,
Providence, RI, 1994, pp. 115–128.
[8] P. Deligne. “Théorème de Lefschetz et critères de dégénérescence de suites
spectrales”. fr. Publications Mathématiques de l’IHÉS 35 (1968), pp. 107–
126.
[9] P. Deligne. “Théorie de Hodge. II.” French. Publ. Math., Inst. Hautes Étud.
Sci. 40 (1971), pp. 5–57.
[10] A. Fujiki. “Closedness of the Douady spaces of compact Kähler spaces”.
Publ. Res. Inst. Math. Sci. 14.1 (1978), pp. 1–52.
[11] A. Fujiki. “On automorphism groups of compact Kähler manifolds”. Invent.
Math. 44.3 (1978), pp. 225–258.
[12] A. Fujiki. “On the Douady space of a compact complex space in the
category C . II”. Publ. Res. Inst. Math. Sci. 20.3 (1984), pp. 461–489.
[13] H. Grauert and R. Remmert. Coherent analytic sheaves. Vol. 265. Grund-
lehren der mathematischen Wissenschaften [Fundamental Principles of
Mathematical Sciences]. Springer-Verlag, Berlin, 1984, pp. xviii+249.
[14] P. A. Griffiths. “Periods of integrals on algebraic manifolds. III. Some
global differential-geometric properties of the period mapping”. Inst. Hautes
Études Sci. Publ. Math. 38 (1970), pp. 125–180.
[15] A. Grothendieck. Revêtements étales et groupe fondamental (SGA 1).
Vol. 224. Lecture notes in mathematics. Springer-Verlag, 1971.
[16] H. Hironaka. “Flattening theorem in complex-analytic geometry”. Amer.
J. Math. 97 (1975), pp. 503–547.
[17] M. Kashiwara and T. Kawai. “Hodge structure and holonomic systems”.
Proc. Japan Acad. Ser. A Math. Sci. 62.1 (1986), pp. 1–4.
[18] M. Kashiwara and T. Kawai. “The Poincaré lemma for variations of
polarized Hodge structure”. Publ. Res. Inst. Math. Sci. 23.2 (1987), pp. 345–
407.
[19] J. Kollár. “Higher direct images of dualizing sheaves. II”. Ann. of Math.
(2) 124.1 (1986), pp. 171–202.
[20] T. Mochizuki. L2 -complexes and twistor complexes of tame harmonic
bundles. Apr. 22, 2022. arXiv: 2204.10443v1 [math.CV].
[21] T. Peternell. “Modifications”. In: Several complex variables, VII. Vol. 74.
Encyclopaedia Math. Sci. Springer, Berlin, 1994, pp. 285–317.
[22] C. Sabbah and C. Schnell. “Degenerating Complex Variations of Hodge
Structure in Dimension One”. CoRR (2022). arXiv: 2206.08166 [math.AG].
[MHM] C. Sabbah and C. Schnell. The MHM Project. 2018. url: http://www.
cmls . polytechnique . fr / perso / sabbah . claude / MHMProject / mhm .
html.
[23] M. Saito. “A young person’s guide to mixed Hodge modules”. In: Hodge
theory and L2 -analysis. Vol. 39. Adv. Lect. Math. (ALM). Int. Press,
Somerville, MA, 2017, pp. 517–553.
26 REFERENCES