Sie sind auf Seite 1von 29

RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY

HELENE ESNAULT AND MICHAEL GROECHENIG


arXiv:1707.00752v1 [math.AG] 3 Jul 2017

Abstract. We prove that the monodromy of a cohomologically rigid integrable connec-


tion (E, ) on a smooth complex projective variety X is integral. This answers positively
a special case of a conjecture by Carlos Simpson. To this aim, we prove that the mod p
reduction of a rigid integrable connection (E, ) has the structure of an isocrystal with
Frobenius structure. We also prove that rigid integrable connections with vanishing p-
curvatures are unitary. This allows one to prove new cases of Grothendiecks p-curvature
conjecture.

1. Introduction
Let (X, OX (1)) be a smooth complex projective variety. Simpson constructed quasi-
projective coarse moduli spaces MdR (X/C)(r), MB (X/C)(r), MDol (X/C)(r) of rank r in-
tegrable connections, local systems, P -semi-stable Higgs bundles, where P -semi-stability
for a Higgs bundle (V, ) is defined with respect to the Euler characteristic P (V, n) =
r1 (X, V OX OX (n)) Z[n], n N ([Sim94]). He showed they are real analytic iso-
morphic spaces, where the Zariski open of irreducible integrable connections corresponds to
the Zariski open of irreducible local systems and the Zariski open of stable Higgs bundles.
In particular, an isolated point on one moduli is isolated on all moduli, and called a rigid
connection (or Higgs bundle, or local system). Viewed as a Higgs bundle, a rigid connection
(E, ) is fixed under the C -deformation (V, ) (V, t ), t C , and thus is a complex
variation of Hodge structure ([Sim92, Lem. 4.1]). Simpson conjectures more.
Conjecture 1.1 (Motivicity). An irreducible rigid connection (E, ) on X is of geometric
origin.
This means that there is a non-trivial Zariski open U , X such that (E, )|U is a
subquotient of a Gau-Manin connection Ri f C for some smooth projective morphism
f : Y X, and some i N.
One restricts the question to irreducible connections as the moduli point of a reducible
connection only sees its semi-simplification. Furthermore, MdR (X/C)(1) acts non-trivially

Date: July 3rd, 2017.


The first author is supported by the Einstein program and the ERC Advanced Grant 226257, the second
author was supported by a Marie Sklodowska-Curie fellowship: This project has received funding from the
European Unions Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie
Grant Agreement No. 701679.

.
1
2 HELENE ESNAULT AND MICHAEL GROECHENIG

via tensor product on MdR (X/C)(r), and has dimension b1 (X), the first Betti number of X.
So the existence of rigid connections forces b1 (X) to be 0. In this case, MdR (X/C)(r) is a
finite disjoint union of the moduli schemes MdR (X/C, LdR )(r) parametrizing the irreducible
rank r connections with fixed determinant LdR MdR (X/C)(1), which is of finite order.
So [(E, )] MdR (X/C, LdR ) is isolated if and only it is in MdR (X/C)(r). In general, one
uses the following definition.
Definition 1.2. An irreducible rank r connection (E, ) with determinant LdR is rigid if
its moduli point [(E, )] MdR (X/C, LdR )(r) is isolated.
One defines similarly MB (X/C, LB )(r), MDol (X/C, LDol )(r) and rigid irreducible local
systems, resp. Higgs bundles with determinant LB , resp. LDol . Isolated points correspond
to each other in the three moduli. Our results hold as well in this more general situation.
We indicate in Remarks 3.2 and 4.6 the slight changes needed in the proofs for this case as
well.
If X has dimension 1, rigid local systems do not exist. They do exist only over stacky
versions of P1 , that is one has to relax the projectivity assumption, and the analogous
conjecture (requesting in addition quasi-unipotent monodromies at infinity) is proven by
Katz [Kat96, Thm. 9.4.1], by completely classifying the rigid connections. For SL(2)-
connections, the conjecture follows from the classification theorem of [CS08].
As factors of monodromy representations defined over Z are defined over number rings,
the motivicity conjecture implies the following conjecture.
Conjecture 1.3 (Integrality). A rank r irreducible rigid connection (E, ) on X is integral.
The connection being rigid, it has monodromy in GL(r, K), where K is a number field.
Integrality means that the monodromy stabilizes an OK -lattice.
The motivicity conjecture is proved for SL(3)-connections with integral monodromy in
[LS16, Thm. 0.3], again by complete classification. In loc.cit., the motivicity and integral-
ity conjectures are precisely formulated as Conjectures 0.1 and 0.2, along with a descent
property for rigid lisse `-adic sheaves ([Sim92, Thm. 4]).
The first aim of this article is to give a positive answer to a weaker form of the integrality
conjecture 1.3. One says that an irreducible connection is cohomologically rigid if in addition
to being rigid, the corresponding moduli point in MdR (X/C)(r), resp. MdR (X/C, LdR )(r)
is smooth. It is equivalent to saying that HdR 1 (X, End0 (E, )) = 0. In dimension 1, one

knows that all rigid connections are cohomologically rigid ([BE04, Thm. 4.10], and [Kat96,
Thm. 1.1.2] for the `-adic analog from which it initially comes). As this is a cohomological
condition, it propagates to MB and MDol and the corresponding moduli points are rigid
and smooth.
Theorem 1.4. Let X be a smooth connected projective complex variety. Then irreducible
cohomologically rigid connections are integral.
We explain the Leitfaden of the proof. At first, we discuss the p-curvatures of rigid con-
nections in general. A particular case of geometric origin of (E, ) is when the connection
has finite monodromy, or equivalently when there is a finite etale cover f : Y X such
that f (E, ) is trivial. This implies that if (XS , (ES , S )) is a model of (X, (E, )) over a
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 3

scheme of finite type S over Z, then for all closed points s of a possibly smaller non-empty
Zariski open S 0 S, (Es , s ) has vanishing p-curvature. Grothendiecks p-curvature con-
jecture predicts the converse, namely the vanishing of the p-curvatures for all closed points
of some open of definition forces the characteristic 0 connection to have finite monodromy
([Kat72, I]). The evidence is so far provided by solvable and by Gau-Manin connections,
see [EK16, Introduction] for a recent summary.
On the other hand, the Gau-Manin connections in characteristic p have nilpotent
p-curvatures (see [Kat72, 3.1]), thus connections of geometric origin have nilpotent p-
curvatures as well. The first main result of this note is the following theorem.
Theorem 1.5. Let X be a smooth connected projective complex variety. Let (E, ) be a
rigid connection. Then there is a scheme S of finite type over Z over which (X, (E, ))
has a model (XS , (ES , S )) such that for all closed points s S, (Es , s ) has nilpotent
p-curvature.
Another way to formulate the same result is to say that the restriction (EKs , Ks ) of
connection (ES , S ) to the p-adic variety XKs , where Spec(Ks ) is a p-adic point of S with
residue field s, is an isocrystal on Xs .
A natural generalization of Grothendiecks p-curvature conjecture is the prediction of the
converse, namely the nilpotency of the p-curvatures for all closed points of some open of
definition should force the characteristic 0 connection to be motivic. We do not discuss at
all this very deep problem.
Gau-Manin connections over a finite field s, viewed as an isocrystal on Xs , have a
Frobenius structure. The second main result of this note is the following theorem, a version
of which is proven in dimension 1 by Crew [Cre17, Thm. 3].
Theorem 1.6. Let X be a smooth connected projective complex variety. Let (E, ) be a
rigid connection. Then there is a scheme S of finite type over Z over which (X, (E, )) has
a model (XS , (ES , S )) such that the isocrystal (EKs , Ks ) has a Frobenius structure.
One reformulates Theorem 1.5 and Theorem 1.6.
Theorem 1.7. An irreducible rigid connection (E, ) on a smooth complex projective va-
riety X has the property that for all models (XS , (ES , S )) over a scheme S of finite type
over Z, there is a non-empty subscheme S 0 S, such that that for all Kv -points of S 0 , where
Kv is a p-adic field, the connection (EKv , Kv ) is an isocrystal with Frobenius structure on
Xs , where s is the residue field of the valuation v.
In fact we will prove a slightly stronger statement. As one would expect from [Ogu77,
1.6.4], the isocrystal above is induced by a filtered Frobenius crystal (with a W (Fpf )-
endomorphism structure). This shows that the isocrystals with Frobenius structure stem-
ming from Theorem 1.7 are associated to a crystalline representation 1 (XKv ) GL(W (Fpf )
for some natural number f 1 called the period. These statements are consequences of the
theory of Lan-Sheng-Zuo, see Remark 4.8. By comparing their construction with Faltingss
p-adic Simpson correspondence, one shows that this representation is rigid over 1 (XKv ),
see Theorem 5.3. This enables one to prove that we have, on the level P GL-representations,
f = 1 for infinitely many p (see Corollary 5.11). By relating the period f to Simpsons
correspondence, one proves the second main theorem of this note.
4 HELENE ESNAULT AND MICHAEL GROECHENIG

Theorem 1.8. Let X be a smooth connected projective complex variety. Then irreducible
rigid connections which have vanishing p-curvature for all closed points s in a non-empty
open subscheme S 0 of S have unitary monodromy.
One of the key ingredients of Theorem 1.4 is the existence of `-adic companions con-
structed in [AE16, Thm. 4.2] associated to the isocrystals with Frobenius structure of
Theorem 1.6. One then has to show that they are irreducible and rigid viewed as repre-
sentations of 1 (XKs ). Under the cohomological rigidity assumption, the question reduces
to a computation of weight, L function and base change. This is the only reason why we
have to assume in Theorem 1.4 cohomological rigidity, as we could not prove rigidity of the
`-adic companions under the sole rigidity assumption. See Section 7.
We finish the article with further various remarks on rigid connections with vanishing
p-curvatures. In Proposition 8.2 we conclude that if rigid unitary connections have finite
monodromies, then all rigid connections with vanishing p-curvature have finite monodromy.
In Proposition 8.3 we show that if all cohomologically rigid connections have vanishing p-
curvature, then all cohomologically rigid connections have finite monodromy. We also
mention that Theorem 1.4 shows that the main theorem of Langer-Simpson [LS16] is true
unconditionally for cohomologically rigid connections, and that their precise description
of which Gau-Manin appear in their solution to Conjecture 1.1 for SL(3)-connections
forces Grothendiecks p-curvature conjecture to be true for cohomologically rigid SL(3)-
connections on smooth projective varieties. See Subsection 8.1 and Remark 8.4.
We now describe the ideas of the proof of Theorem 1.5. As already mentioned, Simp-
son proves in [Sim92, Lem. 4.1] that a rigid connection is a complex variation of Hodge
structure. To this aim, he uses the full strength of his correspondence between irreducible
integrable connections and P -stable Higgs bundles. He argues first that the corresponding
Higgs bundle is rigid, then by letting Gm act on it, that is is a Hodge bundle, then he
returns to the initial connection via the correspondence. In characteristic p > 0 we have
Ogus-Vologodskys de Rham-Higgs correspondence [OV07] for connections with nilpotent
curvature, while submitted to the choice of the W2 -lift of the underlying variety. The lifta-
bility of Xs is at hand in our situation, but our aim is to show that the p-curvature of
s is nilpotent, so one cant use it for the purpose of proving Theorem 1.5. We also have
the second authors thesis [Gro16], where a full de Rham-Higgs correspondence is defined,
yet only for curves. Again it depends on an extra choice. One can remark that the exis-
tence of those extra choices in the two theories show that they can not be in all generality
compatible with Simpsons correspondence.
However, using crucially the work of Braverman-Bezrukavnikov [BB07], a Beauville-
Narasimhan-Ramanan correspondence is proved in [Gro16, Prop. 3.15], to the effect that
there is an equivalence of categories between connections on a characteristic p > 0 variety
and the data consisting of pairs of spectral varieties together with a module under an
Azumaya algebra on it. This is a key point of the proof of Theorem 1.5, as it enables one to
go from connections to linear objects and back. We detail the proof in Theorem 2.4. The
second main ingredient (see Proposition 2.7) is deduced from Ogus-Vologodskys theorem
[OV07, Cor. 2.9] after which a W2 -lift of the characteristic p > 0 variety Xs defines a
splitting of the Azumaya algebra along the (p 1)-th infinitesimal order of the zero section
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 5

in the cotangent bundle. One constructs an infinitesimal deformation of (Es , s ) on Xs in


the direction of the base of the Hitchin map (see (2.2)) to the order precisely (p 1), which
contradicts, as p gets large, the assumption that(E, ) is an isolated point in MdR (X/C)(r)
(see Section 3).

The proof of Theorem 1.6 relies on the work of Lan-Sheng-Zuo (see [LSZ13, Thm. 1.4]),
related to Faltingss p-adic Hodge theory. Their theory is central to our approach. They
associate to what they call a periodic Higgs-de Rham flow on Xs an isocrystal on Xs with
Frobenius structure. We define a Higgs-de Rham flow showing that the Cartier inverse
of a rigid stable Higgs bundle with nilpotent Higgs field is a stable rigid connection with
nilpotent p-curvature, see Proposition 4.3. We prove its periodicity using Theorem 1.5. See
Section 4.

Acknowledgement. The article rests on Simpsons theory, his definition and study of
complex rigid integrable connections, his program. It is a great pleasure to thank him for
the discussions we had when we formulated the first step of the Leitfaden of proof of this
article, namely when we proposed to study the p-curvatures of rigid connections and prove
that they are nilpotent (see Theorem 1.5). We thank him for his warm encouragements.
We thank Kang Zuo for his numerous explanations on the theory of Higgs-de Rham flows
he developed and his essential help. Theorem 1.6 relies on [LSZ13]. We thank Tomoyuki
Abe for the numerous discussions we had on isocrystals with Frobenius structure. Our
work [AE16] plays an important role in the proof of Theorem 1.4. We thank Mark Kisin
for ongoing enlightening discussions on topics related to the content of this article. We
thank Carlos Simpson and Mark Kisin for pointing out an error in a very early version of
Propositions 8.2 and 8.3. We thank Adrian Langer for his interest, for exchanges on our
wish to prove nilpotency of the p-curvatures of rigid connections and for discussions around
the topic presented here. His beautiful lecture in Berlin on [LS16] awoke our interest. We
thank Marco dAddezio, Alexander Beilinson, Javier Fresan, Phung Ho Hai, Moritz Kerz,
Joao Pedro dos Santos, Simon Pepin Lehalleur, Takeshi Saito for discussions during the
preparation of this work. The first author thanks the Vietnam Institute for Advanced
Study in Mathematics for hospitality, and all the participants at the workshop in Tuan
Chau where a tiny piece of the material of this article has been presented.

2. Characteristic p connections and their canonical deformations


Let (Z, OZ (1)) be a smooth projective variety defined over a perfect field k of character-
istic p. We denote by Z 0 is Frobenius twist, by F : Z Z 0 the relative Frobenius. In the
applications, Z is the fibre Xs at a closed point s of a scheme S of finite type over Z over
which a smooth complex projective variety X is defined, and k is the residue field of s. In
particular, k is finite.

By [Lan14, Thm 1.1] there is a quasi-projective coarse moduli space MdR (Z/k)(r) of P -
stable integrable connections. The p-curvature () H 0 (Z, F 1Z 0 End(E)) of any inte-
grable connection (E, ) on Z is flat under the tensor product of the canonical connection
on F 1Z 0 and of End() on End(E) ([Kat70, Prop. 5.2.3]). In particular its characteristic
6 HELENE ESNAULT AND MICHAEL GROECHENIG

polynomial has coefficients in global symmetric forms on Z 0 :

(2.1) (()) = det(() + Id) = r a1 r1 + . . . + (1)r ar ,


ai H 0 (Z 0 , Symi (1Z 0 )).
By the work of Laszlo-Pauly [LP01, Prop. 3.2], Bezrukavnikov-Braverman [BB07, Section 4],
Groechenig [Gro16, Defns. 3.12, 3.16], Chen-Zhu [CZ15, Section 2.1], (()) defines a so-
called Hitchin map
(2.2) dR : MdR (Z/k)(r) AZ 0 ,r
which is known to be proper ([Lan14, Thm. 3.8]), and where the affine variety AZ 0 ,r is
defined by
(2.3) AZ 0 ,r (T ) = ri=1 H 0 (Z 0 , Symi (1Z 0 )) k T
for any k-algebra T . As already mentioned in the introduction, if we do assume there are
rigid connections, then b1 (X) = 0, thus H 0 (X, 1X ) = 0, thus by base change, after perhaps
shrinking S,
(2.4) AZ 0 ,r (T ) = ri=2 H 0 (Z 0 , Symi (1Z 0 )) k T.
By definition, () is nilpotent if and only if dR ([(E, )]) = 0, where [(E, )] is the
moduli point of the connection (E, ).
On the other hand, by [BB07, Section 3.2], F DZ , where DZ is the sheaf of differential
operators spanned by OZ and the vector fields TZ with the standard relation f f =
(f ) for sections f OZ , TZ , is equal to 0 DZ 0 , there 0 : TZ 0 Z 0 is the projection
from the cotangent bundle and DZ 0 is a sheaf of Azumaya algebras over its center OT Z 0 .
One uses this to transfer the study of DZ -modules on Z to the study of DZ 0 -modules on
T Z 0 . In the following QCohZ (A) denotes the category of quasi-coherent sheaves over a
quasi-coherent sheaf of OZ -algebras A on Z.
Lemma 2.1. The following functors are equivalences of categories
F 0
QCohZ (DZ ) QCohZ 0 (F DZ ) QCohT Z 0 (DZ 0 ).
More generally, for a smooth morphism Z S we have equivalences of categories
F 0
QCohZ (DZ/S ) QCohZ 0 (F DZ/S ) QCohT Z 0 (DZ 0 /S 0 ).
Proof. This follows from the facts that F and 0 are affine morphisms, and DZ and DZ 0
are quasi-coherent sheaves of algebras. We then apply the following well-known fact:
f
Claim 2.2. Let U V be an affine morphism and B a quasi-coherent sheaf of algebras on
U . Then the functor f induces an equivalence of categories QCohU (B) and QCohV (f B).
This assertion can be checked by a direct local verification, or can be deduced as a
corollary of the Barr-Beck theorem [Bor94, Section 4] The condition of f being affine implies
that f is a conservative and exact functor, and hence the assumptions of the Barr-Beck
theorem are met. 
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 7

We recall the Beauville-Narasimhan-Ramanan (BNR) correspondence proven in [Gro16,


Prop. 3.15]. To this aim, one introduces ([Gro16, 3.1]), for any a : S AZ 0 ,r , where S is a
k-scheme, a finite cover
0 / Z 0 k S
Za0

# 
S
defined by the equation (2.1) = 0.
Remark 2.3. Since it is central to our approach, we recall the definition of the spectral
variety. Consider the quasi-coherent sheaf of algebras given by symmetric forms Sym 1Z 0
with its natural grading. We adjoin a formal variable of degree 1 and obtain the quasi-
coherent sheaf of algebras Sym 1Z 0 []. Recall that the points of the Hitchin base AZ 0 ,r are
defined by (2.3). The following more detailed description is useful.
(a) The vector space of degree r sections in H 0 (Z 0 , Sym 1Z 0 []) is isomorphic to AZ 0 ,r (k).
0
The tautological section H 0 (Z 0 , 1Z 0 ) is defined by the OZ 0 -linear homomorphism
OZ 0 0 0 1Z 0 = 1Z 0 OZ 0 Sym TZ 0 which is equal to the identity on the factor 1Z 0 OZ 0
TZ 0 = End(1Z 0 ) and equal to zero on the other factors. Pullback along 0 post-composed
with the specialization 7 defines a morphism of algebras
H 0 (Z 0 , Sym 1Z 0 []) H 0 (T Z 0 , 0 Sym 1Z 0 ).
Similarly, one obtains for a k-scheme S a morphism
H 0 (Z 0 k S, pZ 0 Sym 1Z 0 []) H 0 (T Z 0 k S, pT 0 0 Sym 1Z 0 )
Z

where p? : ? k S ? is the projection.


(b) For a AZ 0 ,r (S) we define the spectral cover Za0 to be the closed subscheme of
T Z 0 k S defined by the sheaf of ideals
(r a1 ++(1)r ar )
Ker (Sym TZ 0 ) k OS (Sym TZ 0 ) k OS


in ( 0 k Id) OT 0 k S = (Sym TZ 0 ) k OS . Then Za0 Z k a is a finite morphism.


Z
The closed subscheme of T Z 0 defined in (b) above is precisely the zero locus defined by
the degree r section (r a1 + + (1)r ar ) H 0 (Z 0 , Sym 1Z 0 []).
It is called the spectral cover to a. It is applied in particular when a = dR ((E, )).
Theorem 2.4 (BNR correspondence for integrable connections). Let S be a k-scheme.
The groupoid of rank r integrable connections (E, ) on Z k S, relative to S, is equivalent
to the groupoid of pairs
(a : S AZ 0 ,r , M )
where M is a DZ 0 -module on Za0 T Z 0 , such that 0 M is locally free of rank pd r.
Remark 2.5. Recall that we have a quasi-coherent sheaf of OT Z 0 -algebras DZ 0 . We say
that on TZ 0 k S a quasi-coherent pT 0 DZ 0 -module M is scheme-theoretically supported on
Z
a closed subscheme i : X , T Z 0 k S, if there exists a quasi-coherent i pT 0 DZ 0 -module
Z
8 HELENE ESNAULT AND MICHAEL GROECHENIG

f, which is then uniquely defined, such that i M


M f ' M as a p DZ 0 -module. The scheme-
TZ 0
theoretic support of M as a pT 0 DZ 0 -module only depends on M as a quasi-coherent sheaf
Z
on T Z 0 k S. For any closed immersion X , Y , T Z 0 k S, if M is scheme-theoretically
supported on X, then also on Y . The main part of the proof below is devoted to improving
an a priori bound for the scheme-theoretic support.
Proof of Theorem 2.4. The connection defines the structure of a pZ DZ -module on E over
Zk S. Therefore we obtain a F pZ DZ -module F E, which can be written as pushforward of
a pT 0 DZ 0 -module M on T Z 0 k S along the canonical projection 0 : T Z 0 k S Z 0 k S
Z
(see Lemma 2.1). As remarked above, the scheme-theoretic support of M depends only
on its OT Z 0 k S -module structure, which is induced by the p-curvature () : F E
pZ 0 1Z 0 OZ 0 S F E.
k
The pair F (E, ()) is a Higgs bundle of rank rpd on Z 0 k S, where d = dim(Z). Let
b AZ 0 ,rpd (S) be the characteristic polynomial of the Higgs field (). Then the Higgs
bundle F (E, ()) is scheme-theoretically supported on the closed subscheme
Zb0 T Z 0 k S.
However this upper bound for the scheme-theoretic support is far from being optimal.
d
We construct a degree r polynomial a AZ 0 ,r (S), such that b = ap (for the multiplication
as polynomials), and such that M is scheme-theoretically supported on Za0 . It suffices to
show that M is scheme-theoretically supported on Za0 , Zb0 , T Z 0 k S.
Since F (E, ) is a Higgs bundle of rank pd r, b AZ 0 ,pd r (S), M has scheme-theoretic
d
support Zb0 . We next show that there exists a AZ 0 ,r (S), such that b = ap , and M is
scheme-theoretically supported on Za0 .
It is clear that if there is at most one such a AZ 0 ,r (S) with this property since a and b
are monic polynomials. Etale descent for symmetric forms on Z 0 allows one to construct a
etale locally, and to prove that M is scheme-theoretically supported on Za0 etale locally.
Let x Z 0 k S(k sep ) be a geometric point. We denote by Oxh the corresponding henselian
local ring. Pulling back the finite morphism Zb0 Z 0 k S along Spec Oxh Z 0 k S we
obtain the spectrum of a product of henselian local rings Spec R = Spec N
Q
i=1 Ri . Since
DZ 0 is an Azumaya algebra, the pullback DZ 0 |Spec R splits, that is, is isomorphic to the sheaf
of matrix algebras Mpd (OSpec R ).
Classical Morita theory (see [Lam99, Thms 18.11 & 18.2]) implies that every quasi-
coherent Mpd (OSpec R )-module is isomorphic to a unique (up to a unique isomorphism)
p d d
Mpd (OSpec R )-module of the shape F OSpec R OSpec p with the canonical M (O
R =F pd Spec R )-
action.
Applying the push-forward functor 0 we obtain that the Higgs bundle (F (E, ())|Spec Oxh
d
splits as a direct sum 0 F p . Let a be the characteristic polynomial of the Higgs bundle
d
0 F on Spec Oxh . We have b|Spec Oxh = ap (as polynomials), and F is scheme-theoretically
d
supported on Za0 . Since F p = M , this also holds for M .
Since Spec Oxh is an inverse limit of etale neighbourhoods of x, we obtain an etale mor-
h
Z 0 k S, such that the Azumaya algebra DZ 0 splits already when pulled
phism U
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 9

back to U Z 0 k S Zb0 . Repeating the argument above we obtain a degree n section of


h Sym Z 0 S/S [] (where is a formal variable of weight 1)
n1
X
n
a= + (1)i ani i ,
i=0
d
such that ap= h b.
d
By unicity of solutions to the equation ap = b (where a and b are monic homogeneous
elements of weight n) in Sym 1Z 0 S/S [] we obtain a AZ 0 ,r (S) with the required property.
We conclude from the local strict henselian support property that M is scheme-theoretically
supported on Za0 . This finishes the proof. 
Remark 2.6. By construction, for (E, ) a rank r integrable connection on Z, the point
a AZ 0 (k) from Theorem 2.4 has coordinates ai H 0 (Z 0 , Symi (1Z 0 )) from (2.1).
We consider the Gm -action
(2.5) T Z 0 k Gm
m / T Z 0

0
& 0
Z
m
defined by the conic structure on T Z 0 . We use the notation m : V k Gm T Z 0 for the
restriction of m to any subscheme V T Z 0 . For any natural number n 1, we define
Vn = V k Spec(k[t]/(t 1)n ) V k Gm , with Gm = Spec(k[t, t1 ]), inserted in the
diagram
(2.6) Vn
m / T Z 0

0
" 
Z0
We also denote by r : Vn V the retraction, base changed from Spec(k[t]/(t 1)n )
Spec(k).
Proposition 2.7. Let Z be a smooth projective variety defined over a perfect characteristic
p > 0 field. Let (E, ) be an integrable connection on Z with Hitchin invariant a =
dR ((E, )). Then if Z lifts to W2 (k), we have the equality of Brauer classes
[m DZ 0 ] = [r DZ 0 ] on (Za0 )p .
2 (T Z 0 , O
Proof. By [OV07, Prop. 4.4], the Brauer class [DZ 0 ] Het T Z 0 ) of the Azumaya
algebra DZ 0 is the image () of the tautological one-form
0
H 0 (T Z 0 , 1Z 0 ) H 0 (T Z 0 , 1T Z 0 )
(see Remark 2.3) by the connecting homomorphism of the etale exact sequence
d log w C
(2.7) 0 OT Z 0 F O(T

F Z 1 (1(T Z 0 )0 ) 1T Z 0 0
Z 0 )0

on T Z 0 , where we think of T Z 0 as the Frobenius twist of (T Z 0 )0 via w : T Z 0 (T Z 0 )0 ,


where C is the Cartier operator, and F : (T Z 0 )0 T Z 0 is the Frobenius homomorphism,
10 HELENE ESNAULT AND MICHAEL GROECHENIG

and Z 1 (1(T Z 0 )0 ) are the closed forms. We now replace T Z 0 and (T Z 0 )0 in (2.7) by their
product with Gm over k, and the differential forms over k by the ones over Gm of the
product varieties. This yields on (Za0 )p
m DZ 0 = (m ), r DZ 0 = (r ).
On the other hand,
m r Ker H 0 ((Za0 )p , 1(Za0 )p /Spec(k[t]/(t1)p ) ) H 0 (Za0 , 1Za0 ) .


Thus (m r ) has support in Zp0 , T Z k G0m , where Z 0 , T Z 0 is the zero-section.


We conclude
(m r ) Het
2
(Zp0 , OZ0 ).
p

Thus
(m r ) = (m |Zp0 r |Zp0 ) = m (|Zp0 ) r (|Zp0 ) Het
2
(Zp0 , OZ0 ).
p

As k is perfect, a lifting of Z to W2 (k) is equivalent to a lifting of Z0 to W2 (k). By [OV07,


Cor. 2.9], one has the vanishing
(|Zp0 ) = 0.
This finishes the proof. 
Remark 2.8. One can show that the choice of a W2 -lift in Proposition 2.7 above, induces
a canonical equivalence of categories
(2.8) QCoh(Z 0 ) (DZ 0 ).
QCoh(Z 0 ) (m DZ 0 ) =
a p a p

To see this one applies [BB07, Prop. 3.11] instead of [OV07, Prop. 4.4] in the argument
above (note that the authors of loc. cit. call two Azumaya algebras equivalent, if their
categories of modules are equivalent). Applying [BB07, Cor 3.12] to the special case of a
diagonal morphism, we obtain that for 1-forms 1 , 2 on Z 0 we have a canonical equiva-
lence of categories QCoh(D1 +2 )
= QCoh(D1 D2 ). Putting these two refined assertions
together, and evoking the splitting associated by OgusVologodsky to a W2 -lift [OV07,
Cor. 2.9] we obtain a canonical equivalence of categories as in (2.8).
We denote by : AZ 0 k Gm AZ 0 the action defined for t Gm by multiplication by
ti on H 0 (Z 0 , Symi 1Z 0 ).
Theorem 2.9. Let Z be a smooth projective variety defined over a perfect characteristic
p > 0 field k. Let (E, ) be an integrable connection on Z with Hitchin invariant a =
dR ((E, )). Then assuming that Z lifts to W2 (k), there exists an integrable connection on
m
Z k T , for T = Spec(k[t]/(t 1)p , with spectral variety (Za0 )p T Z 0 k T , with Hitchin
invariant (a k T ), and which restricts to (E, ) on Za0 , T Z 0 .
Proof. Let M be the pT Z 0 DZ 0 k T |Za0 associated to (E, ) via the correspondence of The-
orem 2.4. Then r M is a pT Z 0 r DZ 0 -module, thus by Proposition 2.7, r M can be viewed
as a pT Z 0 m DZ 0 -module on (Za0 )p (non-canonically). We apply again Theorem 2.4 to con-
clude to the existence of an integrable connection on Z k T with the properties of the
theorem. This finishes the proof. 
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 11

3. p-Curvature and Proof of Theorem 1.5


The aim of this section is to prove Theorem 1.5.
Let X be a smooth complex projective variety. We define
rig
(3.1) MdR (X/C)(r) MdR (X/C)(r)
to be the closed subscheme of rigid points of the quasi-projective moduli of P -stable in-
rig
tegrable connections. As MdR (X/C)(r) is quasi-projective, MdR (X/C)(r) is 0-dimensional
rig
and projective, and MdR (X/C)(r) is the disjoint union of MdR (X/C)(r) and of its open
rig
and closed complement MdR (X/C)(r) \ MdR (X/C)(r) which does not contain any isolated
point.
We fix a model XS of X over a scheme S of finite type over Z. By [Lan14, Thm 1.1],
one has quasi-projective moduli MdR (XS /S)(r) S. For any locally noetherian S-scheme
T , one has a morphism T : MdR (XS /S)(r) S T MdR (XT /T )(r). If T is a geometric
point, T induces an isomorphism on geometric points on both sides.
rig
If follows that shrinking S, MdR (XS /S)(r) is the disjoint union of MdR (XS /S)(r), which
rig
is closed and 0-dimensional over S, and of MdR (XS /S)(r) \ MdR (XS /S)(r), which is open
and closed, and does not contain any component which is 0-dimensional over S. Replacing
rig
S by an etale cover, we may assume that the finitely many complex points of MdR (X/C)(r)
rig
extend as S-points in MdR (XS /S)(r)(S). We choose one such connection (ES , S ) defining
rig
one moduli point [(ES , S )] MdR (XS /S)(r)(S).
We shrink S so XS /S is smooth and S/Z is smooth. So for any closed point s S, with
residue field k(s), one has lifts Spec(W2 (k(s))) S, and each such defines a lift XW2 (k(s))
of Xs . We consider the point T = Spec(k(s)[t]/(t1)p ) S factoring through s, where p is
the characteristic of s, and the integrable T -connection (E, ) constructed in Theorem 2.9
for Z = Xs , for a choice of XW2 (k(s)) and for (E, ) = (ES , S ) S s. Then (E, ) defines
rig
a point [(E, )] MdR (XS /S)(r)(T ). Indeed, the geometric points of T are above s, and
(E, )|Xs = (E, ) is P -stable. By construction, one has
(3.2) dR (T ([E, )]) = (a k(s) T ).

By definition of , if a 6= 0, for any closed subscheme s T 0 T with T 0 6= T , one has


(a k(s) T 0 ) 6= (a k(s) T ). However, the nilpotency order of each connected component
rig
of MdR (X/S) is bounded above, independently of p. Thus for p large, there is a closed
rig rig
subscheme s T 0 T with T 0 6= T such that MdR (XS /S)(r)(T 0 ) = MdR (XS /S)(r)(T ).
Hence the relation (3.2) forces a to be equal to 0. This finishes the proof.

Remark 3.1. The moduli space MdR (XS /S)(r) and MdR (Xs /s)(r) are not fine. But for
s a closed point of S, the Hitchin map on MdR (Xs /s)(r) enables one to distinguish points
on MdR (XS /S)(r) S s via s .

Remark 3.2. If we fix a finite order rank 1 connection LdR , and replaces MdR (X/C)(r) by
MdR (X/C), LdR )(r) etc., as mentioned in the introduction, then in the proof of Theorem 1.5
one has to define A0Z,r by the formula 2.4, not by the formula 2.3. Then the proof is identical.
12 HELENE ESNAULT AND MICHAEL GROECHENIG

4. Frobenius structure
The aim of this section is to prove Theorem 1.6 and Theorem 1.7.
rig
Let X be a smooth projective variety over C. We define MdR (X/C)(r) MdR (X/C)(r)
as in the proof of Theorem 1.5 and more generally, for any A1 (C), we define
Mrig (X/C)(r) M (X/C)
to be the 0-dimensional subscheme the points of which being the rigid integrable -connections.
rig
Let MHod (r) A1 be the restriction of MHod (X/C)(r) A1 to be the union over of all
M (X/C)(r). Recall that the notation [ ] denotes a moduli point.
rig
Lemma 4.1. The morphism MHod (X/C)(r) A1 is finite, flat, and splits as
((V,),)7
rig
MHod (X/C)(r) rig
= MDol (X/C)(r) C A1 A1 .
For [(N, D)] a section, then [(N, D)]1 MdR (X/C)(r) is the moduli point of a complex
variation of Hodge structure, with F -filtration F i F i1 . . . F 0 = N with Griffiths
transversality : F i 1X OX F i1 and [(N, D)]0 MDol (X/C)(r) is the moduli point
i 1 i1
of its associated Higgs bundle V = grF E, grF : i (grF E X OX grF E) .
Proof. By construction MHod (X/C)(r) A1 Gm Gm splits as

= ((E,),)7(E,) ((E,),)7
MHod (X/C)(r) A1 Gm MdR (X/C)(r) k Gm Gm ,
where MdR (X/C)(r) is the fibre at = 1. On the other hand, by [Sim97, Thm. 9.1],
at a complex point x MDol (X/C)(r), the fibre at = 0, MHod (X/C)(r) etale locally
isomorphic to the product of MDol (X/C)(r) with A1 . This finishes the proof of the first
part. As for the second part, this is an application of [Sim92, Lem. 4.1] (see [Sim97,
Lem. 7.2]). 

We take a scheme S of finite over Z such that X has a smooth model XS . For every
: S A1 , Langers construction [Lan14, Thm 1.1] yields a coarse moduli space of semi-
stable -connections M (XS /S) defined over S. In particular we can apply this to the
case
= pA1 : S A1 A1
and obtain an S-model of Simpsons Hodge moduli space MHod (XS /S) S A1 .
rig
We may assume that the finitely many moduli points in MdR (X/C)(r) of rigid con-
nections (E, ) have models (ES , S ) such that (Es , s ) is stable for all s S. Those
models define finitely many A1S -connections (NS , DS ) with moduli points [(NS , DS )]
MHod (XS /S)(r), with the property
rig
(4.1) N i i N i i
i=1 [(NS , DS )]|1S = i=1 [(ES , S )] = MdR (XS /S)(S),

where 1S : S A1S is the constant section defined by = 1. This defines


rig
(4.2) N N
i=1 [(NS , DS ]|0S = i=1 [(VS , S )] = MDol (XS /S)(S),
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 13

where 0S : S A1S is the constant section defined by = 0. By definition, for all closed
points s S (Vs , s ) MDol (Xs /s)(r)(k(s)) where k(s) is the residue field of s. We define
rig
(4.3) MHod (XS /S)(r) = N
i=1 [(NS , DS )] MHod (XS /S)(r)
rig
to be the union of those finitely many sections. Thus points in MHod (XS /S)(r) are iso-
1
lated in the fibre M (X /)(r) above their image AS . Vice versa, if a S- point of
rig
MHod (XS /S)(r) is isolated in its fibre M (X /)(r), it lies in MHod (XS /S)(r).
Furthermore, finiteness implies that by further shrinking S if necessary, we have the
following property: if for a given A1 , and a given closed point s S, [(NSi , DSi )]|s =
[(NSj , DSj )]|s , then i = j. We finally further shrink so as to assume that S Spec(Z) is
smooth.
In the sequel, we apply the theory of Higgs-de Rham flows developed by Lan-Sheng-Zuo,
see [LSZ13] for a complete reference. Our first aim is to construct a Higgs-de Rham flow
([LSZ13, Defn. 1.1]) using Lemma 4.1.
Lemma 4.2. Let k be a perfect field of positive characteristic p, and let Z/k be a smooth
projective k-variety. We denote by w : Z 0 Z the canonical isomorphism of schemes which
is the pull-back of the absolute Frobenius of Spec(k). Then a stable Higgs bundle (V, ) on
Z is rigid if and only if the stable Higgs bundle w1 (V, ) is rigid, and a stable integrable
connection (E, ) is rigid if and only if the stable integrable connection w1 (E, ) is rigid.
Proof. Recall that (V, ) is not rigid if and only if there exists an irreducible k-scheme
C of finite type, with dim C > 0, such that there exists an C-family of Higgs bundles
(VC , C ), C : VC 1Xk C/S VC , and two closed points c0 , c1 C(k 0 ) defined over a
finite field extension k 0 /k, such that (VC , C )c0 is isomorphic to (V, )k0 , and (VC , C )c0 and
(VC , C )c1 are not isomorphic over the algebraic closure k.
Since k is perfect, for each k-scheme C there exists a unique (up to a unique isomor-
phism) k-scheme C, such that (C)0 is isomorphic to C as a k-scheme. We have a canonical
isomorphism of schemes (again, not necessarily respecting the k-structure)
wZk C : Z 0 k C Z k C.
1
And the functor wZ kC
induces an equivalence between C-families of stable Higgs bundles
on Z 0 and C-families of stable Higgs bundles on Z. This concludes the proof in the Higgs
case, and also in the de Rham case, arguing precisely similarly. 
In the sequel, we simplify the notation of (2.3) and denote by A0 what there is denoted
by AZ 0 ,r . We apply it to Xs for a model XS /S of X and s S a closed point. For T be
a k(s)-scheme, we say that a scheme-theoretic point a : T A0 is OV-admissible, if the
0 T X 0
spectral variety Xs,a s k(s) T (see Remark 2.3) factors through the (p 1)-st order
infinitesimal neighbourhood of the zero section Xs0 , T Xs0 . We assume until the end of
the section that that the characteristic p of k(s) is r + 2. Recall that then the functors
C and C 1 are defined in [OV07, Section 2.4].
Proposition 4.3. There exists a positive integer D, depending only on XS /S, such that
for any closed point s S with char k(s) > D, and any rigid stable Higgs bundle (Vs , s ),
C 1 (Vs , s ) is a stable rigid integrable connection.
14 HELENE ESNAULT AND MICHAEL GROECHENIG

Proof. The stability is proven in [Lan14, Cor. 5.10] (it is written for semi-stability, but the
argument holds for stability as well). The main point is the rigidity assertion, which we
prove now. Let T be a k(s)-scheme. By the BNR correspondence (Theorem 2.4) we may
describe a T -family flat connections (ET , T ) in terms of a pair (a : T A0 , M ) where M
is a DXs0 |Xs,a 0 0
0 -module on the spectral variety X
s,a , T Xs k(s) T .
If a A0 (T ) is OV-admissibble, we may apply OgusVologodskys result that D splits on
the (p 1)-st order neighbourhood of Xs0 , T Xs0 (see [OV07, Cor. 2.9]), and hence obtain
for every OV-admissible a A0 (T ) that the stack of stable flat connections (E, ) on Xs
with dR ((E, )) = a (defined in (2.2)) is equivalent to the stack of stable stable Higgs
bundles (V, ) on Xs0 with ((V, )) = a. We denote this equivalence by Ca , respectively
Ca1 .
There exists a linear function R(r), with the following property: let m be a positive
integer, and let A0(m) be the m-th order neighbourhood of 0 A0 . If p 1 > R(r) m, then
every scheme theoretic point a : T A0(m) is OV-admissible.
Let D0 be a positive integer, such that D0 is bigger than the degrees of the finite morphism
rig rig
MDol (XS /S) S and MdR (XS /S) S. Let k(s) B k(s) be an augmented Artinian
local algebra, such that we have a B-deformation (VB , B ). The characteristic polynomial
of B defines a point : Spec(B) A0 . Then factors through the (D0 1)-st order
0
infinitesimal neighbourhood A0(D 1) . To see this we observe that we have a commutative
diagram

Spec B / M rig (Xs /ks )
Dol


& 
A0 .

The morphism factors through the connected component of the moduli space correspond-
ing to the isolated point [(Vs , s )]. As a scheme it is non-reduced of order D0 . This shows
that factors through the (D0 1)-st order infinitesimal neighbourhood.
Let m be a positive integer such that m > D0 , and assume p 1 > R(r)m, such that
every scheme theoretic point a : T A0(m) is OV-admissible.
We assume by contradiction that (E, ) = C 1 (Vs , s ) is not rigid as a local system.
This implies that there exists a deformation (ET , T ), parametrised by an augmented
k(s)-scheme Spec k(s) T Spec k(s), such that the corresponding Hitchin invariant
dR ((ET , T )) does not factor through A0(m1) .
(m1)
We denote by Tt be the (m 1)-st order neighbourhood of t = Im(Spec(k(s))) in T .
By construction the family (ET (m1) , T (m1) ) has the property that dR ((ET (m1) , T (m1) ))
t t t t
(m1)
is a morphism Tt A0 ,
and therefore it factors through T A0(m1) ,but not through
A(k1) for k < m.
For every p 1 > R(r)m, we can apply the equivalence of categories Ca to construct
(m1)
a Tt -deformation (VT (m1) , T (m1) ) of (Vs , s ), such that (T (m1) ) does not factor
t t t
0
through A0(D 1) . This contradicts the choice of D0 , and shows that we may choose D =
R(r)m + 1. 
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 15

We fix one closed point s of the scheme S constructed at the beginning of this section.
We also fix a W2 (k(s))-point of S. We define a map
(4.4) : {1, . . . , N } {1, . . . , N }
where N is defined in (4.1) in the following way. Fixing i {1, . . . , N }, we set (NSi , DSi )0S S
rig
C = (Vsi , si ) MdR (X/C)(r). One first defines
0 0
(4.5) (Vs i , si ) = w (Vsi , si ),
0 0
which by Lemma 4.2 is rigid stable, then one defines C 1 (V i , i ) which by Proposition 4.3
is a stable rigid integrable connection. Thus, there is a uniquely defined (i) {1, . . . , N }
such that
0 0
(4.6) C 1 (Vs i , si ) = (Es(i) , s(i) ).
Lemma 4.4. The map is a bijection.
Proof. Clearly, we can reverse the argument: starting with (Esj , js ), then C(Esj , js ) again
is stable and rigid by Lemma 4.2, and thus w1 C(Esj , js ) as well. 
Fixing i N, 1 i N , we define a Higgs-de Rham flow (for the definition, see [LSZ13,
Defn. 1.1]) using as follows:
(i) (i) 2 (i) 2 (i)
(4.7) (Es , s ) (Es , s )
8 7

'
.% . .
(i) (i)
(Vsi , si ) (Vs , s )
The downwards arrows are obtained by taking the graded associated to the restriction to
Xs of the Hodge filtration on the rigid connections.
Lemma 4.5. 1) The Higgs-de Rham flow (4.7) is periodic of period fi which is the
order of the -orbit of i.
2) It does not depend on the choice of the W2 (k(s))-point of S chosen.
Proof. The point 1) is just by definition. As for 2): We fix a W2 (k(s))-lift of Xs . Its Kodaira-
Spencer class endows the set of equivalence classes of W2 (k(s))-lifts of Xs with k(s)-points
of an affine space A. Combining Lemma 4.3 with the operator C 1 , one obtains a A-family
of isomorphisms of moduli spaces
rig rig
MdR (Xs /s)(r) k(s) A ' MDol (Xs /s)(r) k(s) A.
Since the moduli spaces are 0-dimensional, the resulting bijection of closed points is inde-
pendent of the chosen W2 -lift. 
We fix now a W (k(s))-point of S, yielding XW (k(s)) . Recall that the category
M F[0 w],f (XW /W )
[LSZ13, Rmk. 2.4, Variant 2] is the category of formal connections on the formal scheme
XW (k(s)) defined as the completion of XW (k(s)) along XW (k(s)) W (k(s)) k(s), which consists
16 HELENE ESNAULT AND MICHAEL GROECHENIG

of filtered connections {H, , F il} which satisfy Griffiths transversality together with a
Frobenius structure W (Fpf ) EnddR (H, , F il).
Proof of Theorem 1.6. Recall that we have chosen a finite type scheme S over Z, such that
every rigid connection has an S-model. Furthermore, we may assume that for every s S,
every rigid connection (Es , s ) over Xs is the restriction of a unique S-model of a rigid
rig
connection. In particular we have that MdR (XS /S)red S is an isomorphism of schemes
on every connected component. This shows that over every s S with residue field k(s)
we have a bijection
rig rig
MdR (Xs /s)(r)(k(s)) = MdR (XS /S)(r)(W (k(s))).
We use this property below to conclude periodicity of the Higgs-de Rham flow starting
with (VW i i
(k(s)) , W (k(s)) ) over W (k(s)) from periodicity over k(s) (Lemma 4.5). One applies
the main theorem [LSZ13, Thm. 1.4] to conclude that there is a uniquely defined object
(H, , F il) M F[0 w],f (XW (k(s)) /W (k(s))) such that the restriction of (H, ) to Xs is
(Esi , is ). Then necessarily, the moduli point
[(H, )] MdR (XW (k(s)) /W (k(s)))(r)(W (k(s)))
rig
defined by (H, ) lies in MdR (XW (k(s)) )(r)(W (k(s))). It is by the preparation of S equal
i i
to (E , ) over some finite extension of Frac(W (k(s)) in C. 
Proof of Theorem 1.7. One combines Theorem 1.5 with Theorem 1.6. 
Remark 4.6. If we fix a finite order rank 1 connection LdR , and replaces MdR (X/C)(r)
by MdR (X/C, LdR )(r) etc., as mentioned in the introduction, then in the proof of Theo-
rem 1.6, one slightly changes the argumentation as follows. If LdR is the trivial connection
(OX , d), the Higgs-de Rham flow (4.7) is well defined between MDol (Xs /s, (OXs , 0))(r) and
MdR (Xs /s, (OXs , d))(r), and the proof is identical. If LdR is not trivial, and has order d, one
replaces MDol (Xs /s)(r) by the disjoint union of MDol (Xs /s, Li Dol,s )(r) and MdR (Xs /s)(r)
i
by MdR (Xs /s, LdR,s )(r) for 0 i d 1, and the proof is identical.
Corollary 4.7. Let X be a smooth projective variety over C. Let C , X be a complete
intersection of ample divisors. Let XS be a model on which all rigid connections are defined
and on which C has a model CS . Then there is a non-empty open subscheme S 0 , S
such that for all closed points s S, all W (k(s))-points of S 0 , there is a projective mor-
phism fs : Ys Cs (with Ys a smooth scheme), such that the F -overconvergent isocrystal
(EKs , Ks )|CKs on Cs is, over a finite extension of k(s), a subquotient of the Gau-Manin
F -overconvergent isocrystal Ri fs Qp for some i.
Proof. Let (E, ) be an irreducible rigid connection. Recall from the introduction that
its determinant is finite. Then by the classical Lefschetz theorem [Lef50], (E, )|C is
irreducible, and of course has finite determinant. By Theorem 1.7, (Ms := (EKs , Ks ))|CKs
is an irreducible isocrystal with Frobenius structure with finite determinant. By [Abe13,
Thm. 4.2.2], p1 Ms p2 Ms on Cs s Cs is a subquotient of Ri gs! Qp where gs : Chts
Cs s Cs is the Shtuka stack. According to [Abe13, Cor. 2.3.4], which states that an
admissible stack admits a proper surjective and generically finite cover by a smooth and
projective scheme, we can realize p1 Ms p2 Ms as a subquotient of some Rn hs Qp for some
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 17

projective morphism hs : Ys Cs s Cs , with Ys smooth projective. We restrict gs to


Cs s x for a closed point x Cs . This implies that Ms s x is a subquotient of Ri fs Qp
for fs = gs Cs s Cs (Cs s x), where : Cs s x Cs is defined by x s. 
Remark 4.8. We conclude this section by recalling the Fontaine-Lafaille-Faltings equiv-
alence between the category M F[0 w],f (XW (k(s)) /W (k(s))) and the category of represen-
tations of 1 (XKs ) with values in GL(r, W (Ffp )) where Ks = Frac(W (k(s)), as explained
in [LSZ13, p. 3,Thm. 3.2, Variant 2]. Thus starting with (Ei , i ) rigid over X/C, for
i = 1, . . . , N , choosing s, one constructs (H, , F il) M F[0 w],f (XW (k(s)) /W (k(s))) in the
Proof of Theorem 1.6. This corresponds to a p-adic representation
(4.8) i,s : 1 (XKs ) GL(r, W (Fpf )).

5. Identification of the p-adic representation associated to a rigid


connection
The aim of this section is to prove that i,s is geometrically rigid.
Definition 5.1. Let be a profinite group, and : GL(r, F ) a continuous and abso-
lutely irreducible representation, where F is a topological field. One denotes by
Def : ArtF Sets
the functor sending a finite-dimensional commutative local F -algebra A, with the canonical
topology as a finite-dimensional vector space over F , to the set of isomorphism classes of
continuous representations 0 : GL(r, A), such that there exists a finite field extension
F 0 /F , an F -morphism A F 0 , and an isomorphism (0 )F 0 ' F 0 . We say that is rigid,
if Def is corepresented by B ArtF .
Remark 5.2. For a local field F , and its ring of integers OF , we say that a continuous rep-
resentation : GL(r, OF ) is rigid and absolutely irreducible, if the associated residue
representation GL(r, kF ) is rigid and absolutely irreducible.
We freely use the notations of the preceding sections. If A is a field, a representation of
an abstract group G in GL(r, A) is said to be absolutely irreducible if the representation
G GL(r, A) GL(r, ) is irreducible, where is any algebraically closed field containing
A. Similarly with GL(r) replaced by P GL(r). If G is finitely generated, one defines the
moduli scheme MB (G, P GL(r)) of P GL(r)-representations of G, which is also a coarse
moduli scheme of finite type defined over Z. An isolated point is called rigid. For any ring
A, a P GL(r)-representation which is defined over A is said to be rigid if the corresponding
A-point of MB (G, P GL(r)) is isolated. Recall that we had a model XS of X over which
all rigid connections are defined. The field of functions Q(S) is by definition embedded
in C. We denote by Q(S) its algebraic closure in C. We denote by 1top the topological
fundamental group and by 1 the etale one.
The aim of this section is to prove the following theorem.
Theorem 5.3. There exist infinitely many closed points s S, with infinitely many prime
numbers char(k(s)), such that i,s |1 (XK ) is irreducible and rigid, and is in addition coho-
s
mologically rigid if (Ei , i ) is.
18 HELENE ESNAULT AND MICHAEL GROECHENIG

We start with general facts. We emphasize that the following lemma is based on Defini-
tion 1.2 of rigidity.
Lemma 5.4. Let be a finitely generated abstract group, and K be an algebraically
closed field of characteristic 0. We denote by : GL(r, K) an irreducible repre-
sentation. Then, is rigid, if and only if the corresponding projective representation
proj : P GL(r, K) is rigid.
Proof. We show first that rigidity of proj implies rigidity of . We assume by contradiction
that is not rigid. Then there exists a representation : GL(r, R), where R is a
discrete valuation ring with residue field K, such that the diagram
/ GL(r, R)

# 
GL(r, K)
commutes, which is a non-trivial deformation. Since proj is rigid, the associated projective
representation proj has to be constant. In particular we conclude that proj Frac(R) is
equivalent to proj after base change. This implies that there exists a character
: Frac(R),
such that
Frac(R) ' ( Frac(R)) .
Taking determinants, we see that r is trivial, which implies that is already defined over
K. By irreducibility, there is also an isomorphism
' ( Frac(R))
defined over Frac(R). We conclude the proof by observing that ' implies Tr() =
(g)Tr() for all g . Thus we have (g) = 1.
Vice versa, let us assume that is rigid. Let = det : K be the determinant
of . As above we consider a non-trivial deformation eproj : P GLn (R) of proj . The
obstruction of lifting eproj to a homomorphism e: GLn (R) with det e ' lies in
H 2 (, n (R)). Since n (R) = n (K), and the obstruction vanishes over the residue field
(indeed, proj is the projectivization of ), we see that the obstruction vanishes also over R.
This shows the existence of an R-deformation of , which is non-trivial, since the associated
projective presentation is non-trivial. 
Recall that for a profinite ring A, an abstract representation : 1top (X) GL(r, A)
factors through a continuous profinite representation : 1 (X) GL(r, A), similarly for
a P GL(r, A) representation. In the next lemma we emphasise that we are using a slightly
non-standard definition of rigid representations (see Definition 1.2).
Lemma 5.5. Let : 1top (X) GL(r, Fq ) be an absolutely irreducible representation. Let
[] MB (X/C, det())(r)(Fq ) be its moduli point. Then [] lies in MBrig (X/C, det())(r)(Fq )
if and only if the continuous representation is rigid. The analogous assertion holds for
projective representations.
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 19

Proof. We deduce this from the fact that for a finite-dimensional Fq -algebra A, one has
a canonical bijection between continuous morphisms 1 (X) GL(r, A) and morphisms
1top (X) GL(r, A). This shows that Def (Fq ) is represented by the formal scheme which is
the formal completion of MB (X/C, det())(r)Z Fp at the point []. The latter is equivalent
to the spectrum of an artinian Fq -algebra if and only if [] is an isolated point. That is, if
and only if is rigid. Definition 5.1 of rigidity for continuous representations of profinite
groups allows us to conclude the proof. 
Let Kpmon be a number field, such that the topological monodromy of every rank r
irreducible rigid projective representation has values in P GL(r, Kpmon ), and Kmon be a
number field, such that the topological monodromy of every rank r irreducible rigid repre-
sentation has values in GL(r, Kmon ). The index p stands for projective. Since there are only
finitely many irreducible rigid representations, and 1top (X) is finitely presented, there exists
M OKpmon , such that 1
Q every such representation is defined over P GL(r, OKpmon [M ]).
We write Opmon,M = OKpmon , , where ranges over the places of Kpmon , such that
(M ) = 1. As a topological group, Opmon,M is profinite.
Proposition 5.6 (Simpson, [Sim92], Theorem 4). Let : 1top (X) P GL(r, OKpmon [M 1 ])
be an absolutely irreducible rigid P GL(r)-representation. Then there is a finite Galois ex-
tension L/Q(S), such that Opmon,M extends to a projective representation

1top (X) / 1 (XL )


( 
P GL(r, Opmon,M ).

Proof. Let : 1 (X) P GL(r, Opmon,M ) be the profinite representation associated to .


We apply Simpsons theorem loc. cit., with the slight difference that we use here directly
a projective representation in the assumption. For the readers convenience, we sketch
Simpsons argument in this context. We choose finitely many generators A1 , . . . , AM of
1top (X) and use them to embed
R(1top (X))(Opmon,M ) , P GL(r, Opmon,M ))M ,
where R(1top (X))(Opmon,M ) is the set of Opmon,M -points of the scheme of representations
defined by the image of the Ai fulfilling the relations between the Ai . Thus R(1top (X))
is endowed with the profinite topology. To = Gal(Q(S)/Q(S)) one assigns the
representation : 1 (X) P GL(r, Opmon,M ), c 7 (c 1 ). Thus there is an open
U
subgroup U such that for U , the representation is equal to . We set L = Q(S) .
This yields the factorization
1 (X) / 1 (XL )

( 
P GL(r, Opmon,M ).
Continuity is checked as in loc. cit.. 
20 HELENE ESNAULT AND MICHAEL GROECHENIG

In the following we denote by q a power of a prime p. We use the shorthand Zq for


W (Fq ).
Lemma 5.7. There exist infinitely many prime numbers p, such that
(a) every rigid and absolutely irreducible Zq -representation of 1top (X) of rank r is de-
fined over Zp , and similarly for rigid absolutely irreducible projective representations
(in particular every place over p in Kmon and Kpmon splits completely);
(b) every rigid and absolutely irreducible representation : 1top (X) GL(r, Fp ) is ob-
tained as reduction modulo p of a representation 1top (X) GL(r, OKmon [M 1 ]),
and similarly for projective representations;
(c) there exists a closed point s S with k(s) = Fp , which is the specialisation of a
morphism Q(S) Qp ;
(d) for every s as in (b), a rigid and absolutely irreducible projective representation of
1 (XQp ) P GLn (Zp ) descends to 1 (XQp ).
We call such a p a good prime.
Proof. We write S = Spec R. We denote by R1 = Z[1 , . . . , m ] a finitely generated
subalgebra of C containing R, OKmon and OKpmon , as well as the normalization of S
inside all of the (finitely many) field extensions L/Q(S) constructed in Proposition 5.6. By
Casselss embedding theorem [C76, Thm. I], there are infinitely many prime numbers p,
such that the fraction field Q(1 , . . . , m ) can be embedded in Qp , and the generators i
are sent to p-adic units. For such a p, the induced morphism OKmon Zp is injective. This
shows (a).
Claim (b) is automatic by choosing very large prime numbers: since the variety of
top
1 (X)-representations is of finite type over Z, the subscheme of rigid representations
is finite over a dense open of Spec(Z), thus there can only be finitely many primes where
isolated points exist that do not have a Kmon -model.
Moreover we have a non-trivial morphism R Zp , hence the composition R Fp
defines the required Fp -rational point s in (b).
Claim (c) follows from (b) and Proposition 5.6. At first we choose an abstract isomor-
phism of fields C ' Qp , and view a rigid and absolutely irreducible projective representation
of 1 (XQp ) as one of 1 (X).
We know that is obtained from a rigid absolutely irreducible representation defined
over Opmon [M 1 ]. By Proposition 5.6 the associated projective Opmon,M -representation
descends to XL . By tensoring along Opmon,M Zp (using that p splits completely in
Kpmon by (a)), we see that the projective representation itself descends to XL . However,
by construction our p also splits in L and therefore we see that descends to XQp . 

In the proof of the following corollary we will use a variant of Higgs bundles. A P GLr -
Higgs bundle on X consists of a P GLr -torsor P and a section H 0 (X, 1X End0 (P )),
satisfying the integrality condition = 0.
Corollary 5.8. Let p be a good prime. Then there exists a finite field extension F/Qp , such
that every irreducible rigid P GL(r, Zp )-representation of 1 (XQp ) descends to XF . Every
irreducible rigid P GL(r, Qp )-representation of 1 (XQp ) can be obtained in this way.
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 21

Proof. The first assertion is elementary. There are only finitely many rigid representations,
each of which descends to a finite field extension. It is sufficient to take F to be the
compositum of these extensions.
The second assertion follows from Faltingss p-adic Hodge theory. The projective Qp -
representation associated to a projective Zp -representation is small ([Fal05, Defn. 2]) and
hence Faltingss correspondence [Fal05, Thm. 5] can be applied. We remark that Falt-
ingss theorem also holds for projective representations, and P GL-Higgs bundles. One
can deduce this from the tensor functor defined in [LZ17, Thm. 2.1(iv)], but it can also
be shown directly. Indeed, Faltings establishes this equivalence of categories locally (see
[Fal05, Thm. 3]) and then obtains the global statement. The sheaf of categories of P GL-
Higgs bundles can be obtained via descent from the category of Higgs bundles: one locally
quotients by the action of rank 1 Higgs bundles via tensor products. The same procedure
allows one to produce the sheaf of categories of projective lisse Qp -sheaves on XQp . Since
Faltingss correspondence is compatible with twisting by small rank 1 Higgs bundles and
small Qp -representations, we obtain a correspondence between small P GL-Higgs bundles
on XQp and projective lisse Qp -sheaves on XQp .
It sends a rigid irreducible projective representation to a rigid stable P GL(r)-Higgs
bundle on XQp . But there are exactly as many rigid stable P GL(r)-Higgs bundles as there
are rigid irreducible projective representations of the topological fundamental group. 

The LanShengZuo correspondence only relates periodic Higgs bundles with crystalline
representations. A priori we therefore do not know that the representation thereby assigned
to a rigid Higgs bundle is again rigid. This problem is solved in the sequel by relating the
LanShengZuo correspondence to Faltingss correspondence.
Let (Ei , i ) be a rank r stable rigid Higgs bundle on Xs , where s and p are as in Corol-
lary 5.8. We recall the notation i,s : 1 (XQp ) P GL(r, W (Fpfi,s )) of Remark 4.8. We let
m (Ei , i ) denote the periodic Higgs bundle corresponding to m (i,s ) by the LanSheng
Zuo correspondence. Here is the shift operator defined in (4.4) on Higgs bundles, which
corresponds to the Frobenius action on the coefficients W (Fpfi,s ) for i,s .

Lemma 5.9. There exists an mi,s , such that under Faltingss p-adic Simpson correspon-
dence the representation i,s is isomorphic mi,s (Ei , i ) as representations defined over
Frac(W (Fpfi,s )).

Proof. We fix one i,s which we for short denote by . We define the Zp -representation
fi,s 1
M
big = i
i=0

of 1 (XQp ). It corresponds to a Higgs bundle (Ebig , big ) of period 1 by means of Lan


ShengZuos correspondence.
We compare this correspondence with Faltingss one. First we remark that Faltingss
correspondence relates small generalized Qp -representations of 1 (XF ) and small Higgs
bundles on XF . The class of representations is preserved by deformations inside generalized
representations, which allows us to test rigidity. Furthermore, every Higgs bundle with
22 HELENE ESNAULT AND MICHAEL GROECHENIG

nilpotent Higgs field is small (smallness is defined via the characteristic polynomial in the
Qp -theory), and every Zp -representation induces a small Qp -representation.
By definition, big corresponds to a Frobenius crystal, and therefore, by [Fal05, Section 5,
Ex.], Faltingss correspondence associates to big Q the Higgs bundle (Ebig , big ). This
example is thus compatible with the LanShengZuo correspondence.
Let Fm : 1 (XQp ) GL(r, Qp ) denote the representation associated by Faltings to the
stable nilpotent Higgs bundle m (E, ). It follows from Proposition 4.3 that m (E, ) is
again rigid. By construction, Fm is absolutely irreducible and rigid, since it is associated
to a stable rigid Higgs bundle under Faltingss correspondence. Furthermore we have an
isomorphism
fi,s 1 fi,s 1
M M
i
big Z Q = Z Q ' Fm .
i=0 m=0
This implies that each factor on the left hand side is isomorphic to a unique factor on the
right hand side. In particular, we obtain the requested isomorphism i,s Qq ' Fmi,s . 
Corollary 5.10. For a good prime p, the representation i,s is isomorphic to a rigid rep-
resentation
1 (XQp ) GL(r, Qp ).
Proof of Theorem 5.3. Let N be the number of stable rigid Higgs bundles on X as in (4.1).
Let p be a good prime. We know from Faltingss p-adic Simpson correspondence that there
are at most N rigid absolutely irreducible representations 1 (XQp ) GL(r, Qp ) which are
small for all good primes p. And we have seen in Lemma 5.9 that they are induced by
LanShengZuos representations i,s .
Then 1 (X) = 1 (XQp ) and Corollary 5.8 yields that all rigid and irreducible P GL(r, W (Fq ))-
representations of the topological fundamental group descend to XQp as projective represen-
tations. This shows that the projectivization proj
i,s is equivalent to one of the rigid projective
top
representations of 1 (X) P GL(r, Zp ) descended to XQp . Here we use that two abso-
lutely irreducible projective representations defined over Zq are equivalent if and only if
they are equivalent over Qq . This shows that proj
i,s |1top is rigid and absolutely irreducible.
This implies that the representation |1 (XQ ) is rigid as a projective representation (ar-
p
guing as in Lemma 5.7). 
We recall that denotes LanShengZuos shift functor for periodic Higgs bundles. On
the level of crystalline representations it corresponds to the Frobenius-twist of a represen-
tation. The following assertion falls short of showing fi,s = 1 for good primes p. Instead
we show that this assertion holds on the level of P GLr -Higgs bundles.
Corollary 5.11. For every good prime p and every closed point s S with char(k(s)) = p
we have for every rigid stable Higgs bundle (E, ) defined over Xs that (E, ) and (E, )
are equivalent as P GLr -Higgs bundles.
Proof. Recall that Zpfi,s is the ring of definition of i,s . We have seen in the proof of
Theorem 5.3 that proj
i,s is equivalent to a projective representation defined over Zp . This
implies the assertion. 
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 23

6. Rigid connections with vanishing p-curvature have unitary monodromy


The aim of this section is to prove Theorem 1.8.
Let (E, ) be a rigid connection with vanishing p-curvature. We would like to show that
for (gr(E), KS) one has KS = 0 (see Lemma 4.1). By the definition (4.7) of the Higgs-
de Rham flow, the downwards arrows come from the Simpson correspondence, while the
upwards arrows come from taking the canonical connections with p-curvature 0. Therefore
we have (gr(E), KS) = (C 1 (E, )) = (V, 0). By Corollary 5.11, we know that (V, 0)
and (gr(E), KS) are equivalent as projective Higgs bundles. This shows that KS =
id gr(E) , where is a 1-form on Xs . However, we also know that KS is nilpotent, hence we
must have KS = 0.

7. Integrality
The aim of this section is to prove Theorem 1.4.
We use the standard notations. We choose a good prime p, with underlying closed point
s. Given a rigid connection (E, ), we constructed the isocrystal with Frobenius structure
on XKv in Theorem 1.6. For notational convenience, in this section, we denote it by F.
For the readers convenience, we summarise the defining properties of `-adic companions.
Let E be an irreducible isocrystal with Frobenius structure on a smooth variety Y of finite
type over Fq of characteristic p > 0. To every closed point y Y , one attaches the
characteristic polynomial PE,y (t) = det(1 tF ry |Ey ) Qp [t], where F ry is the absolute
Frobenius at y acting on Ey := iy E, and where iy : y Y is the closed embedding. Similarly
for a lisse Q` -sheaf V on Y , for every closed point y Y , one attaches the characteristic
polynomial PV,y [t] = det(1 tFy |Vy ) Q` [t], where Fy is the geometric Frobenius Fy acting
on Vy , where y y an Fp -point above y.
Definition 7.1 (See [AE16], Defn. 1.5, [Dri16], Section 7.4.). Let : Qp ' Q` be an abstract
isomorphism of fields. We say that an irreducible lisse Q` -sheaf V is a Q` -companion of
an irreducible isocrystal E with Frobenius structure and determinant of finite order, if for
every closed point y Y , one has an equality of characteristic polynomials
(PE,y (t)) = PV,y [t] Q` [t].
We shall use in the sequel that Q` -companions exist by [AE16, Thm. 4.2] (see [Dri16,
Thm. 7.4.1] for a summary of the existence of companions, the `-part being Drinfelds key
theorem, and see also [Ked17, Thm. 5.3] for later work in progress).
By Cebotarev s density theorem, an Q` -companion is irreducible, has finite order de-
terminant, is unique up to isomorphism, which is thus a scalar. Furthermore, by Abes
Cebotarevs theorem [Abe13, Prop. A.3.1], if two irreducible isocrystals with Frobenius
structure are not isomorphic, their Q` -companions are not isomorphic either.
So we fix one isomorphism : Qp = Q` and denote by V the Q` -companion of our
F -isocrystal (E, ) above. It corresponds to an irreducible continuous representation
Vs : 1 (Xs ) GL(r, Q` ) with finite determinant, thus precomposing by the (surjective)
specialization homomorphism sp : 1 (XKs ) GL(r, Q` ), it defines an irreducible `-adic
24 HELENE ESNAULT AND MICHAEL GROECHENIG

lisse sheaf VKs : 1 (XKs ) GL(r, Q` ) on XKs , and the underlying geometric representation
VKs : 1 (XKs ) GL(r, Q` ) on XKs .
Recall from the introduction that a connection (E, ) on X over C is called cohomolog-
ically rigid if H 1 (X, End0 (E, )) = 0, that is if it is rigid and in addition its moduli point
[(E, )] MdR (X/C)(r) is smooth.
Proposition 7.2. 1) The representation VKs is irreducible.
2) If in addition (E, ) is cohomologically rigid, then there are infinitely many good
primes p such that
H 1 (XKs , End0 (VKs )) = 0.
Proof. We prove 1). Since sp restricts to the specialization 1 (XKs ) 1 (Xs ), where
s s is a geometric point with residue field Fp , and is still surjective, we just have to show
that Vs is geometrically irreducible. On the other hand, Theorem 5.3 together with the
construction of Theorem 1.7 shows that the isocrystal is compatible with finite base change
s0 s. Since an irreducible `-adic sheaf which is not geometrically irreducible splits over
a finite base change s0 s (see e.g. [Del12, (1.3.1)], this shows that Vs is geometrically
irreducible.
We prove 2). We consider the L- functions L(Xs , End0 (Vs )) and L(Xs , End0 (F)) for the
lisse Q` -sheaf End0 (Vs ) and for the isocrystal with Frobenius structure End0 (F) ([Del73,
5.2.3] and [Abe13, 4.3.2]). From the existence of the product formula for both of them
we know that they are equal. On the other hand, as Vs and F, thus a fortiori End0 (Vs )
and End0 (F) have weight 0 (see [Laf02, Prop.VII.7 (i)], corrected in [EK12, Cor. 4.5]), the
dimension of H 1 (Xs , End0 (Vs )) over Q` and of Hcrys
1 (X , End0 (F)) over Q are computed
s p
as the number of weight 1 eigenvalues counted with multiplicities in the L-function. This
shows that both are the same. On the other hand,
1
Hcrys (Xs , End0 (F)) = HdR
1
(XKs , End0 (E, )) Ks Qp .
By [Kat70, Thm. 8.0], there is a dense open S 0 , S on which one has base change for
de Rham cohomology. By Lemma 5.7, S 0 contains infinitely many good primes. Thus we
conclude that on those good primes
1
0 = dimKs HdR (XKs , End0 (E, )) = dimC HdR
1
(X, End0 (E, )).
Thus dimQ` H 1 (Xs , End0 (Vs )) = 0.
It remains to see that the specialization homomorphism
H 1 (Xs , End0 (Vs )) H 1 (XKs , End0 (VKs ))
is an isomorphism, which is true by local acyclicity and proper base change [Art73, Cor. 1.2].

Corollary 7.3. The composite representation
VK
: 1top (X) 1 (X) = 1 (XKs )
s
GL(r, Q` )
defines a Q` -point of MBrig (X/C)(r) which is cohomologically rigid, and any Q` -point of
MBrig (X/C)(r) which is cohomologically rigid arises in this way.
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 25

Proof. Let A Z` be a subring of finite type such that factors though A : 1top (X)
GL(r, A), and : A , C be a complex embedding. Set C = A . Then
1
Han (X, End0 (C )) = Han
1
(X, End0 (A )) A C.
Here an stands for the analytic topology. On the other hand, by comparison between
analytic and etale cohomology, one has
1
Han (X, End0 (A )) A Q` = Het
1
(X, End0 ()) = H 1 (XKs , End0 (VKs )) = 0.
This proves the first part. It also shows that once the good prime is fixed, the number of
the Vs is at most the number of cohomologically rigid connections.
rig
Vice versa, given all cohomologically rigid points of MdR (X/C)(r), their restrictions to
XKs for some Spec(Ks ) S define isocrystals with Frobenius structure (Theorem 1.7),
which are cohomologically rigid for an infinity of good primes (proof of Proposition 7.2),
and pairwise different. Thus the `-companions Vs are pairwise different as well. This
implies that the number of Vs is precisely the number of cohomologically rigid connections
and finishes the proof. 

Proof of Theorem 1.4. Let K : 1top (X) GL(r, K) be a rigid representation, where K is
a number field. Then it has values in GL(r, OK,T ) where T is a finite set of finite places of
K and OK,T is the subring of K consisting of all elements which localize OKv for all places
v / T . If T is not empty, we take ` to be a prime dividing one of the places t of T , then
we take a good prime p 6= `. The representation Vs has values in GL(r, Z` ), where Z` Q`
is the subring of algebraic integers, so the same holds true for VKs and in Corollary 7.3.
This shows that K has values in GL(r, Z), where Z Q is the subring of algebraic integers.
This finishes the proof. 
Remark 7.4. We use cohomological rigidity as opposed to rigidity in the proof of Theo-
rem 1.4 in order to prove that VKs in Corollary 7.3 yields a cohomologically rigid . If one
could prove that is rigid just under the rigidity assumption, the integrality proof would
go through, as the construction of the isocrystal in Theorem 1.7 is done in all generality.
It is possible that such a statement could be deduced from a more general assertion about
companions. One would like to compare for V a Q` -companion of an isocrystal E with
Frobenius structure the differential graded Lie-algebras governing the deformation theories
of the respective cohomologies Rcrys (Xs , End0 (E)) and R(XKs , End0 (V)) (with respect
to the isomorphism : Qp ' Q` ). Since these differential graded Lie algebras govern the
formal deformation theories (see [Lur11]), such a comparison would imply that E is rigid if
and only if V is.

8. Some remarks
8.1. SL(3)-rigid connections. Theorem 1.4 implies that the main theorem of Langer-
Simpson in [LS16] is true unconditionally, in other words cohomologically rigid SL(3)-
connections verify Conjecture 1.1. As already mentioned in the introduction, their theorem
is more precise as it explicitly describes the Gau-Manin connections that contain the rigid
SL(3)-connections as subquotients.
26 HELENE ESNAULT AND MICHAEL GROECHENIG

8.2. Vanishing of global symmetric forms. A smooth projective variety X without


global non-trivial i-th symmetric differential for all i, has the property that all integrable
connections are rigid and have finite monodromy, see [BKT13, Thm. 0.1]. The proof uses p-
analytic methods to show integrality, in addition to positivity theorems ultimately stemming
from complex Hodge theory, together with L2 -methods. It would be nice to understand at
least part of the theorem by mod p and p-adic methods as presented in this article.
8.3. Geometric origin for ball quotients. This application has been pointed out to us
by C. Simpson. Theorem 1.4 implies that [LS16, Cor. 6.11] holds unconditionally. We recall
their set-up for the convenience of the reader. Let P U (2, 1) be a lattice which we further
assume to be torsion-free and cocompact. The corresponding ball quotient X = B2 / is
known to be a smooth projective variety. We denote by the standard representation
of = 1top (X), which is known by Weils theorem [Wei62] to be cohomologically-rigid
(see [KKP12, Ex. 1.2.4]). Conditional to integrality it was shown in loc. cit. that is of
geometric origin. Integrality is a consequence of Theorem 1.4 and therefore we conclude
that the integrality assumption of loc. cit. can be dropped.
Corollary 8.1. The standard representation is of geometric origin.
8.4. Motivicity of the isocrystal from good curves to the whole variety. The
existence of Frobenius structure implies that the isocrystal defined by an irreducible rigid
(E, ) on XKv is motivic on curves Cs , Xs where C , X is a dimension 1 smooth
complete intersection of ample divisors. See Corollary 4.7. This raises the problem of
extending the motivicity from Cs to Xs . Said concretely: can one find a morphism fs :
Ys Xs which has the property that any irreducible isocrystal with Frobenius structure
and finite determinant on Xs is a subquotient of the Gau-Manin isocrystal Ri fs OXs /Ks
for some i 0, where Ks = Frac(W (k(s))? Beyond the study of rigid connections, it would
perhaps enable one to think of the construction of `-to p-companions (see [AE16, Thm. 4.2]
for p-to-`-companions and the references there.)
8.5. Some remarks on rigid connections and the p-curvature conjecture. We prove
two little propositions which are related to the p-curvature conjecture for rigid connections.
Proposition 8.2. Let X be a smooth projective complex variety such that all rigid uni-
tary connections have finite monodromy. Then rigid connections (E, ) with a model
(XS , (ES , S )) over a scheme S of finite type over Z with p-curvature 0 at all closed points
s S have finite monodromy.
Proof. We just apply Theorem 1.8.

Proposition 8.3. Let X/C be a smooth projective variety. Then if any cohomologically
rigid connection (E, ) on X has a model (XS , (ES , S )) over a scheme S of finite type
over Z with p-curvature 0 at all closed points s S, the monodromy of all cohomologically
rigid connections (E, ) is finite.
Proof. Let : 1top (X) GL(r, C) be a cohomologically rigid representation. By Theo-
rem 1.4 the monodromy lies in GL(r, OL ) for some number field L C. If : L C is
another complex embedding, then the groups Han1 (X, End0 ()) and H 1 (X, End0 ( )) are
an
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 27

equal, thus are vanishing. Thus is cohomologically rigid. This implies that
where runs through the Galois group G of L has mondromy in GL(r|G|, Z) and is unitary
by Theorems 1.4 and 1.8. Thus, applying [Kat72, Prop. 4.2.1.3], one concludes that the
mondromy of is finite, in particular the one of is finite.

Remark 8.4. In [And04, Thm.16.2.1], Andre proved that an irreducible subquotient of a
Gau-Manin connection f : Y X verifies the p-curvature conjecture if f has one complex
fibre with connected motivic Galois group. On the other hand, [LS16, Thm.4.1] together
with Remark 8.1 imply that cohomologically rigid SL(3)-connections all are subquotients
of Gau-Manin connections coming from families of product abelian varieties. Those have
a connected motivic Galois group. We conclude that the p-curvature conjecture is true for
the cohomologically rigid SL(3)-connections.

References
[AC13] Abe, T., Caro, D.: Theory of weights in p-adic cohomology, arXiv:1303.0662 (2013).
[Abe13] Abe, T.: Langlands correspondence for isocrystals and existence of crystalline companions for
curves, arXiv:1303.0662 (2013).
[AE16] Abe, T., Esnault, H.: A Lefschetz theorem for overconvergent isocrystals with Frobenius structure,
http://www.mi.fu-berlin.de/users/esnault/preprints/helene/123_abe_esn.pdf (2016).
[ACV03] Abramovich, D., Corti, A., Vistoli, A.: Twisted bundles and admissible covers, Comm. in Algebra
31 (2003) 8, 35473618.
[And04] Andre, Y.: Sur la conjecture des p-courbures de Grothendieck-Katz et un probleme de Dwork, in
Geometric aspects of Dwork theory, Vol. I II 55-112, Walter de Gruyter (2004).
[Art73] Artin, M.: Theoreme de changement de base par un morphisme lisse, et applications, in SGA4,
3, Exp. XVI, Lect. Notes in Math. 305 (1973), 206250.
[BB07] Bezrukavnikov, R., Braverman, A.: Geometric Langlands Correspondence for D-modules in Prime
Characteristic: the GL(n) Case, Pure and Applied Mathematics Quaterly 3 (2007) 1, 153179.
[BE04] Bloch, S., Esnault, H.: Local Fourier transforms and rigidity for D-modules, Asian Journal
of Mathematics, special issue dedicated to Armand Borel 8 (4) (2004), 587606. Erratum to
Thm. 4.10: http://www.mi.fu-berlin.de/users/esnault/preprints/helene/71_cor.pdf
[Bor94] Borceux, F.: Handbook of Categorical Algebra: Volume 3, Sheaf Theory. Vol. 3 (1994), Cambridge
University Press.
[BKT13] Brunebarbe, Y., Klingler, B., Totaro, B.: Symmetric differentials and the fundamental group,
Duke Math. J. 162 (2013) no 14, 27972813.
[C76] Cassels, J. W. S: An embedding theorem for fields, Bulletin of the Australian Mathematical Society
14.02 (1976): 193-198.
[CZ15] Chen, T.-H., Zhu, X.: Non-abelian Hodge theory for algebraic curves in characteristic p, Geom.
Funct. Anal. 25 (2015) 6, 17061733.
[CS08] Corlette, K., Simpson, C.: On the classification of rank-two representations of quasi-projective
fundamental groups, Compos. Math. 114 (2008) 5, 12711331.
[Cre17] Crew, R.: Rigidity and Frobenius Structure, Documenta math. 22 (2017), 287296.
[Del73] Deligne, P.: Theoreme dintegralite, Appendix to Le niveau de la cohomologie des intersections
completes by N. Katz, Expose XXI in SGA 7, Lecture Notes in Math. 340, 363400, Springer-
Verlag, New York, 1973.
[Del12] Deligne, P.: Finitude de lextension de Q engendree par les traces de Frobenius, en caracteristique
finie, Mosc. Math. J. 12 (2012), no. 3, 497514, 668.
[Del15] Deligne, P.: Comptage de faisceaux `-adiques, in De la geometrie aux formes automorphes (I), en
lhonneur du soixantieme anniversaire de Gerard Laumon, Asterisque 369 (2015), 282312.
28 HELENE ESNAULT AND MICHAEL GROECHENIG

[Dri16] Drinfeld, V.: On the pro-semisimple completion of the fundamental group of a smooth variety over
a finite field, https://arxiv.org/pdf/1509.06059.pdf (2016).
[EK12] Esnault, H., Kerz, M.: A finiteness theorem for Galois representations of function fields over finite
fields (after Deligne), Acta Mathematica Vietnamica 37 4 (2012), 531562.
[EK16] Esnault, H., Kisin, M.: D-modules and finite monodromy, Selecta Mathematica, to appear in the
volume dedicated to Alexander Beilinson, preprint 2016, 9 pages, http://www.mi.fu-berlin.de/
users/esnault/preprints/helene/124_esn-kis.pdf
[Fal05] Faltings, G.: A p-adic Simpson correspondence, Advances in Math. 198 (2005), 847-862.
[Gro16] Groechenig, M.: Moduli of flat connections in positive characteristic, Math. Res. Let. 23 (2016)
4, 9891047.
[Gro63] Grothendieck, A.: Elements de geometrie algebrique. III. Etude cohomologique des faisceaux
coherents. II, Publ. Math. de lInst. des Hautes Etudes Sci. 17 (1963) 91.
[HL97] Huybrechts, D., Lehn, M.: The Geometry of Moduli Spaces of Sheaves, Aspects of Mathematics,
E31 (1997), Vieweg Verlag.
[Kat70] Katz, N.: Nilpotent connections and the monodromy theorem: applications of the result of Turrit-
tin, Publ. math. Inst. Hautes Etudes Sci. 39 (1970), 175232.
[Kat72] Katz, N.: Algebraic Solutions of Differential Equations (p-curvature and the Hodge Filtration),
Invent. math. 18 (1972), 1118.
[Kat96] Katz, N.: Rigid local systems, Princeton University Press (1996).
[Ked17] Kedlaya, K.: Etale and crystalline companions, http://kskedlaya.org/papers/companions.pdf
(2017), in progress.
[KKP12] Kim, I., Klingler, B., Pansu, O.: Local quaternionic rigidity for complex hyperbolic lattices, J. de
lInst. de Math. de Jussieu 11 (2012), 133159.
[Kli13] Klingler, B.: Symmetric differentials, Kahler groups and ball quotients, Invent. math. 192 (2013),
257286.
[Laf02] Lafforgue, L.: Chtoucas de Drinfeld et correspondance de Langlands, Invent. math. 346 (2010) 3,
641668.
[Lam99] Lam, T. Y.: Lectures on modules and rings, Graduate Texts in Mathematics, vol.189, Springer-
Verlag, New York (1999).
[LSZ13] Lan, G., Sheng, M., Zuo, K.: Semistable Higgs bundles, periodic Higgs bundles and representations
of algebraic fundamental groups, to appear in J. Eur. Math. Soc., https://arxiv.org/pdf/1311.
6424.pdf (2013).
[LSZ15] Lan, G., Sheng, M., Zuo, K.: Non-abelian Hodge theory in positive characteristic via exponential
twisting, Math. Res. Let. 22 (2015) 3, 859879.
[Lan04] Langer, A. Semistable sheaves in positive characteristic, Annals of math. (2004), 251-276.
[Lan14] Langer, A.: Semistable modules over Lie algebroids in positive characteristic, Doc. Math. 19
(2014), 509540.
[LS16] Langer, A., Simpson, C.: Rank 3 rigid representations of projective fundamental groups, https:
//arxiv.org/pdf/1604.03252.pdf (2016).
[LP01] Laszlo, Y., Pauly, C.: On the Hitchin morphism in positive characteristic, Int. Math. Res. Notices
(2001) 3, 129143.
[Lef50] Lefschetz, S.: LAnalysis situs et la geometrie algebrique, Gauthier-Villars,Paris (1950).
[LZ17] Liu, R., Zhu, X: Rigidity and a Riemann-Hilbert correspondence for p-adic local systems, Invent.
math., 207 (1) (2017), 291343.
[Lur11] Lurie, J: Derived algebraic geometry X: Formal moduli problems, math.harvard.edu/?lurie/
papers/DAG-X.pdf (2011).
[Ogu77] Ogus, A.: F-crystals and Griffiths transversality, In Proceedings of the International Symposium
on Algebraic Geometry (1977), Kyoto, 1544.
[OV07] Ogus, A., Vologodksy, V.: Nonabelian Hodge theory in characteristic p, Publ. math. Inst. Hautes
Etudes Sci. 106 (2007), 1138.
[Sim92] Simpson, C.: Higgs bundles and local systems, Publ. math. Inst. Hautes Etudes Sci. 75 (1992),
595.
RIGID CONNECTIONS, F -ISOCRYSTALS AND INTEGRALITY 29

[Sim94] Simpson, C.: Moduli of representations of the fundamental group of a smooth projective variety,
Publ. math. Inst. Hautes Etudes Sci. 79 (1994), 47129.
[Sim97] Simpson, C.: The Hodge filtration on nonabelian cohomology, Algebraic Geometry-Santa Cruz
1995, Proc. Symp. Pure Math. 62 (1997) 2, 217281.
[Wei62] Weil, A.: Discrete subgroups of Lie groups, II, Annals of maths. 75 (1962), 97123.

Freie Universitat Berlin, Arnimallee 3, 14195, Berlin, Germany


E-mail address: esnault@math.fu-berlin.de

Freie Universitat Berlin, Arnimallee 3, 14195, Berlin, Germany


E-mail address: groemich@zedat.fu-berlin.de